IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: FUNDAMENTAL THEORY AND APPLICATIONS, VOL. 48, NO.

6, JUNE 2001

699

A New Eigenfilter Based on Total Least Squares
Error Criterion
Soo-Chang Pei, Fellow, IEEE and Chien-Cheng Tseng, Member, IEEE

Abstract—In this paper, a new eigenfilter based on total least
squares error criterion is investigated. The filter coefficients are
obtained from the elements of the eigenvector corresponding to
minimum eigenvalue of a real, symmetric and positive definite matrix. Four features of new method are given below. First, the computation of filter coefficients of new eigenfilter is more numerically stable than that of the least-squares method whose solution
is obtained by solving matrix inverse. Second, new eigenfilter does
not need a reference frequency point for normalization as done in
traditional eigenfilter. Third, the solution of the new eigenfilter is
closer to the solution of the least-squares method than one of the
conventional eigenfilter. Fourth, the proposed method is easy to incorporate with linear constraints and can be extended to design
equiripple and two dimensional linear phase filters. Several design
examples are used to illustrate the effectiveness of this new design
approach.
Index Terms—Eigenfilter,
least-squares criterion.

least-squares

design,

total

I. INTRODUCTION

C

ONVENTIONALLY, we often use the well-known
McClellan–Parks–Rabiner (MPR) computer program
and standard linear programming technique to design linear
phase finite impulse respones (FIR) filters according to the
Chebyshev criterion which minimizes the maximum error in
frequency response [1], [2]. The minimax designs usually give
the designers a smallest length filter for a given specification.
However, it is difficult for the MPR algorithm to incorporate
both a time- and frequency-domain constraint. And, linear
programming technique needs a large memory space and considerable computation time. Thus, a number of researchers have
considered linear phase FIR filter design using least-squares
optimality criterion.
From literature survey, two well-documented least squares
approaches for FIR filter designs are inverse matrix (IM) method
and eigen-approach. References [3]–[8] are examples of publications which include descriptions of using two methods to design various filters. The IM methods are based on solving a set of
linear simultaneous equations by matrix inversion [3], [4], and
the eigen-approaches are based on the computation of an eigenvector of an appropriate real, symmetric, and positive-definite
matrix [5]–[8]. Compared with minimax design, the advantage

of least squares design is that it is easy to add constraints and it
requires simple computation. So far, least squares approach has
been widely used to design various filters in multirate applications and image processing [9], [10].
The purpose of this paper is to design linear phase FIR filters
using total least squares (TLS) error criterion which has been
successfully used to solve many engineering problems such
as acoustic radiation problem, adaptive filtering, beamformer
and harmonic retrieval etc., [11]–[13]. The filter coefficients
based on TLS criterion are obtained from the elements of the
eigenvector corresponding to minimum eigenvalue of a real,
symmetric and positive definite matrix. Due to this, the total
least squares filter design is referred to as the new eigenfilter
approach. The main difference between conventional and
new eigenfilters is that conventional method needs to specify
the reference point in frequency domain, but new approach
does not require this. Moreover, three unique features of
new eigenfilter are as follows. First, the computation of filter
coefficients of new eigenfilter is more numerically stable
than the least-squares method whose solution is obtained by
solving IM. Second, the solution of the new eigenfilter is
closer to the solution of the least-squares method than one
of the conventional eigenfilter. Third, it is easy for the new
eigenfilter to incorporate time domain constraints and other
linear constraints as for the traditional eigenfilter. Moreover,
the new eigenfilter can be extended to design equiripple and
two dimensional linear phase filters.
The paper is organized as follows. In Section II, the linear
phase FIR filter designs using conventional IM method and
eigen-approach are briefly reviewed. In Section III, a new
eigenfilter based on total least squares error criterion is developed. In Section IV, we extend the new eigenfilter approach to
design FIR filters in conjunction with general linear constraints.
The notch filter with controlled null width is presented to show
the effectiveness of our method. In Section V, we use iterative
weighted total least squares method to design equiripple linear
phase FIR filters. Finally, the new eigen-approach is modified
to design two dimensional quadrantally symmetric FIR filters
and concluding remarks are made.
II. CONVENTIONAL LEAST SQUARES FILTER DESIGN

Manuscript received February 24, 2000; revised November 15, 2000. This
paper was recommended by Associate Editor P. P. Vaidyanathan.
S.-C. Pei is with the Department of Electrical Engineering, National Taiwan
University, Taipei, Taiwan, R.O.C.
C.-C. Tseng is with the Department of Computer and Communication Engineering, National Kaohsiung First University of Science and Technology, Kaohsiung, Taiwan, R.O.C.
Publisher Item Identifier S 1057-7122(01)04289-1.

A. Problem Statement
A causal

1057–7122/01$10.00 © 2001 IEEE

-th order FIR filter can be represented as

(1)

the optimal solution (9) The actual value of can be obtained by matrix inversion or by solving simultaneous linear equations (10) For eigenfilter design. the optimal value Because that minimizes must satisfy 0. Obviously. and scalar are band. the minimum of . we are interested in only the minimum eigenvector of a symmetric matrix. symmetric and positive-defNote that ocinite matrix. whereas the column vector (3) and (4) then we rewrite (2) as (5) The notation denotes the vector or matrix transpose.700 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: FUNDAMENTAL THEORY AND APPLICATIONS. where is related to the impulse response of the The coefficient is a function of the filter order . Defining filter. The design procedure of these eigenfilters can be summarized a unified formulation as follows. the final solution is given by (14) Therefore. Depending on whether is symmetric or antisymmetric. the problem is how we can find such that the magnitude response in (5) fits the desired magnitude response as well as possible. Thus. differentiators and Hilbert transformers [5]–[7]. but excluding the transition where is the region . The magnitude responses of these four types of filters can be expressed as (2) is an appropriate trigonometrical function. the straint is simply the eigenvector of solution vector corresponding to its minimum eigenvalue in view of the well-known Rayleigh’s principle [14]. we should scale the soproperly to satisfy lution [5]–[7]. approximate the Step 2: Make magnitude response . 6. . To achieve this specification approximation. JUNE 2001 This filter is said to have linear phase if phase response is linear is even or odd. To avoid trivial solutions. and positive definite matrix so that it is often referred to as eigenfilter approach in the literature. Several interesting design examples of such a least squares FIR filters can be found in [3]. vector . Conventional Least Squares Filter Design The conventional least squares approach corresponds to minimizing the following objective function [Tufts and Francis. From (6) (8) (13) is a real. where is the multivariable derivative of . eigenfilter approach has been used to design various types of filters such as lowpass filters. we will design linear phase FIR filter using another least squares error measure. and symmetric matrix. 1970] as (12) where matrix (6) . Now. The solution in this case is the eigenvector corresponding to the minimum eigenvalue of a real. 48. we find the optimal solution by minimizing the following quadratic error measure (11) Using (5). VOL. approximate Step 3: Since we have made . the simultaneous linear equations can be solved by a computationally efficient method. we can rewrite B. Conventional Eigenfilter Design In the following. So far. we obtain four whether types of real coefficient linear phase filters [2]. The matrix (7) is a quadratic function of . like Cholesky decomposition [16]. real. or iterative power method [16]. C. this computation can be performed efficiently by the conjugate gradient method [15]. Under this condition. satisfying Step 1: Choose the reference point . the concurs at 1 is added. Various least squares error measures will be used through this paper. Because is a positive-definite. NO. Some consideration of this choice can be found in [5]–[7]. and in frequency. symmetric.

we have (19) where and are given by (20) Fig.” When the point perplane does not fall on the plane . Since the solution vector is simply the . We claim that the new eigenfilter is optimal in total least squares error sense. For a given frequency . Third. Now. Substituting (16) into (18). but the normalization of vector to satisfy new eigenfilter is to scale minimum eigenvector vector to the . the notes a point in . Second. the error between them can be measured in several ways. but new eigenfilter does of not require this choice. we summarize the design procedure of new eigenfilter as follows. A New Eigenfilter Design The well-known linear phase filter design problem is to find a filter weight such that the desired magnitude response is equal to the actual magnitude response of the filter. substituting (17) into (18). structural identification and harmonic retrieval etc. but the size of the matrix conventional eigenfilter is of new eigenfilter is . Note that the itive definite matrix. In [11]. Step 1: Compute the matrix Step 2: Calculate the minimum eigenvector of the matrix . (b) Type 2 error.PEI AND TSENG: A NEW EIGENFILTER BASED ON TOTAL LEAST SQUARES ERROR CRITERION 701 The least squares filter design problem means that the optimal filter weight is obtained by minimizing the squared errors (18) . conventional eigenfilter needs to specify the reference point. Two typical ones with geometric interpretation are shown in Fig. Geometric interpretation of two error measures at frequency (a) Type 1 error. we obtain which is same as the (6). Now. the minimum of corresponding to the minimum the eigenvector of the matrix is also a real. 1. Now.. the filter design problem can be reto fall on the hystated as “we want the point for all . Next. First. the normalization of conventional eigenfilter is to scale minimum eigen. 1. symmetric and poseigenvalue. The final desired solution is equal to . but excluding the transition where is the region band. . One error (type 1) is given by (16) the other (type 2) is given by (17) occurs at Based on Rayleigh’s principle. form . beamformer. i. Three main differences between conventional eigenfilter and new eigenfilter are listed below. this least squares error measure is equal to the conventional one. two types of least squares errors will be investigated in detail. to the form Step 3: Normalize the solution vector .e. A NEW EIGENFILTER BASED ON TOTAL LEAST SQUARES ERROR CRITERION A. (15) . Thus. ! = III. the least squares error of type 2 is named as total least squares error which have been successfully used to solve many engineering problems such as acoustic radiation problem. the size of the matrix . In the space generated by and for each . we call this least squares minimum eigenvector of matrix design as new eigenfilter design. it is clear that the expression denotes dea hyperplane.

Although the above two facts are not concerned with the performance of designed filters. Fig. where the region [0. we first study the relation between and . we consider the design of high order filters with large transition band. Step 5: Find the minimum eigenvector And. Moreover. case 1 and case 2 [2]. 0. we have Step 6: Use fact 2 to calculate matrix .25 . we obtain (23). Second. we will compare the performance of three least squares methods. It is clear that the design results of least squares method is very bad in the stopband. The proof is completed. From this result.3 . 48. In the following. Step 7: Find the minimum eigenvector to the form And. and the matrix Fact 2: The relation between the matrix is given by There are four cases of FIR filters with exactly linear phase. However. Fig. The least squares solution is obtained using eigenfilter is “inv. and eigenfilters are designed using “eig. we only consider case 1 filter and are given by whose elements of matrix (23) Proof: From (20). C. normalize the solution vector . The elements of are given Step 1: Calculate matrix by Step 2: Compute vector whose elements can be written as Step 3: Find the solution of conventional least squares filter . This example is performed with the MATLAB Language in an IBM PC compatible computer by using the unified design procedure. The final solution is written as . the solution of eigenfilter is given by . 2(a) shows the magnitude responses of 32.m” of MATLAB.4 . but only two of these could be applied to design lowpass filters. but two eigenfilter approaches still work well. because the amplitude must be satisfied exactly response at the reference frequency for the conventional eigenfilter design. and 0. the computamatrix tion of filter coefficients of new eigenfilter is more numerically . The reference point . design. of matrix . and the matrix Fact 1: The relation between the matrix is given by: (21) Proof: From (13). the solution of the new eigenfilter is closer to the solution of the conventional least squares method than the one of the conventional eigenfilter. Step 4: Use fact 1 to calculate matrix of matrix .m. . A Unified Design Procedure . VOL. Example 1: Lowpass Filter Design: Consider the problem of designing a lowpass filter with the following desired amplitude response (25) don't care (22) Substituting (7) into (22). .2 .. This is due to ill conditioning of whose determinant is 6 10 . we develop a unified design procedure to obtain the solutions of three least squares approaches at the same time. that is. it is clear that the specification and are well satisfied for three least squares methods. three least squares approaches with order 0. two experiments chosen in the conventional are made. we obtain (21). 2(b) shows the 148. Then. the stopband attenuation of conventional eigenfilter is slightly worse than the other two methods. Now. The proof is completed.e. NO.702 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: FUNDAMENTAL THEORY AND APPLICATIONS. . magnitude responses of three approaches with order 0. they provide us a way to program three least squares methods in a unified procedure as follows. i. the distances between the solutions of three methods are listed as follows: Thus. Design Example In the example. 6. JUNE 2001 B.” First. we have (24) Substituting (7) into (24). Here. Thus.

and N 32.186 . = = = ] = = stable than that of the least-squares method whose solution is obtained by solving matrix inverse. Because the solution ventional eigenfilter depends on the choice of reference frequency point . The dashed line and dotted line are almost overlap. and N 32. (b) The magnitude response of a lowpass filter with ! 0. (a) The magnitude response of a lowpass filter with ! least squares design (dashed line).4 . so the distance for all reference frequencies.000 572 when fredistance 0. The solid line and dotted line almost overlap. The specification is chosen as ! 0. ! .2 . and the new eigenfilter design by the dotted line. conventional Fig.3 .2 and 0. and N 148.000 267. . ! 0. the conventional eigenfilter design by the solid line. The conventional least squares design is represented by the dashed line.2 . of conFinally. we see that the has minimum value 0.3 . However.25 . 2(c) shows distance for various reference frequencies in the distance when the specification is chosen as 32. From the result.PEI AND TSENG: A NEW EIGENFILTER BASED ON TOTAL LEAST SQUARES ERROR CRITERION 703 (a) (b) (c) = = = 0. new eigenfilter design (dotted line). This means that the solution of the new eigenfilter is closer to the solution of the conventional least squares method than the one of the conventional eigenfilter for all choice of reference frequency . conventional eigenfilter design (solid line). the distance is equal quency is always greater than to 0. ! 0. Fig.3 . (c) The distance a a for various reference frequency points ! in the range [0. a remark is made. 2. it is useful to investigate the relation between and reference frequency . the range = j 0 j 0. ! 0.

48. A typical application of notch filter is to remove the 60-Hz interference in the recording of ECG signal. because QR decomposition has less computation load than SVD. Then. the detail can be found in [5]. the null space of Finally. or iterative power method [16]. The frequency response of an ideal notch filter has unit gain for all frequency except notch frequency in which gain is zero [17]. . we will employ a case-one to design the notch filter.” Based on full rank assumption of . Fact 3: Let singular value decomposition (SVD) [14] of matrix be (29) (26) and diagonal matrix where unitary matrices matrix and is an vector. upper triangular matrix with size . For the conventional least squares filter design. we summarize the entire procedure of the proposed method as follows: Step 1: Use SVD or QR decomposition to find the orand thonormal basis of null space of matrix construct the matrix . The basic idea of solving this problem is as the following form: to rewrite the constraint (27) . Moreover. Three computation issues are stated as follows: First. However. Third. Thus. Second. a zero matrix with size . Moreover. Fact 4: Let QR decomposition [14] of matrix where . NO. JUNE 2001 IV. Now. Then. QR decomposition is a better candidate when both decomposition programs are at hand. In the linear case. the closed-form solution can be obtained by using the well-known Lagrange multiplier method [17]. the procedure to find the solution of the conventional constrained eigenfilter design is slightly complicated. .704 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: FUNDAMENTAL THEORY AND APPLICATIONS. identity matrix. the design problem becomes Minimize the space of . Example 2: Notch Filter Design: In this example. The final desired solution is given by . are with size . Finally. Note that where is an is the number of the linear constraints which is usually smaller than the number of coefficients . we have that is a matrix and is a vector. we will use notch filter design to demonstrate the effectiveness of proposed design algorithm described in the above. [18]. Any of the following two facts will help us to solve it [16]. As to the new eigenfilter design which is our main focus. the remaining problem is how lution null matrix be (30) where unitary matrix Subject to where Then. of matrix Step 2: Find the minimum eigenvector . we are interested in only the minimum eigenvector of symmetric matrix. the desired optimal soeigenvector of matrix is equal to . new constrained eigenfilter design only requires to find the minimum eigenvector of matrix with size . VOL. where the constraint columns of form an orthonormal basis of the null space of matrix . Hence. this computation can be performed efficiently by the conjugate gradient method [15]. for eigen-approach. the FIR filter of even order . the opof this simplified problem is the minimum timal solution . Thus. the general form of the constraints can be stated as to find the orthonormal basis of null space of matrix . the problem is reduced to Minimize Subject to (28) The key step of our method is that “all the vector satisfying can be expressed as . unconstrained eigenfilter design with size needs to find the minimum eigenvector of matrix . NEW EIGENFILTER DESIGN WITH LINEAR CONSTRAINT The main advantage of the least squares approach is that various time and frequency constraints can be incorporated. we obtain . Thus. Step 3: Calculate the optimal solution to the form Step 4: Normalize the solution vector . where is a the problem described in (28) can be simplified as Minimize which is an unconstrained optimization problem. Due to orthonormal condition. 6. Here. we assume is a full rank matrix in order to avoid redundant constraints. constrained eigenfilter has less computation load in searching minimum eigenvector.

it can be shown that the linear constraints can be written in the standard form Note that for all = (32) is a matrix given by (33) . (c) The magnitude response of a notch filter using linear programming method for L 1.PEI AND TSENG: A NEW EIGENFILTER BASED ON TOTAL LEAST SQUARES ERROR CRITERION 705 (a) (b) (c) = = = Fig. L 1 (solid line). 3. L 3 (dashed line). . (a) The magnitude response of a notch filter using new eigenfilter method. = relation of the parameter in (2) is . for for (31) Now. and Moreover. the following constraints are introduced: . to obtain zero gain at notch frequency and control the null width. (b) The magnitude response of a notch filter using Lagrange multiplier method for L 1. the optimal filter coefficient can be chosen such that the is as close as desired response amplitude response defined by After some manipulation. L 5 (dotted line).

55 . NO. it has been reported that the weighted least squares technique will produce an equiripple design if a suitable least squares frequency response weighting function is used [19]–[21]. Let 0. 3]). JUNE 2001 where vectors So far. using proposed method for It is clear that the frequency response is satisfactory. and 1. and the parameter affects the convergence and convergent speed [19]. Fig. When the number of constraints increases. On the other hand.02). 32. Step 4: Use (35) to obtain optimal solution of filter coefand calculate new amplitude response ficients and function . Step 3: Use (36). VOL. we will present an approach to design equiripple eigenfilter based on weighted total least squares error criterion. For the conventional least squares filter design. and various . The resultant ampliters are chosen as tude response after eight iterations is shown in Fig.706 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: FUNDAMENTAL THEORY AND APPLICATIONS.0213. 3(a) shows the magnitude response of a notch filter 0. Step 5: Use (38) to check whether the errors are equiripple. Hence. the overall design procedure can be summarized as follows: Step 1: Specify the initial weighting function as uniform weighting. the linear programming method is a minimax design and Lagrange multiplier approach is a conventional least squares design.45 .0213.5 . the amplitude response of conventional eigenfilter can be made approximately equiripple by employing the similar iterative techniques [5]. the weighted total least squares error can be written as (34) where and are given by (35) Based on Rayleigh’s principle. Once the errors become equiripple. let us see some numerical examples. The design parameeven order 0. then stop iteration. Basically. we employ the case-one FIR filter of and 32.02 and 1. Otherwise. the new eigenfilter method has better frequency response fitting for all frequency band except at the range around the notch frequency . for (39) Step 2: Use (35) to obtain optimal solution of filter coefficients and calculate initial amplitude response and function .5. Finally. (37) to compute new weighting function . the notch width increases accordingly. to design this filter. 48. multiplier and linear programming methods for 0. 6. Although optimum solution can be obtained for any given least squares weighting function. we find the minimum eigenvector of the matrix and scale it to the form . 3(b) and (c) show the magnitude responses of a notch filter using the well-known Lagrange 32. It is clear that our approach is comparable to the MPR method. and go to Step 3. several novel iterative algorithms for deriving the weighting function have been developed such as Lawson’s algorithm and Lim’s method [19]. And. the linear constraints of notch filter design has been formulated as standard form. For comparison. is also depicted in Fig. then the stop criterion is given by (38) where is a small positive number (say 0. This is due to the fact that both of these methods are based on the least squares criterion. desired amplitude response is specified in (25) with 0. i. then the weighting function used in the th iteration may be expressed as (36) where is given by (37) is the amplitude response at th iteration. there is no known analytical method for deriving the weighting which will produce an equiripple design. Now.e. Fig. it is obvious that proposed method almost has the same frequency response as Lagrange multiplier method. we consider the equiripple lowpass filter design whose 0. 4(a). then is the desired optimum solution. the weighting function does not alter anymore with further iteration. Moreover. Compared with equiripple design using linear programming method. let Example 3: Equiripple Lowpass Filter Design: In this example. A NEW EQUIRIPPLE EIGENFILTER DESIGN In the literature. If equiripple design is achieved. From this result. For comparison. From (17).. In this section. If we define and .5 . The peak ripple magnitude of the result is 0. If function used in the th iteration. The is the envelope of the argument function which is formed by joining together all the extremal points of the error function using straight lines (see [19. function is the weighting an iterative procedure is adopted. the designed result using the MPR program. the weighted total least squares eigenfilter can be easily modified to . Fig. 4(b) with peak ripple magnitude 0. V.

i..e. where and are given by (47) are obtained by miniNow. the minimum of corresponding to the minimum the eigenvector of the matrix is also real. The matrix tive definite. NEW TWO DIMENSIONAL EIGENFILTER DESIGN As conventional eigenfilter. sponse (48) (49) occurs at Based on Rayleigh’s principle. VI. the proposed new eigen-approach is easily modified to design two–dimensional (2-D) quadrantally symmetric FIR filters [22]. it is easily shown that the magnitude response of is given by (42) and The dimension (43) (44) where relation between index and are given by if and if and For the sake of convenience. the optimal filter coefficients mizing the total least squares error between the magnitude reand the desired response . symmetric and posieigenvalue. After some and the filter length is manipulation. i. satisfies the half-plane (41) (a) . All computation issue of the new 2-D eigenfilter .e.PEI AND TSENG: A NEW EIGENFILTER BASED ON TOTAL LEAST SQUARES ERROR CRITERION 707 design equiripple FIR filters with arbitrary complex frequency response and in the discrete coefficient space. The 2-D linear phase has a freFIR digital filter with transfer function quency response (40) where the impulse response symmetric condition.. (b) The magnitude response of a lowpass filter using the MPR program. 4. we rewrite lowing vector form in the fol(b) (45) in which the vectors are (46) Fig. (a) The magnitude response of a lowpass filter using weighted total least squares method.

1989. [9] P. (a) REFERENCES (b) Fig. Rabiner and B. vol. Dec. Pei and J. 40. symmetric and positive definite matrix. W. Thus. 1998. JUNE 2001 be satisfied exactly for conventional 2-D eigenfilter method. 1128–1139.. The maximum ripple magnitudes in the passband and stopband are 0. 2122–2130. Huffel. respectively. pp. Its maximum ripple magnitudes in the passband and stopband are 0. Soewito. is same as the 1-D case described in the above. “Least squared error FIR filter design with transition bands. Oct. 1983. 1993. and E.” IEEE Trans. Tufts and J. T. Matrix Computations. and L. Example 4: 2-D Lowpass Filter Design: In this example. However. Van Loan. New York: Academic. [12] S. [16] E. 506–526. R. . this 2-D eigenfilter is also optimal in total least squares error sense. This is owing that the amplitude response at the reference point must [1] J. Baltimore. Signal Processing. “A survey of conjugate gradient algorithms for solution of extreme eigen-problems of a symmetric matrix.” IEEE Trans.. “The design of arbitrary FIR digital filters using the eigenfilter method. 11–23. 1997. Nov. Shyu. Mar. Aug.” IEEE Trans. 46. linear-phase FIR filter designs using the total least squares error criterion have been investigated. Nguyen. Rabiner. [18] T. vol. Linear Algebra and Its Applications. Apr. Philadelphia. 1457–1461. 1991. H. VII. 1327–1340. 1989. [4] C.. P. Speech. 319–329. This topic will be investigated in the future. Nguyen. vol. pp. vol.1041 and 0. Several design examples have been used to illustrate the effectiveness of this new design approach. Shyu. vol. C. 487–494. Feng. [10] S. V. Sarkar. Sept. “Total least mean squares algorithm.708 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: FUNDAMENTAL THEORY AND APPLICATIONS. Vaidyanathan and T. 1980. pp. A. Englewood Cliffs. Huffel and J. “Eigenfilter design of higher order digital differentiators. pp. Press. Multirate Systems and Filter Banks. Symp. and L. Apr. Yang. “Designing notch filter with controlled null width. H. 41. the new 2-D eigenfilter is slightly better than the conventional 2-D eigenfilter. And. Acoust. pp. Theory and Application of Digital Signal Processing. Philadelphia. [13] D. (a) The magnitude response of a 2-D lowpass filter using new eigenfilter method.. it is interesting to study IIR filter designs based on total least squares error criterion. 1987. 1975. [17] M. pp. Gold. Q. J.” IEEE Trans. AU-18. VOL. 1970. CONCLUSION In this paper. NJ: Prentice-Hall. 331–334. pp. Jan. [7] . vol. Pei and J. Circuits and Sysems. Golub and C. Circuits Syst. pp. Signal Processing. R. 5(b) [22]. For comparison. S. vol. Jiao. and R. Bao. vol. Z. it is also easy to incorporate with linear constraints in time and frequency domain. The desired amplitude response is specified by (50) don't care otherwise The resultant magnitude response is shown in Fig. [11] S. 1993. Gopinath. T. pp. vol. [6] S. 0). Er.” IEEE Trans. Francis. but new eigenfilter is not imposed on this constraint. [3] D. 505–511. Signal Processing.” IEEE Trans. C. The filter coefficients are obtained from the elements of the eigenvector corresponding to minimum eigenvalue of a real. “Eigenfilter: A new approach to least-squares FIR filter design and applications including Nyquist filters. (b) The magnitude response of a 2-D lowpass filter using conventional eigenfilter method. “Designing digital lowpass filters: Comparison of some methods and criteria. [14] G. AU-21. NJ: Prentice-Hall. C. 5(a). McClellan. NO. H. Vandewalle.. 37. 1993. Acoust. Video Technol.” Signal Processing. 3. [8] T. T. only linear phase FIR filters are considered here. IEEE Int.” IEEE Trans. Audio Electroacoust.” IEEE Trans.” in Proc. W.. “Design of two-dimensional FIR eigenfilters for sampling-structure conversion. 1550–1556. [2] L.2353. P. 35. Speech. the design result of the conventional 2-D eigenfilter approach is also depicted in Fig. Englewood Cliffs. Recent Advances in Total Least Squares Techniques and Errors-in-Variables Modeling. W. “Unified eigenfilter approach: With applications to spectral/spatial filtering. pp. 24. Vaidyanathan. Arvas. Burrus. vol. V. Moreover. Chen. PA: SIAM. pp. J. June 1992. Signal Processing. 6. CAS-34. 37. 1973. The reference point chosen is zero frequency (0. Signal Processing. Dec. and the new eigenfilter approach is easily modified to design 2-D linear phase filters. PA: SIAM. The Total Least Squares Problem: Computational Aspects and Analysis. Strang. Z. a 2-D circularly symmetric lowpass FIR filter of length (11. 11) is designed using total least squares method. Q. A. 1991.” IEEE Trans. 5. vol.. K. pp. ISCAS’93. 1988. “Design of FIR Hilbert transformers and differentiators by eigenfilters. Audio Electroacoust. Thus. MD: The Johns Hopkins Univ.2253. 48. “A computer program for designing optimum FIR linear phase digital filters. This design is not only optimal in total least squares error sense. Circuits Syst. [15] X. May 1993. Circuits Syst. [5] P. 158–162.” IEEE Trans. the approximation errors can be made equiripple by using iterative weighted total least squares method.1081 and 0. Parks.

C. respectively. Pei and J.O. Taipei. Laakso.O. and R. He received B. Dr.S. 42. 37.. 2504–2509.PEI AND TSENG: A NEW EIGENFILTER BASED ON TOTAL LEAST SQUARES ERROR CRITERION [19] Y. H. “Eigenfilter approach for the design of allpass filters approximating a given phase response. image processing. C. the Distinguished Research Award from the National Science Council from 1990 to 1998. Taiwan. 1994. respectively. Ltd. R. Santa Barbara in 1972 and 1975.S. and Ph. Taiwan. Signal Processing. and electronic commerce. pattern recognition. 1994. from National Taiwan University. all in electrical engineering. Taiwan. in 1949.O. I. Taipei.. pp. R. Yang. vol. Lim. [21] T. Taipei. Shyu. From 1971 to 1975. 24–34. and the National Taiwan University. in 1990 and 1995.O.E. C. Nguyen.C. and the Academic Achievement Award in Engineering from the Ministry of Education in 1998. His research interests include digital signal processing. “2-D FIR eigenfilters: A least squares approach. he was a Research Assistant at the University of California. pp. Signal Processing. 40. degree. Mar. Koilpillai. R.” IEEE Trans. 1990.D.” IEEE Trans. and 1995 to 1998. in 1970 and M. in 1965.D. Taiwan. “A weighted least squares algorithm for quasiequiripple FIR and IIR digital filter designs. vol. Circuits Syst. Q.C.” IEEE Trans. J.S. and Ph.. Chen. 42. Pei received the National Sun Yet-Sen Academic Achievement Award in Engineering in 1984. 709 Chien-Cheng Tseng (S’90–M’95) was born in Taipei. T. He is currently a Professor in the EE Department at National Taiwan University. pp. Sunder and V. From 1995 to 1997. Taiwan. vol.” IEEE Trans. 551–558.O. Taoyuan. respectively.C. Santa Barbara. Taipei.E. He was an Engineering Officer in the Chinese Navy Shipyard from 1970 to 1971. optical information processing. in 1988. He was the Professor and Chairman in the EE Department of Tatung Institute of Technology. He has been the President of the Chinese Image Processing and Pattern Recognition Society in Taiwan from 1996–1998. Ramachundran. Lee. and R.. and the M.E.C. [22] S. from 1981 to 1983. He is currently an Associate Professor in the Department of Computer and Communication Engineering. Sept. 2257–2263. R.E. he was an Associate Research Engineer at Telecommunication Laboratories. the Outstanding Electrical Engineering Professor Award from the Chinese Institute of Electrical Engineering in 1998. His research interests include digital signal processing. and is a member of Eta Kappa Nu and the Optical Society of America. Taiwan. R. K. (honors). from the Tatung Institute of Technology.. degrees from the University of California. Taiwan. 1992. Chunghwa Telecom Company. degrees from the National Taiwan University. H.S. Kaohsiung. vol. . D. pp.. He received the B.. [20] S. “Design of equiripple nonrecursive digital differentiators and Hilbert transformers using a weighted least squares technique. Sept. Jan. C. J. Signal Processing. and laser holography. National Kaohsiung First University of Science and Technology. Taiwan. Soo-Chang Pei (S’71–M’86–SM’89–F’00) was born in Soo-Auo.