Clock Statistics: A TutorialDon PercivalApplied Physics LaboratoryUniversity of Washington, SeattleMotivating Example: I• consider following measurements:30-40-11010 1.160 0-10 -1.160 256 512t (days)− top: X =time (phase) difference between clock 55 andtUSNO time scale at day t (adjusted for systematic drift)(1)− bottom: X = X −X ∝ fractional frequency deviatet t−1taveraged over one day1X - X (ns) X (ns)t t-1 t13parts in 10Motivating Example: II• clock statistics used to summarize performance(1)− if X constant, clock 55 agrees with time scale (essentially)t(1)−X has stochastic (noise-like) fluctuationst− statistics used to quantify fluctuations• sample statistics(1)N−11− mean: µˆ = Xtt=0N(here N =512=# of measurements)(1)N−12 1 2− variance: σˆ = (X −µˆ)t=0 tN−σˆ (standard deviation) is measure of spread• easiest to interpret µˆ &ˆ σ if data taken to be independentsamples from Gaussian (i.e., normal) distribution2Motivating Example: III• Q: is Gaussian assumption reasonable?• comparison of histogram to probability density function:0.20.10.0-10 -5 0 5 10x (ns)− Gaussian assumption seems reasonable3PDFsMotivating Example: IV• Q: is independent assumption reasonable?• under Gaussianity, uncorrelatedness implies independence• sample autocorrelation sequence measures uncorrelatedness:(1) (1)N−τ−1(X −µˆ)(X −µˆ)t=0 t t+τρˆ = ,τ =1, 2,...,N− 1τ(1)N−12(X −µˆ)tt=0• can interpret ρˆ as correlation ...
Applied Physics Laboratory University of Washington, Seattle
Motivating Example: I
•consider following measurements: 30
-40
-110 10
0
-10
0
256 t (days)
512
1.16
0
-1.16
−top:Xt=time (phase) difference between clock 55 and USNO time scale at dayt(adjusted for systematic drift) (1) −bottom:Xt=Xt−Xt−1∝fractional frequency deviate averaged over one day
1
Motivating Example: II
•clock statistics used to summarize performance −ifXt(1)constant, clock 55 agrees with time scale (essentially) −Xt(1)has stochastic (noise-like) fluctuations −statistics used to quantify fluctuations •sample statistics −mean:µˆ =N1tN=−10Xt(1) (hereN=512 =# of measurements) −airav:ecnσˆ2=N1tN=−10(Xt(1)−ˆ)2 µ −σ(standard deviation) is measure of spread ˆ • ˆeasiest to interpretµ& ˆσif data taken to be independent samples from Gaussian (i.e., normal) distribution
2
Motivating Example: III
•Q: is Gaussian assumption reasonable?
•comparison of histogram to probability density function:
•Xt(1)well-modeled as uncorrelated Gaussian deviates (sometimes called Gaussian white noise) •theory saysµ&σˆ2are sufficient statistics for summarizing ˆ statistical information about clock 55
•implies ‘random walk’ model for time difference dataXt •seems we need little more than what is taught in ‘Statistics 101’
5
Reality Bites!
•alas, other clocks do not have such simple statistical properties
30
-40
-110 -110
-135
-160 10
0
1
0
-1 1
0
-1 1
0
-10 -1 0 256 512 0 32 t (days)τ(days) •µˆ & ˆσ2not sufficient summaries for clock in middle plot 6
Overview of Remainder of Tutorial
•discussion of models for interpreting clock statistics
−models specified via spectrum (spectral density function) −while white noise & random walk models depend onµ&σ2, more comprehensive models depend onµand spectrum −in simplest case, spectrum itself depends on 2 parameters ∗σ2, a parameter setting overall level of spectrum ∗α, a so-called ‘power law’ parameter
•clock statistics based upon 2 variance decompositionslook at
−spectral analysis −wavelet analysis
7
The Spectrum
•letXtbe a stochastic process, i.e., collection of random vari-ables (RVs) indexed byt •suppose further thatXtis stationary •implies certain theoretical properties do not change with time •in particular, its varianceσ2=var{Xt}is the same for allt •spectrumSX(·) decomposesσ2across frequenciesf: /2 var{Xt}=1SX(f)df −1/2 hereffrequency with units of cycles per unit timeis a Fourier (e.g., cycles per day for process sampled once per day)
8
Physical Interpretation of Spectrum via Filtering
•letaube a filter, and formYt=u∞=−∞auXt−u •Ythas spectrumSY(f) =A(f)SX(f), where A(f) =∞aue−i2πf u2is squared gain function u=−∞ •ifaunarrow-band of bandwidth ∆faboutf, i.e., 1 A(f) =20,∆f, foth−er∆w2fise≤,|f| ≤f+∆2f then have following interpretation forSX(f): 1/2SY(f)df=−11//22A(f)S var{Yt}=X(f)df≈SX(f) −1/2
9
Spectrum for White Noise Process
•simplest stationary process is white noise •tis white noise process if −E{t}=µfor allt(usually takeµ=0), whereE{t}denotes expected value of RVt −var{t}=σ2for allt −tandtare uncorrelated for allt=t •spectrum for white noise is justS(f) =σ2 •note that
1/2−11//22σ2df=σ2=var{t}, S(f)df= −1/2 as required