Scholars' Mine

Doctoral Dissertations Student Theses and Dissertations

Summer 2012

Power allocation and linear precoding for wireless

communications with finite-alphabet inputs
Mingxi Wang

Follow this and additional works at: https://scholarsmine.mst.edu/doctoral_dissertations

Part of the Electrical and Computer Engineering Commons

Department: Electrical and Computer Engineering

Recommended Citation
Wang, Mingxi, "Power allocation and linear precoding for wireless communications with finite-alphabet
inputs" (2012). Doctoral Dissertations. 1915.
https://scholarsmine.mst.edu/doctoral_dissertations/1915

This thesis is brought to you by Scholars' Mine, a service of the Missouri S&T Library and Learning Resources. This
work is protected by U. S. Copyright Law. Unauthorized use including reproduction for redistribution requires the
permission of the copyright holder. For more information, please contact [email protected].
 POWER ALLOCATION AND LINEAR PRECODING

FOR WIRELESS COMMUNICATIONS WITH

FINITE-ALPHABET INPUTS

by

MINGXI WANG

A DISSERTATION

Presented to the Faculty of the Graduate School of the

MISSOURI UNIVERSITY OF SCIENCE AND TECHNOLOGY

In Partial Fulfillment of the Requirements for the Degree

DOCTOR OF PHILOSOPHY

in

ELECTRICAL ENGINEERING

2012

Approved by

Chengshan Xiao, Advisor

Yahong Rosa Zheng
Kurt Kosbar
Randy Moss
Xiaoqing (Frank) Liu
 iii

PUBLICATION DISSERTATION OPTION

This dissertation consists of the following four published or to be submitted

papers, formatted in the style used by the Missouri University of Science and Tech-

nology, listed as follows:

Paper I, M. Wang, C. Xiao, and W. Zeng, “Linear Precoding for MIMO

Multiple Access Channels with Finite Discrete Inputs,” has been published in

IEEE Trans. Wireless Commun., vol. 10, pp 3934-3942, Nov. 2011. (pages 6-36)

Paper II, W. Zeng, M. Wang, C. Xiao, and J. Lu, “On the Power Alloca-

tion for Relay Networks with Finite-alphabet Constraints,” has been published in

Proc. IEEE GLOBECOM, Miami, FL, Dec. 2010. (pages 37-55)

Paper III, W. Zeng, Y. R. Zheng, M. Wang, and J. Lu, “Linear Precoding for

Relay Networks: A Perspective on Finite-Alphabet Inputs,” has been accepted by

IEEE Trans. on Wireless Commun., Dec. 2011. (pages 56-95)

Paper IV, M. Wang, Y. R. Zheng, and C. Xiao, “Practical Linear Precoder

Design for Finite Alphabet MIMO-OFDM with Experiment Validation,” to be sub-

mitted to IEEE Trans. Veh. Technol., 2012. (pages 96-127)

 iv

ABSTRACT

This dissertation proposes a new approach to maximizing data rate/throughput

of practical communication system/networks through linear precoding and power allo-

cation. First, the mutual information or capacity region is derived for finite-alphabet

inputs such as phase-shift keying (PSK), pulse-amplitude modulation (PAM), and

quadrature amplitude modulation (QAM) signals. This approach, without the com-

monly used Gaussian input assumptions, complicates the mutual information analysis

and precoder design but improves performance when the designed precoders are ap-

plied to practical systems and networks.

Second, several numerical optimization methods are developed for multiple-

input multiple-output (MIMO) multiple access channels, dual-hop relay networks,

and point-to-point MIMO systems. In MIMO multiple access channels, an iterative

weighted sum rate maximization algorithm is proposed which utilizes an alternating

optimization strategy and gradient descent update. In dual-hop relay networks, the

structure of the optimal precoder is exploited to develop a two-step iterative algorithm

based on convex optimization and optimization on the Stiefel manifold. The proposed

algorithm is insensitive to initial point selection and able to achieve a near global op-

timal precoder solution. The gradient descent method is also used to obtain the

optimal power allocation scheme which maximizes the mutual information between

the source node and destination node in dual-hop relay networks. For point-to-point
 v

MIMO systems, a low complexity precoding design method is proposed, which max-

imizes the lower bound of the mutual information with discretized power allocation

vector in a non-iterative fashion, thus reducing complexity.

Finally, performances of the proposed power allocation and linear precoding

schemes are evaluated in terms of both mutual information and bit error rate (BER).

Numerical results show that at the same target mutual information or sum rate, the

proposed approaches achieve 3-10dB gains compared to the existing methods in the

medium signal-to-noise ratio region. Such significant gains are also indicated in the

coded BER systems.

 vi

ACKNOWLEDGMENTS

The research work reported in this dissertation was carried out with the De-

partment of Electrical and Computer Engineering at Missouri University of Science

and Technology, Rolla, Missouri, USA, from August 2009 to May 2011.

First of all, I would like to thank my Ph.D. advisor Professor Chengshan Xiao.

I greatly appreciate his support and guidance for my course and research work. His

attitudes and visions towards work and other things have been a role model for me

during my study.

Besides my advisor Professor Xiao, I would like to thank Professor Yahong

Rosa Zheng. During my Ph.D. study, she gave me valuable help with my course work

and research. I also worked with her on several projects and papers from which I

learned a lot.

I want to thank the other members of my advisory committee, Professors.

Kurt Kosbar, Randy Moss, and Xiaoqing Liu. They have made various valuable

comments and suggestions to help me improve my research work. They were also

considerate in arranging their time for my comprehensive exams and final defense.

In addition, it is my pleasure to acknowledge my appreciation to some of my

co-authors, especially Mr. Weiliang Zeng and Mr. Yongpeng Wu, with whom I

worked on several theoretical problems. I enjoyed the collaborations and I am glad

that we have made productive publications. I also cherish the time I spent with all

the other members in the research lab.

Last but not least, I dedicate this dissertation to my parents and relatives in

China. I would like to express my sincere gratitude for their love and support.
 vii

TABLE OF CONTENTS

Page
PUBLICATION DISSERTATION OPTION . . . . . . . . . . . . . . . . . . . iii
ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
LIST OF ILLUSTRATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii

SECTION
1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 BACKGROUND AND PROBLEM STATEMENT . . . . . . . . . 1
1.2 SUMMARY OF CONTRIBUTIONS . . . . . . . . . . . . . . . . . 2
1.3 REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

PAPER
I. LINEAR PRECODING FOR MIMO MULTIPLE ACCESS CHANNELS
WITH FINITE DISCRETE INPUTS . . . . . . . . . . . . . . . . . . . . . . . 6
ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 SYSTEM MODEL AND EXISTING GAUSSIAN INPUTS RESULTS . 11
3 CONSTELLATION-CONSTRAINED CAPACITY REGION . . . . . . 13
4 WEIGHTED SUM RATE MAXIMIZATION . . . . . . . . . . . . . . . 15
4.1 NECESSARY CONDITIONS . . . . . . . . . . . . . . . . . . . . . 15
4.2 ITERATIVE ALGORITHM FOR WSR MAXIMIZATION . . . . 17
5 ITERATIVE DETECTION AND DECODING FOR MAC . . . . . . . 20
6 NUMERICAL RESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . 23
7 CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
8 APPENDIX: PROOF OF PROPOSITION 1 . . . . . . . . . . . . . . . 32
9 REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
 viii

II. ON THE POWER ALLOCATION FOR RELAY NETWORKS WITH FI-

NITE ALPHABET CONSTRAINT . . . . . . . . . . . . . . . . . . . . . . . . 37

ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

1 INTRODUCTION AND RELATED WORK . . . . . . . . . . . . . . . 38

2 SYSTEM MODEL AND PRELIMINARIES . . . . . . . . . . . . . . . 40

3 POWER ALLOCATION FOR FINITE ALPHABET INPUTS . . . . . 44

3.1 ANALYSIS FROM INFORMATION THEORETIC PERSPECTIVE 44

3.2 ANALYSIS FROM OPTIMIZATION THEORETIC PERSPECTIVE 45

3.3 ITERATIVE DETECTION AND DECODING . . . . . . . . . . . 48

4 NUMERICAL RESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . 49

5 CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

6 REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

III. LINEAR PRECODING FOR RELAY NETWORKS: A PERSPECTIVE

ON FINITE-ALPHABET INPUTS . . . . . . . . . . . . . . . . . . . . . . . . 56

ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

2 SYSTEM MODEL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

3 MUTUAL INFORMATION FOR RELAY NETWORKS . . . . . . . . 65

4 PRECODER DESIGN TO MAXIMIZE MUTUAL INFORMATION . 67

4.1 OPTIMAL LEFT SINGULAR VECTORS . . . . . . . . . . . . . 67

4.2 OPTIMAL POWER ALLOCATION . . . . . . . . . . . . . . . . . 68

4.3 OPTIMIZATION OVER RIGHT SINGULAR VECTORS . . . . . 71

4.4 TWO-STEP APPROACH TO OPTIMIZE PRECODER . . . . . . 73

5 SIMULATION RESULTS . . . . . . . . . . . . . . . . . . . . . . . . . 75

5.1 MUTUAL INFORMATION PERFORMANCE . . . . . . . . . . . 75

5.2 CODED BER PERFORMANCE . . . . . . . . . . . . . . . . . . . 81

6 CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

7 APPENDIX A: PROOF OF PROPOSITION 1 . . . . . . . . . . . . . 85

 ix

7.1 DEFINITION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

7.2 JACOBIAN OF THE MMSE MATRIX . . . . . . . . . . . . . . . 85

8 REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

IV. PRACTICAL LINEAR PRECODER DESIGN FOR FINITE ALPHABET

MIMO-OFDM WITH EXPERIMENT VALIDATION . . . . . . . . . . . . . . 96

ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

2 LINEAR PRECODER DESIGN FOR FINITE ALPHABET . . . . . . 100

3 PROPOSED SIMPLIFIED LINEAR PRECODER DESIGN . . . . . . 104

4 APPLICATIONS OF LINEAR PREDING TO MIMO-OFDM . . . . . 113

5 TESTBED, EXPERIMENT SETUP, AND RESULTS . . . . . . . . . . 116

5.1 TESTBED AND EXPERIMENT SETUP . . . . . . . . . . . . . . 116

5.2 FIELD EXPERIMENTS . . . . . . . . . . . . . . . . . . . . . . . 118

5.3 EXPERIMENT RESULTS . . . . . . . . . . . . . . . . . . . . . . 120

6 CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

7 ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . 123

8 REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

SECTION

2 CONCLUSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

3 PUBLICATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

VITA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
 x

LIST OF ILLUSTRATIONS

Figure Page

PAPER I

1 MIMO uplink transmitters with precoding. . . . . . . . . . . . . . . . . . . 20

2 Iterative receiver of MIMO multiple access channel. . . . . . . . . . . . . . 21

3 Convergence of sum rate maximization algorithm with BPSK inputs. . . . 24

4 Sum rate of 2-user MAC with BPSK inputs. . . . . . . . . . . . . . . . . . 25

5 Sum rate of 2-user MAC with QPSK inputs. . . . . . . . . . . . . . . . . . 27

6 Capacity region of 2-user MAC with BPSK inputs, when SNR = 5dB. . . . 28

7 Capacity region of 2-user MAC with QPSK inputs, when SNR = 10dB. . . 29

8 BER of 2-user MAC with BPSK inputs. . . . . . . . . . . . . . . . . . . . 30

9 BER of 2-user MAC with QPSK inputs. . . . . . . . . . . . . . . . . . . . 30

PAPER II

1 Block diagram of the relay network with iterative receiver at the destination. 49

2 Mutual Information for a fixed channel with QPSK inputs. . . . . . . . . . 51

3 Bit error performance comparison for optimal power allocation, modified

MMSE and modified network beamforming. . . . . . . . . . . . . . . . . . 51

4 Average mutual Information over Rayleigh fading channels with QPSK inputs. 52

PAPER III

1 Evolution of the mutual information with BPSK inputs when SNR is 0 dB. 76

2 Cumulative distribution of the optimized mutual information with BPSK

inputs when SNR is 0 dB. . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

3 Evolution of the mutual information with QPSK inputs when SNR is 5 dB. 78

4 Cumulative distribution of mutual information with QPSK inputs when

SNR is 5 dB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

5 Mutual information versus the SNR with BPSK inputs. . . . . . . . . . . . 80

 xi

6 Mutual information versus the SNR with QPSK inputs. . . . . . . . . . . . 80

7 Block diagram of the relay network with precoding at the source and itera-
tive receiver at the destination. . . . . . . . . . . . . . . . . . . . . . . . . 82

8 Bit error performance with BPSK inputs and 3/4 channel coding rate. . . 83

9 Bit error performance with QPSK inputs and 2/3 channel coding rate. . . 83

PAPER IV

1 Scatter plot of the precoded BPSK and 8PSK signals using 2 × 2 maximum
diversity and modulation diversity . . . . . . . . . . . . . . . . . . . . . . . 106

2 Mutual information vs. SNR . . . . . . . . . . . . . . . . . . . . . . . . . . 110

3 Amplitudes of precoded elements in a 2 × 2 MIMO-OFDM system. The size

of the dots is proportional to the amplitude of precoder coefficients. Only
eight subcarriers are shown. . . . . . . . . . . . . . . . . . . . . . . . . . . 112

4 System diagram of 2×2 MIMO-OFDM system. . . . . . . . . . . . . . . . 113

5 Frame structure of MIMO-OFDM signaling . . . . . . . . . . . . . . . . . 113

6 Transmitter setup architecture . . . . . . . . . . . . . . . . . . . . . . . . . 117

7 Receiver setup architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

8 Floor plan of the indoor NLOS environment (9.1 m × 7.0 m) . . . . . . . . 119

9 Estimated channel impulse responses of indoor same room NLOS environment120

 xii

LIST OF TABLES

Table Page

PAPER I

1 The algorithm for weighted sum rate maximization with finite discrete inputs 18

PAPER II

1 The algorithm for optimal power allocation coefficients . . . . . . . . . . . 47

PAPER III

1 Algorithm 1: Optimization of power allocation vector . . . . . . . . . . . . 71

2 Algorithm 2: Optimization of right singular vectors . . . . . . . . . . . . . 73

3 Algorithm 3: Two-step optimization algorithm . . . . . . . . . . . . . . . 73

PAPER IV

1 The values of qmod corresponding to different modulations . . . . . . . . . 105

2 The CPU time (seconds) of designing precoders for the 4 × 4 channel H2 . 108

3 Description of Key Equipments . . . . . . . . . . . . . . . . . . . . . . . . 116

4 BER results using List FSD for indoor same room NLOS experiments . . . 121
 1 INTRODUCTION

1.1 BACKGROUND AND PROBLEM STATEMENT

Linear precoding for multiple-input multiple-output (MIMO) communications

has been a popular research topic in the last decade. Traditional precoding methods

can be mainly classified into three categories [1]: (a) diversity- driven designs [2]; (b)

rate-driven designs [3, 4] and (c) mean squared error (MSE) minimization or signal-

to-noise ratio (SNR) maximization [5]. The diversity-driven designs employs pairwise

error probability analysis to maximize diversity order [2]. Yet it may not achieve the

highest coding gain. The rate-driven designs utilize ergodic or outage capacity to

optimize the precoder. Such approach rely on Gaussian input assumptions, whereas

the transmitted signals in practical digital communication systems are non-Gaussian

distributed, drawn from discrete constellation sets such as phase-shift keying (PSK),

pulse-amplitude modulation (PAM), or quadrature amplitude modulation (QAM).

Therefore, precoders obtained based on Gaussian input assumptions may lead to

performance loss when the inputs are actually replaced by finite discrete inputs. For

the third categories, the MSE minimization and SNR maximization approaches [5]

may not necessarily provide minimum bit error rate (BER) or maximum date rate

for practical finite alphabet systems.

To overcome the aforementioned drawbacks, several works have recently re-

ported that designing precoders for point to point MIMO system by maximizing

mutual information with finite-alphabet inputs can achieve higher mutual informa-

tion [1, 6, 7] and lower bit error rate (BER) [1] than employing other optimization

criteria. The performance benefits of these approaches come from optimization of

 2

mutual information formulated directly with finite-alphabet input constraints, com-

pared to using other indirect methods such as maximizing channel capacity with

Gaussian inputs, maximizing diversity order, minimizing SINR, or minimizing MSE.

To maximize the mutual information with finite-alphabet input, this disser-

tation develops several numerical optimization methods for multiple-input multiple-

output (MIMO) multiple access channels, dual-hop relay networks, and point-to-point

MIMO systems. In MIMO multiple access channels, an iterative weighted sum rate

maximization algorithm is proposed which utilizes an alternating optimization strat-

egy and gradient descent update. In dual-hop relay networks, the structure of the

optimal precoder is exploited to develop a two-step iterative algorithm based on con-

vex optimization and optimization on the Stiefel manifold. The proposed algorithm is

insensitive to initial point selection and able to achieve a near global optimal precoder

solution. The gradient descent method is also used to obtain the optimal power allo-

cation scheme which maximizes the mutual information between the source node and

destination node in dual-hop relay networks. For point-to-point MIMO systems, a low

complexity precoding design method is proposed, which maximizes the lower bound

of the mutual information with discretized power allocation vector in a non-iterative

fashion, thus reducing complexity.

1.2 SUMMARY OF CONTRIBUTIONS

This dissertation consists of a couple of journal publications and conference

papers as listed in the publication list. My contributions that are published or under

review are:

1. Linear precoding for MIMO multiple access channels with finite discrete

inputs: The constellation-constrained capacity region for the MIMO MAC is derived

with an arbitrary number of users. Due to the non-concavity of the objective func-

tion, a set of necessary conditions for the optimization problem are obtained through
 3

Karush-Kuhn-Tucker analysis. An iterative algorithm is proposed to utilize alter-

nating optimization strategy. In particular, each iteration of the algorithm involves

the gradient descent update with backtracking line search. Numerical results show

that when inputs are digital modulated signals and the signal-to-noise ratio is in the

medium range, our proposed algorithm offers considerably higher sum rate than non-

precoding and the traditional method which maximizes Gaussian-input sum capacity.

Furthermore, a low-density parity-check coded system with iterative detection and

decoding for MAC is presented to evaluate the bit error rate performance of pre-

coders. BER results also indicate that the system with the proposed linear precoder

achieves significant gains over the non-precoding system and the precoder designed

for Gaussian inputs.

2. On the power allocation for relay networks with finite-alphabet constraints:

An optimal power allocation scheme is proposed to maximize the mutual informa-

tion for the relay networks under discrete-constellation input constraint. Numerical

examples show that significant gain can be obtained compared to the conventional

counterpart for nonfading channels and fading channels. At the same time, we show

that the large performance gain on the mutual information will also represent the large

gain on the bit-error rate, i.e., the benefit of the power allocation scheme predicted

by the mutual information can indeed be harvested and can provide considerable

performance gain in a practical system.

3. Linear precoding for relay networks with a perspective on finite-alphabet

inputs: This paper exploits the structure of the optimal precoder that maximizes the

mutual information and develops a two-step algorithm based on convex optimization

and optimization on the Stiefel manifold. By doing so, the proposed algorithm is

insensitive to initial point selection and able to achieve a near global optimal precoder

solution. Besides, it converges fast and offers high mutual information gain. These
 4

advantages are verified by numerical examples, which also show the large performance

gain in mutual information also represents the large gain in the coded bit-error rate.

4. Practical linear precoder design for finite alphabet MIMO-OFDM with

experiment validation. A low complexity precoding method is proposed for practical

multiple-input multiple-output (MIMO) orthogonal frequency-division multiplexing

(OFDM) systems. Based on the two-step optimal precoder design algorithm that

maximizes the lower bound of the mutual information with finite-alphabet inputs, the

proposed method simplifies the precoder design by fixing the right singular vectors

of the precoder matrix, eliminating the iterative optimization between the two steps,

and discretizing the search space of the power allocation vector. For a 4×4 channel,

the computational complexity of the proposed precoder design is reduced to 3% and

6% of that required by the original two-step algorithm with Quadrature Phase Shift

Keying (QPSK) and 8PSK, respectively. The proposed method achieves nearly the

same mutual information as the two-step iterative algorithm for a large range of

SNR region, especially for large MIMO size and/or high constellation systems. The

proposed precoding design method is applied to a 2×2 MIMO-OFDM system with

2048 subcarriers by designing 1024 precoders for extended channel matrices of size

4×4. A transceiver test bed implements these precoding matrices in comparison

with other existing precoding schemes. Indoor experiments are conducted for fixed-

platform non-line-of-sight (NLOS) channels, and the data processing results show

that the proposed precoding method achieves the lowest BER compared to maximum

diversity, classic water-filling and channel diagonalization methods.

1.3 REFERENCES

[1] C. Xiao, Y. R. Zheng, and Z. Ding, “Globally optimal linear precoders for finite
alphabet signals over complex vector Gaussian channels,” IEEE Trans. on Signal
Process., vol. 59, pp3301-3314, July 2011.
 5

[2] Y. Xin, Z. Wang, and G. B. Giannakis, “Space-time diversity systems based

on linear constellation precoding,” IEEE Trans. Wireless Commun., vol. 2, pp.
294-309, Mar. 2003.

[3] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. New
York: Wiley, 2006.

[4] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, “Capacity limits of

MIMO channels,” IEEE J. Seleted. Areas Commun., vol.21, pp.684-702, Jun.
2003.

[5] D. P. Palomar, J. Cioffi and M. A. Lagunas, “Joint Tx-Rx beamforming design

for multicarrier MIMO channels: A unified framework for convex optimization,”
IEEE Trans. Signal Process., vol. 51, pp.2381-2401, Sep 2003.

[6] A. Lozano, A. M. Tulino, and S. Verdu, “Optimum power allocation for paral-
lel Gaussian channels with arbitrary input distributions,” IEEE Trans. Inform.
Theory, vol.52, pp.3033-3051, July 2006.

[7] F. Perez-Cruz, M. R. D. Rodrigues, and S. Verdu, “MIMO Gaussian channels

with arbitrary inputs: optimal precoding and power allocation,” IEEE Trans.
Inform. Theory, vol.56, pp.1070-1084, Mar. 2010.
 6

PAPER

I. LINEAR PRECODING FOR MIMO MULTIPLE ACCESS

CHANNELS WITH FINITE DISCRETE INPUTS

Mingxi Wang, Weiliang Zeng and Chengshan Xiao, Fellow, IEEE

ABSTRACT—In this paper, we study linear precoding for multiple-input multiple-

output (MIMO) multiple access channels (MAC) with finite discrete inputs. We

derive the constellation-constrained capacity region for the MIMO MAC with an

arbitrary number of users and find that the boundary can be achieved by solving

the problem of weighted sum rate maximization with constellation and individual

power constraints. Due to the non-concavity of the objective function, we obtain

a set of necessary conditions for the optimization problem through Karush-Kuhn-

Tucker analysis. To find the optimal precoding matrices for all users, we propose

an iterative algorithm utilizing alternating optimization strategy. In particular, each

iteration of the algorithm involves the gradient descent update with backtracking

line search. Numerical results show that when inputs are digital modulated signals

and the signal-to-noise ratio is in the medium range, our proposed algorithm offers

considerably higher sum rate than non-precoding and the traditional method which

maximizes Gaussian-input sum capacity. Furthermore, a low-density parity-check

coded system with iterative detection and decoding for MAC is presented to evaluate

the bit error rate (BER) performance of precoders. BER results also indicate that

the system with the proposed linear precoder achieves significant gains over the non-

precoding system and the precoder designed for Gaussian inputs.

 7

1 INTRODUCTION

The problem of linear precoding for multiple-input multiple-output (MIMO)

multiple access channels (MAC) has been investigated in the literature in the last

few years. The existing methods were proposed based on different criteria. Most

of them employed information theoretical analysis which finds the capacity region

of MIMO MAC. It is well known that the input signals achieving the boundary of

MIMO MAC capacity region are Gaussian distributed, and the capacity region only

depends on input covariance matrices [1,2,3]. The optimal input covariance matrices

can be found by maximizing the weighted sum rate, which is a convex optimization

problem with Gaussian inputs [4,5,6]. In particular, when only sum rate maximization

is considered, an effective algorithm called iterative water-filling (WF) algorithm is

developed in [3]. Other criteria in linear procoding design of MAC are also utilized.

For instance, [7] minimizes the mean square error (MSE) assuming a linear receiver

structure, and [8] maximizes the signal-to-interference and noise ratio (SINR) for an

iterative linear receiver.

However, there are some drawbacks of the aforementioned methods. Regard-

ing the optimization techniques using capacity with Gaussian signals, the transmit-

ted signals in practical digital communication systems are not Gaussian distributed,

but rather drawn from certain constellation sets such as phase-shift keying (PSK),

pulse-amplitude modulation (PAM), or quadrature amplitude modulation (QAM).

Therefore, precoders obtained based on Gaussian input assumptions may lead to per-

formance loss when the inputs are actually replaced by finite discrete inputs. On the

other hand, MSE minimization and SINR maximization approaches may not neces-

sarily provide minimum bit error rate (BER) or maximum date rate.
 8

To overcome the shortcomings of optimization through Gaussian capacity,

MSE, SINR, etc., the mutual information with finite discrete inputs has been em-

ployed for precoding design recently. This approach conforms to such an information

theoretical principle that the mutual information with certain input constraints de-

termines the potential achievable data rate of a communication system. In point-to-

point communication scenarios, the mercury/waterfilling power allocation and general

linear precoders are developed [9, 10, 12, 13, 14, 11]. Similar problems are also investi-

gated in relay networks [15,16] and broadcast channels [17,18]. For 2-user single-input

single-output (SISO) MAC, [19] found the optimal angle of rotation and designed the

code pairs based on trellis coded modulation. These papers have shown that adopt-

ing mutual information with finite discrete inputs for precoder design can achieve

considerable performance gains compared with existing methods which are based on

Gaussian input assumptions.

To the best of our knowledge, little research has been done on MIMO MAC

precoding based on mutual information with finite discrete inputs. In this paper,

the maximum mutual information with uniformly distributed finite discrete inputs

is referred to as constellation-constrained capacity [19, 20], while the maximized mu-

tual information with Gaussian inputs is called Gaussian capacity. We derive the

constellation-constrained capacity region for the MIMO MAC with an arbitrary num-

ber of users and find that the boundary can be achieved by solving the problem of

weighted sum rate maximization with constellation and individual power constraints.

Since the weighted sum rate is no longer a concave function of precoding matrices as

opposed to the case of Gaussian inputs [3], we obtain a set of necessary conditions

through Karush-Kuhn-Tucker (KKT) analysis [21]. To find optimal linear precoders

for all users, we propose an iterative algorithm utilizing alternating optimization strat-

egy with gradient descent update method. Furthermore, the backtracking line search

is adopted to determine the step size for fast convergence. Our method is guaranteed
 9

to local optimum, and we resort to multiple run of the algorithm with random ini-

tializations to search for a best possible final solution. Numerical results show that

the proposed algorithm converges fast under various signal-to-noise ratios (SNRs). In

addition, when inputs are digital modulated signals, and the SNR is in the medium

regime, our proposed algorithm offers much higher sum rate than non-precoding and

the traditional power allocation method which maximizes Gaussian-input sum capac-

ity.

Besides the sum rate and constellation-constrained capacity region, bit error

rate of a system over MAC is of significant interest in practice. We thus present a

multiuser system with low-density parity-check (LDPC) coding and linear precoding

for all transmitters. At the receiver, the soft maximum a posteriori (MAP) multiuser

detector and LDPC channel decoders are adopted to iteratively exchange the soft

information. Simulations show that the system with the proposed precoder achieves

significant SNR gains over the non-precoding system and the system with the precoder

designed under Gaussian assumptions.

Throughout the paper, we denote matrices with boldface upper-case letters,

and vectors with boldface lower-case letters. The superscripts (·)t and (·)h represent

transpose and conjugate transpose operations, respectively. In addition, kak and

kAkF means the Euclidean norm of vector a and the Frobenius norm of matrix A,

respectively. The determinant of matrix A is |A|, and log stands for the logarithm

with base 2. The symbol C denotes the complex number field, while E takes the

expectation of a random variable or function.

The rest of the paper is organized as follows. Section 2 describes the model

of MIMO MAC and a brief overview of the existing results on capacity region and

optimal linear precoding with Gaussian input signals. The constellation-constrained

capacity region of MIMO MAC with finite discrete inputs is derived in Section 3.

Section 4 discusses necessary condition of the weighed sum rate maximization problem
 10

and the details of the iterative algorithm. Section 5 presents the MIMO system over

MAC with iterative detection and decoding. Numerical results are provided in Section

6, and Section 6 draws the conclusions.

 11

2 SYSTEM MODEL AND EXISTING GAUSSIAN INPUTS RESULTS

Consider a K-user communication system with multiple antennas at trans-

mitters and the receiver over multiple access channels. The signal model is given

by

y = H1 G1 x1 + H2 G2 x2 + · · · + HK GK xK + v = HGx + v (1)

t
where H = [H1 , H2 , · · · , HK ], and x = [xt1 , · · · , xtK ] . G = Bdiag{G1 , · · · , GK },

where Bdiag means a block diagonal matrix. Therefore, H and G can be viewed as

the equivalent channel matrix and block diagonal precoding matrix for all users. Sup-

pose there are Nr antennas at the receiver, and each user has Nt transmit antennas.

The symbol Hi ∈ CNr ×Nt represents the complex channel matrix between the i-th

transmitter and the receiver. We assume that the receiver knows the channels of all

users, and each transmitter knows its own channel state information. Throughout

this paper, we constrain each user’s precoding matrix to be a square matrix, which

is denoted as Gi ∈ CNt ×Nt . The vector x ∈ CNt K×1 contains signals of all transmit-

ters, and y ∈ CNr ×1 is the received signal. The receiver noise v ∈ CNr ×1 is a zero

mean circularly symmetric complex Gaussian vector with covariance matrix σ 2 I, i.e.,

v ∼ CN (0, σ 2 I).

Assume all signal vectors xi of different users are independent from one an-

other, and elements of each xi are independent and identically distributed (i.i.d) with

unit energy, i.e., E[xi xhi ] = INt . The symbols in xi can be digital modulated signal

points such as PSK or QAM signals. The covariance matrix of transmitted signal of

user i is Qi = E[Gi xi xhi Ghi ] = Gi Ghi .

We now briefly review existing results of MIMO MAC based on Gaussian

inputs. For the K-user MAC, it is well known that the capacity region is the convex
 12

hull of the union of capacity pentagons, and the boundary of the capacity region
P
can be fully characterized by maximizing the weighted sum rate K i=1 µi Ri for all
PK
nonnegative µi . Assuming that µ1 ≥ · · · ≥ µK ≥ 0, and i=1 µi = K, then the

optimal covariance matrices which maximize the capacity region can be found through

solving the following optimization problem [2, 3]:

   
 K
X  K−1
X  l
X 
   
max µK log I + Hi Qi Hhi  + (µi − µi+1 ) log I + Hl Ql Hhl  (2)
Q1 ,··· ,QK    
i=1 i=1 l=i

subject to Tr(Qi ) ≤ Pi , Qi  0, i = 1, · · · , K, (3)

in which Qi is hermitian and positive semidefinite. The above problem is a convex

optimization problem, which can be solved by efficient numerical methods [21, 22].
 13

3 CONSTELLATION-CONSTRAINED CAPACITY REGION

We derive the constellation-constrained capacity region with finite discrete

inputs for MIMO MAC in this section. Let the set A and its complement Ac partition

all users into two groups, where A = {i1 , i2 , · · · , iK1 } ⊆ {1, 2, · · · , K}, and Ac =
t
{j1 , j2 , · · · , jK2 }, K1 + K2 = K. With the assumptions of xA = [xti1 , xti2 , · · · , xtiK ] ,
1
t
and xAc = [xtj1 , xtj2 , · · · , xtjK ] , it is known that the achievable rate region of K-user
2

MAC is the closure of the convex hull of the rate vectors (R1 , R2 · · · , RK ), which

satisfies the following constraints [1]:

X
Ri ≤ I (xA ; y|xAc ) , ∀A ⊆ {1, 2, · · · , K} (4)
i∈A

for some independent input distributions p (x1 ), p (x2 ), · · · , p (xK ).

In practical digital communication systems over multiple access channels, trans-

mitted signals are often equiprobably drawn from certain discrete constellations such

as PSK, PAM, or QAM. Assuming that Mi is the number of constellation points in

each component of xi , then the number of all possible vectors of xi is Ni = MiNt .

Assuming that HA = [Hi1 , Hi2 , · · · , HiK1 ] and GA = Bdiag{Gi1 , Gi2 , · · · , GiK1 }, we

have the following proposition, which generalizes the achievable rates of 2-user MAC

in [19].

Proposition 1. When the discrete signals xi of all users are independent and uni-

formly distributed, I(xA ; y|xAc ) is given as follows:

NA
1 X
I (xA ; y|xAc ) = log NA − Ev
NA i=1
" NA
!#
X −kHA GA xiA − xkA +vk2 + kvk2
log exp (5)
k=1
σ2
 14

Q
where v ∼ CN (0, σ 2 I), and NA = i∈A Ni . The symbol xiA represents one possible

signal vector from xA , and the number of all possible constellation points in vector xA

is NA .

Proof. From the definition I (xA ; y|xAc ) = H (y|xAc ) − H (y|xA , xAc ), we can

prove (5). Details can be found in Appendix 8.

According to [23], the boundary of the constellation-constrained capacity re-

gion can be characterized by the solution of the weighted sum rate optimization prob-

lem. Without loss of generality, we assume the weights µ1 ≥ · · · ≥ µK ≥ µK+1 = 0,

i.e., decoding user K first and user 1 last. Then, the weighted sum rate maximization

with finite discrete inputs is equivalent to the following optimization problem:

K
X
max g(G1, · · · , GK ) = ∆i f (G1 , · · · , Gi ). (6)
G1 ,··· ,GK
i=1

subject to Tr(Gi Ghi ) ≤ Pi , i = 1, 2, · · · , K (7)

where ∆i = µi − µi+1 ; i = 1, · · · , K; and f (G1 ,· · · ,Gi ) = I (x1 ,· · · ,xi ; y|xi+1,· · · ,xK ),

which can be obtained by (5). When µ1 = · · · = µK = 1, the weighted sum rate

maximization problem reduces to sum rate maximization.

 15

4 WEIGHTED SUM RATE MAXIMIZATION

In this section, we solve the problem of weighted sum rate maximization with

finite discrete inputs described in (6) and (7). We obtain a set of necessary conditions

for the optimization problem and then propose an iterative algorithm to find optimal

precoding matrices.

4.1 NECESSARY CONDITIONS

In general, the objective function g(G1, · · · , GK ) is nonconcave on precoding

matrices {G1 , · · · , GK }. Thus, the weighted sum rate maximization with finite dis-

crete inputs is not concave, and we can only find a set of necessary conditions for this

optimization problem, as given in the following proposition.

Proposition 2. The solution for the weighted sum rate maximization described in

(6) and (7) satisfies:

K
X
νi Gi = ∆j Hhi HAj GAj EiAj (8)
j=i


νi Tr Gi Ghi − Pi = 0 (9)

Tr(Gi Ghi ) − Pi ≤ 0 (10)

νi ≥ 0 (11)

for all i = 1, 2, · · · , K. Since the set Aj = {1, 2, · · · , j}, we have HAj = [H1 , H2 , · · · , Hj ]

and GAj = Bdiag{G1 , G2 , · · · , Gj }. The symbol EiAj ∈ CNt j×Nt stands for the i-th

column block of the minimum mean square error (MMSE) matrix [24] of EAj , which

is defined as

 h i h ih
EAj = E xAj − E xAj |y, xAcj xAj − E xAj |y, xAcj . (12)
 16

Proof. The Lagrangian for (6) and (7) is given by

K
X 

L(G, λ) = −g(G1, · · · , GK )+ λi Tr Gi Ghi − Pi (13)
i=1

∂L
in which λi ≥ 0, i = 1, · · · , K. Define the gradient ∇Gi L = ∂G∗
as in [25], then the

KKT conditions are as follows:

∇Gi L = −∇Gi g(G1 , · · · , GK ) + λi Gi = 0 (14)



λi Tr Gi Ghi − Pi = 0 (15)

Tr(Gi Ghi ) − Pi ≤ 0 (16)

λi ≥ 0 (17)

for all i = 1, 2, · · · , K.

Due to the relation between the mutual information and MMSE [26, 27], the

gradient of f (G1 , · · · , Gj ) can be found as:

log e h
∇GAj f (G1 , · · · , Gj ) = H HAj GAj EAj . (18)
σ 2 Aj

For j ≥ i, Gi is the i-th block of diagonal submatrices of GAj . We can write Gi

as (ei ⊗ INt )GAj (ehi ⊗ INt ), in which ei is the i-th row of the j-dimensional identity

matrix I. Then, we have

∇Gi f (G1 , · · · , Gj ) = (ei ⊗ INt )∇GAj f (G1 , · · · , Gj )(ehi ⊗ INt )

log e h
= 2
Hi HAj GAj EiAj (19)
σ
 17

where EiAj = EAj (ehi ⊗ INt ) ∈ CNt j×Nt is the i-th column block of the MMSE matrix

EAj . Substituting (19) to (14) and letting νi = λi σ 2 / log e, we can prove (8).

4.2 ITERATIVE ALGORITHM FOR WSR MAXIMIZATION

From (8), it can be seen that the optimal precoders of different users depend

on one another. A common approach to multidimensional optimization problem is

the alternating optimization method which iteratively optimizes one variable at a

time with others fixed [28, 7, 29]. We adopt this method to maximize the weighed

sum rate with finite discrete inputs. During each iteration of the algorithm, only one

user’s precoding matrix Gi is updated while others are fixed. For i-th user at n-th
e (n) based on the gradient of g(G1, · · · , GK ) with respect
iteration, we first generate G i

to Gi as follows

e (n) = G(n) + t∇G g(G(n) , · · · , G(n) )

G (20)
i i i 1 K

where t is the step size. e (n) k2 > Pi , we project G

If kG e (n) to the feasible set
i F i

Tr(GGh ) ≤ Pi to obtain the update [27]:

h i+ p
(n+1) e (n) e (n) /kG
e (n) kF .
Gi = G i = Pi G i i (21)
Tr(GGh )≤Pi

For fast convergence, we use backtracking line search [21] to determine the

step size t in gradient update. The two parameters in backtracking line search are

α, β with α ∈ (0, 0.5) and β ∈ (0, 1). Detailed steps of the proposed algorithm are

shown in Table 1.

Due to the non-concavity of the weighted sum rate g(G1 , · · · , GK ), the pro-

posed algorithm can only find local optimum. To reduce the chance of being trapped

in local maxima, we run the iterative algorithm with random initialization multiple

times and choose the one with maximal weighted sum rate to be the final solution [13].
 18

Table 1. The algorithm for weighted sum rate maximization with finite discrete inputs
(0)

initialize Gi with Tr Gi Ghi = Pi , i = 1, 2, · · · , K.
repeat
(n) (n) (n)
compute g (n) = g(G1 , · · · , GK ), and EAj for j = 1, · · · , K.
for i = 1 : K
(n) (n) e PK (n)
∇Gi g(G1 , · · · , GK ) = log σ2
h
j=i ∆j Hi HAj GAj (EAj )
i (n)
.
set step size t := 1.
do
Ge (n) = G(n) + t∇G g(G(n) , · · · , G(n) ).
i i 1 K
(n+1) √ (n)
i
(n) (n) 2
Gi e e e
= Pi Gi /kGi kF , if kGi kF > Pi .
compute g (n+1) based on
(n) (n) (n+1) (n) (n)
G(n+1) = Bdiag{G1 , · · · , Gi−1 , Gi , Gi+1 , · · · , GK }.
t := βt.
(n) (n)
while g (n+1) < g (n) +αtk∇Gi g(G1 ,· · · ,GK )k2F .
end
(n) (n)
until the g(G1 , · · · , GK ) converges or n reaches maximum iteration number.

We note that the complexity of the proposed algorithm is mainly due to com-

putations of g(G1 , · · · , GK ) and MMSE matrix EAj , j = 1, 2, · · · , K. When input

signals are Gaussian, the weighted sum rate has a simple analytical expression as in

(2). However, the computation of weighted sum rate with finite discrete inputs de-

mands more consideration and higher complexity. From (5) and (6), we can see that

the computation of g(G1 , · · · , GK ) involves summation of all possible transmitted

vectors from all users, and thus its complexity grows exponentially with Nt ·K. Since

it is generally very difficult to obtain a closed-form expression of g(G1 , · · · , GK ), we

use Monte Carlo simulation method to estimate its value. Such an approach has been

adopted in [13, 14] dealing with single user.

Similar to computing g(G1 , · · · , GK ), we argue that the MMSE matrix EAj

Q
can also be estimated via Monte Carlo simulation method. When NAj = i∈j Ni and
 19

p(xAj = xiAj ) = 1/NAj , the MMSE estimate of xAj is given by

N
Aj PNAj l l
X l=1 xAj p(y|xAj = xAj , xAj )
c
x̂Aj = E(xAj |y, xAcj ) = xlAj p(xlAj |y, xAcj ) = PNAj (22)
i
i=1 p(y|xAj = xAj , xAj )
l=1 c

where
!
1 ky−HAjGAj xlAj −HAcj GAcj xAcj k2
p(y|xAj = xlAj , xAcj ) = exp − . (23)
(πσ 2 )Nr σ2

Substitute (22) and (23) to (12), the MMSE matrix can be formulated to the expec-

tation of a function of complex Gaussian vector v as:

h ihP i
 PNA l NA k h 
1
NA
X  l=1 xAj qm,l (v) k=1 (xAj ) qm,k (v)

E A j = I Nt j − Ev hP i2 (24)
NA m=1   NA 

i=1 qm,i (v)

where the function qm,l (v) is defined as

   
m l
kHAj GAj xAj − xAj + vk2
qm,l (v) = exp − . (25)
σ2

Therefore, we can randomly generate Gaussian vectors v to obtain an estimate of

EAj by (24) and (25).

 20

5 ITERATIVE DETECTION AND DECODING FOR MAC

We have discussed the precoder design with finite discrete inputs from the

information theoretical perspective in previous sections. Yet, another major concern

in practical communication systems is the bit error rate or frame error rate. Therefore,

we provide a transmission scheme for the multiple access channels with multiple

antennas in this section. More specifically, all the transmitters adopt the LDPC

channel coding and linear precoders discussed in Section 4. At the receiver side,

the iterative processing technique involving the soft detection and channel decoder is

employed to achieve good performance. This type of transceiver structure has been

reported to have promising performance in various applications [30, 31, 32, 33].

Fig. 1. shows a bank of parallel transmitters of K users. The i-th user trans-

mits blocks of information bits, and each block ui is encoded by the LDPC encoder.

The coded bits ci are then interleaved and fed into the modulator. Since a squared

precoder is considered at each transmitter, we split the stream of the modulated

symbols into Nt independent streams by a serial to parallel converter. Finally, the

symbol xi is multiplied by the individual precoding matrix Gi , and transmitted to

the space through Nt antennas. We note that all users can separately use identical

LDPC encoders and interleavers, while the linear precoders may differ from each other

according to Section 4.

s1
x1
u1 LDPC c1 b1
User 1 Encoder Π Mod S/P G1

xK sK
uK LDPC cK bK
User K Encoder Π Mod S/P GK

Figure 1. MIMO uplink transmitters with precoding.

 21

The iterative receiver is given in Fig. 2. The information intended for all users

is iteratively exchanged between a MIMO MAC soft detector and a bank of LDPC

soft channel decoders. Within each iteration, the soft multiuser detector, dealing

with both cross-antenna and multiuser interferences, generates the extrinsic informa-

tion LE (bi ), i = 1, · · · , K, based on the received signal y and the priori information

LA (bi ), i = 1, · · · , K. The log likelihood ratio (LLR) LE (bi ) is then interleaved and

fed into the i-th user’s LDPC decoder as the intrinsic information LA (ci ). Those

soft decoding methods of LDPC codes, such as the log domain sum product algo-

rithm [34], should be adopted to exploit the redundancy among coded bits ci and

compute the intrinsic information LD (ci ). After channel decoding, LD (ci ) is sub-

tracted by LA (ci ), and interleaved to become the priori knowledge LA (bi ) of the

MAC detector for use in the next iteration. At the final iteration, hard decisions are

made upon LD (ci ), i = 1, · · · , K to obtain estimate of information of all users.

LE(b1) LA(c1) û1

Π−1 LDPC
Decoder
Hard
decision

Π
y MIMO MAC LA(b1) LE(c1) LD(c1)
Soft detector

LE(bK) LA(cK) ûK

Π−1 LDPC
Decoder
Hard
decision

Π
LA(bK) LE(cK) LD(cK)

Figure 2. Iterative receiver of MIMO multiple access channel.

There are two main categories of soft multiuser detection methods: linear and

non-linear detection. An recent overview of iterative linear detection can be found

in [35]. Although the linear processing has the advantage of low complexity and

easy implementation, its performance is usually non-optimal. In addition, when the

number of total transmit antennas Nt K exceeds the number of receiver antennas

 22

Nr , the linear detection may not have the sufficient capability to tackle the MAC

interference.

On the other hand, the non-linear detections include the MAP method [30,36],

soft interference cancelation [36], sphere decoding [19], and Markov chain Monte Carlo

(MCMC) approach [37], etc. Among these methods, the MAP detection can achieve

the optimal performance while others are sub-optimal or near-optimal. Since our

main goal in this section is to verify that the precoder designed via weighed sum rate

maximization can also provide excellent BER performance, we choose to implement

the MAP method for the MIMO MAC soft detector in Fig. 2 to obtain optimal BER

results. For each received vector y, the extrinsic LLR LE (bi ) can be given as [30]

P h
i
1 t
b∈Bk,+1 p (y|b) exp b L
2 [k] A
b[k]
LE (bi,j ) = ln P h
i (26)
1 t
b∈Bk,−1 p (y|b) exp b L
2 [k] A
b[k]

where bi,j means the j-th bit of the i-th user’s bit vector bi , with 1 ≤ i ≤ K,

and 1 ≤ j ≤ Mc Nt , assuming that all users employ the same modulation and the

number of the constellation points of each modulated symbol is Mc . The vector

t
b = [bt1 , · · · , btK ] with length Mc Nt K contains the interleaved bits from all users.

The vector bt[k] denotes the subvector of b with the k-th element omitted, in which


k = (i − 1) Mc Nt + j. The vector LA b[k] with (Mc Nt K − 1) elements represents

the priori information of b[k] . The sets Bk,+1 and Bk,−1 denote the sets of 2Mc Nt K−1

bit vectors b with the k-th element equaling to +1 and −1, respectively. The channel

likelihood function in (26) is given by

 
1 ky − Hxk2
p (y|b) = p [y|x = map (b)] = exp − (27)
(πσ 2 )Nr σ2

where x = map (b) means the mapping from the bit vector b to symbol vectors x,

including modulation and S/P conversions of all users.

 23

6 NUMERICAL RESULTS

In this section, we provide numerical results of the constellation-constrained

capacity region and sum rate with finite discrete inputs for 2-user MAC. We also

simulate BER performance of the system described in Section 5. Assume that there

are two receiver antennas and two transmit antennas for each user. Suppose the

maximum individual power P1 = P2 = P , and all users adopt the same modulation

scheme. Then, the signal to noise ratio can be defined as SNR = P/σ 2 , when the

channels are normalized. In our simulations, we choose the noise power σ 2 = 1.

For illustrative purpose, we consider an example of two fixed channel matrices

for two users, which are given by

   
 1.3898 0.1069j   1.2247 0 
H1 =   , H2 =  .
−0.1069j 0.2138 0 0.707

Each channel matrix has normalized power with Tr Hi Hhi = Nr = 2, as in [10].

Fig. 3 plots the convergence behavior of the sum rate maximization algorithm

in Table 1 with BPSK inputs. At each SNR, we run the algorithm with random

initializations 10 times and choose the one with the largest sum rate at the end of

iterations. From the figure, we can see that the proposed algorithm usually converges

after 15 iterations under different SNRs. For backtracking line search, the typical

range of parameters α and β are α ∈ (0.01, 0.3), and β ∈ (0.1, 0.8) [21]. We choose

α = 0.1 and β = 0.5, for the algorithm in Section 4. The Monte Carlo simulation

number for both the sum rate and MMSE matrix is set to 500. In general, a limited

simulation number in Monte Carlo method leads to a certain degree of estimate errors.

This is the reason why there are small ripples of sum rate in Fig. 3.
 24

3.5
SNR = 5dB
SNR = 0dB
3
SNR = −5dB
2.5

Sum Rate
2

1.5

0.5

0
0 10 20 30 40 50 60 70 80
Iteration Index

Figure 3. Convergence of sum rate maximization algorithm with BPSK inputs.

Fig. 4 shows the sum rate of various precoding schemes with BPSK modula-

tion. We implement the Gaussian-input sum capacity maximization method [2] and

replace the inputs with BPSK inputs, which is denoted as “BPSK, Gaussian sum

capacity max”. In this method, the optimal input covariance matrices {Q1 , · · · , QK }

are obtained by maximizing Gaussian-input sum capacity described in (2) and (3).

After using standard convex optimization tool [22] to solve this problem, we de-

compose each covariance matrix as Qi = Vi Σi Vih , and choose the precoder to be

1
Gi = Vi Σi2 . Then, we replace Gaussian inputs to BPSK signals and calculate the

sum rate of this precoding scheme using (5). A noticeable result of Gaussian-input

sum capacity maximization method is that within certain SNR range, each user’s

covariance matrix Qi has only one positive eigenvalue due to the water-filling policy.

For example, when SNR is 20 dB, the covariance matrices obtained by the traditional
 25

method are
   
 98.63 11.61j   2.64 −16.04j 
Q1 =   , Q2 =  .
−11.61j 1.37 16.04j 97.36

With eigenvalue decomposition Qi = Vi Σi Vih , we find that Σ1 = Σ2 = diag {100, 0},

and the precoding matrices are as follows:

   
 −9.93 0   −1.63 0 
G1 =   , G2 =  .
1.17j 0 −9.87j 0

In this case the precoder acts as beamforming by allowing each user to transmit only

one modulated symbol in vector xi . Therefore, although the traditional method can

achieve sum capacity with the ideal Gaussian inputs assumption, usually it fails to

serve as the optimal strategy for practical finite discrete inputs.

4
Sum Rate (bits/s/Hz)

Gaussian, Iterative WF
Gaussian, No Precoding
1
BPSK, Optimal Precoding
BPSK, No Precoding
BPSK, Gaussian Sum Capacity Max
0
−20 0 20 40 60 80
SNR (dB)

Figure 4. Sum rate of 2-user MAC with BPSK inputs.

 26

For comparison purpose, we plot the Gaussian-input sum capacity achieved by

iterative water-filling [3] and sum rate of Gaussian inputs without precoding. From

the numerical results, we have several observations: our proposed algorithm, denoted
q
Pi
as the “optimal precoding”, outperforms the non-precoding (i.e., Gi = N t
INt ) with

finite discrete inputs for a wide SNR range from −15 dB to 20 dB. For instance,

the SNR gain of the optimal precoding over non-precoding to achieve the sum rate

3 bits/s/Hz is about 8 dB. When SNR approaches infinity, the sum rate of both

methods saturate at 4 bits/s/Hz, which is determined by the constellation size, the

number of users and transmit antennas. In addition, we find that when the SNR

is less than 0 dB, the optimal precoding with BPSK inputs obtains the same sum

rate as the iterative WF with Gaussian inputs. For the SNR range below 5 dB, our

method also has performance gains compared to sum rate of Gaussian inputs without

precoding or power allocation.

The precoding matrices obtained via optimal precoding method when SNR

= 5 dB are  
 1.22 − 0.048i −0.21 + 1.20i 
Gopt
1 =  
0.12 − 0.17i 0.37 − 0.10i

and  
 −0.56 − 0.81i 0.66 − 0.04i 
Gopt
2 =  .
0.44 + 0.36i 0.47 + 1.10i

The sum rate results regarding QPSK inputs are shown in Fig. 5. We have

similar observations as the BPSK inputs, in the sense that the proposed precoding

with QPSK inputs achieves higher sum rate than non-precoding for the SNR range

from 0 dB to 25 dB, and obtains the same sum rate of iterative WF with Gaussian

inputs when SNR is less than 5 dB. At the target sum rate of 3 bits/s/Hz, the optimal

precoding achieves SNR gains of about 5 dB and 40 dB, compared to non-precoding

and Gaussian-input sum capacity maximization method, respectively.

 27

12

10

Sum Rate (bits/s/Hz)

6

Gaussian, Iterative WF
Gaussian, No Precoding
2
QPSK, Optimal Precoding
QPSK, No Precoding
QPSK, Gaussian Sum Capacity Max
0
−20 0 20 40 60 80
SNR (dB)

Figure 5. Sum rate of 2-user MAC with QPSK inputs.

The constellation-constrained capacity region of optimal precoding with BPSK

inputs when SNR = 5 dB is illustrated in Fig. 6. When inputs are BPSK signals,

the rate regions achieved by non-precoding and Gaussian-input sum capacity max-

imization schemes are also plotted. The pentagons are determined by equations in

Proposition 1. The curve of optimal precoding is obtained by varying µ1 and µ2

using weighted sum rate maximization algorithm with discrete inputs in Section 4.

We can see that the constellation-constrained capacity region of optimal precoding is

much larger than the rate regions of non-precoding and Gaussian-input sum capacity

maximization method.

Fig. 7 plots regions of QPSK signals when SNR = 10 dB. Similar to the results

of BPSK signals, the rate regions achieved by non-precoding and Gaussian-input sum

capacity maximization are inside the constellation-constrained capacity region with

QPSK inputs, which is obtained by the proposed precoding algorithm.

 28

2.5
BPSK, Optimal Precoding
BPSK, Gaussian Sum Capacity Max
BPSK, No Precoding
2

R2 (bits/s/Hz)
1.5

0.5

0
0 0.5 1 1.5 2
R1 (bits/s/Hz)

Figure 6. Capacity region of 2-user MAC with BPSK inputs, when SNR = 5dB.

In addition to the sum rate and rate region, we also simulate the BER perfor-

mances of the 2-user MAC system described in Section 5. The LDPC encoder and

decoder simulation package [38] is used. The length of each codeword is set to 9600,

and coding rate is 3/4. All users separately employ identical LDPC codes and pseu-

dorandom interleavers. At the receiver side, the multiuser detector uses MAP method

expressed in (26), and LDPC decoders adopt the sum-product algorithm [34] with 30

iterations. The soft information exchange between the MAP multiuser detector and

LDPC channel decoders. The iteration number between MAP detector and LDPC

decoder is set to 5.

Fig. 8 plots the BER curves of BPSK inputs under the fixed channels. The

results include optimal precoding, non-precoding, and Gaussian-input sum capacity

maximization method, which show that that the optimal precoding achieves signif-

icant gains over the other methods. From the sum rate results in Fig. 4, we can

see that when the channel coding rate is 3/4 (sum rate of 3 bits/s/Hz), the SNRs of
 29

5
QPSK, Optimal Precoding
4.5
QPSK, Gaussian Sum Capacity Max
QPSK, No Precoding
4

3.5

R2 (bits/s/Hz)
3

2.5

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3 3.5 4
R1 (bits/s/Hz)

Figure 7. Capacity region of 2-user MAC with QPSK inputs, when SNR = 10dB.

the optimal precoding and non-precoding schemes are about 3 dB and 11 dB, respec-

tively. From information theoretical perspective, these SNR limits are the minimum

acceptable SNRs for error-free communication. At the target BER of 10−4 , the SNR

of optimal precoding is about 4 dB, which is close to the limit predicted by the sum

rate vs. SNR curve. In addition, the performance gains of optimal precoding over

non-precoding and Gaussian-input sum capacity maximization well match the results

observed in Fig. 4. The comparison of sum rate and BER results justifies that the

precoding method of using sum rate with finite discrete inputs can not only maxi-

mize the achievable information rate, but also achieve excellent system performance

in terms of bit error rates.

Fig. 9 shows the BER performance with QPSK inputs. We have similar

observations as the results of BPSK signals. The simulations indicate that to achieve

the BER of 10−4 , the optimal precoding method outperforms non-precoding by 6 dB,
 30

0
10
BPSK, Optimal Precoding
BPSK, No Precoding
−1
10 BPSK, Gaussian Sum Capacity Max

−2
10

BER
−3
10

−4
10

−5
10

−6
10
0 10 20 30 40 50 60
SNR (dB)

Figure 8. BER of 2-user MAC with BPSK inputs.

which is nearly the same amount of SNR gain when the sum rate is 6 bits/s/Hz in

Fig. 5.

0
10
QPSK, Optimal Precoding
QPSK, No Precoding
−1 QPSK, Gaussian Sum Capacity Max
10

−2
10
BER

−3
10

−4
10

−5
10

−6
10
0 10 20 30 40 50 60
SNR (dB)

Figure 9. BER of 2-user MAC with QPSK inputs.

 31

7 CONCLUSION

In this paper, we studied the linear precoder design for MIMO MAC with

finite discrete inputs. From the information theoretical perspective, we derived the

constellation-constrained capacity region. We then found a set of necessary conditions

of weighted sum rate maximization with individual power constraints and proposed

an iterative algorithm to obtain the optimal precoding matrices for all users. The

convergence behavior of our algorithm has been verified by simulations. We have

shown that when inputs are digital modulated signals, and SNR is in the medium

range, our precoding method offers significantly higher sum rate than non-precoding

and the existing Gaussian-input sum capacity maximization approach. An LDPC

coded system with iterative detection and decoding for MAC was further presented

to evaluate the BER performance of such precoders. BER simulation results indicated

that the optimal precoding achieves significant SNR gain against the non-precoding

system and the system with Gaussian-input sum capacity maximization precoders.
 32

8 APPENDIX: PROOF OF PROPOSITION 1

The priori probabilities of xA and xAc are p (xA = xiA ) = N1A and p xAc = xkAc =
1
Q Q
NAc
, where N A = i∈A N i and N A c =
i∈Ac Ni . Based on the Gaussian vector

channel model y = HA GA xA + HAc GAc xAc + v, the conditional probability density

function of y can be written as

 

1 ky−HA GA xiA1 −HAc GAc xiA2c k2
p y|xA = xiA1 , xAc = xiA2c = exp − . (28)
(πσ 2 )Nr σ2

The conditional entropy H(y|xAc ) can be calculated as

NAc
X


H (y|xAc ) = p xAc = xiAc H y|xAc = xiA2c
i2 =1
NAc Z
1 X


=− p y|xAc = xiA2c log p y|xAc = xiA2c dy
NAc i =1
2
" #
N Ac Z NA
1 X 1 X

=− p y|xA = xiA1 , xAc = xiA2c
NAc i =1 NA i =1
2 1
" N
#
1 X A
k1 i2


log p y|xA = xA , xAc = xAc dy. (29)
NA
k1 =1

Substituting (28) to the above equation and assuming y − HA GA xiA1 − HAc GAc xiA2c =

v, we have

NA
" NA
!#
1 X X 1 kHA GA xiA1 −xkA1 +vk2
H(y|xAc ) = log NA − Ev log exp − ,
NA i =1
1
(πσ 2 )Nr k1 =1
σ2

(30)

where v is a complex Gaussian vector with probability density function p (v) =

 
1 kvk2
N
(πσ2 ) r
exp − σ 2 .
 33

Similarly, we can get H(y|xA , xAc ) as follows:

  
1 kvk2
H (y|xA , xAc ) = Ev log exp − 2 . (31)
(πσ 2 )Nr σ

From (30) and (31), we can prove Proposition 1.

 34

9 REFERENCES

[1] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. New
York: Wiley, 2006.

[2] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, “Capacity limits of

MIMO channels,” IEEE J. Seleted. Areas Commun., vol.21, pp.684-702, Jun.
2003.

[3] W. Yu, W. Rhee, S. Boyd, and J. M. Cioffi, “Iterative water-filling for Gaussian
vector multiple-access channels,” IEEE Trans. Inf. Theory, vol. 50, pp. 145-152,
Jan. 2004.

[4] M. Kobayashi and G. Caire, “Iterative waterfilling for the weighted rate sum
maximization in MIMO MAC,” in Proc. IEEE Signal Processing Advances in
Wireless Communications (SPAWC 2006), Cannes, July 2006.

[5] M. Kobayashi and G. Caire, “An iterative water-filling algorithm for maximum
weighted sum-rate of Gaussian MIMO-BC,” IEEE J. Sel. Areas Commun., vol.
24, no. 8, pp. 1640-1646, Aug. 2006.

[6] H. Viswanathan, S. Venkatesan, and H. Huang, “Downlink capacity evaluation

of cellular networks with known-interference cancellation,” IEEE J. Sel. Areas
Commun., vol. 21, no. 5, pp. 802-811, June 2006.

[7] S. Serbetli and A. Yener, “Transceiver optimization for multiuser MIMO sys-
tems,” IEEE Trans. Signal Process., vol. 52, no. 1, pp. 214-226, Jan. 2004.

[8] X. Yuan, C. Xu, and X. Lin, “Precoder design for multiuser MIMO ISI channels
based on iterative LMMSE detection,” IEEE J. Sel. Topics Signal Process., vol.
3, pp. 1118-1128, Dec. 2009.

[9] A. Lozano, A. M. Tulino, and S. Verdú, “Optimum power allocation for paral-
lel Gaussian channels with arbitrary input distributions,” IEEE Trans. Inform.
Theory, vol.52, pp.3033-3051, July 2006.

[10] C. Xiao and Y. R. Zheng, “On the mutual information and power allocation
for vector Gaussian channels with finite discrete inputs,” in Proc. IEEE Glocal
Telecommunications Conference (Globecom), New Orleans, USA, Nov. 30 - Dec.
4, 2008.

[11] M. Payaró and D. P. Palomar, “On optimal precoding in linear vector Gaussian
channels with arbitrary input distribution, ” in Proc. of 2009 IEEE international
conference on Symposium on Information Theory (ISIT), Seoul, Korea, June,
2009.
 35

[12] M. Lamarca, “Linear precoding for mutual information maximization in MIMO

systems,” in Proc. of 6th International Symposium on Wireless Communication
Systems, Siena, Italy, 2009.

[13] F. Pérez-Cruz, M. R. D. Rodrigues, and S. Verdú, “MIMO Gaussian channels

with arbitrary inputs: optimal precoding and power allocation,” IEEE Trans.
Inform. Theory, vol.56, pp.1070-1084, Mar. 2010.

[14] C. Xiao, Y. R. Zheng, and Z. Ding, “Globally optimal linear precoders for finite
alphabet signals over complex vector Gaussian channels,” IEEE Trans. on Signal
Process., vol. 59, pp3301-3314, July 2011.

[15] W. Zeng, M. Wang, C. Xiao, and J. Lu, “On the power allocation for relay
networks with finite-alphabet constraints,” in Proc. IEEE Glocal Telecommuni-
cations Conference (Globecom), Miami, USA, Dec. 2010.

[16] W. Zeng, Y. R. Zheng, M. Wang, and J. Lu, “Linear precoding for relay networks:
a perspective on finite-alphabet inputs,” submitted to IEEE Trans. on Wireless
Commun., Dec. 2010.

[17] N. Deshpande and B. S. Rajan, “Constellation constrained capacity of two-

user broadcast channels,” in Proc. IEEE Glocal Telecommunications Conference
(Globecom), Honolulu, HI, USA, 30 Nov. - 4 Dec. 2009.

[18] R. Ghaffar and R. Knopp, “Near optimal linear precoder for multiuser MIMO for
discrete alphabets,” in Proc. IEEE International Conference on Communications
(ICC), Cape Town, South Africa, May 23-27, 2010.

[19] J. Harshan and B. S. Rajan, “On two-user Gaussian multiple access channels
with finite input constellations,” IEEE Trans. Inform. Theory, vol.57, pp.1299-
1327, Mar. 2011.

[20] E. Biglieri, Coding for wireless channels, Springer-Verlag New York, Inc, 2005.

[21] S. P. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University

Press, 2004.

[22] M. Grant and S. Boyd. CVX: Matlab software for disciplined convex program-
ming, version 1.21., http://cvxr.com/cvx, Jan. 2011.

[23] D. N. C. Tse and S. Hanly, “Multi-access fading channels: Part I: Polymatroid

structure, optimal resource allocation and throughput capacities,” IEEE Trans.
Inform. Theory, vol. 44, pp. 2796-2815, Nov. 1998.

[24] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory.

Englewood Cliffs, NJ: Prentice-Hall, 1993.

[25] A. Hjørungnes, “Complex-valued matrix derivatives: with applications in signal

processing and communications,” Cambridge University Press, 2011.
 36

[26] D. Guo, S. Shamai, and S. Verdú, “Mutual information and minimum mean-
square error in Gaussian channels,” IEEE Trans. Inform. Theory, vol. 51, pp.
1261-1282, Apr. 2005.

[27] D. P. Palomar and S. Verdú, “Gradient of mutual information in linear vector

Gaussian channels,” IEEE Trans. Inform. Theory, vol.52, pp.141-154, Jan. 2006.

[28] D. Bertsekas and J. Tsitsiklis, Parallel and Distributed Computation: Numerical

Methods. Englewood Cliffs, NJ: Prentice-Hall, 1989.

[29] A. Soysal and S. Ulukus, “Optimum power allocation for single-user MIMO and
multi-user MIMO-MAC with partial CSI,” IEEE J. Sel. Areas Commun., vol.25,
pp.1402-1412, Sept. 2007.

[30] B. Hochwald and S. T. Brink, “Achieving near-capacity on a multiple-antenna

channel,” IEEE Trans. Commun., vol. 51, pp. 389-399, Mar. 2003.

[31] M. Tüchler, R. Koetter, and A. Singer, “Turbo equalization: principles and new
results,” IEEE Trans. Comm., vol. 50, pp. 754-767, May 2002.

[32] X. Wang and V. Poor, “Iterative (turbo) soft interference cancellation and de-
coding for coded CDMA,” IEEE Trans. Commun., vol.47, pp.1046-1061, Jul.
1999.

[33] J. Ylioinas and M. Juntti, “Iterative joint detection, decoding, and channel es-
timation in Turbo-coded MIMO-OFDM”, IEEE Trans. Veh. Technol., vol. 58,
pp.1784-1796, May. 2009.

[34] X-Y. Hu, E. Eleftherious, D-M. Arnold, and A. Dholakia, “Efficient implemen-
taions of the sum-product algorithm for decodng LDPC codes,” in Proc. IEEE
Glocal Telecommunications Conference (Globecom), San Antonio, USA, Nov. 25
- 29, 2001.

[35] L. Rasmussen, “Linear detection in iterative joint multiuser decoding,”, in Proc.

of 3rd International Symposium on Communications, Control and Signal Pro-
cessing (ISCCSP), St Julians, Malta, Mar. 12-14, 2008.

[36] L. Xu, S. Chen, and L. Hanzo, “EXIT chart analysis aided Turbo MUD designs
for the rank-deficient multiple antenna assisted OFDM uplink,” IEEE Trans. on
Wireless Commun., vol.7, pp.2039-2044, Jun. 2008.

[37] S. Henriksen, B. Ninness, and S. R. Weller, “Convergence of Markov-chain

Monte-Calro approaches to multiuser and MIMO detection,” IEEE J. Seleted.
Areas Commun., vol.26, pp.497-505, Apr. 2008.

[38] M. Valenti, The Coded Modulation Library, http://www.iterativesolutions.com.

 37

II. ON THE POWER ALLOCATION FOR RELAY NETWORKS

WITH FINITE ALPHABET CONSTRAINT

Weiliang Zeng, Mingxi Wang, Chengshan Xiao, Fellow, IEEE, and Jianhua Lu

ABSTRACT—In this paper, we investigate the optimal power allocation scheme for

relay networks with finite-alphabet constraints. It has been shown that the previous

work utilizing various design criteria with the Gaussian inputs assumption may lead to

significant loss for a practical system with finite constellation set constraint, especially

when signal-to-noise ratio (SNR) is in medium-to-high regions, or when the channel

coding rate is medium to high. An optimal power allocation scheme is proposed to

maximize the mutual information for the relay networks under discrete-constellation

input constraint. Numerical examples show that significant gain can be obtained

compared to the conventional counterpart for nonfading channels and fading channels.

At the same time, we show that the large performance gain on the mutual information

will also represent the large gain on the bit-error rate (BER), i.e., the benefit of the

power allocation scheme predicted by the mutual information can indeed be harvested

and can provide considerable performance gain in a practical system.

 38

1 INTRODUCTION AND RELATED WORK

Cooperative relaying has been shown to provide reliable high data rate services

in wireless networks without the need of multiple antennas at each node. These

benefits can be further exploited by utilizing judicious cooperative strategies see [1,

2, 3, 4, 5, 6, 7, 8, 9, 10, 11] and the references therein.

The existing design methods may be categorized into two groups: i) diversity

oriented designs; and ii) transmission rate oriented designs. The first group usually

achieves the steepest asymptotic slope (the highest diversity order) on the outage

probability versus SNR curve, however, it may not obtain the highest possible coding

gain, such as the distributed space-time coding (DST) in [2, 3] and the relaying se-

lection scheme in [4, 5]. The second group often optimizes the performance with the

Gaussian inputs assumption, for example, maximizing output SNR [6, 7, 8, 9], mini-

mizing mean square error (MSE) [10, 6] and maximizing channel capacity [6, 7, 11].

Although Gaussian inputs are capacity-achieving signaling, they can never be

realized in practice. Rather, the inputs must be drawn from a finite constellation

set (such as pulse amplitude modulation (PAM), quadrature amplitude modulation

(QAM) and phase shift keying (PSK) modulation) in a practical communication sys-

tems, which may significantly depart from the Gaussian idealization [12, 13, 14]. Yet,

no solution has been found in the above work for the power allocation that maximizes

the potential transmission rate, i.e., mutual information, with non-Gaussian inputs

over the relay networks, and this is exactly the concern of this study.

For multiple-input multiple-output (MIMO) systems, it has been shown that

the design from the standpoint of finite alphabet can result in significant performance

improvement [12]. In this paper, we will see similar performance gains achieved in re-

lay networks over the existing schemes utilizing the design criteria such as SNR, MSE,
 39

and channel capacity. At the same time, it has been validated that the large per-

formance gain predicted by the mutual information with finite-alphabet constraints

can indeed be harvested and will lead to considerable performance improvement in a

practical system.

The rest of the paper is organized as follows. In Sec. II, we introduce the relay

network model and the main problem. In Sec. III, we analyze the solution structure

of the power allocation problem from the information theoretic and optimization

theoretic point of view. We also utilize the receiver structure that has the near-

capacity performance to validate the effectiveness of the proposed scheme. Sec. IV

presents numerical results, followed by the conclusions in Sec. V.

Notation: Throughout this paper, we use boldface upper-case letters to denote

matrices, boldface lower-case letters to denote column vectors, and italics to denote

scalars. The superscripts (·)T and (·)H stand for transpose and conjugate transpose,

respectively; [A]i,j and [A]:,j denote the (ith, jth) element and jth column of matrix

A, respectively; and kck denotes Euclidean norm of vector c. I denotes the identity

matrix with the appropriate dimensions; diag(c) denotes the diagonal matrix with

diagonal elements given by the vector c; Tr (A) denotes the trace operation; < denotes

the real part of complex number; and E denotes statistical expectation. Likewise, all

logarithms are to the base 2.

 40

2 SYSTEM MODEL AND PRELIMINARIES

Consider a relay network with one transmit-and-receive pair, where the source

node (s) attempts to communicate to the destination node (d) with the assistance of

k relays (r1 , r2 , · · · , rk ). The link from the source to the ith relay is denoted as hi , the

link from the ith relay to the destination is denoted as gi , and the direct link from

the source to the destination is denoted as h0 .

We model the source-relay (S-R), source-destination (S-D) and relay-destination

(R-D) links as the quasi-static flat-fading channels, which are applicable to the scenar-

ios of narrow-band transmissions in a low-mobility environment. We assume that the

ith relay knows its own channels hi and gi , and the destination obtains full knowledge

of S-D, S-R and R-D channels.

We consider the average power constraint of each node for each time slot,

e.g., the power used at the source and the ith relay should be less than Ps and Pr ,

respectively.

The data transmission is over two time slots using two hops. The symbols

transmitted by the source node in the first and second time slot are denoted as x1 and

x2 , respectively. They may be chosen from some complex-valued finite constellation

C. We assume that E[xi ] = 0 and E[|xi |2 ] = 1 for i = 1, 2. During the first time slot,
√
the source node sends α0 Ps x1 . Let yri and yd,1 be received signals at the ith relay

node and the destination, respectively, which are given by

p
yri = α0 Ps hi x1 + nri , (1)
p
yd,1 = α0 Ps h0 x1 + nd,1 , (2)
 41

where nri and nd,1 are complex additive white Gaussian noise at the ith relay and the

destination with zero mean and unit variance ∼ CN (0, 1).

q  
The ith relay node normalizes the received signal by a factor of E |yri |2

(so that the average energy is unity) and retransmits the signal

s
Pr
ti =  αi yri , i = 1, 2, · · · , k (3)
E |yri |2

√
during the second time slot. At the same time, the source node sends αk+1 Ps x2 .

Then the destination node receives a superposition of the relay transmissions and the

source transmission during the second time slot according to

k
X p
yd,2 = gi ti + αk+1 Ps h0 x2 + nd,2
i=1
k
s
X Ps Pr p
= α0 αi hi g i x1 + Ps αk+1 h0 x2 + v, (4)
i=1
1 + Ps |α0 hi |2

Pk Pr |αi gi |2
where the effective noise v ∼ CN (0, Nd ) with Nd = 1+ i=1 1+Ps |α0 hi |2 . We normalize
1/2
yd,2 by a factor w = Nd in order to simplify the presentation. Finally, the effective

input-output relation for the two-hop transmission can be summarized as

y = Gx + n, (5)

where y = [yd,1 yd,2 /w]T is the received signal vector, and G is the effective channel

matrix given by

 
√
 Ps α0 h0 0 
G= P q q . (6)
k Ps Pr Ps
α αhg α h
i=1 (1+Ps |α0 hi |2 )w2 0 i i i w 2 k+1 0
 42

x = [x1 x2 ]T is the transmitted signal vector, and n ∈ C2×1 is the channel noise vector,

assumed independent and identically distributed (i.i.d.) complex Gaussian with zero

mean and unit variance, i.e., n ∼ CN (0, I). Equivalently, we can rewrite the effective

channel G as HP with the channel related matrix

 
√
 Ps h0 0 ··· 0 0 
H= √ √ √ , (7)
0 Ps Pr h1 g1 · · · Ps Pr hk gk Ps h0

and the power allocation related matrix

 T
 γ0 γ1 · · · γk 0 
P=  , (8)
0 0 · · · 0 γk+1

q

where γ0 = α0 ; γi = α0 αi / 1 + Ps |α0 hi |2 w 2 , ∀i = 1, · · · , k; and γk+1 = αk+1 /w.

We should note that the coefficients α0 , α1 , · · · , αk+1 are complex value, and

the rationale of introducing them in the model is in fact quite intuitive. First, they

can be used to control the average transmit power of each node at each time slot,

which requires that

αi αi∗ ≤ 1 ∀i = 0, · · · , k + 1. (9)

Hence the average power used at the ith relay node is |αi |2 Pr . Second, the choice of

the angles can be used to cancel the phases introduced from the channels and ensure

that the signal components are added constructively at the receiver, i.e., arg αi =

− (arg hi + arg gi ), i = 1, · · · , k, [7, 8, 10]. We also set arg α0 = arg αk+1 = − arg h0 ,

since the optimal choice of both angles can be realized by rotating the input constel-

lations equivalently. What is left is the choice of their magnitude, so we will treat the

power allocation coefficients α0 , α1 , · · · , αk+1 as a set of real numbers in the sequel.

Our power allocation scheme is thus the design of coefficients α0 , α1 , · · · , αk+1

to maximize the mutual information with finite-alphabet constraints. Note that for
 43

the proposed algorithm to be effectively implemented in practice, a low-rate feedback

should be allowed from the destination to the source and relay nodes. The feedback is

needed in the cooperative system since antennas are not located at a single terminal

as in a MIMO system. This may result in small penalty on system performance, but

the cost is often compensated by a significant performance gain at high SNR [1, 14].
 44

3 POWER ALLOCATION FOR FINITE ALPHABET INPUTS

We consider the conventional equiprobable discrete signaling constellations

such as M-ary PSK, PAM, or QAM, where M is the number of points in the signal

constellation. The mutual information between x and y, with H and P known at the

receiver, is I (x; y) given by (10) [12],

M 2 ( M2 )
1 X X  
I (x; y) = log M − En log exp −kHP (xm − xk ) + nk2 + knk 2
.
2M 2 m=1 j=1
(10)

where x contains two symbols, taken independently from the M-ary signal

constellation.

The problem that we pose is the determination of the coefficients α0 , α1 , · · · , αk+1

that maximizes the mutual information I (x; y) with given input distributions while

satisfying individual power constraints, i.e.,

max I (x; y) (11)

α0 ,··· ,αk+1

subject to:

αi ≤ 1 ∀i = 0, · · · , k + 1. (12)

3.1 ANALYSIS FROM INFORMATION THEORETIC PERSPECTIVE

Applying the chain rule for mutual information [15], we have

I (x; y) = I (x2 ; y) + I (x1 ; y|x2 ) , (13)

where I (x2 ; y) is the mutual information between x2 and y, and I (x1 ; y|x2 ) is the

conditional mutual information between x1 and y given x2 . The vector y is defined

 45

in (5), which can also be written as:

y = [G]:,1 x1 + [G]:,2 x2 + n. (14)

From (14), we can verify that

I (x; y) = I (x2 ; y) + I (x1 ; y|αk+1 = 0) , (15)

where I (x1 ; y|αk+1 = 0) is the mutual information between x1 and the received signal

y if the source node does not transmit at the second time-slot.

Since I (x2 ; y) > 0 for h0 , αk+1 6= 0, it follows that to maximize the mutual

information I (x; y), the source should always transmit at the second time slot [11].

For the same reason, the source should transmit at the first time slot. Based on the

above discussions, we can state the following lemma:

Lemma 1. The power allocation in the two time slots at the source node is nonzero,

i.e., α0 , αk+1 6= 0, if the channel between the source and the destination node h0 is

nonzero.

3.2 ANALYSIS FROM OPTIMIZATION THEORETIC PERSPECTIVE

We should note that the constraint (12) is convex in the coefficient matrix

C = diag([α0 α1 · · · αk+1 ]). The cost function (11), however, is nonconcave in the

power allocation coefficient αi for the general case [16]. In the sequel, we capitalize

on the relationship between the mutual information and the minimum mean square

error (MMSE) matrix to obtain the power allocation scheme for arbitrary input dis-

tributions.
 46

 
Theorem 1. The optimal power allocation coefficients α0? α1? · · · αk+1
?
that solves

(11) subject to (12) satisfy:


∂I (x; y) 
= λi (16)
∂αi αi =α?
i

λi (αi? − 1) = 0 (17)

λi ≥ 0 (18)

with
k+1
∂I (x; y) X ∂I (x; y) ∂γj
= , (19)
∂αi j=0
∂γ j ∂αi

and

∂I (x; y)  
= < HH HPE i+1,1 , ∀i = 0, · · · , k (20)
∂γi
∂I (x; y)  
= < HH HPE k+2,2 (21)
∂γk+1

where E is the MMSE matrix defined by

n o
E = E [x − E (x|y, H, P)] [x − E (x|y, H, P)]H (22)

Proof. The possible solution to (11) subject to (12) is characterized by the

Karush-Kuhn-Tucker theorem [17], which gives necessary conditions, known as the

KKT or first order conditions. To investigate stationary points of the problem (11)

we formulate the Lagrangian

k+1
X
L (P, λ) = −I (x; y) + λi (αi − 1) , (24)
i=0

in which the Lagrangian multipliers λi , i = 1, · · · , k+1, are chosen to satisfy the power

constraints. Then the first order conditions are given by (16) to (18). The partial
 47

derivative ∂I/∂αi , i = 0, · · · , k + 1, can be proved by employing the techniques

developed in [18, 19] for derivatives of mutual information, the techniques developed

in [20] for matrix differentiation, and the chain rule for multiple variables.

Typically, it is involved to calculate the MMSE matrix (22), especially for large

input dimensions M. But we have been able to estimate the matrix E using Monte

Carlo methods. Hence, we can solve this problem using gradient-based methods in

Table 1 according to the gradient of mutual information (19). Since the cost function

(11) is nonconcave for the general case, it is possible that (11) has local maxima.

Therefore, we should perform the algorithm with multiple initial points and keep the

power allocation coefficients offering the largest mutual information.

Table 1. The algorithm for optimal power allocation coefficients

initialize 0 ≤ αi ≤ 1, i = 0, 1, · · · , k + 1.
repeat
for i = 1 : k
set step size t = 1.  
(n)
(n) ∂I αi
(n)
compute g = ∂αi , in (19).
(αi )
(n+1) (n) (n)
αi = αi + t · g (n) (αi ).
(n+1) (n+1)
set αi = 1, if αi > 1.
choose t by backtracking line search.
end  
(n) (n) (n)
until I α0 , α1 , · · · , αk converges or n reaches maximum iteration number.

Finally, the destination node will notify the source and each relay node of

its assigned transmission power. In this way, the instantaneous mutual information

is maximized for each set of channel realizations. Notice that the resulting opti-

mal power allocation scheme is significantly different from the existing ones in the

conventional setting both due to the presence of finite-alphabet constraints and the
 48

multiplexing structure capitalized in the relay networks. The results in section 4 show

the significant gains obtained compared to the existing methods.

3.3 ITERATIVE DETECTION AND DECODING

To evaluate the advantage of the proposed method in a more practical may,

we utilize the “turbo principle” at the destination node [21,22], which is illustrated in

Fig. 1. The signal sequence b at the source node is encoded by the capacity achiev-

able codes, e.g., low-density parity-check (LDPC) codes, interleaved, and mapped

according to the conventional equiprobable discrete signaling constellations. Then it

is divided into x1 and x2 , and transmitted at the two time slots, respectively.

At the receiver, the maximum a posteriori (MAP) detector takes channel ob-

servations y and a priori knowledge Le (b) from the decoder and computes new in-

formation Le (c) for each of the coded bits. In this way, the extrinsic information

between the MAP detector and decoder is exchanged in an iterative fashion until

desired performance is achieved. It has been shown that the iterative processing is

very effective that can achieve near-capacity performance.

 49

4 NUMERICAL RESULTS

Computer simulation was carried out to validate the performance of the pro-

posed scheme. For the sake of completeness, in all the figures, we show the per-

formance corresponding to MMSE strategy with local power constraint in [10] and

network beamforming in [7]. To ensure a fair comparison, we also show the per-

formance of modified MMSE strategy and modified network beamforming (sending

different symbols at the second time slot, rather than sending the same symbols or

sleeping). We consider a three-relay network with the same transmit power at the

source and each relay node, i.e., Ps = Pr = P , which is indicated by the horizontal

axis in the following figures.

We look first at a fixed (non-fading) system with the channel coefficient h0 =

0.5, h = [0.7, −0.7, 1j] and g = [0.9j, 2.1, 0.3]. The instantaneous mutual infor-

mation that can be achieved by different schemes is shown in Fig. 2, in which the

Relay Channel

interleaver modulation
Power Channel
LDPC Allocation Related
Mod
Encoder

Feedback
AWGN

hard decision deinterleaver

LDPC -1 MAP
Decoder Detector

interleaver
-

Figure 1. Block diagram of the relay network with iterative receiver at the destination.
 50

information symbol x is modulated as quadrature phase shift keying (QPSK). From

Fig. 2, we have several observations. First, the performance loss of the MMSE and

network beamforming is small in the low SNR region, and large in the high SNR

region. This is because both schemes maximize the power gain, which is much im-

pressive compared to the degree-of-freedom gain at low SNR [23]. At high SNR,

however, a degree-of-freedom gain is much more important, which is not provided by

the original MMSE and network beamforming schemes [7, 10], since the source node

transmits the same symbols for the two time slots. Hence, the mutual information is

bounded by 1 bit/s/Hz. Second, the modified MMSE and modified network beam-

forming are not bounded by 1 bit/s/Hz, since we remove the constraint that sends

the same symbols at the two time slots. Moreover, the proposed power allocation

method results in significant gain on mutual information when SNR is in medium-

to-high regions, or when the channel coding rate is medium to high. For example, it

is about 4dB and 8dB better than the modified network beamforming and modified

MMSE scheme when the channel coding rate is 3/4.

In Fig. 3, we show that the large performance gain on the mutual information

will also represent the large performance gain on the BER. To validate the advantage

of the power allocation scheme, we realize the simulation model illustrated in Sec. 3.3.

The coding length is 1800; the coding rate is 3/4; and the iteration between the MAP

detector and the LDPC decoder is 5. Then we compare the optimal power allocation

with modified MMSE and modified network beamforming. It is worth noting that the

benefit of the power allocation scheme predicted by the mutual information can indeed

be realized, and the power allocation coefficients that are “blessed” by the mutual

information formula (10) can provide considerable performance gain in a practical

system.

Then we work on the Rayleigh fading channel, and consider the average mu-

tual information achieved by different methods. We assume the channels of S-R

 51

1.8

1.6

1.4

Mutual Information (bits/s/Hz)

1.2

0.8

0.6

0.4
MMSE in [10]
Modified MMSE
0.2 Netw. Beamf. in [7]
Modified Netw. Beamf.
Opt. Power Alloc.
0
−5 0 5 10 15 20 25 30
SNR (dB)

Figure 2. Mutual Information for a fixed channel with QPSK inputs.

−1
10

−2
10
limit 12.9dB
limit 9dB

limit 17dB

7 dB
BER

−3
10

10 dB

−4
10

Modified MMSE
Modified Netw. Beamf.
Opt. Power Alloc.
−5
10
8 10 12 14 16 18 20 22
SNR (dB)

Figure 3. Bit error performance comparison for optimal power allocation, modified
MMSE and modified network beamforming.
 52

and R-D have the same average SNR, and are 10dB better than the S-D channel,

considering the practical deployment of the relay nodes. Fig. 4 depicts the average

mutual information of the relay network with QPSK inputs. The MMSE and network

beamforming saturates very quickly, while the modified ones perform much better.

However, they still have about 3dB to 15dB loss compared to the optimal power

allocation when the channel coding rate is 3/4.

1.8

1.6

1.4
Mutual Information (bits/s/Hz)

1.2

0.8

0.6

0.4
MMSE in [10]
Modified MMSE
0.2 Netw. Beamf. in [7]
Modified Netw. Beamf.
Opt. Power Alloc.
0
−5 0 5 10 15 20 25 30
SNR (dB)

Figure 4. Average mutual Information over Rayleigh fading channels with QPSK
inputs.
 53

5 CONCLUSION

In this paper, we have studied the optimal power allocation for dual-hop wire-

less relay networks. In contrast with the previous work utilizing various design crite-

ria with the unrealistic Gaussian inputs assumption, the proposed scheme attempts

to maximize the mutual information for the relay networks from the standpoint of

discrete-constellation inputs. To determine the optimal power allocation policy, we

capitalized on the relationship between mutual information and MMSE. Numerical

examples have shown that significant gains can be obtained compared to the con-

ventional counterpart for nonfading channels and fading channels, especially when

SNR is in medium-to-high regions, or when the channel coding rate is medium to

high. Likewise, it has been shown that the large performance gain on the mutual

information can represent the large gain on the bit-error rate, i.e., the benefit of the

power allocation scheme predicted by the mutual information can indeed be realized

and can provide considerable performance gain in a practical system.

 54

6 REFERENCES

[1] Y. Hong, W. Huang, F. Chiu, and C. Kuo, “Cooperative communications in

resource-constrained wireless networks,” IEEE Signal Process Mag., vol. 24,
no. 3, p. 47, 2007.

[2] J. N. Laneman and G. W. Wornell, “Distributed space-time-coded protocols for

exploiting cooperative diversity in wireless networks,” IEEE Trans. Inf. Theory,
vol. 49, no. 10, pp. 2415–2425, Oct. 2003.

[3] Y. Jing and B. Hassibi, “Distributed space-time coding in wireless relay net-
works,” IEEE Trans. Wireless Commun., vol. 5, no. 12, p. 3524, 2006.

[4] A. Bletsas, A. Khisti, D. P. Reed, and A. Lippman, “A simple cooperative di-

versity method based on network path selection,” IEEE J. Sel. Areas Commun.,
vol. 24, no. 3, pp. 659–672, March 2006.

[5] W. Zeng, C. Xiao, Y. Wang, and J. Lu, “Opportunistic relaying for multi-antenna
cooperative decode-and-forward relay networks,” in Proc. IEEE ICC, 2010.

[6] A. S. Behbahani, R. Merched, and A. M. Eltawil, “Optimizations of a MIMO

relay network,” IEEE Trans. Signal Process., vol. 56, no. 10, pp. 5062–5073,
2008.

[7] Y. Jing and H. Jafarkhani, “Network beamforming using relays with perfect
channel information,” IEEE Trans. Inf. Theory, vol. 55, no. 6, pp. 2499–2517,
Jun. 2009.

[8] G. Zheng, K. Wong, A. Paulraj, and B. Ottersten, “Collaborative-relay beam-

forming with perfect CSI: Optimum and distributed implementation,” IEEE Sig-
nal Process Lett., vol. 16, no. 4, pp. 257–260, 2009.

[9] A. Gershman, N. Sidiropoulos, S. Shahbazpanahi, M. Bengtsson, and B. Otter-

sten, “Convex optimization-based beamforming: From receive to transmit and
network designs,” IEEE Signal Process Mag., Special Issue on Convex Optim.
for Signal Process, 2010.

[10] N. Khajehnouri and A. Sayed, “Distributed MMSE relay strategies for wireless
sensor networks,” IEEE Trans. Signal Process., vol. 55, no. 7, p. 3336, 2007.

[11] R. Nabar, H. Bolcskei, and F. Kneubuhler, “Fading relay channels: Performance

limits and space-time signal design,” IEEE J. Sel. Areas Commun., vol. 22, no. 6,
pp. 1099–1109, 2004.
 55

[12] C. Xiao and Y. R. Zheng, “On the mutual information and power allocation for
vector gaussian channels with finite discrete inputs,” in Proc. IEEE Globecom,
Nov. 2008, pp. 1–5.

[13] A. Lozano, A. Tulino, and S. Verdu, “Optimum power allocation for parallel
gaussian channels with arbitrary input distributions,” IEEE Trans. Inf. Theory,
vol. 52, no. 7, pp. 3033–3051, 2006.

[14] G. Caire and K. Kumar, “Information theoretic foundations of adaptive coded

modulation,” Proc. IEEE, vol. 95, no. 12, pp. 2274–2298, 2007.

[15] T. Cover and J. Thomas, Elements of information theory, 2nd ed. Wiley, 2006.

[16] F. Pérez-Cruz, M. Rodrigues, and S. Verdú, “Optimal precoding for digital sub-
scriber lines,” in Proc. IEEE ICC, 2008.

[17] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge university

press, 2004.

[18] D. Guo, S. Shamai, and S. Verdú, “Mutual information and minimum mean-
square error in gaussian channels,” IEEE Trans. Inf. Theory, vol. 51, no. 4, pp.
1261–1282, 2005.

[19] D. Palomar and S. Verdú, “Gradient of mutual information in linear vector gaus-
sian channels,” IEEE Trans. Inf. Theory, vol. 52, no. 1, pp. 141–154, 2006.

[20] J. Magnus and H. Neudecker, Matrix differential calculus with applications in

statistics and econometrics, 2nd ed. Wiley, 1999.

[21] B. Hochwald and S. Ten Brink, “Achieving near-capacity on a multiple-antenna

channel,” IEEE Trans. Commun., vol. 51, no. 3, pp. 389–399, 2003.

[22] R. Koetter, A. Singer, and M. Tuchler, “Turbo equalization,” IEEE Signal Pro-
cess Mag., vol. 21, no. 1, pp. 67–80, 2004.

[23] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. Cam-

bridge University Press, 2005.
 56

III. LINEAR PRECODING FOR RELAY NETWORKS: A

PERSPECTIVE ON FINITE-ALPHABET INPUTS

Weiliang Zeng, Yahong Rosa Zheng, Mingxi Wang, and Jianhua Lu

ABSTRACT—This paper considers the precoder design for dual-hop amplify-and-

forward relay networks and formulates the design from the standpoint of finite-

alphabet inputs. In particular, the mutual information is employed as the utility

function, which, however, results in a nonlinear and nonconcave problem. This paper

exploits the structure of the optimal precoder that maximizes the mutual information

and develops a two-step algorithm based on convex optimization and optimization

on the Stiefel manifold. By doing so, the proposed algorithm is insensitive to initial

point selection and able to achieve a near global optimal precoder solution. Besides,

it converges fast and offers high mutual information gain. These advantages are ver-

ified by numerical examples, which also show the large performance gain in mutual

information also represents the large gain in the coded bit-error rate.
 57

1 INTRODUCTION

Relaying technology is promising to provide reliable communication, high

throughput, and broad coverage for wireless networks. These benefits can be achieved

with judicious designs that exploit network configurations and/or relaying strategies

(see [1, 2, 3, 4, 5, 6, 7, 8, 9] and references therein). The existing design methods for

improving the end-to-end transmission rate and reliability of the relay networks may

be categorized into two groups: diversity oriented designs and data-rate oriented de-

signs. The diversity oriented designs try to obtain the highest diversity order, by

reaching the steepest asymptotic slope of either the outage capacity or uncoded bit

error rate (BER) versus signal-to-noise ratio (SNR) curve, but are not necessary to

achieve the highest coding gain. Examples of such designs include the distributed

space-time coding in [1, 2] and the precoder design for maximal diversity in [3, 4].

The data-rate oriented designs try to obtain the highest data rate of each end-to-

end source-destination pair by maximizing the output SNR [5, 6] or maximizing the

network capacity [5, 7, 8, 9] with Gaussian input assumption.

This paper considers two-hop relay networks employing simple relays that are

equipped with single antenna and adopt the amplify-and-forward (AF) strategy [1].

It focuses on the design of precoder and the selection of relay node that maximize the

mutual information of one source-destination pair. The channel model of such a relay

network becomes very similar to that of a standard multiple-input multiple-output

(MIMO) channel and the existing precoding methods developed for MIMO settings

are applicable to the relay networks. However, most of the existing works in precoder

design use extensively the Gaussian input assumption so that the mutual information

between the transmit and receive signals is a simple and elegant function of the

precoder and channel matrix and the optimization problem becomes easier to solve
 58

than that of finite-alphabet inputs. A practical system usually utilizes finite-alphabet

constellations, such as pulse amplitude modulation (PAM), phase shift keying (PSK)

modulation, or quadrature amplitude modulation (QAM). The mutual information of

the discrete-input systems depart significantly from that of the Gaussian inputs [10,

11, 12, 13]. Therefore, applying the precoder/relaying schemes designed for Gaussian

inputs to discrete-input systems results in a significant performance loss from those

designed directly for finite-alphabet inputs [14].

Several recent works target on the maximization of mutual information with

finite-alphabet inputs and complex precoding matrices. Lozano et. al. [10] constrains

the channel and precoding matrices to be diagonal such that the mutual information

becomes a concave function of the squared precoder and the global maximal can be

solved. The work in [15] discovers the linear relationship between the gradients of mu-

tual information and the minimum mean squared error (MMSE) versus SNR curves

and proposes a gradient descent (GD) method to solve for the precoder iteratively.

The GD method is then further explored in [16]. However, as will be shown later in

this paper, the GD method is very sensitive to initial point selection and can easily get

stuck in a local maximal because the mutual information is a non-concave function

of the precoder. More recently, the work in [13] derives the asymptotic mutual infor-

mation expressions for relay networks in large-system regime with several parameters

approaching infinity and provides necessary conditions that the optimal precoder sat-

isfies. Hessian and concavity results are developed in [17] for real-valued channels, but

with no precoder design. A possible design is presented in [18] that uses the analysis

with real-valued channel assumption, without rigorous proof, in the complex-valued

channel case. The work in [19] proposes a two-layer iterative algorithm that finds the

precoding matrix by iterations between the bottom layer (i.e., the precoding matrix)

and top layer (i.e., the compound channel-precoding matrix) using the concavity of
 59

the mutual information on the compound channel-precoding matrix. The global op-

timal solutions can be achieved for some combinations of discrete constellations and

MIMO configurations.

This paper proposes a novel two-step iterative algorithm and a new framework

for linear precoder design with finite-alphabet inputs by exploring the structure of

the precoder that maximizes the mutual information. The proposed method first

separates precoder and channel matrices, by the singular value decomposition (SVD),

into product of the left singular vectors, diagonal power allocation matrix, and right

singular vectors. Then making use of the results that the left singular vectors of

the optimal precoder coincide with the right singular vectors of the effective channel

matrix [20, Appendix 3.B] and that the mutual information is a concave function on

the squared singular values of the linear precoder, the proposed algorithm maximizes

the mutual information by first designing the power allocation matrix with a given

set of the right singular vectors then optimizing the right singular vectors with the

obtained power allocation matrix. The algorithm iterates through the two steps until

the mutual information is maximized.

The success of the two-step iteration is based on the result that the mutual

information is a concave function on the squared singular values of the linear pre-

coder for a complex-valued channel. Although optimizing the right singular vectors

is extremely difficult even for real-valued channel and precoder matrices [21], we re-

formulate the complex-valued problem on the complex Stiefel manifold and solve it by

the gradient method with projection. The proposed two-step algorithm is applied to

a two-hop relay network and the maximization of the end-to-end mutual information

with multiple relay nodes is also considered by relay selection. Our simulation exam-

ples achieve 68% and 38% of mutual information improvement over no precoding for
 60

Binary PSK (BPSK) and Quadrature PSK (QPSK) systems, respectively, and sim-

ilar performance gain is expected to be achievable when applied to standard MIMO

channels.

The proposed precoder design algorithm has several advantages over the exit-

ing works. First, the proposed algorithm is applicable to general AF relay networks

and complex-valued MIMO systems with arbitrary combinations of constellations

and antenna configurations. It contains the unitary matrix based precoder design

method [3, 4]) as its special cases. It can also handle the special combinations of the

discrete constellations and MIMO configurations addressed in [19] and achieve almost

the same performance. Second, the proposed algorithm is insensitive to initial point

selection and it converges much faster than the GD method. Third, the proposed

algorithm, although suboptimal, can reach most of the maximal capacity predicted

by Gaussian inputs at high probability when using finite-alphabet inputs and random

initial points.

The remainder of this paper is organized as follows: Section II introduces the

system model for the two-hop relaying scheme. The properties of mutual information

are addressed in Section III. Section IV proposes the precoding scheme to maximize

the mutual information, which includes the design of the left singular vectors, the

power allocation matrix by convex optimization, and the right singular vectors using

optimization on the complex Stiefel manifold. Section V presents several numerical

examples to demonstrate the performance gain of this scheme over the existing ones.

Finally, Section VI offers conclusions.

Notation: Real and complex spaces are denoted by R and C,

I respectively.

Boldface uppercase letters denote matrices, boldface lowercase letters denote column

vectors, and italics denote scalars. The superscripts (·)T , (·)∗ , (·)H , and (·)+ stand

for transpose, complex conjugate, Hermitian, and Moore-Penrose pseudoinverse op-

erations, respectively. The scalar with subscript ci denotes the i-th element of vector
 61

c, whereas [A]i,j and [A]:,j denote the (i, j)-th element and j-th column of matrix

A, respectively; Diag(a) represent a diagonal matrix whose nonzero elements are

given by the elements of vector a; vec (A) represents the vector obtained by stacking

the columns of A; I and 0 represent identity matrix and zero matrix of appropriate

dimensions, respectively. The Kronecker matrix product is represented by A ⊗ B;

Tr (A) denotes the trace operation; E denotes statistical expectation; log and ln are

used, respectively, for the base two logarithm and natural logarithm.
 62

2 SYSTEM MODEL

Consider a relay network with one transmit-and-receive pair, where the source

node communicates to the destination node with the assistance of k relays, r1 , r2 , · · · , rk .

All nodes are equipped with a single antenna and operate in the half-duplex mode.

The relaying transmission system is assumed to be flat fading. The channel gain

from the source to the destination is denoted by h0 , and those from the source to

the i-th relay and from the i-th relay to the destination are denoted as hi and gi ,

respectively. These channel gains are assumed to remain unchanged during a period

of observation. The relay system adopts the two-hop AF protocols [7] combined with

single-relay selection [22]. The signals are transmitted in blocks with block length

2L, where L ≥ 1. The selected relay node receives in the first period of length L and

transmits in the second period of length L.

 T
The original signal at the source node is denoted by x = xTa xTb , where

xa = [x1 , · · · , xL ]T and xb = [xL+1 , · · · , x2L ]T with xl being the symbol at the l-th

time slot for l = 1, · · · , 2L. It is assumed to be equally probable from a discrete

constellation set, such as PSK, PAM, or QAM, with a unit covariance matrix, i.e.,
 
E xxH = I.

The original signal is processed by a precoding matrix before being transmitted

 T
from the source node. The precoded data s = sTa sTb can be written as

   
 sa   xa 
s=  = P  (1)
sb xb

where P ∈ C2L×2L is a matrix to be designed to improve the end-to-end performance.

√
The source node transmits the signal Ps sa with power Ps during the first

L time slots. Let yi and ya be the received signals at the i-th relay node and the
 63

destination, respectively. We have

p
yi = Ps hi sa + ni (2)
p
ya = Ps h0 sa + na (3)

where ni and na denote the independent and identically distributed (i.i.d.) zero-

mean circularly Gaussian noise vector with unit variance at the i-th relay and the

destination, respectively.

Assume that the i-th relay node is selected for the information forwarding in

the second time slot and that the relay knows only the second-order statistics of hi .

The selected relay node scales the received signal by a factor b, which guarantees that
 

the average transmit power of the i-th relay b2 Tr E yi yiH /L is less than or equal

to the power constraint Pr . If the channel gains are assumed to have unit variance,
p
then b can be chosen as Pr / (1 + 2Ps ).
√
At the same time, the source node sends the signal Ps sb . Hence, the des-

tination node receives the superposition of the relay transmission and the source

transmission during the second time slot:

p p p
yb = bgi yi + Ps h0 sb + nb = Ps bhi gi sa + Ps h0 sb + ne (4)

where nb denotes the noise vector of the destination at the second time slot, and ne

denotes the effective end-to-end noise with complex Gaussian distribution CN (0, Nd I)

and Nd = 1 + b2 |gi |2 .
√
For convenience of presentation, yb is normalized by a factor w = 1/ Nd , and
 T
the received signal vector for the two time slots is denoted as y = yaT wybT . Thus,

the effective input-output relationship for the two-hop transmission with precoding
 64

is summarized as

y = Hs + n = HPx + n (5)

 T
where x is the original transmitted signal vector; n = nTa wnTe ∼ CN (0, I); H is

the effective channel matrix of the two-hop relay channel,

 
p  h0 I 0 
H= Ps   (6)
wbhi gi I wh0 I

which is full rank for any nonzero channel gain h0 .

The precoding matrix P is thus designed to maximize the mutual information

with finite-alphabet inputs. Note that to effectively implement the proposed algo-

rithm in practice, a low-rate feedback shall be allowed from the destination to the

relays and to the source, respectively, for delivering information of node selection and

precoding. The feedback will be discussed in Section 4.4.

 65

3 MUTUAL INFORMATION FOR RELAY NETWORKS

Consider conventional equiprobable discrete constellations such as M-ary PSK,

PAM, or QAM, where M is the number of points in the constellation set. The mutual

information between the input x and the output y, with the equivalent channel matrix

H and the precoding matrix P known at the receiver, is I (x; y) given by [12]

M 2L M 2L
1 X X
I (x; y) = log M − En log exp(−dmk ) (7)
2LM 2L m=1 k=1

where dmk is kHP (xm − xk ) + nk2 − knk2 and k · k denotes the Euclidean norm of

a vector. Both xm and xk contain 2L symbols, taken independently from the M-ary

signal constellation.

The mutual information I (x; y) is fully determined by the distribution of

kHP (xm − xk ) + nk2 for m, k ∈ {1, · · · , M 2L }, which remain unchanged when a

unitary transform U is applied on the output signal y because a unitary matrix is an

isometry for the Euclidean norm. That is,

I (x; y) = I (x; Uy) . (8)

However, if the linear transform is applied on the input signal, then I (Ux; y) may

change to a different value, even though the transmit power is not altered. That is,

I (Ux; y) 6= I (x; y) . (9)

Note that the property of mutual information for the discrete input vector is

different from the case of Gaussian inputs. For Gaussian inputs, the mutual infor-

mation I G (x; y) is unchanged when either the transmitted signal x or the received
 66

signal y is rotated by a unitary matrix:

I G (x; y) = I G (Ux; y) = I G (x; Uy) . (10)

The case of finite inputs does not follow the same rule; thus, a new opportunity is

available here to improve system performance by transforming the input signal.

The 2L × 2L complex precoding matrix P implements linear transform on x,

thus can exploit property (9) to maximize the mutual information of relay systems

with finite-alphabet inputs. The optimization problem is formulated as:

maximize I (x; y)
  
(11)
subject to Tr E ssH = Tr PPH ≤ 2L

which is difficult to solve because it is nonconcave over P [21]. The next section

provides a new algorithm for this problem.

 67

4 PRECODER DESIGN TO MAXIMIZE MUTUAL INFORMATION

This section examines the properties of the precoding matrix P under finite-

alphabet inputs and presents several new results for complex-valued relay channels.

These results are the foundation for the development of the proposed algorithm, which

maximizes the mutual information.

4.1 OPTIMAL LEFT SINGULAR VECTORS

The first step is to characterize the dependence of mutual information I (x; y)

on the precoding matrix P. Given the signal constellation and the SNR, I (x; y) is a

function of the following variable:

h
i
2 H H H H H H H
kHP (xm − xk ) + nk = Tr êmk êmk P H HP + 2< êmk P H n + nn (12)

where êmk = xm − xk , and < denotes the real part of a complex number; I (x; y)

changes based on the distribution of kHP (xm − xk ) + nk2 , which depends on P

through PH HH HP: the first term of the right-hand side depends on P through

PH HH HP; the second term êH H H

mk P H n is a Gaussian random variable determined

by its zero mean and its variance êH H H

mk P H HPêmk , which also depends on P through

PH HH HP; the last term is independent of the precoding matrix. Therefore, I (x; y)

depends on P through PH HH HP, which is called the compound channel-precoding

matrix.
H
Consider the SVD of the 2L × 2L channel matrix H = UH Diag(σ)VH ,

where UH and VH are unitary matrices, and the vector σ contains nonnegative

entries in decreasing order. We also decompose the precoding matrix by SVD as

 68

√ H
P = UP Diag( λ)VP , where the vector λ is nonnegative constrained by the trans-
H
mit power. To simplify notation, we define U = UP and V = VP , where U and V

are called left and right singular vectors of P, respectively.

For a given matrix PH HH HP and a specific mutual information value, the

structure of the precoding matrix can be used to minimize the transmit power Tr(PPH ),

as shown in Appendix 3.B of [20], by letting the left singular vectors of P coincide
√
with the right singular vectors of H, that is, P = VH Diag( λ)V. Similar results

based on real-valued channels are reported in [21, 18]. In other words, for a given

matrix PH HH HP and a specific power constraint, letting the left singular vectors

of P equal to the right singular vectors of H maximizes the mutual information for

general channel conditions and arbitrary inputs.

Adopting the optimal left singular vectors U = VH , the channel matrix (5)

can be further simplified by (8) to

√
y = Diag(σ)Diag( λ)Vx + n. (13)

It is clear from (13) that the mutual information is now dependent only on the squared

singular values of the precoder λ and V. We use the notations I(λ) and I(V) to de-

velop a two-step optimization algorithm that first maximizes the mutual information

via optimal power allocation λ, and then via optimal right singular vectors V.

4.2 OPTIMAL POWER ALLOCATION

Given the right singular vectors of the precoder, the optimization problem (11)

is addressed over λ:

maximize I(λ)


subject to Tr PPH = 1T λ ≤ 2L (14)

λ0
 69

where 1 and 0 denote the column vector with all entries being one and zero, respec-

tively.

We extend the Hessian and concavity results of real-valued case in [17, Theo-

rem 5] to a complex-valued channel and precoder

Proposition 1. Based on the simplified channel model in (13), the mutual informa-

tion is a concave function of λ; that is, the Hessian of mutual information satisfies

HλI (λ)  0. Moreover, the gradient and Hessian of the mutual information are

given, respectively, as

∇λI(λ) = R · vec Diag2 (σ)VEVH (15)

and

1   n ∗
o
HλI(λ) = − R I ⊗ Diag2 (σ) E Φ̃(y) ⊗ Φ̃(y)
2
h √ √ i T
Diag( λ)Diag (σ) ⊗ Diag( λ) R Diag−1 (λ)
2

1  2
 n ∗
o
− R I ⊗ Diag (σ) E Ψ̃(y) ⊗ Ψ̃(y)
2
h √ √ i
Diag( λ) ⊗ Diag( λ)Diag2 (σ) KRT Diag−1 (λ) (16)

where E is the minimum mean square error (MMSE) matrix, defined as

n o
E , E [x − E (x|y)] [x − E (x|y)]H ;

and Φ̃(y) = VΦ(y)VH and Ψ̃(y) = VΨ(y)VT with Φ(y) and Ψ(y) being the MMSE

matrix and companion MMSE matrix conditioned on a realization of the received

signal y, defined by

n o
Φ(y) , E [x − E{x|y}] [x − E{x|y}]H |y
 70

and
n o
T
Ψ(y) , E [x − E{x|y}] [x − E{x|y}] |y

2
respectively. And R ∈ R2L×4L is a reduction matrix with entries given by [R]i,2L(j−1)+k =

δijk .

Proof. See Appendix A.

The concavity property in Proposition 1 ensures that a global optimal power

allocation vector can be found given right singular vectors V. In addition, the gradient

and Hessian results in (15) and (16) permit either the steepest descent or Newton-

type algorithms to solve for the global optimal power allocation vector. However, the

existing general-purpose solvers for convex problem (e.g., CVX [23]) fail to address

this problem because of the complexity of the objective function I (x; y). Therefore,

a specialized interior-point method is developed here.

The first step is to re-write problem (14), making the inequality constraints

implicit in the objective function:

P
2L
minimize f (λ) = −I (λ) + φ(−λi ) + φ(1T λ − 2L) (17)
i=1

where φ(u) is the logarithmic barrier function approximating an indicator as whether

constraints are violated:




−(1/t) ln(−u), u < 0
φ(u) =


+∞, u≥0

in which the parameter t > 0 sets the accuracy of the approximation [24].
 71

Based on Proposition 1, the gradient of the objective function (17) is written

as

 
2 H

1 1
∇λf (λ) = −R · vec Diag (σ)VEV − q− (18)
t 2L − 1T λ

where qi = 1/λi is the i-th element of vector q. Thus the steepest descent direction

is chosen as

∆λ = −∇λf (λ).

Combining this search direction with the backtracking line search conditions [24],

we establish Algorithm 1 for the optimal power allocation vector, which ensures the

convergence because of concavity. The details of the algorithm are shown in Table 1.

Table 1. Algorithm 1: Optimization of power allocation vector

Step 1. Given a feasible vector λ, t := t(0) > 0, α > 1, tolerance  > 0.

Step 2. Compute the gradient ∇λf (λ) as (18) and the descent direction ∇λf (λ).
Step 3. If k∆λk2 is sufficiently small, then go to Step 6; else go to Step 4.
Step 4. Choose γ so that f (λ + γ∆λ) < f (λ) by backtracking line search.
Step 5. Set λ := λ + γ∆λ. Go to Step 2.
Step 6. Stop if 1/t < , else t := αt, and go to step 2.

4.3 OPTIMIZATION OVER RIGHT SINGULAR VECTORS

Now we consider maximizing the mutual information over the right singular

vectors V for a given λ:

maximize I(V)
(19)
subject to VH V = V VH = I.
 72

This unitary-matrix constrained problem can be formulated as an unconstrained op-

timization in a constrained search space:

minimize g(V)

with domain restricted to the Stiefel manifold St(n) [25]:

dom g = {V ∈ St(n)}

and

I n×n |VH V = I
St(n) = V ∈ C

where the function g(V) is defined as −I(V). For each point V ∈ St(n), the search

direction ∆V on the tangent space has been suggested in [26] to minimize the objec-

tive function,

∆V = −∇V g(V) = ∇V I(V) − V(∇V I(V))H V (20)

where ∇V I(V) is the gradient of mutual information with respect to V, given by

Diag2 (σ)Diag(λ)VE.

Note that moving along the descent direction on the tangent space may cause

the unitary property being lost. Therefore, it needs to be restored in each step via

I n×n , its projection π(W) on the Stiefel

projection. For an arbitrary matrix W ∈ C

manifold is defined as the point closest to W in the Euclidean norm:

π(W) = arg min kW − Qk2 .

Q∈St(n)

If the SVD of W is W = UW ΣVW , the projection can be expressed by UW VW [27,

Sec. 7.4.8].
 73

Combining the search direction and the projection with the backtracking line

search condition, Algorithm 2 in Table 2 is developed to maximize the mutual infor-

mation over the right singular vectors V.

Table 2. Algorithm 2: Optimization of right singular vectors

Step 1. I n×n such that VH V = I.

Given a feasible V ∈ C
Step 2. Compute the descent direction ∆V as (20). Set the step size γ := 1.
Step 3. If k∆Vk2 = Tr{(∆V)H ∆V} is small, then stop; else go to Step 4.
Step 4. Choose γ so that g(π(V + γ∆V)) < g(V) by backtracking line search.
Step 5. Set V := π(V + γ∆V). Go to Step 2.

4.4 TWO-STEP APPROACH TO OPTIMIZE PRECODER

Table 3 shows a complete two-step approach named Algorithm 3, which max-

imizes the mutual information over a generalized precoding matrix P by combining

Algorithm 1 and Algorithm 2.

Table 3. Algorithm 3: Two-step optimization algorithm

Step 1. Initialization: Set the left singular vectors of the precoder U := VH .

Specify a feasible λ and V.
Step 2. Update power allocation vector. Run Algorithm 1 given V.
Step 3. Update right singular vectors: Run Algorithm 2 given λ.
Step 4. Repeat Step 2 and Step 3 until convergence.

The proposed algorithm converges to the globally optimal solution in the low

SNR region because mutual information is maximized by optimizing λ [16]. For

 74

medium to high SNR, the proposed method, theoretically, converges to a local maxi-

mum. However, extensive numerical examples show that different initial points have

limited effect on the algorithm (see Sec. 5); that is, the two-step method achieves

near global optimal performance.

The destination node applies Algorithm 3 to each relay node and calculates

the corresponding achievable mutual information and precoder. Denote the mutual

information of the i-th relay node as Ii . The best relay node is then selected by

Rs = arg max Ii .
i=1,··· ,k

Next, the index of the selected node and the corresponding precoder are transmitted

via a feedback channel from the destination to the relay and to the source node,

respectively.
 75

5 SIMULATION RESULTS

This section examines the efficacy of the linear precoder on relay networks by

several examples. These examples consider a single-relay network with a block length

L = 1 and channel coefficients h0 = 0.4, h1 = 1.2, and g1 = −0.9. The same transmit

power is assumed at the source and relay node (i.e., Ps = Pr = P ).

The proposed two-step iterative algorithm is first tested from different feasible

starting points. A general 2 × 2 unitary matrix group can be expressed as [28]

   
α1 −θ
 e 0   cos ψ e
sin ψ 
V= · .
0 eα2 −eθ sin ψ cos ψ
| {z } | {z }
V1 V2

For the simplified channel model in (13), the mutual information remains unchanged

under the rotation of the diagonal unitary matrix V1 . Thus, only the structure of V2

is used to generate feasible right singular vectors V.

5.1 MUTUAL INFORMATION PERFORMANCE

We first consider two different starting points: Case A: λ = [0.5; 0.5], ψ = π/6,

and θ = π/4; Case B: λ = [0.2; 0.8], ψ = π/10, and θ = π/10. Figure 1 illustrates the

convergence for BPSK inputs when the SNR is 0 dB. For comparison, the figure also

shows the mutual information corresponding to the cases of no precoding, optimal

precoder with Gaussian inputs [5, 8], and optimal precoding with the GD method

[15, 16]. The GD method is influenced by its initial point selection and converges to

different mutual information levels with different initial points. The proposed two-

step iteration of Algorithm 3, however, is insensitive to the initial points and converges

to almost the same value, which increases the mutual information to 0.52 bps/Hz or
 76

68% over that of no precoding. It also approaches the upper bound that is achieved by

Gaussian inputs. Moreover, the progress of the proposed method exhibits a staircase

shape, where each stair is associated with either the shift between the optimizations

of the power allocation vector and the right singular vector or the iteration for the

parameter t within Algorithm 1.

0.55

0.5

0.45
Mutual Information (bits/s/Hz)

0.4

0.35

0.3

0.25

0.2 New Two−Step Algorithm: Case A

New Two−Step Algorithm: Case B
0.15 Gradient Descent Method: Case A
Gradient Descent Method: Case B
0.1 BPSK, No Precoding
Gaussian, Maximum Capacity
0.05
0 5 10 15 20
Iteration Index

Figure 1. Evolution of the mutual information with BPSK inputs when SNR is 0 dB.

The cumulative distribution of the optimized mutual information from differ-

ent initial points for the proposed two-step algorithm and the GD method are depicted

in Fig. 2, which is obtained by maximizing the mutual information via 5,000 uni-

formly random initial points that are feasible to the considered problem. From Fig.

2, the two-step algorithm has a narrow and sharp curve, and the difference between

the highest and lowest optimized mutual information is less than 0.003 bps/Hz; that
 77

is, the iterative algorithm is able to obtain a near global optimal solution even if the

problem is nonconcave. In contrast, the GD method depends highly on the initial

point selection, and the difference between the highest and lowest optimized mutual

information is as wide as 0.3 bps/Hz. It means the GD method may provide a so-

lution even much worse than that of no precoding if the initial point is not chosen

carefully.

100

90

80
Cumulative Distribution

70

60

50

40

30

20

10 New Two−Step Algorithm

Gradient Descent Method
0
0.1 0.2 0.3 0.4 0.5 0.6
Mutual Information (bits/s/Hz)

Figure 2. Cumulative distribution of the optimized mutual information with BPSK

inputs when SNR is 0 dB.

Similar performance is also observed for QPSK with SNR given by 5 dB, as

shown in Fig. 3 and 4. The convergence of the proposed two-step algorithm is

achieved in 15 iterations, similar to the case of BPSK. It also achieves 38% improve-

ment of mutual information over no precoding. From Fig. 4, the proposed algorithm
 78

obtains mutual information within 98% of the maximum capacity of Gaussian inputs

with 69% of the initial point selections, while the GD method reaches within 70% of

the maximum capacity of Gaussian inputs with only 5% of the initial point selections.

1.2

1.1

1
Mutual Information (bits/s/Hz)

0.9

0.8

0.7

0.6

0.5 New Two−Step Algorithm: Case A

New Two−Step Algorithm: Case B
0.4 Gradient Descent Method: Case A
Gradient Descent Method: Case B
0.3 QPSK, No Precoding
Gaussian, Maximum Capacity
0.2
0 5 10 15 20 25
Iteration Index

Figure 3. Evolution of the mutual information with QPSK inputs when SNR is 5 dB.

Figures 5 and 6 show the mutual information of the proposed algorithm versus

SNR for BPSK and QPSK inputs, respectively, in comparison with other schemes

such as no precoding, maximum diversity design in [3], maximum coding gain design

in [3, 4], and maximum capacity design assuming Gaussian inputs in [5, 8]. Figures 5

and 6 suggest the following observations:

First, when the elements of the transmitted signal x are drawn from BPSK

or QPSK, the mutual information for relay networks is bounded by 1 bps/Hz and 2

bps/Hz, respectively, which are achieved by all precoding schemes when SNR is high.
 79

100

90

80

Cumulative Distribution 70

60

50

40

30

20

10 New Two−Step Algorithm

Gradient Descent Method
0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
Mutual Information (bits/s/Hz)

Figure 4. Cumulative distribution of mutual information with QPSK inputs when

SNR is 5 dB.

Second, the precoder design based on maximizing capacity with the assump-

tion of Gaussian inputs may result in a significant loss for systems employing discrete

inputs. This loss comes from differences in designing the power allocation vector and

the right singular vectors. For Gaussian inputs, allocating more power to the stronger

subchannels and less to the weaker subchannels is helpful to maximize capacity. This

strategy, however, does not work for finite-alphabet inputs, because the mutual in-

formation with finite inputs is bounded, therefore little incentive can be gained by

allocating more power to the subchannels already close to saturation. Moreover, the

right singular vectors for Gaussian-input systems is an arbitrary unitary matrix [see

eq. (10)]. Systems with finite inputs, however, have to carefully select the right

singular vectors to maximize the mutual information.

 80

0.8

Mutual Information (bps/Hz)

0.6

0.4
BPSK, New Two−Step Algorithm
BPSK, No Precoding
BPSK, Maximum Diversity
0.2 BPSK, Maximum Coding Gain
BPSK, Maximum Capacity
Gaussian, Maximum Capacity
Gaussian, No Precoding
0
−10 −5 0 5 10 15 20 25
SNR (dB)

Figure 5. Mutual information versus the SNR with BPSK inputs.

2
Mutual Information (bps/Hz)

1.5

QPSK, New Two−Step Algorithm

QPSK, No Precoding
0.5 QPSK, Maximum Diversity
QPSK, Maximum Coding Gain
QPSK, Maximum Capacity
Gaussian, Maximum Capacity
Gaussian, No Precoding
0
−5 0 5 10 15 20
SNR (dB)

Figure 6. Mutual information versus the SNR with QPSK inputs.

 81

Third, although the maximum coding gain method [3] performs better than

the maximum diversity and no precoding methods, it is valid only for the case of

block length L = 1 and QPSK inputs or the case of L = 1 and M-QAM inputs

[4]. In comparison, the proposed two-step method can be used for an arbitrary

block length L and input type. Our algorithm exploits the degrees of freedom in

the optimal left singular vectors, the optimal power allocation vector, and the local

optimal right singular vectors simultaneously. Therefore, providing significant gains

of mutual information in a wide range of SNR. For example, with input BPSK and

3/4 channel coding rate, the performance of the proposed method is about 4 dB,

5.5 dB, and 6 dB better than those of the maximum coding gain, no precoding, and

maximum capacity methods, respectively.

Last, the proposed method achieves mutual information very close to maxi-

mum capacity with Gaussian inputs when the channel coding rate is below 0.6 for

both BPSK and QPSK; it also outperforms the case of Gaussian inputs with no

precoding when the channel coding rate is below 0.9.

5.2 CODED BER PERFORMANCE

To evaluate the coded BER of the proposed method, channel coding is used

at the source node and turbo principle [29, 30] is used at the destination node, as

illustrated in Fig. 7. Note that the interleaver is not shown in the block diagram

because of the usage of the low-density parity-check, or LDPC, codes [31]. The

signal sequence b is encoded by the LDPC encoder and mapped according to the

conventional equiprobable discrete signaling constellations. It is then divided into xa

and xb and transmitted at the two time slots, respectively; the selected relay node

simply amplifies and forwards the signal xa in the second time slot.

At the receiver of the destination node, the maximum a posteriori (MAP)

detector takes channel observations y and the a priori information Le (b) from the
 82

modulation
Precoding Effective
LDPC Matrix Relay Channel
Mod
Encoder

Feedback
AWGN

hard decision

LDPC MAP
Decoder Detector

Figure 7. Block diagram of the relay network with precoding at the source and
iterative receiver at the destination.

decoder and computes the extrinsic information Le (c) for each coded bit. Thus,

the extrinsic information between the MAP detector and the decoder is exchanged

iteratively until the desired performance is achieved [29].

In our simulation of the end-to-end system of Fig. 7, we use the block length

for relaying transmission L = 1. The LDPC encoder and decoder modules are derived

from [32] with coding rate 1/2. The channel coding length 2400, the coding rate 3/4

and 2/3 for BPSK and QPSK, respectively, and five iterations between the MAP

detector and the LDPC decoder. Figures 8 and 9 show that the designed precoder

that maximizes the mutual information provides large performance gains in the coded

BER over other schemes including on precoding, the maximum coding gain, and

maximum capacity methods. Note that the precoder designed by maximum capacity

with Gaussian input offers worse BER performance than no precoding in practical

BPSK and QPSK systems.

 83

−1
10

−2
10

BER −3
10

−4
10 New Two−Step Algorithm
No Precoding
Maximum Diversity
Maximum Coding Gain
Maximum Capacity
−5
10
2 4 6 8 10 12 14 16 18 20
SNR (dB)

Figure 8. Bit error performance with BPSK inputs and 3/4 channel coding rate.

−1
10

−2
10
BER

−3
10

−4
10 New Two−Step Algorithm
No Precoding
Maximum Diversity
Maximum Coding Gain
Maximum Capacity
−5
10
8 10 12 14 16 18 20 22
SNR (dB)

Figure 9. Bit error performance with QPSK inputs and 2/3 channel coding rate.
 84

6 CONCLUSION

This paper has proposed a new two-step iterative algorithm for linear precoder

design that maximizes mutual information with finite-alphabet inputs. By setting

the left singular vectors of the precoding matrix equal to the right singular vectors

of the channel matrix, the mutual information has proved to be a concave function

on the squared singular values of the linear precoder for a complex-valued channel.

Consequently, the proposed algorithm first optimizes the mutual information via the

power allocation matrix or singular values of the linear precoder with a given set of

right singular vectors, then solves the second maximization problem for right singular

vectors with the obtained power allocation matrix, and iterates through the two

steps until convergence. The second optimization problem is difficult and we have

reformulated it on the complex Stiefel manifold and have solved it by the gradient

method with projection. The proposed two-step linear precoder design algorithm has

several advantages: being able to handle general complex-valued channels and system

configurations, insensitive to initial point selection, fast convergence, and high mutual

information gain. A numerical example of a two-hop relay network with amplify-

and-forward strategy and Binary/Quadrature Phase Shift Keying (BPSK/QPSK)

constellations is presented to demonstrate the performance gains of the proposed

linear precoder algorithm.

 85

7 APPENDIX A: PROOF OF PROPOSITION 1

To prove Proposition 1, we first introduce the definitions of gradient, Jacobian,

and Hessian of a complex matrix [33].

7.1 DEFINITION

The gradient matrix with respect to a complex-valued matrix Z is defined as

∂f
∇Z f ,
∂Z∗

where the (i, j)-th element of the gradient matrix is [∇Z f ]ij = ∂f /∂ [Z∗ ]ij .

Let F(Z, Z∗ ) be a complex matrix function of Z and Z∗ ; the Jacobian matrices

of F with respect to Z and Z∗ are then given by:

∂vec(F) ∂vec(F)
DZ F , and DZ ∗ F , . (21)
∂vecT (Z) ∂vecT (Z∗ )

Let Z1 and Z2 be two complex-valued matrices, and let f be a real-valued

scalar function of Z1 and Z2 . The complex Hessian matrix of f with respect to Z1

and Z2 is defined by:

 T
∂ ∂f
HZ1 ,Z2 f , DZ1 (DZ2 f ) = . (22)
∂vecT (Z1 ) ∂vecT (Z2 )

7.2 JACOBIAN OF THE MMSE MATRIX

The gradient of mutual information as the first-order characteristic has been

derived in [15]:

∇P I (x; y) = HH HPE (23)

 86

where E is the MMSE matrix defined by

n o
H
E , E [x − E (x|y)] [x − E (x|y)] (24)

The MMSE estimate of x (the conditional mean) is

M
X
2L  
p(y|x)
E (x|y) = xl p(xl |y) = Ex x (25)
l=1
p(y)

and the conditional probability density function of the received signal y is calculated

as
1 2


p(y|x) = exp −ky − HPxk . (26)
π 2L

The Jacobian of the MMSE matrix DP∗ E is now introduced as a Lemma for

deriving the second-order characteristic function of the mutual information. Note that

Lemma 2 hold for the general complex-valued channel case, which agrees with [17,

Theorem 3] for real-valued signal model.

Lemma 2. The Jacobian of the MMSE matrix E with respect to P∗ is given by



DP∗ E = −E {Φ∗ (y) ⊗ Φ(y)} K I ⊗ PT HT H∗


− E {Ψ∗ (y) ⊗ Ψ(y)} I ⊗ PT HT H∗ (27)

where K is a unique 4L2 ×4L2 permutation matrix with vec ΦT (y) = K·vec (Φ(y)).
 87

Proof. The (i, j)-th element of the MMSE matrix E is given by [E]ij =
 ∗
E xi xj − E {E{xi |y}E{xj |y}}, from which it follows that

Z
∂[E]ij ∂p(y)
∗
=− ∗
E{xi |y}E{x∗j |y}dy
∂[P]kl ∂[P]kl
Z
∂E{xi |y}
− p(y) E{x∗j |y}dy
∂[P]∗kl
Z
∂E{x∗j |y}
− p(y)E{xi |y} dy. (28)
∂[P]∗kl

Employing Dy ky − HPxk2 = (y − HPx)H yields

Dy p(y|x) = −p(y|x)(y − HPx)H (29)


Dy p(y) = E{Dy p(y|x)} = −E p(y|x)(y − HPx)H . (30)

The gradient of probability density function ∇P p(y) can then be written as

∇P p(y) = E [∇P p(y|x)]

 
= E p(y|x)HH (y − HPx)xH

= −HH E [Dy p(y|x)]H xH (31)

and the gradient of conditional expectation ∇P E{xi |y} becomes

 
p(y|x)
∇P E{xi |y} = ∇P E xi
p(y)
 
H [Dy p(y|x)]H xH
H
= −E xi
p(y)
(  )
p(y|x)HH E [Dy p(y|x)]H xH
+ E xi
[p(y)]2

1 
= H − E xi [Dy p(y|x)]H xH
H
p(y)

 H H

+ E{xi |y}E [Dy p(y|x)] x . (32)
 88

The Jacobian Dy E{x|y} is derived as follows:

 
p(y|x)
Dy E{x|y} = Dy E x
p(y)
 
Dy p(y|x) · p(y) − p(y|x) · Dy p(y)
=E x
[p(y)]2
   
−p(y|x) · (y − HPx)H p(y|x) · (y − HPE{x|y})H
=E x +E x
p(y) p(y)
   
H p(y|x) H H p(y|x) H H H
= E xx P H −E x E{x|y} P H
p(y) p(y)
  
= E xxH |y − E{x|y}E{x|y}H PH HH

= Φ(y)PH HH . (33)

Following the similar steps in (33), the Jacobian can also be obtained:

Dy∗ E{x|y} = Ψ(y)PH HH . (34)

Substituting (31) and (32) into (28) yields

Z
∂[E]ij 
∗
= − E{xi |y}E{x∗j |y}eTk HH E [Dy p(y|x)]H x∗l dy
∂[P]kl
Z

+ eTk HH E xi [Dy p(y|x)]H x∗l E{x∗j |y}dy
Z

+ E{xi |y}eTk HH E x∗j [Dy p(y|x)]H x∗l dy (35)

where ek and el are, respectively, the k-th and l-th columns of an identity matrix

with appropriate dimensions.

 89

The first term in (35) can be rewritten as

Z

− E{xi |y}E{x∗j |y}eTk HH E [Dy p(y|x)]H x∗l dy
Z
(a) ∂p(y)E{x∗l |y}
= − E{xi |y}E{x∗j |y}eTk HH dy
∂vec(y∗ )
Z
(b) ∗ T H
∂E{xi |y}E{x∗j |y}
= p(y)E{xl |y}ek H dy
∂vec(y∗ )
Z
(c)
= p(y)E{x∗l |y}E{x∗j |y}eTk HH HPΨ(y)ei dy
Z
+ p(y)E{x∗l |y}E{xi |y}eTk HH HPΦ(y)ej dy (36)

where (a) follows from the definition of Jacobian matrix (21) and the fact E{x∗l |y} =

E {x∗l p(y|x)/p(y)}; (b) is the result of integration by parts; (c) is from the result of

(33) and (34).

Similarly, the second and third terms in (35) can be re-expressed, respectively,

as

Z

eTk HH E xi [Dy p(y|x)]H x∗l E{x∗j |y}dy
Z
= − eTk HH HPΦ(y)ej E{xi x∗l |y}p(y)dy (37)

and

Z

E{xi |y}eTk HH E x∗j [Dy p(y|x)]H x∗l dy
Z
= − eTk HH HPΨ(y)ej E{x∗j x∗l |y}p(y)dy. (38)
 90

Substituting (36), (38), and (37) into (35) yields

Z
∂[E]ij  
=− eTk HH HPΦ(y)ej p(y) E{xi x∗l |y} − E{xi |y}E{x∗l |y} dy
∂[P]∗kl
Z
 
− eTk HH HPΨ(y)ej p(y) E{x∗j x∗l |y} − E{x∗j |y}E{x∗l |y} dy

= −Ey eTi Φ(y)el eTj ΦT (y)GT HT H∗ ek

− Ey eTj Ψ∗ (y)el eTi Ψ(y)GT HT H∗ ek
h 
T T T ∗

i
= −Ey K Φ(y) ⊗ (Φ (y)P H H )
i+(j−1)2L,k+(l−1)2L
h  i 
∗ T T ∗
− Ey Ψ (y) ⊗ Ψ(y)P H H (39)
i+(j−1)2L,k+(l−1)2L

2 ×4L2
where K ∈ R4L is a unique permutation matrix that satisfies [34]:

vec ΦT (y) = K · vec (Φ(y)) . (40)

From ∂[E]ij /∂[P]∗kl = [DP∗ E]i+(j−1)2L,k+(l−1)2L , we have

n
h io
T T T ∗
D E = −KEy Φ(y) ⊗ Φ (y)P H H
P∗
n h io
∗ T T ∗
− Ey Ψ (y) ⊗ Ψ(y)P H H
 
= −KEy {Φ(y) ⊗ Φ∗ (y)} I ⊗ PT HT H∗
 
− Ey {Ψ∗ (y) ⊗ Ψ(y)} I ⊗ PT HT H∗ . (41)

Applying of the property of permutation matrix K [34], K (A ⊗ B) = (B ⊗ A) K,

the proof is complete.

Proof of Proposition 1. Let us denote the remaining part of the precoder,

√
Diag( λ)V, as G, which allows the power allocation vector λ to be rewritten as

λ = R · vec(GGH ). (42)
 91

According to (22), the Hessian of the mutual information can be obtained from

HλI (λ) = Dλ [DλI (λ)] . (43)

Now DλI (λ) can be calculated as

(a)

DλI(λ) = DGGH I(λ) · RT = vecT Diag2 (σ)VEVH RT
(b)

= vecT (E) VH ⊗ VT Diag2 (σ) RT (44)

where (a) follows from the results in [15, Theorem 2], and (b) follows from the property

[34]:


vec (ATB) = BT ⊗ A vec(T) (45)

in which A, T, and B are matrices with appropriate dimensions. The result in (43)

can then be written as




HλI(λ) = Dλ R V∗ ⊗ Diag2 (σ)V vec (E) = R V∗ ⊗ Diag2 (σ)V DλE. (46)

From the chain rule of the Jacobian [33], it yields

DG∗ E = DλE · DG∗ λ + Dλ∗ E · DG∗ λ∗ = 2DλE · DG∗ λ (47)

where DG∗ λ can be derived from the definition of Jacobian (21):



DG∗ λ = DG∗ R · vec GGH = R (I ⊗ G) K. (48)
 92

Because DG∗ λ has full column rank, it is possible to invert the transformation in (47)

to obtain

1 1

Dλ E = (DG∗ E) (DG∗ λ)+ = (DG∗ E) · GH ⊗ I KT RT Diag−1 (λ) . (49)
2 2

Plugging (27) and (49) into (46) yields the Hessian result provided in (16),

which can be readily identified as negative semi-definite.

 93

8 REFERENCES

[1] J. N. Laneman and G. W. Wornell, “Distributed space-time-coded protocols

for exploiting cooperative diversity in wireless networks,” IEEE Trans. Inform.
Theory, vol. 49, no. 10, pp. 2415–2425, Oct. 2003.
[2] Y. Jing and B. Hassibi, “Distributed space-time coding in wireless relay net-
works,” IEEE Trans. Wireless Commun., vol. 5, no. 12, pp. 3524–3536, Dec.
2006.
[3] Y. Ding, J. Zhang, and K. Wong, “The amplify-and-forward half-duplex coop-
erative system: Pairwise error probability and precoder design,” IEEE Trans.
Signal Process., vol. 55, no. 2, pp. 605–617, Feb. 2007.
[4] ——, “Optimal precoder for amplify-and-forward half-duplex relay system,”
IEEE Trans. Wireless Commun., vol. 7, no. 8, pp. 2890–2895, Aug. 2008.
[5] A. S. Behbahani, R. Merched, and A. M. Eltawil, “Optimizations of a MIMO
relay network,” IEEE Trans. Signal Process., vol. 56, no. 10, pp. 5062–5073, Oct.
2008.
[6] A. B. Gershman, N. D. Sidiropoulos, S. Shahbazpanahi, M. Bengtsson, and B. Ot-
tersten, “Convex optimization-based beamforming: From receive to transmit and
network designs,” IEEE Signal Process Mag., vol. 27, no. 3, pp. 62–75, Mar. 2010.
[7] R. Nabar, H. Bolcskei, and F. Kneubuhler, “Fading relay channels: Performance
limits and space-time signal design,” IEEE J. Sel. Areas Commun., vol. 22, no. 6,
pp. 1099–1109, Aug. 2004.
[8] R. Mo and Y. Chew, “Precoder design for non-regenerative MIMO relay sys-
tems,” IEEE Trans. Wireless Commun., vol. 8, no. 10, pp. 5041–5049, Oct. 2009.
[9] W. Zeng, C. Xiao, Y. Wang, and J. Lu, “Opportunistic cooperation for multi-
antenna multi-relay networks,” IEEE Trans. Wireless Commun., vol. 9, no. 10,
pp. 3189–3199, Oct. 2010.
[10] A. Lozano, A. Tulino, and S. Verdú, “Optimum power allocation for parallel
Gaussian channels with arbitrary input distributions,” IEEE Trans. Inform. The-
ory, vol. 52, no. 7, pp. 3033–3051, Jul. 2006.
[11] G. Caire and K. Kumar, “Information theoretic foundations of adaptive coded
modulation,” Proc. IEEE, vol. 95, no. 12, pp. 2274–2298, Dec. 2007.
[12] C. Xiao and Y. R. Zheng, “On the mutual information and power allocation for
vector Gaussian channels with finite discrete inputs,” in Proc. IEEE Globecom,
New Orleans, LA, 2008.
 94

[13] C. Wen, K. Wong, and C. Ng, “On the asymptotic properties of amplify-and-
forward MIMO relay channels,” IEEE Trans. Commun., vol. 59, no. 2, pp. 590–
602, Feb. 2009.

[14] W. Zeng, C. Xiao, M. Wang, and J. Lu, “Linear precoding for relay networks
with finite-alphabet constraints,” in IEEE Int. Conf. Commun. (ICC), Kyoto,
Japan, 2011.

[15] D. Palomar and S. Verdú, “Gradient of mutual information in linear vector Gaus-
sian channels,” IEEE Trans. Inform. Theory, vol. 52, no. 1, pp. 141–154, Jan.
2006.

[16] F. Pérez-Cruz, M. Rodrigues, and S. Verdú, “MIMO Gaussian channels with

arbitrary inputs: Optimal precoding and power allocation,” IEEE Trans. Inform.
Theory, vol. 56, no. 3, pp. 1070 –1084, Mar. 2010.

[17] M. Payaró and D. P. Palomar, “Hessian and concavity of mutual information,

differential entropy, and entropy power in linear vector Gaussian channels,” IEEE
Trans. Inform. Theory, vol. 55, no. 8, pp. 3613–3628, Aug. 2009.

[18] M. Lamarca, “Linear precoding for mutual information maximization in MIMO

systems,” in Proc. Int. Symp. on Wireless Commun. Syst. (ISWCS), 2009, pp.
26–30.

[19] C. Xiao, Y. R. Zheng, and Z. Ding, “Globally optimal linear precoders for finite
alphabet signals over complex vector Gaussian channels,” IEEE Trans. Signal
Process., vol. 59, no. 7, pp. 3301–3314, Jul. 2011.

[20] D. Palomar and Y. Jiang, MIMO Transceiver Design via Majorization Theory.
Now Publishers Inc., 2006.

[21] M. Payaró and D. P. Palomar, “On optimal precoding in linear vector Gaussian
channels with arbitrary input distribution,” in Proc. IEEE Int. Symp. Inform.
Theory (ISIT), Seoul, Korea, 2009, pp. 1085–1089.

[22] A. Bletsas, A. Khisti, D. P. Reed, and A. Lippman, “A simple cooperative di-

versity method based on network path selection,” IEEE J. Sel. Areas Commun.,
vol. 24, no. 3, pp. 659–672, Mar. 2006.

[23] M. Grant and S. Boyd, “CVX: Matlab software for disciplined con-
vex programming, version 1.21,” retrieved Apr. 2011. [Online]. Available:
http://cvxr.com/cvx

[24] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University

Press, 2004.

[25] P. Absil, R. Mahony, and R. Sepulchre, Optimization Algorithms on Matrix Man-

ifolds. Princeton University Press, 2008.
 95

[26] J. H. Manton, “Optimization algorithms exploiting unitary constraints,” IEEE

Trans. Signal Process., vol. 50, no. 3, pp. 635–650, Mar. 2002.

[27] R. Horn and C. Johnson, Matrix Analysis. Cambridge University Press, 1985.

[28] F. Murnaghan, Lectures on Applied Mathematics: The Unitary and Rotation

Groups. Washington, DC: Spartan Books, 1962, vol. 3.

[29] B. Hochwald and S. Ten Brink, “Achieving near-capacity on a multiple-antenna

channel,” IEEE Trans. Commun., vol. 51, no. 3, pp. 389–399, Mar. 2003.

[30] R. Koetter, A. Singer, and M. Tuchler, “Turbo equalization,” IEEE Signal Pro-
cess Mag., vol. 21, no. 1, pp. 67–80, Jan. 2004.

[31] B. Lu, X. Wang, and K. Narayanan, “LDPC-based space-time coded OFDM sys-
tems over correlated fading channels: Performance analysis and receiver design,”
IEEE Trans. Commun., vol. 50, no. 1, pp. 74–88, Jan. 2002.

[32] M. Valenti, “Iterative solutions coded modulation library (ISCML).” [Online].

Available: http://www.iterativesolutions.com/Matlab.htm

[33] A. Hjørungnes, Complex-Valued Matrix Derivatives: With Applications in Signal

Processing and Communications. Cambridge University Press, 2011.

[34] J. Magnus and H. Neudecker, Matrix Differential Calculus with Applications in

Statistics and Econometrics, 3rd ed. John Wiley & Sons, 2007.
 96

IV. PRACTICAL LINEAR PRECODER DESIGN FOR FINITE

ALPHABET MIMO-OFDM WITH EXPERIMENT VALIDATION

Mingxi Wang, Yahong Rosa Zheng and Chengshan Xiao

ABSTRACT—A low complexity precoding method is proposed for practical Multiple-

Input Multiple-Output (MIMO) Orthogonal Frequency-Division Multiplexing (OFDM)

systems. Based on the two-step optimal precoder design algorithm that maximizes

the lower bound of the mutual information with finite-alphabet inputs, the proposed

method simplifies the precoder design by fixing the right singular vectors of the pre-

coder matrix, eliminating the iterative optimization between the two steps, and dis-

cretizing the search space of the power allocation vector. For a 4×4 channel, the

computational complexity of the proposed precoder design is reduced to 3% and 6%

of that required by the original two-step algorithm with Quadrature Phase Shift Key-

ing (QPSK) and 8PSK, respectively. The proposed method achieves nearly the same

mutual information as the two-step iterative algorithm for a large range of SNR re-

gion, especially for large MIMO size and/or high constellation systems. The proposed

precoding design method is applied to a 2×2 MIMO-OFDM system with 2048 sub-

carriers by designing 1024 precoders for extended channel matrices of size 4×4. A

transceiver test bed implements these precoding matrices in comparison with other

existing precoding schemes. Indoor experiments are conducted for fixed-platform non-

line-of-sight (NLOS) channels, and the data processing results show that the proposed

precoding method achieves the lowest BER compared to maximum diversity, classic

water-filling and channel diagonalization methods.

 97

1 INTRODUCTION

Linear precoding for multiple-input multiple-output (MIMO) communications

has been a popular research topic in the last decade. Traditional precoding methods

include maximizing the channel capacity with Gaussian inputs [1], maximizing the

diversity order [2], maximizing the signal to interference-plus-noise ratio (SINR) [3],

minimizing the mean square error (MSE) [3], etc. More recently, several works

have found that designing precoders by maximizing mutual information with finite-

alphabet inputs can achieve higher mutual information [4,5,6,7,8,9,10] and lower bit

error rate (BER) [11] than employing other optimization criteria. The performance

benefits of these approaches [4,5,6,7,8,9,10,10,11] come from optimization of mutual

information with finite-alphabet input constraints, compared to using other indirect

methods such as maximizing channel capacity with Gaussian inputs, maximizing di-

versity order, minimizing SINR, or minimizing MSE.

However, applying precoding to practical MIMO systems encounter several

obstacles. First of all, the algorithms developed to find the optimal precoders [10,11,

12, 13, 14] are computationally complicated because of their iterative algorithm and

multiple computations of gradients through Monte Carlo simulations. For instance,

[10] utilized the gradient descent algorithm to directly find the precoding matrix, and

[14] proposed a two-step iterative algorithm to exploit the precoder structure. Even

offline calculation of precoders is prohibitive for large MIMO dimensions and high

constellations. Second, the optimal precoders are often designed for a specific SNR

value. Their sensitivity to SNR estimation errors is largely ignored in the literature.

Third, the soft maximum a posteriori (MAP) detector [15] is often employed for

performance evaluation [11,12] and good BER performance have been demonstrated.

However, MAP-based iterative receiver has the highest complexity and is thus difficult
 98

to implement in practice. Whether the performance gain of precoders can be leveraged

by practical suboptimal receivers is still questionable. In addition, the demonstrated

performance gains in [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14] are all based on simulations

over either fixed channels or statistical models of fading channels. The feasibility

of employing precoders in practical communication systems and their performances

under real-world wireless channels are still unknown.

In this paper, we first propose a low complexity precoding design method to

simplify the two-step precoder design algorithm that maximizes the lower bound of

the mutual information with finite-alphabet inputs [14]. The proposed method em-

ploys fixed right singular vectors that are modulation diversity matrices designed for

different modulation schemes, thus eliminating the need for the iterative optimiza-

tion between the two steps — one for the right singular vectors, one for the power

allocation vector. Furthermore, the proposed method discretizes the search space

of the power allocation vector and further reduces the complexity of the iterative

design of the power allocation. For a 4×4 channel, the computational complexity

of the proposed precoder design is reduced to 3% and 6% of that required by the

two-step algorithm for Quadrature Phase Shift Keying (QPSK) and 8PSK, respec-

tively. Moreover, the proposed method achieves nearly the same mutual information

as the original two-step iterative algorithm for a large range of SNR region for high

constellation or under large-sized MIMO channels. It also outperforms the maximum

diversity, classic waterfilling, and channel diagonalization methods in most part of

the SNR region.

The proposed precoding design method is applied to a 2×2 MIMO-OFDM

system with 2048 subcarriers for QPSK, 8PSK, and 16QAM (Quadrature Amplitude

Modulation). For each modulation scheme, we design 1024 precoders for the extended

channel matrices of size 4×4 that combines the 2×2 MIMO with 2 subcarriers, thus
 99

leveraging frequency diversity gain along spatial diversity gains. Moreover, the pre-

coders are designed for one specified SNR value but are applied to all SNR scenarios,

further reducing the computational cost on precoder design.

The designed precoding matrices are implemented in a Field Programable

Gate Array (FPGA)-based transceiver test bed employing a rate 3/4 low-density

parity-check (LDPC) code, where the baseband transmitter and digital Intermediate

Frequency (IF) circuits are implemented in Altera’s Stratix III FPGA. Two baseband

receiver algorithms are realized: the soft MMSE linear equalizer with interference can-

cellation [16, 17] in Altera’s Stratix IV FPGA, and the fixed-complexity list sphere

decoding (FSD) algorithm [18, 19] in MATLAB. Although various MIMO-OFDM

testbeds [20, 21, 22, 23, 24, 25, 26] are reported in the literature for wideband commu-

nications, they focus mainly on supporting spatial multiplexing [21, 22, 23, 24, 25] and

transmit diversity schemes [20], or space time block codes [21, 23], or polarization

diversity [26]. To the best of our knowledge, this is the first test bed that implements

precoder designs in both transmitter and receiver.

With the MIMO-OFDM test bed, we conduct indoor non-line-of-sight (NLOS)

experiment to acquire channels, design precoders offline, and verify performance with

precoding. The data processing results of the experiments show that the proposed

precoding method achieves the lowest BER compared to maximum diversity, classic

water-filling and channel diagonalization methods. Besides, the maximum diversity

outperforms the classic water-filling and channel diagonalization methods.

Throughout the paper, we denote vectors with boldface lower-case letters and

matrices with boldface upper-case letters. The superscripts (·)h and (·)+ represent

conjugate transpose and pseudoinverse operations, respectively. In addition, the sym-

bol C stands for the complex number field, and diag{a} denotes a diagonal matrix

with elements of vector a.

 100

2 LINEAR PRECODER DESIGN FOR FINITE ALPHABET

Consider a MIMO system with baseband equivalent model

y = HGx + n, (1)

where the vector x ∈ CNi ×1 contains transmitted signals of Ni symbols. The matrix

H ∈ CNo ×Ni is the complex channel matrix, and G ∈ CNi ×Ni is the designed precoding

matrix. The receiver noise n ∈ CNo ×1 is a zero mean circularly symmetric complex

Gaussian vector with covariance matrix σ 2 I, i.e., n ∼ CN (0, σ 2 I).

To design the precoding matrix G, we first decomposite the channel matrix H

via singular value decomposition (SVD), thus:

h
H = UH ΣH VH (2)

where UH and VH are unitary matrices, and ΣH is a diagonal matrix.

Similarly, the precoding matrix G can also be decomposed into three compo-

nents as:

h
G = UG ΣG VG (3)

in which UG and VG are left and right singular vectors of G, respectively. The

diagonal matrix ΣG has all singular values. The decomposition in (3) is a general

form that incorporates the four precoding schemes we discuss in this paper.

Scheme 1 is the channel diagonalization precoder, also known as parallel de-

composition [27, 28], which sets VG and ΣG to identity matrices, and UG = VH .

 101

The precoding matrix is then given by

GD = VH , (4)

which simply diagonalizes the channel.

Scheme 2 is the classic water-filling precoder which maximizes the channel

capacity assuming that the input signal x is Gaussian distributed. The MIMO channel

capacity with Gaussian input assumption is [27]

 
 1 
C = max log I + 2 HQH 
h
(5)
Tr(Q)=P σ

where P is the transmit power and Q is the covariance matrix after precoding such

that

h
Q = VH ΣWF VH (6)

where the diagonal matrix ΣWF can be solved by water-filling [1,27]. Since Q = GGh ,

the classic water-filling precoding matrix is

1
GW F = VH ΣW
2
F (7)

Scheme 3 is the maximum diversity precoder which maximizes diversity and

coding gains for two-dimensional constellations without Gaussian input assumptions

[2]. It consists of two matrices in (3):

GM D = VH VM D (8)

in which VM D is a matrix obtained through the algebraic construction and it is

designed to be the same for all types of modulations. Specifically, when Ni is a power
 102

of two, VM D is a unitary matrix given by

 
1 α1 ... α1Ni −1
 
 
1  1 α2 ... α2Ni −1 

VM D =√  .. .. ..  (9)
Ni  . . .

 
 
Ni −1
1 αNi . . . αN i

in which

 
4(k − 1) + 1
αk = exp j , k = 1, 2, · · · , Ni (10)
2Ni

Scheme 4 is the optimal precoding which maximizes the channel mutual in-

formation directly with finite-alphabet inputs. Based on the signal model in (1), the

mutual information between the input x and output y is given by [11]

 
M N M N
1 Xi Xi
I (x; y) [H, G] = Ni log M − En log exp (−dmk ) (11)
M Ni m=1 k=1

where dmk = (kHGemk + nk2 − knk2 )/σ 2 , and emk = xm − xk . The symbol En [·]

takes expectation over the noise n. The signal xm is a vector with each of its elements

drawn from the M-ary signal constellation.

The calculation of the mutual information in (11) involves mathematical ex-

pectation, which is often estimated through Monte Carlo simulations [10] that is

computationally expensive. To reduce the high computation complexity caused by

Monte Carlo simulations of the expectation, a lower bound of mutual information is

derived in [29] as

M i M
Xi
N  H N 
1 X cmk cmk
IL (H, G) = Ni log M − (1/ ln 2 − 1) No − N log exp − (12)
M i m=1 2σ 2
k=1
 103

where cmk = HG(xm − xk ). It is shown that the lower bound (12) is a close approx-

imation with a constant shift of the accurate mutual information (11) under various

channel conditions [29].

Employing this lower bound and setting UG = VH in the precoder design

yields a two-step iterative algorithm [14]. After initializations, the algorithm alter-

natively updates ΣG and VG in the two steps until the convergence. During each

iteration, the first step is to optimize the power allocation vector λ = diag(ΣG ) given

VG :

maximize IL (λ) (13)

subject to 1T λ = Ni (14)

λ0 (15)

where IL (λ) is the lower bound in (12). Given the λ obtained in the first step, the

second step is to optimize VG :

maximize IL (VG ) (16)

h
subject to VG VG = I (17)

As shown later in Section 3, the computational complexity of the optimal

precoder design using he lower bound is still prohibitively high. Moreover, the com-

plexity grows exponentially with the modulation level M and the size of precoding

matrix Ni . Thus, lowering the overall computational complexity further is necessary

to make practical use of the precoder design.

 104

3 PROPOSED SIMPLIFIED LINEAR PRECODER DESIGN

To reduce the complexity of the precoder design, we propose a non-iterative

approach to design precoder for finite-alphabet inputs. The proposed precoder G

contains three components as follows

G = VH ΣG Vmod (18)

where Vmod is a unitary matrix designed for different modulation constellations and

it is referred to as modulation diversity matrix in this paper. The matrix ΣG is a

diagonal matrix for power allocation policy depending on both the channel condition

and modulation constellation. The non-iterative precoder design approach first fixes

Vmod and then solves ΣG .

The modulation diversity matrix Vmod in (18) is different from the maximum

diversity matrix [2], because Vmod is tailored for the constellation used for the trans-

mitted signal vector x. Specifically, we propose the following structure for the design

of the modulation diversity matrix

 
1 β1 ... β1Ni −1
 
 
1  1 β2 . . . β2Ni −1 

Vmod =√   (19)
Ni  .. .. .. 
 . . . 
 
Ni −1
1 βNi . . . βN i

where

 
2(k − 1) + qmod
βk = exp jπ , k = 1, 2, · · · , Ni (20)
Ni
 105

The parameter qmod is a rotation angle depending on the modulation type. For

MPSK modulations, qmod is ( 12 )M −1 , where M is the constellation size. The rotation

1
angle qmod is 2
for all QAM such as 4QAM, 16QAM and 64QAM. In this case,

the constellation expansion is maximized. Table 1 lists values of qmod for different

modulations.

Table 1. The values of qmod corresponding to different modulations

Modulation BPSK QPSK 8PSK 16PSK 16QAM 64 QAM
1 1 1 1 1
qmod 1 2 4 8 2 2

Actually, the modulation diversity matrix for 2 × 2 MIMO system can be

easily derived from the values of ν of Table I in [11]. The corresponding modulation

diversity matrix for 4 × 4 MIMO system can be obtained with the extension of 2 × 2

MIMO systems.

Comparing (20) and (9), we see that the modulation diversity matrix is iden-

tical to the maximum diversity for QAM and QPSK modulations when Ni is a power

of 2. However, these two methods are different from each other for other modulations.

For instance, the 2 ×2 maximum diversity matrix for all types of modulations is given

by
 
j π4
1  1 e 
VM D = √   (21)
2 1 ej 4
5π
 106

On the other hand, the 2 × 2 modulation diversity matrices for BPSK and 8PSK

modulations are
   
j π2 j π8
1  1 e  1  1 e 
Vmod−BP SK = √  , Vmod−8P SK = √  . (22)
2 1 ej 2
3π
2 1 ej 8
9π

For QPSK and 16QAM modulations, the modulation diversity matrices are the same

as the maximum diversity matrix.

BPSK, Maximum Divesity BPSK, Modulation Divesity

2 2

1 1
Imag Part

0 0

−1 −1

−2 −2
−2 −1 0 1 2 −2 −1 0 1 2

8PSK, Maximum Divesity 8PSK, Modulation Diversity

2 2

1 1
Imag Part

0 0

−1 −1

−2 −2
−2 −1 0 1 2 −2 −1 0 1 2
Real Part Real Part

Figure 1. Scatter plot of the precoded BPSK and 8PSK signals using 2 × 2 maximum
diversity and modulation diversity

In Fig. 1, we plot the constellations of precoded BPSK and 8PSK inputs

using 2 × 2 modulation diversity and maximum diversity matrices. It is seen that

the scatter plot of the modulation diversity for BPSK inputs is the same as the

QPSK constellation, which is different from the result obtained by maximum diversity.

Comparing the scatter plots of 8PSK inputs, we notice that there are 33 and 64 points

for maximum diversity and modulations diversity, respectively. This indicates that
 107

one of the shortcomings of the maximum diversity is that the original 64 possible signal

points in x collapse to less points in the scatter plot after precoding. The modulation

diversity, on the other hand, preserves all the signal points in the constellations and

converts them to a complex Gaussian-like scatter plot.

After Vmod is fixed, ΣG can be found by exhaustive search based on the lower

bound of the mutual information. Specifically, we partition [0, 1] into K uniform

segments. Assume that the diagonal matrix ΣG = diag{σ1 , . . . , σNi } and each σi is

chosen from the set {0, K1 , K2 , . . . , K

K
}. Thus, there are totally (K + 1)Ni combinations

for all the Ni diagonal elements in ΣG . For each combination of {σ1 , . . . , σNi }, we
P i
normalize their values as σ̃i = σi / N 2
i=1 |σi | so that the power of the resulting pre-

coding matrix G is Ni . Each normalized diagonal matrix ΣGk along with the fixed

VM D corresponds to a precoding matrix Gk = VH ΣGk Vmod . By computing the lower

bound of IL H, Gk , k = 1, . . . , (K + 1)Ni , and choosing the Gk with the maximum

lower bound, we find the approximate optimal precoder. We note that the proposed

discretization of the power allocation diagonal matrix can be viewed as an extension

to the well known on/off power allocation technique.

From the afore mentioned procedure, we note that the proposed simplified

precoding is a non-iterative method. Moreover, it avoids the computation of the

derivative of IL (H, G) that is more time consuming than computing IL (H, G) itself,

especially when the size of the precoding matrix Ni is large. Therefore, the proposed

method has a lower complexity than the two-step iterative algorithm in [14].

To verify the complexity and performance merits of the proposed precoders,

we consider a constant 2 × 2 channel and a complex 4 × 4 channel:

 
 2 1 
H1 =   (23)
1 1
 108

Table 2. The CPU time (seconds) of designing precoders for the 4 × 4 channel H2
QPSK 8PSK
SNR Two-step Proposed Two-step Proposed
(dB) Iterative Algorithm Method Iterative Algorithm Method
-10 177.3064 3.0144 19655 743.6521
-5 163.5079 3.0518 10885 753.4039
5 82.2510 3.0593 11692 755.4079
10 106.9154 3.0773 12707 755.8201
15 178.5296 3.1564 12204 779.5273

and
 
0.1897 + 0.6602i 0.5417 + 0.5341i 0.8600 + 0.5681i 0.8998 + 0.4449i
 
 
 0.1934 + 0.3420i 0.1509 + 0.7271i 0.8537 + 0.3704i 0.8216 + 0.6946i 
 
H2 =  
 
 0.6822 + 0.2897i 0.6979 + 0.3093i 0.5936 + 0.7027i 0.6449 + 0.6213i 
 
0.3028 + 0.3412i 0.3784 + 0.8385i 0.4966 + 0.5466i 0.8180 + 0.7948i

(24)

in which the 2 × 2 channel H1 is also used in [4, 11, 14]. The SNR is defined as

SNRi = Tr(Hi Hhi )/(No σ 2 ), for i = 1, 2, [11].

Table 2 lists the CPU time of obtaining precoders using two-step iterative

algorithm [14] and the proposed method for the 4 × 4 channel H2 . The codes for

both methods are mostly written in MATLAB except that the part of computing

lower bound IL (H, G) is implemented in C++. The simulations are executed on an

Intel E8600 3.33 GHz duo core processor. In the proposed method, K is set to be 4.

Thus the total number of calculated lower bounds is 54 = 625. The results in Table 2

show that the proposed method consumes much less time than the two-step iterative

algorithm for QPSK and 8PSK modulations under various SNRs. For example, when

SNR = 10 dB, the CPU time of running the proposed method for QPSK and 8PSK

modulations is about 3% and 6% that of the two-step iterative algorithm, respectively.

 109

The mutual information results of different modulations for channel H1 and

H2 are shown in (a) - (f) of Fig.2, where for comparison, maximum diversity, classic

water-filling and channel diagonalization are also provided. For the 2 × 2 channel

with QPSK and 8PSK inputs, the proposed method performs almost the same as

the two-step iterative algorithm in the low and medium SNR regions. As the SNR

increases, its performance becomes apart from the two-step iterative algorithm but

still better than that of the maximum diversity method. On the other hand, it is

seen from (c) - (e) in Fig. 2 that the proposed method and the two-step iterative

algorithm have nearly the same performances for a large range of SNR region under

the 2 × 2 channel with 16QAM inputs, and the 4 × 4 channel with QPSK and 8PSK

inputs. Due to the extremely high computation complexity, the two-step iterative

algorithm has not been simulated for the 4 × 4 channel with 16QAM inputs. Yet,

Fig. 3.2(f) shows that the proposed method achieves higher mutual information than

the maximum diversity for the 16QAM. In all cases, the proposed method performs

better than the classic water-filling and channel diagonalization methods especially

in the medium and high SNR regions.

To leverage the benefit of precoding in practical MIMO-OFDM systems, we

further propose to apply linear precoding to extended channel matrices that combines

Nr ×Nt MIMO and KG subcarriers. Specifically, we divide the total of Nf subcarriers

into Ng groups and perform linear precoding for the extended (Nr · KG ) × (Nt · KG )

channel in each group. For example, when Nt = 2, Nt = 2, KG = 2, and Nf = 2048,

precoders of size 4 × 4 are designed for 1024 groups. We group the symbols from

the two subcarriers and two transmitted data streams into a 4×1 column vector as

the equivalent channel input. Each signal vector is then multiplied by the 4×4 linear

precoding matrix. By using a larger-size precoder, higher precoding gain can be

achieved than that of precoding without subcarrier grouping.

 110

4 8
Max Diversity
Two−Step Iterative Algorithm
3.5 7
Proposed Method
Channel Diagnolization
3 6
Mutual Information Classic WF

Mutual Information
2.5 5

2 4

1.5 3

Max Diversity
1 Two−Step Iterative Algorithm 2
Proposed Method
0.5 Channel Diagnolization 1
Classic WF
0 0
−15 −10 −5 0 5 10 15 20 −20 −15 −10 −5 0 5 10 15 20
SNR (dB) SNR (dB)

(a) 2×2 channel with QPSK inputs (b) 4×4 channel with QPSK inputs

6 12
Max Diversity
Two−Step Iterative Algorithm
5 10 Proposed Method
Channel Diagnolization
Classic WF
4 8
Mutual Information

Mutual Information
3 6

2 4
Max Diversity
Two−Step Iterative Algorithm
1 Proposed Method 2
Channel Diagnolization
Classic WF
0 0
−10 −5 0 5 10 15 20 25 −20 −15 −10 −5 0 5 10 15 20
SNR (dB) SNR (dB)

(c) 2×2 channel with 8PSK inputs (d) 4×4 channel with 8PSK inputs

8 16
Max Diversity
7 14 Proposed Method
Channel Diagnolization
6 12 Classic WF
Mutual Information

Mutual Information

5 10

4 8

3 6
Max Diversity
2 Two−Step Iterative Algorithm 4
Proposed Method
1 Channel Diagnolization 2
Classic WF
0 0
−10 −5 0 5 10 15 20 25 −20 −15 −10 −5 0 5 10 15 20 25
SNR (dB) SNR (dB)

(e) 2×2 channel with 16QAM inputs (f) 4×4 channel with 16QAM inputs

Figure 2. Mutual information vs. SNR

 111

In addition, an interleaver is employed for the subcarrier grouping so that

each interleaved group contains both statistically strong and weak frequency tones;

thus diversity gain is achieved by leveraging both spatial and frequency correlation

of the MIMO channels. To illustrated this, we assume that Hk represents the k-th

frequency tone of all the channel matrices, the matrix Hk has the same distribution
√ 1 1
as fk ΨRX2
Hw,k ΨRX2
[30] for Rayleigh fading, where ΨRX and ΨT X are the antenna

spatial correlation matrices due to angle spreads at the transmitter and receiver,

respectively. The matrix Hw,k is an Nr × Nt random matrix with i.i.d ∼ CN (0, 1)

elements. The scalar fk is the channel spectrum function given by:

L−1
X  
2πi(k − 0.5)
fk = 1 + 2 ai cos , k = 1, · · · , Nf (25)
i=1
Nf

and

L−i
X
ai = ΨISI (l, l + i) (26)
l=1

in which ΨISI is the inter-tap correlation matrix [30]. The average power of the k-th

frequency tone is

 
Tr E(Hk Hhk ) = fk · Tr(ΨRX ) · Tr(ΨT X ). (27)

According to (25) and (26), the low frequency tones (when k is small) have stronger

average power than the high frequency tones (when k is around Nf /2). Thus, the

objective of the interleaver is is that the interleaved frequency tones in different groups

have similar descending patters on the average power.

Figure 3 gives an example which indicates the effect of subcarrier grouping for

a 2 × 2 MIMO-OFDM system. For comparison purpose, signals without subcarrier

grouping are shown in Fig. 3.3(a). In Fig. 3.3(b), each group has signals from two
 112

subcarriers, i.e., KG = 2. It is seen that each group contains strong and weak signals

when interleaving is employed in the subcarrier grouping. In this way, both spatial

and frequency diversity are utilized.

16 16

14 14

12 12

10 10

8 8

6 6

4 4

2 2

0 0

(a) No subcarrier grouping (b) Precoding with subcarrier grouping

Figure 3. Amplitudes of precoded elements in a 2 × 2 MIMO-OFDM system. The

size of the dots is proportional to the amplitude of precoder coefficients. Only eight
subcarriers are shown.
 113

4 APPLICATIONS OF LINEAR PREDING TO MIMO-OFDM

In this section, we present the MIMO-OFDM system that employs the designed

4 × 4 linear precoding matrices for verification purposes. Specifically, we will discuss

the data frame structure, system diagram, and receiver algorithms.

Figure 4. System diagram of 2×2 MIMO-OFDM system.

BPSK Seq 1 Preamble 1 Preamble 2 Data Block 1 Data Block 2

Tx 1
144 128CP 256 128CP 256 128CP 2048 128CP 2048

BPSK Seq 2 Preamble 3 Preamble 4 Data Block 3 Data Block 4

Tx 2
144 128CP 256 128CP 256 128CP 2048 128CP 2048

Figure 5. Frame structure of MIMO-OFDM signaling

 114

The system diagram of the 2×2 MIMO-OFDM testbed is shown in Fig. 4,

where the transmitter employs a signaling frame structure shown in Fig. 5. The

144-bit BPSK sequence at each transmitter antenna is a frame header for frame and

symbol synchronization. The two BPSK sequences are two different pseudo-random

noise (PN) sequences generated by linear feedback shift registers with different initial

seeds. The two preambles at each antenna in Fig. 5 serve as training sequences for

channel estimation. Each preamble is consisted of a 256-bit length Zadroff-Chu (ZC)

sequence [31] and a 128-bit length cyclic prefix (CP). The CP length is chosen so that

it is greater than the length of the channel impulse response (CIR). The preambles

have the following structure [32, 33]:

   
∗
 Preamble 1 Preamble 2   Schu −Schu 
 =  (28)
∗
Preamble 3 Preamble 4 Schu Schu

where Schu is a 256-bit ZC sequence. Two OFDM data blocks follow the preambles,

and each data block consists of 2048 data symbols preceded by 128 symbols of CP.

As shown in Fig. 4, the data blocks at the transmitter are generated by two

raw source bit streams b1 and b2 encoded by a LDPC encoder with 3/4 coding rate

and codeword length of 2040 bits. We employ the LDPC channel codes specified

by the latest WiMAX standard [34], which has also been realized in the software

package in [35]. After LDPC encoding, each codeword is added 8 bits of zeros at the

end to form a data block of length 2048, which is then interleaved and modulated

to symbols. The supported modulation schemes in our implementation are QPSK,

8PSK, or 16QAM.

After precoding, the symbols of each stream are fed into a subcarrier grouper,

which is discussed in Section 3. After the subcarrier grouping, each data block is

converted to time domain signals by inverse fast Fourier transform (IFFT), followed

by CP insertion and preamble insertion. The digital up convertor (DUC) is used

 115

to up-sample the baseband signals and modulate them into the 17.5 MHz IF. The

output of the DUC module is fed to the digital to analog converter (DAC) directly.

At the receiver, the ADC bandpass sampling rate is 125 MHz and the IF is

70 MHz. The digital down convertor (DDC) module down-samples the IF signals to

baseband I and Q signals. At each receive antenna, the frame synchronizer captures

the received frame based on the time domain correlations of the local PN sequence.

After the start of the frame is located, the received preambles are used for channel

estimation which uses frequency domain LS method [32, 33]. The received preambles

are first converted to frequency domain with a 256-point FFT. For each subcarrier

of the preamble, LS estimation is performed to obtain an initial estimation of a 2×2

channel matrix. Then we convert the estimated channel matrices to time domain

with a 256-point IFFT and pass them through a rectangular window whose width

is larger than the length of the channel impulse response. Finally, we convert the

windowed CIR to frequency domain channel response Ĥi using a 2048-point FFT.

A turbo receiver [36] is employed to iteratively exchange log likelihood ratio

(LLR) between the MIMO MMSE-IC soft detector [16, 17] and the LDPC channel

decoder for each bit stream. Alternatively, the list fixed sphere decoder (FSD) [19,

18] is also implemented. Since the soft MIMO detector group received signals from

two subcarriers to detect symbols, the effective channel matrices Hei = ĤiGi , i =

1, 2, · · · , 1024, are used in both MMSE-IC and FSD schemes. The MIMO detector

generates extrinsic LLR, based on yi and the a priori information from the LDPC

channel decoders. The extrinsic information of MIMO detector is then interleaved and

sent to the LDPC decoder as its a priori input. By employing soft channel decoding

methods such as log domain sum-product algorithm [37], the LDPC decoder computes

its extrinsic information as output, which is fed back to the MIMO detector as the a

priori information. After two turbo iterations, hard decisions are made at the output

of the LDPC decoder.

 116

5 TESTBED, EXPERIMENT SETUP, AND RESULTS

In this section, we present the setup of the MIMO-OFDM testbed and discuss

the experiment results.

5.1 TESTBED AND EXPERIMENT SETUP

The testbed was built using the equipment listed in Tabel 3. The baseband

processing was implemented on Altera Stratix III FPGA development kits [38]. The

FPGA-DSP development kit features Stratix III EP3SL150F1152 high-performance

FPGA, and ADC/DAC daughter board. Baseband signal processing was implemented

in Altera FPGA and daughter board. The detailed architecture setup of the trans-

mitter and receiver are shown in Fig. 6 and Fig. 7.

Table 3. Description of Key Equipments

Function Equipment
FPGA platform for Tx Altera Stratix III FPGA Development kit
with daughter board
FPGA platform for Rx Altera Stratix IV FPGA Development kit
with daughter board
Receive antennas Pharad’s wearable antennas, BW-800-900-D
RF up-converter NuWaves, RF2-3000UCV1
Transmitter power amplifiers RF Bay, MPA-10-40
(optional)
Receiver low noise amplifier RF Bay, LNA-0915
(optional)
RF down-converter NuWaves, RF200-2500RV1
External clock for FS725 Rubidium Frequency Standard
frequency reference

In the transmitter shown in Fig. 6, we implemented the baseband modules and

digital up-converter (DUC) in the Stratix III FPGA and utilized the DAC daughter
 117

board to convert two data streams to analog IF signals. At each stream of the

transmitter, the IF signals were fed into the low pass filter SLP-50 with 50 MHz

bandwidth to reduce the out of band noise. Then the NuWave RF2-3000UCV1 RF up-

converter was employed to up-convert the IF signal to 915 MHz RF signals. The MPA-

10-40 RF was chosen for power amplification, before the signals being transmitted

through wearable antennas. The computer was used to program the FPGA and to

configure the RF up-converters. The transmit power can be adjusted manually by

changing the attenuation level of the RF up-converter. The FS725 rubidium clock

serve as an external reference clock for FPGA board and RF up-converters.

Figure 6. Transmitter setup architecture

At the receiver side, the signals were received by antennas and then down-

converted to 70 MHz IF signals by the RF front-end. The 2-channel 14-bit analog to

digital converter (ADC) daughter board was employed to convert signals into digital
 118

Figure 7. Receiver setup architecture

signals via bandpass sampling. The sampling frequency was also 125 MHz and the

folded IF frequency was 55 MHz.

5.2 FIELD EXPERIMENTS

Our experiments were conducted in three steps. Step 1 was channel acquisi-

tion without precoding. We transmited packets using identity matrices as precoders,

received the signals and estimated the channels. Step 2 was offline precoder design

for all subcarriers based on the estimated channels. After the precoding matrices

were computed at a fixed SNR, they were quantized and stored in read-only mem-

ories (ROMs) of the transmitter FPGA. Step 3 was performance verification with

precoding. A particular precoding method was selected and multiple frames were

transmitted with the assumption that the channels are static. We change the atten-

uation level of the RF upconverter at the transmitter to obtain different values of

signal transmission power. Under a specific modulation scheme, we use the manual

gain controller of the RF downconverter at the receiver side to adjust the spectra

of the received signals. Experiment data was logged at the DDC output through
 119

Signal Tap II logic analyzer of Quartus II and baseband receiver was implemented in

MATLAB.

We conducted indoor non-line-of-sight (NLOS) experiments in Room 208 of the

Emerson Electric Company Hall at the Missouri University of Science and Technology

campus. The transmitter and receiver are located in the same room with dimensions

of 9.1 m by 7.0 m. In addition, a metal bookshelf is placed between the transmitter

and receiver to block the line-of-sight communication path. The transmitter and

receiver antennas are located at half of the height of the blocking bookshelf above the

floor. The floor plan is shown in Fig. 8, where the ceiling is meters above floor.

Figure 8. Floor plan of the indoor NLOS environment (9.1 m × 7.0 m)

 120

CIR for h11 CIR for h12

600 600

500 500

400 400

Amplitude

Amplitude
300 300

200 200

100 100

0 0
0 5 10 15 20 0 5 10 15 20
Tap index Tap index

CIR for h21 CIR for h22

600 600

500 500

400 400
Amplitude

Amplitude
300 300

200 200

100 100

0 0
0 5 10 15 20 0 5 10 15 20
Tap index Tap index

Figure 9. Estimated channel impulse responses of indoor same room NLOS environ-
ment

5.3 EXPERIMENT RESULTS

Figure 9 plots a snapshot of the estimated channel impulse responses for the

indoor same room NLOS environment, where hij represents the channel from j-th

transmit antenna to i-th receive antenna. It is observed that the length of all the

channels are less than 20 taps. Our extensive experiments have shown similar char-

acteristics of the channel impulse responses. Therefore, the width of the time domain

window for channel estimation are set to 20.

Table 4 lists BER results of using list FSD method for data processing. The

experiments include different modulations for precoding schemes 1, 2, and 3 discussed

in Section 2, and the proposed precoding method in Section 3. From the results of

QPSK and 8PSK shown in Table 4, it is seen that the proposed method achieved low-

est BER among all the implemented precoding method in our system. In addition,
 121

Table 4. BER results using List FSD for indoor same room NLOS experiments
Mod Transmitter Classic Channel Maximum Proposed
attenuation WF diagonalization diversity method
(dB)
18 0.3896 0.3776 0.07614 0.00784
16 0.3806 0.1003 0 0
QPSK
14 0.3578 0.05833 0 0
10 0.3858 0 0 0
18 0.4149 0.4419 0.2015 0.121
16 0.4031 0.3425 0.08562 0.0488
8PSK
14 0.4116 0.2091 0.05403 0.03475
10 0.3700 0.2170 0.009695 0.002505

the maximum diversity outperformed the classic water-filling and channel diagonaliza-

tion. The performance gains of the the proposed method and the maximum diversity

over the other two schemes are because of the fact that these two approaches take into

account the specific structure of finite-alphabet inputs and offer higher mutual infor-

mation. Such observations from the experiments are similar to the simulation results

in [14], which includes the optimal precoding and the maximum diversity methods

with MAP detection. For the list FSD in our receiver, the node distribution [19] for

QPSK was ns = [1, 1, 4, 4], and the number of survival candidates to calculate LLR

was NL = 16. For 8PSK, the distribution was set to ns = [1, 1, 8, 8] and 16 candidates

with the minimum distances were used for LLR calculation.

 122

6 CONCLUSION

In this paper, a practical linear precoding method for MIMO-OFDM sys-

tem has been proposed that simplifies the two-step algorithm maximizing the lower

bound of the mutual information with finite-alphabet inputs. The proposed algorithm

achieves similar performance as the original two-step algorithm but requires only 3-

6% of the computational complexity, thus making the precoding design feasible to

practical systems.

The precoding design has been applied to a 2×2 MIMO-OFDM system with

2048 subcarriers by designing 1024 precoding matrices of size 4 × 4 based on the

extended channel matrices that combine the 2×2 MIMO with 2 subcarriers. Experi-

mental results of indoor NLOS experiments have showed that the proposed precoding

method has achieved the lowest BER in comparison to maximum diversity, classic

water-filling and channel diagonalization methods.

 123

7 ACKNOWLEDGMENTS

This work was supported in part by Office of Naval Research under Grant

N00014-09-1-0011 and by National Science Foundation under Grant ECCS-0846486

and CCF-0915846. The authors would like to thank Mr. Weiliang Zeng for providing

MATLAB codes of iterative two-step algorithm. We also acknowledge Mr. Bing

Han and Drs. Fei Ren and Huang Lou for their help with the FPGA transceiver

implementation and field experiments.

 124

8 REFERENCES

[1] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. New
York: Wiley, 2006.

[2] Y. Xin, Z. Wang, and G. B. Giannakis, “Space-time diversity systems based on

linear constellation precoding,” IEEE Trans. Wireless Commun., vol. 2, pp. 294-
309, Mar. 2003.

[3] D. P. Palomar, J. Cioffi and M. A. Lagunas, “Joint Tx-Rx beamforming design

for multicarrier MIMO channels: A unified framework for convex optimization,”
IEEE Trans. Signal Process., vol. 51, pp.2381-2401, Sep 2003.

[4] D. P. Palomar and S. Verdú, “Gradient of mutual information in linear vector

Gaussian channels,” IEEE Trans. Inform. Theory, vol.52, pp.141-154, Jan. 2006.

[5] A. Lozano, A. M. Tulino, and S. Verdú, “Optimum power allocation for paral-
lel Gaussian channels with arbitrary input distributions,” IEEE Trans. Inform.
Theory, vol.52, pp.3033-3051, July 2006.

[6] C. Xiao and Y. R. Zheng, “On the mutual information and power allocation
for vector Gaussian channels with finite discrete inputs,” in Proc. IEEE Glocal
Telecommunications Conference (Globecom), New Orleans, USA, Nov. 30 - Dec. 4,
2008.

[7] C. Xiao and Y. R. Zheng, “Transmit precoding for MIMO systems with partical
CSI and discrete-constellation inputs,” in Proc. IEEE International Conference on
Communications (ICC), Dresden, Germany, June 14-18, 2009.

[8] M. Payaró and D. P. Palomar, “On optimal precoding in linear vector Gaussian
channels with arbitrary input distribution,” in Proc. IEEE International Sympo-
sium on Information Theory, Seoul, Korea, Jun. 28 - Jul. 3, 2009.

[9] M. Lamarca, “Linear precoding for mutual information maximization in MIMO

systems,” in Proc. of 6th International Symposium on Wireless Communication
Systems, Siena, Italy, 2009.

[10] F. Pérez-Cruz, M. R. D. Rodrigues, and S. Verdú, “MIMO Gaussian channels

with arbitrary inputs: optimal precoding and power allocation,” IEEE Trans. In-
form. Theory, vol.56, pp. 1070-1084, Mar. 2010.

[11] C. Xiao, Y. R. Zheng, and Z. Ding, “Globally optimal linear precoders for fi-
nite alphabet signals over complex vector Gaussian channels,” IEEE Trans. Signal
Process., vol. 59, pp. 3301-3314, July 2011.
 125

[12] M. Wang, C. Xiao, and W. Zeng, “Linear precoding for MIMO multiple access
channels with finite discrete inputs,” IEEE Trans. Wireless Commun., vol. 10, pp
3934-3942, Nov. 2011.

[13] W. Zeng, Y. R. Zheng, M. Wang, and J. Lu, “Linear precoding for relay networks:
a perspective on finite-alphabet inputs,” IEEE Trans. on Wireless Commun., vol.
11, pp. 1146-1157, Mar 2012.

[14] W. Zeng, C. Xiao, M. Wang, and J. Lu, “Linear precoding for finite-alphabet
inputs over MIMO fading channels with CSI,” IEEE Trans. Signal Process., vol.
60, July 2012 (scheduled to appear).

[15] B. Hochwald and S. T. Brink, “Achieving near-capacity on a multiple-antenna

channel,” IEEE Trans. Commun., vol. 51, pp. 389-399, Mar. 2003.

[16] L. Boher, R. Rabineau, and M. Helard, “An efficient MMSE equalizer implemen-
tation for 4×4 MIMO-OFDM systems in frequency selective fast varing channels,”
in IEEE Proc. Personal, Indoor and Mobile Radio Communications (PIMRC),
Athens, Greece, Sept 3-7, 2007.

[17] L. Boher, R. Rodrigue and M. Helard, “FPGA implementation of an iterative

receiver for MIMO-OFDM Systems,” IEEE J. Sel. Areas Commun., vol. 26, pp.
857-866, Aug. 2008.

[18] L. G. Barbero and J. S. Thompson, “Fixing the complexity of the sphere decoder
for MIMO detection,” IEEE Trans. Wireless Commun., vol. 7, pp. 2131-2142, Jun.
2008.

[19] L. G. Barbero and J. S. Thompson, “Extending a fixed-complexity sphere de-

coder to obtain likelihood information for Turbo-MIMO systems,” IEEE Trans.
Veh. Technol., vol. 57, pp. 2804-2814, Sep. 2008.

[20] H. Sampath, S. Talwar, J. Tellado, V. Erceg, and A. Paulraj, “A fourth-

generation MIMO-OFDM broadband wireless system: design, performance, and
field trial results,” IEEE Commun. Mag., vol. 40, pp. 143-149, Sep. 2002.

[21] C. Dubuc, D. Starks, T. Creasy, and H. Yong, “A MIMO-OFDM prototype for

next-generation wireless WANs,” IEEE Commun. Mag., vol. 42, pp. 82-87, Dec.
2004.

[22] A. van Zelst and T. C. W. Schenk, “Implementation of a MIMO OFDM based-

wireless LAN system,” IEEE Trans. Signal Process., vol. 52, pp. 483-494, Feb.
2004.

[23] H. Yu, M. Kim, E. Choi, T. Jeion, and S. Lee, “Design and prototype develop-
ment of MIMO-OFDM for next generation wireless LAN,” IEEE Trans. Consumer
Electron., vol. 51, pp. 1134-1142, Nov. 2005.
 126

[24] T. Haustein, A. Forck, H. Gäbler, V. Jungnickel, and S. Schiffermüller, “Real-

time signal processing for multiantenna systems: algorithms, optimization, and
implementation on an experimental test-bed,” EURASIP Journal on Applied Signal
Processing, 2006, Article ID 27 573.

[25] S. Haene, D. Perels, and A. Burg, “A real-time 4-stream MIMO-OFDM

transceiver: system design, FPGA implementation, and characterization,” IEEE
J. Sel. Areas Commun., vol. 26, pp. 877-889, Aug. 2008.

[26] C. Gómez-Calero, L. C. Navarrete, L. Haro, and R. Martiı́nez, “A 2×2 MIMO

DVB-T2 system: design, new channel estimation scheme and measurements With
polarization diversity,” IEEE Trans. on Broadcast., vol. 57, pp. 195-203, June 2011.

[27] E. Telatar, “Capacity of multi-antenna Gaussian channels,” Eur. Trans.

Telecomm. ETT, vol. 10, pp. 585-596, Nov. 1999.

[28] A. Goldsmith, “Wireless communications,” Cambridge University Press, 2005.

pp.301

[29] W. Zeng, C. Xiao, and J. Lu, “A low-complexity design of linear precoding for
MIMO channels with finite-alphabet inputs,” accepted by IEEE Wireless Commun.
Lett., Dec. 2011.

[30] C. Xiao and Y. R. Zheng, “On the ergodic capacity of MIMO triply selective
Rayleigh fading channels,” IEEE Trans. Wireless Commun., vol.7, pp.2272-2279,
June 2008.

[31] D. Chu, “Polyphase codes with good periodic correlation properties,” IEEE
Trans. Inform. Theory, vol.18, pp. 531-532, July 1972.

[32] A. N. Mody and G. L. Stuber, “Receiver implementation for a MIMO OFDM

system,” in Proc. Global Communications Conference (Globecom), Taipei, Taiwan,
Nov. 17-21, 2002.

[33] G. L. Stuber, J. Barry, S. McLaughlin, Y. (G.) Li, M. A. Ingram, and T. Pratt,

“Broadband MIMO-OFDM wireless communications,” Proceedings of IEEE, vol.
92, pp. 271-294, Feb. 2004.

[34] Part 16:Air Interface for Broadband Wireless Access Systems, IEEE Std
802.16TM -2009, May. 2009.

[35] M. Valenti, The Coded Modulation Library, http://www.iterativesolutions.com.

[36] J. Ylioinas and M. Juntti, “Iterative joint detection, decoding, and channel es-
timation in Turbo-coded MIMO-OFDM,” IEEE Trans. Veh. Technol., vol. 58, pp.
1784-1796, May. 2009.
 127

[37] X-Y. Hu, E. Eleftherious, D-M. Arnold, and A. Dholakia, “Efficient implemen-
taions of the sum-product algorithm for decodng LDPC codes,” in Proc. IEEE
Glocal Telecommunications Conference (Globecom), San Antonio, USA, Nov. 25 -
29, 2001.

[38] Altera. Stratix III 3SL150 development kit reference manual [Online]. Available:
http://www.altera.com/products/devkits/altera/kit-siii-host.html
 128

SECTION

2 CONCLUSIONS

This dissertation proposed a new approach to maximizing data rate/throughput

of practical communication system/networks through linear precoding and power al-

location. By maximizing the mutual information with finite-alphabet inputs such as

PSK, PAM and QAM signals, this approach improves performance when the designed

precoders are applied to practical systems and networks.

Several numerical optimization methods were developed for MIMO multiple

access channels, dual-hop relay networks, and point-to-point MIMO systems. In

MIMO multiple access channels, an iterative weighted sum rate maximization algo-

rithm was proposed which utilized an alternating optimization strategy and gradient

descent update. In dual-hop relay networks, the structure of the optimal precoder

was exploited to develop a two-step iterative algorithm based on convex optimiza-

tion and optimization on the Stiefel manifold. The gradient descent method was

also used to obtain the optimal power allocation scheme in dual-hop relay networks.

For point-to-point MIMO systems, a low complexity precoding design method was

proposed, with discretized power allocation vector in a non-iterative fashion, thus

reducing complexity.

Performances of the proposed power allocation and linear precoding schemes

were evaluated in terms of both mutual information and BER. Numerical results

showed that at the same target mutual information or sum rate, the proposed ap-

proaches achieved 3-10 dB gains compared to the existing methods in the medium

signal-to-noise ratio region. Such significant gains were also indicated in the coded

BER systems.
 129

The contributions of my PhD research work are summarized in five journal

papers and seven conference papers, among which, three journal papers and one

conference paper are included in this dissertation.

 130

3 PUBLICATIONS

[1] M. Wang, Y. R. Zheng, and C. Xiao, “Practical linear precoder design for finite
alphabet MIMO-OFDM with experiment validation,” to be submitted to IEEE
Trans. Veh. Technol..
[2] Y. Wu, M. Wang, C. Xiao, Z. Ding, and X. Gao, “Linear precoding for finite
alphabet signals in multi-antenna broadcast channels,”, accepted by IEEE Int.
Conf. Commun. (ICC), Ottawa, Canada, Jun., 2012.
[3] W. Zeng, Y. R. Zheng, M. Wang, and J. Lu, “Linear precoding for relay networks:
a perspective on finite-alphabet inputs,” IEEE Trans. Wireless Commun., vol.
11, pp 1146 - 1157, Mar. 2012.
[4] W. Zeng, C. Xiao, M. Wang, and J. Lu, “Linear precoding for finite-alphabet in-
puts over MIMO fading channels with statistical CSI”, accepted by IEEE Trans.
Signal Process., Feb. 2012.
[5] M. Wang, W. Zeng, and C. Xiao, “Linear precoding for MIMO multiple access
channels with discrete-constellation inputs,” accepted by IEEE Global Commun.
Conf. (Globecom), Houston, TX, USA, Dec. 2011.
[6] W. Zeng, C. Xiao, M. Wang, and J. Lu, “On the linear precoder design for
MIMO channels with finite-alphabet inputs and statistical CSI,” accepted by
IEEE GLOBECOM, Houston, TX, USA, Dec. 2011.
[7] M. Wang, C. Xiao, and W. Zeng, “Linear precoding for MIMO multiple access
channels with finite discrete inputs,” IEEE Trans. Wireless Commun., vol. 10,
pp 3934 - 3942, Nov. 2011.
[8] Y.Wu, M. Wang, C. Xiao, Z. Ding, and X. Gao, “Linear precoding for MIMO
broadcast channels with finite-alphabet constraints,” submitted to IEEE Trans.
Wireless Commun., Nov. 2011.
[9] S. Subedi, H. Lou, F. Ren, M. Wang, and Y. R. Zheng, “Validation of the triply
selective fading channel model through a MIMO test bed and experimental re-
sults,” in Proc. IEEE Military Communi. Conf. (MILCOM), Nov. 2011.
[10] W. Zeng, C. Xiao, M. Wang, and J. Lu, ”Linear precoding for relay networks with
finite-alphabet constraints,” in Proc. IEEE Int. Conf. Commun. (ICC), Kyoto,
Japan, Jun. 2011.
[11] W. Zeng, M. Wang, C. Xiao, and J. Lu, “On the power allocation for relay
networks with finite-alphabet constraints,” in Proc. IEEE Global Commun. Conf.
(Globecom), Miami, FL, USA, Dec. 2010.
 131

VITA

Mingxi Wang received his B.S degree in Telecommunication Engineering from

Beihang University (formerly: Beijing University of Aeronautics and Astronautics),

China, in 2006, and the M.S degree in Signal and Information Processing from the

same university in 2009. He began his Ph.D study in August 2009 at the Department

of Electrical and Computer Engineering at Missouri University of Science and Tech-

nology (formerly: University of Missouri-Rolla), USA. His research interests include

wireless communications, signal processing, MIMO, OFDM, linear precoding and it-

erative receiver. He received his Ph.D. degree in Electrical Engineering from Missouri

University of Science and Technology in Aug. 2012.