IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-22, NO.

2, APRIL 1974 117

ACKNOWLED GMENT [2] —, "A fast Fourier transform for high-speed signal processing,"
IEEE Trans. Comput., vol. C-20, pp. 843—846, Aug. 1971.
[3] M. L. Groginsky and G. A. Works, "A pipeline fast Fourier
The authors wish to thank M. J. Britt for contributions transform," IEEE Trans. Comput., vol. U-19, pp. 1015—1019,
to the initial design of the radix-2 processor (in particular Nov. 1970.
[4] G. D. Bergland, "Fast Fourier transform hardware implementa-
the serial vector multiplier). tions—an overview," IEEE Trans. Audio Electroacoust., vol.
AU-17, pp. 104—108, June 1969.
[5] R. C. Singleton, "A short bibliography on the fast Fourier
REFERENCES transform," IEEE Tran.s. Audio Electroacsust., vol. ATJ-17,
pp. 166—169, June 1969.
[1] M. J. Corinthios, "The design of a class of fast Fourier transform [6] M. C. Pease, "An adaptation of the fast Fourier transform for
computers," IEEE Tran.s. Comput., vol. C-20, pp. 617—623, parallel processing," J. Ass. Comput. Mach., vol. 15, pp. 252—
June 1971. 264, Apr. 1968.

Linear Programming Design of hR Digital Filters with Arbitrary

Magnitude Fnt0
LAWRENCE R. RABINER, MEMBER, m, NANCY Y. GRAHAM, AND
HOWARD D. HELMS, SENIOR MEMBER, IEEE

Abstract—This paper discusses the use of linear programming dure [8] have been recently proposed for designing hR
techniques for the design of infinite impulse response (IIR) digital
filters. One difficulty with almost all of these procedures
filters. In particular, it is shown that, in theory, a weighted equiripple
approximation to an arbitrary magnitude function can be obtained is that convergence of the optimization procedure that
is used to design the filter is not guaranteed, and even
in a predictable number of applications of the simplex algorithm of
linear programming. When one implements the design algorithm, when the procedure converges, the optimality of the re-
certain practical difficulties (e.g., coefficient sensitivity) limit the
sulting filter is also not guaranteed. In this paper a fre-
range of filters which can be designed using this technique. How- quency domain hR filter design procedure is discussed
ever, a fairly large number of hR filters have been successfully
designed and several examples will be presented to illustrate the which uses linear programming techniques (the simplex
range of problems for which we found this technique to be useful. algorithm) to choose filter coefficients to approximate an
arbitrary magnitude characteristic. Theoretical conver-
gence of the optimization algorithm is guaranteed, and
INTRODUCTION the resulting filter can be shown to be optimal in the given

A LARGE NUMBER of techniques are available for

designing infinite impulse response (IIR) digital
design sense (e.g., minimum absolute weighted error). The
optimization algorithm itself has been designed to mini-
filters [1], [2]. The techniques of impulse invariant de- mize the number of applications of the simplex algorithm.
sign, bilinear transformation, and matching poles and The next section presents the design procedure with a
zeros [3], for transforming a given analog filter to an discussion of the practical aspects of implementing the
"equivalent" digital filter are well known and widely method. Following this, several representative filter de-
used. These techniques, however, are limited in that they signs are given. Finally, some discussion is given as to
are generally applied only to the case of transforming practical limitations in using the method.
standard analog filters—e.g., low-pass, bandpass, band- THEORY
stop, or high-pass filters. When one is interested in design-
ing a digital filter with a nonstandard frequency response, Let H(z) be the transfer function of an hR digital
i.e., one which has not been exhaustively studied, then filter. Assume H (z) has the form
some algorithmic (as opposed to closed form) design pro-
cedure is generally used. Several frequency domain (e.g., H(z) = —--- =
/ Nz m az
(1)
[4]—[7]) and one time domain algorithmic design proce- D(z)
where the numerator polynomial N (z) is of rnth degree,
Manuscript received 1)ecember 13, 1973. and the denominator polynomial D (z) is of nth degree.
L. R. Rabiner and N. Y. Graham are with Bell Laboratories, The a0 term in (1) can be set to 1.0 without any loss in
Murray Hill, N. J. 07974.
H. D. Helms is with Bell Laboratories, Holmdel, N. J. 07733. generality. The magnitude response of the filter is ob-
118 IEEE TRANSACTIONS ON ACOUSTICS, SPEECU, AND SIGNAL PROCESSING, APRIL 1974

n
tamed by evaluating (1) on the unit circle (i.e., for m

z = exp [1w]),' and taking its magnitude, thus giving [co + 2c, cos (wi) ]/[d0 + 2d cos (wi)]. (10)
i1
H(exp[3w]) I
= N(exp [1w]) (Again the d0 term in (10) can be set to 1.0 without loss
D(exp[jw]) of generality.) Equation (10) shows that the magnitude
squared function of the filter is a ratio of trigonometric
rn fl
= polynomials. It is also seen that both N (w), the numerator
I b exp [—jwi]/ a exp [—jwi] polynomial, and D(w), the denominator polynomial, are
1=0 i0
linear in the unknown filter coefficients (cd and {d1}. It is
(2) now shown how linear programming techniques can be used
In many frequency domain filter design problems it is de- to determine the c1's and d1's such that H(exp [1w]) 2
sired that the magnitude of the resulting filter approxi- approximates a given magnitude squared characteristic
F (w) where the peak weighted error of approximation is
mate a given magnitude function, ill (exp [1w II), to within
a tolerance G(w,), where G(w,) is a monotically increas- minimized—i.e., the weighted error is an equiripple func-
ing function of for fixed w.2 Thus, the resultant approxi- tion.
mation problem is to choose the filter coefficients (the If we let F(w) be the desired magnitude squared char-
a's and b1's) to minimize the quantity consistent with acteristic, then the approximation problem consists of
that constraint inequality
N(exp [1w]) — M(exp
__________
D(exp [jw])
[3w])
finding the filter coefficients such that

G(w,ô). (3) —€(w) < 1 (w)

€(w) is a tolerance function on the error which
where
—
F(w) <€(w) (11)

Inequality (3) is generally evaluated over a union of dis-

joint subintervals of the band 0 w allows for unequal weighting of errors as a function of
The above approximation problem is a nonlinear one frequency. Since F(w) and (w) are generally specified
in that the filter coefficients enter into the constraint functions of frequency, (or depend on some parameter in
a manner explained below), (11) can be expressed as a
equation nonlinearity. Although various techniques have
been proposed for solving this nonlinear problem (e.g., set of linear inequalities in the ci's and d1's by writing it
[4] and [5]), a linear approximation problem can be in the form
defined by considering the magnitude squared function of I(w) — b(w)F(w) <s(w)(w)
the filter. From (1) we get the relation
—(w) + (w)F(w) <e(w)(w) (12)
N(z) ZT(z1) or
H(z)H(z') (4)
D(z)D(z)
N(w) — D(w)[F(w)+ (w)] <0 (13)
= ( bz)
1=0
(,fl

j=0
b5z+2)/(
fl

1=0
fl
az1) ( ar+)
j=O
—N(w) + D(w)[F(w) — s(w)] <0.
The additional linear inequalities
(14)

(5)
m n
—I(w)<0 (15)
=
1=—rn
cz'/ (6) —(w) <0 (16)
where completely define the approximation problem.
Thus, the question of whether or not there exists a digi-
c = c_ i = 1,2,.. (7) tal filter with magnitude squared characteristic F (w) and
tolerance function e (w) is equivalent to the question of
= d_1 i = 1,2,. .,n. (8)
whether or not there exists a set of filter coefficients satis-
The magnitude squared function of the ifiter is obtained fying the system of constraints defined by (13)— (16). The
by evaluating (6) on the unit circle giving question can be answered by using linear programming
techniques [9], First, an auxiliary variable is subtracted
H(exp [1w]) 2 from the left side of each constraint, forming the new set
of constraints
= H(z)H(z ) z=expjoj
= N(w) (9) N(w) — D(w)[F(w) + €(w)] — <0
D (w) (17)
—N(w) + (w)[F(w) (w)] -- v <0 (18)
1 The quantity is the normalized frequency variable (i.e., the
sampling period, T, is assumed to be 1.0). —N(w) — v < 0 (19)
2 It should be noted that the function G(w,) is generally deter-
mined as soon as M(exp {jw]) is specified by the designer, as will be
seen in the examples later in the paper. — v < 0. (20)
RABINER et al.: DESIGN OF TIE DIGITAL FILTERS 119

The objective function IH(eJw)12

I + K8

zv (21) I—K8

is chosen to be minimized under the constraints of (17)—

(20). Clearly a solution to constraints (13)—(16) exists
if and only if the minimum value of z under constraints S
w
0
(17)—(20) is zero. If the minimum value of v is 0, then a
solution exists to the approximation problem and the filter (a) (b)
coefficients may be obtained directly as the output of the
linear programming routine. If v > 0, then no solution to
the approximation problem exists, and either F (w), or
F(w) C (w)
or both must be modified in order to obtain a
I + K232
solution.
To illustrate the above procedure, consider the design
of a low-pass filter. If we let be the peak approximation
error in the stopband, and Ko (K is a constant expressing
the ratio of passband to stopband ripple) be the peak — — I — _________
approximation error in the passband, then the magnitude 0
0 WpW 7r
Iw
U)5
I)
iT
function for an equiripple error approximation is as shown (C) (d)
in Fig. 1 (a). The quantity is unknown and is to be Fig. 1. Specifications for designing an equiripple error low-pass
minimized in the ultimate design program. (Of course in filter. (a) Bounds on the approximation. (b) Bounds on the square
of the approximation. (c) Function obtained by averaging the
this case the resulting filter is an elliptic filter which can upper and lower bounds on the square of the approximation. (d)
be readily designed in closed form, but we are only using Error bounds obtained by subtracting Fig. 1(c) from Fig. 1(b).
this as an example of how to apply the design technique.)
The passband cutoff frequency is o, = 2irF and the stop- manner. If we let dl" denote the minimum value of for
band cutoff frequency is w, = 2irF. The magnitude which the approximation problem has a solution, then 5*
squared function of the filter is the square of the response satisfies the inequality
in Fig. 1 (a) and is shown in Fig. 1 (b). This magnitude
1
squared function can be viewed as a weighted equiripple (24)
approximation of the function F (w) [shown in Fig. 1 (c) , K+1
with peak approximation error e (o,) [shown in Fig. 1 (d) , since the sum of passband ripple (K*) and stopband
defined by ripple (*) must be less than or equal to 1.0 because
otherwise the passband and stopband would no longer be
F(w)=l+K2l 0<w<o, well defined. Based on (24), a binary search may be used
w<w<1r 22 to locate , as in the following procedure. (This proce-
=o2/2 dure is different from the usual binary search in that the
search is performed on the log of ô instead of itself.)
Step 1: Let and &J denote initial upper and lower
(23) bounds for For example, choose ô+' = (K + 1),
10—8 (since S < 10_8 is unrealistic). Initialize at (56_i) 1/2,
It is easily verified that H(exp [joI) 2 of Fig. 1 (b) is the geometric mean of the initial upper and lower bounds.
less than or equal to F(w) + s(oi) and greater than or
(Note that the geometric mean of two quantities is equal
equal to F(w) — (c) in both the passband and the stop- to the arithmetic mean of the logs of these quantities.)
band.
To determine the smallest value of such that the filter Step 2: Solve the linear programming problem (17)—
approximation problem has a solution (i.e., to find the c (21) with this value of ô. If z 0 a solution to the ap-
of the elliptic filter) an iterative procedure must be used proximation problem exists and &'' < . In this case set
= . Otherwise no solution exists for this value of tI
since 5 enters into the design constraints in a nonlinear
and ô < '; in which case set &..
Step 3: Set l = (oo )1/2 and repeat Step 2. rfhis pro-
The linear programming problem defined above (by constraints cedure is iterated until a predetermined accuracy criterion
(17)—(20) and objective function z = v) may be solved by a straight-
forward application of the simplex (or revised simplex) algorithm. in locating &'' is satisfied.
However, since the number of constraints is generally much larger
than the number of filter coefficients, it is of course much more It should be noted that, when 5* is small (<102) (as
efficient to apply the simplex algorithm to the dual linear program- in practical filter design problems), choosing ô to be the
ming problem. Furthermore, since the range of values of filter co-
efficients, for the examples of interest here, is between —1 and +1, geometric mean of the upper and lower bounds for 5* will
a change of variables to c,' c + 1, d' = d + 1 is performed result in a smaller number of iterations required to achieve
before application of the simplex algorithm, which does not allow
variables to assume negative values. relative accuracy in locating * than required by the usual
 120 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, APRIL 1974

TABLE I
RESULTS ON PER DESIGN OF SEVERAL LOW-PASS FILTERS

No. of Total Run

Filter No. n m F,, F, K —20 log,, ' —20 log,0 &+ Iterations Time (s)

1 4 4 0.30 0.35 5.8 37.9 37.8 11 40.6

2 4 4 0.15 0.18 2.0 29.8 29.9 11 43.5
3 4 4 0.10 0.15 12.0 43.1 42.9 11 44.2
4 4 4 0.10 0.14 6.5 38.7 38.6 11 46.9
5 4 4 0.10 0.13 3.4 34.0 33.9 11 48.2
6 4 4 0.10 0.12 1.7 28.5 28.4 11 50.6
7 6 6 0.20 0.23 8.5 48.3 48.4 11 134.3
8 6 6 0.20 0.25 71.9 61.6 61.8 11 141.7

10
0
—10

—20 nm=6
—30
8 0.0008252
K 7 1.879
-40

00
I— -60
-50
N
z
-70
,0 -80
-J -90
-100
-110
-120

0 0.20 0.5 0.5

NORMALIZED FREQUENCY

Fig. 2. Log magnitude response of a sixth order low-pass filter designed using linear programming methods.

binary search in which 5 is chosen to be the arithmetic lie inside or on the unit circle. (If zeros lie on the unit
mean of the upper and lower bounds. In particular, the circle, only half of the pairs are retained.)
number n of iterations required for a maximum relative The extension of the above procedure to other types of
error p (where p is defined as filters other than low-pass filters is straightforward and
— will not be discussed here. In the next section we present
p= (25) examples of several filters which were designed using the
above iterative procedure.
with +1 and L1 being the final upper and lower bounds FILTER EXAM1LES
determined in step 3 above) can be shown to be of the
form In order to test out the procedure a number of filters
were designed. Table I gives the results for a set of 8 low-
(+) — log,0 (&)\1
log10 pass filters. The data in this table correspond to the 8
n= + 1 (26)
[iog2 (log10 (f) — log15 (&_f))j low-pass filters designed by Swanton [10] using linear
programming methods in a sequence of individual numer-
(si) — log,0 ator and denominator optimization iterations. These data
= 1og2 (si))] + 1 (27) are for low-order filters (4th or 6th order). The values for
[ (low log10 (1 + p) F and F. are the corresponding filter cutoff frequencies
where [y] denotes the largest integer y. where F = w,,/2ir and F, = w,/2r. The quantity 20 log10 *
Theoretically the value of 3* may be bounded as tightly is the theoretical stopband attenuation for the elliptic
as desired by using a large number of iterations thereby filter and 20 log10 & is the actual attenuation for the filter
reducing the tolerance (difference between + and &) at designed using the linear programming technique described
much as desired. In practice, a relative error of p = 0.01 = above. Table I also gives the number of iterations on the
1 percent on the deltas is sufficient for most problems. delta's (for a 1 percent tolerance on ) and the overall
The values of c and d associated with the final value run time on a Honeywell 6000 Computer. it is seen from
of 3÷ are used in the polynomials in z (6), which are then Table I that the resulting filters meet approximately the
factored. A z-transform, which is stable, and minimum same specifications as the equivalent elliptic filter. To
phase is obtained by retaining only zeros and poles which further illustrate these results, Fig. 2 shows the log magni-
RABINER et at.: DESIGN OF hR DIGITAL FILTERS 121

V
z
Ui

F-
2
.5

F,
0-J

0.50
NORMALIZED FREQUENCY

Fig. 3. Magnitude response of an eighth order bandpass filter

designed using linear programming techniques.
(a)
tude response of the eighth filter in the table which
achieved a ô = + of 0.0008144 or a stopband attenuation
of 61.8 dB.
As seen in Table I the run time for the simple 4th order z -PLANE
examples was about 45 s whereas for the two 6th order
examples it was about 138 s. These times could be signifi-
cantly reduced by an improved initial guess of , or by
relaxing the convergence criterion (the 1 percent tolerance
on the deltas). These were not done for the examples in
Table I in order to verify that the procedure would con-
verge without any starting information.
In addition to low-pass filters, several 3, 4 and 5 band (b)
filters, and wide-band differentiators have been designed
with this procedure. Figs. 3—7 show the frequency re-
sponses of some typical filters which were designed. Fig. 3
gives an example of a standard bandpass filter of eighth
order (i.e., n = m = 8). This filter took 11 iterations to
(0
achieve the desired 1 percent accuracy, and required 187 s Ui
-J
a-
of processor time. The value of K was 1.0 giving a value .5
of of 0.0051. The two filter stopbands were from 0.0 to a)
2
0.15 and from 0.35 to 0.5. Fig. 3 shows the magnitude >-
4-J
response of the filter including the tolerances in each of U,
F,
the bands and the filter band edges. a.
Fig. 4 shows the frequency response of a 4 band filter 0
with a stopband followed by 2 disjoint passbands, fol- F,

lowed by another stopband. The filter is of eighth order

and the tolerances in all the bands were the same. It re-
quired 11 iterations for convergence of the delta to within
1 percent and took 238 s of computation. Fig. 4(a) shows 0 .'O 0 30 040
NORMAl 71 (1 FRI UtIINCY
the log magnitude response of the filter; Fig. 4 (b) shows
the positions of the poles and zeros in the z-plane; and (a)

Fig. 4(c) shows the group delay response of this filter. Fig. 4. Log magnitude response, r-plane pole—zero diagram,
and group delay response of an eight order-four band filter.
The pole and zero positions of the filter are obtained by
factoring the denominator and numerator polynomials of
the magnitude squared function of the filter and assign- Fig, 5 shows an example of a 5 band filter with 2 pass-
bands and 3 stopbands. An eight order filter gave a delta
ing poles inside the unit circle to the resulting denominator,
and zeros inside or on the unit circle to the resulting nu- of 0.035 with equal weighing in each of the five bands.
merator. (Generally the zeros of the magnitude squared The design procedure took 11 iterations and required
function will be on the unit circle in pairs. Thus one of 224 s to design.
each pair of zeros on the unit circle is assigned to the Fig. 6 shows another 5 band filter with the arbitrary
resulting filter.) specifications:
 122 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, APRIL 1974

and with equal tolerances on the error in each band. The

resulting filter was of 10th order and had a S of 0.0303.
Even though the filter met specifications in all the bands,
-a the magnitude squared function blew up between the
a fourth and fifth bands because of a pole which was on the
Ui
a
D unit circle. The simple expedient of moving the pole
I—
a
C:,
slightly inside the unit circle gave a stable filter whose
ci
log magnitude response and pole—zero positions are plotted
0
0
-1
in Fig. 5. It should be noted that the presence of a pole
on the unit circle in one of the unconstrained regions of
the frequency scale is not a violation of the design proce-
dure and is perfectly acceptable to the optimization
0.15 0.20 0.30 0.35 procedure. Of course the resulting filter cannot tolerate
NORMALIZED FREQUENCY such a situation.
Fig. 5. Log magnitude response of an eight order filter with 5 Finally, Fig. 7 shows the error response of a differen-
alternating stopbands and pasebands. tiator of fourth degree designed using the technique. In
- this case the error criterion was a minimum relative error
criterion. The resulting value of S was 0.00000763 when
the desired band for differentiation was from 0 to 0.45.
—10 (In this case S represents the peak relative error for the

-D
z
—20 -

—30 -
N / differentiator.) This example required 11 iterations and
took 51 s of computation time.
DISCUSSION
Lii
C
I-. The preceding examples have shown that the linear
z0 —40 -
ci
programming method does, in many cases, give reason-
able solutions for 1111 filters which approximate arbitrary
0
0
—50 -
-j magnitude specifications with arbitrary weighing of the
—60 - error function. In this section we discuss what we believe
are the practical limitations of this technique.
—70 - One of the maj or difficulties with the proposed method
is that one is forced to work with magnitude squared
—80 characteristics to solve for the filter coefficients. Kaiser
aos OiO 0.16 0.20 0.26 0.30 0.36 040 050 [11] has shown that an extreme coefficient sensitivity
NORMALIZED FREQUENCY problem exists for sharp cutoff filters when implemented
(a) in the direct form. This coefficient sensitivity is aggra-
vated by using magnitude squared functions rather than
the magnitude function itself. Thus the results derived
Z— PLANE
by Kaiser, along with our own practical experience indi-
cate the following.
1) Sharp cutoff filters are difficult to design w'ith this
procedure. Thus, if the width of a transition band is small,
the coefficient sensitivity will make the procedure un-
stable.
2) High-order filters are difficult to design. Filters with
order greater than about 12 cannot readily be designed
with double precision arithmetic on a 36 bit word length
computer since the high order polynomial coefficients are
(b)
extremely sensitive to small changes in the filter specifi-
Fig. 6. Log magnitode response and z-plane pole—zero positions cations.
of a tenth order five band filter with arbitrary magnitude speci-
fiCations. 3) Filters with deltas on the order of 10 or less can-
not generally be designed even with double precision
H(exp [j2irfj) = 0.5 0.00 I 0.06 arithmetic since, a tolerance of the magnitude function
on the order of l0— in a band implies a tolerance of
= 0.75 0.10 f 0.16 the magnitude squared function on the order of 10°
= in that band. The attainment of such a high degree of
0.0 0.20 <f < 0.26 numerical precision is of course limited by the precision
= 1,0 0.30 <f < 0.36 capabilities of the computer. Experience has shown that
the linear programming routine required to solve for the
= 0.0 0.40 f < 0.50 ifiter coefficients must be implemented in double precision
RABINER et al.: DESIGN OF lIE DIGITAL FILTERS 123

ERROR FUNCTION

n=rn4
0.00000763 - SO.O0O00763

— 0.00000763

NORMALIZED FREQUENCY
Fig. 7. Error function of a fourth order differentiator with minimum relative error over the band 0 f 0.45.

arithmetic in order to obtain sufficient accuracy for deltas ACKNOWLED GMENT

in the range 1O > 1O, whereas for deltas <10, The authors wish to thank J. F. Heller of Bell Labo-
sufficient accuracy is almost impossible to consistently ratories, Whippany, N. J. and W. L. Roach of Bell
obtain even with double precision arithmetic on a 36-bit Laboratories, Holmdel, N. J. for their comments and
computer. suggestions.
4) Experience has also shown that for filter design
problems with 8 degrees of freedom in numerator and REFERENCES
denominator polynomials, the number of pivots within a [1] J. F. Kaiser, "Digital filters," in System Analysis by Digital
Computer, F. F. Kuo and J. F. Kaiser, Ed. New York: Wiley,
given application of the simplex algorithm may be ex- 1966.
ceedingly large, say >400. [2] B. Gold and C. M. Rader, Digital Processing of Sianals. New
York: McGraw-Hill, 1969.
Fortunately, despite the above practical limitations, [3] H. M. Golden, "Digital filter synthesis by sampled-data trans-
there is a large class of problems where the linear program- formation," IEEE Trans. Audio Electroacoust., vol. AU-16,
ming technique can be used to advantage. One of the key pp. 321—329, Sept. 1968.
[4] K. Steiglitz, "Computer-aided design of recursive digital
properties here is the guaranteed theoretical convergence filters," IEEE Trans. Audio Electroacoust., vol. AU-iS, pp.
and the clear statement of optimality of the resulting 123—129, June 1970.
[5] A. G. Deczky, "Synthesis of recursive digital filters using the
approximation. Also, compilers which permit arbitrary minimum p-error criterion," IEEE Trans. Audio Electroacoust.,
precision arithmetic may circumvent the above limita- vol. AU-20, pp. 257—263, Oct. 1972.
[6] P. Thajchayapong and P. J. Rayner, "Recursive digital filter
tions, although at the cost of greatly increased computa- design by linear programming," IEEE Trans. Audio Electro-
tion time. acoust., vol. ATJ-21, pp. 107—112, Apr. 1973.
[7] F. X. Brophy and A. C. Salazar, "Recursive digital filter
synthesis in the time domain," IEEE Trans. Acoust., Speech,
SUMMARY and Signal Processing, vol. ASSP-22, pp. 45—56, Feb. 1974.
[8] C. S. Burrus and T. W. Parks, "Time domaio design of re-
A new technique for designing hR filters which can cursive digital filters," IEEE Trans. Audio Electroacoust., vol.
approximate arbitrary magnitude specifications was pre- [9] AU-IS, pp. 137—141, June 1970.
H. L. Loeb, "Algorithms for Chebyshev approximations using
sented. The technique sets up a linear programming prob- the ratios of linear forms," J. Soc. md. App?. il/lath., vol. 8,
lem which is solved iteratively for the best approximation pp. 458—465, Sept. 1960.
[101 D. Swanton, "Linear programming design of recursive digital
to the given specifications. Several examples of filters de- filters," M.Sc. Thesis, McGill Univ., Mar. 1973.
signed using this technique were given and the ultimate [ii] J.ofF.linear
Kaiser, "Some practical considerations in the realization
digital filters," in Proc. 3rd Annu. Allerton Conf. on
limitations of the procedure were discussed. Circuit and System Theory, 1965, pp. 621—623.