IEEE

TRANSACTIONS

ON ACOUSTICS, SPEECH, AND

SIGNAL

PROCESSING, VOL. ASSP-32, NO. 6 , DECEMBER 1984

1243

theprematuretermination,
By theassumptioninSection
I,
f ( z ) should be the common factor of F1( z ) and F z ( z ) , or we
have D ( z ) = f ( z ) c ( z ) where c ( z ) is apolynomialfunction.
Then we can continue to test all zeros of f ( z ) byapplying
Property 3. Therefore,witbProperty2,thesufficientand
necessary conditions for all zeros
of f ( z ) c ( z ) being inside or
on the unit circle are that it is alwayspossible to obtain all
the real and positive Kis, for 0 i n - 1.

<<

Q.E.D.

Property 5 191: If noprematureterminationoccurs,then

D ( z ) has r zeros outside the unit circle and n - r zeros inside
the unit circle, where r is the number of Kis in (5), which are
negative.

(3,E, 9).

No prematureterminationoccurs;therefore,
zeros of D ( z ) are inside the unit circle and set
f ( z ) = 1.
S t e p 2:
@k(w>=

all

w6+2ws+5w4+4w3+5w2+2w+l
8w6 + 12w5 + I 2 w 4 - 12w2 - 12w - 8

No premature termination occurs and there are two negative

1 180 49
32
567
numbersin(Ko,K1,K2,K3,K4,Ks)=(i~,--;r,~.,~,-~,
therefore, it has one pair of complex and one pair of
real roots inD ( z ) .

-7%);

REFERENCES

[ 11 E. A. Guillemin, The Mathematics of Circuit Analysis. New York:

111. A SIMPLE TRANSFORMATION
ONADISCRETE SYSTEM
Wiley, 1962.
[ 2 ] M. E. Van Valkenburg, Modern Network Synthesis. New York:
Now let us apply a transformation
Wiley, 1962.
[3] .S. Barnett,Mabicesin Control lheory. New York: 1971.
w + w-l
z-f
[4] H. W. Schussler, A stability theorem for discrete systems, IEEE
2
Dans. Acoust., Speech, Signal Processing, vol. ASSP-24, pp. 8789, Feb. 1976.
or
[ 5 ] R. Gnanasekaran, A note on the new 1-D and 2-D stability theorems for discrete systems, IEEE Trans. Acoust., Speech, Signal
Processing, vol. ASSP-29, pp. 1211-1212, Dec. 1981.
D ( z ) -P D(w) = wnD
[ 6 ] N. K. Bose, Implementation of a new stability testfortwo dimensional filters, IEEETrans. Acoust., Speech, Signal ProcessThen a real rootzl, with -1 z1 1 , maps into a pair of coming, vol. ASSP-25, pp. 117-120, Apr. 1977.
plex roots on the unit circle in the W-plane and a pair of com- [ 71 J. Szczupak, S. K. Mitra, and E. I. Jury, Some new results on discrete system stability, IEEE lirans. Acoust., Speech, Signal Pro)
the unit circle will map into a set of
plex roots ( z 2 , z z inside
cessing, vol. ASSP-25, pp. 101-102, Feb. 1977.
four roots (w;, w y , w;,w;*)
of symmetryabouttheunit
[ 81P. Steffen, An algorithm for testing stability of discrete systems,
circle in the W-plane, two of them are inside the unit circle,
IEEE Pans. Acoust., Speech, Signal Processing, vol. ASSP-25, pp.
and the others are outside the unit circle. For a stable system
454-456, Oct. 1977.
w_e performthecontinuedfractionexpansionin@kn(w)
=
191 F. R. Gantmacher, The Theory ofMatrices, vols. 1, 2. New York:
D(w)/Dd(w), since F2(w) = 0 for D ( w ) polynomial. Two posiChelsea, 1959.
tive entries in [Ki} will be created by a real
root z1 in D ( z ) ;
__I

(+;-).
~

< <

two positive and two negative entries in

by a pair of complex roots ( 2 2 , z g ) .

[ K i ] will be created

I v . AN ALGORITHM TO TEST AND AN EXAMPLE

Step 1: Apply the continued fraction expansion on
@,(z).
If it ends prematurely, apply Property 3, and continue the expansion. The resulting K:s will be all positive. Then all zeros
of the polynomialf ( z ) ,which causes the first premature termination of &(z), appear on the unit circle. It is easy to determine the number of roots on z = 1 (or -1) and the pairs of
complex roots on the unit
circle bytesting f ( z ) . Of course,
no premature termination occurs for
a stable system and we
set f ( z )= 1 .
S t e p 2: Obtain E ( z ) = D ( z ) / f ( z ) ,which has all its zeros inside the unit circle, and ap@y another continued fraction expansion on @&(w)= E(w)/Ed (w), until all new 21 entries can
eventuallybeobtained,where
1 denotesthedegree
of E@).
If 2n1 negative and 2(n1 + n 2 ) positive entries appear after the
kth premature termination, which should also appear k times
nl pairs of
before the kth premature termination, there exist
complex and n 2 real roots of multiplicity ( k + 1 ) in E ( z ) . If
noterminationoccurs,
itimplies that all roots of E ( z ) are
distinct inside the unit
circle. Then the number 211 in { K l } ,
which is negative, corresponds to the number
lI of pairs of
complex roots of E ( z ) and the number 1 - 211 corresponds t o
the number of real roots inE ( z ) .
Example /8]: Let D ( z ) = 2z3 + 2 z 2 + z.
Step I: Then

All positive entries are obtained

in [8], i.e., ( K O , K 1 , K 2=)

A New Algorithm to Compute the Discrete

Cosine Transform
BYEONG GI LEE
Abstract-A newalgorithm is introduced for the 2m-point discrete
cosine transform. This algorithm reduces the number
of multiplications
to about half of those requiredby the existing efficient algorithms, and
it makes the system simpler.

INTRODUCTION
During the past decade, the discrete cosine transform (DCT)
[ 11 has foundapplicationsinspeechandimageprocessing.
Various fast algorithms have been introduced for reducing the
number of multiplications involved in the transform [ 21 -[ 61.
Inthiscorrespondence
we proposeanadditionalalgorithm
which not only reduces the number of multiplications but also
has a simpler structure. We refer to this algorithm as the FCT
(fastcosinetransform),sinceit
is similar to the FFT (fast
Fourier transform). The number
of real multiplications it requires is about half thatrequiredbytheexistingefficient
algorithms.
ALGORITHM
DERIVATION
We denote the DCT of the data sequence x(k), k = 0 , 1 , . *
N - I,byX(n),n=O,l;..,N1. T h e n w e h a v e [ 1 1
Manuscript received August 15, 1983; revised February 2 9 , 1 9 8 4 .
The author is with the Granger Associates, Santa Clara, CA 95051.

0096-35 18/84/1200-I 243$01 .OO 01984 IEEE

IEEE
TRANSACTIONS
ON
ACOUSTICS,
SPEECH,
AND
SIGNAL
PROCESSING,
VOL.
ASSP-32,

1244

NO. 6 , DECEMBER 1984

k=O,l;*-,N- 1

and

n=O
N

,
l

-1

because
(2k+ 1)2 ( N / 2 ) = (2k + 1) =
'2 N
C2

n=O,l;.*,Nwhere

e(n)=

[ ;;fi,

if n

Thus
(2)

0.

rewritten
be
ascan
(1 2)

=o,

otherwise.

n=O

Therefore,
we
haveindecomposed
N-point
IDCT
the
(5) into
(8b) the sum of two N/2-point
IDCT's in (1 8). By repeatingthis
n=O
further.IDCT
process,
the
decompose
we can
= o, 1, . . . , N/2 - 1, forms an ~ / 2 -IDCT,
~
~
i
We
~
can
~
also decompose the DCT in a similar manner. AlterClearly, g ( k ) ,
natively, the DCT can be obtained by "transposing" the IDCTsince
i.e., reversing the direction of the arrows in the flow graph of
(2k+1)2n = (2k+l)n = C(2k+ 1)n
c2 N
CN
2 (NI2)
IDCT,
the
since
(9)
DCTorthogonal
is an transform.
the
EXAMPLE
We rewrite h ' ( k ) in
With N = 8, ( 17)-( 19) yield
N/2-1 A
A
(2k + 1)n
h'(k)=
X'(2n + 1) CZ(N/2)
(lo)
G ( n ) = X(2n),
N/2-1

form

n=O

which is another N/2-point IDCT. Since

(2k+1) C(2k+1)(2n+l)
=C(2k+1)2n
(2k+1)2(n+l)(ll)
2c2N
2N
2N
+2
',
we

+ 1) + X(2n -

l),

n = 0, 1 ,2 , 3( 2 0 b )

and
3

g(k)( 2=1 4

G ( n ) C $ 2 k +l ) n ,

n= n

N/2-1

2cy;+l) h ' ( k ) =

H ( n ) = X(2n

X(2n

+ 1) Cp;+l)an

n=O

NI2-1

X(2n

+ 1) C p i + 1 ) 2 ( n + 1 () 1. 2 )

n=O

x(k) = g ( k ) + (1/(2C,26+'))h(k),

(22a)

x(7-k)=g(k)-(l/(2C126++1))h(k), k = 0 , 1 , 2 , 3 .

(' 3,

Equations (20) and (22) respectively form the first and the last
stages of the flow graph in Fig. 1. By repeating the above steps

IEEE TRANSACTIONS ON ACOUSTICS,

SPEECH,

AND
SIGNAL

PROCESSING, VOL. ASSP-32, NO. 6, DECEMBER 1984

1245

cosinetransform, IEEE Pans. Cornput., vol.C-25,pp.764765, July 1976.

B. D. Tseng and W. C . Miller, On computing the discrete cosine
transform, IEEE nuns. Cornput., vol. C-27, pp. 966-968, Oct.
1978.
W. H. Chen, C. H. Smith, and S. C. Fralick, A fast computational
algorithm for thediscretecosinetransform, IEEE Trans. Commun.,vol. COM-25, pp. 1004-1009, Sept. 1977.
M. J. Narasimha and A. M. Peterson, On the computation of discrete cosine transform, IEEE Pans. Commun., vol. COM-26, pp.
934-936, June 1978.
J. Makhoul, A fast cosine transform in one and two dimensions,
IEEE naris. Acoust., Speech,
Signal Processing,vol. ASSP-28, pp.
27-34, Feb. 1980.

Fig. 1.

On the Interrelationships Among a Classof Convolutions

JAE CHONLEE

TABLE I

AND

CHONGKWAN UN

Abstract-In this paper some interrelationships among a class of circularoperationsareinvestigatedbased

on matrixformulation. It is
shown that a class of convolutions representing forward/backward and
convolution/correlation of two periodicsequencesmayberelated
to
each other in terms of discrete transforms having the circular convolution property. The results obtained are useful in efficient realization of
adaptive digital filtersusing fast transforms.

I. INTRODUCTION
The need for computing convolution of two functions arises
8
1024
1668
256
2690
2817
inmany diverseapplications.Theseincludedigitalfiltering,
spectrum analysis, time delay estimation, computation of dis6146
9
6401
512
2304
3844
crete Fourier transform (DFT) using circular correlation, mul10
1024
tiplicationoflargeintegers,polynomialtransforms,and
so
8708
5120
13826
14337
forth [ 1 1, [ 21. In computation of .various convolutions, the
11
2048
11264
19460
30722
31745
fast convolution approach using efficient computational algorithms of discrete transforms has proven to be useful[3].
12
4096
67506
21576
43012
69633
Recently,discretetransformsbasedonnumbertheoretic
concepts have received considerable attention as a method for
efficientanderror-freecomputationofdigitalconvolutions
on(21), we obtaintheFCTflowgraphforaneight-point
121. Unlike thefastFouriertransform(FFT),thenumber
IDCT as shown in Fig. 1.
theoretic transform (NTT) does not cause roundoff errors
in
arithmetic operations. Particularly, the Fermat number transCONCLUDINGREMARKS
form that is one of the NTTs requires only word shifts and
.additions, butnotmultiplications,northestorageof
basis
It follows from Fig. 1 that the flow graphs of the FCT and
FFT are similar. The number
of real multiplications thus apfunctions.Accordingly,
theNTThas severaldesirableproppears to be (N/2)log2N for an N-point FCT withN=
2 m , which erties in carrying out various convolution operations in comis about half the number required by existing efficient algoparison to the FFT.
rithms. The number of additions, however, is slightly higher
In this correspondence, we consider a class of convolutions
and given by (3N/2)log2N- N + 1. See Table I for a compar- that include forward and backward convolutions of two peri[ 41.
A
ison
with
algorithm
the in
odicsequencesandalsoforwardandbackwardcorrelations.
If Fig. 1 we also note that the input sequence
X ( n ) is in bit- Based on matrixformulation, we studytheirinterrelationreversed order. The order of the output sequence x(k) is gen- ships. Particularly, we show that they may be related to each
erated in the following manner: starting with the set
( 0 , l ) , other through a discrete transform such as DFT and NTT.
form a set by adding the prefix 0 t o each element, and then
11. INTERRELATIONSHIPS AMONGA C.LASS OF
obtain the rest of the elements by complementing the existing
CONVOLUTIONS
(00, 01, 1 1, 1.0), and by
ones. This process results in the set
repeatingit we obtain (000, 001, 011, 010, 111, 110, 100,
Here we discuss a class of circular operations based on ma101).Thus, wehave theoutputsequencex(O),x(l),
x ( 3 ) , trix formulation. In the following discussion it is assumed that
x(2), ~ ( 7 1~, ( 6 1x(4),
,
x(5) for the case N = 8; see Fig. 1 .
various arithmetic operations including matrix operations are
7

128

708

448

1154

1217

REFERENCES
[ l ] N. Ahmed, T. Natarajan, and K. R. Rao,Discretecosinetransform,IEEE nuns. Compur.,vol. C-23, pp. 90-94, Jan. 1974.
[2] M. R. Haralick, A storage efficient way to implementthe discrete

Manuscript received May 18,1983;revised May 17,1984.

The authors are with the Communications Research LaFoTatory, Department of Electrical Engineering, Korea Advanced Institute of Science and Technology, Chongyangni, Seoul, Korea.

0096-3518/84/1200-1245$01.00 O 1984 IEEE