Introduction To Time and Frequency Metrology
Introduction To Time and Frequency Metrology
Introduction To Time and Frequency Metrology
REVIEW ARTICLE
2567
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2568 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
Either type of oscillator can be used to construct a clock. III. DEFINITION OF TIME INTERVAL OR FREQUENCY
The periodic output is converted to a series of pulses ~using
Time interval is one of the base units of the International
a mechanical escapement or an electrical zero-crossing de-
System of Units ~the SI system!. The current definition of the
tector, for example!, and the resulting ‘‘ticks’’ are counted in
second was adopted by the 13th General Conference on
some way. While the frequency of an oscillator is a function Weights and Measures ~CGPM! in 1967:
of its mechanical or physical properties, the initial reading of
The second is the duration of 9 192 631 770 periods of
a clock is totally arbitrary—it has to be set based on some
the radiation corresponding to the transition between
external definition. This definition cannot be specified from
the two hyperfine levels of the cesium 133 atom.3
the technical point of view—for historical and practical rea-
sons the definition that is used is based on the motion of the This definition of the base unit can be used to derive the
unit of frequency, the hertz. Multiples of the second, such as
earth.1
the minute, hour, and day are also recognized for use in the
The vast majority of oscillators currently in use are sta-
SI system. Although these common units are defined in
bilized using quartz crystals. Even the quartz crystals used as terms of the base unit just as precisely as the hertz is, they do
the frequency reference in inexpensive wrist watches can not enjoy the same status in the formal definition of the
have a frequency accuracy of 1 ppm and a stability a factor international system of units.4
of 10 or more better than this. Substantially better perfor- The definition of the SI second above implies that it is
mance can be achieved using active temperature control. realized using an unperturbed atom in free space and that the
They are usually the devices of choice when lowest cost, observer is at rest with respect to the atom. Such an observer
robust design, and long life are the most important consider- measures the ‘‘proper’’ time of the clock, that is the result of
ations. They are not well suited for applications ~such as a direct observation of the device independent of any con-
primary frequency standards! where frequency accuracy or ventions with respect to coordinates or reference frames.
long-term frequency stability are important. There are two Practical timekeeping involves comparing clocks at dif-
ferent locations using time signals, and these processes intro-
reasons for this. The mechanical resonance frequency de-
duce the need for coordinate frames. International atomic
pends on the details of the construction of the artifact and is
time ~TAI! is defined in terms of the SI second as realized on
therefore hard to replicate. In addition, the resonance fre- the rotating geoid,5 and TAI is therefore a coordinate ~rather
quency is usually affected by fluctuations in temperature and than a proper! time scale.
other environmental parameters so that its long-term stability This distinction has important practical consequences. A
is relatively poor. A number of very clever schemes have perfect cesium clock in orbit around the earth will appear to
been developed to mitigate these problems, but none of them observers on the earth as having a frequency offset with re-
can totally eliminate the sensitivity to environmental pertur- spect to TAI. This difference must be accounted for using the
bations and the stochastic frequency aging that are the char- usual corrections for the gravitational redshift, the first- and
acteristics of an artifact standard.2 second-order Doppler shifts, etc. These frequency correc-
Atomic frequency standards use an atomic or molecular tions are important in the use of the global positioning sys-
transition as the discriminator. In the passive configuration, tem ~GPS! satellites, as we will discuss below.
the atoms are prepared in one of the states of the clock tran- A. Realization of TAI
sition and are illuminated by the output from a separate os-
TAI is computed retrospectively by the International Bu-
cillator. The frequency of the oscillator is locked to the maxi- reau of Weights and Measures ~BIPM!, using data supplied
mum in the rate of the clock transition. An atomic frequency by a world-wide network of timing laboratories. Each par-
standard therefore requires a ‘‘physics package’’ that pre- ticipating laboratory reports the time differences, measured
pares the atoms in the appropriate state and detects the fact every five days ~at 0 h on those modified Julian day numbers
that they have made the clock transition following the inter- which end in 4 or 9!, between each of its clocks and its
action with the oscillator. It also requires an ‘‘electronics laboratory time scale, designated UTC~lab!. To make it pos-
package’’ that consists of the oscillator, the feedback cir- sible to compare data from different laboratories, each labo-
cuitry to control its frequency, and various synthesizers or ratory also provides the time differences between UTC~lab!
frequency dividers to convert the clock frequency to standard and GPS time, measured using an algorithm defined by the
output frequencies such as 5 MHz or 1 Hz. Since the transi- BIPM and at times specified in the BIPM tracking schedules.
tion frequencies are determined solely by the atomic struc- ~The GPS satellites are used only as transfer standards, and
the satellite clocks drop out of the data.! Approximately 50
ture in principle, these standards would seem to be free of
laboratories transmit data from a total of about 250 clocks to
many of the problems that limit the accuracy of artifact stan-
the BIPM using these methods.
dards. This is true to a great extent, but the details of the Using an algorithm called ALGOS, the BIPM computes
interaction between the atoms and the probing frequency and the weighted average of these time-difference data to pro-
the interaction between the physics and electronics packages duce an intermediate time scale called echelle atomique libre
affect the output frequency to some degree, and blur the ~EAL!. The length of the second computed in this way is
distinction between a pure atomic frequency standard and compared with data from primary frequency standards and a
one whose frequency is determined by a mechanical artifact. correction is applied if needed so as to keep the frequency as
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2569
close as possible to the SI second as realized by these pri- is inserted as the last second of the day ~usually on the last
mary standards. The resulting steered scale is TAI. As of day of either June or December!, after 23:59:59, which
December, 1998, the fractional frequency difference, would normally have been the last second of the day. Its
f (EAL)2 f (TAL)57.13310213. At that same epoch, the name is 23:59:60, and the next second is 00:00:00 of the next
fractional frequency offset between TAI and the SI second as day. The first problem is that it is difficult to define a time
realized on the rotating geoid was estimated by the BIPM to tag for an event that happens during the leap second. This
be time does not even exist on conventional clocks, for ex-
ample, and a time tag of 23:59:60 might be rejected as a
d5 ~ uTAI2u0 ! /u05 ~ 24610! 310215. format error by a digital system. Furthermore, there is no
The determination of the uncertainty for the estimate of d is natural way of naming the leap second in systems that do not
complex. It includes the uncertainties ~usually the ‘‘type B’’ use the hh:mm:ss nomenclature. Many computer systems fall
uncertainty, which we define later! associated with the data into this category because they keep time in units of seconds
from the different primary frequency standards; extrapolat- ~and fractions of a second! since some base date.
ing these data to a common interval also requires a model for Calculating the length of a time interval that crosses a
the stability of EAL. The details are presented in reports leap second presents additional ambiguities. An interval
published by the BIPM.6 based on atomic time or on some other periodic process will
differ from a simple application of a calculation based on the
difference in the UTC times, since the latter calculation ef-
B. Coordinated universal time and leap seconds fectively ‘‘forgets’’ that a leap second has happened once it
It is simple in principle to construct a timekeeping sys- has passed.
tem and a calendar using TAI. When this scale was defined, There is no simple way of realizing a time scale that is
its time was set to agree with UT2 ~a time scale derived from both smooth and continuous and simultaneously tracks UT1.
the meridian transit times of stars corrected for seasonal This is because the intervals between leap seconds are not
variations! on 1 January 1958. Unfortunately, the length of exactly constant and predictable, and there is no simple al-
the day defined using 86 400 TAI seconds was shorter than gorithm that could predict them automatically very far into
the length of the day defined astronomically ~called UT1! by the future. Each leap second is announced several months in
about 0.03 ppm, and the times of the two scales began to advance by the International Earth Rotation Service in Paris.8
diverge immediately. If left uncorrected, this divergence Thus, changing the SI definition of the second ~which would
would have continued to increase over time, and its rate produce difficulties in the definitions of other fundamental
would have slowly increased as well as the rotational rate of constants anyway! could reduce the frequency of leap sec-
the earth continued to decrease. onds, but would not entirely remove the need for them. The
Initially, this divergence was removed by introducing problem would return in the future anyway as the length of
frequency offsets between the frequency used to run atomic the astronomical day increased.
clocks and the definition of the duration of the SI second. It is possible to avoid the difficulties associated with leap
Small time steps of 0.05 or 0.1 s were also introduced as seconds by using a time scale that does not have them, such
needed.7 This steered time scale was called coordinated uni- as TAI or GPS time. There are a number of practical diffi-
versal time ~UTC!—it was derived from the definition of the culties with this solution, such as the facts that most timing
SI second, but was steered ~that is, ‘‘coordinated’’! so that it laboratories disseminate UTC rather than TAI, and that GPS
tracked the astronomical time scale UT1. time may not satisfy legal traceability requirements in some
This method of coordinating UTC turned out to be awk- situations. Practical realizations of this idea often require an-
ward in practice. The frequency offsets required for coordi- cillary tables of when leap seconds were inserted into UTC.
nation were on the order of 0.03 ppm, and it was difficult to
insert these frequency offsets into real-world clocks. On 1
IV. CESIUM STANDARDS
January 1972, this system was replaced by the current sys-
tem which uses integral leap seconds to realize the steering Cesium was chosen to define the second for a number of
needed so that atomic time will track UT1. The rate of UTC practical reasons. The perturbations on atomic energy levels
is set to be exactly the same as the rate of TAI, and the are quite small in the low-density environment of a beam,
divergence between UTC and astronomical time is removed and atomic beams of cesium are particularly easy to produce
by inserting ‘‘leap’’ seconds into UTC as needed. The leap and detect. The frequency of the hyperfine transition that
seconds are inserted so as to keep the absolute magnitude of defines the second is relatively high and the linewidth can be
the difference between UTC and UT1 less than 0.9 s. These made quite narrow by careful design. At the same time the
leap seconds are usually added on the last day of June or transition frequency is still low enough that it can be ma-
December; the most recent one was inserted at the end of 31 nipulated using standard microwave techniques and circuits.
December 1998 and the time difference TAI-UTC became In spite of these advantages, constructing a device that
exactly 32 s when that happened. The leap seconds are al- realizes the SI definition of the second is a challenging and
ways positive, so that UTC is behind atomic time, and this expensive undertaking. Although the frequency offsets pro-
trend will continue for the foreseeable future. duced by various perturbations ~such as Stark shifts, Zeeman
Although this system of leap seconds is simple in con- shifts, Doppler shifts,...! are small in absolute terms, they are
cept, it has a number of practical difficulties. A leap second large compared to the frequency stability that can be
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2570 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
achieved in a well-designed device. A good commercial ce- usually measured by observing the frequencies of field-
sium frequency standard, for example, might exhibit frac- dependent transitions between other m F states.
tional frequency fluctuations of 2310214 for averaging The microwave radiation that induces the clock transi-
times of about one day. The frequency of the same device tion is applied in two separate regions separated by a drift
might differ from the SI definition by 1310213 or more space ~Ramsey configuration10!. The atoms experience two
~sometimes much more!, and this frequency offset may kinds of transitions: ‘‘Rabi’’ transitions due to a single inter-
change slowly with time as the device ages. ~This frequency action with the microwave field, and ‘‘Ramsey’’ transitions
offset is what remains after corrections for the perturbations due to a coherent interaction in both regions. The Rabi tran-
mentioned above have been applied. If no corrections are sitions have a shorter interaction time and therefore a larger
applied, the fractional frequency offset is usually dominated linewidth. The magnitude of the magnetic field in the C re-
by Zeeman effects, which can be as large as 1310210.! gion is usually chosen so that the Rabi transitions between
Constructing a device whose accuracy is comparable to the different sub levels will be well resolved. A typical value
its stability is a difficult and expensive business, and only a for a conventional thermal-beam standard would be about
few such primary frequency standards exist. It is usually im- 531026 T. It is possible to use a lower value of magnetic
possible to reduce the systematic offsets to sufficiently small field which does not completely resolve the Rabi transitions.
values, and the residual offset must be measured and re- While this reduces the Zeeman correction, it complicates the
moved by means of a complicated evaluation process that analysis of the Ramsey line shapes.
often takes several days to complete.9 The operation of the The length of this drift space and the velocity of the
standard may be interrupted during the evaluation, so that atoms determine the total interaction time, which in turn sets
many primary frequency standards cannot operate continu- the fundamental linewidth of the device. The microwave fre-
ously as clocks. In addition, the systematic offsets may quency is adjusted until the transition rate in the atoms is a
change slowly with time even in the best primary standard, maximum; the transitions are detected using a second inho-
so that evaluating their magnitudes is a continuing ~rather mogeneous magnetic field, which is created by the ‘‘B’’
than a one-time! effort. magnet. The gradient of the B field is usually configured so
A simple oven can be used to produce a beam of cesium as to pass only those atoms that are in the final state of the
atoms. The oven consists of a small chamber and a narrow clock transition ~a ‘‘flop-in’’ configuration!. This configura-
slit through which the atoms diffuse. The operating tempera- tion eliminates the shot noise in the detected beam due to
ture varies somewhat, but is generally less than 100 °C, so atoms that do not make the clock transition in the C region.
that the thermal mean velocity of the cesium atoms is some- However, it is more difficult to test and align, since there is
what less than 300 m/s. Operating the oven at a higher tem- no signal at all until everything is working properly.
perature increases the flux of atoms in the beam and there- The state-selection process in a conventional cesium de-
fore the signal to noise ratio of the servo loop to control the vice is passive—it simply throws away those atoms that are
local oscillator, but the velocity of the atoms also increases not in the proper state. There are 16 magnetic sub levels in
as the square root of the temperature. This decreases the the ground state: 9 for F54 and 7 for F53 ~that is, 2F
interaction time and increases the linewidth by the same fac- 11 in both cases!. All of them have essentially the same
tor, but the more serious problem is usually that the collima- energy and are therefore roughly equally populated in a ther-
tion of the beam is degraded. The decrease in interaction mal beam. As a consequence, only 1/16 of the flux emerging
time resulting from raising the temperature can be offset by from the oven is in a single F, m F state, and only this small
using only the low-frequency ‘‘tail’’ of the velocity distribu- fraction of the atoms is actually used to stabilize the oscilla-
tion. tor. This unfavorable ratio can be improved using active state
Conventional frequency standards use an inhomoge- selection—that is, by optically pumping the input beam so as
neous magnetic field ~the ‘‘A’’ magnet! as the state selector. to put nearly all of the atoms into the initial state of the clock
The atoms enter the A magnet off axis and are deflected by transition. A similar technique can be used to detect the at-
the interaction between the inhomogeneous magnetic field oms that have undergone the clock transition in the interac-
and the magnetic dipole moment of the atom. The geometry tion region. This design increases the number of atoms that
is arranged so that only atoms in the F53 state emerge from are used in the synchronization process and therefore the
the magnet—atoms in the other state are blocked by a me- signal to noise ratio of the error voltage in the servo loop that
chanical stop. It is not practical to make the magnetic field of controls the oscillator. In addition, it removes some of the
the A magnet large enough to separate the m F sub levels of offsets and problems associated with the fields due to the A
the F53 state. and B magnets. Although the optical pumping techniques
The atoms that emerge from the A magnet enter an in- that are needed to realize this design have been known for
teraction region ~the ‘‘C’’ region!, where they are illumi- many years,11 realizing a standard using these principles was
nated by microwave radiation in the presence of a constant not practical until recently when lasers of the appropriate
magnetic field, which splits the hyperfine levels because of wavelengths became available.
the Zeeman effect. Although the transition between the F Another technique for improving the performance of the
53 m F 50 to F54 m F 50 is offset somewhat as a result of device is to increase the interaction time in the C region by
this field, the sensitivity to magnetic field inhomogeneities is increasing its length or by decreasing the velocity of the
reduced, and the clock transition is independent of the mag- atoms. The ultimate version of this idea may be an ‘‘atomic
netic field value in first order. The actual value of the field is fountain’’ in which atoms are shot upward through the inter-
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2571
action region, reverse direction, and fall back down through of 1%! is used to lock the incident microwave frequency to
the same region a second time. This idea was proposed more the clock transition.
than 40 years ago,12 but has only recently become feasible as The beauty of this design is that the microwave transi-
a result of developments in laser cooling and trapping. The tion modulates an optical intensity, and changes in the vis-
fountain idea has a number of advantages over one-way de- ible photon flux are much easier to detect than the same flux
signs. The atoms are moving more slowly and the interaction of microwave photons would be. On the other hand, the de-
time can be longer as a result. In addition, the fact that atoms crease in the transmitted light through the cell must be de-
traverse the interaction region in both directions can be used tected against the shot noise in the full background light
to cancel a number of systematic offsets that would other- from the lamp. Using a discriminator based on the visible
wise have to be estimated during the evaluation process. flux from the lamp makes the device simpler and cheaper,
but it also makes for poorer stability and larger and more
variable frequency offsets than would be found in a good
V. OTHER ATOMIC STANDARDS
cesium-based device. This is because of pressure shifts in the
Cesium is not the only atom that can be used as a refer- cells, light shifts due to the lamp, a large dependence of the
ence for a frequency standard; transitions in rubidium and clock frequency on the ambient magnetic field, and other
hydrogen are also commonly used. In order to relate their effects. The values of many of these parameters must be
frequencies to the SI definition, devices based on rubidium chosen as a compromise between accuracy and stability. A
or hydrogen must be calibrated with respect to primary higher-intensity lamp, for example, increases the signal to
cesium-based devices if accuracy is important to their use. noise ratio in the resonance-detection circuit, but also in-
As we have discussed above, this is true for commercial creases the light shift ~the ac Stark effect on the clock tran-
cesium devices as well if the accuracy of the output fre- sition!. A typical rubidium-stabilized device might have a
quency is to approach the frequency stability of the device. stability and an accuracy 100 times ~or more! poorer than a
In spite of this similarity, there are real differences among device using cesium in the standard beam configuration. In
the devices. terms of the Allan deviation ~to be defined later!, the perfor-
Rubidium devices usually use low-pressure rubidium va- mance is generally about 3310211/ t 1/2 for averaging times
por in a cell filled with a buffer gas such as neon or helium ~that is, values of t! between about 1 and 104 s; the long-
rather than the atomic-beam configuration used for cesium. term frequency aging is typically on the order of 5
The clock transition is the F51, m F 50 to F52, m F 50 310211/month. On the other hand, they are usually substan-
transition in the 2 S 1/2 ground state of 87Rb, its frequency is tially cheaper, lighter, and smaller than standard cesium de-
about 6.83 GHz. The two states of the clock transition are vices.
essentially equally populated in thermal equilibrium. Both the advantages and the limitations of rubidium de-
The atoms are optically pumped into the F52 state. vices are due to the implementation in an optically pumped
There are a number of ways of doing this. One common cell with a buffer gas rather than to any inherent difference
method uses a discharge lamp containing 87Rb whose light is between cesium and rubidium. A cesium-based device that
passed through an absorption cell containing 85Rb. The light used a cell rather than a beam, for example, might have
emerging from the lamp could induce optical frequency tran- many of the advantages of both systems, and such a device
sitions from both the F51 and F52 states to higher excited has been under development for some time. Furthermore, a
levels. However, there is a coincidence between transitions fountain based on rubidium would have a number of advan-
in 85Rb and 87Rb. The effect of this coincidence is that the tages over a cesium-based device. Although the clock tran-
light which would have induced transitions originating from sition has a lower frequency in rubidium, the atom–atom
the F52 state is preferentially absorbed in passing through collision cross section in the cold beam of a fountain is sub-
the 85Rb filter, whereas the component that interacts with the stantially smaller than in cesium13 so that the beam flux can
F51 lower state is not. When the filtered light enters the be greater. The short-term stability will be improved as a
cell, it preferentially excites the F51 state of the clock tran- result; a number of other systematic offsets would also be
sition because the light that would interact with the F52 smaller. ~Improving the short-term stability is a very valu-
state has been preferentially absorbed in the 85Rb cell. The able advantage, since it eases the requirements on the stabil-
excited atoms decay back to the two ground sub levels. ity of the local oscillator which interrogates the atoms and
Those atoms which decay back to the F51 state are prefer- which acts as a flywheel between cycles of the fountain.!
entially excited again, while those atoms that decay back to As with cesium and rubidium, the clock transition in a
the F52 state tend to remain there. The result is to build up hydrogen maser is a hyperfine transition in the ground state
a nonequilibrium distribution in these two sub levels. As in of the atom. The F50 and F51 hyperfine states are sepa-
cesium, a magnetic field is applied to split the hyperfine sub rated using an inhomogeneous field as in a cesium device,
levels and to reduce the sensitivity to any residual inhomo- but instead of using the beam geometry of a cesium-based
geneities and fluctuations in the ambient field. When the mi- device, the atoms in the initial state of the clock transition
crowave clock frequency is applied, the F51, m F 50 state is enter a bulb that is inside of a microwave cavity.14 In a
partially repopulated by the induced transition between the passive maser, the atoms in the cavity are probed with a
two clock states, the absorption of the incident light from the microwave signal at 1.42 GHz, and the interaction with the
lamp increases again, and the transmitted intensity drops. atoms produces a phase shift in the reflected signal that is
This decrease in transmitted intensity ~typically on the order dispersive about line center. In an active maser, the losses of
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2572 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
the cavity are small enough so that the device oscillates at can be isolated from environmental perturbations, or how
the hyperfine transition frequency. well the effects of these perturbations can be estimated and
The frequency of oscillation of an active maser is a func- removed.2 Quartz–crystal oscillators can be surprisingly
tion of both the cavity resonance frequency and the atomic good in this respect. Even cheap wrist watches have oscilla-
transition frequency, so that some means of stabilizing the tors that may have a frequency accuracy of about 1 ppm and
cavity is usually required. Collisions with the walls of the a stability 10 or 50 times better than this. The residual fre-
bulb also affect the frequency of the device. Various strate- quency fluctuations are largely due to fluctuations in the am-
gies are used to minimize these effects, including applying bient temperature, and stabilizing the temperature can im-
special ‘‘nonstick’’ coatings to the inside of the bulb and prove the performance.
actively tuning the resonance frequency of the cavity so that
it is exactly on atomic line center.
In spite of these techniques, hydrogen masers usually VII. MEASURING TOOLS AND METHODS
have frequency offsets that are large compared to the stabil- The oscillators that we have discussed above generally
ity of the device. ~The fractional frequency offset of a hydro- produce a sine-wave output at some convenient frequency
gen maser is often of order 5310211, while the fractional such as 5 MHz. ~This frequency may also be divided down
frequency fluctuations at one day can be as small as 2 internally to produce output pulses at a rate of 1 Hz. Crystals
310216.! In addition to being significant, these offsets usu- designed for wrist watches and some computer clocks often
ally change slowly with time in ways that are difficult to operate at 32 768 Hz to simplify the design of these 1 Hz
predict. Nevertheless, the frequency stability over short peri- dividers.! A simple quartz–crystal oscillator might operate
ods ~usually out to at least several days! can be much better directly at the desired output frequency; atomic standards
than the best cesium device. Hydrogen masers are therefore relate these output signals to the frequency appropriate to the
the devices of choice when the highest possible frequency reference transition by standard techniques of frequency
stability is required and cost is not an issue. multiplication and division. The measurement system thus
Finally, a number of other transitions have been used to operates at a single frequency independent of the type of
stabilize oscillators. Some of the most stable devices are oscillator that is being evaluated. The choice of this fre-
those using a hyperfine transition in the ground state of an quency involves the usual trade-off between resolution,
alkali-like ion,15 such as 199Hg1. 16 This ion has a structure which tends to increase as the frequency is made higher, and
similar to cesium, with a single 6s electron outside of a the problems caused by delays and offsets within the mea-
closed shell. The clock transition is the hyperfine transition surement hardware, which tend to be less serious as the fre-
in the ground state analogous to the transition in cesium; it quency is made lower.
has a frequency of 40.5 GHz. Using a linear ion trap to Measuring instruments generally have some kind of dis-
confine the ions, frequency standards with a stability of 2 criminator at the front end—a circuit that defines an event as
310214/ t 1/2 ~where t is the averaging time in seconds! have occurring when the input signal crosses a specified reference
been reported15 and devices based on other ions are being voltage in a specified direction. Examples are 1 V with a
developed.17 These devices have the advantage that the in- positive slope in the case of a pulse, or 0 V with a positive
teraction time can be much longer than in a conventional slope in the case of a sine-wave signal. The trigger point is
beam configuration. chosen at ~or near! a point of maximum slope, so as to mini-
mize the variation in the trigger point due to the finite rise
time of the wave form.
VI. QUARTZ CRYSTAL OSCILLATORS
The simplest method of measuring the time difference
Each of the devices we have discussed above is imple- between two clocks is to open a gate when an event is trig-
mented using a quartz–crystal oscillator whose frequency is gered by the first device and to close it on a subsequent event
stabilized to the atomic transition using some form of feed- from the second one.18 The gate could be closed on the very
back loop. The details of the loop design vary from one next event in the simplest case, or the Nth following one
device to another, but the loop will always have a finite could be used, which would measure the average time inter-
bandwidth and therefore a finite attack time. The internal val over N events. The gate connects a known high-
oscillator is essentially free running for times much shorter frequency oscillator to a counter, and the time interval be-
than this attack time, and the stability of all atomic-stabilized tween the two events is thus measured in units of the period
oscillators is therefore limited at short periods to the free- of this oscillator. The resolution of this method depends on
running stability of the internal quartz oscillator. ~The the frequency of this oscillator and the speed of the counter,
boundary between ‘‘short’’ and ‘‘long’’ periods varies from while the accuracy depends on a number of parameters in-
one device to another, of course. A typical value for many cluding the latency in the gate hardware and any variations
commercial devices is probably on the order of seconds.! in the rise time of the input wave forms. The resolution can
An obvious simplification would be to use the bare be improved by adding an analog interpolator to the digital
quartz–crystal oscillator and to dispense altogether with the counter, and a number of commercial devices use this
stabilizer based on the atomic transition. The stability of method to achieve sub-nanosecond resolution without the
such a device is clearly the same as an atomic-based system need for a reference oscillator whose frequency is 1 GHz or
at sufficiently short times. The stability at longer times will greater.
depend on how well the quartz crystal frequency reference In addition to these obvious limitations, time-difference
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2573
measurements using fast pulses have additional problems. ogy ~NIST! system is only a few ps—about a factor of 10 or
Reflections from imperfectly terminated cables may distort 20 poorer than its resolution.
the edge of a sharp pulse, and long cables may have enough Heterodyne methods are well suited to evaluating the
shunt capacitance to round it by a significant amount. In frequency stability of an oscillator, but they often have prob-
addition to distorting the wave forms and affecting the trig- lems in measuring time differences because they usually
ger point of the discriminators, these reflections can alter the have an offset, which is an unknown number of cycles of the
effective load impedance seen by the oscillator and pull it off input frequencies. Thus the time difference between two
frequency. Isolation and driver amplifiers are usually re- clocks measured using the NIST mixer system is offset with
quired to minimize the mutual interactions and complicated respect to measurements made using a system based on the 1
reflections that can occur when several devices must be con- Hz pulse hardware by an arbitrary number of periods of the 5
nected to the same oscillator, and the delays through these MHz input frequency ~i.e., some multiple of 200 ns!. To
amplifiers must be measured. These problems can be ad- further complicate the problem, this offset can change if the
dressed by careful design, but it is quite difficult to construct power is interrupted or if the system stops for any other
a direct time-difference measurement system whose mea- reason.
surement noise does not degrade the time stability of a top- The offset between the two measurement systems must
quality oscillator, and other methods have been developed be measured initially, but it is not too difficult to recover it
for this reason. Averaging a number of closely spaced time- after a power failure, since the step must be an exact multiple
difference measurements is usually not of much help because of 200 ns. Using the last known time difference and fre-
these effects tend to be slowly varying systematic offsets, quency offset, the current time can be predicted using a
which change only slowly in time. simple linear extrapolation. This prediction is then compared
Many measurement techniques are based on some form with the current measurement, and the integer number of
of heterodyne system. The sine-wave output of the oscillator cycles is set ~in the software of the measurement system! so
under test can be mixed with a second reference oscillator that the prediction and the measurement agree. This constant
which has the same nominal frequency, and the much lower is then used to correct all subsequent measurements. The
difference frequency can then be analyzed in a number of lack of closure in this method is proportional to the fre-
different ways. If the reference oscillator is loosely locked to quency dispersion of the clock multiplied by the time inter-
the device under test, for example, then the variations in the val since the last measurement cycle, and the procedure will
phase of the beat frequency can be used to study the fast unambiguously determine the proper integer if this time dis-
fluctuations in the frequency of the device under test. The persion is significantly less than 200 ns. This criterion is
error signal in the lock loop provides information on the easily satisfied for a rubidium standard if the time interval is
longer-period fluctuations. The distinction between ‘‘fast’’ less than a few hours; the corresponding time interval for
and ‘‘slow’’ would be set by the time constant of the lock cesium devices is generally at least a day.
loop. Given the difficulties of making measurements using 1
This technique can be used to compare two oscillators Hz pulses, it is natural to consider abandoning them in favor
by mixing a third reference oscillator with each of them and of phase measurements of sine waves. The principal reason
then analyzing the two difference frequencies using the time- for not doing this now is that almost all of the methods that
interval counter discussed above. In the version of this idea are used to distribute time are designed around 1 Hz ticks.
developed at NIST,19 this third frequency is not an indepen- As we will see below, extracting 1 Hz ticks from a GPS
dent oscillator, but is derived from one of the input signals signal is a straightforward business in principle, while phase
using a frequency synthesizer. The difference frequency has comparisons between a local 5 MHz oscillator and the same
a nominal value of 10 Hz in this case. The time interval GPS signal have a number of awkward aspects and ambigu-
counter runs with an input frequency of 10 MHz and can ities. Very similar problems also make measurements on 1
therefore resolve a time interval of 1026 of a cycle. This is Hz pulses the method of choice in two-way satellite time
equivalent to a time-interval measurement with a resolution transfer.
of 0.2 ps at the 5 MHz input frequency.
All of these heterodyne methods share a common advan- VIII. NOISE AND MEASURES OF STABILITY
tage: the effects of the inevitable time delays in the measure-
A. Introduction to the problem
ment system are made less significant by performing the
measurement at a lower frequency where they make a much We now turn to looking at ways of characterizing the
smaller fractional contribution to the periods of the signals performance of clocks and oscillators. The simple notions of
under test. Furthermore, the resolution of the final time- Gaussian random variables will turn out to be inadequate for
interval counter is increased by the ratio of the input frequen- this job, because many of the noise processes in clocks are
cies to the output difference frequencies ~5 MHz to 10 Hz in not even close to Gaussian. In particular, the common no-
the NIST system!. This method does not obviate the need for tions of mean and standard deviation ~which completely
careful design of the front-end electronics—if anything the characterize a Gaussian distribution! must be applied with
increased resolution of the back-end measurement system great care—while their values usually exist in a formal sense,
places a heavier burden on the high-frequency portions of the they do not have the nice properties that we usually think of
circuits and the transmission systems. As an example, the when these parameters are used to characterize a true Gauss-
stability of the National Institute of Standards and Technol- ian variable. In particular, calculating either the mean or the
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2574 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
standard deviation of one of these data set results in esti- correlated—the variances estimated in this way can
mates that are not stationary and that depend on the length of even be negative in this case despite the fact that this is
the time series used for the analysis. impossible on physical grounds.
When we speak of the time or frequency of a clock in ~b! We are trying to evaluate ~or synchronize! a device
the following discussion, we almost always mean relative using data that we receive from a distant calibration
values. Time is measured as the difference ~in seconds! be- source over a noisy channel. Even though the reference
tween the device under discussion and a second identical device may be much more stable than the device under
noiseless one, and frequency is the fractional frequency dif- test, its data are degraded by the channel noise.
ference between the same two devices. If the nominal fre- It would be very helpful to be able to separate the
quency of the devices is v 0 , and if the phase difference contributions to the variance due to the channel and the
between them at some instant is f, then the fractional fre- local clock in this configuration. This separation is im-
quency difference, y, is given in terms of the time derivative portant for two reasons. In the first place, it tells us
of the phase by which part of the overall system is limiting its overall
performance, and where we should spend time and ef-
ḟ fort to improve things. In the second place, separation
y5 . ~1!
2pv0 of variance plays a central role in designing methods
Using this convention, frequencies are dimensionless quanti- for synchronizing clocks when the calibration data are
ties. We will also use the term frequency aging to denote degraded in this way.
slow changes in the fractional frequency ~where slow means This situation arises in a number of difference con-
long with respect to 1/y and certainly with respect to the texts. We discuss later two common situations: syn-
much smaller 1/v 0 !. The unit of frequency aging is s21. chronizing local clocks to GPS signals which are de-
graded by selective availability and other effects and
synchronizing clocks using time signals transmitted
over packet networks such as the Internet.
B. Separation of variance
Noiseless measuring systems and perfect clocks are
C. Measurement noise
scarce commodities, and we must use real devices in a test
system and extract from these data the performance param- All of the instruments that are used for time and fre-
eters of each of the devices. One of the most important as- quency measurements have some kind of discriminator at the
pects of this job is the concept of separation of variance— input—something that is triggered when the input wave form
that is identifying which part of the system is responsible for crosses some preset threshold. Any noise on the input signal
the observed fluctuations in the data. This idea most often combines with the internal noise of the circuit to produce a
arises in two contexts: jitter in the trigger point. These effects will produce a fluc-
tuation in the measured time difference between two clocks
~a! We are trying to evaluate the performance of a device
that has nothing to do with the clocks themselves—it arises
or system and we would like to be able to separate its
purely in the measurement process and there is no corre-
performance from the noise due to the system that we
sponding jitter in the underlying frequency difference be-
are using as a reference. Both systems can be physical
tween the two devices. It is especially important that we
clocks or pseudo clocks, such as time scales. This job
develop techniques that can identify the source of these fluc-
is easy if the reference is much less noisy than the
tuations in the data in applications where the goal is to syn-
device under test; if this is not the case then the ‘‘three-
chronize the local clock to a second standard. Adjusting the
cornered hat’’ method can be used:20
frequency of the local clock to compensate for this measured
Suppose we have three devices whose individual
time jitter is almost always the wrong thing to do, as we will
variances are s 2i , s 2j , and s 2k . We wish to estimate
see below.
these parameters, but we can only compare the devices
If the input data are a sine wave with amplitude A and
to each other using pair-wise comparisons. The mea-
frequency v, then additive noise at the input to the discrimi-
sured variances of these pair-wise comparisons are s 2i j ,
nator of amplitude V n will result in a time jitter whose am-
s 2jk , and s 2ki . If the variances of the three devices are
plitude is approximately
roughly equal, then we can estimate the three indi-
vidual variances by Vn
Dt> . ~3!
s2ij1s2ki2s2jk Av
s2i 5 ,
2 We usually assume that in each discriminator the equivalent
noise voltage at the input is not correlated with the corre-
s2jk1s2ij2s2ki sponding input signal, so that the time jitter in each channel
s2j 5 , ~2! has no mean offset relative to the corresponding noise-free
2
trigger point. Consecutive measurements of the time differ-
s2jk1s2ki2s2ij ence are then randomly distributed about a mean value that
s2k 5 .
2 represents the actual time difference between the two clocks.
These estimates can be misleading if the variances are If the instantaneous noise-induced offset in the trigger point
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2575
at some epoch t is e (t), then the independence between the quency of the device whose autocorrelation function is a
signal and the noise in each channel implies that delta function. By analogy with the argument above, we
would call this white frequency noise.
^ e ~ t ! e ~ t1 t ! & 5 ^ e 2 ~ t ! & d ~ t ! . ~4! The frequency fluctuations in the oscillator result in time
Since the autocorrelation of e (t) is a delta function, its fluctuations as well. If we started the device at some time 2T
power spectrum is constant for all frequencies, and measure- in the past, and if the instantaneous frequency at any time is
ment noise of this type is often called white phase noise for y( t ), then the time difference at some epoch t between this
this reason. ~Although white phase noise is often thought of oscillator and a second noiseless but otherwise identical de-
as a Gaussian variable, this is not necessarily true. While the vice would be given by
properties described above are necessary for the statistical
distribution of the phase noise to be Gaussian, they are not X~ t !5 E
2T
t
y ~ t !dt. ~5!
sufficient to guarantee this. Nevertheless, although the indi-
vidual noise-induced offsets may or may not be Gaussian, The integration of the instantaneous frequency offset to pro-
consecutive estimates of their mean will exhibit a Gaussian duce a time difference has an important consequence for the
distribution as a result of the central limit theorem of statis- power spectrum of the resulting time fluctuations. If the fluc-
tics.! tuations in y( t ) are dominated by white frequency noise,
Note that Eq. ~4! is only an approximation—we cannot then its spectral density is a constant at all frequencies ~keep-
make t either very small or very large and so the autocorre- ing in mind the comment above that the infinities implied by
lation is in fact bounded at small times by the finite band- this formal description are not a problem because of the fi-
width of the measurement system and at long times by the nite limits both in time and in frequency of a real-world
length of the experiment. A real-world spectrum will there- situation!. That is, the Fourier expansion of y( t ), the fre-
fore be not quite ‘‘white’’ for very short or very long mea- quency offset of the oscillator as a function of time, is given
surement intervals but this is usually not important. The fi- in terms of the Fourier frequency V by
E
nite response time of the overall system limits the
importance of very high-frequency noise, and other sources y ~ t !5 Y ~ V ! e iV t dV, ~6!
of noise always become important at longer averaging times.
If the white phase noise is superimposed on a constant where
frequency offset between the two devices, then consecutive
u Y ~ V ! u 5C ~7!
measured time differences will scatter uniformly about a line
whose slope is given by this offset frequency. Because of and C is a constant. Keep in mind that the variable y repre-
this uniform scatter, the frequency difference between the sents the real fractional frequency difference between two
two oscillators can be estimated using the usual least-squares nominally identical physical oscillators. Unfortunately, this
machinery. Although the white phase noise will contribute to quantity is not a constant. The expression y( t ) expresses the
the uncertainty of this estimate, it will not introduce a bias— variation in y in the time domain as a function of epoch t,
frequency estimates computed from different blocks of data while Y (V) expresses this variation in the Fourier frequency
will scatter uniformly about the true frequency offset be- domain as a function of Fourier frequency V. Also note that
tween the two devices. The statistical distribution of these Eq. ~7! cannot be literally correct for a true noise process.
frequency estimates can be calculated using the standard We will discuss this point in more detail below.
methods of propagation of error. Using Eqs. ~5!, ~6!, and ~7!, we can see that the power
spectral density of X(t) is not white but rather varies as 1/V 2
with an amplitude determined by C in Eq. ~7!. Thus there is
a fundamental difference between white phase noise and
D. Frequency noise processes
white frequency noise. The former results in time-difference
Although white phase noise arises in the measurement fluctuations that still have a white spectrum, while the inte-
process and is not due to the oscillator itself, there are also gration implied by the relationship between time and fre-
noise sources within the oscillator that we must consider. As quency causes white frequency noise to have a ‘‘red’’
we discussed above, an atomic clock ~or any other oscillator Fourier-frequency dependence when its effect on time-
for that matter!, is comprised of two basic components: a difference measurements is considered. This is an important
frequency reference ~such as an atomic transition or a vibra- conclusion because it means that we can distinguish between
tional mode in a piece of quartz! and an oscillator that inter- frequency noise and phase noise by examining the shape of
rogates this reference frequency and is locked to it by means the Fourier spectrum of the time-difference data. ~The form
of some kind of feedback loop. As with all circuits, there will of the dependence on Fourier frequency distinguishes the
inevitably be some noise in this feedback loop ~due to shot two cases—not the magnitude of the spectral density itself.!
noise in the number of atoms that are interrogated or Johnson Unlike the white phase noise case, if there is a determin-
noise in the circuit resistors!, and this noise will produce istic frequency offset between two devices whose noise spec-
jitter in the lock point of the oscillator. In the best of all tra are characterized by white frequency noise, it is no longer
possible cases, this noise will not be correlated with the sig- true that the time differences between the devices will scatter
nals in the feedback loop. Using exactly the same argument uniformly about a simple straight line determined from this
as above, this noise will produce jitter in the output fre- offset. A least-squares fit of a straight line to these time
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2576 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
differences will not result in a statistically robust estimate of Since the aging and the frequency specify the time evo-
this frequency offset for this reason. ~We mean by this that lution of the frequency and the time, respectively, white fre-
the mean difference between these estimates and the under- quency aging is sometimes called random-walk frequency
lying deterministic frequency offset is zero, and that the noise in the literature and white frequency noise is some-
mean and variance are characteristics of the data set and are times called random-walk phase noise.
independent of the sample size or of which portion of the There is another way of thinking of the differences
data is used for computing the estimate.! However, consecu- among the types of noise processes. If a time-difference mea-
tive frequency estimates ~i.e., the first difference of the time surement process can be characterized as being limited by
differences normalized by the time interval between them! pure white phase noise, then there is an underlying time dif-
are distributed uniformly about the mean frequency offset, ference between the two devices. The measurements scatter
and the average of these first differences will produce a sta- about the true time difference, but the distribution of the
tistically robust estimate of the deterministic frequency offset measurements ~or at worst the mean of a group of them! can
with a well-defined variance. be characterized by a simple Gaussian distribution with only
These results illustrate a more general conclusion, two parameters: a mean and a standard deviation. We can
namely that there is no single optimum method for estimat- improve our estimate of the mean time difference by averag-
ing clock parameters. The best strategy depends on the un- ing more and more observations, and this improvement can
derlying noise type. Unfortunately, there is no simple opti- continue forever in principle. There is no optimum averaging
mum strategy for some types of noise ~flicker processes, for time in this simple situation—the more data we are prepared
example!, and the best we can do is an approximate analysis to average the better our estimate of the mean time difference
whose predictions are not statistically robust. will be. As we discussed above, a least-squares fit of a
It is clear that these ideas could be extended to higher- straight line to these time differences can be used to compute
order derivatives. If we had an oscillator in which the fre- a statistically robust estimate of the frequency difference be-
quency aging, d( t ), had uncorrelated fluctuations about a tween the two devices.
~possibly nonzero! mean value, then, by analogy with Eq. The situation is fundamentally different for a measure-
~5!, the frequency at any time would be given by ment in which one of the contributing clocks is dominated by
E
zero-mean white frequency noise. Now it is the frequency
y ~ t ! 5y ~ 0 ! 1 d~ t !dt. ~8! that can be characterized ~at least approximately! by a single
parameter—the standard deviation. If the time difference be-
Using the same argument as above, this white frequency ag- tween two identical devices at some epoch is X(t), we would
ing would result in fluctuations in the frequency difference estimate the time difference a short time in the future using
between this clock and an identical noiseless one that varied
X ~ t1 t ! 5X ~ t ! 1y ~ t ! t , ~9!
as 1/V 2 relative to the spectral density of the aging itself.
The time differences between the two devices would have a where y(t) is the instantaneous value of the white frequency
spectrum that varied as 1/V 4 relative to the aging, since the noise, and
aging must be integrated once to compute the frequency and
a second time to compute the time difference. Again, we ^ y ~ t ! & 50, ~10!
could distinguish the different contributions by examining
the shape of the Fourier spectrum of the time differences. ^ y 2~ t ! & 5 e 2. ~11!
Using the same sort of argument as above, when the Since y(t) has a mean of 0 by assumption, the time differ-
noise spectrum of the device is dominated by white fre- ence at the next instant is distributed uniformly about the
quency aging, we cannot make a statistically robust estimate current value of X, and the mean value of X(t1 t ) is clearly
of the frequency of a device by averaging the consecutive X(t). In other words, for a clock whose performance is
frequency estimates, since those estimates are no longer dis- dominated by zero-mean white frequency noise, the opti-
tributed uniformly about a mean value. The best we can do mum prediction of the next measurement is exactly the cur-
now is to estimate a mean value for the frequency aging. rent measurement with no averaging. Note that this does not
The previous discussion used the power spectrum of the mean that our prediction is that
time differences as the benchmark, but it is also common in
the literature to find a discussion in which the relative depen- X ~ t1 t ! 5X ~ t ! , ~12!
dencies on Fourier frequency are described with respect to
the power spectrum of the frequency fluctuations. The rela- but rather that
tionship between frequency and phase specified in Eq. ~5! is ^ X ~ t1 t ! 2X ~ t ! & 50, ~13!
unchanged, of course—only the benchmark spectrum is dif-
ferent. White phase noise would then be characterized as which is a much weaker statement because it does not mean
having a relative dependence proportional to the square of that our prediction will be correct, but only that it will be
the Fourier frequency in this case. A white frequency-aging unbiased on the average. This is, of course, the best that we
process would be described as having a relative dependence can do under the circumstances. The frequencies in consecu-
on Fourier frequency of 1/V 2 , where both of these spectra tive time intervals are uncorrelated with each other by defi-
would be evaluated relative to the power spectrum of the nition, and no amount of past history will help us to predict
frequency fluctuations. what will happen next. The point is that this is the opposite
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2577
extreme from the discussion above for white phase noise, ing the two-sample or Allan variance. Suppose we have mea-
where the optimum estimate of the time difference was ob- sured the time difference between two clocks at times t 1 and
tained with infinite averaging of older data. t 2 5t 1 1 t . If we want to predict the time difference at a time
Clearly, there must be an intermediate case between t 3 5t 2 1 t , then we would use the two measurements to es-
white phase noise and white frequency noise, where some timate the frequency difference between the two clocks dur-
amount averaging would be the optimum strategy, and this ing the first interval
domain is called the ‘‘flicker’’ domain. Physically speaking,
X ~ t 2 ! 2X ~ t 1 !
the oscillator frequency has a finite memory in this domain. y 125 ~14!
Although the frequency of the oscillator is still distributed t
uniformly about a mean value of 0, consecutive values of the and we would combine this estimate with the measurement
frequency are not independent of each other and time differ- at t 2 to estimate the value at time t 3 :
ences over sufficiently short times are correlated. Both the
frequency and the time differences have a short-term X̂ ~ t 3 ! 5X ~ t 2 ! 1y 12t 52X ~ t 2 ! 2X ~ t 1 ! , ~15!
‘‘smoothness’’ that is not characteristic of a true random
where we have assumed ~for lack of any better choice! that
variable. Flicker phase noise is intermediate between white
the frequency in the interval between t 2 and t 3 is the same as
phase noise and white frequency noise, and we would there-
the value in the previous interval. If this is not the case, the
fore expect that the spectral density of the time difference
prediction error will be proportional to the difference be-
data would vary as the inverse of the Fourier frequency rela-
tween the actual frequency in the interval between t 2 and t 3 ,
tive to the spectral density of the phase fluctuations them-
which is y 23 , and our estimate of it, which is that it was the
selves.
same as the average frequency over the previous interval
The same kind of discussion can be used to define a
flicker frequency noise that is midway between white fre- e 5X ~ t 3 ! 2X̂ ~ t 3 ! 'y 232y 12 . ~16!
quency noise and white aging ~or random walk of fre-
quency!. The underlying physical effect is the same, except Expressed in terms of time difference measurements, the pre-
that now it is the frequency aging that has short-period cor- diction error would be proportional to
relations. We could think of a flicker process as resulting X ~ t 3 ! 22X ~ t 2 ! 1X ~ t 1 !
from a very large number of very small jumps—not much ~17!
happens in the short term because the individual jumps are t
very small, but the integral of them eventually produces a and the mean-square value of this quantity, usually desig-
significant effect. The memory of the process is then related nated as s 2y ( t ), is called the two-sample ~or Allan! variance
to this integration time. for an averaging time of t. ~The variance is actually defined
Data that are dominated by flicker-type processes are as one half of this mean-square value so that it will be con-
difficult to analyze. They appear deterministic over short pe- sistent with other estimators when the input data are charac-
riods of time, and there is a temptation to try to treat them as terized by white phase noise.!
white noise combined with a deterministic signal—a strategy If the frequency of the oscillator is absolutely constant
that fails once the coherence time is reached. On the other over these time intervals, then the expressions in Eqs. ~16!
hand, they are not quite noise either—the correlation be- and ~17! are nonzero only because of the white phase noise
tween consecutive measurements provides useful informa- that contributes to the measurements of X @and indirectly
tion over relatively short time intervals, and short data sets through Eq. ~14! to our estimates of y#. The numerators of
can be well characterized using standard statistical measures. Eqs. ~14! or ~17! are independent of t in this case so that the
However, the variance at longer periods is much larger than Allan variance for measurements dominated by white phase
the magnitude expected based on the short-period standard noise will vary as 1/t 2 .
deviation. If the frequency of the oscillator varies with time, then
The finite-length averages that we have discussed can be the numerator in Eq. ~14! will have a dependence on t. At
realized using a sliding window on the input data set. This is least when t is sufficiently small, this dependence must con-
simple in principle but requires that previous input data be tain only non-negative powers of t, because it must be that
stored; an alternative way of realizing essentially the same
transfer function is to use a recursive filter on the output X ~ t1 t ! 2X ~ t ! →0 as t →0. ~18!
values. This method is often used in time scales, since these The Allan variance must therefore decrease more slowly than
algorithms are commonly cast recursively. Both the recur- 1/t 2 in this case and may actually increase with t if the
sive and nonrecursive methods could be used to realize av- expansion of the numerator in Eq. ~17! contains a sufficiently
erages with more complicated transfer functions. These large power of t.
methods are often used in the analysis of data in the time The Allan variance measures predictability in the rather
domain.21 narrow sense of Eq. ~15!. However, it does not make a clear
distinction between oscillators that have stochastic frequency
noise and those that may have deterministic frequency varia-
E. Two-sample Allan variance
tion ~a constant frequency aging, for example!, even though
The ability to predict the future performance of a clock such a deterministic parameter might be included in a modi-
based on past measurements can be quantified by consider- fied version of Eq. ~15!. It also does not give any information
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2578 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
about frequency accuracy—arbitrarily large frequency off- tral density can be approximated by a series of straight-line
sets will make no contribution to Eqs. ~16! or ~17!. segments on a log–log plot in this case, and the approxima-
A constant frequency offset will result in time differ- tion that only a single type of noise dominates the spectrum
ences that increase linearly with time, and a constant fre- in each one of the corresponding Fourier frequency/
quency aging results in a quadratic dependence of the time averaging time domains is well justified in practice. We have
differences. It is tempting to remove either ~or both! of these already mentioned the five most common noise processes in
effects using standard least-squares methods before comput- clock data, and their properties are summarized in Table I.
ing the Allan variance. As we have discussed above, least- This relationship between Allan variance, spectral den-
squares estimators may produce biased estimates if they are sity, and noise type has three important limitations. The first
used on data containing appreciable variance due to random- is that the slope of the simple Allan variance that we have
walk or flicker processes; the resulting Allan variance esti- defined above cannot distinguish between flicker phase
mates are likely to be too small as a result. modulation and white phase modulation—in both cases the
slope is 1/t 2 . This defect can be remedied by defining a
F. Relationship between the Fourier decomposition modified Allan variance22 ~MOD AVAR! which replaces the
and the Allan variance time differences in Eq. ~17! @or the frequencies of Eq. ~16!#
The Allan variance and Fourier decomposition views of with averages over a number of adjacent intervals. The prin-
the noise process are always valid in principle, but they are cipal advantage of this modification is that it provides a way
particularly useful if the measurements are dominated by a of distinguishing between white phase noise and flicker
single noise process, since both parameters have particularly phase noise, as is shown in Table I. A relative of the modi-
simple forms in these special cases. Specifically, if the slope fied Allan variance is the time variance, usually called
of the power spectral density as a function of frequency on a ‘‘TVAR’’ and designated s 2x ( t ). 23 It is an estimate of the
log–log plot has a slope of a, that is if the spectral density of time dispersion resulting from the corresponding frequency
the frequency fluctuations is such that variance, and it is normalized so as to be consistent with the
conventional definition of the variance when the data are
S y ~ V ! 5h a V a , ~19! dominated by white phase noise. It is defined by
where h a is a constant scale factor and a is an integer such
that t2
s 2x 5 mod s 2y ~ t ! . ~23!
3
22< a <1, ~20!
then the two-sample ordinary Allan variance we have de- It has the same statistical properties as the modified Allan
fined above @as one half of the mean square of Eq. ~17!#, has variance, but it is somewhat easier to interpret in applications
a slope of m when plotted as a function of averaging time on where time dispersion rather than frequency dispersion is of
a log–log plot. That is, primary concern. Using TVAR, it is also somewhat easier to
identify the onset of the domain in which white frequency
s 2y ~ t ! 5b m t m ~21! noise dominates the spectrum by simply looking at the plot
with of the variance as a function of averaging time. This is be-
cause it is easier to see the point at which the slope of TVAR
m 52 a 21. ~22!
changes from 0 to 11 than it is to identify the point at which
Measuring the slope of the Allan variance as a function of the slope of MOD AVAR changes from 22 to 21 ~see Table
averaging time therefore determines the slope of the spectral I!.
density and the noise type. This is a substantial advantage, The second limitation is that the relationship between the
because computing the Allan variance is usually a much sim- Allan variance and the slope of the power spectral density
pler job than computing the full power spectrum. This pro- depends on the fact that the input data are really noise and do
cedure is most useful when the spectrum is dominated by a not have any coherent variation. The most common excep-
single noise type in each range of Fourier frequency and tion to this assumption is the case of deterministic frequency
averaging time, which is quite often the case for many clocks aging—the relatively constant drift in the frequency of hy-
and oscillators. Both the Allan variance and the power spec- drogen masers, for example, or a nearly diurnal frequency
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2579
fluctuation in quartz–crystal oscillators driven by changes in other words, the Fourier decomposition of a white noise pro-
the ambient temperature. Both the Allan variance and the cess is not a smooth straight line but is itself a random vari-
spectral density function are well defined in these cases, but able whose statistical properties can be derived from the
the simple relationships we have described are no longer original time series. ~Although more sophisticated analyses
valid. The spectral density function of a periodic signal is, of are usually used, the spectral densities are linear combina-
course, not a smoothly varying function but a peak at the tions of the time series data, and the standard methods of
corresponding frequency, while the Allan variance oscillates propagation of error can be applied to estimate the variance
at the period of the perturbation and no longer has a simple of the spectral density.!
monotonic slope either. It is common practice to reduce the variance in the Fou-
The third limitation is that the variance is a statistical rier coefficients by averaging a number of adjacent values.
parameter, and it does not give any information on worst- The simplest method is a straight average which effectively
case performance. This problem is not unique to the Allan weights the adjacent points equally, but more complicated
variance, of course—all statistical estimators share this fail- ‘‘triangular’’ functions ~such as 1/6, 2/3, 1/6 or 0.23, 0.54,
ing. It is possible, however, to define an uncertainty in an 0.23! are also used. These averages provide a trade-off of
Allan variance estimate;24,25 while this does not speak to stability against frequency resolution. The practice of aver-
worst-case performance, it does provide some information aging adjacent estimates using a sliding window is math-
on the width of the distribution function—that is on the ematically equivalent to convolving the Fourier coefficients
spread that might be expected from measurements on an en- with the square or triangular-shaped function derived from
semble of identical oscillators. It is also important to remem- the weights. This convolution in turn is equivalent to multi-
ber that both mod avar and tvar are computed using an av- plying the original time series by a ‘‘windowing’’ function
erage over adjacent frequency estimates, and the resulting that varies smoothly from a maximum at the midpoint of the
magnitudes are not the same as what would be obtained from data set to zero at the end points.
the same device measured without that averaging. The choice of windowing function is particularly impor-
tant when analyzing time series that have significant power
G. Uncertainty of the Allan variance at discrete frequencies, since the windowing process can af-
fect both the magnitude and shape of these ‘‘bright lines.’’
The uncertainty in the Allan variance estimate for a More generally, the windowing operation can cause ambigu-
given averaging time is proportional to the number of differ- ities when the width of the windowing function ~in frequency
ences that contribute to it. If the entire data set spans a time space! contains statistically significant variation in the power
T, for example, then an estimate using a time difference of t spectral density of the data, since quantitative estimates of
will be comprised of about T/2t estimates. This number the spectral density will depend on the shape of the window
clearly decreases with averaging time, so that the uncertainty in this case.29 Analyzing a data set that contains a number of
in each estimate must increase as the averaging time in- discrete spectral components requires considerable care in
creases. In the limit we will have only one estimate for an the choice of a windowing function because a poor choice
averaging time of t 5T/2, and we will have almost no infor- may result in frequency aliases—spurious peaks in the spec-
mation at all about its uncertainty as a result. To make mat- tral density corresponding to frequencies that are not really
ters worse, consecutive estimates of the frequency have a present in the time series. It is common practice to remove
correlation that depends on the noise type, so that simple- such bright lines ~using simple least squares, for example!
minded estimates of the number of degrees of freedom can before estimating the spectral density. This process is called
be misleading even for small averaging times. ‘‘pre-whitening.’’
The summation of Eq. ~17! that is required to compute
the Allan variance can be considered to be a convolution of
the input data set X(t j ) with the series 1, 22, 1, 0,... . If the I. Other noise estimators
series of time differences are dominated by noise processes
~rather than deterministic effects! then it is likely that the A number of applications require an estimator that pro-
series will be at least approximately stationary. If this is the vides some measure of the worst-case ~rather than the aver-
case, then we can replace the ordinary convolution with a age! performance of an oscillator. The most common appli-
circular one,26 and this change removes the difficulty in the cation is the use of oscillators to synchronize the
calculation for large time differences. If the data are not sta- multiplexors in communications systems. This hardware is
tionary initially, then it may be possible to pre-whiten them used to combine data from several circuits into one higher-
using digital filter techniques or by removing a low-order speed data stream using time-division multiplexing ~in which
polynomial first.27 each input circuit is given a periodic time slice!. This com-
bined stream is de-multiplexed at the other end of the line
using receiving hardware that routes the data to the appro-
H. Uncertainty using Fourier decomposition
priate output line based on the same time-slice logic. A cer-
The uncertainties in Fourier frequency estimates of noise tain amount of frequency difference between the hardware at
processes are discussed in the literature.28 Although Eq. ~7! the end points can be accommodated by ‘‘slip’’ buffers, but
~and other Fourier estimates of noisy series! accurately char- these have finite capacity, and the maximum time dispersion
acterizes the power spectral density of white noise in the and frequency difference between the two end stations must
mean, there are significant uncertainties in the estimates. In be controlled to prevent overflowing these buffers.30
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2580 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
This requirement leads naturally to the concept of ance ~or its square root, the standard deviation! combined
MTIE—the maximum peak-to-peak time-interval error in with the number of degrees of freedom in the estimate. If
some observation time, T.31 A number of different versions needed, covariances are also specified. These terms have
of MTIE exist—in some cases the reference for the measure- their conventional statistical definitions.
ment is the input frequency to the network element and in Type B uncertainties are also characterized by a variance
other cases it is an external standard reference frequency. ~or its square root, the standard deviation!. The variances are
In addition to specifying T, measurements of MTIE usu- approximations to the corresponding statistical parameters,
ally also specify some observation period for the measure- the existence of which is assumed. If needed, covariances are
ments of the time dispersion. This is often implemented as a included in the same way.
sliding window of width t ,T. The value of MTIE for the The combined uncertainty is computed in the usual way
complete data set is the maximum of the values for each of as the simple sum of the variances, and the combined stan-
the windows. dard deviation is the square of the sum of the squares of the
The variation of MTIE with t can provide some insight contributing standard deviations. ~The guide recognizes that
into the character of the frequency offset between the two in some applications it may be appropriate to multiply the
clocks. A static frequency offset, for example, produces a combined uncertainty by an additional ‘‘coverage’’ factor;
value of MTIE that increases strictly linearly with t. The this factor must be specified when it is used.!
converse is not necessarily true, however. Random-walk pro- The procedure of simply summing the variances to com-
cesses can produce almost the same effect. pute the combined uncertainty is justified if the contributions
Another estimator that is often used in telecommunica- are statistically independent of each other—an assumption
tions work is ZTIE. This estimator uses averages of the first that is easy to make but hard to prove. In principle, it is
difference of the time data @Eq. ~14!# instead of the second possible to deal with data where the observations are not
differences @Eq. ~17!# as in the Allan variance. As with statistically independent of each other, but this is difficult in
MTIE, the value of the estimator is the maximum value of practice because the correlation coefficients are usually not
the statistic over the averaging window. well known. Furthermore ~as we discussed above!, summing
Both ZTIE and MTIE are sensitive to static frequency variances ~or averaging data! that are dominated by flicker or
offsets, whereas AVAR ~in all its variations! is not. A clock random-walk noise processes usually does not produce esti-
with an arbitrarily large and constant frequency offset is a mates that are statistically robust. As a specific case in point,
‘‘good’’ clock by AVAR standards—while its time disper- while adding more and better clocks to EAL certainly helps
sion will eventually become very large, its performance is matters, the accuracy of TAI is still evaluated by using data
predictable. The same clock is not nearly so useful by MTIE from primary frequency standards.
or ZTIE standards because its static frequency offset will If the combined uncertainty is computed as the simple
produce repeated overflows of the slip buffers when it is used sum of variances of all types then identifying a particular
as the reference for a communications switch. The fact that contributor as being of type ‘‘A’’ or ‘‘B’’ does not affect the
these overflows are regular and predictable is beside the final result, since only the magnitude of each contribution is
point. In addition, MTIE is more sensitive to transients—it significant. However, it is useful in that it makes it easier to
will give a better estimate of the worst-case performance understand how the result was obtained. Many time and fre-
because of this, but it can also become saturated by a single quency applications depend more on the stability ~or predict-
event that will mask its response to the remainder of the data. ability! of a device or on its frequency accuracy. In these
applications the question of type A versus type B is less
important than the magnitude of the Allan variance ~for
IX. REPORTING THE UNCERTAINTY OF A stability/predictability! or MTIE ~for accuracy!.
MEASUREMENT
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2581
the one-way propagation delay of either signal can be delay is dominated by a diurnal effect that can be canceled
significant—perhaps as much as tens of milliseconds for almost perfectly by averaging the data for 24 h.37 This is an
WWV and a few milliseconds for WWVB. Short-wave sig- important example of the difference in the requirements for
nals tend to travel as both a ground wave and a sky wave, time and frequency distribution. Although the uncertainty in
which is reflected from the ionosphere ~possibly multiple the transit time of a WWVB signal might be as large as 1 ms,
times!. The interference between these two signals and the the variation in the delay averaged over 24 h is probably
variations in the height of the ionospheric reflecting layer almost 1000 times smaller than this value.!
result in significant variation in the path delay. These prob- The relatively simple models used to correct transmis-
lems are less serious for lower-frequency signals, but even sions from high-frequency radio stations ~like WWV! are not
these signals show a strong diurnal variation in the path de- very accurate, and they are only adequate for users with very
lay, which tends to vary most rapidly near sunrise and sun- modest timing accuracy requirements on the order of 1 ms.
set. This diurnal variation can be almost completely removed This is a result of the inadequacy of the models, however,
by averaging the observations for 24 h ~leaving a much and is not a fundamental limitation of the technique. Trans-
smaller uncertainty on the order of microseconds! but many missions from the global positioning system ~GPS! satellites,
oscillators have too much flicker or random-walk frequency for example, can be corrected to much higher accuracy using
noise to support averaging for this long a time. Even with much more complex models of the satellite orbit and the
this level of averaging, the uncertainties are too large for ionospheric and atmospheric delays.38 In addition, multiple-
more demanding applications, and other techniques have wavelength methods ~to be described below!, can also be
been developed to meet these needs. Roughly speaking, the used to estimate the delay through dispersive media.
design goal for any distribution system is to transmit the data The fundamental limitation to modeled corrections is of-
to a user without degrading the accuracy or stability of the ten due to the fact that the adequacy of the model cannot be
source. This goal is not easy to meet now, and it will grow evaluated independently of the transmission itself. The GPS
even more difficult in the future as clocks continue to im- system is a notable exception because of the redundancy pro-
prove. vided by being able to observe several ~almost always at
Although time and frequency are closely related quanti- least 4 and sometimes as many as 12! satellites simulta-
ties, there are important differences between them, and these neously, and this technique is especially powerful when they
differences affect how they are distributed and what uncer- are located at different azimuths with respect to the antenna.
tainties are associated with these processes. Time distribu- When modeling is not sufficiently accurate to estimate
tion is, of course, directly affected by the delay in the trans- the path delay, users must measure the delay with an uncer-
mission channel between the clock and the user, and tainty small enough for the task at hand. There are a number
uncertainties in this delay enter directly into the error budget. of methods that can be used for this task, and we will de-
A frequency, on the other hand, is a time interval ~rather than scribe them in the following sections. Many applications use
an absolute time! so that the uncertainty in its distribution is more than one of them to estimate different components of
limited more by temporal fluctuations in the transmission the delay. The full accuracy of time transfer between a global
delay rather than by its absolute magnitude. In other words, positioning system satellite and an earth station, for example,
the time delay of a channel must be accurately known if that can only be realized by combining a model of the tropo-
channel is to carry time information, while the delay need spheric delay with a measurement of the delay through the
only be stable if frequency information is to be transmitted. ionosphere using multiple wavelength dispersion. Using GPS
This distinction is not of purely academic interest; many signals to transfer time between two earth stations often ben-
common channels have large delays that are difficult to mea- efits from common-view subtraction as well.
sure but which change slowly enough with time that they
A. Delay measurements using portable clocks
may be thought of as constant for appreciable periods; such
channels are clearly better suited to frequency-based appli- The simplest method for measuring the time delay along
cations. a path is to synchronize a clock to the transmitter and then to
The simplest way of dealing with the transmission delay carry it to the far end of the path where its time is compared
is to ignore it, and many users with modest timing require- to the signals sent from the transmitter over the channel to be
ments do just that. Many users of the time signals broadcast calibrated. Many timing laboratories used this technique in
by short-wave stations ~NIST radio stations WWV and the past. This method is limited in practice by the finite fre-
WWVH, for example!, fall into this group. A somewhat quency stability of the portable clock while it is being trans-
more sophisticated strategy is to approximate the actual de- ported along the path and by uncertainties in a number of
lay using some average value—perhaps derived from some relativistic corrections that may have to be applied. These
simple parameter characterizing the distance between the corrections will probably be needed if the transport velocity
user and the transmitter.36 Users of the signal from NIST is high enough that time dilation must be considered, if the
radio station WWVB, for example, often do not measure the path is long enough that the Sagnac correction is important,
transmission delay, but rather model it based on average at- or if points along the path are at different heights above the
mospheric parameters and the physical distance between geoid so that gravitational frequency shifts must be included.
themselves and the transmitter. ~Many users of WWVB are While this method can be used to calibrate a static transmis-
interested in frequency calibrations and do not model the sion delay, it is too slow and cumbersome to evaluate the
delay at all. Instead they use the fact that the variation in the temporal fluctuations in the delay through real transmission
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2582 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
media ~due to changes in temperature or other environmental tainty is much greater because of variations in the reciprocity
parameters, for example!. of the path—that is, in the lack of equality of the delays over
Even though various electrical or electromagnetic distri- the inbound and outbound circuits. Note that jitter in the
bution methods are usually faster and more convenient, the overall delay itself does not produce a problem if it does not
portable clock method remains quite useful because it tends affect the reciprocity, but it may be difficult to exploit this
to be affected by a very different set of systematic uncertain- fact in a real-time system, which often uses the measured
ties than the other methods to be described below, and it may delay of a previous round-trip measurement to correct the
provide estimates of these uncertainties that are difficult to current one. While this strategy is simple in principle, it re-
achieve in any other way.39 sults in a system that is sensitive both to jitter in the overall
delay itself and to its asymmetry.
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2583
are often limited in practice more by fluctuations in the de- received copies using an exhaustive brute-force method.
lays through the hardware than by problems in the path itself. When this procedure succeeds, it locks the local clock to the
time broadcast by the satellite modulo 1 ms ~with an offset
due to the transmission delay!, and allows the receiver to
XI. GLOBAL POSITIONING SYSTEM begin searching for the 50 b/s navigation data. Once this data
The signals from the global positioning system satellites stream has been found, the receiver can read the satellite
are widely used for high-accuracy time transfer using both time from the navigation message, use it to construct a local
one-way and common-view methods. A similar system, op- copy of the P code, and begin looking for a cross-correlation
erated by Russia, is called GLONASS. We will first describe peak using this faster code. This cross-correlation process
the satellites and the characteristics of the signals they broad- fails when a nonauthorized receiver processes the Y code,
cast, and we will then discuss how these messages can be and these receivers usually drop back automatically to using
used for time and frequency distribution. the C/A code only in this case.
The GPS system uses 24 satellites in nearly circular or- The pseudo-random codes are Gold codes,48 which are
bits whose radius is about 26 600 km. The orbital period of implemented using shift registers that are driven at the chip-
these satellites is very close to 12 h, and the entire constel- ping frequency of the code. The details of the design of the
lation returns to the same point in the sky ~relative to an code generators, the correlation properties of the codes, and
observer on the earth! every sidereal day ~very nearly 23 h the relationship between the codes and the satellite time are
56 m!. discussed in the GPS literature. A particularly clear exposi-
The satellite transmissions are derived from a single os- tion is in Ref. 49.
cillator operating at a nominal frequency of 10.23 MHz as The navigation message contains an estimate of the time
measured by an observer on the earth. In traveling from the and frequency offsets of each satellite clock with respect to
satellite to an earth-based observer, the signal frequency is the composite GPS time, which is computed using a
modified by two effects—a redshift due to the second-order weighted average of the clocks in the satellites and in the
Doppler effect and a blueshift due to the difference in gravi- tracking stations. This composite clock is in turn steered to
tational potential between the satellite and the observer. UTC~USNO!, which is in turn steered to UTC as computed
These two effects produce a net fractional blueshift of about by the BIPM. The time difference between GPS time and
4.4310210 ~38 ms/day!, and the proper frequencies of the UTC~USNO! is guaranteed to be less than 100 ns, and the
oscillators on all of the satellites are adjusted to compensate estimate of this offset, which is transmitted as part of the
for this effect, which is a property of the orbit and is there- navigation message, is guaranteed to be accurate to 25 ns. In
fore common to all of them. In addition to these common practice, the performance of the system has almost always
offsets, there are two other effects—the first-order Doppler substantially exceeded its design requirements.
shift and a frequency offset due to the eccentricity of the GPS time does not incorporate additional leap seconds
orbit, which vary with time and from satellite to satellite. beyond the 19 that were defined at its inception; the time
The primary oscillator is multiplied by 154 to generate differs from UTC by an integral number of additional leap
the L1 carrier at 1575.42 MHz and by 120 to generate the seconds as a result. This difference is currently 13 s, and will
L2 carrier at 1227.6 MHz. These two carriers are modulated increase as additional leap seconds are added to UTC. The
by three signals: the precision ‘‘P’’ code—a pseudo-random accuracy statements in the previous paragraph should there-
code with a chipping rate of 10.23 MHz and a repetition fore be understood as being modulo 1 s. The number of leap
period of 1 week, the coarse acquisition ~‘‘C/A’’! code with seconds between GPS time and UTC is transmitted as part of
a chipping rate of 1.023 MHz and a repetition rate of 1 ms, the navigation message, but is not used in the definition of
and a navigation message transmitted at 50 bits/s. Under nor- GPS time itself. Advance notice of a future leap second in
mal operating conditions, the C/A is present only on the L1 UTC is also transmitted in the navigation message.
carrier. The GLONASS system is similar to GPS in general con-
The satellites currently in orbit also support ‘‘anti- cept. In its fully operational mode, there will be 24 satellites
spoofing.’’ ~AS! When the AS mode is activated ~which is orbiting in 3 planes 120° apart. At the present time ~October,
almost always at the present time!, the P code is encrypted 1998! there are 13 operational satellites. The GLONASS P
and the result is called the Y code. The P and Y codes have code has a chipping frequency of 5.11 MHz and a repetition
the same chipping rate, and this code is therefore often re- period of 1 s; the C/A code chipping frequency is 511 kHz
ferred to as P(Y ). and repeats every millisecond. All satellites transmit the
Each satellite transmits at the same nominal frequencies same codes, but the carrier frequencies are different
but uses a unique pair of codes which are constructed so as to ~FDMA—frequency division multiple access!. These fre-
have very small cross correlation at any lag and a very small quencies are defined by
auto-correlation at any nonzero lag ~CDMA—code division f 1 5160210.5625k MHz ~24!
multiple access!. The receiver identifies the source of the
signal and the time of its transmission by constructing local and
copies of the codes and by looking for peaks in the cross
f 2 5124610.4375k MHz, ~25!
correlation between the local codes and the received ver-
sions. Since there are only 1023 C/A code chips, it is feasible where k is an integer that identifies the satellite. These rather
to find the peak in the cross correlation between the local and widely spaced frequency allocations complicate the design of
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2584 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
the front end of the receiver—especially the temperature sta- the L1 signal then it usually corrects for the ionospheric
bility of the filters that are usually needed to reject strong delay using a parameter transmitted in the navigation mes-
out-of-band interference. On the other hand, cross talk be- sage.! The result is an estimate of the time difference be-
tween signals from the different satellites is much less of a tween the local clock and GPS or GLONASS system time.
problem; interference from a single-frequency source is also The time difference between the local clock and satellite
less of a problem with GLONASS, since the interfering sig- system time can be used directly to discipline a local oscil-
nal is less likely to affect the signals from the entire constel- lator or for other purposes. However, these signals have de-
lation. terministic fluctuations due to imperfections in the broadcast
In the original design, k was a unique positive integer for ephemeris which is used by the receiver to correct the pseu-
each satellite. The resulting frequencies interfered with radio dorange. In addition, signals from the GPS constellation are
astronomy measurements, and a set of lower frequencies usually also degraded by the frequency dithering of selective
with negative values of k will be phased in over the next few availability. These effects are attenuated or removed in com-
years. In addition, satellites on opposite sides of the earth mon view; if the stations are not too far apart, the common-
may be assigned the same value for k. This would not view method also attenuates fluctuations in the delay due to
present a problem for ground-based observers, since they the troposphere and the ionosphere since the paths to the two
could not see both of them simultaneously, but satellite- stations often have similar delays and the fluctuations in
based systems may require additional processing ~similar to these delays often have significant correlation as well.
the tracking loops in GPS receivers! to distinguish between To facilitate time transfer among the various national
the two signals. laboratories and timing centers, schedules for observing the
As with GPS, the satellites transmit the offset between signals from the GPS and GLONASS satellites are published
the satellite clock and GLONASS system time, but the rela- by the International Bureau of Weights and Measures
tionship to UTC is not as well defined. Signals from the ~BIPM!,50 and all timing laboratories adhere to them as a
GLONASS satellites are therefore usually used in common minimum. ~Not all timing laboratories can receive GLO-
view, since the time of the satellite clock cancels in this NASS signals at the present time.! Other users often observe
procedure. this schedule as well so as to facilitate the evaluation of their
The signals from the GPS constellation are usually af- local clocks in terms of national time scales. The comparison
fected by anti-spoofing ~AS!, which encrypts the P code, and with national time scales may be required for legal traceabil-
by selective availability ~SA!, which dithers the primary os- ity purposes, and using GPS time directly may not satisfy
cillator and degrades the frequency stability of the signal. these legal requirements.
~The implementation of SA may also dither parameters in the The need for common-view schedules arose initially be-
ephemeris message which produce a dither in the computed cause the first GPS receivers could only observe one satellite
position of the satellite. This aspect of SA is usually at a time. Newer receivers can observer many satellites si-
switched off.! From the point of view of a nonauthorized multaneously ~as many as 12 for the newest units!, and this
user, the primary effect of AS is to deny access to the P makes it possible to eliminate formal schedules in principle.
code, since it can only be processed by a keyed receiver Each station records the difference between its local clock
when it is transmitted in its encrypted Y-code form. An au- and GPS or GLONASS system time using all of the satellites
thorized user would see this encryption as authenticating the in view, and the time difference between any two stations is
transmission, since only a real satellite can produce valid computed after the fact by searching the data sets from both
encrypted messages. Hence the name anti-spoofing. The de- stations for all possible common-view measurements. In ad-
scription of the GLONASS P code has not been published dition to eliminating the need for formal tracking schedules,
and is officially reserved for military use, but the GLONASS this method can be used to construct common views between
C/A code is not degraded. two stations that are so far apart that they never have a sat-
The operation of the receiver is basically the same in ellite in common view. A receiver in Boulder, Colorado, for
either system. The receiver generates a local copy of the example, might have a common view link to Europe using
pseudo-random code transmitted by each satellite in view, one satellite and a simultaneous second link to Asia using a
and finds the peak in the cross correlation between each of second one. These data could be combined to produce a
these local copies and the received signal. The local code common-view link between Europe and Asia that was inde-
generator is driven from a local reference oscillator, and the pendent of the intermediate clocks in Boulder and on the
time offset of the local clock needed to maximize the cross satellites.
correlation is the pseudo range—the raw time difference be-
tween the local and satellite clocks. Using the contents of the XII. MULTIPLE FREQUENCY DISPERSION
navigation message, the receiver corrects the pseudo range
for the travel time through the atmosphere, for the offset of In practice, it is difficult to arrange things so that the two
the satellite clock from satellite system time, etc. ~If the re- receivers in a common-view arrangement are exactly equally
ceiver can process both the L1 and L2 frequencies, then the distant from the transmitter, and common-view data must be
receiver can also estimate the delay through the ionosphere corrected for this path difference. If the transmission path is
by measuring the difference in the pseudo ranges observed through the atmosphere or some other refractive medium,
using the L1 and L2 frequencies. The details of this calcu- then knowing the difference in the physical path lengths is
lation are presented below. If the receiver can process only not enough—it is also necessary to know the refractivity of
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2585
the medium ~the difference between the actual index of re- easy way to estimate the correction due to the troposphere,
fraction and unity!. If the refractivity is dispersive ~that is, if and its effect is usually ignored at the present time as a
it depends on the frequency of the carrier used to transmit the result. This is not adequate when the phase of the carrier is
information!, and if the dispersion is separable into a part used as in geodetic measurements, and more sophisticated
that depends on the carrier frequency and a part that is a estimates of the troposphere are required in those
function of the physical parameters of the path, then it is applications.51!
often possible to compute the refractivity by measuring the Although the time difference arising from the dispersion
dispersion, even when both of these parameters vary from may be small, it is multiplied by the refractivity divided by
point to point along the path. If the length of the physical the dispersion, a ratio that is larger than unity. ~The value is
path is L then the transit times measured using two different about 1.5 for the L1/L2 frequencies used by GPS.! The re-
carrier frequencies f 1 and f 2 will be sult is that the ionospheric refractivity makes a significant
E E
contribution to the time of flight of a signal from a satellite;
@ 11n 1 ~ r !# L F~ f 1!
t 15 dr5 1 G ~ ... ! dr ~26! ignoring this contribution is only adequate for common-view
c c c observations over short base lines where the effect cancels in
and the subtraction. Furthermore, the noise in the dispersion
measurement (t 1 2t 2 ) is amplified by this same ratio, so that
t 25 E @ 11n 2 ~ r !#
c
L F~ f 2!
dr5 1
c c
E G ~ ... ! dr, ~27!
the price for removing the refractivity of the path is an in-
crease in the noise in the measurements by roughly a factor
of 3 compared to a single-frequency measurement using only
where the quantities n 1 (r) and n 2 (r) are the refractivities at L1.
the two frequencies at a coordinate point r along the path.
These refractivities are decomposed into a known function F
that depends only on the frequency of the carrier and a sec- XIII. EXAMPLES OF TIME TRANSFER USING GPS
ond known function G that depends only on the parameters
of the medium such as its density, pressure, temperature,... . In this section we illustrate the principles of time transfer
The integrals must be evaluated along the transmission path. using signals from the GPS satellites. ~The same principles
This would be difficult or impossible for a real-world path would apply to time transfer using GLONASS signals.52!
through the atmosphere, since the required parameters are Since the entire GPS constellation usually has AS enabled,
not known and cannot be easily measured. Since the integral the P code is not available to most time-transfer users. ~Even
of G is independent of the carrier frequency, we can elimi- when AS is turned on, the P code can still be used for dis-
nate G in terms of the difference in the two measured transit persion measurements. A cross correlation of the P-code sig-
times nals on L1 and L2 can be performed without knowing the
actual P code—the fact that it is the same on both carriers is
E G ~ ... ! dr5
c ~ t 1 2t 2 !
L @ F ~ f 1 ! 2F ~ f 2 !#
. ~28!
enough.! All of the time-difference results in this section are
based on measurements using the C/A code or the carrier
itself. In addition, the code-based data were acquired with
Substituting Eq. ~28! into Eq. ~26! single-frequency receivers, which cannot measure the L1-L2
L F~ f 1! L n1 dispersion and must estimate or measure the contribution of
t 1 5 1 ~ t 1 2t 2 ! 5 1 ~ t 1 2t 2 ! . ~29! the ionosphere using some other method. This is generally
c F ~ f 1 ! 2F ~ f 2 ! c Dn
done using the parameters in the navigation message trans-
The first term on the right-hand side is the time delay due to mitted by the satellite, although other techniques exist.53
the transit time of the signal along the path of physical length The C/A code cross-correlation process results in a se-
L, and the second term is the additional delay due to the ries of ticks with a period of 1 ms, and decoding the ephem-
refractivity of the medium. Note that this second term de- eris message allows the receiver to identify the ticks in this
pends only on the measured time difference between the sig- stream that are associated with integer GPS seconds.54 When
nals at the two frequencies and on the known frequency de- the arrival times of these ticks are corrected for the offset of
pendence of the refractivity. It does not depend on the the satellite clock from GPS time, for the path delay and for
physical length L itself or on the details of the spatial depen- the other effects described above, the result is a string of 1
dence of G, and we would have obtained exactly the same Hz pulses that are synchronized to GPS time. ~In order to
result if only a small piece of the path L 8 ,L was actually apply these corrections, the receiver must parse the data
dispersive. This is an important conclusion, because it means stream transmitted by the satellite.! The time differences be-
that a receiver can estimate the delay through the ionosphere tween these pulses and the corresponding ticks from a local
by measuring the difference in the arrival times of identical clock are then measured using a standard time interval
messages sent using the L1 and L2 frequencies. Note that counter. Single-channel receivers can only perform this mea-
this method fails if the medium is refractive without being surement on one satellite at a time, but newer ‘‘all in view’’
dispersive—that is, if the index of refraction differs from receivers can measure the time difference between the tick of
unity but is independent of the carrier frequency. This is the the local clock and GPS time using up to 8 or 12 satellites
situation with the effect of the troposphere on microwave simultaneously. In the simplest version of this scheme, the
frequencies near 1 GHz, and other means must be found to receiver measures the time difference between the local
model the refractive component of its delay. ~There is no clock and a composite 1 Hz tick. This tick is positioned
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2586 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
FIG. 1. The time difference between a local cesium clock and GPS time FIG. 2. The time differences measured in a short-base line common-view
measured using data from satellite 9 only. The observed variation is due to experiment between two multichannel GPS receivers. The two receivers
the effects of SA on the GPS signal; the contributions of the clock and were connected to the same clock and the antennas were about 1 m apart.
receiver to the variance are negligible on this scale. The average time difference between the two receivers was computed using
all satellites in view.
using data from all of the satellites that are being tracked, but
the better receivers can provide independent time differences shown in the figure is due to the difference in the lengths of
between the local clock and GPS time using each satellite. the two cables between the receivers and the common clock.
This is done by providing a single 1 Hz tick to the time These data have an amplitude of 1.7 ns rms, and they are
interval counter together with a data stream that specifies the much closer to white phase noise—at least at short periods.
offset between the time of the physical tick and GPS time as The data shown in Fig. 2 represent a best-case result: the
computed using data from each satellite being tracked. Un- reference clock is the same for both units and the antennas
less the two stations are very close together, this latter capa- are very close together so that almost all propagation effects
bility is necessary for common-view measurements, as we are the same for both of them and therefore cancel in the
will show below. common-view difference. Likewise, the fluctuations in the
Figure 1 shows an example of these time-difference clock itself are also common to both channels. However, the
measurements. The plot shows the time difference between a advantages of common view are still significant even over
local cesium standard and GPS time measured using data longer base lines and different clocks. Figure 3 shows data
from a single satellite—SV 9. The time of the local clock is from a common-view experiment between NIST in Boulder,
offset from GPS time by about 500 ns, and this explains the Colorado, and the United States Naval Observatory ~USNO!
mean value of the data that are displayed. The time disper- in Washington, DC. Even though these stations are about
sion in the figure is totally due to the effects of SA on the
GPS signal—the contributions due to receiver noise and to
the frequency dispersion of the clock are negligible on this
scale. These data are not white phase noise at periods shorter
than a few minutes, and the mean time difference over a few
minutes is not stationary. Averages over longer periods begin
to approximate white phase noise, but the contribution of SA
to the variance does not become comparable to the noise in
the clock itself until the data are averaged for several days.
~Unfortunately, averaging white phase noise only decreases
the standard deviation of the mean at a rate proportional to
the square root of the averaging time, and this improvement
does not even begin until we enter the domain where the
variance is dominated by white phase noise.!
As we pointed out above, the common-view procedure
cancels the contribution of the satellite clock to the variance
of the time differences, and we would therefore expect that
this procedure would cancel SA. This expectation is con- FIG. 3. The time differences between a receiver at NIST in Boulder, Colo-
firmed by the data in Fig. 2. This figure shows the time rado, and a second receiver at USNO in Washington, DC, about 2400 km
away. As in the previous figure, the average time difference between the two
differences between two identical receivers whose antennas
stations was computed using all satellites in common view at each time. The
were located about 1 m apart. The two receivers used the members of the common-view ensemble change slowly during the observa-
same local clock as a reference; the mean time difference tion period.
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2587
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2588 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
a sensitivity of the antenna or its connecting cable to the not as simple as they first appear to be for reasons to be
diurnal fluctuation in the ambient temperature. The fre- outlined below, and the full advantage of using several
quency signatures of the two effects differ by only 1 c/y, and clocks can be realized only when they are combined in a
are therefore difficult to separate. Diurnal temperature fluc- more complex time-scale procedure.
tuations are much less repeatable from day to day, so that The problem with simple majority voting ~or with com-
these effects cannot be modeled so effectively in this way. paring the time of a ‘‘master’’ clock with a number of sub-
The tracking schedules published by the BIPM include sidiary ones! is that the stability of the frequency of almost
an advance of 4 min every day so that tracks on consecutive any commercial clock ~even a cesium-based device! is better
days are observed in approximately the same multipath en- than its accuracy by three orders of magnitude or more.
vironment. This method does not remove or correct for the Comparisons among clocks are almost always realized using
effect of multipath—it simply converts it from a fluctuating time-difference measurements, and the large and stable fre-
signal to a nearly static offset which is different for each quency differences between nominally identical devices re-
satellite. This is not necessarily a desirable result, since there sult in linearly increasing time differences as well. ~Many
are already satellite-to-satellite offsets due to errors in the clocks have deterministic frequency aging as well, and this
ephemerides. further complicates the issue by adding a quadratic depen-
dence to the time-difference data.! It is no longer a simple
XIV. DELAY MEASUREMENTS IN VIRTUAL CIRCUITS matter to detect a hardware failure in this situation. It would
be nice to think that the deterministic frequency offsets of an
We now turn to a discussion of time distribution using
ensemble of clocks would scatter uniformly about the true SI
communications networks. Everything that we have said so
frequency, but there is no reason to believe that this is true in
far about measuring the transmission delay assumed implic-
general. In any case, it is much less likely to be true for a
itly that the path between the transmitter and receiver had the
small group of devices. Thus neither a simple average of the
well-defined characteristics that we intuitively associate with
times nor of the frequencies of the ensembles of devices has
a physical circuit or a well-defined path through the atmo-
good statistical properties. The machinery of a time scale
sphere. These ideas are still useful in characterizing most
paths through the atmosphere, but transmission circuits on provides a natural way of dealing with this problem.
cable and fiber networks are increasingly packet switched A second advantage of a time scale is in providing a
rather than circuit switched. This means that there is no end- framework for combining a heterogeneous group of devices
to-end physical connection between the transmitter and the so as to achieve a time scale that has the best properties of
receiver. Packets are sent from the transmitter to the receiver each of them. For example, a time scale might be used to
over a common high-bandwidth physical channel, which is combine the excellent short-term stability of a hydrogen ma-
shared among many simultaneous connections using time- ser with the better long-term performance of a cesium-based
domain multiplexing ~or some other equivalent strategy!. device. Finally, a time scale can provide a flywheel for
Each packet passes through a series of routers which read the evaluating the performance of primary frequency standards
destination address and direct the packet accordingly. Data or other devices such as prototype standards. Such evalua-
appear to enter at one end and emerge at the other, but there tions can be done using only one clock, of course, but there
need not be a physical connection between the two. In an is no way of identifying the source of a glitch in that con-
extreme case, consecutive packets may travel over very dif- figuration.
ferent paths because of congestion or a hardware failure. As we will show below, the ensemble-average time and
This variability complicates any attempt to measure the ef- frequency have many of the properties of a physical clock. In
fective delay of the circuit. particular, the ensemble time and frequency must be set ini-
These problems are likely to become more important as tially using some external calibration. Conceptually, this is
multiplexed digital circuits become more widely used. A done by measuring the time difference and frequency offset
number of strategies have been proposed for dealing with between each member clock and the external standard, al-
these problems. A shorter path is likely to have smaller vari- though other methods in which a single member clock is
ability as well, since it would probably have fewer routers calibrated and the other clocks are then compared to it are
and gateways, and this suggests that the absolute delay might often used in practice. These initial time and frequency off-
be used as an estimate of its variability. It is tempting to sets can be used to define the corresponding parameters of
assert that the absolute magnitudes of these problems will be the ensemble or an additional administrative offset can be
smaller on higher-speed channels, but this is by no means a added to them. This offset might be used to steer the scale to
sure thing. a primary standard, for example. As we discussed above,
TAI has been steered from time to time in this way using
data from primary frequency standards.
XV. ENSEMBLES OF CLOCKS AND TIME SCALES
A. Clock models
You will probably want to have more than one clock if
you are serious about time keeping. Although a second clock As we discussed above, one of the problems with not
can provide a backup when the first one definitively fails, using a formal time-scale algorithm but simply comparing
three clocks provide the additional ability to detect partial clocks using time differences, is the deterministic frequency
failures or occasional bad data points using a simple differences between nominally identical devices. A natural
majority-voting scheme. Unfortunately, such schemes are way to deal with these deterministic effects is to include
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2589
them in the definition of a model of how we expect the time B. AT1 algorithm
differences to evolve, and to compare our observations with
Several versions of the AT1 algorithm have been used at
the predictions of the model in order to detect problems.
NIST for many years. Although it works quite well with real
Since the physical devices that make up an ensemble change
data, I discuss it in detail not because it is theoretically op-
as old devices fail and new ones are added as replacements,
timum but because it is relatively simple to implement. In
it is customary to define the model parameters of each mem-
addition, it illustrates the general ideas that underlie a num-
ber clock with respect to the ensemble average rather than
ber of other algorithms including ALGOS, which is used by
with respect to any particular master device. The goal is to
the BIPM to compute EAL and TAI.
make the characterization of each of the clocks in the en-
The starting point for a discussion of AT1 is the obser-
semble independent of the parameters of the others.
vation that it is difficult to make robust estimates of the
The deterministic performance of most clocks can be
deterministic frequency aging for most clocks. Even at the
very well characterized using three parameters: x, the time
difference, y, the frequency difference, and d, the frequency largest values of t ~where the effects of frequency aging are
aging, where each of these parameters characterizes mea- greatest!, these deterministic effects are usually swamped by
surements between the clock and the ensemble. These defi- the flicker and random-walk frequency noises in the device.
nitions imply that the ‘‘ensemble’’ is a clock too, although Almost the only exceptions are the very best hydrogen ma-
we have not yet specified how it can be read, set or charac- sers, whose frequency noise is so small that robust estimates
terized. In fact, different algorithms address this problem in of the deterministic frequency aging are possible, and rather
different ways. poor quartz oscillators ~which are not normally used as mem-
Using these three parameters, we can construct a model bers of ensembles anyway! whose frequency aging is so
of the deterministic performance of the clock. Assuming that large that it is readily observable even for small averaging
we have the values of the three model parameters at some times. Thus the assumption that d(t)50 for all t is consistent
time t, the values at some later time t1 t are given by with the data for almost all clocks, except for the best hydro-
gen masers. For those clocks we would use a constant for the
x ~ t1 t ! 5x ~ t ! 1y ~ t ! t 1 21 d ~ t ! t 2 , aging that is determined outside of the ensemble algorithm.
A typical value for such a maser would be of order
y ~ t1 t ! 5y ~ t ! 1d t , ~30! 10221 s21.
If we model the frequency aging as a noiseless constant
d ~ t1 t ! 5d ~ t ! . value then this is equivalent to assuming that the frequency
fluctuations are dominated by white noise processes. This is
It is possible to extend these equations to include higher- a reasonable assumption provided that t is not too large. The
order terms, but this is not necessary ~or even useful! for maximum value of t that still satisfies this requirement varies
most devices. These model equations assume that there is no with the type of device; a typical value for a cesium oscilla-
deterministic variation in the frequency aging—an assump- tor would be less than a day or two. On the other hand, it
tion that is justified more by experience than by rigorous must not be made too small either, because otherwise the
analysis. time difference data will be dominated by the white phase
In addition to these deterministic parameters, our previ- noise of the measurement process. ~This is not necessarily a
ous analysis of oscillator performance would lead us to ex- bad thing, but taking full advantage of the data in this case
pect that our data would also have stochastic contributions to requires a change in strategy. This point is discussed below
the variation of each of these parameters. For simplicity, we in connection with techniques used for transmitting time us-
could add terms for white phase noise, white frequency ing the Internet.! As specific examples, the NIST time scale,
noise, and white frequency aging. These names are chosen to which uses measurement hardware with a noise floor of sev-
parallel the deterministic description in Eqs. ~30!—we can eral ps uses t 52 h, while the BIPM computation of AL-
also use the alternative names we have defined above. ~As GOS, which uses the much noisier GPS system to measure
we discussed above, the variance in each parameter has con- the time differences between laboratories, uses t 55d.
tributions both from its own white noise term and from all of As we pointed out above, the initial values of the time
the noise terms below it. The combined phase or frequency and frequency of the ensemble must be known a priori. The
noises are very definitely not white.! epoch associated with these initial values is t 0 . The algo-
The next step would be to compare the predictions of our rithm then continues with periodic measurements of the time
model equations with measurements and to test if the differ- differences between each clock and one of them, which is
ences are consistent with the assumed model for the noise. designated as the reference clock or working standard. These
Our measurements usually consist of pairwise time differ- measurements are made at epochs t 1 ,t 2 ,...,t n . These mea-
ences between our physical clocks, and there is no obvious, surements are usually equally spaced in time, but this is more
universal prescription for using these data to evaluate a for convenience than for any formal requirement of the al-
model of the type specified in Eqs. ~30!. There are two dif- gorithm.
ficulties: the first is that the time difference data are ambigu- If at some epoch t k the time difference and frequency
ous to within an overall additive constant and the second is offset between the jth clock and the ensemble were x j (t k )
that we have not specified a model for how the ensemble and y j (t k ), respectively, then AT1 models the time differ-
clock itself is measured or how it evolves. ence at some later epoch t k11 as being given by
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2590 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
x̂ j ~ t k11 ! 5x j ~ t k ! 1y j ~ t k ! t , ~31! follows, since it simplifies the notation and emphasizes the
fact that clock m is not a special clock with respect to the
where ensemble—only with respect to the time-difference hard-
t 5t k11 2t k . ~32! ware.! If we assume that the fluctuations in the time-
difference measurements are not correlated with each other,
Note that the right side of Eq. ~31! is an estimate based on the N estimates of the time of clock m with respect to the
the two model parameters of the clock using a simple linear ensemble are characterized by a well-defined mean, which
model. As we discussed above, this is appropriate for all but we can compute using
the best hydrogen masers. We would use the first of Eqs. 30
with a constant value for d for these devices. This constant
frequency aging does not really have any effect at this point.
x m5 (j w j x̂ mj . ~34!
Since the algorithm normally operates with a fixed interval The weights w j are arbitrary at this point, subject only to the
between computations, the effect of adding frequency aging usual normalization requirement that
is to add a constant offset to each of the predictions. Such an
offset can be absorbed into the starting values without loss of
generality. (j w j 51. ~35!
The mth clock in the ensemble is the working standard.
It is characterized using the same model as any other clock If the clocks are all identical then a simple assumption would
@i.e., Eq. ~31! with j5m#. The working standard is not nec- be to set all of the weights equal to 1/N, so that Eq. ~34!
essarily the clock with the best statistical properties—since it reduces to a simple average, but this is usually not the opti-
may be awkward to change this designation in a running mum choice in a real ensemble.
ensemble, it is often the clock that is likely to be the most Equation ~34! can be equally well thought of as defining
reliable or to live the longest. the time of the ensemble with respect to the working stan-
At each epoch we measure the time differences between dard, clock m, computed by averaging the independent esti-
the working standard and each of the other clocks using a mates obtained from the time difference between clock m
time interval counter or any of the other methods we de- and each of the member clocks. The time of clock j with
scribed above. These measured time differences are X jm (t), respect to the ensemble can then be calculated by combining
where j takes on all possible values except for m. Combining Eq. ~34! with the measured time difference X jm :
the predicted time of each clock with respect to the ensemble x j 5X jm 1x m . ~36!
using Eq. ~31! with the corresponding measured time differ-
ence between that clock and the working standard, we can These values are then used as the starting point for the next
estimate the time difference between clock m and the en- measurement cycle.
semble via clock j using The difference between Eq. ~31!, which predicted the
time of clock j with respect to the ensemble based on its
x̂ mj ~ t k11 ! 5x̂ j ~ t k11 ! 2X jm ~ t k11 ! , ~33! previous state, and Eq. ~36!, which gives the value of clock j
based on the data from the current measurement cycle, is the
where the superscript j on the left side indicates that the prediction error, e j . The average of this prediction error is a
estimate has been made using data from clock j. Note that measure of the quality of the clock in predicting the time of
Eq. ~33! could equally well have been thought of as a defi- the ensemble. The average of this value can be used to obtain
nition of the time of the ensemble with respect to physical an estimate for the weight to be used for clock j in Eq. ~34!
clock m estimated using data from clock j. Also note that, as
expected, Eq. ~33! reduces to Eq. ~31! when j5m since C
w j5 , ~37!
X mm (t)[0. Thus Eq. ~33! applies to the working standard ^ 2j &
e
just as it does to any other clock in the group.
If Eq. ~31! is an accurate model of the time of clock j, where C is chosen so that the weights satisfy the normaliza-
then Eq. ~33! is an unbiased estimate of the time of the tion condition, Eq. ~35!. Defined in this way, the weights are
working standard with respect to the ensemble since the de- related to the TVAR statistic, which we defined above. As in
terministic time and rate offset of clock j have been properly our previous discussions, a good clock is a predictable
estimated. Note that clock m may have a deterministic time clock—not necessarily one that has a small deterministic fre-
or frequency offset with respect to the ensemble—there is no quency offset, because we can always remove that offset
requirement that its parameters be the same as those of the using Eq. ~31!.
ensemble. Its statistical parameters are also not important for The dispersion in the prediction error can also be used to
this procedure to work. test for a clock failure. That is, on each cycle we could ex-
If there are N clocks in the ensemble, there are N21 amine the statistic
independent time-difference measurements X jm , and the e 2j 2 ^ e 2j &
same number of equations identical to Eq. ~33! with all pos- ~38!
^ e 2j &
sible values of j except j5m. The previous state of clock m
is used to predict its time with respect to the ensemble using and decide that clock j has a problem if this statistic is
Eq. ~31!. ~Alternatively, we could use Eq. ~33! for this pur- greater than some threshold s max . We might drop clock j
pose, setting X mm 50. We adopt this latter strategy in what from the ensemble calculation in this case and repeat the
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2591
computation @Eq. ~34!# with its weight set to 0. If the statistic model parameters computed on the previous step are com-
is close to s max , we could choose to decrease the weight of bined with the current measurements—older data need not
the clock gradually so that there is no abrupt step in perfor- be saved for this computation.
mance exactly at s max . The AT1 time scale currently used at Several versions of this algorithm have been used at
NIST sets s max54, and linearly decreases the weight of a NIST for many years.64,65 The AT1 time scale provides op-
clock starting when its average prediction error reaches 75% timum performance when the noise in the measurement sys-
of this value. tem and the clocks is no more divergent than white phase
This procedure is a more formalized version of a simple noise. Unfortunately, clocks exhibit flicker and random-walk
majority-voting scheme. Its advantage is that the systematic frequency noise at longer term, and the scale itself tends to
differences between the clocks have been removed by the become dominated by this type of noise at longer periods ~on
model equations and therefore do not affect the error- the order of months!. The ALGOS algorithm used to com-
detection process. Furthermore, the averaging procedure is pute EAL ~and from that TAI and UTC! tends to have the
applied to quantities that have no first-order deterministic same problem; as we discussed above, the solution in that
variation. The working standard enters into these calcula- case is to steer TAI using data from the primary frequency
tions exactly as any other member of the ensemble, and its standards. This limitation has been addressed in the design of
data can still be rejected by the prediction-error test just dis- AT2.66
cussed even though it continues to play a central role in the The AT1 algorithm has other limitations. It cannot deal
time-difference measurements. with a device ~such as a primary frequency standard! that
If the clock frequencies were noiseless and constant then operates only occasionally. We would like to be able to char-
that would be the end of the story. The time-difference mea- acterize such a device using only a well-defined frequency—
surements would be composed of deterministic frequency its time with respect to the ensemble is not interesting or
and time offsets which would be accounted for by the model important. In fact, since it operates only intermittently, its
equations, and the residual time difference data would be time with respect to the ensemble may not be well defined.
characterized by pure white phase noise, which would be As we mentioned above, the data from a primary frequency
optimally attenuated by the averaging procedure. Unfortu- standard can be incorporated into the ensemble frequency by
nately, the frequency of the clock is characterized by white administrative adjustment as is done with TAI.
frequency noise at intermediate times and by flicker and The AF167 procedure presents another solution to this
random-walk fluctuations at longer times. As we showed problem. Using an algorithm analogous to AT1, it estimates
above, these frequency fluctuations do not produce time dif- the frequency ~rather than the time! of the working standard
ferences whose variation is characterized as pure white phase using the measured frequency differences between it and the
noise, and the average in Eq. ~34! is no longer completely other members of the ensemble. Data from a primary fre-
justified. quency standard can be incorporated into this scale in a natu-
These frequency fluctuations can be accommodated by ral way, and the fact that the standard operates only intermit-
replacing the static frequency parameter y j in Eq. ~31! with a tently is not a problem. Since it is quite similar to AT1 in its
parameter that is allowed to evolve with time. In order to basic design, it will also become dominated by flicker and
compute this dynamic frequency, the short-term frequency random-walk processes at longer periods—probably on the
estimates order of 1–2 yr.
In addition to ALGOS, there are a number of other al-
x j ~ t k11 ! 2x j ~ t k ! gorithms that are similar to AT1. These algorithms have
f j ~ t k11 ! 5 ~39! been widely discussed in the literature.68 Other algorithms
t
are based on Kalman filters.69,70 This formalism has some
are averaged using a simple exponential low-pass filter to advantages in dealing with clocks that operate irregulary; it
produce the frequency estimate used in the model equation can also provide more robust definitions of the ensemble
time and frequency than is possible using AT1 or its rela-
tives. On the other hand, Kalman filter algorithms often have
y j ~ t k ! 1 a f j ~ t k11 ! startup transients before the filter enters the steady-state re-
y j ~ t k11 ! 5 , ~40!
11 a gime. These transients can result in instabilities when the
membership in the ensemble changes frequently.
where the time constant a is set to the transition at the upper A fundamental problem with many time scales is the
end of the domain in which white frequency fluctuations way the weights of the member clocks are calculated. The
dominate the spectrum of the clock noise. The time constant AT1 algorithm and its relatives assign a weight to each clock
parameter is actually a dimensionless constant less than 1; based on its prediction error—a good clock is one that is
the effective time constant is t/a. The value of the time predictable over the time interval between measurement
constant is set to attenuate white frequency noise at interme- cycles. ~We have already seen this principle in the discussion
diate periods; the gain of the filter approaches unity at very of the Allan variance.! However, a clock whose weight is
long periods. As we discussed above, the primary reason for increased because of this will tend to pull the ensemble
using an exponential filter of this type is that it can be real- through Eq. ~34!. A small increase in weight thus tends to
ized using the same kind of iteration as is used for the other produce positive feedback, and a clock that starts out a little
parameters of the model. At each step in the process the bit better than the other members of the ensemble can
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2592 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
quickly take over the scale. The problem, of course, is that nificant long-term frequency aging combined with very small
Eqs. ~31! and ~36! are correlated, so that the prediction error short-term prediction errors. The very good short-term per-
is always systematically too small. The correlation clearly formance means that these clocks will be given a very high
gets worse as the weight of a clock increases. weight in an AT1-type procedure, and this high weight will
This correlation can be addressed by administratively then adversely affect the long-term stability of the scale due
limiting the maximum weight any clock can have, either to to the frequency aging of these devices. One solution is to
an absolute maximum value ~as in ALGOS! or to a maxi- incorporate a constant frequency-aging term that is deter-
mum percentage of the scale ~as in the NIST implementation mined administratively using procedures outside of the scale
of AT1!. It is also possible to estimate the magnitude of the computations ~using data from primary frequency standards,
correlation and correct for it.71 for example!. Another solution is to modify the algorithm to
These methods limit the maximum weight that any clock include explicitly different weights for short-term and long-
can have in the ensemble, even if that large weight would term predictions.
have been deserved on statistical grounds. The reason is that The time and frequency of the scale itself can be made
one of the motivations for using an ensemble is safety and almost totally insensitive to a change in the membership of
robustness in case of a failure. An ensemble whose time is the ensemble, provided that the initial parameters of a new
effectively computed using only one clock will have a seri- clock are accurately determined. These initial parameters are
ous problem when that high-weight clock fails or has a data usually determined by running a new clock in the scale with
glitch. These problems can be more easily addressed in its weight administratively kept at 0.
scales that are not computed in real time ~such as TAI, which One of the practical difficulties in running an ensemble
is computed about one month after the fact!. It is much easier is that there is no physical clock that realizes its time or
to recognize a problem retrospectively and to administra- frequency. This deficiency must be remedied if the ensemble
tively remove a clock from the ensemble when later data time is used to control a physical device. One solution is to
shows that it was starting to behave erratically. steer a physical clock so that its time and frequency track the
Since the weights are normalized using Eq. ~35!, any corresponding ensemble parameters. There are a number of
method that limits the maximum weight that a good clock different ways of realizing this steering: the algorithm can
can have implicitly gives poorer clocks more weight than emphasize time accuracy or frequency smoothness, but not
they deserve. The price for keeping good clocks from taking both. An algorithm that emphasized time accuracy would use
over the scale is therefore a scale that is not as good as it large frequency adjustments and time steps to keep the
could be because poorer clocks have more weight than they steered clock as close as possible to ensemble time. An al-
deserve based on statistical considerations. gorithm that emphasized smoothness would limit the magni-
A similar effect happens when a member clock has a tude of frequency adjustments and never use time steps. The
large prediction error and is dropped from the ensemble as a choice between the two extremes ~or the relative importance
result. The contribution that this clock would have made to of accuracy and smoothness in a mixed solution! cannot be
the ensemble is divided among the remaining members, with specified a priori—the choice depends on how the data are
the result that the ensemble time ~and, indirectly, its fre- to be used. Historically, steering algorithms at NIST have
quency! may take a step when this happens. This can be tended to place a higher value on frequency smoothness than
especially serious in a small ensemble with one clock that is on accuracy, with the inevitable result that the time disper-
significantly poorer than the others. If one of the good clocks sion of UTC~NIST! is larger than it would be if time accu-
is dropped because it fails a test based on its prediction error, racy were emphasized more strongly.
its weight will be distributed among the remaining members.
If the weights of the other clocks are already limited by the
XVI. SYNCHRONIZING A CLOCK USING TIME
administrative constraints that we discussed above, the result SIGNALS
is to transfer the weight of this good clock to the poorest
clock in the ensemble. Finally, we turn to the question of how best to use a time
The methods used to calculate the frequency and the signal that we receive from a distant source of uncertain
weight of each clock are effectively trying to separate the accuracy over a channel whose delay may be variable. In the
observed variance in the time difference data into determin- simplest distribution system, the receiver is a simple,
istic and stochastic components. This separation, which is ‘‘dumb’’ device. Its internal state ~whether time or fre-
implicitly based on a Fourier decomposition of the variance, quency! is simply reset whenever a signal is received. The
can never be perfect. These methods tend to have the most signal may be corrected for transmission delay or some other
difficulty distinguishing between deterministic frequency ag- systematic correction might be used, but the important point
ing and random-walk frequency noise because the two have is that the previous state of the receiver makes little or no
similar ‘‘red’’ Fourier decompositions. This is the reason contribution to the process. This may be the best strategy if
AT1 and similar time scales often do not include a term for the uncertainty in the receiver state is so high as to make it
deterministic frequency aging into the model equations for effectively worthless, but it is important to recognize that
the time differences of the clocks—it is very difficult to es- this is not universally true, and that the state of the receiver
timate these parameters in the presence of the random-walk before the new information is received also contains infor-
noise in the data. This can be a serious problem in an en- mation that may be useful.72
semble containing hydrogen masers, which tend to have sig- The performance of the receiver clock can be character-
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2593
ized in terms of deterministic and stochastic components. availability, which is a pseudo-random process that looks
The deterministic aspects of its performance can be used to like noise in the short term. In the usual state of affairs, the
predict the data that will be received from the transmitter at frequency of the local clock is more stable than the time
some future time. This prediction is never perfect, of course, difference measurements would indicate—in other words
because the deterministic components may not be known these fluctuations in the measured time difference do not
precisely, because even if they are known at one instant they imply a corresponding fluctuation in the frequency of the
are likely to evolve in time, and because the stochastic com- local clock. The optimum strategy is clearly to leave the
ponents of the model are always present and add noise to all local clock alone in the short term, since its free-running
measurements. Although this prediction might not be as ac- stability is better than the time-difference data would indi-
curate as we would like, it nevertheless can often provide an cate. This implies some form of averaging of the measure-
important constraint on the received data. ments, with the details depending on the application.
The most useful situation arises when the spectrum of When the variance in the time-difference data can be
the noise in the receiver clock has little or no overlap with modeled as white phase noise, then the optimum strategy is
the spectrum of the noise in the transmission channel.73 This clearly to make a large number of rapid-fire time-difference
happy situation supports the concept of ‘‘separation of measurements and to average these data. This process can
variance’’—the ability to distinguish between a ‘‘true’’ sig- continue for as long as the variance continues to look like
nal, which should be used to adjust the state of the receiver white phase noise. For synchronization over packet networks
clock, and a ‘‘false’’ signal which is due to an un-modeled such as the Internet, it is quite common to find periods of up
fluctuation in the channel delay and which should therefore to 15–30 s when this assumption is valid, and it is a simple
be ignored. In an extreme situation, this strategy could be matter to make 50 ~or more! time-difference measurements
used to detect a failure in the transmitter clock itself. during this period. ~The same idea can be used to improve
Decisions based on separation of variance are probabi- the measurement noise in a time-scale system. In that case as
listic ~rather than deterministic! in nature, and will only be well, the short-term variance is well modeled as white phase
correct in an average sense at best. Whether or not they noise.! These data would be averaged to form a single point,
improve matters in a particular situation depends on the va- which was then used to discipline the local clock using a
lidity of the assumption of separation of variance and on the model such as Eqs. ~30! or ~31!. Although the average cost
accuracy of the various variance estimators. Since the esti- of the synchronization process depends only on the number
mators must be based on the data themselves, a certain of calibration cycles, clustering the measurements in this
amount of ambiguity is unavoidable. way may improve the performance of the synchronization
This method obviously fails in those spectral regions algorithm by optimally averaging the white phase noise of
where separation is impossible. There is no basis for decid- the measurement process. Although the average cost of the
ing whether the observed signal is ‘‘real’’ and should be used synchronization process would be the same if the measure-
to adjust the local clock or represents transmission noise and ments were spread out uniformly in time, the performance
should be ignored. If the resulting ambiguity results in unac- would not necessarily be as good because the increased in-
ceptably large fluctuations, then either the receiver or the terval would put the measurements outside of the time do-
channel must be improved. The crucial point is that neither main where the fluctuations are dominated by white phase
the receiver clock nor the transmission channel need be good noise.76 While the cost of realizing this synchronization pro-
everywhere in the frequency domain—it is perfectly ad- cess increases linearly with the number of calibration cycles
equate if they have complementary performance characteris- that are used in each group, the variance only improves as
tics, and this may in fact be the ‘‘optimum,’’ least-cost alter- the square root of this number, so that incremental improve-
native for achieving a given accuracy in the distribution of ments in the performance of the algorithm become increas-
time or frequency. ingly expensive to realize.
These considerations explain why it may not be opti- At some longer time, the variance of the averaged time
mum to use the signal received over a noisy channel ‘‘as is.’’ differences drops below the variance of the free-running
It may be possible to remove the channel noise if the re- clock, and that is the point at which some form of correction
ceived data can be used to discipline an oscillator whose to the time of the local clock is appropriate. ~Note that the
inherent noise performance can be characterized by a vari- crossover is not determined by the frequency offset of the
ance that separates cleanly from the noise added by the chan- local clock but by the stochastic noise in this parameter.
nel in the sense we have discussed. These ideas can be used Deterministic offsets may be a nuisance in practice, but they
to filter the intentional degradations imposed on the GPS can always be removed in principle and should not play a
satellites because of selective availability74 and to improve role in a variance analysis. Furthermore, if the variance of
the distribution of time on widearea networks such as the the local clock can be characterized as white frequency
Internet.75 noise, then averaging consecutive, closely spaced frequen-
These strategies can become quite complicated in detail, cies will yield a statistically robust frequency estimate. The
but all of them tend to share a number of general character- domain in which white frequency noise dominates the vari-
istics. Individual measurements of the time difference are ance is often quite small, so that frequency averaging often
usually dominated either by white phase noise arising from does not produce an appreciable improvement in the perfor-
the measurement process itself or by something like selective mance of the synchronization process.! This crossover point
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2594 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
is never absolute, however. The local clock will always con- would be required to remove diurnal and similar load-related
tribute some information to the measurement process—and fluctuations in the network delay.
even if this information is only used to detect a gross failure
of the distant clock or of the calibration channel it should
probably never be completely ignored.77
In the usual situation, the noise in the data from the XVII. DISCUSSION
distant clock continues to decrease ~relative to the local I have described some of the principles that are used to
clock! at longer averaging times, so that the weight attached design clocks and oscillators, and I have introduced the sta-
to the time-difference data in steering the local clock in- tistical machinery that is used to describe time and frequency
creases by a corresponding factor. This is certainly true for data. This machinery can be used to characterize the oscilla-
the degradation imposed on the GPS satellites by selective tors themselves and also to understand the degradation that
availability, and it is also a common feature of noise in a results from transmitting these data over noisy channels with
communications channel. Since both of these fluctuations are varying delays.
bounded in time, they make contributions to the Allan vari- Time and frequency measurements will continue to play
ance that must decrease at least as fast as the averaging time an important role in science, engineering, and life in general.
for sufficiently long intervals. Several new requirements are likely to be most important in
These considerations tend to imply relatively slow cor- the near term:
rection loops—the time constant for steering a good com- ~a! The need for higher-accuracy frequency standards for
mercial cesium standard to a GPS signal is usually a day or applications ranging from research in atomic physics
longer. The corresponding time constant for a computer and astrophysics to synchronizing high-bandwidth
clock synchronized using messages transmitted over the In- communications channels. Devices such as cesium or
ternet is usually a few hours. Although such loops maximize rubidium fountains and trapped-ion devices are already
the separation of variance and uses both the local hardware in the prototype stage, and devices that can potentially
and the calibration channel in a statistically optimum way, realize the SI second with an uncertainty of signifi-
they have problems nevertheless. The very slow response of cantly less than 10215 are already being designed.
the correction loop means that it takes a long time to settle There is no reason to believe that the trend towards
down when it is first turned on and also after any transient. A improving the accuracy and stability of frequency stan-
step of the local clock either in frequency or in time may dards will stop, or that applications for these improved
standards ~both in science and in engineering! will not
persist for quite some time before it is finally removed. The
follow their realization.
loop may never reach steady-state operation in the extreme
~b! Making use of these new devices will require improve-
case. More frequent calibrations and shorter time constants
ments in distribution methods. Methods based on the
are usually used when worst-case ~rather than long-term av- phase of the GPS carrier may be adequate—at least
erage! performance is the principal design metric. initially, but other methods will probably be needed as
Another situation which requires more frequent calibra- well. Digital transmission methods—especially meth-
tions and shorter time constants occurs when the local oscil- ods using optical fibers—are likely to become more
lator has a relatively rapid variation in frequency that cannot important for many of these applications.
be characterized as a noise process. An example would be a ~c! If these devices are to make a contribution to TA1, then
large sensitivity to ambient temperature. Such an oscillator incorporating them will also require a better under-
will tend to have large nearly diurnal frequency variations, standing of the transformation from the proper time
and much of the variation will often be concentrated near the realized by the device to coordinate time—the exact
start and end of the working day. Although the time disper- shape of the geoid, for example, especially at real-
sion averaged over 24 h will tend to be small, the maximum world laboratories, which are often located at places
dispersion during the day may be unacceptably large, and ~such as Boulder, Colorado! which have inhomoge-
relatively frequent calibrations may be required to keep the neous and inadequately documented crustal structure.
time dispersion within specified limits. ~d! Applications such as electronic commerce, digital no-
taries, and updating distributed databases need methods
The opposite problem happens when it is the channel
for simple, reliable, authenticated, and legally traceable
that has a large diurnal fluctuation in the delay ~as with
distribution of time information with moderate accu-
WWVB, for example!. The optimum strategy is clearly to
racy on the order of milliseconds. The demand for
increase the interval between calibrations so as to reduce the these services is growing exponentially, and the servers
sensitivity of the measurements to these spurious fluctua- must be protected against attacks by a significant num-
tions. This is not too hard to do in the case of WWVB, since ber of malicious users. These requirements for security
it is usually used with local oscillators that are stable enough and reliability are not unique to digital time services, of
to support the required increase in averaging time. The course, but addressing them while maintaining the ac-
analogous problem in computer networks is more difficult to curacy of the time messages presents some unique
deal with, since the clock circuits in many computers are not problems, which may not be completely solved by sim-
stable enough to support the increase in averaging time that ply scaling up the current designs.
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine 2595
ACKNOWLEDGMENT 20
J. A. Barnes, Proc. IEEE 54, 207 ~1966!.
21
G. E. P. Box and G. M. Jenkins, Time Series Analysis: Forecasting and
The work on the algorithms and statistics used for net- Control ~Holden–Day, San Francisco, 1970!.
22
work synchronization was supported in part by Grant Nos. S. R. Stein, Frequency and Time, Their Measurement and Characteriza-
NCR-941663 and NCR-911555 from the NSF to the Univer- tion, Precision Frequency Controls ~Academic, New York, 1985!, Vol. 2,
Chap. 12.
sity of Colorado. 23
D. W. Allan, M. A. Weiss, and J. L. Jesperson, Proceedings of the 45th
Annual Symposium of Frequency Control, 1991, pp. 667–678. IEEE
General References Catalog No. 91CH2965-2.
24
P. Lesage, and C. Audoin, IEEE Trans. Instrum. Meas. IM-22, 157
The following references provide more detailed informa- ~1973!.
tion about many of the subjects discussed in this article. 25
P. Lesage and T. Ayi, IEEE Trans. Instrum. Meas. IM-33, 332 ~1984!.
26
They are not specifically referenced in the text. D. A. Howe, Proceedings of the 49th Symposium on Frequency Control,
1995, pp. 321–329. IEEE Catalog No. 95CH3446-2.
27
D. A. Howe and K. J. Lainson, Proceedings of the 50th Symposium on
Applied Optimal Estimation, edited by Arthur Gelb ~MIT Press, Cambridge, Frequency Control, 1996, pp. 883–887. IEEE Catalog No. 96CH35935.
28
MA, 1992!. See especially Chaps. 4 and 9 on Kalman filters and real- R. B. Blackman and J. W. Tukey, The Measurement of Power Spectra
time control. ~Dover, New York, 1959!, pp. 15–22.
C. Hackman and D. B. Sullivan, Am. J. Phys. 63, 306 ~1995!. This resource 29
F. L. Walls, D. B. Percival, and W. R. Irelan, Proceedings of the 43rd
letter, together with a number of other papers on time and frequency Annual Symposium on Frequency Control, 1989, pp. 336–341. IEEE pub-
metrology, was reprinted in 1996 by the American Association of Phys- lication 89CH2690-6.
ics Teachers, College Park, MD. The title of the reprint volume is, 30
P. Kartaschoff, Proc. IEEE 79, 1019 ~1991!.
‘‘Time and Frequency Measurement.’’ 31
W. C. Lindsey, F. Ghazvinian, W. C. Hagmann, and K. Dessouky, Proc.
A. Leick, GPS Satellite Surveying ~Wiley, New York, 1990!. IEEE 73, 1445 ~1985!.
F. G. Major, The Quantum Beat: The Physical Principles of Atomic Clocks 32
P. Giacomo, Metrologia 18, 43 ~1982!.
~Springer, New York, 1998!. 33
P. Giacomo, Metrologia 24, 49 ~1987!.
D. B. Sullivan, D. W. Allan, D. A. Howe, and F. L. Walls, Characterization 34
Published by the International Organization for Standardization ~ISO!,
of Clocks and Oscillators, NIST Technical Note No. 1337. ~U.S. Gov- Geneva, Switzerland, 1993.
ernment Printing Office, Washington, DC, 1990!. This is a book of 35
Published annually by the International Bureau of Weights and Measures
reprints that contains some of the papers cited in the specific reference ~BIPM!, Sevres, France.
section. J. Vanier and C. Audoin, The Quantum Physics of Atomic Fre- 36
B. E. Blair, Time and Frequency Dissemination, An Overview of Prin-
quency Standards, in 2 volumes. Bristol: ~Hilger and Institute of Phys- ciples and Techniques, NBS Monograph No. 140 ~National Bureau of
ics, London, 1989!. Standards, Boulder, CO, 1974!.
37
G. Kamas and M. A. Lombardi, Time and Frequency Users Manual, NIST
Specific References Special Publication 559 ~National Institute of Standards and Technology,
Boulder, CO, 1990!.
38
W. Lewandowski, and C. Thomas, Proc. IEEE 79, 991 ~1991!.
1
D. D. McCarthy, Proc. IEEE 79, 915 ~1991!. 39
P. Wheeler, D. Clamers, A. Davis, P. Koppang, A. Kubig, and W. Powell,
2
F. L. Walls, and J. J. Gagnepain, IEEE Trans. Ultrason. Ferroelectr. Freq. Proceedings of the 26th Precise Time and Time Interval Planning and
Control 39, 241 ~1992!. Applications Meeting, 1995, pp. 51–62. This is NASA Conference Pub-
3
Le Systeme International d’Unites, 7th ed. ~Bureau International des Poids lication No. 3302, and is available from NASA Goddard Space Flight
et Mesures, Sevres, France, 1998!. Center, Greenbelt, MD 20771.
4
A. O. McCoubrey, Guide for the Use of the International System of Units, 40
J. Jesperson, in Ref. 29, pp. 186–192.
NIST Special Publication No. 811 ~NIST, Gaithersburg, MD, 1991!. 41
D. Kirchner, Proc. IEEE 79, 983 ~1991!.
5
T. J. Quinn, Proc. IEEE 79, 894 ~1991!. 42
G. de Jong and M. C. Polderman, in Ref. 39, pp. 305–318.
6
C. Thomas, BIPM Report No. 98/2, Sevres, France, 1998. 43
M. Kihara and A. lmaoka, Proceedings of the 7th European Time and
7
Annual Report of the BIPM Time Section ~Bureau International des Poids
Frequency Forum, 1993, pp. 83–88, Available from EFTF Secretarial,
et Mesures, Sevres, France, 1995!, Vol. 8, pp. 17–19.
8 Rue de l’Orangerie 8, CH-2000, Neuchatel, Switzerland.
Central Bureau of the IERS, Observatoire de Paris, Avenue de 44
J. A. DeYoung, W. J. Klepczynski, J. A. Davis, W. Powell, P. R. Pearce,
l’Observatoire, 61, 75014 Paris, France. See also Ref. 6, p. 5.
9 C. Hackmann, D. Kirchner, G. de Jong, P. Hetzel, A. Bauch, A. Soering,
R. E. Drullinger, W. D. Lee, J. H. Shirley, and J. P. Lowe, IEEE Trans
P. Grudler, F. Baumont, J. Ressler, and L. Veenstra, in Ref. 39, pp. 39–50.
Instrum. Meas. 44, 120 ~1995!; R. E. Drullinger, J. P. Glaze, J. P. Lowe, 45
L. Primas, R. Logan, Jr., and G. Lutes, in Ref. 29, pp. 202–212.
and J. H. Shirley, ibid. 40, 162 ~1991!. 46
J. Levine, M. Weiss, D. D. Davis, D. W. Allan, and D. B. Sullivan, J. Res.
10
N. F. Ramsey, Phys. Today 33, 25 ~1980!. See also the references to
Ramsey’s original papers at the end of this article. Natl. Inst. Stand. Technol. 94, 311 ~1989!.
47
11
A. Kastler, J. Opt. Soc. Am. 47, 460 ~1957!. W. Lewandowski, G. Petit, and C. Thomas, IEEE Trans. Instrum. Meas.
12
J. R. Zacharias, Phys. Rev. 94, 751 ~1954!. IM-42, 474 ~1993!.
13
K. Gibble, and S. Chu, Phys. Rev. Lett. 70, 1771 ~1993!.
48
R. Gold, IEEE Trans. Inf. Theory 33, 619 ~1967!.
49
14
D. Kleppner, H. C. Berg, S. B. Crampton, N. F. Ramsey, R. F. Vessot, H. Understanding GPS, Principles and Applications, edited by Elliott D. Ka-
E. Peters, and J. Vanier, Phys. Rev. 138, 972 ~1965!. plan ~Artech, Boston, 1996!, Chap. 4.
15
W. M. Itano, Proc. IEEE 79, 936 ~1991!.
50
Annual Report of the BIPM Time Section ~Bureau International des Poids
16
J. D. Prestage, G. J. Dick, and L. Maleki, IEEE Trans. Instrum. Meas. 40, et Mesures, Sevres, France 1995!, Vol. 8, pp. 55–61 and 81–103.
51
132 ~1991!. Y. Bar-Sever, P. Kroger, and J. Borjesson, J. Geophys. Res. 103, 5019
17
J. J. Bollinger, D. H. Heinzen, W. M. Itano, S. L. Gilbert, and D. J. ~1998!; J. Davis, G. Elgered, A. Niell, and C. Kuehn, Radio Sci. 28, 1003
Wineland, IEEE Trans. Instrum. Meas. 40, 126 ~1991!. ~1993!.
52
18
D. A. Howe, D. W. Allan, and J. A. Barnes, Proceedings of the 35th W. Lewandowski et al., Proceedings of the 11th Eurpoean Frequency and
Annual Symposium on Frequency Control, 1981, pp. A1–A47. This is Time Forum, 1997 pp. 187–193.
53
document AD-A110870, available from the National Technical Informa- D. Davis, M. A. Weiss, K. Davies, and G. Petit, Proceedings of the 4th Ion
tion Service, 5285 Port Royal Road, Springfield, VA 22161. GPS Intl. Technical Mtg., Albuquerque, NM, 1991, pp. 89–95.
19 54
S. Stein, D. Glaze, J. Levine, J. Gray, D. Hilliard, D. Howe, and L. Erb, Interface Control Document, Navstar GPS Space Segment/Navigation
Proceedings of the 36th Annual Symposium on Frequency Control, 1982, User Interfaces, ICD-GPS-200, Seal Beach, CA, Rockwell International
pp. 314–320. This is document AD-A130811, available from National Corporation, 1987.
Technical Information Service, 5285 Port Royal Road, Springfield, VA 55
J. Levine, IEEE Ultrason. Ferroelectr. Freq. Control ~submitted!.
56
22161. L. E. Young, D. C. Jefferson, S. M. Lichten, R. L. Tjoelker, and L.
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp
2596 Rev. Sci. Instrum., Vol. 70, No. 6, June 1999 Judah Levine
Downloaded 18 Sep 2001 to 132.163.136.56. Redistribution subject to AIP license or copyright, see http://ojps.aip.org/rsio/rsicr.jsp