Searching for Periodic Signals in Time Series Data 1. Least Squares Sine Fitting 2. Discrete Fourier Transform 3. Lomb-Scargle Periodogram 4. Pre-whitening of Data 5. Other techniques • Phase Dispersion Minimization • String Length • Wavelets Period Analysis How do you know if you have a periodic signal in your data? What is the period? Try 16.3 minutes: 1. Least-squares Sine fitting Fit a sine wave of the form: V(t) = A·sin(t + ) + Constant Where  = 2/P,  = phase shift Best fit minimizes the 2: 2 = di –gi)2/N di = data, gi = fit Advantages of Least Sqaures sine fitting: • Good for finding periods in relatively sparse data Disadvantages of Least Sqaures sine fitting: • Signal may not always be a sine wave (e.g. eccentric orbits) • No assessement of false alarm probability (more later) • Don‘t always trust your results This is fake data of pure random noise with a  = 30 m/s. Lesson: poorly sampled noise  almost always can give you a period, but it it not signficant Any function can be fit as a sum of sine and cosines FT() =  Xj (t) e–it N0 j=1 A DFT gives you as a function of frequency the amplitude (power) of each sine wave that is in the data Power: Px() = | FTX()|2 1 N0 Px() = 1 N0 N0 = number of points [( Xj cos tj + Xj sin tj) ( )] 2 2 Recall eit = cos t + i sint X(t) is the time series 2. The Discrete Fourier Transform The continous form of the Fourier transform: F(s) =  f(x) e–ixs dx f(x) = 1/2  F(s) eixs ds eixs = cos(xs) + i sin (xs) This is only done on paper, in the real world (computers) you always use  a discrete Fourier transform (DFT) The Fourier transform tells you the amplitude of sine (cosine) components to a data (time, pixel, x,y, etc) string Goal: Find what structure (peaks) are real, and what are artifacts of sampling, or due to the presence of noise A pure sine wave is a delta function in Fourier space t P Ao FT  Ao 1/P x  A constant value is a delta function with zero frequency in Fourier space: Always subtract off „dc“ level. 0  f(u)(x–u)du = f *  f(x): (x): Useful concept: Convolution Useful concept: Convolution (x-u) a1 a2 g(x) a3 a2 a3 a1 Convolution is a smoothing function In Fourier space the convolution is just the product of the two transforms: Normal Space Fourier Space f*g F  G Convolution DFT of a pure sine wave: So why isn‘t it a ‐function? width Side lobes It would be if we measured the blue line out to infinity: But we measure the red points. Our sampling degrades the delta function and  introduces sidelobes In time space In Fourier Space = × t 0 × * t 1/P * = 1/P t Window Data 16 min period sampled regularly for 3 hours The longer the data window, the narrower is the width of the sinc function  window:  3 2 N1/2 T A  = error of measurement T = time span of your observations A = amplitude of your signal N = number of data points Error in the period (frequency) of a peak in DFT: A more realistic window And data Alias periods: Undersampled periods appearing as another period Alias periods: P  –1 false P  –1 alias P  –1 true = + Common Alias Periods: P  –1 false (1 day) –1 P  –1 true = + P  –1 false (29.53 d) –1 P  –1 true = + P  –1 false (365.25 d) –1 P  –1 true = + day month year Nyquist Frequency If T is your sampling rate which corresponds to a frequency of fs, then  signals with frequencies up to fs/2 can be unambiguously reconstructed.  This is the Nyquist frequency, N: N < fs/2 e.g. Suppose you observe a variable star once per night. Then the  highest frequency you can determine in your data is 0.5 c/d = 2  days When you do a DFT on a sine wave with a period = 10,  sampling = 1: 0=0.1, 1/t = 1 positive  negative  positive  negative  Nyquist  frequency =0=0 =0 00 0 0 The effects of noise: 2 sine waves ampltudes of 100 and 50 m/s. Noise added at different levels =10 m/s =50 m/s =100 m/s =200 m/s Real Peaks 3. Lomb-Scargle Periodograms Power is a measure of the statistical significance of that frequency (period): 1 2 Px() = [Xj sin tj–] 2 j Xj sin2 tj– [Xj cos tj–] 2 j Xj cos2 tj– j + 1 2 tan(2) = sin 2tj)/cos 2tj)j j Scargle, Astrophysical Journal, 263, 835, 1982 Power: DFT versus Scargle DFT Scargle Amplitude (m/s) Scargle Power DFTs give you the amplitude of a periodic signal in the data. This does not change with more data.  The Lomb‐Scargle power gives you the statistical significance of a period. The more data you have  the more significant the detection is, thus the higher power with more data N=40 N=100 Frequency (c/d) N=15 False Alarm Probability (FAP) The FAP is the probability that random noise will produce a peak with  Lomb‐Scargle Power the same as your observed peak in a certain  frequency range FAP ≈ 1 – (1–e–P)N Unknown period: Where P = Scargle Power N = number of independent  frequencies in the frequency  range of interest Known period: FAP ≈ e–P In this case you have only one  independent frequency Scargle Power (significance) is increased by lower level of noise and/or more data points False Alarm Probability (FAP) FAP ≈ e–P The probability that noise can produce the highest peak over a range ≈ 1 – (1–e–P)N The probability that noise can produce the this peak exactly at this frequency =  e–P Depth [%] 1.2 Duration [h] ? Orbital period [d]  Semi mayor axis[AU] Number of detections 1 Target field No. 8 Host star K...(?) Magnitude(B.E.S.T.) 11.25 Radius[Rsun] 0.65‐0.85  Radius of planet ? [RJup] 0.71‐0.89? DSS1/POSSI Why is the FAP Impotant? Example: A transit candidate from BEST To confirm you need radial  velocity measurements, but  you do not have a period… 16 one hour observations made with the 16 one-hour observations made with the 2m coude echelle Least Square sine fit yields of 2.69 days Published (Dreizler et al. 2003) Radial Velocity Curve of the transiting planet  OGLE 3 Wrong Phase (by 180 degrees) for a transiting planet! Discovery of a rapidly oscillating Ap star with 16.3 min period FAP ≈ 10–5  CrB Period = 11.5 min FAP = 0.015 Period = 16.3 min FAP = 10–5 Lesson: Do not believe any  FAP < 0.01 My limit: < 0.001 Better to miss a real period  than to declare a false one Small FAP does not always mean  a  real signal How do you get the number of independent Frequencies? First Approximation: Use the number of data points N0 Horne & Baliunas (1986, Astrophysical Journal, 302, 757): Ni = –6.362 + 1.193 N0 + 0.00098N0 2 = number of independent frequencies Determining FAP: To use the Scargle formula you need the number of  independent frequencies. Use Scargle FAP only as an estimate. A more valid determination of the FAP  requires Monte Carlo Simulations: Method 1: 1. Create random noise at the same level as your data 2. Sample the random noise in the same manner as your data 3. Calculate Scargle periodogram of noise and determine highest peak in frequency range of interest 4. Repeat 1.000-100.000 times = Ntotal 5. Add the number of noise periodograms with power greater than your data = Nnoise 6. FAP = Nnoise/Ntotal Assumes Gaussian noise. What if your noise is not  Gaussian, or has some unknown characteristics? Use Scargle FAP only as an estimate. A more valid determination of the FAP  requires Monte Carlo Simulations: Method 2: 1. Randomly shuffle the measured values (velocity, light, etc) keeping the times of your observations fixed 2. Calculate Scargle periodogram of random data and determine highest peak in frequency range of interest 3. Reshuffle your data 1.000-100.000 times = Ntotal 4. Add the number of „random“ periodograms with power greater than your data = Nnoise 5. FAP = Nnoise/Ntotal Advantage: Uses the actual noise characteristics of your data FAP comparisons  Scargle formula using N as number of data points Monte Carlo Simulation Using formula and number of data points as your independent frequencies may  overestimate FAP, but each case is different. Number of measurements needed to detect a signal of a certain amplitude. The FAP of  the detection is 0.001. The noise level is  = 5 m/s. Basically, the larger the  measurement error the more measurements you need to detect a signal. Amplitude = level of noise 1086421 Amplitude/error N = 20 Lomb-Scargle Periodogram of 6 data points of a sine wave: Lots of alias periods and false alarm probability (chance that it is due to noise) is 40%! For small number of data points do not use Scargle, sine fitting is best. But be cautious!! Comparison of the 3 Period finding techniques Least squares sine fitting: The best fit period (frequency) has the lowest 2 Discrete Fourier Transform: Gives the power of each frequency that is present in the data. Power is in (m/s)2 or (m/s) for amplitude Lomb-Scargle Periodogram: Gives the power of each frequency that is present in the data. Power is a measure of statistical signficance Amplitude(m/s) 4. Finding Multiple Periods in Data: Pre-whitening What if you have multiple periods in your data? How do you find these and make  sure that these are not due to alias effects of your sampling window. Standard procedure: Pre‐whitening. Sequentially remove periods from the data  until you reach the level of the noise Prewhitening flow diagram: Find highest peak in  DFT/Scargle (Pi) Fit sine wave to data  and subtract fit Calculate new DFT Are more peaks  above the level of  the noise? Save Pi yes no Stop, publish all Pi Noise level Alias Peaks False alarm probability ≈ 10–14 False alarm probability ≈ 0.24 Raw data After removal of dominant period http://www.univie.ac.at/tops/Period04/ Useful program for pre‐whitening of time series data: • Program picks highest peak, but this may be an  alias • Peaks may be due to noise. A FAP analysis will  tell you this Phase Amplitude 5. Other Techniques: Phase Dispersion Minimization Wrong Period Correct Period Choose a period and phase the data. Divide phased data into M bins and compute the  standard deviation in each bin. If 2 is the variance of the time series data and s2 the total  variance of the M bin samples, the correct period has a minimum value of  = s2/2 See Stellingwerf, Astromomical Journal, 224, 953, 1978 PDM is suited to cases in which only a few observations are  made of a limited period of time, especially if the light curves  are highly non‐sinsuisoidal In most cases PDM gives the same answer as DFT, Scargle periodograms. With  enough data it should give the same answer PDM Scargle 5. Other Techniques: String Length Method Phase the data to a test period and minimize the distance between adjacent points Wrong period longer string length Lafler & Kinman, Astrophysical Journal Supplement, 11, 216, 1965 Dworetsky, Monthly Notices Astronomical Society, 203, 917, 1983 5. Other Techniques: Wavelet Analysis This technique is ideal for finding signals that are aperiodic, noisy,  intermittent, or transient. Recent interest has been in transit detection in light curves „You have to be careful that you do not fool  yourself, and unfortunately, you are the easiest  person in the world to fool“ Richard Feynmann