Anda di halaman 1dari 10

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO.

2, FEBRUARY 2012

411

On the Minimum Differential Feedback for Time-Correlated MIMO Rayleigh Block-Fading Channels
Leiming Zhang, Student Member, IEEE, Lingyang Song, Member, IEEE, Meng Ma, and Bingli Jiao, Member, IEEE

AbstractIn this paper, we investigate the differential channel state information (CSI) feedback problem for a general multiple input multiple output (MIMO) system over time-correlated Rayleigh block-fading channels. Specically, we derive the closedform minimum differential feedback rate in the presence of channel estimation errors and quantization distortion. With the feedback-channel capacity constraint, we further study the ergodic capacity in a periodic feedback system in terms of the minimum differential feedback rate and the feedback interval. Through theoretical analysis, we nd that there exists an optimal differential feedback interval to achieve the maximum ergodic capacity. Analytical results are veried by simulations in a practical periodic differential feedback system employing waterlling precoder and Lloyds quantization algorithm. Index TermsMIMO, differential feedback rate, temporal correlation, feedback interval, ergodic capacity.

I. I NTRODUCTION

HANNEL state information (CSI) feedback problems have been intensively studied due to its potential benets to the multiple input multiple output (MIMO) systems [1]. CSI can be utilized by a variety of channel adaptive techniques (e.g., water-lling, beamforming, zero-forcing, interference alignment, etc.) at the transmitter side to enhance the spectral efciency as well as the robustness, especially for systems operating in the frequency division duplexing (FDD) mode. As the capacity of the feedback channel is normally limited, the innite feedback of CSI is hard to realize in practice. Therefore, it is important to investigate how to decrease the amount of feedback signalling overhead to meet the uplink channel constraint. As a result, CSI feedback reduction has attracted lots of attention in recent years [1], [2]. Specically, when the wireless channel experiences timecorrelated fading [3], typically represented by a Markov

Paper approved by D. J. Love, the Editor for MIMO and Adaptive Techniques of the IEEE Communications Society. Manuscript received July 28, 2010; revised April 24, 2011 and August 11, 2011. This work was partially supported by the National Natural Science Foundation of China 60972009 and 61061130561, the National Science and Technology Major Projects of China 2009ZX03003-011, 2010ZX03005-003 and 2009ZX03003-001, and by the Specialized Research Fund for the Doctoral Program of Higher Education of China 20100001120026. The materials in this paper have been presented in part at the IEEE Globecom Conference, Dec. 2010. The authors are with the State Key Laboratory of Advanced Optical Communication Systems and Networks, School of Electronics Engineering and Computer Science, Peking University, Beijing, P. R. China, 100871 (email: {leiming.zhang, lingyang.song, mam, jiaobl}@pku.edu.cn). Digital Object Identier 10.1109/TCOMM.2012.011311.100455

random process [4][6], the amount of CSI feedback can be signicantly reduced. In [7], a number of feedback reduction schemes were reviewed, considering the lossy compression scheme exploiting the properties of fading process as the best choice. In [8] and [9], switched codebook and rotation codebook schemes were proposed with differential feedback. In [10], the time-correlated fading channel was modeled as a nite-state Markov chain, and the feedback rate was reduced by ignoring the states occurring with small probabilities. In [11] and [12], a predictive vector quantization scheme was proposed, assuming that the previous quantization CSI was known. In [13], variable-length codes were applied to decrease the feedback rate. Despite so much research on practical feedback reduction schemes, to the best knowledge of the authors, the lower bound of the feedback compression as well as the required minimum differential feedback rate to guarantee the accuracy of quantization CSI have not yet been well studied for time-correlated MIMO Rayleigh block-fading channels. In order to optimize the feedback design, the relationship between the system capacity and the feedback rate of CSI was studied in [14][17]. Lower and upper bounds of feedback rate that obtain positive capacity gain in comparison with openloop systems were reported in [18] and [19]. The optimal feedback rate in a memoryless periodic feedback scheme was studied in [20] when the CSI is fed back independently at each time. However, the differential feedback with temporal correlation is not taken into account. Without loss of generality, in this paper, we consider a general point-to-point MIMO system with periodic differential CSI feedback over timecorrelated Rayleigh block-fading channels, and investigate the relationship between the capacity and the differential feedback rate with feedback channel capacity constraint. The main contributions of the present paper can be briey summarized as follows: 1) We derive the minimum differential feedback rate expression for time-correlated MIMO Rayleigh blockfading channels by taking into account both the channel estimation errors and the channel quantization distortion, which are assumed as independent and identically distributed (i.i.d.) complex Gaussian variables. 2) We investigate the relationship between the ergodic capacity and the feedback interval with feedback channel capacity constraint in a periodic differential feedback

c 0090-6778/12$31.00 2012 IEEE

412

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO. 2, FEBRUARY 2012

system. We also prove that there exists an optimal feedback interval to achieve the maximum ergodic capacity. 3) We present the approximate optimal feedback intervals, and verify the theoretical results in a practical differential feedback system using water-lling precoder and Lloyds quantization algorithm. The rest of the paper is organized as follows. In Section II, we describe the system model. In Section III, the minimum differential feedback rate is derived, and the relationship between the ergodic capacity and differential feedback interval is studied. In Section IV, we provide some simulation results. In Section V, we draw the main conclusions. The derivations are given in the appendices. Notation: Bold uppercase (lowercase) letters denote matrices (vectors), () , ()+ and det() represent transpose, Hermitian transpose and determinant operator, respectively, log() stands for the logarithm function, represents Frobenius norm, Round () denotes a function that rounds to the nearest integer and [] is used for expectation over random variables. II. S YSTEM M ODEL The system model is illustrated in Fig. 1, where the downlink channel is modeled as a time-correlated MIMO Rayleigh block-fading channel, and the feedback channel is modeled as a capacity-limited and error-free channel. In this paper, we consider a differential feedback scheme, i.e., the receiver only feeds back the differential CSI to the transmitter when the previous quantization CSI is known at both transmitter and receiver. We also take the channel estimation errors and channel quantization distortion into account. A. Time-Correlated MIMO Rayleigh Block-Fading Channel Model We consider MIMO Rayleigh block fading channels, where the channel fading matrix remains constant within a fading block and varies from one to another. There are transmit antennas and receive antennas. The received signals can be expressed in a vector form y = Hx + n0 , (1) where y = [1 , 2 , , ] denotes a 1 received signal vector, H is a channel fading matrix with independent 2 entries obeying complex Gaussian distribution (0, ), x = [1 , 2 , , ] represents a 1 transmitted signal vector, and n0 is a 1 noise vector whose entries are i.i.d. 2 complex Gaussian variables satisfying (0, 0 ). The time-correlated channel can be represented by the rstorder Autoregressive model (AR1) [6], and the channel fading matrix can be written as (2) H = H1 + 1 2 W , where H denotes the -th channel fading matrix, W is a noise matrix, which is independent of H1 , and the 2 entries are i.i.d. complex Gaussian variables with (0, ). The parameter is time correlation coefcient, which is given by the zero-order Bessel function of rst kind, i.e., = 0 (2 ) [3], where stands for the maximum

Doppler frequency and denotes the time interval between consecutive feedback messages. In the block-fading system, the feedback interval can be calculated as = / , where represents the duration of every block. The CSI is estimated by the receiver using orthogonal pilots. Without loss of generality, in this paper, the maximum likelihood (ML) criterion is employed to perform channel estimation, and the resulting estimated channel matrix can be expressed in an equivalent form as [23] H = H + He , (3) where H denotes the channel estimation matrix, whose entries are i.i.d. complex Gaussian variables following a distribution 2 (0, ), H is the actual channel fading matrix, He stands for the channel estimation error matrix, which is independent of H, with i.i.d. entries following a complex Gaussian dis2 2 2 2 tribution (0, ), and = . With ML channel estimation, the estimation error He can be decomposed into two terms as ) ( 2 (4) He = 1 2 H, with =H
2 2 H,

(5)

where is independent of H)with the entries following a ( 2 2 2 ( ) distribution 0, . The detailed proof is given 2 in Appendix A. Thus, from (5), we have H=
2 2 H + .

(6)

B. CSI Feedback Model We consider a limited and error-free feedback channel. Through CSI quantization, we can get [18] H = H + E, (7) where H represents the feedback channel output, E denotes an independent additive noise matrix with i.i.d. complex Gaussian entries following a distribution (0, ), and represents the average channel quantization distortion constraint. From (7), the quantized CSI can be rewritten as H=
2 2 H + ,

(8)

( where is independent of H and the entries) i.i.d. comare 2 ( ) plex Gaussian variables with 0, 2 . The proof is similar to Appendix A. In this paper, we consider the differential feedback, where only the differential CSI will be sent back to the transmitter, assuming that the previous channel quantization matrix H1 is known at both the receiver and the transmitter. The differential CSI can be formulated as ) ( (9) H = Di H , H1 , where H represents the differential CSI between H and 1 , and Di() denotes the differential function. H

ZHANG et al.: ON THE MINIMUM DIFFERENTIAL FEEDBACK FOR TIME-CORRELATED MIMO RAYLEIGH BLOCK-FADING CHANNELS

413

S'
H

Hn | Hn

cd

Hd

Fig. 1.

System model of the differential feedback over time-correlated MIMO Rayleigh block-fading channels.

Furthermore, we assume that the CSI feedback channel has a capacity constraint per fading block. When the CSI is quantized to bits and the feedback interval is set as blocks, the average feedback rate satises the inequality / . Therefore, the feedback interval can be calculated by = , (10) where denotes the smallest integer larger than . C. Ergodic Capacity of Pilot-assisted MIMO Systems At the transmitter, the feedback CSI can be used for precoding (e.g., water-lling, zero-forcing, beam-forming, interference alignment [21], [22], etc.) to improve the system performance. In this subsection, we utilize the water-lling precoder, which is a common and well-known power allocation scheme to achieve the capacity. However, the analysis and conclusions are also valid for other precoder designs. With water-lling precoder, the channel quantized CSI is decomposed using singular value decomposition (SVD) at the transmitter as (11) H = UV+ , where U and V represent two unitary matrixes, and is a non-negative and diagonal matrix composed of eigenvalues of H. For the pilot-assisted MIMO system with ML channel estimation, the closed-loop ergodic capacity employing waterlling algorithm can be obtained with the help of [20],[23] [ ] ( ( )) = H,H log det I + J J+ F1 , (12) 1 where J = ] HVZ, Je = He VZ, F = 2 I + [ Je Je Je + J , denotes the number of transmitted symbols, represents the amplitude of signal symbol, and Z stands for a diagonal matrix determined by the water-lling algorithm, which is given by [20] { )1 2 2 ( 2 , 2 , , 1 2 = 0, otherwise (13) 2 2 = 2 ,
=1

where , , = 1, 2, , , are entries of , is a cut-off value chosen to meet the power constraint, and 2 denotes the power constraint. It can be observed from (7) and (12) that the closed-loop ergodic capacity is determined by H and H, and the loss of the capacity is mainly caused by the distortion. Hence, the ergodic capacity is a negative-correlated function in conjunction with the distortion of CSI feedback [14], [17]. III. M INIMUM D IFFERENTIAL F EEDBACK R ATE In this section, we derive the minimum differential feedback rate of the time-correlated MIMO Rayleigh block-fading channels to guarantee the accuracy of CSI. The minimum differential feedback rate is determined by the Shannons Rate Distortion Theory of continuous-amplitude sources [25]. When the ( 1)-th channel quantization matrix H1 is known at both the receiver and the transmitter, the minimum differential feedback rate can be expressed as ) [ ( )] } { ( = inf H ; H H1 : H ; H , (14) ) ( ; H H1 where inf {} denotes inmum function, H represents the mutual information between H and H given ) ( 2 H1 , and H , H = H H is the channel quantization distortion, which is the measurement of the CSI loss. Since the entries of H, H and H are i.i.d. complex Gaussian variables, the minimum differential feedback rate can be rewritten as { ( ) [ ( )] } = inf ; 1 : ; , (15) where = represents the average quantization distor tion, and , , and 1 denote the entries of H , H , 1 , respectively. and H Lemma 1: If the estimation channel coefcient and the are independent and idenquantization channel coefcient tically distributed complex Gaussian variables, given an average quantization distortion constraint and the ( 1)-th

414

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO. 2, FEBRUARY 2012

channel quantization coefcient 1 , the mutual information ) ( ; 1 can be calculated as ( ) ( )] ) ( 2 2 2 2 1 2 2 2 2 + 1 + 2 log + , 2

(16) 2 2 where and denote the variances of and , respectively, and is the time correlation coefcient. The proof of Lemma 1 is based on Shannons Rate Distortion Theory of continuous amplitude sources, and the details can be found in Appendix B. As , and 1 are complex Gaussian variables, the minimum value of the mutual information is indeed achievable [25]. By combining (15) and (16), the minimum differential feedback rate of the time-correlated MIMO block-fading channels is given by ( ) ) ( 2 2 1 2 2 2 + = log 2 ( )] 2 2 2 1 + 2 + . (17) From (17), we nd that the minimum differential feedback rate is a function of the quantization distortion, time correlation coefcient, and the channel estimation variance. Note that the minimum differential feedback rate in (17) is the lower bound of feedback compression with temporal correlation in the block-fading MIMO channels. Given the accuracy constraint of feedback CSI (i.e., the distortion ), the minimum feedback rate can be easily obtained from (17). Furthermore, as the ergodic capacity increases when the distortion decreases, we can investigate the feedback design scheme by minimizing the distortion of the feedback CSI in order to maximize the ergodic capacity. It is obvious that the CSI distortion in the differential feedback model is relevant to the feedback rate and temporal correlation. In addition, large feedback rate and high temporal correlation could reduce the distortion, and thus improve the ergodic capacity. Given the feedback channel capacity constraint in the periodic feedback system, indicated by (2) and (10), both the feedback rate and temporal correlation are related to the feedback interval. When the feedback interval increases, the feedback rate increases but temporal correlation decreases. Intuitively, there may exist an optimal feedback interval to realize the maximum ergodic capacity by minimizing the distortion of CSI. In the following, we investigate the relationship between the distortion of CSI and feedback interval, and prove the existence of the optimal feedback interval. From (17), can be calculated as ( 2 )2 1 2 2 2 (18) = ( 2 )2 . 2 2 2

we consider the impact of the feedback delay on the distortion. Combining (18) and (38) in Appendix B, the distortion is given by ( 2 )2 ( ) ( )2 2 2 2 2 2 2 2 2 2 =2 ( 2 )2 + 2 2 2 2 2 ) ) 2 ( 2 ( 2 2 + 1 + ( ) 1 2 2 4 2 (19) = ( 2 )2 + . 2 2 2 2

2 2 From (19), we can see that given and , is a function of and . In a periodic feedback system with feedback capacity constraint, indicated by (2) and (10), both and are related to the feedback interval . Therefore, the distortion can be expressed as a function of , and thus (19) can be rewritten as ) ( 2 0 (2 ) 12 4 2 + , (20) ( ) = ( 2 )2 2 2 2 2 0 (2 )

where the temporal correlation function ( ) = 0 (2 ) is a zero-order Bessel function of its rst kind. From (20), when is large enough, the temporal correlation tends to 0. Then, we have
2 lim ( ) = .

(21)

The reason is that when is large enough, the temporal correlation decreases to zero, which implies that the feedback quantization channel is completely outdated, and it also becomes independent of the estimated channel. As a result, the 2 distortion of CSI approaches . On the other hand, when 0 < < , we have 0 < ( ) < 1 and 1 < 2 . Hence, the rst term of (20) on the right-hand side is negative, and we have 2 ( ) < . (22) As stated before, with the increase of , the feedback rate increases, but the temporal correlation decreases and feedback delay becomes worse. Intuitively, there exists an optimal to minimize the distortion ( ). We prove the existence of the optimal by nding the extreme value of ( ), and the details of the proof can be found in Appendix C. It is hard to obtain the closed form solution of the optimal since there are Bessel function and Exponential function in the equation. Therefore, we calculate the approximate as + 3 3 + 27 + 3 81 2 + 123 2 2 =Round 6 3 3 + 27 3 81 2 + 123 2 2 , (23) 6
2

In a practical communication system, the feedback channel is causal, which implies that H is only used in the next +1 . With the causal feedback constraint, feedback period H

ZHANG et al.: ON THE MINIMUM DIFFERENTIAL FEEDBACK FOR TIME-CORRELATED MIMO RAYLEIGH BLOCK-FADING CHANNELS

415

14

1.1

12

10

0.9

R (bits)

8
2 e=0.05, d = 0.1, no differential compression 2=0.05, d = 0.1, differential compression e 2=0, d = 0.1, differential compression e 2 e=0.05, d = 0.2, no differential compression 2=0.05, d = 0.2, differential compression e 2 e=0, d = 0.2, differential compression

0.8 d (T) 0.7 C =0.5


fb

0.6

Cfb=1 0.5 C =2
fb fb

C =4 0.8 1 0.4 0 5 10 15 T 20 25 30

0.2

0.4

0.6

Fig. 2. The relationship between the minimum differential feedback rate and 2 time correlation for = 2, = 2, = 0, 0.05, and = 0.1, 0.2.
4 5 3 4 4 4 2 .

Fig. 3. The relationship between the distortion of CSI and the feedback 2 2 interval for = 2, = 2, = 1, and = 1.2.

where =

)/(

ln 2 2

( 4 )/( ) ) ln 2 , = 1 4 2

and = The detailed derivation of the approximate value is given in Appendix D. In the derivation, we use polynomials to approximate the Bessel function and Exponential function. To further verify the theoretical analysis, numerical results of the relationship between the distortion and the feedback interval in (20) are provided in Fig. 3. In addition, the comparison of approximation and the simulated result is given in Tab. I. IV. S IMULATION R ESULTS In this section, we rst provide the simulation results for the derived minimum differential feedback rate expression. Then, we discuss the relations between the ergodic capacity and the feedback interval in a periodic differential feedback system with feedback channel capacity constraint. Finally, we verify our theoretical results by a practical differential feedback system employing water-lling precoder and Lloyds quantization algorithm. All simulations are performed for a general point-to-point MIMO system over time-correlated block fading channels. For simplicity and without loss of generality, we use = 2 transmit antennas and = 2 2 receive antennas, and the channel variance is set as = 1. A. Minimum Differential Feedback Rate Figure 2 illustrates the minimum differential feedback rate versus the temporal correlation. The simulation parameters are 2 selected as follows. The value of channel estimation error is selected as 0 for the case of ideal channel estimation and 0.05 for a typical case of practical channel estimation in a medium SNR case (e.g., SNR = 6 dB) [23]. The quantization distortion is selected as 0.1 and 0.2 approximately corresponding to 10-bit and 14-bit quantization for 2 2 MIMO channels [20], respectively. For comparison, we also include the nondifferential compression results in Fig. 2, which is dened as a strategy that feeds back CSI independently with the previous

quantization CSI. The lower bound of the non-differential feedback rate can be found in[18] and [20]. From Fig. 2, we can see that when temporal correlation increases, it results in signicant reduction of feedback rate by using differential compression. In addition, the impact of estimation error and quantization distortion is also illustrated in Fig. 2. For lower quantization distortion, higher feedback rate is required. It can be also observed from Fig. 2 that with more estimation errors, the feedback rate increases. Especially, when the time correlation coefcient is = 1, the channel coefcient is a constant. If there is no estimation error, we get that H = H = H1 = H1 is also a constant. For this case, the previous quantization CSI is sufcient to preserve the accuracy of CSI at the transmitter, and thus no differential feedback bits are required. However, if the estimation error exists, the estimation channel H = H1 is no long a constant. Therefore, additional differential feedback bits are required to guarantee the accuracy of CSI at the transmitter. These observations can be well explained by (17). B. Ergodic Capacity and Feedback Interval In this subsection, we give the simulation results of the relationship between the ergodic capacity and the feedback intervals. For simplicity, we assume that the block size is = 100 with duration of 1 ms, and the power of the pilots takes = 10% of the total transmit power, which is a reasonable value in practice [23]. We select a relatively smaller value of SNR, which is 0 dB, and the Doppler frequency is 9.26 Hz, corresponding to 5 km/h moving speed and 2 GHz carrier frequency. Note that higher SNR and more pilot power result in lower estimation error, which also causes lower differential feedback rate. Given = 100, = 10%, and SNR = 0, we 2 can obtain that the variance of channel estimation is = 1.2 with the help of [23]. Figure 3 illustrates the distortion of CSI in (20) versus the feedback interval with different feedback capacity constraint = {0.5, 1, 2, 4} for every fading block. From Fig. 3, we can see that the CSI distortion is a convex function

416

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO. 2, FEBRUARY 2012

TABLE I A PPROXIMATION OF THE OPTIMAL FEEDBACK INTERVAL Approximation Values = 0.5 =1 =2 =4 8.9839 7.9729 6.8602 5.7563

1.75 1.7 1.65 1.6 1.55 1.5 1.45 1.4 1.35 Theoretical, Cfb=1 Lloyd, C =1
fb

Numerical results in Fig.3 9 7 6 4

2 2 Note: = = 2, = 1, = 1.2, = 9.26 Hz and = 0.001 s.

1.8 C =4
fb

1.75 1.7 Cerg (bits/s/Hz) 1.65 1.6 1.55 1.5 1.45 1.4

Cfb=2 C =1
fb fb

erg

(bits/s/Hz)

9 8 7 6

Thereotical, C =0.5
fb

C =0.5 2 4 6 8 T 10

Lloyd, C =0.5
fb

12

14

Fig. 5. The relationship between the ergodic capacity and feedback interval with Lloyd algorithm in AR1 model for = 2, = 2, SNR = 0 dB, = 100, and = 9.26 Hz.

8 T

10

12

14

16

Fig. 4, we can see that as increases, the ergodic capacity also increases. However, the absolute increment becomes smaller, which implies that it is necessary to limit the feedback channel transmission rate since little gain can be achieved when becomes very large. C. Differential Feedback System with Lloyds Quantization Algorithm In order to verify our theoretical results, we design a differential feedback system using Lloyds quantization algorithm ) ( [26], [27]. In this subsection, we use Di H , H1 = H H1 as a differential function. However, the analysis of the paper is actually based on Shannons Rate-Distortion Theory, which can be seen as a theoretical lower bound of feedback rate of any practical differential feedback scheme. The feedback procedure is given as follows. Firstly, a differential quantization codebook Cd is generated ofine using Lloyds algorithm according to the statistics of { } H H1 at both the transmitter and the receiver, where H1 is generated from (8). Secondly, the receiver calculates the current differential CSI H = H H1 . Thirdly, this current differential CSI is quantized as in the differential codebook with the smallest Euclidean distance to it. Fourthly, the corresponding codeword index is sent back to the transmitter. Finally, the transmitter recovers the channel quantization matrix by H = H1 + . In Fig. 5 and Fig. 6, we give the simulation results of ergodic capacity employing Lloyds algorithm in AR1 and Jakes model [28], respectively. As shown in Fig. 5 and Fig. 6, with the increase of , the ergodic capacity from both Lloyd algorithm (dashed curves) and theoretical calculation (solid curves) increases in the smaller region, but decreases in the larger region. It clearly shows that there exists an optimal feedback interval to maximize the ergodic capacity, which veries our theoretical analysis. Figure 5 and 6 also show that the ergodic capacity of Lloyds algorithm is lower than the theoretical ones, but they

Fig. 4. The relationship between the ergodic capacity and feedback interval for = 2, = 2, SNR = 0 dB, = 100, and = 9.26 Hz.

of feedback intervals given a feedback constraint and there exists an optimal feedback interval to minimize the distortion. This veries the analysis in Section III. We can also see that the optimal feedback intervals are = {9, 7, 6, 4} for = {0.5, 1, 2, 4}, respectively. In addition, Tab. I shows that the approximate values of the optimal feedback interval calculated by (23) are close to the numerical results. From Tab. I, we can see that the gap between approximations and the numerical results will be larger with increasing. This is mainly caused by the loss of the accuracy in (55). As increases, = (2 ) will increase, which degrades the approximation accuracy as it ignores the high-order terms. In Fig. 4, we show the relations between ergodic capacity and the feedback interval with the feedback capacity constraint = {0.5, 1, 2, 4} using water-lling precoder. It clearly shows that the ergodic capacity is a monotonic concave function of the feedback interval, and there exists an optimal feedback interval which achieves the maximum ergodic capacity. The reason is given as follows. When increases from a small region, it begins to provide larger feedback rate, and thus, improves the quality of feedback information. When goes towards a relatively larger region, the temporal correlation gradually decreases and the feedback delay becomes larger, which causes the feedback information outdated and therefore impairs the performance. Note that the relations between and in Fig. 4 is consistent with the analysis in section III, and similar optimal values of can be also found in Fig. 3. Additionally, from

ZHANG et al.: ON THE MINIMUM DIFFERENTIAL FEEDBACK FOR TIME-CORRELATED MIMO RAYLEIGH BLOCK-FADING CHANNELS

417

1.7 1.65 1.6 1.55 Cerg (bits/s/Hz) 1.5 1.45 1.4 1.35 1.3 1.25

Theoretical, Cfb=1 Lloyd, C =1


fb

Theoretical, C =0.5
fb

Lloyd, C =0.5
fb

A PPENDIX A P ROOF OF (6) Substituting (3) into (5), it yields ) ( 2 2 = 1 2 H He , 2

(24)

8 T

10

12

14

16

2 2 where and are the variances of the entries of H and H, respectively. Since the entries , of H, and ,, of He are i.i.d. complex Gaussian variables, the entries , of are also i.i.d. variables. Therefore, we only need to prove the onedimensional model. For simplicity, the foot labels are omitted. From (24), we can get ) ( 2 2 (25) = 1 2 . 2

Fig. 6. The relationship between the ergodic capacity and feedback interval with Lloyd algorithm using Jakes model for = 2, = 2, SNR = 0 dB, = 100, and = 9.26 Hz.

From (25), as is independent of using ML channel estimation, the variance of can be calculated by ( ) 2 2 2 2 . (26) = 2 As a result, the distribution of is given by ( ) 2 2 2 . 0, 2

go close to each other with increases. The reasons are given as follows. The gap between the Lloyds algorithm and theoretical derivation is caused by the imperfection of the Lloyds codebook, because Lloyds method is optimal only in the sense of minimizing a variables quantization error, but not in data sequence compression. In other words, the method of Shannons Rate-Distortion Theory can obtain even lower quantization error than Lloyds method. However, the effect of the imperfection on capacity calculation reduces when increases, because the quantization errors of both Lloyds algorithm and Shannons Rate-Distortion Theory tend to zero with sufcient feedback bits in the region of larger . In this case, the performance loss of the ergodic capacity is mainly caused by the channel time-variation (e.g., the feedback delay and the decrease of temporal correlation). As a result, the curves of Lloyds algorithm and theoretical calculation get converged in the larger region.

(27)

In the next, we give the proof that is independent of . As a complex Gaussian variable, can be written as = + , where = ) 1, and , are i.i.d. Gaussian variables ( 2 satisfying 0, 2 . Similarly, can be written as = + ) ( 2 , where and are i.i.d. Gaussian variables with 0, 2 . ( ) We then consider the conditional probability when is given. For the real part, the probability can be expressed as () = () () ()
1 2 (2 )

= V. C ONCLUSIONS In this paper, we have derived the minimum differential feedback rate for the time-correlated Rayleigh block-fading channels considering channel estimation errors, which is the lower bound of feedback compression with temporal correlation. Furthermore, provided the feedback-channel constraint per fading block, the relationship between the ergodic closed-loop capacity and the feedback interval is investigated in this paper. We nd that the ergodic closed-loop capacity is a monotonic concave function of the feedback interval and there exists an optimal feedback interval to achieve the maximum ergodic capacity. Simulation results of a practical differential feedback system with water-lling precoder (Please note that other precoder can still be readily applied.) and Lloyds quantization algorithm are provided to validate our theoretical results.

) ( 2) ( 2 exp ()2 1 2 exp 2 2 ) ( . 2 1 exp 2 2

Therefore, we have

1 exp () = 2 2 (2 ) 2

2 2

)2 . (28)

2 2 (2 ) 2

Similarly, the imaginary part can be written as ( )2 2 2 1 () = exp 2 2 2 . 2 2 ( ) (2 ) 2 2


(29)

Combining (28) and (29), given , the conditional distribution of can be calculated by ( ) 2 2 2 2 2 , . (30) 2

418 Since = , the conditional distribution of given 2 can be represented as ( ) 2 2 2 . 0, (31) 2


2

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO. 2, FEBRUARY 2012

From (27) and (31), we nd that the distribution of does not change whether is given or not. Hence, is independent of . Finally, the independent property between and H is proved. A PPENDIX B P ROOF OF L EMMA 1 From (3), we have = + . (32)

Then, (41) can be rewritten as ) ( 2 2 2 1 2 1 + 2 1 + 2 + 2 ( ) 2 (1 ) + 2 log . (42) 2 As 1 , 1 , and are complex Gaussian variables, and independent with each other, (42) can be further written as ( 2 2 2 1 2 1 + 2 1 + 2 + 2 ; ) 2 2 2 2 1 2 1 + 2 + 2 ( ) 2 + 2 log . (43) 2 According to the rate distortion theory of continuous amplitude sources [25], (43) achieves the minimum value when 1 , 1 , and are independent Gaussian variables, and we have [ ( ) ] 4 4 ) 2 1 2 1 ( 2 2 log 1 + 2 + 4 2 + 2 2 ) ( 2 (44) + 2 log . 2 From (44), we can easily obtain ( ) ( )] ) ( 2 2 2 2 1 2 2 2 2 log + 1 + 2 + . 2

From (2), the one-dimensional AR1 channel model can be rewritten as a scalar form = 1 + 1 2 . (33) Substituting (33) into (32) yields ) ( = 1 + 1 2 + . From (6), we have 1 =
2 2 1 + 1 ,

(34)

(35)

where is independent on , as proved in Appendix A. Substituting (35) to (34), it yields ( ) 2 = 1 + 1 + 1 2 + . (36) 2 From (7), we have 1 = 1 + 1 . Substituting (37) into (36), we obtain = 1 + 1 + 1 + 2 2
2 2

(45)

(37)

A PPENDIX C P ROOF OF E XISTENCE OF THE O PTIMAL For simplicity, we assume that = 2 is a continuous variable. Thus, the time correlation coefcient can be written as = 0 (2 ) = 0 (), and the feedback interval can be calculated as a nearest integer of / = / (2 ). From (20), the distortion can be rewritten as ) ( 2 4 0 () 1 2 2 () = (46) + , ( 2 )2 2 2 2 0 () 2

1 2 + .

(38) When 1 is given, the conditional mutual information can be written as ( ) ( ) ( ) ; 1 = 1 , 1 . (39) Substituting (38) into (39), it yields ( ) 2 = 2 1 + 1 + 1 2 + ( ) 1 . (40) ) ( Considering the inequality 1 ( ), and the identical equation ( ) = (1 ), (40) can be expressed as ) ( 2 2 + 2 1 + 1 + 1 (1 ) . (41)

where = / (2 ), and () is a continuously differentiable function of . Then, the rst-order derivative of () is given by ( 4) 2 0 () 2 [ ( ) () = ( )2 2 2 1 1 () ( 2 )2 2 2 0 () 2 ( )2 2 3 ln 2 0 () 0 () , (47) 2

ZHANG et al.: ON THE MINIMUM DIFFERENTIAL FEEDBACK FOR TIME-CORRELATED MIMO RAYLEIGH BLOCK-FADING CHANNELS

419

where 1 () = 0 () [17], and () denotes a rst kind -th order Bessel function. As lim 0 () = 1 and 1 () = 0, the rst-order derivative

of () can be calculated as [ ( ( 2 ) )] 4 ln 2 1 2 2 2 < 0. ()0 = ( )2 ( 2 )2 1 2 (48) By taking a special value (there may exist many other values), = 3 , which enables the value of the rst-order Bessel 2 ( ) ( ) function larger than the zero-order one, i.e., 1 3 > 0 3 , 2 2 we have ( 4) ( ) 3 2 2 2 0 3 [ ( 3 ) 2 2 2 ()= 3 > ( ( 2 )2 ( ) )2 2 2 1 2 3 2 2 2 0 3 2 2 ( )2 ( ) 2 2 3 ln 2 1 0 2 2 ( 4) ( ) [ ( 3 ) ] 3 2 2 2 0 3 2 2 2 1 ln 2 2 2 > . ( ( 2 )2 ( ) ) 2 3 3 2 2 0 2 2 2

As it is difcult to nd the closed-form solution of (53), we calculate the approximate value of . The Bessel function can be approximately calculated by [30] 0 () = 1 25 ( )2 ( )4 ( )6 + 150 350 10 10 10 ( )8 ( )10 +350 126 + . 10 10

(54)

When 1 (i.e., 1 implies low moving speed), using Taylor expansion, the exponential function can be approximated by 2 =1 + ln 2 + ( ln 2) 2 ( ln 2) 3 + 2 6 4 5 6 ( ln 2) 4 ( ln 2) 5 ( ln 2) 6 + + + . + 24 120 720 (55)
2 3

(49)
3 2

Utilizing Taylor expansion, we have 2 = 1 + 3 ln 2 + 2 ( )2 1 3 + . Therefore, we obtain 2! 2 ln 2 ( 3 ) 2 2 2 1 ln 2 > 3 ln 2 ln 2 > 0. (50) Substituting (50) to (49), we have ()= 3 > 0. 2 (51)

Substituting (54) and (55) into (53), we have {( ) ) ( 4 4 5 3 ln 2 3 ln 2 2 + 4 1 + 4 4 4 2 ( ) 4 93 21 1 2 + + 4 4 + 6 ( ln 2) 200 400 ) ( } 1 3 3 ( ln 2) ln 2 5 + = 0. + 24 50

(56)

As () is a continuous function on , combining (48) and (51), we can easily conclude that there exists a to make 3 () = 0 when 0 < < 2 . Thus, the existence of the optimal = Round ( / (2 )) is proved.

Note that the accuracy of the solution is dependent on the orders of the polynomial in the left term of (56). However, higher order causes more complexity to obtain the solution. For simplicity, we limit the solution of the 3rd-order equation ) ( ) ( 4 4 5 3 ln 2 3 = 0. (57) 2 + 4 1 + 4 4 4 2 Using Cardans formula, the real root can be obtained ( 3 1 27 3 + 3 + + 81 2 + 123 = 3 2 2 ) 3 27 3 + 3 + 81 2 + 123 . 2 2 (58) where = ( ln 2 ) 4 5 3 , 4 4 4 2 2 . Accordingly, ( )/ = (
4 4

A PPENDIX D A PPROXIMATION OF THE O PTIMAL Assuming a continuous variable = 2 , the distortion (20) has been expressed as a function of in (46). In addition, the rst-order derivative of () has been obtained in (47). In this section, we rst try to derive the approximation of the optimal , and then calculate the approximation as a nearest integer of / = / (2 ). When the rst-order derivative of () is equal to zero, an optimal exists, i.e., () = 0. (52) Combining (47) and (52), the optimal satises ( )2 2 ( ) 3 2 2 1 1 () ln 2 0 () 0 () = 0. 2

)/( ) ln 2 1 2

and = the approximation of the optimal , which is the nearest integer of /2 , can be calculated by + 3 3 + 27 + 3 81 2 + 123 2 2 = Round 6 3 3 + 27 3 81 2 + 123 2 2 . 6 (59)

(53)

420

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO. 2, FEBRUARY 2012

R EFERENCES
[1] D. J. Love, R. W. Heath. Jr., W. Santipach, and M. L. Honing, What is the value of feedback for MIMO channels? IEEE Commun. Mag., vol. 42, no. 10, pp. 5459, Oct. 2004. [2] D. J. Love, R. W. Heath, V. K. N. Lau, D. Gesbert, B. D. Rao, and M. Andrews, An overview of limited feedback in wireless communication systems, IEEE J. Sel. Areas Commun., vol. 26, no. 8, pp. 13411365, Oct. 2008. [3] W. C. Jakes, Microwave Mobile Communications. Wiley, 1974. [4] K. Huang, B. Mondal, R. W. Heath. Jr., and J. G. Andrews, Markov models for limited feedback MIMO systems, in Proc. 2006 IEEE ICASSP. [5] K. Huang, R. W. Heath, Jr., and J. G. Andrews, Limited feedback beamforming over temporally-correlated channel, IEEE Trans. Signal Process., vol. 57, no. 5, pp. 19591975. May 2009. [6] K. E. Baddour and N. C. Beaulieu, Autoregressive modeling for fading channel simulation, IEEE Trans. Wireless Commun., vol. 4, pp. 1650 1662. July 2005. [7] T. Eriksson and T. Ottosson, Compression of feedback for adaptive transmission and scheduling, Proc. IEEE, pp. 23142321, Dec. 2007. [8] B. Mondal and R. W. Heath, Jr., Channel adaptive quantization for limited feedback MIMO beamforming systems, IEEE Trans. Signal Process., vol. 54, no. 12, pp. 47174729, Dec. 2006. [9] T. Kim, D. J. Love, B. Clerckx, and S. J. Kim, Differential rotation feedback MIMO system for temporally correlated channels, in Proc. 2008 IEEE GLOBECOM. [10] K. Huang, B. Mondal. R. W. Heath, Jr., and J. G. Andrews, Multiantenna limited feedback for temporally-correlated channels: feedback compression, in Proc. 2006 IEEE GLOBECOM. [11] C. Simon, R. de. Francisco, D. T. M. Slock, and G. Leus, Feedback compression for correlated broadcast channels, 2007 IEEE Symp. Commun. Veh. Technol. Benelux. [12] K. Kim, H. Kim, and D. J. Love, Utilizing temporal correlation in multiuser MIMO feedback, in Proc. 2008 IEEE ACSSC. [13] C. Simon and G. Leus, Feedback reduction for spatial multiplexing with linear precoding, in Proc. 2007 IEEE ICASSP. [14] D. Zhang, J. Xie, G. Wei, and J. Zhu, Capacity and feedback rate in closed-loop MIMO wireless systems, in Proc. 2004 IEEE ICCS. [15] G. M. Gvensen and A. O. Yilmaz, An upper bound for limited rate feedback MIMO capacity, IEEE Trans. Wireless Commun., vol. 8, no. 6, pp. 27482754, June 2009. [16] K. N. Lau, Y. Liu, and T. A. Chen, On the design of MIMO blockfading channels with feedback-link capacity constraint, IEEE Trans. Commun., vol. 52, no. 1, pp. 6270, Jan. 2004. [17] A. D. Dabbagh and D. J. Love, Feedback rate-capacity loss tradeoff for limited feedback MIMO systems, IEEE Trans. Inf. Theory, vol. 52, no. 5, pp. 21902202, May 2006. [18] D. Zhang, G. Wei, J. Zhu, and Z. Tian, On the bounds of feedback rates for pilot-assisted MIMO systems, IEEE Trans. Veh. Technol., vol. 56, no. 4, pp. 17271736, July 2007. [19] W. Li, M. Ma, and B. Jiao, Comments on On the bounds of the feedback rates for pilot-assisted MIMO systems, IEEE Trans. Veh. Technol., vol. 58, no. 8, pp. 46574660, Oct. 2009. [20] W. Li, M. Ma, and B. L. Jiao, Optimization of feedback for adaptive MIMO transmissions over time varying channels, Wireless Personal Commun., vol. 54, no. 4, pp. 579590, July 2009. [21] V. R. Cadambe and S. A. Jafar, Interference alignment and degrees of freedom of the K-user interference channel, IEEE Trans. Inf. Theory, vol. 54, no. 8, pp. 34253441, Aug. 2008. [22] J. Thukral and H. Bolcskei, Interference alignment with limited feedback, in Proc. 2009 IEEE ISIT. [23] D. Samardzijia and N. Mandayam, Pilot-assisted estimation of MIMO fading channel response and achievable data rates, IEEE Trans. Signal Process., vol. 51, no. 11, pp. 28822890, Nov. 2003. [24] Y. Sun and M. L. Honing, Asymptotic capacity of multicarrier transmission with frequency-selecive fading and limited feedback, IEEE Trans. Inf. Theory, vol. 54, no.7, pp. 28792902, July 2008. [25] R. J. McEliece, The Theory of Information and Coding, 2nd edition. Cambridge University Press, 2002. [26] S. P. Lloyd, Least-square quantization in PCM, IEEE Trans. Inf. Theory, vol. 28, pp. 129137, Mar. 1982. [27] P. Xia, S. Zhou, and G. B. Giannakis, Achieving the Welch bound with difference sets, IEEE Trans. Inf. Theory, vol. 51, no. 5, pp. 19001907, May 2005.

[28] C. X. Wang, M. Patzold, and D. F. Yuan, Accurate and efcient simulation of multiple uncorrelated Rayleigh fading waveforms, IEEE Trans. Wireless Commun., vol. 6, no. 3, pp. 833839, Mar. 2007. [29] R. P. Millane and J. L. Eads, Polynomial approximations to bessel functions, IEEE Trans. Antennas Propag., vol. 51, no. 6, pp. 1398 1400, June 2003. [30] F. B. Gross, New approximations to 0 and 1 bessel functions, IEEE Trans. Antennas Propag., vol. 43, no. 8, pp. 904907, Aug. 1995. Leiming Zhang (S10) received the B.S. degree in Electronic Engineering from Nankai University with the outstanding undergraduate thesis award, in 2007. He is currently pursuing his Ph.D. degree in Peking University with the major in Wireless Communication and Signal Processing. His research interests include adaptive transmission techniques in multi-antenna systems and interference alignment.

Lingyang Song (S03-M06) received the B.S. degree in communication engineering from Jilin University, China in 2002 and a PhD in differential space time codes and MIMO from the University of York, UK, in 2007, where he received the K. M. Stott Prize for excellent research. He worked as a postdoctoral research fellow at the University of Oslo, Norway, until rejoining Philips Research UK in March 2008. Now, he is a full professor with School of Electronics Engineering and Computer Science, Peking University, China. He was a visiting research fellow at Harvard University, US, and the University of York, UK. His main research interests include adaptive signal processing, estimation and optimization theory, cognitive and collaborative communications, channel coding, wireless mesh/sensor networks, broadband wireless access and future communication systems. He is co-inventor of a number of patents and author or co-author of over 100 journal and conference papers. He is currently on the Editorial Board of the International Journal of Communications, Network and System Sciences, the Journal of Network and Computer Applications, and the International Journal of Smart Homes, and a guest editor of Elsevier Computer Communications and EURASIP Journal on Wireless Communications and Networking. He serves as a member of Technical Program Committee and Co-chair for several international conferences and workshops. He is a member of the IEEE and IEEE ComSoc. Meng Ma was born in Harbin, China, in 1978. He received B.S. and Ph.D. degrees in Electrical Engineering from Peking University, in 2001 and 2007, respectively. From January 2009 to January 2010, He was a visiting scholar at Commonwealth Scientic and Industrial Research Organization (CSIRO), Sydney, Australia. He is currently a lecturer with the School of Electronics Engineering and Computer Science, Peking University, Beijing, China. His research interests include transmitter and receiver signal processing in OFDM or CDMA systems, channel estimation and information theory. Bingli Jiao received B.S. and M.S. degree from Peking University, China in 1983 and 1988, respectively, and received Ph.D. from University of Sarrbruecken, F. R., Germany in 1995. Then, he became an associate professor and professor with Peking University in 1995 and 2000, respectively. His current interests include communication theory and techniques and sensor design.