0% found this document useful (0 votes)
42 views76 pagesLO Chap Multirate
Uploaded by
HIMANICopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
42 views76 pagesLO Chap Multirate
Uploaded by
HIMANICopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 76
3
Multirate Processing
of Digital Signals
Ronald E. Crochiere
Lawrence R. Rabiner
AT&T Bell Laboratories
3.0 INTRODUCTION
‘As more and more signals in the real world are represented, stored, and transmitted
in digital formats, the importance of being able to process a signal digitally from its
inception to its final destination grows. The general theory of digital signal processing
has emerged in the past two decades and has grown to prominence in the engineering
‘community, as evidenced by the publication of several key texts and reprint col-
lections [1-9]. Almost all of the theory presented in those references deals with
processing signals at a fixed sampling rate. However, in the past few years; there has
been an increasing need for a deeper understanding of how to process digital signals
in systems that require more than one sampling rate. An entire subfield of digital signal
processing —multirate signal processing—has developed to meet this need [10]
‘This chapter presents the main ideas and concepts of multirate digital signal
processing, with particular emphasis on digital techniques for changing the sampling
rate of a signal. We begin our discussion with a thorough review of the Nyquist
sampling theorem and its interpretations in terms of modulated signals. We then show
how a continuous-time signal can be reconstructed from its digital samples. This
x0 + (0) = 2 > xf:
js
Flot 270] 6.16)
Equation (3.16) shows that the Fourier transform of the digital signal is the sum
of frequency-shifted and scaled versions of the Fourier transform of the continuous
signal.
‘We can now see quite clearly the effect of different choices of ty = 2x/T on
the resulting Fourier transform of the digital signal. These possibilities are illustrated
in Fig. 3.4. Figure 3.4(a) shows the Fourier transform of the continuous signal and
Fig. 3.4(b) shows the resulting digital Fourier transform when Qy > 4. The
individual terms for € = O and € = +1 of Eq. (3.16) are shown in this igure. Figure
3.4(6) shows the resulting discrete-time Fourier transform when Qe = 47F.. In this
case the individual terms of Eq. (3.16) come right up to each other in frequency. This
case is referred (0 as critical sampling of the signal. Finally, Fig. 3.4(@) shows the
resulting discrete-time Fourier transform when 9 < 4F...In this case the individual
terms in Eq, (3.16) overlap in frequency, and the resulting digital frequency response,
in general, bears no simple, direct relationship to the continuous frequency response
of Fig. 3.4(2). In this case we say that the digital signal is an aliased representation
of the continuous signal.
‘The implications of the three cases of sampling discussed above are summarized
in a simple and straightforward manner in the Nyquist sampling theorem:
Ifa continuous-time signal s.(0) has a bandlimited Fourier transform X-(M) that satisfies
the condition |X.(Q)| = 0 for 2 = 2nf., then x.(0) can be uniquely reconstructed,
without error, from equally spaced samples x(n) = x(nT), > nn
Toast Tosa!
igtateanaog towne
ste conerter “ter
Figure 3.5. Sampling aod reconsuvction of a continuous-time signal.
To implement this process, an ideal digital-to-analog converter is required to get
2(95-(0) from x(n). Assuming that we do not worry about the reliability of suct an
ideal converter, the reconstruction formula for x.(1) is
20 = [1 xeloselvhe ~ 9 dr G.17)
and applying Eqs. (3.4) and (3.5) gives
xe) = xtohu(t ~ a) Gus)
If we use the reconstruction method, the relationship between the reconstructed
signal x-(#) and the set of samples x(n) is given by the convolutional formula
x) =D x(m\in(t = aT) (2.19)
“The ideal lowpass filter hy(t) has the frequency-domain characteristics
aio = {5 Gaile a7 6.20 Sec. 3.2 Sampling Rate Conversion 131
‘giving the ideal impulse response
h(i) el -e<1 ad -
(3.61)
= 5 hw + date | "P| - )
Similarly, by considering the transform relationships of the individual integer
decimation and interpolation systems, the output spectrum ¥(w") can be determined
in terms of the input spectrum X(w) and the frequency response of the filter Ho").
From Eq. (3.46) we see that V(o") can be expressed in terms of X(w) and H(w") as
V(o") = H(wK(o"D) (6.62)
and from Eq. (3.35) ¥(') can be expressed in terms of V(w") a5
Yo) = B'S vie" - 270)
= E'S ae’ - 20 MOX(wL - 2O/M) 0.83)
‘When H(w”) closely approximates the ideal characteristic of Eq. (3.54), we see that
this expression reduces to
Lies : uw
ven - [fe Lim), torle'| = ain, 7 ‘]
(3.64)
3.2.4 General Form of Digital Sampling Rate Conversion
It is possible t0 generalize the discussion of Sections 3.2.1-3.2.3 so as to give a
canonic form of a digital system for sampling rate conversion. The form of this
canonic system is illustrated in Fig. 3.10, in which an input signal x(n), sampled at
rate F, = 1/T, is sent toa linear, time-varying, digital system with impulse response
a(n) to give the output signal y(m), with new sampling rate F/ = 1/7". If we assume
‘ret
la) —> diitaigcat ey
stor —>] ota |e vim gure 310 (a) Dicect ital
conversion of x(n) t0 ym) and () a
Aime varying filter interpretation of the
1 process. a2 Multirate Processing of Digital Signals Chap. 3
periods of y(m) and x(n) can be expressed as the rational
TLR LM
ne 6.65)
with M and L integers, then the digital system response gq(n) becomes the response
at output sample time m to an input at sample time [mAf/L ] — n, where [u] again
denotes the integer less than or equal t0 u.
Since the system is linear, each output sample y(m) can be expressed as alinear
‘combination of input samples. A general form [10] for this expression is
yim) = 3 setois(| 24] = ) (3.66)
where ga(n) is periodic in m with period Lyi
ml) = Baral), 7 6.67)
‘Thus the system ga(n) belongs to the class of linear, periodically time-varying sys-
tems.
Consider now several specific cases of sampling rate conversion systems. First
consider the trivial case T’ = Tor L = M = I, in which case Eq. (3.66) reduces to
the simple time-invariant digital convolution equation, i.c.,
ym) = 3 g(nx(m ~ n) 0.68)
since the period of gq(n) is 1 and the integer part of m — n is the same as m~ n.
Next consider the case of sampling rate reduction (decimation) by an iateger
factor M. Jn this case we get
v(m) = B anlade(mbt ~ 0) 13.69)
where gq(n) = g(n) = k(n) for all m and n, with A(n) the lowpass filter impulse
response of the system of Fig. 3.7. Although gq (7) is nota function of m for this case,
itcan readily be shown that the overall system of Eq. (3.69) is not time-invarian: [10]
Next consider the case of sampling rate increase (interpolation) by an integer
factor L. By comparing Eqs. (3.66) and (3.52), we see that the form of gq(n is
Sula) = A[nL + ((m)).], for all mand n G.70)
and also that g_(n) is periodic in m with period L.
Finally, if we consider the general case of sampling rate conversion by the
rational fraction M/L, then from Eq. (3.61) we get the result
Sn(n) = nL + ((mM)).], for all m and n en)
‘where (2) is the time-invariant unit sample response of the lowpass filter at the
sampling rate LF,. Sec. 3.3 Practical Structures for Decimators and interpolators, 3
3.3 PRACTICAL STRUCTURES FOR DECIMATORS,
AND INTERPOLATORS
Tis easy to understand the need for studying structures for realizing sampling rate
‘conversion systems by examining the simple block diagram of Fig. 3.9(b), which can
be used to convert the sampling rate of a signal by a factor L/M. As discussed in
Section 3.2, the theoretical model for this system first increases the signal sampling
rate by a factor L (by filling in Z ~ 1 zero-valued samples between each pair of
samples of x(n) to give the signal w(K), then filters w(t) (to eliminate the images of
X(q)) by a standard lineae, time-invariant, lowpass filter 4(&) to give v(b), and then
‘compresses the sampling rate of o(k) by a factor M (by retaining 1 of each M samples
of v(W). A direct implementation of this system is grossly inefficient since the lowpass
filter A(R) is operating at the high sampling rate on a signal for which L — 1 out of
each L input values are zero, and the values of the filtered output are required oaly
‘once each M samples. For this example, we can directly apply this knowledge in
implementing the system of Fig. 3.9(b) in a more efficient manner, as'will be dis-
ccussed in this section,
3.3.1 Signal Flow Graphs
To precisely define the sets of operations necessary to implement these digital sys-
tems, we will strongly rely on the concepts of signal fow graph representation (10]
Signal flow graphs provide a graphical representation of the explicit set of equations
that are used to implement such systems. Furthermore, manipulating the flow graphs
in a pictorial way is equivalent to manipulating the mathematical equations.
Figure 3.11 illustrates an example of a signal flow graph of a direct-form finite
impulse response (FIR) digital filter. The input branch applies the external signal x(n)
to the network, and the output of the network y(n) is identified as one of the node
wou 210) Out
a vot
non
nea
aio)
of Rid Figure 3.11 Direct-form structure for
an FIR digital er. ms Multirate Processing of Digital Signals Chas. 3
values. Branches define the signal operations in the structure such as delays, gains,
and sampling rate expanders and compressors. Nodes define the connection points and
summing points. The signal entering a branch is taken as the signal associated with
the input node value ofthe branch. The node value of a branch is the sum of all branch
signals entering the node.
From the signal flow graph in Fig. 3.11 we can immediately write down the
network equation as
y(n) = x(n)h(0) + x(n — DAI) +--+ + x(n N+ DAW = 1)
‘An important concept in the manipulation of signal flow graphs is the principle of
‘commutation of branch operations. Two branch operations commute if the order of
their cascade operation can be interchanged without affecting the input-to-output
response of the cascaded system. Thus interchanging commutable branches :n a
network is one way of modifying the network without affecting the desired input-to-
output network response. This operation will be used extensively in constructing
efficient structures for decimation and interpolation, as we will see shorily
‘Another important network concept on which we rely heavily is that of trins-
position and duality [10]. Basically, a dual system is one that performs a com-
plementary operation to that of an original system; and it can be constructed from the
original system through the process of transposition. We have already seen an example.
of dual systems, namely, the integer decimator and interpolator (Fig. 3.7a and Fig.
3.8a) for the case M = L. s
Basically the transposition operation is one in which the direction of all branches
in the network are reversed and the roles of the input and output of the rietwork are
interchanged. Furthermore, all branch operations are replaced by their trans;ose
operations. In the case of linear time-invariant branch operations, such as gains and
delays, these branch operations remain unchanged. Thus, for example, the transpose
of the direct-form structure of Fig. 3.11 is the transposed direct-form structure shown
in Fig. 3.12. Also it can be shown that for the case of linear time-invariant systems
output mo) Input
vn =——+ aa}
no
neat
nar
e 1 A Figure 3.12 Trassposed direet-fora
FIR flier sauesure, Sec. 3.3 Practical Structures for Decimators and Interpolators 145:
the input-to-output system response of a system and its dual are identical (e.g., it ean
be verified that the networks of Fig. 3.11 and Fig. 3.12 have identical ¢ystem
functions).
For the time-varying systems this is not necessarily the case. For example, the
transpose of a sampling rate compressor is a sampling rate expander, and the transpose
of a sampling rate expander is a sampling rate compressor, as shown in Fig. 3.13.
Clearly these systems do not have the same system response.
Trane
J EN
tee NOL.
Figure 3.13 Traspostions of
co sampling rate compressor and expander.
By extending the concepts of transposition rigorously, we can also show that the
transposition ofa network that performs a sampling rate conversion by the factor L/M
is a network that performs a sampling rate conversion by the factor M/L. This is
illustrated in Fig. 3.14.
opted pee fw bee et mim Looe] far Loe
fe fet] tir eet fae Le
fe bef na
o
Figure 3.14 Transpostios of a decimator and s generalized L/M sampling rate changer.
3.3.2 Direct-Form FIR Structures for Integer Changes
in Sampling Rates
Consider the model of an M-to-1 decimator as shown in Fig. 3.15(a). According to
this model the filter s(n) operates at the high sampling rate F, and M — 1 out of every
‘M output samples ofthe filters are discarded by the M-to-1 sampling rate compressor.
In particular, if we assume thatthe filter h(n) is an N-point FIR filter realized with a
direct-form structure, the network of Fig. 3.15(b) results. The multiplications by (0), 146 ‘Multirate Processing of Digital Signals Chap. 3
‘A(1), ...., AON — 1) and the associated summations in this network must be pe--
formed at the rate F.
‘A more efficient realization of the above structure can be achieved by noting that
the branch operations of sampling rate compression and gain can be commuted. By
performing a series of commutative operations on the network, we obtain the modified
network of Fig. 3.15(c). The multiplications and additions associated with the
coefficients (0) to h(N — 1) now occur at the low sampling rate F/M and therefore
the total computation rate in the system has been reduced by a factor Mf. For every M
samples of x(n) that are shifted into the structure (the cascade of delays), one outpat
xk] tod | yim
4 am
x0) f+ vim
wo
1
ato vin)
no
m2,
L
nin
‘ee Figure 3.15 Generation of an effceat
diect-form sructure of an M-t0-1
decimate. Sec. 33 Practical Structures for Decimators and interpolators 147
sample y(m) is computed. Thus the structure of Fig. 3.15(c) is seen to be a direct
realization of Eq. (3.29).
‘An efficient structure for the 1-to-L integer interpolator, using an FIR filter, can
be derived in a similar manner. We begin with the cascade model for the interpolator
shown in Fig. 3.16(a). In this case however, if h(m) is realized with the direct-form
structure of Fig. 3.11 we are faced with the problem of commuting the 1-to-L sampling
tale expander with a series of unit delays. One way around this problem is to realize
‘h(m) with the transposed direct-form FIR structure as shown in Fig. 3.12. The
sampling rate expander can then be commuted into the network as shown by the Series
xt te en} re
p10)
Py te a
aot
m2) fe
ROOF uw, Figure 3.16 Steps inthe generation of
an effcieat sructe of 2 I-o-L
‘e ‘ncexpoato, 148 Muttirate Processing of Digital Signals Cnap. 3
of operations in Fig. 3.16. Since the coefficients (0), (1), . .., R(V ~ 1) in Fig.
3.16(c) are now commuted to the low sampling rate side of the network, this structure
requires a factor of L times less computation than the structure in Fig. 3. 16(t.
An altemative way of deriving the structure of Fig. 3.16(c) is by a direct
transposition of the network of Fig. 3.15(c) (leting L = M). This is a direct con-
sequence of the fact that decimators and interpolators are duals. A further property of
transposition is that for the resulting network, neither the number of multipliers nor
the rate at which these multipliers operate will change [15]. Thus if we are given a
network that is minimized with respect to its multiplication rate, then its transpose will
also be minimized with respect to its multiplication rate.
3.3.3 Polyphase FIR Structures for Integer Decimators
and Interpolators
‘A second general class of structures of interest in multirate digital systems is
the polyphase networks. We will find it convenient to first derive this structure for
the L-to-I interpolator and then obtain the structure for the decimator by trans-
posing the interpolator structure.
In Section 3.2 it was shown that a general form for the input-to-outpur time-
domain relationship for the I-to-L interpolator is
G.72)
y(n)
where
Buln) = h(nk + ((rr)),) for all m and n (3.73)
isa periodically time-varying filter with period L. Thus to generate each output sample
y(m), m= 0,1, 2,..., L— 1, a different set of coefficients gq (n) is used. After
7 outputs are generated, the coefficient pattern repeats; thus y(L) is generated using
the same set of coefficients go(n) as y(0), y(L + 1) uses the same set of coefiicients
8:() as y(1), and so on.
‘Similarly the term [m/L | in Eq. (3.72) increases by 1 for every L samples of
_»(mm). Thus for output samples y(L), y(L + 1),.. . ,y(2L — 1) the coefficients ga (7)
are multiplied by samples x(1 — 1). In general, for output samples y(rL), y(rL + 1),
= +++ ¥(L +L ~ 1), the coefficients gq(n) are multiplied by samples x(r — n)
‘Thus we see that x(n) is updated at the low sampling rate F., whereas y(m) is evaluated
atthe high sampling rate LF.
‘An implementation of the -to-L interpolator based on the computation of Eq.
(3.72) is shown in Fig. 3.17(a). The way in which this structure operates is as follows.
‘The partitioned subsets go(n), s(n), . . . » gr-i(a), of h(m) can be identified with L
separate linear time-invariant filters that operate atthe low sampling rate F,. To make
this subtle notational distinction between the time-varying coefficients and the time-
invariant filters, we wil refer to the time-invariant filters respectively as po(n). p(n),
+ Pe-a(r). Thus
Poln) = Bo(n), for p = 0, 1, 2,.
L-1andalln (3.74) Sec. 3.3 Practical Structures for Decimators and Interpolators 149
sar ra
ao ae Lee mi
no Lee
ato) |_2al0) ze
| ion te
=r)
os
‘
r
raid
voy te vm
rate
LT Le fe
vate
ni) PL TS
Yous tmd
eee
A “
ewe 217 Polyphase smc foe 4
Horns
‘These filters p(n) will be referred to as the polyphase filters. Furthermore, by com-
‘ining Eqs. (3.73) and (3.74) we see that
pol) = h(n +p), for p=0,1,2,...,L—1andalln (3.75)
For each new input sample x(n), there are L output samples (see Fig. 3.17). The
‘output from the upper path yo(m) has nonzero values for m = nln
. - , which correspond to system outputs y(nL), n = 0, +1
the next path y;(m) is nonzero for m = nb + 1,n = 0, #1, +2,
delay of one sample at the high sampling rate. Thus y,(m) corresponds to the inter-
polation output samples y(nl. + 1), n= 0, =1,... In general, the output of the 150 Multirate Processing of Digital Signals Chap. 3
pth path, ys(m) corresponds to the interpolation output samples y (nL + p), n = 0,
‘1, .... For each input sample x(n), each of the L branches of the polyphase
network contributes one nonzero output that corresponds to one of the L outputs of the
network. In the polyphase interpolation network of Fig. 3.17(a), the filtering is
performed at the low sampling rate, and thus it is an efficient structure. A simple
‘manipulation of the structure of Fig. 3.17(a) leads to the equivalent network of Fig.
3.17(b), in which all the delays are single sample delays.
‘The individual polyphase filters p,(n), p = 0, 1, 2,...,L ~ Uhaveanumber
of interesting properties. This is a consequence of the fact that the impulse resporses
elt), p= 0, 1, 2,...,L— 1, correspond to decimated versions of the impulse
response of the prototype filter h(m) (decimated by a factor of L according to Eq. 3.73
or 3.75). Figure 3.18 illustrates this for the case L = 3 and for an FIR filter (m) with
N= 9 taps. Part (a) shows the samples of /t(m) for A(m) symmetric about m = 4,
‘Thus h(m) has a flat delay of 4 samples. The filter po(n) has three samples correspond-
ing to (0), h(3), (6) = (2). Since the point of symmetry ofthe envelope of p(n)
it has a flat delay of $ samples. Similarly, pi(n) has samples (1), (4),
'5(1) = (1), and because its zero reference (2 = 0) is offset by } sample (with respect
tom = 0) itas a lat delay of 1 sample. Thus different fractional sample delays and
consequently different phase shifts are associated with the different filters p,(n), as
seen in Fig. 3.18(b). These delays are compensated for by the delays that occur atthe
high sampling rate LF in the network (see Fig. 3.17). The fact that different phases
are associated with different paths of the network is, of course, the reason for the term
polyphase network,
‘A second property of the polyphase filters is shown in Fig. 3.19. The frequency
response of the prototype filter A(m) approximates the ideal lowpass characteristic
H,(a) shown in Fig. 3.19(a). Since the polyphase filters p,(n) are decimated versions
of A(m) (decimated by L), the frequency response 0 = w = /L of Hj(«) scales to
the range 0. w' = 2 for P,,(w") as seen in Fig. 3.19, where P,(c') is the ideal
characteristic that the polyphase filter p,(n) approximates. Thus the polyphase filters
approximate allpass functions and each value of p, p = 0, 1,2,...,L— 1, corre-
sponds to a different phase shift.
“The polyphase filters can be realized in a variety of ways. If the prototype filter
‘h(m) is an FIR filter of length N, then the filters p,(n) will be FIR filters of length N/L.
In this case itis often convenient to choose N to be a multiple of L so that all of the
polyphase filters are of equal length. These filters may be realized by any of the
‘conventional methods for implementing FIR filters such as the direct-form structure
‘or the methods based on fast convolution [10]. If a direct-form FIR structure is used
for the polyphase filters, the polyphase structure of Fig. 3.17 will require the same
‘multiplication rate as the direct-form interpolator structure of Fig. 3.16.
By transposing the structure of the polyphase 1-to-L interpolator of Fig. 3.17(b),
‘we get the polyphase M-to-1 decimator structure of Fig. 3.20, where L is replaced by
M. Again the filtering operations of the polyphase filters occur atthe low sampling rate
side of the network, and they can be implemented by any of the conventional struc~
tures discussed above
In the preceding discussion for the 1-to-L interpolator, we have identified the
coefficients of the polyphase filters p(n) with the coefficient sets gq (n) of the time- Sec. 3.3 Practical Structures for Decimators and Interpolators 151
eis) Pel)
2a) Ate)
ag heah + Pala)
Phase of 2, (a)
Figure3.18 Mlustration ofthe properties
o of polyphase network Stes. alm
152 Multirate Processing of Digital Signals Chap. 3
te
° wr 7
”
patel
° +
agin
fe To
fw Le Te
ee rg Be ee
fe et ca
f ee lM
Figure 3.19 Ideal frequency respoase of
the polyphase nerwork filters.
oh
‘igure 3.20 Polyphase structure for an
M1 decimate.
varying filter model. In the case of the M-to-1 decimator, however, this identification
cannot be made directly. According to the time-varying filter model, discussed in
Section 3.2, the coefficients gq(n) for the M-to-1 decimator are
Bn (n) = a(n) = h(n),
for all n and m 76)
Alternatively, according to the transpose network of Fig. 3.20, the coefficients ofthe
‘M-to-1 polyphase decimator are
Pelt) = h(a +p), forp=0,1,2,...,M~1,andalln (3.77) Sec. 3.3 Practical Structures for Decimators and Interpolators 153,
where p denotes the pth polyphase filter. Thus the polyphase filters p,(n) for the
‘M-to-I decimator are equal to the time-varying coefficients gq(n) of the transpose
(interpolator) of this decimator.
From a practical point of view itis often convenient to implement the polyphase
structures in terms of a commutator model. By careful examination of the interpolator
structure of Fig. 3.17 we can see that the outputs of each of the polyphase branches
contributes samples of y(m) for different time slots. Thus the I-to-L sampling rate
expander and delays can be replaced by a commutator, as shown in Fig. 3.21. The
Commutator rotates in a counterclockwise direction starting with the zeroth-polyphase
branch at time m
rato
ato) ein
rita} mn =0
Dyled ~e
no wa
| pated
reso
Peale)
5, f, er,
Figure 3.21 Countrclockwise commutator model for & tov intecpolatoc.
A similar commutator model can be developed for the M-to-1 polyphase deci-
‘mator by starting with the structure of Fig. 3.20 and replacing the delays and M-to-1
sampling rate compressors with a commutator. This leads to the structure of Fig. 3.22.
‘Again the commutator rotates in a counterclockwise direction starting with the zeroth
polyphase branch at time m = 0.
3.3.4 FIR Structures with Time-Varying Coefficients
for Interpolation/Decimation by a Factor of L/M
In the previous two sections we have considered implementations of decimators and
interpolators using the direct-form and polyphase structures for the case of integer
changes in the sampling rate. We obtained efficient realizations of these structures by
commuting the filtering operations to occur at the low sampling rate. For the case of
‘network that realizes a change in sampling rate by a factor of L/M, itis difficult to 156 Multirate Processing of Digital Signals Chap. 3
sale
ol vel
alah
m0 sind
stn xin
c pain |
Kurta |
outa
A me eM
Figure 3.22 Counterclockwise commutator model for an M-to- decimatoe.
achieve such efficiencies. The difficulty is illustrated in Fig. 3.23. If we realize the
|1-t0-L interpolation part of the structure using the techniques described earlier, then
we are faced with the problem of commuting the M-to-1 sampling rate compressor into
the resulting network (Fig. 3.23a). If we realize the decimator part of the structure
first, then the l-to-L sampling rate expander must be commuted into the structure (Fig,
3.23b). In both cases difficulties arise and we are faced with a network that cannot be
implemented efficiently simply using the techniques of commutation and trans-
position.
Efficient structures exist for implementing a sampling rate converter with a ratio
in sampling rates of L/M, and in this section we discuss one such class of FIR
structures with time-varying coefficients [10]. This structure can be derived from the
time-domain input-to-outputrelation of the network, as derived in Section 3.2, namely
yn) = eno(| FI - /) 78)
where
Bm(t) = AnL + ((mM)),), for all m and all n (3.79)
and h() corresponds to the lowpass (or bandpass) FIR prototype filter. It will be
convenient for our discussion to assume that the length of the filter h(k) is a multiple
of L, ie,
N= QL 6.80)
where Q is an integer. Then all of the coefficient sets g,,(n),m = 0,1,2,...,L—1
contain exactly Q coefficients. Furthermore, gq(n) is periodic in m with period L,
0, #1, =2,.. G81)
Bm) = Sarre (Mt), See. 3.3 Practical Structures for Decimators and Interpolators 155
stn)
vie
sate PY er Lf fi Lei
Figure 3.23 Possible realizations of an /M sampling rte converter.
‘Therefore, Eq. (3.78) can be expressed as
ym) = & centro | z | a (82)
=
Equation (3.82) shows that the computation of an output sample y(mm) is obtained as
‘a weighted sum of Q sequential samples of x(n) starting at the x((mM/L) sample and
going backward in n sequentially. The weighting coefficients are periodically time-
varying, so the ((m);, coefficient set gay, (7), m = 0, 1,2,... ,Q — 1, is used for
the mth output sample. Figure 3.24 illustrates this timing relationship for the n — 0
He
rim | (LI) | suming
> | BY | om a
° ° 3 sinh
2 : 2
2 2 °
‘ 3 : a6,
: 3 2
* = 7 o 1 2 3 +o”
a o
Figure 3.24 Timing relationships berween y(m) and x(n) forthe case M = 2, 186 Multirate Processing of Digital Signals Chap. 3
term in Eq. (3.82) and for the case M = 2 and L = 3. Figure 3.24(a) shows the index
values of y(n), x(lmM/L1), and gya,(0) for m = 0, 1, 2,.. . , 6. Figure 3.24(b)
illustrates the relative timing positions of the signals y(m) and x(n) drawn on an
absolute time scale. By comparison of parts (a) and (b) we can see that the value
x({mi/L) always represents the most recent available sample of x(n), i.e., y(0) and
(1) are computed on the basis of x(0 — n). For y(2) the most recent available value
‘of x(n) is x(1), for y(3) itis x(2), and so on,
Based on Eq. (3.82) and the preceding description of how the input, output, and
coefficients enter into the computation, the structure of Fig. 3.25 is suggested for
realizing an L/M sampling rate converter. The structure consists of the following:
1. AQ sample “shift register” operating at the input sampling rate F, which stores
sequential samples of the input signal
2. A direct-form FIR structure with time-varying coefficients (gya(n), n= 0, 1,
2, ...,Q~ 1) that operates at the output sampling rate (L/M)F,
3. A series of digital “hold-and-sample” boxes that couple the two sampling rates.
‘The input side of the box “holds” the most recent input value until the next input
value comes along; the output side of the box “samples” the input values at times
n= mid/L. For times when mM/L isan integer (i.e., input and output sampling
times are the same), the input changes first and the output samples the changed
ing
oi n
martin
penllion: Mant, 1
(4)
st tS Hm
oe Beem
a
S sa
“(t*)
ate -
Da Seon, (= 1)
*([4]-0+1) BN
g
‘ut
: wi,
Figure 3.25 ficient sructare for realizing an L/M sampling rte converter. Sec. 3.3 Practical Structures for Decimators and Interpolators 187
It should be clear that the structure of Fig. 3.25 is an efficient one for imple-
‘menting an (L/M) sampling rate converter since the filtering operations are all per-
formed at the output sampling rate with the minimum required number of coefficients
used to generate each output.
Figure 3.26 shows a diagram of a program configuration to implement this
structure in a block-by-block manner. The program takes in a block of M samples of
the input signal, denoted as x(n"), n’ = 0, 1,2, ... ,M — 1, and computes a block
of £ output samples y(m') m’ = 0, 1,2,..., L — 1. For each output sample time
m', m'=0, 1, 2,...,L~ 1, the Q samples from the state-variable buffer are
multiplied respectively with Q coefficients from one of the coefficient sets gq-(n") and
the products are accumulated to give the output y(m"). Each time the quantity
[m'M/L, increases by 1, one sample from the input buffer is shifted into the state-
variable buffer. (This information can be stored in & control array.) Thus after L output
values are computed, Mf input samples have been shifted into the state-variable buffer
and the process can be repeated for the next block of data. In the course of processing
one block of data (M input samples and L output samples), the state-variable buffer
is sequentially addressed L times and the coefficient storage buffer is sequentially
addressed once. A program that performs this computation can be found in [16].
x vie)
Inout butter Fp] Ovo tter
ie samt! (Comper
pace
we
fo] 2. ot] a.
Se
(eee! ah) gin Paty
ection storage
[eset ot Omen echt
igure 3.26 Block diagram of a program structure to implement the signal flow
‘raph of Fig. 3.25 ina biock-by-block maaner.
3.3.5 Comparison of Structures
{In Section 3.3 so far, we have discussed three principal classes of FIR structures for
decimators and interpolators. In Section 3.5 we will discuss multistage cascades of
these structures and show how such cascading can lead to additional gains in com-
putational efficiency when conversion ratios are large. A natural question to ask at this
Point is which of these methods is most efficient. The answer, unfortunately, is
‘nontrivial and is highly dependent on the application being considered. Some insight
and direction, however, can be provided by observing a few general properties of the
classes of structures discussed here. 188 Multirate Processing of Digital Signals Chap. 3
‘The direet-form structures have the advantage that they can be easily modified
to exploit symmetry in the system function to gain an additional reduction in com-
putation by a factor of approximately two. The polyphase structures have the aivan-
tage thatthe filters pp(n) can be easily realized with efficient techniques such as the
fast convolution methods based on the FFT. As such, this structure has been ‘ound
useful for filter banks [10]. The structures with time-varying coefficients are particu-
larly useful when considering conversions by factors of L/M.
3.4 DESIGN OF FIR FILTERS FOR DECIMATION
AND INTERPOLATION
In the previous discussion we have assumed that the filter 4(K) approximates some
ideal lowpass (or bandpass) characteristic. Consequently the effectiveness of these
systems is directly related to the type and quality of design of this digital filter. The
purpose of this section is to review digital filter design techniques that are especially
applicable to the design of the digital filter in sampling rate-changing systems.
‘The filter design problem is essentially onc of determining suitable values of i(k)
to meet given performance specifications on the filter. Such performance specifica-
tions can be made on the time response f(b) or the frequency response of the filter
(a) defined as
H(o) = ¥ hWe™ = H(2) |e (3.83)
Before proceeding to a discussion of filter design techniques for decimato:s and
interpolators, it is important to consider the ideal frequency domain and the time-
domain criteria that specify such designs. It is also important to consider, in more
detail, the representation of such filters in terms of a single prototype filter or as a set
of polyphase filters. Although both representations are equivalent, it is sometimes
easier to view filter design criteria in terms of one representation or the other.
3.4.1 Relationship Between the Prototype Filter
and Its Polyphase Representation
{As discussed in Section 3.3, the coefficients, or impulse responses, of the polyphase
filters correspond to sampled (and delayed) versions of the impulse response of the
prototype filter. For a I-to-L. interpolator there are J. polyphase filters and they are
defined as (see Fig. 3.18)
Pon) =h(p+nb), p=0,1,2,...,L—l,andalln (3.84)
Similarly, for an M-to-1 decimator there are M polyphase filters in the polyphase
structure and they are defined as
Pn) =h(p+ nM), p=0, 1,2, »M-\l,andalin (3.85)
Taken as a set, the samples p,(n)(p = 0,1, ..., L — 1, for an interpolator or Sec. 3.4 Design of Fi Filters for Decimation and Interpolation 159
p= 0,1,...,M~ 1, for a decimator) represent all of the samples of A(K). Since
the development of the filter specifications is identical for both cases (1-to-L inter-
polators and M-to-1 decimators) we will consider only the case of interpolators. The
results for decimators can then simply be obtained by replacing L by M in the
appropriate equations.
‘The samples i(k) can be recovered from p,(n) by sampling rate expanding the
sequences p(n) by a factor L. Each expanded set is then delayed by p samples and
the L sets are then summed to give /(&) (the reverse operation to that of Fig. 3.18).
If we let p, (8) represent the sampling rate expanded set
ro(k/L), k=0, +L, £2L,
Belk) = 1G" carve (3.86)
then (A) can be reconstructed from p(k) via the summation
HW = > Ale p) (6.87)
‘The z-transform #1 (2) of the prototype filter can similarly be expressed in terms of the
transforms of the polyphase filters P,(z). It can be shown that
H@ = 5 2°R 4) (3.88)
Finally, the z-transform P,(z) can be expressed in terms of H(z) according to the
following derivation. If we define a sampling function 6,(R) such that
1, k=p.ptlp+2,.
30 = {i ph (3.39)
= Scone 090)
then the sampling rate expanded sequences ,(K) in Eq, (3.86) can be expressed as
Belk) = 6, (Ok (k) = ord § efatmalt (G.91)
‘The z-transform P,(z) can then be expressed in the fon
B= 3 ple = 3 plo + ae (3.92)
ni ta saan of raaindn kp a,
ne = 3 Ble 6.93)
Combining Eqs. (3.91) and (3.93), we get
20-73
fet
5! nWenett-aity tent @.94)
a 160 Muttirate Processing of Digital Signals Chap. 3
Letting z = e and rearranging terms gives
lay = 2S gtemente $ perenne
TS ferro Hw —2eO/L, p= 01,2)... ,L-1
6.95)
Equation (3.95) shows the relationships of the Fourier transforms of the polyphase
filters to the Fourier transform of the prototype filter.
3.4.2 Ideal Frequency-Domain Characteristics
for Interpolation and Decimation Filters
In the previous sections we have assumed thatthe filter h(k) approximates some ideal
lowpass (or bandpass) characteristic. We will elaborate on these “ideal” characteristics
in somewhat more detail in the next two sections. In practice itis also necessary to
specify a performance criterion to measure (in a consistent manner) how closely an
sctual filter design approximates this ideal characteristic. Since different design ‘ech-
niques are often based on different criteria, we will consider these eriteria as they
atise.
Recall from the discussion in Section 3.2 that the interpolator filter A(&) must
approximate the ideal lowpass characteristic defined as
rorya (Gels al
0. otherwise 6.96)
where the subscript / refers to the “ideal” characteristic.
By combining Eqs. (3.95) and (3.96) its possible to derive the equivalent ideal
characteristic P, (a) that are implied in the polyphase fillers. Note thatthe frequency
Variable «refers to A(m) whereas the frequency variable @ = o' refers to the
polyphase filters p,(n). Because of the constraint imposed by Eq. (3.96), oaly the
© = 0 tem in Bq. (3.95) is nonzero, and the equation simplifies tothe form
Lait
Pasleo) = 7 eM Hiy(w/L) sn
sett, 9 =0,1,2,...,L-1
Equation (3.97) shows that the “ideal” polyphase filters P,,(n) should approximate
allpass filters with linear phase shifts corresponding to fractional advances of p/L
samples (p = 0, 1, 2,... . £ ~ 1) (ignoring any fixed delays that must be intro-
duced in practical implementations of such filters)
Tn some cases it is known that the spectrum of x(n) does not occupy its full
bandwidth. This property can be used to advantage in the filter design, and we will
see examples of this in the next section on cascaded (multistage) implementations of
sampling rate changing systems. If we define «, as the highest frequency of interest
in X(o), ie.,
IX@)|<¢ forw>|ol>o 3.98) Sec. 3.4 Design of FIR Filters for Decimation and interpolation 161
where ¢ is a small quantity (relative to the peak of |X (a) |), as shown in Fig. 3.27 (for
L = 5). In this case, the ideal interpolator filter has to’ remove only the (L— 1)
repetitions of the band of X(«) where |X(a}| > ¢. Thus inthe frequency domain, the
ideal interpolator filter satisfies the constraints
no = {+ 0S |o'l