1542

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL.

35.

NO.

12, DECEhlBER 1988

On the Design of Optimal Equiripple FIR Digital Filters for Data TransmissionApplications
HENRY SAMUELI. MEMBER, IEEE
Abstract -An improved linear programming algorithm is presented for the design of FIR digital filters. This algorithm avoids the numerical ill-conditioning problems which commonly occur due to the necessity to sample the frequency response on a very dense grid of points for high-order filters. This new technique is applied to the design of equiripple FIR Nyquist filters and equiripple FIR transmit and receive matched filters for data transmission applications. Nyquist filters have the property that their impulse response has uniformly spaced zero-crossings which is a requirement that modem filters must possess in order to satisfy a zero intersymbol interference constraint. Matched filters are typically used in the transmit and receive sections of data transmission modems. They have the property that their cascade is a Nyquist filter and thereby satisfies the zero intersymbol interference constraint. The design algorithm for the matched filters consists of a two-stage process of linear programming to obtain a Nyquist filter with nonnegative frequency response followed by standard spectral factorization techniques to extract the nonlinear phase transmit and receive filters.

However it is well known [4] that in order to minimize the bit-error-rate in the presence of white Gaussian noise in the communication channel the optimum partitioning of the transmit and receive filters is the so called matched filter pair whereby the transmit filter frequency response H,( w ) and the receive filter frequency response HR(w ) satisfy
H A W ) =H,*(o) (2 ) where * denotes complex conjugate. Therefore the data transmission filter design problem consists of deriving a pair of low-pass filters whose frequency responses are complex conjugates of one another and whose cascade is a Nyquist filter. Several techniques have been proposed for the design of digital matched filters and digital Nyquist filters. In [5] Mueller presents a numerical solution to an eigenvalue problem for designing FIR Nyquist digital filters which have maximum spectral energy concentration within the transmission bandwidth. In [6] Chevillat and Ungerboeck extend these results and present nonlinear optimization techniques for designing a pair of FIR digital matched filters which satisfy the zero intersymbol interference constraint of (1) when cascaded and have maximum spectral energy concentration within the transmission bandwidth. For designing FIR data transmission filters with equiripple stopbands, linear programming has been used successfully as demonstrated by Salazar and Lawrence in [7] and Liang, DeFigueiredo, and Lu in [3]. Although linear programming is a very flexible and powerful technique for designing digital filters it does suffer from some numerical ill-conditioning problems for high-order filter design. This is primarily due to the fact that the inequalities which specify the linear program are obtained by evaluating the frequency response equation of the FIR filter on a dense grid of frequency points (the number of grid points is typically chosen as 8 or 16 times the filter length). Therefore, for high-order filters an excessive number of inequality constraints are obtained which can cause numerical difficulties in the pivoting operations of the revised simplex method. A discussion of some of these numerical issues is presented by Steiglitz in [2]. In [7] Salazar and Lawrence present some specific bounds relating the numerical accuracy of the approximation problem to the number of frequency grid points chosen. In this paper we propose a simple modification to the standard linear programming approach to FIR filter design which avoids the necessity for a dense grid of frequency points. As a result, high-order filters can be designed using commonly available FORTRAN subroutine packages for linear programming without the need for any special precautions. To illustrate the results the new algorithm is applied to the design of FIR Nyquist filters and FIR transmit and receive matched filters for data transmission applications. In Section I1 some of the properties of matched filters and Nyquist filters are reviewed and the data transmission design problem is formulated. The modified linear programming algorithm is discussed in Section 11-D. In Section I11 several numerical examples are presented to illustrate the overall design process. It should be pointed out that in a recent paper [E] Saram&i and Neuvo present an iterative technique for the design of equiripple FIR Nyquist filters which avoids the use of linear programming and appears to >bevery well-behaved numerically. The algorithm consists of a frequency-domain optimization using the well-known Parks-McClellan program [9] in conjunction with a time-domain optimization involving the solution of a

I. INTRODUCTION
Linear programming is a very popular technique for linear phase FIR filter design when time-domain constraints are imposed on the filter coefficients [1]-[3]. An important application of filters with time-domain constraints is in the transmit and receive portions of data communication systems. The primary function of the transmit filter is to constrain the transmitted spectrum to accommodate a bandlimited channel. The theoretical minimum bandwidth (for which zero intersymbol interference results) is the Nyquist bandwidth, R / 2 , where R is the transmitted baud rate in symbols per second. The primary function of the receive filter is to reject out-of-band noise and thus maximize the signal-to-noise ratio. The time-domain constraints on these filters result from the requirement for zero intersymbol interference whereby the impulse response of the cascade of the transmit and receive filters must have uniformly spaced zero-crossings. Data transmission filters whose impulse responses have uniformly spaced zero-crossings are known as Nyquist filters. In a digital implementation of a Nyquist filter the sampling rate F, is generally chosen to be an integer multiple of the baud rate R and thus the impulse response coefficients satisfy the constraints

where C is a nonzero constant, no is a positive integer, and M is the oversampling factor, i.e., M = F , / R . Linear phase Nyquist digital filters have the additional constraint of requiring the impulse response coefficients to be symmetric about the point n = no. The zero intersymbol interference constraints of (1) apply to the cascade of the transmit and receive filters and no explicit restrictions are imposed on their individual impulse responses.
Manuscript received July 6, 1987; revised March 26. 1988. This work was supported in part by the University of California MICRO program in cooperation with TRW, Inc. under Grant 86-110. This paper was recommended by Associate Editor R. Ansan. The author is with the Integrated Circuits and Systems Laboratory, Electrical Engineering Department, University of California. Los Angeles, CA 90024. IEEE Log Number 8823897.

0098-4094/88/1200-1542$01.00 01988 IEEE

IEEI, l-KANSAC'rIONS ON CIKCIJI'IS AND SYSTkMS. VOL.

35, NO. 12, DECEMBER 1988

1543

system of linear equations. Thls algorithm requires some additional modifications to be suitable for the design of FIR transmit and receive matched filters due to the requirement for a Nyquist filter design with nonnegative frequency response (see Section 11-A), however it does appear to be a useful alternative to linear programming for designing equiripple Nyquist filters. 11. DESIGN OF DATA TRANSMISSION FILTERS A . Properties of Mutched Filters Consider an N-tap FIR transmit filter with a transfer function
H,( z)
= U,,

filter (with a passband gain normalized to unity) satisfies the zero intersymbol interference constraint of (1)if and only if
M-1

HA,( e J

( ~ + 2 ~ ~ / M =)l .

( 7)

k=O

+ u,z-' + a,z- + . . . + a N -'z-(,-')

. . . +u,~,zN-'.

(3)

In order to satisfy the matched filter requirement of (2) the transfer function of the receive filter must be H~(z) =a,+ulz1+a,z2+ (4)

Thus the frequency response in the band [0,a/M] of a Nyquist filter is uniquely determined by the frequency response in the band [a/M, n]. It should be noted that Nth-band digital filters [14] which are primarily used for decimation and interpolation in multirate digital signal processing applications have identical time-domain constraints as Nyquist filters, and hence they also possess property (7). In [14] h4intzer shows that for linear phase FIR digital filters, property (7) implies that the central coefficient of the impulse response has a value of h, = 1/M. An ideal Nyquist filter has the frequency response

Furthermore, as a consequence of (2), the cascade of H7( z ) and HR(z) has a frequency response which is a purely real and nonnegative function of frequency. To make the receive filter physically realizable, a delay of N - 1 samples is inserted in (4) to obtain H R ( z ) = u A , - , + a N - , z - 1 + a , ,z-'+ . . . + u o z ~ " - ' ) ' (5) Thus the coefficients of the transmit and receive filters are identical except that they are reversed in order. Notice that there is no requirement that either the transmit or receive filter be linear phase, however the cascade of the two filters must have a linear phase characteristic since the product of H7( z) and HR(z) results in a polynomial with symmetric coefficients. One final property to note regarding H,(z) and HR(z) of (3) l ) , transfer funcand ( 5 ) is that since HR(z)= H T ( ~ - l ) ~ - ( N - the tion of the cascade of the two filters, denoted H N ( z ) ,has the form

In a general equiripple FIR filter design specification the passband edge a , , ,the stopband edge u,,, the passband ripple a,, and the stopband ripple 8, are independently specified. However, for a Nyquist filter specification, the critical parameters arc the stopband edge, which determines the bandwidth required for transmission, and the stopband ripple, which determines the out-of-band signal attenuation. The passband edge and the passband ripple only determine the envelope of the transmitted spectrum which is generally not of much importance. In traditional "raised-cosine" Nyquist filter designs [4], an excess bandwidth factor a is specified which results in the symmetric passband and stopband edge frequencies a , , = ( l - a ) ~ / M and ~,=(l+a)a/M
(9)

thus if zk is a zero of HN(z) then so is l/zk. Therefore, the zeros of HN (z) occur in mirror image pairs about the unit circle in the z-plane and all zeros on the unit circle must be of even order. Furthermore, HN (z) must have even degree (odd length impulse response) since it is obtained by multiplying two equal order polynomials. Hence, the procedure for designing the transmit and receive matched filters can be partitioned as follows. First, design an equiripple linear phase FIR Nyquist digital filter, with a nonnegative frequency response, and then use standard spectral factorization techniques to extract the transmit and receive matched filters. This two-stage process of linear phase filter design followed by spectral factorization has been applied many times to a variety of related filter design problems. Herrmann and Schuessler [lo] and Kamp and Wellekens [ll]applied this two-stage approach to the design of minimum phase FIR filters. Smith and Bamwell [12] and Wackersreuther [13] also used a similar twostage technique for the design of FIR filters for perfect reconstruction filter banks.

however, in an equiripple Nyquist digital filter design the passband and stopband edge frequencies are not necessarily symmetric (except for the case where M = 2). In fact, the passband edge frequency and the passband ripple are parameters whch should not be specified at all in an equiripple FIR Nyquist filter design. Equation (7) necessarily implies that the response in the passband [O,w,,] is bounded by unity plus or minus the sum of the peak ripple in each of the M - 1 slices of the stopband [a,, a]. Hence, for an equiripple stopband, the passband ripple 8, is bounded by
8,

< (M-1)8,.

(10)

C. Linear Program Formulution

The frequency response of an N-tap ( N odd) linear phase FIR digital filter is given by
(N-1)/2

H(e'") = h o t

1 2h,,coswn.
n=l

(11)

After incorporating the Nyquist filter time-domain constraints the frequency response simplifies to
( N - 1)/2

HN ( el"')

= 1/M

+
nez
n=l

2h,, cos a n

(12)

B. Properties of Nyquist Filters

As a direct consequence of Nyquist's fundamental theorem for samling a bandlimited waveform, necessary and sufficient conditions on the frequency response of a filter are obtained for which it is possible to transmit data with zero intersymbol interference. These results are summarized in [4] and can be restated in the context of Nyquist digital filters as follows. A Nyquist digital

where Z is the set of indices corresponding to the zero-valued coefficients of the impulse response. The number of indices in the set Z is [ N/2 MI where [XI denotes integer part of x, thus the number of unconstrained filter coefficients is Nu

( N - 1)/2

[ N/2M].

(13)

1544

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. 3 5 , NO.

12, DECEMBER 1988

The classical approach to specifying a linear program for the design of linear phase FIR filters [1]-[3] is to formulate a set of inequalities by evaluating (12) on a dense grid of frequency points in the passband and stopband of the filter. For Nyquist filter design only the stopband response needs to be constrained thus the following linear program is obtained: Minimize 6 subject to

Parks-McClellan FIR filter design program [9] whereby all the extremal frequencies are refined after each iteration. The details of the modified algorithm are briefly described as follows. The maximum number of extremal frequencies in the frequency response of an odd-length N-tap linear phase FIR digital filter is given by N,
=

( N + 1)/2

(15)

n=l nSZ

where Ng is the number of frequency grid points. Notice that the left-hand-side of inequality (14a) does not have 6 added to it and hence the frequency response is constrained to be nonnegative as is required for matched filter design. To design a conventional Nyquist filter with an equiripple stopband characteristic centered about zero, simply add 6 to the left-hand-side of (14a). The linear program in (14) has Nu + 1 unconstrained variables and 2Ng inequality constraints. It can also be written in its dual form which is a linear program with 2Ng nonnegative variables and Nu 1 equality constraints. The dual program is the standard form of a linear program. It can be solved using the revised simplex method which is commonly available as a subroutine package on many mainframe computers. Typically, the number of grid points Ng is chosen to be 8 or 16 times the filter length N and the frequency grid points wk are chosen to be uniformly spaced in the frequency bands of interest. Since the linear program is constrained on a finite set of frequency points, the actual extrema of the resulting frequency response will, in general, fall between the grid points w k . Thus the lower extrema of the stopband frequency response of the Nyquist filter will not exactly equal zero and the corresponding unit circle double zeros of the transfer function " ( z ) will split apart. This will cause a slight mismatch in the resulting transmit and receive filters. This effect can be minimized by selecting a large va!ue for Ng thereby creating a very dense frequency grid, however a large value for Ng can cause numerical instability problems in the simplex algorithm. This problem is exacerbated as the filter length is increased. A simple modification to the linear programming technique is presented in the next section which avoids the necessity for selecting a large value for Ng without sacrificing the accuracy of the final result.

where Ne includes the extremal frequencies at w = 0 and w = T but does not include the passband and stopband edge frequencies U , , and w,.Since the stopband width of an ideal Nyquist filter specified by (8) is M - 1 times the passband width, it is reasonable to expect that an optimal equiripple Nyquist filter will have the property that the number of stopband extrema will be M - 1 times the number of passband extrema and thus the number of stopband extremal frequencies can be approximated by (16) For example, for a 23-tap Nyquist filter with an oversampling factor of M = 4, the maximum number of extremal frequencies is 12 and thus we would expect 3 extremal frequencies in the passband and 9 in the stopband. The number of stopband extremal frequencies specified by (16) takes on integer values when N is restricted to the values N = 2Mk -1, k=1,2,. . .. (17) Extensive numerical examples over a wide range of Nyquist filter specifications have verified that the approximation of (16) (subject to (17)) is indeed satisfied with equality. It is interesting to note that number of stopband extrema predicted by (16) and (17) coincides exactly with the number of unconstrained filter coefficients Nu of (13). The linear program in (14) has Nu + 1 unconstrained variables and based on the previous discussion only N, 1 inequalities are required (including one at w = u s ) ,thus the matrix formulation of the linear program in (14) is exactly square since Nu = N,. For the first iteration, the N, + 1 frequency grid points are chosen to be uniformly spaced between w, and T. The linear program is then solved and the resulting frequency response is evaluated to determine the new stopband extremal frequencies. In order to accurately determine the new extremal frequencies, Newton's method is used to search for the zeros of the derivative of the frequency response. This is accomplished by evaluating the derivative on a course grid of points (typically 2 N points) to approximately locate its zero-crossings. Then several iterations of Newton's root-finding algorithm are used to very accurately determine the zero-crossing frequencies which are then used as the frequency grid points for the next iteration of the linear program. Since the stopband extrema alternate between maxima and minima, the stopband edge frequency os is used as the first grid point in (14b) to obtain a constraint on the maximum stopband response and then the next extremal frequency is used in (14a) to obtain a constraint on the minimum stopband response. Equations (14b) and (14a) are then alternatively evaluated at each of the remaining extremal frequencies to obtain the complete set of constraint equations for the next iteration of the linear program. After each iteration the stopband attenuation of the resulting filter is evaluated and the algorithm terminates when the stopband attenuation does not significantly change (0.01 dB was the selected threshold). Although convergence of this iterative algorithm cannot be guaranteed, extensive numerical examples over a wide range of input parameters have demonstrated this algorithm to be very reliable. N, Ne( M
-

1)/M= ( N

+ 1)( M

1)/2M.

D. Modification to Linear Programming Algorithm

Ideally, the solution to the linear program in (14) could be obtained if only one inequality constraint were specified for each extremal frequency of the stopband response. The remaining inequalities are all redundant. Unfortunately the locations of the stopband extremal frequencies are not known a priorz, however by using an iterative approach to the problem the locations of the extremal frequencies can be successively refined and therefore the number of inequality constraints required in the linear program of (14) can be significantly reduced. Thus the one-step solution to the problem is abandoned in favor of an iterative approach whereby the minimum grid density is used for each iteration of the linear program and after each iteration the frequency grid points ok in (14) are adjusted to coincide with the computed extrema of the stopband response. This is the same basic principle that is used in the Remez exchange algorithm of the

IEEE TRANSACTIONS O N CIRCUITS AND SYSTEMS, VOL.

35, NO. 12, DECEMBER 1988

OL

154s
I

E. Spectral Factorization Techniques

The second step in the design of the equiripple FIR transmit and receive matched filters is the spectral factorization of the Nyquist filter polynomial HN( z ) designed by the techniques presented in Section 11-D. There are several well-known approaches to this problem. The first approach uses brute-force polynomial root-finding techniques to determine all of the roots of H N ( z ) . Then H,(r) is formed from half of the unit circle double-zeros and half of the remaining reciprocal pairs of zeros. Considerable flexibility is available in choosing which individual roots out of each reciprocal pair of roots is chosen to form H , ( z ) . If only roots inside the unit circle are chosen then H7( z ) will be minimum phase. If only roots outside the unit circle are chosen then H 7 ( z ) will be maximum phase. Approximately linear phase can be obtained by alternating roots from inside and outside the unit circle [12]. Once H 7 ( z ) has been formed then HR( z ) is formed out of the remaining roots. For low-order polynomials commonly available subroutine packages for root-finding work quite well. For higher order filters, the burden on the root-finding program can be considerably reduced by taking advantage of the fact that the locations of all the unit circle double zeros of HN ( z ) are known U priori, i.e., they correspond to the stopband zeros of the frequency response. Thus H , v ( z ) can first be deflated by extracting the unit circle zeros and then a standard root finding program can be applied to the remaining low-order polynomial. This technique was used in 1 1 1 1 to design a 320-tap minimum phase FIR filter. For the special case of minimum and maximum phase transmit and receive matched filter designs an alternate approach to spectral factorization can be used which entirely avoids any polynomial root-finding. This technique, known as cepstral deconvolution, is based on the fact that the magnitude response of the matched filters is known (i.e., IH7(eJw)l = I H R ( e J w )= l I H,v ( e J W ) l ' / ' ) ,therefore the phase response of the corresponding minimum phase filter can be determined by Hilbert transform relations, i.e., the log-magnitude and phase responses are Hilbert transforms of each other. The FFT provides a convenient numerical tool for implementing the cepstral deconvolution. The details of this algorithm are presented in [15]. 111. DESIGN EXAMPLES

-fin -50

0.0

t
01

0.2

03

04

05

FREQUENCY (cycles/sornple)

Fig. 1. Amplitude response of 47-tap FIR Nyquist filter with overaainphng factor M = 4 and excess bandwidth factor a=0.23. Expanded wew o f stopband response on a linear scale illustrates the nonnegative characteristic.

03

I 1
-011 '
" "

'

"

"

"

"

"

"

" " "

10 15 SAMPLE NUMBER

20

Fig. 2. Impulse response coefficients of minimum phase 24-tap FIR matched filter with oversampling factor M = 4 and excess bandwidth [actor a = 0.23

I
0.0
0 1

I
0.2
03

a4

05

FREQUENCY (cycle3/sample)

All programs were written in Fortran and compiled using the VS Fortran compiler on an IBM 3090 mainframe computer. Double precision arithmetic (64-bit floating point) was used in all computations. The IMSL Fortran subroutine library was used for linear programming (routine ZX4LP) and polynomial factorization up to order 100 (routine ZPOLR) [16]. Example 1: This example is based on the same specifications as the matched filters designed in [6]. The filter has length N = 24, oversampling factor M = 4, and excess bandwidth factor a = 0.23. Thus a 47-tap FIR Nyquist filter with M = 4, a = 0.23 and a nonnegative frequency response was designed using the techniques discussed in Section 11-D. Convergence (to within .01 dB) was obtained in 6 iterations of the linear programming algorithm and required 0.32 s of CPU time. The amplitude response of the Nyquist filter is shown in Fig. 1. An expanded view of the stopband on a linear scale is also shown in Fig. 1 which clearly illustrates the nonnegative equiripple behavior. The symmetric Nyquist filter polynomial was then factored to obtain a minimum phase transmit matched filter and a maximum phase receive matched filter. The impulse response coefficients of the transmit filter are shown in Fig. 2 and the corresponding

Fig. 3. Magnitude response of 24-tap FIR matched filter with oversampling factor M = 4 and excess bandwidth factor a = 0.23.

magnitude response is shown in Fig. 3. The minimum stopband attenuation is 23.4 dB which is approximately 4 dB greater than the minimum stopband attenuation for the filters designed in [6]. Example 2: In this example a conventional FIR Nyquist filter was designed which does not have the constraints of having a nonnegative frequency response. This example is based on the same specifications as the FIR Nyquist filter designed in [3] and [SI. The filter has length N = 39, oversampling factor M = 4, and excess bandwidth factor a = 0.15. The only modification required for the linear program formulation is to add 6 to the left-hand side of inequality (14a). The amplitude response of the resulting filter is shown in Fig. 4. The minimum stopband attenuation is 34.3 dB and the peak passband ripple is 0.44 dB, hence t h s design is virtually identical to the design obtained in [E] using the modified Parks-McClellan program and it achieves approximately 1.3 dB additional stopband attenuation as compared to the design obtained in [3] using conventional linear programming techniques.

1546
0

IEEE TRANSACTIONS ON CIRCUITSAND SYSTEMS, VOL. 35, NO.

12, DECEMBER 1988

IV.

CONCLUSIONS

n m
Y

-10 -20

r 3

5 9

-30
-40

-50

0 . 0

0.1

0.2

0.3

0.4

0.5

FREQUENCY (cycleo/?lomplr)

Fig. 4. Amplitude response of 39-tap FIR Nyquist filter with oversampling factor M = 4 and excess bandwidth factor a = 0.15. Stopband is not constrained to be nonnegative.

-4C

- <n

In h s paper an iterative linear programming technique has been presented for the design of FIR digital filters. The iterative technique avoids the numerical ill-conditioning problems associated with conventional linear programming techniques which result from the necessity to use a dense frequency grid for high-order filter designs. This new technique is applied to the design of equiripple FIR Nyquist filters and equiripple FIR transmit and receive matched filters for data transmission applications. The use of linear programming insures that the designs are optimal in the sense that they achieve the maximum possible stopband attenuation for a given filter order and stopband edge frequency. The design algorithm for the matched filters consists of two-stage process of linear programming to design a Nyquist filter with a nonnegative frequency response followed by standard spectral factorization techniques to extract the nonlinear phase transmit and receive filters. The design software consists primarily of commonly available Fortran subroutine packages for linear programming and polynomial factorization and is numerically well behaved and very accurate.

I -

0.0

0.1

0.2

0.3

0.4

0.5

FREQUENCY (cycles/romple)

Fig. 5 .

Magnitude response of 100-tap FIR matched filter with oversampling factor M = 4 and excess bandwidth factor a = 0.1.

REFERENCES
L. R. Rabiner, The design of finite impulse response digital filters using linear programming techniques, Bell Syst. Tech. J., vol. 51, pp. 1177-1198, July-Aug. 1972. K. Steiglitz, Optimal design of FIR digital filters with monotone passband response, IEEE Trans. Acourt., Speech, Signal Processing, vol. ASSP-27, pp. 643-649, Dec. 1979. J. K. Liang, R. J. P. DeFigueiredo, and F. C. Liu, Design of optimal Nyquist, partial response, Nth band, and nonuniform tap spacing FIR digital filters using linear programming techniques, IEEE Trans. Circuits Syst., vol. CAS-32, pp. 386-392, Apr. 1985. R. W. Lucky, J. S a l z and E. J. Weldon, Jr., Principles of Data Communication. New York: McGraw-Hill, 1968. K. H. Mueller, A new approach to optimum pulse shaping in sampled systems using time-domain filtering, Bell Syst. Tech. J., vol. 52, pp. 123-729, May-June 1973. P. R. Chevillat and G. Ungerboeck, Optimum FIR transmitter and receiver filters for data transmission over band-limited channels, IEEE Trans. Commun., vol. COM-30, pp. 1909-1915, Aug. 1982. f A. C. Salazar and V. B. Lawrence, Design and implementation o transmitter and receiver filters with periodic coefficient nulls for digital systems, in Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing, pp. 306-310,1982. T. SaramZiki and Y. Neuvo, A class of FIR Nyquist (Nth-band) filters with zero intersymbol interference, IEEE Trans. Circuits Syst., vol. CAS-34, pp. 1182-1190, Oct. 1987. J. H. McCleIlan, T. W. Parks,and L. R.Rabiner, A computer program for designing optimum FIR linear phase digital filters, IEEE Trans. Audio, Electroacourt., vol. AU-21, pp. 506-526, Dec. 1973. 0. Herrmann and H. W. Schuessler, Design of nonrecursive digital filters with minimum phase, Electron. Lett., vol. 6, no. 11, pp. 329-330, May 1970. Y. Kamp and C. J. Wellekens, Optimal design of minimum phase FIR filters, IEEE Trans. Acousr., Speech, Signal Processing, vol. ASSP-31, pp. 922-926, Aug. 1983. M. J. T. Smith and T. P. Barnwell, 111, A procedure for designing exact reconstruction filter banks for tree structured subband coders, in Proc. IEEE Int. Conf. Acourt., Speech, Signal Processing, pp. 27.1.1-27.1.4, Mar. 1984. G. Wackersreuther, Some new aspects of filters for filter banks, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-34, pp. 1182-1200, Oct. 1986. F. Mintzer, On half-band, third-band, and Nth-band FIR filters and their design, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASP-30, pp. 734-738, Oct. 1982. G. A . Mian and A. P. Nainer, A fast procedure to design equiripple minimum-phase FIR filters, IEEE Trans. Circuits Syst., vol. CAS-29, pp. 327-331, May 1982. I M S L Library Reference Manual, International Mathematical and Statistical Libraries, Inc., Houston, TX, 1982.

I
20

30

40

50

60

70

80

STOPEAND ATTENUATION (dB)

Fig. 6. Design chart relating FIR filter length to stopband attenuation and excess bandwidth factor a for optimal equiripple transmit and receive matched filters with oversampling factor of M = 4.

Example 3: This example illustrates the efficiency of the iterative linear programming technique. A 100-tap transmit and receive matched filter pair were designed with an oversampling factor M = 4 and an excess bandwidth factor a = 0 . 1 . Thus a 199-tap FIR Nyquist filter with M = 4, a = .1 and a nonnegative frequency response was designed using the iterative linear programming techniques of Section 11-D. The linear programming design converged in 7 iterations and required 3.05 s of CPU time. The minimum phase transmit filter was then obtained via cepstral deconvolution using the algorithm presented in [15]. The resulting magnitude response is shown in Fig. 5. The optimal equiripple stopband behavior is evident and thus the grid density problems normally associated with the design of high-order filters using conventional linear programming techniques have been eliminated. A large number of examples were tested to verify the convergence properties of the iterative linear programming algorithm discussed in Section 11-D. The data obtained from these examples were compiled into a design chart which is shown in Fig. 6. This chart specifies the number of taps required for an optimal equiripple matched filter as a function of the excess bandwidth factor Q and the desired stopband attenuation for an oversampling factor of M = 4.