SPECTRAL MEASURES OF PARMA SEQUENCES By
Transkrypt
SPECTRAL MEASURES OF PARMA SEQUENCES By
doi:10.1111/j.1467-9892.2007.00541.x SPECTRAL MEASURES OF PARMA SEQUENCES By Agnieszka WyŁomańska Wrocław University of Technology First Version received February 2006 Abstract. The aim of this article is to give a complete description of the spectral measure of periodic autoregressive moving-average (PARMA) system in terms of its coefficients. In the analysis we use the spectral theory of strongly harmonizable sequences presented in Hurd [Journal of Multivariate Analysis (1989) Vol. 29, pp. 53–67] and the form of the unique bounded solution of ARMA model with periodic coefficients. As an application of the theoretical results, we present some examples of the spectral measures for PARMA models. Keywords. PARMA models; periodically correlated process; spectral measure; strongly harmonizable sequence. 1. INTRODUCTION Let (x(n)), n 2 Z, where Z is the group of integers, be a sequence in a Hilbert space H ¼ L2(X, F, P), and let Rx(n, m) ¼ (x(n), x(m)) be its correlation function. The dual group of Z is identified here with T ¼ [0, 2p) with addition modulo 2p. Definition 1 (Hurd, 1989). The sequence (x(n)), n 2 Z is called strongly harmonizable if there is a measure F on T2 (called spectral measure of (x(n))) such that Rx ðn; mÞ ¼ Z 2p Z 0 2p eiðnsmtÞ F ðds; dtÞ: 0 Definition 2 (Dehay, 1994). A sequence (x(n)), n 2 Z is called periodically correlated (PC) with period T, if for every n, k 2 Z the following hold mx ðnÞ ¼ EðxðnÞÞ ¼ mðn þ T Þ and Rx ðn þ k; kÞ ¼ Rx ðn þ k þ T ; k þ T Þ: In the sequel we will be assuming that mx(n) 0. Any PC sequence is strongly harmonizable and its spectrum is supported on the sets 0143-9782/08/01 1–13 JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd., 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA 02148, USA. 2 A. WYŁOMAŃSKA Ln ¼ ðs; tÞ 2 T2 : t ¼ s þ 2np ; T where n ¼ 0; 1; . . . ; T 1: The spectral measure F of a PC sequence with period T has the following form (see Gladyshev, 1961; Dehay and Hurd, 1994; Makagon, 2001) F ðds; dtÞ ¼ T 1 X Cn ðds; dtÞ; n¼0 where Cn are supported on the sets Ln, n ¼ 0, 1, 2, . . . , T 1. In this article, we study the spectral measures of certain harmonizable sequences which admit a moving-average representation of the form xðnÞ ¼ 1 X ak ðnÞnnk ; ð1Þ k¼1 where (nn) is an orthonormal basis in Mx ¼ spfx(n) : n 2 Zg. Of course, each sequence can be written in this form in infinite ways, depending on the choice of (nn). However, very often this representation is somewhat unique, for example when x(n) is required to be a bounded solution to ARMA systems with varying coefficients (Makagon et al., 2004; Weron and Wyłomańska, 2004). In the latter case, (nn) is its forward or backward innovation sequence and the representation (1) is one-sided. This article is therefore motivated by the needs to describe the spectra of harmonizable solutions to such systems. We extend the results obtained in Hurd et al. (2002) for AR(1) models with periodic coefficients. Moreover, we use the form of the unique bounded solution of PARMA sequences (Makagon et al., 2004; Weron & Wyłomańska, 2004) to calculate the formulas of the spectral measures for such systems. The article is organized as follows. In Section 2 the conditions of existence of the unique, bounded, periodically correlated solution of general PARMA model are obtained. The complete description of periodically correlated PARMA spectral measure expressed in terms of the coefficients is formulated in Section 3 and the main results are obtained in Theorems 2 and 3. Some examples of the spectral measures of PARMA models are shown in Section 4. In the Appendix we present the proofs of Theorems 1, 2 and 3. 2. Definition 3 (Sakai 1991). linear equations xðnÞ p X k¼1 PARMA(p,q) MODELS A PARMA(p,q) system with period T is a system of bk ðnÞxðn kÞ ¼ q1 X aj ðnÞnnj ; j¼0 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 n 2 Z; ð2Þ SPECTRAL MEASURES OF PARMA SEQUENCES 3 where (bk(n)), k ¼ 1, . . . , p and (aj(n)), j ¼ 0, . . . , q 1, are periodic (with period T) sequences of complex numbers, bp(n) 6¼ 0 for all n 2 Z and (nn) is a sequence of uncorrelated complex random variables with mean zero and unit variances. Throughout the article, the L2-closed linear space generated by a sequence of random variables nn, n 2 Z, is denoted by Mn. The space Mn with norm kÆk is a Hilbert space. Any sequence (x(n)) in Mn that satisfies eqn (2) is called a solution. The solution (x(n)) is bounded if supn2Zkx(n)k2 < 1. The special case of system (2) was considered in Makagon et al. (2004) where a complete description of the unique bounded PC solutions to PARMA(1, q) models is obtained. Let us denote 2 b1 ðnÞ 6 1 6 0 Bn ¼ 6 6 . 4 .. 0 b2 ðnÞ 0 1 .. . ... ... ... .. . 0 ... 3 bp1 ðnÞ bp ðnÞ 0 0 7 7 0 0 7; .. 7 .. . 5 . 1 0 Bnk ¼ Bn Bn1 Bk ; n k: Let Yn ¼ ½Yn ; 0; . . . ; 00 and xðnÞ ¼ ½ xðnÞ; xðn 1Þ; . . . ; xðn p þ 1Þ0 ; where Yn ¼ q1 X ap ðnÞnnp p¼0 and [. . .]0 denotes the column vector. Then (see Boshnakov (1997)) xðnÞ ¼ Bn xðn 1Þ þ Yn : Consequently xðnÞ ¼ Bnnkþ1 xðn kÞ þ k1 X Bnnþ1s Yns ; k > 0; s¼0 where Bnj ¼ I if j>n. Since all matrices Bnj are invertible 1 xðnÞ ¼ ðBnþk nþ1 Þ xðn þ kÞ k X 1 ðBnþs nþ1 Þ Ynþs ; k > 0: s¼1 Therefore we obtain xðnÞ ¼ p X j¼1 Uj ðn k þ 1; nÞxðn k j þ 1Þ þ k1 X U1 ðn s þ 1; nÞYns ; ð3Þ s¼0 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 4 A. WYŁOMAŃSKA where Uj ðl; kÞ ¼ Bkl ð1; jÞ for l < k. On the other hand we have: xðnÞ ¼ p X Wj ðn þ 1; n þ kÞxðn þ k j þ 1Þ k X W1 ðn þ 1; n þ sÞYnþs ; ð4Þ s¼1 j¼1 where Wj ðl; kÞ ¼ ðBkl Þ1 ð1; jÞ for l < k. Theorem 1 shows that the bounded solution of the PARMA system (2) can be written in causal (forward) or backward form under conditions I and II, respectively. The conditions are defined as follows: Condition I. 2 1 q1þs X X W1 ðn þ 1; n þ jÞajs ðn þ jÞ < 1 sup n2Z s¼2qj¼maxð1;sÞ and inf k1 p X jWj ðn þ 1; n þ kÞj2 ¼ 0; n 2 Z: j¼1 Condition II. 2 1 min ð0;q1sÞ X X sup U1 ðn j þ 1; nÞasþj ðn þ jÞ < 1 n2Z s¼0 j¼s and inf k1 p X jUj ðn k þ 1; nÞj2 ¼ 0; n 2 Z: j¼1 Theorem 1. If the sequence (x(n)) is the unique bounded PC solution of system (2) then it has the following form: xðnÞ ¼ 8 1 P > > > < > > > : q1þs P W1 ðn þ 1; n þ jÞajs ðn þ jÞnsþn ; if I holds, s¼2q j¼maxð1;sÞ q1sÞ 1 P minð0;P U1 ðn þ j þ 1; nÞasþj ðn þ jÞnns ; s¼0 j¼s 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 ð5Þ if II holds. 5 SPECTRAL MEASURES OF PARMA SEQUENCES 3. SPECTRA OF PERIODICALLY CORRELATED SOLUTIONS OF PARMA(p,q) MODELS Because of eqn (5) we can suppose MX ¼ Mn. Let us denote V as a forward shift operator V(nn) ¼ nnþ1. The operator V is unitary and the sequence x(n) ¼ VnQn, where 8 q1þs 1 P P > > W1 ðn þ 1; n þ jÞajs ðn þ jÞns if I holds, > < s¼2q j¼maxf1;sg Hn ¼ 1 minf0;q1sg > P P > > U1 ðn þ j þ 1; nÞajþs ðn þ jÞns if II holds. : s¼0 j¼s The operator V is an isometry, therefore the correlation function of x(n) and x(n þ k) is given by: Rx ðn; n þ kÞ ¼ ðHn ; V k Hnþk Þ: When Condition I holds, we obtain 1 X Rx ðn; n þ kÞ ¼ q1þs X q1þsk X s¼2q j¼maxf1;sg l¼maxf1; skg W1 ðn þ 1; n þ jÞajs ðn þ jÞW1 ðn þ k þ 1; n þ k þ lÞalsþk ðn þ k þ lÞ: On the other hand, when condition II is fulfilled, we have Rx ðn; n þ kÞ ¼ minf0;q1kþsg 1 minf0;q1sg X X X s¼0 j¼s l¼kþs U1 ðn þ j þ 1; nÞajþs ðn þ jÞU1 ðn þ l þ k þ 1; n þ kÞalþks ðn þ k þ lÞ: Let D be an unitary mapping D : nn ! einÆ. D maps MX into L2([0, 2p), dt), where dt is the normalized Lebesgue measure on [0, 2p), that is the Lebesgue measure divided by 2p. If we denote fn ¼ D(x(n)), then system (3) takes the following form fn ðÞ ¼ p X bk ðnÞfnk ðÞ þ k¼1 q1 X aj ðnÞeiðnjÞ ; n 2 Z: j¼0 Thus we have 8 q1þs 1 P P > in > W1 ðn þ 1; n þ jÞajs ðn þ jÞeis > < e s¼2q j¼maxf1;sg DðxðnÞÞðÞ ¼ q1sg 1 minf0;P > P > > U1 ðn þ j þ 1; nÞajþs ðn þ jÞeis : ein s¼0 if I holds, if II holds. j¼s If we denote 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 6 A. WYŁOMAŃSKA gn ðkÞ ¼ 8 1 P > > > < > > > : q1þs P W1 ðn þ 1; n þ jÞajs ðn þ jÞeisk s¼2q j¼maxf1;sg 1 minf0;q1sg P P if I holds, ð6Þ U1 ðn þ j þ 1; nÞajþs ðn þ jÞe isk if II holds, j¼s s¼0 we can see that x(n) is unitary equivalent to fn ðkÞ ¼ eink gn ðkÞ in L2 ð½0; 2pÞ; dtÞ: Lemma 1. The correlation function of PC sequence with period T has the following form Z 2p T 1 X ik 2pl T Bn ðkÞ ¼ Rx ðn þ k; kÞ ¼ e eins cl ðdsÞ; 0 l¼0 where 2pl : cl ðdsÞ ¼ Cl ds; ds þ T Proof. Bn ðkÞ ¼ Z 2p 2p e 0 ¼ Z Z 0 2p Z iðnþkÞsikt F ðds; dtÞ ¼ 0 2p 0 2p eikðstÞ eins 0 Z T 1 X Cl ðds; dtÞ ¼ l¼0 Z 2p 0 T 1 X l¼0 eikðstÞ eins F ðds; dtÞ 2pl eik T Z 2p eink cl ðdkÞ: u 0 Corollary 1. The measures (ck) (k ¼ 0,1, . . . , T 1) given in Lemma 1 satisfy the following condition Z 2p T 1 1X 2ikpl eink cl ðdkÞ ¼ e T Bn ðkÞ: ð7Þ T 0 k¼0 The measures (ck), k ¼ 0,1, . . . , T 1 are called spectrum of the PC sequence (x(n)). Theorem 2. Let (bk(n)), k ¼ 1, 2, . . . , p and (aj(n)), j ¼ 0, 1, . . . , q 1 be periodic (in n) sequences with period T and condition I or II holds. Let (ck), k ¼ 0, 1, . . . , T 1 be spectrum of the unique solution of PARMA(p,q) system defined in eqn (2). Then the measures (ck) are absolutely continuous with respect to the normalized Lebesgue measure dk on [0, 2p) and 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 SPECTRAL MEASURES OF PARMA SEQUENCES 7 T 1 dck ðkÞ X 2pl 2pl ^ ^glk k þ ¼ ; gl k þ dk T T l¼0 where ^ gj ðkÞ ¼ T 1 2pinj 1X gn ðkÞe T T n¼0 and gn(k) is given in eqn (6). Now we consider a PARMA(1,q) system given by: xðnÞ bn xðn 1Þ ¼ q1 X ap ðnÞnnp ; n 2 Z; ð8Þ p¼0 where the coefficients and innovations have the same properties like in the general model. If the sequence (x(n)) is the unique bounded PC solution of system (3) then it has the following form (Theorem 2.2 in Makagon et al., 2004): 8 q1þs 1 P P > 1 > ðBnþj > nþ1 Þ ajs ðn þ jÞnnþs ; if jP j > 1 < s¼2q j¼maxf1;sg xðnÞ ¼ 1 minf0;q1sg > P P > > ðBnnþjþ1 Þajþs ðn þ jÞnns ; if jP j < 1, : s¼0 j¼s where P ¼ b1b2 bT and Blk ¼ bk bkþ1 bl (with the convention Blk ¼ 1 for k > l). PARMA(1,q) models satisfy Theorem 3: Theorem 3. Let fn(k) be unitary equivalent to the unique bounded solution (x(n)) of PARMA(1,q) sequence through the mapping D [that means fn(k) ¼ D(x(n))(k)], then the following formula holds: fn ðkÞ ¼ eink ½1 P eiT k 1 Gn ðkÞ þ eink Rn ðkÞ; where Gn ðkÞ ¼ and Rn ðkÞ ¼ 8 T þq2 q1 P P n > > Bnkþjþ1 aj ðn þ j kÞeikk > < k¼q1 j¼0 1 q1 P n > TP > > Bnkþjþ1 aj ðn : k¼0 j¼0 if jP j < 1, þ j kÞeikk if jP j > 1, 8 q2 s PP > > Bnnsþjþ1 aj ðn þ j sÞeiks > < if jP j < 1, > > : if jP j > 1. s¼0 j¼0 P q1þs P > q2 s¼0 j¼1 1 iks ðBnþj nþ1 Þ ajþs ðn þ jÞe 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 8 A. WYŁOMAŃSKA 4. Example 1. EXAMPLES Let us consider AR(1) system given by the equation: xðnÞ bxðn 1Þ ¼ nn ; for |b| < 1. It is a PAR(1) system with period T ¼ 1. Using Corollary 1 and formula (7) we obtain the following: Z 2p Bn ð0Þ ¼ eink c0 ðdkÞ: 0 Because the measure c0 is absolutely continuous with respect to the normalized Lebesque measure on [0, 2p), then from Theorem 2 we have: Z 2p Z 2p Bn ð0Þ ¼ eink ^ eink g0 ðkÞg0 ðkÞdk: g0 ðkÞ^ g0 ðkÞdk ¼ 0 0 Using Theorem 3 we can see that R0 ðkÞ 0; f0 ðkÞ ¼ ð1 beik Þ1 : G0 ðkÞ 1; From the relation between f0(k) and g0(k) we conclude that g0(k) ¼ (1 beik)1. Thus we have: Z 2p Bn ð0Þ ¼ eink j1 beik j2 dk: 0 If we take n ¼ 1 we can write Z 2p Z ik ik 2 e j1 be j dk ¼ 2 B1 ð0Þ ¼ 0 p 0 cosðkÞ pð1 þ b2 Þ p dk ¼ : 1 þ b2 2b cosðkÞ bj1 b2 j b If an estimate of B1(0) is available, then the above equation can be solved to get an estimate of the parameter b. In this case, the spectral measure c0 ðkÞ ¼ 1 j1 beik j2 : . Example 2. Let us consider PAR(1) system with period T ¼ 2 given by: xðnÞ bn xðn 1Þ ¼ nn : Let suppose |P| ¼ |b0b1| < 1. Using results presented in Theorem 3 we obtain for n ¼ 0,1: Rn ðkÞ 0; Gn ðkÞ ¼ 1 þ bn eik ; fn ðkÞ ¼ eink ½1 P e2ik 1 Gn ðkÞ; which gives gn(k) ¼ (1þbneik)/(1Pe2ik). 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 SPECTRAL MEASURES OF PARMA SEQUENCES 9 To obtain the spectral measures of the sequence (x(n)) we use the relation between c0 and c1 (defined in Lemma 1) and the normalized Lebesgue measure on [0, 2p) presented in Theorem 2: c0 ðkÞ ¼ c1 ðkÞ ¼ ¼ jb0 b1 j2 þ j2 þ b0 eik þ b1 eik j2 4j1 b0 b1 e2ik j2 ¼ 2 þ 2ðb0 þ b1 Þ cosðkÞ þ b20 þ b21 ; 2 4b0 b1 cosð2kÞ þ 2b20 b21 eik ðb1 b0 Þð2 þ b0 eik þ b1 eik Þ þ eik ðb1 b0 Þð2 b0 eik b1 eik Þ 4j1 b0 b1 e2ik j2 cosðkÞðb0 b1 Þ : 1 2b0 b1 cosð2kÞ þ b20 b21 To obtain the coefficients b0 and b1 we use formula (7) from Corollary 1 and substitute theoretical correlations with estimated ones. For the sample of length NT from PARMA models with period T the estimators of Bn(k) have the following form T 1 X ^ n ðkÞ ¼ 1 B xðjT þ k þ nÞxðjT þ kÞ: N j¼0 We then obtain 2pðb0 þ b1 Þb0 b1 1 ðb0 b1 Þ 2 ^ 1 ð0Þ þ B ^ 1 ð1Þ; ¼B 2pðb0 b1 Þb0 b1 1 ðb0 b1 Þ2 ^ 1 ð0Þ B ^ 1 ð1Þ: ¼B We conclude that b0 ¼ . ^ 1 ð0Þ B b ^ 1 ð1Þ 1 B The coefficient b1 can be obtained from the following equation: 0 ! ^ B1 ð0Þ 2p þ 1 b31 ¼ @1 ^ 1 ð1Þ B ^ 1 ð0Þ B ^ 1 ð1Þ B !2 1 ^ 1 ð0Þ þ B ^ 1 ð1ÞÞ: b4 A ð B 1 APPENDIX: PROOFS Proof of Theorem 1. Suppose that I holds and (x(n)) is bounded solution of eqn (2) then for every n 2 Z there is a sequence kr such that lim r p X jWj ðn þ 1; n þ kr Þj2 ¼ 0: j¼1 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 10 A. WYŁOMAŃSKA Hence, from eqn (4), we conclude that for every n xðnÞ ¼ lim r kr X W1 ðn þ 1; n þ sÞYnþs : s¼1 From formula (9) in Lemma 2.1 in Makagon et al. (2004) we have that kr X W1 ðn þ 1; n þ jÞðYnþj ; nk Þ ¼ j¼1 minðkr ;q1þknÞ X W1 ðn þ 1; n þ jÞanþjk ðn þ jÞ: j¼maxð1;knÞ Letting r ! 1, we obtain that ðxðnÞ; nk Þ ¼ lim r kr X W1 ðn þ 1; n þ jÞðYnþj ; nk Þ j¼1 q1þkn X ¼ W1 ðn þ 1; n þ jÞanþjk ðn þ jÞ; j¼maxð1; knÞ if k > 1 q þ n, and (x(n),nk) ¼ 0 if k 1 q þ n. Therefore from formula (6) in Lemma 2.1 in Makagon et al. (2004) we obtain xðnÞ ¼ 1 X 2 4 q1þkn X 3 W1 ðn þ 1; n þ jÞanþjk ðn þ jÞ5nk ; k¼2þnq j¼maxð1;knÞ which after substituting s ¼ k n gives the thesis. This also shows that the solution is unique. The variance of x(n) is given by 2 1 q1þs X X : W ðn þ 1; n þ jÞa ðn þ jÞ kxðnÞk2 ¼ 1 js s¼2qj¼maxð1;sÞ Suppose now Condition II holds. If inf r p X jUj ðn r þ 1; nÞj2 ¼ 0; j¼1 then for every n there is a sequence kr ! 1, such that lim r p X jUj ðn kr þ 1; nÞj2 ¼ 0: j¼1 If the sequence (x(n)) is a bounded solution of eqn (2), then from eqn (3) it follows that xðnÞ ¼ lim r 0 X U1 ðn þ j þ 1; nÞYnþj : j¼1kr Using formula (6) Lemma 2.1 in Makagon et al. (2004) we conclude that 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 SPECTRAL MEASURES OF PARMA SEQUENCES xðnÞ ¼ " 1 minð0;q1sÞ X X 11 # U1 ðn þ j þ 1; n þ 1Þasþj ðn þ jÞ nns : j¼s s¼0 Moreover 2 1 minð0;q1sÞ X X U1 ðn j þ 1; nÞasþj ðn þ jÞ : kxðnÞk ¼ j¼s s¼0 2 u Proof of Theorem 2. T 1 T 1 T 1 T 1 X X 1X 1X ^l ðkÞe2iplðnþpÞ=T ^j ðkÞe2ipjn=T g g ei2pnk=T gnþp ðkÞgn ðkÞ ¼ ei2pnk=T T n¼0 T n¼0 l¼0 j¼0 ¼ ¼ T 1 X T 1 T 1 X 1X e2pilp=T ^ e2ipn=T ðkþjlÞ gl ðkÞ^ gj ðkÞ T j¼0 l¼0 n¼0 T 1 X e2pilp=T ^gl ðkÞ l¼0 ¼ T 1 X e2pilp=T l¼0 ¼ T 1 X T 1 X T 1 1X ^ gj ðkÞe2ipn=T ðjðlkÞÞ T j¼0 n¼0 T 1 1X gl ðkÞ gn ðkÞe2ipn=T ðlkÞ ^ T n¼0 e2pilp=T ^glk ðkÞ^ gl ðkÞ: l¼0 If we use the equivalence between x(n) and fn(Æ) ¼ einÆgn(Æ) we obtain Z 2p eipk ck ðdkÞ ¼ 0 ¼ ¼ T 1 1X e2ipnk=T Bp ðnÞ T n¼0 T 1 1X e2ipnk=T T n¼0 Z 2p eipk 0 ¼ ¼ Proof of Theorem 3. eiðnþpÞkink gnþp ðkÞgn ðkÞdk 0 e2pilp=T ^ gl ðkÞdk glk ðkÞ^ 2p eipðk2pl=T Þ ^ gl ðkÞ^ glk ðkÞdk 0 T 1 Z X l¼0 2p l¼0 T 1 Z X l¼0 T 1 X Z 2p eipk ^gl ðk þ 2pl=T Þ^ glk ðk þ 2pl=T Þdk: u 0 If |P| < 1, then we have 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 12 A. WYŁOMAŃSKA gn ðkÞ ¼ 1 minf0;q1sg X X s¼0 ¼ Bnnþjþ1 ajþs ðn þ jÞeiks j¼s q2 X s X Bnnþjsþ1 aj ðn þ j sÞeiks þ q2 X s X Bnnþjsþ1 aj ðn þ j sÞeiks þ s¼0 j¼0 ¼ q2 X s X Bnnþjþ1 ajþs ðn þ jÞeiks s¼q1 j¼s s¼0 j¼0 ¼ 1 q1s X X q1 1 X X Bnnþjsþ1 aj ðn þ j sÞeiks s¼q1 j¼0 Bnnþjsþ1 aj ðn þ j sÞeiks s¼0 j¼0 þ þq2 X q1 1 TX X BnnþjNT kþ1 aj ðn þ j NT kÞeikðNT þkÞ N ¼0 k¼q1 j¼0 ¼ ½1 P eik 1 Gn ðkÞ þ q2 X s X Bnnþjsþ1 aj ðn þ j sÞei ks: s¼0 j¼0 For |P| > 1 we obtain gn ðkÞ ¼ 1 X q1þs X nþj 1 ðBnþ1 Þ ajs ðn þ jÞeiks s¼2q j¼maxf1;sg ¼ q2 q1s X X s¼0 ¼ nþj 1 ðBnþ1 Þ ajþs ðn þ jÞeiks j¼1 1 q1þs X X s¼1 nþj 1 ðBnþ1 Þ ajs ðn þ jÞeiks j¼s q1 1 X T X X nþjþNT þk 1 ðBnþ1 Þ aj ðn þ j þ NT þ kÞeikðNT þkÞ N ¼0 k¼1 j¼0 q2 q1s X nþj 1 X ðBnþ1 Þ ajþs ðn þ jÞeiks s¼0 ¼ 1 X j¼1 P N eikTN N ¼0 q1 T X X 1 ikk ðBnþjþk nþ1 Þ aj ðn þ j þ kÞe k¼1 j¼0 q2 q1s X nþj 1 X ðBnþ1 Þ ajþs ðn þ jÞeiks s¼0 j¼1 ¼ ½1 P eikT 1 q1 T X X ðBnþjþkT Þ1 aj ðn þ j þ kÞeikðkT Þ nþ1 k¼1 j¼0 q2 q1s X nþj 1 X ðBnþ1 Þ ajþs ðn þ jÞeiks s¼0 j¼1 ¼ ½1 P eikT 1 q1 T 1 X X 1 ikk ðBnþjþk nþ1 Þ aj ðn þ j kÞe k¼0 j¼0 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1 SPECTRAL MEASURES OF PARMA SEQUENCES q2 q1s X X s¼0 13 nþj 1 ðBnþ1 Þ ajþs ðn þ jÞeiks j¼1 ¼ ½1 P eikT 1 Gn ðkÞ q2 q1s X X s¼0 1 iks ðBnþj : nþ1 Þ ajþs ðn þ jÞe j¼1 ink Because fn(k) ¼ e gn(k), then we obtain the thesis. u NOTE Corresponding author: Agnieszka Wyłomańska, Institute of Mathematics and Computer Science, Wrocław University of Technology, Janiszewskiego 14a, 50370 Wrocław, Poland. E-mail: [email protected] REFERENCES Boshnakov, G. (1997) Periodically correlated solutions to a class of stochastic difference equations. In Stochastic Differential and Difference Equations (eds I. Csiszár and G. Michaletzky). Boston: Birkhauser, 1–9. Dehay, D. (1994) Spectral analysis of the covariance kernel of almost periodically correlated processes. Stochastic Processes Applications 50, 315–30. Dehay, D. and Hurd, H. (1994) Representation and estimation for PC and almost PC random processes, In Cyclostationarity in Communications and Signal Processing (ed. W. Gardner). New York: IEEE Press, 295–329. Gladyshev, E. (1961) Periodically correllated random sequences. Soviet Mathematics Doklady 2, 385–88. Hurd, H. (1989) Representation of strongly harmonizable periodically correlated processes and their covariance. Journal of Multivariate Analysis 29, 53–67. Hurd, H., Makagon, A. and Miamee, A. (2002) On AR(1) models with periodic and almost periodic coefficients. Stochastic Proceedings and Applications 100, 167–85. Makagon, A. (2001) Characterization of the spectra of a periodically correlated processes. Journal of Multivariate Analysis 78, 1–10. Makagon, A., Weron A. and Wyłomańska, A. (2004) Bounded solutions of ARMA models with varying coefficients. Applied Mathematics 31, 273–85. Sakai, H. (1991) On the spectral density matrix of a periodic ARMA process. Journal of Time Series Analysis 12, 72–82. Weron, A. and Wyłomańska, A. (2004) On ARMA(1,q) models with bounded and periodically correlated solutions. Probability Mathematics and Statistics 24, 165–72. 2007 The Author Journal compilation 2007 Blackwell Publishing Ltd. JOURNAL OF TIME SERIES ANALYSIS Vol. 29, No. 1