Design of Optimum High Order Finite Wordlength Digital FIR Filters with Linear Phase

Similar documents
4214 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 54, NO. 11, NOVEMBER 2006

EFFICIENT REMEZ ALGORITHMS FOR THE DESIGN OF NONRECURSIVE FILTERS

Minimax Design of Complex-Coefficient FIR Filters with Low Group Delay

Design of Biorthogonal FIR Linear Phase Filter Banks with Structurally Perfect Reconstruction

Design of FIR Nyquist Filters with Low Group Delay

Maximally Flat Lowpass Digital Differentiators

LINEAR-PHASE FIR FILTER DESIGN BY LINEAR PROGRAMMING

Design of Coprime DFT Arrays and Filter Banks

A Width-Recursive Depth-First Tree Search Approach for the Design of Discrete Coefficient Perfect Reconstruction Lattice Filter Bank

DESIGN OF QUANTIZED FIR FILTER USING COMPENSATING ZEROS

Iterative reweighted l 1 design of sparse FIR filters

Closed-Form Design of Maximally Flat IIR Half-Band Filters

Filter Design Problem

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 49, NO. 2, FEBRUARY H(z) = N (z) D(z) =

A DESIGN OF FIR FILTERS WITH VARIABLE NOTCHES CONSIDERING REDUCTION METHOD OF POLYNOMIAL COEFFICIENTS FOR REAL-TIME SIGNAL PROCESSING

Efficient algorithms for the design of finite impulse response digital filters

Design of Stable IIR filters with prescribed flatness and approximately linear phase

FIR BAND-PASS DIGITAL DIFFERENTIATORS WITH FLAT PASSBAND AND EQUIRIPPLE STOPBAND CHARACTERISTICS. T. Yoshida, Y. Sugiura, N.

Two Dimensional Linear Phase Multiband Chebyshev FIR Filter

Research Article Design of Finite Word Length Linear-Phase FIR Filters in the Logarithmic Number System Domain

T ulation and demodulation. It has applications not only

FIR filter design is one of the most basic and important

POLYNOMIAL-BASED INTERPOLATION FILTERS PART I: FILTER SYNTHESIS*

On the Frequency-Domain Properties of Savitzky-Golay Filters

Filter Banks with Variable System Delay. Georgia Institute of Technology. Atlanta, GA Abstract

ISSN (Print) Research Article. *Corresponding author Nitin Rawal

Digital Signal Processing Lecture 9 - Design of Digital Filters - FIR

Analysis of Finite Wordlength Effects

Computer-Aided Design of Digital Filters. Digital Filters. Digital Filters. Digital Filters. Design of Equiripple Linear-Phase FIR Filters

Design of Orthonormal Wavelet Filter Banks Using the Remez Exchange Algorithm

H Optimal Nonparametric Density Estimation from Quantized Samples

Using fractional delay to control the magnitudes and phases of integrators and differentiators

Efficient signal reconstruction scheme for timeinterleaved

Part 4: IIR Filters Optimization Approach. Tutorial ISCAS 2007

Shifted-modified Chebyshev filters

A REALIZATION OF FIR FILTERS WITH SIMULTANEOUSLY VARIABLE BANDWIDTH AND FRACTIONAL DELAY. Håkan Johansson and Amir Eghbali

OPTIMIZED PROTOTYPE FILTER BASED ON THE FRM APPROACH

Multirate Digital Signal Processing

SYNTHESIS OF BIRECIPROCAL WAVE DIGITAL FILTERS WITH EQUIRIPPLE AMPLITUDE AND PHASE

ONE-DIMENSIONAL (1-D) two-channel FIR perfect-reconstruction

ON THE REALIZATION OF 2D LATTICE-LADDER DISCRETE FILTERS

Butterworth Filter Properties

Analysis of methods for speech signals quantization

Low-delay perfect reconstruction two-channel FIR/IIR filter banks and wavelet bases with SOPOT coefficients

On Information Maximization and Blind Signal Deconvolution

Linear Programming Algorithms for Sparse Filter Design

Design of Narrow Stopband Recursive Digital Filter

Citation Ieee Signal Processing Letters, 2001, v. 8 n. 6, p

Neural Network Algorithm for Designing FIR Filters Utilizing Frequency-Response Masking Technique

FAST FIR ALGORITHM BASED AREA-EFFICIENT PARALLEL FIR DIGITAL FILTER STRUCTURES

FINITE PRECISION EFFECTS 1. FLOATING POINT VERSUS FIXED POINT 3. TYPES OF FINITE PRECISION EFFECTS 4. FACTORS INFLUENCING FINITE PRECISION EFFECTS

Lapped Unimodular Transform and Its Factorization

DISCRETE-TIME SIGNAL PROCESSING

A New Algorithm for Designing FIR-Filters with Small Word Length

ECE 8440 Unit 17. Parks- McClellan Algorithm

PS403 - Digital Signal processing


Magnitude F y. F x. Magnitude

DESIGN OF ALIAS-FREE LINEAR PHASE QUADRATURE MIRROR FILTER BANKS USING EIGENVALUE-EIGENVECTOR APPROACH

3 Finite Wordlength Effects

DIGITAL SIGNAL PROCESSING UNIT III INFINITE IMPULSE RESPONSE DIGITAL FILTERS. 3.6 Design of Digital Filter using Digital to Digital

SHAPE OF IMPULSE RESPONSE CHARACTERISTICS OF LINEAR-PHASE NONRECURSIVE 2D FIR FILTER FUNCTION *

COMPARISON OF CLASSICAL CIC AND A NEW CLASS OF STOPBAND-IMPROVED CIC FILTERS FORMED BY CASCADING NON-IDENTICAL COMB SECTIONS

Optimal Design of Real and Complex Minimum Phase Digital FIR Filters

A Unified Approach to the Design of Interpolated and Frequency Response Masking FIR Filters

Determining Appropriate Precisions for Signals in Fixed-Point IIR Filters

Problem Set 9 Solutions

Organization of This Pile of Lecture Notes. Part V.F: Cosine-Modulated Filter Banks

EECS 123 Digital Signal Processing University of California, Berkeley: Fall 2007 Gastpar November 7, Exam 2

Stable IIR Notch Filter Design with Optimal Pole Placement

-Digital Signal Processing- FIR Filter Design. Lecture May-16

GENERALIZED DEFLATION ALGORITHMS FOR THE BLIND SOURCE-FACTOR SEPARATION OF MIMO-FIR CHANNELS. Mitsuru Kawamoto 1,2 and Yujiro Inouye 1

Filter Banks for Image Coding. Ilangko Balasingham and Tor A. Ramstad

Filters. Massimiliano Laddomada and Marina Mondin. Abstract

Dominant Pole Localization of FxLMS Adaptation Process in Active Noise Control

Difference between Reconstruction from Uniform and Non-Uniform Samples using Sinc Interpolation

Design of High-Performance Filter Banks for Image Coding

Design of Infinite Impulse Response (IIR) filters with almost linear phase characteristics via Hankel-norm approximation

Enhanced Steiglitz-McBride Procedure for. Minimax IIR Digital Filters

A UNIVERSAL SQUARE-ROOT NYQUIST (M) FILTER DESIGN FOR DIGITAL COMMUNICATION SYSTEMS

DSP. Chapter-3 : Filter Design. Marc Moonen. Dept. E.E./ESAT-STADIUS, KU Leuven

Perfect Reconstruction Two- Channel FIR Filter Banks

Quantization and Compensation in Sampled Interleaved Multi-Channel Systems

Quadrature-Mirror Filter Bank

DESIGN OF INFINITE IMPULSE RESPONSE (IIR) FILTERS WITH ALMOST LINEAR PHASE CHARACTERISTICS

Towards Global Design of Orthogonal Filter Banks and Wavelets

REAL TIME DIGITAL SIGNAL PROCESSING

CHOICE OF THE WINDOW USED IN THE INTERPOLATED DISCRETE FOURIER TRANSFORM METHOD

STOCHASTIC INFORMATION GRADIENT ALGORITHM BASED ON MAXIMUM ENTROPY DENSITY ESTIMATION. Badong Chen, Yu Zhu, Jinchun Hu and Ming Zhang

Design of Higher Order LP and HP Digital IIR Filter Using the Concept of Teaching-Learning Based Optimization

On the Use of A Priori Knowledge in Adaptive Inverse Control

ONE can design optical filters using different filter architectures.

1. FIR Filter Design

UNIVERSITY OF OSLO. Faculty of mathematics and natural sciences. Forslag til fasit, versjon-01: Problem 1 Signals and systems.

A METHOD OF ADAPTATION BETWEEN STEEPEST- DESCENT AND NEWTON S ALGORITHM FOR MULTI- CHANNEL ACTIVE CONTROL OF TONAL NOISE AND VIBRATION

DSP-CIS. Chapter-4: FIR & IIR Filter Design. Marc Moonen

DESIGN OF LINEAR-PHASE LATTICE WAVE DIGITAL FILTERS

On the steady-state mean squared error of the fixed-point LMS algorithm

LAB 6: FIR Filter Design Summer 2011

A HIGH-SPEED PROCESSOR FOR RECTANGULAR-TO-POLAR CONVERSION WITH APPLICATIONS IN DIGITAL COMMUNICATIONS *

Transcription:

EURASIP SIGNAL PROCESSING 1 Design of Optimum High Order Finite Wordlength Digital FIR Filters with Linear Phase Gennaro Evangelista Abstract A novel iterative quantization procedure for the design of finite wordlength linear phase FIR filters of high order and minimum frequency domain error is proposed: a one by one increased number of filter coefficients is quantized where the augmented frequency domain error is re-minimized in each case This new approach achieves a smaller frequency domain error than rounding of the optimal non-quantized coefficients It is also successfully applied to the design of Lth band filters with minimum frequency domain error FIR filters up to a filter order 1500 are designed Keywords Design of high order FIR filters, Constrained Least Squares (CLS) design, finite wordlength coefficients I Introduction Conventional design methods of FIR filters with linear phase minimize a norm of the frequency domain error E(Ω) = H(Ω) D(Ω) (1) on a prescribed approximation domain B, being a subset of [0, π] H(Ω) is the real-valued amplitude response and D(Ω) a given real-valued target function (In this contribution no weighting function is considered) Two error norms are commonly used: the -error norm (maximum error) E(Ω) = max E(Ω) (2) and the L 2 -error norm (mean squared error) E(Ω) 2 2 = 1 E(Ω) 2 dω (3) The coefficients of a filter with minimum -error norm are computed by the well known MPR-program [16] The filter coefficients for a minimum L 2 -error norm are obtained by solving a system of linear equations [3] Both algorithms yield infinite wordlength coefficients However the realization of a digital filter with finite expenditure allows inevitably only filter coefficients of finite precision If a digital signal processor with floating point arithmetic is used, the effects of a finite coefficient wordlength w can (nearly always) be neglected However for a fast and cheap implementation of a digital filter (eg as ASIC) a fixed point arithmetic has to be used (as assumed in the following) and the effects of finite coefficient wordlength must be considered This is especially true for very long filters with applications to sharp cut-off or sample rate alteration [24] The author is with SIEMENS ICM, Munich email: gennaroevangelista@mchsiemensde This work was performed during his time as research assistant with Digital Signal Processing Group at the Ruhr-Universität Bochum, Germany and was supported by Deutsche Forschungsgemeinschaft under contract GO 849/1-1 The most widespread method of finding finite wordlength coefficients is the direct quantization (DQ) method in which the infinite wordlength coefficients, obtained by one of the aforementioned filter design methods [16], [19], are rounded to yield quantized coefficients As a result the respective error norm increases, possibly beyond the maximally allowed value To achieve a smaller error norm the coefficient wordlength and/or the filter order must be increased [9], leading to an additional expenditure As the finite wordlength coefficients found by the direct quantization method are not optimal in sense of the error norm (since this method does not take into account the coefficient wordlength), finite wordlength filters of original order and coefficient wordlength may exist with smaller error norm Several contributions deal with the problem of finding the optimal finite wordlength coefficients of FIR filters with minimum -error norm by using mixed integer linear programming [10], [14], [15], simulated annealing [2] and optimal or local search methods [4], [11], [25] All these methods have nothing in common with the conventional infinite wordlength FIR filter design method [16] and are very time-consuming or, even worse, do not converge for high filter orders A more promising approach, where the coefficients are quantized one by one and the L 2 -error norm is reminimized each time, is presented in [13] An application of this approach to the design of FIR filters with minimum -error norm requires an algorithm for re-minimization of the -error norm A modification of the MPR algorithm [16] can not be recommended because the MPR algorithm can only influence as many extrema as free variables (number of non-quantized filter coefficients) exist, and the - error norm increases significantly with a decreasing number of extrema [6], [5] This contribution focuses on the design of linear-phase FIR filters of high order and minimum -error norm (2) A filter with minimum -error norm is further on referred to as optimal Instead of a modified MPR algorithm a modified Constrained Least Square (CLS)-algorithm [1] is proposed for re-minimization The filter coefficients are assumed to be identical with the impulse response h(k) as it is the case for a direct form implementation The coefficients are restricted to take only discrete values on the quantization grid [ 1 + q, 1 + 2q,, 1 q, 1] with q = 2 w+1 First a procedure similar to the method in [13] is outlined, where the filter coefficients are iteratively quantized Then the modified CLS-algorithm is presented which is used for the re-minimization of the -error norm after the quantization of one somehow selected coefficient The

2 EURASIP SIGNAL PROCESSING design examples show the power of this new approach II Iterative Quantization Procedure The proposed Iterative Quantization (IQ) consists of 3 steps Steps 2 and 3 are repeated iteratively In each iteration one additional coefficient is quantized So the set of quantized coefficients increases from iteration to iteration until all coefficients are quantized The IQ procedure is outlined subsequently (Fig 1): 1 Initialization: The -error norm is minimized with the MPR [16] yielding the non-quantized coefficient set {h 0 (k)} The set {h 0 Q (k f)} of quantized coefficients and the set {k f } of indices of quantized coefficients are void 2 Selection and Quantization of an Additional Coefficient: Out of the actual set {h i 1 (k)} of non-quantized coefficients one coefficient h(k S ) is selected, rounded to the next quantization step and assigned to the next set {h i Q (k f)} of quantized (fixed) coefficients The index k S is shifted from the set {k} of indices of non-quantized coefficients to the set {k f } of indices of quantized coefficients Hence, from one iteration (i 1) to the next i, the set {h i Q (k f)} increases by one coefficient, while the set {h i (k)} diminishes by one coefficient 3 Re-Minimization of -Error Norm: The -error norm increased due to the preceding coefficient quantization is re-minimized Note that only the set {h i (k)} can be used for re-minimization The iterative procedure ends as soon as the set {h i (k)} is void error: min k h(k) = min h Q(k) h(k) (4) k The IQ method is suboptimal but leads with a great probability to a smaller -error norm than the direct quantization of the coefficients from the MPR-program (DQMPR), because the -error norm is constantly reminimized considering the already quantized coefficients III Re-Minimization of -Error Norm For the re-minimization of the -error norm considering the set {h i Q (k f)} of fixed filter coefficients a modified CLS-algorithm is used The CLS-algorithm [1] minimizes the L 2 -error norm (3) while the -error norm (2) is constrained not to exceed prescribed bounds To use the CLSalgorithm for the minimization of the -error norm these prescribed bounds must be chosen as tighten as possible The combination of the Iterative Quantization with the modified CLS-algorithm for the re-minimization of the - error norm is subsequently called IQCLS First the original CLS-algorithm [1] is described and then its modification A CLS-Algorithm for the Design of FIR Filters The amplitude response H(Ω) of a zero-phase FIR filter of even order n and even symmetry 1 is with 2 n/2 H(Ω) = h(0) + 2 cos(kω)h(k) = a T h (5) k=1 Initialization: Minimization of L -error {h 0 (k)} with {kf },{h 0 Q (kf )} void i=1 i=0 a T = [1, 2 cos(ω),, cos( n 2 Ω)], (6) h = [h(0),, h( n 2 )]T (7) With (5) the error E(Ω) of (1) is E(Ω) = a T h D(Ω) (8) Selection of oneh(ks ) out of {h i-1 (k)}: k {k }, {k}={k} k S f Quantization of h(k hq (k S ) {h Q i (kf ) } S ) S i=i+1 Since E(Ω) is real and a scalar: E(Ω) 2 = E 2 (Ω) = E(Ω) T E(Ω) (9) Introducing (9) with (8) in (3) leads to an alternative expression of the L 2 -error norm [3] E(Ω) 2 2 = h T Ah 2h T b + 1 D 2 (Ω)dΩ, (10) Fig 1 Re-Minimization of L -error {h i (k)} Flow chart of the Iterative Quantization (IQ) Procedure where A = 1 b = 1 aa T dω, (11) ad(ω)dω (12) There exists a large number of possible selection criteria for step 2 The most successful selection strategy [6], [5] chooses the coefficient with the smallest absolute rounding 1 Equivalent equations can easily be found for all other types of linear phase FIR-filters [19] 2 Boldface capital letters are used for matrices, boldface lower case letters for vectors

EVANGELISTA: DESIGN OF DIGITAL FIR FILTERS 3 The minimization of (10) with respect to h leads to the starting solution of the CLS-algorithm: h start = A 1 b (13) If all local extrema of the amplitude response of h start do not exceed the prescribed upper and lower bounds R lo (Ω) and R up (Ω) on B, no iteration is necessary and h start is the final solution of the CLS-algorithm Otherwise the amplitude response is set to the upper or lower bound, respectively, at all extremal frequencies Ω ν, where H(Ω) violates these bounds: { H(Ω ν ) =! Rup (Ω ν ), if H(Ω ν ) > R up (Ω ν ), (14) R lo (Ω ν ), if H(Ω ν ) < R lo (Ω ν ), or in matrix notation with B = a T (Ω ν ) Bh = r (15), r = R up,lo (Ω ν ) (16) Hence, in this case the extended L 2 -error norm, the Lagrange-function [8] E Lag = E(Ω) 2 2 m T (Bh r) (17) with the Lagrange-multipliers m, has to be minimized The minimization of the Lagrange-function (17) in conjunction with (10) with respect to h leads to Ah 05B T m = b (18) Eqs (15) and (18) commonly represent a set of linear equations with the solution [1]: m = 2(BA 1 B T ) 1 (r BA 1 b) (19) h = A 1 (b + 05B T m) (20) The amplitude response of h is calculated and if its extrema are beyond the bounds at the (new) extremal frequencies Ω ν, the procedure starting at eq (14) is repeated The CLS-algorithm stops if all extrema of the amplitude response are within the bounds (convergence) or the set {Ω ν } of extremal frequencies does not change anymore (no convergence, because the bounds are to tight) B Modification of the CLS-Algorithm To consider fixed coefficients h Q (k f ) in the CLSalgorithm the amplitude response (5) is expressed depending of a vector h f containing the fixed coefficients h Q (k f ) and a vector h u containing all non-quantized coefficients to be used for re-minimization: H(Ω) = a T u h u + a T f h f (21) Replacing H(Ω) with (21) in (1) yields where E(Ω) = a T u h u D Mod (Ω), (22) D Mod (Ω) = D(Ω) a T f h f (23) As the expressions of the modified error (22) and the original error (1) and (5) are the same replacing a with a u, h with h u and D(Ω) with D Mod (Ω), the modified and original CLS-algorithm are also the same by using the same replacements So A (11), b (12) and B (16) must be replaced with A Mod = 1 a u a T u dω, (24) b Mod = 1 a u D Mod (Ω)dΩ, (25) B Mod = a T u (Ω ν ) (26) The matrix A Mod to be inverted in the modified CLSalgorithm is of lower order than the matrix A of the original CLS-algorithm So the consideration of the fixed coefficients h f leads to a saving in computational expenditure The upper and lower bounds are introduced as follows R up (Ω) = D(Ω) + δ, (27) R lo (Ω) = D(Ω) δ, (28) with δ as input parameter of the modified CLS-algorithm δ must be iteratively decreased for the re-minimization of the -error norm (step 3 of the IQ) until the modified CLS-algorithm does not converge anymore The smallest value of δ with convergence of the modified CLS-algorithm is the minimal -error norm A major advantage of the IQCLS algorithm is that is based on a conventional design method (the CLS-algorithm [1]): All experience, especially with the design of filters of high order, can be applied IV Examples This section focuses on the design of lowpass filters, because lowpass filters are widely used and similar results are also obtained for all other types of frequency selective filters A first target function D(Ω) is { L, for 0 Ω ΩP, D(Ω) = (29) 0, for Ω S Ω π, where the factor L in D(Ω) is chosen in such a way that the quantization grid [ 1 + q, 1 + 2q,, 1 q, 1] for the coefficients h Q (k f ) is optimally exploited In Fig 2 the magnitude resonse H(Ω)/L of a lowpass filter designed with IQCLS is depicted Although already 7 of 11 coefficients are quantized with 8 bit, H(Ω)/L shows nearly equi-ripple behaviour

4 EURASIP SIGNAL PROCESSING Magnitude / db Magnitude / db 02 01 0 01 02 0 005 01 015 02 025 03 035 04 35 40 45 Ω / π 50 05 06 07 08 09 1 11 12 13 14 15 Ω / π Fig 2 20 log 10 H(Ω)/L with IQCLS for n = 20, Ω P = 04π, Ω S = 06π, 7 of 11 coefficients quantized with w = 8: passband (above), stopband (below) A Design of FIR Lowpass Filters FIR lowpass filters were designed for different specifications (29) and wordlengths w using the IQCLS algorithm The following observations can be reported: 13 14 15 16 17 18 19 21 n=50 n=60 n=70 n=80 n=90 n=100 0 5 10 15 20 25 30 35 40 45 50 Number of quantized coefficients Fig 3 Remaining log -error norm after IQCLS re-minimization for Ω P = 019, Ω S = 0208π, L = 5, w = 8 For a small number of quantized coefficients the remaining -error norm increases scarcely after re-minimization in step 3 of the IQCLS algorithm (s figure 3) This is due to the fact that for IQCLS re-minimization enough nonquantized coefficients are left (eg for n = 90, 100 in figure 3) In the opposite case a smaller final -error norm can sometimes be achieved by quantizing all remaining nonquantized coefficients directly The number of coefficients to be quantized in one step can be deduced of the -error norm of figure 3 (eg for n = 50, 70, 80 as indicated by the dashed lines) The overall increase of the -error norm during the IQCLS is smaller for larger coefficient wordlength w As expected the IQCLS algorithm (with adequate direct quantization of the remaining coefficients as in figure 3) outperforms the DQMPR approach, if the difference between the minimax errors of the DQMPR and MPR is large (see figures 4, 5, 6) 10 15 25 35 50 100 150 200 Filter order n Fig 4 Log -error norm with IQCLS ( ), DQMPR ( ) and MPR (+) for Ω P = 019, Ω S = 0208π, L = 5, w = 8 For a given wordlength w and specification D(Ω) a filter order n max exists beyond which the -error norm does not improve anymore for both methods, IQCLS and DQPMR (see figure 5: n max 250 for w = 8, figure 6: n max 180 for w = 6) This result is consisting with [9] 10 40 50 60 70 80 90 100 110 120 100 200 300 400 500 600 700 800 900 Filter order n w=8 w=12 w=16 Fig 5 Log -error norm with IQCLS ( ), DQMPR ( ) and MPR (+) for Ω P = 019, Ω S = 0208π, L = 5 FIR filters were successfully designed using the IQCLS algorithm (implemented in Matlab) up to a filter order of n = 1500 Beyond this upper bound for n the Choleskydecomposition [7], [8] used to invert A Mod does not work properly The IQCLS method needs considerably more computation time than the DQMPR (eg 750 re-minimizations of - error are required for n = 1500 in IQCLS while in DQMPR there is only one minimization of -error)

EVANGELISTA: DESIGN OF DIGITAL FIR FILTERS 5 10 295 15 5 25 31 315 32 325 33 335 34 35 50 100 150 200 Filter order n 345 2 3 4 5 6 7 8 9 10 Upsampling factor L Fig 6 Log -error norm with IQCLS ( ), DQMPR ( ) and MPR (+) for Ω P = 019, Ω S = 0208π, L = 5, w = 6 Fig 7 Log -error norm with IQCLS, ( ), CMPR ( ) and MPR (+) for Ω P = 096 π, n = 40L, non-quantized coefficients L B Design of Lth-Band Filters The IQCLS algorithm can also be applied to the design of Lth-band (Nyquist-)filters [17] with minimum -error norm The zero phase impulse response h(k) of Lth-band filters must hold h(νl) = { 1, for ν = 0, 0, for ν Z{0} (30) Lth-band filters are advantageously used as anti-imaging filters in interpolators with upsampling factors L [24] An appropriate target function D(Ω) for Lth-band filters is given by [22]: { L, 0 Ω ΩP, D(Ω) = 0, L ν Ω p Ω L ν + Ω p, (31) with ν = 1,, L 1 Obviously D(Ω) has L 1 stopbands The task of designing Lth-band filters with a minimum -error norm is addressed by [2], [12], [15], [20], [21] using methods with high computational expenditure The use of the IQCLS algorithm makes the design very fast: The coefficients given by (30) are all treated as fixed coefficients in step 2 and the re-minimization of the -error norm by the modified CLS algorithm is done once in step 3 yielding the optimal, non-quantized coefficients Of course, the design procedure can be continued in order to have all coefficients quantized, as treated above The -error norm of Lth-band filters for different upsampling factors L and design methods is depicted in figure 7 The coefficients h(k) resulting from the MPR-program with D(Ω) according to (31) violate (30) Hence the result is not an Lth band filter The corresponding -error norm is only depicted as lower bound of the achievable - error norm The forced correction of the coefficients h(k) from the MPR to hold (30) is done by the CMPR As can clearly be seen the IQCLS algorithm leads to a considerably smaller -error norm than the CMPR [6] V Conclusion and Outlook The design of linear phase FIR filters with finite wordlength coefficients and minimum -error norm using an Iterative Quantization approach combined with a modified CLS algorithm (IQCLS) has been presented The IQCLS algorithm leads to a smaller -error norm than the DQMPR (direct quantization of the coefficients h(k) resulting from the MPR program) for the case, where a great difference between the -error norms of MPR and DQMPR exists The IQCLS has successfully been applied to filters up to a filter order of n = 1500 It is also very suitable for the design of Lth-band filters with minimum -error norm For a further reduction of the -error norm the direct quantization of one or more remaining coefficients at the end of the IQCLS (see sect II) might be replaced by applying local search algorithms [11] The IQCLS can also be easily applied to the design of filters with different coefficient wordlengths This leads to simple and fast multiplications for those coefficients with very small wordlengths Especially the coefficients quantized in the first iteration steps of the IQCLS require only a small wordlength [6] (see also figure 3) The design of filter with solely power-of-two coefficients or with coefficients in CSD-code representation [18] is possible, too With a small modification, the IQCLS can also be used for the design of finite wordlength linear phase FIR filters and Lth-band filters with minimum L 2 -error norm Hence, with the IQCLS a very flexible and powerful filter design tool is given Because the selection criterion (4) used in the second step of the IQ is not proven to be optimal, further research could deal with finding other more appropriate selection criteria Acknowledgement The author wishes to express his appreciation to Prof Dr-Ing H G Göckler and the unknown reviewers for their careful review of the manuscript and valuable suggestions

6 EURASIP SIGNAL PROCESSING 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 [2] N Benvenuto, M Marchesi, and Aurelio Uncini, Applications of Simulated Annealing for the Design of Special Digitals Filters, IEEE Transactions on Signal Processing, 40:323-332, February 1992 [3] 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 [4] C-L Chen and A N Willson Jr, A Trellis Search Algorithm for the Design of FIR Filters with Signed-Powers-of-Two Coefficients, IEEE Transactions on Circuits and Systems II, 46:29-39, January 1999 [5] M Erdogan, Vergleichende Untersuchung zum Entwurf linearphasiger FIR-Filter mit quantisierten Koeffizienten, Diplomarbeit, Ruhr-Universität Bochum, 1999 [6] G Evangelista, Zum Entwurf digitaler Systeme zur asynchronen Abtastratenumsetzung, PhD Thesis, Bochum, 2000 [7] F R Gantmacher, Matrizentheorie, Berlin: Springer, 1986 [8] G Hämmmerlin and K-H Hoffman, Numerische Mathematik, Berlin: Springer, 1991 [9] U Heute, Über Realisierungsprobleme bei nicht-rekursiven Digitalfiltern, PhD Thesis, Erlangen, 1975 [10] 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 [11] 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 [12] J K Liang, R J P DeFigueiredo, and F C Lu, Design of Optimal Nyquist, Partial Response, Nth Band, and Nonuniform Tap Spacing FIR Filters using Linear Programming Techniques, IEEE Transactions on Circuits and Systems, 32:386-392, April 1985 [13] Y C Lim and S R Parker, A Discrete Coefficient FIR Digital Filter Design Based Upon an LMS Criteria, Int Symposium Circuits and Systems, Rome, Italy, 3:796-799, May 1982 [14] Y C Lim and S R Parker, FIR Filter Design Over a Discrete Powers-of-Two Coefficient Space, IEEE Transactions on Acoustics, Speech, Signal Processing, 31:583-591, June 1983 [15] Y C Lim and B Liu, Design of Cascade FIR Filters with Discrete Valued Coefficients, IEEE Transactions on Acoustics, Speech, Signal Processing, 36:1735-1739, November 1988 [16] 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 [17] F Mintzer, On Half-band, Third-Band, and Nth Band FIR- Filters and Their Design, IEEE Trans on Acoustics, Speech, Signal Processing, ASSP-30:734-738, October 1982 [18] S K Mitra and J F Kaiser, Handbook for Digital Signal Processing, p 206, J Wiley & Sons: New York, 1993 [19] T W Parks and C S Burrus, Digital Filter Design, New York: John Wiley & Sons, 1987 [20] 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 [21] T Saramäki and Y Neuvo, A Class of FIR Nyquist (Nth- Band) Filters with Zero Intersymbol Interference, IEEE Transactions on Circuits and Systems, 34:1182-1190, October 1987 [22] R W Schafer and L R Rabiner, A Digital Signal Processing Approach to Interpolation, Proceedings of the IEEE, 61:692-702, June 1973 [23] I W Selesnick, M Lang and C S Burrus, Constrained Least Square Design of FIR Filters without Specified Transition Bands, IEEE Transactions on Signal Processing, 44:1879-1892, August 1996 [24] P P Vaidyanathan, Multirate Systems and Filter Banks, Englewood Cliffs: Prentice Hall, 1993 [25] Q Zhao and Y Tadakoro, A Simple Design of FIR Filters with Powers-of-Two Coefficients, IEEE Transactions on Circuits and Systems, 35:566-570, May 1988