350 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 35, NO.

2, MARCH 1989

Nonparametric Time Series Analysis for

Periodically Correlated Processes

Ahstruct -Correlation functions of continuous-time periodically corre- for every (s, t ) E R X R . For the remainder of this discus-
lated processes may be represented by a Fourier series with coefficient sion we shall take rn( t ) = 0.
functions { B k ( 7 ) , k E Z ) , where T is the usual lag parameter. We show
that the usual estimator for stationary covariances, formed from a single
If X( t ) is periodically correlated, then B( t , 7 ) =
sample path of the process, can be simply modified to provide a consistent R ( t + 7,t ) is periodic in t with period T for fixed 7,and
(in quadratic mean) estimator for any of the coefficient functions resulting so the Fourier series representation
from the aforementioned representation. Further, if the process is Gauss-
ian and B k ( 7 ) is a Fourier integral with respect to a density function
gk (A), we show that a two-dimensional periodogram, formed from a single
sample function, may be smoothed along a line of constant difference
frequency to produce a consistent estimator for g k ( A ) . This natural
extension of the well-known procedure for stationary processes provides a
method for nonparametric spectral analysis of periodically correlated pro-
cesses. is suggested where the coefficient functions B k ( t ) are given
by
I. INTRODUCTION
AND BACKGROUND

PP1 ERIODICALLY correlated stochastic processes [l],

are nonstationary, yet exhibit many of the proper-
ties of stationary processes. The periodic nature of the
nonstationarity has produced a number of interesting syn- assuming only that B( -,7)E LJO, TI. Convergence of (2)
onyms for periodically correlated processes. They have depends on B ( t , T), and we note that B ( t , 7 ) is not neces-
been called cyclostationary [3], [4], periodically nonstation- sarily continuous in the variable t because periodically
ary [5], [6], periodically stationary [7], periodic nonstation- correlated processes are not necessarily continuous in
ary [SI, and processes with periodic structure [9], [lo]. quadratic mean. A simple example is provided by the
Periodically correlated processes are, through several rep- periodically correlated process formed by X ( t ) = P( t ) Z (t ) ,
resentations, equivalent to possibly infinite-dimensional where Z ( t ) is a wide-sense stationary process with contin-
stationary vector processes [l], [4], [8], [ll]. In addition, a uous covariance and P ( t ) is a discontinuous periodic
second-order process is periodically correlated if and only function. However, under reasonably weak conditions each
if it can be made stationary by an independent uniformly B k ( 7 ) is a Fourier transform; if B ( - ,7 )E L,[O,TI for

distributed random time shift [12]; this is a paraphrasing every 7,the following conditions are equivalent [13]:
of the more technical statements presented in [12].
A continuous-time stochastic process { X ( t , U ) , t E R }
defined on a probability space ( Q , I;, P) and having fi- 1) Bo( T ) is continuous at T = 0;
nite second moments is called periodically correlated 2) B ( . , 7 )is i-continuous in L,[O,TI norm;
3) B,(T)= /?mexp(iAT)dGk(A) where
(Gladyshev, [l], [2]) with period T if

E { X ( s ) X * ( t ) }= ~ ( s t ,) = R ( s + T , t + T ) (la)
j-: I dG, ( A ) I I Bo(0) Jm
=
-m
dGo(A ) , k E 2.

+
E { X ( t ) } = r n ( t ) = rn(t T ) Ob)
For wide-sense stationary processes, one can construct
estimators for the covariance and spectral density func-
tions that are consistent in quadratic mean [14], [15]. In
Manuscript received April 18, 1987; revised April 27, 1988. This work
was supported in part by the Office of Naval Research under Contract this paper, we show that the usual estimator for stationary
N00014-86-C-0227. covariances, formed from a single sample path of the
The author is with Harry L. Hurd Associates, 309 Moss Run, Raleigh, process, can be simply modified to provide consistent (in
NC 27614, and with the Center for Stochastic Processes, Department of
Statistics, University of North Carolina, Chapel Hill, NC 27514. quadratic mean) estimators for the sequence of coefficient
IEEE Log Number 8926616. functions Bk( T ) appearing in (2).

0018-9448/89/0300-0350$01.00 01989 IEEE

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 18,2024 at 06:55:52 UTC from IEEE Xplore. Restrictions apply.
 HURD: NONPARAMETRIC TIME SERIES ANALYSIS 351

Specifically, we give, for a real periodically correlated ever, we now must deal with sequences of functions and
process, conditions in which the estimator sometimes with interactions between members of the se-
quences.
The consistency result may be interpreted as follows. In
smoothing gk(A , A ) to estimate gk(A), one may expect the
smoothing to produce a bias, as in the stationary case. The
consistency result says that the smoothing may be chosen
to produce a compromise between bias and variation so
that estimators having any prescribed variance may be

1 . e x p [ F ] d t , T < O (4b)
obtained provided A is chosen sufficiently large. In this
sense, by adjusting the smootlung as A is increased, one
may expect a better estimate by takmg a larger sample.
Evidently, it is the continuing nature of the periodic non-
converges in quadratic mean to B k ( 7 ) as A + m . For stationarity that permits this to occur.
k = 0, (4a) and (4b) reduce to the familiar estimator of the Early work along these lines was done by Gudzenko [5]
covariance or asymptotic covariance who briefly discussed the Fourier series representation ( 2 )
and gave conditions for consistent estimation of Bk( 7).
1 Kampt de Ftriet [17] was motivated, in part, by periodi-
jgA
I?( A , T ) = 2 - IT1x( t + T ) x(t ) dt. cally correlated processes in his approach to the definition
of c2rrelation and spectrum for nonstationary processes; if
If B ( t , 7 ) is defined symmetrical9 by B ( t , T ) = R ( t + E { Bo( A , 7 ) ) converges to some Bo( T ) as A + cc for some
r/2, t - 7 / 2 ) , then the estimators B,(A, 7 ) may be given nonstationary process X(t), then he defines B,(T) as the
by a single expression; in this paper we have chosen to correlation of the process and similarly for the spectrum.
keep the definition of B ( t , T ) consistent with [l],[2], and T h s work, however, was concerned primarily with the
[ l l ] to use directly the interpretation obtained from the presentation of this idea along with a few examples and
representation of periodically correlated processes and their not with the issue of consistent estimation. Processes hav-
covariances. ing this property were called asymptotically stationary by
If X ( t ) is Gaussian and the Bk(T ) are Fourier integrals Parzen [18], but the question of consistency was left open.
with respect to density functions, In a subsequent paper, Parzen [19] addressed the problem
of consistent estimators for the spectral density of the
process Y ( t ) where one is given samples of the process
X(t) = g(t)Y(t), and g(t) is assumed to have a general-
ized harmonic analysis in the sense of Wiener [20], [29];
however, consistent estimation of the asymptotic covari-
then consistent estimators for the density functions can be
ance and spectrum of X(t) was not treated. We note that
constructed in a manner that is a natural extension of the
some related work concerning processes having periodic
stationary case [14], [15]. Specifically, the finite Fourier
variances was reported by Herbst [21].
transform
More recently, Gardner [16], [22], [23] has developed a
number of techniques for estimating cyclic autocorrela-
tions and cyclic spectral densities which, for periodically
correlated processes, may be identified with B k ( 7 ) and
gk(A) as before; the question of estimator variance is
is used to form a shifted periodogram (or a cyclic peri- discussed, but conditions sufficient for consistency are not
odogram in Gardner's nomenclature [16]) given. Gardner treats cyclostationary processes in [22] from
the probabilistic viewpoint whereas the estimation tech-
1
gk(A, A ) =--l,(A)l,*(A -2~k/T) ( 6 ) niques [16], [23] are treated from the approach that the
2 mA quantities to be estimated are defined by time averages.
This paper, based in part on the author's dissertation
yhich, for fixed A , is shown to be the Fourier transform of [24], uses the probabilistic set-up; we give sufficient condi-
B , ( A , T ) and an asymptotically unbiased but inconsistent tions on the processes to ensure both the existence of
estimator of gk(A). For Gaussian processes we show that sample path integrals and the consistency of the estima-
consistent estimators of the gk(A) may be constructed by tors. Our results also contribute to the issue of consistent
smoothing the shifted periodogram, a well-known [14], [15] estimation of the asymptotic covariance and spectrum
procedure in the stationary case. This provides a method because the periodically correlated processes discussed here
for nonparametric spectral analysis of periodically corre- are asymptotically stationary in the sense of Parzen [18]
lated processes. We note here that the methods of proof (see also Gardner [25]), and the asymptotic covariance and
are direct extensions of those used by Parzen [14]; how- spectral density are B,(T) and gO(A).

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 18,2024 at 06:55:52 UTC from IEEE Xplore. Restrictions apply.
 352 IEEE TRANSACTIONS ON INFORMATIONTHEORY, VOL. 35, NO. 2, MARCH 1989

11. ADDITIONAL INTERPRETATION

FOR odogram
X ( t ) HARMONIZABLE 1
A new meaning can be given to the distributions G,(A) g ( A , A,, A,) = 5--&A(~l)LV2)

in the event X ( t ) is harmonizable in the sense of L o k e restricted to s,; the estimators for g k ( ~ are) formed by
[261 (also known as strongly harmo*zable); in this case the smoothing g( A , A,, A,) along the fine S,. This illuminates
covariance is a Fourier transform of a finite measure the contrast with the stationary case in whch the smooth-
ing of g ( A , A,, A,) is along the diagonal A, = A,.
R(s,t )=E{ X(s)X*(t)}

= J J exp(iX,s - i ~ , t ) r (, d ~ ,d , ~ , ) (7)
R R III. CONSISTENCY OF j , ( ~7 ),
where
We assume throughout that the process X ( t ) is real
valued, has zero mean, is jointly measureable, and has
uniformly bounded fourth moments, E { X4((t)} I M.
These two assumptions are frequently used and, for exam-
The treatment of weakly harmonizable processes [27],[28] ple, ensure that (4) exists as a quadratic mean integral, that
is unnecessary here as we only wish to note the following R ( t + T , t ) is jointly measurable in t and 7,and that X ( t )
relationship between r y and the distribution functions also has bounded nth moments for n 2 4. The bounded
Gk(A)* moments also produce
If X(t ) is periodically correlated with period T , then the
support of r y is contained in a set S=U,Sk of equally
spaced lines parallel to the main diagonal [ l l ] ,[24] where

The distribution functions Gk(A ) may be identified with (13)

the restriction of r y to the k th line Sk and
and

for every r . Whenever we address a periodically correlated

from the preceding, it follows that if X( t ) is harmonizable,
process X ( t ) , we take the period to be T. Finally, (4)
then ( 2 ) converges absolutely. Finally, a harmonizable
implies that we must have a sample path of length A , and
periodically correlated process (or sequence) has the repre-
for convenience we assume this sample is on the interval
sentation [l],[4], [8],[ll]
0 s t I A . The following proposition describes the bias
produced by the use of the shortened interval of integra-
tion in (4) and shows, by setting k = 0, that periodically
correlated processes considered here are asymptotically
stationary [18] and the asymptotic covariance is Bo(7).
where the a k ( t )are jointly wide-sense stationary, are band Proposition I : If X ( t ) is periodically correlated, then
limited in the sense that the frequency support of each is
limited to the interval [0,27r/T), and have cross covari-
ances determined by

where € ( A ,T ) is bounded for all A and T . Further, if

B ( . , T ) is in L,([O,TI X R ) , then there is a constant K
= j:‘il/rexp
(2: )
[ i A ( s- t ) ] dGk_j -+ A . ( 1 2 ) such that

/ p m m l ~ ( A , ~K
) l. d ~ ~ (16)
Therefore, the distribution function Gk(A) on the k th line
S , determines all the cross-spectral distributions Gp9(A ) Proof: For fixed T > 0, the process X ( t + T ) X ( t ) is
for w l c h k = q - p . In this manner, the family { G k ( A ) , jointly measureable and has uniformly bounded second
A E R , k E 2 } determines the doubly indexed family moments, and thus
{ Gp9(X),A E [0,27r/T), p E Z , q E Z }, and conversely.
In view of the identification of G k ( A )with the restric-
tion of r y to s k , the shifted periodogram g , ( A , A) given
by (6) may be interpreted as the two-dimensional peri-

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 18,2024 at 06:55:52 UTC from IEEE Xplore. Restrictions apply.
 HURD: NONPARAMETRIC TIME SERIES ANALYSIS 353

This expected value may be expressed as are true, then

l.i.m.k,(A,I-) = B,(I-) (23)
lcTR(t+I-,r)exp A+CC

A where 1.i.m. means limit in quadratic mean.

+lJm;-'R(t+I-,r)exp
1 Remark: The result holds for any T for which a) or b)
are true; the result holds for all I- if c) or d) are true.
= - mT
B , ( I - ) + ~ J1~ - - ~ R ( ~ + T , ~ ) Proof: We will provide the proof for I- 2 0; the proof
A mT for T < 0 follows in the same manner by adjusting the

.exp [7- i2akt

1 dl
interval of integration as indicated in (4b). If the quantity

where m is the largest integer in (A - T ) / T .Since A - I- =

+
mT a where 0 I a 2 T, then (18) can be rewritten as (15) = k,( A , .) - E { B,( A , .)} (24)
with
converges in quadratic mean to zero as A + 00, we may
use the triangle inequality and Proposition 1 to conclude
that

Finally, it may be seen that

where M is the bound on the fourth moments of X(t). For

I- < 0, (15) is again obtained but with
converges to zero as A + 00.
€(A,T)=- uB,(I-)+ That condition a) is sufficient for E { J ; ( A ) } + 0 fol-
lows from
and the inequality (20) remains true. The second assertion 1
is established by integrating the two leftmost quantities of E { J,2(A)} I
A
JmIRZ,-7(U,4 IdUdU. (26)
(20) with respect to T over ( - 00, CO), noting that R( t + I-,t )
-m -m

is jointly measureable in t and I-. 0 For condition b), let < > 0 be given and choose A, such
From the preceding result it follows directly that that IRz,-7('+ ' 9 < €I2for 1' > Define the sets
j k ( A ,I-) is asymptotically unbiased as E = [0, A ] x [0, A ] and F = E n {(U, U ) : IU - U I < Ao}.
Then
lim
A+m
~ ~ E {/ k , ~( A ,1 I-) } - B , ( I-) I = o

for any k and T. We now define the process

Z ( t , I - )= X ( t + T ) X ( t ) - R ( t + T , t ) (21)

so that Z( t , I-) has correlation function

R&, U ) = E { Z ( U , I - ) Z ( U ? I-)}. (22)

With this notation y e can proceed to answer the consis-

tency question for Bk(A, I-).
Proposition 2: If X(t) is periodically correlated and if 4MA,A
any of the conditions a)-d): I €/2+ ~

A2 <E
-
(27)
a) /00m./00mlRZ,-7(tl, t2)1 dtl dt2 < OO;
b) +
limu+, R , , ( t U, t) = 0 uniformly in t; provided A>8A,M/r where M is the bound on the
c) X ( t ) is Gaussian and 100mlzB2(t,I-)d t d ~
< 00; fourth moments of X(t). For condition c) we use the
d) X( t ) is Gaussian and lim. ,B( t U, t ) = 0 uni-
~ + fourth-moment property for real jointly Gaussian variates,
formly in t ; noting that one of the three terms is negated by the mean

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 18,2024 at 06:55:52 UTC from IEEE Xplore. Restrictions apply.
 354 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 35, NO. 2, MARCH 1989

of Z ( t , T ) , to obtain where
up = Bp( u + c) B n p p( u + f )
eexp [22mp(au b ) / T ] +
.exp[i2m(n--p)(du+e)/T],
I~(U,t,,t,)IIK1<0O, (32)

r,, t 2 ) Idu
J-OOOOl€(u, K, <0 , (33)

I $[ /OR/OAR2(U + 7 ,U + 7 )dudu
Ill2
and K,, K , may be chosen to be independent of t , and t,.
Proof: We first denote Z ( t ) as the integrand appear-
. [iAiAR2( U, u ) dudu ing in (31). Then by defining t;=t,+ mT where m =
Ill2 [ ( t 2 - t , ) / T ] , the number of whole intervals of length T in
the interval [t,, t,], we obtain
+f [ / O A i A R 2u( + 7 ,U ) dudu
Ill2 ~ , , ( ut,, , t 2 ) = J r i z ( t )dt + J r z ~ ( tdt)
.[
tl ti
+ 7 )dudu I l l 2.
/Dd/OAR2(U ,U (28)
=mT c
00

p'-OO
up+/t*I(r)dt
ti
In the first of the preceding four integrals, the transforma-
tion s = u - U , t = U yields
J A J A R 2 (u + 7 ,U + 7 )dudu p=-m
0 0 00

-(t2-t;) u p + p ( ' ) d f . (34)

<1-Ldsi2AdtB2(t+7,s) p=-m

The term €(U, t,, t 2 ) appearing in (31) is identified from

I ( 1 + [ 2 A / T ] ) J w d s J r d t B 2 ( t + 7 ,s) (29) (34) as
-m 0
00
where [ X I denotes the greatest integer not exceeding X.
€(U,tl,t2) =J'2z(f)dt-(t2-t;) up (35)
The other three integrals appearing in (28) are also bounded ti p=-m
by the last quantity in (29) (which is independent of 7);
hence the required integral is bounded by where ((U, t,, t,) = 0 if ( t 2 - t,)mod(T) = 0. To establish
bounds on IC( U , t,, t2)1, we observe
'2 ( 1 + [ 2 A / T ] ) CO
A, J00dsptBz(t+7,s),

which is O( A - ' ) as required.

For condition d) we use the fact that for any > 0 there
is an A , such that Is(> A , implies IR(t + s, t)l < (c/2)'12
for all t . Hence ( 36)
J R , , , ( ~+ s , t ) J ~ ( t + +
11 +) R ( t + s , t ) I
s ~ , t T
where for fixed t ; , t 2 the functions f and g may both be
+ I R ( t + s + 7 , f ) R ( t+ s , t + .)I taken to be nonnegative and bounded by 1 for t E [0, TI. It
follows that
-
<€ (30)
for all t if Is1 > A , + 171. 0
The following lemma is required in the proof of Proposi-
tion 3.
Lemma I : If R ( t + 7,t ) is the covariance of a per-
I 2 [ i T B 2 (t + au + b , u + c) dt
+
iodically correlated process and satisfies joOmjzR2(t
7,t ) dtd7 = K < 00, then for t , > I , ,

J,,(u, t,, t 2 ) = J t 2 B ( t + au
11
+ b, u + c ) B ( t+ du+ e , u + f )
[
. p y t + du + e , U + f ) dt

I 2 [ i7MdtI1l2[p d t ] 1 / 2 = 2MT (37)

r2
.exp [ - i 2 m n t / T ] dt
m
= ( t 2- tl> C up + t,, t z ) (31) where M is the bound on the fourth moment of X ( t ) . To
show that €(U,t,, t 2 ) E L,( - 00, CO), we again use (37) to
p=-m

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 18,2024 at 06:55:52 UTC from IEEE Xplore. Restrictions apply.
 HURD: NONPARAMETRIC TIME SERIES ANALYSIS 355

obtain Proof: The proof yill be gven for r, 2 7,20 only.

From the definition of B k ( A ,7 )
(38)

(39)
Lp=-m J
exists because B ( . , u ) E L,[O,T] for every U from the
from which it may be seen that the last term in (43) adds
general assumptions. From the hypothesis
to zero with a similar term in the fourth-moment expan-

’LA-“LA
sion for Gaussian variates. Therefore, (43) reduces to
/-WmiTB2(t,7 )d t d ~
=K <CO,

it follows that A2
- T2B(t + 7,,
S -1 + 71 - T2)B( 1, S - t )

J-m

<
- [ (O0 S 2 (U + c ) du]l’z[ (“ S 2 (U + f ) d ~ ] ~ ”

This completes the proof for fixed t , , t,. 0

where the meaning of Il and I , is clear and we have used
We observe that if t , and t , in (31) are Bore1 measur-
B ( t , 7 ) = R ( t + 7,t ) in (44).We will show that the integral
able functions of U , then the result is still true. This
I , evaluates to the term in (41) involving F(u,T ~T, ~ ) the ;
result is used in Proposition 3 to show that if X ( t ) is
integral I , evaluates to the term involving G ( U, T,, T ~ ) ,but
Gaus$an, the rate of convergence of the covariance
~

we will omit this part of the proof. The evaluation of I , is

cov[B,(A, T,), B,(A, T,)] may be determined. This permits
facilitated by partitioning the region of integration E =
us, in the next section, to determine the rate of conver-
gence of the smoothed estimators for the density functions [O,A - ‘I1 [O, A - ‘21 into three sets
&(A).
Proposition 3: If X ( t ) is a Gaussian periodically corre-
A , = { ( s , t ) E E :t > s }
lated process for which A, = { (s,t ) E E : s 2 t and s I t + 7, - T , }
J_P I T I B ( t , 7 ) I d t d T < 0 0 ,
000
A,= {(s,t) EE:s 2 t + ~ , - ~ , } .

then for any 0 I71, T~ IA, We denote by .I1, J,, and J3 the three corresponding
integrals which sum to I,.
Acov [ i j ( A , d,i k ( A , d] To evaluate J,, the transformation U = s - t , U = s re-
sults in
= /-AA U, 1
J1= - J O
A2 u=-A+T,
/” - T2 + ‘B( u - u
o=o
+ T,, u + 7,- 7,)
* [ F ( u , 71972) + G ( U , 7 1 , 7 2 ) ] dU +0
. B ( U - U , u)exp [ - i277(ju - ku + k u ) / T ] dudu. (45)
where
Now using Lemma 1, the inner integral becomes
c
W

F ( u ,71972) = Bp(U + 71 + d B j - , - p ( U )
p=-CC ( A - 7, + u ) F ( u ,T,, r2) +c(u,O, A - 7, + U), (46)
eexp [ - i27r(ju +p u ) / ~ ] (42a) which yields

f(U,O, A - TI+ U)
du (47)
A

and where the function U’(U, T,, 1 for U E [ - A ,

7,) I A] We identify the expression (1- (7,- u ) / A ) in (47) with
and increases to 1 as A + 00. UA(u,T,, 7,) for U in the interval [ - A T~,O]. +

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 18,2024 at 06:55:52 UTC from IEEE Xplore. Restrictions apply.
 356 IEEE TRANSACTIONS ON INFORMATIONTHEORY, VOL. 35, NO. 2, MARCH 1989

The same transformation used to evaluate 5, yields results yield a procedure of spectral analysis for periodi-
1 A - T ~ + u
cally correlated processes. We now show that the shfted
J =-
T - 7
B ( u - U + r 2 ,U 7,- r 2 ) + periodogram is an unbiased but inconsistent estimator for
A2Lf;o’Lu gk(A).
. B ( U - U , u)exp [ - i 2 a ( j u - ku k u ) / ~ dudu
] + Proposition 4: If X ( t ) is periodically correlated, the
shifted period2gram g,(A, A) defined by (6) is the Fourier
(48)
transform of B , ( A , r ) ; if X ( t ) is Gaussian and B ( t , r ) E
J3 = -
A’
1
/“ -
u = T ~ - ~/ UA
,=
B( U - U + r,, U + 7,- r 2 ) (L,[O,TI x R ) , then gk(A , A ) is an asymptotically unbi-
ased but inconsistent estimator of gk(A) and
.B( U - U, u)exp [ - i 2 m ( j u - ku + k u ) / T ] dudu. lim var[g,(A, A)]
“+CC

i
(49)
g,(A)g,(A -2mk/T), A + mn/T
Using Lemma 1, these reduce to
= g,(A)go(A -2%‘k/T) (53)
AJ, = J’,” - ‘( 1 );
- F( U ,r,, 7,) du + Ig,-,(%‘n/T) 12, A = mn/T.

Proof: To show the Fourier transform relationship, we

+f - “c( U , U , A - r2 + U ) du (50) write

1
=nAs,s, X ( t ) X ( s) exp [ - iA( t - s)]
A A
We identify the expression (1- ( r 2 / A ) )in (50) and (1- ( 7,
+ u ) / A ) in (51) with UA(u,r l , r 2 ) for U in the intervals
over which the integrals (40) and (41) are performed. We .exp [ - i2mks/T] dtds. (54)
define U, = 0 for U in the complement of [ - A r2,A - r,]. +
Finally, we can write Breaking the integral over [0, A ] X [0, A ] into the sum of
two integrals, I , for U = t - s < 0 and I, for U 2 0, gives
A ( J , + J~ + J , ) = /”
- A
U,(., 71, 7 2 ) ~ ( U , 7 1 , 72) dU
K2
+-
A 1
(52)
I
211f-A
B, ( A , U ) exp ( - i~ U ) du
=-

where K , can be made bounded and independent of A ; 1

that is, I =- B,(~,u)exp(-i~u)du,

which establishes the result

I j A 1 c ( U ,U ,A
“ - 1
r1) du = K 2 To show that g,(A, A ) is asymptotically unbiased, we
72 - 71 write
where the last inequality results from (33) of Lemma 1. 0 1

OF gk(A )
IV. ESTIMATION
E { gk( A A ) }
9 = 2, /-A { j k ( A , ‘I}exp ( - ZA7) ”

In the stationary situation the periodogram is known to

be an unbiased but inconsistent estimator for the spectral
density function [14], [15]. However, the procedure of
smoothing the periodogram (or smoothing the covariance (55)
estimator) can produce consistent estimators; a similar
result is shown here to be true for periodically correlated As A + 00, the first term converges to gk(A) by the
processes. We show that a consistent estimator for any Lebesgue convergence theorem and the second term con-
member of the family { gk(A), k E Z } may be formed by verges to 0 from Proposition 1. This shows gk(A,A) is
smoothing the shifted periodogram (6), or in other words, asymptotically unbiased.
by smoothing the two-dimensional periodogram along a To show g,(A, A ) is an inconsistent estimator, we must
line S , of constant difference frequency. In this sense, our show var[g,(A, A)] does not converge to 0 as A + 00.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 18,2024 at 06:55:52 UTC from IEEE Xplore. Restrictions apply.
 HURD: NONPARAMETRIC TIME SERIES ANALYSIS 351

From the definition of g,(A, A), The same reasoning leads to

-exp[- i ( h s + At -2mkt/T
+ X U - XU + 2 ~ k v / T ) drdtdudu
]

- [ qp(s,
27rA o o
t)
.exp(innr/T)+ ~

A (60)

for h = m n / T and as above, c 3 ( A , r ) and c 4 ( A , r ) are

integrable functions of r over ( - CO, CO).In t h s case,
.exp [ - iXs + iXt - i 2 m k t / T ] drdt
lim I2 = gn-,( r n / T ) g , - , ( - m n / T )
A+a,
One of the three terms in the fourth moment expansion for
jointly Gaussian random variables cancels the negative =Ig,-hn/T) l2 (61)
term, giving where the last equality follows from the definition of
B k ( r ) for real periodically correlated processes. If X #
var [ g,(A, A ) ] = 11 + 12 m n / T for integer n , then lim,,,I,=O from the
Riemann-Lebesgue lemma. Combining these results pro-
= ~
1
(274,
/”/”/”J A R( s ,
0 0 0 0
u )R ( t ,U ) duces (53) and the general inconsistency of the shifted
periodogram estimator. Note that gk(X) = 0 for k # 0 if
X ( t ) is stationary, so (53) gives the known result [14] for
.exp[ - i h ( s - U +u -t ) the stationary case
- i2mk(t - u ) / T ] dsdtdudu
8,2(V, AZO
lim v a r [ g , ( ~ , ~ )=]
1 A-m (2g,2(X), X = 0.
+-JA/“/”/”R(s,u)R(t, U)
(274, 0 0 0 0
In the remainder of t h s section we show that consistent
estimators of the densities g,(A) may be formed by
.exp [ - i X ( s + u - t - u )
- i2mk(t - u ) / T ] dsdtdudu. &(A, A) = 2?r-/”
1
- A k ( B , u ) j , ( A , u)exp(- i X u ) du

The approach used in (54) and (55) to show that

E { g , ( A , A)} converges to g,(h) may be applied to evalu-
ate Zl and I,. We find that Il may be written as
where BA is a positive function of A such that BA+ 0 and
I l = - / :1
AIBo(r)(l-;) AB, + CO as A -+ CO. The function k ( . ) is called a covari-
(2d2 ance averaging kernel and is taken to be even, bounded, in
L,( - CO, CO) and k(0) = 1; thus k ( .) is also in L2(- CO, CO).
-exp(-iXr)+- The function K ( - ) is the L, Fourier transform of k( .).
A It is possible to give a precise statement concerning the
rate at which the bias of & ( A , A) tends to zero as A + CO;
such a statement appears in [24] and is similar to the
statement made in [14] for the stationary case. We proceed,
however, to establish the consistency of the smoothed
.exp (iXr - i 2 n k r / T ) + ~

A (58) estimators and use this result to produce a bound on the

rate at which the bias must tend to zero. The follow-
where € , ( A ,7 ) and € , ( A ,r ) are integrable functions of r
ing proposition shows that for any j , k and A,, A,,
over (- 00,oo). It follows that cov[g,(A, XI), & ( A , A,)] is O([AB,]-’) where BA is de-
fined above; by setting j = k and X,=X,, we see that
lim I , = go( A) go ( X - 2 a k / T ) . var[$,(A, A)] -,0 as A -+CO so the consistency is estab-
(59)
A-cc lished.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 18,2024 at 06:55:52 UTC from IEEE Xplore. Restrictions apply.
 358 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 35, NO. 2, MARCH 1989

Proposition 5: If X ( t ) is a Gaussian periodically corre- Next, to determine a bound for the term containing
lated process for which F( U ,U,, U,), we define

~exp(-iA,u,+iA,u,) du,du,. (69)

Using the Schwarz inequality for sums in the expression
(42a) for F(u,U,, U,) we obtain

. S ( U + U, - u 2 ) S (U ) d ~ dd ~~ ,, . (70)
where The transformation U, = U, - U,, U, = BAu2produces
- W

S ).1 =/'B2(t, U) dt = 1 IBk(U ).,1

0 - W

Proof: From the definition (62) of & ( A , A ) and (41)

we may write
* I k( BAUl+ U,) IJ" dus( U + u J s ( U)
-A

[ si(A , ik( A,

* s ( u +u , ) s ( u ) du,du,du. (71)
For any fixed BA, the Schwarz inequality for L2(- 00, co)
produces

The hypothesis that S( .) E L,( - co,co), along with

Fubini's theorem establishes
where UA(u,U,, U,) 11 for U, U,, u2 in [ - A , A ] and
lim, ~ U,( U, U,, u 2 ) = 1. We first evaluate the integral J-Wmd%J-;S(u + u , ) s ( u ) du
containing the term O ( A - ' ) ; this integral is bounded by
m
= / _ _ d u S ( u)JW s ( u + U,) du, < 00, (73)
-w

so that we finally obtain

and under the transformation u1 = BAul, U, = BAu2, the

bound becomes
s ( U + U,) s ( U ) du < 00 . (74)

A similar exercise for the term containing G ( u , U,, u2)

produces the same bound for ZG.We have thus established
Since k(.) E L,( - 00, co), this term is O([B,A]-') and that cov[g,(A, A,), & ( A , A,)] is O([ABA]-') and so in
converges to 0 as A -,00. _ Y n ~ , ._ ,.-
Darticular v a r l k ( A , A)1 is 0([ABA1-').
I _
Therefore, under
,

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 18,2024 at 06:55:52 UTC from IEEE Xplore. Restrictions apply.
 HURD: NONPARAMETRIC TIME SERIES ANALYSIS 359

the stated conditions the smoothed shifted periodogram [ l l ] H. L. Hurd, “Representation of harmonizable periodically corre-
lated processes and their covariances,” J . Multivariate Anal., vol.
converges in quadratic mean to gk(A). 0
29, no. 1, p,p. 53-67, 1989.
[12] -, “Staiionarizing properties of random shifts,’’ SIAM J . Appl.
Math., vol. 26, no. 1, pp. 203-211, 1974.
V. ACKNOWLEDGMENT [13] __ , “Periodically correlated processes with discontinuous corre-
lation functions,” Theory Prob. Appl.,
.. vol. 19, no. 4, ..
pp. 834-838,
1974.
The author the constmc- [14] E. Parzen, “On consistent estimates of the spectrum of a stationary
tive comments of the referees and of Professor Stamatis time series,” Ann. Math. Statist., vol. 28, no. 2, pp. 24-43, 1957.
Cambanis of the Center for Stochastic Processes at Chapel [15] M. Rosenblatt, Random Processes. London: Oxford Press, 1962
Hill. [16] w. A. Gardner, “Measurement of spectral correlation,” IEEE
Trans. Acoust., Speech, Signal Processing, vol. ASSP-34, pp.
1111-1123, 1986.
[17] J. Kampt de Ftriet, “Correlation and spectrum of asymptotically
stationary random functions,” Math Studia, vol. 30, pp. 55-67,
REFERENCES 1962
~~

E. Parzen, “Spectral analysis of asymptotically stationary time

E. G. Gladyshev, “Periodically correlated random sequences,” Sou. series,” Bull. Inst. Int. Statist., vol. 39, livrakon 2, pp. 87-103,
Math., vol. 2, pp. 385-388, 1961. 1962.
-, “Periodically and almost periodically correlated random -, “On spectral analysis with missing observations and ampli-
processes with continuous time parameter,” Theory Prob. Appl., tude modulation,” Sankhya A , vol. 25, part 4, pp. 180-189, 1963.
vol. 8, pp. 173-177, 1963. N. Wiener, “Generalized harmonic analysis,” Acta Math., vol. 55,
[31 W. R. Bennett, “Statistics of regenerative digital transmissi,m,” Bell pp. 117-258, 1930.
Syst. Tech. J . , vol. 37, pp. 1501-1542, 1958. L. J. Herbst, “The statistical Fourier analysis of variance,” J . Roy.
[41 W. A. Gardner and L. E. Franks, “Characterization of cyclosta- Statist. Soc. B., vol. 27, pp. 159-165, 1965.
tionary random processes,” IEEE Trans. Inform. Theory, vol. IT-21, W. A. Gardner, Introduction to Random Processes with Application
pp. 4-14, 1975. to Signals and Systems. New York: Macmillan, 1986.
L. I. Gudzenko, “On periodically nonstationary processes,” Ra- __ , Statistical Spectral Analysis: A Nonprobabilistic Theoty.
diotek. Elektron., vol. 4, no. 6, pp. 1062-1064, 1959. Englewood Cliffs, NJ: Prentice-Hall, 1988.
V. A. Markelov, “Axis crossings and relative time of existence of a H. L. Hurd, “An investigation of periodically correlated stochastic
periodically nonstationary random process,” Sou. Radiophys., vol. processes,” Ph.D. dissertation, Duke Univ., Durham, NC, 1969.
9, no. 4, pp. 440-443, 1966. W. A. Gardner, “Stationarizable random processes,” IEEE Trans.
A. Papoulis, Probability, Random Variables and Stochastic Pro- Inform. Theory, vol. IT-24, no. 1, pp. 8-22, 1978.
cesses. New York: McGraw-Hill, 1965. M. Ldve, Probability Theory. New York: Van Nostrand, 1965.
H. Ogura, “Spectral representation of periodic nonstationary ran- H. Niemi, “Stochastic processes as Fourier transformations of
dom processes,” IEEE Trans. Inform. Theory, vol. IT-17, pp. stochastic measures,” Ann. Acad. Sei. Fenn. AI Math., vol. 591,
143-149, 1971. pp. 1-47,1975.
[91 W. M. Brelsford, “Probability predictions and time series with D. K. Chang and M. M. Rao, “Bimeasures and nonstationary
periodic structure,” Ph.D. dissertation, Johns Hopkins Univ., Balti- processes,” in Real and Stochastic Analysis, M. M. Rao, Ed. New
more, MD, 1967. York: Wiley, 1987, ch. 1.
W. M. Brelsford and R. H. Jones, “Time series with periodic N. Wiener, The Fourier Integral and Certain of Its Applications.
structure,” Biometrika, vol. 54, pp. 403-407, 1967. New York: Dover, 1933.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 18,2024 at 06:55:52 UTC from IEEE Xplore. Restrictions apply.