Professional Documents
Culture Documents
Lossy Transmission of Correlated Sources Over Two
Lossy Transmission of Correlated Sources Over Two
Lossy Transmission of Correlated Sources Over Two
net/publication/325008932
CITATIONS READS
0 23
3 authors, including:
Some of the authors of this publication are also working on these related projects:
Performance Analysis of Linear Block Codes over the Queue-Based Channel View project
All content following this page was uploaded by Jian-Jia Weng on 14 May 2018.
Abstract—Achievability and converse results for the lossy these results were extended to noisy TWCs by also adopting
transmission of correlated sources over Shannon’s two-way the interactive protocol of [7]. To date, the performance of
channels (TWCs) are presented. A joint source-channel coding lossy transmission under Shannon’s set-up of simultaneous
theorem for independent sources and TWCs for which adaptation
cannot enlarge the capacity region is also established. We further user transmissions is not fully known.
investigate the optimality of scalar coding for TWCs with discrete In the first part of this paper, we establish achievability and
modulo additive noise as well as additive white Gaussian noise. converse theorems for the lossy transmission of two correlated
Comparing the distortion of scalar coding with the derived sources over TWCs under Shannon’s scenario. A TWC is
arXiv:1805.01974v1 [cs.IT] 5 May 2018
bounds, we observe that scalar coding achieves the minimum viewed as two one-way channels with associated states, and
distortion over both families of TWCs for independent and
uniformly distributed sources and independent Gaussian sources. two-dimensional distortion regions for the TWC are derived.
For independent sources and for TWCs whose capacity region
are not enlarged by adaptation coding, we further find that the
Index Terms—Two-way channels, lossy joint source-channel
coding, source-channel separation, uncoded transmission. achievability and converse parts are matched, resulting in a
complete joint source-channel coding theorem that shows that
the TWC system can be treated as two parallel one-way sys-
I. I NTRODUCTION tems with separate source and channel coding. Based on these
Two-way channels (TWCs) were first introduced by Shan- results, we investigate the performance of scalar coding for
non in [1]. Transmission over such channels makes the best two important classes of additive-noise TWCs: q-ary discrete
use of channel resources since two users can exchange their additive-noise TWCs [6] and additive white Gaussian noise
source data on the same frequency band simultaneously. Also, (AWGN) TWCs [4]. For these channels, adaptation coding
as the channel inputs of both users can be generated by in- does not enlarge the capacity region. Scalar coding (also
teractively adapting to the past received signals, the receivers’ known as single-letter coding or uncoded transmission [9]) is
knowledge about the transmitted source can be refined, which particularly interesting because it is the simplest one among
may increase the rate of successful data recovery. Apart all possible transmission schemes. We analyze the distortion
from the (asymptotically) lossless transmission considered by incurred by the scalar coding scheme and compare it with the
Shannon, it is also natural to investigate the lossy counterpart distortion lower bounds obtained from the converse theorem.
in which data reconstruction is allowed within a tolerable As expected, it is observed that scalar coding is sub-optimal
distortion. Such coding schemes may play an important role with a performance deteriorating as the correlation between
in improving the efficiency of data transmission over resource- the two sources increases. However, when the two sources
limited networks. In this paper, we investigate the performance are independent, we show that scalar coding is optimal for
of lossy transmission over noisy TWCs. uniform sources under the Hamming distortion measure over
In the literature, TWCs have been studied from different the discrete additive TWC and for Gaussian sources under
points of view. For lossless transmission, TWCs are either the squared error distortion measure over the AWGN-TWC.
viewed as a part of the multiple access channels (MACs) with These results are extensions of their well-known counterparts
feedback [2] or related to the compound MACs with correlated for one-way point-to-point systems (e.g., see [9] and references
side information at the receiver [3]. Inner and outer bounds for therein).
the transmission rate over noisy TWCs were derived based on The paper is organized as follows. In Section II, the system
these channel models. Furthermore, from the channel capacity model of TWCs is introduced. Achievability and converse
perspective, it has been found that adaptation is not always results on lossy transmission over TWCs are also presented.
useful [4]-[6]. In contrast to lossless transmission, studies In Section III, the performance of lossy transmission over
regarding the lossy counterpart are limited. The first lossy TWCs with discrete additive noise is investigated. Analogous
transmission problem over error-free TWCs appeared in [7], in results are obtained for the AWGN-TWC system in Section IV.
which only one user can use the channel at each time instant. Finally, concluding remarks are given in Section V.
Interactive source coding for noiseless TWCs was considered
II. L OSSY T RANSMISSION OVER T WO - WAY C HANNELS
to establish a rate distortion (RD) region of TWC. In [8],
A. System Model
The authors are with the Department of Mathematics and Statis- For i = 1, 2, let Ui , Xi , and Yi denote random variables
tics, Queen’s University, Kingston, ON K7L 3N6, Canada (email: jian-
jia.weng@queensu.ca, fa@queensu.ca, linder@mast.queensu.ca). corresponding to terminal i’s source, channel input, and chan-
This work was supported in part by NSERC of Canada. nel output, respectively. Let Ui , Xi , and Yi respectively denote
U1K X1N X2N U2K (fi1 , fi2 , . . . , fiN ). The associatedP
expected distortion is given
K
T1 TWC T2 by Dfi , E[d(UiK , ÛiK )] = K −1 n=1 E[d(U1n , Û1n )]. The
Û2K Y1N Y2N Û1K code pairs (f1 , g2 ) and (f2 , g1 ) then form an overall coding
scheme for the two-way source-channel (TWSC) system.
Fig. 1: The block diagram of two-way communications.
Definition 1. A distortion pair (D1 , D2 ) is achievable at rate
r for the TWSC system if there exists a sequence of encoding
their alphabets. We consider a correlated source transmission and decoding functions with limK→∞ K/NK = r such that
problem over Shannon’s TWCs as shown in Fig. 1, where two lim supK→∞ E[d(UiK , ÛiK )] ≤ Di for i = 1, 2.
terminals want to exchange the correlated sources U1 and U2 Definition 2. The distortion region of a rate-r TWSC system
within desired distortions via a memoryless and noisy TWC is defined as the convex closure of the set of all achievable
governed by the channel input-output transition probability distortion pairs.
p(y1 , y2 |x1 , x2 ). One special feature of such transmission
is that channel inputs can be generated by adapting to the
B. Necessary and Sufficient Lossy Transmission Conditions
previously received signals, which may improve the quality
of reconstruction. Let Ri denote the channel coding rate of In [3, Section VIII], achievability and converse theorems for
terminal i for i = 1, 2. In [1], Shannon derived inner and outer (asymptotically) lossless transmission of correlated sources in
bounds for the capacity region of the TWC. Both bounds are of the sense that limK→∞ p(U K 6= Û K ) = 0 are derived. Here,
the same form but with different input distribution restrictions. we present a similar result for lossy transmission in which
In particular, the bounds are of the form: nonzero expected single-letter distortion is considered. We take
Shannon’s viewpoint: a TWC can be viewed as two one-way
R1 ≤ I(X1 ; Y2 |X2 ), channels with state variables and these states are known at the
(1)
R2 ≤ I(X2 ; Y1 |X1 ), receiver but not at the transmitter [1]. For transmitting over
where I(Xi ; Yj |Xj ) denotes conditional mutual information, such a one-way channel, it is natural to treat the correlated
and X1 and X2 are independent inputs in the inner bound, source at the receiver as side information for the source at the
while in the outer bound X1 and X2 are arbitrarily correlated. transmitter as in the setting of the Wyner-Ziv coding problem
(i)
Note that the inner bound is proved by a standard coding [11]. Following this perspective, let R(i) (D) and RWZ (D)
scheme which does not use adaptation. A channel symmetry denote the standard and Wyner-Ziv RD functions of Ui for
condition for which the two regions coincide was also provided i = 1, 2, which are respectively given by (e.g., [10])
in [1, Sec. 12]. This result shows that adaptation coding cannot R(i) (D) = min I(Ui ; Ûi )
enlarge the capacity region of symmetric TWCs. p(ûi |ui ):E[d(Ui ,Ûi )]≤D
For a positive integer blocklength K, let UiK ,
and
(Ui1 , Ui2 , . . . , UiK ) denote the source sequence of terminal i,
(i)
i = 1, 2. Here, U1K and U2K are Q distributed according to the RWZ (D) = min min I(Ui ; W |Uj )
K p(w|ui ) gj :Uj ×W →Ui
product probability distribution m=1 ps (u1m , u2m ), where E[d(Ui ,gj (Uj ,W ))]≤D