EFFICIENT COMPUTATION FOR EIGENVALUE DECOMPOSITION AND SINGULAR VALUE DECOMPOSITION OF MATRICES

Abstract

For eigenvalue decomposition, a first set of at least one variable is derived based on a first matrix being decomposed and using Coordinate Rotational Digital Computer (CORDIC) computation. A second set of at least one variable is derived based on the first matrix and using a look-up table. A second matrix of eigenvectors of the first matrix is then derived based on the first and second variable sets. To derive the first variable set, CORDIC computation is performed on an element of the first matrix to determine the magnitude and phase of this element, and CORDIC computation is performed on the phase to determine the sine and cosine of this element. To derive the second variable set, intermediate quantities are derived based on the first matrix and used to access the look-up table.

Description

CLAIM OF PRIORITY

This application is a continuation of and claims the benefit of priority from U.S. patent application Ser. No. 11/096,839 (now allowed), entitled “Efficient Computation for Eigenvalue Decomposition and Singular Value Decomposition of Matrices” and filed Mar. 31, 2005, which claims the benefit of priority from U.S. Provisional Patent Application Ser. No. 60/628,324, entitled “Eigenvalue Decomposition and Singular Value Decomposition of Matrices Using Jacobi Rotation” and filed Nov. 15, 2004, both of which are assigned to the assignee of this application and are fully incorporated herein by reference for all purposes.

BACKGROUND

1. Field

The present invention relates generally to communication, and more specifically to techniques for decomposing matrices.

2. Background

A multiple-input multiple-output (MIMO) communication system employs multiple (T) transmit antennas at a transmitting entity and multiple (R) receive antennas at a receiving entity for data transmission. A MIMO channel formed by the T transmit antennas and the R receive antennas may be decomposed into S spatial channels, where S≦min {T, R}. The S spatial channels may be used to transmit data in a manner to achieve higher overall throughput and/or greater reliability.

The MIMO channel response may be characterized by an R×T channel response matrix H, which contains complex channel gains for all of the different pairs of transmit and receive antennas. The channel response matrix H may be diagonalized to obtain S eigenmodes, which may be viewed as orthogonal spatial channels of the MIMO channel. Improved performance may be achieved by transmitting data on the eigenmodes of the MIMO channel.

The channel response matrix H may be diagonalized by performing either singular value decomposition of H or eigenvalue decomposition of a correlation matrix of H. The singular value decomposition provides left and right singular vectors, and the eigenvalue decomposition provides eigenvectors. The transmitting entity uses the right singular vectors or the eigenvectors to transmit data on the S eigenmodes. The receiving entity uses the left singular vectors or the eigenvectors to receive data on the S eigenmodes.

Eigenvalue decomposition and singular value decomposition are computationally intensive. There is therefore a need in the art for techniques to efficiently decompose matrices.

SUMMARY

Techniques for efficiently decomposing matrices are described herein. According to an embodiment of the invention, a method is provided in which a first set of at least one variable (e.g., cosine c₁, sine s₁, and magnitude r) is derived based on a first matrix to be decomposed and using Coordinate Rotational Digital Computer (CORDIC) computation. A second set of at least one variable (e.g., variables c and s) is derived based on the first matrix and using a look-up table. A second matrix of eigenvectors is then derived based on the first and second sets of at least one variable.

According to another embodiment, an apparatus is described which includes a CORDIC processor, a look-up processor, and a post-processor. The CORDIC processor derives a first set of at least one variable based on a first matrix to be decomposed. The look-up processor derives a second set of at least one variable based on the first matrix and using a look-up table. The post-processor derives a second matrix of eigenvectors based on the first and second sets of at least one variable.

According to yet another embodiment, an apparatus is described which includes means for deriving a first set of at least one variable based on a first matrix to be decomposed and using CORDIC computation, means for deriving a second set of at least one variable based on the first matrix and using a look-up table, and means for deriving a second matrix of eigenvectors based on the first and second sets of at least one variable.

According to yet another embodiment, a method is provided in which CORDIC computation is performed on an element of a first matrix to determine the magnitude and phase of the element. CORDIC computation is also performed on the phase of the element to determine the sine and cosine of the element. A second matrix of eigenvectors is then derived based on the magnitude, sine, and cosine of the element.

According to yet another embodiment, an apparatus is described which includes means for performing CORDIC computation on an element of a first matrix to determine the magnitude and phase of the element, means for performing CORDIC computation on the phase of the element to determine the sine and cosine of the element, and means for deriving a second matrix of eigenvectors based on the magnitude, sine, and cosine of the element.

According to yet another embodiment, a method is provided in which intermediate quantities are derived based on a first matrix to be decomposed. At least one variable is then derived based on the intermediate quantities and using a look-up table. A second matrix of eigenvectors is derived based on the at least one variable.

According to yet another embodiment, an apparatus is described which includes a pre-processor, a look-up table, and a post-processor. The pre-processor derives intermediate quantities based on a first matrix to be decomposed. The look-up table provides at least one variable based on the intermediate quantities. The post-processor derives a second matrix of eigenvectors based on the at least one variable.

According to yet another embodiment, an apparatus is described which includes means for deriving intermediate quantities based on a first matrix to be decomposed, means for deriving at least one variable based on the intermediate quantities and using a look-up table, and means for deriving a second matrix of eigenvectors based on the at least one variable.

According to yet another embodiment, a method is provided in which multiple iterations of Jacobi rotation are performed on a first matrix of complex values with multiple Jacobi rotation matrices. Each Jacobi rotation matrix is derived by performing eigenvalue decomposition using CORDIC computation, a look-up table, or both. A unitary matrix with orthogonal vectors is then derived based on the multiple Jacobi rotation matrices.

According to yet another embodiment, an apparatus is described which includes means for performing multiple iterations of Jacobi rotation on a first matrix of complex values with multiple Jacobi rotation matrices and means for deriving a unitary matrix with orthogonal vectors based on the multiple Jacobi rotation matrices. Each Jacobi rotation matrix is derived by performing eigenvalue decomposition using CORDIC computation, a look-up table, or both.

According to yet another embodiment, a method is provided in which multiple matrices of complex values are obtained for multiple transmission spans. Multiple iterations of Jacobi rotation are performed on a first matrix of complex values for a first transmission span to obtain a first unitary matrix with orthogonal vectors. Each iteration of the Jacobi rotation utilizes eigenvalue decomposition using CORDIC computation, a look-up table, or both. Multiple iterations of the Jacobi rotation are performed on a second matrix of complex values for a second transmission span to obtain a second unitary matrix with orthogonal vectors. The first unitary matrix is used as an initial solution for the second unitary matrix.

According to yet another embodiment, an apparatus is described which includes means for obtaining multiple matrices of complex values for multiple transmission spans, means for performing multiple iterations of Jacobi rotation on a first matrix of complex values for a first transmission span to obtain a first unitary matrix with orthogonal vectors, and means for performing multiple iterations of the Jacobi rotation on a second matrix of complex values for a second transmission span to obtain a second unitary matrix with orthogonal vectors. Each iteration of the Jacobi rotation utilizes eigenvalue decomposition using CORDIC computation, a look-up table, or both. The first unitary matrix is used as an initial solution for the second unitary matrix.

Various aspects and embodiments of the invention are described in further detail below.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows a processing unit for eigenvalue decomposition of a 2×2 Hermitian matrix.

FIG. 2 shows a CORDIC processor within the processing unit in FIG. 1.

FIG. 3 shows a look-up processor within the processing unit in FIG. 1.

FIG. 4 shows a process for efficiently performing eigenvalue decomposition of the 2×2 Hermitian matrix.

FIG. 5 shows an iterative process for performing eigenvalue decomposition of an N×N Hermitian matrix.

FIG. 6 shows an access point and a user terminal in a MIMO system.

DETAILED DESCRIPTION

The word “exemplary” is used herein to mean “serving as an example, instance, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

The decomposition techniques described herein may be used for single-carrier and multi-carrier communication systems. Multiple carriers may be obtained with orthogonal frequency division multiplexing (OFDM), some other multi-carrier modulation techniques, or some other construct. OFDM effectively partitions the overall system bandwidth into multiple (K) orthogonal frequency subbands, which are also called tones, subcarriers, bins, and frequency channels. With OFDM, each subband is associated with a respective subcarrier that may be modulated with data. For clarity, much of the following description is for a single-carrier MIMO system.

A MIMO channel formed by multiple (T) transmit antennas and multiple (R) receive antennas may be characterized by an R×T channel response matrix H, which may be given as:

$\begin{array}{cc}\underset{\_}{H}=\left[\begin{array}{cccc}{h}_{1,1}& h_{1,2}& \dots & h_{1,T}\\ h_{2,1}& h_{2,2}& \dots & h_{2,T}\\ ⋮& ⋮& \ddots & ⋮\\ h_{R,1}& h_{R,2}& \dots & h_{R,T}\end{array}\right],& \mathrm{Eq}.\left(1\right)\end{array}$

where entry for h_i,j, for i=1, . . . , R and j=1, . . . , T, denotes the coupling or complex channel gain between transmit antenna j and receive antenna i.

The channel response matrix H may be diagonalized to obtain multiple (S) eigenmodes of H, where S≦min {T, R}. The diagonalization may be achieved by performing either singular value decomposition of H or eigenvalue decomposition of a correlation matrix of H.

The eigenvalue decomposition may be expressed as:

R=H^H·H=V·Λ·V^H,  Eq. (2)

where R is a T×T correlation matrix of H;

V is a T×T unitary matrix whose columns are eigenvectors of R;

Λis a T×T diagonal matrix of eigenvalues of R; and

“^H” denotes a conjugate transpose.

The unitary matrix V is characterized by the property V^H·V=I, where I is the identity matrix. The columns of the unitary matrix are orthogonal to one another, and each column has unit power. The diagonal matrix Λ contains possible non-zero values along the diagonal and zeros elsewhere. The diagonal elements of Λ are eigenvalues of R. These eigenvalues are denoted as {λ₁, λ₂, . . . , λ_S} and represent the power gains for the S eigenmodes. R is a Hermitian matrix whose off-diagonal elements have the following property: r_i,j=r_j,i*, where “*” denotes the complex conjugate.

The singular value decomposition may be expressed as:

H=U·Σ·V^H,  Eq. (3)

where U is an R×R unitary matrix of left singular vectors of H;

Σ is an R×T diagonal matrix of singular values of H; and

V is a T×T unitary matrix of right singular vectors of H.

U and V each contain orthogonal vectors. Equations (2) and (3) indicate that the right singular vectors of H are also the eigenvectors of R. The diagonal elements of Σ are the singular values of H. These singular values are denoted as {σ₁, σ₂, . . . , σ_S} and represent the channel gains for the S eigenmodes. The singular values of H are also the square roots of the eigenvalues of R, so that σ_i=√{square root over (λ_i)} for i=1, . . . , S.

A transmitting entity may use the right singular vectors in V to transmit data on the eigenmodes of H, which typically provides better performance than simply transmitting data from the T transmit antennas without any spatial processing. A receiving entity may use the left singular vectors in U or the eigenvectors in V to receive the data transmission on the eigenmodes of H. Table 1 shows the spatial processing performed by the transmitting entity, the received symbols at the receiving entity, and the spatial processing performed by the receiving entity. In Table 1, s is a T×1 vector with up to S data symbols to be transmitted, x is a T×1 vector with T transmit symbols to be sent from the T transmit antennas, r is an R×1 vector with R received symbols obtained from the R receive antennas, n is an R×1 noise vector, and ŝ is a T×1 vector with up to S detected data symbols, which are estimates of the data symbols in s.

TABLE 1
Transmit Spatial
Processing	Received Vector	Receive Spatial Processing
x = V · s	r = H · x + n	ŝ = Σ−1 · UH · r
		ŝ = Λ−1 · VH · HH · r

Eigenvalue decomposition and singular value decomposition of a complex matrix may be performed with an iterative process that uses Jacobi rotation, which is also commonly referred to as Jacobi method or Jacobi transformation. The Jacobi rotation zeros out a pair of off-diagonal elements of the complex Hermitian matrix by performing a plane rotation on the matrix. For a 2×2 complex Hermitian matrix, only one iteration of the Jacobi rotation is needed to obtain the two eigenvectors and two eigenvalues for this matrix. For a larger complex matrix with dimension greater than 2×2, the iterative process performs multiple iterations of the Jacobi rotation to obtain the desired eigenvectors and eigenvalues, or singular vectors and singular values, for the larger complex matrix. Each iteration of the Jacobi rotation on the larger complex matrix uses the eigenvectors of a 2×2 submatrix, as described below.

Eigenvalue decomposition of a simple 2×2 Hermitian matrix R_2×2 may be performed as follows. The Hermitian matrix R_2×2 may be expressed as:

$\begin{array}{cc}{\underset{\_}{R}}_{2\times 2}=\left[\begin{array}{cc}{r}_{1,1}& r_{1,2}\\ r_{2,1}& r_{2,2}\end{array}\right]=\left[\begin{array}{cc}A& B·{}^{{\mathrm{j\theta }}_b}\\ B·{}^{-{\mathrm{j\theta }}_b}& D\end{array}\right],& \mathrm{Eq}.\left(4\right)\end{array}$

where A, B, and D are arbitrary real values, and θ_b is an arbitrary phase.

The first step of the eigenvalue decomposition of R_2×2 is to apply a two-sided unitary transformation, as follows:

$\begin{array}{cc}\begin{array}{c}{\underset{\_}{R}}_{\mathrm{re}}=\left[\begin{array}{cc}1& 0\\ 0& {}^{{\mathrm{j\theta }}_b}\end{array}\right]·\left[\begin{array}{cc}A& B·{}^{{\mathrm{j\theta }}_b}\\ B·{}^{-{\mathrm{j\theta }}_b}& D\end{array}\right]·\left[\begin{array}{cc}1& 0\\ 0& {}^{-{\mathrm{j\theta }}_b}\end{array}\right]\\ =\left[\begin{array}{cc}A& B\\ B& D\end{array}\right],\end{array}& \mathrm{Eq}.\left(5\right)\end{array}$

where R_re is a symmetric real matrix containing real values and having symmetric off-diagonal elements at locations (1, 2) and (2,1).

The symmetric real matrix R_re is then diagonalized using a two-sided Jacobi rotation, as follows:

$\begin{array}{cc}\begin{array}{c}{\underset{\_}{\Lambda }}_{2\times 2}=\left[\begin{array}{cc}\cos \phi & -\sin \phi \\ \sin \phi & \cos \phi \end{array}\right]·\left[\begin{array}{cc}A& B\\ B& D\end{array}\right]·\left[\begin{array}{cc}\cos \phi & \sin \phi \\ -\sin \phi & \cos \phi \end{array}\right]\\ =\left[\begin{array}{cc}{\lambda }_1& 0\\ 0& {\lambda }_2\end{array}\right],\end{array}& \mathrm{Eq}.\left(6\right)\end{array}$

where angle θ may be expressed as:

$\begin{array}{cc}\phi =\frac{1}{2}{\tan }^{-1}\left(\frac{2B}{D-A}\right).& \mathrm{Eq}.\left(7\right)\end{array}$

A 2×2 unitary matrix V_2×2 of eigenvectors of R_2×2 may be derived as:

$\begin{array}{cc}\begin{array}{c}{\underset{\_}{V}}_{2\times 2}=\left[\begin{array}{cc}1& 0\\ 0& {}^{-{\mathrm{j\theta }}_b}\end{array}\right]·\left[\begin{array}{cc}\cos \phi & \sin \phi \\ -\sin \phi & \cos \phi \end{array}\right]\\ =\left[\begin{array}{cc}\cos \phi & \sin \phi \\ -{}^{-{\mathrm{j\theta }}_b}·\sin \phi & {}^{-{\mathrm{j\theta }}_b}·\cos \phi \end{array}\right].\end{array}& \mathrm{Eq}.\left(8\right)\end{array}$

The two eigenvalues λ₁ and λ₂ may be computed directly from the elements of R_re as follows:

$\begin{array}{cc}{\lambda }_{1,2}=\frac{1}{2}\left(A+D\right)\pm B·\sqrt{1+{\left(\frac{D-A}{2B}\right)}^2}.& \mathrm{Eq}.\left(9\right)\end{array}$

Equation (9) is the solution to a characteristic equation of R_2×2. In equation (9), λ₁ is obtained with the plus sign for the second quantity on the right hand side, and λ₂ is obtained with the minus sign for the second quantity, where λ₁≧λ₂.

The elements of V_2×2 may be computed directly from the elements of R_2×2, as follows:

$\begin{array}{cc}r=\sqrt{{\left(\mathrm{Re}\left\{r_{1,2}\right\}\right)}^2+{\left(\mathrm{Im}\left\{r_{1,2}\right)\right)}^2},& \mathrm{Eq}.\left(10a\right)\\ c_1=\frac{\mathrm{Re}\left\{r_{1,2}\right\}}{r}=\cos \left(\angle r_{1,2}\right),& \mathrm{Eq}.\left(10b\right)\\ s_1=\frac{\mathrm{Im}\left\{r_{1,2}\right\}}{r}=\sin \left(\angle r_{1,2}\right),& \mathrm{Eq}.\left(10c\right)\\ g_1=c_1-{\mathrm{js}}_1,& \mathrm{Eq}.\left(10d\right)\\ \tau =\frac{r_{2,2}-r_{1,1}}{2·r},& \mathrm{Eq}.\left(10e\right)\\ x=\sqrt{1+{\tau }^2},& \mathrm{Eq}.\left(10f\right)\\ t=\frac{1}{\tau +x},& \mathrm{Eq}.\left(10g\right)\\ c=\frac{1}{\sqrt{1+t^2}},& \mathrm{Eq}.\left(10h\right)\\ s=t·c=\sqrt{1-c^2},\mathrm{if}\left(r_{2,2}-r_{1,1}\right)<0,& \mathrm{Eq}.\left(10i\right)\\ \mathrm{then}{\underset{\_}{v}}_{2\times 2}=\left[\begin{array}{cc}{v}_{1,1}& v_{1,2}\\ v_{2,1}& v_{2,2}\end{array}\right]=\left[\begin{array}{cc}c& -s\\ g_1·s& g_1·c\end{array}\right],& \mathrm{Eq}.\left(10j\right)\\ \mathrm{else}{\underset{\_}{v}}_{2\times 2}=\left[\begin{array}{cc}{v}_{1,1}& v_{1,2}\\ v_{2,1}& v_{2,2}\end{array}\right]=\left[\begin{array}{cc}s& c\\ g_1·c& -g_1·s\end{array}\right],& \mathrm{Eq}.\left(10k\right)\end{array}$

where r_1,1, r_1,2 and r_2,1 are elements of R_2×2, r is the magnitude of r_1,2, and <r_1,2 is the phase of r_1,2, which is also denoted as θ=<r_1,2.

Equation set (10) performs a complex Jacobi rotation on the 2×2 Hermitian matrix R_2×2 to obtain the matrix V_2×2 of eigenvectors of R_2×2. The computations in equation set (10) are derived to eliminate the arc-tangent operation in equation (7) and the cosine and sine operations in equation (8). However, the square root and division operations in equation set (10) present their own implementation difficulties, so a simpler implementation is desirable. Equation set (10) is computed once for eigenvalue decomposition of the 2×2 Hermitian matrix R_2×2. A complex matrix larger than 2×2 may be decomposed by performing eigenvalue decomposition of many 2×2 submatrices, as described below. Thus, it is highly desirable to compute equation set (10) as efficiently as possible in order to reduce the amount of time needed to decompose the larger complex matrix.

FIG. 1 shows a block diagram of an embodiment of a processing unit 100 capable of efficiently computing equation set (10) for eigenvalue decomposition of the 2×2 Hermitian matrix R_2×2. Processing unit 100 includes a CORDIC processor 110, a look-up processor 120, and a post-processor 130. Processing unit 100 receives elements r_1,1, r_1,2 and r_2,1 of matrix R_2×2 and provides elements v_1,1, v_1,2, v_2,1, and V₂₂ of matrix V_2×2. Within processing unit 100, CORDIC processor 110 receives element r_1,2 of R_2×2 and computes the magnitude r, phase θ, cosine c₁, and sine s₁ of element r_1,2. Look-up processor 120 receives elements r_1,1 and r₂₂ of R_2×2 and the magnitude r of element r_1,2 from CORDIC processor 110 and computes variables c and s. Post-processor 130 receives elements a r_1,1 and r_2,2, variables c₁ and s₁ from CORDIC processor 110, and variables c and s from look-up processor 120 and computes elements v_1,1, v_1,2, v_2,1, and v_2,2 of matrix V_2×2.

CORDIC Processor

Variables r, c₁, s₁ and hence g₁ in equation set (10) may be efficiently computed using CORDIC processor 110. A CORDIC processor implements an iterative algorithm that allows for fast hardware calculation of trigonometric functions such as sine, cosine, magnitude, and phase using simple shift and add/subtract hardware. Variables r, c₁ and s₁ may be computed in parallel to reduce the amount of time needed to perform eigenvalue decomposition. The CORDIC processor computes each variable iteratively, with more iterations producing higher accuracy for the variable.

A complex multiply of two complex numbers, R=R+jR_im and C=C_re+jC_im, may be expressed as:

$\begin{array}{cc}\begin{array}{c}Y=R·C,\\ =\left(R_{\mathrm{re}}+jR_{\mathrm{im}}\right)·\left(C_{\mathrm{re}}+jC_{\mathrm{im}}\right),\\ =\left(R_{\mathrm{re}}·C_{\mathrm{re}}-R_{\mathrm{im}}·C_{\mathrm{im}}\right)+j\left(R_{\mathrm{re}}·C_{\mathrm{im}}+R_{\mathrm{im}}·C_{\mathrm{re}}\right),\end{array}& \mathrm{Eq}.\left(11\right)\\ \mathrm{where}Y=Y_{\mathrm{re}}+jY_{\mathrm{im}},& \\ Y_{\mathrm{re}}=R_{\mathrm{re}}·C_{\mathrm{re}}-R_{\mathrm{im}}·C_{\mathrm{im}},\mathrm{and}& \mathrm{Eq}.\left(12a\right)\\ Y_{\mathrm{im}}=R_{\mathrm{re}}·C_{\mathrm{im}}+R_{\mathrm{im}}·C_{\mathrm{re}}.& \mathrm{Eq}.\left(12b\right)\end{array}$

The magnitude of Y is equal to the product of the magnitudes of R and C. The phase of Y is equal to the sum of the phases of R and C.

The complex number R may be rotated by up to 90 degrees by multiplying R with a complex number C_i having the following form: C_i=1±jK_i, where C_i,re=1 and C_i,im=±K_i. K_i is decreasing powers of two and has the following form:

K_i=2⁻ⁱ,  Eq. (13)

where i is an index that is defined as i=0, 1, 2, . . . .

The complex number R may be rotated counter-clockwise if the complex number C_i has the form C_i=1+jK_i. The phase of C_i is then <C_i=arctan (K_i). Equation set (12) may then be expressed as:

Y_re=R_re−K_i·R_im=R_re−2⁻ⁱ·R_im, and  Eq. (14a)

Y_im=R_im+K_i·R_re=R_im+2⁻ⁱ·R_re.  Eq. (14b)

The complex number R may be rotated clockwise if the complex number C, has the form C_i=1−jK_i. The phase of C_i is then <C_i=−arctan (K_i). Equation set (12) may then be expressed as:

Y_re=R_re+K_i·R_re+2⁻ⁱ·R_im, and  Eq. (15a)

Y_im=R_im−K·R=R_im−2⁻ⁱ·R_re.  Eq. (15b)

The counter-clockwise rotation in equation set (14) and the clockwise rotation in equation set (15) by the complex number C_i may be achieved by shifting both R_im and R_re by i bit positions, adding/subtracting the shifted R_im to/from R_re to obtain Y_re, and adding/subtracting the shifted R_re to/from R_im to obtain Y_im. No multiplies are needed to perform the rotation.

Table 2 shows the value of K_i, the complex number C_i, the phase of C_i, the magnitude of C_i, and the CORDIC gain g_i for each value of i from 0 through 7. As shown in Table 2, for each value of i, the phase of C_i, is slightly more than half the phase of C_i−1. A given target phase may be obtained by performing a binary search and either adding or subtracting each successively smaller phase value θ_i. Index i denotes the number of iterations for the binary search, and more iterations give a more accurate final result.

TABLE 2
			Phase of
			Ci,	Magnitude
			θi =	of	CORDIC
i	Ki = 2−i	Ci = 1 + jKi	arctan (Ki)	Ci	Gain, gi
0	1.0	1 + j1.0	45.00000	1.41421356	1.414213562
1	0.5	1 + j0.5	26.56505	1.11803399	1.581138830
2	0.25	1 + j0.25	14.03624	1.03077641	1.629800601
3	0.125	1 + j0.125	7.12502	1.00778222	1.642484066
4	0.0625	1 + j0.0625	3.57633	1.00195122	1.645688916
5	0.03125	1 + j0.03125	1.78991	1.00048816	1.646492279
6	0.015625	1 + j0.015625	0.89517	1.00012206	1.646693254
7	0.007813	1 + j0.007813	0.44761	1.00003052	1.646743507
. . .	. . .	. . .	. . .	. . .	. . .

Since the magnitude of C_i is greater than 1.0 for each value of i, multiplication of R with C_i results in the magnitude of R being scaled by the magnitude of C_i. The CORDIC gain for a given value of i is the cumulative magnitude of C_i for the current and prior values of i. The CORDIC gain for i is obtained by multiplying the CORDIC gain for i−1 with the magnitude of C_i, or g_i=g_i−1·|C_i|. The CORDIC gain is dependent on the value of i but converges to a value of approximately 1.647 as i approaches infinity.

In equation set (10), r is the magnitude of element r_1,2 and B is the phase of element r_1,2. The magnitude and phase of r_1,2 may be determined by CORDIC processor 110 as follows. A variable {tilde over (r)}_1,2 is formed with the absolute values of the real and imaginary parts of r_1,2, or {tilde over (r)}_1,2=abs (Re {r_1,2})+j abs (Im {r_1,2}). Thus, {tilde over (r)}_1,2 sits on the first quadrant of an x-y plane. The phase θ is initialized to zero. {tilde over (r)}_1,2 is then iteratively rotated such that its phase approaches zero.

For each iteration starting with i=0, {tilde over (r)}_1,2 is deemed to have (1) a positive phase if the imaginary part of {tilde over (r)}_1,2 is positive or (2) a negative phase if the imaginary part of {tilde over (r)}_1,2 is negative. If the phase of {tilde over (r)}_1,2 is negative, then {tilde over (r)}_1,2 is rotated counter-clockwise by θ_i (or equivalently, θ_i is added to the phase of {tilde over (r)}_1,2) by multiplying {tilde over (r)}_1,2 with C_i=1+jK_i, as shown in equation set (14). Conversely, if the phase of {tilde over (r)}_1,2 is positive, then {tilde over (r)}_1,2 is rotated clockwise by θ_i (or equivalently, θ_i is subtracted from the phase of {tilde over (r)}_1,2) by multiplying {tilde over (r)}_1,2 with C_i=1−jK_i, as shown in equation set (15). {tilde over (r)}_1,2 is thus updated in each iteration by either a counter-clockwise or clockwise rotation. The phase θ is updated by (1) adding θ_i to the current value of θ if θ_i was added to the phase of {tilde over (r)}_1,2 or (2) subtracting θ_i from the current value of θ if θ_i was subtracted from the phase of {tilde over (r)}_1,2. θ thus represents the cumulative phase that has been added to or subtracted from the phase of {tilde over (r)}_1,2 to zero out the phase.

The final result becomes more accurate as more iterations are performed. Ten iterations are typically sufficient for many applications. After all of the iterations are completed, the phase of {tilde over (r)}_1,2 should be close to zero, the imaginary part of {tilde over (r)}_1,2 should be approximately zero, and the real part of {tilde over (r)}_1,2 is equal to the magnitude of r_1,2 scaled by the CORDIC gain, or r=Re {{tilde over (r)}_1,2}/g_i. The final value of θ is the total phase rotation to zero out the phase of {tilde over (r)}_1,2. The phase of {tilde over (r)}_1,2 is thus equal to −θ. The phase θ may be represented by a sequence of sign bits, z₁ z₂ z₃ . . . , where z_i=1 if was subtracted from θ and z_i=−1 if θ_i was added to θ.

The computation of the magnitude and phase of r_1,2 may performed as follows. First the variables are initialized as:

i=0,  Eq. (16a)

x₀=abs(Re{r_1,2}),  Eq. (16b)

y₀=abs(Im{r_1,2}), and  Eq. (16c)

θ_tot(i)=0.  Eq. (16d)

A single iteration of the CORDIC computation may be expressed as:

$\begin{array}{cc}{z}_i=\mathrm{sign}\left(y_i\right)=\{\begin{array}{cc}1& \mathrm{if}y_i\ge 0\\ -1& \mathrm{if}y_i<0,\end{array}& \mathrm{Eq}.\left(17a\right)\\ x_{i+1}=x_i+z_i·2^{-i}·y_i,& \mathrm{Eq}.\left(17b\right)\\ y_{i+1}=y_i-z_i·2^{-i}·x_i,& \mathrm{Eq}.\left(17c\right)\\ {\theta }_i=\arctan \left(2^{-i}\right),& \mathrm{Eq}.\left(17d\right)\\ {\theta }_{\mathrm{tot}}\left(i+1\right)={\theta }_{\mathrm{tot}}\left(i\right)-z_i·{\theta }_i,\mathrm{and}& \mathrm{Eq}.\left(17e\right)\\ i=i+1.& \mathrm{Eq}.\left(17f\right)\end{array}$

In equations (17b) and (17c), a counter-clockwise rotation is performed if the phase of x_i+jy_i is positive and z_i=1, and a clockwise rotation is performed if the phase of x_i+jy_i is negative and z_i=−1. After all of the iterations are completed, the magnitude is set as r=x_i+1 and the phase is set as θ=θ_tot(i+1). The scaling by the CORDIC gain may be accounted for by other processing blocks.

In equation set (10), c₁ is the cosine of r_1,2 and s₁ is the sine of r_1,2. The cosine and sine of r_1,2 may be determined by CORDIC processor 110 as follows. A unit magnitude complex number R′ is initialized as R′=x₀′+jy₀′=1+j0 and is then rotated by −θ. For each iteration starting with i=0, the complex number R′ is rotated (1) counter-clockwise by θ_i by multiplying R′ with C_i=1+jK_i if sign bit z_i indicates that 9 was subtracted from 9 or (2) clockwise by θ_i by multiplying R′ with C_i=1jK_i if sign bit z_i indicates that θ_i was added to θ. After all of the iterations are completed, c₁ is equal to the real part of the final R′ scaled by the CORDIC gain, or c₁=x_i′/g_i, and s₁ is equal to the imaginary part of the final R′ scaled by the CORDIC gain, or s₁=y_i′/g_i.

The computation of the cosine and sine of r_1,2 may be performed as follows. First the variables are initialized as:

i=0,  Eq. (18a)

x₀′=1,  and Eq. (18b)

y₀′=0.  Eq. (18c)

A single iteration of the CORDIC computation may be expressed as:

x_i+1′=x_i′−z_i·2⁻ⁱ·y_i′,  Eq. (19a)

y_i+1′=y_i′+z_i·2⁻ⁱ·x_i′, and  Eq. (19b)

i=i+1.  Eq. (19c)

In equations (19a) and (19b), for each iteration i, R′ is rotated in the direction indicated by sign bit z_i. After all of the iterations are completed, the cosine is set as c₁=x_i+1′ and the sine is set as s₁=y_i+1′. The scaling by the CORDIC gain may be accounted for by other processing blocks.

The cosine c₁ and the sine s₁ may also be computed in parallel with the magnitude r by initializing the variable R′ to a CORDIC gain scaled version of 1+j0, or R′=(1+j0)/g where g is the CORDIC gain for the number of iterations to be performed. At each iteration, the CORDIC rotation performed for the magnitude r is determined, and an opposite CORDIC rotation is performed on the variable R′. For this scheme, it is not necessary to determine the phase θ.

FIG. 2 shows a block diagram of CORDIC processor 110 for computing variables r, c₁, and s₁ in equation set (10). CORDIC processor 110 includes a CORDIC unit 210 that computes the magnitude and phase of element r_1,2 and a CORDIC unit 230 that computes the cosine and sine of element r_1,2. CORDIC units 210 and 230 may be operated in parallel.

Within CORDIC unit 210, a demultiplexer (Demux) 208 receives element r_1,2, provides abs (Re {r_1,2}) as x₀, and provides abs (Im {r_1,2}) as y₀. A multiplexer (Mux) 212a receives x₀ on a first input and x_i from a delay element 219a on a second input, provides x₀ on its output when i=0, and provides x_i on its output when i>0. The output of multiplexer 212a is x_i for the current iteration. A shifter 214a receives and shifts x_i to the left by i bits and provides a shifted x_i. A multiplexer 212b receives y_o on a first input and y_i from a delay element 219b on a second input, provides y₀ on its output when i=0, and provides y_i on its output when i>0. The output of multiplexer 212b is y_i for the current iteration. A shifter 214b receives and shifts y_i to the left by i bits and provides a shifted y_i. A sequencer 222 steps through index i and provides appropriate controls for the units within CORDIC processor 110. A sign detector 224 detects the sign of y_i and provides sign bit z_i, as shown in equation (17a).

A multiplier 216a multiplies the shifted x_i with sign bit z_i. A multiplier 216b multiplies the shifted y_i with sign bit z_i. Multipliers 216a and 216b may be implemented with bit inverters. A summer 218a sums the output of multiplier 216b with x_i and provides x_i+1 for the current iteration (which is also x_i for the next iteration) to delay element 219a and a switch 220. A summer 218b subtracts the output of multiplier 216a from y_i and provides y_i+1 for the current iteration (which is also y_i for the next iteration) to delay element 219b. Switch 220 provides x_i+1 as the magnitude r after all of the iterations are completed.

Within CORDIC unit 230, a multiplexer 232a receives x₀′=1 on a first input and x_i′ from a delay element 239a on a second input, provides x₀′ on its output when i=0, and provides x_i′ on its output when i>0. The output of multiplexer 232a is x_i′ for the current iteration. A shifter 234a receives and shifts x_i′ to the left by i bits and provides a shifted x_i′. A multiplexer 232b receives y₀′=0 on a first input and y_i′ from a delay element 239b on a second input, provides y₀′ on its output when i=0, and provides y_i′ on its output when i>0. The output of multiplexer 232b is y_i′ for the current iteration. A shifter 234b receives and shifts y_i′ to the left by i bits and provides a shifted y_i′.

A multiplier 236a multiplies the shifted x_i′ with sign bit z_i from detector 224. A multiplier 236b multiplies the shifted y_i′ with sign bit z_i. Multipliers 236a and 236b may also be implemented with bit inverters. A summer 238a subtracts the output of multiplier 236b from x_i′ and provides x_i+1′ for the current iteration (which is also x_i′ for the next iteration) to delay element 239a and a switch 240a. A summer 238b adds the output of multiplier 236a with y_i′ and provides y_i+1′ for the current iteration (which is also y_i′ for the next iteration) to delay element 239b and a switch 240b. After all of the iterations are completed, switch 240a provides x_i+1′ as cosine c₁, and switch 240b provides y_i+1′ as sine s₁.

Look-Up Table

In equation set (10), variables c and s are functions of only τ, and intermediate variables x and t are used to simplify the notation. abs(τ) ranges from 0 to ∞, variable c ranges from 0.707 to 1.0, and variable s ranges from 0.707 to 0.0. A large range of values for τ is thus mapped to a small range of values for c and also a small range of values for s. Hence, an approximate value of r should give good approximations of both c and s.

A look-up table (LUT) may be used to efficiently compute variables c and s based on the dividend/numerator and the divisor/denominator for r. The use of the look-up table avoids the need to perform a division to compute r in equation (10e), a square root to compute x in equation (10f), a division to compute t in equation (10g), a division and a square root to compute c in equation (10h), and a multiply to compute s in equation (10i). Since division and square root operations are computationally intensive, the use of the look-up table can greatly reduce the amount of time needed to perform eigenvalue decomposition. The computation of variables c and s using the look-up table may be performed as follows.

The dividend is equal to r_2,2−r_1,1, and the absolute value of the dividend is represented as a binary floating-point number of the form 1.b_n1 b_n2 b_n3 . . . ×2^mn, where b_ni represents a single bit of the mantissa for the dividend and m_n is the exponent for the dividend. The binary floating-point number for the dividend may be obtained by left shifting the dividend one bit at a time until the most significant bit (MSB) is a one (‘1’), and setting m_n equal to the number of left bit shifts.

The divisor is equal to 2·r and is a positive value because r is the magnitude of r_1,2. The divisor is also represented as a binary floating-point number of the form 1.b_d1 b_d2 b_d3 . . . ×2^md, where b_di represents a single bit of the mantissa for the divisor and m_d is the exponent for the divisor. The binary floating-point number for the divisor may be obtained by left shifting the divisor one bit at a time until the MSB is a one (‘1’), and setting m_d equal to the number of left bit shifts.

The fractional bits of the mantissa for the dividend (which are b_n1, b_n2, b_n3, . . . ), the fractional bits of the mantissa for the divisor (which are b_d1, b_d2, b_d3, . . . ), and the difference in the exponents for the dividend and the divisor (which is Δm=m_d−m_n) are intermediate quantities that are used as an input address for the look-up table. The look-up table returns stored values for variables c and s based on the input address.

In general, the look-up table may be of any size. A larger size look-up table can provide greater accuracy in the computation of variables c and s. In a specific embodiment, the look-up table has a size of 2K×16, an 11-bit input address, a 8-bit output for variable c, and a 8-bit output for variable s. The 11-bit input address is composed of three fractional bits b_n1, b_n2 and b_n3 for the dividend, three factional bits b_d1, b_d2 and b_d3 for the divisor, and five bits for the exponent difference Δm. The range of values for the 5-bit exponent difference was determined by computer simulation. The minimum observed exponent difference was −17 and the maximum observed exponent difference was +14. Since variables c and s are positive values, it is not necessary to store the sign bit for each of these variables in the look-up table. Furthermore, since variable c is always greater than 0.707, the leftmost fractional bit is always equal to ‘1’ and does not need to be stored in the look-up table. The look-up table thus stores the next eight leftmost fractional bits for variable c (i.e., excluding the leftmost fractional bit) and the eight leftmost fractional bits for variable s. A 10-bit signed value for variable c can thus be obtained with an 8-bit unsigned value provided by the look-up table for c. A 9-bit signed value for variable s can be obtained with an 8-bit unsigned value provided by the look-up table for s.

FIG. 3 shows a block diagram of look-up processor 120 for computing variables c and s in equation set (10). Within look-up processor 120, a difference computation unit 312 receives elements r_1,1 and r_2,2 of matrix R_2×2, computes the difference between these elements as r_2,2−r_1,1, and provides the difference as the dividend for τ. A format converter 314 converts the dividend into a binary floating-point number 1.b_n1 b_n2 b_n1, . . . ×2^mn, provides the fractional bits {b_n} of the mantissa to a look-up table 320, and provides the exponent m_n to an exponent difference computation unit 318. A format converter 316 receives variable r as the divisor, converts the divisor into a binary floating-point number 1.b_d1 b_d2 b_d3 . . . ×2^md, provides the fractional bits m_d of the mantissa to look-up table 320, and provides the exponent m_d to unit 318. Unit 318 computes the difference of the exponents for the dividend and the divisor as Δm=m_d−m_n and provides the exponent difference Δm to look-up table 320. Units 312, 314, 316, and 318 form a pre-processor that generates the intermediate quantities for look-up table 320.

Look-up table 320 receives the fractional bits {b_n} for the dividend, the fractional bits {b_d} for the divisor, and the exponent difference Δm as an input address. Look-up table 320 provides the stored values for variables c and s based on the input address. An output unit 322 appends a ‘1’ for the leftmost fractional bit for variable c, further appends a plus sign bit (‘+’), and provides the final value of c. An output unit 324 appends a plus sign bit (‘+’) for variable s and provides the final value of s.

Look-up table 320 may be designed to account for the CORDIC gain in the computation of r, c₁, and s₁ so that the elements of V_2×2 have the proper magnitude. For example, since the magnitude r is used to form the address for look-up table 320, the CORDIC gain for the magnitude r may be accounted for in the addressing of the look-up table. In another embodiment, look-up table 320 stores a rotation sequence for a CORDIC processor, which then computes variables c and s with the rotation sequence. The rotation sequence is the sequence of sign bits z_i and may be stored using fewer bits than the actual values for variables c and s. However, the CORDIC processor would require some amount of time to compute variables c and s based on the rotation sequence.

FIG. 4 shows a process 400 for efficiently performing eigenvalue decomposition of the 2×2 Hermitian matrix R_2×2. Variables r, c₁, and s₁ are computed based on element r_1,2 of R_2×2 and using CORDIC computation (block 412). Variables c and s are computed based on elements r_1,1 and r_2,2 of R_2×2 and variable r and using a look-up table (block 414). The computation in block 412 may be performed in parallel with the computation in block 414 by different hardware units to speed up the eigenvalue decomposition. The four elements of matrix V_2×2, which contains the eigenvectors of R_2×2, are then derived based on variables c₁, s₁, c and s, as shown in equation (10j) or (10k) (block 416).

In the description above, the variables r, c₁ and s₁ are derived with a CORDIC processor and the variables c and s are derived with a look-up table. The variables r, c₁ and s₁ may also be derived with one or more look-up tables of sufficient size to obtain the desired accuracy for r, c_i and s₁. The variables r, c₁ and s₁ and the variables c and s may also be computed in other manners and/or using other algorithms (e.g., power series). The choice of which method and algorithm to compute each set of variables may be dependent on various factors such as the available hardware, the amount of time available for computation, and so on.

Eigenvalue Decomposition

Eigenvalue decomposition of an N×N Hermitian matrix that is larger than 2×2, as shown in equation (2), may be performed with an iterative process. This iterative process uses the Jacobi rotation repeatedly to zero out off-diagonal elements in the N×N Hermitian matrix. For the iterative process, N×N unitary transformation matrices are formed based on 2×2 Hermitian sub-matrices of the N×N Hermitian matrix and are repeatedly applied to diagonalize the N×N Hermitian matrix. Each N×N unitary transformation matrix contains four non-trivial elements (elements other than 0 or 1) that are derived from elements of a corresponding 2×2 Hermitian sub-matrix. The resulting diagonal matrix contains the real eigenvalues of the N×N Hermitian matrix, and the product of all of the unitary transformation matrices is an N×N matrix of eigenvectors for the N×N Hermitian matrix.

In the following description, index i denotes the iteration number and is initialized as i=0. R is an N×N Hermitian matrix to be decomposed, where N>2. An N×N matrix D_i is an approximation of diagonal matrix Λ of eigenvalues of R and is initialized as D₀=R. An N×N matrix V_i is an approximation of matrix V of eigenvectors of R and is initialized as V₀=I.

A single iteration of the Jacobi rotation to update matrices D_i and V_i may be performed as follows. First, a 2×2 Hermitian matrix D_pq is formed based on the current D_i as follows:

$\begin{array}{cc}{\underset{\_}{D}}_{\mathrm{pq}}=\left[\begin{array}{cc}{d}_{p,p}& d_{p,q}\\ d_{q,p}& d_{q,q}\end{array}\right],& \mathrm{Eq}.\left(20\right)\end{array}$

where d_p,q is the element at location (p,q) in D_i; and

pε{1, . . . , N}, qε{1, . . . , N}, and p≠q.

D_pq is a 2×2 submatrix of D_i, and the four elements of D_pq are four elements at locations (p,p), (p,q), (q,p) and (q,q) in D_i. The values for indices p and q may be selected in a predetermined or deterministic manner, as described below.

Eigenvalue decomposition of D_pq is then performed as shown in equation set (10) to obtain a 2×2 unitary matrix V_pq of eigenvectors of D_pq. For the eigenvalue decomposition of p_pq, R_2×2 in equation (4) is replaced with p_pq, and V_2×2 from equation (10j) or (10k) is provided as V_pq.

An N×N complex Jacobi rotation matrix T_pq is then formed with matrix V_pq T_pq is an identity matrix with the four elements at locations (p,p), (p,q), (q,p) and (q,q) replaced with the (1,1), (1,2), (2,1) and (2,2) elements, respectively, of V_pq. T_pq has the following form:

$\begin{array}{cc}{\underset{\_}{T}}_{\mathrm{pq}}=\left[\begin{array}{ccccccc}1& & & & & & \\ & \ddots & & & & & \\ & & v_{1,1}& \dots & v_{1,2}& & \\ & & ⋮& 1& ⋮& & \\ & & v_{2,1}& \dots & v_{2,2}& & \\ & & & & & \ddots & \\ & & & & & & 1\end{array}\right],& \mathrm{Eq}.\left(21\right)\end{array}$

where v_1,1, v_1,2, v_2,1 and v_2,2 are the four elements of V_pq. All of the other off-diagonal elements of T_pq are zeros. Equations (10j) and (10k) indicate that T_pq is a complex matrix containing complex values for v_2,1 and v_2,2. T_pq is also called a transformation matrix that performs the Jacobi rotation.

Matrix D_i is then updated as follows:

D_i+1=T_pq^H·D_i·T_pq. Eq. (22)

Equation (22) zeros out two off-diagonal elements d_p,q and d_q,p at locations (p,q) and (q,p), respectively, in D_i. The computation may alter the values of other off-diagonal elements in D_i.

Matrix V_i is also updated as follows:

V_i+1=V_i·T_pq  Eq. (23)

V_i may be viewed as a cumulative transformation matrix that contains all of the Jacobi rotation matrices T_pq used on D_i.

The Jacobi rotation matrix T_pq may also be expressed as a product of (1) a diagonal matrix with N−1 ones elements and one complex-valued element and (2) a real-valued matrix with N−2 ones along the diagonal, two real-valued diagonal elements, two real-valued off-diagonal elements, and zeros elsewhere. As an example, for p=1 and q=2, T_pq has the following form:

$\begin{array}{cc}\begin{array}{c}{\underset{\_}{T}}_{\mathrm{pq}}=\left[\begin{array}{ccccc}c& -s& 0& \dots & 0\\ g_1s& g_1c& 0& \dots & 0\\ 0& 0& 1& & ⋮\\ ⋮& ⋮& & \ddots & 0\\ 0& 0& \dots & 0& 1\end{array}\right]\\ =\left[\begin{array}{ccccc}1& 0& 0& \dots & 0\\ 0& g_1& 0& \dots & 0\\ 0& 0& 1& & ⋮\\ ⋮& ⋮& & \ddots & 0\\ 0& 0& \dots & 0& 1\end{array}\right]·\left[\begin{array}{ccccc}c& -s& 0& \dots & 0\\ s& c& 0& \dots & 0\\ 0& 0& 1& & ⋮\\ ⋮& ⋮& & \ddots & 0\\ 0& 0& \dots & 0& 1\end{array}\right],\end{array}& \mathrm{Eq}.\left(24\right)\end{array}$

where g₁ is a complex value and c and s are real values. The update of D_i in equation (22) may then be performed with 12(N−2)+8 real multiplies, and the update of V_i in equation (23) may be performed with 12N real multiplies. A total of 24N−16 real multiples are then performed for one iteration. Besides the T_pq structure, the number of multiplies to update D_i is reduced by the fact that D_i remains Hermitian after the update and that there is a 2×2 diagonal sub-matrix after the update with real-valued eigenvalues as the diagonal elements.

Each iteration of the Jacobi rotation zeros out two off-diagonal elements of D_i. Multiple iterations of the Jacobi rotation may be performed for different values of indices p and q to zero out all of the off-diagonal elements of D_i. The indices p and q may be selected in a predetermined manner by sweeping through all possible values.

A single sweep across all possible values for indices p and q may be performed as follows. The index p is stepped from 1 through N−1 in increments of one. For each value of p, the index q is stepped from p+1 through N in increments of one. An iteration of the Jacobi rotation to update D_i and V_i may be performed for each different combination of values for p and q. For each iteration, D_pq is formed based on the values of p and q and the current D_i for that iteration, V_pq is computed for D_pq as shown in equation set (10), T_pq is formed with V_pq as shown in equation (21) or (24), D_i is updated as shown in equation (22), and V_i is updated as shown in equation (23). For a given combination of values for p and q, the Jacobi rotation to update D and V, may also be skipped if the magnitude of the off-diagonal elements at locations (p,q) and (q,p) in D_i are below a predetermined threshold.

A sweep consists of N·(N−1)/2 iterations of the Jacobi rotation to update D_i and V_i for all possible values of p and q. Each iteration of the Jacobi rotation zeros out two off-diagonal elements of D_i but may alter other elements that might have been zeroed out earlier. The effect of sweeping through indices p and q is to reduce the magnitude of all off-diagonal elements of D_i, so that D_i approaches the diagonal matrix Λ. V_i contains an accumulation of all Jacobi rotation matrices that collectively give D_i. V_i thus approaches V as D_i approaches Λ.

Any number of sweeps may be performed to obtain more and more accurate approximations of V and Λ. Computer simulations have shown that four sweeps should be sufficient to reduce the off-diagonal elements of D_i to a negligible level, and three sweeps should be sufficient for most applications. A predetermined number of sweeps (e.g., three or four sweeps) may be performed. Alternatively, the off-diagonal elements of D_i may be checked after each sweep to determine whether D_i is sufficiently accurate. For example, the total error (e.g., the power in all off-diagonal elements of D_i) may be computed after each sweep and compared against an error threshold, and the iterative process may be terminated if the total error is below the error threshold. Other conditions or criteria may also be used to terminate the iterative process.

The values for indices p and q may also be selected in a deterministic manner. As an example, for each iteration i, the largest off-diagonal element of D_i is identified and denoted as d_p,q. The iteration is then performed with D_pq containing this largest off-diagonal element d_p,q and three other elements at locations (p,p), (q,p), and (q,q) in D_i. The iterative process may be performed until a termination condition is encountered. The termination condition may be, for example, completion of a predetermined number of iterations, satisfaction of the error criterion described above, or some other condition or criterion.

Upon termination of the iterative process, the final V_i is a good approximation of V, and the final D_i is a good approximation of Λ. The columns of V_i may be used as the eigenvectors of R, and the diagonal elements of D_i may be used as the eigenvalues of R. The eigenvalues in the final D_i are ordered from largest to smallest because the eigenvectors in V_pq for each iteration are ordered. The eigenvectors in the final V_i are also ordered based on their associated eigenvalues in D_i.

Except for the first iteration, the computation of T_pq and the updates of D_i and V_i do not have to proceed in a sequential order, assuming that the computations do not share the same hardware units. Since the updates of D_i and V_i involve matrix multiplies, it is likely that these updates will proceed in a sequential order using the same hardware. The computation of T_pq for the next iteration can start as soon as the off-diagonal elements of D_i have been updated for the current iteration. The computation of T_pq may be performed with dedicated hardware while V_i is updated. If the Jacobi rotation computation is finished by the time the V_i update is done, then the D_i update for the next iteration can start as soon as the V_i update for the current iteration is completed.

FIG. 5 shows an iterative process 500 for performing eigenvalue decomposition of an N×N Hermitian matrix R, where N>2. Matrices V_i and D_i are initialized as V₀=I and D₀=R, and index i is initialized as i=1 (block 510).

For iteration i, the values for indices p and q are selected in a predetermined manner (e.g., by stepping through all possible values for these indices) or a deterministic manner (e.g., by selecting the index values for the largest off-diagonal element) (block 512). A 2×2 matrix D_pq is then formed with four elements of matrix D_i at the locations determined by indices p and q (block 514). Eigenvalue decomposition of D_pq is then performed as shown in equation set (10) to obtain a 2×2 matrix V_pq of eigenvectors of D_pq (block 516). This eigenvalue decomposition may be efficiently performed as described above for FIGS. 1 through 4.

An N×N complex Jacobi rotation matrix T_pq is then formed based on matrix V_pq, as shown in equation (21) or (24) (block 518). Matrix D_i is then updated based on T_pq as shown in equation (22) (block 520). Matrix V_i is also updated based on T_pq, as shown in equation (23) (block 522).

A determination is then made whether to terminate the eigenvalue decomposition of R (block 524). The termination criterion may be based on the number of iterations or sweeps already performed, an error criterion, and so on. If the answer is ‘No’ for block 524, then index i is incremented (block 526), and the process returns to block 512 for the next iteration. Otherwise, if termination is reached, then matrix D_i is provided as an approximation of diagonal matrix Λ, and matrix V_i is provided as an approximation of matrix V of eigenvectors of R (block 528).

For a multi-carrier MIMO system (e.g., a MIMO system that utilizes OFDM), multiple channel response matrices H(k) may be obtained for different subbands. The iterative process may be performed for each channel response matrix H(k) to obtain matrices D_i (k) and V_i(k), which are approximations of diagonal matrix Λ(k) and matrix V(k) of eigenvectors, respectively, of R(k)=H^H(k)·H(k).

A high degree of correlation typically exists between adjacent subbands in a MIMO channel. This correlation may be exploited by the iterative process to reduce the amount of computation to derive D_i (k) and V_i(k) for all subbands of interest. For example, the iterative process may be performed for one subband at a time, starting from one end of the system bandwidth and traversing toward the other end of the system bandwidth. For each subband k except for the first subband, the final solution V_i(k−1) obtained for the prior subband k±1 may be used as an initial solution for the current subband k. The initialization for each subband k may be given as: V₀(k)=V_i(k−1) and D₀(k)=V₀^H(k)·R(k)·V₀(k). The iterative process then operates on the initial solutions of D₀(k) and V₀(k) for subband k until a termination condition is encountered.

The concept described above may also be used across time. For each time interval t, the final solution V_i(t−1) obtained for a prior time interval t−1 may be used as an initial solution for the current time interval t. The initialization for each time interval t may be given as: V₀(t)=V_i(t−1) and D₀(t)=V₀^H(t)·R(t)·V₀(t), where R(t)=H^H/(t)·H(t) and H(t) is the channel response matrix for time interval t. The iterative process then operates on the initial solutions of D₀(t) and V₀(t) for time interval t until a termination condition is encountered.

Singular Value Decomposition

The iterative process may also be used for singular value decomposition (SVD) of an arbitrary complex matrix H larger than 2×2. H has a dimension of R×T, where R is the number of rows and T is the number of columns. The iterative process for singular value decomposition of H may be performed in several manners.

In a first SVD embodiment, the iterative process derives approximations of the right singular vectors in V and the scaled left singular vectors in U·Σ. For this embodiment, a T×T matrix V_i is an approximation of V and is initialized as V₀=I. An R×T matrix W_i is an approximation of U·Σ and is initialized as W₀=H.

For the first SVD embodiment, a single iteration of the Jacobi rotation to update matrices V_i and W_i may be performed as follows. First, a 2×2 Hermitian matrix M_pq is formed based on the current W_i. M_pq is a 2×2 submatrix of W_i^H·W and contains four elements at locations (p,p), (p,q), (q,p) and (q,q) in W_i^H·W_i. The elements of M_pq may be computed as follows:

$\begin{array}{cc}{\underset{\_}{M}}_{\mathrm{pq}}=\left[\begin{array}{cc}{m}_{1,1}& m_{1,2}\\ m_{1,2}^{*}& m_{2,2}\end{array}\right],& \mathrm{Eq}.\left(25a\right)\\ m_{1,1}={{\underset{\_}{w}}_p}^2=\sum _{l=1}^Rw_{l,p}^{*}·w_{l,p},& \mathrm{Eq}.\left(25b\right)\\ m_{2,2}={{\underset{\_}{w}}_q}^2=\sum _{l=1}^Rw_{l,q}^{*}·w_{l,q},\mathrm{and}& \mathrm{Eq}.\left(25c\right)\\ m_{1,2}={\underset{\_}{w}}_p^H·{\underset{\_}{w}}_q=\sum _{l=1}^Rw_{l,p}^{*}·w_{l,q},& \mathrm{Eq}.\left(25d\right)\end{array}$

where w_p is column p of W_i, w_q is column q of W_i, and w_l,p is the element at location (l,p) in W_i. Indices p and q are such that pε{1, . . . , T}, qε{1, . . . , T}, and p#q. The values for indices p and q may be selected in a predetermined or deterministic manner, as described above.

Eigenvalue decomposition of M_pq is then performed as shown in equation set (10) to obtain a 2×2 unitary matrix V_pq of eigenvectors of M_pq. For this eigenvalue decomposition, R_2×2 is replaced with M_pq, and V_2×2 is provided as V_pq.

A T×T complex Jacobi rotation matrix T_pq is then formed with matrix V_pq. T_pq is an identity matrix with the four elements at locations (p,p), (p,q), (q,p) and (q,q) replaced with the (1,1), (1,2), (2,1) and (2,2) elements, respectively, of V_pq T_pq has the form shown in equations (21) and (24).

Matrix V_i is then updated as follows:

V_i+1=V_i·T_pq.  Eq. (26)

Matrix W_i is also updated as follows:

W_i+1=W_i·T_pq.  Eq. (27)

The iterative process is performed until a termination condition is encountered, which may be a predetermined number of sweeps or iterations, satisfaction of an error criterion, and so on. Upon termination of the iterative process, the final V_i is a good approximation of V, and the final W_i is a good approximation of U·Σ. When converged, W_i^H·W_i=Σ^T·Σ and U=W_i·Σ⁻¹, where “^T” denotes a transpose. For a square diagonal matrix, the final solution of Σ may be given as: {circumflex over (Σ)}=(W_i^H·W_i)^1/2. For a non-square diagonal matrix, the non-zero diagonal values of {circumflex over (Σ)} are given by the square roots of the diagonal elements of W_i^H·W_i. The final solution of U may be given as: Û=W_i·{circumflex over (Σ)}⁻¹.

The left singular vectors of H may be obtained by performing the first SVD embodiment and solving for scaled left singular vectors H·V=U·Σ and then normalizing. The left singular vectors of H may also be obtained by performing the iterative process for eigenvalue decomposition of H·H^H.

In a second SVD embodiment, the iterative process directly derives approximations of the right singular vectors in V and the left singular vectors in U. This SVD embodiment applies the Jacobi rotation on a two-sided basis to simultaneously solve for the left and right singular vectors. For the second SVD embodiment, a T×T matrix V_i is an approximation of V and is initialized as V₀=I. An R×R matrix U_i is an approximation of U and is initialized as U₀=I. An R×T matrix D_i is an approximation of Σ and is initialized as D₀=H.

For the second SVD embodiment, a single iteration of the Jacobi rotation to update matrices V_i, U_i, and D_i may be performed as follows. First, a 2×2 Hermitian matrix X_p1q1 is formed based on the current D_i. X_p1q1 is a 2×2 submatrix of D_i^H·D_i and contains four elements at locations (p₁,p₁), (p₁,q₁) (q₁,p₁) and (q₁,q₁) in D_i^H·D_i. The four elements of X_p1q1 may be computed as follows:

$\begin{array}{cc}{\underset{\_}{X}}_{p_1q_1}=\left[\begin{array}{cc}{x}_{1,1}& x_{1,2}\\ x_{1,2}^{*}& x_{2,2}\end{array}\right],& \mathrm{Eq}.\left(28a\right)\\ x_{1,1}={\underset{\_}{d}p_1}^2=\sum _{l=1}^Rd_{l,p_1}^{*}·d_{l,p_1},& \mathrm{Eq}.\left(28b\right)\\ x_{2,2}={{\underset{\_}{d}}_{q_1}}^2=\sum _{l=1}^Rd_{l,q_1}^{*}·d_{l,q_1},\mathrm{and}& \mathrm{Eq}.\left(28c\right)\\ x_{1,2}={\underset{\_}{d}}_{p_1}^H·{\underset{\_}{d}}_{q_1}=\sum _{l=1}^Rd_{l,p_1}^{*}·d_{l,q_1},& \mathrm{Eq}.\left(28d\right)\end{array}$

where d_p1 is column p₁ of D_i, d_q1 is column q₁ of D_i, and d_l,p1 is the element at location (l,p₁) in D_i. Indices p₁ and q₁ are such that p₁ε{1, . . . , T}, q₁ε{1, . . . , T}, and p₁≠q₁. Indices p_i and q_i may be selected in a predetermined or deterministic manner.

Eigenvalue decomposition of X_p1q1 is then performed as shown in equation set (10) to obtain a 2×2 matrix V_p1q1 of eigenvectors of X_p1q1. For this eigenvalue decomposition, R_2×2 is replaced with X_p1q1, and V_2×2 is provided as V_p1q1. A T×T complex Jacobi rotation matrix T_p1q1 is then formed with matrix V_p1q1 and contains the four elements of V_p1q1 at locations (p₁,p₁) (p₁,q₁) (q₁,p₁) and (q₁,q₁). T_p1q1 has the form shown in equations (21) and (24).

Another 2×2 Hermitian matrix Y_p2q2 is also formed based on the current D_i. Y_p2q2 is a 2×2 submatrix of D_i·D_i^H and contains elements at locations (p₂,p₂), (p₂,q₂), (q₂,p₂) and (q₂,q₂) in D_i·D_i^H. The elements of Y_p2q2 may be computed as follows:

$\begin{array}{cc}{\underset{\_}{Y}}_{p_2q_2}=\left[\begin{array}{cc}{y}_{1,1}& y_{1,2}\\ y_{1,2}^{*}& y_{2,2}\end{array}\right],& \mathrm{Eq}.\left(29a\right)\\ y_{1,1}={{\underset{\_}{\overset{~}{d}}}_{p_2}}^2=\sum _{l=1}^Td_{p_2,l}·d_{p_2,l}^{*},& \mathrm{Eq}.\left(29b\right)\\ y_{2,2}={{\underset{\_}{\overset{~}{d}}}_{q_2}}^2=\sum _{l=1}^Td_{q_2,l}·d_{q_2,l}^{*},\mathrm{and}& \mathrm{Eq}.\left(29c\right)\\ y_{1,2}={\underset{\_}{\overset{~}{d}}}_{p_2}·{\underset{\_}{\overset{~}{d}}}_{q_2}^H=\sum _{l=1}^Td_{p_2,l}·d_{q_2,l}^{*},& \mathrm{Eq}.\left(29d\right)\end{array}$

where {tilde over (d)}_p2 is row p₂ of D_i, {tilde over (d)}_q2 is row q₂ of D_i, and d_p2,l is the element at location (p₂, l) in D_i. Indices p₂ and q₂ are such that p₂ε{1, . . . , R}, q₂ε{1, . . . , R}, and p₂≠q₂. Indices p₂ and q₂ may also be selected in a predetermined or deterministic manner.

Eigenvalue decomposition of Y_p2q2 is then performed as shown in equation set (10) to obtain a 2×2 matrix U_p2q2 of eigenvectors of Y_p2q2. For this eigenvalue decomposition, R_2×2 is replaced with Y_p2q2, and V_2×2 is provided as U_p2q2. An R×R complex Jacobi rotation matrix S_p2q2 is then formed with matrix U_p2q2 and contains the four elements of U_p2q2 at locations (p₂,p₂), (p₂,q₂), (q₂,p₂) and (q₂,q₂). S_p2q2 has the form shown in equations (21) and (24).

Matrix V_i is then updated as follows:

V_i+1=V_i·T_p1q1.  Eq. (30)

Matrix U_i is updated as follows:

U_i+1=U_i·S_p2q2.  Eq. (31)

Matrix D_i is updated as follows:

D_i+1=S_p2q2^H·D_i·T_p1q1.  Eq. (32)

The iterative process is performed until a termination condition is encountered. Upon termination of the iterative process, the final V_i is a good approximation of {tilde over (V)}, the final U_i is a good approximation of U, and the final D_i is a good approximation of {tilde over (Σ)}, where {tilde over (V)} and {tilde over (Σ)} may be rotated versions of V and Σ, respectively. V_i and D_i may be unrotated as follows:

{circumflex over (Σ)}=D_i·P,  and Eq. (33a)

{circumflex over (V)}=V_i·P,  Eq. (33b)

where P is a T×T diagonal matrix with diagonal elements having unit magnitude and phases that are the negative of the phases of the corresponding diagonal elements of D_i. {circumflex over (Σ)} and {circumflex over (V)} are the final approximations of Σand V, respectively.

For a multi-carrier MIMO system, the iterative process may be performed for each channel response matrix H(k) to obtain matrices V_i(k), U_i(k), and D_i(k), which are approximations matrices V(k), U(k), and Σ(k), respectively, for that H(k). For the first SVD embodiment, for each subband k except for the first subband, the final solution V_i(k−1) obtained for the prior subband k−1 may be used as an initial solution for the current subband k, so that V₀(k)=V_i(k−1) and W₀(k)=H(k)·V₀(k). For the second SVD embodiment, for each subband k except for the first subband, the final solutions V_i(k−1) and U_i(k−1) obtained for the prior subband k−1 may be used as initial solutions for the current subband k, so that V₀(k)=V_i(k−1), U₀(k)=U_i(k−1), and D₀(k)=U₀^H(k)·H(k)·V₀(k). The concept may also be used across time or both frequency and time, as described above.

System

FIG. 6 shows a block diagram of an embodiment of an access point 610 and a user terminal 650 in a MIMO system 600. Access point 610 is equipped with N_ap antennas and user terminal 650 is equipped with N_ut antennas, where N_ap>1 and N_ut>1. For simplicity, the following description assumes that MIMO system 600 uses time division duplexing (TDD), and the downlink channel response matrix H_dn(k) for each subband k is reciprocal of the uplink channel response matrix H_up(k) for that subband, or H_dn(k)=H(k) and H_up(k)=H^T(k).

On the downlink, at access point 610, a transmit (TX) data processor 614 receives traffic data from a data source 612 and other data from a controller 630. TX data processor 614 formats, encodes, interleaves, and modulates the data and generates data symbols, which are modulation symbols for data. A TX spatial processor 620 multiplexes the data symbols with pilot symbols, performs spatial processing with eigenvectors or right singular vectors if applicable, and provides N_ap streams of transmit symbols. Each transmitter unit (TMTR) 622 processes a respective transmit symbol stream and generates a downlink modulated signal. N_ap downlink modulated signals from transmitter units 622a through 622ap are transmitted from antennas 624a through 624ap, respectively.

At user terminal 650, N_ut antennas 652a through 652ut receive the transmitted downlink modulated signals, and each antenna provides a received signal to a respective receiver unit (RCVR) 654. Each receiver unit 654 performs processing complementary to the processing performed by transmitter units 622 and provides received symbols. A receive (RX) spatial processor 660 performs spatial matched filtering on the received symbols from all receiver units 654a through 654ut and provides detected data symbols, which are estimates of the data symbols transmitted by access point 610. An RX data processor 670 processes (e.g., symbol demaps, deinterleaves, and decodes) the detected data symbols and provides decoded data to a data sink 672 and/or a controller 680.

A channel estimator 678 processes received pilot symbols and provides an estimate of the downlink channel response, Ĥ(k), for each subband of interest. Controller 680 may decompose each matrix Ĥ(k) to obtain {circumflex over (V)}(k) and {circumflex over (Σ)}(k), which are estimates of V(k) and Σ(k) for H(k). Controller 680 may derive a downlink spatial filter matrix M_dn(k) for each subband of interest based on {circumflex over (V)}(k), as shown in Table 1. Controller 680 may provide M_dn(k) to RX spatial processor 660 for downlink spatial matched filtering and {circumflex over (V)}(k) to a TX spatial processor 690 for uplink spatial processing.

The processing for the uplink may be the same or different from the processing for the downlink. Data from a data source 686 and signaling from controller 680 are processed (e.g., encoded, interleaved, and modulated) by a TX data processor 688, multiplexed with pilot symbols, and further spatially processed by TX spatial processor 690 with {circumflex over (V)}(k) for each subband of interest. The transmit symbols from TX spatial processor 690 are further processed by transmitter units 654a through 654ut to generate N_ut uplink modulated signals, which are transmitted via antennas 652a through 652ut.

At access point 610, the uplink modulated signals are received by antennas 624a through 624ap and processed by receiver units 622a through 622ap to generate received symbols for the uplink transmission. An RX spatial processor 640 performs spatial matched filtering on the received data symbols and provides detected data symbols. An RX data processor 642 further processes the detected data symbols and provides decoded data to a data sink 644 and/or controller 630.

A channel estimator 628 processes received pilot symbols and provides an estimate of either H^T(k) or U(k) for each subband of interest, depending on the manner in which the uplink pilot is transmitted. Controller 630 may receive Ĥ^T(k) for multiple subbands and decompose each matrix Ĥ^T(k) to obtain Û(k). Controller 680 may also derive an uplink spatial filter matrix M_up(k) for each subband of interest based on Û(k). Controller 630 provides M_up(k) to RX spatial processor 640 for uplink spatial matched filtering and Û(k) to TX spatial processor 620 for downlink spatial processing.

Controllers 630 and 680 control the operation at access point 610 and user terminal 650, respectively. Memory units 632 and 682 store data and program codes used by controllers 630 and 680, respectively. Controllers 630 and/or 680 may perform eigenvalue decomposition and/or singular value decomposition of the channel response matrices obtained for its link.

The decomposition techniques described herein may be implemented by various means. For example, these techniques may be implemented in hardware, software, or a combination thereof. For a hardware implementation, the processing units used to perform decomposition may be implemented within one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, micro-controllers, microprocessors, other electronic units designed to perform the functions described herein, or a combination thereof.

For a software implementation, the decomposition techniques may be implemented with modules (e.g., procedures, functions, and so on) that perform the functions described herein. The software codes may be stored in a memory unit (e.g., memory unit 632 or 682 in FIG. 6 and executed by a processor (e.g., controller 630 or 680). The memory unit may be implemented within the processor or external to the processor.

Headings are included herein for reference and to aid in locating certain sections. These headings are not intended to limit the scope of the concepts described therein under, and these concepts may have applicability in other sections throughout the entire specification.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A computer-program storage apparatus for decomposing a matrix comprising a memory unit having software codes stored thereon, the software codes being executable by one or more processors and the software codes comprising:

code for deriving a first set of at least one variable based on a first matrix to be decomposed and using Coordinate Rotational Digital Computer (CORDIC) computation;

code for deriving a second set of at least one variable based on the first matrix and using a look-up table; and

code for deriving a second matrix of eigenvectors based on the first and second sets of at least one variable.

2. The computer-program storage apparatus of claim 1, wherein the code for deriving the first set of at least one variable comprises:

code for performing CORDIC computation on an element of the first matrix to determine magnitude and phase of the element; and

code for performing CORDIC computation on the phase of the element to determine sine and cosine of the element, and

wherein the first set of at least one variable comprises the sine and cosine of the element.

3. The computer-program storage apparatus of claim 1, wherein the code for deriving the second set of at least one variable comprises:

code for deriving intermediate quantities based on the first matrix; and

code for deriving the second set of at least one variable based on the intermediate quantities and using the look-up table.

4. The computer-program storage apparatus of claim 3, wherein the code for deriving the intermediate quantities comprises:

code for deriving a divisor for an intermediate variable based on a first element of the first matrix;

code for converting the divisor into a first binary floating-point number;

code for deriving a dividend for the intermediate variable based on second and third elements of the first matrix;

code for converting the dividend into a second binary floating-point number; and

code for forming the intermediate quantities based on the first and second floating-point numbers.

5. A computer-program storage apparatus for decomposing a matrix comprising a memory unit having software codes stored thereon, the software codes being executable by one or more processors and the software codes comprising:

code for deriving intermediate quantities based on a first matrix to be decomposed;

code for deriving at least one variable based on the intermediate quantities and using a look-up table; and

code for deriving a second matrix of eigenvectors based on the at least one variable.

6. The computer-program storage apparatus of claim 5, wherein the code for deriving the intermediate quantities based on the first matrix comprises:

code for deriving a divisor for an intermediate variable based on a first element of the first matrix;

code for converting the divisor into a first binary floating-point number;

code for deriving a dividend for the intermediate variable based on second and third elements of the first matrix;

code for converting the dividend into a second binary floating-point number; and

code for forming the intermediate quantities based on the first and second floating-point numbers.

7. The computer-program storage apparatus of claim 6, wherein the code for forming the intermediate quantities based on the first and second floating-point numbers comprises:

code for deriving a first intermediate quantity based on a mantissa of the first floating-point number;

code for deriving a second intermediate quantity based on a mantissa of the second floating-point number; and

code for deriving a third intermediate quantity based on exponents of the first and second floating-point numbers.

8. The computer-program storage apparatus of claim 7, wherein the code for deriving the at least one variable comprises:

code for forming an input address for the look-up table based on the first, second, and third intermediate quantities; and

code for accessing the look-up table with the input address.

9. A computer-program storage apparatus for decomposing a matrix comprising a memory unit having software codes stored thereon, the software codes being executable by one or more processors and the software codes comprising:

code for performing Coordinate Rotational Digital Computer (CORDIC) computation on an element of a first matrix to determine magnitude and phase of the element;

code for performing CORDIC computation on the phase of the element to determine sine and cosine of the element; and

code for deriving a second matrix of eigenvectors based on the magnitude, sine, and cosine of the element.

10. The computer-program storage apparatus of claim 9, wherein the CORDIC computation on the element to determine the magnitude and phase of the element and the CORDIC computation on the phase of the element to determine sine and cosine of the element are performed in parallel.

11. The computer-program storage apparatus of claim 9, wherein the CORDIC computation on the element to determine the magnitude and phase of the element and the CORDIC computation on the phase of the element to determine sine and cosine of the element are performed for a predetermined number of iterations.

12. A computer-program storage apparatus for decomposing a matrix comprising a memory unit having software codes stored thereon, the software codes being executable by one or more processors and the software codes comprising:

code for performing a plurality of iterations of Jacobi rotation on a first matrix of complex values with a plurality of Jacobi rotation matrices, each Jacobi rotation matrix being derived by performing eigenvalue decomposition of a correlation submatrix using Coordinate Rotational Digital Computer (CORDIC) computation, a look-up table, or both the CORDIC computation and the look-up table; and

code for deriving a first unitary matrix with orthogonal vectors based on the plurality of Jacobi rotation matrices.

13. The computer-program storage apparatus of claim 12, wherein the code for performing the plurality of iterations of the Jacobi rotation comprises, for each iteration:

code for forming the submatrix based on the first matrix;

code for decomposing the submatrix using the CORDIC computation, the look-up table, or both the CORDIC computation and the look-up table to obtain eigenvectors of the submatrix;

code for forming a Jacobi rotation matrix with the eigenvectors; and

code for updating the first matrix with the Jacobi rotation matrix.

14. The computer-program storage apparatus of claim 13, wherein the code for decomposing the submatrix comprises:

code for deriving a first set of at least one variable based on the submatrix and using the CORDIC computation;

code for deriving a second set of at least one variable based on the submatrix and using the look-up table; and

code for deriving the eigenvectors of the submatrix based on the first and second sets of at least one variable.

15. The computer-program storage apparatus of claim 12, further comprising:

code for deriving a diagonal matrix of eigenvalues based on the plurality of Jacobi rotation matrices.

16. The computer-program storage apparatus of claim 12, further comprising:

code for deriving a second matrix of complex values based on the plurality of Jacobi rotation matrices; and

code for deriving a second unitary matrix with orthogonal vectors based on the second matrix.

17. The computer-program storage apparatus of claim 16, further comprising:

code for deriving a diagonal matrix of singular values based on the second matrix.

18. The computer-program storage apparatus of claim 12, further comprising:

code for deriving a second unitary matrix with orthogonal vectors based on the plurality of Jacobi rotation matrices.

19. The computer-program storage apparatus of claim 18, further comprising:

code for deriving a diagonal matrix of singular values based on the plurality of Jacobi rotation matrices.

20. A computer-program storage apparatus for decomposing a matrix comprising a memory unit having software codes stored thereon, the software codes being executable by one or more processors and the software codes comprising:

code for obtaining a plurality of matrices of complex values for a plurality of transmission spans;

code for performing a plurality of iterations of Jacobi rotation on a first matrix of complex values for a first transmission span to obtain a first unitary matrix with orthogonal vectors, wherein each iteration of the Jacobi rotation utilizes eigenvalue decomposition of a correlation submatrix using Coordinate Rotational Digital Computer (CORDIC) computation, a look-up table, or both the CORDIC computation and the look-up table; and

code for performing a plurality of iterations of the Jacobi rotation on a second matrix of complex values for a second transmission span to obtain a second unitary matrix with orthogonal vectors, wherein the first unitary matrix is used as an initial solution for the second unitary matrix, wherein the first and second matrices are among the plurality of matrices, and wherein the first and second transmission spans are among the plurality of transmission spans.

21. The computer-program storage apparatus of claim 20, further comprising:

for each remaining one of the plurality of matrices of complex values, code for performing a plurality of iterations of the Jacobi rotation on the matrix of complex values to obtain a unitary matrix with orthogonal vectors, wherein another unitary matrix obtained for another one of the plurality of matrices is used as an initial solution for the unitary matrix.

22. The computer-program storage apparatus of claim 20, further comprising:

code for selecting the plurality of matrices in sequential order for decomposition.

23. The computer-program storage apparatus of claim 20, wherein the plurality of transmission spans correspond to a plurality of frequency subbands in a multi-carrier communication system.

24. The computer-program storage apparatus of claim 20, wherein the plurality of transmission spans correspond to a plurality of time intervals.

25. The computer-program storage apparatus of claim 20, wherein the plurality of matrices of complex values are a plurality of channel response matrices for a plurality of frequency subbands.