LOW COMPLEXITY EQUALIZATION FOR AFDM IN DOUBLY DISPERSIVE CHANNELS

Page created by Robert Byrd
 
CONTINUE READING
LOW COMPLEXITY EQUALIZATION FOR AFDM IN DOUBLY DISPERSIVE CHANNELS

                                                                               Ali Bemani1 , Nassar Ksairi2 , Marios Kountouris1

                                                        Communication Systems Department, EURECOM, Sophia Antipolis, France
                                                           1

                                                      Mathematical and Algorithmic Sciences Lab, Huawei France R&D, Paris, France
                                                       2

                                                 Email: ali.bemani@eurecom.fr, nassar.ksairi@huawei.com, marios.kountouris@eurecom.fr

                                                                                                                   to achieve full diversity in doubly dispersive channels [1] as
arXiv:2203.01875v2 [cs.IT] 7 Mar 2022

                                                                    ABSTRACT
                                        Affine Frequency Division Multiplexing (AFDM), which                       opposed to existing chirp-based waveforms, for instance [3,
                                        is based on discrete affine Fourier transform (DAFT), has                  4]. Furthermore, AFDM has similar performance in terms
                                        recently been proposed for reliable communication in high-                 of bit error rate (BER) with orthogonal time frequency space
                                        mobility scenarios. Two low complexity detectors for AFDM                  (OTFS) [5]. However, AFDM outperforms OTFS in terms of
                                        are introduced here. Approximating the channel matrix as                   pilot overhead and multiuser multiplexing overhead [2, 6].
                                        a band matrix via placing null symbols in the AFDM frame                       In this paper, we propose two low complexity detection
                                        in the DAFT domain, a low complexity MMSE detection is                     algorithms for AFDM taking advantage of its inherent chan-
                                        proposed by means of the LDL factorization. Furthermore,                   nel sparsity. By placing some null symbols - zero padding the
                                        exploiting the sparsity of the channel matrix, we propose a                AFDM frame - in the DAFT domain, the channel matrix can
                                        low complexity iterative decision feedback equalizer (DFE)                 be approximated as a band matrix. Using the approximated
                                        based on weighted maximal ratio combining (MRC), which                     band matrix and inspired by the equalizer in [7] for OFDM
                                        extracts and combines the received multipath components                    systems, we first design a low complexity MMSE detector
                                        of the transmitted symbols in the DAFT domain. Simula-                     based on LDL factorization [8]. The overall complexity of the
                                        tion results show that the proposed detectors have similar                 proposed algorithm is linear in the number of subcarriers and
                                        performance, while weighted MRC-based DFE has lower                        quadratic in the bandwidth of the band matrix, which in turn
                                        complexity than band-matrix-approximation LMMSE when                       depends on the the maximum delay and maximum Doppler
                                        the channel impulse response has gaps.                                     shift. Second, we propose a low complexity iterative decision
                                                                                                                   feedback equalizer (DFE) based on weighted maximal ratio
                                            Index Terms— AFDM, affine Fourier transform, doubly                    combining (MRC) of the channel impaired input symbols re-
                                        dispersive channels, detector, MMSE, DFE, MRC.                             ceived from different paths. The overall complexity of the
                                                                                                                   second algorithm is also linear in the number of subcarriers
                                                               1. INTRODUCTION                                     and quadratic in the number of paths. We show that these two
                                         Next-generation wireless systems (e.g., B5G/6G) are evolv-                detectors have similar performance between them, whereas
                                        ing to cater to a wide range of applications and services re-              when the channel is sparse in the delay domain, weighted
                                        quiring reliable communication in high-mobility scenarios.                 MRC-based DFE detector exhibits lower complexity.
                                        This calls for new waveform design able to cope with time-                                     2. SYSTEM MODEL
                                        varying channels. In this setting, existing waveforms, in par-
                                                                                                                   The AFDM block diagram is given in Fig. 1. Modulation
                                        ticular orthogonal frequency division multiplexing (OFDM),
                                                                                                                   is produced by using DAFT at the transmitter and receiver.
                                        lose subcarrier orthogonality, thus resulting in inter-carrier in-
                                                                                                                   DAFT is a discretized version of AFT [1, 9–11] and its kernel
                                        terference and deteriorated system performance.                                                   2  1        2
                                                                                                                   is equal to e−ı2π(c2 m + N mn+c1 n ) where c1 and c2 are the
                                             Affine frequency division multiplexing (AFDM) has re-
                                                                                                                   AFDM parameters tuned to provide full delay-Doppler rep-
                                        cently been proposed as a promising waveform for commu-
                                                                                                                   resentation of the channel in the DAFT domain. It has been
                                        nication in time-varying channels [1, 2] showing significant
                                                                                                                   shown that tuning c1 using the the maximum Doppler shift
                                        performance gains over OFDM. AFDM employs multiple or-
                                                                                                                   normalized with respect to the subcarrier spacing, and setting
                                        thogonal information-bearing chirps generated using the dis-
                                                                                                                   c2 to be an arbitrary irrational number or a rational number
                                        crete affine Fourier transform (DAFT). A key feature is that
                                                                                                                   sufficiently smaller than 1/2N , enables AFDM to achieve full
                                        its chirp pulse parameters can be adapted to the channel char-
                                                                                                                   diversity in doubly dispersive channels [1].
                                        acteristics, making a complete delay-Doppler representation
                                        of the channel in the DAFT domain. This enables AFDM                       2.1. Modulation and Demodulation
                                            This work has been supported by a Huawei France-funded Chair towards   Consider a set of quadrature amplitude modulation (QAM)
                                        Future Wireless Networks.                                                  symbols xk , k = 0, 1, 2, ..., N − 1. AFDM maps xk to sn
Modulation                                                       Demodulation                                                        Heff
                                                                                                                                                                                   0           x
                                                                                                                                                                                   0           x
                                                                                                                                                                                   0           x
                                                                                                                Q+1
                                                                                                                                                                                   0           x
                                                                                                                                                                                   d           x
                                  Add                            Remove
   S\P   ΛH       IFFT      ΛH              P\S          S\P              Λc1       FFT        Λc2   P\S                                                                           d           x
          c2                 c1
                                  CPP                            CPP
                                                                                                                                                                                   d           x
                                                                                                                                                                                   d           x
                                                                                                                                                                                   d           x
                                                                                                                                                                                   d           x
                                                                                                                                                                                   d           x
                                                                                                                                                                                   d           x
                                                                                                                                                                                           =
                                                                                                                                                                                   d           x
  Fig. 1: AFDM modulation/demodulation block diagram.                                                                                                                              d
                                                                                                                                                                                   d   x
                                                                                                                                                                                               x
                                                                                                                                                                                               x
                                                                                                                                                                                   d           x

using inverse DAFT (IDAFT) as follows:                                                                                                                                             d
                                                                                                                                                                                   d
                                                                                                                                                                                               x
                                                                                                                                                                                               x
                                                                                                                                                                                   d           x
        N −1                                                                                                                                                                       d           x
      1 X                2  1        2                                                                                                                                             d           x
sn = √       xm eı2π(c2 m + N mn+c1 n ) ,                                 n = 0, · · · , N − 1.                                                                                    d           x
      N m=0                                                                                                                                                                        d           x
                                                                                                                                                                                   d           x
                                                                                                       (1)                                                                         d           x

 To make the channel lie in a periodic domain, a chirp-                                                                                                                            0           x

periodic prefix (CPP) should be added to the modulated                                                                   Fig. 2: Truncated parts of x and Heff
signal, defined as
   sn = sN +n e−ı2πc1 (N
                                        2
                                            +2N n)
                                                     ,         n = −M, · · · , −1                     (2)    sparse structure i.e., there are exactly L = P non-zero en-
                                                                                                             tries in each row and column if νi is integer and if we set
                                                                                                                                                         PP
where M is any integer greater than or equal to the value in                                                 c1 = 2νmax
                                                                                                                      2N
                                                                                                                         +1
                                                                                                                            . Indeed, in this case Heff = i=1 hi Hi where
samples of the maximum delay spread of the wireless chan-
                                                                                                                                             2                 2
                                                                                                                                   2π                              −p2 ))
                                                                                                                             (
nel. After transmission over the channel, the received samples                                                                   eı N (N c1 li −qli +N c2 (q                q = (p + loci )N
                                                                                                               Hi (p, q) =
are                     X∞                                                                                                       0                                          otherwise
                  rn =      sn−l gn (l) + wn               (3)                                                                                                                                     (7)
                                    l=0                                                                      where loci = (νi + (2νmax + 1)li )N [1]. For fractional
where wn ∼ CN (0, N0 ) is an additive Gaussian noise and                                                     νi , it can be shown that Hi has in each row and column a
gn (l) is the impulse response of the time-varying channel at                                                peak surrounded by approximately 2kν non-zero entries de-
time n and delay l, given by                                                                                 creasing rapidly as we move away from it. Hence, if we set
                         P
                         X                                                                                   c1 = 2(bνmax2N c+kν )+1
                                                                                                                                     , Heff can be approximated as having
                gn (l) =    hi e−ı2πfi n δ(l − li )       (4)                                                L = (2kν + 1)P non-zero entries per row and column (i.e.,
                                    i=1
                                                                                                             each received symbol can be approximately expressed as a
where P ≥ 1 is the number of paths, δ(·) is the Dirac delta                                                  linear combination of only a few input symbols). We pro-
function, and hi , fi and li are the complex gain, Doppler shift                                             pose below two low complexity detection algorithms leverag-
(in digital frequencies), and the integer delay associated with                                              ing the channel matrix sparsity.
the i-th path, respectively. We define νi , N fi , where νi ∈
[−νmax , νmax ] is the Doppler shift normalized with respect to                                                          3. DETECTION ALGORITHMS
the subcarrier spacing. We assume that the maximum delay
of the channel satisfies lmax , max(li ) < N .                                                               The first step is to place some null symbols that allow to ap-
The DAFT domain output symbols are obtained by                                                               proximate the truncated part of Heff as a band matrix. This
                       N −1
                                                                                                             also simplifies the input-output relation as the modular oper-
                     1 X                 2  1        2                                                       ation is no longer needed, as shown in Fig. 2. Note that these
           ym      =        rn e−ı2π(c2 m + N mn+c1 n ) .                                             (5)
                     N n=0                                                                                   symbols do not entail extra overhead as they can serve not
                                                                                                             only the proposed detection algorithms but also embedded pi-
Discarding the CPP, the input-output relation can be written                                                 lot aided channel estimation. Due to the structure of Heff and
in matrix form as                                                                                            Hi , the number of the null guard symbols should be greater
                                                                                                             than Q = (lmax + 1)(2(αmax + kν ) + 1) − 1. Taking into ac-
                            y =Ar = Heff x + Aw                                                       (6)    count the zero padding, the vector of DAFT domain received
where A = Λc2 FΛc1 is the DAFT matrix, F is the discrete                                                     samples writes as
                                                       √                                                                             y = Heff x + w e                    (8)
Fourier transform (DFT) matrix with entries e−ı2πmn/N / N ,
                    2
Λc = diag(e−ı2πcn , n = 0, 1, . . . , N −1), Heff = AHAH ,                                                   where x and Heff are the truncated parts of x and Heff , re-
and H is the matrix representation of the channel. The ele-                                                  spectively (see Fig. 2). They can be expressed using the ma-
ments of y, r, x and w ∼ CN (0, N0 I) are similarly related                                                  trix T = [IN ]Q−(αmax +kν ):N −(αmax +kν )−1,: as x = Tx and
to yk , rk , xk and wk , respectively. Since A is a unitary                                                  Heff = Heff TH . Using LMMSE equalization based on (8)
matrix, we = Aw and w have the same statistics. Heff has                                                     for detection requires O(N 3 ) flops, which can be prohibitive
for large N . We thus propose two detectors with lower com-            Transmitted symbols:     x0                       x1                         x2                      x3                       x4                  x5
                                                                                                             h02                        h13                     h24                     h35                 h46                 h57
plexity. The first is a low-complexity LMMSE based on a
                                                                                                       h01                        h12                     h23                     h34                      h45                  h56
band approximation of Heff . The second is a weighted MRC-                                      h00                      h11                        h22                     h33                      h44                 h55

based DFE exploiting the sparse representation of the com-                 Received symbols:    y0                       y1                         y2                      y3                       y4                  y5                 y6   y7

munication channel provided by AFDM.                                                            b20                      b21                        b22                     b23                      b24                 b25

                                                                                                       b10                        b11                     b12                     b13                      b14                  b15

3.1. Low complexity MMSE detection                                                                                 b00                        b01                     b02                     b03                 b04                 b05

                                                                                Detection:     W-MRC                 W-MRC                      W-MRC                   W-MRC                   W-MRC                   W-MRC
To recover the data symbols x, considering (8), the following
MMSE equalization is used
              x̂ = HH          H
                    eff (Heff Heff + N0 IN )
                                            −1
                                               y.            (9)          Estimated symbols:    x̂0                      x̂1                        x̂2                     x̂3                      x̂4                 x̂5

Although (9) involves matrix inversion, the matrix M =              Fig. 3: Weighted MRC operation for N = 8 with a 3-path
Heff HH eff + N0 IN is a Hermitian band matrix with lower           channel with Q = 2 where W-MRC stands for weighted MRC
and upper bandwidth Q. Thus, M−1 can be computed us-                and hji = Heff (i, j).
ing LDL factorization. Algorithm 1 can be performed to
efficiently equalize the received signal. The computational         for the combining. Considering the structure of Heff , it can
  Algorithm 1: Low complexity MMSE detection                        be seen that each received symbol yk is given by
 1   Construct the matrix Heff = Heff TH                                                                                       L−1
     Construct the band matrix M = Heff HH
                                                                                                                               X
 2                                          eff + N0 IN                                                yk =                                   Heff (k, pik )xpik                                                                (10)
 3   Compute the LDL factorization of M = LDLH                                                                                 i=0
      where L is a lower triangular matrix with Q sub
                                                                    where pik is the column index of the i-th path coefficient in
      diagonals and D is a diagonal matrix
                                                                    row k of matrix Heff . Let bik be the channel impaired in-
 4   Solve the triangular system Lf = y
                                                                    put symbol xk in the received samples yqki after canceling
 5   Solve the diagonal system Dg = f
 6   Solve the triangular system LH d = g                           the interference from other input symbols, where qki is the
                                                                    row index of the i-th path coefficient in column k of ma-
 7   Calculate x̂ = HHeff d
                                                                    trix Heff . In each iteration, assuming estimates of the in-
cost of the proposed detection algorithm is evaluated in terms      put symbols xk are available either from the current iteration
of complex additions (CAs), complex multiplications (CMs)           (for pjqi < k, j = 0, ..., L − 1) or previous iteration (for
                                                                                k
and complex divisions (CDs). The first step does not need           pjqi > k,            j = 0, ..., L − 1), bik can be written as
any complex operation since Heff is truncated from Heff .             k

In step 2, every element of Heff HHeff requires at most Q + 1
                                                                                                                           j      (n)
                                                                                                       X
                                                                                         bik = yqki −       HH        i
                                                                                                                eff (qk , pq i )x̂pj
CMs and Q CAs. Considering that Heff HH    eff is Hermitian and                                                                pj i k
                                                                                                                                                                                  k
                                                                                                                                                                                                    qi
                                                                                                                                                                                                     k

2 (Q + 3Q)N CMs, 2 (Q + Q)N CAs, and QN CDs. Steps
1    2                 1   2                                                                                                    q
                                                                                                                                 k

4 and 6 can be solved by band forward and backward sub-             where superscript (n) denotes the n-th iteration. Considering
stitutions [8] and each of them has QN CMs and QN CAs.              bik for all paths i = 0, 1, ..., L − 1, weighted MRC (as op-
Step 5 can be solved using N CDs since D is a diagonal              posed to pure MRC [12]) can be performed. The output of
matrix and the last step requires (Q + 1)N CMs and QN               the weighted MRC for estimating xk is given by
CAs. Thus, the algorithm requires (Q2 + 6Q + 2)N CMs,
                                                                                                                                                   gk
(Q2 + 4Q + 1)N CAs and (Q + 1)N CDs, which amounts                                                                         ck =                                                                                                 (12)
to (2Q2 + 11Q + 4)N complex operations in total.                                                                                                dk + γ −1
3.2. Weighted MRC-based DFE detection                               where                                                      L−1
                                                                                                                               X
                                                                                                        gk ,                                    HH     i      i
                                                                                                                                                 eff (qk , k)bk ,                                                               (13)
As mentioned in Section 2, Heff has L non-zero entries per
                                                                                                                                i=0
column. This feature enables us to propose a weighted MRC-
based detector where each data symbol is detected from the                                                                     L−1
                                                                                                                               X
weighted MRC of its L channel-impaired received copies.                                                      d,                                |HH     i
                                                                                                                                                 eff (qk , k)|
                                                                                                                                                              2
                                                                                                                                                                                                                                (14)
Fig. 3 shows an example of this detector for AFDM with                                                                         i=0

N = 8 and a 3-path channel with Q = 2. The proposed                 and γ is the signal-to-noise ratio (SNR). Let D(.) denote the
detector is iterative, where in each iteration, the estimated in-   decision on the symbol estimate ck , i.e, x̂nk = D(ck ). In this
ter symbol interference is canceled in the branches selected        paper, we consider x̂nk = ck . The estimated symbols are then
35
used for the next iteration. The algorithm continues until the                10 -2

maximum number of iterations is reached or the updated input
                                                                                                                                                                    30

                                                                                                                                      Average number of iteration
symbol vector is close enough to the previous one as summa-
                                                                                                                                                                    25

rized in Algorithm 2.                                                                                                                                               20

                                                                        BER
                                                                                                                                                                    15

 Algorithm 2: Weighted MRC-based DFE detection                                10 -3
                                                                                                                                                                    10

      Data: Heff , d, y, x̂0 = 0                                                                                                                                    5

 1    for n = 1 : niter do                                                                                                                                          0
                                                                                 10 -3                 10 -2        10 -1      10 0                                 10 -3        10 -2     10 -1         10 0
 2        for k = 0 : N-Q-1 do
 3            for i = 0 : L-1 do
                               X                                                      (a) BER variation vs. .                                                              (b) Iterations vs. .
                                                            (n)
                 bik = yqi −              HH     i    j
                                           eff (qk , pq i )x̂   j
                         k                                  p i
                                                                        Fig. 4: BER variation and the average number of iterations
                                                        k
                                  j                          q
                                 p i k                             k
                                  q
                                    k                                                       100
 5           end
                  PL−1
 6           gk = i=0 HH         i      i
                           eff (qk , k)bk
                     gk                                                                     10-1
 7           ck = dk +γ −1
               (n)               (n)
 8            x̂k = ck or x̂k = D(ck )

                                                                                      BER
 9       end                                                                                10-2
10       if ||x̂(n) − x̂(n−1) || <  then EXIT;
11    end
                                                                                                               AFDM, MMSE
                                                                                            10-3
    Computing the complexity of Algorithm 2 is straightfor-                                                    AFDM, MRC-based DFE
                                                                                                               AFDM, low-complexity MMSE
ward as it has only scalar operation. From step 3 to step 11, it                                               OFDM, MMSE

requires L2 CMs, L2 CAs and 1 CD. Therefore, its total com-                                 10-4
                                                                                                               OFDM, low-complexity MMSE [7]

plexity is niter (2L2 +1)(N −Q). In simulations, we observed                                       0                   5                                            10               15             20
that the algorithm typically converges within 15 iterations. In                                                                 SNR in dB

the longer version of the article, convergence of x̂(n) to the          Fig. 5: BER performance comparison between AFDM and
LMMSE estimate x̂ defined in (9) is proved.                             OFDM systems using different detectors
    The complexity of the two proposed algorithms is re-
markably smaller than maximum likelihood (ML) and linear
                                                                        constant, and the algorithm converges within 14 iterations, as
MMSE detectors, which have exponential O(|A|N ) and cu-
                                                                        shown in Fig. 4b. In Fig. 5, we plot the BER performance of
bic O(N 3 ) complexity, respectively, with A representing the
                                                                        AFDM and OFDM for LMMSE, low-complexity MMSE [7]
QAM alphabet. Moreover, Algorithm 2 has lower complexity
                                                                        and weighted MRC-based DFE detectors. First, we observe
than Algorithm 1 when the channel impulse response has
                                                                        that AFDM outperforms OFDM, thanks to achieving full di-
gaps. This is due to the fact that its complexity only depends
                                                                        versity and as every information symbol is received through
on the number of non-zero elements in each column of Heff ,
                                                                        multiple independent non-overlapping paths. Second, we ob-
i.e L, instead of Q ≥ L.
                                                                        serve that both proposed detection algorithms have similar
                 4. SIMULATION RESULTS                                  performance between them, while conventional MMSE de-
                                                                        tection has slightly better performance at the cost of higher
In this section, we simulate the uncoded BER performance of
                                                                        complexity.
AFDM over doubly dispersive channels. The following pa-
rameters are used: carrier frequency fc = 4 GHz, number
of subcarriers N = 128, and AFDM frame length 330 µs.                                                                5. CONCLUSION
Path delays are fixed, and considering Jakes Doppler spec-
trum for each channel realization, the Doppler shift of the i-th        We proposed two low complexity detection algorithms for
path is generated using νi = νmax cos(θi ), where θi is uni-            zero-padded AFDM. First, a low complexity MMSE detector
formly distributed over [−π, π] with 4-QAM signaling. The               which makes use of band LDL factorization was derived.
maximum Doppler shift is νmax = 1, which corresponds to a               Second, an iterative weighted MRC-based DFE detector,
maximum speed of 810 km/h. Fig. 4a shows the BER perfor-                which exploits the channel sparsity, was proposed. Our
mance of AFDM using the proposed weighted MRC-based                     results showed that both detectors have comparable perfor-
DFE detector for different values of  at SNR = 20 dB. We               mance as exact LMMSE while their complexity order is
can see that below  = 0.01, the performance remains almost             linear, instead of cubic, in the number of subcarriers.
6. REFERENCES                                [12] T. Thaj and E. Viterbo, “Low complexity iterative rake
                                                                      decision feedback equalizer for zero-padded otfs sys-
 [1] A. Bemani, N. Ksairi, and M. Kountouris, “AFDM: A                tems,” IEEE Trans. on Vehicular Technology, vol. 69,
     full diversity next generation waveform for high mo-             no. 12, pp. 15 606–15 622, 2020.
     bility communications,” in 2021 IEEE International
     Conference on Communications Workshops (ICC Work-
     shops), 2021, pp. 1–6.

 [2] A. Bemani, G. Cuozzo, N. Ksairi, and M. Koun-
     touris, “Affine frequency division multiplexing for
     next-generation wireless networks,” in 2021 Interna-
     tional Symposium on Wireless Communication Systems
     (ISWCS), 2021, pp. 1–6.

 [3] R. Bomfin, M. Chafii, A. Nimr, and G. Fettweis, “A ro-
     bust baseband transceiver design for doubly-dispersive
     channels,” IEEE Trans. on Wireless Communications,
     vol. 20, no. 8, pp. 4781–4796, 2021.

 [4] R. Bomfin, M. Chafii, and G. Fettweis, “Low-
     complexity iterative receiver for orthogonal chirp divi-
     sion multiplexing,” in 2019 IEEE Wireless Communica-
     tions and Networking Conference Workshop (WCNCW).
     IEEE, 2019, pp. 1–6.

 [5] R. Hadani, S. Rakib, M. Tsatsanis, A. Monk, A. J. Gold-
     smith, A. F. Molisch, and R. Calderbank, “Orthogo-
     nal time frequency space modulation,” in IEEE Wireless
     Communications and Networking Conference (WCNC),
     2017, pp. 1–6.

 [6] P. Raviteja, K. T. Phan, and Y. Hong, “Embedded
     pilot-aided channel estimation for OTFS in delay–
     doppler channels,” IEEE Trans. on Vehicular Technol-
     ogy, vol. 68, no. 5, pp. 4906–4917, 2019.

 [7] L. Rugini, P. Banelli, and G. Leus, “Simple equalization
     of time-varying channels for OFDM,” IEEE communi-
     cations letters, vol. 9, no. 7, pp. 619–621, 2005.

 [8] G. H. Golub and C. F. Van Loan, “Matrix computations,”
     Johns Hopkins University Press, 3rd edition, 1996.

 [9] J. J. Healy, M. A. Kutay, H. M. Ozaktas, and J. T. Sheri-
     dan, Linear canonical transforms: Theory and applica-
     tions. Springer, 2015.

[10] S.-C. Pei and J.-J. Ding, “Relations between frac-
     tional operations and time-frequency distributions, and
     their applications,” IEEE Trans. on Signal Processing,
     vol. 49, no. 8, pp. 1638–1655, 2001.

[11] S. Pei and J. Ding, “Closed-form discrete fractional and
     affine Fourier transforms,” IEEE Trans. on Signal Pro-
     cessing, vol. 48, no. 5, pp. 1338–1353, 2000.
You can also read