Wireless Communication (EEL 6509 Spring 2013) Project Report; UFID#06078296

BER Analysis of Radio Receiver Equalization Algorithms using BPSK Signaling.

Ankan Roybardhan, Dept. of ECE, University of Florida, Gainesville, FL.
AbstractWireless Communication inherently comes with multipath propagation of the transmitted signal and Doppler Spread due to movement of Mobile user. The depressiveness of the channel brings Intersymbol Interference (ISI) which has been recognized as major drawback to high speed data transmission over wireless channels. This affects the transmitted signal and this calls for advanced adaptive equalization techniques at the receiver end in addition to effective pulse shaping techniques at the transmitter. In this project, in order to mitigate the effects of ISI, two of the famous Adaptive Equalization Algorithms has been discussed namely, Zero Forcing (ZF) and Minimum Mean Square Error Equalization (MMSE), to analyze the BER performance for a BPSK transmitted signal. We finally evaluate and exhibit the advantage of one algorithm over another and several trade-offs in a Rayleigh Fading Channel. Index Terms BER, BPSK, Equalization, Intersymbol Interference, MMSE, Rayleigh Fading, Zero Forcing.

I. INTRODUCTION propagation of the transmitted signal is a common effect to the transmitted signal, because typically the wireless medium is hostile in nature and it changes rapidly with time. So, the transmitted symbols in the signals interfere constructively or destructively at the receiver, and the phase of the received signal is different from what sent. This destructive interference is termed as Fading. Equalization techniques compensates for Intersymbol Interference within the channel. Intersymbol Interference occurs also because the transmitted signal passes through the channel, and any wired or wireless channel is band-limited in frequency domain. So limiting a signal in frequency domain spreads in time domain, reason why successive symbols interfere with each other. ISI which occurs due to non-linear frequency response of the channel causes symbols to interfere with the successive symbols in the data stream. Due to this intermixing of the successive symbols in the data stream, the receiver decision device cannot efficiently perform slicing or thresholding activity. Therefore it stands imperative from the receiver design point of view, to implement good ISI mitigation mechanism in the radio receivers. One of the ways to instrument this is to implement Adaptive Equalization Algorithm in Equalizers [5][9]. The Equalization Algorithms are termed as Adaptive since the channel is always changing, so the equalizer configuration should be fast changing and adapting in conjunction to the channel. The channel impulse response is obtained by performing

ULTIPATH

several channel sounding techniques. This channel can be modeled as a Finite Impulse Response (FIR) Filter whose coefficients are constantly modified based on channel variations. In essence, an equalizer is another FIR filter and to get an ISI free response, when convolved (in time domain) with the channel impulse response, it should output similar characteristics of an Impulse Function (Direc Delta Function) to give a flat composite received frequency response and linear phase response [5]. In [2], it has been found out that automatic equalization technique which uses Training sequences in the test pulse is being sent prior to actual data transmission, and the filter is set to pass the transmitted signal through this channel. Primary shortfalls of this technique is, the filter has to be modified on each training pulses only and secondly, long training pulses were required for the filter parameters to converge to an optimal value. Also this estimation process from the fast changing channel could behave differently when actual data transmission is initiated rather than sending separate training pulses. A novel way proposed in [1] was to implement a constantly monitoring Adaptive Equalization/Equalizer which could adapt to the changes in the channel and readjust fast to instrument optimal equalizer Impulse Response. Adaptive Equalizations has been developed for Single Carrier High Speed transmission over radio channels, also applies to VHF Bands [4]. Two of the famous Adaptive Equalization Algorithms is studied here viz. Zero Forcing (ZF) and Minimum Mean Square Error (MMSE) which can be implemented into any Linear or Non-Linear Equalizers. A BPSK (Binary Phase Shift Keying) signal is passed through the channel with channel impulse response . There are two basic assumptions that we make here to make the flow of work easy- BPSK and QPSK fetch the same bit error rate, so the BER performance will be the same, except for the fact that hardware implementation is tougher for QPSK and secondly, we consider that the Channel State Information (CSI) is known to the transmitter, which typically might not be the case. (CSI is basically a gain matrix, whose elements are basically gains for each multipath). In that situation, we can only estimate the CSI with some sophisticated statistical estimation process. These are inherently one of the most famous and simple algorithms which are implemented in the receivers. The former follows the Nyquist Criterion where at every sampling instances, the all other symbols/pulses except one being transmitted/sampled is driven forcefully to Zero (hence the name). In latter case, involves in deriving a Cost Function (CF) i.e. the Expected value of the error function, to

Wireless Communication (EEL 6509 Spring 2013) Project Report; UFID#06078296

be minimum by differentiating the CF with respect to Filter Weight Vector , and setting to zero. MMSE technically is used for compensating for Doppler Spread in channel and Maximal Ratio Diversity Combining (a receiver diversity concept to mitigate or reduce depth and duration of fading experiences in Rayleigh Fading Channels) in addition to removing ISI [3]. The paper is composed in the following way; Section II consists of an overall view of mathematics behing Equalization and a brief classification is presented in Section III. In Section IV, the adaptive algorithms are discussed and mathematical modeling and interpretations has been presented. There are a few definitions related to modeling the algorithms, so the definitions are shown using matrix notations. These notations are used for simulation as well. Basic Simulation and analysis of the result has been shown in Section V and a conclusion is drawn with the scope of future work in Section VI. II. EQUALIZATION MECHANISM OVERVIEW - MATHEMATICAL MODELING.

domain) with the received signal y(t). The output of the equalizer is We segregate to T(t), which is nothing but the composite impulse response convolved with that of the equalizer. Now, as mentioned in Section I, the Equalizer Impulse Response can be represented as FIR filter with some filter co-efficients called the weights of the filter. So, the impulse response can be written as some weight wn times the impulse response sampled at every integer multiple (n) of T seconds symbol duration For simplicity in calculation, if we consider noise to be ineffective here (nb(t)=0), then the function T(t) should be simply an impulse function or simply Unity in frequency domain

Or

Fig.1. Communication System with Equalizer at receiver

which means, the Equalizer is nothing but Inverted Channel FIR filter. So our main motive is to measure the channel impulse response during ongoing data transmission and modify the Equalizer Impulse response based on changing the filter coefficients obtained from (iii). By doing this i.e. substituting delta function in eq. (ii) considering nb(t) =0, we obtain , which is output of equalizer equals the transmitted signal. An important observation is if the channel is Frequency Selective, the equalizer enhances the frequency components with small amplitudes and attenuates the strong amplitudes [5]. This idea has been presented in subsequent sections and is potentially a problem in Zero Forcing equalizations which is discussed in Section V in details. This is where MMSE implementation stands better in theory. III. TYPES OF EQUALIZERS The Equalizers, in general, are classified based of its use of the output of the equalizer to control the response of the equalizer. Also, the power budget, Radio Propagation Characteristics and Cost Computational Mechanisms are important for choosing a specific structure and algorithm. The equalizers are classified broadly into two main types [4][5]1. 2. Linear Equalizers If output d(t) as in Fig.1. is used in feedback path to modify the equalizer weight. Non Linear Equalizers If output is not used in not fed back to adapt the subsequent output of the equalizer.

A simple communication system has been shown in Fig.1.[5]. The original signal x(t) is fed into the channel (shown in dotted box). f(t) is the composite impulse response of the Baseband Modulator, Radio Channel (with a real/complex channel impulse response emulating Rayleigh Fading Channel), Receiver RF system. Additive White Gaussian Noise (AWGN) nb(t) is fed in due to thermal noise at receiver circuitry. Now if we model it mathematically, the received signal y(t) at the receiver (input to the equalizer) is given by

where is the complex conjugate of the composite impulse response. Now we have the equalizer characterized by the impulse response heq(t) which is convolved (in time

These equalizers can be further subdivided based on its structural implementations [4]-

Wireless Communication (EEL 6509 Spring 2013) Project Report; UFID#06078296

1. 2.

Transversal Lattice

Finally we can implement various equalization algorithms under any of the above types. The algorithm does not depend on the type of structure it is implemented upon. These algorithms are

1. 2. 3.

Zero Forcing (ZF) Minimum Mean Square Error (MMSE) also known as Least Mean Square Estimation (LMS). Recursive LMS.

Transversal Equalizer (LTE) structure. So, there exists a tradeoff again. 3. MisadjustmentSpecial regards to MMSE implementation, here we have to check for the amount of deviation of the evaluated Mean Square Error(MSE) with the MSE which is optimal. 4. Numeric Complexity- Since there is a filter block representation of the equalizers; there are places where the numeric values are rounded off. So these rounding off introduce what is called Round Off Noise source at every such location, which makes the evaluation of Filter transfer function difficult. For the purpose of simulation and analysis, we take two of the most important Adaptive Algorithms, Zero Forcing and Minimum Mean Square Estimation. A) Zero Forcing Algorithm The idea of ZF algorithm is based on Nyquist criteria. From Fig.3. we observe that in time domain, a peak occurs at every NT seconds. The equalizer co-efficients are chosen such that the composite impulse response and the equalizer impulse response is forcefully driven to zero at these all sampling points except at NT samples. As mentioned in previous sections | | The Channel can be modeled as below

We focus our discussion our work in first two algorithms only-ZF and MMSE.

Fig.2. Types of Equalizers

DFE-Decision Feedback Equalization MLSD-Maximum Likelihood Symbol Detection MLSE-Maximum Likelihood Signal Estimation

To received x(t) at receiver, we can detect X(f) at receiver and then take Inverse Fourier Transform. As mentioned in Section II, we use Matrix Notation for - y(t):Y; x(t):X; Hch(t):H; nb(t):N(only one element in noise matrix) For Real Channel Matrix H :

IV. ALGORITHMS FOR ADAPTIVE EQUALIZATION Various Equalization Algorithms has been named in previous section, but there exists some tradeoff to pick any specific algorithm based on requirement and implementation ease. Following are few of the important aspects which requires attention while choosing any adaptive equalization algorithm [5]1. Convergence Rate- The filter weights are evaluated from equation (iii) iteratively. So, faster the convergence of this computation, the equalizer can fast adapt to the hostile channel by quick evaluation of optimal filter weights. 2. Complexity in Computation The simple or complex equalizer structures are implemented either by using Transversal or Lattice Structures, where delays are provided by putting 1 symbol period delay shift registers. Essentially, Lattice structures as mentioned in Section III, have faster convergence, but it brings complexity in implementation as compared to Linear

For Complex Channel Matrix H: [ ]

Fig.3. Nyquist Criteria Sinc Function

If , then the term in equation (ix), which means Zero Forcing amplifies noise in the receiver. This is the primary issue in ZF algorithm where in some frequencies, the

Wireless Communication (EEL 6509 Spring 2013) Project Report; UFID#06078296

small amplitude (almost negligible and close to Noise Floor) component are amplified to a major extent than large components. Another problem in this implementation is, even if the channel impulse response is a finite length FIR Filter, for complete removal of ISI, the equalization impulse response should be infinitely long [6]. Only application is, this algorithm performs best in High SNR conditions such as local hard wired telephone networks. Another approach to ZF is to consider the following way ( is the FIR filter in z-domain)-

The weight vector, whose elements are the co-efficients of the FIR filter Equalizer, is as below[ ]

From (x) & (xii) the output of the equalizer can be shown to be (Note: All the equations has been written in Matrix Notation) The Cost Function is the basically the Expected Value of the Error Function squared, which is difference between the desired signal/symbol (exactly the same symbol as transmitted which is x(t) or symbol which resembles similar properties as the transmitted symbols) and one received

It is pretty clear if H(z) tends to very low value, the FIR response approaches infinite amplification of noise, which is not desired. A) Minimum Mean Square Error (MMSE) Algorithm Before the MMSE architecture is explained, some of the notations are explained below, will be required in modeling a communication system and evaluating a Cost Function or Error Function-

b) Algorithm to find Optimal Filter Co-efficients

From the Error Function equation (xv), we can the modulus squares of the error function| | The cost function is therefore given by taking the Expected Value of the eq.(xvi)[ ] [| | ] [ [| | ] ] [ ] [ ]

Now, we define two Matrices viz. p which is the cross correlation matrix. This is formed by taking the Expected Value of the Matrix multiplicative terms of (input vector and the received symbol vector indexed to time k)Fig.4. Equalizer Block Diagram k=Time Index meaning weights must change with time based on Channel.

a) Evaluation of Error Function

The received symbol/signal can be written as below in matrix notation [ ]

k=Time Index for which the received signal changes and the weights are updated based on this.

This term can be substituted in the last part of the eq. (xvi). We define another matrix RNN which is called as Input Correlation Matrix. Input is to the Equalizer again index to time k

] [( )]

The above figure models the equalizer by depicting tapped delay lines with adjustable gains at each tap. Output of the equalizer is given by

This matrix is basically taking each received symbol at the equalizer and correlated to every other symbol. Due to fading and ISI, each received should have some correlation with other . If not every term other than the diagonals should converge to zero. In in terms of random processes, the major diagonal terms are the mean square values of each input samples and all other cross terms are the Autocorrelation

Wireless Communication (EEL 6509 Spring 2013) Project Report; UFID#06078296

terms which is because of multipath delayed response of the channel [5]. If are Wide Sense Stationary, then the elements RNN and p are second order statistics which are stationary in time (Mean Function is also Zero) [5]. The Mean Square Error is [ ]

Clearly it converges to ZF. ii) For Low SNR Conditions || ||

For MMSE Equalization, the Cost Function is differentiated with respect to weights and setting equal to zero

|| || This is the set of optimum filter co-efficients because the error is now minimized and is made orthogonal to xk. Putting (xx) into eq. (xix) [ ]

It shows the same performance as Matched Filter. A very important concept to consider is, MMSE should be implemented in FIR filter form in contrast to IIR filters because of the fact that IIR Filters have poles, which can cause instability at those points, whereas FIR filters do not have poles except at z=0, hence easy to design. But the FIR filter has to be truncated at some point because it is not possible practically to implement infinite length FIR filter. So we can truncate it from L1 to L2.

Another approach to MMSE Linear Equalizer design is to minimize the noise variance in the following way-

Fig.5. Signal in AWGN and through Channel

We need to have | | or
| |

V. SIMULATION AND ANALYSIS OF EQUALIZATION ALGORITHMS In theory ZF algorithm is comparatively in-efficient due to the fact that it introduces non-uniform amplifications and unnecessarily it amplifies the Signal Noise which is certainly not permitted. MMSE, on the contrary, is preferred over ZF, since it uses a better approach by statistically minimizing the Expected value of the error between the sent signal and the output of the equalization. Practically, Mean Square Error is minimized by Stochastic Gradient Algorithm also called Least Mean Square (LMS) Algorithm [5]. The simulation of these equalizations is done using Matlab [7] and BER performance is compared with the theory established in the literature. For keeping simplicity in the simulation part, the prefiltering for pulse shaping is neglected. The simulation is divided into several parts and subparts, where first the simulation is explained based on ZF algorithm and the MMSE. Further subdivisions are based on several steps in the algorithms from generation of BPSK Signals to evaluation of Eb/N0 for specific number of taps (filter co-efficients) in the equalizer structure. A QPSK Signaling could also been used since both QPSK and BPSK gives the same error probability as one set of symbols are completely orthogonal to the other set. So from hardware perspective of analysis, BPSK signal is preferred. For the purpose of simulation, there are certain

|| ||

|| ||

The MMSE equalizer converges to two types of receiver system in following casesi) For High SNR Conditions || ||

|| ||

|| ||

Wireless Communication (EEL 6509 Spring 2013) Project Report; UFID#06078296

areas to be discussed. Followings are some of few assumptions/considerations taken while composing the simulation code1. The Channel Matrix H is known at receiver (see eq.viii) In contrast, most of the time due to fast adaptive mechanisms, the Channel State Information (CSI) can be known to receiver but it tough to feed back to transmitter. Hence pre-coding at transmitter is difficult, typical problem in MIMO Systems. No pulse shaping filter is considered at the transmitter, to remove ISI. This is a logical consideration since, we study about the sole performance of the receiver equalizer. Noise Variance is , since the thermal noise can be modeled as Gaussian Random Distribution which is completely determined by its mean (which is zero) and variance. Bit Error Rate for BPSK Signal is given by ( )

We pass it through the channel with impulse response hch(t) and finally at receiver, we divide the received signal with the exact channel impulse response function. This is analogous to the equalization (Zero forcing). Based on the assumptions made earlier, we consider that the CSI is successfully relayed to the transmitted.

2.

3.

4.

The above expression is for AWGN channel. For Fading Channel| |

Fig.7a. BER performance of BPSK in AWGN + Rayleigh Fading

( Or, [| | ]

The theoretical BER expression given in eq can be modified to the following [ ] This above equation has been used to plot Rayleigh Fading performance (in Theory). The derivation requires the substitution of |h|2chi-square Distribution and then taking the conditional probability over the eq (xxviii) AWGN is plotted with the eq. (xxviii) . It is clear that BER performance drops drastically in fading environment than in AWGN channel for example, for a given BER 10-2, Rayleigh Fading channel would require approx. 10dB higher SNR. The theoretical and simulation results in Rayleigh fading shows to be overlapping at points till the SNR requirement gets higher where simulation with hard decision shows to have a little bit better BER for a given SNR, but the fluctuation is very small.
B) Zero Forcing Equalization We repeat the same process as in section A, till we get the received signal at the equalizer input. The idea here is, when we pass the signal through the channel, we consider we know the CSI, i.e. the Channel Matrix. So, number of elements in the Channel Matrix is the number of taps or the number of excess delay bins occurring due to channel impairment. Therefore, the equalizer is modeled as an FIR filter, should also have the same number of filter weights or taps as that in

The Simulation section is divided into following waysA) BPSK Modulated Signal in Fading In Binary Phase Shift Keying (BPSK) the phase of the constant amplitude modulated signal is switched in between two values. There is radian phase shift between two symbols. The Signal Constellation looks like below-

Fig.6. BPSK Signal Constellation

The general form of BPSK Signal is 01

QPSK Signal gives the same Bit Error Performance as that of BPSK, but differs in the number of bits per symbols being pumped and the hardware implementation is complex as compared to BPSK. In the simulation we generate 105 bits stream of 1s and 0s with equal probabilities. After modulation, the stream is added with White Gaussian Noise with mean zero and variance .

Wireless Communication (EEL 6509 Spring 2013) Project Report; UFID#06078296

7 ]

the channel. We simulate the same using various number of filter co-efficients.

RNN=Input (input to equalizer) co-variance matrix= [ [ ]

Which is in line with eq. (xx). The signal is uncorrelated with the noise at the receiver (E[sk.nk]=0) and the symbols are not correlated at several time instances, which means, the variance of s(m)=1 but the E[sk.sl]=0.

Fig.7b. BER performance of BPSK in AWGN + Rayleigh Fading with ZF Equalization at receiver

The basic idea to find the optimal weight is based on the eq. (iii) where in simulation; the matrix convolution has to be carried out using the function toeplitz. Several observations are; Firstly, the concept of diversity is analogous to the increasing the number of taps of the receiver equalizer. As we go on increasing the number of taps, for a given BER, SNR improvement reduces and this phenomenon is termed as Law of Diminishing Returns. It is not worth keep on increasing the number of taps because we do not get adequate SNR improvement with the same BER, after a specific number of taps. From Fig.7b, it is observe that as the taps approaches 7 and above, the performance improvement saturates, thus implying it is not advisable to increase the number of taps further. Secondly, ZF amplifies the noise as shown in eq. (ix), so again it proves, the performance will never match to that of AWGN, unless the degree of diversity tends to . C) Comparison of MMSE and ZF Equalization Let the signal sent is s(k) (the kth symbol of the stream). The received signal at the equalizer input is given by-

Fig.7c. BER performance of BPSK in AWGN + Rayleigh Fading with ZF Equalization at receiver

Based on eq. (xxix), (xv) and (xvi) the MMSE works in minimizing the Error Function as follows (matrix notation)[ [ [ [ ] ] [ ] ] ]

The simulation has been carried out for only 3 filter coefficients in the equalizer. We observe that there is significant improved performance shown by MMSE equalizer than ZF. For a given BER say 10-2, the Eb/N0 is around 9dB for MMSE and 9.6 dB for ZF, which means MMSE stands 0.6dB better for a given SNR which shows MMSE performs better. We have a close look by comparing the fig.7b or 7c with 7a, for an example BER say 10-3, the SNR requirement for a fading channel was 23dB whereas by implementing these algorithms, the SNR requirements has been cut down to mere 11dB thereby direct 12dB better improvement is shown. There are several other observations from the plots. Firstly, the abscissa of the plot starts from -2dB. This has been done necessarily to include the Shannons Capacity which is 1.69dB. Refer to Fig.7a. a marker has been placed showing for , BER is 0.1306. Following is to proof that the plot is in line with the theory of Shannons theoremIf

The matrices w, p and RNN has been introduced earlier asw=Weight Vector p=Cross Correlation matrix between signal transmitted and received at equalizers input (received signal)= [ ]

This is close to 0.1306 as traced on graph. Hence our simulation is correct. Another point to be noted is draw a trade-off between speed of convergence of the algorithm to the self-noise algorithm.

Wireless Communication (EEL 6509 Spring 2013) Project Report; UFID#06078296

The filter weights are typically adapted by sending Training Pulses (typically <10% of total data bits)[Class notes] After every step of training, the filter weights has to be increases or decresed n steps and the iteration occurs untill optimal weights are not achieved. Now this step size is large, convergence is faster, but this large step size would induce more noise and hence less accuracy. So the step size has to be adjusted to an optimum value. The equalizer weights are adjusted in the following wayNew Weights=Previous weights+(constant)x(Previous Error)x(Current Input Vector)

[6]http://en.wikipedia.org/wiki/Zero_Forcing_Equalizer Zero Forcing Equalizer. [7] Prez Fontn, F., Modeling the wireless propagation channel : a simulation approach with Matlab. Chichester, West Sussex, England ; Hoboken, NJ, USA : Wiley, 2008. [8] Hongsheng Gao, Peter J. Smith, Martin V. Clark, Theoretical Reliability of MMSE Linear Diversity Combining in Rayleigh-Fading Additive Interference Channels, IEEE Transactions on Communications, vol. 46, no. 5, pp., May 1998. [9] Lecture Notes Spring 2013, Wireless Communications Dr. Dapeng Oliver Wu.

VI. CONCLUSION AND SCOPE OF FUTURE WORK Mobile Channels are random and time varying. Equalizers must track the time varying characteristics of this mobile channel. One basic assumption taken for this work is, the Channel State Information is known to the transmitter. This is perhaps one of the toughest tasks in the domain of Wireless Communications since the channel is noisy and fast varying. For the scope of future work, the concepts can be borrowed from MIMO, where transmitter pre-coding can be performed by only estimating the channel. One of the most important points to consider when designing a Linear Filter is that the ISI is only minimized and not forced to complete zero. There are cases when, in case of deep fades, even after equalization process, SNR falls by 1720 dB whereas, ZF drives the SNR to dB to make a complete break away point. Key factor is for having the most optimal performance amongst all Equalizing mechanism, NonLinear Decision Feedback Equalization takes care of such channel and it drives ISI to zero. Any device that can improve upon the MMSE-Linear Equalization has to be non-linear, due to optimality of the MMSE-LE over the class of linear Filters.

VII. ACKNOWLEDGEMENT This project is in requirement towards completion of the course, EEL6509 Wireless Communication. I thank Dr. Dapeng Oliver Wu and the Mr. Pratik Prabhanjan Bramha (TA) for guiding with every little details about this work and providing valuable inputs throughout the time working on this project.

VIII. REFERENCES
[1] Crozier, S.; Falconer, D.; Mahmoud, S., "Short-block equalization techniques employing channel estimation for fading time-dispersive channels," Vehicular Technology Conference, 1989, IEEE 39th , vol., no., pp.142,146 vol.1, 1-3 May 1989. [2] Lucky, R.W., Automatic Equalization for Digital Communication, Bell System Technical Journal. Vol.44,pp.547-588, 1965. [3] Monsen, Peter, "MMSE Equalization of Interference on Fading Diversity Channels," Communications, IEEE Transactions on , vol.32, no.1, pp.5,12, Jan 1984. [4] Proakis, J.G., "Adaptive equalization for TDMA digital mobile radio," Vehicular Technology, IEEE Transactions on , vol.40, no.2, pp.333,341, May 1991. [5] Rappaport, T.S., Wireless Communications Principles and Practice, Prentice Hall, 2002.