lEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-32, NO.

3, J U N E 1984 563

Interpolated Finite Impulse Response Filters

YRJO N E W O , SENIOR MEMBER, IEEE, DONG CHENG-YU, AND SANJIT K. MITRA, FELLOW, IEEE

Abstract-A new approach to implement computationally efficient can be achieved in both the linear and nonlinear phase cases.
finite impulse response (FIR) digital filters is presented. The filter struc- However, inthispaperweconcentrateonthe design and
ture is a cascade of two sections. The first section generates a sparse set analysis of linear phase IFIR filters. The reduction of multi-
of impulse response samples andthe other section generates the remain-
ing samples by using interpolation. The method can be used to imple-
pliers also results in reduced coefficient sensitivitiesand round-
ment most practical FIR fiiters with significant savings in the number off noise levels as well. The number of delaysis approximately
of arithmetic operations. Typically 1/2 to 1/8 of thenumber of multi- thesameasin the correspondingconventionalimplementa-
pliers and adders of conventional FIR filters are required in the imple- tion. The effect of interpolating the impulse response can be
mentation. The saving is achieved both in thelinear phase and the non- analyzed easily in the frequencydomain. This makes it possible
linearphasecases. In addition, the new implementation gives smaller
to develop a simple design procedure that only requires the use
coefficient sensitivities and better roundoff noise properties than
conventional implementations. of a standard FIR filter design program.
There are several other methods that utilize the redundancy
I. INTRODUCTION of thetap coefficients to achieve computationallyeffective
FIR realizations. Thinning of the impulse response by remov-

F INITE impulse response (FIR) digital filters are known to

have some very desirable properties like guaranteed stabil-
ity, absence of limit cycles, and linear phase, if desired. The
ing some of the tap coefficients has been proposed by Smith
andFarden [l] . The method gives someimprovement,but
the design of the filters is complicated and the resulting filter
major drawback is the large number of arithmetic operations
is irregular in structurebecauseofnonuniformtapspacing.
needed in the implementation. The number of multipliers in
Theapproachproposed by Boudreaux and Parks [2] usesa
the direct form implementation is the sameas the length of
theimpulseresponsesequence.In the linearphase case the low-order IIR sectionincascadewithuniformly or nonuni-
numberofmultiplierscan be reduced by approximately 50 formly thinned numerator. Dynamic programming is used to
optimize the performance of the cascade. The basic principle
percent. In both cases the number of adders is approximately
the same as the impulseresponselength.However,practical in thisapproach is somewhatsimilarwithour method, the
FIR filters have an impulse response with a smooth,predictable difference being that Boudreaux and Parks use the IIR section
envelope and do not need the generality provided by standard to performtheinterpolation.Due to pole-zerocancellation,
FIR filter implementations. One can remove quite a few im- the overall filter has a finite length impulse response. However,
pulse response samples and easily find their values again with the presence of feedbacks implies that the structure can have
good accuracy using some type of interpolation scheme. limit cycle and overflow oscillations. The method gives good
In this paperwe describe a novel method for FIR filterdesign, results on filterswithsharpcutoffs. A majordifferencebe-
inwhichthisredundancyinfiltercoefficientshasbeenex- tweenthinnedfiltersandIFIRfilters is in the filter design
ploited resulting in significant savings in the number of arith- process. Our method is based on the frequency domain prop-
erties of digital interpolation as opposed to the direct optimi-
metic operations. The basic idea is to implement the filter as
zation approach of theformer.
a cascade of two FIR sections, where one section generatesthe
Another approach to reduce the computational workload is
sparse set of impulse response values with every Lth sample
via multirate filtering [3], [4], where the internal data rate is
being nonzero, and the other section performs the interpola-
altered by usingdecimationandinterpolation.Redundancy
tion. The interpolator can be often implemented with only a
in the tap gains is reducedbymaking the actual frequency
few simple arithmetic operations. The overall implementation,
shaping at a low rate and thus with a filter with a relatively
to be called the interpolated finite impulse response (IFIR)
wide passband. Multirate filtering is effective in narrow-band
'

filter, typically requires approximately 1/Lth of the multipliers

and wide-band applications. The overall structure is, however,
and adders of a conventional equivalent FIR filter. This saving
more complex than an IFIR filter. As the internal data rate in
IFIR filters is constant there is no danger of internal aliasing,
Manuscript received October 25, 1982; revisedJuly 19, 1983. This
work was supported in part by the National Science Foundation under which is one of the major design considerations in multirate
grants ENG 79-18028 and ECS 82-183 10, and in part by a University of filtering.
California MICRO researchgrant with matching supportfrom Intel In a recent paper [5] Adams and Willson describe an FIR
Corporation and Rockwell International.'
Y. Neuvo is with the Department of Electrical Engineering, Tampere structure composed of a recursive running sum prefilter fol-
University of Technology, Tampere, Finland. lowed by an FIR amplitude equalizer. The prefilter produces
C.-Y. Dong is with the Department of Radio-Electronics, Beijing several zeros on the unit circle without any multiplications.
University, Beijing, People's Republic of China.
S. K. Mitra is with the Department of Electricaland Computer The amplitude equalizer is used to make theoverall filter meet
Engineering, University of California, Santa Barbara, CA 93106. the passband andstopband specifications.The method is

0096-3518/84/0600-0563$01.00 0 1984 IEEE

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on February 05,2025 at 07:24:54 UTC from IEEE Xplore. Restrictions apply.
 564 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL.
ASSP-32, NO. 3, JUNE 1984

applicable to low-pass and high-pass filters with narrow pass-

bands. A moderate saving in the number of adders and multi-
H,(zL) I h;(n)
G(z)

pliers is achieved over conventional FIR filters at the expense Fig. 1. Interpolated finite impulse reponse filter.
of a small increase in the number ofdelays.

IMPULSERESPONSEFILTERS
11. INTERPOLATED
Let us consider a digital filter HM(z) with impulse response
hM(n). We call this the model filter as it will determine the
frequency behavior of the final interpolated impulse response
filter.
If we insert L - 1 zero-valued samples between the original
samples of h M ( n ) ,we obtain the sequence h h ( n )
0 0.5
hM(n/L) n=iL,i=0,*1,+2;..
hb(n) = (1)
otherwise.

The z transform h i ( n ) of (1) is

Hn; (z) = H M ( Z L ) . (2)
/H,(ejLW)l
Thus, the implementation of Hn; (2) is simply obtained from
the implementation of HM(z) by replacing each delay with L n
0 0.5
delays. To generatetheinterpolated impulseresponse we hi(d IHi(eiw)l
cascade H M ( z L ) with an interpolator G(z) as shown in Fig. 1.
The overall frequency response Hi(z) is thus 0 . 2 5 L n f
0 0.5
Hi(z) = H M ( z ~G(z).
) (3) (b)
Fig. 2. Schematic forms of the signals in IFIR filters.(a) Low-pass,
The design of C(z) is most conveniently performed in the fre- L = 2. (b) Bandpass, L = 5.
quency domain. Note that the frequency response ofHh(eiw)
is periodic with a period of 2nl.L. Any of thepassbands in the 111. DESIGNPROCEDURE
interval [0, T ] can be selected to be the desired one. The pur-
Selection of L
poseof the interpolator is thustoattenuatetheunwanted
replicas of the desired passband below the prescribed level S 2 The largest value for L , LMAX,that can be used depends on
of the overall filter.Timeandfrequencydomainbehavior the specificationsof theIFIR filter. We actuallywant the
of the different signals in the various stagesof implementa- model filter to meet the passband and stopband attenuation
tion are illustrated in Fig. 2. It is important to note that in requirements. If the stopband edge frequency of the low-pass
the interpolatedimpulseresponsefilter,thepassbandand IFIR filter is denoted by usL,the maximum value for L is
transition bandwidths are 1/Lth of the corresponding widths
of the model filter.
Let us assume that we want the gain of Hi(z) of (3) to ap-
proximate 1 inthe passbandwithamaximumdeviation of where thebracketsdenotetruncation.Equation (5) ensures
S1 and zero in the stopband with maximum deviation of S 2 . that the model filter stopband edgeis less than T . For high-
The requirements of the overall filter are thus pass fdters, we replace us^ by T - WSH, where W ~ His the
stopband edge frequency of the highpassfilter.Theabove
1 - S 1 < /HM(e'LW)G(eiw)]< 1 + S 1 in the equation can also be used for bandpass filters if the passband is
passband
(4a) centered around r / k , where k is a positive integer. In this case
jHM(ejL ")G(eiU) I <a2 stopband.
the
in is replaced by (usl- ws2)/2,where u s l and os2 are
the desired stopband edge frequencies. For generalbandpass
(4b)
cases L M A Xis the largest L that makes us1 and us2fit into
Our main interest is in those cases where HM(eiL ") to a great the same division of T/L.
extent determines the passband and the transition band behav- In practice it is recommended that a somewhat smaller value
iors. In addition, we assume that the passbands of HM(eiL ") than L M A X be selected as the requirements for the interpolator
are reasonably well separated. Themagnitude of G(eiw)should become otherwise more stringent if it has to select one of two
have only a slight effect on the passband shape, which if neces- very close passbands of H M ( ~ ~ ) .
sarycan be compensatedbypredistortingthespecifications
for HM(z). We also require HM(z) to provide thestopband. Interpolator Structures
attenuation (4b). This makes the requirements for G(z) very The design of the interpolator G(z) can now be formulated
mild. G(z) has just to provide the attenuation of the unwanted as the design of a multistopband FIR filter having passband
replicas
Authorized of use
licensed thelimited
desired passband.
to: Malaviya matching
National Institute of Technology Jaipur. Downloaded the wanted
on February 05,2025 atpassband of H
07:24:54 UTC fromMIEEE
( ~Xplore.
and
~ )stopbands
Restrictions over-
apply.
 NEUVO et al.: INTERPOLATED FIR FILTERS 565

lapping the unwanted replicas of the passband. The McClellan- produces one zero on two of theundesired passbands assuming
Parks algorithm [6] can be used to solve this design problem thatthe desiredpassband is centeredat w o . The third un-
However, a simplermethod of choosingalow-order FIRsection wanted passbandcanbe attenuated by using Gz(z). In the
with one or more zeros at the desired stopbands often gives low-pass case, G,(z) of (1 1) interpolates linearly the center
quite good results, This method also provides a saving in the value and G,(z) of (9) generates the first and third value for
number of arithmetic operations needed to implement G(z) if the impulse response of H M ( z ~having
) three zero-valued sam-
L can be factored out into a product of small integers. In this ples between the samples of H M ( ~ )Extension
. of the method
case we can effectively utilize the periodic nature of the fre- to larger valuesof Lis straightforward.
quency response of digital filters to generate several stopbands
at right positions. Design Steps
Let the passband of the overall filter H j ( z )be centered atw o . The design of IFIR fdters can now be summarized as follows.
For L = 2, H M ( . ~has
) passbands centered at wo and n - w o 1) From the given filter stopband edge frequencies, calculate
where we have assumed that w o < n/2. The simplest possible LMAXand select a suitable L < L M A X . After L hasbeen
form of G(z) for low-pass implementationis given by selected, the positions of the unwanted repetitionsof the pass-
Go(z)= $(l t z - l ) band are known.
2 ) Design the interpolator C(z) to attenuate these repetitions
which has a zero at z = - 1. With this interpolator, the impulse of the passband to orbelow the stopband level.
response of the composite structure is 3 ) Design the model filterHM(z). The band edge frequencies
hi(n) = $ [h&(n)+ h&(n - I)] . (7) of the model filter are obtained by multiplying the edge fre-
quenciesof H M ( Z ~in) theinterval [ 0 , (n/L)]by L.The
Hence, this is the zero-order interpolator with L = 2. For high- model filteramplitudespecificationscannow be calculated
pass cases, one can use from the specifications of the IFIR filter by compensating for
GI ( z ) = $( 1 - z - >.
~ (8) the effect of the interpolator G(z). If the passband is suffi-
cientlynarrow, nocompensation is required.The design of
For bandpass cases, as well as for better performance in low- the model filter can be done using any FIR filter design pro-
pass and high-pass cases, a second-order FIR interpolator gram. We used the program of McClellan andparks [6], which is
very easy to modifyfor the variablepassband and stopband
(9) speifications.
The above procedure is for narrowband low-pass, bandpass,
can be used. Gz ( z ) has a zero pair on the unit circle at n rt w o. and high-pass filters, To design wide-band low-pass filters as
Here well as bandstopfdters,wecanfirst design a narrow-band
IFIR filter Hj(z) which is the complementary fdterof the final
1
hi(n) = - [h&(n)+ 2 cos w o h&n - 1) t h&(n - 2)] . (I 0) filter H(z) [7]. The realizationfor H ( z ) then acquires the
K form
Note that wo is the center frequency of the desired passband. H(z) = z-@-1)/2 - Hi(Z) (1 2)
If wo = 0 and K = 4 we get the linear interpolator as can be
seen from (10). The realization of Hi(z) as shown in Fig. 2(a) where the length of the fdter.is odd.
is for linear interpolation.
For L = 3, the passbands of H M ( z ~are) centered at wo and IV. COMPUTATIONAL
SAVINGS
( 2 ~ 1 3rt) wo where w o is the passband in the interval [0, n/3]. We now compare the amount of computations required by
For low-pass, high-pass, and bandpass cases with w o = n / 3 or the IFIR filter with that of an equivalent conventional FIR
2n/3, the interpolator of (9) can be used. For more general filter meeting the same frequency domain specifications. Two
bandpass cases, one can use two sections of (9) in cascade to examplesof the IFIR implementations areshownin Fig. 3.
attenuate the two unwanted passbands. If the desiredpass- Assuming that the hardware needed to implement the inter-
band has a center frequency close to n/2,the use of L = 2 can polator is small, we notice that the number of multipliers and
be difficult or impossible as the two passbands of H M ( z 2 )are adders in the structure is approximately the sameas in the
close to each other and may in fact overlap. In this case the model fdter. In IFIR filter the passband and stopband gains
use of L = 3 provides good separation as the center frequen- are the same as in the model filter, but the passband and stop-
cies will be approximately at n/6, n/2, and 5n/6. Correspond- band widths are only 1/Lth of those of the model filter. Thus,
inglytherecanbeproblemsimplementingbandpassfilters the effect of the interpolation of the impulse response is to
with center frequencies close to n/3 or 2n/3 if L = 3 . Then shrink the passband and transition bands without any signifi:
either L = 2 or L = 4 should be used. cant increase in the number of arithmetic operations.
In the case of a very narrow passband the design may call for The length of an FIR filter required to meet given specifica-
a large value of L. In this case it is recommended that L be tions is approximately [8]
selected as composite number and then build the interpolator
insteps.Thus,for L = 4, the use of two cascadedsections N=
- 20 lOg,o a-
13 +

with L = 2 is preferred. For the first section the use of 14.6 AF

1 where ti1 and 8, are the passband and stopband ripples, and
G3(Z) = - (1 -t 2 cos 2w0z-2 -t 2 - 4 ) (1 1) AF isonthe
Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded relative
February transition
05,2025 at 07:24:54bandwidth
UTC from IEEE Xplore. Restrictions apply.
K
 566 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING,
VOL.
ASSP-32, NO. 3, JUNE I 984

A~=LNM-L. (1 7)
If the number of zeros per unwanted passband is increased by
a factor of P,the numbers J i n (14) and (16) get multiplied by
P. However, for relatively narrow-bandfilterspecifications,
NM tends to be large and the requirements for the interpolator
at the same time can be met with very few arithmetic opera-
tions. Thus, the IFIR filterrequiresapproximately 1/Lth of
the multipliers andaddersofacorrespondingconventional
FIR filter.
Themodelfilterandtheinterpolatorblockscan also be
merged into one block in the implementation of an IFIR filter.
Fig. 4 shows this alternative structure of a low-pass IFIR filter
with L = 2. The number of multipliers is still the same as in
the structure of Fig. 3(a) but the number of addershasin-
creased as well as the number of multipliers with coefficients.
(The transpose of the structure of Fig. 4 may find use in CCD
implementations, for example.)

V. FINITEWORDLENGTHPROPERTIES
OF IFIR FILTERS

Both the coefficient sensitivities and output roundoff noise

properties of the lFIR filter are next analyzed and shown to
be better than those of the conventional FIR fdter.
An upper bound for the standard deviation of the error in
the frequency response due to coefficient quantization is given
by [91
(b)
Fig. 3. IFIR filter structures. (a) Low-passwith L = 2. (b) Bmdpass
with L = 4 and two-stage interpolation.

where Q is the quantization step size and the filter length N is

0 s - 0 p
AF=-. odd.
2n
IntheimplementationoftheIFIRfilter, wecansafely
Thus, theIFIR filterrequiresapproximately 1/Lthofthe assume that the interpolator does not increase the coefficient
multipliers and adders of an equivalent conventionalFIR filter. sensitivity. This is a reasonable assumption as in the neighbor-
In addition t o neglecting the effect of the interpolator, this hood of the passband, where the gain of the interpolator is
comparisondoesnottakeintoaccountthe oscillatoryper- approximately one, the distances to the zeros of the interpo-
formancecharacteristicsofoptimal FIR filters. As thiscan lator are large. Closer to the zerosoftheinterpolatorthe
slightly favor either the IFIR or the conventional FIR filter, required stopband attenuation is exceeded so much that the
comparison based on (13) doesillustrateadequately the coefficient sensitivity is no problem. On the unwantedreplicas
computational savings. of the passband, thr: attenuation caused by the interpolator is
If L = 2 J , J = 1, 2, . . . , the total number of multipliers for easy to check using quantizedinterpolatorcoefficients, if
the type of structure shown in Fig. 3(b) is necessary. Based on these arguments it is sufficient t o analyze
only the effect of quantizing the coefficient of f i ~ ( z ~It )is.
easy to see that the Coefficient sensitivities of H M ( z ~will)
have the same upper bound for the standard deviation as the
where NM is the length of the modelfilter. A total of Jmulti- model fdter H M ( z )has.
4
pliers in the interpolator take a value of for low-pass filters. If we now compare the standard deviation ( a c o ~ v of ) the
In the calculation of the number of the multipliers the gain coefficient quantization error of a conventional filter of length
constants of the interpolator subsections are merged with the LN with that of an IFIR fdter (uIFIR) derived from a model
tap gains of the actual fdter. A comparable frequency response filter of lengthN, we find, using (18), that they are related as
can be obtained with a linear phase conventional FIR filter.
The number ofmultipliers in the conventional filter is.then ~ C O N V L~IFIR (1 9)
for large N.
The output roundoff noise variance (IJCONV) at the output
of a conventional filter of length
L N is approximately
The number A I of two input adders in the IFIR filter and&
in the conventionalfilter are L N Q2
AI=NM-1+2J (1 6 )
%ONV NN -2 -12

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on February 05,2025 at 07:24:54 UTC from IEEE Xplore. Restrictions apply.
 NEUVO et al.: INTERPOLATED FIR FILTERS 5 67

FREQUENCY RESPONSE IN DE

iI
-10.0

-20.0,

-30.0 .-
--
3 -40.0
-.
*C -50.0

--
$3 -eo.o
-*
-70.0
C
Fig. 4. Low-pass IFIR fiiter with L = 2. The linear phase interpolation --
is built into the structure. 3 -eo.o
-so.o t
if N is reasonably large. An IFIR filter derived from a model
filter of lengthof N has the outputnoise variance 0.i 0.2 0.3 0.4 0.5

N o r m a l i z e df r e q u e n c y i n rad/sec

(a)

where g(i) are the impulse response coefficients of the inter-

polator and C comes from the noise generated in the inter-
polatorsection. If themodel filter is designed to have a
peak magnitude of ,one in the passband and the interpolatoris
O.

-20.0
,
1 \
FREQUENCY RESPONSE IN DE

scaled so that the maximum gain of the IFIR filter is one in

-30.0
the passband, itis very easy to see that the summation ofg 2 ( i )
produces a number that is less than one in all practical IFIR
filters.In the low-pass case, with G,(z) of (9) as the inter- C -50.0
polator and K = 4, the summation over g 2 ( i ) yields 3/8. In $ -60.0
the low-pass case with L = 4 the cascade of (9) and (11) yields
11/64. If the interpolator of (9)is used for the bandpasscase,
K = 4 cos wo, and the noise gain of the interpolatoris less than
one if oo< 0.37 IT or wo 2 0.63 R. .In the low-pass case C = -90.0 t
2.5 if a scaling multiplier of one half is in front of the inter- I I
I 4
I
I
I
I I
polator and if we assumethat bothit and the multiplierhaving
0.1 0.2 0.3 0.4 0.5
the value 1/2 produce the same noise as the other multipliers.
N o r m a l i z e df r e q u e n c yi nr a d / s e c
In bandpass cases,as wo increasesfromzero, C increases
gradually and is 4 at wo = r/3. Similar behavior is observed if (b)
wo decreases from R.
FREQUENCY RESPONSE IN DB
For large N and assuming that the noise gain of the inter-
polator is one we get
UCONV LUIFIR. (22)
This is often a somewhatpessimistic estimate for the IFIR filter --
-30.0
as the noise attenuation of the interpolator more than com-
.-
pensates for the extranoise produced by the multipliers in the B -40.0
interpolator. C -50.0 --
.*
--
VI. DESIGNEXAMPLES 23 -eo.o
We have designed several IFIR filters with varying specifica- :C: -70.0
.-

--
tions. Some illustrative examples are given next. The different -eo.o 5
steps in the design process are shown in Fig. 5. Fig. 5(a) shows
for comparison an optimal linear phase low-pass reference fdter
withpassband edge at fp = 0.0404, stopband edge at f, = 0.1 0.2 0.3 0.4 0.5
0.0556, and length 99. The passband and stopband have been
N o r m e l i z e df r e q u e n c yi nr a d / s e c
designed with equal weights. We decided to use L = 2 for the
(c)
IFIR filter. The model filter oflength 49 is shownin Fig. 5(b).
Fig. 5. Low-pass example withL = 2. Frequency responses of (a) refer-
It has the same stopband attenuation as the reference filter. ence filter of length 99, (b) model fdter of length 49, (c) IFIR filter
The shape of the passband has been predistorted to compen- of length 99.

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on February 05,2025 at 07:24:54 UTC from IEEE Xplore. Restrictions apply.
 568 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING,
VOL.
ASSP-32, NO. 3, J U N E 1984

FREQUENCY RESPONSE IN O B
sate for the effect of the interpolator, although the need for
compensation is questionable.It can be seen from Fig.5(c) 0 .

that the design meets essentially the same specifications as the -5.0

reference filter. An IFIR filter where both the passband and -10.0
stopband of the model filter have been predistorted to com-
-15.0
pensate for the effect of the interpolator is shown in Fig. 6 .
Comparing the response of Fig. 5(c) with that ofFig. 6(b) shows D - 2 0 . 0 Q

that there is no significant difference between them. The im- C -25.0 I

plementation for these IFIR filters is shown in Fig. 3(a). The d -30.0
referencefilterrequires 50 multiplications,98additionsand 3

98 delays. TheIFIR filtersrequire 26 multipliersofwhich

:: -35.0
c
one has the value of 0.5, 50 additions, and 98 delays. -40.0

As an example of bandpassdesign, we considered a filterwith -45.0

passband edge frequencies offpl = 0.03, f p 2 = 0.05, and stop-

band edge frequenciesof = 0.02, f s z = 0.06. The ripple
0.1 0.2 0.3 0.4 0.5
weights are 2, I , and 2. The optimal linear phase FIR filter of
Normalized f r e q u e n c y in r a d / s e c
length 111 is shown in Fig. 7(a) and the corresponding IFIR
filter of the same length is shown in Fig. 7(b). The IFIR filter (a)
has been designed using a model filter of length 27. The un-
wanted passbands have been attenuated by one zero at the cen- FREQUENCY RESPONSE IN OR
ter of eachpassband. The implementation is shown in Fig. 3(b).
The number of multipliers, adders, and delays in the reference
-10.0
filterare 56, 110, and 110, respectively.The IFIR filterre-
quires 18 multiplications, 30 additions, and 110 delays. -20.0

The final example is a very narrow-band low-pass filter taken -30.0

from [2]. The passband edge is at fp = 0.001 and the stopband -40.0
--
Q

edge at f, = 0.025. Thepassbandand stopband have equal DC - 5 0 . 0 --

weights in the design. A reference fdter of length 65 is shown II(

-60.0
--
in Fig. 8(a) and a corresponding IFIR realization derived from 3
--
a length 9 model filter using L = 8 is shown in Fig. 8(b). The -70.0
C
IFIR filter exceeds the performance of the reference filter. In 01 --
$
this case the reference filter requires 33 multipliers, 64 adders,
and 64 delays. The IFIR filter has a three stage interpolator
and requires eight multipliers, of which three have a value of
0.5, 14 adders, and 78 delays.
-so. 0

t
0.1 0.2 0.3 0.4 0.5

All of the design examples are summarized in Table I which N o r m a l i z e df r e q u e n c y i n r a d / s e c

also includes the resultsofathinnedfilter design [2]. Note (b)

that the thinned filter gives a comparable result to the IFIR Fig. 6 . IFIR filter of length 99 with both passband and stopband com-
filter.Thisindicates thatbothmethods effectively remove pensated for the attenuation of the interpolator. (a) Model filter. (b)
IFIR filter.
the redundancy from the tapcoefficients.
An exact comparison with the low-pass fdter of Adams and
Willson [5] has not been included. They achieved aproximately VII. CONCLUDINGREMARKS
reduction in thenumber of arithmeticoperations. As Based on interpolation of the impulse response an efficient
their stopband edge frequency was at os= 0.14 71, the IFIRim- new method for implementing FIR filtershasbeen derived.
plementationcoulduse,forexample, I, = 4 resulting in a The IFIR filter is a cascade of a prefilter with sparse impulse
much more efficient implementation. responseandafrequencyresponseperiodicwith 27rIL (L
We have also implemented IFIR filtersusingIntel 2920 signal integer) followed by an interpolator section filling the missing
processor [lo]. For example an IFIR filter of length 37 with impulse response samples and attenuating the
unwanted
L = 2 requires 80 instructions with seven bit coefficients. A repetitions of the desired passband. A simple design procedure
correspondingconventionalFIRfilterthatmeetsthesame has been advanced based on a frequency domain analysis.
specificationsrequires 116 instructions.TheIFIRfilter has It hasbeenshown that IFIR filtersrequireapproximately
about half of the arithmetic operations of the FIR filter. In l/Lth of the adders andmultipliers and, in addition, have 1/Lth
addition,due to smaller coefficient sensitivities, multipliers of the output roundoff noise level and l/$th of the coeffi-
could be expressedwith one bit smaller accuracy. The IFIR cient sensitivity of an equivalent conventional FIR filter.
filter had 4 dB lower output roundoff noise level than the FIR Comparisonwith two recursive implementationsof FIR
filter. filtersindicatedthatIFIRfdters are at least as effective in

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on February 05,2025 at 07:24:54 UTC from IEEE Xplore. Restrictions apply.
 NEUVO e t a l . : INTERPOLATEDFIRFILTERS 5 69

FREQUENCYRESPONSE I N OR FREQUENCYRESPONSE I N OE

0 .

-20.0 1I I
-10.0

-20.0
n
t
-30.0 t
C -50.0

E -70.0

:
-90.0 I I I I
I
gC
-60.0

-90.0
t
L
0.1 0.2 0.3 0.4 0.5 0.1 0.2 0.3 0.4 0.5

N o r m a l i z e df r e q u e n c yi nr a d / s e c N o r m a l d z e df r e q u e n c yi nr a d / s e c

(a) (a)

FREQUENCY RESPONSE I N O R FREQUENCY RESPONSE I N O R

--1o.o

-20.0

-30.0

3
C -50.0 C -50.0
.*I

2 -60.0
3
E -70.0
C
9
: z
-80.0

-so. 0 -so. 0
I ---4--- t
0.1 0 .2 0.3 3.4 0.5 0. i 0.2 0.3 0.4 0.5

N o r m a l i r e df r e q u e n c yI nr a d / s e c N o r m a l i z e df r e q u e n c yi nr a d / s e c

(b) co)
Fig. 7. Bandpass example. (a) Conventional linear phase filter of length Fig. 8.Narrow-band low-pass example. (a) Conventionallinearphase
111. (b) Corresponding IFIR fiter of the same length with L = 4. fiiter of length 65. (b) IFIR fiter of length 79 derived from length 9
model with L = 8.

TABLE I
DESIGN
COMPARISON OF EXAMPLES

Filter Passband Stopband Ne. of No. of No. of

Specifications Implementation Error (dB) Error (dB) Delays Additions
Multiplicationsa

low-pass optimal FIR [Fig. 5(a)] -0.203 98

-32.9 98 50
fb = 0.0404 IFIR, L = 2 [Fig. 5 ( ~ ) ] -0.321 50
-29.1 98 25 + (1)
f,= 0.0556 IFIR, L = 2 [Fig. 6(b)] -0.412 50
-32.7 98 25 + (1)
bandpass optimal
FIR [Fig. 7(a)] -1.544 -27.8 110
56 i 10
fpl = 0.03, fp2-Z 0.05
fsl = 0.02,& -
0.06 IFIR, L = 4 [Fig. 7(b)] 18 -1.646 30-27.3 110

narrow-band
low-pass FIR
optimd [Fig. 8(a)l -37.8
-0.111 33 64 64
fp = 0.001 IFIR,L = 8 [Fig. 8(b)] -0.0358
78
-36.8 5 14 + (3)
t0.885 [2] thinned
fs = 0.025 -34.8 NA NA 6

aNo. or multipliers of value 1/2 is given in parentheses.

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on February 05,2025 at 07:24:54 UTC from IEEE Xplore. Restrictions apply.
 570 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSp-32, NO. 3, JUNE 1984

terms of the required numberofarithmeticoperationsand Yrjo Neuvo (SY70-M’74-SM’82)was born in

naturallyeliminate the finitearithmeticproblemsassociated Turku, Finland, on July 21, 1943. He received
the Diploma Engineer and Licentiate of Tech-
with the recursive sections. nology degrees from the Helsinki University of
The IFlR concept can be extended to two-dimensional FIR Technology, Helsinki, Finland, in1968and
fdters with the same types of advantagesas achieved in one 1971, respectively, andthe Ph.D. degree in
electrical engineering fromCornell University,
dimension [ l l ]. Extension to IIR filterscan also bemade. Ithaca, NY, in 1974.
In IIR filterstheorderof alow-passfilter (Butterworth, He held various research and teaching posi-
Chebyshev, or elliptic) meeting given passband and stopband tions at the Helsinki University of Technology,
the Academy of Finland, and Cornell Univer-
specifications is a function of the transition ratiok sity, from 1968 to 1976. Since 1976 he has been a Professor of electri-
cal engineering at the Tampere University of Technology,Tampere,
tan U- P
Finland. During the academic year 1981-1982 he was with the Univer-
k =-
2 sity of California, Santa Barbara as a Visiting Professor. His main re-
(23) searchinterests are in the areas of digital filters and microcomputer
tan -
U S
systems.
2 Dr. Neuvo is a member of PhiKappa Phi and the Finnish Academy of
Technical Sciences,
where upand osare respectively the passband and stopband
edge frequencies [12]. According to (23) we canmakethe
filter narrower without any significant change in the number
of arithmetic operations if we keep the ratio w P / o sconstant.
Thus, considering the number of arithmetic operations, there Dong Cheng-yu graduated in 1964fromthe
seems t o be no major advantage obtainablein using interpolated Department of Radio-Electronics, Beijing Uni-
IIR (IIIR) filters derived from conventional IIR filters. How- versity, People’s Republic of China.
Since 1964 he has been with the Department
ever, the use of IIIR fdters may result in reducing the finite of Radio-Electronics, Beijing University, in
wordlength effects. teaching and research positions. His research
has involved signal detection, laser, sonar, and
ACKNOWLEDGMENT radar systems. He was a visiting research scholar
at the University of California, Santa Barbara,
The simulation of fdters using Intel 2920 signal processor from 1980 to 1982. Atpresent he is engaged in
was carried on equipment donated by Intel Corporation. research in digital signal processing.
Mr. Dong is a member of Beijing Institution of Electronics.
REFERENCES
[ l ] M. V. Smith and D. C. Farden, “Thinning the impulse response of
FIR digital filters,” inProc. I981 IEEEInt. Conf Acoust., Speech,
Signal Processing, Atlanta, GA, Mar. 1981, pp. 240-242.
[2] G. F. Boudreaux and T. W. P.arks, “Thinning digital filters: A Sanjit K. Mitra (S159-M’63-SM’69-F’74) IB
piecewise exponentialapproximationapproach,” IEEE Trans. ceived the B.S. (Honors) degree in physics from
Acoust., Speech, Signal Processing, vol. ASSP-31, pp. 105-113, the Utkal University, India, in 1953, the M.Sc.
Feb. 1983. (Tech.) degree in radio physics and electronics
[3] M. G . Bellanger, J. L. Daquet, andG. P. Lepagnol, “Interpolation, from the University of Calcutta, India, in 1956,
extrapolationandreduction of computation speed in digital andthe M.S. and Ph.D. degrees in electrical
filters,” IEEE Trans. Acoust., Speech, Signal Processing, vol. engineering from the University of California,
ASSP-22, PP. 231-235, Aug. 1974. Berkeley, in 1960 and 1962,respectively.
[4] R. E. Crochiere and L. R. Rabiner, “Optimum FIR digital filter Hejoinedthefaculty of the University of
implementations fordecimation, interpolation and narrowband California, Davis in 1967 and transferred to the
filtering,” IEEE Trans. Acoust., Speech, Signal Processing, vol. University of California Santa Barbara campus
ASSP-23, pp. 444-456, Oct. 1975. in 1977 as a Professor of Electrical and Computer Engineering. From
[5] J. W. Adams and A. N. Willson, Jr., “A new approach t o FIR July 1979 to June 1982 he served as the Chairman of the Department
digital filters with fewer multipliersand reduced sensitivity,” IEEE of Electrical and Computer Engineering at UCSB. He is a Consultant to
Trans. Circuits Syst., vol. CAS-30, pp. 277-283, May 1983. the Lawrence Livermore National Laboratory, Livermore, CA, and has
[6] J. H. McClellan, T. W. Parks, and L. R. Rabiner, “A computer served asa consultant totheAmpex Corporation,Fairchild Semi-
program for designing linear phaseFIR filters,”IEEE Trans. Audio conductors, GM Delco Electronics, Siliconix, the U.S. Army, and World
Electroacoust., vol. AU-21, pp. 506-526, Dec. 1973. Bank. He has held visiting appointments at the Indian Institute of
[7] R. M. Golden, “Digital filters,” in Modern Filter Theory and Technology, New Delhi, India, Kobe University, Japan, the University
Design, G. C. Temes and S. K. Mitra, Eds. New York: Wiley, of Erlangen-Nuernberg, West Germany,and the AustralianNational
1973, ch. 12. University, Canberra. Be has published a number of papers in active
[8] J. F. Kaiser, “Nomecursive digital filter design using the Io-sinh and passive networks and digital filters, and is the author of two texts
window function,” in Digital Signal Processing I1 New York: and editor/co-editor of three other books. He is the Consulting Editor
IEEE Press, 1975, pp. 123-126. for the ElectricalfComputer Science and Engineering Series of the Van
[9] L. R. Rabiner and B. Gold, Theory and Application of Digital Nostrand Reinhold Company,NY.
Signal Processing. Englewood Cliffs, NJ: Prentice-Hall, 1975, Dr. Mitra is a Fellow of the American Association for the Advancement
sec. 5.30, pp. 346-349. of Science and a member of the American Society for Engineering Educa-
101 2920 Analog Signal Processor Design Handbook, Intel Corpora- tion, Sigma Xi, and Eta Kappa Nu. He was an Associate Editor of the
tion, Santa Clara, CA 1980. IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, a member of the Adminis-
111 M. V. Thomas, Y. Neuvo, and S. K. Mitra, “Two-dimensional trative Committeeof the IEEE Circuitsand Systems Society, and a member
interpolated finite @pulse response filters,” in Proc. I983 IEEE of the editorial boards of the IEEE Press and PROCEEDINGS OF THE IEEE.
Int. Symp. Circuits Syst., Newport Beach, CA, May 1983, pp. He was the General Chairman of the 1974 IEEE International Symposium
104-106. on Circuits and Systems, and Technical Program Chairman of the 1983
121 L.R. Rabiner and B. Gold, Theory and Application o f Digital IEEE International Symposium on Circuits and Systems. He is also the
Signal Processing. Englewood Cliffs, NJ: Prentice-Hall 1975, recipient of the 1973 F. E. Terman Award of the American Society for
sec. 4.10, pp. 238-252. Engineering Education.

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on February 05,2025 at 07:24:54 UTC from IEEE Xplore. Restrictions apply.