Introductiotl

Several analyses of the roundof€-noise output from a

digital filter with fixed dynamic range (i,e., implemented
Roundof f-Noise using fixed-point arithmetic) have recently appeared in the
literature [ 11-[4]. Comparisons between theroundoff-
Analysis for Fixed- noise outputs of different circuit configurations for a digi-
tal filter have been of particular interest because of the
Point Digital Filters desire to maximize some measureof the output signal-to-
noise ratio. A general framework for these comparisons
Realized in Cascade has been provided by the author [ 2 ] , and two configura-
tions forthe direct form of a digital filter have been
or Parallel F o r m analyzed and compared using this approach. In this paper
the techniques in [2] are applied to two configurations
each forthe cascade andparallel forms, and thesequential
LELAND B. JACKSON, Member, IEEE
ordering of the second-order sections comprising the
cascade form is also investigated.
Bell Telephone Laboratories, Inc.
Murray Hill, N. J.
Background
The interaction between theroundoff-noise output
froma digital filter and the associated dynamic-range
limitations is investigated in [2] for the case of uncor-
Abstract
related rounding errors from sample to sample and from
The roundoff-noiseoutputsfrom two transposeconfigurations,each one error source to another. The spectrumof the output
for the cascade and parallel forms of a digital filter, are analyzed for roundoff noise from fixed-point implementations is
the case of uncorrelated roundoff noise and fixed dynamicrange. readily shown to be of the form
Corresponding transpose configurations are compared on the basis of
the variance, or total average power, and the peak spectral density of
I
N,(w) = uo2 kj’ Gj’(w) 1’ (1)
j
the output roundoff noise. In addition to providing general computa-
tional techniques to be employed in choosing an appropriate configura- where theGj’(w) are scaled transfer responses fromcertain
tion for the digital filter, these results also indicate useful “rules of “summation nodes” in the digital circuit to the filter out-
put. go2is the variance of the rounding errors from each
thumb” relating to this choice of configuration. Included are indica-
multiplier (or other rounding point), and the k / are inte-
tions of good (althoughnot necessarily optimum) sequential orderings
gers indicating the number of error inputs to the respec-
and pole-zero pairings for the second-order sections comprising the
tive summation nodes.
cascade form. Computationalresults are presented which indicate that Defining Fi’(w) to be the scaled transfer response from
the analysis is quite accurate and useful. the input to the ith “branch node” at which a dynamic-
range constraintis required, constraints of the form

I[ 1
for p 2 1 are then derived, where F([ is the L, norm of
the response Fil(w) defined by

It canbe shown [ 5 ] that for Fi’(w) continuous(as is

always the case fordigital filters), the limit of (3) asp+ cc
exists and is given by

Manuscript received January 12, 1970.

where us is theradiansamplingfrequency 2r/T. The
This paper is taken in part from a thesis submitted by the author
in partial fulfillment of the requirements for the degree of appropriate value of p in (2) is determined by assumed
Doctor
conditions on thespectra of the input signals to the filter.
of Science to the Department of Electrical Engineering, Stevens In-
stitute of Technology, Hoboken, N. J. [l]. The effect of (2) is to bound the maximum signal ampli-
The author is now with Rockland Systems Corp., Blauvelt, N. Y . tude (for deterministic inputs) or the maximum average
10913. power (for random inputs) at the ith branch node.
IEEE
ELECTROACOUSTICS
TRANSACTIONS
AND
AUDIO
ON VOL. AU-18, NO. 2 JUNE 1970 107
 As implied above, the prime is used to indicate that the
filter has been scaled to satisfy (2). It is intuitively clear
that topreserve the greatestpossible signal-to-noise ratio,

-
this scaling should reduce the magnitude of Fi’(w) no more
than is required (orincrease it as much as possible, as the
case maybe). In otherwords, Fi’(w) should actually
satisfy u In)

IlFi’Ilp = 1. (5)
This condition willbe satisfied if the scaling factors si,
defined by
Fi’(U) = SiFi(U), (6%)

are given by
si = 1/llFil\p* (6b) Fig. 1. Generaldigital-filter model.

The unscaled responses Fi(w) and Gj(w), as well as the

overall filter response H(w), are depicted in Fig. 1 where
H(w) = .FI*(ejw*)
The Parallel Form
Fi(w) = Fi*(ejwT) (7) Two transpose configurations for the parallel form of a
G ~ ( w=) Gk*(e+”’). digital filter are shown in Figs. 2 and 3. The parallel form
corresponds to an H*(z) of the general form [6], [7]
This notation will be used throughout this paper. Thatis,
I
for any z transform A*(z) which converges for zI = 1, the
corresponding Fourier transform is given by
A(w) = A*(efuT). However, in order to satisfy (5) in the general case, addi-
tional scaling multipliers p i have been added to the con-
The sequences {ej(n)) in Fig. 1 represent the roundoff- figurations in Figs. 2 and 3 (called forms 1P and 2P,
noise “inputs” atthe summationnodes; while the se- respectively). Hence, in form IP, H*(z) is actually realized
quences { vi(n>] arethetotal“outputs”fromthose as
branchnodes at which dynamic-rangerestraintsare
required.
A state-variabledescription is then employed in [ 2 ]
to formulate the general concept of “transpose configura-
where the scaling multipliers pla are given by
tions” for a digital network and to illustrate the useful-
I
ness of this concept in digital-filter synthesis. A particu- P l i = I P S < = 1/11I P F i ; l p
larly important result is that for a given unscaled con-
and
figuration with transpose responses Fi(w) and Gj(w), as
described above, the responses Fit(w) and Gjt(w) for the 1yki’ IY~</PI~ == 0, 1. (11)
corresponding transpose configurationare given by
The transfer function in(10) obviously equals that in (9),
F i t ( w ) = Gi(w), but now the transferresponses Fi’(w) [from thefilter input
to thenodes marked by (*)I also satisfy (5). The unscaled
and transfer functions lPFi*(z)are given by
Gjt(w) = F ~ ( w ) . (8)

Hence, although the overall transfer responses H(w) for

these twoconfigurationsare the same, theirroundoff- where
noise outputs and/or dynamic-range limitations can be
quite different, in general. The transpose configuration is pi*(z) = flziz-2 + p1iz-1 + 1,
obtained by simply reversing the direction of all branches and thus
in the given networkconfiguration, and the poles and
zeros of the network are thus realized in reverse order in
the transpose configuration.

108 IEEE TRANSACTIONS ON AUDIO AND ELECTROACOUSTICS JUNE 1970

Fig. 2. Parallel form 1P with scaling. Fig. 3.
w
Parallel form 2P withscaling.

The sealing for parallel from2P is shown in Fig. 3. The in the multipliers. In particular, scaling may be restricted
transfer function H*(z) for this formis also given by (lo), t o powers of 2 in a binary system, in which case the over-
except that thescaling multipliers pli are replaced by all transfer function can be maintained exactly and no
extramultiplicationhardware is required (just shifting
pzi = 1/2PSi = / j P P F i / j p (134
operations). In this case, ( 5 ) is not satisfied generally, but
and now, of course, we can ensurethat
2yki‘ ZYki/P2i k = 0, 1. (13b) 3 < j/Fi’llP I 1. (15)
Thetransfer responses to theform-2Pbranchnodes Since such finite-accuracy considerations arenot expected
marked by (*) are to influence significantly the noise comparisons to follow,
wewill assume forthepurpose of analysis that ideal
scaling is performed as indicated in (1 1) and (13), and thus
that ( 5 ) is indeed satisfied.
The roundoff-noise output fromeach of the two paral-
where
lel forms is readily analyzed, as one would expect, be-
Yi”’(2) = yli’x-’ + yoit cause this output noise is simply the sum of the noise pro-
duced by the individual second-order sections comprising
and thus Z ~ F i ’ (indeed
~) satisfies ( 9 , as required. the filter. Referring to Fig. 2, the transfer responses
Note that in form 1P the scaling is accomplished via the Gj’(w) for form lP, corresponding to the roundoff-noise
pli multipliers and is then compensated for in the ykil inputs e&), are easily seen to be
multipliers; whereas in form 2P, the scaling is performed
in the Y k i ’ multipliers and is then compensated for in the
pzi mutlipliers. Note also that scaling has increased the
required number of constant multipliers in these parallel
IPGM+~’(w) = 1.
forms to 2N+M+ 1 from 2N+ 1, in the general case. We
will find that this applies to the cascade forms as well. Assuming that the coefficients P k j , lykjl, and plj for all
Practically speaking, however, it is assumed that in most k and j are nonintegers, the weights kj’ corresponding to
cases scaling will be performed with only a few (extra) bits the Gj’(w) are simply

JACKSON: ROUNDOFF-NOISE ANALYSIS FOR FIXED-POINT DIGITAL FILTERS 109

 1pJC,11+: = s+1 We now compare the output roundoff noise for form
where [ ] denotes the "integer part of," and we have 1P given by (20)with that for form 2Pgiven in (24).First
assumed that if the order ( N ) of the filter is odd, the Mth note that if N is odd, the noise contribution of the one
section is the one (degenerate) first-order section. As indi- degenerate first-order section (theMth section)is the
cated previously, with all noninteger coefficients the kj' same for both forms. This is seen by noting that, in this
for each of the configurations to be considered satisfy case, lPk.TIli/ = 2, while
= 2pk,pf'

kj' = 25 + ill + 1. (18) ( l p G . l r ' ( W ) ) J v odd = ( w G . v ' ( W ) ) . v odd

i (25)
Using the fact that M = [(N+1)/2], (17) is readily shown
to satisfy (18).
We next note that N = M+ [N/2] and hence that
The effect of scaling on the transfer responses l p G i / ( ~ )
is seen by substituting (1 1) into (16) to yield
(26)

Therefore, defining
or, from (12), X l ( W ) = uo2

Turning to parallel form 2P, shown in Fig. 3, we note 2pLyy(W)

=iyl(w)+uo2
that only one noise input e j ( n ) has been shown for each
second-order section, rather than two as would be indi-
Consider now the variance, or total average power, of
cated by the number of summationnodes per section.
the output roundoff noise, given by
This is done for convenience, and it is possible because
the two summation nodes are separated only by a delay
and, therefore, actually comprise a singleThe
summation. (29)
corresponding transfer responses G,'(w) are then given by
(since N,(w) is a nonnegative real function). From (28)
2p~i'(W) = --
Pzj
^ ,\
j = 11 2 , . . . , 1' 1. (21) and the definition of I] i ! 2 in (3), it follows that
Pi(0)
But from (13) through (15) this is just

j = I, 2 , . . . , M. (22a) whereas

Note the differencebetween(19b) and (22a). The transfer +

1 ~ 2 p N y l ~=1 ~ ~ N 1 1 1 1 uo2
[N!21

j=1
4
JP~
Ii-'12.
I ! -1yj
11
1
2'I

P I ~j
1,
1 2
(30b)
response for the input e.,l+l(n) is, of course, just
Hence, letting
2pG.~+{ (w) 1 (22b)
as in (16). (31)
The noise weights kj' for form 2P are
2PlCj' = 4 j = 1 2 ' . ,Jl- 1
, l ' we have the followingsimpleresult forthe case p =q = 2:
2pk.v' = 4 - 2 [ 2 x - N ] (23) < [ N / 2 ]* form 2 P
(32)
*pk.,;+I = x+1
and again (23) is readily shown to satisfy (18). The total Note, however, that for p =2, the output noise variances
roundoff noise in the output of form 2Pis then IINylilfor these two forms candiffer at most by a factor of

110 IEEE TRANSACTIONS

AND ON AUDIO ELECTROACOUSTICS
JUNE 1970
4/3, or 1.25 dB, and thus it really makes little difference The Cascade Form
in this case whether form 1P or form is2Pchosen.
The two most commonly employed (transpose) con-
The comparison of (30a) and (30b) in the case of p = w
figurations forthecascadeform of a digital filter are
is not so obvious as forp = 2, but it is basically the same.
shown in Figs. 4 and 5 and are designated as forms 1D
Let aojbe a frequencyat which the maximum of 1 l/Pj (a)\
and 2 0 , respectively. (The letter D indicates that each
is achieved; that is,
second-order section is realized in direct form, as opposed
to other possible configurations [l].) The cascade form
(33) corresponds to anH*(z) of the general form [ 6 ] , [7]

I
Because of the very peaked nature of l/Pj (a) relative to 1
1y j ( w ) / , we maythenapproxi.mate the L, norm of
y j ( w > / P ~(a)by
However, in the scaled versions of forms ID, I= 1,2,
H*(z) is realizedas
(34)

and with this approximation(31) becomes

where for form 1D (letting IDSM+I = ao)

1010’ = lDsl
Hence, in this case
1Dsi+l
pnej wej (35b)
lCYyi*’(Z) = -a;*(z) i = 1,2, * * * ,M ;
lDsi

and (32) again applies. As before, it really makes little

while for form 2D (letting 2 ~ = ~1) o
differencewhich parallel form is chosen.However, ex-
perience has shown that usually ,z6j> 1, and hence form ZDSi
1P is generally to be preferred over form 2P forp = a. ZQli*’(X) = ---aCri*(z) i = 1, 2, . . * ,M
PDSi-1
The other norm of N,(w) to be considered is the L,
norm, i.e., the maximum of N,((J) over all (J. Here we 2ao‘ = -- .
a0

1
actually compare bounds on IN,lloofor forms 1P and 2P, 2DSM
ratherthanexpressionsorapproximations
itself. Although comparisons based on bounds
do not, of course, necessarily hold for
for

l~lV,~lc0
l Nvllm
for Nv , The correspondingtransfer responses to the branchnodes
itself, experi- marked by (*) are then
mental results supporting the application of these com-
parisons to llN,ll ,are presented in a latersection.
Again employing Nl(w) as defined by (27) in (20) and
(24), the following bounds on the L, norms of N,(w) for
forms 1P and 2P are readily obtained: Z D F i ’ ( a ) = ZDSi
Cxj(W)
__ n
j=1 P j ( w )
i = 1,2, * ’ * ,&I

where wedefine (.) = 1. Hence,from the definition

of si in (6), the LDF~’(w)
in (39) satisfy ( 9 , as required.
Scaling is accomplished in cascade forms 1D and 2 0
via the altered multipliers lahi’ within the second-ordersec-
j=1
tions, ratherthan by inserting multipliers between the
where ,*ej is given by(3 1). Utilizing the approximation for second-order sections, for the following reasons. In form
p s f 3 j in (35), the conclusions containedin (32) apply in this 1D the insertion of additional multipliers between the
case as well. As indicated previously, we usually find that second-order sections would require that an additional
z,Oj> 1 and almost always that ,&> 1. overflow constraint be added at the output of each sec-
11
Therefore, for all four cases considered (i-e., NullTfor tion. In a large number of cases (although not in every
r = 1, w and p = 2, w ) there is really little advantage to be case), these additional overflow constraintswould sig-
gained by using form 1P over form 2P (orvice versa); but nificantly reduce the potential signal-to-noise ratio which
as a general “rule of thumb,” form 1P is somewhat to be could otherwise be realized in the filter. In form 20, on
preferred. The situation is much more interesting in the the other hand,it is readily apparent thateven if the addi-
case of the cascadeform,where large differences are tional overflow constraints (now required at the inputs to
possible between the roundoff-noise outputs of the form- the second-order sections) donot actually change the
1 andform-2 configurations. Inaddition, there is the required scaling, the total rounding error in the outputsof
question of how to sequentially order the cascade-form the ali and azi multipliers must be greater with scaling
sections. (and rounding)between sections, rather thanwithin them.

JACKSON: ROUNDOFF-NOISE ANALYSIS FOR FIXED-POINT DIGITAL FILTERS 111

 (*I I*l
u(ni ylnl

I U I

Fig. 4. Cascade form 1 D with scaling.

Fig. 5. Cascade form 2D with scaling.

I
II F((w1- I +Gk(U+---q

The transfer responses (?/(a)for form 1D are given by ,&’,’(u) i s seen by substituting the expressions for laal(u)
in (38) into (40), which yields

or, from(39) and (6),

Assuming that the coefficients l a k i and P k i for all k and i
are nonintegers, the corresponding weights kj’ for form
1D are
mkl‘ =3 j = 1, 2, ‘ ’ * , M.
1Dkj’ =3 j = 2,3, . . ‘ ,114 - 1 In cascade form 2 0 , shown in Fig. 5, wehave again
(41) chosen to represent the roundoff-noise inputs to each
1ok.w’ = 5 - [ 2 X -N]
second-order section by a single error source ej(n). The
lDkl>r,l’ = 3- [2M - N ]
corresponding Gj’(o) are then
where we have assumed, for the present, that if the order
( N ) of the filter is odd, the Mth section is the one first-
order section. Equation (41) is readily shown to satisfy
(18). The effect of scaling on the transfer responses

112 IEEE TRANSACTIONS ON AUDIO AND ELECTROACOUSTICS JUNE 1970

 But from (38) and (39)becomes
this and hence that

j=1,2,.*-,M.
The transfer response for the input eM+l(n)is, of course,
just
ZDGM+~’(W) == 1. (45b)
(49)
The noise weights kj’ for form 2 0 are
2Dkjl=5 j=1,2,.-.,M-l
where we have employed the fact that
2DkM’ =5 - 2[2M - N ] (46)
2DkM+< = 1. 2Dt kM-Z+; = 2Dkl’ 1 = 1, 2, * * ‘ , fl!f.
The variance, or total average power, of the output Comparing ]lNu]llfor form 1D as given in (47a) with
roundoff noise from forms 1D and 2 0 is derived from that for form2 0 , in (49), we see that these twoexpressions
(1) and (42) through (46) to be differ in only two respects: first, the functions appearing
in the Lp and Lz norms are reversed; and second, the kj’
M
correspond to forms 1 D and2 0 , respectively. From (41)
11 lDNvII1 = 60’ { l D k M + l ) f a0 lDk/
and (46), the kj’ for these two forms are related in the
following way:
1Dkl’ = 2Dk< -2
1Dkj‘ = z~kj’ j = 2, 3, * * , l%f - 1
1DkM’ = ZDkM’ + [2hf - N ] (50)

1DkM+1’ = 2DkM+1‘ + 2 - [2M - N ] .

Therefore, for p = 2 , /iN,Ill differs for forms 1D and 2 0 ,
only by virtue of the kj’ for j = 1, M , and M+ 1. And for
There are two separate questionsto be considered in the N even, only ki and kM+i are different for these two
minimization of (47): first, which of these two forms is forms. Hence, as was the case for the parallel form, it
best; and second, what should be the sequential ordering really makes little difference which form is used when p =2
of the M sections. We investigate the first question in this (and we are considering llN,l[ 1).
section, and the second in the next. In thecase ofp = 00,however, there can bea significant
Cascade forms D1 and 2 0 are related via the conceptof
transposeconfigurations,as previously described. In
differencebetween // 1 d V l l 1 and 11
2DtNulI duetothe
reversal of the L, and Lz norms in (47a) and (49). Ne-
particular, assuming that the ai(.) and pi(.) appearing in glecting the small difference due to the different k / for
(47) are the same for both forms, form 2 0 corresponds to j = 1, M+ 1, and (for N odd) M , the choice between forms
the section-by-section transpose of form 1D. Therefore, 1D and 2Dt is determined by the relative magnitudes of
the sequential ordering of the individual sections in form pqOj and 4 p O j for p = UJ and q= 2, where
1D is not reversed in form 20. There is, of course, another
form-2 configuration (denoted as form 2Dt) which cor-
responds to the ouera2Z transpose of the given form-ID
configuration, and likewise, a form-1 configurationwhich
corresponds to the overall transpose of the given form-2D Note thatthis definition for pnOj in the case of the D forms
configuration. The sequential ordering of the individual is consistent with that given in (31) for the P forms be-
sections of forms 1D and 2 0 is thus reversed in forms cause in both cases
1Dt and 20,.
Since we will separately investigate the ordering of the Psej = I I ~ ~ =~ I ~ ~ ~ ~ (52)
~ Q ~ ~ ~ ~
112t~j/I~I/2t~jll~.

sections of the cascade forms, we may here compare the In particular, then,
noise output from form 1D with that from form 2 0 or
2Dt. We will find it most convenient to consider form 2 0 ,
in the case of p = 2. Denoting the numerator and denom-
inator factors of the Dt forms by ait(.) and pit(#), respec-
for p = co and q=2.
tively, we then note that
Although it wasespecially convenient to compare form
pit((.) = PM-~+~(w) i = 1, 2, * * * ,M 1D with form 2 0 , in the case of p-2, another useful
ROUNDOFF-NOISE
JACKSON: FOR ANALYSIS FIXED-POINT DIGITAL FILTERS 113
comparisoncan be made using form 2 0 when p = 03. IlNYll,
Noting that r=l r - m

p.2 EITHER FORM ID

we may rewrite (47) in thefollowing form: IIF/ Ilp

p = a FORM 2 D EITHER

Fig. 6. General rules for selection of cascade-form con-

figuration.

Hence, neglecting the difference in (54a) and (54b) due to

the different kj’ for j = 1, M+ 1, and (for N odd) M , the
primary difference between these two expressions results
from the occurrence of the extra factors aj(u)multiplying
,,Gj(w) in the former and ~ D F ~ (inwthe ) latter for all j . (55c)
The implication of the extra factorsaj(w) in (54) for the
case o f p = 03 is as follows. In (54a) the factorsa,(.) affect
the Lz norms, while in (54b)they affect the L, norms. But where, as before,
the L, norm of a function “concentrates” exclusively on
the maxima of that function, whereas the L, norm of a
function reflects the rms value of that function over all
argument values. Therefore, the effect of the extra a j ( w )
in(54b) results from thealteration of the maxima of and
lDF,(w) in O ~ ~ ( W ) ~ D while
F ~ ( Uin) ;(54a), the effect concerns
the difference in~ D G ~ (and w ) a , ( ~ ) ~ ~ Gover
~ ( wall
) w.
Intuitively, one expects that th.e former effectis po-
tentially much greater; that is,in some cases the extra Therefore, it is now the case of p = co for which the only
a j ( w ) should affect the L, norms in (54b) much more than difference between these bounds for forms 1D and 2 0 ,
the Lz norms in (54a). The experimental evidence pre- lies in the kj’ forj= 1, M+ 1, and (for N odd) M . And for
sented later will support thisconclusion. In particular,we p = 2, the attenuation provided by the extra factors a,(.)
will find that when the a j ( w ) provide significant attenua- in (55a) and (55b) should, when significant, benefit from
tion in theneighborhood of the maxima of thecor- 1D over form 2 0 .
responding responses lDFj(w), [ lNy:l can be significantly The general results of this section are thussummarized
less for form 2 D than for form 1D (when p = cc ). On the as follows.Significantdifferencesin ~ ~ N for
u ~cascade
~ T
other hand, when the a j ( w ) do not provide such attenua- forms 1D and 2 0 (or 2 D t ) can occur in the cases of r = 1,
tion, there is not a great difference between the liNull1 for p = cc and r = a3 ,p = 2. In the first case, form 2 0 tends to
these two forms. be superior; whereas in the latter case, form 1D tends to
Turning now to the consideration of IINyIIE(Le., the be superior. For p = Y+ 1, r = 1, oc ; however, there seems
maximum of N,(w)over all w), the analysis goes through to be little difference between these two forms, A ch.art
exactly as above, but with the L2norms being replaced by incorporating these conclusions is given in Fig. 6.
L, norms. As in the case of the parallel forms, the com-
parisons are actually based on bounds for 1 ,N,Iim,rather
than on llNu!l itself, with intuition and experimental evi- Sequential Ordering of Cacsade Sections
dence supportingthisapproach. By analogy with(54) In this section we present a heuristic discussion of the
and (49), these bounds on llNu!l for forms 1D , 2 0 , and sequential ordering of the second-order sections compris-
2 D t are readily derived to be ing the cascade D forms, based upon theanalytical results
of the preceding section. We must not only consider the
ordering ofgiven second-order factors a,*(z)lPi*(z), but
also the pairing of each numerator factor ai*(z) with a
denominatorfactor pZ*(z). An analytical technique to
determine the best sequential ordering, other than a com-

114 IEEE TRANSACTIOXS ON AUDIO AND ELECTROACOUSTICS JUNE 1970

plete enumeration and evaluationof all possible permuta- IlNYll,
tions, has not yet been devised, but it is possible to state r=l r = w
some generalrules which can be quite helpful in the design DECREASING P i
of a given digital filter. In view ofthe fact that there are up p s 2 ----__-
WITH FORM 2 0
INCREASiNG P i
DECREASING P i
to M ! possible pairings of the ai*(z) and pi*(z) and M ! WITH FORM ID
possible permutations of each setof ai*(z)/pi*(z),it is IlP DECREASING PI
important todetermine such general rules, where possible.
P = OD INCREASING PI --- - - --
WITH FORM 20
Consider the relations in (54) and (55) for llNu\],.,r= 1 INCREASING Pi
and r= CQ , respectively. Each term of these expressions WITH FORMID

corresponds to the noise contribution from one section of

Fig. 7. General rules for sequential ordering of cas-
the filter (i.e., the jthsection). In each of these terms, the cade-form sections.
transfer response ,DFj(w) [containing the factors ai(w)/
pi(.) for thej- 1 sections preceding that section] appears
as the argument of an L, norm, while the response , ~ G j ( o ) Fig. 8. Alternative description of general
[containing the factors a%{w)/pi(w) for the M - j sections rulesinFigs. 6 and 7 [use either (A) or
following that section] appears as the argument of an
Lr+,norm. This, of course, just reflects the fact that it is
the sections preceding agiven section which determinethe
overflow scaling for that section, while the succeeding sec-
tions filter the roundoff noise produced by that section.
Therefore, for p # r+ 1 the primary difference between
different sequential orderings of the M sections most likely
results from the different characteristics of the L, and Lz
norms, as discussed in the precedingsection. In particular,
since the L, norm is much moresensitive to themaxima of
its argument functionthan theL, norm, one wouldexpect
that the preferred sequential orderingshouldminimize
r =I
(in some sense) the peakednature of those functions which 1-00

appear as arguments of the L, norms of (54) and (55).

This would,in turn, indicate that theai(w)/pi(w) should be
ordered from “most peaked” to “least peaked”, or vice
versa, depending upon whether r = co or p = co , respec- K
\
tively. p=CO /
\
\
Since it is the LZand L, norms with which we are con-
cerned, a reasonable measure of the peakedness of the
cri(w)/pi(w) for our purposes is simply

= ,!l:! ;!I:/ tenuation in the neighborhood of the peaks of the re-

sponses d j ( w ) and z ~ G j ( w ) this
, attenuation should be
applied to themost peaked of these two sets of responses.
We would then expect fromthe above discussion that for This, in turn, implies (from the preceding discussion for
p # r = w , a good (but not necessarily optimum) ordering p#r+ 1) that for p =r+ I, a good ordering for the M
for the M sections of the cascade form is provided by sections of cascade form 1D should result from having P i
having 6idecrease (i.e., decreasingpeakedness)with increase with increasing i; while for cascade form 2 0 , Pi
increasing i; while for r # p = 00,6i should increase with should decrease withincreasing i. This and thepreceding
increasing i. Our experimentalresults support this general rules for sequential ordering are contained in the chart of
rule. Note that because of the numerator factors cri(w), Fig. 7.
Pi is not necessarily proportional to theQ of the section, It is instructive to note that something like the above
although thisis often the case. rule for p = r+ 1 was to be expected from theresults of the
For p=r+l, however, the factors ai(w)/pi(w) for all preceding section, where each form-1D configuration was
i # j appear in one L, norm or another, and hence the out- found to be comparable to the corresponding transpose
put roundoff noise in this case should be less sensitive to configuration ( 2 D J in this case. Note fromFigs. 6 and 7,
the ordering of the M sections of the cascade form than however, that this rule is opposite to that for p#r+ 1
for p#r+ 1. There will still, of course, be some effect due since, for p # r = co , form 1D with decreasing Pi is indi-
to thegrouping of the ai(w)/pi(u)within these norms; but cated ; while for r#p = 00,form 2 0 with increasing Piis
the primaryeffect isprobably due,as before, to the attenu- indicated. An alternate, and perhapsclearer, presentation
ation provided by the extra factors q ( w ) which occur in of these results is given inthe chartsof Fig. 8, which show
the L, norms for form I D and the Lr+lnorms for form the recommended combination(s) of form and ordering
2 0 . In particular, when the q ( w ) provide significant at- for each p , r pair.

JACKSON:ANALYSIS
ROUNDOFF-NOISE DIGITAL
FOR FIXED-POINT FILTERS 115
 Having considered the sequential orderingof the section t‘”‘”
responses ai(w)/pi(w), we now addressthe question of how
best to pair the numerator factors ai(w) with the denom-
inator factorsp(i.) to produce theseresponses. There are,
of course, many cases where this question does not arise
including, for example, the cases of Butterworthor
Chebyshev (type-1) low-pass or high-pass filters, where all
zeros occur at z= 1 or z= - 1 (assuming that the bilinear
z transform has been employed). In these cases the ~ ( w ) Fig. 9. Proper
pairing
of
poles
and
zeros
in cascade
form
for
sixth-order
are equal for all i (except for one first-order factor when
band-rejectionfilter.
N is odd), and there is no choice to be madein pairing the
a i ( w ) and pi(.) (assuming that the first-order factors, if
present, are paired together).
When the second-order factors a,(w) are notequal, how-
Fig. 10. Properpairing of polesandzerosincascadeformfor
ever, there is almostalwaysa significant effect on the sixth-order bandpars filter.
roundoff-noise output due to the pairing of these factors

t
with the pi(.). The reason i s apparent from (54) and ( 5 9 , Irn(z1

where the ai(w)occur onlyin ratio with the corresponding

pi(w), and the ratios ai(w)/pi(w) for all i appear in each
term of the summations. Since the normsof 2M- 1 differ-
ent combinationsof the a i ( w ) / p i ( w ) occur in each relation
for llNul’rin (54) and ( 5 9 , it is most reasonableto assume
that the minimum llNulI T will result when the individual
ai(w)/Pi(w) are minimized (in some sense). And since

for a l l p k 1, it is reasonable to minimize the L, norms of

the individual a,(w)/pi(o), insofar as this is possible.
Although the above discussion may seem inconclusive either cascade or parallel form, computes the scaling re-
as it stands, the proper pairing of the ai(w) and pi(w) is quired for that form to satisfy the overflow constraints in
really quite evident from such considerations in most cases (5) (for p = 2 or forp = cc ) and then predicts the resul.ting
of practical interest. Consider, for example, the z-plane roundoff-noise spectrum N,(w) from (1). The other is the
diagram of Fig. 9 for a sixth-order band-rejection filter. FilterSimulationProgram(FSP), whichsimulates the
The dottedlines indicate the best pairing of the zeros cor- (scaled) digital filter in the appropriate form using sinus-
responding to the ai(w) with the poles corresponding to oidal or white-noise input signals and estimates the power-
the pi(.) This pairing obviously minimizes the L, norms density spectrum Nu(@)of the output roundoff noise. In
of the ai(w)/pi(w)for all i and thus, mostlikely, the values other words, NAP implements the synthesis and analysis
Ip
of l:ai/pil for all i and p , as well. Almost as obvious are procedures developed in thispaper, and FSP tests the
the pairings of Fig. 10 for a sixth-order elliptic bandpass validity of these results for specific filters and inputsignals.
filter. The two “higher-Q” pole pairs are combined with In both NAP and FSP the output noise density N,(w)
the nearest zero pairs, leaving the real zeros to be com- is normalized with respect to N O =go2 (Le., the density or
bined with the “lower-Q” pole pair. This should minimize variance of the white noise from a single rounding opera-
the maximum Ilai/pilix over i. Experimental evidence of tion). That is, instead of N,(w) we actually compute
the effect of proper pole-zero combination is presented in
the next section. (57)

Hence N,(w) and norms thereof arealways given in deci-

Computer Implementation: An Example
bels relative to No. This normalization is very helpful in
Two digital computer programs have been written to relating the results of NAP and FSP, for eliminates
it any
implement and test the analytical resultsof the preceding (direct) reference to the inputsignal level. In [I] NAP and
sections. These programs are written in FORTRAN IV and FSP aredescribed in greater detail, and several representa-
have been run successfullyon theGE635 digital computer tive examples of their operation are given.
at Bell Telephone Laboratories, Inc., Murray Hill, N. J. One of these examples is thesixth-order Chebyshev
The first is the Noise Analysis Program (NAP), which, (type-2) band-rejection filter (BRF) described in Table I.
given the transferfunction H*(z) for a digital filter in The unscaled coefficientsfor both thecascade and parallel

116 IEEE TRANSACTIONS ON AUDIO AND ELECTROACOUSTICS JUNE 1970

 TABLE I
Sixth-Order Chebyshev-2 Band-RejectionFilter

ao=0.76091619;
yo=1.3142206

Section
1 1‘ 2’

0,75829007 P0.84506679
li 0.90352914
Pli .7636952 - 1
-1.5334490
-1.4427789 - 1.7636952
-1,4427789
01% 1 .o 1 .o 1 .o 1 .o 1 .o
aii -1.8118373 - 1.6545862 -1
.7442502 - 1.6545862 - 1.8118373
7 cci/Pi j m
1.65
Ilai/fiAi* 1.280
1 ,056 .602 1.076 1 1 ,066
7 Pi 1.56 2.11
Yli 0.09494903 0.17123073 - -
Yoi -0.28023324
-0.10898306 -0.16408810 - -
-.

Fig. 1 1 . Overall frequency

response of sixth-order Cheby- TABLE II
shev-2 BRF in parallel form.
Sixth-Order
Chebyshev-2 Band-Rejection
Filter
(A)
0.00 Variance ( r = 1) Peak Noise ( r = m )

-10.00
p= m p =2 p= m p =2

-3 -20.00
F1 F2
....................................
Parallel
F1 F2
F2

1
Ft F1 F2

2
[ -30.00
(P)Forms 24.1*
25.2
13.8**
14.3 35.6*
37.0
23.0%
24.4

-400
.0 ( D ) Forms
(123)
(321)
34.2
23.5** 22.1
21.0
2320413..42.86*
22.7**
34.8
34.8 31.2
32533.367.50. 0. 0
(213) 33.6
23.5”
21.3
22.4 32333.357.63. 0. 0
-50.00
0. looo. 2000. 3000. 4000. 5000. (312) 21.7
(231)
22.0
23.8
34.0
20.3*
24.5
23.8
31.6 36.0
23.3
3 5 . 0 32.5
37.0
35.0
23.1
35.5

FREQUENCY ( H Z ) (132) 20.8

22.3
23.7
34.4 3 5 . 0 34.5** 30.7
23.1

( B)
2 w . 0 , , , 1 , 1 1 1 1 , ,

designed from these specifications using a FORTRAN-IV

computerprogramdevelopedand written by R. M.
Golden,’ J. F. Kaiser, and E. J. Sitar of Bell Telephone
Laboratories, Inc. The plots were generated on a Strom-
berg-Carlsen 4060 microfilm plotter using another sub-
FREQUENCY (Hz)
routine due to Kaiser and Sitar.
Predictions of the output roundoff-noise spectra N,(w)
were made by NAP for parallel forms 1P and 2P and all
orderings of cascade forms 1D and 2 0 with p = 2 and
p = tc, . The results are summarized in Table 11, where the
forms of this filter are given in Table I, as well as the Lz
predicted \lNz/\lT for r = 1, m are given in decibels relative
and L, norms of the cascade-form responses ai(o)/pi(w)
to No for all cases. The form-1 results are indicated by
and thecorresponding ratiosPi. The effect of proper pole-
“Fl” column headings, and form-2 by “F2.”
zero pairing is also illustrated by reversing the pairings of
The ordering (123) corresponds to decreasing Pi, Le.,
Sections 1 and 2 toyield Sections 1’ and 2’, as described
decreasing peakedness. Note from Table I1 that there is
in Table I. The pairings in Sections 1, 2, and 3 are in ac-
not much difference between the different cascade-form
cordance with the discussion contained in the’preceding
orderings, although for p = r + 1 form 2 0 is somewhat
section, and illustrated in Fig. 9.
better when Section 1 is first and form 1D is better when
The frequency response (both magnitude and phase)
of the BRFis shown in Fig. 11. The specifications for the
filter are 2.26-dBpassband ripple, 25-dB stopbandat- 1 Now withTechnology Service Corporation,SantaMonica,
tenuation, and a transition ratio of 0.53. The filter was Calif.

JACKSON: ROUNDOFF-NOISE ANALYSIS FOR FIXED-POINT DIGITAL FILTERS 117

 24.00

v
5 22.00
20.00
g 18.03

2 166.00
B 14.03
I *-Oo
2 10.00
8.00
0. 1ooO. 2ooO. 3003. 4ooo. 5000.
FREQUENCY ( H Z )
( 4
28.03

2 24.03

1i 20*00

@ 120.0
16*00

3 8.00
E l I / / ' I I I

4 . 0 3
c3003.
. 2000.1003.
t ~ ~ ~ 4ooo.
1 ~ 5000.I I
FREQUENCY (Hz) FREQURiCX ( H z )
(W (B)
Fig. 12. Predicted
roundoff-noise densities for sixth-order Fig. 13. Predicted
roundoff-noise densities for sixth-
BRF in parallel forms 1P (A) and 2P (B) with p = m . order BRF in parallel forms 1 P (A) and 2P (B) with p = 2.

Fig, 14. Predicted

roundoff-noise densities for
sixth-order Fig. 15. Predicted
roundoff-noise densities for
sixth-order
BRF incascadeforms 1D (A) and 2D (B) with p = m ( 1 2 3 BRF incascade forms 1D (A) and 2 0 (B) with p = 2 ( 1 2 3
ordering). ordering).

[A)
400.0 ( 4

5
v

E 30.00

z 20.00

B
k
, I
I 1 1 I ,
0.00
0.
/ 1o/oO. / ! / 3000./ I 4000.I /
2000. 5000.
mqmm ( H z )
(B)

118 IEEE TRANSACTIONS ON AUDIO A N D ELECTROACOUSTICS JUNE 1970

 TABLE 111
Sixth-Order
Chebyrhev-2 Band-Rejection
Filter (Pairing Reversed
in
Sections 1 and 2)

Variance (Y = 1) Peak Noise ( r = m )

Cascade
Forms
(0) . p Le p =2 p= m p =2
F1 F2 1 F1 F2 F1 F2 F1 F2

(1’2’3)
(32’1’)
3226283.0.5.3.96
27.7*
29.619.9*
37.7*
24.9
28.1
40.4
1 39.4
37.7*
26.7*
29.0
37.6
(2’1’3) 27.6*
23.8
30.7
20.4 35.8
28.0
39.8
38.2
(31’2’) 20.9
23.5
28.235.8 30.6
27.1
38.4
39.5
(2’31’) 28.4
21.6
32.932.5 4462. 0. 1 31.2 41.6
(1’32’) 2237. 401. 91. 3. 1 32.9
32.6
42.1
45.8

Note: I(N,II,/No in decibels.

Section 1 is last, in agreement with the general rules in 1 D,however, the Fi(w)are not wide-band, and the nois-,
Fig. 7. This relative insensitivity to ordering is explained densities are, therefore, muchless for p = 2 than for p = co
by the relatively small variationin Pi for Sections1,2, and due to scaling.
3, as given in TableI. As an example ofthe effect ofproper pole-zero pairing,
The largest differences in the predicted N,llT for the 11 we now consider the corresponding results for the BRF
BRF occur between forms 1D and 2D when p # r+ 1 and when the pairings of the ai(@) and pi(w) for Sections 1 and
between the P and D forms when p = r+ 1 = 2. Note from 2 are reversed to form Sections 1’ and 2’. The resulting
Table I1 that these differences are of the order of 7 to 12 11 1
Nu/ are given in Table I11 for all orderingsof the cascade
dB. In agreement with Fig. 6, form 2 0 is superior for forms. As before, the ordering (1’2’3) corresponds to
r # p = 00,while form 1D is best when p # r = co . Parallel decreasing Pi,but now the variationin Piis much greater.
form 1P is comparableto the best cascade form for each Note that the results in Table I11 are in complete agree-
p , r pair except p = r+ 1=2, where it (1P) is definitely ment with the general rules in Figs. 6 and 7 concerning
superior. form and ordering.
The above differences in [lNul],are readily apparent Our main point is, however, that the lowest llN,I[,. for
from the predicted densities in Figs. 12 through 15. Figs. each p , r pair in Table 111 is significantly greater (3 to 4
12 and 13 showthe predicted N,(w) for the parallel forms, dB) than that for the cascade forms in Table I1 in every
and Figs. 14 and 15 givethe predicted N,(w) for the (123)- case except for p = r+ 1=2. And in the latter case, the
ordered cascade forms.Note that theshapes of the form- results are comparable. Hence, proper pole-zero pairing
2 0 densities are very similar to those for the P forms. In is indeed important in the synthesis of this digital filter,
particular, the densities for form I D resemble the wide- as expected.
band response of the BRF itself, whereas those for forms The results of four representative FSP simulations of
20, lP,and 2P are narrow-bandin character and achieve the sixth-order BRF are presented in Table IV and Figs.
their peak values in the BRF stopband. 16 through 19. The cascade-form results correspond to
WithLI scaling ( p = co) the peak values llN,/lm are the (123) ordering (with proper pole-zero pairing). The
comparable in every case, but because of the wide-band scaling appropriate to the input signal is employed (i.e.,
character of I~Ny(w), the total power I] Null for form 1 D p = 00 for sinusoidal inputs and p = 2 for white noise).
is much greater than for the other forms. For L2 scaling The inputsignal amplitudes, peakoutput roundoff errors,
( p = 2), however, the peak density for form D1 is much less estimated error means, and peak signal levels at overflow-
than for form 2 0 although their total noise powers are constrained branch nodes arelisted in Table IV in quanti-
very comparable. The peak density for parallel form 1P zation-step units (assuming rounding to the nearest inte-
is almost the sameas for form 1D when p = 2, but because ger). The estimated and predictederror variances are
of its narrow-band character the total noise power for the given in decibels relative to No. Parzen lag windows have
parallel form ismuch less in this case. been employed in each case.
Another implication of the wide-band response of the Note first of all the excellent agreement between the
BRF is that theform-2D noise densities are notmuch less estimated noise densities in Figs. 16 through 19 and the
for p = 2 than forp = cc , in contrast with otherexamples correspondingpredictions in Figs. 12(A), 13(A), 14(B),
presented in [l]. The reason for this is that the transfer and 15(A), respectively. For example, note that the peak
responses F2(w)are all wide-band in this case, and hence densities and those at dc and 5000 Hz agree quite closely
the scaling for wide-band inputs( p = 2) is not muchdiffer- with their predictedvalues, The estimated error variances,
ent from that for narrow-band inputs ( p = a). For form or average noise powers, in these cases are all within 0.9

JACKSON: ANALYSIS
ROUNDOFF-NOISE DIGITAL
FOR FIXED-POINT FILTERS 119
 (4

0
0.
.
1000.
0
2000.
0
3000. 4ooo.
1 ~
5000.
~ ~ ~ ~ ~ ~ ~ ~

FREQUENCY (Hz)
(B) (B)

Fig. 16. Estimated densities of filter

output and roundoff Fig. 17. Estimated densities of filter
output and roundoff
noise from simulation of sixth-order BRF in parallelfrom 1P noise from simulation of sixth-order BRFin parallelform1P
with 1 3 6 9 - H z input, p = g. (A) Actual filter outputincluding with white-noise input, p =2. (AI Actual filter output including
roundofferror. (B) Roundoff error in actual filter output, roundofferror.(B) Roundoff error in actualfilter output.

TABLE 1V

Simulation Sixth-Order
of Chebyshev-2 Band-Rejection
Filter (2048
Samples)
___ _____
Pre-
Syl- Form Input
Ing
Am- Fre-
pli- quency
tude or PF
Peak Error
Error ance
Mean
F:ri $$!
Peak
ante Data
(PI (**) (dB)

LI 1P 1024
sine 1369Hz
15.54
-0,0253
24.16
24.1 372
L2 1P 1024
noise 1076
0.0218
4.545
13.8
13.45
4.0
LI 20 1024
sine 1369Hz
17.99
-0.284
24.42
23.5 1025
L2 1D noise 1024 11.88
0.0707
22.05
22.1
4.0 909
..___

dB of their predicted values. This agreement is perhaps at overflow-constrained branchnodesare close to the
surprising in the first and third cases because the input input signal amplitude (1024). The one case (form 2 0 ,
signal is a single sinusoid (at 1369 Hz) and significant LI) where the peak data exceeded1024 (plus the peak
correlation might be expected in theoutput roundoff error) resulted from the residual transient remaining in the
error as a result. However, the correlation in the indi- data after the 200-sample initialization period.Otherwise,
vidual roundoff-error inputs seems to average out in the the overflow constraints appear to have been satisfied i,n
total output from thefilter. Cases can, of course, be gen- all cases.
erated with significant correlation being apparent in the
form of harmonics of the sinusoid [I], but these cases are
Summary
the exception, not the rule.
There is also good agreementbetween our analytical The roundoff-noise outputs from two transpose con-
results andother measured data.The estimated error figurations each for the cascade and parallel forms of a
means are very close to zero in all cases, and the peak data digital filter have been analyzed and compared using the

120 ELECTROACOUSTICS
AND AUDIO IEEEON
TRANSACTIONS JUNE 1970
 (6) (6)
40.00 28.00

-8
2 30.00 h

8 24.00
v

8 20.00
8
20.00

16.00

i
s8
E
10.00

O*O0 1 1 1
I
BE
l2.00
8.00

4.00
-10.00 o. 1000. 2000. 3000. 4000. 5000. 0.00
0. 1000. 2000. 3000. 4000. 5000.
FREQUENCY (Hz) FREQuEir’CY (Hz)

Fig. 18. Estimateddensities offilter output and roundoff Fig. 19. Estimateddensities of filter output androundoff
noise from simulation of sixth-order BRF in cascade form 2 0 noise from simulation of sixth-order BRF in cascade form 10
with 1 3 6 9 - H z input, p = m. (A) Actuol filter output including with white-noise input, p = 2. (A) Actuol niter output including
roundoff error. (6) Roundoff errror in actual fllter output. roundoff error. (6) Roundoff error in octuol niter output.

techniques developed in [2]. The spectrum of the output sons based on the bounds in (60), rather than on lINylIw
roundoff noise from these four configurations has been itself, have been justified by the excellent experimental
shown to be of the form agreement reported here and in [ 11. A heuristic discus-
sion of the sequential ordering and pole-zero pairing of
the cascade-form sections has also been presented on the
basis of (59) and (60). From these results, the following
general “rules of thumb” may be stated.
1) Parallel form 1P is generally preferable to form 2P,
where Fi(w) is the (unscaled) transfer response from the
although the maximum difference in l/NullTfor these two
input to the ith branch node, G j ( w ) is the (unscaled) trans-
forms (for a given p ) is only of the order of 4/3, or 1.25
fer response from thejth summation node to the output,
dB.
kj’ is the number of noise sources inputting (directly) to 2) For p = r+ 1 there is little difference in IINy/i, be-
thejth summation node,M is the numberof second-order
tween any given form-1Dconfiguration andthecor-
sections comprising the cascade or parallel form, and go2
responding transpose configuration (form 2Dt).
is the varianceof the noise from each rounding operation.
3) For p # r + l , however, cascade form2 0 is generally
Hence, the variance, or total average power,of the output
superior to form 1D for p = M , and vice versa for r = c4 ;
roundoff noise is simply
and the difference can be quitesignificant.
4) Good(butnot necessarily optimum) sequential
orderings forthe M sections of the cascade form areindi-
cated in the chartsof Figs. 7 and 8, where they are stated
in terms of the variation in
11 1
while the peakspectral density Nu]oo is bounded by

= l;w/ l;
with increasing i.
Correspondingtransposeconfigurationshave been 5) Thenumeratorfactors ai(w) and
denominator
compared on the basis of (59) and (60), and the compari- factors pi(.) of the cascade D forms should be paired so

ANALYSIS
ROUNDOFF-NOISE
JACKSON: FILTERS DIGITAL
FOR FIXED-POINT 121
as to minimize the individual liai/pi!i a, insofar as this is period T and/or is of insufficientmagnitude 111. The gen-
possible. eral rules concerning circuit configuration have also been
The analytical results obtained do not provide ageneral found to be very helpful and effective.
rule relating to the choice between parallel form 1P and
one of the cascade D forms. The experimental results do
References
indicate, however, that form 1P is usually comparable to
or somewhat better than the best D form(s). Therefore, 111 L. B. Jackson, “An analysis of roundoff noise in digital filters,”
Sc.D. dissertation, Dept. of Elec. Engrg., Stevens Institute of
from the viewpoint of roundoff noise alone, the simplest Technology, Hoboken, N. J., 1969.
and most reliable choice of a digital-filter configuration [2] ---, “On the interaction of roundoff noise and dynamic range
would seem to be parallel form 1P. However, the cascade in digital filters,” Bell Sys. Tech. J., vol. 49, no. 2, February
1970.
D forms often have other advantages which make them [3] R. Edwards, J. Bradley, and J. B. Knowles, “Comparison of
most desirable [SI, especially when the zeros of H*(z) lie noise performance of programming methods in the realization
on the unit circle (1 zI = 1). In these cases, the D-form of digital filters,” Proc. Symp. on Conzputer Processing in Com-
nzunicatiorzs, PIB-MRI SymposiaSer.,vol. 19, 1969.
options of form, sequential ordering, and pole-zero pair- [4]C. Weinstein and A. V. Oppenheim, “Acomparisonof roundoff
ing contained in (2) through (5) must be considered. noise in floating point and fixed point digital filter realizations,”
Finally, two computer programs written to implement Proc. E E E (Letters), vol. 57, pp. 1181-1 183, June 1969.
[5] J. R. Rice, TheApproximation of Functions. Reading, Mass.:
and test the analytical results have been briefly described. Addison-Wesley, 1964, pp. 4-10.
A representative example has been included to illustrate [6] C. M.Rader and B. Gold, Digital Processtng of Signals. New
the operation of these programs and, more importantly, York: McGraw-Hill, 1969.
[ 7 ] J. F. Kaiser,“Digital filters,” in SystemAnalysis by Digital
to demonstrate the effectiveness of the analysis and syn- Computer, F. F. Kuoand J. F. Kaiser, Eds. New York:
thesis procedures upon which they are based. These ex- Wiley, 1966, ch. 7 , pp, 218-285.
perimental results indicate thatthe uncorrelated-error [8] L. B. Jackson, J. F. Kaiser,and H. S. McDonald,“Anap-
proach to the implementation of digital filters,” ZEEE T r a m
analysis is quiteaccurate, even for sinusoidal inputs, Audio and Electroacoustics, vol. AU-16: pp. 413-421, September
unless the input is periodic in a multiple of the sampling 1968.

€3.
Leland Jackson (S162-M’65) wasborn in Atlanta, Ga.,on July 23, 1940. He received
the S.B. and S.M. degrees in 1963 from the Massachusetts Institute of Technology,
Cambridge, and the Sc.D.degree in 1970 from Stevens Institute of Technology, Hobo-
ken, N. J., all in electrical engineering.
In 1961 and 1962 he was associated with BellTelephone Laboratories, Inc., under the
M.I.T. cooperative program in electrical engineering. From. 1964 to 1966 he was em-
ployed by Sylvania Electronic Systems, Inc., Mountain View, Calif., where he studied
and developed signal processing techniques for ionospheric research. From 1966 to
1970 he was a member of the technical staffofBell Telephone Laboratories, Inc.,
Murray Hill, N. S.,where he was primarily concerned with the analysis and synthesis
of digital filters. In 1970 hejoinedRockland Systems Corp., Blauvelt, N. Y.,as
Vice President for Engineering.
Dr. Jackson is a member of Tau Beta Pi, Eta Kappa Nu, andSigma Xi.

122 IEEE TRAKSACTIOKS ox AUDIO ANDELECTROACOUSTICS JUNE 1970