EEN536 Digital Signal Processing Case Study-IFIR 1

Interpolated FIR (IFIR) Filters: A Case Study

Douglas A. Mann, Graduate Member, IEEE

x[n] G(z) L I(z) y[n]

Abstract—The method of interpolated finite impulse response
(IFIR) filter design is presented and applied to narrowband FIR
filter design. Based on available literature, techniques for Fig. 1. Interpolated finite response filter
achieving optimal IFIR design are explored. Computational gains
are given for each of the techniques discussed.
stage uses an interpolator, or image filter I(z) to
Index Terms—IFIR, Narrowband, Optimization, Interpolation attenuate the unwanted repetitions, resulting in the
desired filter response. As L is increased, the order
of the model filter decreases, while the order of the
I. INTRODUCTION
image filter increases. Depending of the selection of
F inite impulse response (FIR) filters are often
considered more preferable to infinite impulse
the stretching factor L , the reduction in the number
of multipliers needed for implementation can be up
response (IIR) filters due to the desirable to 80% [3].
characteristics of linear phase and guaranteed
stability. However, the computational requirements
of a FIR filter for a given specification are usually II. NARROWBAND FILTER DESIGN
greater than an IIR filter with the same
requirements. This is especially true in narrowband
FIR filter design, where filter order can become A. FIR Filter Design
prohibitively large. It has been proposed by Neuvo
et al. [1] that an alternative approach to narrowband In this section, a narrowband low-pass FIR filter is
FIR filter design is to use two filters in a cascade designed using the Remez exchange algorithm, also
arrangement as shown in Fig. 1. This method allows known as the Parks-McClellan algorithm. First, the
for a reduction in the number of multipliers needed minimum order of a linear phase low-pass FIR filter
in the implementation, while still meeting the is calculated using the Kaiser formula given by
specifications of the narrowband filter. The benefit
of this approach is that the combined order of the "20 log10 ( # p# s ) " 13
N! (1)
two filters is less than that achieved using a single 14.6($ s " $ p ) / 2%
optimized [8] narrowband filter. The first stage
consists of a model or prototype filter representing where ! p and ! s are the maximum passband and
the “stretched” version of the desired filter. The stopband ripples, and ! p and ! s are the passband
stretching factor L determines increased width of
the model filter passband relative to the desired and stopband edge frequencies. Using this
filter. As L increases, the passband width increases approximation for N, an optimized FIR filter can be
by L , and filter order decreases by a factor of L . designed using the Parks-McClellan algorithm [8].
The model filter is then upsampled by L , producing Consider the example where ! p = 0.015" ,
the desired narrowband filter and repeated images
, and ! p = 0.001 . Using these
! s = 0.020" ! p = 0.001
as shown in the plot in Fig. 6. The second
specifications in (1), the length N equals 1288. A
similar approximation of N is calculated with the
Manuscript received November 14, 2008. D. A. Mann is a Graduate kaiserord function in Matlab, resulting a larger N
Student in Music Engineering Technology Department at the University of
Miami, Coral Gables, FL 33143 USA (e-mail: doug.a.mann@ gmail.com).
equal to
EEN536 Digital Signal Processing Case Study-IFIR 2

Fig. 2: Low-pass optimal equiripple FIR filter with N=1451 Fig. 3. Model filter with L=20 and N=73

1451. Using the Matlab approximation of N=1451, 1) Expansion Factor Selection

the optimal low-pass equiripple FIR shown in Fig. 2
is designed and verified against the cutoff frequency The expansion or “stretching” factor L must be less
specifications. The total number of adders for than or equal to 2 and greater or equal than LMAX .
implementation is equal to the order N. The number
of multipliers needed for implementation is related L ! 2 ! LMAX (4)
to the length by:
#! %
LMAX = # % (5)
! N + 1$ $" s &
Mult = ceil # (2)
" 2 &%
It has been also recommended [1] to select L
The number of multipliers for this design is 726, somewhat smaller than LMAX for allow for less
which can be significantly reduced using the IFIR stringent interpolator design. As the expansion
design approach discussed in the next section.
factor increases, the order of the model filter
decreases. Consequently, a higher order image filter
B. IFIR Filter Design is needed to remove replicas. In this example, an
arbitrary expansion factor L = 20 will be used,
In this section we use the same design parameters as where LMAX = 50 .
section A, but using this IFIR approach. The
interpolated FIR is described by the equation: 2) Design of Model Filter

H FIR (z) = G(z L )I(z) (3) Using the same specifications for passband and
stopband ripple as the FIR designed in section II, a
where G(z L ) is the upsampled prototype or “model” model filter is designed using a passband equal to
filter, and I(z) is the interpolator, or “image” filter. L! p and stopband equal to L! s . The order estimate
As outlined in [1], the process of designing an IFIR for this model filter is 73 and frequency response is
filter is as follows: shown in Fig. 3. It can be observed that the model
filter resembles the desired filter in Fig. 2, but with
1) Select a suitable expansion factor L the specifications relaxed by a factor of L .
2) Design model filter G(z)
3) Upsample model filter G(z) by L to create G(z L )
4) Design the image filter I(z) to remove replicas
EEN536 Digital Signal Processing Case Study-IFIR 3

Fig. 4a. (top) Input sequence Fig. 4b. (bottom) Input sequence upsampled by 4 Fig. 6. Model filter with L=20 and N=73 after upsampling

The upsampled model filter G(z L ) , for L=20 is

shown in Fig. 6.

4) Design the Image Filter

The image filter is designed using the ripple

specifications used in example FIR design with
2
stopband ! si = " ! s , and passband equal to ! p .
L
The Kaiser approximated order is 112. The
frequency response is shown in Fig. 7. Convolving
these two filters in the time domain results in the
Fig. 5. Frequency response of (a) and (b) final IFIR filter shown in Figure 8. Note that the
passband and stopband edges are the same with
3) Upsample the Model Filter stopband attenuation of 65dB.

The model filter is upsampled by the expansion 5) Computational Complexity

factor to get G(z L ) using the following heuristic:
The computational complexity is often stated in
!# G(n / L) n = iL, i = 0, ±1, ±2... terms of the number of multipliers needed for
G '(z) = " (6)
#$ 0 otherwise implementation of a digital filter. The number of
multipliers needed for implementation of the FIR
G '(z) = G(z L ) (7) filter is 726. In comparison, the model filter requires
37 multipliers and the image filter requires 57
The result of upsampling in the time domain is multipliers for a total of 94 multipliers. The number
shown in Fig. 4a, where there are L ! 1 zeros of adders is 185. The reduction of complexity is
inserted between each sample. The frequency computed using
response after upsampling gives a periodic response
with period 2! / L on the interval [0, ! ] . The N FIR ! N IFIR
%computational reduction = (8)
N FIR
replicas from this periodicity can be seen in Fig. 5b.
EEN536 Digital Signal Processing Case Study-IFIR 4

TABLE I
COMPUTATIONAL REQUIREMENTS

L NG NI NG+NI MULT ADD %REDUCTION

2 726 8 734 368 734 49.41
3 484 12 496 249 496 65.82
4 363 16 379 190 379 73.88
5 291 20 311 156 311 78.57
6 242 25 267 134 267 81.60
7 208 29 237 119 237 83.67
8 182 34 216 109 216 85.11
9 162 39 201 101 201 86.15
10 146 44 190 96 190 86.91
11 132 50 182 92 182 87.46
12 121 56 177 89 177 87.80
13 112 62 174 88 174 88.01
14 104 68 172 87 172 88.15
15 97 74 171 86 171 88.22
16 91 81 172 87 172 88.15
17 86 88 174 88 174 88.01
Fig. 7. Image filter with L=20 and N=112 18 81 96 177 89 177 87.80
19 77 104 181 91 181 87.53
20 73 112 185 93 185 87.25

For the previous example, the optimal stretching

factor is L = 15 . The resulting orders of the image
and model filters are 74 and 97, respectively. The
total number of multipliers and adders are 87, and
171 respectively. The computational reduction
improves to 88.2%, which is considered optimal.
The optimal expansion factor using these
specifications is illustrated in Fig. 9, which plots
expansion factor against combined filter order for
the image and model filters. As the expansion factor
increase past the optimal value, it increases
Fig 8. Interpolated FIR Filter with L=20 and N=185
approaching the order of the FIR filter.
The combined order using the IFIR method is 185
with an arbitrary expansion factor L = 20 . B. Multiple Stretching Factors
Compared the optimal equiripple FIR, the
computational reduction (8) for L = 20 is 87.3%. It has been suggested [7] that more demanding
designs could benefit from using a second IFIR
implementation for the image filter. The structure
III. IMPROVEMENTS
shown in Fig. 10 using two stretching factors L and
A. Selection of Optimal Stretching Factor L1 could be utilized to realize this design. The
optimal solution for L and L1 is obtained by
Due to the dependency of both the image filter solving the simultaneous equations
and model filter complexity on the stretching factor,
there exists an optimal L , where the combined !1 2# 2#
+ ! =0
complexity is minimized. According to [7], the (" s ! " p ) *$ 2# ' -
2
* 2# $ 2# ' -
2

L1 ,& ! "s ) ! " p / , L ! &% L ! " s )( ! " p /

optimal stretching factor can be calculated using %
+ L ( . + 1 .

$ 2! ' !1 2"
LOpt = & ) (9) + =0
*$ 2" ' -
2
&% " p + " s + " s # " p )( * 2" $ 2" ' -
,&% L ! # s )( ! # p / , L ! &% L ! # s )( ! # p /
+ . + 1 . (10)
EEN536 Digital Signal Processing Case Study-IFIR 5

x[n] G(zL) I1(zL1) I2(z) y[n]

Fig. 10. Two stage interpolator

1) Choose expansion factor L , design model and

image filters, g(n) and i(n) using optimal FIR
methods.
2) Choose a rounding constant r . A recommended
starting point is r = 0.01 .
3) Round g(n) and i(n) to obtain gr (n) and ir (n) by

hr (n) = r ! round(h(n) / r) (12)

Fig. 9. Expansion factor vs. combined order of model and image filter

The first interpolator is used to suppress the 4) Upsample the rounded model filter gr (n) by
images as done in section II. However, the expansion factor L
upsampling of I1 (z L1 ) by L1 creates images as well. 5) Cascade the upsampled rounded model filter and
The final interpolator I 2 (z) is used to suppress the the rounded image filter. Apply a sharpening
polynomial to the result, as discussed in [10][11].
images produced by I1 (z L1 ) . The determination of
Start with n = 1 and m = 1 , and increase these values
optimal L and L1 was not computed, but analysis if the specification is not satisfied.
in [7] suggests that the computational reduction is
margin if not nonexistent for reasonable image filter 1) Rounding and Sharpening
orders. It should be noted that the two-variable IFIR
method could be extended further to construct a 3- The multiplier-free design method is applied to
variable structure. The gains from using these the design example given in the FIR and IFIR
highly complex design structures are only realized design sections. Using the same optimal expansion
for very high order narrowband filters. Whether or factor L = 15 , the model and image filters are
not computational gains can be achieved can be designed. A rounding constant of 0.01 is used, and
evaluated using the coefficients of the model and image filters are
1 rounded according to (12). The rounded model filter
! s = 0.5(1 " )! (11)
# p is upsampled by a factor of 15. The resulting filter
is then convolved in the time domain, or multiplied
where the ratio of ! s to ! p can be used to construct in the frequency domain with the rounded image
a region of benefit diagram referred to as the IFIR filter. Finally a sharpening function is applied
triangle in [7]. according to [10][11], with n = 1 and m = 1 , which
gives the equation
C. Multiplier-Free IFIR Design
H SH = H 2 (3 ! 2H ) (13)
The concept of interpolated FIR design is
extended [9] to construct a “multiplier-free” design. In this specific example, the frequency response
The approach is based on the optimization of the was better using n = m = 3 where the sharpening
filter coefficient values, where the values are function is given by
represented in signed powers-of-two or canonic
signed digits. The design process involves the H SH = H 4 (35 ! 84H + 70H 2 ! 20H 3 ) (14)
following steps:
EEN536 Digital Signal Processing Case Study-IFIR 6

with no compromise in the performance of the low-

pass filter. In some cases, depending on the filter’s
complexity, further gains may result from using
multiple stage IFIR design techniques as mentioned
in section IIIB. The use of rounding and sharpening
methods allow for narrowband IFIR filter design
without the use multipliers, which can be helpful
when designing for more primitive digital signal
processors.

REFERENCES
[1] Y. Neuvo, C. Y. Dong, and S. K. Mitra, “Interpolated finite impulse
response filters,” IEEE Trans. Acoust., Speech, Signal Processing, vol.
Fig. 10. IFIR response after rounding and sharpening
ASSP-32, pp. 563-570, June 1984.
[2] T. Saramaki, Y. Neuvo, and S. K. Mitra, “Design of computationally
efficient interpolated FIR filters,” IEEE Trans. Circuits Syst., vol. CAS-
The resulting multiplier-free design in shown in 35, pp. 70-88, Jan. 1988.
Fig. 10. Using r = 0.01 , the passband and transition [3] R. Lyons, “Interpolated narrowband lowpass FIR filters,” IEEE Signal
Process. Mag., pp. 50-57, Jan. 2003.
bands meet the specifications, but the stopband [4] P. P. Vaidyanathan, Multirate Systems and Filter Banks, ser. Prentice-
attenuation is only 30 dB in the worst case. Using a Hall Signal Processing Series, A. V. Oppenhein, Ed. Englewood Cliffs,
NJ: Prentice Hall, 1993.
smaller rounding constant of 0.001 gives a stopband [5] J. L. Webb, D. Munson “A new approach to designing computationally
attenuation of 50dB for the worst case. The result of efficient interpolated FIR filters” IEEE Trans. Signal Processing, vol.
44, pp. 1923-1931, Aug. 1996.
using large rounding constant is more passband [6] S. K. Mitra, Digital Signal Processing: A Computer-Based Approach,
New York, NY: McGraw Hill, 2006, pp. 427-578.
distortions, requiring higher order sharpening [7] A. Mehrnia, J. A. Willson, “On optimal IFIR filter design,” IEEE
functions to get desired results. Using rounding Proceedings of the 2004 International Symposium on Circuits and
Systems, vol. 3, pp. 133-136, May 2004.
coefficient of 0.001, the number of non-zero [8] J. H. McClellan, T. W. Parks, and L. R. Rabiner, “A computer program
coefficients are 59 and 79 for the image and model for designing optimum FIR linear phase digital filters,” IEEE Trans.
Audio Electroacoust., vol. AU-21, pp. 506-526, Dec. 1973.
filters respectively. It should be noted that the [9] G. Jovanovic-Dolecek, S. K. Mitra, “Multiplier-free FIR filter design
designs given in [9], the passband is much wider based on IFIR structure and rounding,” IEEE 48th Midwest Symposium
on Circuits and Systems., vol. 1, pp. 559-562, 2005.
than this design example. The computational gains [10] A. Bartolo, B. D. Clymer, R. C. Burges, and J. P. Turnbull, “An efficient
method for FIR filtering based on impulse response rounding,” IEEE
from performing integer multiplications, which can Trans. On Signal Processing, vol. 46, No. 8, August 1998, pp. 2243-
be implemented using shift-add operations, have to 2248.
[11] J. Kaiser, R. Hamming, “Sharpening the response of a symmetric
be compared to the additional computational nonrecursive filter by multiple use of the same filter,” IEEE Trans.
overhead of recursively using the same filter in the Acoust., Speech, Signal Process., vol. ASSP-25, no. 5, pp. 415-422, Oct.
1977.
sharpening function. The reduction in complexity
with respect to the combined order of the model and
image filters is 90.5%. This may not result in an
overall improvement due to the rounding and
sharpening overhead.

IV. CONCLUSION
The interpolated finite response filter can offer
significant gains in the computational complexity of
narrowband filter implementation. For a general
IFIR design problem an optimal selection of the
expansion factor results in the greatest
computational reduction. In the given example, a
reduction in computation of 88.2% was realized,