Detection and measurement of periodicities in biology is often a problem, because most of the biological time series are short, i.e. only a few cycles are available. Therefore the usual spectral approach for an analysis cannot be used with great success and special procedures must be developped (Sollberger, 1970). We have chosen two of these methods for a comparison: the numerical signal averaging of De Prins and Cornelissen (1975) and the pergressive Fourier analysis of Blume (1955, 1965, 1975). Both methods assume the time series model f(t)=s(t)+e(t) (signal plus noise), and they provide estimates of period length, amplitude and phase of the signal. In order to compare both methods we have taken into account the number of necessary numerical operations (computation time) and the quality of the estimates.To test this quality, we have used computer simulated data of the above type as well as circadian rhythms of Euglena gracilis. A comparison of the results for various conditions can be used for a decision which method should be used in a future analysis.
METHODSA. PERGRESSIVE FOURIER ANALYSIS. Blume's pergressive Fourier analysis is closely related to the complex demodulation, a well-known procedure in time series analysis (Tukey, 1961;Granger and Hatanaka, 1964;Bingham et al., 1967;Welch, 1967). Complex demodulation is briefly described by the following steps, (i) Multiply the time series X(t) with a periodic complex exponential X jU (t):=e-i / /t X(t), (ii) use a narrow-band, low-pass filter L to get Zjult) := L (X^t)), (iii) with we get estimates of the 'average' amplitude an.d 'average' phase of the frequency band f(/u) d/i at time t. Assuming that a periodicity of length T:=ju~' is present in a time series i.e., that there is a major peak at frequency /* in the spectrum, then it can be shown (Blume, 1955(Blume, , 1965Bingham et al., 1967) that the phase ç», (¿i) of Z/u(t) is linear in t and can be used to estimate p, the underlying frequency. 409 Downloaded by [North Carolina State University] at 23:06 15 March 2015Performing the pergressive Fourier analysis, we define a segment of the time series including L points and pass it through the total series with stepwidth of Q points. So we get a set of complex demodulates Zfil) with /i=27m/L, n=0...., L/2. Blume has now stated a heuristic principle for the choice of the phase function under the above set belonging to the periodicity in question. Having selected this phase function, we can calculate the length T of the period by T=L/R with R defined by the slope 27TR/L of the phase function.B. NUMERICAL SIGNAL AVERAGING. The term signal averaging is used to the procedure of looking for the Buys-Ballot-filter i.e., a linear nonsymmetric filter with gapped weights p Y (t) := P-1 ,£ X (t -Kj), t = 1 K, which gives an optimal signal. The Buys-Ballot-filter has the properties that for a set of discrete frequencies -A k = 2?ik/K phase shift is zero and in the limit P-co the frequenciesÀ K are isolated (De Prins and Cornelissen, 1975;Koopmans, 1974). Therefore by choic...