A Fast LMMSE Channel Estimation Method for OFDM Systems

The present channel estimation methods generally can be divided into two kinds. One kind is based on the pilots [6‐9], and the other is blind channel estimation [10‐12] which does not use pilots. Blind channel estimation methods avoid the use of pilots and have higher spectral efficiency. However, they often suffer from high computation complexity and low convergence speed since they often need a large amount of receiving data to obtain some statistical information such as cyclostationarity induced by the cyclic prefix. Therefore, blind channel estimation methods are not suitable for applications with fast varying fading channels. And most practical communication systems such as World Interoperability for Microwave Access (WIMAX) system adopt pilot assisted channel estimation, so this paper studies the first kind.

For the pilot-aided channel estimation methods, there are two classical pilot patterns, which are the block-type pattern and the comb-type pattern [4]. The block-type refers to that the pilots are inserted into all the subcarriers of one OFDM symbol with a certain period. The block-type can be adopted in slow fading channel, that is, the channel is stationary within a certain period of OFDM symbols. The comb-type refers to that the pilots are inserted at some specific subcarriers in each OFDM symbol. The comb-type is preferable in fast varying fading channels, that is, the channel varies over two adjacent OFDM symbols but remains stationary within one OFDM symbol. The comb-type pilot arrangement-based channel estimation has been shown as more applicable since it can track fast varying fading channels, compared with the block-type one [4, 13]. The channel estimation based on comb-type pilot arrangement is often performed by two steps. Firstly, it estimates the channel frequency response on all pilot subcarriers, by lease square (LS) method, LMMSE method, and so on. Secondly, it obtains the channel estimates on all subcarriers by interpolation, including data subcarriers and pilot subcarriers in one OFDM symbol. There are several interpolation methods including linear interpolation method, second-order polynomial interpolation method, and phase-compensated interpolation [4].

In [14], the linear minimum mean square error (LMMSE) channel estimation method based on channel autocorrelation matrix in frequency domain has been proposed. To reduce the computational complexity of LMMSE estimation, a low-rank approximation to LMMSE estimation has been proposed by singular value decomposition [6]. The drawback of LMMSE channel estimation [6, 14] is that it requires the knowledge of channel autocorrelation matrix in frequency domain and the signal to noise ratio (SNR). Though the system can be designed for fixed SNR and channel frequency autocorrelation matrix, the performance of the OFDM system will degrade significantly due to the mismatched system parameters. In [15], a channel estimation exploiting channel correlation both in time and frequency domain has been proposed. Similarly, it needs to know the channel autocorrelation matrix in frequency domain, the Doppler shift, and SNR in advance. Mismatched parameters of the Doppler shift and the delay spread will degrade the performance of the system [16]. It is noted that the channel estimation methods proposed in [6, 14‐16] can be adopted in either the block-type pilot pattern or the comb-type pilot pattern.

When the assumption that the channel is time-invariant within one OFDM symbol is not valid due to high Doppler shift or synchronization error, the intercarrier interference (ICI) has to be considered. Some channel estimation and signal detection methods have been proposed to compensate the ICI effect [17, 18]. In [17], a new equalization technique to suppress ICI in LMMSE sense has been proposed. Meanwhile, the authors reduced the complexity of channel estimator by using the energy distribution information of the channel frequency matrix. In [18], the authors proposed a new pilot pattern, that is, the grouped and equispaced pilot pattern and corresponding channel estimation and signal detection to suppress ICI.

In this paper, the OFDM system framework based on comb-type pilot arrangement is adopted, and we assume that the channel remains stationary within one OFDM symbol, and therefore there is no ICI effect. We propose a fast LMMSE channel estimation method. The proposed method has three advantages over the conventional LMMSE method. Firstly, the proposed method does not require the knowledge of channel autocorrelation matrix and SNR in advance but can achieve almost the same performance with the conventional LMMSE channel estimation in terms of the normalized mean square error (NMSE) of channel estimation and bit error rate (BER). Secondly, the proposed method needs only fast Fourier transform (FFT) operation instead of the inversion operation of a large dimensional matrix. Therefore, the computational complexity can be reduced significantly, compared with the conventional LMMSE method. Thirdly, the proposed method can track the changes of channel parameters, that is, the channel autocorrelation matrix and SNR. However, the conventional LMMSE method cannot track the channel. Once the channel parameters change, the performance of the conventional LMMSE method will degrade due to the parameter mismatch.

The paper is organized as follows. Section 2 describes the OFDM system model. Section 3 describes the proposed fast LMMSE channel estimation. We analyze the mean square error (MSE) of the proposed fast LMMSE channel estimation and the MSE of the conventional LMMSE channel estimation in Section 4. The simulation results and numerical results of the proposed algorithm are discussed in Section 5 followed by conclusion in Section 6.

The channel impulse response in time domain can be expressed as

where is the complex gain of the th path in the th OFDM symbol period, is the Kronecker delta function, is the delay of the th path in unit of sample point, and is the number of resolvable paths. Assume that different paths are independent from each other and the power of the th path is . The channel is normalized so that The channel response in frequency domain is the FFT of and it is given by

where denotes points FFT operation. The channel autocorrelation matrix in frequency domain can be expressed as

where denotes expectation. Denote the vector form of the channel autocorrelation matrix by and we have . It is easy to find that the matrix is a circulant matrix. Therefore, as in [20], the eigenvalues of are given by

The formula (8) can be equivalently written as

We can easily obtain from (7) and (9) that the number of nonzero eigenvalues of is equal to the total number of resolvable paths, (see Appendix A). It is known by us that the rank of a square matrix is the number of its nonzero eigenvalues. Therefore the rank of is and is a singular matrix since . The matrix does not have the inverse matrix and has only the Moore-Penrose inverse matrix. However, the rank of the matrix is N (see Appendix A), where is an by identity matrix. Therefore, the matrix is not singular and has the inverse matrix.

Let

denote the channel frequency response at pilot subcarriers of the th OFDM symbol, and let

denote the vector of received signal at pilot subcarriers of the th OFDM symbol after FFT. Denote the pilot signal of the th OFDM symbol by . The channel estimate at pilot subcarriers based on least square (LS) criterion is given by

The LMMSE estimator at pilot subcarriers is given by [6]

where is channel autocorrelation matrix at pilot subcarriers and is defined by , where denotes Hermitian transpose. It is easy to verify that the matrix is circulant, the rank of is equal to and the rank of is equal to . The signal-to-noise ratio (SNR) is defined by and is a constant depending on the signal constellation. For 16QAM modulation and for QPSK and BPSK modulation If the channel autocorrelation matrix and SNR are known in advance, needs to be calculated only once. However, the autocorrelation matrix and SNR are often unknown in advance and time varying. Therefore the LMMSE channel estimator becomes unavailable in practice. To solve the problem, we propose the fast LMMSE channel estimation algorithm. The algorithm can be divided into three steps. The first step is to obtain the estimate of channel autocorrelation matrices and . Firstly, we obtain the least square (LS) channel estimation at pilot subcarriers in time domain, and it is given by

Secondly, the most significant taps (MSTs) algorithm [21] has been proposed to obtain the refined channel estimation in time domain. The MST algorithm deals with each OFDM symbol by reserving the most significant paths in terms of power and setting the other taps to be zero. The algorithm can reduce the influence of AWGN and other interference significantly, compared with the LS method. However, the algorithm may choose the wrong paths and omit the right paths because of the influence of AWGN and other interference. Thus, we will improve the algorithm of [21] by processing several adjacent OFDM symbols jointly. We calculate the average power of each tap for adjacent OFDM symbols, and it is given by

Then we choose the most significant taps from and reserve the indeces of them into a set . Finally, the refined channel estimation in time domain, , is given by

Denote the first row of the matrix by . Then can be given from (7) by

where is a 1 by vector with each entry

Since the matrix is circulant, can be acquired by circle shift of . The second step is to obtain the estimate of SNR. The estimate of SNR, , is given by

The third step is to obtain the estimate of the matrix , . We refer to the matrix as the LMMSE matrix in this paper. Since is a circulant matrix and is a circulant matrix, the product of and is also a circulant matrix. Therefore, we need only to compute the estimate of the first row of the LMMSE matrix. Denote the first row of LMMSE matrix by . The estimate of , , is given by (see Appendix B)

where denotes points IFFT operation. Therefore the estimated LMMSE matrix can be obtained from circle shift of . The channel estimation in frequency domain at pilot subcarriers for the th OFDM symbol can be given by

The proposed fast LMMSE algorithm avoids the matrix inverse operation and can be very efficient since the algorithm only uses the FFT and circle shift operation. The proposed fast LMMSE algorithm can be summarized as follows.

Step 1.

Obtain the LS channel estimation of pilot signal in time domain, , by formula (14).

Step 2.

Calculate the average power of each tap for OFDM symbols, , by formula (15). Then, we choose the most significant taps from and reserve it as , by formula (18).

Step 3.

Obtain the estimate of SNR, , by formula (19).

Step 4.

Obtain the estimate of the first row of the LMMSE matrix, , by formula (20).

Step 5.

Obtain the estimation of the LMMSE matrix, , by circle shift of . Then, the channel estimation in frequency domain at pilot subcarriers can be obtained by formula (21).

It is noted that the estimation of the LMMSE matrix requires only points FFT operation and circle shifting operation, which reduce the computational complexity significantly compared with the conventional LMMSE estimator since it requires the inverse operation of a large dimension matrix.

Denote the MSE of LMMSE algorithm by , where SNR is the true SNR, and SNR design is the designed SNR.

The MSE of LMMSE algorithm at pilot subcarriers for matched SNR can be derived as [22]

where is the first row of the matrix and denotes Hermitian transpose.

The MSE of LMMSE algorithm on pilot subcarriers for mismatched SNR can be derived as [22]

where is the first row of the matrix and denotes Hermitian transpose.

Let us denote the MSE of the proposed fast LMMSE algorithm by , where SNR is the true SNR, and is the estimated SNR or the designed SNR.

The MSE of the proposed fast LMMSE algorithm is given by

where , . If the number of the chosen OFDM symbol to obtain the estimated average power for each tap, , is large, we can replace with in (24), then, (24) can be further derived as

If the improved MST algorithm chooses ( ) paths, where is number of resolvable paths of the dispersive channel, and the chosen paths contain all the channel paths without omission, then (25) can be further written as

where is the channel delay of the th resolvable path, and is the power of the th path,

Similarly, the MSE of the proposed fast LMMSE algorithm for mismatched SNR is given by

where , . is the channel delay of the th resolvable path, and is the power of the th path,

It is noted that since the channel is assumed to be normalized, the MSE of the proposed fast LMMSE algorithm and the MSE of the conventional LMMSE are equal to their normalized mean square errors (NMSEs), respectively. In addition, for the sake of performance comparison between the above analysis of NMSE and the NMSE obtained by computer simulation, we define the NMSE obtained by simulation as follows:

where denotes the channel estimate at the th pilot subcarrier in the th OFDM symbol, obtained by LMMSE algorithm or the proposed fast LMMSE algorithm, and denotes the number of OFDM symbols in the simulation.

Figure 3 shows the magnitude of the first row of the channel autocorrelation matrix , Since the channel autocorrelation matrix is circulant, it is enough to show the first row of the channel autocorrelation matrix. Observe that the magnitude of varies approximately periodically, and the period is 13 pilot subcarriers. Since the channel power intensity profile is negative exponential distributed, the period of the first row of the channel autocorrelation matrix is decided by the delay of the second path. The delay of the second path is , that is, 10 sample points. According to (7), the period is . It is noted that the parameter should be replaced by in (7). Therefore, the period is about 13, as shown in Figure 3. Figure 4 shows the magnitude of the first row of the LMMSE matrix with SNR of 5 dB, 10 dB, and 20 dB, respectively. Since the LMMSE matrix is also circulant, it is sufficient to depict the first row of the LMMSE matrix. Observe that the value of the first row of the LMMSE matrix is symmetry, and the center point is 64. The first row of the LMMSE matrix is approximately periodic, and the period is about 13 pilot subcarriers. Observe that the value of the first row of the LMMSE matrix varies insignificantly when SNR changes from 5 dB to 20 dB. In addition, the local maximum values of the curves correspond to strong correlation between pilot subcarriers, and the local minimum values correspond to weak correlation between pilot subcarriers.

Figure 5 shows the NMSE of channel estimation of LMMSE algorithm versus that of the proposed fast LMMSE algorithm by computer simulation and numerical method, respectively. The numerical results of LMMSE algorithm and the proposed fast LMMSE algorithm are obtained by (22) and (26), respectively. The simulation results are obtained by (30). We replace in (30) with for LMMSE algorithm and replace with for the proposed LMMSE algorithm, respectively. For the proposed fast LMMSE algorithm, the number of OFDM symbols chosen to obtain the average power of each tap, , is 20, and the number of chosen paths, , is 10. The number of OFDM symbols in the simulation, , is 5000, for both LMMSE algorithm and the proposed fast LMMSE algorithm. Observe that the NMSE of the proposed fast LMMSE algorithm is very close to that of LMMSE algorithm in theory over the SNR range from 0 dB to 25 dB. In addition, for LMMSE algorithm the numerical result is verified by the simulation. For the proposed fast LMMSE algorithm, the simulation result approaches the numerical result well, except that the simulation result is a little higher than the numerical result at low SNR. Observe that both the proposed fast LMMSE algorithm and LMMSE algorithm are superior to LS algorithm. For instance, the LMMSE algorithm has about 16 dB gain over the LS algorithm, at the same MSE over the SNR range from 0 dB to 25 dB.

Figure 6 shows the normalized mean square error (NMSE) of LMMSE algorithm with matched SNR and mismatched SNRs versus SNR, by simulation and numerical method, respectively. Firstly, we give a necessary illustration of the curves obtained by numerical method. For the curves with matched SNR, we use (22) to calculate the MSEs under different SNRs, by numerical method. For the curves with mismatched SNRs, that is, designed SNRs, we use (23) to obtain the results, by numerical method. Secondly, for the curves with mismatched SNRs obtained by computer simulation, we use the designed SNR (predetermined and invariable) instead of the true SNR in (13) to obtain the channel estimation of pilot subcarriers. Observe that the analysis results are verified by computer simulation well, for the designed SNR of 5 dB, 10 dB, and 20 dB, respectively. For the case of the designed SNR of 5 dB, the MSE approaches the curve of matched SNR well within the range from 0 dB to about 10 dB. However, when the SNR increases, an MSE floor of about occurs. Similar trend can be found for the case of designed SNR of 10 dB. Observe that the curve of designed 20 dB approaches the curve with matched SNR well within the SNR range from 0 dB to 25 dB. Therefore, if we only know the channel autocorrelation matrix and do not know the SNR, the above results suggest that we use a higher designed SNR in (13) when performing channel estimation.

Figure 7 shows the NMSE of the proposed fast LMMSE algorithm with matched SNR and mismatched SNRs versus SNR, by simulation and numerical method respectively. Firstly, we give a brief illustration of the curves obtained by numerical method. For the curve with matched SNR, we use (26) to obtain the results. For the curves with mismatched SNRs, that is, designed SNR, we use (28) to obtain the numerical results. To verify the numerical results, we perform computer simulation for each case with different designed SNR. In the computer simulation, step 3 in the proposed fast LMMSE algorithm is modified by letting the estimated SNR, , be the designed SNR. For instance, if we choose the designed SNR to be 10 dB, will be set to be 10 dB in step 3 of the proposed fast LMMSE algorithm instead of using formula (19) to obtain . For the computer simulation, the number of OFDM symbols chosen to obtain the average power of each tap, , is 20, and the number of chosen paths, , is 10. The number of OFDM symbols in the simulation, , is 5000. Observe that the analysis results are verified by computer simulation well, for the designed SNR of 5 dB, 10 dB, and 20 dB, respectively. For the case of the designed SNR of 5 dB, the MSE approaches the curve of matched SNR well within the range from 0 dB to about 10 dB. However, when the SNR increases, an MSE floor of about occurs. Similar trend can be found for the case of designed SNR of 10 dB. Observe that the curve of designed 20 dB approaches the curve of matched SNR well within the SNR range from 0 dB to 25 dB.

Figure 8 shows the BER of LS, LMMSE, the proposed fast LMMSE, and perfect channel estimation, respectively. We adopt linear interpolation to obtain the channel frequency response at all subcarriers after the channel frequency response at pilot subcarriers is obtained by LS, LMMSE, and the proposed fast LMMSE estimator. Once the channel frequency response is obtained, we use maximum likelihood detection to obtain the estimated signal . In addition, the perfect channel estimation refers to that the channel frequency response is known by the receiver in advance. Observe that the BERs of LMMSE estimator is very close to that of the proposed fast LMMSE estimator over the SNR range from 0 dB to 25 dB. And they are about 1 dB worse than the perfect channel estimator, over the SNR ranging from 0 dB to 25 dB. The LMMSE estimator and the proposed LMMSE estimator are about 3-4 dB better than the LS estimator at the same BER over the SNR ranging from 0 dB to 25 dB.

Figure 9 shows the BER performance of the LMMSE channel estimation with matched SNR and the LMMSE channel estimation with designed SNRs. The LMMSE channel estimator with designed SNR refers to that we use a predetermined and unchanged SNR in (13) instead of the true SNR. Observe that the BERs of the LMMSE with designed SNR of 5 dB, 10 dB, and 20 dB are almost overlapped with each other within the lower SNR range from 0 dB to 15 dB. However, when SNR increases from 15 dB to 25 dB, the BER of the LMMSE estimator with higher designed SNR is better than that of the lower designed SNR. The results are consistent with the NMSEs in Figure 4. Therefore, a design for higher SNR is preferable as for mismatch in SNR.

Figure 10 shows the BER of the proposed fast LMMSE estimator with estimated SNR and the proposed fast LMMSE estimator with designed SNRs. It is noted that the proposed fast LMMSE estimator with estimated SNR refers to our proposed algorithm summarized in Section 3. The proposed fast LMMSE estimator with designed SNR refers to that we modify the step 3 of the proposed algorithm by using a predetermined and unchanged SNR instead of using formula (19) to obtain the estimated SNR. Observe that the BERs of the proposed fast LMMSE estimator with designed SNR of 5 dB, 10 dB, and 20 dB are almost overlapped with each other within the lower SNR range from 0 dB to 15 dB. However, when SNR increases from 15 dB to 25 dB, the BER of the proposed fast LMMSE estimator with higher designed SNR is better than that of the lower designed SNR. Thus, a design for higher SNR is preferable as for mismatch in SNR.