Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
ON CIRCUITS
AND
SYSTEMS,
VOL.
277
IEEE
I. INTRODUCTION
FUNDAMENTAL goal in digital filter design is to
m inimize the computational complexity of the filter
realization. It is virtually always desirable to m inimize the
quantities M , b, and A, where M denotes the number of
m u ltipliers, b denotes the number of bits used for the
m u ltiplier coefficients, and A is the number of adders. The
relative importance of the above quantities depends on the
specific application, so that various complexity measures
are in use. Typical complexity measures are: Mb, M , and
M+.A.
Mb is often used to characterize the complexity of digital
(b)
Fig. 1. Linear phase FIR digital filter structure. (a) Even length. (b)
Odd length.
II.
2B
DBf
tpi
1),2,
for even L
for odd L.
278
IEEE TRANSACTIONS
-pgg-f
-1
(4
-qypyj$-1
(b)
STAGE
1 STAGE
2 STAGE
---
STAGE
L-l
ON CIRCUITS
AND
SYSTEMS,
VOL.
CAS-30, NO. 5,
MAY
1983
(4
p( dw) =
sin(wL/2)
e-jo(L-1)/2
sin (a/2)
The first null occurs at: o,,,,t = 27r/L. All of the nulls of
the prefilters frequency response must of course lie in the
stopband, which implies that: L < 27r/o,. This constraint
may be refined by noticing that the transition band of the
equalizer can be effectively widened if the first null of the
prefilters response is placed just slightly above ws. This
causes the prefilter to provide a large amount of attenuation near the stopband edge, so that the equalizer is not
required to work very hard until the frequency of the
prefilters first sidelobe. According to the above discussion,
the length of the RRS prefilter should be chosen as follows:
Lp = ISLT(2 77/o,}
where
L,, = RRS prefilter length
ADAMSAND
WILLSON:
FIR
DIGITAL
219
FILTERS
PREFILTER:
- - -
TABLE I
HARDWA~REQUIREMENTSSUMMARYFO~THEEXAMPLE
EQUALIZER:
-10
DELAYS
d6
ADDERS
MULTIPLIERS
PAEFILTER
14
EQUALIZER
23
23
12
TOTAL
37
25
12
35
35
18
CONVENTIONAL
FILTER
-40
-50
02lr
04x
06K
08K
lt
RADIANS/SAMPLE
(4
0
-10
-20
d6
-30
-50
027T
04a
0.6K
(b)
-20
d6
-30
-40
-50
027r
0477
RADlANSlSAMPLE
06K
OSli
noise contributed
by each multiplier
is often
used.
RADIANS/SAMPLE
(4
Coefficient Quantization
F ig. 4(a)-4(c) illustrate the effects of coefficient quantization on the a m p litude responses of the proposed and
conventional filters. The proposed filter clearly exhibits
superior performance with respect to coefficient-quantizaF ig. 3(a) shows overlays of the individual RRS prefilter tion sensitivity in this case. Moreover, there are some
and equalizer a m p litude responses. Notice that not only fundamental reasons why this should be true in general. A
does the prefilter allow a wider transition band for the general analysis of the sensitivity of the frequency response
equalizer, but also the stopband attenuation requirements to coefficient quantization is developed in the next section.
on the equalizer are .relaxed for those frequencies in the
IV. SENSITIVITY O F THE FREQUENCY RESPONSE T O
vicinity of nulls on the prefilter response.F ig. 3(b) provides
COEFFICIENT QUANTIZATION
the overall a m p litude response of the prefilter-equalizer
cascade. As a basis for comparison, a conventional filter
In practice the filters m u ltiplier coefficients must be
was designed to meet the same specifications. The required representedby finite-length computer words. Typically each
Fig. 3. (a) Individual amplitude responses for the prefilter and the
equalizer. (b) Prefilter-equalizer cascade amplitude response. (c) Amplitude response of the conventional filter.
EEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-30, NO. 5, MAY 1983
a
CONVENTIONAL:
- - - -
PROPOSED:
(kt
-::s
(Y+-+
(a)
CONVENT?ONAL :
PROPOSED:
- - - -
-30
-40
-50
027r
0.4K
0.6K
RADIANS/SAMPLE
(W
-10
CONVENTIONAL:
PROPOSED :
-c::,L$
RADIANS/SAMPLE
-10
- - - -
0.6R
Clearly, the prefilter attenuates the equalizers quantization error spectrum for frequencies in the stopband. Fig. 5
)'\
shows block diagram representations of the above relations
'I ..,'i . ...,..
for the conventional, and the prefilter-equalizer cascade
-40
filters. Fig. 6(a) shows .plots of lH(ej)l and IE(ej)l for
the conventional filter example discussed in Section III,
-50
with
8-bit coefficients. Similarly, Fig. 6(b) illustrates
0
0.677
02x
04H
0.6X
IH( t+)l and 1P( ej)E( ej)l for the prefilter-equalizer
RADIANSISAMPLE
cascade version of this example, again using g-bit coeffi(4
cients.
The attenuation of the coefficient-quantization error
Fig. 4. Amplitude responses of the proposed and conventional filters
with coefficients quantized to: (a) 6-bits, (b) I-bits, (c) IO-bits.
spectrum is clearly evident in the prefilter-equalizer cascade
for frequencies in the stopband.
By employing Parsevals theorem it is easy to demonideal coefficient value is rounded to the nearest represenstrate
that the total coefficient-quantization error energy at
table number, so that the coefficient-quantization error
all
frequencies,
including those in the passband, is expected
incurred is no more than half the quantization stepsize. We
to
be
less
in
the
prefilter-equalizer cascade than in the
let h(n) and h(n) denote the infinite precision and finite
conventional
filter.
Parsevals theorem relates the total
precision coefficients, respectively. Then the quantization
energy
in
the
error
spectrum
to the energy in the coeffierror e(n) is given by: e(n) = h(n)- h(n). We let H(ej),
cient-quantization
error
sequence
as follows.
H(@), and E(ej) denote the Fourier transforms of the
sequences {h(n)}, (h(n)}, and {e(n)}, respectively. As the
Fourier transformation is a linear operator, then we also
t e(n).
&J_
IE(@)l *da=
77
have: E(ej) = H(ej)H(ej). Assuming that the e(n)
n=l
ADAMS
AND
WILLSON:
FIR
DIGITAL
281
FILTERS
-\
pi(ejW) I:
tE(ej)I
\
I
RADIANS/SAMPLE
(4
IH'(ej")l:,
IP(e
jw).E(ejw)l:
RADIANS/SAMPLE
(b)
Fig. 6. Amplitude response of the filter with quantized coefficients and
the error spectrum. (a) Conventional filter with 8-bit coefficients. (b)
Prefilter-equalizer
network with 8-bit coefficients. (Notice that the
level of the error spectrum is below -50 dB for frequencies in the
stopband so that the effect on the overall amplitude response is very
small.)
IE(ej)[do)
77
for even M
for odd M
where
example,
the filter
designer
may
have
been
forced
to
APPENDIX
As for any digital filter, the details of the implementation of the RRS must depend on the specific digital signal
processing environment in which it is used. In particular,
the scaling of the gain of the RRS would normally depend
on many factors, including: the desired overall system gain,
the nature of the signal being filtered (its spectral composition and a m p litude), and also various hardware constraints
(for example, scaling is usually not even an issue for digital
filters implemented in a general purpose computer with
floating-point arithmetic).
282
IEEE TRANSACTIONS
I
-1
Fig. 7.
AND
SYSTEMS,
VOL.
l/L
ON CIRCUITS
I
1
APPENDIX
G(ej)
W( ej)
For lowpass filters the unmodified Parks-McClellan computer program [4] uses G,(ej) and R(ej) as follows:
O<W<W P
1,
Gi(ej)=
o
u,Qw<v
1,
]W(ej)*{~(ej)~~(ei)-Gd(ejw)}]
A(.?@) = Iw(ej).P(ejw).{~(ejo)-Gd(ejw)/B(ejw)}l
A(ej) = I~(ejO).{~(ejw)-Zb(ejw)}l
where
*(
eja)
=
i
G(e+)
lP(eiw>l,
OdW<O,
KIF(
w,<w<?l
= Q(ej),
ejO)I,
O<C.O<7T.
FIR DIGITAL
283
FILTERS
Equalizer
Conventional Filter
-0.12163029
0.023 19508
-0.02155842
- 0.00969477
- 0.00605022
0.01019741
0.05381588
0.07186058
0.11563688
0.0826 1736
0.14487070
0.15673977
0.01084545
0.007335 12
0.00876 169
0.00942935
0.00898000
0.00704726
0.00336417
0.00222462
0.00969186
0.01891595
0.02952445
0.04103817
0.052825 16
0.06417377
0.07436218
0.08270440
0.08861664
0.09168579
REFERENCES
[II
PI
+
John W. Adams (S75-M81) was born in Santa
Monica, CA, on February 8, 1954. He received
the B.S., MS., Engr., and Ph.D. degrees in electrical engineering from the University of California, Los Angeles, in 1976, 1976, 1978, and
1982, respectively.
In 1978 he joined the Radar Systems Group of
the Hughes Aircraft Company, where he now
holds the position of Senior Staff Engineer in the
Systems Engineering Department.
His current
research interests are in the areas of digital signal
processing and synthetic aperture raaar.
Dr. Adams is a member of Phi Beta Kappa, Tau Beta Pi and Sigma Xi.
+
Alan N. Willson, Jr. (s66-M67-SM73-F78)
was born in Baltimore, MD, on October 16,
1939. He received the B.E.E. degree from the
Georgia Institute of Technology, Atlanta, GA, in
196 1, and the M.S. and Ph.D. degrees from
Syracuse University, Syracuse, NY, in 1965 and
1967, respectively.
From 1961 to 1964 he was with the IBM
Corporation, in Poughkeepsie, NY. He was an
Instructor in Electrical Engineering at Syracuse
University from 1965 to 1967. From 1967 to
1973 he was a Member of the Technical Staff at Bell Laboratories,
Murray Hill, NJ. Since 1973 he has been on the faculty of the University
of California, Los Angeles, where he is now Professor of Engineering and
Applied Science, in the Electrical Engineering Department. In addition,
he served the UCLA School of Engineering and Applied Science as
Assistant Dean for Graduate Studies, from 1977 through 1981. He has
been engaged in research concerning the stability of distributed circuits,
properties of nonlinear networks, theory of active circuits, digital signal
processing, and analog circuit fault diagnosis. He is editor of the book
Nonlinear Networks: TheoT and Analysis, IEEE Press, 1974.
Dr. Willson is a member of Eta Kappa Nu, Sigma Xi, Tau Beta Pi, the
Society for Industrial and Applied Mathematics, and the American Society
for Engineering Education. From June 1977 to June 1979 he served as
Editor of the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, and during
1980 he was Vice-President of the IEEE Circuits and Systems Society. He
now holds the office of President-Elect of the IEEE Circuits and Systems
Society. He is the recipient of the 1978 Guillemin-Cauer
Award of the
IEEE Circuits and Systems Society, for co-authoring
the best paper
published in their TRANSACTIONS during the previous
year. He is the
recipient of the 1982 George Westinghouse Award of the ASEE, and the
1982 Distinguished Faculty Award of the Engineering Alumni Association.