IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMSII: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 48, NO.

8, AUGUST 2001

Least Mean Squared Error-Design of Complex FIR Filters with Quantized Coecients
Gennaro Evangelista
Abstract The design of complex FIR lters with quantized coecients according to a discrete mean squared error criterion in the frequency domain is revisited. In the rst approach the coecients are quantized step-by-step by recursively solving a system of linear equations. The second approach is a generalised form of the rst one: The mean squared error is repeatedly minimized considering a growing number of xed, quantized coecients. The rst approach requires the same computational burden as the design of unquantized coecients, in contrast to the second where the higher exibility is paid with a higher expenditure. For a considerable number of test cases the rst approach and one variant of the second achieves a smaller mean squared error than the direct quantization method. Keywords LMS lter design, complex FIR lters, nite wordlength coecients

I. Introduction Conventional FIR lter design methods, which minimize a given error criterion, yield the innite wordlength coecients. The error E() in the frequency domain is given by the dierence of the frequency response Hc () of the lter and a prescribed complex target function Dc (): E() = W (){Hc () Dc ()}, (1)

weighted with the real function W (). Two error criteria are commonly used: the minimax criterion, which minimizes the maximum of the absolute error on a prescribed approximation domain B: ||E()|| = max |E()|, (2)

and the least mean squared error (LMSE) LMSE criterion minimizing the discrete mean squared error (MSE) ||E()||2 = 2
X

|E( )|2

(3)

for an arbitrary set { | B, = 1, . . . , } of disjunct frequencies. For the minimax criterion the coecients are computed by the well known MPR-program [17], for the LMSE criterion by solving a system of linear equations [20]. For the implementation of an FIR lter in direct form, very eciently done with digital signal processors or ASICs, the coecients (and with it the impulse response) must be of nite wordlength wL . The simplest way to get nite wordlength coecients is the direct quantization method, where the innite wordlength coecients, obtained by one of the aforementioned lter design methods [17], [20], are rounded to yield quantized coecients. As a result the corresponding errors increase and the nite wordlength coecients are not optimal in sense of the error
Manuscript received February 16, 2000; revised August 16, 2001. This work was supported by Deutsche Forschungsgemeinschaft under contract GO 849/11. This paper was recommended by Associate Editor T. Stathaki. The author was with the Digital Signal Processing Group, Ruhr-University Bochum, D44780 Bochum, Germany. He is now with SIEMENS ICM, D- 81667 Munich, Germany (email: gennaro.evangelista@mch.siemens.de).

criterion, since these lter design methods do not take into account nite coecient wordlength. However general experience predicts the existence of nite wordlength lters of the same order that result in smaller approximation errors. Several contributions deal with the problem of nding the optimal nite wordlength coecients for the minimax criterion. For nite coecient wordlength mixed integer linear programming [11], [15], [16], simulated annealing [2] and optimal or local, search methods [4], [12], [24] are applied. All these methods have nothing in common with the conventional innite wordlength FIR lter design method [17] and are usually very time-consuming for large lter orders. However, the design of nite wordlength FIR lters according to an LMSE criterion is less examined, although this error criterion is sometimes more suitable than the minimax criterion: if, for instance, in case of multirate systems the signal-to-noise ratio, as the ratio of passband and stopband energies, is an important criterion for the lter performance. An LMSE lter design minimizes the energy of the approximation error in the stopband and achieves a better signal-to-noise-ratio than a lter designed according to a minimax criterion [1]. An approach to an LMSE-design of nite wordlength FIR lters is presented in [14], where a tree search algorithm in connection with a reoptimization procedure of the MSE is used. This contribution focuses on the design of nite coecient wordlength FIR lters according to an MSE criterion. To this end an LMSE complex FIR (CFIR) lter design [20] of innite coecient wordlength is presented for the most general case of a complex impulse response and a complex target function. Then a rst approach, closely related to the direct quantization method and with the same computational burden, is developed. This method is a special case of a more general design method, which is thereafter introduced. This second, more general approach is similar to the IIR lter design method used in [10] and was successfully applied in [6], [7], [5] to the design of nite wordlength lters according to a constrained least squared error criterion [1], [23]. In this approach the coecients are quantized successively: In each iteration step one coecient is chosen (according to a certain selection criterion) and quantized. The remaining unquantized coecients are then used for a re-minimization of the mean squared error. Therefore the reminimization algorithm of [14] is generalized to complex FIR lters. The here proposed lter design procedures are less time consuming than the above tree search algorithm, but the resulting MSE usually exceeds that obtained by the approach of [14]. Both presented approaches are an extension of the wellestablished LMSE lter design [20]. Hence, in contrast to most nite coecient wordlength lter design methods the knowledge about properties of the design for innite wordlength is fully exploited. II. LMSE Design of CFIR Filters The lter design problem according to an LMSE criterion [20] is to nd the coecient set, which minimizes the MSE in (3). In the most general case of an CFIR lter the impulse

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMSII: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 48, NO. 8, AUGUST 2001

response hc (k), being identical to the direct form coecient set, is complex-valued. Then the frequency response of a causal lter is given by Hc () =
n X k=0

The minimization of the MSE with respect to S h = b,

h yields
(23)

hc (k)ejk

(4)

where the solution h represents the coecient vector of the lter with the minimal MSE ||Emin ()||2 = dT d hT S h, 2 (24)

with the lter order n. Introducing (4) into (1) yields the error at the frequency B:
(

E( ) = W ( )

n X k=0

hc (k)ej k Dc ( ) .

(5)

Using the column vector e 1 with e[] = E( ) eq. (5) is rewritten as e = W (F hc dc ) = eR + jeI (6) with

obtained by applying (23) to (22). Note, that although a complex approximation problem has been formulated with dc () and hc (k) being complex, all matrices and vectors in (22) and (23) consist only of real components. In case of real lters with hI = 0, (22) reduces to ||E()||2 = hT AhR 2hT bR + dT d R R 2 with the optimal solution hR given by AhR = bR . (26) (25)

= diag{W ( )},

= 1, . . . , , W ( ) R, = 1, . . . , , 0 k n, 0 k n, = 1, . . . , ,

(7) (8) (9) (10)

F [, k] = ej k , hc [k] = hc (k), dc [] = Dc ( ),

and the real (imaginary) part eR (eI ) of the error vector With eR and eI the MSE (3) is ||E()||2 = |e|2 = eT eR + eT eI . 2 R I Next (11) is expressed in terms of

e.

(11)

As expected, the restriction on real lters leads to a bisection of the order of the system of linear equations to be solved. Note, that the vector d still contains the imaginary part dI = 0. For the design of linear phase FIR lter the symmetry of the impulse response can be used for a further reduction of the order of the system of linear equations, which can be modied such, that d contains only the real-valued amplitude response D0 ( ). In the case of an CFIR nite wordlength lter the quantized coecients h usually dier from the optimal set h leading to a greater MSE than in the optimal case (24). The dierence of the actual MSE (22) to the minimum one (24) can be expressed in terms of the dierence coecient vector h = h h with (23) as ||E()||2 ||Emin ()||2 = hT S h, 2 2 (28) (27)

hc F dc
With some rearrangement ||E()||2 2 =

= = =

hR + j hI , FR + j FI , dR + j dI .

(12) (13) (14)

hT AhR + hT AhI + 2hT BhI 2hT bR R I R R 2hT bI + dT dR + dT dI (15) I R I

T FR W T W FR + FIT W T W FI T FR W T W FI + FIT W T W FR

is obtained, where the matrices

A B
and the vectors

= =

(16) (17)

which represents a quadratic form. Hence, this dierence is positive, if S is a real, symmetric, positive denite matrix [8]. With the denition of S in (20), (16) and (17) the rst two properties are easily proven. For the last property (S positive denite) it must be shown that xT Sx > 0 for any arbitrary x = 0. Therefore, rst (20) with x = (xR , xI )T is applied resulting

xT Sx = xT AxR + xT AxI + 2xT B T xR . R I I

Replacing

(29)

bR bI

= =


T FR W T W dR + FIT W T W dI T FR W T W dI FIT W T W dR

(18) (19) yields (20) (21)

A, B by (16), (17) and introducing yR,R = W FR xR , yR,I = W FR xI , yI,R = W FI xR , yI,I = W FI xI ,

(30)

introduced for convenience, are real-valued. With

S=

A B BT A


,h =


hR hI


xT Sx

(yR,R yR,I )T (yR,R yR,I ) +(yI,R + yI,I )T (yI,R + yI,I )

X

b=
the MSE (15) is

bR bI

,d =

dR dI

(yR,R [] yR,I [])2 + (yI,R [] + yI,I [])2 .

=1

||E()||2 = hT Sh 2hT b + dT d. 2

(22)

1 Boldface capital letters are used for matrices, boldface lower case letters for vectors. An element of a matrix A or vector a is represented by A[k, l] or a[k], respectively, where k and l indicate the row and column number.

The non-negativity of the last expression is obvious. Thus, S is positive semi-denite [8]. If a positive semi-denite matrix is not singular, it is positive denite [8]. Hence, S is positive denite, if an optimal solution h = 0 (23) exists, which is the only case of interest.

EVANGELISTA: LEAST MEAN SQUARED ERROR-DESIGN OF COMPLEX FIR FILTERS

Being real, symmetric and positive denite jected to a Cholesky-decomposition [8]:

can be sub(31)

S = RT R. R

computation. The computation of the nite wordlength coecients begins just the same as for innite wordlength using (36) with = m: y[m] = b[m] + R[m, m]h[m]. For quantized coecients |y[m]|2 is minimized, if


is a upper triangular (m + 1) (m + 1) matrix [(m+1) is the length of h in (20), hence: m=2n+1] with positive main diagonal elements: R[, ] > 0, 0 m. Applying (31) to (28) leads to h
T

(40)

hQ [m] = rd

S h

= =

R Rh
T

 b[m] = hD [m]. R[m, m]

(41)

y T y =

m X

(y[])2

(32)

So hQ [m] is the same as for DQLS (39). Next, is set to (m-1) in (36) yielding y[m 1] = 1] + R[m 1, m]hQ [m] b[m +R[m 1, m 1]h[m 1], (42)

=0

with

y = Rh = Rh Rh.

(33)

Hence, to minimize (32), the length of the vector y must be minimized. Replacing S in (23) with (31) gives Rh = (RT )1 b = b. Introducing the latter equation into (33) yields y = Rh b, or for each component of y , respectively, y[] = + b[]
m X =

where hQ [m] is already xed by (41). |y[m1]|2 is minimal, if h[m 1] is chosen as close as possible to the value, which leads to y[m 1] = 0, hence:


(34) hQ [m 1] = rd (35)

1] R[m 1, m]hQ [m]  b[m . R[m 1, m 1]

(43)

This procedure is recursively repeated for the remaining coecients by decreasing until = 0, yielding
(

R[, ]h[],

0 m,

(36)

hQ []

= =

rd
n

m X R[, ] b[] hQ [] R[, ] =+1 R[, ]

which can be used to nd recursively the optimal solution of (23). With y[] = h[] = 0, 0 m, follows from (36):
m X R[, ] b[] h[]. h[] = R[, ] =+1 R[, ]

rd hQ []

(44)

(37)

This system of linear equations is solved by starting with = m, then successively decreasing by one, using the already obtained h[], > , until the complete, optimal solution for innite coecient wordlength is found. Note, that the solution (37) could have directly derived with (23) by using the Cholesky-decomposition (31) of S . However, the recursive approach using (36) is essential for the CFIR design method with inherent coecient quantization, presented in the next section. III. Recursive Quantized Least Squares CFIR Design (RQLS) Now coecient quantization is introduced into the recursive LMSE design method of CFIR lters of section II. To this end, (27) is replaced with h = hQ h = 0. (38)

for 0 m with the previous calculated, quantized hQ []. For obvious reasons this lter design method is referred to as Recursive Quantized Least Squares (RQLS). As a dierence of RQLS to DQLS in the summations of (39) (39) h[] is replaced with hQ []. Note, that DQLS (39) and RQLS (44) need the same number of additions, multiplications and rounding operations and that the result of RQLS is generally a relatively optimum set hQ with a MSE greater or equal to the MSE of the optimally quantized coecients, obtained only applying appropriate (more timeconsuming) search algorithms. To compare the MSE of DQLS and RQLS the additional MSE |y |2 (32) is deduced for both as a function of the rounding error: erd [] = x[] rd {x[]}, (45) with x[] = hQ [] or x[] = hD and |erd []| q/2 (quantization step q = 2wL +1 ). First the additional MSE |y |2 of RQLS is derived. Therefore hQ [] (44) is introduced into y[] (36) yielding yQ [] = = + R[, ]hQ [] + b[]
 n o
m X =+1

R[, ]hQ []


Obviously, hQ represents the optimum set of quantized coecients, if and only if |y | in (32) is minimal. In the direct quantization method (DQLS) the optimal coef cients h[] (37) are rounded to yield quantized coecients
(

R[, ] rd hQ [] hQ []

(46)

hD [] = rd

m X R[, ] b[] h[] R[, ] =+1 R[, ]

with hQ [] from (44). With (45) and (46) follows:

m X

(39)

(yQ [])2 =

m X =0

R2 [, ]e2 []. rd

(47)

=0

for 0 m with rd as rounding operator. In the following approach the rounding operation is introduced in the computation process, instead of rounding all coecients after their

The additional MSE of DQLS is found with y[] (33) and hD [] = hD [] h[] = erd [] (48)

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMSII: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 48, NO. 8, AUGUST 2001

as

m X

(yD []) =

m X =0

m X =

!2

R[, ]erd []

(49)
0 1

=0

In case of RQLS (47) each rounding provides one non-negative contribution, weighted with R2 [, ], to |y |2 . In case of DQLS (49) the m(m+1)/2 contributions of the rounding errors, including mixed terms erd []erd [], can partly compensate each other. So it is accidental, which of both methods achieves a smaller MSE. But for large lter orders it seems to be more probable, that RQLS yields better results than DQLS, because the dierence between the number of error contributions, m(m + 1)/2 for DQLS and (m + 1) for RQLS, gets larger for increasing m. Only for large lter orders n and large wordlengths wL a simple stochastic model can be used with erd [] uniformly q q distributed over [ 2 , 2 ], R[, ] uniformly distributed over ]0, RMax ] and R[, ], > , uniformly distributed over [RMax , RMax ]. With these assumptions it can be shown that |y |2 is scarcely greater for DQLS than for RQLS. But the difference in |y |2 between both methods decreases the better the above assumptions (n and wL large) apply. This is not surprising, because the better the assumptions apply the closer is the case of innite wordlength. In the more interesting range of wL the above assumptions are no longer valid. For better insight some examples are discussed in the next section. FIR Examples

Mean Squared Error (log.)

8 30

40

50

60 filter order n

70

80

90

Fig. 2. MSE (log.) of bandstop lters for innite wordlength (dashed), DQLS (dash-dotted) and RQLS (solid) with wL = 6

Mean Squared Error (log.)

The MSEs are plotted in Figs. 1 and 2 for real, linear phase bandstop lters with the desired amplitude response D0 () =
8 < 1,

0, : 1,

0 0.49, 0.5 0.8, 0.9 0.95,

(50)

using three design methods: 1) LMSE design with innite wordlength (dashed line), 2) DQLS (dash-dotted line) 3) RQLS (solid line), with equidistantly distributed over B, = 100, m = n + 1 for n even, m = n+1 for n odd. 2 2

8 20

30

40

50 60 filter order n

70

80

90

Fig. 3. MSE (log.) of lowpass lters for innite wordlength (dashed), DQLS (dash-dotted) and RQLS (solid) with wL = 7

2 Mean Squared Error (log.)

8 30

40

50

60 filter order n

70

80

90

Fig. 1. MSE (log.) of bandstop lters for innite wordlength (dashed), DQLS (dash-dotted) and RQLS (solid) with wL = 7

For bandstop lters with D0 () (50) RQLS achieves nearly always a smaller MSE than DQLS. The dierence in the MSE between both methods increases with the lter order n.

In addition, numerous linear phase FIR lowpass lters have been designed with dierent transition band widths and lter orders n (e.g. Fig. 3). The dierences in the MSE between RQLS and DQLS are negligibly small for all tested lowpass lters. The MSE dierence is zero up to a certain lter order n (e.g. n = 53 in Fig. 3). Beyond this lter order RQLS leads to a little smaller or the same MSE than DQLS. For n > n0 ( 70 in Fig. 3) RQLS is sometimes worse than DQLS, but this is of no interest, because the MSE of the nite wordlength lters tends to decrease with increasing lter order n up to this certain lter order n0 . Beyond n0 the MSE saturates or oscillates about a constant value. So it does not make sense to enlarge the lter order beyond n0 to achieve a lter with smaller MSE. Similar results have been reported in [9] in conjunction with equiripple FIR lter design. Following observations concerning n0 are reported: 1) n0 increases with increasing wL . 2) n0 is larger for more stringent lter specications (narrow pass band, narrow stop band lters and/or narrow dont care regions). For the innite wordlength lowpass lters it is observed that

EVANGELISTA: LEAST MEAN SQUARED ERROR-DESIGN OF COMPLEX FIR FILTERS

there are region of n (Figs. 3) where the MSE is nearly constant (similar to [3] for an integral squared error and specied transition bands). These regions are the wider the tighter the dont care region of D0 (). Moreover, it resembles to certain properties of equiripple FIR designs in conjunction with the existence of extra ripple lter [19].
0

IV. Iterative Quantized Least Squares CFIR Design (IQLS) The RQLS method of the last section is a special case of a more general nite wordlength design method, where the successive quantization of coecients is performed in a dierent order governed by certain selection strategies. A similar approach is used in [10] for the design of IIR lters. This approach consisting of 3 steps is outlined subsequently and, further on, referred to as Iterative Quantized Least Squares (IQLS). Note, that steps 2 and 3 are repeated iteratively. 1. Initialization: The MSE is minimized with (37) yielding an CFIR lter with unquantized coecients, {h0 []} = {h[]}. The set u {h0 [Q ]} of quantized coecients is void. Q 2. Selection and Quantization of an Additional Coefcient: Out of the actual set {hi1 [u ]} of unquantized coecients u one coecient is selected, rounded and assigned to the next set {hi [Q ]} of quantized (xed) coecients. From one itQ eration step (i 1) to the next i, the set {hi [Q ]} increases by one coecient, while Q the set {hi [u ]} diminishes by one coecient. u 3. Re-Minimization of MSE: The MSE increased due to the preceding coecient quantization is re-minimized. Note, that only the set {hi [u ]} u can be used for re-minimization. The iterative procedure ends as soon as the set {hi [u ]} is void. u A. Re-Minimization of MSE Subsequently, re-minimization of the MSE in step 3 using the non-quantized coecients {hi [u ]} is described. To this u end, the MSE of (22) is expressed in terms of the vectors hu , consisting of hi [u ], and hQ with the hi [Q ]: u Q ||E()||2 2 = =

20

40 Mean Squared Error (log.)

60

80

100

120

140

200

400

600

800 1000 filter order n

1200

1400

1600

1800

Fig. 4. MSE (log.) of linear phase lowpass lters for innite wordlength (dashed), DQLS (dash-dotted) and RQLS (solid) with wL = 28, = 1000

The proposed LMSE CFIR lter design methods are applicable to large lter orders n. In Fig.4 for lter order up to 1700 and = 1000) the MSE of the design for innite wordlength decreases up to n = 1600. Beyond n = 1600 the MSE increases meaning that the design does not work properly anymore. In Fig.5 the magnitude response of a lter with n = 1600 is plotted. (S to be inverted in (23) is a 800800-matrix.)
50

hT Sh 2hT b + dT d hu T Su hu + 2hu T Smix hQ + hQ T SQ hQ 2hu T bu 2hQ T bQ + dT d, (51)

50

100 20 log |H()|

150

where bu and bQ are the parts of b related to hu and hQ , the rows and columns of S related to hu form Su , the rows and columns of S related to hQ form SQ , the rows of S related to hu and the columns of S related to hQ form Smix . The MSE (51) is minimized by setting to zero its derivation with respect to the unknown vector hu yielding

200

Su hu = bu S mix hf = bmod .

(52)

250

300

350

0.1

0.2

0.3

0.4

0.5 /2

0.6

0.7

0.8

0.9

Fig. 5. Log. magnitude response of a lowpass lter for innite wordlength with n + 1 = 1600, = 1000

With RQLS a method was presented, which achieves a smaller MSE than DQLS for a large percentage of test cases, even if the dierence is sometimes negligible as for lowpass lters. For both methods the same computational burden is needed. Hence, even if the solution of RQLS is not globally optimal, it may advantageously be combined with global search methods, e.g. to provide better starting values.

This equation is of the same kind as (23) for the original LMSE lter design. Thus, introducing xed coecients into the MSE reduces the order of the system of linear equation (52) by the number of xed coecients hi [Q ] compared with the original Q LMSE design (23). Hence, the knowledge of coecients leads to computational savings. Su has the same properties as S : it is a real, symmetric and positive denite matrix and can be subjected to a Cholesky-decomposition. Note, that (52) is also obtained, if the xed coecients are used to modify the target function (see appendix). To conclude, the relationship of RQLS and IQLS is elaborated. If always hi1 [u ] with the largest possible u is selected u and quantized in step 2, IQLS and RQLS lead to the same set of quantized coecient. In this case Su of IQLS is found by removing the last column and row of Su of the previous iteration. The matrix Ru from the Cholesky-decomposition is likewise obtained by removing the last column and row, respectively, of

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMSII: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 48, NO. 8, AUGUST 2001

the previous Ru [8]. Thus, the system of linear equations (52) to be solved is not changed in IQLS, only the last equations belonging to hi [Q ] are ignored. The contribution Smix hQ of Q hi [Q ] is shifted to the right side of the system of linear equaQ tions (52). Hence, in each iteration step i IQLS solves the same system of linear equations taking into account hi [Q ]. Since Q the computational burden of IQLS is much higher than that of RQLS, the latter approach is always preferable as long as no other selection criterion is required. B. Coecient Selection Criteria Other possible selection criteria in step 2 of IQLS are 1) smallest absolute quantization error, 2) greatest absolute quantization error, 3) smallest absolute relative quantization error, 4) greatest absolute relative quantization error, 5) smallest absolute sensitivity of the additional MSE, 6) greatest absolute sensitivity of the additional MSE, 7) smallest relative sensitivity of the additional MSE, 8) greatest relative sensitivity of the additional MSE with the quantization error hi1 [u ] = u hi1 [u ] rd hi1 [u ] . The selection criterion is indicated by u u the respective number (e.g. IQLS3 means to IQLS with selection criterion 3). C. FIR Examples DQLS, RQLS and IQLS1-8 have been applied to the design of bandstop lters specied by D0 () (50) and to dierent linear phase FIR lowpass lters and compared with respect to the MSE: IQLS1 achieves for half of the test cases a slightly smaller MSE than RQLS. The MSE for IQLS2 is nearly always greater than for RQLS. The results of IQLS3 (IQLS4) are similar to those of IQLS1 (IQLS2). The MSE of IQLS5 (IQLS6) deviates little from MSE of IQLS1 (IQLS2). Hence, the absolute quantization error |hi1 [u ]| exerts a great inuence to the absolute sensiu tivity of the additional MSE. Concluding, only IQLS1 provides an alternative to RQLS. But IQLS1 needs considerably more computation for the selection of one coecient in step 2 and especially for the repeating minimization of the MSE in step 3 (in total m minimization procedures). V. Conclusion The design of complex FIR lters with quantized coecients has been investigated. The presented methods are based on a discrete mean squared error criterion in the frequency domain. To this end, rst an LMSE CFIR design procedure for FIR lters with unquantized complex-valued coecients hc (k) has been developed along the lines of [20]. A rst approach, referred to as RQLS, is closely related to the direct quantization method (DQLS: innite wordlength LMSE design with subsequent rounding of the coecients). In RQLS the coecient quantization is included in solving the system of linear equations. For both approaches, the additional mean squared error is deduced in terms of rounding errors. For a considerable number of test cases RQLS achieves a smaller mean squared error than DQLS, while requiring the same computational burden. Next a method, which quantizes successively the coecients, was presented (IQLS). For an appropriate selection criterion IQLS and RQLS lead to the same solution, but with signicantly higher computational expenditure for IQLS. From nu-

merous test cases with dierent selection criteria it appeared, that only IQLS1 is an alternative for RQLS or DQLS. All investigated design procedures do not nd the global optimum for a given coecient wordlength. However, the results may be used as improved initial values for subsequent discrete optimization. Most interestingly, the proposed LMSE-FIR lter design methods (implemented in MATLAB) are also applicable to large lter orders n up to 1600 in the case of real linear phase FIR lters. Further research can deal with nding more appropriate selection criteria for step 2 and testing the alignment method, where not only one but more coecients are quantized at a time in step 2 and the mean squared error is re-minimized only once. Parallel research on the design of N th band lters (N > 2) [18] has shown that this alignment method with the coecients xed by  1 , for k = 0, N h0 (k) = (53) 0 , for k = N, Z, leads to N th band lters with a smaller MSE than an LMSE design with appropriate frequency domain specications and no time domain constraints, followed by the correction of the coecients to hold (53). Additionally, it can be examined, if IQLS is more suitable to the design of lters with powers-of-two coecients [2], [4], [15], [24], and if an approach similar to RQLS could be applied to the design of nite coecient wordlength lters according to an integral squared error [3]. Appendix The minimization of the MSE considering a set of xed coefcients presented in subsection IV-A can also be interpreted as an original LMSE design with modied target function Dc (). Therefore rst the frequency response (4) is expressed in terms of the real and imaginary part of the impulse response hc (k): Hc () =
n X k=0

hc,R (k)ejk + j

n X k=0

hc,I (k)ejk .

(54)

Then the sums are each separated into two parts, one for the already quantized coecients, one for the remaining coecients: Hc () = +j
X
R ku

R hc,R (ku )ejku +

X
R kQ

R hc,R (kQ )ejkQ

X
I ku

I hc,I (ku )e

I jku

+j

X
I kQ

I hc,I (kQ )ejkQ .

The obtained Hc () is applied to E( ) in (1): E( ) = W ( )

X
I ku

8 <X :
R ku

R hc,R (ku )ejku

+j with
mod Dc ( )

I I hc,I (ku )ejku

mod Dc ( )

9 = ;

(55)

Dc ( ) j
X
I kQ

X
R kQ

R hc,R (kQ )ejkQ

I hc,I (kQ )ejkQ .

(56)

EVANGELISTA: LEAST MEAN SQUARED ERROR-DESIGN OF COMPLEX FIR FILTERS

In (55) a complex error E() is derived, whose MSE is given by (11), which is minimized in the same way as for the original LMSE in section II, but with an appropriate matrix S of reduced order as Su in (52). Hence, the LMSE design considering xed coecients can be interpreted as an original LMSE design mod but with the modied target function Dc () in (56). Acknowledgement The author wishes to express his appreciation to Prof. Dr.Ing. H. G. Gckler for his careful review of the manuscript and o valuable suggestions. References
[1] J. W. Adams, FIR Digital Filters with Least Squares Stopbands Subject to Peak-Gain Constraints, IEEE Transactions on Circuits and Systems, 39:376-388, April 1991. N. Benvenuto, M. Marchesi, and Aurelio Uncini, Applications of Simulated Annealing for the Design of Special Digital Filters, IEEE Transactions on Signal Processing, 40:323-332, February 1992. C. S. Burrus, A. W. Soewito, R. A. Gopinath, Least Squared Error FIR Filter Design with Transition Bands, IEEE Transactions on Signal Processing, 40:1327-1340, June 1992. C.-L. Chen and A. N. Willson Jr., A Trellis Search Algorithm for the Design of FIR Filters with Signed-Powers-of-Two Coecients, IEEE Transactions on Circuits and Systems II, 46:29-39, January 1999. M. Erdogan, Vergleichende Untersuchung zum Entwurf linearphasiger FIRFilter mit quantisierten Koezienten, Diplomarbeit, Ruhr-Universitt a Bochum, 1999. G. Evangelista, Zum Entwurf digitaler Systeme zur asynchronen Abtastratenumsetzung, Ph.D. Thesis, Ruhr-Universitt Bochum, 2000. a G. Evangelista, Design of Optimum High Order Finite Wordlength Digital FIR Filters with Linear Phase, EURASIP Signal Processing. Accepted for Publication. F. R. Gantmacher, Matrizentheorie, Berlin: Springer, 1986. U. Heute, Uber Realisierungsprobleme bei nicht-rekursiven Digitalltern, Ph.D. Thesis, Universitt Erlangen-N rnberg, 1975. a u Z. Jing and A. T. Fam, Design of Finite Wordlength IIR Filters by Successive Discretization and Reoptimization, Int. Electrical, Electronics Conference, Toronto, Canada, 2:516-519, September 1983. D. M. Kodek, Design of Optimal Finite Wordlength FIR Digital Filters using Integer Programming Techniques, IEEE Transactions on Acoustics, Speech, Signal Processing, 28:304-308, June 1980. D. M. Kodek and K. Steiglitz, Comparison of Optimal and Local Search Methods for Designing Finite Wordlength FIR Digital Filters, IEEE Transactions on Circuits and Systems, 28:28-32, January 1981. J. K. Liang, R. J. P. DeFigueiredo, and F. C. Lu, Design of Optimal Nyquist, Partial Response, N th Band, and Nonuniform Tap Spacing FIR Filters using Linear Programming Techniques, IEEE Transactions on Circuits and Systems, 32:386-392, April 1985. Y. C. Lim and S. R. Parker, A Discrete Coecient FIR Digital Filter Design Based Upon an LMS Criteria, Int. Symposium Circuits and Systems, Rome, Italy, 3:796-799, May 1982. Y. C. Lim and S. R. Parker, FIR Filter Design Over a Discrete Powersof-Two Coecient Space, IEEE Transactions on Acoustics, Speech, Signal Processing, 31:583-591, June 1983. Y. C. Lim and B. Liu, Design of Cascade FIR Filters with Discrete Valued Coecients, IEEE Transactions on Acoustics, Speech, Signal Processing, 36:1735-1739, November 1988. J. H. McClellan, T. W. Parks and L. R. Rabiner, A Computer Program for Designing Optimum FIR Linear Phase Digital Filters, IEEE Transactions on Audio Electroacoustics, 21:506-526, 1973. F. Mintzer, On Half-band, Third-Band, and N th Band FIR-Filters and Their Design, IEEE Trans. on Acoustics, Speech, Signal Processing, ASSP-30:734-738, October 1982. S. K. Mitra and J. F. Kaiser, Handbook for Digital Signal Processing, p. 206, J. Wiley & Sons: New York, 1993. T. W. Parks and C. S. Burrus, Digital Filter Design, New York: John Wiley & Sons, 1987. H. Samueli, On the Design of Optimal Equiripple FIR Digital Filters for Data Transmission Applications, IEEE Transactions on Circuits and Systems, 35:1542-1546, December 1988. T. Saramki and Y. Neuvo, A Class of FIR Nyquist (N th-Band) Filters with a Zero Intersymbol Interference, IEEE Transactions on Circuits and Systems, 34:1182-1190, October 1987. I. W. Selesnick, M. Lang and C. S. Burrus, Constrained Least Square Design of FIR Filters without Specied Transition Bands, IEEE Transactions on Signal Processing, 44:1879-1892, August 1996. Q. Zhao and Y. Tadakoro, A Simple Design of FIR Filters with Powers-ofTwo Coecients, IEEE Transactions on Circuits and Systems, 35:566-570, May 1988.

[2]

[3]

[4]

[5]

[6] [7]

[8] [9] [10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19] [20] [21]

[22]

[23]

[24]