Efficient joint detection

Abstract

Data signals are received over a shared spectrum in a code division multiple access communication format as a combined signal. The combined signal is sampled as a plurality of received vector versions. A plurality of system matrices and an associated covariance matrix using codes and estimated impulse responses of the data signals is produced. Each system matrix corresponds to a received vector version. The system and covariance matrices are extended and approximated as block circulant matrices. A diagonal matrix of each of the extended and approximated system and covariance matrices are determined by prime factor algorithm-fast Fourier transform (PFA-FFT)without division of the matrix. The received vector versions are extended. A product of the diagonal matrices and the extended received vector versions is taken. An inverse block discrete Fourier transform is performed by a PFA-FFT a on a result of the product to produce the estimated data.

Description

BACKGROUND

FIG. 1 is an illustration of a wireless communication system 10. The communication system 10 has base stations 12₁ to 12₅ (12) which communicate with user equipments (UEs) 14₁ to 14₃ (14). Each base station 12 has an associated operational area, where it communicates with UEs 14 in its operational area.

In some communication systems, such as frequency division duplex using code division multiple access (FDD/CDMA) and time division duplex using code division multiple access (TDD/CDMA), multiple communications are sent over the same frequency spectrum. These communications are differentiated by their channelization codes. To more efficiently use the frequency spectrum, TDD/CDMA communication systems use repeating frames divided into timeslots for communication. A communication sent in such a system will have one or multiple associated codes and timeslots assigned to it.

Since multiple communications may be sent in the same frequency spectrum and at the same time, a receiver in such a system must distinguish between the multiple communications. One approach to detecting such signals is multiuser detection (MUD). In MUD, signals associated with all the UEs 14, are detected simultaneously. For TDD/CDMA systems, one of the popular MUD techniques is a joint detection technique using block linear equalizer (BLE-JD). Techniques for implementing BLE-JD include using a Cholesky or an approximate Cholesky decomposition. These approaches have high complexity. The high complexity leads to increased power consumption, which at the UE 14 results in reduced battery life.

Accordingly, it is desirable to have computationally efficient approaches to detecting received data.

SUMMARY

K data signals, or bursts, are transmitted over a shared spectrum in a code division multiple access communication format. A combined signal is received and sampled over the shared spectrum, as a plurality of received vector versions. The combined signal includes the K transmitted data signals. A plurality of system matrices and an associated covariance matrix using codes and estimated impulse responses of the K data signals is produced. Each system matrix corresponds to a received vector version. The system and covariance matrices are extended and approximated as block circulant matrices. A diagonal matrix of each of the extended and approximated system and covariance matrices are determined by prime factor algorithm-fast Fourier transform (PFA-FFT) without division of the matrices. The received vector versions are extended. A product of the diagonal matrices and the extended received vector versions is taken. An inverse block discrete Fourier transform is performed by a PFA-FFT on a result of the product to produce the estimated data of the K data signals.

BRIEF DESCRIPTION OF THE DRAWING(S)

FIG. 1 is a wireless communication system.

FIG. 2 is a simplified transmitter and an efficient joint detection receiver.

FIG. 3 is an illustration of a communication burst.

FIGS. 4 a and 4b are a flow chart of a preferred embodiment for efficient joint detection.

FIG. 5 is an illustration of a data burst indicating extended processing areas.

FIG. 6 is a block diagram of a preferred implementation of efficient joint detection.

FIG. 7 is a simplified receiver having multiple antennas.

FIG. 8 is a simplified receiver sampling the received signal using fractional sampling.

FIG. 9 is a simplified receiver having multiple antennas and using fractional sampling.

FIG. 10 is a block diagram of a preferred implementation of efficient joint detection for fractional sampling or receive diversity.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT(S)

FIG. 2 illustrates a simplified transmitter 26 and receiver 28 using efficient joint detection in a TDD/CDMA communication system, although efficient joint detection is applicable to other systems, such as FDD/CDMA. In a typical system, a transmitter 26 is in each UE 14 and multiple transmitting circuits 26 sending multiple communications are in each base station 12. The joint detection receiver 28 may be at a base station 12, UEs 14 or both.

The transmitter 26 sends data over a wireless radio channel 30. A data generator 32 in the transmitter 26 generates data to be communicated to the receiver 28. A modulation/spreading/training sequence insertion device 34 spreads the data with the appropriate code(s) and makes the spread reference data time-multiplexed with a midamble training sequence in the appropriate assigned time slot, producing a communication burst or bursts.

A typical communication burst 16 has a midamble 20, a guard period 18 and two data fields 22, 24, as shown in FIG. 3. The midamble 20 separates the two data fields 22, 24 and the guard period 18 separates the communication bursts to allow for the difference in arrival times of bursts transmitted from different transmitters 26. The two data fields 22, 24 contain the communication burst's data.

The communication burst(s) are modulated by a modulator 36 to radio frequency (RF). An antenna 38 radiates the RF signal through the wireless radio channel 30 to an antenna 40 of the receiver 28. The type of modulation used for the transmitted communication can be any of those known to those skilled in the art, such as quadrature phase shift keying (QPSK) or M-ary quadrature amplitude modulation (QAM).

The antenna 40 of the receiver 28 receives various radio frequency signals. The received signals are demodulated by a demodulator 42 to produce a baseband signal. The baseband signal is sampled by a sampling device 43, such as one or multiple analog to digital converters, at the chip rate of the transmitted bursts. The samples are processed, such as by a channel estimation device 44 and an efficient joint detection device 46, in the time slot and with the appropriate codes assigned to the received bursts. The channel estimation device 44 uses the midamble training sequence component in the baseband samples to provide channel information, such as channel impulse responses. The channel information is used by the efficient joint detection device 46 to estimate the transmitted data of the received communication bursts as soft symbols.

The efficient joint detection device 46 uses the channel information provided by the channel estimation device 44 and the known spreading codes used by the transmitter 26 to estimate the data of the desired received communication burst(s).

Although efficient joint detection is explained using the third generation partnership project (3GPP) universal terrestrial radio access (UTRA) TDD system as the underlying communication system, it is applicable to other systems. That system is a direct sequence wideband CDMA (W-CDMA) system, where the uplink and downlink transmissions are confined to mutually exclusive timeslots.

The receiver 28 receives a total of K bursts that arrive simultaneously, within one observation interval. For the 3GPP UTRA TDD system, each data field of a time slot corresponds to one observation interval. For a frequency division duplex (FDD) CDMA system, the received signals are continuous, i.e., not in bursts. To handle the continuous signals, FDD systems divide the received signals into time segments prior to applying efficient joint detection.

A code used for a k^th burst is represented as c^(k). The K bursts may originate from K different transmitters or for multi-code transmissions, less than K different transmitters.

Each data field of a communication burst has a predetermined number of transmitted symbols, Ns. Each symbol is transmitted using a predetermined number of chips, which is the spreading factor, Q. Accordingly, each data field has N_S×Q chips. After passing through the wireless radio channel, which can introduce a delay spread of up to W−1 chips, the observation interval at the receiver is of length Q×N_S+W−1 chips.

The symbol response vector b^(k) as the convolution of the channel response vector h^(k) with the corresponding spreading code c^(k) is per Equation 1.b^(k)=h^(k)∘c^(k)  Equation 1∘ denotes the convolutional operator. The length of b^(k) is SF+W−1.

Using the symbol response vectors, the system matrix A is defined as per Equation 2.

$\begin{array}{cc}A=\left[\begin{array}{ccccc}B& & & & \\ & \stackrel{\stackrel{\_}{\updownarrow Q}}{B}& & & \\ & & B& & \\ & & & \ddots & \\ & & & & B\end{array}\right]& \mathrm{Equation}2\end{array}$

The size of the matrix is (N_s•SF+W−1)×N_s•K. A is a block Toeplitz matrix.

Block B is defined as per Equation 3.B=[b⁽¹⁾b⁽²⁾ . . . b^(K)]  Equation 3The received vector sampled at the chip rate can be represented by Equation 4.r=Ad+n  Equation 4

The size of vector r is (N_s•SF+W−1) by 1. This size corresponds to the observation interval.

Data vector d of size N_s•K by 1 has the form of Equation 5.d=[d₁^Td₂^T . . . d_Ns^T]^T  Equation 5

The sub-vector d_n of size K by 1 is composed of the n^th symbol of each user and is defined as Equation 6.d_n=[d_n⁽¹⁾d_n⁽²⁾ . . . d_n^(K)]^T, n=1, . . . , N_s  Equation 6

The vector n of size (N_s•SF+W−1) by 1 is the background noise vector and is assumed to be white.

Determining d using an MMSE solution is per Equation 7.d=R⁻¹(A^Hr)  Equation 7(•)^H represents the Hermetian function (complex conjugate transpose). The covariance matrix of the system matrix R for a preferred MMSE solution is per Equation 8.R=A^HA+σ²I  Equation 8σ² is the noise variance, typically obtained from the channel estimation device 44, and I is the identity matrix.

Using block circulant approximation and block DFT using PFA-FFTs, d in Equation 7 can be determined per Equation 9.

$\begin{array}{cc}\begin{array}{c}\underset{\_}{d}=F\left(R^{-1}\right)F\left(A^H\underset{\_}{r}\right)\\ F^{-1}\left({\Lambda }^{-1}{\Lambda }_{A}F\left({\underset{\_}{r}}_c\right)\right)\end{array}& \mathrm{Equation}9\end{array}$ F(•) and F⁻¹(•) indicate the block-DFT function and the inverse block-DFT, respectively. The derivation of the block diagonal matrices Λ and Λ_A is described subsequently. Instead of directly solving Equation 9, Equation 9 can be solved using the LU decomposition and the forward and backward substitution of the main diagonal block of Λ. Alternately, Equation 9 can be solved using Cholesky decomposition.

FIG. 4 is a flowchart for a preferred method of determining the data vector d using fast joint detection. The system matrix A is determined using the estimated channel response vector h^(k) and the spreading code c^(k) for each burst, 48. R, the covariance matrix of the system matrix, is formed, 49. The system matrix A and its covariance matrix R are extended block square matrices. The extended A is of size D•Q by D•K and the extended R is of size D•K by D•K, respectively. D is chosen as per Equation 10.

$\begin{array}{cc}D\ge \lceil N_s+\frac{W-1}{Q}\rceil & \mathrm{Equation}10\end{array}$ Both extended matrices are approximated to block circulant matrices, A_c and R_c, 50. Because of the extension of A and R, the received vector, r, is extended to the vector r_c of size D•SF×1 by inserting zeros, 51. The block-diagonal matrix, Λ, is determined by taking a block DFT using PFA-FFT of the first block column of R_c, 52.

The block DFT of matched filtering F(A^Hr) is approximated by F(A_c^Hr_c). It is calculated by taking a block DFT using PFA-FFT of A_c and r_c, 53. Due to the block-diagonal structure of Λ and Λ_A, the blocks F(d)⁽ⁱ⁾, i=1, . . . , D, are of size K by 1 in F(d). They are determined by performing on the main diagonal blocks, Λ⁽ⁱ⁾ of Λ, LU decomposition, Λ⁽ⁱ⁾=L⁽ⁱ⁾U⁽ⁱ⁾, forward substitution, L⁽ⁱ⁾y⁽ⁱ⁾=Λ_A^(i)HF(r_c)⁽ⁱ⁾, 55, and backward substitution, U⁽ⁱ⁾[F(d)]⁽ⁱ⁾=y⁽ⁱ⁾, 56. L⁽ⁱ⁾ is a lower triangular matrix. U⁽ⁱ⁾ is an upper triangular matrix. Λ_A⁽ⁱ⁾ is the i^th main diagonal block of size SF by K in Λ_A and F(r_c)⁽ⁱ⁾ is the i^th block of size Q×1 in F(r_c). Λ_A is the block DFT using PFA-FFT of the first column of A_c and F(r_c) is the block DFT using PFA-FFT of the vector, r_c. The estimated data vector, d, is determined by an inverse block-DFT of F(d), 57.

Although Equation 9 is a MMSE based solution, fast joint detection can be applied to other approaches, such as a zero forcing approach as per Equation 11.Rd=(A^HA)d=A^Hr  Equation 11As shown in Equation 11, in the zero forcing solution, the σ²I term is deleted from Equation 8. The following is a derivation for the MMSE solution, although an analogous derivation can be used for a zero forcing solution.

To reduce the complexity in determining F(A^Hr), a block DFT using PFA-FFT approach taking advantage of the block-Toeplitz structure of A may be used as shown in Equation 2. First, by repeating B, we extend A to a block-square matrix of size D•Q×D•K, to use all of the chip symbols in the observation interval. The extended A is composed of D² blocks of size Q×K. The extended A is approximated to the block-circulant matrix A_c.

A_c can be decomposed into three matrices per Equation 12.A_c=F_(Q)^HΛ_AF_(K)  Equation 12F_(n)=FI_n is a block DFT using PFA-FFT matrix of size D•n by D•n. denotes a Kronecker product. I_n is an identity matrix of size n×n and F is a DFT matrix of size D×D, whose elements f_il, i and l=1, 2, . . . , D are per Equation 13.

$\begin{array}{cc}{f}_{\mathrm{il}}=\frac{1}{\sqrt{D}}\exp \left(-j\frac{2\pi \mathrm{il}}{D}\right)& \mathrm{Equation}13\end{array}$ D is the length of the DFT and F^HF=I. I is an identity matrix of size D×D.

The block diagonal matrix Λ_A is of size D•Q×D•K and has the form per Equation 14.

$\begin{array}{cc}{\Lambda }_A=\left[\begin{array}{cccc}{\Lambda }_A^{\left(1\right)}& 0& \dots & 0\\ 0& {\Lambda }_A^{\left(2\right)}& \dots & 0\\ ⋮& ⋮& \ddots & ⋮\\ 0& 0& \dots & {\Lambda }_A^{\left(D\right)}\end{array}\right]& \mathrm{Equation}14\end{array}$

Each of its entries Λ_A⁽ⁱ⁾, i=1, 2, . . . , D, is a Q by K block per Equation 15.

$\begin{array}{cc}{\Lambda }_A^{\left(i\right)}=\left[\begin{array}{cccc}{\lambda }_{1,1}^{\left(A,i\right)}& \dots & \dots & {\lambda }_{1,K}^{\left(A,i\right)}\\ ⋮& \ddots & & ⋮\\ ⋮& & \ddots & ⋮\\ {\lambda }_{Q,1}^{\left(A,i\right)}& \dots & \dots & {\lambda }_{Q,K}^{\left(A,i\right)}\end{array}\right]& \mathrm{Equation}15\end{array}$

Alternatively, the main diagonal blocks, Λ_A⁽ⁱ⁾, i=1, 2, . . . , D, can be computed by, per Equation 16.[Λ_A^(1)TΛ_A^(2)T . . . Λ_A^(D)T]^T=(F_(Q)A_c(:,1:K))  Equation 16A_c(:,1:K) denotes the first block column of A_c. Namely, the first K columns of A_c. F_(SF)A_c(:,1:K) can be calculated by Q•K parallel non-block DFTs of length D, using PFA-FFTs.

Due to the extension of A, the received vector r is also extended by inserting zeros, becoming vector r_c of size D•Q by 1.

Using the above, F(A^Hr) is approximated to F(A_c^Hr_c). It can be written as Equation 17.F(A_c^Hr_c)=F_(K)^HΛ_A^HF_(Q)r_c  Equation 17

The covariance matrix R of size N_s•K×N_s•K has the block-square matrix form shown in Equation 18.

$\begin{array}{cc}R=\left[\begin{array}{cccccccccccc}{R}_0& R_1^H& \dots & R_L^H& \dots & \dots & \dots & \dots & 0& \dots & 0& 0\\ R_1& R_0& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ ⋮& R_1& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ ⋮& ⋮& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ R_L& ⋮& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ 0& R_L& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ ⋮& 0& \ddots & R_L& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ ⋮& ⋮& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & 0& ⋮\\ ⋮& ⋮& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & R_L^H& 0\\ ⋮& ⋮& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& R_L^H\\ ⋮& ⋮& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ 0& 0& \dots & \dots & \dots & \dots & \dots & \dots & R_L& \dots & R_1& R_0\end{array}\right]& \mathrm{Equation}18\end{array}$ L is defined per Equation 19.

$\begin{array}{cc}L=\lceil \frac{Q+W-1}{Q}\rceil & \mathrm{Equation}19\end{array}$ Each entry, R_i, in the R matrix is a K by K block and 0 is a K by K zero matrix. Due to the size of the extended A, the matrix R is also extended to size D•K by D•K by inserting zeros.

The extended R is approximated to a block-circulant matrix, R_c, of size D•K by D•K per Equation 20.

$\begin{array}{cc}{R}_c=\left[\begin{array}{cccccccccccc}{R}_0& R_1^H& \dots & R_L^H& \dots & \dots & \dots & \dots & 0& \dots & R_2& R_1\\ R_1& R_0& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ ⋮& R_1& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & R_L& ⋮\\ ⋮& ⋮& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& R_L\\ R_L& ⋮& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ 0& R_L& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ ⋮& 0& \ddots & R_L& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ ⋮& ⋮& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & 0& ⋮\\ R_L^H& ⋮& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & R_L^H& 0\\ ⋮& R_L^H& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& R_L^H\\ ⋮& ⋮& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & ⋮& ⋮\\ R_1^H& R_2^H& \dots & \dots & \dots & \dots & \dots & \dots & R_L& \dots & R_1& R_0\end{array}\right]& \mathrm{Equation}20\end{array}$

The block circulant matrix R_c, is decomposed into three matrices per Equation 21.R_c=F_(K)^HΛF_(K)  Equation 21F_(K)=FI_K is a block DFT using PFA-FFT matrix of size D•K×D•K. denotes a Kronecker product. I_K is an identity matrix of size K×K and F is a DFT matrix of size D×D as described in Equation 13.

The block diagonal matrix Λ of size D•K by D•K has the form per Equation 22.

$\begin{array}{cc}\Lambda =\left[\begin{array}{cccc}{\Lambda }^{\left(1\right)}& 0& \dots & 0\\ 0& {\Lambda }^{\left(2\right)}& \dots & 0\\ ⋮& ⋮& \ddots & ⋮\\ 0& 0& \dots & {\Lambda }^{\left(D\right)}\end{array}\right]& \mathrm{Equation}22\end{array}$ Each of its entries, Λ⁽ⁱ⁾, i=1, 2, . . . , D, is a K by K block, per Equation 23A.

$\begin{array}{cc}{\Lambda }^{\left(i\right)}=\left[\begin{array}{cccc}{\lambda }_{1,1}^{\left(i\right)}& \dots & \dots & {\lambda }_{1,K}^{\left(i\right)}\\ ⋮& \ddots & & ⋮\\ ⋮& & \ddots & ⋮\\ {\lambda }_{K,1}^{\left(i\right)}& \dots & \dots & {\lambda }_{K,K}^{\left(i\right)}\end{array}\right]& \mathrm{Equation}23A\end{array}$ Alternatively, the main diagonal blocks, Λ⁽ⁱ⁾, i=1, 2, . . . , D, can be computed per Equation 23B.[Λ^(1)TΛ^(2)T . . . Λ^(D)T]^T=(F_(K)R_c(:,1:K)),  Equation 23BR_c(:,1:K) denotes the first block column of R_c. Namely, the first K columns of R_c. F_(K)R_c(:,1:K) can be calculated by K² parallel non-block DFTs of length D, using PFA-FFTs.

The estimated data vector, d, in Equation 7 is preferably approximated per Equation 24A.

$\begin{array}{cc}\begin{array}{c}\underset{\_}{d}=R^{-1}{A}^H\underset{\_}{r}\\ \approx R_c^{-1}{A}_c^H{\underset{\_}{r}}_c\\ =F_{\left(K\right)}^H{\Lambda }^{-1}{\Lambda }_A^H{F}_{\left(Q\right)}{\underset{\_}{r}}_c\end{array}& \mathrm{Equation}24A\end{array}$ The block diagonal matrix Λ⁻¹ is per Equation 24B.

$\begin{array}{cc}{\Lambda }^{-1}=\left[\begin{array}{cccc}{\Lambda }^{{\left(1\right)}^{-1}}& 0& \dots & 0\\ 0& {\Lambda }^{{\left(2\right)}^{-1}}& \dots & 0\\ ⋮& ⋮& \ddots & ⋮\\ 0& 0& \dots & {\Lambda }^{{\left(D\right)}^{-1}}\end{array}\right]& \mathrm{Equation}24B\end{array}$ The inversion of Λ requires an inversion of K×K matrices, Λ⁽ⁱ⁾, i=1, 2, . . . , D.

Equation 24A can be rewritten as Equation 25.F(d)=Λ₋₁Λ_A^HF(r_c)  Equation 25F(r_c) is per Equations 26A and 26B.F(r_c)≡F_(Q)r_c  Equation 26AF(d_c)≡F_(K)r_c  Equation 26BDue to the block-diagonal structure of Λ⁻¹ and Λ_A^H, Equation 25 can be efficiently calculated as follows. The terms of Equation 25 are partitioned into D blocks, as per Equation 27.

$\begin{array}{cc}\left[\begin{array}{c}{F\left(\underset{\_}{d}\right)}^{\left(1\right)}\\ {F\left(\underset{\_}{d}\right)}^{\left(2\right)}\\ ⋮\\ {F\left(\underset{\_}{d}\right)}^{\left(D\right)}\end{array}\right]={\left[\begin{array}{cccc}{\Lambda }^{{\left(1\right)}^{-1}}& & & \\ & {\Lambda }^{{\left(2\right)}^{-1}}& & \\ & & \ddots & \\ & & & {\Lambda }^{{\left(D\right)}^{-1}}\end{array}\right]\left[\begin{array}{cccc}{\Lambda }_A^{\left(1\right)}& & & \\ & {\Lambda }_A^{\left(2\right)}& & \\ & & \ddots & \\ & & & {\Lambda }_A^{\left(D\right)}\end{array}\right]}^H\hspace{1em}\left[\begin{array}{c}{F\left({\underset{\_}{r}}_c\right)}^{\left(1\right)}\\ {F\left({\underset{\_}{r}}_c\right)}^2\\ ⋮\\ {F\left({\underset{\_}{r}}_c\right)}^{\left(D\right)}\end{array}\right]& \mathrm{Equation}27\end{array}$

Each block in Equation 27 is solved separately, per Equation 28.F(d)⁽ⁱ⁾=Λ⁽ⁱ⁾⁻¹Λ_A^HF(r_c)⁽ⁱ⁾  Equation 28

F(d)⁽ⁱ⁾ is a K by 1 vector. Λ⁽ⁱ⁾ is a K by K matrix, per Equation 22. Λ_A⁽ⁱ⁾ is a Q by K matrix, per Equation 14. F(r_c)⁽ⁱ⁾ is a Q by 1 vector and is composed of elements (1+(i−1)Q) through (i•Q) of F(r_c).

To avoid the direct inversion of Λ⁽ⁱ⁾, Equation 28 can be solved by using LU decomposition and forward and backward substitution. Equation 28 is rewritten per Equation 29.Λ⁽ⁱ⁾F(d)⁽ⁱ⁾=Λ_A^(i)HF(r_c)⁽ⁱ⁾  Equation 29Λ⁽ⁱ⁾ is decomposed per Equation 30.Λ⁽ⁱ⁾=L⁽ⁱ⁾U⁽ⁱ⁾  Equation 30L⁽ⁱ⁾ is a lower triangular matrix and U⁽ⁱ⁾ is an upper triangular matrix.

Using LU decomposition, Equation 28 is represented as, per Equation 31.L⁽ⁱ⁾U⁽ⁱ⁾F(d)⁽ⁱ⁾=Λ_A^(i)HF(r)⁽ⁱ⁾  Equation 31F(d)⁽ⁱ⁾ in Equation 31 is solved by forward substitution, per Equation 32, and backward substitution, per Equation 33.Forward Substitution: L⁽ⁱ⁾y⁽ⁱ⁾=Λ_A^(i)HF(r)⁽ⁱ⁾  Equation 32Backward Substitution: U⁽ⁱ⁾[F(d)]⁽ⁱ⁾=y⁽ⁱ⁾  Equation 33

Finally, d is determined for all blocks as per Equation 34.

$\begin{array}{cc}\underset{\_}{d}=F^{-1}\left(\underset{\_}{d}\right)=F^{-1}\left(\left[\begin{array}{c}{F\left(\underset{\_}{d}\right)}^{\left(1\right)}\\ {F\left(\underset{\_}{d}\right)}^{\left(2\right)}\\ ⋮\\ {F\left(\underset{\_}{d}\right)}^{\left(D\right)}\end{array}\right]\right)& \mathrm{Equation}34\end{array}$

FIG. 6 is a block diagram of a preferred implementation of efficient joint detection in a TDD/CDMA system. Using the received vector, r, r_c is formed by inserting zeros, then a block DFT 100 of r_c is performed per Equation 26, to produce F(r_c).

Using the received training sequences, the channel impulse responses for each transmitted burst, h^(k), is determined by an estimate channel response block 102. Using each channelization code, c^(k) and the channel impulse response, h^(k), the system matrix, A, is determined by compute block matrix A block 104 per Equation 2.

To determine Λ_A, the system matrix A is extended by extend block 132, to use all received chips in the observation interval. The first block column of the block-circulant matrix A_c is determined by selecting the first K columns of the extended A matrix by first block column block 114. By taking a block DFT using PFA-FFT 118 using PFA-FFT, Λ_A is determined.

To determine Λ, R is first determined by compute R block 140. For an MMSE solution, R=A^HA+σ²I is used; for a zero forcing solution, R=A^HA is used. Due to the size of the extended A, R is also extended by extend block 134. The first block column of the extended R matrix is determined by selecting the first K columns of the extended R matrix by first block column block 108. The first block column of the extended R matrix is circularized by a circularize block column block 110. It becomes the first block column of a block-circulant R_c. By taking a block DFT using PFA-FFT by block DFT block 112, Λ is determined.

To efficiently compute the estimated data vector d, Λ_A, Λ, and F(r_c) are divided into blocks Λ_A⁽ⁱ⁾, Λ⁽ⁱ⁾, and F(r_c)⁽ⁱ⁾, i=1, 2, . . . , D, respectively, exploiting the block-diagonal structures of Λ_A and Λ, by partition block 136. The complex conjugate transpose of Λ_A⁽ⁱ⁾, Λ_A^(i)H, is determined by a transpose block 130. A multiplier 128 multiplies Λ_A^(i)H by F(r_c)⁽ⁱ⁾. Λ⁽ⁱ⁾ is decomposed using LU decomposition by a LU decomposition block 126, per Equation 30. By performing forward and backward substitution, per Equations 31-33, using forward and backward substitution blocks 124 and 122, respectively, F(d)⁽ⁱ⁾ is determined. By repeating the LU decomposition and forward and backward substitution D times, F(d) is found. Taking an inverse block DFT using PFA-FFT of F(d) by block inverse DFT block 120, d is estimated.

FIGS. 7, 8 and 9 are simplified diagrams of receivers applying efficient joint detection to multiple reception antennas and/or fractional (multiple chip rate) sampling. A receiver 28 with multiple reception antennas is shown in FIG. 7. Transmitted bursts are received by each antenna 40₁ to 40_m (40). Each antennas' version of the received bursts are reduced to baseband, such as by demodulators 42₁ to 42_m. The baseband signals for each antenna are sampled by sampling devices 43₁ to 43_m to produce a received vector, r₁ to r_m, for each antenna 40. The samples corresponding to the midamble are processed by a channel estimation device 144 to produce channel response matrices, H₁ to H_m, for each antenna 40. The received data vector, d, is determined by an efficient joint detection device 142 using the received vectors and the channel response matrices.

A receiver 28 sampling using fraction sampling is shown in FIG. 8. Transmitted bursts are received by the antenna 40. The received bursts are reduced to baseband, such as by a demodulator 42. The baseband signal is sampled by a sampling device 43 to produce factional samples as received vectors, r₁ to r_m. Each received vector represents chip rate samples sampled at a fraction of a chip offset. To illustrate, for twice the chip rate sampling, two received vectors r₁ and r₂ are produced. Each of those vectors has samples spaced by half a chip in time. Samples corresponding to the midamble are processed by a channel estimation device 144 to produce channel response matrices, H₁ to H_m, for each set of fractional samples. The received data vector, d, is determined by an efficient joint detection device 142 using the received vectors and the channel response matrices.

A receiver 28 with multiple reception antennas and using fractional sampling is shown in FIG. 9. Transmitted bursts are received by each antenna 40₁ to 40_i(40). Each antennas' version of the received bursts are reduced to baseband, such as by demodulators 42₁ to 42_i. The baseband signals for each antenna are sampled by sampling devices 43₁ to 43_j to produce a received vectors, r₁ to r_m. The received vectors for each antenna correspond to each multiple of the chip rate samples. The samples corresponding to the midamble are processed by a channel estimation device 144 to produce channel response matrices, H₁ to H_m, for each antenna's fractional samples. The received data vector, d, is determined by an efficient joint detection device 142 using the received vectors and the channel response matrices.

In applying efficient joint detection to either receive diversity, fractional sampling or both, the received communication bursts are viewed as M virtual chip rate received bursts. To illustrate, for twice the chip rate sampling and two antenna receive diversity, the received bursts are modeled as four (M=4) virtual chip rate received bursts.

Each received burst is a combination of K transmitted bursts. Each of the K transmitted bursts has its own code. The channel impulse response vector of the k^th out of K codes and the m^th out of the M virtual received bursts is h^(k,m). h^(k,m) has a length W and is estimated from the midamble samples of the burst of the k^th code of the m^th virtual received burst.

Each of the N data symbols of the burst of the k^th code is per Equation 35.d^(k)=[d₁^(k)d₂^(k) . . . d_N^(k)]^T, 1≦k≦K  Equation 35

The code of the k^th burst is per Equation 36.c^(k)=[c₁^(k)c₂^(k) . . . c_Q^(k)]^T, 1≦k≦K  Equation 36

The symbol response of the k^th code's contribution to the m^th virtual burst, b^(k,m) is per Equation 37.b^(k,m)=h^(k,m)c^(k)  Equation 37

The length of the symbol response is Q+W−1. Q is the spreading factor. The system matrix, A^(m), for each m^th received burst is per Equation 38.

$\begin{array}{cc}{A}^{\left(m\right)}=\left[\begin{array}{ccccc}{B}^{\left(m\right)}& & & & \\ & \stackrel{\stackrel{\_}{\updownarrow Q}}{B^{\left(m\right)}}& & & \\ & & B^{\left(m\right)}& & \\ & & & \ddots & \\ & & & & B^{\left(m\right)}\end{array}\right]& \mathrm{Equation}38\end{array}$

Each block B^(m) is of size (Q+W−1) by K and is per Equation 39.B^(m)=└b^(1,m)b^(2,m) . . . b^(K,m)┘  Equation 39

The overall system matrix A is per Equation 40.

$\begin{array}{cc}A=\left[\begin{array}{c}{A}^{\left(1\right)}\\ A^{\left(2\right)}\\ ⋮\\ A^{\left(M\right)}\end{array}\right]& \mathrm{Equation}40\end{array}$

As shown in Equation 38, each sub-system matrix A^(m) is block Toeplitz. The overall received vector of the M virtual bursts is of size M(NQ+W−1) and is per Equation 41.r=[r₁^Tr₂^T . . . r_M^T]^T  Equation 41The m^th received vector r_m is of size NQ+W−1 by 1.

Equation 42 is a model for the overall received vector.r=Ad+n  Equation 42n is the noise variance.

Each m^th received virtual burst is per Equation 43.r_m=A^(m)d+n_m  Equation 43n_m is the noise variance for the m^th received virtual burst.

To solve for the data vector d in Equation 42, a block linear equalizer with either a zero forcing or minimum mean square error (MMSE) approach may be used per Equation 44.{circumflex over (d)}=R⁻¹A^Hr  Equation 44R is the covariance matrix.

For a zero forcing solution, R is per Equation 45.

$\begin{array}{cc}R=\sum _{m=1}^M{A}^{{\left(M\right)}^H}{A}^{\left(m\right)}=A^{H}A& \mathrm{Equation}45\end{array}$

For a MMSE solution, R is per Equation 46.

$\begin{array}{cc}R=\sum _{m=1}^M{A}^{{\left(m\right)}^H}{A}^{\left(m\right)}+{\sigma }^{2}I=A^{H}A+{\sigma }^{2}I& \mathrm{Equation}46\end{array}$

The covariance matrix for either the zero forcing or MMSE solution is a block Toeplitz. To apply a discrete Fourier transform to the block-Toeplitz A^(m) matrix, a block-circulant approximation of A^(m), A_c^(m) is used. To make A^(m) a block-square matrix, A^(m) is extended. The extended A^(m) matrix is then approximated to a block circulant matrix A_c^(m).

The A_c^(m) matrix is composed of D by D blocks. Each block is of size Q by K. Accordingly, the size of A_c^(m) becomes DQ by DK. To include all the elements of A^(m), D is chosen to be an integer larger than D_min as determined per Equation 47.

$\begin{array}{cc}{D}_{\min }=\lceil N+\frac{\left(W-1\right)}{Q}\rceil & \mathrm{Equation}47\end{array}$ ┌•┐ represents a round up to an integer function.

The covariance matrix R is a block-square matrix of size NK by NK with blocks of size K by K. For R to be compatible with the extended A_c^(m) matrix, R is extended to the size DK by DK by zero-padding and approximating the extended R to a block circulant covariance matrix R_c. For the received vector, r^(m), to be compatible with A_c^(m) and R_c, r^(m) is extended to a DQ by 1 vector, r_c^(m) by zero padding.

After extending the received vectors, r^(m), the overall received vector is per Equation 48.

$\begin{array}{cc}{\underset{\_}{r}}_c={\left[\begin{array}{cccc}{\underset{\_}{r}}_c^{{\left(1\right)}^T}& {\underset{\_}{r}}_c^{{\left(2\right)}^T}& \dots & {\underset{\_}{r}}_c^{{\left(M\right)}^T}\end{array}\right]}^T& \mathrm{Equation}48\end{array}$

Each block-circulant matrix A_c^(m) is diagonalized to a block-diagonal matrix by block discrete Fourier transform matrices per Equation 49.A_c^(m)=F_(Q)^HΛ_A^(m)F_(K)  Equation 49

F_(Q) is per Equation 50.F_(Q)=FI_Q  Equation 50

F_(K) is per Equation 51.F_(K)=FI_K  Equation 51F is a discrete Fourier transform matrix of size D by D and is an n by n identity matrix. Λ_A^(m) is a block diagonal matrix of the form of Equation 52.

$\begin{array}{cc}{\Lambda }_A^{\left(m\right)}=\left[\begin{array}{cccc}{\Lambda }^{\left(1,m\right)}& 0& \dots & 0\\ 0& {\Lambda }_A^{\left(2,m\right)}& \dots & 0\\ ⋮& ⋮& \ddots & ⋮\\ 0& 0& \dots & {\Lambda }_A^{\left(D,m\right)}\end{array}\right]& \mathrm{Equation}52\end{array}$

Λ_A^(l,m) for l=1, . . . , D, and m=1, . . . , M is a non-zero block of size Q by K. 0 is a zero matrix of size Q by K having all zero elements.

Λ_A^(l,m) is alternately computed per Equation 53.Λ_A^(m)=diag^B(F_(Q)A_c^(m)(:,1:K))  Equation 53

A_c^(m)(:,1:K) is the first block column of A_c^(m). The first block column having K columns. To determine Λ_A^(m), preferably F_(Q)A_c^(m)(:,1:K) is determined by QK parallel non-block DFTs of length D, using PFA-FFTs. The block circulant matrix R_c is also preferably diagonalized to the block diagonal matrix Λ_R by a block DFT matrix F_(K)=FI_K as per Equation 54.R_c=F_(K)^HΛ_RF_(K)  Equation 54

The block diagonal matrix Λ_R is composed by blocks Λ_R^(l), l=1, . . . , D of size K by K in its main diagonal block, per Equation 55.

$\begin{array}{cc}{\Lambda }_R=\left[\begin{array}{cccc}{\Lambda }_R^{\left(1\right)}& 0& \dots & 0\\ 0& {\Lambda }_R^{\left(2\right)}& \dots & 0\\ ⋮& ⋮& \ddots & ⋮\\ 0& 0& \dots & {\Lambda }_R^{\left(D\right)}\end{array}\right]& \mathrm{Equation}55\end{array}$

Another approach to determine Λ_R is per Equation 56.Λ_R=diag^B(F_(K)R_c(:,1:K))  Equation 56R_c(:,1:K) is the first block column of R_c. F_(K)R_c(:,1:K) is preferably determined using K² parallel non-block DFTs of length D. In one implementation, the K² parallel non-block DFTs are implemented using K² parallel non-block prime factor algorithm fast Fourier transforms (PFA-FFTs) of length D.

Preferably to perform the block equalization of Equation 44, the matched filtering is approximated per Equation 57.

$\begin{array}{cc}\begin{array}{c}{A}^H\underset{\_}{r}\approx A_c^H{\underset{\_}{r}}_c\\ =\sum _{m=1}^M{A}^{{\left(m\right)}^H}{\underset{\_}{r}}_c^{\left(m\right)}\\ =\sum _{m=1}^M{\left(F_{\left(Q\right)}^H{\Lambda }_A^{\left(m\right)}{F}_{\left(K\right)}\right)}^H{\underset{\_}{r}}_c^{\left(m\right)}\\ =F_{\left(K\right)}^H\sum _{m=1}^M{\Lambda }_A^{{\left(m\right)}^H}{F}_{\left(Q\right)}{\underset{\_}{r}}_c^{\left(m\right)}\end{array}& \mathrm{Equation}57\end{array}$

The block-diagonalization of A_c^(m) is per Equation 58.A_c=└A_c^(1)TA_c^(2)T . . . A_c^(M)T┘^T  Equation 58

The estimation of the data vector {circumflex over (d)}, is per Equation 59.

$\begin{array}{cc}\begin{array}{c}\underset{\_}{\hat{d}}=R^{-1}{A}^H\underset{\_}{r}\\ \approx R_c^{-1}{A}_c^H{\underset{\_}{r}}_c\\ ={\left(F_{\left(K\right)}^H{\Lambda }_R{F}_{\left(K\right)}\right)}^{-1}{F}_{\left(K\right)}^H\sum _{m=1}^M\left({\Lambda }_A^{{\left(m\right)}^H}{F}_{\left(Q\right)}{\underset{\_}{r}}_c^{\left(m\right)}\right)\\ =F_{\left(K\right)}^H{\Lambda }_R^{-1}{F}_{\left(K\right)}{F}_{\left(K\right)}^H\sum _{m=1}^M\left({\Lambda }_A^{{\left(m\right)}^H}{F}_{\left(Q\right)}{\underset{\_}{r}}_c^{\left(m\right)}\right)\\ =F_{\left(K\right)}^H{\Lambda }_R^{-1}\sum _{m=1}^M\left({\Lambda }_A^{{\left(m\right)}^H}{F}_{\left(Q\right)}{\underset{\_}{r}}_c^{\left(m\right)}\right)\\ =F_{\left(K\right)}^H\underset{\_}{y}\end{array}& \mathrm{Equation}59\end{array}$

The vector y is of size DK by 1 and is per Equation 60.

$\begin{array}{cc}\begin{array}{c}\underset{\_}{y}={\Lambda }_R^{-1}\sum _{m=1}^M\left({\Lambda }_A^{\left(m\right)}{F}_{\left(Q\right)}{\underset{\_}{r}}_c^{\left(m\right)}\right)\\ ={\left[{\underset{\_}{y}}^{{\left(1\right)}^T}{\underset{\_}{y}}^{{\left(2\right)}^T}\dots {\underset{\_}{y}}^{{\left(D\right)}^T}\right]}^T\end{array}& \mathrm{Equation}60\end{array}$ y^(l), l=1, . . . D is a vector of size K by 1.

Preferably to determine y, F_(Q)r_c^(m) is determined using Q parallel non-block DFTs of a length D. In one implementation, the Q parallel non-block DFTs are implemented using Q parallel non-block PFA-FFTs of length D. Λ_R⁻¹ is a block diagonal matrix having blocks of size K by K in the main diagonal and is per Equation 61.

$\begin{array}{cc}{\Lambda }_R^{-1}=\left[\begin{array}{cccc}{\Lambda }_R^{\left(1\right)-1}& 0& \dots & 0\\ 0& {\Lambda }^{\left(2\right)-1}& \dots & 0\\ ⋮& ⋮& \ddots & ⋮\\ 0& 0& \dots & {\Lambda }_R^{\left(D\right)-1}\end{array}\right]& \mathrm{Equation}61\end{array}$ Each Λ_R^(l)−1, l=1, . . . , D, is a block of size K by K.

Preferably using the block diagonal structure of Λ_R⁻¹, y^(l) is determined by the Cholesky decomposition of Λ_R^(l) and forward and backward substitution in parallel. Alternately, Λ_R^(l) is directly inverted.

To perform the Cholesky decomposition, a vector

$\sum _{m=1}^M\left({\Lambda }_A^{{\left(m\right)}^H}{F}_{\left(Q\right)}{\underset{\_}{r}}_c^{\left(m\right)}\right)$ is divided into D blocks of a size K by 1, per Equation 62.

$\begin{array}{cc}\begin{array}{c}\underset{\_}{x}=\sum _{m=1}^M\left({\Lambda }_A^{{\left(m\right)}^H}{F}_{\left(Q\right)}{\underset{\_}{r}}_c^{\left(m\right)}\right)\\ ={\left[{\underset{\_}{x}}^{{\left(1\right)}^T}{\underset{\_}{x}}^{{\left(2\right)}^T}\dots {\underset{\_}{x}}^{{\left(D\right)}^T}\right]}^T\end{array}& \mathrm{Equation}62\end{array}$ A Cholesky factor G^(l) of Λ_R^(l) is determined using a factorization, per Equation 63.Λ_R^(l)=G^(l)G^(l)H  Equation 63Using the Cholesky factor G^(l), each y^(l) is determined by forward and backward substitution separately per Equations 64, 65 and 66.Λ_R^(l)y^(l)=G^(l)G^(l)Hy^(l)=x^(l)  Equation 64Forward Substitution: Find z^(l) in G^(l)z^(l)=x^(l), where z^(l)=G^(l)Hy^(l)  Equation 65Backward Substitution: Find y^(l) in z^(l)=G^(l)Hy^(l)  Equation 66

By performing a block inverse DFT of y, the data vector d is estimated as {circumflex over (d)}. Preferably, the block inverse DFT is implemented using K parallel non-block inverse PFA-FFTs of a length D.

FIG. 10 is a block diagram of a preferred implementation of efficient joint detection in a TDD/CDMA system. Although FIG. 10 illustrates using two sets of samples, the figure can be extended to other multiple sets. Using the received vector for each set of chip rate samples, r₁ and r₂, r_c⁽¹⁾ and r_c⁽²⁾ is formed by inserting zeros, by Extend blocks 232₁ and 232₂, respectively. A block DFT 200₁ and 200₂ using PFA-FFT of r_c⁽¹⁾ and r_c⁽²⁾ is then performed, to produce F_(Q)r_c⁽¹⁾ and F_(Q)r_c⁽²⁾.

Using the received training sequences, the channel impulse responses for each chip rate version of each transmitted burst, h^(k)(1) and h^(k)(2), is determined by estimate channel response blocks 202₁ and 202₂. Using each channelization code, c^(k) and the channel impulse response, h^(k)(1) and h^(k)(2), each system matrix, A⁽¹⁾ and A⁽²⁾, is determined by compute sub-system matrix blocks 204₁ and 204₂ per Equations 37 and 38.

To determine Λ_A⁽¹⁾ and Λ_A⁽²⁾, each system matrix, A⁽¹⁾ and A⁽²⁾, is extended by extend blocks 231₁ and 231₂. The first block column of each block-circulant matrix, A⁽¹⁾ and A⁽²⁾, is determined by selecting the first K columns of the extended A^(m) matrix by first block column blocks 214₁ and 214₂. By taking block DFTs 218₁ and 218₂, Λ_A⁽¹⁾ and Λ_A⁽²⁾ are determined using a PFA-FFT.

To determine Λ_R, the first block column of R is determined by compute first block column R block 240. The first column of R is extended by extend block 234. The first block column of the extended R is determined by a first block column determining device 208. The first block column of the extended R matrix is circularized, R_c, by a circularize block column block 210. By taking a block DFT by block DFT block 212, Λ_R^{( )} is determined using a PFA-FFT.

To efficiently compute the estimated data vector d, each of Λ_A⁽¹⁾, Λ_A⁽²⁾ and F(r_c⁽¹⁾) and F(r_c⁽²⁾) as well as Λ_R are used. Each of F(r_c⁽¹⁾), F(r_c⁽²⁾), Λ_A⁽¹⁾, Λ_A⁽²⁾ and Λ_R is divided into D blocks, by partition block 236. The complex conjugate transpose of each of Λ_A^(1,i) and Λ_A^(2,i), Λ_A^(1,i)H and Λ_A^(2,i)H, where i is the i^th block, is determined by transpose blocks 230, 231. A multiplier 228 multiplies Λ_A^(2,i)H by F_(Q)(r_c⁽²⁾)⁽ⁱ⁾. A multiplier 229 multiplies Λ_A^(1,i)H by F_(Q)(r_c⁽¹⁾)⁽ⁱ⁾. A summer 225 sums the multiplied results per Equation 62. Λ_R⁽ⁱ⁾ is decomposed using Cholesky decomposition 226, per Equation 63. By performing forward and backward substitution, per Equations 65 and 66, using forward and backward substitution blocks 224 and 222, respectively, F(d) is determined. Taking an inverse block DFT using PFA-FFT of F(d) by block inverse DFT block 220 using a PFA-FFT, {circumflex over (d)} is estimated.

Claims

1. A method of estimating a transmitted data vector d from a received vector r comprising: receiving a wireless communications signal and producing a vector r from the signal, where r has a form r=Ad+n and where A is a block Toeplitz matrix and n is a noise vector;producing a covariance matrix R of the form AHA+σ2I for a minimum mean square error block linear equalizer (MMSE-BLE) based solution, where AH is a Hermetian of A, σ2 is a noise variance and I is an identity matrix or AHA for a zero forcing block linear equalizer (ZF-BLE) based solution;extending the A matrix and R matrix;approximating the extended A and R matrices as block circulant matrices;determining a diagonal matrix of each of the extended and approximated A and R matrices, using the block column of the extended and approximated A and R matrices;extending r;taking a Fourier transform of r;taking products of the diagonal matrices and the extended r;summing the products; andestimating d using an inverse Fourier transform and the summed products.

2. The method of claim 1 wherein the A and R matrices and r are extended to be compatible with a prime factor algorithm fast Fourier transform.

3. The method of claim 1 further comprising performing LU decomposition on the diagonal of the R matrix and a forward substitution and a backward substitution devices for producing an inverse Fourier transform of d.

4. The method of claim 1 further comprising performing Cholesky decomposition on the diagonal of the R matrix and a forward substitution and a backward substitution devices for producing an inverse Fourier transform of d.