國
立
交
通
大
學
電機與控制工程學系
碩
士
論
文
正交時空區塊碼之頻域等化及其基於訓練機制
之多輸入多輸出的擇頻通道估測
On Frequency-Domain Equalization with Training-Based Channel
Estimation for Orthogonal Space-Time Block Coded System via
MIMO Frequency-Selective Fading Channels
研 究 生:楊傑智
楊傑智
指導教授:林清安 教授
正交時空區塊碼之頻域等化及其基於訓練機制之多輸入多輸出的擇頻通道估測
On Frequency-Domain Equalization with Training-Based Channel Estimation for Orthogonal Space-Time Block Coded System via MIMO Frequency-Selective Fading Channels
研 究 生:楊傑智
Student:Chieh-Chih Yang
指導教授:林清安 Advisor:Ching-An Lin
國 立 交 通 大 學
電 機 與 控 制 工 程 學 系
碩 士 論 文
A ThesisSubmitted to Department of Electrical and Control Engineering College of Electrical Engineering and Computer Science
National Chiao Tung University in partial Fulfillment of the Requirements
for the Degree of Master
in
Electrical and Control Engineering
August 2007
Hsinchu, Taiwan, Republic of China
國 立 交 通 大 學
國 立 交 通 大 學
國 立 交 通 大 學
國 立 交 通 大 學
博碩士論文全文電子
博碩士論文全文電子
博碩士論文全文電子
博碩士論文全文電子檔
檔
檔
檔著作權授權書
著作權授權書
著作權授權書
著作權授權書
(提供授權人裝訂於紙本論文書名頁之次頁用)
本授權書所授權之學位論文,為本人於國立交通大學電機與控制工程
系所 ______組, 95 學年度第____學期取得碩士學位之論文。
論文題目:正交時空區塊碼之頻域等化及其基於訓練機制之多輸入多
輸出的擇頻通道估測
指導教授:林清安
■ 同意
本人茲將本著作,以非專屬、無償授權國立交通大學與台灣聯合大學
系統圖書館:基於推動讀者間「資源共享、互惠合作」之理念,與回
饋社會與學術研究之目的,國立交通大學及台灣聯合大學系統圖書館
得不限地域、時間與次數,以紙本、光碟或數位化等各種方法收錄、
重製與利用;於著作權法合理使用範圍內,讀者得進行線上檢索、閱
覽、下載或列印。
論文全文上載網路公開之範圍及時間:
本校及台灣聯合大學系統區域網路
■ 立即公開
校外網際網路
■ 立即公開
■
■
■
■
全文電子
全文電子
全文電子
全文電子檔
檔
檔送交國家圖書館
檔
送交國家圖書館
送交國家圖書館
送交國家圖書館
授 權 人
授 權 人
授 權 人
授 權 人:
:
:楊傑智
:
楊傑智
楊傑智
楊傑智
親筆簽名
親筆簽名
親筆簽名
親筆簽名:
:
:______________________
:
______________________
______________________
______________________
中華民國
中華民國
中華民國
中華民國
年
年
年
年
月
月
月
月
日
日
日
日
第 1 頁,共 4 頁國 立 交 通 大 學
國 立 交 通 大 學
國 立 交 通 大 學
國 立 交 通 大 學
博碩士紙本論文著作權授權書
博碩士紙本論文著作權授權書
博碩士紙本論文著作權授權書
博碩士紙本論文著作權授權書
(提供授權人裝訂於全文電子檔授權書之次頁用)
本授權書所授權之學位論文,為本人於國立交通大學電機與控制工程
系所 ______組, 95 學年度第____學期取得碩士學位之論文。
論文題目:正交時空區塊碼之頻域等化及其基於訓練機制之多輸入多
輸出的擇頻通道估測
指導教授:林清安
■ 同意
本人茲將本著作,以非專屬、無償授權國立交通大學,基於推動讀者
間「資源共享、互惠合作」之理念,與回饋社會與學術研究之目的,
國立交通大學圖書館得以紙本收錄、重製與利用;於著作權法合理使
用範圍內,讀者得進行閱覽或列印。
本論文為本人向經濟部智慧局申請專利(未申請者本條款請不予理會)
的附件之一,申請文號為:____________________,請將論文延至
____年____月____日再公開。
授 權 人
授 權 人
授 權 人
授 權 人:
:
:楊傑智
:
楊傑智
楊傑智
楊傑智
親筆簽名
親筆簽名
親筆簽名
親筆簽名:
:
:______________________
:
______________________
______________________
______________________
中華民國
中華民國
中華民國
中華民國
年
年
年
年
月
月
月
月
日
日
日
日
第 2 頁,共 4 頁國家圖書館
國家圖書館
國家圖書館
國家圖書館
博碩士論文電子
博碩士論文電子
博碩士論文電子
博碩士論文電子檔
檔
檔
檔案上網授權書
案上網授權書
案上網授權書
案上網授權書
(提供授權人裝訂於紙本論文本校授權書之後)
ID:GT009412501
本授權書所授權之論文為授權人在國立交通大學電機與控制工程系所
95 學年度第___學期取得碩士學位之論文。
論文題目:正交時空區塊碼之頻域等化及其基於訓練機制之多輸入多
輸出的擇頻通道估測
指導教授:林清安
茲同意將授權人擁有著作權之上列論文全文(含摘要),非專屬、無
償授權國家圖書館,不限地域、時間與次數,以微縮、光碟或其他各
種數位化方式將上列論文重製,並得將數位化之上列論文及論文電子
檔以上載網路方式,提供讀者基於個人非營利性質之線上檢索、閱
覽、下載或列印。
※ 讀者基於非營利性質之線上檢索、閱覽、下載或列印上列論文,應依著作權法相
關規定辦理。
授權人
授權人
授權人
授權人:
:
:楊傑智
:
楊傑智
楊傑智
楊傑智
親筆簽名
親筆簽名
親筆簽名
親筆簽名:
:
:_______________
:
_______________
_______________
_______________
民國
民國
民國
民國
年
年
年
年
月
月
月
月
日
日
日
日
第 3 頁,共 4 頁正交時空區塊碼之頻域等化及其基於訓練機制之多輸入多輸出的
擇頻通道估測
學生:楊傑智 指導教授:林清安
國立交通大學電機與控制工程學系(研究所)碩士班
摘
摘
摘
摘
要
要
要
要
針對正交時空區塊碼應用於多輸入多輸出之擇頻通道上的問題,吾人於本文中提出了一 個嶄新且完整的推導方式;另著墨於此架構下所能達到的最大之多重增益-包括來自於 天線以及多重路徑所貢獻者。藉由多次的模擬實驗結果發現,吾人可以得到一個於實際 考量下為合理的假設,進而由此推導錯誤率分析。錯誤率分析的推導結果證實了吾人所 提出之架構確實是有能力達到完整的多重增益。另一方面,藉由接收端的頻域等化機制 以及所提出之傳輸架構,吾人可以進一步找到基於訓練機制時最佳的通道估測所需之相 對應的訊號設計模式。本論文之結果為線性等化器能否達到完整的多重增益提供了一個 正面的佐證。
On Frequency-Domain Equalization with Training-Based Channel
Estimation for Orthogonal Space-Time Block Coded System via MIMO
Frequency-Selective Fading Channels
Student:Chieh-Chih Yang Advisor:Ching-An Lin
Department(Institute)of Electrical and Control Engineering
National Chiao Tung University
Abstract
We propose an instructive derivation for the generalized block-level orthogonal space-time block encoder, capable of achieving full spatial diversity via frequency- selective fading environment provided that channel order is known. Instead of dealing with special case and then extending the results intuitively, we provide an alternative by starting with the general signal model with multiple transmit and multiple receive antennas, from which a general form of block-level orthogonality is established. In particular, transmit diversity with more than two transmit antennas can be achieved without compromise by means of frequency-domain equalization, in contrast to the QO-STBC-based approach. Pairwise error probability analysis is derived, under certain assumption which is numerically supported by simulation results, for analytical verifications of our claim on full diversity, inclusive of transmit-receive diversity and the multipath one. Moreover, the encoder structure enables us to generalize a training-based channel estimation technique, originally proposed for flat-fading scenario, to the frequency-selective fading scenario. Surprisingly we even obtain similar optimality criteria for optimal training block design which in our case, the signal block are fixed as OSTBC-based and the design derivation reduces to derive optimal power constraint over the training blocks. The optimality criteria for the training blocks are easy to satisfy when randomness of signal constellation is not a concern. Simulation results validate our discussion of the behaviors of the least-squares and linear MMSE channel estimates.
Acknowledgement
I would like thank my instructor, Prof. Ching-An Lin, for his teaching, enlightenment and support over the past two years. The thesis would never be complete without his dedication to the revisions and his provision of ingenuity. The valuable joint advice from Prof. Yuan-Pei Lin, Prof. Wing-Kin Ma and Dr. Jwo-Yu Wu serves no less than the above efforts. Especially, issues regarding simulation accuracy and rigorousness of proofs are greatly improved with their assistances. Moreover, the generous helps from the seniors of our laboratory and my friends see me through many difficulties, and I am cordially grateful to all of them. Above all, I would like thank my family for their encouragement.
Contents
摘要 摘要 摘要 摘要 ...i Abstract ...ii Acknowledgement...iiiList of Figures ...vii
Chap. 1 Introduction ... 1
1.1 Motivation and Background ... 1
1.2 Thesis Overview ... 3
1.3 Notations... 4
Chap. 2 System Model ... 5
2.1 System Configuration ... 5
2.1.1 BGOSTBC Encoder... 5
2.1.2 BGOSTBC Encoder (with CP Insertions) ... 7
2.2 Block Transmission and IBI-free Model upon CP Removal... 8
2.2.2 Block Transmission via MIMO Frequency-Selective Fading Channels ... 8
Chap. 3 Frequency-Domain Equalization... 11
3.1 Overall Frequency-Domain Received Signal Model... 11
3.1.1 Space-Time Combining Using Signal DFT’s ... 12
3.1.2 Frequency-Domain Equalization with Perfect CSI ... 13
Chap. 4 Pairwise Error Probability Analysis... 17
4.1 PEP Analysis for Suboptimal Detection Problem ... 17
Chap. 5 Training-Based Channel Estimation... 28
5.1 Least-Squares Channel Estimate ... 28
5.1.1 Optimality Criterion for LS Estimate under a Power Constraint... 31
5.1.2 Selection of training signal blocks for optimal LS channel estimate... 33
5.2 Linear Minimum-Mean-Square Error Estimate ... 34
5.2.1 Optimality Criterion for Linear MMSE Estimate under a Power Constraint ... 36
Chap. 6 Simulation Results ... 40
6.2.1 Arbitrary Training: NMSE Performance ... 41
6.2.2 Optimal/Orthogonal Training: NMSE Performance... 41
Chap. 7 Conclusion... 42
Appendix ... 43
A1. Proof of Block-Level Orthogonality part I... 43
A2. Proof of block-level orthogonality part II ... 44
A3. OSTBC construction ... 45
List of Figures
Fig. 1 Transmission scheme based on BGOSTBC encoder and the associated FDE
based on training-based channel estimation. ... 48
Fig. 2 A BGOSTBC encoder employing K=4, Mt=3 configuration... 48
Fig. 3-1 Condition numbers generated for 106 independent simulation runs under
(N,Mr,Mt,L,K)=(10,1,3,1,4)... 49
Fig. 3-2 Condition numbers generated for 106 independent simulation runs under
(N,Mr,Mt,L,K)=(16,1,3,1,4)... 49
Fig. 3-3 Condition numbers generated for 106 independent simulation runs under
(N,Mr,Mt,L,K)=(16,1,3,2,4)... 50
Fig. 3-4 Condition numbers generated for 106 independent simulation runs under
(N,Mr,Mt,L,K)=(10,2,6,1,8)... 50
Fig. 3-5 Condition numbers generated for 106 independent simulation runs under
(N,Mr,Mt,L,K)=(10,2,6,3,8)... 51
Fig. 3 Compare NMSE of LS and linear MMSE estimates, with arbitrary training. 51
Fig. 4 Compare NMSE of arbitrary training and optimal/orthogonal training... 52
Chapter 1
Introduction
1.1 Motivation and Background
Orthogonal space-time block code (OSTBC) was introduced in [1] for combating channel fading by exploiting diversities in multiple-input-multiple-output (MIMO) antenna configuration. By virtue of orthogonal design, the symbol decoding relies on only linear processing with relatively lower computational complexity than its trellis counterpart, provided that the channel state information is available. Hence both blind channel estimation and joint symbol detection using OSTBC architecture are well studied in flat-fading scenario.
Various transmission schemes and algorithms based on suitably manipulated orthogonality conditions have been proposed for flat-fading environment[19],[20]. In [19], joint signal detection for a flat-fading MIMO communication under OSTBC scheme was studied, where the exponent, embedded with OSTBC structure, of the closed-form log-likelihood was to be minimized in the maximum likelihood (ML) sense. A non-polynomial time (NP-hard) ML exhaustive search was cleverly reformulated and relaxed as a convex optimization problem which had been well studied theoretically and numerically. The relaxation is known as semidefinite relaxation (SDR) with corresponding numerical algorithm tailored in [21]. The resultant SDR-ML proposed by Ma et al. performs substantially better than the cyclic ML method[22]. It is noteworthy that the reformulation itself exploited the orthogonality of OSTBC, by which a near optimum signal detection with relatively lower computational cost than the optimal ML/sphere decoding was achieved. In [20], a closed-form channel estimation technique based on a variant of the generic orthogonality of OSTBC was developed. This variant was first derived by Alex B. Gershman et al. in [23], where real parts and imaginary parts of the received signal model were separated with discretion. The closed-form channel estimation, unlike the other approaches such as subspace method, suffers from sign ambiguity only due to its special real-valued formulation of signal models. Several
sufficient antenna configurations for unique channel estimate were studied numerically. The uniqueness of the channel estimate was determined by the discrepancy between the eigenvectors of a certain objective which was formed by exploiting the orthogonality. For those configurations where uniqueness failed, i.e., algebraic multiplicity of eigenvectors greater than 1, the authors proposed a diagonal precoder to alienate the eigenvectors. Benefited from the orthogonality, this can be surprisingly easily done by assigning the power weighting coefficients with sufficient discrepancy in an ad-hoc manner.
However, it remains challenging for the above mentioned methods to be effectively extended to frequency-selective fading environment which is a more practical consideration. By effective extension we mean the algorithm on which the estimation or decoding is based should enjoy either full or partial diversity due to orthogonal nature of the coding itself. To apply the OSTBC structure to frequency- selective case without compromising the code orthogonality, E. Lindskog and A. Paulraj [7] cleverly combine the cyclic-prefix (CP) mechanism with time-reversal operation on symbol blocks, bringing the space-time concept to “block-level”. It was then incorporated with Alamouti scheme[8] and known as time-reversed Alamouti-like (TR-Alamouti) scheme[9]. The TR-Alamouti scheme is unique in that it enjoys full 2-fold transmit diversity and nearly full transmission rate, when neglecting the CP overheads, at the same time. A general block-level orthogonal space-time block code was first proposed by Z. Liu et al.[14] for consideration regarding preserving orthogonality over frequency- selective channels with more than two transmit antennas, where the generalized complex orthogonal design (GOSTBC) [1] was adopted with ZP assistance for mitigating channel distortion. On the other hand, there have been extensive studies based on the 2-fold diversity scheme with frequency-domain equalization (FDE), such as [11],[12] and [13]. As for a general FDE scheme having more than 2 transmit antennas, a compromising method based on quasi-orthogonal STBC (QO-STBC) [10] was proposed in [2], where the nearly full rate was preserved at the cost of achieving only partial spatial diversity. Achieving full spatial diversity, particularly based on FDE and cast into the general structure in [14], was reported in [26]. However, the derivation in [26] is based on the 2-fold structure as a start and then generalized to multiple-antenna scenario intuitively by introducing the block-level concept in [14]. It is as instructive as
Aside from the derivation issue, we will counterbalance the deduction in [26] that the “CP-only” scheme cannot exploit full multipath diversity, by giving a PEP analysis. Although a premise which is numerically supported by simulations must be met for justification of our PEP analysis, it does shade some light on extending the discussions in [17] to a more general scheme.
Also, optimal training design for MIMO communications in either flat-fading or frequency-selective fading environments is an important topic in practice. A pilot symbol-aided linear MMSE-based training scheme with optimal/orthogonal training is considered in [15]. In [16], the discussion in [15] was generalized to take advantage of space-time diversity. Nonetheless, none of which particularly considered the OSTBC class, and hence the optimal designs were quite involved. It will be shown that with the structure of training blocks fixed as OSTBC, the optimal training design can be simplified to optimal power allocation design.
1.2 Thesis Overview
In this thesis, we introduce a generalized FDE technique based on structure of [14] and by extending the training-based channel estimation approach in [3] we arrive at similar design criteria for optimal/orthogonal training as those obtained in [3] which considers flat-fading scenario. We propose an extended block-level OSTBC scheme capable of achieving full spatial and multipath diversities over frequency-selective fading channels when more than two transmit antennas involved by using FDE. What differs from [26] is that instead of starting from the 2-fold special case, we provide a new and instructive derivation based on general multiple-antenna signal model. The proposed scheme has nearly 1/2 symbol rate when discarding CP overheads. Pairwise error probability (PEP) analysis for demonstrating full spatial and full multipath diversities will be given. It will also be shown that since the signal model resembles that in flat-fading scenario, the optimal training designs such as those developed by M. Biguesh et al.[3] can be generalized to the frequency-selective fading scenario in a straightforward manner. The extended training-based channel estimation is optimal in least-squares (LS) sense in frequency-domain under a given power constraint, provided that a power criterion is satisfied. Also a time-domain linear MMSE channel estimation technique can be developed in a similar fashion as [3]. Adopting the Alamouti scheme for the construction
of the proposed encoder with proper changes in dimension and scalar factor makes the 2-fold diversity scheme a special case.
In the next chapter, we will propose a transmission scheme and outline that entire system model. Post-processing including the subsequent FDE for data transmission and the training-based channel estimation for training mode will be introduced in chapter 3 and 5, respectively. Also, the optimality conditions for achieving lowest possible NMSE will be derived accordingly. In chapter 4, we will derive our PEP analysis. Simulation results of NMSE vs. SNR for training-based channel estimation are discussed in chapter 6.
1.3 Notations
The following notations are adopted throughout the thesis: P*, PT and PH denote conjugate, transpose and conjugate-transpose of matrix P, respectively. A⊗B stands for the kronecker product of matrix A and B. Let Re{P} and Im{P} stand for the real and imaginary parts of matrix P, respectively. Let Tr{P} and vec(P) denote the trace and the vectorization of matrix P, respectively. For y∈ , Diag(y)N ∈ N N× stands for an N by N diagonal matrix with y on its main diagonal. For A∈N N× , Diag(A)∈ N N× stands for the vector whose ith
entry is the ith
diagonal entry of A. For a matrix N M×
∈
A ,
[ ]
ijA denotes the entry at the ith
row-and-jth
column position of A . P(i:j,m:n) denotes an extracted submatrix consists of from ith
to jth
rows and from mth to nth
columns of matrix P. P(:,m:n) indicates that all rows ranging from mth to nth
columns are referred. Similarly define P(i:j,:). The symbol F is preserved for N by N normalized discrete-time Fourier transform (DFT) matrix, with the (m,n)th
entry of F being [ ] 2 ( 1)( 1) 1 m n j N mn N e π − − − = F , 1≤m n, ≤N.
Chapter 2
System Model
In this chapter, we will introduce a transmission scheme based on a block-level extension of OSTBC. As a preliminary, we will also review some basic properties of orthogonal design of space-time block codes, which are to be employed for introducing block-level generalized orthogonal space-time block code (BGOSTBC) encoder as depicted in Fig. 2. The mechanism of CP insertion and CP removal for combating inter-block interference (IBI) will be reviewed at the bottom of this chapter.
2.1 System Configuration
As depicted in Fig. 1, the overall system consists of a BGOSTBC encoder followed by CP insertions at the transmitting end while the receiving end comprises CP removal, DFT and the subsequent channel estimation plus FDE. Let Mt and Mr denote the number of transmit antennas and receive antennas, respectively. Assume that the channel order is L (L+1 taps) for all the subchannels and known a priori. We assume the CP length is exactly L. Let N be the block length and N ≥ + . The information L 1 symbol blocks to be transmitted are accumulated over K blocks, each one of which will be sent by a certain transmit antenna during a specific time epoch with CP insertion. The block ordering is set up according to the proposed BGOSTBC encoder which occupies 2K time epochs for transmitting KN symbols. Each time epoch lasts for N+L symbol periods, where the redundant L symbol duration accounts for the CP insertion. At the receiving end, the received signal blocks over the entire 2K time epochs are buffered after CP removals. Then the equivalent IBI-free received blocks are Fourier transformed. With the outputs at the DFT system block available, we can either acquire channel estimation in training mode or perform FDE on the transmitted information symbols.
In this subsection, the referred time epoch lasts for N symbol periods only since no CP insertion involved. Now suppose we have collected K symbol blocks, c1, , c2 , cK. Let C = c1 c2 cK, where
N k ∈
c denotes the kth
information symbol block to be transmitted, whose nth
symbol is denoted by c nk( ), 0≤ ≤n N −1, 1≤ ≤k K. The transmission scheme of the BGOSTBC encoder can be regarded as a forward transmission mode followed by a reversed transmission mode, each of them occupying exactly K epochs. In the latter mode each incoming signal block is time-reversed and conjugated prior to transmission, while in the former mode transmission is carried out with blocks unmodified.
Now, let us define the encoder output. The encoder has single serial input and parallel Mt outputs. Let
( )p N m ∈
s denote the signal block transmitted from the mth
encoder output path during the pth
time epoch. Collecting encoder output over 2K time epochs and across Mt encoder output paths yields a matrix of dimension 2NK by Mt, whose ((p-1)N+1:pN,m)th
block-entry is ( )p m
s by definition. The so obtained matrix is essentially a block-level extension of generalized complex orthogonal design[1]. See also [14]. Let us define G as the overall output at the BGOSTBC encoder stacking across B 2K epochs and herein list some properties of it:
* 2NK Mt 1 , k k K B A k B k k × = ⊗ + ⊗ ∈
∑
G ΧΧΧΧ c ΧΧΧΧ d (2.1) t 2K M where (0) (1) ( 1) . ( ) ((- ) ), 0 1. , , 1 . , 1 . , 1 , 1 . , 1 , k k t k l l k t k l l k k l t t T k k k k k k N A B M T A A T A A M T B B T B B T A B M M d d d N d n c n n N k K k l K k l K k l K k l K × × = − = ≤ ≤ − ∈ ≤ ≤ ≤ = ≤ = − ≤ ≠ ≤ ≤ = ≤ = − ≤ ≠ ≤ = d I I 0 Χ Χ ΧΧ ΧΧ Χ Χ Χ Χ Χ ΧΧ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ ΧΧ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ ΧΧ Χ Χ Χ 1 k l, K. ≤ ≤ Specificallyand k k k k t K K K A A B A K K K K M × × × = = 0 I X X I 0 0 Χ ΧΧ Χ G , (2.2) where Mt K k A × ∈
G are generic constituent matrices of certain OSTBC design of dimensions depending of choices of K and Mt. See also Appendix A3 and [1] for a detailed description about the construction of constituent matrices of OSTBC. The time-reversal, d nk( )=ck((- ) )n N , is carried out via the modulo-N operation defined as (- )n N =N− for 1n ≤ ≤n N− and (- )1 n N = for n = 0. Note that 0 Χ ΧΧ ΧΧ ΧΧ ΧAk, Bk ∈
t
2K M×
, 1≤ ≤k K , are constituent matrices of a certain GOSTBC determined by choice of (K,Mt). ΧΧΧΧAk and ΧΧΧΧ are non-overlapping matrices consisting of only ones Bk and zeros up to sign changes, and are responsible for designating space-time ordering of the forward transmission mode and the reversed one, respectively, i.e., ( )p
m s takes ck if k A pm
ΧΧΧΧ is one, for 1≤ ≤p K . Similarly, s( )mp takes * k
d if ΧΧΧΧBkpm is one, for
1 2
K + ≤ ≤p K .
2.1.2 BGOSTBC Encoder (with CP Insertions)
Note that in Fig. 2, (K,Mt) = (4,3) and CP insertions following the encoding have been taken into account. The mth
encoder output path is followed by a CP insertion whose output is connected to the mth
transmit antenna, for all values of m. CP is inserted prior to the transmission of each symbol block for mitigating the channel distortion. From here on each time epoch lasts for N+L symbol periods. Since the constituent matrices simply serve as designating the space-time ordering for signal blocks, CP insertions directly apply to ( )p, 1 2
m ≤ ≤p K
s , 1≤m≤Mt. Let M=N+L. Then the signal blocks collected at the output of the CP insertions over 2K time epochs and across Mt transmit antennas can be represented in matrix from as
(
)
(
)
(
)
2 CP * CP CP 1 = k k B K B K A k B k k= ⊗ ⊗ + ⊗ ∑
G I I G I c I d Χ Χ Χ Χ Χ Χ Χ Χ * 2MK Mt 1 = k k , K A k B k k × = ⊗ + ⊗ ∈ ∑
ΧΧΧΧ c ΧΧΧΧ d (2.3) whereM CP * * M CP ( ) M N CP ; ;
, the CP insertion matrix.
k k k k L N L L N × − × ∈ ∈ ∈ c I c d I d 0 I I I
As a building block, next we review the transmission of a single signal block with CP through a frequency-selective fading channel.
2.2 Block Transmission and IBI-free Model upon CP Removal
Let the sequence sm(pM), sm(pM +1), , sm(pM +M −1)
denotes the symbols
transmitted from mth
antenna during pth
time epoch. Let ( )p
[
( ) ( 1) m sm pM sm pM + s M ( 1)T m s pM +M − ∈ . By noticing ( ) ( ) CP p p m = m s I s , we know that ( )p m s takes ck if k A pm Χ ΧΧΧ is one, for 1≤ ≤p K . Similarly, s( )mp takes * k d if ΧΧΧΧBkpm is one, for 1 K + ≤ p≤2K.
2.2.1 MIMO Frequency-selective Fading Channel and Noise Model
We consider a frequency-selective channel where the channel impulse response of L+1 taps between mth transmit antenna and jth receive antenna is defined as
jm h L+1 (0) (1) ( )T jm jm jm h h h L ∈
. Throughout the thesis, we assume independence between channel taps for all hjm, 1≤ ≤j Mr, 1≤m≤Mt . Assume that hjm is circular symmetric Gaussian distributed, i.e., 2
1
( , )
jm σh L+
h ∼ CN 0 I . Hence the real and imaginary parts of each of the entries of hjm are i.i.d. zero-mean Gaussian with variance 0.5 2
h
σ each, i.e., (0, 0.5 )2 h σ
N . Assume that the channel remains fixed during 2K time epochs. Let the sequence ηj(pM), (ηj pM +1),, (ηj pM +M −1) denotes the additive noise samples, circular symmetric Gaussian distributed, at the jth receiver during the pth
time epoch. Hence, the noise vector defined as ( )p ( ) j ηj pM η M ( 1) ( 1)T j pM j pM M η + η + − ∈ assumes ( , 2 ) w M σ 0 I CN .
Let the sequence v pMj( ), (v pMj +1),, (v pMj +M −1) denote the received symbols at the jth receiver during the pth time epoch. Define ( )p ( ) ( 1)
j =v pM v pMj j + v M ( 1)T j v pM +M − ∈ . For 0≤ ≤n M − , 11 ≤ ≤j Mr, 1 0 ( ) ( ) ( ) ( ). t M L m j jm j m l v pM n h l s pM n l η pM n = = +
∑∑
+ − + + We can write ( )p j v as ( ) ( ) ( 1) ( ) 1 t M p tr p IBI p p m m j jm jm j m − = =∑
+ + v H s H s η , where (0) 0 0 0 (1) (0) 0 0 (1) ; ( ) 0 (0) 0 0 0 ( ) (1) (0) 0 ( ) (2) (1) 0 0 ( ) (2) 0 0 0 0 0 ( ) 0 0 0 jm jm jm jm tr jm jm jm jm jm jm jm jm jm jm jm IBI jm jm h h h h h L h h L h h h L h h h L h h L H H . 0 0 0 After CP removal, we have the IBI-free received signal model, ( )p ( )p j N L× N j y 0 I v = N ( ) ( 1) ( 1)T j j j y pN y pN y pN N + + − ∈ , where y ij( ), pN ≤ ≤i pN +N − , 1
are the received symbols at the jth receiver during the pth time epoch after CP removals.
(
)
( ) ( ) ( ) ( 1) 1 ( ) ( 1) ( 1) t p j j j j M tr p IBI p p m m N L N jm N L N jm N L N j m v pM L v pM L v pM M − × × × = = + + + + − =∑
+ + y 0 I H s 0 I H s 0 I η( ) ( ) ( ) 1 ( ) ( ) 1 ( ) (0) 0 0 0 0 0 0 ( ) 0 0 0 0 0 0 ( ) (0) = , t t jm jm M jm L N L L p p m j m N jm jm M p p jm m j m h L h h L h L h × − = = = + +
∑
∑
0 I s w I H s w where ( )p ( )p j N L× N j w 0 I η and (0) 0 0 0 ( ) (1) (1) (0) 0 (0) 0 ( ) . ( ) ( 1) (0) 0 ( ) 0 0 0 0 ( ) ( 1) (0) jm jm jm jm jm jm jm jm jm jm jm jm jm jm jm h h L h h h h h L h L h L h h L h L h L h − − H Let ( ) N 1 1 = T T jm jm × − −N L ∈ h h 0 . We note that Hjm is an N by N circulant matrix with its first column being hjm. It follows that, after CP removal, the signal received at the jth
receive antenna on the pth
time epoch is IBI-free and given by, for 1≤ ≤p 2K ,
( ) ( ) ( ) 1 . t M p p p j jm m j m= =
∑
+ y H s w (2.4)Let us define the following terms which will be used through out the following chapters: t t r t M (L+1) 1 2 M N 1 2 M M (L+1) 1 2 t t r T T T T j j j jM T T T T j j j jM T T T T all M ∈ ∈ ∈ h h h h h h h h h h h h . (2.5)
Chapter 3
Frequency-Domain Equalization
In this chapter, we propose a generalized FDE scheme based on block-level extension of generalized complex orthogonal designs (GOSTBC). The extended block-level OSTBC scheme with more than two transmit antenna is capable of achieving full transmit- receive diversity using FDE over frequency-selective channels. This shows an alternative to the block-level extension resorting to QO-STBC[2], which can indeed achieve perfect FDE in more-than-two transmit antenna scenario but at the cost of additional hardware complexity accounting for adders. Aside from this complexity drawback, its spatial diversity is halved.
3.1 Overall Frequency-Domain Received Signal Model
In the subsequent discussion, we will only derive the received signal model at the jth receive antenna as all receivers have the same signal model except for different channel impulse responses. First recall that ( )p
m
s in (2.4) takes ck if ΧΧΧΧAkpm is one, for 1≤ ≤p K and takes *
k
d if ΧΧΧΧBkpm is one, for K+ ≤ ≤1 p 2K . With the above observations, we have the following received signal model after collecting the IBI-free received signal blocks over 2K time epochs (Without loss of generality, we have assumed the transmission started from time epoch index 1 and collect the received blocks all the way up to index 2K)
( ) ( )
( ) (
)
(
)
(
)
( ) ( )
( ) (
)
(
)
' (1) (2) ( ) ( 1) (2 ) (1) (2) ( ) ( 1) (2 ) ' 2 1 t T T T K T K T K T j j j j j j M T T T T T T K K K K jm m m m m m j m + + = = ⊗ + ∑
y y y y y y I H s s s s s w(
)
' 2 1 (:, ) , t M K jm B j m m = =∑
I ⊗H G +w (3.1) where '( ) ( )
(1) T (2) T(
(2 )K)
T T j = j j j time epochs.
3.1.1 Space-Time Combining Using Signal DFT’s
Define ( )p ( )p, 1 2 j yj ≤ ≤p K Y F , '(
)
' 2 j IK ⊗ yj Y F and '(
)
' 2 j IK ⊗ wj W F . Soperforming block-wise DFT on y in (3.1) leads to 'j
( ) ( )
(
) (
)
(
)
(
)
(
)
(
)
' (1) (2) ( ) ( 1) (2 ) ' 2 2 2 1 (:, ) t T T T K T K T K T j j j j j j M K K jm B K j m m + = = =∑
I ⊗ I ⊗H G + I ⊗ w Y Y Y Y Y Y F F(
)
(
)
' 2 1 (:, ) . t M K jm B j m m = =∑
I ⊗ FH G +W (3.2) See the tailored version for 2-fold diversity in [9],[26]. Since Hjm is circulant, we candecompose it by the DFT matrix as follows: H ,
(
)
,jm jm = Λjm Λ =jm Diag N
H F F Fh
for all values of j and m. Let Xk Fck. Since d is obtained from k c by conjugated k time reversal, we have[24, pp. 123-124], * *
k k =
d X
F , 1≤ ≤k K . Hence we can rewrite (3.2) as
(
)
(
)
(
)
(
)
(
)
(
)
(
)
{
}
(
)
(
)
{
}
' ' 2 1 * ' 2 1 1 * ' 1 1 ' * 1 1 (:, ) (:, ) (:, ) (:, ) (:, ) (:, ) (:, ) . t t k k t k k t k k M j K jm B j m M K K jm A k B k j k m M K A jm k B jm k j k m M K k k A jm B jm j k m m m m m m m m = = = = = = = = ⊗ Λ + = ⊗ Λ ⊗ + ⊗ + = ⊗ Λ + ⊗ Λ + = ⊗ Λ + ⊗ Λ +∑
∑∑
∑∑
∑∑
I G I c d c d Χ Χ ΧΧ ΧΧ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Χ Y F W F W F F W W X X Based on (2.2), we have(
)
(
)
' 1 ' * 1 1 1 (:, ) . (:, ) t k k M K k A jm NK j j k m NK A jm k m m × = = × ⊗ Λ = + ⊗ Λ + ∑∑
G 0 0 Y W G X X (3.3)(
)
(
1 *)
1 1 1 (:, ) (:, ) t k k M K k A jm NK j j k m NK A jm k m m × = = × ⊗ Λ = + + ⊗ Λ ∑ ∑
G 0 0 Y W G X X 1 * 1 1 1 (:, ) . (:, ) t k k M K A K k jm jm j k m K A m m × = = × =∑ ∑
⊗ Λ + ⊗ Λ + 0 0 X G W G (3.4)Note that the definition of Yj follows from the reversed transmission mode by virtue of
GOSTBC. Let
(
)
(
1)
1 (:, ) * (:, ) 1 t Ak K K Ak M m jk jm m jm m × × = ⊗ Λ + ⊗ Λ ∑
0 0 O G G , 1 2 T T T T K X X X X and 2KN KN 1 2 j j j jK × Λ O O O ∈ . We can then rewrite (3.4) more compactly as, for 1≤ ≤j Mr,
j = Λj + j
Y X W. (3.5)
Equation (3.5) represents the frequency-domain input-output relation between the jth receive antenna and the transmit antennas, where Λ Xj, and Wj are embedded with knowledge of channel state information (CSI), of transmitted signal blocks and of channel noise at the jth
receive antenna, respectively. Let 1 2 r
T T T T M = Y Y Y Y r 2KNM
∈ . By stacking the received signals in (3.5) across Mr receive antennas, we arrive at the following MIMO channel model in frequency domain
= Λ + Y X W, (3.6) where 1T 2T Tr T M = W W W W and 2KNMr KN 1 2 r T T T T M × ΛΛ Λ Λ ∈ is the equivalent frequency-domain MIMO channel matrix for block transmission.
3.1.2 Frequency-Domain Equalization with Perfect CSI
Let X= Λ+Y and W = Λ+W and assume that rank
( )
Λ =KN . Hence Λ Λ is H invertible, and the Moore-Penrose pseudo-inverse of Λ can be written as Λ = +(
Λ ΛH)
−1Λ ∈ H KN 2KNM× r . So by pre-multiplying both sides of (3.6) by the linearX=X+ W, (3.7)
where X contains the decoded DFT’s of the transmitted signal blocks. Therefore the transmitted blocks can be recovered through FDE. It is seen from (3.7) that by nonsingularity of Λ Λ , FDE can be achieved up to a zero-mean equalization error term H
W when perfect CSI is available. The proposed BGOSTBC achieves block-level orthogonality in frequency domain, that is, Λ Λ is a diagonal matrix and diagonally H loaded with diversity weightings. More precisely, we have the following results.
Theorem 3.1 1 1 2 r t M M H H K jm jm j= m= Λ Λ = ⊗ Λ Λ I
∑∑
.(Proof: See Appendix A1.)
Remark 3.1 With perfect CSI, the FDE can be carried out efficiently. By efficiency we mean the involved computation for matrix inverse
(
Λ ΛH)
−1reduces to merely computing the reciprocals of each of the diagonal entries of Λ Λ . This H can be seen from Theorem 3.1 that Λ Λ is diagonal. H
Remark 3.2 The double summation of the elementary square block of Λ Λ H implicitly indicates that the full MrMt-fold spatial diversity is achieved through the proposed FDE scheme (See also chapter 5 for an explicit validation of full diversity via PEP analysis). Notice that had the encoder been realized with QO-STBC, the system would enjoy halved spatial diversity by carefully decoupling the received signals upon DFT transform[2]. The QO-STBC is more redundant than OSTBC in nature by inspection and relies on additional arithmetic operations for exploiting quasi-orthogonality at the receiving end, with approximately full transmission rate though.
Remark 3.3 Equation (3.3) is hereby regarded as a building block as we will exploit more of it towards the discussion of training-based channel estimation in chapter 5. Such discussion is as practical as essential to applying the OSTBC to the frequency-selective fading environment. Our proposed scheme can be adopted to
scenario in a quite straightforward manner, in contrast to [16], with similar conclusion on optimal training design obtained in [3].
Remark 3.4 By adopting the Alamouti scheme for the construction of G with B proper changes in dimension and scalar scaling factor makes the 2-fold diversity FDE (Alamouti-like) scheme a special case.
Remark 3.5 Notice that the FDE behind (3.7) is based on the premise that
( )
rank Λ =KN , and hence
(
H)
rank Λ Λ =KN . This in turn requires that
1 1 r t M M H jm jm j m rank N = = Λ Λ =
∑ ∑
be satisfied as we can see from Remark 3.1. Notice that each of the diagonal entries of Λ Λ is nonnegative. To proceed further, let Hjm jmus define ( )0 ( )0 ( 1)T N jm jm = hjm hjm hjm N − ∈ h h F , where h jm( )n
is the nth entry of the DFT, 0≤ ≤n N− , 11
r
j M
≤ ≤ and 1≤m≤Mt.
Theorem 3.2 If there exist at least one pair of values
(
j m*, *)
, 1≤ j* ≤Mr, *1≤m ≤Mt, such that * *( ) 2
0 j m
h n > , 0≤ ≤n N − , then 1 rank
( )
Λ =KN holds.
Let us denote the sufficient condition in Theorem 3.2 as (c1.R). That is, if there is a subchannel whose transfer function is free from zeros at the frequencies
2 n N
j
e π , 0≤ ≤n N − , then 1 Λ Λ is nonsingular. The condition is very weak and is H
generically satisfied. It is interesting to consider the following special case. Suppose
( ) * * 2 0 j m h n > , for 0≤ ≤n N− while 1 * 1, for 1 jm = N× ≤ ≠j j ≤Mr h 0 , 1 ≤ * t
m ≠m ≤M . Consequently equation (3.4) becomes
{
}
1 K k j jk j k= =∑
O + Y X W with(
*)
(
1)
* * * * * 1 (:, ) (:, ) Ak K K Ak m H jk j m m j m × × ⊗ Λ + ⊗ Λ 0 0 O G G .By stacking across all Mr receive antennas, we have Λ with the block-level orthogonality that 2
(
* * * *)
H H
K j m j m
Λ Λ = I ⊗ Λ Λ , which clearly lacks of spatial diversity, compared to (3.6) whose subchannels are all active. Intuitively, the decoded signal X = Λ Y+ becomes much more erroneous without the spatial
diversity than with any. The above scenario arises when our MIMO channel model degenerates to a single-input-single-output (SISO) frequency-domain equivalent. So far, even though we consider a scenario which is more likely than (c1.R) to occur, such situation is far from common, not to mention that only one pair of values
(
*, *)
, 1 * and 1 * r t j m ≤ j ≤M ≤m ≤M such that * *( ) 2 0 j m h n > , 0≤ ≤n N − . 1 This is why we call (c1.R) a weak condition as it can be easily satisfied in practice.Notice that the reason why we are able to recover the transmitted symbols even under the above mentioned harsh situations lies in that the space-time redundancy transmitted a signal block through every transmit antennas over different time epochs. Even though the MIMO channels degenerates to an SISO channel, all the transmitted blocks can still manage to the single receiver, provided that the reception lasts for 2K time epochs.
Chapter 4
Pairwise Error Probability Analysis
In this chapter, we will derive upper bounds for the average pairwise error probability (PEP) for the FDE-based scheme. The essential assumption under which the derivation is justified would be stated with numerical support. Our perspective reveals that at high SNR, the proposed system dose have the potential of delivering maximum possible spatial as well as multipath diversity.
4.1 PEP Analysis for Suboptimal Detection Problem
First, we derive an upper bound for the average PEP, assuming that the decoding consists of a linear equalization followed by an symbol-wise quantization into the signal constellation A . See also [11] and [17]. To see that, let us formulate the detection problem as follows. Based on (3.6), we have Y=Ωc+W with the matrix Ω
(
)
2KNMr KNK
×
Λ I ⊗F ∈ being responsible for yielding the frequency-domain output of
the time-domain input signal blocks KN
1 2 T T T T K ∈ c c c c . Hence we can address the ML-detection problem as follows:
KN 2 argmin ML ∈ = − c c c A Y Ω , (4.1)
where cML is the optimal decoded symbol block in ML sense. However, this ML exhaustive search yields infeasible computational cost of order AKN for practical values of K and N, where A denotes size of the employed constellation. So we seek a suboptimal linear equalization approach. Note that W is a white Gaussian vector since DFT serves as a unitary operation only. Let us define Z ΛHY∈KN
, ΛH ∈KN
W W and ΘΛHΩ∈KN KN× . Then we can rewrite (3.6) as
We assume rank
( )
Λ =KN . Since IK ⊗F is of full rank, rank( )Θ =KN by definition of Θ . Hence the inverse Θ−1 exists and can be used as a linear equalizer. Let Z =Θ−1Z denote the output of the linear equalizer Θ−1, then from (4.2) we readily arrive atZ = +c Θ−1W. (4.3)
Notice that the diversity gain weighting has been normalized at the output of Θ−1 and hence the detection problem can be formulated with respect to the employed constellation. Based on signal model of (4.3), we obtain an suboptimal detection problem: KN 2 argmin , ∈ = − c c c A Z (4.4)
where c is an estimate of c after equalization. Note that the so obtained estimate minimizing the metric in (4.4) is suboptimal since the underlying noise is no longer white and its covariance depends on the matrix Θ−1ΛH. We see that minimizing the metric in (4.4) amounts to a symbol-wise hard-decision into A since
2 − c
Z is the sum of KN nonnegative terms and each of which can be minimized with respect to an entry of c , i.e., minimization on symbol-level. Therefore, using Θ−1 as an equalizer and (4.4) as the detector amounts to having a linear equalization scheme followed by a hard decision on each entry of Z into the constellation A . It is noteworthy that the computational cost of (4.4) is linear in composite block length KN, which is computationally cheaper than that of (4.1).
The PEP analysis considers the probability that a symbol block c∈AKN is transmitted while another c is detected in the minimum-distance perspective. Given the channel realization h , and hence the matrix Λ , the conditional PEP is defined as all
[
]
2 2
Pr c→ c hall =Prc Z- < c-Z Λ. (4.5)
Note that proving that the average PEP has maximum diversity is equivalent to proving that the error rate performance exhibits maximum diversity, by virtue of the union
the Euclidean distance between c and c , as well as the corresponding normalized difference, respectively. Let Re
{
(
1)
H}
ξ Θ− W e [18, pp. 508], then (4.5) becomes
[
]
Pr Pr . 2 all ξ → = > Λ c c h D (4.6)As the variable ξ , when conditioned on channel realizations, is a zero-mean Gaussian random variable, the error probability in (4.6) is completely determined by the variance
2 ξ
σ . To compute 2 ξ
σ , first let us write W =
(
I2KMr ⊗F w)
for notational purpose, where w is related to the white noise 'j w in (3.1) by '(0 : 1)T T T j KN =w − w w , ( )* ' mod : 2 1 j KN KN −N = −
w w . Then we can factorize
(
Θ−1W)
H e as H e w , where(
)
(
1)
2 r H H KM −Λ ⊗ e I eΘ F . The variance is computed as follows:
{
}
{
(
)
}
{
}
{
}
{ }{ }
{
}
{ }{ }
{
}
{ }{ }
(
)
(
{ }{ }
)
{
}
2 2 2 1 2 2 2 Re Re Re Re Im Im 2 Re Re Im Im . H H T T T T E E E E E E ξ σ = ξ Λ = − Λ = Λ = Λ + Λ + Λ e e e e e e W Θ w w w w wNotice that since w is white, the entries of w are i.i.d. circular symmetric Gaussian, with real part and imaginary part of each of the entries being (0, 0.5 2)
w
σ
N distributed.
Also notice that Re{ }w T Re
{ }
e and Im{ }T Im{ }
ew are statistically independent of each other by virtue of the circular nature of w . For notational purpose, let [ ]yi denotes the ith
entry of a vector y . So we can rewrite the above equation as
{ }
{ }
{
}
{
{ }{ }
}
{ }[
]
{ }
[
{ }]
{ }
{ }
{
[
{ }]
}
{ }
{
[
{ }]
}
2 2 2 2 2 2 2 1 1 2 2 2 2 2 2 1 1 Re Re Im Im Re Re Im Im Re Re Im Im r r r r T T KNM KNM i i i i i i KNM KNM i i i i i i E E E E E ξ σ = = = = = Λ + Λ = + Λ = Λ + Λ∑
∑
∑
∑
e e e e e e w w w w w w
{ }
{ }
(
)
(
)
(
)
2 2 2 2 2 1 1 2 2 2 2 2 1 2 2 2 2 2 2 1 Re Im 2 1 2 1 2 1 , 2 r r r r KNM KNM w i i i i w H H w KM H H w K KM σ σ σ σ = = − + = + = = Λ ⊗ = ⊗ Λ ⊗ ∑
e∑
e e I e I I e F F F Θwhere we have used independence between each of the entries of Re w due to w { } being white and zero-mean. Notice that 1 H
(
H)
K
−Λ = I ⊗F Λ+
Θ . Similar reasoning
applies to Im w{ }. We know that the PEP in (4.6) can be expressed in terms of the
Q-function, ( ) 1 22 2 t Q πe dt α α
∫
∞ − , as[
]
(
)
(
)
(
)
(
)
2 2 2 2 2 2 2 2 Pr 4 2 , 2 r r all H H w K KM H w K KM F Q Q Q ξ σ σ σ + + → = = ⊗ Λ ⊗ ≤ ⊗ Λ ⊗ c c h I I e I I F F F F D D Dwhere we have invoked the inequality that Ae 2 ≤ AF e 2 = AFfor any matrix A
and a unit vector e. Since multiplication by unitary matrix dose not change Frobenius norm, we arrive at
[
]
2 2 Pr . 2 all w F Q σ + → ≤ Λ c c h D Then by using ( ) 1 2 2exp( 2)Q α ≤ −α , we have the PEP of interest upper bounded as
2
Let us define the following terms:
(
)
(
)
(
)
( ) 1 ( ) 1 (1) (2) ( ) (1) (2) ( ) 1 1 1 2 1 2 2 2 2 1 (:, ) ; ; (:, ) ; ; ; k k t t t T m A jk N K T m K jk N A M jk jk jk jk M jk jk jk jk T H H H H H H j K j K j K jM j K j m m × × − − − ⊗ ⊗ ⊗ Λ Λ Λ ⊗ Λ Λ Λ ⊗ Λ Λ Λ ⊗ Λ Λ A I 0 0 A I A A A A A A A A I I I I G G ϒ ϒ(
)
1(
)
1(
)
1 2 2 2 t . T H H H K j K jM − − − Λ ⊗ Λ Λ Λ ⊗ Λ Λ Λ I I (4.8)From (4.8), Λ+
(
Λ ΛH)
−1ΛH can be rewritten as+ , Λ = A A ϒ ϒ (4.9) 11 1 11 1 1 1 1 1 where ; ; ; . r r r t r t r r M M K M M K M M M M A A A A A A A A A A ϒ ϒ ϒ ϒ ϒ ϒ Hence, 2 2 2 2 F F F F + Λ ≤ +
A A ϒ ϒ . Notice that A A is solely determined by
the encoder structure, and by (4.8) it can be easily verified that 2 2
F = F ϒ ϒ and 2 1 2 1 1 1 1 2 . r t r t M M M M H H jm jm jm F j m j m F K − = = = = = ⋅ Λ Λ Λ
∑∑ ∑∑
ϒ (4.10)With the definition of hjm
in Remark 3.5, (4.10) can be rewritten as
( ) ( ) 2 1 2 2 2 1 1 0 1 1 2 r t r t M M N jm F M M j m n jm j m h n K N h n − = = = = = = ⋅
∑∑ ∑
∑∑
ϒ ( ) 1 1 2 0 1 1 2 . r t M M N jm n j m K h n N − − = = = = ⋅ ∑ ∑∑
(4.11)Now, let us define an N by N matrix D as ps
( ) ( ) ( ) 2 1 1 2 1 1 2 1 1 0 1 1 r t r t r t M M jm j m M M jm j m ps M M jm j m h h h N = = = = = = = −
∑∑
∑∑
∑∑
D . (4.12)We assume rank
( )
Λ =KN, or equivalently1 1 r t M M H jm jm j m rank N = = Λ Λ =
∑ ∑
. Hence, we see that Dps is invertible. Hence, by defining2 E F C A A and from (4.9) up to (4.12) we arrive at 2 2 4 1 E ps F F K C N + − Λ ≤ ⋅ ⋅ D . (4.13)
Before we proceed further, let us make one more assumption deduced from the rank premise: condition number of Dps, denoted as K
( )
Dps , is upper bounded by a finite real number Ku ∈ for all possible channel realizations, i.e.,( )
Dps ≤ u for all Dps.( )
max 1 1 1 1 min 1 1 1 1 max min r t r t r t r t M M M M H H jm jm jm jm j m j m M M M M H H jm jm jm jm j m j m Diag Diag λ λ = = = = = = = = Λ Λ Λ Λ Λ = = Λ Λ Λ Λ ∑∑
∑∑
∑∑
∑∑
K ,where λmax
( )
Λ and λmin( )
Λ denote the maximal and minimal eigenvalues of matrix Λ respectively. Hence K( )
Dps =K( )
Λ . By regarding the simulation results from Fig. 3-1. to Fig. 3-5, we see thati) For all possible system configurations (N,Mr,Mt,L,K), the probability of singular occurrence rank
( )
Λ <KN is very low in practice. As one can perform as many simulations as possible to find the above highly plausible, our rank premise is reasonable. Hence K( )
Dps < ∞.ii) Given known channel order L and desired transmission rate, selecting high values for Mt and Mr can effectively suppress K
( )
Dps .iii) It agrees with the intuition that as N increases, the probability of K
( )
Dps → ∞ drops. This is because each of the diagonal entries, which are correlated, of Dps is sum of nonnegative terms.From the above three observations we deduce that K
( )
Dps can be universally upper bounded by a finite real number in practice.On the other hand, suppose B∈n n× is a nonsingular diagonal matrix, then for any full rank A∈n m× , m<n, such that
( ) ( ) 2 1 m F ≤ n +n B A K , 1 2 ( ) 2 F F + − ≤ B BA . Proof: [ ] [ ]
(
( ))
( ) 1 1 2 2 2 2 1 0 0 2 2 2 2 1 . n n ii ii i i F F F F F F n n m − − − + − = = ⋅ + =∑
∑
≤ ≤ ≤ B B B B BA B B B A K (4.15)Therefore, by assuming (4.14) and along with (4.15), there exist infinitely many full rank
{ }
Ψ ∞ ∈ N N L× −( ) satisfying Ψ 2 ≤ m such that D−1 2 ≤(
D Ψ)
+ 2 forall channel realizations. Now let us find such a matrix Ψ among these possible candidates so that any (N-L) of the rows of Ψ are linearly independent. Note that this additional constraint can be very easily satisfied and hence does not in any way contradict (4.14). Then, (4.13) can be written as
(
)
2 2 4 E ps F F K C N + Λ ≤ ⋅ ⋅ D + Ψ . (4.16)Let β
{
n n0, ,1,nL−1}
be the set of indexes corresponding to the smallest L composite frequency responses such that( )
( )
( )
( )
(
)
2 2 1 1 1 1 2 2 2 0 1 1 1 1 1 1 1 1 for and , r t r t r t r t r t M M M M jm p jm q p q j m j m M M M M M M jm jm jm L j m j m j m h n h n n n h n h n h n β β = = = = − = = = = = = ≤ ∈ ∈ ≤ ≤ ≤ ∑∑
∑∑
∑∑
∑∑
∑∑
(4.17)where 0≤ni ≤N−1 and 0≤ ≤ − Define i L 1. Dβ as an N by N diagonal matrix whose entries are given by
( ) 2 1 1 1 , if . 0 , if r t nn M M jm j m nn n h n n β β β β = = = ∈ = ∈
∑∑
D D (4.18)Then we construct an (N-L) by N matrix Ψβ by the following two-fold step:
i) Remove all the rows, whose row indexes belong to set β , of Ψ to obtain an (N-L) by (N-L) matrix Ψ .
ii) Insert L zero columns to Ψ−1 so that after the insertion, the inserted columns have indexes belong to set β . Let Ψβ denote the so obtained matrix. Note that the constraint on linear independence of any (N-L) of rows of Ψ ensures existence of Ψ−1.