IEEE TRANSACTIONS ON AUDIO AND ELECTROACOUSTICS, VOL. AU-2 1 , NO.

6 , DECEMBER 1973

49 1

would require less hardware than the other possibilities. Another means of evaluating these different realizations is to ex&ine the effect of coefficient roundoff error and multiplier truncation noise. These have not yet been considered. The techniques discussed in this paper can be easily adapted for use in realizing transfer functions in the s-domain. All the results discussed are equally valid exif we replace z- by s-l in the mathematical pressions, and replace delays by integrators,adders by summing amplifiers, and multipliers by amplifiers in the diagrams.
Conclusion

References

A technique similar to that of [6] for realizing a digital filter in a ladder like structure is proposed. It is especially suitable for realizing a transfer function having all its zeros at the origin although it can easily be used t o realize any general transfer function. With a slight modification this technique becomes suitable for realizing transferfunctionsobtained via the bilinear transform from all pole s-domain transfer functions. Two other forms are obtained, one of which is . topologically similar to the coupled form [ll]

F. F. Kuo and J. F. Kaiser, Eds., System Analysisby Digital Computer. New York: Wiley, 1967, ch. 7. C. Rader and B. Gold, Digital filter design techniques in the frequency domain, Proc. IEEE, vol. 55, pp. 149171, Feb. 1967. R. M. Golden, Di ita1 filter synthesis by sampled data transformation, I8EE Trans. Audio Electroacoust., vol. AU-16, pp. 321-329, Sept. 1968. A. Fettweis, Some principles of desi ning digital filters imitating classical filter structurei, I ~ E E Trans. Circuit . 314-316 Mar. 1971, Theory vol. CT-18 S. K. Ihitra and R! !&erwood, Canonic realizations of digital filters using thecontinued fraction expansion, IEEE Trans. Audio Electroacoust., vol. AU-20,. pp. 185__ 194, Aug. 1972. -, Digital ladder networks, IEEE Trans. Audio Electroacoust.. vol. AU-21. DD. 30-36. Feb. 1973. R. Edwards, J.Bradley, a n x j . B. Knowles, Comparison of noise performances of programming methods in the realization of digital filters, in Proc. Symp. Comput. Processing Commun., PIB-MRI Symp. Ser., vol. 19. Brooklyn, N.Y.: Polytechnic Press, 1969, pp. 295-311. 0. Hermann and W. Schussler, On the accuracy problem in., the design of nonrecursive digital filters, Arch. Elek. Ubertragung, vol. 24, pp. 525-526, 1970. R. E. Crochiere,Digitalladder structuresand coefficient sensitivitv. IEEE Trans.Audio Electroacoust.. vol. AU-20, pp. 248-246, Oct.~1972. [ 10 J J. Tow and Y. L. Kuo, Coupled biquad active filters, in 1972 Proc. IEEE Symp. - - Theory. Int. Circuit pp. - . __ 164-167. [ 1 3 B. Gold and C. Rader, DigitalProcessingof Signals. New 1 York: McGraw-Hill, 1969, pp. 45-46, 101-103. [12 J M.E. Van Valkenburg, Introduction to Modern Network Synthesis. New York: Wiley, 1960, pp. 86-93.

Digital Lattice and Ladder Filter Synthesis

multiplies and delaysis obtained. An internal scaling procedure is also introduced that will be of importance for optimizing one of the lattice forms for finite word length implementation.

Introduction

Techniques for synthesizing digital ladder structures [ 11, [ 21 have recently been studied because of their apparent frequency response insensitivity with respect to themore standard direct, parallel, and cascade studies by Fettweis 141, [SI forms [ 3 ] . Recent Abstract-There is evidence that in addition to standard digital show that digital ladder filters can be patterned after filterformssuch as the direct, parallel, and cascade forms, Crochiere has demondigital lattice and ladder filters may play an important role in classical analog structures. stratedthat,at least foronefilter,the coefficient finite word length implementation problems. In this paper, techniques developed in detail efficiently sensitivity of the digital ladder filter is at least several are for synthesizing digital lattice and ladder filters from any stable bits less than a cascade implementation of the same direct form. In one form, a lattice filter canonic in terms of filter [6] , It is certainly desirable to know how to directly transform a direct form into a corresponding ladderorlatticestructure.Notablesteps in this diManuscript received May 15, 1973; revised August 9, 1973. This work was supported under ONR Contract N00014-674- , rection have been made by Mitra and Sherwood [ 1 1 , 0118. Unfortunately,theirresultsare restrictive in A. H. Gray, Jr. is with the Department of Electrical Engineer- [2] ing and Computer Science, University of California, Santa that only certain classes of stable filters can be Barbara, Calif. 93106. synthesized.addition, procedure In their requires J. D. Markel is with Speech Communications Research Laboratory, Santa Barbara, Calif. 93109. the solving of a set of linear simultaneous equations

492

AND AUDIO ON IEEE TRANSACTIONS

ELECTROACOUSTICS, DECEMBER 1973

in order to obtain tap the gains (gain terms for implementing the numerator without additional delays). In this paper, a general solution to the synthesis of digital lattice and ladder structures from direct forms is presented. In form, resultant one the filter is canonical in thenumber of multipliersand delays. The procedure follows from an application of an inner product formulation first developed by Itakura by Markel andSaito [7] andlaterexpandedupon and Gray [ 8 ] . The solution is easily developed and programmed as a recursive procedure for the filterdesign. Furthermore, the procedure is extremely efficient. No simultaneous equation solution for tap gains is necessary. What falls out of the procedure is twoa multiplier lattice form. A trivial transformation results in a three-multiplier ladder form. From the two-multiplier form new a generalized onemultiplier form, canonic in both multipliers and delays, is developed. With the generalized one-multiplier model,a new technique is presented for optimally scaling the filter structure with respect t o finite word length implementation.
Lattice Filter Formulation

Vm

-Pm,m
2Bm (zlvrn

(7)
(8)

Pm -1

( 2 ) = Pm ( 2 ) -

for rn = M , M - 1,. . . , 1with v 0 = P ~ , ~ . The only division that arises is in (6). This will never be a division by zero, for it is known that the h-parameters will always have a magnitude less than one for a stable filter [ 7 ] , [SI. Should any k , equal plus or minus one, or have a magnitude larger than one, then AM((z) will not have all its roots within the unit circle and the filter will be unstable. In terms of the new variables, PM( 2 ) is equivalent to

Therefore, equivalent an representation given as

of G ( z ) is

It remains to be shown that this formulation results in a digital lattice structure. First, the general proof of these equations will be presented.
Proof of Formulation

The results stated here are essentially applications of the inner product formulation published elseG = G ( z )= PM ( z ) / A M ( z ) (1) where [ 71 , [ 81. where & ( 2 ) and A ( z ) are Mth order polynomials in M First, we define 2 - , of the form R ( z ) =[ A M ( Z ) A M ( ~ - ) I - , (11) m Pm = Pm (2) = Pm,nZ-n (2) which will be positive on the unitcircle. An inner, or
n=O

We wish to design a digital filter that will implement the direct formtransfer function

and A, =A,
m

scalar, product of functions of z can be then defined by

am,n.z-n

(3)

n=o

< F ( z ) G ( z ) )= ,

de R(e)F*(ej@)G(e)2n

(12)

where rn = M . Without loss of generality, it is assumed that the leading coefficient of A M ( z ) is one, a M , 0 = 1. If the filter is implemented in the direct form, then the parameters representing the filter are 2M in general, P M , O 7 P M , l 7 * . 7 P M , M 7 aM,l 7 a M . 2 7

...

~ M , M

The 2M + 1 parameters of the lattice filter thatwill be designed are the M k-parameters h , , k , . . . , k w - l and the M + 1 tap parameters v o , v l , . . . , v M . These parameters are recursively obtained by starting from A M ( 2 ) and P M ( 2 ) as

where the asterisk denotes a complex conjugate. Using this notation, it can be shown [SI that by defining zB,(z) = A , ( l / z ) z - , both the polynomials A, ( 2 ) and B , ( z ) are orthogonal to the powers of z- , from the 1st through the mth; that the polynomials zB, ( z ) form an orthogonal set; and that the polynomials are related bythe recursion matrix

zB, ( z ) = A , ( l / ~ ) z - ~
hm - 1
=

(4) for rn
(5)
-1)

0 , 1 , . . . , M - 1 with A o ( z )

1 and B o ( z )=

am,,

2-1.

A m - 1 ( z ) = [ A m ( 2 ) - hrn -1zBm (z)I /(I -

In addition, the coefficients a ,

can be found from

(6)

= (1, , ( z ) ) , A paper, for notational simplicity, capital Throughout the letters will be used to refer to the corresponding z transforms.

(14)

or through a recursion relation

GRAY

MARJSEL: DIGITAL FILTER SYNTHESIS

493

a m + l= a , ( l - k& )

for rn

0,1,

. ,M

1 (15) delay only. By further manipulation it is possible to

with aM = 1. The k-parameters can also be expressed in terms of scalar products as k, = -/3,/am where

obtain a form that is also canonical with respect to first by multiplications. It was evidently shown Itakura and Saito that (18) and (19) can be transp, = ( A , (z),B, (2))= (1,B , (2)). (16) formed into a one-multiplier model. In the next section the one-multipliermodel is derivedwitha Equation (6) is obtained by simply computing the generalization that includes a sign parameter definimatrix inverse from (13) and replacing rn by rn - 1. tion. The potential value of thisgeneralization will Equations (7) and (8) are obtained by noting that if be discussed in a later section. v, is defined as P,,, , then P, - zB,v, results in a 1 since zB, has a polynomialreducedinorderby OneMultiplier Implementation final coefficient of unity for all rn (from (4) and the fact that = 1). Here we define a new set of sign parameters, e o , From the recurrence relations, it can be seen that E 1 , . . , eM- 1 . Each of these will be equal t o plus one each P,(z) will be of the form of (2) and that each or minus one. How the sign parametersarechosen then A , @ ) will be of the form of (3) with am,o= 1. Equa- willbe the topic of a later section. These are tion (9) is proven from (8) by simply summing used to obtain a set of modified tap parameters, Go, P - P, , from rn = 0 to M with P-, = 0 . v l , . . . , v M , and a set of modified polynomials by the relations

-,

A Two-Multiplier Lattice Filter

In this section, the general form of the filter as a block diagram is first developed. If X = X ( z ) is the inputto G = G ( z ) , and Y = Y ( z ) is theoutput, from (lo),

Y=

v,*
m
=O

zB, X
A M
The recursion relations of (13)then become

The recursions specified by (13) and (14)can be rewritten in the form

z(1 + E,k,)B,+l(z)

= km8,(z)

+E,(z).

(24)

From (23),8, ( z ) is obtained by inspection as

Noting that AMX/AM = X and AoX/AM = zBOX/AM, a block diagram for the filter is shown in Fig. 1. The details of a prototype section G, are obtained from (18) and (19) as shown in Fig. 2. The overall implementation of G(z) with these particular sections is referred t o as a two-multiplier model since each sectioncontainstwo multiplies. By makingproper analogies to classical network theory [9], this filter of filter. It is would be defined as alatticetype trivial, however, to obtainaladderrepresentation from the lattice. Substituting (18) into (19) results . in a three-multiplier form as shown in Fig. 3 By inspection of the figure it is possible to easily reduce it t o a two-multiplier ladder form. As canbe noted from Fig. 1, each of the terms multiplying the tap parameters in the summation of (10) appear explicitly; direct multiplication by these tap parameters and a single summation will fully implement the filter in the form (10). of The forms presented are canonical with respect to

Substituting (25) into (24)and noting that (1- k$ ) = ( 1 + e,k, )(1-e,km) and E & = 1results in
h

z B m + l ( z )=Brn(z)- cmkm [Bm(z)-

~rnArn+l(~)I (26)

Equations (25) and (26) can be implemented as indicated in Fig. 4, where Fig. 4(a) shows the blocks in detail for E , = +1 and Fig. 4(b) shows the blocks in detail for e, = - 1. Note that the general form of , Fig. 1 is still valid with the A , , B,, and v terms hatted since by substituting (20) through (22) into (lo), G(z)is equivalent to

From Fig. 1(with the replacement of hatted terms), it is seenthat each term needed to multiply the modified tap parameters, as indicated in (27), appears explicitly in the model; all that needs to be idded are the multiplicationsby thetap parameters andthe summation indicated.

494

IEEE TRANSACTIONS ON AUDIO ANDELECTROACOUSTICS, DECEMBER 1973

G( 2 )

9
d
Y(3

Fig. 1. Ladder and lattice form implementation.

Fig. 2. Prototype filter for two-multiplier lattice model.

(b)
I

Fig. 3. Prototype filter for three multiplier ladder model.

Fig. 4. Prototype filters for generalized one-multiplier lattice model. (a) E , = +l. (b) E , = -1.

sensitive in changing the gross behavior of the filter and, then, show how to choose the sign parameters I t has been noted that when poles of a filter lie at so as to makethosethecalculationsthat are perlowfrequenciesandarenear the unit circle, direct formedthe with numbers having the largest implementation of a digital filter by using a recursive magnitudes. approach is much worse than factoring the filter into In linear prediction synthesis of speech, it has been a complex pole pair and using a cascade approach [ 31 . observed thatit is those h-parameters having the The meaning of the term much worse is related to largest magnitude that are the mostimportant.In how sensitive the angular location of the complex dealing with voiced speech signals it usually is true that k o and k , have the largest effectontheend pole is to small percentage changes in the filter results. parameters. Filter sensitivity to h-parameters near unit magniDue to the complexity of the problem, no attempt is made here t o derive relationships for sensitivity of tude can be notedby considering the case of two the pole or zero locations with respect to small per- separate filters being implemented, whose only difference lies in the Zth h-parameter.One of the filters centagevariationsin thefilter parameters.Rather, we shall determine which calculations are the most uses h, and the other filter uses k ; = k , +. Ah,. The
Sensitivity

GRAY AND MARKEL: DIGITAL FILTER SYNTHESIS

495

polynomials for the filter might be denoted by A n @ ) and A k ( z ) so that the polynomials are equal for n = 0, 1, . . . , 1. For n larger than 1 they will differ, and the effect of this difference will first show up in the n = 1 + 1 polynomial. I t is shown in Appendix I that the difference of the log magnitudes of Ai+l( e j e )and ( e j e ) will oscillate between the values log [ l + A k , / ( l + k,)] and log [ l - A k , / ( l - k , ) ] . Thus, the closer that the magnitude of k, is t o one, the more significant the effectsof ,nonzero Ah,. As aresult of this, it can be notedthatthe Izparameters are inherently scaled in a desirable fashion, that is, the ones that affect the filter most are the ones with the largest magnitudes. Inthe following sections the problem of scaling is studiedinmore detail.
Relative Sizes of Numerical Values

and

For purposes of normalization we shall define a set the of z transforms U m ( z ) , Vm(z), and W m ( z ) by relations
( z ) = Am ( z ) x ( z ) / [ A M (z)&
vm

These also represent the inputs to the multipliers as seen from Fig. 2. From (15), one should note that the Q, form a sequence that decreases with increasing m. Thus the largest multiplier inputs will be in the blocks with the smallest m, those furthest from theinput. This is useful in speech synthesiswhere k o and k l are the k-parameters closest to unit magnitude. However if the am cover a wide range [ a o >> aM = 11, then the multiplier inputs in the blocks with large m will be a great deal smaller in size than those with small m. In the one-multipliermodels ofFig. 4, thenode values are specified by
i m (2)

X ( z ) = nm&m

un ( 2 ) r

(33)

1,

(28) (29) and

A M (2)

( z ) = Bm

(z)X(Z)/[AM

< z > c m

and

B m (2)
A m (2)

X ( z )= n

m 6 m V m (2).

(34)

and X ( z ) represents the z transform of the input to the filter.Theparticulardefinitionscorrespond to normalized node values andmultiplier input values for the various filter forms. The motivation for these particular normalization constant choices will be presented shortly. While it is impossible to judge the sizesof the time sequences associated with Urn z ) , ( Vm( z ) , and Wm ( z ) for all possible inputs, some statements can be made about their sizes based upon their spectra, their log spectra, and their rms value when the inputis zero mean uncorrelated noise. First, terms in of their spectra, it is shown in Appendix I1 that each of the three the has unit energy or total squared integral when X ( z ) = 1. Second, when the input time sequence is a random, stationary, uncorrelated, and zero mean process having a variance or mean squareequal to o:, thenit is shown in Append,@ I1 that the time sequences associated with Urn( x ) , V, ( x ) , and W , ( z ) will all be random,stationary,zero mean process each having a variance or mean square equal to 02. Thus, in a qualitative sense, one can say that these time sequences, on the average, will have identical rms values. Third, the rather surprising property also shown in Ap( pendix I11 is that the Urn z ) and V m ( z )formtwo separate sequences of mini-max approximationsin terms of their log spectra. Inthe two-multipliermodel of Fig. 2, the node values used in the mth block are

From (26) and Fig. 4,the multiplier input in the mth block is easily seen to be

The product n m G m appears in each of these expressions and is obtained by recursion from (15) and (22) as
M

nm&m

n=m

-1

d(1+n k n ) / ( 1 e

Enhn).

(36)

As will be seen in the next section, appropriate choices of the sign parameters can be utilizedto make this product largest when k-parameters are near one in magnitude, without requiring a monotonically decreasing sequence with increasing m, as was the case with the f i m terms in the two-multiplier model. and As in the case of showing that Um(z),V m ( z ) , W m ( z )represent z transforms of unit energy signals sequences will (when X ( z ) = 1) and that their time have variances equal to u: for random inputs with variance equal to u:, it is also shown in Appendix I1 that the impulse response of the filter contains total a amount of energy given by

496

IEEE TRANSACTIONS ON AUDIO AND ELECTROACOUSTICS, DECEMBER 1973

and that when the input is the random, zero mean, uncorrelatedstationarysequencewith variance 02, the filter output yn will be random and zero mean, with a variance given by
M

signal's magnitude exceeding unity will be givenby 0.0026. To implement this approach, following algorithm the is applied. Assume that k l has the largest magnitude among the k,, m = 0, 1,.. . , M - 1. Define the quantities

u; = u:
n=O

v:(Yn.

(38)
and

Sign Parameter Choices

Numerous criteria could be utilized to pick optimal values of the sign parameters, based upon the user's definition of optimal. In order to obtain an effective filter over differing types of input, we shall choose our criteria in terms of rms values for the case where the input signal is random, uncorrelated, stationary, and zero mean. If the variance of the input signal is , : so that its rms value is u, ; then the results of u Appendix I1 show that the r p s values t f the time ( %equences associated with A , (z)X(z)/AMz ) and B, ( z ) X ( z ) / A M ( z ) both be given by T,&, will the rrns value of the multiplier input in the mth section will be given by rm&irn u,d2/(1 + e,h, ), and that the rms value of the output of the overall filter from (38)is (39)
n=o

I h I). m (42) Then Ql = 1. Each Qm should be as large as possible without exceeding Q1. By combining (39) through (42),

q = d ( 1+ I km I ) / ( 1 m

For m = I - 1, I - 2 , . . . , 0, choose em = sgn (h,) if Qm+l < l/q,. Otherwise choose e = - sgn (h,). , For m = I , I + 1 , . . . ,M - 1,choose e, = - s g n ( h , ) if Q, < l / q , , otherwisechoose e, = sgn (h,).The effect of thesechoices is to maximize Q, at each node without exceeding unity.
An Example

As an illustrative example, the third-orderChebyshev low-pass filter used by Mitra andSherwood [ 21 is synthesized in a lattice form. In this case, M = 3 and

One optimization criterion based upon rms values be largest for the h-parameter is t o require that T,&, nearest unity in magnitude. Then the sign parameters & n be as are recursively found by requiring that , large as possible without exceeding the maximum value. If that maximum occurs for m = 1, then the recursion relation proceeds for rn = I - 1 , l - 2, . . . , 0 and again for m = 1, I + 2, . * ,M - 1. This relation is simple to effect because of the fact that
1

P~(z)0.0154 + 0.0462 Z - ' + 0.0462 z =

- ~

+ 0.0154 z - ~ , (44)
A ~ ( z )1- 1.990 z-' + 1.572 z-' =
0.4583 z (45)

3 .

h 2 = - 0.4583 and v 3 = 0.0154. By simply changing the sign parameter, this ratio can From (4), always be made smaller than or larger than one. zB~(z) 0.4583 + 1.572 2 - l - 1.990 Z-' + x 3 . =Once the sign parameters chosen, are from the aforementioned or some other rule, the rms values of From (6) and ( 8 ) with m = 3, the multiplier inputs are simply evaluated, in addition to the rms value of the output from (39). Scaling of P 2 ( z )= 0.02245782 + 0.0219912 z-' + 0.076846 z 2 the input signal can be carried out by observing the and maximum rms value appearing in the filter and applying some sort of rule to relate rms value to probability A ~ ( z )1- 1.607107469 2 - l + 0.835462647 X-'. = of this peak value exceeding one. For example, if the Repeating the procedure from (4) through ( 8 ) input is Gaussian and scaled so that the peakrms with m = 2, value appearing in 'the filter is 3, then normal distribution tables show thatthe probability of that Fzl 0.835462647

7 1 , &m nrn+&GX

=.\/(I+ ~ , k , ) / ( l - ~,k,).

(40)

In order to illustrate the simplicity of the approach, and also to provide enough details so that a user could checkcomputer a program, all calculations were carried out using only paper, pencil and a calculator (HP-35). First, from (5) and (7),

GRAY AND MARKEL: DIGITAL FILTER SYNTHESIS

491

and
~2

= 0.076846

c3 = 0.0154 c2 = 0.0526956044
$ = 0.6063518248 ,

with the reduced order polynomials

Do = 0.1903083602.
2-l

PI
and

(2) =

- 0.0417441466 + 0.1454909806
(2)=

AI
Using rn
=

1- 0.8755871285 z - ' .
0.8755871285

1in the recursion relations gives

and
=

0.1454909806

with the final polynomials

This completes the design procedure the for onemultipliermodel. Theimplementation is shownin Fig. 5. Inorder to obtain an idea about relative sizes of the numbers being utilized, one can utilize results of the previous section and ApgendAk 11. Ln pzrticular, at stage m the nodes X A , / A M , XB, / A M , and X Z ~ /AMwill all have rms values given by nmfi,,, a , , for a random inputwith variance equal to . : a The values of a aremost easily evaluated from , (15). Starting with a M = 1,
a3 = 1 a 2 = 1.265885102 a 1 = 4.191642467 010 = 17.96311599.

P ~ ( z= 0.0856458833 )
Ao(2) = 1

so that v o = 0.0856458833. This completes the evaluation of the k-parameters and the tap parameters. Since all the k-parameters have magnitudes less than one, the filter is stable. These results define a two-multiplier lattice form. Next, the one-multiplier model with optimal sign choice is presented. To obtain the sign parameters, the algorithm of the preceding section is applied. Since k o has the largest magnitude, (41) and (42) result in 1 = 0,
Qo =

This then gives the results for the node values

0,

n2& u, = 1.640563849 u, nl& u, = 0.4912510573 u, n o G u, = 1.907389831 u,.

The multiplier inputs were shown t o have rms values given by n , c m u x d 2 / ( l + e,k, ). Therefore,
n1fi

and
qo = 3.882719036.

o,d-h,)

172fi2

u,.\/2/(1+ e o k o )= 1.969635389 a , = 1.712719502 a, CJ,.~~)= 1.921253614 a , .

As Qo is more than l / q o , choose eo Therefore

sgn ( k , )

-1.

The rms value of the output in this case is found simply from (39) tobe given by
uy = 0.4777060586 u,.
Conclusions

= Qo/qo =

0.2575514712.

From (42),
q 1 = 3.339954283.

Several different digital latticeandladderforms have been developed startingfromthestandard direct form. The procedure was shown to have several desirable features as follows. Q2 = = 0.8602101393. 1) All stable recursive filters can be transformed From (42), into the formspresented. Other procedures developed thus far have undesirable restrictions such as all poles q 2 = 1.640756072. being required to be in the left half of the z-plane. As Qz is more than 1/q2, choose ea = sgn ( k , ) = - 1. 2) A built-in stability test exists withinthe synthesis Having found the sign parameters, the multipliers process. If any k-parameter magnitude exceeds or x can be obtained from (22): , equals unity,thefilter is unstable.Otherwise, it is stable. 713 = 1 3) The synthesis procedure is very efficient. Only r 2 = 1.4583 on the order of M 2 operations are necessary for the 71, = 0.2399448219 synthesis. No polynomial root solvers or explicit 710 = 0.4500374194. simultaneous equation solvers are necessary. Using these with (20), the modified tap parameters Anewprocedure was also introduced for scaling are obtained as nodes of a particular one-multiplier lattice form. As Q 1 is less than l / q , , choose e l = - sgn ( k , ) Therefore
= - 1.

498

IEEE TRANSACTIONS ON AUDIO AND ELECTROACOUSTICS, DECEMBER 1973

Ylr)

Fig. 5. One-multiplier implementation of the example.

Presently, the application of this scaling procedure is being studiedforfinitewordlength digital filter implementation.
Appendix I

log I AI+, (e'') I - log I A,+,( e i e )I will oscillate between the values log [(l + k i ) / ( l + k l ) ]= log [l + A k l / ( l + k z ) ]

and Let A,(z) satisfy the recursion relation (13) where log [(I- h i ) / ( l - k , ) ] = IOg [l - A k , / ( l - k l ) ] . B ( z ) is from (4) given by , B, (2) = 2-(mC1)A (1/z), m (1-1) Appendix I I

A,,@) = 1,and m = 0, 1,' . ,M - 1. LetAL ( z ) satisfy Let Um( z ) , V ( z ) , and W , ( z ) be as given in (28), a similar set of recursion relations whose only differ- (30). When X ( z ) = 1, one can apply (11)and (12) ence lies in the fact that k , is replaced by hi = hl + to obtain their energies in terms of inner products. Akl. Thus A; (2) A , (2) for m = 0, 1, . . . , 1. Com- In this way we find that bining (13) and (1-1)with z = ej', one finds that on dB the unit circle 1 u,(eje) 12 271 = <~,(z)/.\/ol,,A,(z)/.\/or,)
" .

This can be rewritten as

Ai+,(ej') - 1+ hie-jq A l + l( e i 0 ) 1+ kle-i*

(1-2)

From (14) one should note that this is simply one. In exactly the same manner, one can show that
d8

I V (ej@) , i2

-=

2n

I.

(I + l ) e + 2 arg [ A l ( e i e ) ] . (1-3)

By using the form of A,@) from (3) and the fact that it has all of its zeros within the unit circle, one can show that

One more step is needed in the handling of W , ( z ) , for by the same applications of scalar products one finds that

- ( B , ( 2 ) - EmAm (21, B m ( z ) dq - > 1. dB

E,A,

(2))

2%l (1+ Gnhm 1

The ratio of (1-2) hasamagnitude that oscillates between the values (1 + h i ) / ( l + h l ) when \k is an even multiple of n, and (1- k i ) / ( l - h , ) when \Ir is an oddmultiple of n. As J( is amonotonically increasing function of 8 , these extreme values will all be met. In terms of the log spectra, this yieldsthe result that

If the numerator terms are multiplied out and use is made of (14) and (16), one finds that this quantity also equals unity. As a result,

GRAY

MARKEL: DIGITAL FILTER SYNTHESIS

499

In the same manner, one can obtain the energy of the overall filter output when X ( z ) = 1,by expressing (l), the energy of G ( z ) as a scalar product.From (111,and (12)

TABLE I The r m s Values at Nodesand Multiplier Inputs for One- and

Model
Two multiplier

rms Value

Node values and multiplier inputs

(a, : a

)1/2

If cw(z)is expressed as in (9), the orthogonality of give the polynomials zB, ( z ) can combine with(14) to

I=

One multiplier Node values One multiplier

4rnX/$M a m x/-$.M

n,

(a, 0 :

)1/2

vga,.

(11-2)

zBmX/AM

m=o

Thesymbolic notation H ( z ) X ( z ) represents the x transform of the output of a filter whose input time sequence has a z transform and whose transfer function is H ( z ) . Whenwe speak of the time sequence associated with H ( z ) X ( z )we imply the convolution
m

Sk

n = -m

hnXk-n

(11-3)

pressions for the z transforms of the node values and (35). multiplier inputs by applying (31) through These are summarized in Table I.
Appendix I I I

where hk represents the unit sample response to the filter and its z transform is H ( z ) . Equation (11-3) is considered t o apply, even in cases where the input sequence, xk, does not have a convergent z transform. In such cases, referring t o a time sequence associated with H ( z ) X ( z )will be understood t o imply (11-3), even when X ( x ) is not convergent. If the input sequence is random and uncorrelated, then the output sequence will also be random. Using the symbol E to denote the expected value (expectation or mean) of a random variable, one finds that the mean square of the sequence sk will be given by
m

We shall consider, here, only the U r n @ ) ,given by (28), forin the frequency domain, the only difference between U , ( e j e ) and V, ( e j e )lies in the phase since ~ , ( ~ j e ) = e-j(rn+l)OA* , j m ( e)* (111-1) The ratio of any two successive Urn( z ) terms can be simply expressed as

_____ +1

un (2) r

7 1-( z ) Am + ~ a /am+1
Am ( 2 )

This can be expressed in the frequency domain by ) applying (1-2) with 1 = r n , k ; = 0, so that A ; + , ( e j o = 0,"= E [ $ ] = h n h m E [ x k - n x k - m ] . (11-4) A , ( e j e ) . Combining this result with the recursion ren = - m ,=--oo lation of (18),one obtains If the inputis stationary and uncorrelated,
m

E[xk -nxk -m

1 = 0;hnrn

where \k is given by (1-3)with I = rn. The magnitude of the ratio of(111-2) oscillates be-______ is an even multitween fll + k , )/(1 - k , ) when ple of n, and d(1- k,)/(l + k , ) when \k is an odd multiple of n. From Appendix I, it should be noted that q is a monotonically increasing function of 8 , going from q(0) 0 to \.k (n) = (rn + 1)n. Thus, as 8 = By taking the definitions of Urn( z ) , V ( z ) , and goes from zero to n, the ratio of(111-2)will hit its , Wm(z),along with the unit energy results of(11-1), extreme values at exactly rn + 2 points, two of them one can then use (11-5) to note that when the input being the end points 8 = 0 and 8 = n. time sequence is random, stationary, and uncorrelated, In terms of the log spectrum, this yields the result , the time sequences associated with Urn(z), V ( z ) , and that the difference W , (x) will all have a mean square equal to the input log I U m + l ( e i e I) - log I u m ( e j e ) I mean square and, thus, rms values equal to the input rms value, ax. 'In addition, the rms output value of hits the extremevalues the overall filter is equal t o .,.I where I is found log [(I + krn )/(I- km 11 from (11-2). These results can be utilized with the various ex- exactly rn + 2 times. Thusthe surprising result is

where 02 is the input meansquareand t i n , is the Kronecker delta. In this case, (11-4) can be reduced to a single summation, and by applying z-transform properties one finds

*+

5 00

IEEE TRANSACTIONS AUDIO ON AND

ELECTROACOUSTICS, VOL. AU-21, NO. 6, DECEMBER 1973

shown that U,(z), and similarly V, ( z ) , form minmax sequences. Infact,thisresultcan be used to show that A m ( z ) / G is the mth order polynomial having all its roots inside the unit circle which is the mini-max approximation in termsof the log spectrum to A , + I ( d / G l . The phase of the ratio of (111-2) can also be studied, and it can be shown that the phase of the ratio will hit its extreme values, t sin-? ( k , ) , exactly m + 1 times as 8 goes from 0 to 7. r
References [ l ] S. K. Mitra and R. J. Sherwood, Canonic realizations of digital filters using thecontinuedfractionexpansion, IEEE Trans. Audio Electroacoust., vol. AU-20, pp. 185194, Aug. 1972.

Digital ladder networks, IEEE Trans. Audio Electroacoust., vol. AU-21, pp. 30-36, Feb. 1973. B. Gold and C. M. Radar, Digital Processing of Signals. New York: MzGraw-Hill, 1969. A. Fettweis, IXgital filterstructures related to classical filternetworks, Arch. Elek. Ubertragung., vol. 25, pp. 79-89, Feb. 1971. , Some principles of designing digital filters imitating classical filterstructures. IEEE Trans.CircuitTheorv..,, voll~CT-18, 314-316,Mar. 1971. pp. R. Crochiere, Digital ladder filter structures and coefficient sensitiiity,Res. Lab. Electronics Rep. 103, Mass. Inst. Tech., Cambridge, Oct. 15, 1971. F. Itakura and S. Saito, Digital filtering techniques for speech analysis and synthesis, presented at the 7th Int. Congr. Acoust., Paper 25C-1, Budape::, 1971. J. D. Markel and A. H. Gray, Jr., Onautocorrelation equationswithapplications to speech analysis, IEEE Trans. Audio Electroacoust., vol. AU-21, pp. 69-79, Aor. 2973. x--~- . - E. A. Guillemin, Synthesis of Passive Networks. New York: Wiley, 1957.

Conside

FRANCIS BROPHY and ANDRES C. SALAZAR, Member, IEEE

Abstract-The Pad& approximant technique provides a quick design of recursive digital filters. An added advantage of the technique lies in that spectrum shaping requirements as as well linear phase constraints can be handled easily, even for higher order Titers. This is important in supplying initial guesses of the filter parametersto iterative routines thatwould then seek a locally optimal design solution. These advantages are among those discussed in a partly tutorial presentation of the technique that relates to filter needs found in data transmission systems. In addition, the question of stability is treated and a new criterion is presented.Thecriterion provides sufficient conditions inestablishing stability fora Titer designed by using the Pad6 approximant technique.

that calculates the extremumof an object function of several variables. The function is generally nonlinear and positive definite and indicates the closeness77 of the designed spectrum to the desired spectrum.In some cases, depending on the complexity of the function, the number of iterations or even convergence t o an extremum is dependent on the,initial guess for the a and 0 (feedforward and feedback) parameters. By working in the time domain thedegrees of freedom available can be used t o match a set of time samples exactly, thus reducing the design to the solution of a linear system of equations. While this approach, call the Pad6 approximant, does not lead t o a locally optimalsolution as aniterativetechnique would, it nevertheless provides a viable solution in a fraction of the time. In the following we show how the Pad6 approximant technique can yield a simple digital filterdesign for spectrum shaping networks with linear (or nonlinear, if so desired) phase constraints often required in data transmission systems. The problem of stability is discussed and sufficient conditions are given t o ensure that the design procedure will not lead t o an unstable filter.
II. Pad6 Approximate inDigital Filter Design

I. Introduction

Let H ( w ) denote in the interval [-27rW, 2xWI the bounded filter amplitude characteristic that is to be synthesized. Since H ( w ) E L, [- 2 x W , 2n W ] ,p Z 1it has a Fourier series expansion

The design of spectrum-shaping recursive digital filthe in ters z-plane often requires the use routine of a
Manuscript received December28, 1972; revised June 29, 1973. Theauthorsarewith Bell Laboratories, Holmdel, N.J. 07733.

H ( ~ ) =

hne-jnwi2w

(1)

Pronys method is related to the Pad6 approximant techniquethrough a transformation of variables (see 111 formore details).