Kalman Filter Based Channel Estimation Method For MIMO-OFDM System with Low Complexity

Ganesh Miriyala
Department of Electrical Engineering National Institute of Technology Hamirpur Hamirpur, India gani.india@gmail.com
Abstract—In multi input multi output orthogonal frequency division multiplexing (MIMO-OFDM) systems, a low complexity subspace based time domain channel estimation method is studied. This method works based on parametric channel model, where the response of the time varying channel is considered as a collection of sparse propagation paths. Translate the estimation of channel parameters into an unconstrained minimized problem by considering the channel correlation matrix. To solve this problem, Kalman filter based method which tracks the subspace is proposed, which provides the constant subspace to construct state equation and measurement equation. This method represents a low complexity subspace schemes. The approach can be extended to MC-CDMA systems. The simulation results prove that the Kalman filter method can track faster fading channel, and is more accurate with low complexity. Keywords— Channel estimation, Kalman filter, low rank adaptive filter, subspace tracking and time varying channel.

decomposition should operate in the algorithm flow. Givens rotation and Householder transform are the conventional approaches to QR decomposition with a complexity of O( N 3 ) , for N × N matrix. Hence, the complexity of the QRRecursive Least Square (QR-RLS) method is high as O ( N p r 2 ) , where r is the number of channel paths and N p is the number of pilot subcarriers. Because of low complexity to track the channel variation compared with decomposition or matrix inverse Kalaman filter is presented in [5]-[7]. In this method, it won’t consider the low dimensional signal subspace, i.e., they operate upon high dimensional signal subspace channel matrix, which leads to weaker performance at low signal to noise rate (SNR). A time varying pilot-aided parametric channel estimation method by tracking the subspace in MIMO-OFDM systems with Kalman filter has proposed. For time varying multi path channels, variation of delay is slow, but not of amplitudes, that gives the guarantee to the valid of this adaptive filter. Accuracy of estimation is significantly increased through the adaptive tracking of a low dimensioned signal subspace. The applications of Kalman filter provides fast convergence with low complexity O ( N p r ) . This method is compatible in low complexity and low dimension compared with existing methods. II. A. Channel Model Consider a time varying MIMO-OFDM system with mT transmit, mR receive antennas with K subcarriers. The baseband impulse response of wireless channel between j th transmit antenna and the k th receive antenna is given as h jk (t ,τ ) =
N p −1 m=0



Combination of Multi-input Multi-output (MIMO) with the orthogonal frequency division multiplexing (OFDM) has received a great extent attention in recent times to combat multi path delay spread, and increases the system capacity. MIMIOFDM has turn out to be one of the most promising technique for the 4G (fourth generation) of wireless communication systems. For the reason that the necessity of high mobility for 4G systems, the signals would experience fast time varying channel environment at MIMO-OFDM receivers and strong Doppler frequency spread interface. This requires the receivers which are exactly estimate, real time track the variation of multipath channels with low complexity. In recent years, parametric channel estimation methods have been studied widely in OFDM systems. These methods model the radio channels by a few dominant sparse paths [1]. Hence, parameters of channel paths are estimated before the stage of interpolating over the entire frequency time grid, which are including complex amplitudes and delay. In [2], [3], subspace analysis methods are presented. These methods obtain the channel parameters. But, cumulating the large amount of samples for these methods are very complex. A subspace tracking method is presented in [4]. This method employs to track the delay subspace by using a QR decomposition based Recursive Least Square filter. This QR
†. M.tech student



( jk ) m

( jk ) (t )δ (τ − τ m (t ))


( jk ) Where N p the number of paths, τ ( t ) is m th path delay and

am ( jk ) ( t ) is corresponding amplitude. By consider the constant delays in M symbols provided that variations in M symbols

. are the pilot position values of X The Least Square time varying channel estimated of hLS (n) are given by H % H H % H ˆ (n) = (W % % % −1 % % h LS p X p ( n) X P ( n)WP ) W p X p ( n )Y p ( n ) [W jk ]s .. LS Channel Estimation Consider the same pilot pattern for each transmit antenna.. h jk (n. we can write % % % % % % %% Y p ( n ) = X p ( n ) H p ( n) + ξ p ( n) = X p ( n )W p h p ( n ) + ξ p ( n) % = diag (W % W %L W % ).. N p ) as rank of matrix φn . translate the estimation of time varying channel parameters into an unconstrained minimization problem.) is the convolution of the pulse shaping filter and the matched filter at the receiver.. l ) = N p −1 m =0 This method reduces the total number of delay vector to be estimated and tracked from N p to rn .. the transmit antenna index j will be omitted. 2. (i = 1. by considering only pilot-sub-carriers. .L . g (−τ N p (n))   g (−τ 1 (n))   g (1 − τ 2 (n)) . the complete procedure of the parametric channel estimation for H (n) consists of three steps. α n = [a1 (n) a2 (n) L a Np (n)] is complex gain of the time varying channel. compare with the time resolution (1/B ) of the system.. h jk (n. then the equation can be written as h(n) = φnα n Where g (−τ 2 (n)) . m ) . Because of Doppler frequency shift..1). g (. and they works based on the consideration given below. Finally. which is Kalmn filter. First... We define rn = rank (φn ) ≤ min( L. The rn -dimensional subspace spanned by the columns of the φn is called delay subspace and it’s denoted by L × rn orthonormal basis as U n ( L × rn ) .    g ( L − 1 − τ 1 (n)) g ( L − 1 − τ 2 (n)) . A.1. Then the equation (5) can be written as h ( n) = U n d n Where r × 1 vector dn = Λ V α n H n n IV. Block diagram of parametric channel estimation φn is matrix of L × N p ..L . then the sampled channel response can be expressed as h jk (n. H p % % (n) and H (n) .1.are much smaller than the resolution of the OFDM system. 1. g ( L − 1 − τ N (n))  p   (5) ˆ (n). . Where W p p1 p2 pmT (8) pl s % ] = 1 e − j 2π K [W pi s ... CHANNEL ESTIMATION BASED ON PILOTS ∑a ( jk ) m ( jk ) (n) g (l − τ m (n)). S .1. the time varying channel response at all of the N subcarriers is obtained by smoothing or interpolation. L and X p (n). By considering the pulse shaping filter is used in both transmitter and receiver. which is decided by pulse shaping T filter and multipath delays. P = [ p1 p2 L pT ] being set of T pilot tones in each OFDM symbol. Step1: For low-dimensional signal subspace. the time varying channel parameter of each path are estimated and tracked to improve the accuracy of estimates of Least Square results. l = 0. L − 1)] . Delay-Subspace In (2).... l = 1.L . And then we get for all antennas as h i . g (1 − τ N p (n))   g (1 − τ 1 (n)) φn =   ..l K % % ( n) s = 1.l = 1 K e j 2π sl − K (s = 0. In (4). mT . i = 1.... These delays are not integers in general so the channel tap power will leak to other samples [8]. The frequency response of the channel in the n th OFDM symbol can represented as H jk = W jk h jk (n) (3) T Where h jk (n) = [h jk (n. .. l = 0. Because of channel estimation method for each transmit antenna is same. Step2: Solve the step1 problem via subspace tracking with the adaptive algorithm. S − 1. if the number of time delays and multi-paths are known..... τ ( jk ) ( t ) are the sampled time delays. L − 1 (2) Because of the time delays are unknown.. CHANNEL ESTIMATION BASED ON SUBSPACE TRACKING As shown in Fig. 2. 0). III. inter-carrier interference ( jk ) is neglected. the time varying channel response at each pilot-subcarrier is computed with Least Square (LS) method. LS T Figure 1. (6) (7) . By using the concept of singular value decomposition (SVD) φn = U n Λ nVnH . Then. L − 1) (9) B. The presented method with adaptive algorithm is used in second step. Consider L as maximum length of the channel. 2... 2.L ... am ( t ) assumed to be quasi-fixed over one OFDM symbol.

0) = I r . U n n n   Estimator Output: ˆ ∗ (n)U ˆ ∗ H (n) x(n). So. and mainly focused on low-dimensional subspace and to avoid the matrix inverse or decomposition. the cost function is given by 2 J (U n . ζ e (0) = c1 0  For each OFDM symbol n = 1. Tracking the signal subspace of R is guaranteed by solving the global convergence of J (U n ) . n) = K (n. hLS (n) can be represented as hLS (n) = U n d n . and % (n) is minimal error vector when adapting U Where e ζ (18) (12) Minimizing the equation (12). Using Kalman filter to reach U n ∗ ˆ y (n) = U (n) x(n) . n n n (11) ˆ∗) expanding J (U n ˆ ∗ ) = E[h H (n)h (n)] . Iff U n = U n . p p ( n) / X p ( n ) [4]. then 2 ˆ ∗ ) = E ζ H ζ  − E ζ H U ˆ ∗ ˆ ∗H  J (U (15) n  n n  n nU n ζ n  < N pσ ζ ∗ ˆ adaptively. c1 = 10 and c2 = 0. the tracking operation of each OFDM symbol is 2 low as O ( N p r ) when compare to QR-RLS with O ( N p r ) . n − 1). the estimation of φn is equal to tracking the signal subspace U n . d n n LS estimator for subspace U n . A.By the assumption of that the total number of channel paths L is obtained from the minimum description length (MDL) method as mentioned in [4]. and J (U n ) has global minima where U n is equal to signal subspace. Therefore. such that U point U n n Un n then % ( n) = φ α + ζ h LS p. Comparision of Kalman Filter and QR-RLS Methods Kalman filter based method is used to obtain the signal subspace.L ˆ ∗ (n − 1) x(n). n − 1) y (n) ]   y (n) K (n. In order to get the least mean square error (MSE) of estimation. which is nothing but the projection of hLS (n) to the signal ˆ = U H h (n) . we define the new cost function ˆ = arg J (U ) U n n Un H  = arg E   hLS (n) − U nU n hLS (n)  Un The channel estimation algorithm based on subspace with Kalman filter is given as Initialization: I  ∗ K (1. r ∈ £ any r distinct number In order to get the better convergence. We can solve this problem by three following methods to analyze their computational complexity. By ˆ ∗ ∈ £ N × r is exist. x(n) = h (n). a unique ˆ ∗ = arg min J (U ) . y (n) = U LS n H g (n) = [ K (n. The estimates of α n and delays are converted into optimization problem for U n . n − 1) y (n) + ζ e (n − 1)  . Kalman filter for Subspace Tracking ( In (13). update ζ e (n) for each OFDM symbol. U n U n = I r . H K (n + 1. then construct the state equation n ˆ ∗ (n + 1) = U ˆ ∗ ( n) U n n and the measurement equation ˆ ∗ ( n) y ( n) + e % x ( n) = U ( n) n ∗ n 2 H  x ( n) − U ˆ ∗ (n)U ˆ ∗ H ( n)  % % E ( n )e ( n)  n n e  = E    ∗ 2 ˆ )< N σ =J (U n p (16) (17) = U n ∑ U nH + σ ζ2 I N p Where U n consists of r eigenvectors corresponding to main Eigen values of diagonal matrix ∑ . n − 1) − g (n) y (n) K (n. it is proved in ( N ×r [10]. can be achieved by minimizing the (13). x(n) = hLS (n) . and let h LS LS LS LS p r ×r of eigenvectors of R . ˆ ∗ ( n) = U ˆ ∗ (n − 1) +  x(n) − U ˆ ∗ (n − 1) y (n)  g H (n). n is Vander monde matrix of size N p × L . By observing the equation (13). ζ n and d n are uncorrelated. To get the subspace U n . ψ ∈ £ is an arbitrary unitary matrix. we will get the estimator of d n . hKalman (n) = U n n (13) ζ e (n) = c2 ⋅ norm ( hKalman (n) − x(n) ) . d n ) = E  hLS ( n) − U n d n    ˆ . As given in [4]. For this method.75 are adjustable constant factors. U n (0) =  r  . 2. J (U n ) has a stationary point at U n . Unconstrained Minimization Problem for Subspace The Least Square estimation for time varying channel with %(n) is h % ( n) = Y % % pilot-sub-carriers h p LS . B. and it is a collection of rows corresponding to pilot-sub-carriers in φn . ζ n is a vector of AWGN noise with variance σ ζ2 . apply the tracking technique. C. From (6). rψ where U n . subscript p. for given least square results of each hLS (n) .2 E[h H (n)U ˆ ∗U ˆ ∗H J (U n LS LS LS n n h LS ( n)] ˆ ∗U ˆ ∗HU ˆ ∗U ˆ ∗ H h (n)] +E[h H (n)U LS Where φ p . Omit % ( n) = [ h % (0) h % (1) L h % ( N − 1)]T . we can state that the signal subspace tracking as a solution to the minimization problem. By defining correlation matrix as H R = E[hLS (n)hLS (n)] H = U n ΛV H E[α nα n ]V Λ H U nH + σ ζ2 I N p (14) n n n n LS ∗ ∗H ˆ ∗ ˆ ˆ Observe that hLS (n) = U n d n + ζ n .

QPSK modulation scheme. ICI tends to be serious. The MSE of channel estimation versus SNR at fdmax = 40Hz and 200Hz. Kalman filter method reduces the MSE and BER of channel estimation almost one order of magnitude less than QR-RLS method. and optimal pilot design is used. But complexity of Kalman Filter is low. . RLS and LMS methods cannot find the AWGN noise and ICI. But. The BER of channel estimation versus SNR at fdmax = 40Hz. The curves of RLS and Kalman filter are almost super imposed with same error flow and convergence speed. But Kalman filter method consumes lower complexity. ζ n in (11) consists of ICI terms and AWGN noise. The time varying channel model is COST 207 typical urban (TU) environment with multi-paths N p = 6 . SIMMULATION RESULTS Fig 4 and Fig 5 gives the performance of MSE and BER of channel estimation with the variation of signal noise ratio of QR-RLS and Kalman filter respectively with maximum Doppler frequency shift fdmax = 40Hz and 200Hz. L = 16 is the maximum length of channel. [11].4 km/h and 72km/h respectively. the maximum Doppler frequency shift is 40Hz. Here. The MSE of channel estimation versus SNR at fdmax = 40Hz. Fig 2 and Fig 3 gives the performance of mean square error (MSE) and bit error rate (BER) of channel estimation with respect to the signal noise ratio of QR-RLS estimator and Kalman filter with optimal pilots. each OFDM symbol consists of K p = 32 pilots.given in [4]. proposed Kalman filter method has capability to suppress the ICI and gives the better performance in this channel. The performance of the proposed time varying channel estimation method has been done by conducting the simulation for MIMO-OFDM systems with two transmit antennas and two receive antennas. The simulation is performed in Doppler frequency shift of 40Hz and 200Hz with vehicular velocity of 14. The system consists of 256 subcarriers occupying the bandwidth of 2 MHz. Figure 3. we can interpolate the frequency time grid. Figure 2. Figure 5. If fd increases. V. With the estimator output of the time varying channel hKalman (n) . Figure 4. so that once there exists Doppler frequency spread. The BER of channel estimation versus SNR at fdmax = 40Hz and 200Hz.

“A subspace blind channel estimation method for OFDM systems without cyclic prefix. Y. [3] S. Jan. “Optimal training design for MIMO OFDM systems in mobile wireless channels”. 2. IEEE Trans. Mar. vol. vol. 3.pp. vol.2005. K. 1283. Sang-Wook Lee. vol. 9-12. 50. 1995. P.K. 443449. B. 2. “Channel estimation for OFDM transmission in multipath fading channels based on parametric channel modeling. Aronsson. “Channel estimation for OFDM with fast fading channels by modified Kalman filter. Cai. [6] M. 50. M. . Duhamel. Kalman filter method is best one. 2003. pp. Jul. Consumer Electronics. pp. On Broadcasting.” in VTC Fall. vol.Signal Processing.” in PIMRC. 2003. [10] Bin Yang. and Z.Analysis of Pilot-Assisted Channel Estimators for OFDM Systems With Transmit Diversity [J]. 1699-1712. G. 467478. Xu. Systems and Computers. REFERENCES [1] B. 5. Sternad. vol. Letaief. de Courville. F. Schafhuber. “Projection approximation subspace tracking. Jun-Seok Lim. S. The key of this method lies in the idea that the estimate of time varying channel parameters can be translated into an unconstrained minimization problem. Bar-Ness. “Kalman tracking of timevarying channels in wireless MIMO-OFDM systems. IEEE Trans. pp. 2. “Pilot-based channel estimation for OFDM systems by tracking the delay-subspace.” IEEE Trans.51(6): 1615-1624. 2003. U.Li. and M. Matz. Z. Cao. Li. 49. May 2004. 43.” in Signals. vol. pp. vol. Muquet.1283. Oct. D. [9] ongxu Shen. Leus. Simeone. pp. The signal subspace of the time varying channel correlation matrix is tracked by the Kalman filter.CONCLUSION A Kalman filter based time domain pilot aided parametric channel estimation method is presented. [4] O. 2002. [7] D. pp. Jan. Zhifeng Diao. “Channel estimation and prediction for adaptive OFDM downlinks. pp. Barhumi. G. “An improved channel estimation scheme for OFDM systems by tracking the subspace. Moonen.. Koeng-Mo Sung. [11] Y. vol. Kai-Kit Wong and Victor O. 572-579. Cheng. 2003.” IEEE Trans. [5] Ki-Young Han. IEEE. This method consumes low complexity. R. Signal Processing. Commun. apply this method via subspace tracking. 52(2):193-202. Wireless Communications. Wireless Communications.1109-1113. Chengyang Li. Conference Record of the Thirty-Seventh Asilomar Conference. 315-325.” IEEE Trans. IEEE. “Subspace-based blind and semi-blind channel estimation for OFDM systems.” IEEE Trans.” IEEE Trans. [2] B. Roy. pp. 2002. 1261-1265 Nov. 95-107. For Doppler frequency spread channels.” IEEE Trans. Yang. 2001.2003. Li. Y.Signal Processing. [8] I. 2004. Spagnolini. Hlawatsch. Y. To solve this problem.