[0004] BACKGROUND [0005] Due to the increased demands for improved receiver performance, many advanced receivers use zero forcing (ZF) block linear equalizers and minimum mean square error (MMSE) equalizers.

[0006] In both these approaches, the received signal is typically modeled per Equation 1. r = Hd + n Equation 1 [0007] r is the received vector, comprising samples of the received signal. H is the channel response matrix. d is the data vector to be estimated.

In spread spectrum systems, such as code division multiple access (CDMA) systems, d may be represented as data symbols or a composite spread data vector. For a composite spread data vector, the data symbols for each individual code are produced by despreading the estimated data vector d with that code. n is the noise vector.

[0008] In a ZF block linear equalizer, the data vector is estimated, such as per Equation 2. d = (H"H)'H"r Equation 2 [0009] (-)"is the complex conjugate transpose (or Hermetian) operation.

In a MMSE block linear equalizer, the data vector is estimated, such as per Equation 3. d= (H"H+o'l)'H"r Equation 3 [0010] In wireless channels experiencing multipath propagation, to accurately detect the data using these approaches requires that an infinite number of received samples be used, which is not practical. Therefore, it is desirable to use an approximation technique. One of the approaches is a sliding window approach. In the sliding window approach, a predetermined window of received samples and channel responses are used in the data detection. After the initial detection, the window is slid down to a next window of samples. This process continues until the communication ceases.

[0011] By not using an infinite number of samples, an error is introduced into the symbol model shown in Equation 1 and, therefore causes inaccurate data detection. The error is most prominent at the beginning and end of the window, where the effectively truncated portions of the infinite sequence have the largest impact. One approach to reduce these errors is to use a large window size and truncate the results at the beginning and the end of the window.

The truncated portions of the window are determined in previous and subsequent windows. This approach has considerable complexity, especially when the channel delay spread is large. The large window size leads to large dimensions on the matrices and vectors used in the data estimation.

Additionally, this approach is not computationally efficient by detection data at the beginning and at the ends of the window and then discarding that data.

[0012] Accordingly, it is desirable to have alternate approaches to data detection.

[0013] SUMMARY [0014] The present invention has many aspects. One aspect of the invention is to perform equalization using a sliding window approach. A second aspect reuses information derived for each window for use by a subsequent window. A third aspect utilizes a discrete Fourier transform based approach for the equalization. A fourth aspect relates to handling oversampling of the received signals and channel responses. A fifth aspect relates to handling multiple reception antennas. A sixth embodiment relates to handling both oversampling and multiple reception antennas.

[0015] BRIEF DESCRIPTION OF THE DRAWING (S) [0016] Figure 1 is an illustration of a banded channel response matrix.

[0017] Figure 2 is an illustration of a center portion of the banded channel response matrix.

[0018] Figure 3 is an illustration of a data vector window with one possible partitioning.

[0019] Figure 4 is an illustration of a partitioned signal model.

[0020] Figure 5 is a flow diagram of sliding window data detection using a past correction factor.

[0021] Figure 6 is a receiver using sliding window data detection using a past correction factor.

[0022] Figure 7 is a flow diagram of sliding window data detection using a noise auto-correlation correction factor.

[0023] Figure 8 is a receiver using sliding window data detection using a noise auto-correlation correction factor.

[0024] Figure 9 is a graphical representation of the sliding window process.

[0025] Figure 10 is a graphical representation of the sliding window process using a circulant approximation.

[0026] Figure 11 is a circuit for an embodiment for detecting data using discrete Fourier transforms (DFTs).

[0027] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT (S) [0028] Although the features and elements of the present invention are described in the preferred embodiments in particular combinations, each feature or element can be used alone (without the other features and elements of the preferred embodiments) or in various combinations with or without other features and elements of the present invention.

[0029] Hereafter, a wireless transmit/receive unit (WTRU) includes but is not limited to a user equipment, mobile station, fixed or mobile subscriber unit, pager, or any other type of device capable of operating in a wireless environment.

When referred to hereafter, a base station includes but is not limited to a Node-B, site controller, access point or any other type of interfacing device in a wireless environment.

[0030] Although reduced complexity sliding window equalizer is described in conjunction with a preferred wireless code division multiple access communication system, such as CDMA2000 and universal mobile terrestrial system (UMTS) frequency division duplex (FDD), time division duplex (TDD) modes and time division synchronous CDMA (TD-SCDMA), it can be applied to various communication system and, in particular, various wireless communication systems. In a wireless communication system, it can be applied to transmissions received by a WTRU from a base station, received by a base station from one or multiple WTRUs or received by one WTRU from another WTRU, such as in an ad hoc mode of operation.

[0031] The following describes the implementation of a reduced complexity sliding window based equalizer using a preferred MMSE algorithm.

However, other algorithms can be used, such as a zero forcing algorithm. h () is the impulse response of a channel. d (k) is the k th transmitted sample that is generated by spreading a symbol using a spreading code. It can also be sum of the chips that are generated by spreading a set of symbols using a set of codes, such as orthogonal codes. r (.) is the received signal. The model of the system can expressed as per Equation 4. M r (t) = E d (k) h (t-kTc) + n (two < t < oo Equation 4 k=- [0032] n (t) is the sum of additive noise and interference (intra-cell and inter-cell). For simplicity, the following is described assuming chip rate sampling is used at the receiver, although other sampling rates may be used, such as a multiple of the chip rate. The sampled received signal can be expressed as per Equation 5. oxo A=-°o A=-co équations T is being dropped for simplicity in the notations.

[0033] Assuming h () has a finite support and is time invariant. This means that in the discrete-time domain, index L exists such that h (i) = 0 for i < 0 and !'> Z. As a result, Equation 5 can be re-written as Equation 6. L-1 r (j) = I h (k) d (j-k) + n (j) j E f...,-2,-1, 0, 1, 2,...} Equation 6 k=O [0034] Considering that the received signal has M received signals \ r (0),###,r(M-1), Equation 7 results. r=Hd+n where, r=[r(0),###,r(M-1)]T # CM, d = [d(-L+1),d(-L+2),...,d(0),d(1),...,d(M-1)]T # CM+L=1 n = [7Z (0),###,n(M_1)]T # CM Equation 7 [0035] In Equation 7, CM represents the space of all complex vectors with dimension M.

[0036] Part of the vector d can be determined using an approximate equation. Assuming M > L and defining N = M-L + 1, vector d is per Equation 8.

Equation 8 [0037] The H matrix in Equation 7 is a banded matrix, which can be represented as the diagram in Figure 1. In Figure 1, each row in the shaded area represents the vector [h (L-1), h (L-2),..., h (1), h (0)], as shown in Equation 7.

[0038] Instead of estimating all of the elements in d, only the middle N elements of d are estimated. d is the middle N elements as per Equation 9.

#=[d(0),...,d(N-1)]T Equation 9 [0039] Using the same observation for r, an approximate linear relation between r and d is per Equation 10. r=Hd+n Equation 10 [0040] Matrix H can be represented as the diagram in Figure 2 or as per Equation 11. h (0) 0 h (l) h (0) <BR> <BR> <BR> <BR> # h(1) 0<BR> <BR> <BR> <BR> <BR> <BR> <BR> H= h (L-1).. h (0) 0 h (L-1). h (1) <BR> <BR> <BR> <BR> . 0..<BR> <BR> <BR> <BR> <BR> <BR> <BR> <BR> <P> .. h (L-1) Equation 11 [0041] As shown, the first L-1 and the last L-1 elements of r are not equal to the right hand side of the Equation 10. As a result, the elements at the two ends of vector d will be estimated less accurately than those near the center.

Due to this property, a sliding window approach, as described subsequently, is preferably used for estimation of transmitted samples, such as chips.

[0042] In each kth step of the sliding window approach, a certain number of the received samples are kept in r [k] with dimension N+L-1. They are used to estimate a set of transmitted data d [k] with dimension N using equation 10.

After vector d [k] is estimated, only the"middle"part of the estimated vector d [k] is used for the further data processing, such as by despreading. The"lower"part (or the later in-time part) of d [k] is estimated again in the next step of the sliding window process in which r [k+1] has some of the elements r [k] and some new received samples, i. e. it is a shift (slide) version of r [k].

[0043] Although, preferably, the window size N and the sliding step size are design parameters, (based on delay spread of the channel (L), the accuracy requirement for the data estimation and the complexity limitation for implementation), the following using the window size of Equation 12 for illustrative purposes.

N=4NS xSF Equation 12 SF is the spreading factor. Typical window sizes are 5 to 20 times larger than the channel impulse response, although other sizes may be used.

[0044] The sliding step size based on the window size of Equation 12 is, preferably, 2Ns x SF. Ns E {1, 2,...} is, preferably, left as a design parameter. In addition, in each sliding step, the estimated chips that are sent to the despreader are 2Ns x SF elements in the middle of the estimated d [k]. This procedure is illustrated in Figure 3.

[0045] In the sliding window approach described above, the system model is approximated by throwing away some terms in the model. In the following, a technique is described where terms are kept by either using the information estimated in previous sliding step or characterizing the terms as noise in the model. The system model is corrected using the kept/characterized terms.

[0046] One algorithm of data detection uses an MMSE algorithm with model error correction uses a sliding window based approach and the system model of Equation 10.

[0047] Due to the approximation, the estimation of the data, such as chips, has error, especially, at the two ends of the data vector in each sliding step (the beginning and end). To correct this error, the H matrix in Equation 7 is partitioned into a block row matrix, as per Equation 13, (step 50).

H = [H, H,] Equation 13 [0048] Subscript"p"stands for"past", and"f'stands for"future". H is as per Equation 10. Hp is per Equation 14.

Equation 14 [0049] Ho ils per Equation 15.

Equation 15 [0050] The vector d is also partitioned into blocks as per Equation 16.

Equation 16 [0051] d is the same as per Equation 8 and dp is per Equation 17. dp=[d(-L+1) d(-L+2) ### d(-1)]T # CL-1 Equation 17 [0052] du vis per Equation 18. df=[d(N) d(N+1) ### d(N+L-2)]T # CL-1 Equation 18 [0053] The original system model is then per Equation 19 and is illustrated in Figure 4. r=Hpdp+Hd+H+n Equation 19 [0054] One approach to model Equation 19 is per Equation 20. r =Hd+ñl where #=r-Hpdp and 31=Hfdf+n Equation 20 [0055] Using an MMSE algorithm, the estimated data vector d is per Equation 21.

#=gd#H(gd##H+#1)-1# Equation 21 [0056] In Equation 21, gd is chip energy per Equation 22.

E{d(i)d*(j)}=gd#ij Equation 22 [0057] r is per Equation 23. r =r-Hpdp Equation 23 [0058] dp, is part of the estimation of d in the previous sliding window step. El is the autocorrelation matrix of ñl, i. e., El =E{#1#1H}. If assuming Hf df and n are uncorrelated, Equation 24 results. l gdHfHf + E {nn} Equation 24 [0059] The reliability of d depends on the sliding window size (relative to the channel delay span L) and sliding step size.

[0060] This approach is also described in conjunction with the flow diagram of Figure 5 and preferred receiver components of Figure 6, which can be implemented in a WTRU or base station. The circuit of Figure 6 can be implemented on a single integrated circuit (IC), such as an application specific integrated circuit (ASIC), on multiple IC's, as discrete components or as a combination of IC ('s) and discrete components.

[0061] A channel estimation device 20 processes the received vector r producing the channel estimate matrix portions, Hp, H and Hf, (step 50). A future noise auto-correlation device 24 determines a future noise auto-correlation factor, gdH HH, (step 52). A noise auto-correlation device 22 determines a noise auto-correlation factor, E {nnH}, (step 54). A summer 26 sums the two factors together to produce E,, (step 56).

[0062] A past input correction device 28 takes the past portion of the channel response matrix, H p, and a past determined portion of the data vector, dp, to produce a past correction factor, Hp#p, (step 58). A subtractor 30 subtracts the past correction factor from the received vector producing a modified received vector, r, (step 60). An MMSE device 34 uses Xj, H, and r to determine the received data vector center portion d, such as per Equation 21, (step 62). The next window is determined in the same manner using a portion of d as dp in the next window determination, (step 64). As illustrated in this approach, only data for the portion of interest, d, is determined reducing the complexity involved in the data detection and the truncating of unwanted portions of the data vector.

[0063] In another approach to data detection, only the noise term is corrected. In this approach, the system model is per Equation 25. r=Hd+n2, where n2=Hpdp+Hfdf +n Equation 25 [0064] Using an MMSE algorithm, the estimated data vector d is per Equation 26.

#=gd#H(gd##H+#2)-1r Equation 26 [0065] Assuming Hpdp, Hfdf andnare uncorrelated, Equation 27 results.

E, =GdHpHpH+gdHfHfH+E{nnH} Equation 27 [0066] To reduce the complexity in solving Equation 26 using Equation 27, a full matrix multiplication for HpHpH and H H H are not necessary, since only the upper and lower corner of Hp and Hf, respectively, are non-zero, in general.

[0067] This approach is also described in conjunction with the flow diagram of Figure 7 and preferred receiver components of Figure 8, which can be implemented in a WTRU or base station. The circuit of Figure 8 can be implemented on a single integrated circuit (IC), such as an application specific integrated circuit (ASIC), on multiple IC's, as discrete components or as a combination of IC ('s) and discrete components.

[0068] A channel estimation device 36 processes the received vector producing the channel estimate matrix portions, Hp, H and Hf, (step 70). A noise auto-correlation correction device 38 determines a noise auto-correlation correction factor, gdHpHpH + gdHf HfH, using the future and past portions of the channel response matrix, (step 72). A noise auto correlation device 40 determines a noise auto-correlation factor, E {nnH}, (step 74). A summer 42 adds the noise auto-correlation correction factor to the noise auto-correlation factor to produce EZ, (step 76). An MMSE device 44 uses the center portion or the channel response matrix, H, the received vector, r, and E2 to estimate the center portion of the data vector, d, (step 78). One advantage to this approach is that a feedback loop using the detected data is not required. As a result, the different slided window version can be determined in parallel and not sequentially.

[0069] Discrete Fourier Transform Based Equalization [0070] The sliding window approach described above requires a matrix inversion, which is a complex process. One embodiment for implementing a sliding window utilizes discrete Fourier transforms (DFTs), as follows. Although the preferred implementation of the DFT based approach is with a MMSE algorithm, it can be applied to other algorithms, such as a zero forcing (ZF) based algorithm.

[0071] A matrix ACir E CNXN, for some integer N, is a circulant matrix if it has the following form per Equation 28.

Equation 28 [0072] This kind of matrix is expressed using the DFT and the IDFT operators, such as per Equation 29.

ACir = #(Acir[:,1])FN where, Acir[:,1]=(a0,a1,...,aN)T # CN, i.e. it is the first column of matrix Acir Equation 29 Columns other than the first column can be used if properly permuted. FN is the N-point DFT matrix which is defined as, for any x E CN, as per Equation 30.

Equation 30 Fun'ils the N-point inverse DFT matrix which is defined as, for any x E CN, as per Equation 31.

Equation 31 AN () is a diagonal matrix, which is defined as, for any x e Con, as per Equation 32.

AN (x) = diag (FNx) Equation 32 [0073] The inverse of matrix Acir is expressed, such as per Equation 33.

Acir-1=FN-1#N-1(Acir[:,1])FN Equation 33 [0074] The following is an application of a DFT based approach to the data estimation process using the sliding window based chip level equalizer. The first embodiment uses a single receiving antenna. Subsequent embodiments use multiple receiving antennas.

[0075] The receiver system is modeled as per Equation 34.

Equation 34 h zu is the impulse response of the channel. d (k) is the k th transmitted chip samples that is generated by spreading symbols using a spreading code. r(#) is the received signal. n () is the sum of additive noise and interference (intra-cell and inter-cell).

[0076] Using chip rate sampling and h () having a finite support, which means, in discrete-time domain, there is an integer L such that h (i) = 0 for i < 0 and i#L, the sampled received signal can be expressed (Tc is dropped for simplicity of the notations), as per Equation 35.

Equation 35 [0077] Based on M (At > L) received signals r (0),###,r(M-1), Equation 36 results. r=Hd+n where r =[r(0),###,r(M-1)]T # CM, <BR> <BR> <BR> d=[d(-L+1),d(-L+2),...,d(0),d(1),...d(M-1)]T # CM+L=1<BR> <BR> <BR> <BR> <BR> <BR> n=[n(0),###,n(M-1)]T # CM Equation 36 [0078] As illustrated by Equation 36, the H matrix is Toeplitz. As described subsequently in the application for multiple chip rate sampling and/or multiple receive antennas, the H matrix is block Toeplitz. Using the block Toeplitz property, discrete Fourier transform techniques can be applied. The Toeplitz/block Toeplitz nature is produced as a result of the convolution with one channel or the convolution of the input signal with a finite number of effective parallel channels. The effective parallel channels appear as a result of either oversampling or multiple receive antennas. For one channel, a single row is essentially slide down and to the right producing a Toeplitz matrix.

[0079] The statistics of the noise vector are treated as having the autocorrelation property, per Equation 37.

Efn nH} =a-zI Equation 37 [0080] The left hand side of equation (5) can be viewed as a"window"of continuous input signal stream. To estimate the data, an approximated model is used. In this approximated model, the first L-1 and the last L-1 elements of vector d are assumed to be zero before applying the MMSE algorithm and the reset M-L +1 elements of d forms a new vector d = [d (0),..., df M-L + 1)] T. The approximated model can be expressed explicitly as per Equation 38. r=Hd+n Equation 38 [0081] After the vector d is estimated, only the middle part of it is taken for de-spreading. Subsequently, the window of observation (i. e. the received signal) is slid by (M-L+1)/2 elements and the process is repeated.

Figure 9 is a graphical representation of the sliding window process, as described above.

[0082] Using MMSE algorithm, the estimated data is expressed per Equation 39. d = R-IHHr where R=#H#+#2I Equation 39 [0083] In Equation 39, neither the matrix R nor the matrix H is circulant to facilitate a DFT implementation. To facilitate a DFT implementation, for each sliding step, the approximated system model per Equation 40 is used.

Equation 40 In Equation 40, only the first L-1 elements (equations) are approximations of those of Equation 36.

[0084] The matrix H is replaces by a circulant matrix, such as per Equation 41.

Equation 41 [0085] The system model, for each sliding step, is per Equation 42. r =Hcird+n with d = [d (O),..., d (M-1)] T E CMx' Equation 42 The vector d in Equation 42, due to the new model, is different than the vector d in Equation 36. Equation 42 adds additional distortion to the first L-1 element of Equation 39. This distortion makes the two ends of the estimated vector d inaccurate. Figure 10 is a graphical representation of the model construction process.

[0086] Using approximated model per Equation 42, the MMSE algorithm yields the estimated data as per Equation 43. d = Rcir-1HcirHr where Rir r=HcirHHcir+#2I Equation 43 Both HcirH and Rcir are circulant and RC, r is of the form per Equation 44.

Equation 44 [0087] Applying the properties of circulant matrices, the estimated data is per Equation 45.

#=FM-1#M-1(Rcir[:,1])#M(HcirH[:,1])FMr Equation 45 [0088] Figure 11 is a diagram of a circuit for estimating the data per Equation 45. The circuit of Figure 11 can be implemented on a single integrated circuit (IC), such as an application specific integrated circuit (ASIC), on multiple IC's, as discrete components or as a combination of IC ('s) and discrete components.

[0089] The estimated channel response H is processed by an H determination device 80 to determine the Toeplitz matrix H. A circulant approximation device 82 processes H to produce a circulant matrix HC, r. A Hermetian device 84 produces the Hermetian of H, H.. Using HC, r, HCH, r and the noise variance #2, Rcir is determined by a RC, r determining device 86. Using a first column of HH, a diagonal matrix is determined by a AM (hoir determining device 88. Using a first column of RC, r, an inverse diagonal matrix is determined by a AM (RCir [:, l]) determination device 90. A discrete Fourier transform device 92 performs a transform on the received vector, r. The diagonal, inverse diagonal and Fourier transform result are multiplied together by a multiplier 96. An inverse Fourier transform device 94 takes an inverse transform of the result of the multiplication to produce the data vector d.

[0090] The sliding window approach is based on an assumption that the channel is invariant within each sliding window. The channel impulse response near the beginning of the sliding window may be used for each sliding step.

[0091] One preferred approach for determining the window step size NSS and window size M is per Equation 46, although others may be used.

NSS=2Nsymbol#SF and M = 4Nsymbol x SF Equation 46 Nsymbol#{1, 2,...} is the number of symbols and is a design parameter which should be selected, such that M > L. Since M is also the parameter for DFT which may be implemented using FFT algorithm. M may be made large enough such that the radix-2 FFT or a prime factor algorithm (PFA) FFT can be applied. After the data is estimated, 2Nsymbol# SF samples are taken to process despreading starting from Nsymbol X SFth sample. Figure 11 is an illustration of taking the samples for despreading.

[0092] Multiple Receive Antenna Equalization [0093] The following is an embodiment using multiple receive antennas, such as Kreceive antennas. Samples of the received vector and estimates of the channel impulse response are taken for each antenna independently. Following the same process as for the single antenna embodiment, each antenna input, rk, is approximated per Equation 47. rk Hcir1kd+nk for k = 1,..., K Equation 47 or in block matrix form per Equation 48 Equation 48 [0094] Equations 49 and 50 are estimates of the auto-correlation and cross-correlation properties of the noise terms.

E {nknk} =#2I for k=1,...,K Equation 49 and E{nknjH}=0 for k#j Equation 50 [0095] Applying MMSE algorithm, the estimated data can be expressed as per Equation 51.

Equation 51 R, ir is still a circulant matrix and the estimated data can be determined per Equation 52.

Equation 52 [0096] If the receive antennas are positioned close to each other, the noise terms may be correlated in both time and space. As a result, some degradation in the performance may result.

[0097] Multiple Chip Rate Sampling (Oversampling) Equalization [0098] The following describes embodiments using a sliding window based equalization approach with multiple chip rate sampling. Multiple chip rate sampling is when the channel is sampled at a sampling rate which is an integer multiple of the chip rate, such as two times, three times, etc. Although the following concentrates on two times per chip sampling, these approaches can be applied to other multiples.

[0099] Using a sliding window of width of N chips and two time chip rate sampling, our received vector is r=[r0,r1,...,r2N-1]T. This vector may be rearranged and separated into an even received vector re = [rO, r2,..., r2N 2] T and an odd received vector ro=[r1,r3,...,r2N-1]T, with r= [re, rO] T. Without loss of generality, the data transmission model is per Equation 53.

Equation 53 Equation 53 separates the effective 2-sample-per-chip discrete-time channel into two chip-rate discrete-time channels.

[00100] The matrices He and Ho in Equation 53 are, correspondingly, the even and odd channel response matrices. These matrices are constructed from the even and odd channel response vectors he and ho, which are obtained by sampling the channel response at 2 samples per chip and separating it into the even and odd channel response vectors.

[00101] The channel noise is modeled as white with a variance o-2, as per Eqaution 54.

E [neneH] = E [nonoH] = a. zI Equation 54 [00102] If the channel is an additive white Gaussian noise (AWGN) channel and the received data is provided directly from the sampled channel, then Equation 55 results.

E[nenoH]=0 Equation 55 [00103] As a result, the problem is mathematically similar to the case of the chip-rate equalizer for 2 receive antennas with uncorrelated noise, as previously described. However, the received antenna signals in many implementations are processed by a receive-side root-raised cosine (RRC) filter before being provided to the digital receiver logic for further processing.

Following such processing, the received noise vector is no longer white, but has a raised-cosine (RC) autocorrelation function. RC is the frequency-domain square of a RRC response. Since the RC pulse is a Nyquist pulse, Equation 54 holds, however Equation 55 does not. The (i) th element of the matrix def 1 #cross = E[nenoH] is per Equation 56.

#² 1/#2E[nenoH](i,j)=xRC(#i-j#+0.5) Equation 56 XRc is the unity-symbol-time normalized RC pulse shape.

[00104] Properties of A are it is real, symmetric and Toeplitz; it is not banded and has no zero entries and its entries do get smaller and tend to 0 as they get farther and farther away from the main diagonal.

[00105] En represent the cross-correlation matrix of the total noise vector and is per Equation 57.

Equation 57 [00106] Exact Solution [00107] The exact solution to the problem of linear minimum mean- square estimation of d from the observation of r is per Equation 58. dMMSE =(HH#n-1H+I)-1HH#n-1r where y = HHS"lr is the whitening matched filtering (WMF) #MMSE=(HH#n-1H+I)-1y is the linear MMSE equalization Equation 58 [00108] Neither HH#n-1 nor HH#n-1H+I are Toeplitz and neither can be made Toeplitz through elemental unitary operations (e. g. row/column re- arrangements), due to the structure of E,,. Accordingly, DFT-based methods based on circulant approximations of Toeplitz matrices cannot be applied here and an exact solution is highly complex.

[00109] Two embodiments for deriving an efficient algorithm for solving this problem are described. The first embodiment uses a simple approximation and the second embodiment uses an almost-exact solution.

[00110] Simple Approximation [00111] The simple approximation ignores the correlation between ne and no, Across = 0. As a result, the same approach as multiple chip-rate receive antennas is used.

[00112] The complexity of this simple approximation approach is as follows. N-chip data blocks are considered. For rough approximation, anN-point DFT complexity, given by NlogN operations per second (ops), is assumed.

Additionally, N-point vector multiplications are assumed to take N ops and vector additions are ignored.

[00113] The complexity of the DFT-based approach can be roughly partitioned into 2 components: the processing which has to be performed on every received data set and the processing which is performed when the channel estimate is updated, which is typically done one to two orders of magnitude less frequently then the former operation.

[00114] For processing performed on each received data set, the following operations are performed: 2 N-point DFTs to transform the received vector into the frequency domain; 2 N-point vector multiplications (multiply each received vector by the appropriate"state"vector"); and one more DFT to transform the result back into time domain. Thus, the approximate complexity is per Equation 59.

Cl,, = 3N log N + 2N Equation 59 [00115] For processing performed when the channel response is updated, the following operations are performed: 2 DFT operations, 6 N-point vector multiplies and a vector division, which need to be taken 10 times the operations of a vector multiply. Thus, the complexity of this step is roughly given per Equation 60.

Cl,, = 2N log N + 16N Equation 60 [00116] Almost Exact Solution [00117] For the almost-exact solution which uses a block-Toeplitz solution, the vector and matrices are rearranged in their natural order, such that the vector r is given by r=[r0,r1,...,r2N-1]T. Equation 61 is the natural order model. r=HbTd+n where HbT is defined as Equation 61 he, i is the ith row of He and ho, i is the ith row of Ho. Gi is a 2xN matrix whose Ist row is he, i and whose 2nd row is ho, i. Using Gi oxyl as the row-x, column-y element of Gi, HbT is block-Toeplitz as illustrated in Equation 62.

Gi[x,y]=Gj[x,y+(i-j)] provided that 1 < y + (i-j) # N Equation 62 [00118] The block-Toeplitz structure of HbT follows immediately from the Toeplitz structure of He and Ho and the row-rearrangement. From the Toeplitz structure of I and #cross, the autocorrelation matrix of the noise in the re-defined problem is also block Toeplitz. Because this matrix is also symmetric, it can be rewritten per Equation 63. where E i j are 2x2 matrices with the property that S Equation 63 [00119] Subsequently, block-circulant approximations to the block- Toeplitz matrices are produced. Since the HbT matrix is also banded, the block circulant approximation of HbT is then obtained directly. However, EbT is not banded and therefore it is not possible to produce a block-circulant approximation directly from it. Since the elements of AbT tend to 0 as they get farther away from the main diagonal, a banded approximation to EbT is per Equation 64. where #i,j are 2x2 matrices with the property that #i,j = ##i-j# if #i-j##Bn and 1,, j = 0 otherwise Equation 64 The noise-covariance-bandwidth, Bn, is a design parameters that is selected. Due to the decay properties of the RC pulse shape, it is likely to be only several chip.

Now £bT is banded block-Toeplitz and a circulant approximation to it is produced.

[00120] The circulant approximations of HbT and £bT are Hbc and EbC respectively. Wn denotes the n-point DFT matrix, that is if x is an n-vector, then xf=Wnx is the DFT of x. A block-circulant matrix C is of the form of Equation 65. where Ci is an NxN matrix and therefore C is an MNxMN matrix Equation 65 [00121] C can also be written as Equation 66.

C=WM#N-1#M#N(C)WM#N where N is the block-N-DFT matrix defined as WM#N = WM#IN Equation 66 AM-N (C) is a block diagonal matrix that depends on C and is given by Equation 67.

Equation 67 A, (C) is an NxN matrix. To completely specify A, (C), #i,(k,l) denotes the (k,l) th def element of A, (C) and is defined as #(k,l)=[#1,(k,l),32,(k,l),...,#M,(k,l)]T# ci,(k,l) denotes dey the (k,l)th element of C and is defined as c(k,l)=[c1,(k,l), c2,(k,l),...,cM,(k,l)]T. #(k,l) is the M-point DFT of c (k, l) and is per Equation 68. k (k, l)=WMc(k,l) Equation 68 [00122] Equations 66-68 specify the block-DFT representation of square block circulant matrices. N2 DFTs are required to compute AMXN (C).

[00123] The MMSE estimator is rewritten per Equation 69. dMMSE-HHlrn +HHH) 1P Equation 69 [00124] The MMSE estimator form as per Equation 68 has several advantages. It requires only a single inverse matrix computation and thus in the DFT domain only a single vector division. This provides a potentially significant savings as divisions are highly complex.

[00125] The almost-exact solution has two steps in the most preferred embodiment, although other approaches may be used. Every time a new channel estimate is obtained, the channel filter is updated, (HH (En +HHH)-1 is determined). For every data block, this filter is applied to the received data block. This partition is utilized because the channel is updated very infrequently compared to the received data block processing and therefore significant complexity reduction can achieved by separating the overall process into these two steps.

[00126] The DFT of E is the DFT of the pulse shaping filter multiplied by the noise variance C2. Since the pulse shaping filter is typically a fixed feature of the system, its DFT can be precomputed and stored in memory and thus only the value #2 is updated. Since the pulse-shaping filter is likely to be close to the"ideal" (IIR) pulse shape, the DFT of the ideal pulse shape can be used for En reducing the complexity and is also far away from the carrier.

[00127] To channel update step, the following is performed: 1. The"block-DFT"of H needs to be computed. Since the block is of width 2, it requires 2 DFTs. The result is a Nx2 matrix whose rows are the DFTs of he and ho.

2. The"block-DFT"of HHH is computed by finding element-by-element autocorrelations and the crosscorrelation of he and ho. This required 6N complex multiplies and 2N complex adds: the products of N 2x2 matrices are computed with there own Hermitian transposes.

3. The block-DFT of Sn iS added, which requires 3N multiplies (scale the stored block-DFT of the RRC filter by C2) and 3N adds to add the block-DFT of the two matrices.

4. An inverse of Eo + HHH is taken in the block-DFT domain. To do this an inverse of each of the N 2x2 matrices is taken in the block-DFT domain. To estimate the total number of operations, consider a Hermitian matrix The inverse of this matrix is given per Equation 70.

Equation 70 Accordingly, the complexity of computing each inverse involves 3 real multiplications and 1 real subtraction (roughly 1 complex multiply) and 1 real division.

5. The result are block-multiplied by the block-DFT of HH, which, takes a total of 8N multiplies + 4N adds (since HH is not Hermitian).

[00128] To summarize, the following computation are required: 2 N-point DFTs ; 18N complex multiplies (17 N-point vector multiplies + N stand-alone multiplies); 11N complex adds (11 N-point vector adds); and 1N real divisions.

[00129] The complexity of processing a data block r of 2N values (N chips long) involves: 2 N-point DFTs ; one product of the N-point block-DFTs (filter and data), which required 8N complex multiplies and 4N complex adds; and 1 N-point inverse DFTs.

[00130] To summarize, the following is required: 3 N-point DFTs ; 8N complex multiplies (8 N-point vector multiplies); and 4N complex adds (4 N-point vector adds).

[00131] Multiple Chip Rate Sampling and Multiple Receive Antenna Equalization [00132] The following are embodiments using multiple chip rate sampling and multiple receive antennas. For L receive antennas, 2L channel matrices- one"even"and one"odd"matrix for each antenna result. The channel matrices for Ith antenna are denoted as Hi, e and Hi, o and ho ein and hl, o, n denote the nth row of such matrix. Each channel matrix is Toeplitz and with the appropriate re- arrangement of rows the joint channel matrix is a block-Toeplitz matrix, per Equation 71.

Equation 71 The matrices Gi are the Toeplitz blocks of HbT. Each Gi is a 2LxN matrix.

[00133] Estimating the vector d from the received observations r can be modeled per Equation 72. r HbTd+n Equation 72 [00134] The MMSE estimation formulation is per Equation 73. dMMSE HbT (n + HbTHbT) Ir Equation 73 'n iS the covariance of the noise vector n. The form that the solution of Equation 73 depends on the assumptions that are made for En. The introduction of the multiple receive antenna introduces an additional spatial dimension. Although the interplay of temporal and spatial correlations can be extremely complex, it can be assumed that the spatial correlation properties of the noise do not interplay with the temporal correlation properties, except as a direct product of the two, as per Equation 74.

#n=#n,1ant # #sp Equation 74 Sn tant is the noise covariance matrix of the noise observed at a single antenna as per Equation 57. En tant iS of dimension 2Vx2N. Sgpis the normalized synchronous spatial covariance matrix, i. e. it is the covariance matrix between the L noise samples observed at the L antennas at the same time normalized to have 1's on the main diagonal 0 denotes the Kroenecker product.

[00135] S is a 2LNx2LN Hermitian positive semi-definite matrix, which is block-Toeplitz with 2Lx2L blocks. To estimate the data, four preferred embodiments are described: an exact solution; a simplification by assuming that the L receive antenna have uncorrelated noise; a simplification by ignoring the temporal correlation of the noise in the odd and even streams from the same antenna; and a simplification by assuming that all 2L chip-rate noise streams are uncorrelated.

[00136] The complexity of DFT-based processing using the circulant approximation may be partitioned into two components: the processing of channel estimation which need not be done for every new data block and the processing of data itself which is performed for every data block. In all four embodiments, the complexity of processing data involves: 2L forward N-point DFTs ; 2LN complex multiplies; and 1 inverse N-point DFT. The complexity of processing the channel estimate varies for each embodiment.

[00137] In the case of the exact MMSE solution, the complexity of computing the"MMSE filter"from the channel estimate is as follows : 2L N-point DFT's ; N 2Lx2L matrix products + N 2Lx2L matrix additions to compute (; n + HbTHbTH) ; N 2Lx2L matrix inverses to compute the inverse of gFn + HbTHbTH) ; andN2Lx2L matrix products to produce the actual filter.

[00138] A major contributor to the overall complexity of this process is the matrix inverse step in which an inverse of 2Lx2L matrices has to be taken. It is precisely this complexity that can be reduced by various assumptions on the uncorrelated nature of the noise, as follows : 1. If it is assumed that the noise is uncorrelated both temporally (odd/even samples) and spatially (across antennas), then En reduces to a diagonal matrix and the problem is identical to single-sample-per-chip sampling with 2L antennas with spatially uncorrelated noise. As a result, the operation of matrix inverse simply reduces to a division since all the matrices involved are Toeplitz.

2. If it is assumed that the noise is spatially uncorrelated, then the matrix inverses involved are those of 2x2 matrices.

3. If it is assumed that a temporal uncorrelation of odd/even streams but a spatial noise correlation is retained, the matrix inverses involved are LxL.

* * *