Professional Documents
Culture Documents
X, MONTH, YEAR 1
Abstract— User cooperation is a powerful tool to combat fading argue here that, when we are interested in the high signal-
and increase robustness for communication over wireless chan- to-noise ratio (SNR) behavior of the average distortion, in
nels. Although it is doubtless a promising technique for enhancing most cases it suffices to consider separate source and channel
channel reliability, its performance in terms of average source
distortion is not clear since source-channel separation theorem encoders and decoders as well as cooperative strategies whose
fails under the most common non-ergodic slow fading channel parameters are optimized jointly. We specifically consider
assumption, when channel state information is only available at distortion exponent (∆) [2] as our performance measure,
the receiving terminals. This work sheds some light on the end-to- which is defined as the exponential decay rate of the end-
end performance of joint source-channel coding for cooperative to-end expected distortion in the high SNR limit.
relay systems in the high SNR regime. Considering distortion
exponent as a figure of merit, we propose various strategies Bandwidth ratio, which is the ratio between the channel
for cooperative source and channel coding that significantly and source bandwidths, or the number of channel uses per
improve the performance compared to the conventional scheme source sample, plays a crucial role in the achievable system
of source coding followed by cooperative channel coding. We performance. In [1], [12]–[14], we consider a MIMO block
characterize the optimal distortion exponent of a full-duplex fading channel, and analyze its distortion exponent as a
relay channel for all bandwidth ratios. For the half-duplex relay
channel, we provide an upper bound which is tight for small and function of the bandwidth ratio. We observe that the diversity-
large bandwidth ratios. We consider the effect of correlated side multiplexing tradeoff (DMT) is a useful tool in characterizing
information on the distortion exponent as well. the distortion exponent, and propose various strategies that
Index Terms— Block fading channel, cooperative transmission, utilize the DMT curve of the MIMO channel to improve the
distortion exponent, diversity-multiplexing tradeoff, relay chan- achievable distortion exponent. The three schemes proposed
nel. in [1] are layered source coding with progressive transmis-
sion (LS), its hybrid digital-analog extension called hybrid
I. I NTRODUCTION LS (HLS), and broadcast strategy with layered source (BS).
W IRELESS communication technology has entered a These schemes adapt to channel variations by utilizing layered
new era where it evolved from a system offering compression followed by variable rate channel coding that
mainly voice service to one that provides services with rich provides unequal error protection. Furthermore, [1] establishes
multimedia content. The increased demand for different ser- a close relationship between the maximal diversity of the
vices at the application layer results in higher transmission rate system and its highest achievable distortion exponent. For a
and reliability requirements at the physical layer. However, MIMO system with Mt transmit and Mr receive antennas and
high data rate and high reliability are two conflicting design L fading blocks, where the highest possible diversity gain is
parameters. Accordingly, we need an end-to-end performance LMt Mr , the distortion exponent is also bounded by LMt Mr
measure for the overall system. and this bound is achievable provided that the bandwidth ratio
In this paper, we are interested in transmission of a con- is high enough.
tinuous amplitude (analog) source over a quasi-static fading Our goal in this paper is to characterize the distortion expo-
relay channel within the delay requirement of the underlying nent of cooperative relay systems for all bandwidth ratios. User
application. The performance measure we consider is the cooperation [3], [4] is a popular spatial diversity technique
end-to-end average distortion between the source and its which utilizes relays to increase diversity. Simple yet effective
reconstruction at the destination. As we argued in detail in cooperation schemes such as amplify-and-forward (AF) and
[1], Shannon’s source-channel separation theorem does not decode-and-forward (DF) [5] can result in diversity gains
hold when the channel state information (CSI) is not available similar to multiple antenna transmission. On the other hand,
at the transmitters since the channel is no more ergodic. the full DMT of cooperative transmission might not follow
Therefore the system we consider requires a joint source- the MIMO tradeoff or the best known achievable strategies fall
channel code design for optimal end-to-end performance. We short of MIMO [6], [7]. Also, for a cooperative system, it is not
clear how to utilize the best DMT achieving relaying strategies
The material in this paper was presented in part at IEEE Workshop on and combine them with layered source coding in a manner
Signal Processing Advances in Wireless Communications (SPAWC), New
York, NY, June 2005. similar to [1] in order to improve the distortion exponent. In
This work is partially supported by NSF grant No. 0430885 and No. the half-duplex relay case, where the relay cannot receive and
0635177. transmit simultaneously, matching of layered source coding
The authors are with Department of Electrical and Computer
Engineering, Polytechnic University, 11201, Brooklyn, NY. Email: to different channel transmission techniques becomes even
dgundu01@utopia.poly.edu, elza@poly.edu more challenging. For example, as we show in Section III,
2 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. X, NO. X, MONTH, YEAR
d∗ (r)
known at the corresponding receivers, but not at the transmit-
ters. We also assume an independent additive noise at each
receiver modeled as circularly symmetric complex Gaussian 2
AF or DF Cooperative
with variance 1/2 in each dimension. Each transmitter has protocol
the same power constraint and hence all average received
SNRs are equal. Since our analysis will focus on high SNR
exponents, considering path loss and asymmetric systems will
not change our results. The relay, depending on the physical
system constraints, might operate on the half-duplex or the 1
full-duplex modes. A half-duplex relay cannot transmit and y = br
the transmitters, but also on the joint compression, channel Fig. 2. A geometric interpretation illustrating the optimal multiplexing gain
coding and transmission techniques used. As discussed in and distortion exponent for a single layer source-channel coding system.
Section I, Shannon’s source-channel separation theorem does
not hold in our model due to slow fading and lack of CSI
at the transmitters. The problem of finding the minimum
possible expected distortion for any given channel SNR is an For single rate source and channel coding, we assume a
open problem. However, our main focus is on the high SNR channel code of rate R bits per channel use. The corresponding
behavior of the expected distortion. To capture this behavior, source coding rate is N R/K = bR bits per source sample.
we define the distortion exponent [2] as The rate R can be optimized to obtain the single rate distortion
exponent.
log ED
∆=− lim . (3) For bandwidth ratio b, we denote the distortion exponent
SNR→∞ log SNR achieved by single rate source and channel coding with direct
We want to find the maximum achievable distortion exponent transmission (DT) by ∆DT,1 , and by simple AF and DF
of this cooperative system for a given bandwidth ratio b. cooperation with a half-duplex relay by ∆AF,1 or ∆DF,1 ,
Note here the similarity between distortion exponent and respectively.
diversity gain which is defined as the exponential decay rate Theorem 3.1: For single rate source and channel coding, we
of the outage probability Pout (SNR), as a function of the have the following distortion exponents for direct transmission,
average SNR. In general, for a family of codes with rate AF and DF relaying.
R = r log SNR, r is defined as the multiplexing gain of the
family, and
log Pout (SNR) b
d(r) = − lim (4) ∆DT,1 = , (5)
SNR→∞ log SNR b+1
2b
as the diversity advantage [31]. The diversity gain d∗ (r) is ∆DF,1 = ∆AF,1 = . (6)
b+4
defined as the supremum of the diversity advantage over Proof: We follow an approach similar to [1], [8], where
all possible code families with multiplexing gain r. The the expected distortion for single rate transmission is written
relationship between the distortion exponent ∆ and d∗ (r), also as
known as the DMT, will be clarified in the next section.
III. S INGLE R ATE S OURCE AND C HANNEL C ODING AND ED(R, SNR) = (1−Pout (R, SNR))·D(bR)+Pout (R, SNR),
D ISTORTION E XPONENT U PPER B OUND (7)
In this section, we first illustrate the distortion exponent where Pout (R, SNR) is the outage probability at channel rate
analysis for single rate direct transmission from the source R and average received signal to noise ratio SNR, while
to the destination, and compare it with simple cooperation D(R) is the distortion-rate function of the source. We assume
protocols. Our results show that, depending on the bandwidth that, in case of an outage, the receiver simply outputs the
ratio, additional diversity provided by cooperation may not average value of the source distribution, hence obtains the
necessarily lead to an improvement in the distortion exponent. highest possible distortion of 1. In order to achieve a vanishing
Hence a careful design of compression and channel coding expected distortion with increasing SNR, we need to increase
strategies, as well as cooperation protocols is essential for rate R with SNR. Scaling R faster than O(log SNR) would
improving the distortion exponent. We next provide an upper result in outage with probability 1, since the instantaneous
bound on the distortion exponent, and motivate layered source channel capacity scales as log SNR. Thus we assign R =
compression to achieve this upper bound. r log SNR, where r ∈ [0, 1] is the multiplexing gain. Then
4 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. X, NO. X, MONTH, YEAR
1.4 Rn
R1 R2 Analog transmission
1.2 t1 (N − K) t2 (N − K) tn (N − K) K
1 (b) HLS strategy with n layers for b > 1/M∗ .
0.8 R1 , SN R1
0.6
Upper bound Rn , SN Rn
0.4
Single rate, DMT upper bound
(c) BS strategy with n layers. Layers are transmitted simultaneously
0.2 Single rate, AF/DF
with total power allocated among them.
Single rate, DT
0
0 1 2 3 4 5 Fig. 4. Illustration of various transmission strategies analyzed in this paper.
Bandwidth ratio, b
Fig. 3. Distortion exponent of single rate transmission with direct transmis- system, whose distortion exponent upper bound is given in [1].
sion and simple AF/DF cooperation.
d∗ (r)
their corresponding distortion exponents as a function of the
bandwidth ratio.
2 AF or DF
Cooperative
protocol
A. Layered Source with Progressive Coding (LS)
In this strategy, we consider layered source compression
followed by progressive transmission of the compressed lay-
∆LS
ers over the channel. In layered source coding, each layer AF
1
contains the refinement information for the preceding layers. LS
∆DT
By adjusting the compression and transmission rates, our goal
is to provide more protection for the most important source
layers, thereby improving the distortion exponent. Also, we
Direct
will observe that cooperation can be used as a form of unequal transmission (DT)
error protection. r
Suppose that the source encoder has n layers and the N 1
2
1
channel uses are divided among these layers such that layer
iPis transmitted in ti N channel uses, where ti ≥ 0, and Fig. 5. A geometric interpretation of the LS strategy in the case of multiple
N source layers. Straight lines climb the DT curve, while the dashed ones climb
i=1 ti = 1 (see Fig. 4-a). We will initially consider same
the cooperation curve.
transmission scheme for all layers. Let Ri bits per channel
use be the channel transmission rate of the i-th layer. We
denote the outage probability of layer i with the underlying transmission protocol. Then in the high SNR regime, the
i
transmission scheme as Pout and the distortion achieved when expected distortion expression in (11) can be written as:
first i layers are successfully decoded as DiLS . Then using n h i
the successive refinability of the Gaussian source [29], for . X ∗ ∗
ED(R, t, SNR) = SNR−d (ri+1 ) − SNR−d (ri )
i = 1, . . . , n, we have i=0
à i ! Pi
X Pi · SNR−b k=1 tk rk
LS
Di = D b tk Rk = 2−b k=1 tk Rk , n
. X ∗ Pi
k=1 = SNR−d (ri+1 ) SNR−b k=1 tk rk (13)
i=0
with D0LS = 1. Note that, due to successive refinement source
coding, layer i+1, . . . , n are useless when any of the preceding Pi
.
=SNRmax0≤i≤n (−d (ri+1 )−b k=1 tk rk ),
∗
N channel uses
S tx. layer 1 S tx. layer n
(a)
S tx. layer 1 R tx. layer 1 S tx. layer k R tx. layer k S tx. layer (k+1) S tx. layer n
(c)
Fig. 6. Illustration of LS strategy with (a) only direct transmission, (b) only simple cooperative relaying, and (c) with both cooperative relaying and direct
transmission.
corresponding layers. This interpretation allows us to observe Gaussian source as well as real video using RCPC channel
that, increasing the number of layers improves the distortion coding.
exponent even though the total slope of the lines is equal to Visualizing this approach on Fig. 5, we observe that, we can
b. achieve a higher distortion exponent, i.e., climb to a higher
In [1], we proved that, in the limit of infinite layers, for point, using direct transmission DMT for the last few layers.
a transmission scheme with piecewise linear DMT curve, Once we pass the intersection point of DT and AF/DF DMT
allocating the channel equally among the layers is optimal curves, using the cooperative tradeoff is more advantageous.
in terms of the distortion exponent, that is ti = 1/n, for This is equivalent to operating on the tradeoff curve that is
i = 1, . . . , n. This enabled us to obtain an explicit result obtained by picking the maximum diversity gain among the
for the distortion exponent of communication over MIMO two protocols at each multiplexing gain, i.e., d∗AF −DT (r) =
channels in the limit of infinite layers, which serves as an max{d∗DT (r), d∗AF (r)}. The portion of the layers that will be
upper bound for LS scheme with any finite number of layers. transmitted directly depends on the bandwidth ratio. When the
Since AF/DF relaying protocols have linear DMT curves, the bandwidth ratio is low, it is better to transmit all layers directly
same principles can be used to find the distortion exponent since the channel bandwidth is scarce and we would rather not
for LS, when all layers are transmitted using AF or DF. The use it for cooperation. On the other hand, if the bandwidth ratio
following corollary gives a counterpart to Theorem 3.1 for is higher, we can transmit the lower layers with cooperation
infinite layer LS. thus obtaining higher robustness for these important layers.
Corollary 4.1: LS with infinite layers results in the follow- Corollary 4.2: Consider LS strategy where each layer is
ing distortion exponents for direct transmission and AF/DF transmitted either with AF/DF relaying or directly as in Fig.
cooperation protocols: 6-c. The distortion exponent in the limit of infinite layers is
½
∆LS
DT = 1 − e−b , (18) LS 1 − e−b if b ≤ ln 3
∆AF −DT = (20)
∆LS
AF
LS
= ∆DF = 2(1 − e−b/4 ). (19) 2 − 4 · 3−3/4 e−b/4 if b > ln 3.
Half−duplex relay channel the threshold effect. Another technique in the literature that
2
successfully mitigates the threshold effect is hybrid digital-
1.8 analog transmission. Mittal and Phamdo [32] introduced hy-
1.6 brid digital-analog transmission to improve the end-to-end
distortion performance when transmitting over an unknown
Distortion exponent, ∆
1.4
noise channel or broadcasting to users with different noise
1.2
variances. For practical construction of hybrid schemes, see
1 [33]- [35] and references therein.
0.8 In [12], we considered transmitting source samples in a
pure analog fashion, i.e., simply by power scaling. Although
0.6 Upper bound
LS upper bound this is optimal for a point-to-point single antenna system
0.4 LS with DDF (infinite layers) for b ≥ 1 [13], when the relay is present, pure analog
LS with AF and DT(infinite layers)
0.2 LS with AF (or DF) (infinite layers)
transmission cannot utilize the additional diversity introduced.
0
LS with DT(infinite layers) Hybrid digital-analog transmission has the potential to both
0 0.5 1 1.5 2 2.5
Bandwidth ratio, b
3 3.5 4
mitigate threshold effect and utilize the inherent diversity in
the system, and hence improve the distortion exponent as well
Fig. 7. Distortion exponent vs. bandwidth ratio for LS using different
[10].
cooperation protocols. For b ≤ 1, we use the hybrid scheme of [10] which is
based on dividing the source samples into two groups, where
one group is coded, the other one is simply scaled, and the
redundancy codes for the source and space-time codes for the two groups are superimposed for transmission. This strategy
joint source-relay transmission. achieves ∆ = b, which is equal to ∆U B for b ≤ 1 and thus,
Theorem 4.3: Consider a LS strategy that uses DDF proto- this hybrid scheme for direct transmission is optimal for the
col for all the layers. In the limit of infinite layers the distortion relay scenario as well.
exponent is For b > 1, we propose hybrid digital-analog transmission
½ √ with layered source (HLS) strategy that combines LS with
−1 + 1 + 2b if 0 ≤ b ≤ 3/2
∆LSDDF = (22) analog transmission [1]. This can be considered as a combi-
2 − e3/4 e−b/2 if 3/2 ≤ b.
Proof: In this case, the distortion exponent cannot be nation of the hybrid digital-analog scheme of [10] and our LS
obtained directly from [1], since the tradeoff curve is not a strategy. In HLS, we divide N channel uses into two portions.
piecewise linear function anymore. Appendix I provides the The first portion, composed of N −K channel uses, is reserved
details. for layered source coding similar to LS, where the second
In Fig. 7 we plot the distortion exponent vs. bandwidth portion, composed of K channel uses, is used for uncoded
ratio for LS coupled with various transmission strategies transmission as illustrated in Fig. 4-b. Just as in the LS, we
and the upper bound ∆U B of Section III. We can observe assume that the source samples are compressed into n layers
the benefit of AF/DF type simple cooperation over DT at where layer k contains the successive refinement information
high bandwidth ratios. Although the distortion exponent of for layers 1, . . . , k − 1. These source layers can be transmitted
cooperative schemes converge to 2 as b → ∞, while DT using direct transmission or cooperation. Let s̄ ∈ CK be
distortion exponent is limited to 1, the rate of convergence is the reconstruction of the source vector s = [s1 , . . . , sK ]
very slow. LS with AF-DT as in Corollary 4.2 performs better upon successful reception of all the LS layers. We denote
than each strategy alone. As expected, DDF protocol has a the reconstruction error as e ∈ CK where e = s − s̄. We
considerable improvement over the other strategies, however, transmit this error during the second portion of K channel
there is still a significant gap from the upper bound. We uses where each component of the error vector is transmitted
also included the distortion exponent upper bound for the LS without coding in an analog fashion. We simply scale the error
scheme, by considering LS transmission over 2 × 1 MISO vector such that the transmit power constraint is satisfied. This
system. The gap between the DDF tradeoff curve and the DMT analog transmission is directly from the source terminal to the
upper bound for multiplexing gains above 1/2 reflects in the destination without cooperation. This is justified since relaying
distortion exponent as well. We also note that LS distortion does not improve the distortion exponent behavior of analog
exponent upper bound is still far away from ∆U B of Section transmission as discussed above.
III. The destination first attempts to decode all the digitally
transmitted layers, and in case of successful reception of all
the layers, it forms the estimate s̄ + ẽ, where ẽ is the linear
B. Hybrid Digital-Analog Transmission with Layered Source MMSE estimate of e based on the received signal during K
(HLS) channel uses reserved for analog transmission. This analog
In general, digital transmission suffers from the threshold portion is neglected unless all digitally transmitted layers can
effect, that is, the system performance degrades drastically be successfully decoded at the destination. Similar to Section
when the instantaneous channel capacity falls below the tar- IV-A, we can find the HLS distortion exponent as stated in
get rate. LS strategy introduces source scalability and thus the following theorem.
adaptivity to the channel variations to partially overcome Theorem 4.4: For HLS where each layer is transmitted
8 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. X, NO. X, MONTH, YEAR
Half−duplex relay channel above would correspond to transmitting to users through time-
2
sharing. However, as proved in [38], superposition coding
1.8 performs better than time-sharing for broadcast channels.
1.6 Assuming successive decoding at the receiver, information
1.4
intended for a bad channel state can also be received when
Distortion exponent, ∆
used. Let R = [R1 , . . . , Rn ]T be the channel rate vector. If and Tse also consider the problem of transmitting multiple
we define the average distortion achieved by receiving the first streams over a MIMO channel with different multiplexing and
k layers as DiBS , we can write diversity gains [40]. They coin the term successive refinability
à i ! of the diversity-multiplexing tradeoff curve for the cases when
X
DiBS = D b Rk , (27) the reliability of receiving layers up to k is equal to the
k=1 reliability one would achieve sending only these layers. Hence,
having a successively refinable DMT curve is equivalent to
where we use the successive refinability of the source. Then
successive decoding diversity gain lying on the DMT curve.
we can write the expected distortion for n layer BS strategy
We next investigate various cooperation protocols and power
using successive decoding as
allocation among layers to study the successive decoding
n
diversity gain. We will show that, a power allocation rule
X
ED(R, SNR) = d,i+1
(P̄out d,i
− P̄out )DiBS , (28) similar to the one used for the SIMO/MISO systems in [1] will
i=0
successively refine the diversity-multiplexing tradeoff curve of
d,0 d,n+1
some cooperation protocols we study, while extension to other
with P̄out = 0, P̄out = 1, and D0BS = 1. Similar to Section protocols is not straightforward.
IV-A, we can argue that the rate of each layer should scale as 1) Amplify-and-Forward (AF): We first consider the AF
O(log SNR) with increasing signal-to-noise ratio in order to cooperation protocol for each of the transmitted layers. In BS
have vanishing expected distortion. Let the multiplexing gain with AF, N uses of each channel realization are divided into
vector be r = [r1 , . . . , rn ]T . We have R = r log SNR, and two equal portions. While the source transmits a superposition
the high SNR approximation for expected distortion can be of the channel codewords of n layers with an appropriate
obtained as power allocation in the first half, the relay simply amplifies
n
. X d,i+1 Pi and forwards its received signal within its power constraint in
ED(r, SNR) = P̄out SNR−b j=1 rj . (29) the second half. Let the received signals at the relay and the
i=0
destination at time instant i be yr [i] and yd [i], respectively.
We define the successive decoding diversity gain of each We illustrate the BS strategy with AF cooperation protocol in
layer at the destination based on the outage probability of Fig. 9-b. We have,
d,i .
successive decoding as dsd (ri ), i.e., P̄out = SNR−dsd (ri ) . n
X p
Then we can rewrite the expected distortion as below. yr [i] = h2 SNRj xs,j [i] + zr [i], for i = 1, . . . , N/2
n
X Pi j=1
.
ED(r, SNR) = SNR−dsd (ri+1 ) SNR−b j=1 rj
.(30) Xn
p
i=0 yd [i] = h1 SNRj xs,j [i] + zd [i], for i = 1, . . . , N/2
Based on this expression, since the slowest decay rate will j=1
dominate the expected distortion, we can find the distortion yd [i] = h3 βyr [i − N/2] + zd [i], for i = N/2 + 1, . . . , N
exponent of BS strategy for the given power and multiplexing
where zr [i] and zd [i] are independent additive circularly sym-
gain allocation and the underlying transmission policy as
metric complex Gaussian
p noise terms with variance 1/2 in
X i each dimension, SNRj xs,j [i] is the source signal transmit-
∆ = min dsd (ri+1 ) + b rj . (31) ted at time instant i, corresponding to the j-th source layer, and
0≤i≤n
j=1 βyr [i − N/2] is the relay signal transmittedP at time i. Due to
N
Recall that in LS, the DMT curve of the underlying trans- power P constraint at the source, we have N1 i=1 xs,j [i] = 1,
n
mission scheme is sufficient to determine the best achievable while j=1 SNRj ≤ SNR. For the relay to satisfy its power
distortion exponent. Similarly, when BS is used, once we constraint, we need
s
are given the successive decoding diversity-multiplexing gain SNR
tradeoff curve of the underlying transmission scheme obtained β= 2
. (32)
|h2 | SNR + 1
by the best power allocation among layers, we can find the
corresponding highest distortion exponent achieved by BS Theorem 4.5: The optimal distortion exponent of AF coop-
strategy. Let d∗ (r) be the original DMT curve for the given eration protocol with BS, in terms of the number of layers n,
transmission scheme obtained when transmitting a single code- and bandwidth ratio b, is
word with full power. We can bound the successive decoding 8(4 − b)
diversity gain of the i-th layer as dsd (ri ) ≤ d∗ (r1 + . . . + ri ), ∆BS
AF,n = 2− . (33)
since successful transmission of the i-th layer with BS requires 16 + (32 − b2 )( 4b )n−1
successful transmission of all the preceding layers, which is This is an increasing function of n, and in the limit of infinite
equivalent to transmitting a total rate of (r1 +. . .+ri ) log SNR. layers, we obtain
In [1], we proposed a power allocation scheme for ½
b/2 if 0 < b < 4
MISO/SIMO systems such that dsd (ri ) = d∗ (r1 + . . . + ri ) ∆BS
AF = lim ∆ BS
AF,n = (34)
n→∞ 2 if 4 ≤ b.
for all i = 1, . . . , n. This condition suggests that all layers can
simultaneously operate on the optimal DMT curve for all mul- Proof: We defineP
the total power allocated to layers k, k+
n
tiplexing gain allocations. In an independent study, Diggavi 1, . . . , n as SNRk , i=k SNRi for k = 1, . . . , n. We have
10 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. X, NO. X, MONTH, YEAR
N channel uses
Source tx. layer 1
SNR1 = SNR. In order to ensure successive refinement of It is easy to check that, this is achieved by the following
the AF DMT curve we use an exponential power allocation multiplexing gain allocation, leading to eqn. (33).
strategy similar to the one used in [12], [13]. For k = 2, . . . , n,
2(4 − b)
and multiplexing gain vector r = [r1 , . . . , rn ] let r1 = ¡ ¢n−1 ,
16 + (32 − b2 ) 4b
SNRk = SNR1−2r1 −···−2rk−1 −²k−1 , (35) µ ¶i−1
b
for some 0 < ²1 < · · · < ²n−1 . Then we have ri = r1 , for i = 2, . . . , n.
4
SNRi = SNR1−(2r1 +···+2ri−1 +²i−1 ) − SNR1−(2r1 +···+2ri +²i ) ,
(36)
2) Decode-and-Forward (DF): Next we consider decode-
for i = 1, . . . , n − 1, and
and-forward (DF) as our cooperation protocol. Recall that the
SNRn = SNR1−(2r1 +···+2rn−1 +²n−1 ) . DF protocol of [5] achieves the same DMT curve as of AF.
Although in the LS strategy, this equivalence of the DMT
We show in Appendix III that, with ²n−1 → 0, the curves would suggest an equality of the achievable distortion
successive decoding diversity gain of the AF protocol becomes exponents, in the BS strategy, we cannot claim this directly,
dAF ∗
sd (ri ) = dAF (r1 + · · · + ri ) = 2(1 − 2r1 − · · · − 2ri ), for since we need the equivalence of the successive decoding
i = 1, . . . , n. Hence, AF protocol is successively refinable in DMT curves. In our analysis, we will generalize the DF
the DMT sense. protocol in [5] to unequal division of the channel between
Once we have the successive decoding diversity gain for the source and the relay transmissions, and find the conditions
AF protocol, we can rewrite the distortion exponent from Eqn. under which the resulting DMT is successively refinable. In
(31). the generalized DF protocol we consider, the source transmits
∆ = min {2 − 4(r1 + · · · + ri+1 ) + b(r1 + · · · + ri )} , for the first αN channel uses (α ≥ 1/2), and for the rest of
0≤i≤n
the (1 − α)N channel uses only the relay terminal transmits if
where we define rn+1 = 1 − (r1 + · · · + rn ). The optimal it can decode the source information. We assume that the relay
distortion exponent for this transmission scheme is obtained reencodes the original message using an independent Gaussian
when all the terms in the above minimization are equal. codebook. Note that α does not depend on the instantaneous
GUNDUZ, ERKIP: SOURCE AND CHANNEL CODING FOR COOPERATIVE RELAYING 11
fading state, so this can be considered as a static protocol. we obtain the successive decoding diversity gain dsd (ri ) for
Lemma 4.6: The DMT for the generalized DF protocol for layer i as
fixed α is characterized by ( µ ¶+
³ ´ 1
r
, 2 − 2r dsd (ri ) = min 2 1 − (r1 + · · · + ri ) ,
min 2 − 1−α α if r ≤ 1 − α α
¡ ¢
d∗gDF (r) = 2r 1−r
min 2 − α , α if 1 − α ≤ r ≤ α µ ¶+
1
0 if r > α. 1 − (r1 + · · · + ri )
(37) α
µ ¶+ )
Proof: Proof can be found in Appendix IV. 1
Note that, for α = 1/2 we get the classical DF protocol + 1− (r1 + · · · + ri−1 ) , (43)
1−α
with DMT d∗ (r) = 2(1 − 2r) for r ∈ [0, 1/2]. On the other
hand, for α = 2/3, we have d∗ (r) = 2 − 3r for r ∈ [0, 2/3] where we define x+ , max{x, 0}.
which achieves the best diversity gain over all generalized Proof: Proof can be found in Appendix V.
DF schemes for multiplexing gains less than 1/3. For each The following corollary which directly follows from Lemma
multiplexing gain r ≥ 1/3, the diversity gain is maximized 4.6 and Lemma 4.7, states that the classical DF protocol (α =
for a different channel allocation α. 1/2) indeed achieves the same distortion exponent as AF.
Now we consider using BS strategy coupled with the gener- Corollary 4.8: Classical DF protocol with α = 1/2 is
alized DF protocol. While the source transmits a superposition successively refinable in terms of the DMT curve. Thus, with
of the channel codewords for the n source layers in the first the same number of source layers, it can achieve the same
αN channel uses, the relay terminal tries to decode as many distortion exponent as AF. Therefore ∆BS BS
DF,n = ∆AF,n and
layers as possible, and retransmits those layers in the rest of the BS BS BS BS
∆DF = ∆AF , where ∆AF,n and ∆AF are given by (33)-(34).
(1 − α)N channel uses. Note that, while the diversity gain is Corollary 4.9: Consider the generalized DF protocol with
maximized for a different channel allocation for each r ≥ 1/3, α = 2/3. This cooperation protocol is successively refinable
we cannot use a different α for each source layer due to the in terms of the DMT curve if
half-duplex constraint on the relay. This constrains us to a
fixed channel allocation for transmission of all the layers. rk ≥ r1 + · · · + rk−1 , (44)
We assume successive decoding at both the relay and the for all k = 1, . . . , n.
destination. Let yr and yd be the received signals at the relay Proof: Under condition (44), using Lemma 4.7 we
and the destination, respectively. Assuming m source layers can show that the successive decoding diversity gain of DF
can be successfully decoded at the relay, the channel model protocol with α = 2/3 is
for generalized DF protocol with BS strategy can be written
as below. d∗sd (rk ) = 2 − 3(r1 + · · · + rk ), (45)
X n
p if r1 + · · · + rk ≤ 2/3 and 0 elsewhere. By Lemma 4.6, this
yr [i] = h2 SNRsk xs,k [i] + zr [i], (38)
is equal to d∗ (r1 + · · · + rk ), where d∗ (r) is the diversity gain
k=1
Xn of the generalized DF protocol with α = 2/3.
p
yd [i] = h1 SNRsk xs,k [i] + zd [i], (39) This example shows that, successive refinability of a DMT
k=1 curve may not hold for all cooperation protocols, under all
for i = 1, . . . , αN , and multiplexing gain allocations. In [41], authors prove that the
m DMT curve of two parallel fading channels is not succes-
X p
yd [i] = h3 SNRrk xr,k [i] + zd [i], (40) sively refinable. Although we do not give a formal proof, we
k=1
conjecture that generalized DF is not successively refinable
unless α = 1/2. This suggests that even though generalized
for i = αN +1, . . . , N , where zr [i] and zd [i] are
p i.i.d. sadditive DF for α = 2/3 has higher DMT than α = 1/2, it is
white
p Gaussian noise terms as before, and SNRi xs,k [i],
not clear whether there exists a power allocation resulting in
SNRrk xr,k [i] are the source and the relay signals, respec-
successive refinability of the DMT, or in general, whether a
tively, corresponding to the k-th source layer transmitted at
power allocation results in a successive decoding diversity gain
time i. Due P to power constraints at the PNsource and the relay,
1 N 1 that is higher than the DMT with α = 1/2. We will revisit this
we have P i=1 xs,k [i] = 1 and Pn i=1 xr,k [i] = 1 for all
N
n
N observation later when we discuss static cooperation protocols
k, while k=1 SNRsk ≤ SNR and k=1 SNRrk ≤ SNR.
in conjunction with BS.
Lemma 4.7: Consider the generalized DF protocol coupled
3) Direct transmission combined with AF/DF relaying:
with BS strategy with n source layers as explained above. Let
Similar to Section IV-A sending important (the first few)
r = [r1 , . .P
. , rn ] be the multiplexing rate
Pn vector. rWe define
s n s r source layers using cooperation and the rest directly may im-
SNRk = i=k SNRi and SNRk = i=k SNRi . For the prove the distortion exponent further since direct transmission
power allocation at the source terminal
DMT is better than AF/DF cooperation for high multiplexing
s 1
SNRk = SNR1− α (r1 +···+rk−1 −²k−1 ) , (41) gains. For BS, rather than having separate time slots for
cooperation and direct transmission as we did in Section IV-A
and the power allocation at the relay terminal
for LS, we need to superimpose multiple layers corresponding
r 1
SNRk = SNR1− 1−α (r1 +···+rk−1 −²k−1 ) , (42) to cooperation and direct transmission while ensuring good
12 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. X, NO. X, MONTH, YEAR
Distortion exponent ( ∆)
1.4
The overall diversity gain of this protocol is characterized by
d∗DF −DT (r) = max(1 − r, 2 − 4r) for a multiplexing gain of 1.2
Full−duplex relay channel the increase in distortion exponent with the introduction of
2
the relay terminal is mainly due to the increase in channel
1.8 diversity. We show that the availability of a correlated side
1.6 information at the destination and/or the relay would not
further improve either the upper bound or the achievable
Distortion exponent ( ∆)
1.4
distortion exponents of the schemes we have considered.
1.2
Suppose that the destination and the relay have access to
1 correlated side information sequences S1 and S2 , respectively.
0.8 We assume that Si = S + Wi , i = 1, 2, where Wi is a zero-
mean complex Gaussian with variance σi2 , and is independent
0.6
of S and each other. No side information is equivalent to
0.4 Upper bound
BS with DF (infinite layers)
having σi2 → ∞. Suppose both S1 and S2 are made available
0.2 HLS with DF (infinite layers) to all three terminals. Note that, this can only reduce the
0
LS with DF (infinite layers) achievable expected distortion, hence increase the distortion
0 0.5 1 1.5 2 2.5
Bandwidth ratio (b)
3 3.5 4
exponent. Under this assumption, we can now use conditional
distortion-rate function DS|S1 ,S2 (R) of the source S given
Fig. 11. Distortion exponent vs. bandwidth ratio for the full-duplex
(S1 , S2 ), which for the Gaussian source and side information,
cooperative relay channel. is equal to
µ ¶
σ̃ 2
DS|S1 ,S2 (R) = 2−R , (52)
Theorem 5.2: The distortion exponent of full-duplex DF 1 + σ̃ 2
with n-layer BS is given as σ2 σ2
where σ̃ 2 = σ21+σ22 .
µ ¶ 1 2
1 − b/2 We observe from (52) that, availability of side information
∆BS
f d−DF,n = 2 1 − . (50)
1 − (b/2)n+1 does not change the exponential dependence of the rate-
distortion function on R, it only reduces the effective variance.
Proof: Proof is given in Appendix VIII. The proof uses
However, constant scaling in the distortion-rate function does
successive decoding diversity of full-duplex DF which is found
not change the distortion exponent. Therefore, providing the
in Appendix IX.
side information (S1 , S2 ) to all three terminals does not im-
In the limit of infinite layers, ∆BS
f d−DF,n approaches the prove the upper bound or the achievable distortion exponents
upper bound in Theorem 3.2, and we have the following
of the schemes we have considered. We can conclude that side-
corollary.
information does not help in increasing the distortion exponent
Corollary 5.3: BS strategy coupled with full-duplex DF
of the strategies considered in this paper. However, when
relaying has optimal distortion exponent in the limit of infinite
channel SNR is finite, side information aware joint source-
source coding layers. We have
½ channel coding techniques have the potential to reduce the
b if b < 2, expected distortion [46].
∆BS
fd = lim ∆ BS
f d−DF,n = (51)
n→∞ 2 if b ≥ 2.
Fig. 11 summarizes the distortion exponents of BS, HLS VII. C ONCLUSION
and LS all employing a full duplex DF cooperation protocol.
In this work, we consider the problem of joint source chan-
nel coding over slowly fading cooperative relay systems. Our
VI. C ORRELATED S IDE I NFORMATION focus is on the high SNR behavior of the expected distortion
The original motivation behind user cooperation is to im- characterized by the distortion exponent. Even though single
prove the reliability of wireless uplink channel [3], [4]. How- layer source coding followed by cooperative transmission im-
ever, when source transmission is considered, the availability proves the achievable distortion exponent for high bandwidth
of correlated side information at the destination and/or at the ratios, its performance is quite limited for arbitrary bandwidth
relay terminals should be taken into account since it has the ratios. We propose strategies based on concatenation of layered
potential to improve the end-to-end distortion performance source coding and channel coding with various cooperation
[45]. This may be especially relevant for sensor applications protocols to improve the distortion exponent.
where user cooperation has been proposed as an effective tool There is a close relation between the DMT performance
to improve the sensor network reliability. of the underlying communication system and the achievable
Under the assumption of correlated side-information at the distortion exponent. Our techniques attempt to utilize the DMT
relay terminal, source-channel separation theorem fails even curve most efficiently to maximize the achieved distortion
when there is no fading. We investigate joint source-channel exponent. In the LS strategy, where compressed source layers
and relaying strategies for lossy transmission over an AWGN are transmitted progressively in time with the help of a
channel in [46] and for lossless transmission over discrete cooperating relay, time and rate allocation among the layers is
memoryless channel in [47], however our focus here is on the optimized. We show that this optimization is geometrically
quasi-static fading model, and the effect of side information equivalent to climbing up the underlying DMT curve. An
on the achievable distortion exponent. We next argue that extension of LS is HLS, where an additional error message is
14 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. X, NO. X, MONTH, YEAR
transmitted in analog fashion without coding, directly from the We solve this equation for M .
source terminal. HLS strictly improves the distortion exponent p
n n2 + 2bn(n + 1)
compared to LS while the improvement decays with increasing M = − + ,
bandwidth ratio. Both LS and HLS can easily be adapted to n +√1 n+1
n→∞ −1 + 1 + 2b.
use any cooperative protocol for each layer. In BS multiple −−−−→
source layers are transmitted simultaneously by an appropriate However, note that since the part of the tradeoff curve for
power and rate allocation. Unlike LS and HLS, since BS 0 ≤ r ≤ 1/2 is a straight line, above is valid only if M ≤ 1,
superimposes layers, benefits of using a cooperation strategy which corresponds to b ≤ 3/2. For b > 3/2, we use the rest of
with a better DMT is not immediately apparent, and one needs the bandwidth ratio to climb up the straight line characterized
to consider the effect of successive decoding on the DMT. by d∗ (r) = 2(1 − r). This can be done
Coupling BS with a combination of direct transmission and ¡ using¢ the results of
[1] by assuming a bandwidth ratio of b − 23 and adding 1
cooperative relaying, we prove that it achieves the optimal to the corresponding distortion exponent.
distortion exponent of the half-duplex system for b ≤ 1 and
b ≥ 4. We also analyze the distortion exponent of full-duplex A PPENDIX II
relay system, and show that BS with full-duplex decode-and- P ROOF OF T HEOREM 4.4
forward protocol can achieve the optimal distortion exponent
The expected distortion for HLS with n layer source coding
for all bandwidth ratios. Finally, we consider the availability of
can be written as
correlated side information at the destination and/or the relay
Pn−1 HLS ¡ i+1 i
¢
terminals, and show that the side information does not effect ED(R, SNR) = i=0 Di · Pout − Pout
the distortion exponents of the cooperation strategies we study. R
+ Ac DaHLS (h1 )p(h)dh, (53)
Overall, our results show that layered source coding is an
0 n+1
effective technique to improve the distortion exponent for where Pout = 0, Pout = 1, D0HLS = 0,
communication over fading wireless channels, and to achieve a à i
!
HLS
¡ ¢X
higher distortion exponent, we need to use better channel codes Di =D b−1 tk Rk for i = 1, . . . , n (54)
and cooperation strategies that improve DMT, and carefully k=1
match the code rates and cooperation scheme to the source
DnHLS
coder. DaHLS (h1 ) = , (55)
1 + |h1 |2 SNR
.. = DnHLS SNR−(1−γ)
Ac ∩R+
.
M SNR SNR SNR−θ dγdβdθ,
−γ −β
bt1 = (1 + M ) . .
n = DnHLS SNR−µ ,
Summing up both sides, we get where
Xn µ ¶
M M n(n + 1) µ = inf (1 − γ)+ + γ + β + θ,
bti = n+ · (γ,β,θ)∈Ac ∩R3+
i=1
n n 2
≥ 1, (57)
M 2 (n + 1)
b = M+ . +
since (1 − γ) + γ ≥ 1 for any γ ≥ 0.
2n
GUNDUZ, ERKIP: SOURCE AND CHANNEL CODING FOR COOPERATIVE RELAYING 15
Eqn. (56) tells us that the SNR exponent of the last term is Then the successive decoding outage probability for layer j
at least 1. We will assume that it behaves as SNR−1 to find a can be written as
½ ¾
lower bound to the overall distortion exponent. Using d∗ (r), d,j 1
the DMT curve of the particular cooperation protocol utilized Pout = P r IAF,j < Rj ,
2
for the source layers, and defining d∗ (rk+1 ) = 1, we can ½ ¾
wSNRj + u 2rj
write the high SNR approximation for the expected distortion = Pr < SNR ,
of HLS as wSNRj+1 + u
©
n−1 h i = P r w(SNR1−2r1 −···−2rj−1 −²j−1 −
. X ∗ ∗ ª
ED(R, t, SNR) = SNR−d (ri+1 ) − SNR−d (ri ) SNR1−2r1 −···−2rj−1 −2rj −²j ) < u(SNR2rj − 1) ,
i=0 . © ª
Pi Pn = P r wSNR1−2r1 −···−2rj−1 −²j < uSNR2rj ,
· SNR−(b−1) k=1 tk rk + SNR−(b−1) k=1 tk rk −1 n o
.
n = P r wSNR/u < SNR2(r1 +···+rj +²j ) ,
. X ∗ Pi
½
= SNR−d (ri+1 )−(b−1) k=1 tk rk |h2 |2 |h3 |2 SNR2
.
i=0 = P r |h1 |2 SNR +
. Pi 1 + |h2 |2 SNR + |h3 |2 SNR
= SNRmax0≤i≤n [−d (ri+1 )−(b−1) k=1 tk rk ].
∗
o
< SNR2(r1 +···+rj ) ,
An analysis similar to Section IV-A with infinite number of
layers gives us the distortion exponents for AF/DF and DDF as ²j → 0. Comparing this last equality with the outage
protocols. probability expression of the AF protocol for a single stream
in [5], we conclude that in the high SNR regime,
A PPENDIX III d,j . ∗
Pout = SNR−dAF (r1 +···+rj ) , (59)
S UCCESSIVE D ECODING D IVERSITY G AIN OF AF
and we have
We first find the maximum mutual information of the j-th
d,j . d,j−1 d,j
layer for the AF protocol assuming all previous layers have P̄out = max(P̄out , Pout ), (60)
been decoded and subtracted. Layers j + 1, . . . , n will be . d,j
= Pout . (61)
treated as noise. First, we define
n Therefore dAF ∗
sd (ri ) = dAF (r1 + · · · + ri ), and the diversity-
X p
xs,j , SNRj xs,j . multiplexing tradeoff curve of the AF protocol with the pro-
i=j
posed exponential power allocation is successively refinable.
Then the equivalent channel can be written as A PPENDIX IV
· ¸ · ¸
yd [i] h1 p P ROOF OF L EMMA 4.6
= SNRj xs,j [i] +
yd [i + N/2] h3 βh2 The proof closely resembles the proofs in [6], so we will
| {z } | {z }
yd [i] A briefly give the outline. Outage probability at the destination
is the sum of two probabilities: probability of outage given the
· ¸ xs,j+1 [i]
relay can decode the message at the end of the first portion
h1 0 1 0 zr [i] ,
(58) and probability of outage given the relay cannot decode:
h3 βh2 h3 β 0 1 zd [i]
| {z } zd [i + N/2] Pout = P r(outage | relay decodes )P r(relay decodes)
B | {z }
z[i] +P r(outage|relay cannot decode)
·P r(relay cannot decode).
where the noise covariance matrix is E[zz† ] =
diag(SNRj+1 , 1, 1, 1). The mutual information for source Then we have
layer j is maximized for circularly symmetric complex ©
P r(outage | relay decodes ) = P r α log(1 + |h1 |2 SNR)+
Gaussian input, and is found to be ª
(1 − α) log(1 + |h3 |2 SNR) ≤ r log SNR ,
IAF,j = log det(I + SNRj AA† (BE[zz† ]B† )−1 ).
P r(outage | relay cannot decode) =
We can find © ª
· ¸ P r α log(1 + |h1 |2 SNR) ≤ r log SNR .
† |h1 |2 βh1 h∗2 h∗3
AA = , We also have
βh∗1 h2 h3 β |h2 |2 |h3 |2
2
w , |h1 |2 + |h1 |2 |h2 |2 SNR + |h3 |2 (|h1 |2 + |h2 |2 )SNR, P r(outage|relay cannot decode)P r(relay cannot decode)
.
u , 1 + |h2 |2 SNR + |h3 |2 SNR. = SNR−d1 (r) ,
16 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. X, NO. X, MONTH, YEAR
r
where d1 (r) = 2(1 − α) for r < α and 0 elsewhere. On the where
à !
other hand, p r
1 + SNRk kh3 k2
. I( SNRrk xr,k ; ȳd,k ) = log r .
P r(outage | relay decodes )P r(relay decodes) = SNR−d2 (r) 1 + SNRk+1 kh3 k2
Note here that, it is possible that the relay decodes layer
where
k but not all the n layers, and hence not transmit anything
d2 (r) = inf γ + θ, for those uncoded layers. However, in the mutual information
(γ,θ)∈A
expression we assume all the following layers act as noise,
with which in turn gives a lower bound to the mutual information.
Then the successive decoding outage event for layer k at the
A = {(γ, θ) ∈ R2+ : α(1 − γ)+ + (1 − α)(1 − θ)+ ≤ r}.
destination terminal can be written as,
We can show that Okd = {(h1 , h2 , h3 ) : Is,r→d < Rk }.
½ r
2− 1−α if r ≤ 1 − α Recall that d,k
Pout
denotes the probability of outage for layer
d2 (r) = 1−r (62)
α if r > 1 − α. k at the destination, assuming that the decoder successfully
decodes and subtracts the previous k − 1 layers. On the other
Then we have d∗gDF (r) = min(d1 (r), d2 (r)). d,k
hand, P̄out is the overall outage probability of layer k for
successive decoding.
A PPENDIX V For k = 1, . . . , n we have
P ROOF OF L EMMA 4.7 n p
d,k
Pout = P r αI( SNRsk xs,k ; ȳd,k )
Suppose the first
Pnk −p1 layers have been decoded at the p o
r,k
relay. Let x̄s,k = i=k SNRsi xs,i and ȳr,k = h2 x̄s,k + zr . +(1 − α)I( SNRrk xr,k ; ȳd,k ) < Rk (1 − P̄out )
We find the mutual information as n p o
r,k
p +P r αI( SNRsk xs,k ; ȳd,k ) < Rk P̄out
Is→r , αI( SNRsk xs,k ; ȳr,k ),
p Let |h1 |2 = SNR−γ , |h2 |2 = SNR−β , and |h3 |2 = SNR−θ .
= αI( SNRsk xs,k , x̄s,(k+1) ; ȳr,k ) Using the source and the relay power allocation, for k =
p
−αI(x̄s,k+1 ; ȳr,k | SNRsk xs,k ), 1, . . . , n − 1, we have
s 1 +
= α log(1 + SNRk kh2 k2 ) Okr = β:α 1− (r1 − · · · − rk−1 ) − ²k−1 − β
s α
−α log(1 + SNRk+1 kh2 k2 ), 1 +
à s
! − 1− (r1 − · · · − rk ) − ²k − β < rk ,
1 + SNRk kh2 k2 α
= α log s .
1 + SNRk+1 kh2 k2 and
½ ¾
£ 1 ¤+
Similar to the successive decoding outage events at the des- Onr = β : α 1 − (r1 − · · · − rn−1 ) − ²n−1 − β < rn .
α
tination defined in Section IV-C, we define the following
successive decoding outage events for layer k at the relay Similarly, we have, for k = 1, . . . , n − 1,
½ µ
terminal, d
£ 1 ¤+
Ok = (γ, β, θ) : α 1 − (r1 − · · · − rk−1 ) − ²k−1 − γ
α
Okr = {h2 : Is→r < Rk }, £ ¤+
¶
1
[k − 1 − (r1 − · · · − rk ) − ²k − γ
Ōkr = Okr , α
µ
i=1
£ 1 ¤+
+(1 − α) 1 − (r1 − · · · − rk−1 ) − ²k−1 − θ
1−α
and the corresponding outage probabilities ¶ ¾
£ 1 ¤+ ¯
r,k − 1− (r1 − · · · − rk ) − ²k − θ ¯
< rk τk = 0
Pout = P r{h2 : h2 ∈ Okr }, 1−α
r,k [ ½ µ
P̄out = P r{h2 : h2 ∈ Ōkr }. £ 1 ¤+
(γ, β, θ) : α 1 − (r1 − · · · − rk−1 ) − ²k−1 − γ
α
Now we look at the outage event at the destination for layer ¶ ¾
£ 1 ¤+ ¯
k assuming that the P previous
√ layers have been successfully − 1 − (r1 − · · · − rk ) − ²k − γ < rk ¯ τ k = 1 ,
n
decoded. Let x̄r,k = i=k SNRi xr,i and ȳd,k = h1 x̄s,k + α
h3 x̄r,k + zr . The outage event at the destination will depend and
on the outage event at the relay for layer k. Let τk be the £ 1 ¤+
Ond = {(γ, β, θ) : α 1 − (r1 − · · · − rn−1 ) − ²n−1 − γ
random variable denoting the outage event at the relay, i.e., α
£ r1 − · · · − rn−1 ¤+ ¯
we have τk = 1 if layer k can be decoded at the relay, and + (1 − α) 1 − − ²n−1 − θ < rn ¯τk = 0}
τk = 0 otherwise. Then we can write [ 1 − α
p {(γ, β, θ) :
Is,r→d , αI( SNRsk xs,k ; ȳd,k ) £ ¤+ ¯
p r1 − · · · − rn−1
+(1 − α)sk I( SNRrk xr,k ; ȳd,k ), α 1− − ²n−1 − γ < rn ¯τk = 1},
α
GUNDUZ, ERKIP: SOURCE AND CHANNEL CODING FOR COOPERATIVE RELAYING 17
the same as DT for which the DMT is successively refinable, A PPENDIX VIII
we find dDF
sd
−DT
(rj ) = 1 − (r1 + · · · + rj ). P ROOF OF T HEOREM 5.2
For layers j = 1, . . . , k, when there is outage at the relay, The source terminal uses block Markov superposition en-
the outage probability expression at the destination is same as coding and the destination does backward decoding [43], [44].
transmitting all layers with DF protocol, since the latter layers Total N uses of the channel, during which the fading is
have the same total power and all are transmitted only through constant, is divided into B+1 blocks, each of length N/(B+1)
the direct channel. However, when layer j, for j = 1, . . . , k, channel uses. Let wi,b be the message corresponding to layer i
can be decoded at the relay, the destination will combine both in block b, which is compressed at rate bRi with i = 1, . . . , n,
the source and the relay signals for decoding. Assume that the b = 1, . . . , B.
relay transmits all the layers after j up to k as well. This will In the first block, the source transmits a superposition of
only degrade the performance for layer j as those layers act as the codewords corresponding to all layers for b = 1 given by
noise when decoding that layer. Now, when decoding layer j the scaled versions of xs,1 (1, w1,1 ), . . . , xs,n (1, wn,1 ), while
at the destination, the received signal at the destination in the the relay transmits a known signal. The relay then tries to
second half will be composed of layers j, . . . , k transmitted decode as many layers as possible. The number of layers
from the relay, and layers k + 1, . . . , n transmitted from the that the relay can decode depends on the fading level of the
source directly.We have source-relay channel and can be considered as constant for all
k
the B + 1 blocks, since the channel is constant. In the next
X p block, the relay transmits the layers it decodes, superimposing
yd [i] = h3 SNRrm xr,m [i]
m=j
them with the same power allocation as the source, that
n is, the relay transmits a superposition of scaled versions of
X p 0 0 0
+h1 SNRrm xs,m [i] + zd [i], xr,1 (w1,1 ), . . . , xr,k (wk,1 ). Here wj,1 denotes the estimate of
m=j+1 the relay and is equal to wj,1 for j = 1, . . . , k with high
probability as these layers are not in outage. Both the source
for j = N/2 + 1, . . . , N . Then for the second half of the time and the relay use Gaussian codebooks.
slot, the mutual information between the relay codeword for In the second and the following blocks, the source transmits
layer j and the received signal at the destination is found as a superposition of codewords for all layers using a doubly
à ! indexed codebook based on the current and the previous
|h3 |2 SNRj block messages (see Fig. 12). The destination combines the
log 1 + Pk
1 + |h3 |2 m=j+1 SNRm + |h1 |2 SNRk+1 information from the source and the relay, and tries to decode
à Pk ! as many layers as possible starting from block B + 1 and
1 + |h3 |2 m=j SNRm + |h1 |2 SNRk+1 going backwards. For the layers that the relay cannot decode,
= log Pk ,
1 + |h3 |2 m=j+1 SNRm + |h1 |2 SNRk+1 destination decodes them based on its received signal from the
µ ¶ source terminal only. We assume B is large and N/(B + 1) is
1 + |h3 |2 SNRj + (|h1 |2 − |h3 |2 )SNRk+1
= log , integer. Both are valid based on the infinite N assumption.
1 + |h3 |2 SNRj+1 + (|h1 |2 − |h3 |2 )SNRk+1
µ ¶ Further details of encoding and decoding structure can be
1 + |h3 |2 SNRj found in [44].
≥ log ,
1 + |h3 |2 SNRj+1 Let the received signals at the relay and the destination for
block b be yr,b and yd,b , respectively. We have
if |h1 |2 ≤ |h3 |2 , that is, for this case, the mutual information n p
X
is at least as much as transmitting all layers from the relay yr,b = h2 SNRi xs,i (wi,b−1 , wi,b ) + zr , (67)
terminal, as in the pure cooperation protocol. Hence, if the i=1
condition |h1 |2 ≤ |h3 |2 is satisfied, then transmitting layers n p
X
k + 1, . . . , n from the source in the second half, instead yd,b = h1 SNRi xs,i (wi,b−1 , wi,b )
of transmitting from the relay can only increase the mutual i=1
Xn p
information for the layers 1, . . . , k, as the source-destination 0
channel is weaker, and hence has lesser interference compared +h3 SNRi xr,i (wi,b−1 ) + zd , (68)
i=1
to the relay-destination channel. If we have |h1 |2 ≥ |h3 |2 ,
the mutual information term coming from the first half of the where zr and zd are independent additive white Gaussian noise
time slot would be dominant in the high SNR regime. In this vectors
√ with i.i.d.
√ components of zero mean and variance 1/2,
case, since the second half would have no effect on outage SNRi xs,i , SNRi xr,i are the source and the relay signals,
event anyway, the high SNR behavior would be equivalent to respectively, corresponding to the i-th source layer. We assume
transmitting all layers with cooperation. xr,i is 0 if the layer i cannot be decoded at the relay. Due
We conclude that the successive decoding outage probability to power constraints at the sourcePn and the relay, we have
for the first k layers which are transmitted by simple coop- ||xs,i ||2 = ||xr,i ||2 = 1, while j=1 SNRj ≤ SNR.
For each block, both the relay and the destination use
eration is the same as when all the layers are transmitted by
successive decoding starting from the first layer. For the power
cooperation. Hence, the DMT curve is successively refinable
allocation
for these layers as well, and we have dDF sd (rj ) = 2 − 4(r1 +
· · · + rj ) for j = 1, . . . , k. SNRk = SNR1−r1 −···−rk−1 −²k−1 , (69)
GUNDUZ, ERKIP: SOURCE AND CHANNEL CODING FOR COOPERATIVE RELAYING 19
SOURCE
0
xr,1 (1) 0
xr,1 (w1,1 ) xr,1 (w1,B )
RELAY
0
xr,k (1) 0
xr,k (wk,1 ) xr,k (wk,B )
Fig. 12. Block Markov encoding for full duplex relay BS. In the figure, it assumed that the relay can decode, and thus can forward, up to layer k. Also
0
wb,j denotes the relay’s estimate of layer j in block b.
for some 0 < ²1 < · · · < ²n−1 where SNRk was defined in The successive decoding outage event for layer k at the
Section IV-C, we show in Appendix IX that dsd (rk ) = 2(1 − relay terminal is then found as,
r1 − · · · − rk ). Therefore, the full-duplex DMT is successively µ ¶
refinable. 1 + SNRk kh2 k2
Okr = {h2 : log < Rk }, (73)
Assigning the multiplexing gains as below, all the terms in 1 + SNRk+1 kh2 k2
[k
Eqn. (31) will be the same.
Ōkr = Okr , (74)
1 − b/2 i=1
r1 = , (70)
1 − (b/2)n+1 where the corresponding outage probabilities are defined as in
b Appendix IV.
ri = ri−1 , for i = 2, 3, . . . , n, (71)
2 r,k
Pout = {h2 : h2 ∈ Okr }, (75)
and we find the distortion exponent given in (50). r,k
P̄out = {h2 : h2 ∈ Ōkr }. (76)
A PPENDIX IX If the relay can decode up to layer k in the first block, then
S UCCESSIVE D ECODING D IVERSITY G AIN OF it can decode k layers in the following blocks as well, as it has
F ULL -D UPLEX DF a perfect knowledge of the previous block’s information up to
layer k and the mutual information corresponding to layers
The source codeword xs,i for layer i can be considered as 1, . . . , k are the same as in the first block.
the superposition of the two codewords, one for the current The destination uses backward decoding. Hence it starts
2
message xs,i , and the other for the previous block’s message decoding from the last block and decodes as many layers as
1
xs,i , where we dropped the dependence on the messages possible. Since in this block, the source terminal does not send
for simplicity. Both codewords are formed using Gaussian any new information, the destination receives layers decoded
distribution. Without loss of generality, we assume ||x1s,i ||2 = by the relay from both terminals, and the remaining layers
||x2s,i ||2 = 1/2. only from the source terminal. Then it moves backwards to
At each block, the relay only tries to decode the layers the next block. If the destination can decode up to some layer
that it could decode in the previous block. Then, for each j in the previous block, then knowing the message for these
block, when decoding the message corresponding to a source layers for the previous block, it can decode the same number of
layer in that block, the relay already knows the previous layers in the next block as well, since the mutual information
block’s message corresponding to that layer. Thus the mutual for layers 1, . . . , j in the next block would be equal to the
information related to the new message at layer k is previous block. Pn √ Pn √
à ! Let x̄r,j = i=j SNRi xr,i , x̄s,j = i=j SNRi xs,i and
1 + 21 SNRk kh2 k2 ȳd,j = h1 x̄s,j + h3 x̄r,j + zr . The destination tries to decode
log , (72)
1 + SNRk+1 kh2 k2 the layers that it could decode in the previous block. For each
block, when decoding the message corresponding to a certain
where we can omit the 1/2 term in the numerator as that has source layer coming from both the source and the relay, the
no effect on the high SNR behavior. destination already knows the new information sent by the
20 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. X, NO. X, MONTH, YEAR
source related to the decoded layers due to backward decoding. From Laplace’s method [31], for the above multiplexing
Then the mutual information related to the message of layer gain and power allocation among the layers, high SNR behav-
d,j
j, that the receiver can gather is ior of the outage probabilities Pout for j = 1, . . . , n.
³ ´
1+SNRj (kh1 k2 +kh3 k2 ) d,j
log 1+SNRj+1 (kh1 k2 +kh3 k2 ) if relay decodes layer j, Pout = SNR−dsd (rj ) , (77)
³ ´ where
log 1+SNRj kh1 k2 otherwise.
1+SNR kh k2j+1 1 dsd (rj ) = inf γ + β + θ. (78)
(γ,β,θ)∈Ojd ∩R3+
Note that, in the above mutual information expression, if layer
k is decoded at the relay, when writing the noise term in Now we have two minimizing (γ, β, θ) tuples, one for each
the denominator inside the logarithm, we assume that all the set in Ojd ∩ R3+ above. We have
following layers can also be decoded and forwarded by the
relay, and hence act as noise at the destination when decoding γ̃ = β̃ = 1 − r1 − · · · − rj − ²j−1 , and θ̃ = 0, (79)
layer j. This can only degrade the outage performance. Then or,
outage event for layer j at the destination terminal can be
written as, γ̃ = θ̃ = 1 − r1 − · · · − rj − ²j−1 , and β̃ = 0. (80)
Ojd = In both cases, the successive decoding diversity gain dsd (rj )
p p for j = 1, . . . , n as ²j → 0 can be found as
{(h1 , h2 , h3 ) : I( SNRj x1s,j , SNRj xr,j ; ȳd,j ) < Rj },
j
[ dsd (rj ) = 2(1 − r1 − · · · − rj ). (81)
Ōjd = Ojd ,
i=1 Due to successive decoding at the destination, we have
and the corresponding outage probabilities are the same as d,j . d,j−1 d,j
P̄out = max(P̄out , Pout ), (82)
Eqn. (25)-(26). . d,j
= Pout . (83)
For j = 1, . . . , n − 1 we have
½µ ¶ ¾ We conclude that the successive decoding DMT for the expo-
d,j 1 + SNRj (kh1 k2 + kh3 k2 ) rj
Pout = Pr < SNR nential power allocation is equivalent to the DMT of a single
1 + SNRj+1 (kh1 k2 + kh3 k2 ) data stream transmitted at multiplexing gain (r1 + · · · + rk ),
½µ ¶ ¾
r,j 1 + SNRj kh1 k2 rj r,j that is, DMT of full-duplex DF is successively refinable.
· (1 − P̄out ) + P r < SNR P̄out
1 + SNRj+1 kh1 k2
and R EFERENCES
© ª [1] D. Gündüz, E. Erkip, “Joint source-channel codes for MIMO block
n
Pout = P r 1 + SNRn (kh1 k2 + kh3 k2 ) < SNRrn fading channels,” under revision, IEEE Transactions on Information
r,j © ª r,j Theory.
· (1 − P̄out ) + P r 1 + SNRn kh1 k2 < SNRrn P̄out . [2] J. N. Laneman, E. Martinian, G. W. Wornell, and J. G. Apostolopoulos,
“Source-channel diversity for parallel channels,” IEEE Transactions on
Let |h1 |2 = SNR−γ , |h2 |2 = SNR−β , and |h3 |2 = SNR−θ . Information Theory, vol. 51, no. 10, pp. 3518-3539, Oct. 2005.
Then with the power allocation in (69), for j = 1, . . . , n − 1, [3] A. Sendonaris, E. Erkip, and B. Aazhang, “User cooperation diversity,
Part I and Part II,” IEEE Trans. on Communications, vol. 51, pp. 1927-48,
Ojr = {β : (1 − r1 − · · · − rj−1 − ²j−1 − β)+ November 2003.
[4] A. Sendonaris, E. Erkip, and B. Aazhang, “User cooperation diversity,
+(1 − r1 − · · · − rj − ²j − β)+ < rj }, Part II: Implementation aspects and performance analysis,” IEEE Trans.
on Communications, vol. 51, pp. 1939-1948, November 2003.
and [5] J. N. Laneman, D. N. C. Tse, and G. W. Wornell, “Cooperative diversity
in wireless networks: Efficient protocols and outage behavior,” IEEE
Onr = {β : (1 − r1 − · · · − rn−1 − ²n−1 − β)+ < rn }. Trans. Inform. Theory, vol. 50, no. 12, pp. 3062-3080, Dec. 2004.
[6] K. Azarian, H. El Gamal, P. Schniter, ”On the achievable diversity-
Similarly, we have, for j = 1, . . . , n − 1, multiplexing tradeoff in half-duplex cooperative channels,” IEEE Trans.
Info. Theory, vol. 51, no. 12, pp. 4152-4172, Dec. 2005.
Ojd = {(γ, β, θ) : (1 − r1 − · · · − rj−1 − ²j−1 − min(γ, θ))+ [7] M. Yuksel, E. Erkip, “Multi-antenna cooperative wireless systems: A
diversity multiplexing tradeoff perspective,” to appear, IEEE Transactions
+ (1 − r1 − · · · − rj − ²j − min(γ, θ)))+ < rj |β ∈
/ Ōjr } on Information Theory.
[ [8] T. Holliday, A. Goldsmith, “Optimizing end-to-end distortion in MIMO
{(γ, β, θ) : (1 − r1 − · · · − rj−1 − ²j−1 − γ)+ systems,” Proc. of IEEE International Symposium on Information Theory
(ISIT), Adelaide, Australia, Sep. 2005.
+ (1 − r1 − · · · − rj − ²j − γ)+ < rj |β ∈ Ōkr }, [9] B. P. Dunn and J. N. Laneman, “Characterizing source-channel diversity
approaches beyond the distortion exponent,” in Proc. Allerton Conf.
and Communications, Control, and Computing, Monticello, IL, Sept. 2005.
[10] G. Caire, K. Narayanan , “On the SNR exponent of hybrid digital-analog
Ond = {(γ, β, θ) : space time coding,” in Proceedings of Allerton Conference, Allerton, IL,
September 2005.
(1 − r1 − · · · − rn−1 − ²n−1 − min(γ, θ))+ < rn |α2 ∈
/ Ōnr } [11] D. Gündüz, E. Erkip, “Joint source-channel cooperation: Diversity
[ versus spectral efficiency,” in Proc. of IEEE International Symposium
{(γ, β, θ) : on Information Theory (ISIT), Chicago, June 2004.
[12] D. Gündüz, E. Erkip, “Source and channel coding for cooperative
(1 − r1 − · · · − rn−1 − ²n−1 − γ)+ < rn |β ∈ Ōnr }, relaying,” in Proceedings of SPAWC, New York City, June 2005.
GUNDUZ, ERKIP: SOURCE AND CHANNEL CODING FOR COOPERATIVE RELAYING 21
[13] D. Gündüz, E. Erkip, “Source and channel coding for quasi-static [38] T. M. Cover, “Broadcast channels,” IEEE Transactions on Information
fading channels,” in Proceedings of 39th Asilomar Conf., Pacific Grove, Theory, vol. IT-18 pp. 2-14, January 1972.
California, November 2005. [39] S. Shamai, A. Steiner, “A broadcast approach for a single-user slowly
[14] D. Gündüz, E. Erkip, “Distortion exponent of MIMO fading channels,” fading MIMO channel,” IEEE Transactions on Information Theory,
in Proceedings of IEEE Information Theory Workshop, Punta del Este, October 2003.
Uruguay, March 2006. [40] S. Diggavi, D. Tse, “Fundamental limits of diversity-embedded codes
[15] D. Gündüz, E. Erkip, “Distortion exponent of parallel channels,” in Proc. over fading channels,” Proceedings ISIT, Adelaide, Australia, September
of IEEE International Symposium on Information Theory (ISIT), Seattle, 2005.
WA, July 2006. [41] S. Diggavi, D.N.C. Tse, “On opportunistic codes and broadcast codes
[16] K. Bhattad, K. R. Narayanan, G. Caire, “On the distortion exponent of with degraded message sets,” In Proc. of IEEE Information Theory
some layered transmission schemes”, in Proceedings of 40th Asilomar Workshop, Punta del Este, Uruguay, March 2006.
Conf., Pacific Grove, California, Nov. 2006. [42] N. Prasad, M. K. Varanasi, ”High performance static and dynamic
[17] F. Etemadi, H. Jafarkhani, “Optimal layered transmission over quasi- cooperative communication protocols for the half duplex fading relay
static fading channels,” in Proc. of IEEE International Symposium on channel,” in Proc. of IEEE Globecom Conference, San Fransisco, CA,
Information Theory (ISIT), July 2006. Nov. 2006.
[18] C. T. K. Ng, D. Gündüz, A. J. Goldsmith and E. Erkip, “Minimum [43] TM Cover and A. El Gamal, “Capacity theorems for the relay channel,”
expected distortion in Gaussian joint source-channel layered broadcast IEEE Trans. Inform. Theory, vol. IT-25, no. 5, pp. 572-584, Sept. 1979.
coding,” in Proc. of IEEE International Symposium on Information [44] G. Kramer, M. Gastpar, P. Gupta, “Cooperative strategies and capacity
Theory (ISIT), Nice, France, June 2007. theorems for relay networks,” IEEE Transactions on Information Theory,
[19] C. T. K. Ng, D. Gündüz, A. J. Goldsmith and E. Erkip, “Recursive power 51(9):3037-3063, September 2005.
allocation in Gaussian joint source-channel layered broadcast coding,” in [45] S. Draper, G. Wornell, “Side information aware coding strategies for
Proc. of IEEE International Conference on Communications (ICC 2007), sensor networks,” IEEE J. Select. Areas Commun., vol. 22, Aug 2004.
Glasgow, Scotland, June 2007. [46] D. Gündüz, C. T. K. Ng, E. Erkip and A. J. Goldsmith “Source
[20] X. Xu, D. Gündüz, E. Erkip, Y. Wang, “Layered cooperative source transmission over relay channel with correlated relay side information,”
and channel coding,” IEEE International Conference on Communications in Proc. of IEEE International Symposium on Information Theory (ISIT
(2):1200 - 1204, May. 2005. 2007), Nice, France, June 2007.
[21] H. Y. Shutoy, D. Gündüz, E. Erkip and Y. Wang, “Cooperative source [47] D. Gündüz, E. Erkip, “Reliable cooperative source transmission with
and channel coding for wireless multimedia communications,” to appear side information,” IEEE Information Theory Workshop (ITW 2007),
in IEEE Journal of Selected Topics In Signal Processing, special issue Bergen, Norway, July 2007.
on Network Aware Multimedia Processing and Communications, 2007.
[22] A. Kwasinski, Z. Han, K. J. R. Liu, “Cooperative multimedia commu-
nications: Joint source coding and collaboration”, Proceeding of IEEE
Global Telecommunications Conference (GLOBECOM), Saint Louis,
MO, Nov.-Dec. 2005.
[23] A. Kwasinski, Z. Han, K. J. R. Liu, “Joint source coding diversity and
cooperative diversity for multimedia communications”, Proc. of IEEE
Inter. Workshop on Signal Processing Advances for Wireless Comm.
(SPAWC), pp. 129-133, New York, NY, June 2005. Deniz Gündüz [S’02] received the B.S. degree in Electrical and Electronics
[24] A. Kwasinski, K. J. R. Liu, “Optimal unequal error protection with user Engineering from Middle East Technical University, Ankara, Turkey, in 2002
cooperation for transmission of embedded source-coded images” ICIP and the M.S. degree in Electrical and Computer Engineering from Polytechnic
2006, Atlanta, GA, Oct. 2006. University, Brooklyn, NY, in 2004. He is currently working toward his Ph.D.
[25] T. Cover, J. Thomas, Elements of Information Theory, Wiley and Sons, at Polytechnic University.
New York, 1991. In 2004, he was a summer researcher in the laboratory of information
[26] P. Zador, “Topics in the asymptotic quantization of continuous random theory (LTHI) at EPFL in Lausanne, Switzerland. His research interests lie
variables,” Bell Laboratories Technical Memorandum, 1966. in the areas of communications theory and information theory with special
[27] N. Jayant and P. Noll, Digital Coding of Waveforms, Englewood Cliffs, emphasis on joint source-channel coding, cooperative communications and
NJ: Prentice Hall, 1984. cross-layer design.
[28] T. Linder, R. Zamir, “On the asymptotic tightness of the Shannon lower
bound”, IEEE Transactions on Information Theory, pp. 2026-2031, Nov.
1994.
[29] W. Equitz, T. Cover, “Successive refinement of information,” IEEE
Transactions on Information Theory, 37(2): 269-275, Mar. 1991.
[30] L. Lastras, T. Berger, “All sources are nearly successively refinable,”
IEEE Transactions on Information Theory, 47(3): 918-926, March 2000.
[31] L. Zheng, D. Tse, “Diversity and multiplexing: A fundamental tradeoff
in multiple antenna channels,” IEEE Transactions on Information Theory,
Elza Erkip [S’93, M’96, SM’05] received the Ph.D. and M.S. degrees
vol. 49, pp. 1073-96, May 2003.
in Electrical Engineering from Stanford University, and the B.S. degree in
[32] U. Mittal, N. Phamdo, “Hybrid digital-analog joint source-channel codes
Electrical and Electronics Engineering from Middle East Technical University,
for broadcasting and robust communications,” IEEE Transactions on
Turkey. She joined Polytechnic University in Spring 2000, where she is
Information Theory, vol. 48, pp. 1082-1102, May 2002.
currently an Associate Professor of Electrical and Computer Engineering.
[33] H. Coward, T. A. Ramstad, “Quantizer optimization in hybrid digital
Dr. Erkip received the 2004 Communications Society Stephen O. Rice
analog transmission of analog source signals,” in Proc. IEEE Int. Conf.
Paper Prize in the Field of Communications Theory and the NSF CAREER
Acoustics, Speech and Signal Processing, Istanbul, Turkey, Jun. 2000.
award in 2001. She is an Associate Editor of IEEE Transactions on Communi-
[34] M. Skoglund, N. Phamdo, F. Alajaji, “Design and performance of VQ- cations, a Publications Editor of IEEE Transactions on Information Theory and
based hybrid digital-analog joint source–channel codes,” IEEE Trans. Inf. a Guest Editor of IEEE Signal Processing Magazine, Special Issue on Signal
Theory, vol. 48, no. 3, pp. 708720, Mar. 2002. Processing for Multiterminal Communication Systems. She is the Technical
[35] M. Skoglund, N. Phamdo, F. Alajaji, “Hybrid digital-analog source- Area Chair for the ”MIMO Communications and Signal Processing” track of
channel coding for bandwidth compression/expansion,” IEEE Transac- 41st Annual Asilomar Conference on Signals, Systems, and Computers, and
tions on Information Theory, Vol. 52, No. 8, pp. 3757-3763, August the Technical Program Co-Chair of 2006 Communication Theory Workshop.
2006. Her research interests are in wireless communications, information theory and
[36] I. Kozintsev, K. Ramchandran, “Multiresolution joint source-channel communication theory.
coding using embedded constellations for power-constrained time-
varying channels,” In Proc. of IEEE Inter. Conf. on Acoustics, Speech,
and Signal Proc. (ICASSP), Atlanta, GA, May 1996.
[37] S. Shamai, “A broadcast strategy for the Gaussian slowly fading chan-
nel,” Proceedings of ISIT, p. 150, June 1997.