Time-Variant Channel Estimation Using Discrete Prolate Spheroidal Sequences
Thomas Zemen, Member, IEEE, and Christoph F. Mecklenbräuker, Member, IEEE
Abstract—We propose and analyze a low-complexity channel estimator for a multiuser multicarrier code division multiple access (MC-CDMA) downlink in a time-variant frequency-selective channel. MC-CDMA is based on orthogonal frequency division multiplexing (OFDM). The time-variant channel is estimated individually for every flat-fading subcarrier, assuming small intercarrier interference. The temporal variation of every subcarrier over the duration of a data block is upper bounded by the Doppler bandwidth determined by the maximum velocity of the users. Slepian showed that time-limited snapshots of bandlimited sequences span a low-dimensional subspace. This subspace is also spanned by discrete prolate spheroidal (DPS) sequences. We expand the time-variant subcarrier coefficients in terms of orthogonal DPS sequences we call Slepian basis expansion. This enables a time-variant channel description that avoids the frequency leakage effect of the Fourier basis expansion. The square bias of the Slepian basis expansion per subcarrier is three magnitudes smaller than the square bias of the Fourier basis expansion. We show simulation results for a fully loaded MC-CDMA downlink with classic linear minimum mean square error multiuser detection. The users are moving with 19.4 m/s. Using the Slepian basis expansion channel estimator and a pilot ratio of only 2%, we achieve a bit error rate performance as with perfect channel knowledge. Index Terms—Discrete prolate spheroidal sequence, MC-CDMA, OFDM, Slepian basis expansion, time-variant channel estimation.

OFDM transforms the time-variant frequency-selective channel into a set of time-variant frequency-flat subcarriers. We deal with time-variant channels that vary significantly over the duration of a long block of OFDM symbols. Each OFDM symbol is preceded by a cyclic prefix to avoid intersymbol interference. The variation of a wireless channel over the duration of a data block is caused by several impinging wavefronts, each with potentially different Doppler shifts. The Doppler frequencies de, and the scatpend on the velocity, the carrier frequency tering environment. The maximum variation in time of the wireless channel is upper bounded by the maximum normalized one-sided Doppler bandwidth (1) where is the maximum (supported) velocity, is the OFDM symbol duration, and is the speed of light. Under the assumption of small intercarrier interference (ICI), each time-variant frequency-flat subcarrier is fully described through a sequence of complex scalars at the OFDM symbol rate . This sequence is bandlimited by . At the receiver side, we perform block-oriented processing for data detection and channel estimation. Hence, we aim at finding a channel representation that describes a time-variant subcarrier for the duration of a single data block. In [1] and [2], the autocorrelation of the channel, i.e., its second-order statistic, is assumed to be known, and a Kalman filter or a Karhunen–Loéve transform are applied. The autocorrelation is calculated under the assumption of a dense scatterer model in the limit of an infinite number of scatterers [3], [4, Sec. 2.5.2]. This assumption is not fulfilled in practical channels [5]. Furthermore, the actual velocity of the user and the angles of arrival enter the autocorrelation as parameters and must be estimated explicitly. Another approach along these lines is to estimate and track the channel statistics online [6]. We estimate the parameters of a deterministic channel model from a single realization of a randomly fading channel using known pilot symbols. The discrete multipath channel model discussed in [7] is a deterministic channel description that depends nonlinearly on the Doppler frequencies of each propagation path. We want to avoid such a nonlinear estimation problem [8], [9]. We use a basis expansion with an appropriately chosen set of basis functions to describe the variation of the channel in time. Fourier basis functions are applied in [10] and [11] to represent the time-variant channel. We have observed that the



HIS paper describes a solution for the problem of channel estimation in a wireless communication system based on multicarrier (MC) code division multiple access (CDMA) when users are moving at vehicular speed. MC-CDMA is based on orthogonal frequency division multiplexing (OFDM), and the data symbols are spread by a user specific spreading sequence that is applied in the frequency domain. The transmission scheme is block oriented. A block consists of OFDM data symbols with interleaved OFDM pilot symbols to allow pilot-based estimation of the time-variant channel.

Manuscript received December 14, 2003; revised November 23, 2004. This work was suported by Kplus in the I0 project of the ftw. Forschungszentrum Telekommunikation Wien and Siemens AG Austria, PSE PRO Radio Communication Devices (RCD). T. Zemen was with Siemens AG Austria when he performed parts of this work. Parts of this work have been presented at the 37th Asilomar Conference on Signals, Systems, and Computers, November 7–10, 2003, Pacific Grove, CA, and at the 12th European Signal Processing Conference (EUSIPCO), Sep. 6–10, 2004, Vienna, Austria. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Nicholas D. Sidiropoulos. The authors are with the ftw. Forschungszentrum Telekommunikation Wien (Telecommunications Research Center Vienna), 1220 Vienna, Austria (e-mail:; Digital Object Identifier 10.1109/TSP.2005.853104

1053-587X/$20.00 © 2005 IEEE

the properties of the DPS sequences are described. set The pilot placement is defined through the index set (4) Fig.4 m/s.3598 IEEE TRANSACTIONS ON SIGNAL PROCESSING. . The performance results of the Slepian basis expansion are compared with the one obtained with the Fourier basis expansion. In Section VI.e. lower or equal than Smallest integer. We compare the channel estimation square bias and variance of the Slepian basis expansion and the Fourier basis expansion for time-variant flat-fading channels. 53. 9. a . SIGNAL MODEL FOR FLAT-FADING TIME-VARIANT CHANNELS We consider the transmission of a symbol sequence with symbol rate over a time-variant flat-fading channel in order to simplify the initial presentation. 1 shows an example for the pilot set defined in (4). The symbol duration is much longer than the delay spread of the channel . a truncated discrete Fourier transform (DFT). Sec. This rethe infinite set markable property enables parameter estimation without the drawbacks of windowing in the case of the Fourier basis expansion [15. and the matched receive filter. the physical channel. flat-fading channel estimation error for the Slepian basis expansion and the Fourier basis expansion in Section V. Slepian showed in [15] that time-limited parts of bandlimited sequences span a low-dimensional subspace. The pilot symbols are independent identically distributed (i. This estimator does not need detailed knowledge of the channels autocorrelation (i. Complex conjugate of . NO. These DPS sequences have a double orthogonality property: They are orthogonal over the finite set and simultaneously. and we generalize our channel estimator to time-variant frequency-selective channels. The Slepian basis expansion represents bandlimited sequences with a minimum number of basis functions avoiding the deficiencies of the Fourier basis expansion. 3. III.e.4]. however. duration of the transmitted data block with length A. we achieve a bit error rate performance as with perfect channel knowledge. greater or equal than -norm of vector . Numerical simulations are carried out for time-variant channels with Jakes’ Doppler spectrum [17] and compared with the analytical results. We develop a low-complexity channel estimator for a MC-CDMA downlink in a time-variant frequency-selective channel. In Section IV. 5. We show simulation results for a fully loaded MC-CDMA downlink with classic linear minimum mean square error (MMSE) multiuser detection [18]. and the associated Slepian basis expansion for time-variant flat-fading channel estimation is defined.2]. Contributions • We apply the concept of discrete prolate spheroidal sequences for the estimation of time-variant mobile communication channels. Sec. • • • B. column vector with all ones. diag . The basis functions of the Slepian basis expansion are bandlimited to the known maximum variation and time-concentrated for the in time of the channel [16]. has the following major drawback: the rectangular window associated with the DFT introduces spectral leakage [12. The channel in incorporates the transmit equivalent baseband notation filter. VOL. SEPTEMBER 2005 Fourier basis expansion. Finally. 2. II. the MC-CDMA system is introduced.d. Largest integer. identity matrix. Matrix with elements .4]. Diagonal matrix with entries . NOTATION Column vector with elements . The optimal pilot placement is not known to the authors. Conjugate transpose of . i. we conclude in Section VII. The is given by the multiplication of the received sequence symbol sequence and the sampled time-variant channel plus additional circular symmetric complex white Gaussian noise with zero mean and variance (2) We transmit symbols in blocks of length . The energy from low-frequency Fourier coefficients leaks to the full frequency range.) chosen with equal probability from the QPSK symbol for and for . Sec.i. column vector with all zeros. as was shown in [13] and [14].1. Upper left part of with dimension Transpose of . Discrete time is denoted by . We give an analytical and numerical performance evaluation of the time-variant .. Absolute value of . Paper Organization To simplify the initial presentation. The users are moving with 19. 1 for . an effect similar to the Gibbs phenomenon [12. together with spectral leakage. The orthogonal basis is spanned by the so-called discrete prolate spheroidal (DPS) sequences. Using the Slepian basis expansion channel estimator and a pilot ratio of only 2%. we define a signal model for a time-variant flat-fading channel in Section III. When the DFT is truncated. This results in a floor in the bit error rate (BER) for channels with Doppler spread. 0 otherwise. the Doppler spectrum). Each block condata symbols with interleaved pilot symsists of bols (3) The data symbols satisfy for and for . leads to significant phase and amplitude errors at the beginning and at the end of the data block [4]. Analytical expressions are established for time-variant channels with arbitrary Doppler spectra of finite support..4.

we for need an efficient representation of the channel . 3. see [19]. IV. . The Slepian basis in terms of Slepian seexpansion expands the sequence quences (13) (8) for and [15]. Sec. hence (6) where (7) We see that . We drop the explicit dependence of and on and . Thus. 12g defined by (4) for M = 15 and J = 3. The symbols have constant modulus. For channel estimation and equalization at the receiver. The eigenvalues are a measure for this energy concentration expressed by (5). we use the superscript . is the unique sequence that is The DPS sequence band-limited and most time-concentrated in a given interval is the next sequence having maximum with length . Example pilot pattern P = f2. This means that the DPS seand quences are orthonormal on the set . We introduce the term Slepian sequences for the index limited DPS sequences and define the vector with elements for . the DPS basis expansion. The DPS sequences are defined as the real-valued solution of The eigenvalues defined as are identical to those in (8). The eigenvalues orthogonal on the set are clustered near 1 for and rapidly drop to zero for . which we consider fixed system parameters for the remainder of the paper. Therefore.3] of time-limited snapshots of a bandlimited signal is approximately given by (10) For a rigorous proof. For our application. we will use a generic notation for the basis expansion quantities . energy concentration among the DPS sequences orthogonal to . SLEPIAN BASIS EXPANSION Slepian [15] answered the question of which sequences are and simulbandlimited to the frequency range taneously most concentrated in a certain time interval of length . 1. and so on. and to indicate that the expression is applicable to any set of orthonormal basis functions . 2. . which we will pursue in Section IV. we are interested in for the time index set only. If we specifically want to reference the Slepian . All the properties described so far enable parameter estimation without the drawbacks of windowing as in the case of the Fourier basis expansion [15. In Fig. 7. fulfilling of the matrix (11) (5) while being bandlimited to . The theory of time-concentrated and bandlimited sequences developed by Slepian [15] enables a new approach for the time-variant channel estimation problem. More specifically (9) where . the Slepian sequences span an orthonormal basis that allows the representation of a time-limited snapshot of bandlimited sequences through a basis expansion. The Slepian sequences are eigenvectors . The dimension of this basis ex(14) Wherever possible. The solution of this constrained maximization problem are the discrete prolate spheroidal (DPS) sequences [15]. the DPS sequences show that a set where pansion fulfills . Concluding. 3. Sec. We will exploit the knowledge of two parameters only: the maximum Doppler bandwidth and the block length . uniform placement performs well as shown below. The sequences we are seeking will have their maximum energy concentration in an interval with length of orthogonal sequences exists that is exactly bandlimited and simultaneously posses a high (but not complete) time concentration in a certain interval with length . Matrix is (12) where . their energy is normalized .ZEMEN AND MECKLENBRÄUKER: TIME-VARIANT CHANNEL ESTIMATION USING DISCRETE PROLATE SPHEROIDAL SEQUENCES 3599 Fig.3]. the signal space dimension [15. The DPS sequences are doubly orthogonal on the infinite set and the finite set .

VOL. with . we compare its performance to the Fourier basis expansion by means of simulations with the system defined in (2). The true speed of the user is denoted by . . The approximate dimension of the signal d  Me space evaluates to D . the approximate . maximum normalized . With these system parameters. The Slepian and a maximum Doppler sequences are designed for block length M bandwidth  : 1 . = 3 9 10 = 2 = 256 +1 = 3 m for block length M Fig. We define the vector channel knowledge for . and data block length Doppler frequency .3600 IEEE TRANSACTIONS ON SIGNAL PROCESSING. Fig. we will compare the simulation results with analytic results derived in Sections V-B and C. BASIS EXPANSION ERROR ANALYSIS FOR TIME-VARIANT FLAT-FADING CHANNELS To demonstrate the merits of the Slepian basis expansion. Additionally. dimension of the signal space Therefore. V. the dimension of the estimation problem is reduced . 3. 53. By choosing . s The symbol duration s is chosen for later comparison with the MC-CDMA system introduced in Section VI. SEPTEMBER 2005 . The basis expansion parameters in (13) are estimated according to (18) where (19) We use a generic notation since (16)–(18) are valid for any basis expansion. 9. The speed of the user is varied km h m s. The autocorrelation of (20) where and is the zeroth-order Bessel function of the first kind. eigenvalue spectrum is given. normalized Doppler bandwidth  [ ] = 3 9 10 = 256 and maximum A. 2. The Slepian sequences for are depicted in Fig. symbol s . Channel Model and System Assumption The actual realization of the time-variant flat-fading channel is calculated according to the simulation model in the Apis pendix. NO. . we give a numerical example of an actual communication system. which in the range 180 Hz results in a range of Doppler bandwidths In order to highlight the utility of the Slepian basis expansion in terms of dimension reduction. Eigenvalue spectrum  for the Slepian sequences. we can control the mean square error (MSE) of the basis expansion defined as MSE (15) (21) is the one-sided normalized Doppler bandwidth in the range . which is a very considerable saving. by a factor of in our application allows us to obtain The pilot pattern . maximum speed of the user rate km h m s. Slepian sequences u : 1 . 3. (16) with the instantaneous values of the basis functions and the correlation matrix (17) where we have used the fact that . We use the following parameters: carrier frequency GHz.

Using the reconstruction expression (13) and evaluating (15) in the sense of an empirical mean provides the numerical . we calculate the inner product with every basis function. The design goal for a basis and no expansion is to have no amplitude error .ZEMEN AND MECKLENBRÄUKER: TIME-VARIANT CHANNEL ESTIMATION USING DISCRETE PROLATE SPHEROIDAL SEQUENCES 3601 or normalized to the symbol duration . The complex exponential in (23) is shifted by . Similar performance results for the Fourier basis expansion are reported in [21] as well. We define the instantaneous frequency response of the basis expansion estimator according to (23) where .1. Fig. 256 symbols. thus. variance B. Inserting (28) in (25) and calcuand lating the integral numerically. and var depends linearly on the noise and the dimension of the basis expansion. the power spectral density must fulfill for (30) which is ensured by the physical mechanism behind the Doppler effect in a wireless channel. We depict its square bias after an orthog- . Sec. The same behavior is also visible in Fig. The square bias of the Fourier basis expansion slightly decays toward since this frequency coincides with the one of the Fourier basis function and. Basis Expansion Square Bias Using Niedzwiecki’s results from [4. In fact. basis expansion for The simulation results in terms of bias versus normalized are given in Fig. we obtain an analytic expression for the basis expansion square bias. i. is a local minimum. The phase of expressed by . we plot the results after evaluating (27) in Fig.e. we evaluate the basis expansion estimator for the autocorrelation defined in (20). is the instantaneous phase shift of the basis expansion at time index . and for the Fourier basis are obtained by using expansion.. In (23).4] (24) The square bias per symbol bias of the basis expansion estimator can be expressed as the integral over the instantaneous multiplied by the power spectral error characteristic (see [4. which is expansion at time instant . The square bias of the Slepian basis expansion is three magnitudes smaller if compared with the Fourier basis expansion. 6. 4. which corresponds to the Jakes’ spectrum for (28) We will see in Section V-B that bias depends on the actual set of basis functions. Therefore. Real channels do not show Jakes’ Doppler spectrum. 6] and specializing to our application. 4 plots the square bias results under the condition square bias of the the Slepian basis expansion and the Fourier and pilots. To achieve optimum performance with the Slepian basis expansion. Then. 6. we substitute (29) into (16). the realization at time instant is calculated by left multiplying with . we use the autocorrelation (20). and the time The length of the data block is . The results for the Slepian basis expansion in (16).4]) density of bias where is given by (26) The square bias for a block of length bias bias is given by (27) (25) elsewhere. the sum projects the complex exponential onto the basis function at the pilot grid positions. a set of nonorthogonal complex exponential basis with and functions is used.1. This basis spans an approximation of the Slepian subspace by definition. Sec. The channel estimation is performed according to (18). 6. In [22]. we are able to evaluate the basis expansion estimator square bias for any autocorrelation. thus. Sec. with (25). The MSE (15) of the basis expansion can be described by the sum of a square bias and a variance term MSE bias var (22) In order to obtain analytic performance results. 4. We average over 2000 channel realizations of the channel defined in Section V-A. The numerical square bias results are obtained using the signal model defined in Section III for . and . All basis expansions use index is restricted to the same amount of parameters to model the channel . as was proven by measurements in [5]. rameters for our Slepian basis expansion are Only to allow for easier comparison with other publications. the instantaneous phase error error characteristic of the basis expansion is defined as [4. Note that the Slepian basis expansion estimator only exploits for and does not require any that other knowledge about the Doppler spectrum of the time-variant channel for the design of the basis functions. is the instantaneous amplitude response of the basis . together with the Doppler bandwidth analytic results that show a perfect match. The design paand .

and v = 27. Fig. . 5 plots the analytic result for the square bias per symbol (25). We vary the Doppler bandwidth in the range 0    3:8 1 10 . 6. the variance var (32) decreases. 6 evaluates the basis expansion estimator for a signal-todB. Fig.1. The square bias curve of the Slepian basis expansion has the same qualitative behavior toward the block boundaries. 5. NO. we fix in the range of 0 to 30 dB. 7 plots the analytic and simulation results for a varying with number of pilot symbols and dB. and the square bias . SEPTEMBER 2005 Fig. . 6. which we will use later for data detection. 9. Square bias per symbol bias [m] for the Fourier and the Slepian basis = D = 5.") for five basis functions. With increasing . but the . VOL. Using the results from [4. 4 (denoted "Slepian approx.3602 IEEE TRANSACTIONS ON SIGNAL PROCESSING. MSE of the basis expansion estimator with J at an E =N = 10 dB for one user moving with v corresponding to  = 0 . The Slepian basis expansion needs fewer parameters and shows better performance than the spline approximation investigated in [23]. The MSE of the Slepian basis expansions is smaller than the MSE of the Fourier basis expansion. J = 10 pilots. and the distance to the Slepian basis expansion is reduced because of the present noise level of dB. Thus.4]. pilot symbols.8 expansion with dimension D m/s. 53. The square bias of the Fourier basis expansion dominates the MSE (22). . Thus. 27 8 pilot symbols : m= s square bias per symbol of the Slepian basis expansion is several magnitudes lower compared to the Fourier basis expansion. 8 shows that the Fourier basis expansion is biased and that its MSE saturates for increasing at 3. 4 (denoted "Polynomial") for five basis functions. bias for the Fourier and the Slepian basis expansion with dimension D = D = 5 and J = 10 pilots. C. Additionally. Basis Expansion Variance For channel estimation purposes. Fig. The error floor of the Fourier basis expansion can be explained due to the structure of (22). and noise ratio of . we define (31) Note that this definition is different from . its square bias varies heavily over the normalized Doppler range of interest. cf. Fig. both with five basis functions.5 . onalization step in Fig. (50). Fig. The Slepian basis expansion is (practically) unbiased."). This basis is not restricted to time-limited and bandlimited functions. = 10 = 0 . . as shown in Fig. 4. and and vary Finally. var inEquation (32) becomes exact for bias and decreases with the dimension of the basis expansion creases with an increasing number of pilot symbols but does not depend on the block length . we depict the square bias for the polynomial approach of [20] (denoted "Polynomial") and the nonorthogonal complex exponentials used in [11] (denoted "Slepian approx. Sec. 3:8 1 10 . This shows clearly that the square bias per symbol of the Fourier basis expansion is highest at the edges of the data block. A polynomial approach resulting from a Taylor expansion was used in [20]. we can express the basis expansion estimator variance as var (32) Fig.

random interleaving. for symbol set . The base station transmits quaternary phase shift keying with symbol rate (QPSK) modulated symbols .d. 8. In order to accommodate for time-variant channel estimation. the spread signals of all users are added together. However. in order to achieve the same performance as with perfect channel knowledge. chosen with equal probability from the scaled QPSK . and a perfect power control is assumed. We ignore the effects of path loss and shadow fading for (33) Fig. Thus. 9 depicts the downlink transmitter schematically. and the Slepian basis expansion channel estimator. and common pilot with elements are added symbols (34) where (35) and Fig. The elements of for and the pilot symbol vector are i. an -point inverse DFT is performed. The channel varies significantly over the duration of a long data block. . A data block consists of and OFDM pilot symbols. MC-CDMA FOR TIME-VARIANT FREQUENCY-SELECTIVE CHANNELS In this section. (36) contains the stacked data symbols for users. The transmission is block OFDM data symbols oriented. data symbols are distributed over a block of length the . Every OFDM symbol is preceded by a cyclic prefix to avoid ISI. Signal Model In MC-CDMA. if the channel variation in time. measured by the normalized bias (27) starts to dominate the MSE since it is independent of . and QPSK modulation with code rate with Gray labeling.ZEMEN AND MECKLENBRÄUKER: TIME-VARIANT CHANNEL ESTIMATION USING DISCRETE PROLATE SPHEROIDAL SEQUENCES 3603 A.i. The OFDM symbol including . Each symbol is spread by a random with elements . After parallel serial the cyclic prefix has length conversion. and the user index is denoted by . a data symbol is spread by a user-specific spreading code. The data result from the binary information sequence symbols of length by convolutional encoding . Then. We provide simulation results showing that this system needs a pilot ratio of 2% only. where the pilot placement is defined through (4). otherwise. 15g pilot symbols at an E =N = 15 dB for one user moving with v = 27:8 m=s corresponding to  = 3:8 1 10 . As shown in Fig. the data symbols satisfy for and for . MSE of the basis expansion estimator with J 2 f5. we define the signal model for a MC-CDMA downlink in a time-variant channel. transmitted over a time-variant multipath fading channel We denote the chip-rate sampled time-variant impulse response by (37) A time-variant channel impulse response generally introduces intercarrier interference in an OFDM system. . . it is clear that the Slepian basis expansion offers major performance gains compared with the Fourier basis expansion while having the same complexity. the linear MMSE multiuser detector. 7. . There are users in the system. MSE of the basis expansion estimator for J = 10 pilot symbols at an E =N = 0 . and a cyclic prefix of length is inserted [24]. 30 dB for one user moving with v = 27:8 m=s corresponding to  = 3:8 1 10 . 9. VI. spreading sequence Fig. The resulting chips are processed by an inverse DFT to obtain an OFDM symbol. . From this analysis. . allowing for pilot symbol insertion. . the chip stream with chip rate is .

The code symbol estimates are given by After demapping and deinterleaving. where B. The bandlimited property of as well. The received signal vector after these two operations is given by diag (42) and the time-variant effective spreading matrix to express the time-variant linear MMSE filter unbiased (44) . Slepian Basis Expansion Channel Estimator The performance of the receiver crucially depends on the . Comparing (45) with (2). We define where complex additive white Gaussian noise with zero mean is denoted by with elements and covariance . although the channel is time variant.3604 IEEE TRANSACTIONS ON SIGNAL PROCESSING. the intercarrier interference can be neglected for the receiver side processing [25]. we see that the structure of these equations are directly applies identical. For the processing at the receiver side. VOL. a hard decision is . we are able to treat the time-variant channel as constant for the duration of each single OFDM symbol if (38) is fulfilled. In order takes place over to reflect this. stays below a certain threshold. Hence. Multiuser Detector The linear MMSE receiver detects the data using the received vector . Further noise suppression is obtained if we exploit the correlation between the subcarriers . subcarrier (45) where are the elements of [see (34)]. The estimated time-variant frequency response is given by . 9. the spreading matrix . Hence (39) respectively (40) in vector notation. any basis expansion can be used for performance comparisons. 53. SEPTEMBER 2005 Fig. we rewrite (42) as a set of equations for every . performed to obtain the transmitted data bits C. channel estimates for the time-variant frequency response The MC-CDMA signal model describes a transmission that parallel frequency-flat channels. which is assumed to be known for the quency response moment [18]. After the decoder. and the time-variant fre. 9. The time-variant frequency response with elements is defined as the DFT of the time-variant impulse response (41) . we keep the notation regarding the basis expansion generic. Model for the MC-CDMA transmitter and the channel in the time-variant downlink. Doppler bandwidth. This allows us to estimate the time-variant to with the Slepian basis expanfrequency-flat subcarrier sion (18). the code bit estimates are supplied to the decoder. the DFT is still applicable [26]. where The receiver removes the cyclic prefix and performs a DFT. This threshold is fulfilled if the one-sided is much smaller than the normalized Doppler bandwidth normalized subcarrier bandwidth (38) For . and (46) (47) Although we will use the Slepian basis expansion. We define the time-variant effective spreading sequences diag (43) . NO.

2 GHz. as in the Uninote sampling at rate versal Mobile Telecommunications System (UMTS). 10 illustrates the downlink MC-CDMA receiver performance with users in terms of BER versus . It is important to point out that we use a simulation model for the time-variant frequency-selective channel (see the Appendix) that is completely different from the basis expansion model. Simulation Results The realizations of the time-variant frequency-selective channel . 8. any possible effect from residual ICI would be visible in the simulation results. The block length 256. VII. The Fourier basis expansion and the Slepian basis expansion. The K 2 f32. We use the simulation model (52) for every channel tap . nonsystematic. all using D = 5 basis functions. nonre1/2 code with generator polynocursive. D. Downlink MC-CDMA receiver performance in terms of BER versus E =N . and the The system operates at carrier frequency 19. For data transmission. which leads to an error floor higher in Fig. Taking (51) into account when referring to Fig. as is visible for values in Fig. and its essential support . This gives 126 Hz users move with velocity 0. thus. 10 to the channel In order to relate the estimation performance results in Section III. The Slepian basis expansion is specified according to just two system parameters: the maxand the block length . We also plot the receiver performance for a perfectly known channel (denoted "PER") in Fig. Additionally. and the lowest amount of pilots possible J = D = 5 is used. We compare the Slepian basis expansion and the Fourier basis expansion.0026. use the same dimensionality 5. 10. both applied for performance comparison. we proved the robustness of the Slepian basis expansion for a very realistic class of time-variant channels. which results in 3. 64g users are moving with v = 19. sampled at the chip-rate . and multiuser detection can be performed. The square bias of the Slepian basis expansion .4 m/s. which are limited to the class of Fourier basis expansion channels only. Thus. which we use for equalization. Thus. the count. We evaluate the performance of the Slepian basis expansion channel estimation with the lowest number of pilots . where we took the higher pilot energy in the downlink into ac32 users. the range is from 17 dB 34 dB.5 m/s. we establish the relation (51) Fig. The receiver performance is extremely close to the one with perfect channel knowledge when only as few as 2% pilots are used. The autocorrelation for every channel tap is given by (49) which results in the classical Jakes’ spectrum. 10 are in the range 11 dB 28 dB. 8 makes clear that the drastic reduced MSE for the Slepian basis expansion channel estimation leads to pronounced reduction in bit error rates if compared with the Fourier basis expansion.ZEMEN AND MECKLENBRÄUKER: TIME-VARIANT CHANNEL ESTIMATION USING DISCRETE PROLATE SPHEROIDAL SEQUENCES 3605 . values in Fig. A imum Doppler bandwidth lower bound for the model order results naturally. The simulation uses a chip-rate sampled time-variant channel. Finally. We have compared the Fourier and the Slepian basis expansion for the same numerical complexity (same number of unknowns). and the (50) Fig. The illustrated results are obtained by avmial eraging over 400 independent channel realizations. For the simulation results with values for the simulation in Fig. and the and OFDM symbol with cyclic prefix has length of 79. The discrete time indices and defor s . Each OFDM symbol has a length of chips. This is because the Fourier basis expansion is biased. 10. a convolutional. the plot also shows the single user bound. For 64 users. CONCLUSION We showed that the Slepian basis expansion is very suitable to the modeling of a time-variant frequency-selective channel for the duration of a data block. the channel estimates are inserted into (43). The system is designed for 28. which is 1 and a perfectly defined as the performance for one user known channel . four-state. For reference the single-user bound (SUB) and the performance with perfect channel knowledge (PER) are shown as well. 10. The QPSK is defined as symbol energy is normalized to 1. The number of subcarriers 64. rate is used. we have as many pilots as basis possible functions in the basis expansion. Our results are therefore more general than the one presented in [11]. The root mean square delay spread of the power-delay profile in (48) is s. are generated using an exponentially decaying power-delay profile according to COST 259 [27] (48) .4 m/s. The pilot ratio is 2%.

Telecommun. Verdú. Jul. 1969–1986. vol. [4] M. 12. vol. 9/10. IEEE Int. 1996.0023 We presented simulation results for a fully loaded 64 MC-CDMA downlink for users moving at 19. Inf. F. 1996. Tech. Giannakis and C.. Comput. Multi-Carrier Spread-Spectrum Workshop. 57. J.” Ann. Applying the Slepian basis expansion for channel estimation. Englewood Cliffs. vol. no. Hlawatsch. pp. Rappaport. Conf. Multiuser Detection. Vainikainen. the distribution of the channel coefficients is significantly different from a normal distribution. Microwave Mobile Communications. C. 2003. Areas Commun. and R.. invited paper. Matz. Pollak. [22] G. [3] T. Überhuber. no. 1065–1076.. Pacific Grove. no. Commun. [8] G.. Sep. 2003. A. Germany. Nov. Fourier analysis. “Maximum-diversity transmission over doubly selective wireless channels.” IEEE Trans. Sayeed. Press. “Frequency-selective fading channel estimation with a polynomial time-varying channel model. 688–691. pp. pp. May-Jun. vol. 1998. 5.” Int. Leus. number of interfering paths to In the limit . Sayed. B. and uncertainty—V: The discrete case. 14–23.” in Proc. Skog. In this case.” Bell Syst. 4. Moonen. no. VOL. 2003. C.” in Proc. [19] H. “Multi-input multi-output fading channel tracking and equalization using Kalman estimation. Institute for Analysis and Scientific Computing. 451–454. [10] A. SEPTEMBER 2005 compared with the Fourier basis expansion is three magnitudes smaller.. Apr. no. 96. Upper Saddle River.” in Proc. no. Nov. “Prolate spheroidal wave functions. 57. no.. Mar. Giannakis. [2] M. May 2004.4 m/s using a classic linear MMSE multiuser detector. pp. Speech. Niedzwiecki. CA. the model converges to a block fading channel. 1295–1336. D. 52. [6] D. Theory. Jun. REFERENCES [1] C.. For the numerical simulations. Tepedelenlioglu. H. and J. Conf.” Bell Syst. Zemen and C. vol. T. Zemen. pp. J. Proakis and D.3606 IEEE TRANSACTIONS ON SIGNAL PROCESSING. 3246–3250. “Multiuser detection in fast-fading multipath environment. we achieved the same bit error rate performance as with perfect channel knowledge. Borah and B. Oct. vol. vol. Technol. 5. the channel coefficients are normally distributed. . 461–468.” in Proc. IEEE Int. [14] T. “Time-varying FIR equalization of doubly-selective channels. Tech. Therefore. (53) with for where . Schafhuber. 6. Fourier analysis. 50. Veh. no. [20] D. [11] G. M. “Modeling and equalization of rapidly fading channels. R.and frequency-selective channels. and M.” IEEE J. NO. Wehinger and R. 1. J. 2003. B. 2003. 49. Ma and G. Technical University Vienna. Moonen. pp. 1371–1430. vol. Anchorage. vol. Commun. [16] T. C. pp. Oberpaffenhofen. “Basis expansion models and diversity techniques for blind identification and equalization of timevarying channels. Adlard.” IEEE Trans. ACKNOWLEDGMENT The authors would like to thank J. 2003. Digital Signal Processing.and band-limited signals.” IEEE Trans. [13] I. pp. Adaptive Control Signal Process.. 1974. Wireless Communications—Principles & Practice. G. 2/3.” in Proc. [7] P. F.” IEEE Trans. Sep. [12] J. 10. Leus. Technol. 1942–1950. G. 1691–1701. Veh. IEEE. 60–74. vol. Hart. 41. Kivinen. P. 1198–1208. 4. Telecommunication Research Center Vienna (ftw. 53. no. D... no. Pohl. Leus.” IEEE Trans. Dec. no. Giannakis. Aazhang. which helped to improve the paper. 1992.. Syst. “Low-complexity serial equalization of doubly-selective channels. vol. New York: Cambridge Univ. Hoeher. for valuable comments and suggestions. 10. pp. pp. K. Barhumi. O. and K. 1288–1292. Zhao. “Time variant channel equalization for MC-CDMA via fourier basis functions. 2000. “Adaptive Wiener filters for time-varying channel estimation in wireless OFDM systems. “Characterization of Doppler spectra for mobile communications at 5. A. Signals. 873–924. APPENDIX SIMULATION MODEL FOR TIME-VARIANT CHANNELS WITH JAKES’ SPECTRUM We use the model from [28] with a correction for low velocifor ties [29]. [23] Y. 4. AK. pp. and F. “Time-variant channel equalization via discrete prolate spheroidal sequences. 1832–1840. [17] W. Inf. May 2003. NJ: Prentice-Hall. and B.. “Orthogonal multiple access over time. 52. vol. pp. [15] D. B. New York: Wiley.). 1962. 5. S. Sel. Jakes. they would like to thank the anonymous reviewers for their comments. Tozer. Third ed. Zhou. Müller. pp. 5. vol. Mecklenbräuker. [18] S. Slepian. NJ: Prentice-Hall. and R. Identification of Time-Varying Processes. Spain. J. 4. and G. B. We showed that the Slepian basis expansion outperforms the polynomial approach in [20] for normalized Doppler . 1999. and are independent and uniformly distributed for all . May 2002. Landau and H. Manolakis. 1998. Mecklenbräuker. 9. pp. Tsatsanis and G. [9] M. 47. no. Siala. and A. Theory. 16. Signal Process. Aug. Giannakis. “Prolate spheroidal wave functions./Oct. Acoust. pp. Baiona Workshop Signal Process. Fragouli. “Polynomial spline-approximation of Clarke’s model. Jul. Sep. Komninakis. I. Our model ensures a Rayleigh distribution of m s. 37th Asilomar Conf. pp. pp. C. Signal Process. 2002. Barhumi. The pilot ratio for channel estimation was 2% only. we fix the over . Sendonaris. and M. 3. “Maximum a posteriori semi-blind channel estimation for OFDM systems operating on highly frequency selective channels. spreads 0.” IEEE Trans. F. 1998. 2003. vol. vol. New York: Wiley. 41. 1978. 159–176. 862–873.3 GHz. Signal Process. Müller. Zakharov. 49. Commun. “A statistical discrete-time model for the WSSUS multipath channel. The detailed simulation all velocities and even at is defined as follows: model for (52) Finally. K. [5] X.. [21] X. Baiona... In [28]. no. pp.” IEEE Trans. (53) reduces to (54) Because of the central limit theorem and the independence of all and . Siemens Austria PSE PRO Radio Communication Devices (RCD). .” Proc. Jan. is replaced by the common phase . Wesel. G. J. and uncertainty—III: The dimension of the space of essentially time. V. G. 1996.

respectively.. 401–404.” He was a delegate to the Third Generation Partnership Project (3GPP) and engaged in the standardization of the radio access network for UMTS. . vol. 920–928. vol. 17. His research interests include orthogonal frequency division multiplexing (OFDM). 3. 29–48.-Ing. 4041–4071. and software defined radio concepts. He received the Dipl. both with distinction. Phys. from 1997 to 2000. Austria.” J. Mar. Cimini. New York: Wiley.. R. from Vienna University of Technology.. G. degree in electrical engineering from Vienna University of Technology. Correia. and the Dr. “Doppler diversity in MC-CDMA using the Slepian basis expansion model.” IEEE Trans. [28] Y. He has authored approximately 18 papers in international journals and conferences. He received the Dipl-Ing.). Austria. Sep. 51. Zheng and C. Currently. pp. Commun. Wireless Flexible Personalised Communications. Since October 2003. “Time-frequency transfer function calculus (symbolic calculus) of linear time-varying systems (linear operators) based on generalized underspread theory.” IEEE Trans. Bochum. 2001. where he participated in the European framework of ACTS 90 “Future Radio Wideband Multiple Access System (FRAMES). he has lead the “Future Mobile Communications Systems—Mathematical Modeling Analysis and Algorithms for Multi Antenna Systems” project. “Bounds on the interchannel interference of OFDM in time-varying impairments. he has been with the Telecommunications Research Center Vienna. degree from Ruhr-University. Giannakis. time-variant channel estimation. he teaches 3G Mobile Networks at the Vienna University of Technology. and ultra-wideband radio. Vienna. in 1998. [27] L. for which he has also served as a reviewer. Conf. no. He worked for Siemens. Vienna. Vienna. Currently. His current research interests include antenna arrayand MIMO-signal processing for mobile communications. Matz and F. Vienna (ftw. in 1998. Dr.. J. B. Austria. iterative receiver structures. in the field of mobile communications. Mag. May 2000. vol. both in electrical engineering. Hlawatsch. Commun.Germany.. Thomas Zemen (S’03–M’05) was born in Mödling. Jun. Wang and G. Austria. 2001. Since 2000. Li and L. for which he has also served as a reviewer and holds eight patents in the field of mobile cellular networks. From October 2001 to September 2003. he teaches MIMO communications at the Vienna University of Technology. no. He has authored around 60 papers in international journals and conferences. 2003. Since 2005.” in Proc.” IEEE Signal Process. Signal Process. 1998. vol. Mecklenbräuker. 49. no. “Wireless multicarrier communications. pp. Mecklenbräuker (S’88–M’97) was born in Darmstadt. M. degrees (both with distinction) in 1998 and 2004. 3. Zemen and C. Christoph F.-Ing. Vienna. which was funded by the Vienna Science and Technology fund (WWTF). [29] T. [25] Y. in 1970.techn. Vienna. 12th Eur. multiuser detection. pp. working as Researcher in the strategic I0 project. where he was hardware engineer and project manager for the radio communication devices department. “Simulation models with correct statistical properties for Rayleigh fading channels. He joined Siemens Austria. Mecklenbräuker is a member of the IEEE Signal Processing and Antennas and Propagation societies and EURASIP. Xiao. His doctoral thesis on matched field processing received the Gert–Massenberg Prize in 1998. Germany. [26] G. 39. Math. pp. in 1992 and the Dr. he has held a senior research position at the Telecommunications Research Center. 2004. he was delegated by Siemens Austria as a researcher to the mobile communications group at the Telecommunications Research Center Vienna (ftw. 6. in 1967.). F.ZEMEN AND MECKLENBRÄUKER: TIME-VARIANT CHANNEL ESTIMATION USING DISCRETE PROLATE SPHEROIDAL SEQUENCES 3607 [24] Z. Aug.

Sign up to vote on this title
UsefulNot useful