Data Transmission in the Presence of Channel State

Feedback and Outage Probability Constraint

Behrooz Makki Thomas Eriksson
Department of signals and systems Department of signals and systems
Chalmers University of Technology Chalmers University of Technology
Gothenburg, Sweden Gothenburg, Sweden
Email: [email protected] Email: [email protected]

Abstract—Recently, substantial attention has been paid to gain G : [0, ∞) is partitioned into 2N nonoverlapping regions
increase the achievable rates of wireless networks using different Gi = [gi−1 , gi ), i = 1...2N (g0 = 0, g2N = ∞). With the
kinds of limited channel state information (CSI) feedback. This instantaneous channel gain g being in the region Gi , the index
paper aims to investigate the effect of limited CSI feedback on
the achievable rates of power-limited block fading channels con- i is sent back and the transmitter selects a codeword with
strained to have limited outage probability. Considering different power Ti and rate Ri such that the system total performance
power allocation strategies and various fading conditions, the is optimized. Throughput, average rate and outage probability
results are obtained for different outage probabilities and are are some yardsticks that might be considered as optimization
compared with the case where there is no outage limitation. criterion.
Simulation results show that 1) although the outage probability
constraint reduces the channel average rate, its influence dimin- Particularly, in presence of noise-free CSI at the transmitter
ishes as the channel state information feedback or the input and perfect knowledge at the receiver, Liu et. al, [10] and
power increases, 2) the effect of the outage probability constraint Kim and Skoglund [11] have recently obtained the achievable
decreases as the channel fading severity is reduced. rates of power-limited block fading channels using optimal
channel SNR quantization and power allocation. In their work,
I. I NTRODUCTION
in order to maximize the average rate, the channel SNR is
High-rate reliable transmission over wireless channels is an assumed to be constant, equal to its worst case within each
important research problem from both theoretical and practical quantization interval. This is a simple practical approach which
aspects. New generation wireless networks require high-rate results in Gaussian input distributions reaching the maximum
reliable data transmission which, due to unknown channel reliable rates if the channel remains fixed for a long time so
quality variations, is not always possible [1]–[6]. In order that the ergodic capacity is achieved. However, as illustrated
to tackle this problem, there are two different approaches, in the following, by considering the worst case, no data might
namely, automatic repeat request (ARQ) [7], [8] and channel be transmitted in the first quantization region which leads to
state information (CSI) feedback [9]–[13], mainly applied by some unavoidable outage probability. This is the point that
the experts. The ARQ is a well-known approach applied in deteriorates the system efficiency particularly in applications
today networks to increase the transmission reliability in the requiring high reliability in data transmission.
absence of transmitter channel state information. On the other Based on this perspective, the present work studies the effect
hand, as shown in e.g. [9]–[13], having knowledge about the of limited CSI feedback on the achievable rates of power-
channel fading status, it is possible to increase the transmission limited block fading channels constrained to have limited
rates via updating the transmission power and rate relative outage probability. Average rates are obtained for different
to channel quality. The more information available at the power allocation strategies and under different fading condi-
transmitter, the higher rates can be achieved, with the rate tions. Moreover, we will compare our results with the ones
upper bounded by full knowledge channel capacity [13], [14]. obtained in [11]. Simulation results show that although outage
However, due to the signaling overhead caused by reporting probability constraint reduces the channel average rate, its
the channel information, assuming perfect channel knowledge influence diminishes as the channel state information feedback
at the transmitter is an overly optimistic assumption [10]– or the input power increases. We also show that the effect
[12]. This is the main motivation for the present limited CSI of outage probability constraint is higher in severe fading
feedback systems, such as UMTS/WCDMA [15]. conditions while its influence is reduced in the cases where
In a general limited CSI feedback scheme, the receiver the channel fading severity is not too high.
provides the transmitter with some rough measure of channel The rest of the paper is organized as follows. System model
gain, and the transmitter adjusts its rate and power according and problem statement are illustrated in section II. Then, the
to this imperfect information. Rough CSI is normally pro- theoretical concepts are presented in section III. Section IV
duced by channel gain quantization at the receiver where, consists of simulation results and some discussions. Finally,
considering N bits feedback, the set of all possible channel the last section concludes the paper.

978-1-4244-3574-6/10/$25.00 ©2010 IEEE

 g N (0, d 2 ) is successfully decoded, otherwise outage occurs. Therefore,
the probability of successful decoding in each region is
FG (gi ) − FG (gi∗ )
psi = Pr{g ≥ gi∗ |g ∈ [gi−1 , gi )} = , (4)
pi
X × + Y the total rate would be
N
2
 2N
Pr ob{outage} £ Poutage  1
R= pi ps i R i = [FG (gi ) − FG (gi∗ )] log(1 + gi∗ Ti )
EX 2 £ Ttotal
i=1
2 i=1
(5)
Figure 1. Channel model. EX 2 ≤ Ttotal and Pr{outage} ≤ Poutage represent and finally, the outage probability is found by
the power and outage probability constraints, respectively.
N N
2
 2

Pr{outage} = pi (1 − psi ) = (FG (gi∗ ) − FG (gi−1 ))
II. S YSTEM MODEL AND PROBLEM STATEMENT i=1 i=1
(6)
As illustrated in Fig.1, we consider a scalar block fading gi
in which pi = fG (g)dg is the probability of being in
channel in which the input message X multiplied by the gain gi−1
random variable g is summed with an iid Gaussian noise N : the i-th region and FG (g) is the gain cumulative distribution
N (0, δ 2 ) resulting in the output function (cdf).
√ It is common to determine the results obtained under full
Y = gX + N. (1) and no feedback assumptions as two extreme cases. Having
full knowledge, it is well accepted [11], [13] that maximum
The input signal is power limited, i.e. EX 2 ≤ Ttotal , and the
achievable rate can be calculated by
data transmission system is constrained to have limited outage
probability denoted by Pr{outage} ≤ Poutage . The channel gain ∞
1
remains constant for a duration of Tc , generally determined R∞ = fG (g) log(1 + gT (g))dg (7)
2
by the channel coherence time, and then changes indepen- 0
dently according to the fading probability density function
where T (g) is the power allocation function optimally deter-
(pdf) fG (g). It is assumed that there is perfect instantaneous
mined based on the power constraint which can be interpreted
knowledge about the channel gain at the receiver, which is
in two different ways, namely, short- and long-term power
an acceptable assumption under block fading conditions [8].
constraints. The short-term power constraint requires that, for
Further, the feedback channel is supposed to be error-free and,
every single Xi , i = 1, ....2N , the power allocated cannot
with no loss of generality, we consider δ 2 = 1. Similar to
exceed the total input power Ttotal , that is,
the method presented in [11], the receiver can provide the
transmitter with some auxiliary information as follows. Ti ≤ Ttotal , i = 1, ..., 2N (8)
Considering N bits feedback, a quantization function
Under the more relaxed long-term power constraint, the
H(g) = i, g ∈ Gi = [gi−1 , gi ), i = 1...2N , g0 = 0, g2N = ∞ transmitter can adapt the power based on channel conditions
(2) such that the average transmission power does not exceed
is implemented by the receiver and the index i is sent back Ttotal , i.e.,
2N
to the transmitter. Getting i, with no loss of generality, the
channel gain is assumed to be some fixed value gi∗ ∈ [gi−1 , gi ). pi Ti ≤ Ttotal . (9)
Therefore, the transmission rate is selected to be1 i=1

Finally, note that under the full knowledge assumption, the

Ri = max I(Xi ; Yi |g = gi∗ ) transmitted codewords are always decoded successfully and
fXi (x)
 (a) 1
there is no outage probability. On the other hand, with no
= h( gi∗ Xi + N ) − h(N ) = log(1 + gi∗ Ti ) (3) information about the channel realization g, it is selected as
2
some fixed value g ∗ and data transmission is done at rate R0 =
where I(U ; V ) denotes the mutual information between two 1 ∗
2 log(1 + g Ttotal ). The data is successfully decoded if g ≥
random variables U and V , h(U ) is the differential entropy of ∗
g . Consequently, the outage probability constraint changes to
the random variable U , Ti = EXi2 is the input power allocated FG (g ∗ ) ≤ Poutage and the maximum rate is obtained by
to the i-th input distribution and (a) follows from the fact that,
for a fixed power EYi2 = gi∗ Ti + δ 2 , the Gaussian distribution  1
maximizes the differential entropy. If the instantaneous gain R= max { (1 − FG (g ∗ )) log(1 + g ∗ Ttotal )} (10)
g ∗ ≤FG−1
(Poutage ) 2
realization supports the rate, i.e. g ≥ gi∗ , the transmitted data
where, for every given Ttotal , maximization is done over all
1 All results are presented in natural logarithm basis. values of g ∗ which satisfy the outage probability constraint.
Considering different fading conditions and the two mentioned strategy, and so can be implemented for both short- and long-
power constraints, the effect of limited CSI feedback on the term power constraints.
achievable rates of block fading channels constrained to have While short-term power constraint leads to Ti = Ttotal , under
limited outage probability is studied in the following. long-term power constraint, the optimal power allocation can
be found by
III. T HEORETICAL CONCEPTS
1
Considering N bits feedback, long-term power constraint ∂θα Ti + gi−1 = −1
λ , i>1
=0⇒ 1 −ϕ1 (15)
and limited outage probability, the rate optimization problem ∂Ti T1 + g∗ = λp1 , i = 1
1
can be expressed as
N which, due to the fact that the powers cannot be negative, is the

2
maximize { [FG (gi ) − FG (gi∗ )] log(1 + gi∗ Ti )} famous water filling problem [13]. The Lagrange multiplier λ
{gi ,gi∗ ,Ti ,α>0}
⎧ i=1 is obtained according to (9).
⎪
⎪ 2N
 Finally, considering ∂θ α
⎪
⎪ pi Ti ≤ Ttotal ∂gi , optimal quantization boundaries
⎪
⎨ i=1 are determined according to
subject to &
⎪
⎪ log(1 + g1∗ T1 ) − log(1 + g1 T2 )
⎪
⎪ 2N

⎪
⎩ (FG (gi∗ ) − FG (gi−1 )) = α ≤ Poutage p2 T2
i=1 + + λ(T1 − T2 ) = 0 (16a)
(11) fG (g1 ) 1 + g1 T2
which, for every given α, can be represented as pi+1 Ti+1
log(1 + gi−1 Ti ) − log(1 + gi Ti+1 ) +
2N fG (gi ) 1 + gi Ti+1

θα = [FG (gi ) − FG (gi∗ )] log(1 + gi∗ Ti ) + λ(Ti − Ti+1 ) = 0, g1∗ = FG−1 (α) ≤ FG−1 (Poutage ), i > 1
i=1 (16b)
N N
2
 2
 which, along with the power constraint condition, can be
+λ [FG (gi ) − FG (gi−1 )]Ti + β (FG (gi∗ ) − FG (gi−1 ))
implemented to find the maximum average rate. However,
i=1 i=1
(12) since solving such nonlinear water filling-based equations are
difficult, we propose an iterative algorithm instead:
Here, λ and β are the Lagrange multiplier coefficients deter-
mined based on the power and outage probability constraints, Algorithm 1 Average rate optimization
respectively, and α is any positive value less than Poutage .
I. For a given power Ttotal and an outage probability Poutage ,
Taking the derivative with respect to gi ∈ [gi∗ , gi+1
∗
), i ≥ 1, it
consider L, e.g. L = 20, randomly generated gain vectors
can be written ∗ ∗
ḡl = [g0,l g1,l g1,l ...g2N ,l ], g0,l = 0 < g1,l < g1,l <
∂θα ∗ −1
... < g2N ,l = ∞, g1,l ≤ FG (Poutage ) ∀l = 1...L.
= fG (gi ){log(1 + gi∗ Ti ) + λ(Ti − Ti+1 ) − β} (13)
∂gi ∗
II. For each vector, considering gi = gi+1 , i ≥ 1, find the

whose sign does not depend on the value of gi . Therefore, the optimal power allocation and average rate Rl based on
optimal average rate is found by either gi = gi∗ or gi = gi+1∗
. (5), (9) and (15).
∗
However, note that, based on (5), considering gi = gi the III. Determine the vector which results in highest average
 
i-th region has no effect on the average rate and so, it can rate, i.e. ḡn where ∀l = 1...L, Rl ≤ Rn .
be merged with the neighboring regions. Therefore, for i ≥ 1 IV. ḡ1 ← ḡn .
∗
and any 0 ≤ α ≤ Poutage , we have gi = gi+1 . In this way, the V. Generate Q L, e.g. Q = 5, gain vectors ḡknew , k =
outage probability constraint changes to FG (g1∗ ) ≤ Poutage and 1...Q, around ḡ1 . These vectors should also satisfy the
the optimization criterion is simplified to constraints introduced in I.
2

N
VI. ḡk+1 ← ḡknew , k = 1...Q.
θα = pi log(1 + gi−1 Ti ) + ϕ1 log(1 + g1∗ T1 ) VII. Regenerate the remaining ḡl , l = Q + 2, ..., L randomly
∗
i=2 such that g0,l = 0 < g1,l < g1,l < ... < g2N ,l =
N ∗ −1
2
 ∞, g1,l ≤ FG (Poutage ) ∀l = Q + 2, ..., L.
+λ pi Ti , g1∗ = FG−1 (α) ≤ FG−1 (Poutage ), VII. Go to II and continue until convergence.
i=1
ϕ1 = [FG (g1 ) − FG (g1∗ )] (14)
Although time-consuming, the proposed algorithm has been
which simply means that, in order to have maximum aver- shown to be efficient in complex optimization problems deal-
age rate under power and outage probability constraints, the ing with local minima issues [16]. Moreover, note that, the
channel gain should be assumed to be constant, equal to its algorithm has been illustrated for long-term power constraint
worst case within each quantization interval except the first while, as mentioned before, having short-term power con-
one and the outage just happens if g < g1∗ . Finally, note straint the power allocation strategy changes to Ti = Ttotal .
that the argument is valid independent of the power allocation Simulation results are presented in the following.
 1.8 1.8

1.6 1.6

1.4 1.4

1.2 1.2
Average rate

Average rate
1 1

0.8 0.8
Short−term, Poutage=0.2
0.6 0.6 Long−term, P =0.2
outage
Long−term, Poutage=0.2
Short−term, P =0.2
0.4 0.4 outage
Full knowledge, long−term
Full knowledge, long−term
0.2 No knowledge, P =0.2 0.2
outage No knowledge P =0.2
outage
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Input power Input power

Figure 2. Average rate obtained by N = 3 bits channel quality feedback. Figure 3. Average rate obtained by N = 3 bits channel quality feedback.
Outage probability is constrained to be Pr{outage} ≤ 0.2. Gain distribution: Outage probability is constrained to be Pr{outage} ≤ 0.2. Gain distribution:
log-normal, γ = 1, η = 0. Nakagami, μ = 2, w = 1.

IV. S IMULATION RESULTS 1.5

Two log-normal and Nakagami channel gain distributions

are considered to evaluate the results. The parameters of the
distributions are set as
Log-normal: 1
Average rate

1 −
(log(g)−η)2
fG (g) =  e 2γ 2 , g ≥ 0, γ = 1, η = 0 (17)
g 2πγ 2
Nakagami:
0.5
2μμ μ 2 2−bits feedback, no outage constraint [11]
fG (g) = g 2μ−1 e− w g , g ≥ 0, μ = 2, w = 1 (18) 2−bits feedback, Poutage=0.1
Γ(μ)wμ
No knowledge P =0.1
outage
where Γ(.) denotes the standard Gamma function. As men- No knowledge, no outage constraint [11]
tioned in [13], the Log-normal distribution can model severe 0
0 5 10 15 20 25 30
fading conditions, while Nakagami distributions can be used Input power
to model the cases in which fading severity is not too high.
Considering N = 3 bits feedback and the two mentioned
Figure 4. Comparison between the results obtained with or without outage
power allocation strategies, the average rates have been found probability constraint, N = 0 and 2 bits feedback, gain distribution: log-
for the two gain distributions, as illustrated in Fig.2 and normal, γ = 1, η = 0.
3. Outage probability constraint have been selected to be
Pr{outage} ≤ 0.2. Moreover, the results obtained under no
knowledge and long-term full knowledge assumptions have fading severity decreases the effect of feedback informa-
been plotted as two lower and upper bounds, respectively. The tion and adaptive transmission is also reduced (Fig.3).
effect of outage probability constraint can be further studied in • Although optimal power allocation can increase the aver-
Fig.4 and 5; Considering N = 0 and 2 bits feedback and log- age rate, the increment is negligible particularly at high
normal gain distribution, Fig.4 compares the rates found under SNR’s (Fig.2 and 3).
Pr{outage} ≤ 0.1 constraint with the ones obtained when there • Outage probability constraint effect is higher in severe
is no outage probability constraint [11]. Furthermore, having fading conditions (Fig.5, Log-normal pdf’s) while its
fixed input power, Fig.5 demonstrates the effect of outage effect is reduced in the cases where the fading severity is
probability constraint on the average rates of the channel not too high (Fig.5, Nakagami pdf’s). Moreover, although
under no knowledge, i.e. (10), and N = 2 bits CSI feedback hard constraints, i.e. small Poutage in Fig.5, leads to
conditions. significant rate reduction, its influence decreases under
The results emphasize on the facts that : more relaxed constraints.
• While channel side information can result in significant • Increasing the number of bits or the input power, the
rate increment in severe fading conditions (Fig.2), as the effect of outage probability constraint diminishes and it
 [8] C. Shen, T. Liu, and M. P. Fitz, “On the average rate performance
of hybrid-ARQ in quasi-static fading channels,” IEEE Transactions on
Communications, vol. 57, no. 11, pp. 3339–3352, Nov. 2009.
0.25 [9] S. Tatikonda and S. Mitter, “The capacity of channels with feedback,”
IEEE Transactions on Information Theory, vol. 55, no. 1, pp. 323–349,
Jan. 2009.
0.2 [10] X. Liu, H. Yang, W. Guo, and D. Yang, “Capacity of fading channels
with quantized channel side information,” IEICE Transactions on Com-
Average rate

munications, vol. E89-B(2), pp. 590–593, 2006.

0.15 [11] T. T. Kim and M. Skoglund, “On the expected rate of slowly fading
channels with quantized side information,” IEEE Transactions on Com-
munications, vol. 55, no. 4, pp. 820–829, April 2007.
0.1 [12] S. Ekbatani, F. Etemadi, and H. Jafarkhani, “Transmission over slowly
Nakagami, no knowledge, input power=1 fading channels using unreliable quantized feedback,” in Data Compres-
Log−normal, no knowledge, input power=1 sion Conference, March 2007, pp. 353–362.
Log−normal, 2−bits feedback, input power=1 [13] A. J. Goldsmith and P. P. Varaiya, “Capacity of fading channels with
0.05
Nakagami, 2−bits feedback, input power=1 channel side information,” IEEE Transactions on Information Theory,
vol. 43, no. 6, pp. 1986–1992, Nov 1997.
[14] T. M. Cover and M. Chiang, “Duality between channel capacity and
0 0.1 0.2 0.3 0.4 0.5
P rate distortion with two-sided state information,” IEEE Transactions on
outage Communications, vol. 48, no. 6, pp. 1629–1638, June. 2002.
[15] “Technical specification group radio access network, physical layer
Figure 5. Average rate versus outage probability constraint, N = 0 and 2 procedures,” 3rd Generation Partnership Project, Tech. Spec. 3GPP ts
bits channel state information feedback, input power=1. 25.214 V3.10.0, 2002-2003.
[16] B. Makki, M. Noori Hosseini, S. A. Seyyedsalehi, and N. Sa-
dati, “Unaligned training for voice conversion based on a lo-
cal nonlinear principal component analysis approach,” Neural
is possible to reach at (almost) the same performance as computing and applications, Nov. 2009, in press, available at
the ones obtained with no outage probability limitation http://www.springerlink.com/content/n70783321l54jq73/.
(Fig.4). Particularly, in the asymptotic case, i.e. N → ∞,
both rates converge to full knowledge channel capacity.
V. C ONCLUSION
The goal of this paper is to investigate the effect of partial
channel state information feedback on the achievable rates
of power-limited block fading channels constrained to have
limited outage probability. Considering short- and long-term
power allocation strategies and various fading conditions, the
results are obtained for different outage probabilities. Simu-
lation results showed that although this new constraint plays
a great role in severe fading conditions, its effect becomes
negligible as the channel fading severity is reduced. Moreover,
we showed that the effect of outage probability constraint is
reduced as the input power or the channel state information is
increased.
R EFERENCES
[1] B. Makki and T. Eriksson, “Efficient channel quality feedback signaling
using transform coding and bit allocation,” in IEEE 71st Vehicular
Technology Conference, May 2010, accepted.
[2] T. Eriksson and T. Ottosson, “Compression of feedback for adaptive
transmission and scheduling,” Proceedings of the IEEE, vol. 95, no. 12,
pp. 2314–2321, Dec. 2007.
[3] ——, “Compression of feedback in adaptive OFDM-based systems
using scheduling,” Communications Letters, IEEE, vol. 11, no. 11, pp.
859–861, Nov. 2007.
[4] S. Sanayei and A. Nosratinia, “Exploiting multiuser diversity with only
1-bit feedback,” WCNC, vol. 2, pp. 978–983, March 2005.
[5] R1-060018, “Performance evaluation of EUTRA downlink CQI feed-
back schemes,” Motorola, RAN1 meeting 44, Helsinki, Finland, Jan.
2006.
[6] M. Johansson, “Diversity-enhanced equal access-considerable through-
put gains with 1-bit feedback,” IEEE 5th Workshop on Signal Processing
Advances in Wireless Communications, pp. 6–10, July 2004.
[7] J. Perret and D. Tuninetti, “Repetition protocols for block fading
channels that combine transmission requests and state information,”
in IEEE International Conference on Communications, May 2008, pp.
1297–1301.