STATISTICAL FOURIER ANALYSIS The Furier Represetati f a Sequece Accrdig t the basic result f Furier aalysis, it is always pssible t apprximate a arbitrary aalytic fucti defied ver a fiite iterval f the real lie, t ay desired degree f accuracy, by a weighted sum f sie ad csie fuctis f harmically icreasig frequecies. The accuracy f apprximati icreases with the umber f fuctis withi the sum. Similar results apply i the case f sequeces, which may be regarded as fuctis mappig frm the set f itegers t the real lie. Fr a sample f T bservatis y,..., y T 1, it is pssible t devise a expressi i the frm (1) y t X j α j cs( j t) + β j si( j t), wherei j πj/t is a multiple f the fudametal frequecy 1 π/t. Thus, the elemets f a fiite sequece ca be expressed exactly i terms f sies ad csies. This expressi is called the Furier decmpsiti f y t ad the set f cefficiets {α j, β j ; j, 1,..., } are called the Furier cefficiets. Whe T is eve, we have T/; ad it fllws that () si( t) si(), cs( t) cs() 1, si( t) si(πt), cs( t) cs(πt) ( 1) t. Therefre, equati (1) becmes 1 X (3) y t α + α j cs( j t) + β j si( j t) + α ( 1) t. j1 Whe T is dd, we have (T 1)/; ad the equati (1) becmes (4) y t α + X j1 α j cs( j t) + β j si( j t). I bth cases, there are T zer cefficiets amgst the set {α j, β j ; j, 1,..., }; ad the mappig frm the sample values t the cefficiets cstitutes a e-t-e ivertible trasfrmati. I equati (3), the frequecies f the trigmetric fuctis rage frm 1 π/t t π; whereas, i equati (4), they rage frm 1 π/t t π(t 1)/T. The frequecy π is the s-called Nyquist frequecy. 1
STATISTICAL FOURIER ANALYSIS Althugh the prcess geeratig the data may ctai cmpets f frequecies higher tha the Nyquist frequecy, these will t be detected whe it is sampled regularly at uit itervals f time. I fact, the effects the prcess f cmpets with frequecies i excess f the Nyquist value will be cfuded with thse whse frequecies fall belw it. T demstrate this, csider the case where the prcess ctais a cmpet which is a pure csie wave f uit amplitude ad zer phase whse frequecy lies i the iterval π < < π. Let π. The (5) cs(t) cs (π )t cs(π) cs( t) + si(π) si( t) cs( t); which idicates that ad are bservatially idistiguishable. Here, < π is described as the alias f > π. The Spectral Represetati f a Statiary Prcess By allwig the value f i the expressi (1) t ted t ifiity, it is pssible t express a sequece f idefiite legth i terms f a sum f sie ad csie fuctis. Hwever, i the limit as, the cefficiets α j, β j ted t vaish. Therefre, a alterative represetati i terms f differetials is called fr. By writig α j da( j ), β j db( j ), where A(), B() are step fuctis with disctiuities at the pits { j ; j,..., }, the expressi (1) ca be redered as (6) y t X cs( j t)da( j ) + si( j t)db( j ). j I the limit, as, the summati is replaced by a itegral t give (7) y(t) cs(t)da() + si(t)db(). Here, cs(t) ad si(t), ad therefre y(t), may be regarded as ifiite sequeces defied ver the etire set f psitive ad egative itegers. Sice A() ad B() are disctiuus fuctis fr which derivatives exist, e must avid usig α()d ad β()d i place f da() ad db(). Mrever, the itegral i equati (7) is a Furier Stieltjes itegral. I rder t derive a statistical thery fr the prcess that geerates y(t), e must make sme assumptis ccerig the fuctis A() ad B(). S far, the sequece y(t) has bee iterpreted as a realisati f a stchastic prcess. If y(t) is regarded as the stchastic prcess itself, the the fuctis A(), B() must, likewise, be regarded as stchastic prcesses defied ver
the iterval [, π]. A sigle realisati f these prcesses w crrespds t a sigle realisati f the prcess y(t). The first assumpti t be made is that the fuctis da() ad db() represet a pair f stchastic prcesses f zer mea which are idexed the ctiuus parameter. Thus (8) E da() E db(). The secd ad third assumptis are that the tw prcesses are mutually ucrrelated ad that -verlappig icremets withi each prcess are ucrrelated. Thus (9) E da()db(λ) fr all, λ, E da()da(λ) if 6 λ, E db()db(λ) if 6 λ. The fial assumpti is that the variace f the icremets is give by (1) V da() V db() df () f()d. We ca see that, ulike A() ad B(), F () is a ctiuus differetiable fucti. The fucti F () ad its derivative f() are the spectral distributi fucti ad the spectral desity fucti, respectively. I rder t express equati (7) i terms f cmplex expetials, we may defie a pair f cjugate cmplex stchastic prcesses: (11) d() 1 da() idb(), d () 1 da() + idb(). Als, we may exted the dmai f the fuctis A(), B() frm [, π] t [, π] by regardig A() as a eve fucti such that A( ) A() ad by regardig B() as a dd fucti such that B( ) B(). The, there is (1) d () d( ). Frm cditis uder (9), it fllws that (13) E d()d (λ) if 6 λ, E{d()d ()} f()d. 3
STATISTICAL FOURIER ANALYSIS These results may be used t reexpress equati (7) as (14) y(t) Ω (e it + e it ) da() i (eit e it ) db() Ω it {da() idb()} it {da() + idb()} e + e Ω e it d() + e it d (). Whe the itegral is exteded ver the rage [, π], this becmes (15) y(t) e it d(). This is cmmly described as the spectral represetati f the prcess y(t). The Autcvariaces ad the Spectral Desity Fucti The sequece f the autcvariaces f the prcess y(t) may be expressed i terms f the spectrum f the prcess. Frm equati (15), it fllws that the autcvariace y t at lag τ t k is give by (16) Ω γ τ C(y t, y k ) E λ e it d() e iλk d( λ) λ e it e iλk E{d()d (λ)} e iτ E{d()d ()} e iτ f()d. Here the fial equalities are derived by usig the results (1) ad (13). This equati idicates that the Furier trasfrm f the spectrum is the autcvariace fucti. The iverse mappig frm the autcvariaces t the spectrum is give by (17) f() 1 π τ 1 γ + π γ τ e iτ τ1 γ τ cs(τ). This fucti is directly cmparable t the peridgram f a data sequece. Hwever, the peridgram has T empirical autcvariaces c,..., c T 1 i place 4
f a idefiite umber f theretical autcvariaces. Als, it differs frm the spectrum by a scalar factr f 4π. I may texts, equati (17) serves as the primary defiiti f the spectrum. T demstrate the relatiship which exists betwee equatis (16) ad (17), we may substitute the latter it the frmer t give (18) γ τ 1 π e iτ 1 π κ γ κ τ γ τ e iτ d e i(τ κ) d. Frm the fact that Ω π, if κ τ; (19) e i(τ κ) d, if κ 6 τ, it ca be see that the RHS f the equati reduces t γ τ. This serves t shw that equatis (16) ad (17) d ideed represet a Furier trasfrm ad its iverse. The essetial iterpretati f the spectral desity fucti is idicated by the equati () γ f()d, which cmes frm settig τ i equati (16). This equati shws hw the variace r pwer f y(t), which is γ, is attributed t the cyclical cmpets f which the prcess is cmpsed. It is easy t see that a flat spectrum crrespds t the autcvariace fucti which characterises a white-ise prcess ε(t). Let f ε f ε () be the flat spectrum. The, frm equati (16), it fllws that (1) γ πf ε, f ε ()d ad, frm equati (16), it als fllws that () γ τ f ε. f ε ()e iτ d e iτ d These are the same as the cditis that serve t defie a white-ise prcess. Whe the variace is deted by σ ε, the expressi fr the spectrum f the white-ise prcess becmes (3) f ε () σ ε π. 5