Detecting multiple periodicities in observational data with the multi-frequency periodogram. I. Analytic assessment of the statistical significance

Abstract

We consider the "multi-frequency" periodogram, in which the putative signal is modelled as a sum of two or more sinusoidal harmonics with idependent frequencies. It is useful in the cases when the data may contain several periodic components, especially when their interaction with each other and with the data sampling patterns might produce misleading results. Although the multi-frequency statistic itself was already constructed, e.g. by G. Foster in his CLEANest algorithm, its probabilistic properties (the detection significance levels) are still poorly known and much of what is deemed known is unrigourous. These detection levels are nonetheless important for the data analysis. We argue that to prove the simultaneous existence of all $n$ components revealed in a multi-periodic variation, it is mandatory to apply at least $2^n-1$ significance tests, among which the most involves various multi-frequency statistics, and only $n$ tests are single-frequency ones. The main result of the paper is an analytic estimation of the statistical significance of the frequency tuples that the multi-frequency periodogram can reveal. Using the theory of extreme values of random fields (the generalized Rice method), we find a handy approximation to the relevant false alarm probability. For the double-frequency periodogram this approximation is given by an elementary formula $\frac{\pi}{16} W^2 e^{-z} z^2$, where $W$ stands for a normalized width of the settled frequency range, and $z$ is the observed periodogram maximum. We carried out intensive Monte Carlo simulations to show that the practical quality of this approximation is satisfactory. A similar analytic expression for the general multi-frequency periodogram is also given in the paper, though with a smaller amount of numerical verification.