Reduced Complexity Decision Feedback Equalizer for Supporting High Mobility in Wimax

Iulia Ivan, Bertrand Muquet

Sequans Communications, 19 Le Parvis, 92073 Paris La Dfense, France
AbstractIn this paper, we consider the performance of Wimax systems in fast varying environments. While optimal in static channels, we first show that the conventional single tap OFDM equalizer is not satisfying even at moderate terminal speeds of 60km/h. We then study the performance of linear sliding window equalizers in the Wimax context before focusing on decision feedback equalizers (DFE). Based on our analysis, we propose a new DFE equalizer using sliding window that offers a good trade off between performance and complexity. Index TermsWimax, 802.16e, equalization, intercarrier interference, OFDM, OFDMA, time-varying channels, DFE, sliding window

factorization, but this technique still requires inverting the entire matrix. Recently, several interesting structures based on sliding window equalizers have been proposed. The decision feedback equalizer (DFE) proposed in [5] yields very good performance but at the price of inversion of large matrices. Besides, it requires the knowledge of the full channel matrix which may be impossible to accurately estimate in practice. In [6], Schniter proposed a turbo equalizer with sliding window with a small computational complexity and good performance. Though it is a promising approach, it is a real challenge for implementation in chipsets. Indeed this requires to loopback the Viterbi decisions in the equalization block which is problematic in Wimax since it would involve huge amounts of memory and latencies due to the coding structure. In this paper we propose a low complexity equalizer based on a DFE adapted to a sliding window receiver. Although the proposed equalizer performs similarly to the ones previously mentioned, its computational complexity is only linear in the number of OFDM subcarriers making its implementation in a Wimax receiver possible. The paper is organized as follows: Section II presents a model of OFDM in time-variant channels; Section III describes the simulation setup and shows the limitations of the conventional Wimax equalizer. Section IV focuses on linear sliding window equalizers; Section V presents our DFE sliding window equalizer while Section VI draws conclusions. II. SYSTEM MODEL In an OFDM system, an N-size QAM coded input vector s=[s0 s1 sN-1]T is first modulated by the Inverse Fast Fourier Transform (IFFT) before insertion of the cyclic prefix. The components of the resulting vector are then sent sequentially through the channel. The channel is modeled by its time varying equivalent discrete model hl,n, where hl,n stands for the lth channel tap at time n and where the number of taps is assumed to be L. At the receiver, the CP is removed before computing the FFT and it can be shown [3] that the received signal in the frequency domain is given by:
ri = 1 N s i H i ,n +
n=0 N 1

I.

INTRODUCTION

There are currently two versions of the Wimax standard: while 802.16d [1] has been designed for fixed broadband wireless access based on an OFDM waveform, 802.16e [2] has been designed for mobile broadband wireless access with explicit mobility support based on an OFDMA waveform. OFDMA is an extension of OFDM in which the user allocation can be multiplexed both in the time and frequency domains. As a consequence, as in OFDM systems, in static channels, the conventional single-tap equalizer can be used to recover the data, thanks to the cyclic prefix (CP) insertion. However, in highly-mobile environments, the Doppler spread destroys the orthogonally of the subcarriers, yielding intercarrier interference (ICI) due to power leakage among OFDMA subcarriers. Therefore, mobility can be severely limited if the conventional OFDM receiver is used and we focus in this paper on designing equalizers supporting high mobility for a Wimax system. Several papers propose efficient techniques to mitigate the BER performance degradation due to the ICI. In [3], some linear equalizers which perform a direct inversion of the entire channel matrix are presented. Though they provide performance very close to the matched filter bound (MFB) [5], they have a complexity that depends on the cube of the FFT size, which is prohibitive for a practical implementation. Furthermore, they may not be suited to OFDMA systems in which only some subparts of the received FFT signal can be exploited when several users are multiplexed in frequency (possibly with different beamforming schemes). In order to reduce this complexity, several techniques have been proposed. For instance, the band structure of the channel matrix is exploited in [4] to reduce the complexity using an LDLH
The authors would like to acknowledge the support of the European Commission through the FP7 project WiMAGIC (see www.wimagic.eu)

1 N

s k H k ,n e
k =0 k i n=0

N 1

N 1

j 2n ( k i ) N

+ wi

(1)

where Hi,n is the Fourier transform of the channel impulse response hl,n for 0 l<L at the time n that is given by:

978-1-4244-2517-4/09/$20.00 2009 IEEE

H i,n =

1 N

hl ,n e
l =0

L 1

j 2il N

0i< N

(2)

and where wi is a complex additive white Gaussian noise (AWGN) with power spectral density N0. Let us define:

Gi ,k =

1 N

H k ,n e
n =0

N 1

j 2n ( k i ) N

i, k [0, N 1]

(3)

where Gi,k denotes the entry (i,k) of the NxN frequency domain channel matrix G; the received signal vector r=[r0 r1 rN-1]T becomes:

equalization techniques because some of the received subcarriers may not be exploited. This happens for instance if not all the subcarriers of the OFDM symbol are used or if bursts are multiplexed in frequency using different beamforming patterns. This also makes very problematic techniques like turbo equalization in which the Viterbi decoder output shall be fed back to the equalization structure. Indeed since a burst can span over the entire frame, it may be necessary to have entirely received and store the complete frame before being able to decode and feedback the soft information to a turbo equalizer. The scalability notion is illustrated in Table I. As clearly explained in details in [8], the idea is to adapt the FFT size to the channel bandwidth so that the system performs identically regardless of whether the system is deployed using 2.5, 5, 10 or 20MHz channels. In the rest of this study, for speeding up simulations, we have used an FFT size of 256 assuming a bandwidth of 2.5MHz though this FFT size exists only for 802.16d. Thanks to the scalability property, our results can nevertheless be extended to other FFT sizes since the system exhibits by design the same robustness to Doppler variations regardless of the actual channel bandwidth. The channel model used for the simulations is the ITU vehicular channel model A that has been defined for conformance testing of Wimax equipments [7]. Its coefficients are given in Table II assuming a classical Doppler spectrum for each tap. In this paper, the channel has been assumed to be perfectly estimated since we were focusing primarily on the equalization structure. Therefore, the coefficients hl,n or equivalently the matrix G are perfectly known.
TABLE I. Parameters
System bandwidth (MHz) FFT size (N) Number of guard subcarriers,

r = G s + w .

(4)

In the case of a static channel, the matrix G is diagonal and the low complexity one-tap conventional equalizer can be used to detect the symbols:

si =

ri Gi*,i Gi ,i
2

(5)

In the case of time-varying channels, the NxN matrix G is no longer diagonal and the non-diagonal terms correspond to the ICI introduced by the channel variations during the OFDMA symbol. Actually it is well known that the ICI on a subcarrier is mainly due to the closest neighboring subcarriers [5] and therefore the G matrix is strongly structured: it is a banded matrix in which most of the channel energy is located on the main diagonal and on a few diagonals distributed around it. Despite this matrix is essentially sparse, the performance is considerably degraded when the conventional OFDM receiver is used as shown in next section. III. THE WIMAX STANDARD In this document, the study has been conducted in the context of the Wimax system. A very good overview of this system can be found in [8] despite the fact that the standard has slightly changed since that date. Actually the most important points for our study are the OFDMA structure and the scalability notion. The OFDMA structure is illustrated in Fig. 1. The main difference with classical OFDM systems is that the users are no longer multiplexed only in the time domain but both in the time and frequency domains. This may prevent some

OFDMA SCALABILITY PARAMETERS Values 2.5 256 23 22 5 512 46 45 10 1024 92 91 20 2048 184 183

Left
Number of guard subcarrierst,

Right
Subcarrierr frequency spacing OFDMA symbol duration

11.16071429 kHz 100.8 s

TABLE II.
Tap

ITU VEHICULAR A MULTIPATH MODEL[7] ITU Vehicular A

Average delay (ns) Average power (dB)

1 2 3 4 5 6

0 310 710 1090 1730 2510

0.0 -1.0 -9.0 -10.0 -15.0 -20.0

Figure 1. OFDMA frame structure [8]

Finally, all simulations have been conducted assuming a receiver with 2 antennas since all Wimax receivers are embedding nowadays at least 2 receive chains. Note that though this is not covered in this paper, the use of 2Rx antenna may strongly affect the performance of the receiver and that some conclusions drawn in this paper may not apply to SISO receivers. Also simulations have been conducted using uncoded bit error rates and it shall be noted that the area of interest for uncoded BER is between 10e-1 and 10e-3. In Fig. 2, the performance of the conventional single-tap equalizer in Wimax and the mobility limitations of basic Wimax receivers are illustrated in the case of the 64QAM constellation. It can be observed that even at small speeds (v=60 km/h), the performance is strongly affected when ICI is not taken into account: though optimal in static channels, the single-tap equalizer presents irreducible error floors at high speeds. IV. SLIDING WINDOW EQUALIZERS As mentioned in Section I, although the linear equalizers based on a direct inversion of the entire channel matrix perform well [3], they have a very large complexity (O(N3)). A possible approach to reduce the complexity is to take advantage of the banded structure of the channel matrix G using sliding window equalizers. The idea is to estimate a symbol transmitted on a subcarrier by considering only a small slice of the entire channel matrix in which is concentrated almost all the energy corresponding to the symbol of interest while the rest is simply neglected. This approach is then repeated by sliding through all the subcarriers of the received OFDM symbol. That way the dimension of the system to be resolved and subsequently the complexity are significantly reduced without severe performance reduction. For instance, supposing that we want to detect the symbol sk. Since most of the energy carried by symbol sk is located on subcarrier k and its neighbours, we can consider the vector rk:=[rk-D rk+D]T where K=2D+1 is the size of the sliding window. This vector is given by:

N where s:=[s0 sN-1]T and where the size KxN matrix G k is defined as:

Gk D , 0 G k D +1, 0 N G k := Gk + D , 0

Gk D ,1 Gk D , N 1 Gk D +1,1 Gk D +1, N 1 Gk + D ,1 Gk + D , N 1

This approach corresponds to the sliding window equalizer proposed in [5]. Actually it is possible to further reduce the complexity by fully exploiting the banded structure. For this, we can consider that rk is given by: rk = G k s + w k (7)

where sk:=[sk-Q sk+Q]T and where the size KxP matrix Gk is defined as: G k D , k Q G k D +1,k Q G k := Gk + D ,k Q with P=2Q+1. The MMSE estimate of sk is then given by:
H sk = a MMSE,k rk

Gk D ,k Q+1 Gk D ,k +Q Gk D +1,k Q +1 Gk D +1,k +Q Gk + D ,k Q +1 Gk + D ,k +Q

(8)

where aMMSE,k = ( GkGkH + N0Ik )-1gk and where gk stands for the central column of the matrix Gk . That way the complexity corresponding to the inversion of a size N matrix is turned to N inversions of KxP matrices. In Figure 3 we can observe that by reducing the width of the sliding window (i.e P) from N to K+2, where K<<N, the complexity is greatly reduced without major performance penalty compared to [5]. Furthermore, by reducing the size of the window, we also significantly reduce the number of elements of the channel matrix to be estimated which should be favorable to reduce the channel estimation complexity and improve the channel

rk = G s + w k

N k

(6)

Figure 2. BER performance of the conventional OFDM receiver for 64QAM at different mobile velocities

Figure 3. BER performance for sliding window KxP MMSE receiver for K=3, v=360km/h and different values of P

estimation performance. The parameter K has to be chosen as a tradeoff between performance and complexity. It also has to be adapted to the mobile velocity. While we can reduce K to reduce complexity, we can observe that the performances degrade when K is very small. Fig. 4 shows a comparison between a KxP sliding window receiver with P=K+2 for two different mobile velocities: v=240km/h and 360km/h. We can see that when increasing the speed, K has to increase as well in order to obtain the same robustness to mobility: for instance at v=360km/h and for K=7, the performance is the same as for K=3 at only v=240km/h. V. DECISION FEEDBACK EQUALIZER A decision feedback equalizer (DFE) is a nonlinear equalizer that uses previous symbol decisions improve the detection and several DFE structures for OFDM systems with fast varying channels have already been proposed in the literature. In [3] a block MMSE equalizer with successive detection is proposed but its computational complexity is larger than O(N3) which is prohibitive for practical implementation. In [5] a decision feedback ICI cancellation for sliding window of size KxN receiver is proposed with a complexity of O(KN2). In this section, we propose a new DFE structure suited to the sliding window equalizer KxP previously described which is a good trade off between complexity and performance. The principle is to find the symbol with the largest energy (i.e., find the column of the matrix G with the largest norm). Supposing this symbol is sm, it can be detected by using the KxP sliding window MMSE equalizer presented in Section IV, aMMSE,m as: H sm = a MMSE,m rm . (9) After taking a hard decision mHD based on m the received vector r becomes:

detecting the symbols successively in a forward order starting from the beginning: s0,s1,,sm-1 as in [5], we continue in backward order: sm-1,sm-2,,s0 to prevent error propagation. An important advantage of this algorithm is that after a few steps, the size of the sliding window is reduced from KxP to Kx(P+1)/2. Therefore the major computation involved in the detection of each symbol that consists in the calculation of the estimator aMMSE,m. for the linear sliding window equalizer described in previous section is strongly reduced by this DFE approach. In Fig. 5 we compare the BER performance of the DFE presented with the performances of the pure KxP sliding window equalizer and with the performances of the MMSE sliding window KxN with decision feedback proposed in [5]. Even if the equalizer proposed in [5] slightly outperforms the proposed equalizer, its complexity is quadratic instead of linear in N. In addition, it relies on the knowledge of all the elements of the channel matrix and hence the conclusion could be different with channel estimation errors. The worst performances are obtained by using the pure sliding window equalizer though its complexity is almost that of the proposed DFE. Fig. 6 and Fig. 7 compare the BER performance of the DFE presented in Section IV for different values of K with the performances of the conventional one-tap equalizer and with the Matched Filter Bound [5]. At high speeds (v=360km/h), the performances of the DFE are degrading if K is very small, and a compromise performance-complexity has to be made. VI. CONCLUSIONS AND FUTURE WORK In this paper we have proposed a decision feedback equalizer (DFE) based on a sliding window receiver for OFDM systems in time-varying channels. Compared to the other equalizers already proposed in the literature, it offers a very good complexity-performance tradeoff making its implementation doable in a real system. Furthermore, it does not require the knowledge of the full channel matrix but only of the channel coefficients around the main diagonal making the channel estimation simpler. Future research directions could be to study the impact of the channel estimation on the performance of that class of receivers.

HD rnew = rold s m m where m is the column m of the matrix G.

(10)

We then replace the column m of the channel matrix G with a zero vector [3]. We continue in the same manner successively detecting the symbols sm+1,sm+2,,sN-1. When arriving at the last symbol (i.e sN-1), instead of continuing

Figure 4. BER performance for sliding window KxP MMSE receiver for v=360km/h, v=240km/h and different values of K

Figure 5. BER performance for different sliding window KxP MMSE receivers for K=3, v=360km/h and different values of P

Figure 6. BER comparison for different equalizers for v=120km/h

Figure 7. BER comparison for different equalizers for v=360km/h

REFERENCES
[1] IEEE Std 802.16d-2004, IEEE Standard for Local and Metropolitan Area Networks Part 16: Air Interface for Fixed and Mobile Broadband Wireless Access Systems. IEEE Std 802.16e-2005, IEEE Standard for Local and Metropolitan Area Networks Part 16: Air Interface for Fixed and Mobile Broadband Wireless Access Systems Amendment 2: Physical and Medium Access Control Layers for Combined Fixed and Mobile Operation in Licensed Bands and Corrigendum 1. Yang-Seok Choi, Peter J. Voltz, Frank A. Cassare , On Channel Estimation and Detection for Multicarrier Signals in Fast and Selective Rayleigh Fading Channels , IEEE Trans. Commun. , vol. 49, pp. 13571387, Aug 2001.

[4]

[5]

[2]

[6]

[3]

[7] [8]

Luca Rugini, Paolo Banelli and Geert Leus, Simple Equalization of Time-Varying Channels for OFDM, IEEE Communications Letters, Vol. 9, No. 7, July 2005. Xiaodong Cai and Giorgios B. Giannakis, Bounding Performance and Suppressing Intercarrier Interference in Wireless Mobile OFDM, IEEE Trans. on Communications, vol. 51, No. 12, December 2003. Philip Schniter, Low-Complexity Equalization of OFDM in Doubly Selective Channels, IEEE Trans. on Signal Processing, vol. 52, No. 4, April 2004. Rec. ITU-R M.1225. Hassan Yaghoobi, Scalable OFDMA Physical Layer in IEEE 802.16 WirelessMAN, Intel Technology Journal, vol. 08, Issue 03, Aug 2004