Principles of Time Scales Judah Levine Time and

Principles of Time Scales Judah Levine Time and Frequency Division NIST Boulder jlevine@boulder. nist. gov 303 497 -3903 Judah Levine, NIST, CENAM, Oct 2012 1

Outline n Time scale principles – Examples of special cases • AT 1 and EAL • Large Drift or Long averaging • Large measurement noise or near real-time – The general problem • Kalman Solution Adding a steered clock n Steering the time scale n Judah Levine, NIST, CENAM, Oct 2012 2

What and why? A time scale is a procedure for combining the data from several clocks n Inputs: n – (Initial estimates of the statistical characteristics of each member) – Measurements of times or frequencies of all members with respect to a reference device • Reference device need not be special Judah Levine, NIST, CENAM, Oct 2012 3

What and why? A time scale is a procedure for combining the data from several clocks n Outputs: n – ensemble time and frequency – Statistical performance of each member – Update to model for each clock – (Physical realization of ensemble time) Judah Levine, NIST, CENAM, Oct 2012 4

What and why? n Advantages: – Minimize single points of failure – Output does not depend on a single device – Ensemble provides error detection – Get the best of each contributor • Nominally identical clocks may not be equal • Combine clocks with different properties Judah Levine, NIST, CENAM, Oct 2012 5

Partition of input time differences n Noise of the measurement process – Time noise with no frequency aspect Deterministic model of each clock n Stochastic contribution of each clock n Non-statistical glitches for each clock n Judah Levine, NIST, CENAM, Oct 2012 6

TDEV of measurement systems in seconds, common clock into two channels 1. 00 E-12 sec 1. 00 E-13 1. 00 E-14 1. 00 E+02 1. 00 E+03 1. 00 E+04 1. 00 E+05 1. 00 E+06 Averaging time, s Judah Levine, NIST, CENAM, Oct 2012 7

Time Scale Clock Model Each clock in time scale has iterative model: AT 1 Model: j =0 for all j Measurement interval, clock model, and noise parameters are related and must be considered together Judah Levine, NIST, CENAM, Oct 2012 8

Variance in AT 1 clock model In AT 1 model, variance of time differences Is due to pure white frequency noise Frequency drift is constant parameter Judah Levine, NIST, CENAM, Oct 2012 9

AT 1 Algorithm, continued n n Measured time differences represent differences of time states of clocks Frequency estimate has deterministic and white noise contributions – Averaging statistically appropriate • Time constant determined by flicker frequency floor – Frequency estimate ( x/ t) freq. state y(tk ) n Drift parameter determined outside of algorithm – Treated as a constant by AT 1 Judah Levine, NIST, CENAM, Oct 2012 10

Ensemble Time n Computed as weighted average of each clock – Weight derived from prediction error on previous cycles – Sum of weights is 1 Statistically optimum weights Judah Levine, NIST, CENAM, Oct 2012 11

Ensemble Frequency and Drift n AT 1 algorithm does not explicitly calculate these parameters – Ensemble frequency is time evolution of ensemble time – Ensemble frequency drift is time evolution of ensemble frequency Statistically ok over WFM noise domain Statistically difficult, Estimate not robust Judah Levine, NIST, CENAM, Oct 2012 12

Clock Correlation Correction - 1 n n Every clock is a member of ensemble used to evaluate its performance Prediction error is always too small – Weight is biased too large – Error detection is degraded – Positive Feedback loop Judah Levine, NIST, CENAM, Oct 2012 13

Clock correlation Correction - 2 Statistical Weight Adjustment (Tavella, EFTF): Administrative weight limiting: NIST: 30%, EAL: 2. 5/N Weight limiting always degrades the time scale Most serious in small ensemble with very different true weights Judah Levine, NIST, CENAM, Oct 2012 14

Error detection and clock resets Assume clock error if: NIST model: K < 3: no error 3<k<4: k>4: Error is modeled as a single time step with no change in frequency or drift parameters Judah Levine, NIST, CENAM, Oct 2012 15

The frequency drift problem Suppose: Frequency variance no longer white frequency noise AT 1 -type algorithm no longer statistically robust AT 1 -type algorithms cannot be used when t too large and frequency drift has significant variance Judah Levine, NIST, CENAM, Oct 2012 16

Frequency Drift Solutions n Short measurement interval – Frequency variance approximately wfm n Mixed ensemble computed iteratively – Separate computation for clocks with negligible drift n Full Kalman algorithm – Complex and difficult to handle errors Judah Levine, NIST, CENAM, Oct 2012 17

The measurement noise problem Suppose: Measured time differences due to two sources Time state differences no longer time differences Frequency estimator no longer statistically robust AT 1 -type algorithms cannot be used at sufficiently short averaging times Judah Levine, NIST, CENAM, Oct 2012 18

Significant Measurement Noise n Problem important when time differences are noisy or as t 0 – AT 1 algorithm cannot be used for near real-time systems n Measured time differences must be partitioned into measurement noise and clock noise – Measurement noise must not degrade clock parameter estimates Judah Levine, NIST, CENAM, Oct 2012 19

Kalman Solution n Partition variance of measurements based on initial estimates of noise parameters and covariance matrix – Jones and Tryon, TA(NBS) – GPS Composite clock (Brown) – KAS 2 (Sam Stein, Symmetricom) Judah Levine, NIST, CENAM, Oct 2012 20

Summary - 1 n AT 1 -type algorithms assign variance to frequency noise – Measurement noise very small – Frequency drift constant (or 0) n Errors are modeled as simple time steps with no change in parameters Judah Levine, NIST, CENAM, Oct 2012 21

Summary - 2 n AT 1 -type algorithms are appropriate only over a range of averaging times determined from the clock statistics – Lower limit from measurement noise – Upper limit from frequency variance n Kalman-type algorithms can handle more complex noise types – More sophisticated partition of measured variance – Reset/Error detection more difficult to handle • Reset machinery is outside of statistical considerations Judah Levine, NIST, CENAM, Oct 2012 22

Correlations among clocks Time scale algorithms assume variance of clocks is not crosscorrelated n Common-mode effects are a serious problem n – Common time step in high-weight clocks • Wrong clocks are reset Judah Levine, NIST, CENAM, Oct 2012 23

Clock steering Time and frequency of the scale are paper parameters Scale algorithm defines offset of each member relative to the ensemble average No member clock realizes the ensemble-average values Judah Levine, NIST, CENAM, Oct 2012 24

Statistics of a real-time ensemble Interaction between weighting algorithm and clock noise usually results in random walk at longer term Every ensemble requires external data for steering Judah Levine, NIST, CENAM, Oct 2012 25

Steered clock Measurement system and time scale computation Data from clock ensemble Phase stepper Clocks are not steered Steering Control Steered output Judah Levine, NIST, CENAM, Oct 2012 26

Steered Clock Error Signal n Steered clock usually steered based on time: – Simple steering drives x s 0 • Steered clock realizes ensemble time – More complex steering • Steered clock is UTC(lab) steered to UTC – Error signal is UTC(lab)-UTC from Circular T • xs x 0+y(t-t 0)+0. 5*d*(t-t 0)2 Judah Levine, NIST, CENAM, Oct 2012 27

Statistics of the steered output Free-running performance defined by statistics of steered clock reference Time Noise in the reference clock for the phase stepper: 5 10 -13 1/2 = 13 ps @ 12 minutes Steering loop drives steering error to 0 Long-period performance defined by stability of the scale Judah Levine, NIST, CENAM, Oct 2012 28

Types of steering algorithms Time-driven: Minimize time error Frequency driven: Minimize frequency excursions Bang-bang Drift: Frequency and time continuous Steering algorithm set by administrative considerations and by needs of users No Universal “perfect” solution Judah Levine, NIST, CENAM, Oct 2012 29

Judah Levine, NIST, CENAM, Oct 2012 30

Judah Levine, NIST, CENAM, Oct 2012 31

References n Realizing UTC(NIST) at a Remote Location – Metrologia, Vol. 45, page S 23, 2008 Other papers in this volume of Metrologia n The Statistical Model of Atomic Clocks and the Design of Time Scales n – Review of Scientific Instruments, Feb. 2012 Judah Levine, NIST, CENAM, Oct 2012 33