#PAGE_PARAMS# #ADS_HEAD_SCRIPTS# #MICRODATA#

New space-time block codes from spectral norm


Authors: Carlos A. R. Martins aff001;  Mauro Luiz Brandão, Jr. aff002;  Eduardo Brandani da Silva aff003
Authors place of work: Department of Mathematics, UTFPR, Pato Branco, PR, Brazil aff001;  Department of Electrical Engeneering, State University of Maringá, Maringá, PR 87020-900, Brazil aff002;  DMA - UEM, Avenida Colombo 5790 - Campus Universitário, 87020-900-Maringá-PR, Brazil aff003
Published in the journal: PLoS ONE 14(9)
Category: Research Article
doi: https://doi.org/10.1371/journal.pone.0222708

Summary

Current research proposes a natural environment for space-time codes and a new design criterion is obtained for space-time block codes in multi-antenna communication channels. The objective of this criterion is to minimize the pairwise error probability of the maximum likelihood decoder, endowed with the matrix spectral norm. The random matrix theory is used and an approximation function for the probability density function for the largest eigenvalue of a Wishart Matrix is obtained.

Keywords:

Probability theory – Eigenvalues – Graphs – Information theory – Block codes – Antennas – Signal decoders – Probability density

1 Introduction

A consistent theory for communication systems was introduced by Shannon’s [1] classical work, where it was proved that, for any communication channel with capacity C, the data transmission below C may be done efficiently by using an appropriate error-correcting code, or rather, there exists a code such that the error probability may become as small as required. The results of Shannon’s work are valid for channels with only one transmitter antenna and only one receiver antenna, called Single Input and Single Output (SISO) channel. However, during the last decade, we have witnessed a huge demand expansion in telecommunications, and the available technologies will not be sufficient in the future. This is due to the growing need to develop reliable communication systems that allow high rates of data transmission and, consequently, to the study and development of new mathematical methods and structures that provide support for new technologies, mainly wireless communication systems.

Wireless transmissions may be impaired by a number of factors, such as great distances, objects between wireless devices, and other wireless networks, which restrict communication speed and reliability. Low capacities and high error rates of fading channels, plus the growing demand for wireless devices, motivated the development of several techniques to overcome these disadvantages. The works of Foschini-Gans [2] and Telatar [3] proved that the multiple-input and multiple-output (MIMO) communication systems attained capacity of data transmission greater than SISO systems. These results proved the superiority of MIMO systems and called the attention of researchers, originating several real applications. The foundations of space-time codes were established by Tarokh, Seshadri and Calderbank in [4] and space-time trellis codes (STTC) were introduced.

A special technique which explores spatial and time diversity of MIMO channels, called space-time block codes (STBC), was first studied by Alamouti [5], and instantly became a topic of great importance in digital communications. Tarokh, Jafarkhani and Calderbank extended Alamouti’s results in [6]. These works triggered researches on criteria to design good space-time codes for MIMO channels. After the publication of these works, the constant development of the space-time block code theory gave birth to new families of codes, that can be confronted using extensive information theory and construction parameters, such as the pairwise error probability (PEP), the code rate and the diversity order. Based on the above, several design criteria have been previously proposed to design space-time block codes for MIMO channels. The rank and determinant criterion [4] is a criterion for asymptotic SNR used to design full-rate and full-diversity STBCs. For low SNRs, the trace criterion [7], also called Euclidian-distance criterion, can be used to design STBC with low pairwise error probability.

The language of matrices to model the communication systems with multiple antennas, is the natural way [8]. Random matrices are particularly a powerful tool. The theory of random matrices emerged in the work of the mathematical statistician John Wishart in [9], but it gained great visibility in the 1950s with the contributions of the physical mathematician Eugene Paul Wigner, with publications [10], [11] and [12], on the spread of resonance of particles with heavy nucleicores in slow nuclear reactions. Further, the physical mathematician Freeman Dyson formalized the theory in [13], [14] and [15]. The theory of random matrices is used in several areas and problems, such as Riemann hypothesis, stochastic differential equations, statistic physics, chaotic systems, numerical linear algebra, neural networks, information theory, signal processing and in the study of the capacity of data transmission in MIMO channels. The deep mathematical results may be perceived in [16], [17], [18] and [19]. For applications, see [20] and [21]. The study of probability density and distribution functions of eigenvalues is one of the main problems in the random matrices theory. It was attacked by von Neumann, Birkhoff, Smale, Demmel and others. The pdf of the eigenvalues of a Wishart matrix was established in 1939, in [22]. Many researchers studied this issue, for instance, [23], [24], [25] and [26]. Estimations for the largest and smallest eigenvalues are given in [27], [28], [29] and [30]. More recently, [31] and [32] have been published. These results have many applications, and they were used to study the channel capacity of MIMO channels, for examples [21], [3], [33] and [34].

Current research proposes a new approach on STBC. We assume that the space-time block codes are elements of an appropriate normed space of matrices endowed with the spectral norm as the intrinsic norm of this space. From this norm, a new criterion for the design of STBC is proposed. The maximum likelihood decoder is endowed with this norm, and several results from the theory of Random Matrices are used to obtain this criterion. As far as we know, this is the first time that random matrix theory is used to obtain a design criterion for STBC. The usual criteria and models used, even in a very recent work such as [35], neither they assume a natural environment for the space-time block codes, nor do they use the spectral norm.

We may also remark that, since its origin, MIMO has seen deep technological advances. Initially, there was the Point-to-Point MIMO [2], [3], [36], and, subsequently, the more efficient Multi-User Mimo [37], [38], [39], [40]. The main drawback of these technologies is that they are not scalable. Current State-of-Art is the Massive MIMO [41], [42], [43], [44], [45], which is a scalable MIMO technology. There are several theoretical and practical questions about Massive MIMO. A new criterion which allows matrices that cannot be used by other criteria may be a useful tool in this new context. Further, the proposed criterion can be used in all SNR regimes.

The work is organized as follows. Section 2 describes the basis on MIMO and gives the usual design criteria to project STBC. Section 3 introduces the spectral norm and its main properties. We propose a natural environment where STBC inhabits. The maximum likelihood decoder is endowed with the spectral norm. Section 4 exposes the Random Matrices theory, with a focus on the cumulative distribution of eigenvalues of Wishart Matrices. In this section, we obtain an approximation function to the probability density function for the largest eigenvalue of a Wishart Matrix. This approximation will be used to provide a new design criterion to project STBC. In Section 5, we introduce the new criterion to design STBC, called Largest Eingenvalue Criterion, and we give a bound related to it. Section 6 presents a performance analysis between the proposed criterion and the known ones provided for several known STBC codes. We present new examples of codes and family of codes. In Section 7 we apply the new criterion for codes whose matrices are made from blocks.

2 System model and notations

This section fowards a short review of space-time block codes. For more details, see [46]. Matrices are represented by capital letters and vectors by bold lower cases.

Consider a constellation S ⊂ C. A space-time block code (STBC), or simply a code, is defined as a subset of matrices C ∈ S n T × n S = { { s i j } n T × n S : s i j ∈ S }, where the natural numbers nS and nT are the number of time slots and the number of transmit antennas, respectively. Each element of a STBC is called a word. Normally a STBC is represented with only one matrix

where each entry sij is a function of k symbols x1, …, xk codified by the block. In this representation, the entry sij is transmitted by the antenna i at time j. The rate of a STBC is defined as R = k/nS. STBC is full-rate if R = 1.

If we transmit a codeword C ∈ C of a given STBC, at the receiver we will have the following matrix

where Es is the average power by signal in each transmit antenna, and the entries of matrix N are complex additive white Gaussian noises with zero mean and variance N0/2 per real dimension. The matrix H = { h i j } n R × n T, where nR is the number of receiver antennas, is known as the channel matrix. The entry hij of H is the fading coefficient between the transmit antenna j and the receiver antenna i. We assume the Rayleigh model, where hij has normal distribution with zero mean and variance 1/2 per real dimension.

Suppose that the codeword C was transmitted. The procedure of maximum likelihood decoding is to choose X ^, that minimizes ∥ Y - H X ^ ∥ F, where ∥.∥F is the Frobenius norm. In this case we suppose that the channel state information (CSI) is completely known at the receiver. A decoding error occurs if we choose E ∈ C, such that

for some EC. The pairwise error probability, denoted by P(CE) in this case, is the probability of transmitting C and incorrectly decoding E.

Giving two matrices C and C′, we define the matrix A(C, C′) by A(C, C′) = (CC′)(CC′)*, where C* means the transpose conjugate of matrix C. Suppose that A(C, E) has rank r and non-null eigenvalues λ1, …, λr. From [4], one has

where, the coefficients βij are related with the terms hij of H. See [4, page 748].

One of the main results from [4] is a search criterion for STBC. It is currently known as Rank and Determinant Criterion. The expression in (1) is hard to manipulate. Using some approximations, it may be written in a simpler way and the error probability is given by

Therefore, good STBC for wireless channels, when rnR is small (≤ 4), must be searched to minimize (2). The criterion is given by:

  • maximizing the minimum rank r of A(C, C′), on all pairs of distinct codewords;

  • maximizing the product ∏ i = 1 r λ i of eigenvalues of A(C, C′), between all pairs of distinct codewords.

Another important search criterion for STBC is also obtained from (1), when rnR > 4, established in [7]. Supposing the space-time code operates with reasonable SNR, after some approximations, the authors deduce that

In this case, when rnR is large (≥ 4), the search of STBC must minimize (3). The limiting (3) shows that the error probability is dominated by codewords with minimum sum of eigenvalues of A(C, E), that is, trace(A(C, E)). Thus, the minimum sum of all eigenvalues of A(C, E) between all pairs of distinct codewords must be maximized. This criterion is called Trace Criterion and is given by:

  • the minimum rank r of A(C, C′) over all pairs of distinct codewords so that r nR ≥ 4;

  • maximizing the minimum trace ∑ i = 1 r λ i of A(C, C′) between all pairs of distinct codewords with minimum rank.

3 The matrix space where space-time block codes live in

In connection with the two criteria given, it must be observed that most works on STBC deal with the search of new codes. In [4], [7] and other works, a vector and a matrix are seen as the same object, that is, a matrix is a representation of a vector in a space R n or C n. However, from a mathematical point of view, there exist deep analytical, algebraic and geometric differences when a codeword is seen as a vector c = ( c 1 1 c 1 2 ⋯ c 1 n T c 2 1 c 2 2 ⋯ c 2 n T ⋯ c l 1 c l 2 ⋯ c l n T ) or a matrix C = (cij)ij.

For the two criteria given, the two representations are used freely. The Frobenius norm is very useful, since the value ∥ M ∥ F 2 of a matrix M is the square of Euclidian norm of M, seen as a vector. If we consider a convenient matrix space as a natural environment in which space-time codes, gaussian noise and fading matrices live in, and if this matrix space has enough rich analytic, algebraic and geometric structures, we will have powerful mathematical tools to manipulate the matrices. For instance, determinant, rank and trace, extensively used in space-time codes and MIMO research, are all operators on matrix spaces.

Definition 3.1 Let M=M(m,n,C) be the set of all m × n complex matrices. Under matrix addition and multiplication by complex numbers (scalars), M is a vector space. Together with matrix multiplication, it is a matrix algebra, that is, an associative algebra of matrices. The spectral norm on M is the function ∥.∥2: M → [0, ∞), where, for a given A ∈ M, one has

where λmax(A) and σmax(A) are respectively, the largest eigenvalue and the largest singular value of A. The spectral norm has the following fundamental properties for all matrices A and B in M and all scalar α:

  • i)

    A2 ≥ 0

  • ii)

    A2 = 0 ⇔ A = 0

  • iii)

    αA2 = |α|∥A2

  • iv)

    A + B2 ≤ ∥A2 + ∥B2.

  • v)

    AB2 ≤ ∥A2B2

For a general approach on matrix norms and related results, see [47].

Space M, endowed with the spectral norm, is a Banach algebra. From property (iv), the following useful inequality is obtained,

The equivalent definition can be proved

We will also need the following relation between Frobenius and spectral norms.

Proposition 3.1

  • i)

    For all matrix A ∈ M one has

    where r ≤ min{m, n} is the rank of A.

  • ii)

    If A is n × n, then

Definition 3.2 A space-time block code (STBC) C is a finite subset of M.

Definition 3.2 is very generic. To obtain applicable STBC, subsets of M with good geometric and algebraic properties must be considered. Let C be a space-time block code. When a codeword C ∈ C is sent, the received signal is

As decoding rule, once R is received, our decoder will search the closest codeword E ∈ C of R, that is, E, such that ∥RE2 is the minimum. Since ∥ A ∥ 2 2 = λ m a x ( A A * ) = σ m a x 2 ( A ) and H is known, if E is wrongly chosen, one has

where, in the first inequality, we used Proposition 3.1(i).

4 Random matrices and main results

In last section we deduced that

This implies that we need to find the pdf of the largest eigenvalue of NN*. To obtain this result, we have to use the theory of random matrices.

Random matrices were introduced by Eugene Wigner to model the nuclei of heavy atoms. In his works, Wigner realized that the eigenvalues distribution of a matrix with random Gaussian entries coincided with the statistics of fluctuations of the levels of heavy atoms, experimentally obtained. Thus, the pdf of eigenvalues of Random Matrices became an important object.

The set of all random variables z = x + iy, where x and y are iid N(μ, σ2), is denoted by N ˜ ( μ , σ 2 ). The following matrix sets are fundamental to the following.

Definition 4.1 (i) The complex Gaussian set G ˜ ( m , n ) is the family of all m × n complex random matrices with independent and identically distributed (iid) elements which are N ˜ ( 0 , σ 2 ).

(ii) The complex Wishart set W ˜ ( m , n ) is the family of all m × m complex random matrices, which may be written in the form AA*, where A ∈ G ˜ ( m , n ).

(iii) The Gaussian Unitary Ensemble GUE is the set of all symmetric m × m complex matrices with (iid) elements that are N(0, 1/4) in the upper-triangle and iid elements that are N(0, 1/2) on the diagonal.

Now, considering G ˜ ( m , n ), where the elements in the matrices are N ˜ ( 0 , σ 2 ), one has the following result from [48].

Theorem 4.1 Given M ˜ = A ˜ A * ˜ ∈ W ˜ ( m , n ) , where A ˜ ∈ G ˜ ( m , n ), suppose λ1 ≥ λ2 ≥ ⋯ ≥ λm−1 ≥ λm ≥ 0 are the eigenvalues of M ˜. Then, the joint pdf of the eigenvalues of M ˜ is

where

Now, we want the pdf of the largest eigenvalue of a complex Wishart matrix. Results found in literature, for instance, in [49] and [50], are not easy to manipulate. Thus, an approximation to the pdf will be obtained. We begin with the following bound.

Theorem 4.2 If M ˜ ∈ W ˜ ( m , n ) , then f λ m a x ( λ ) satisfies

Proof. From Theorem 4.1, one has

Thus,
where R2 = {(λ2, λ3, ⋯, λm): λ2 ∈ [0, λ]; λi ∈ [0, λi−1], i ∈ {3, ⋯, m}}. Since 0 ≤ λ − λi ≤ λ, then 0 ≤ (λ − λi)2 ≤ λ2, and this may be bounded above by
where R2 = {(λ2, ⋯, λm): λ2 ∈ [0, ∞]; λi ∈ [0, λi−1], i ∈ {3, ⋯, m}}. From Theorem 4.1, we have
and, substituting (6) in the limiting of f λ m a x ( λ ), one has
Finally, using the expression of K ˜ n , m in Eq (7), one has
which concludes the proof.

Now, with the bound above, we may build our approximation function. Since

normalizing the bound of Theorem 4.2, we define the function
Then, g is a pdf on [0, ∞). Using an algebraic computer program, f λ m a x ( λ ) may be plotted for all cases of m and n. Comparing cases of g with f λ m a x, for the same pair (m, n), it may be seen that a translation of g(λ) is a good approximation to f λ m a x ( λ ). Fig 1 provides an example.

Fig. 1. Graphs of exact pdf f λ m a x ( λ ) and of g(λ) for W ˜ ( 3 , 13 ).
Graphs of exact pdf f λ m a x ( λ ) and of <i>g</i>(λ) for W ˜ ( 3 , 13 ).

Thus, a constant d1 = d1(m, n) must be found, such that the translation of g(λ), given by

is an approximation to f λ m a x ( λ ). Fig 2 shows the graphs of f λ m a x ( λ ), g(λ) and ϕ(λ − 10.4) for W ˜ ( 3 , 13 ).

Fig. 2. Graphs of f λ m a x ( λ ), g(λ) and the translation with d1 = 10.4 for W ˜ ( 3 , 13 ).
Graphs of f λ m a x ( λ ), <i>g</i>(λ) and the translation with <i>d</i><sub>1</sub> = 10.4 for W ˜ ( 3 , 13 ).

Table 1 shows the exact pdf f λ m a x ( λ ) for some cases and the translations of g(λ) which fit better, such that ϕ(λ) is the best approximation. Data were obtained by trial and error to minimize the distance

between f λ m a x ( λ ) and ϕ(λ). Table 1 also presents the maximum point of ϕ(t). For simplicity, in Tables 1 and 2, it is assumed that σ2 = 1. Since ϕ(λ) is a translation of g(λ), they have the same maximum.

Tab. 1. Better translation and its maximum point.
Better translation and its maximum point.
Tab. 2. Maximum points and their differences.
Maximum points and their differences.

It is not possible to find analytically the maximum point of f λ m a x ( λ ). However, the maximum point of g(λ) may be found. Given the maximum point of g(λ), we must determine the constant d1 = d1(m, n). Supposing σ2 = 1, the maximum point of g(λ) is λ0 = 2(n + m − 2). Thus, λ0 + d1(m, n) = 2(n + m − 2) + d1(m, n) must coincide with the maximum point of f λ m a x ( λ ). Let h(m, n) be the maximum point of f λ m a x ( λ ), then

and
Using data from Table 2, an expression to d1(m, n) will be obtained by the least squares method. Plotting the data of Table 2, the function describing d1(m, n) may be seen as a plane and its equation is given by
where
and μi are the data of the third column of Table 2. Thus, we must find the constants a, b and c by minimizing:
Then, we need to solve the equation ∇F(a, b, c) = 0, given by
From Table 2, one has
and the solution is given by

Therefore, the translation of g(λ) is

Putting together the results, one has

Theorem 4.3 An approximation to the pdf of the largest eigenvalue of a Wishart matrix NnR×lNl×nR*, with variance σ2 = N0/2, is the given by the pdf

where d1 = d1(nR, l) = 2.53573nR + 0.574893l − 5.40273.

In what follows in the text, we are considering d1 = d1(nR, l), where nR × l is the dimension of matrix N.

Remark 1. Table 3 shows values of d1(m, n), which may be compared with those in Tables 1 and 2. We have

Then, the total variation is
and the explained variation is
Therefore, the coefficient of determination is R2 = 290.364/309.68 = 0.953994, implying that the model explains the observed values with 95% of confidence.

Tab. 3. Values for the translation.
Values for the translation.

5 A new criterion to search STBC

Up to the present, the use of random matrices to obtain a search criterion of STBC for MIMO channels is unknown in the literature. Results in this section will establish one. From Section 3, we need to calculate

Since
where f λ m a x ( λ ) is the pdf of the largest eigenvalue of a Wishart matrix. Theorem 4.3 shows

If 0 ≤ ad1, then

Eq (10) is the probability of the maximum likelihood decoder, when receiving R choose wrongly E, if C was sent. When 0 ≤ a < d1, an error occurred. On the other hand, if ad1,
Hence,
where d1 = d1(nR, l) is given by (9). Therefore, we proved the following:

Theorem 5.1 In a MIMO communication channel, where the codeword C was sent, if the maximum likelihood decoder is endowed with the spectral norm, the error probability of received signal be decoded by the codeword E, given that H is known, is

where d1 = d1(nR, l) = 2.53573nr + 0.574893l − 5.40273.

Up to now we are supposing that H is known, that is, the statistics of H are known. Now, we want to calculate the mean in H, that is,

where p(H) is a pdf of the matrix H.

Theorem 5.1 shows that the term ∥H(EC)∥2 is our main concern, since we need more information on the term Γ ( l + n r - 1 , E s ∥ H ( E - C ) ∥ 2 2 - ( 1 + n R ) 2 d 1 ( 1 + n R ) 2 N 0 ) / Γ ( l + n r - 1 ).

Define fm(x) = Γ(m, x)/Γ(m) for m > 0 fixed and x ≥ 0. A typical example is shown in Fig 3. We know that fm(x) is a fast decreasing function, such that 0 < fm(x) ≤ 1, limx→∞ fm(x) = 0 and limm→∞ fm(x) = 1, for all x.

Fig. 3. Graph of fm(x) for m = 9.
Graph of <i>f</i><sub><i>m</i></sub>(<i>x</i>) for <i>m</i> = 9.

Let t = ∥ H ∥ 2 2 and c = ∥ ( E - C ) ∥ 2 2, then

and from property (v) of spectral norm, one has

Thus, ( 1 + n R ) 2 d 1 E s c < t < ∞ and from the behavior of fm(x), it will be enough to assume the following

The elements of H n R × n T are Gaussian random complex variables with mean zero and variance 1/2. From Theorem 4.3, the pdf of the largest eigenvalue of H n R × n T H n T × n R * is given by

where d2 = d2(nR, nT) = 2.53573nR + 0.574893nT − 5.402373. Prior to proving one of our main results, we need the following proposition.

Proposition 5.1 We have that

where 1F1(a, b; z) is the confluent hypergeometric function.

Proof. From Newton’s binomial formula, one has ( E s c t - ( 1 + n R ) 2 d 1 ) i = ∑ j = 0 i ( - 1 ) j ( i j ) ( E s c t ) i - j ( ( 1 + n R ) 2 d 1 ) j. Then, we need to calculate

Since
the result follows.

Now, we have one of the most important result.

Theorem 5.2 In a MIMO communication channel, given that the codeword C was sent, if the decoder is endowed with the spectral norm, the error probability of received signal be decoded by the codeword E, is

where (x)k represents the Pochhammer symbol, defined by (x)k = x(x − 1)(x − 2) ⋯ (xk + 1).

Proof. From (12) and (13), the probability P ( C → E ) = ∫ D o m ψ ( t ) P ( C → E ∣ H ) ψ ( t ) d t, is given by

Since
one has
From Proposition 5.1,
and the result follows.

Theorem 5.2 presents an approximation for the error probability of the sent codeword C wrongly decoded by E, in a transmission on a MIMO channel with a quasi-static coherent flat Rayleigh fading, where the maximum likelihood decoder is endowed with spectral norm. Thus, to obtain STBC with small error probability, we need to find codes which minimize the expression in Theorem 5.2. For fixed Es, nT, nR, l and N0, (14) is a decreasing function of c. In short.

(Largest Eigenvalue Criterion) To design a space-time block code in a MIMO communication channel with slow Rayleigh fading, we need to determine a finite family of matrices C ⊂ M, such that

is as large as possible.

Using the largest eigenvalue criterion, we obtain an upper bound for the pairwise error probability (PEP). Suppose a codeword C was sent and incorrectly decoded as EC. In this case, if the SNR is finite, the PEP is bounded by [51]

where Δ = CE, γ d = 1 4 σ 2 is a constant value proportional to SNR, and σ2 is the noise variance. Since t r a c e (ΔΔ* ) = ∥Δ∥ F 2, from Proposition 3.1, we have
If μ is the minimum given in (17), we obtain the following upper bound for the PEP of any codewords XE
We will refer to (18) as spectral bound.

6 Performance analysis, comparisons and examples

Let PF be the error probability bound giving in (2), used in the Rank and Determinant Criterion, PT be the bound given in (3), used in the Trace Criterion and PE the expression (14) of Theorem 5.2. They may be seen either as a function of the variable c = cF = ∥ECF = ∏λ, c = cT = |trace(EC)| = ∑λ and c = cE = ∥EC2, respectively for PF, PT and PE, or as functions of variable Es, for fixed c and N0.

i) Considering c as variable, Fig 4 shows, from left to right, the curves of PT, PE and PF, where the parameters are nT = nR = l = r = 2, N0 = 1 and Es = 10.

Fig. 4. Graphs of PT, PE and PF for nT = nR = l = 2.
Graphs of <i>PT</i>, <i>PE</i> and <i>PF</i> for <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 2.

ii) Fig 5 shows, from left to right, the curves of PE and PT, where the parameters are nT = nR = l = r = 3, N0 = 1 and Es = 10.

Fig. 5. Graphs of PE and PT for nT = nR = l = 3.
Graphs of <i>PE</i> and <i>PT</i> for <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 3.

In each case, the variable c is a fundamental parameter to choose good STBC for the criterion given. Figs 4 and 5 show that, if we choose a code C where cE is greater or approximately equal to cF and cT, we will have a very low error probability, and, for the largest eigenvalue criterion, we have much more freedom to choose C.

iii) On the other hand, Figs 6 and 7 shows, from top to botton, the curves for PF, PE and PT, respectively, in function of Es for a fixed c. The parameters are nT = nR = l = r = 2, N0 = 1 and c = 5 in Fig 6. In Fig 7, we have the same parameters, where c = 10.

Fig. 6. Graphs of PT, PE and PF for nT = nR = l = 2, where the SNR is a function of Es, for c = 5.
Graphs of <i>PT</i>, <i>PE</i> and <i>PF</i> for <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 2, where the <i>SNR</i> is a function of <i>E</i><sub><i>s</i></sub>, for <i>c</i> = 5.
Fig. 7. Graphs of PE, PT and PF for nT = nR = l = 2, where the SNR is a function of Es, for c = 10.
Graphs of <i>PE</i>, <i>PT</i> and <i>PF</i> for <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 2, where the <i>SNR</i> is a function of <i>E</i><sub><i>s</i></sub>, for <i>c</i> = 10.

6.1 Examples of STBC

Now we will consider several examples of STBC and their pairwise error probabilities will be calculated.

Example 1. The Alamouti SBTC A for BPSK {−1, 1} constellation is given by

Let us consider also the SBTC A 1 given by

Both SBTC have the same rate. For A, c F = max A , A ′ ∈ A d e t ( A - A ′ ) = 8, thus PF(8) ≈ 0.0004, c E = max A , A ′ ∈ A ∥ A - A ′ ∥ 2 = 2 2 and P E ( 2 2 ) ≈ 0 . 0005. We also have max A , A ′ ∈ A r a n k ( A - A ′ ) = 2. Taking c T = max A , A ′ ∈ A t r a c e ( A - A ′ ) = 4, in Fig 8, we show the PEP for each case, with N0 = 1.

Fig. 8. Graphs of PE, PF and PT for A of Example 1, where nT = nR = l = 2, where SNR is a function of Es.
Graphs of <i>PE</i>, <i>PF</i> and <i>PT</i> for A of Example 1, where <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 2, where SNR is a function of <i>E</i><sub><i>s</i></sub>.

Now, for A 1 one has c T = max B , B ′ ∈ A 1 ∥ B - B ′ ∥ 2 = 4 and PE(4) ≈ 0.00002. On the other hand, c T = max B , B ′ ∈ A 1 t r a c e ( B - B ) = 2 and c F = max B , B ′ ∈ A 1 d e t ( B - B ′ ) = 4, thus, PF(4) ≈ 0.0016. Therefore, for the Alamouti code A, one has similar performances. However, for STBC A 1 the performance of spectral case is better than the Rank-Determinant case. Values are shown in Table 4. In Fig 9, we show the PEP for each case. Since this SBTC does not have full diversity, it would not be used by the two classical criteria. Assuming spectral norm, it can be used.

Tab. 4. Comparisons between the two STC.
Comparisons between the two STC.
Fig. 9. Graphs of PE, PF and PT for A 1 of Example 1, where nT = nR = l = 2, and SNR is a function of Es.
Graphs of <i>PE</i>, <i>PF</i> and <i>PT</i> for A 1 of Example 1, where <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 2, and <i>SNR</i> is a function of <i>E</i><sub><i>s</i></sub>.

Example 2. Considering again the Alamouti STBC A, albeit for QPSK { 1 2 ( 1 + j ) , 1 2 ( - 1 + j ) , 1 2 ( - 1 - j ) , 1 2 ( 1 - j ) } constellation. We have a code with 16 matrices.

We have that max A , A ′ ∈ A d e t ( A - A ′ ) = 8, thus PF(8) ≈ 0.0004, but max A , A ′ ∈ A ∥ A - A ′ ∥ 2 = 2 2 and P E ( 2 2 ) ≈ 0 . 0005. Fig 10 shows the PEP for the two cases, for N0 = 1. We see that, for a SNR > 10, the PEP for the spectral case is much lower than the rank and determinant case.

Fig. 10. Graphs of PE, PF and PT for Example 2, where nT = nR = l = 2, and SNR is a function of Es.
Graphs of <i>PE</i>, <i>PF</i> and <i>PT</i> for Example 2, where <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 2, and <i>SNR</i> is a function of <i>E</i><sub><i>s</i></sub>.

Example 3. Jafarkhani proved that we have an orthogonal STBC only for two, four and eight antennas. Let C be the orthogonal STBC for four antennas, where nT = nR = l = 4, considering the BPSK {−1, 1} constellation. The 16 matrices of C are in the form

With regard to this code, we have c F = max A , A ′ ∈ C | d e t ( A - A ′ ) | = 256, c E = max A , A ′ ∈ C ∥ A - A ′ ∥ 2 = 4 and c T = max A , A ′ ∈ C | t r a c e ( A - A ′ ) | = 8. With these parameters, Fig 11 compares the three cases, in which we have a fast decrease of PEP for PE and PT.

Fig. 11. Graphs of PE, PF and PT for Example 3, where nT = nR = l = 4, and SNR is a function of Es.
Graphs of <i>PE</i>, <i>PF</i> and <i>PT</i> for Example 3, where <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 4, and <i>SNR</i> is a function of <i>E</i><sub><i>s</i></sub>.

Example 4. In [52], Jafarkhani introduced the Quasi-orthogonal STBC. In the case of four antennas, we have the code C, where nT = nR = l = 4, considering the BPSK {−1, 1} constellation. The 16 matrices of C are in the form

With regard to this code, we have c F = max A , A ′ ∈ C | d e t ( A - A ′ ) | = 256, c E = max A , A ′ ∈ C ∥ A - A ′ ∥ 2 = 5 . 6568542 and c T = max A , A ′ ∈ C | t r a c e ( A - A ′ ) | = 8. With these parameters, Fig 12 compares the three cases, in which we have a fast decrease of PEP for PE and PT. In Example 3, we have the same values to cF and cT, and a similar value to cE, albeit now we have a better PEP to PE.

Fig. 12. Graphs of PE, PF and PT for Example 4, where nT = nR = l = 4, and SNR is a function of Es.
Graphs of <i>PE</i>, <i>PF</i> and <i>PT</i> for Example 4, where <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 4, and <i>SNR</i> is a function of <i>E</i><sub><i>s</i></sub>.

The largest eigenvalue criterion may be used to choose a STBC from a given set of matrices, which usually would not be chosen by the other criteria. We can see this in the next examples.

Example 5. Let C be the set of matrices for BPSK {−1, 1} constellation given by the matrices

In this set, we have 12 matrices, where the columns of each one of them is a orthogonal set of vectors. Even though we have neither an orthogonal STBC, nor a full diversity code, the set has several important properties, since it has similar mathematical properties to the Alamouti code of Example 1. It can be used for three antennas, with nT = nR = l = 3, considering the largest eingenvalue criterion. Fig 13 shows PEP for the three cases, where, for this code, we have c F = max A , A ′ ∈ C | d e t ( A - A ′ ) | = 2, c E = max A , A ′ ∈ C ∥ A - A ′ ∥ 2 = 2 . 828427 and c T = max A , A ′ ∈ C | t r a c e ( A - A ′ ) | = 4.

Fig. 13. Graphs of PE, PF and PT for Example 5, where nT = nR = l = 3, and SNR is a function of Es.
Graphs of <i>PE</i>, <i>PF</i> and <i>PT</i> for Example 5, where <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 3, and <i>SNR</i> is a function of <i>E</i><sub><i>s</i></sub>.

Example 6. Let C = ( x i - z ¯ z - x i ), where x ∈ R and z ∈ C. We can prove that for any codeword X of C, ∥ X ∥ 2 = x 2 + | z | 2. For convenient choices of the constellation, we may consider finite families of matrices where, for the differences Δ = XE, ∥Δ∥2 will be as large as desired. On the other hand, since trace(Δ) = 0, then C cannot be used as a STBC, according to trace criterion. For instance, in C, let z = a + ib. For x, a, b ∈ {−1, 1} we have eight rank 2 matrices giving a BPSK code. Fig 14 shows PEP for the two cases. For this code, c F = max A , A ′ ∈ C | d e t ( A - A ′ ) | = 12 and c E = max A , A ′ ∈ C ∥ A - A ′ ∥ 2 = 3 . 4641.

Fig. 14. Graphs of PE and PF for Example 6, where nT = nR = l = 2, and SNR is a function of Es.
Graphs of <i>PE</i> and <i>PF</i> for Example 6, where <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 2, and <i>SNR</i> is a function of <i>E</i><sub><i>s</i></sub>.

Example 7. [53] studies space-time group codes. An important example is the following.

Let G = { ± ( 1 0 0 1 ) , ± ( i 0 0 - i ) , ± ( 0 1 - 1 0 ) , ± ( 0 i i 0 ) } and D = ( 1 - 1 1 1 ). Then C = D G is called the Quaternion STBC, and satisfies the Rank and Determinant Criterion. One has c F = max A , A ′ ∈ C | d e t ( A - A ′ ) | = 8 and c E = max A , A ′ ∈ C ∥ A - A ′ ∥ 2 = 2 . 828427. Fig 15 shows PEP for each case.

Fig. 15. Graphs of PE and PF for Example 7, where nT = nR = l = 2, and SNR is a function of Es.
Graphs of <i>PE</i> and <i>PF</i> for Example 7, where <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 2, and <i>SNR</i> is a function of <i>E</i><sub><i>s</i></sub>.

Now, if D 1 = ( 1 1 1 1 ), then C 1 = D 1 G is a new STBC with ∥ C 1 - C 2 ∥ 2 = 4 2 for all C 1 ≠ C 2 ∈ C 1, but the Rank and Determinant Criterion is not satisfied.

Example 8. Let

This matrix appears in the study of representations of Clifford algebras. If we consider the BPSK {−1, 1} constellation, we have a code for eight antennas, with 256 matrices and nT = nR = l = 8.

For this code we have c F = max A , A ′ ∈ C | d e t ( A - A ′ ) | = 518400, c E = max A , A ′ ∈ C ∥ A - A ′ ∥ 2 = 7 . 2111 and c T = max A , A ′ ∈ C | t r a c e ( A - A ′ ) | = 8. With such parameters, we see that Fig 16 compares the three cases, where we have a fast decrease of PEP for PE and PT. Even for a really big value of cF, in this case, PE is better than the other two cases. We may also consider subsets of this code to obtain new codes.

Fig. 16. Graphs of PF, PT and PE for Example 8, where nT = nR = l = 8, and SNR is a function of Es.
Graphs of <i>PF</i>, <i>PT</i> and <i>PE</i> for Example 8, where <i>n</i><sub><i>T</i></sub> = <i>n</i><sub><i>R</i></sub> = <i>l</i> = 8, and <i>SNR</i> is a function of <i>E</i><sub><i>s</i></sub>.

Remark 2. In Figs 6 to 16, we see that, at high SNR, the proposed Largest Eingenvalue Criterion in better than the Rank-determinant Criterion, but it is overcome by the Trace Criterion in some cases. We recall that the Rank Criterion is obtained supposing a small SNR, see [46]. Thus, the comparisons at high SNR are theoretical. On the other hand, the proposed criterion can be used in all SNR regimes.

7 Codes from block matrices

There are several space-time block codes whose matrices are made of blocks of other matrices. In this section, we will supply some properties that will help us apply the Largest Eingenvalue Criterion for those codes.

Quasi-Orthogonal Space-Time Block Codes (QOSTBC) are a powerful family of codes used in Multiple-Input Multiple-Output (MIMO) communication systems, and they provide transmit diversity with higher code rates than the well-known orthogonal STBC (OSTBC), with a lower decoding complexity than non-orthogonal STBC. We have the BPSK case in Example 4 of last section.

We obtain results for several well-known quasi-orthogonal space-time block codes (QOSTBC) from [52] and [54]. In order to derive the main results, we need the following proposition related to the spectral norm and block matrices. Giving matrices A , B ∈ M = M ( n , C ) = M ( n , n , C ), by ( A 0 0 B ), we denote the block matrix of size 2n × 2n, where 0 = 0n×n ∈ M.

Proposition 7.1 Let A, BM. Then,

  • (a)

    ∥ A * A ∥ 2 = ∥ A A * ∥ 2 = ∥ A ∥ 2 2 ;

  • (b)

    ∥ ( A 0 0 B ) ∥ 2 = max { ∥ A ∥ 2 , ∥ B ∥ 2 } ;

  • (c)

    ∥ ( 0 B A 0 ) ∥ 2 = max { ∥ A ∥ 2 , ∥ B ∥ 2 } ;

  • (d)

    ∥ ( 0 - A A 0 ) ∥ 2 = ∥ A ∥ 2 .

Proof.

  • (a)

    Let ∥ A ∥ 2 2 = λ. Then, λ is the largest eingenvalue of A1 = A*A. We have to show that λ is a singular value for both A1 = A*A and A2 = AA*. Let v ≠ 0 be an eingenvector associated to λ. Thus, A*Av = λv. Applying A1 on both sides, we have:

    which means that λ2 is an eigenvalue for (A*A)*(A*A). Therefore, λ is a singular value for A1. The argument is similar for A2. The result then follows from the definition of the spectral norm.

  • (b)

    Suppose x and y are unit vectors with entries in C, and a, b ≥ 0 are such that a2 + b2 = 1. The vector

    is also a unit vector. Since every unit vector v can be written in this form, it follows that
    which concludes the affirmation.

  • (c)

    Define

    From (a),
    Using (b), it follows that
    concluding the proof.

  • (d)

    It follows directly from (c).

Using Proposition 7.1, we find an upper bound for the spectral norm of block matrices and then use the spectral criterion to get a bound for the PEP for STBCs. Let Δ = XE, where XE are codewords of a given code C. We write

From (18) and Proposition 7.1, we have

We can use this result to obtain an estimate of the spectral bound in (18), writing any code in the block form. It also can be used to derive results for STBCs, defined with block matrices. In the sequence, we will present some examples of how this is done for several well-known codes.

i) (QOSTBC for nT = 4) Consider the Alamouti block

As we have seen in Example 4 of last section, the following STBC for nR = nT = k = 4 is due to Jafarkhani [52], with the following code:

Note that C is a rate one Quasi-orthogonal STBC. This means the ML decoding may be done for groups of symbols independently. Using properties of the spectral norm, one can find an upper bound for pairwise error probability of C. Let A , E ∈ C and let Δ = AE. We have

Supposing we have a real constellation (BPSK for instance), it follows that A = A*, and then,
Thus, using the spectral bound for the PEP (19),
which is lower than PEP for the Alamouti code Cpq, for example.

ii) (QOSTBC for nT = 8) Another QOSTBC, whose codewords are block matrices, is also given by Jafarkhani [52]. Define the block

The code in [52] is given by

Similarly to what we did in the previous QOSTBC, we can compute the spectral norm of the difference Δ of the codewords XE for a real constellation with

Each of the terms Δabcd can be calculated using the norm of the codewords from the last example. We obtain

Thus, using the spectral bound again for the PEP (19),

which is lower than the PEP for the Alamouti code Cpq, and the QOSTBC of the Example (i).

Fig 17 compares the SER for the 4 × 4 QOSTBC presented in (i) and the Alamouti code, showing that the first possesses a lower error probability. In Fig 18, we compare the 8×8 QOSTBC of (ii), with the Alamouti code and the 4 × 4 QOSTBC of (i). In both cases, we assume BPSK modulation and, in both cases, we consider the spectral norm for the comparisons.

Fig. 17. Symbol error probability for the QOSTBC with nT = 4 compared with the Alamouti code and an uncoded channel.
Symbol error probability for the QOSTBC with <i>n</i><sub><i>T</i></sub> = 4 compared with the Alamouti code and an uncoded channel.
All codes use BPSK modulation, one receive antenna and Rayleigh fading.
Fig. 18. Symbol error probability for the QOSTBC with nT = 8 compared with the Alamouti code and the 4 × 4 QOSTBC of (i).
Symbol error probability for the QOSTBC with <i>n</i><sub><i>T</i></sub> = 8 compared with the Alamouti code and the 4 × 4 QOSTBC of (i).
All codes use BPSK modulation, one receive antenna and Rayleigh fading.

8 Conclusions

A natural environment where the space-time codes live in is proposed. A new design criterion for space-time block codes for multi-antenna communication systems on coherent Rayleigh fading channels is obtained. This criterion aims at minimizing the pairwise error probability of the maximum likelihood decoder, endowed with the matrix spectral norm. The random matrix theory is used, and a very useful approximation function for the probability density function of the largest eigenvalue of a Wishart Matrix is given, and an approximation for the pairwise error probability function for the spectral case is obtained. The proposed criterion can be used to choose the best STBC in a given family of matrices. The choice is based on the pairwise error probability, serving as a tool to find new codes. Several known and new STBCs were also analyzed in terms of the largest eigenvalue criterion and comparisons were done with the classical criteria.


Zdroje

1. Shannon C. Mathematical theory of communications. Bell Systems Tech J. 1948;27: 379–423, 623–656. doi: 10.1002/j.1538-7305.1948.tb01338.x

2. Foschini GJ and Gans MJ. On limits of wireless communications in a fading environment when using multiple antennas. Wirel Pers Commun. 1998;6: 311–335. doi: 10.1023/A:1008889222784

3. Telatar IE. Capacity of multi-antenna Gaussian channels. European Trans on Telecomm. 1999;10(6): 585–595. doi: 10.1002/ett.4460100604

4. Tarokh V, Seshadri N and Calderbank AR. Space-time codes for high data rate wireless communication: Performance Criterion and Code Construction. IEEE Trans Inf Theory. 1998;44(2): 744–765. doi: 10.1109/18.661517

5. Alamouti SM. A simple transmit diversity technique for wireless communications. IEEE J Sel Top Comm. 1998;16(8): 1451–1458. doi: 10.1109/49.730453

6. Tarokh V, Jafarkani H and Calderbank AR. Space-time block codes from orthogonal designs. IEEE Trans Inf Theory. 1999;45(5): 1456–1467. doi: 10.1109/18.771146

7. Yuan J, Chen Z and Vucetic B. Performance and design of space-time Coding in fading channels. IEEE Trans Commun. 2003;51(12): 1991–1996. doi: 10.1109/TCOMM.2003.820741

8. Tirkkonen O and Hottinen A. Square-matrix embeddable space-time block codes for complex signal constellations. IEEE Trans Inf Theory. 2002;48(2): 384–395. doi: 10.1109/18.978740

9. Wishart J. The generalized product moment distribution in samples from a normal multivariate population. Biometrika 1928;20A: 32–43. doi: 10.1093/biomet/20A.1-2.32

10. Wigner EP. On the Statistical distribution of the widths and spacings of nuclear resonance levels. Proc Camb Philos Soc. 1951;47: 790–798. doi: 10.1017/S0305004100027237

11. Wigner EP. Characteristic vectors of bordered matrices with infinite dimensions. Ann Math. 1955;62: 548–564. doi: 10.2307/1970079

12. Wigner EP. On the distribution of the roots of certain symmetric matrices. Ann Math. 1958;67: 325–327. doi: 10.2307/1970008

13. Dyson F. A Brownian-motion model for the eigenvalues of a random matrix. J Math Phys. 1962;3: 1191–1198. doi: 10.1063/1.1703862

14. Dyson F. Statistical theory of the energy levels of complex systems, I-III. J Math Phys. 1962;3: 140–156, 157–165, 166–175. doi: 10.1063/1.1703773

15. Dyson F. The threefoldway algebraic structure of symmetry groups and ensembles in quantum mechanics. J Math Phys. 1962;3: 1199–1215. doi: 10.1063/1.1703863

16. Tao T and Vu V. Random Matrices: the distribution of the smallest singular values. Geom Funct Anal. 2010;20: 260–297. doi: 10.1007/s00039-010-0057-8

17. Tao T and Vu V. The Wigner-Dyson-Mehta bulk universality conjecture for Wigner Matrices. Electron J Probab. 2011;16: 2104–2121. doi: 10.1214/EJP.v16-944

18. Tao T and Vu V. Random Matrices: Universality of local eigenvalue statistics. Acta Math. 2011;206: 127–204. doi: 10.1007/s11511-011-0061-3

19. Tao T and Vu V. A Central limit theorem for the determinant of a Wigner matrix. Adv Math. 2012;231(1): 74–101. doi: 10.1016/j.aim.2012.05.006

20. Forrester PJ. Log-gases and random matrices. Princeton. Princeton University Press, 2010.

21. Tulino AM and Verdú S. Random matrix theory and wireless communications. Foundations and Trends in Communications and Information Theory, 2004. doi: 10.1561/0100000001

22. Hsu PL. On the distribution of the roots of certain determinantal equations. Ann Eugen. 1939;9: 250–258. doi: 10.1111/j.1469-1809.1939.tb02212.x

23. Jonsson D. Some limit theorems for the eigenvalue of a sample covariance matrix. J Multivar Anal. 1982;12: 1–38. doi: 10.1016/0047-259X(82)90080-X

24. Marcenko VA and Pastur LA. Distributions of eigenvalues for some sets of random matrices. Math USSR-Sb. 1967;1: 457–483. doi: 10.1070/SM1967v001n04ABEH001994

25. Trotter HF. Eigenvalue distributions of large hermitian matrices Wigner semi-circle law and theorem of Kac, Murdock and Szego. Adv Math.1984;54: 67–82. doi: 10.1016/0001-8708(84)90037-9

26. Wachter KW. The strong limits of random matrix spectra for sample matrices of independent elements. Ann Probab. 1978;6: 1–18. doi: 10.1214/aop/1176995607

27. German S. A limit theorem for the norm of random matrices. Ann Probab. 1980;8: 252–261. doi: 10.1214/aop/1176994775

28. Sugiyama T. On the distribution of the largest root of the covariance matrix. Ann Math Stat. 1967;38, 1148–1151. doi: 10.1214/aoms/1177698783

29. Krishnaiah PR and Cheng TC. On the exact distribution of the smallest roots of the Wishart Matrix using zonal polynomials. Ann Inst Stat Math. 1971;23: 293–295. doi: 10.1007/BF02479230

30. Silverstein JW. The smallest eigenvalue of a large-dimensional Wishart Matrix. Ann Probab. 1985;13: 1364–1368. doi: 10.1214/aop/1176992819

31. Vlok JD. Analytic approximation to the largest eigenvalue distribution of a white Wishart matrix. IET Comm. 2012;6(12): 1804–1811. doi: 10.1049/iet-com.2011.0843

32. Chiani M. Distribution of the largest eigenvalue for real Wishart and Gaussian random matrices and a simple approximation for the Tracy-Widom distribution. J Multivar Anal. 2014;129: 69–81. doi: 10.1016/j.jmva.2014.04.002

33. Alfano G, Lozano A, Tulino AM and S. Verdú. Mutual information and eigenvalue distribution of MIMO rician channels. International Symposium on Information Theory and its Applications, ISITA, 2004.

34. Chiani M, Win MZ and Zanella A. On the capacity of spatially correlated MIMO Rayleigh fading channels. IEEE Trans Inf Theory. 2003;49(10): 2363–2371. doi: 10.1109/TIT.2003.817437

35. Ahmadi A. A new approach to fast decode quasi-orthogonal space-time block codes. IEEE Trans Wirel Commun. 2015;14(1): 165–176. doi: 10.1109/TWC.2014.2334615

36. Raleigh GG and Cioffi JM. Spatio-temporal coding for wireless communication. IEEE Trans Commun. 1998;46(3): 357–366. doi: 10.1109/26.662641

37. Gesbert D, Kountouris M, Heath RW Jr, Chae C and Chae T. Shifting the MIMO paradigm. IEEE Signal Process Mag. 2007;24(5): 36–46. doi: 10.1109/MSP.2007.904815

38. Caire G and Shamai S. On the achievable throughput of a multi-antenna gaussian broadcast channel. IEEE Trans Inf Theory. 2003;49(7): 1691–1706. doi: 10.1109/TIT.2003.813523

39. Viswanath P and Tse DNC. Sum capacity of a vector gaussian broadcast channel and uplink-downlink duality. IEEE Trans Inf Theory. 2003;49(8): 1912–1921. doi: 10.1109/TIT.2003.814483

40. Vishwanath S, Jindal N and Goldsmith A. Duality, achievable rates, sum-rate capacity of gaussian MIMO broadcast channels. IEEE Trans Inf Theory. 2003;49(10): 658–668. doi: 10.1109/TIT.2003.817421

41. Larsson EG, Tufvesson F, Edfors O and Marzetta TL. Massive MIMO for next generation wireless systems. IEEE Commun Mag. 2014;52(2): 186–195. doi: 10.1109/MCOM.2014.6736761

42. Marzetta TL. How much training is required for multiuser MIMO. Proc. 40th Asilomar Conf. Signals, Syst., Comput., Nov. 2006; 359–363.

43. Marzetta TL. Noncooperative cellular wireless with unlimited numbers of base station antennas. IEEE Trans Wirel Commun. 2010;9(1): 3590–3600. doi: 10.1109/TWC.2010.092810.091092

44. Rusek F, Persson D, Lau BK, Larsson EG, Marzetta TL, Edfors O, et al. Scaling up MIMO: opportunities and challenges with very large arrays. IEEE Signal Process Mag. 2013;30(1): 40–60. doi: 10.1109/MSP.2011.2178495

45. Björnson E, Hoydis J and Sanguinetti L. Massive MIMO has unlimited capacity. IEEE Trans Wirel Commun. 2018;17(1): 574–590. doi: 10.1109/TWC.2017.2768423

46. Vucetic B and Yuan J. Space-time coding. Wiley, 2003.

47. Horn RA and Johnson CR. Matrix analysis. Cambridge University Press, 1990.

48. Edelman A. Eigenvalues and condition numbers of random matrices. SIAM J Matrix Anal Appl. 1988;9(4): 543–560. doi: 10.1137/0609045

49. Zanella A and Chiani M. The PDF of the Ith largest eigenvalue of central Wishart matrices and its application to the performance analysis of MIMO channels. GLOBECOM, New Orleans, 2008.

50. Zanella A, Chiani M and Win MZ. On the marginal distribution of the eigenvalues of Wishart Matrices. IEEE Trans Commun. 2009;57(4): 1050–1060. doi: 10.1109/TCOMM.2009.04.070143

51. Biglieri E. Coding for wireless channels. Springer, 2005.

52. Jafarkhani H. A quasi-orthogonal space-time block code. IEEE Trans Commun. 2001;49(1): 1–4. doi: 10.1109/26.898239

53. Hughes BL. Optimal space-time constellations from groups. IEEE Trans Inf Theory. 2003;49(2): 401–410. doi: 10.1109/TIT.2002.807283

54. Grover R, Su W and Pados DA. An 8 × 8 quasi-orthogonal STBC form for transmissions over eight or four antennas. IEEE Trans Wirel Commun. 2008;7(12): 4777–4785. doi: 10.1109/T-WC.2008.070791


Článok vyšiel v časopise

PLOS One


2019 Číslo 9
Najčítanejšie tento týždeň
Najčítanejšie v tomto čísle
Kurzy

Zvýšte si kvalifikáciu online z pohodlia domova

Aktuální možnosti diagnostiky a léčby litiáz
nový kurz
Autori: MUDr. Tomáš Ürge, PhD.

Všetky kurzy
Prihlásenie
Zabudnuté heslo

Zadajte e-mailovú adresu, s ktorou ste vytvárali účet. Budú Vám na ňu zasielané informácie k nastaveniu nového hesla.

Prihlásenie

Nemáte účet?  Registrujte sa

#ADS_BOTTOM_SCRIPTS#