I. Field
The present invention relates generally to communication, and more specifically to techniques for decomposing matrices.
II. Background
A multiple-input multiple-output (MIMO) communication system employs multiple (T) transmit antennas at a transmitting station and multiple (R) receive antennas at a receiving station 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.
A 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 station uses the right singular vectors or the eigenvectors to transmit data on the S eigenmodes. The receiving station uses the left singular vectors or the eigenvectors to receive data transmitted on the S eigenmodes.
Eigenvalue decomposition and singular value decomposition are very computationally intensive. There is therefore a need in the art for techniques to efficiently decompose matrices.
Techniques for efficiently decomposing matrices using Jacobi rotation are described herein. These techniques may be used for eigenvalue decomposition of a Hermitian matrix of complex values to obtain a matrix of eigenvectors and a matrix of eigenvalues for the Hermitian matrix. The techniques may also be used for singular value decomposition of an arbitrary matrix of complex values to obtain a matrix of left singular vectors, a matrix of right singular vectors, and a matrix of singular values for the arbitrary matrix.
In an embodiment, multiple iterations of Jacobi rotation are performed on a first matrix of complex values with multiple Jacobi rotation matrices of complex values to zero out the off-diagonal elements in the first matrix. The first matrix may be a channel response matrix H, a correlation matrix of H, which is R, or some other matrix. For each iteration, a submatrix may be formed based on the first matrix and decomposed to obtain eigenvectors for the submatrix, and a Jacobi rotation matrix may be formed with the eigenvectors and used to update the first matrix. A second matrix of complex values is derived based on the Jacobi rotation matrices. The second matrix contains orthogonal vectors and may be a matrix Vi of right singular vectors of H or eigenvectors of R.
For eigenvalue decomposition, a third matrix Di of eigenvalues may be derived based on the Jacobi rotation matrices. For singular value decomposition (SVD) based on a first SVD embodiment, a third matrix Wi of complex values may be derived based on the Jacobi rotation matrices, a fourth matrix Û with orthogonal vectors may be derived based on the third matrix Wi, and a matrix {circumflex over (Σ)} of singular values may also be derived based on the third matrix Wi. For singular value decomposition based on a second SVD embodiment, a third matrix Ui with orthogonal vectors and a matrix {circumflex over (Σ)} of singular values may be derived based on the Jacobi rotation matrices.
Various aspects and embodiments of the invention are described in further detail below.
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 matrix decomposition techniques described herein may be used for various communication systems such as a single-carrier communication system with a single frequency subband, a multi-carrier communication system with multiple subbands, a single-carrier frequency division multiple access (SC-FDMA) system with multiple subbands, and other communication systems. Multiple subbands may be obtained with orthogonal frequency division multiplexing (OFDM), some other modulation techniques, or some other construct. OFDM partitions the overall system bandwidth into multiple (K) orthogonal subbands, which are also called tones, subcarriers, bins, and so on. With OFDM, each subband is associated with a respective subcarrier that may be modulated with data. An SC-FDMA system may utilize interleaved FDMA (IFDMA) to transmit on subbands that are distributed across the system bandwidth, localized FDMA (LFDMA) to transmit on a block of adjacent subbands, or enhanced FDMA (EFDMA) to transmit on multiple blocks of adjacent subbands. In general, modulation symbols are sent in the frequency domain with OFDM and in the time domain with SC-FDMA. For clarity, much of the following description is for a MIMO system with a single subband.
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:
where entry hi,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, for example, performing either singular value decomposition of H or eigenvalue decomposition of a correlation matrix of H.
The eigenvalue decomposition may be expressed as:
R=HH·H=V·Λ·VH, 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;
Λ is a T×T diagonal matrix of eigenvalues of R; and
“H” denotes a conjugate transpose.
The unitary matrix V is characterized by the property VH·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, . . . , λS} and represent the power gains for the S eigenmodes. R is a Hermitian matrix whose off-diagonal elements have the following property: ri,j=r*j,i, where “*” denotes a complex conjugate.
The singular value decomposition may be expressed as:
H=U·Σ·VH, 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, . . . , σS)} 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 σi=√{square root over (λi)} for i=1, . . . , S.
A transmitting station may use the right singular vectors in V to transmit data on the eigenmodes of H. Transmitting data on eigenmodes typically provides better performance than simply transmitting data from the T transmit antennas without any spatial processing. A receiving station may use the left singular vectors in U or the eigenvectors in V to receive the data transmission sent on the eigenmodes of H. Table 1 shows the spatial processing performed by the transmitting station, the received symbols at the receiving station, and the spatial processing performed by the receiving station. 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.
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 and Jacobi transformation. The Jacobi rotation zeros out a pair of off-diagonal elements of the complex 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 2×2 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 2×2 Hermitian matrix R2×2 may be performed as follows. The Hermitian matrix R2×2 may be expressed as:
where A, B, and D are arbitrary real values, and θb is an arbitrary phase.
The first step of the eigenvalue decomposition of R2×2 is to apply a two-sided unitary transformation, as follows:
where Rre is a symmetric real matrix containing real values and having symmetric off-diagonal elements at locations (1, 2) and (2, 1).
The synmetric real matrix Rre is then diagonalized using a two-sided Jacobi rotation, as follows:
where angle φ may be expressed as:
A 2×2 unitary matrix V2×2 of eigenvectors of R2×2 may be derived as:
The two eigenvalues λ1 and λ2 may be derived based on equation (6), or based on the equation Λ2×2=V2×2H·R2×2·V2×2, as follows:
In equation set (9), the ordering of the two eigenvalues is not fixed, and λ1 may be larger than smaller than λ2. However, if angle φ is constrained such that |2φ|≦π/2, then cos 2φ≧0, and sin 2φ>0 if and only if D>A. Thus, the ordering of the two eigenvalues may be determined by the relative magnitudes of A and D. λ1 is the larger eigenvalue if A>D, and λ2 is the larger eigenvalue if D>A. If A=D, then sin 2φ=1 and λ2 is the larger eigenvalue. If λ2 is the larger eigenvalue, then the two eigenvalues Λ2×2 may be swapped to maintain a predetermined ordering of largest to smallest eigenvalues, and the first and second columns of V2×2 may also be swapped correspondingly. Maintaining this predetermined ordering for the two eigenvectors in V2×2 results in the eigenvectors of a larger size matrix decomposed using V2×2 to be ordered from largest to smallest eigenvalues, which is desirable.
The two eigenvalues λ1 and λ2 may also be computed directly from the elements of Rre, as follows:
Equation (10) is the solution to a characteristic equation of R2×2. In equation (10), 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.
Equation (8) requires the computation of cos φ and sin φ to derive the elements of V2×2. The computation of cos φ and sin φ is complex. The elements of V2×2 may be computed directly from the elements of R2×2, as follows:
where r1,1, r1,2 and r2,1 are elements of R2×2, and r is the magnitude of r1,2. Since g1 is a complex value, V2×2 contains complex values in the second row.
Equation set (11) is designed to reduce the amount of computation to derive V2×2 from R2×2 For example, in equations (11c), (11d), and (11f), division by r is required. Instead, r is inverted to obtain r1, and multiplication by r1 is performed for equations (11c), (11d), and (11f). This reduces the number of divide operations, which are computationally more expensive than multiplies. Also, instead of computing the argument (phase) of the complex element r1,2, which requires an arctangent operation, and then computing the cosine and sine of this phase value to obtain c1 and s1, various trigonometric identities are used to solve for c1 and s1 as a function of the real and imaginary parts of r1,2 and using only a square root operation. Furthermore, instead of computing the arctangent in equation (7) and the sine and cosine functions in equation (8), other trigonometric identities are used to solve for c and s as functions of the elements of R2×2.
Equation set (11) performs a complex Jacobi rotation on R2×2 to obtain V2×2. The set of computations in equation set (11) is designed to reduce the number of multiply, square root, and invert operations required to derive V2×2. This can greatly reduce computational complexity for decomposition of a larger size matrix using V2×2.
The eigenvalues of R2×2 may be computed as follows:
1. 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 the 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 submatrices of the N×N Hermitian matrix and are repeatedly applied to diagonalize the N×N Hermitian matrix. Each unitary transformation matrix contains four non-trivial elements (i.e., elements other than 0 or 1) that are derived from elements of a corresponding 2×2 Hermitian submatrix. The transformation matrices are also called Jacobi rotation matrices. After completing all of the Jacobi rotation, 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 Di is an approximation of the diagonal matrix Λ of eigenvalues of R and is initialized as D0=R. An N×N matrix Vi is an approximation of the matrix V of eigenvectors of R and is initialized as V0=I.
A single iteration of the Jacobi rotation to update matrices Di and Vi may be performed as follows. First, a 2×2 Hermitian matrix Dpq is formed based on the current Di, as follows:
where dp,q is the element at location (p,q) in Di; and
p ε {1, . . . , N}, q ε {1, . . . , N}, and p≠q.
Dpq is a 2×2 submatrix of Di, and the four elements of Dpq are four elements at locations (p,p), (p,q), (q,p) and (q,q) in Di. The values for indices p and q may be selected in various manners, as described below.
Eigenvalue decomposition of Dpq is then performed, e.g., as shown in equation set (11), to obtain a 2×2 unitary matrix Vpq of eigenvectors of Dpq. For the eigenvalue decomposition of Dpq, R2×2 in equation (4) is replaced with Dpq, and V2×2 from equation (111) is provided as Vpq.
An N×N complex Jacobi rotation matrix Tpq is then formed with matrix Vpq. Tpq 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 Vpq. Tpq has the following form:
where v1,1, v1,2, v2,1 and v2,2 are the four elements of Vpq. All of the other off-diagonal elements of Tpq are zeros. Equation (111) indicates that Tpq is a complex matrix containing complex values for v2,1 and v2,2. Tpq is also called a transformation matrix that performs the Jacobi rotation.
Matrix Di is then updated as follows:
Di+1=TpqH·Di·Tpq. Eq (15)
Equation (15) zeros out two off-diagonal elements dp,q and dq,p at locations (p,q) and (q,p), respectively, in Di. The computation may alter the values of the other off-diagonal elements in Di.
Matrix Vi is also updated as follows:
Vi+1=Vi·Tpq. Eq (16)
Vi may be viewed as a cumulative transformation matrix that contains all of the Jacobi rotation matrices Tpq used on Di.
Each iteration of the Jacobi rotation zeros out two off-diagonal elements of Di. 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 Di. 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 may be stepped from 1 through N−1 in increments of one. For each value of p, the index q may be stepped from p+1 through N in increments of one. An iteration of the Jacobi rotation to update Di and Vi may be performed for each different combination of values for p and q. For each iteration, Dpq is formed based on the values of p and q and the current Di for that iteration, Vpq is computed for Dpq as shown in equation set (11), Tpq is formed with Vpq as shown in equation (14), Di is updated as shown in equation (15), and Vi is updated as shown in equation (16). For a given combination of values for p and q, the Jacobi rotation to update Di and Vi may be skipped if the magnitude of the off-diagonal elements at locations (p,q) and (q,p) in Di is below a predetermined threshold.
A sweep consists of N·(N−1)/2 iterations of the Jacobi rotation to update Di and Vi for all possible values of p and q. Each iteration of the Jacobi rotation zeros out two off-diagonal elements of Di 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 Di, so that Di approaches the diagonal matrix Λ. Vi contains an accumulation of all Jacobi rotation matrices that collectively gives Di. Thus, Vi approaches V as Di 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 Di 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 Di may be checked after each sweep to determine whether Di is sufficiently accurate. For example, the total error (e.g., the power in all off-diagonal elements of Di) 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 Di may be identified and denoted as dp,q. Jacobi rotation may then be performed with Dpq containing this largest off-diagonal element dp,q and three other elements at locations (p,p), (q,p), and (q,q) in Di. 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 Vi is a good approximation of V, and the final Di is a good approximation of Λ. The columns of Vi may be provided as the eigenvectors of R, and the diagonal elements of Di may be provided as the eigenvalues of R. The eigenvalues in the final Di are ordered from largest to smallest because the eigenvectors in Vpq for each iteration are ordered. The eigenvectors in the final Vi are also ordered based on their associated eigenvalues in Di.
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 112). A 2×2 matrix Dpq is then formed with four elements of matrix Di at the locations determined by indices p and q (block 114). Eigenvalue decomposition of Dpq is then performed, e.g., as shown in equation set (11), to obtain a 2x2 matrix Vpq of eigenvectors of Dpq (block 116). An N×N complex Jacobi rotation matrix Tpq is then formed based on matrix Vpq, as shown in equation (14) (block 118). Matrix Di is then updated based on Tpq, as shown in equation (15) (block 120). Matrix Vi is also updated based on Tpq, as shown in equation (16) (block 122).
A determination is then made whether to terminate the eigenvalue decomposition of R (block 124). 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 124, then index i is incremented (block 126), and the process returns to block 112 for the next iteration. Otherwise, if termination is reached, then matrix Di is provided as an approximation of diagonal matrix Λ, and matrix Vi is provided as an approximation of matrix V of eigenvectors of R (block 128).
For a MIMO system with multiple subbands (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 Di(k) and Vi(k), which are approximations of diagonal matrix Λ(k) and matrix V(k) of eigenvectors, respectively, of R(k)=HH(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 Di (k) and Vi(k) for the 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 Vi(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)=Vi(k−1) and D0(k)=V0H(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 Vi(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)=Vi(t−1) and D0(t)=V0H(t)·R(t)·V0(t), where R(t)=HH(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. The concept may also be used across both frequency and time. For each subband in each time interval, the final solution obtained for a prior subband and/or the final solution obtained for a prior time interval may be used as an initial solution for the current subband and time interval.
2. Singular Value Decomposition
The iterative process may also be used for singular value decomposition of an arbitrary complex matrix H that is larger than 2×2. The singular value decomposition of H is given as H=U·Σ·VH. The following observations may be made regarding H. First, matrix R=HH·H and matrix {tilde over (R)}=H·HH are both Hermitian matrices. Second, right singular vectors of H, which are the columns of V, are also eigenvectors of R. Correspondingly, left singular vectors of H, which are the columns of U, are also eigenvectors of {tilde over (R)}. Third, the non-zero eigenvalues of R are equal to the non-zero eigenvalues of {tilde over (R)}, and are the square of corresponding singular values of H.
A 2×2 matrix H2×2 of complex values may be expressed as:
where h1 is a 2×1 vector with the elements in the first column of H2×2; and
h2 is a 2×1 vector with the elements in the second column of H2×2.
The right singular vectors of H2×2 are the eigenvectors of H2×2H·H2×2 and may be computed using the eigenvalue decomposition described above in equation set (11). A 2×2 Hermitian matrix R2×2 is defined as R2×2=H2×2H·H2×2, and the elements of R2×2 may be computed based on the elements of H2×2, as follows:
For Hermitian matrix R2×2, r2,1 does not need to be computed since r2,1=r*1,2 Equation set (11) may be applied to R2×2 to obtain a matrix V2×2. V2×2 contains the eigenvectors of R2×2, which are also the right singular vectors of H2×2.
The left singular vectors of H2×2 are the eigenvectors of H2×2·H2×2 and may also be computed using the eigenvalue decomposition described above in equation set (11). A 2×2 Hermitian matrix {tilde over (R)}2×2 is defined as {tilde over (R)}2×2=H2×2·H2×2H, and the elements of {tilde over (R)}2×2 may be computed based on the elements of H2×2, as follows:
Equation set (11) may be applied to {tilde over (R)}2×2 to obtain a matrix {tilde over (V)}2×2. {tilde over (V)}2×2 contains the eigenvectors of {tilde over (R)}2×2, which are also the left singular vectors of H2×2.
The iterative process described above for eigenvalue decomposition of an N×N Hermitian matrix R may be used for singular value decomposition 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 (SVD) 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 Vi is an approximation of V and is initialized as V0=I. An R×T matrix Wi 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 Vi and Wi may be performed as follows. First, a 2×2 Hermitian matrix Mpq is formed based on the current Wi. Mpq is a 2×2 submatrix of WiH·Wi and contains four elements at locations (p,p), (p,q), (q,p) and (q,q) in WiH·Wi. The elements of Mpq may be computed as follows:
where wp is column p of Wi, wq is column q of Wi, and wl,p is the element at location (l,p) in Wi. Indices p and q are such that p ε {1, . . . , T}, q ε {1, . . . , T}, and p≠q. The values for indices p and q may be selected in various manners, as described below.
Eigenvalue decomposition of Mpq is then performed, e.g., as shown in equation set (11), to obtain a 2×2 unitary matrix Vpq of eigenvectors of Mpq, For this eigenvalue decomposition, R2×2 is replaced with Mpq, and V2×2 is provided as Vpq.
A T×T complex Jacobi rotation matrix Tpq is then formed with matrix Vpq. Tpq 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 Vpq. Tpq has the form shown in equation (14).
Matrix Vi is then updated as follows:
Vi+1=Vi·Tpq. Eq (21)
Matrix Wi is also updated as follows:
Wi+1=Wi·Tpq. Eq (22)
For the first SVD embodiment, the iterative process repeatedly zeros out off-diagonal elements of WiH·Wi without explicitly computing HH·H. The indices p and q may be swept by stepping p from 1 through T−1 and, for each value of p, stepping q from p+1 through T. Alternatively, the values of p and q for which |wpH·wq| is largest may be selected for each iteration. The iterative process is performed until a termination condition is encountered, which may be a predetermined number of sweeps, a predetermined number of iterations, satisfaction of an error criterion, and so on.
Upon termination of the iterative process, the final Vi is a good approximation of V, and the final Wi is a good approximation of U·Σ. When converged, WiH·WiΣT·Σ and U=Wi·Σ−1, where “T” denotes a transpose. For a square diagonal matrix, the final solution of Σ may be given as: {circumflex over (Σ)}=(W1H·Wi)1/2. For a non-square diagonal matrix, the non-zero diagonal elements of {circumflex over (Σ)} are given by the square roots of the diagonal elements of WiH·Wi. The final solution of U may be given as: Û=Wi·{circumflex over (Σ)}−1.
For iteration i, the values for indices p and q are selected in a predetermined or deterministic manner (block 212). A 2×2 matrix Mpq is then formed with four elements of matrix Wi at the locations determined by indices p and q as shown in equation set (20) (block 214). Eigenvalue decomposition of Mpq is then performed, e.g., as shown in equation set (11), to obtain a 2×2 matrix Vpq of eigenvectors of Mpq (block 216). A T×T complex Jacobi rotation matrix Tpq is formed based on matrix Vpq, as shown in equation (14) (block 218). Matrix Vi is then updated based on Tpq, as shown in equation (21) (block 220). Matrix Wi is also updated based on Tpq, as shown in equation (22) (block 222).
A determination is then made whether to terminate the singular value decomposition of H (block 224). 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 224, then index i is incremented (block 226), and the process returns to block 212 for the next iteration. Otherwise, if termination is reached, then post processing is performed on Wi to obtain {circumflex over (Σ)} and Û (block 228). Matrix Vi is provided as an approximation of matrix V of right singular vectors of H, matrix Û is provided as an approximation of matrix U of left singular vectors of H, and matrix {circumflex over (Σ)} is provided as an approximation of matrix Σ of singular values of H (block 230)
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·HH.
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 an arbitrary complex 2×2 matrix H2×2=[h1 h2], the conjugate transpose of this matrix is H2×2H=[{tilde over (h)}1 {tilde over (h)}2], where {tilde over (h)}1 and {tilde over (h)}2 are the two columns of H2×2H and are also the complex conjugates of the rows of H2×2. The left singular vectors of H2×2 are also the right singular vectors of H2×2H. The right singular vectors of H2×2 may be computed using Jacobi rotation, as described above for equation set (18). The left singular vectors of H2×2 may be obtained by computing the right singular vectors of H2×2H using Jacobi rotation, as described above for equation set (19).
For the second SVD embodiment, a T×T matrix Vi is an approximation of V and is initialized as V0=I. An R×R matrix Ui is an approximation of U and is initialized as U0=I. An R×T matrix Di 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 Vi, Ui and Di may be performed as follows. First, a 2×2 Hermitian matrix Xp
where dp
Eigenvalue decomposition of Xp
Another 2×2 Hermitian matrix Yp
where {tilde over (d)}p
Eigenvalue decomposition of Yp
Matrix Vi is then updated as follows:
Vi+1=Vi·Tp
Matrix Ui is updated as follows:
Ui+1=Ui·Sp
Matrix Di is updated as follows:
Di+1=Sp
For the second SVD embodiment, the iterative process alternately finds (1) the Jacobi rotation that zeros out the off-diagonal elements with indices p1 and q1 in HH·H and (2) the Jacobi rotation that zeros out the off-diagonal elements with indices p2 and q2 in H·HH. The indices p1 and q1 may be swept by stepping p1 from 1 through T−1 and, for each value of p1, stepping q1 from p1+1 through T. The indices p2 and q2 may also be swept by stepping p2 from 1 through R−1 and, for each value of p2, stepping q2 from p2+1 through R. As an example, for a square matrix H, the indices may be set as p1=p2 and q1=q2. As another example, for a square or non-square matrix H, a set of p1 and q1 may be selected, then a set of p2 and q2 may be selected, then a new set of p1 and q1 may be select, then a new set of p2 and p2 may be selected, and so on, so that new values are alternately selected for indices p1 and q1 and indices p2 and q2.
Alternatively, for each iteration, the values of p1 and q1 for which |dp
Upon termination of the iterative process, the final Vi is a good approximation of {tilde over (V)}, the final Ui is a good approximation of U, and the final Di 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. The computation described above does not sufficiently constrain the left and right singular vector solutions so that the diagonal elements of the final Di are positive real values. The elements of the final Di may be complex values whose magnitudes are equal to the singular values of H. Vi and Di may be unrotated as follows:
{circumflex over (Σ)}=Di·P, and Eq (28a)
{circumflex over (V)}=Vi·P, Eq (28b)
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 Di. {circumflex over (Σ)} and {circumflex over (V)} are the final approximations of Σ and V, respectively.
For iteration i, the values for indices p1, q1, p2 and q2 are selected in a predetermined or deterministic manner (block 312). A 2×2 matrix Xp
Matrix Vi is then updated based on Tp
A determination is then made whether to terminate the singular value decomposition of H (block 336). 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 336, then index i is incremented (block 338), and the process returns to block 312 for the next iteration. Otherwise, if termination is reached, then post processing is performed on Di and Vi to obtain {circumflex over (Σ)} and {circumflex over (V)} (block 340). Matrix {circumflex over (V)} is provided as an approximation of V, matrix Ui is provided as an approximation of U, and matrix {circumflex over (Σ)} is provided as an approximation of Σ (block 342)
For both the first and second SVD embodiments, the right singular vectors in the final Vi and the left singular vectors in the final Ui or Û are ordered from the largest to smallest singular values because the eigenvectors in Vpq (for the first SVD embodiment) and the eigenvectors in Vp
For a MIMO system with multiple subbands, the iterative process may be performed for each channel response matrix H(k) to obtain matrices Vi(k), Ui(k), and Di(k), which are approximations of the matrix V(k) of right singular vectors, the matrix U(k) of left singular vectors, and the diagonal matrix Σ(k) of singular values, respectively, for that H(k). 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 the first SVD embodiment, for each subband k except for the first subband, the final solution Vi(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)=Vi(k−1) and W0(k)=1(k)·V0(k). For the second SVD embodiment, for each subband k except for the first subband, the final solutions Vi(k−1) and Ui(k−1) obtained for the prior subband k−1 may be used as initial solutions for the current subband k, so that V0(k)=Vi(k−1), U0(k)=Ui(k−1), and D0(k)=U0H(k)·H(k)·V0(k). For both embodiments, the iterative process operates on the initial solutions for subband k until a termination condition is encountered for the subband. The concept may also be used across time or both frequency and time, as described above.
For eigenvalue decomposition, as determined in block 416, a third matrix Di of eigenvalues may be derived based on the multiple Jacobi rotation matrices (block 420). For singular value decomposition based on the first SVD embodiment or scheme, a third matrix Wi of complex values may be derived based on the multiple Jacobi rotation matrices, a fourth matrix Û with orthogonal vectors may be derived based on the third matrix Wi, and a matrix {circumflex over (Σ)} of singular values may also be derived based on the third matrix Wi (block 422). For singular value decomposition based on the second SVD embodiment, a third matrix Ui with orthogonal vectors and a matrix {circumflex over (Σ)} of singular values may be derived based on the multiple Jacobi rotation matrices (block 424).
For eigenvalue decomposition, apparatus 500 further includes means for deriving a third matrix Di of eigenvalues based on the multiple Jacobi rotation matrices (block 520). For singular value decomposition based on the first SVD embodiment, apparatus 500 further includes means for deriving a third matrix Wi of complex values based on the multiple Jacobi rotation matrices, a fourth matrix Û with orthogonal vectors based on the third matrix, and a matrix {circumflex over (Σ)} of singular values based on the third matrix (block 522). For singular value decomposition based on the second SVD embodiment, apparatus 500 further includes means for deriving a third matrix Ui with orthogonal vectors and a matrix {circumflex over (Σ)} of singular values based on the multiple Jacobi rotation matrices (block 524).
3. System
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/processor 630. TX data processor 614 formats, encodes, interleaves, and modulates the received data and generates data symbols, which are modulation symbols for data. A TX spatial processor 620 receives and multiplexes the data symbols with pilot symbols, performs spatial processing with eigenvectors or right singular vectors if applicable, and provides Nap streams of transmit symbols to Nap transmitters (TMTR) 622a through 622ap. Each transmitter 622 processes its transmit symbol stream and generates a downlink modulated signal. Nap downlink modulated signals from transmitters 622a through 622ap are transmitted from antennas 624a through 624ap, respectively.
At user terminal 650, Nut antennas 652a through 652ut receive the transmitted downlink modulated signals, and each antenna 652 provides a received signal to a respective receiver (RCVR) 654. Each receiver 654 performs processing complementary to the processing performed by transmitters 622 and provides received symbols. A receive (RX) spatial processor 660 performs spatial matched filtering on the received symbols from all receivers 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 further 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/processor 680.
A channel processor 678 processes received pilot symbols and provides an estimate of the downlink channel response, Ĥ(k), for each subband of interest. Processor 678 and/or 680 may decompose each matrix Ĥ(k) using the techniques described herein to obtain {circumflex over (V)}(k) and {circumflex over (Σ)}(k), which are estimates of V(k) and Σ(k) for the downlink channel response matrix H(k). Processor 678 and/or 680 may derive a downlink spatial filter matrix Mdn(k) for each subband of interest based on {circumflex over (V)}(k), as shown in Table 1. Processor 680 may provide Mdn(k) to RX spatial processor 660 for downlink matched filtering and/or {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. Traffic data from a data source 686 and other data from controller/processor 680 are processed (e.g., encoded, interleaved, and modulated) by a TX data processor 688, multiplexed with pilot symbols, and further spatially processed by a 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 transmitters 654a through 654ut to generate Nut 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 receivers 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/processor 630.
A channel processor 628 processes received pilot symbols and provides an estimate of either HT(k) or U(k) for each subband of interest, depending on the manner in which the uplink pilot is transmitted. Processor 628 and/or 630 may decompose each matrix ĤT(k) using the techniques described herein to obtain Û(k). Processor 628 and/or 630 may also derive an uplink spatial filter matrix Mup(k) for each subband of interest based on Û(k). Processor 680 may provide Mup(k) to RX spatial processor 640 for uplink spatial matched filtering and/or Û(k) to TX spatial processor 620 for downlink spatial processing.
Controllers/processors 630 and 680 control the operation at access point 610 and user terminal 650, respectively. Memories 632 and 682 store data and program codes for access point 610 and user terminal 650, respectively. Processors 628, 630, 678, 680 and/or other processors may perform eigenvalue decomposition and/or singular value decomposition of the channel response matrices.
The matrix decomposition techniques described herein may be implemented by various means. For example, these techniques may be implemented in hardware, firmware, software, or a combination thereof. For a hardware implementation, the processing units used to perform matrix 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 firmware and/or software implementation, the matrix 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 (e.g., memory 632 or 682 in
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.
The present Application for Patent claims priority to Provisional Application Ser. No. 60/628,324, entitled “Eigenvalue Decomposition and Singular Value Decomposition of Matrices Using Jacobi Rotation,” filed Nov. 15, 2004 and Provisional Application No. 60/628,440, entitled “EIGENVALUE Decomposition and Singular Value Decomposition of Matrices Using Jacobi Rotation,” filed Nov. 15, 2004, assigned to the assignee hereof, and hereby expressly incorporated by reference herein.
Number | Name | Date | Kind |
---|---|---|---|
6859747 | Yutkowitz | Feb 2005 | B2 |
7065144 | Walton et al. | Jun 2006 | B2 |
7298805 | Walton et al. | Nov 2007 | B2 |
7327800 | Oprea et al. | Feb 2008 | B2 |
7539253 | Li et al. | May 2009 | B2 |
20020127978 | Khatri | Sep 2002 | A1 |
20030033105 | Yutkowitz | Feb 2003 | A1 |
20030218973 | Oprea et al. | Nov 2003 | A1 |
20030235255 | Ketchum et al. | Dec 2003 | A1 |
20050047515 | Walton et al. | Mar 2005 | A1 |
20050267925 | Clue | Dec 2005 | A1 |
20060056531 | Li et al. | Mar 2006 | A1 |
20060067277 | Thomas et al. | Mar 2006 | A1 |
20060106902 | Howard et al. | May 2006 | A1 |
20060285531 | Howard et al. | Dec 2006 | A1 |
20070249296 | Howard et al. | Oct 2007 | A1 |
Number | Date | Country |
---|---|---|
2005115862 | Jan 2006 | RU |
2005114868 | Dec 2005 | WO |
Number | Date | Country | |
---|---|---|---|
20060155798 A1 | Jul 2006 | US |
Number | Date | Country | |
---|---|---|---|
60628324 | Nov 2004 | US | |
60628440 | Nov 2004 | US |