Uncertainty of Stability Variances Based on Finite Differences

Transcription

1 Uncertainty of Stability Variances Based on Finite Differences C. A. Greenhall* Jet Propulsion Laboratory California Institute of Technology W. J. Riley S ymme t r icom, Inc. Draft 009 Abstract This work gives an algorithm for computing the degrees of freedom of estimators of Allan and Hadamard variances, including their modified versions. A consistent approach is used throughout. Introduction This work gives an algorithm by which one can compute error bars for measurements of frequency stability variances in the presence of powerlaw phase noises. These stability variances fall into two categories: unmodified variances, which use dth differences of phase samples for d = (Allan) or (Hadamard), and modified variances, which use dth differences of averaged phase samples. The corresponding sampling functions that act on phase and frequency are shown in [l]. Each variance is defined as a scaling factor times the expected value of the squared differences. Unbiased estimates of this variance are computed from available phase data by taking time averages of the squared differences. The usual choices for the estimation stride (the time step) are the sample period 70 and the averaging period 7, a multiple of 70. These give respectively the overlapped estimator (OE) and nonoverlapped estimator (NOE) of the stability variance (although nonoverlapped is a misnomer; there is always some overlap between summands). The new algorithm, which computes the equivalent degrees of freedom () of a variance estimator, can replace several ones currently in use with a single, complete, and consistent method. Specifically, this algorithm covers the OE and NOE of the unmodified and modified Allan and Hadamard variances for all common noise types (4 5 a 5 ) at all applicable sample sizes and averaging factors. Previously, for the NOE of Allan and Hadamard variances, sigma confidence limits were generally set by scaling the measured deviation by a noisedependent factor K,/a, where M is the number of summands [][6]. For the OE of Allan variance, empirical approximations [7] were generally used along with chisquared statistics; nonempirical methods for both Allan variance estimators were also published [][4]. Although an computation for the OE of modified Allan (and time) variance was first given in [5], the simpler approach of [6] was generally used; an alternative unpublished approach using the Hadamard formed the basis of the new The work of this author was performed at the Jet Propulsion Laboratory, California Institute of Technology ( JPL ), under a contract with the National Aeronautics and Space Administration. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise, does not constitute or imply its endorsement by the United States Government or JPL.

2 algorithm. As an example of the application of the new algorithm, the Stable program for frequency stability analysis [7] has adopted it (since Version.4) instead of the multiple algorithms previously used for setting confidence limits and error bars. We wish to maintain a clear distinction between a stability variance and its estimators; in particular, we treat both the OE and the NOE of each variance. Even though the OE usually has lower uncertainty than the NOE, the NOE is convenient when phase data are processed in real time or read sequentially from a file. In fact, the TSC 50A Time Interval Analyzer [8], in its averaging mode, computes an NOE of modified Allan variance. The goal of the new algorithm is not closed formulas but fast numerical results whose accuracy is adequate for the purpose at hand. All the calculations are based on the same theoretical principles, with no empirical formulas. For each T, one must choose a dominant phasenoise power law, S, ( f ).fa where a E {,,0,,,, 4) (white PM to random run FM); see [9] for a method of powerlaw identification. The phase noise is assumed to be approximately bandlimited to the Nyquist frequency / (70). Not covered are effects of trend removal, drift removal for Allan variance in particular; the dth phase differences are assumed to have zero mean. One can use Hadamard variance to obtain stability results that are invariant to linear frequency drift. Special longterm stability statistics, such as total Allan variance [0], total Hadamard variance [9], and Theol [], are not covered; these require their own treatments. N Theory of operation Although the presentation of the algorithm is selfcontained, a brief account of the theory behind it may be helpful. The algorithm s output is the equivalent degrees of freedom () of an unbiased estimator V of a stability variance u = EV. Define It has been observed empirically (but not systematically) that ( v / u )V has approximately a xz distribution. Thus, having computed v and observed V, one can obtain confidence intervals of form vv/xz 5 cr 5 v V / z ~from x$ levels < x [7]. The model for phase x ( t )is the 70difference of a pure powerlaw process: x ( t )= ATOW ( t )7 () where w ( t )is a continuoustime process with spectral density for all f > 0, and A is the backward difference operator. Then S, ( f ) is asymptotically proportional to fa as f 0 and has approximate bandwidth / (70); this is the first reason for using w ( t ). Now let ( t )= A$A ( t ), () d where E = TO or T. For E = TO we have z ( t )= A,z ( t ),which leads to an unmodified variance. For E =7 =m~0 we observe from () that A,w ( t )= w ( t ) w ( t TO) = m A,w ( t TO) = m (t TO) = me ( t ), n=o n=o where 5 ( t ) is a discretetime average of m samples of x. In this case, z ( t ) = ma$ ( t ),which leads to a modified variance [6]; this is the second reason for using w ( t ). In either case we

3 assume that z ( t )is a stationary zeromean Gaussian process with autocovariance (ACV) function S, ( t ) = Ez (U t ) z (u). Ignoring the conventional scaling factors, we define the stability variance and its estimator by where the stride S is 70for the OE and 7 for the NOE. The number of terms M depends on the estimator type and the number of data. We have EV = a = s, (0). Then cov [ z ( t ),z (u)]= s; (u t ), and The definition ()) after substitution of ( 5 ), simplifies to The ACV sz ( t )is obtained from () by applying a difference operator of order d generalized autocovariance (GACV) s, ( t )of the powerlaw process w (t) []: sz d ( t )= (ATAT) A,A,s, to the (t). The function s, ( t )is given below [6] for each a. Algorithm for calculation Our purpose is to obtain practical numerical approximations of (6). We give two versions of the algorithm: the simplified version merely truncates the sum in the exact formula; the full version maintains the number of summation terms below a presassigned threshold and avoids catastrophic roundoff errors. They have the same inputs, output, function definitions, and initial step. Some explanation of the approximations is given in Appendix A. Because the results are invariant to time scaling, we may set 7 =, 70= l/m. All arithmetic is to be carried out in doubleprecision floating point. Operations that give signed integers are the floor function x (greatest integer that is 5 x) and ceiling function.r = xj (least integer that is x).. Q Inputs = frequency noise exponent a =,, 0,,,, 4 Noise type = WHPM, FLPM, WHFM, FLFM, RWFM, FWFM, RRFM d = order of phase difference d = : firstdifference variance (included for completeness) d = : Allan variance d = : Hadamard variance Restriction: a d > m = averaging factor 7 / 7 0, positive integer

4 F = filter factor F = : modified variance F = m: unmodified variance S = stride factor (estimator stride = T / S ) S = : nonoverlapped estimator S = m: overlapped estimator N = number of phase data with sample period. TO Output = equivalent degrees of freedom of the variance estimator. Constant and function definitions Set an integer constant J,, (used only in the full version); we suggest J,, = 00. The formal arguments of the following functions have the same names as the input arguments of the main algorithm.. Define the function s, ( t,a ) as follows: a,s ( t, a ) ~tl tn It t4in ~tl [ti5 t6n It 0 lt 4 lt7. (7) The entries with In It( must evaluate to 0 when t = 0.. Define the function s, ( t,00, a ) =,s ( t,a ), 4 5 a Define the function s, ( t,f, a, d ) = ( A I A ~ ) (~t,f s,a ), d =,, ; (9) that is (with dependence on F and a suppressed on the right sides), sz ( t,f, a, ) = 6 s x ( t ) 4sZ ( t ) 4 s X (t SZ ) S, (t ) sx (t ), ( t,f, Q, ) = Osx ( t ) 5sZ ( t ) 5sX (t ) 6s, (t ) 6sX(t ) S, (t ) S, ( t 4. Define the function ( BasicSum(J,M,S,F,a,d)=s;(O,F,a,d) J j=l ( 4 G s, &) ($,F, a, s; ),F,a,d d).

5 .4 Initial steps for both versions. Compute M, the number of summands in the estimator, as follows: if N L, otherwise there are not enough data.. Let J = min ( M,(d )S). Main procedure, simplified version.5 This is just one step: _ BasicSum ( J,M, S, F, a,d ). sz (0, F, a, d ) M () To check the effect of the truncation, one can also,try a larger value of J, say min ( M,6s). Main procedure, full version.6 M Let r =. S There are four cases. The calculations use coefficients from three numerical tables..6. Case. Modified variances: F =, all a This case also applies to unmodified variances when F = m =. If J 5 Jm, _ BasicSum (J,M, S,,a,d ) sz (0,, a, d ) M Else if J > J,, and r d, take ao,al from Table ; then r Jmax Else let m' = (not necessarily an integer); then r _ BasicSum (J", sz (0,, a,d ) Jmm.6. Jmax, m',,a, d ) Case. Unmodified variances, WHFM to RRFM: F = m, a If J I Jmax If m ( d ) 5 J,, then let m' = m else let m' = 00. Then sz (0, m', a, d ) M BasicSum ( J,M, S, m', a, d ) 5 I0

6 Else if J > Jmax and T d, take ao,al from Table ; then r Jmax Else let m' = (not necessarily an integer); then r SP BasicSum (Jmax, Jmax, m', 00, Q, d ) (O,OO, a, d) Jm,.6. Case. Unmodified variances, FLPM: F = m, If J 5 Jmax I _ s? (0, m,,d ) M Q = BasicSum (J,M, S,m,,d ) Remark: For this case, m must be less than about lo6 to avoid roundoff error. Else if J > Jmax and r d, take ao,al from Table (Q = ),bo, bl from Table ; then Jmax (not necessarily an integer); then Else let m' = r _ (bo bl In m)' Jmax BasicSum (J",.6.4 Case 4. Unmodified variances, WHPM: F = m, Jm,, Q m', m',,d ) = This calculation is exact, and can be expressed in closed form. In these formulas, binomial coefficient Let K = IfKSd r.. n! IC! (n I C )!. r Else M where a0 = also given in Table ( a = ). 6 d ( denotes the

7 .7 Tables Table. Coefficients for modified variances d=l d= d= I I Table. Coefficients for unmodified variances d=l d= d= a o $ a z 5. B 4 a a a0 a % z Table. Coefficients for logarithmic denominator, unmodified variances, FLFM ( a = ) d=l bo bl d= bo d= bl bo bl 40 Examples First, we must point out that the new values differ somewhat from older ones because of our choice of phase noise model. Previously (for a < ) the phase was generally assumed to have a pure f a spectrum; here the phase is modeled as the first difference of an process. For example, consider overlapped Allan variance, white FM, 05 phase data. The old results are from [4] (close to those of [7]). 7/70 old new The old and new results are in practical agreement at the larger values of r, where the results matter more. By allowing this mild discrepancy, we were able to make the algorithm simpler and more uniform. Figures and (at the end of the paper) show examples of, computed by the new algorithm, as a function of noise type and of variance type with other parameters fixed. 7

8 5 Conclusion The stability variances considered here can all be put into the same form, namely, the meansquare average of the output of a finitedifference filter acting, not on the phase samples, but on their cumulative sums. With this insight, we have been able to make an algorithm for computing the equivalent degrees of freedom of variance measurements. It covers all the commonly used variances and estimators, and then some. There is now a single unified source for these uncertainty calculations instead of the many papers that each cover only one variance and perhaps only one estimator of it. Complete pseudocode for the new algorithm is given here; software is available on request []. A Appendix: explanat ion of approximations As is, ( 6 ) is unfit for numerical computation. We find empirically that sp ( t )tends rapidly to zero as t increases beyond d. For the accuracy needed here (a few percent), there is no point in allowing j / S to be more than d. Indeed, for sufficiently large t the calculation of sp ( t )blows up from roundoff error, even in double precision, because linear combinations of large sw values are taken to get small s, values. At the very least, one should truncate the sum at j = (d )S, as in the simplified version of the algorithm. we do the The full version of the algorithm uses the following general strategy. If J 5 J,, there are two cases. First, if A4 (d )S then S = m summation ( 6 ). When J > J,, J,,/ (d ) >>. We truncate the sum at ( d )S and approximate it by an integral; this gives dl vd r 0 = (ao r (l;)s:(t)dt?), where These coefficients can be evaluated in advance. Second, if M < (d )S then we do another summation in which J is reduced from M to Jmax and S is reduced proportionately from m. The extra term for j = J in BasicSum makes the sum a trapezoidal approximation to the integral, whether or not the sum is truncated. This method works straightforwardly for Case ; indeed, in this setting the modified variances are simpler than the unmodified ones. In Case, when m is large we compute s, ( t ) using the limiting form s, ( t,m), which is actually sl ( t ). This means that we are treating z ( t ) as w' ( t ), the process w ( t )being differentiable in the meansquare sense. The most troublesome case is the overlapped estimators of the unmodified variances for flicker PM. As S = m 00, s, ( t ) approaches a function with logarithmic singularities. The factor bo bl l n m is an asymptotic form of s, (0). It would be possible (though inconvenient) to add another largem subcase as in Case, but one does not expect flicker PM to be the dominant noise type when m is large. Case 4 is constructed by knowing that the phase samples are accurately uncorrelated when w ( t ) is a Wiener process. The simplified computation () is correct but wasteful because s, ( t ) is a linear combination of hatshaped peaks of width /m. 8

9 References F. Vernotte, 00, Application of the moment condition to noise simulation and to stability analysis, IEEE Transactions on Ultrasonics, Ferroelectrics,and Frequency Control, 49, The unmodified and modified Hadamard variances are called the Picinbono and pulsar variances. P. Lesage and C. Audoin, 97, Characterization of frequency stability: uncertainty due to the finite number of measurements, IEEE Transactions on Instrumentation and Measurement, IM, 576. P. Lesage and C. Audoin, 974, Correction to Characterization of frequency stability: uncertainty due to the finite number of measurements, IEEE Transactions on Instrumentation and Measurement, IM, 0. P. Lesage and C. Audoin, 975, Comments on Characterization of frequency stability: uncertainty due to the number of measurements, IEEE Transactions on Instrumentation and Measurement, IM4, 86 P. Lesage and C. Audoin, 976, Correction to characterization of frequency stability: uncertainty due to the number of measurements, IEEE Transactions on Instrumentation and Measurement., IM5, 70. P. Lesage and C. Audoin, 977, Estimation of the twosample variance with limited number of data, st Annual Symposium on Frequency Control, 8. D. Howe, D. Allan, and J. Barnes, 98, Properties of signal sources and measurement methods, 5th Annual Symposium on Frequency Control, 47. Reprinted in [8], TN460; see also [9], 5. In the flicker PM expression on p. 7, the argument of exp should be raised to the / power; in the flicker FM expression, N should be squared. These corrections have been made in [lo], 4. D. Sullivan, D. Allan, D. Howe, and F. Walls (editors), 990, Characterization of Clocks and Oscillators, NIST Technical Note 7, U. S. Department of Commerce, National Institute of Standards and Technology. Online at go./ timefreq/gener al/tnl7/ S. Stein, 985, Frequency and timetheir measurement and characterization, in E. Gerber and A. Ballato (editors), Precision Frequency Control, Academic Press, New York. Reprinted in [8], TN60. [lo] IEEE Std 9999, IEEE Standard Definitions of Physical Quantities for Fundamental Frequency and Time MetrologyRandom Instabilities. [ll] K. Yoshimura, 978, Characterization of frequency stability: uncertainty due to the autocorrelation of the frequency fluctuations, IEEE Transactions on Instrumentation and Measurement, IM7, 978, 7. [] K. Yoshimura, 989, Degrees of freedom of the estimate of the twosample variance in the continuous sampling method, IEEE Transactions on Instrumentation and Measurement, 8, [] C. Greenhall, 98, A structure function representation theorem with applications to frequency stability estimation, IEEE Transactions on Instrumentation and Measurement, IM,

10 [4] C. Greenhall, 99, Recipes for degrees of freedom of frequency stability estimators, IEEE Transactions on Instrumentation and Measurement, 40, [5] T. Walter, 994, Characterizing frequency stability: a continuous powerlaw model with discrete sampling, IEEE Transactions on Instrumentation and Measurement, 4, [I6 C. Greenhall, 997, The thirddifference approach to modified Allan variance, IEEE Transactions on Instrumentation and Measurement, 46, [7] Stable, Software for frequency stability analysis, Hamilton Technical Services, 95 Woodbury St., S. Hamilton, MA 098, [8] Timing Solutions Corporation, 4775 Walnut St. Suite lb, Boulder, CO , www. timing.com [9] D. Howe, R. Beard, C. Greenhall, F. Vernotte, and W. Riley, 00, A total estimator of the Hadamard function used for GPS operations, Proceedings of the nd Annual Precise Time and Time Interval (PTTI) Meeting, Nov. 000, Reston, VA, [0] D. Howe, 000, The total deviation approach to longterm characterization of frequency stability, IEEE Transactions on Ultrasonics, Ferroelectriqand Frequency Control, 47, 00. [] D. Howe and T. Peppler, 00, Very longterm frequency stability: estimation using a specialpurpose statistic, Proceedings of the 00 IEEE International Frequency Control Symposium, 8. [] Upon request by the second author will send an electronic copy of C source code that implements this algorithm, along with a small Microsoft Windows (r) program that computes the for various cases. 0

11 Overlapping ADEV EDF for N=lOO I 0 00 Averaging Factor, m Figure. Edf vs. averaging factor with powerlaw noise type as a parameter: Allan variance, overlapped estimator, 00 phase samples. 00 Overlapping EDF for N=00 and W FM Noise Averaging Factor, m Figure. Edf vs. averaging factor for three stability variances: overlapped estimator, white FM, 00 phase samples.