H

The Science of Timekeeping

Application Note 1289

• UTC: Official World Time

• GPS: A Time Distribution
Utility
• Time: An Historical and
Future Perspective
• From Laboratory to
Practical Use
• Broad Applications
Across Society

1
2
Table of Contents

Introduction to the Science of Timekeeping .......................................... 8

Precise Timing Applications Pervade Our Society ................................ 10
Clocks and Timekeeping ............................................................................ 14
The Definition of the Second and its General Importance .................. 19
Historical Perspective ................................................................................ 19
An Illustrative Timekeeping Example ..................................................... 22
UTC, Official Time for the World ............................................................. 26
GPS Time and UTC ..................................................................................... 30
Accuracy and Stability of UTC .................................................................. 31
Einstein’s Relativity and Precise Timekeeping ...................................... 36
How to Access UTC .................................................................................... 39
Future Timing Techniques ......................................................................... 44
Global Navigation Satellite System Developments ................................ 47
UTC and the Future .................................................................................... 52
Conclusions ................................................................................................. 53
Acknowledgments ...................................................................................... 54
Appendix A: Time and Frequency Measures Accuracy, Error,
Precision, Predictability, Stability, and Uncertainty ............................. 56
Appendix B: Stability Analysis of Harrison-like Chronometers ........... 66
Appendix C: Time and Frequency Transfer, Distribution
and Dissemination Systems ....................................................................... 72
Glossary and Definitions ............................................................................ 82
Bibliography ................................................................................................ 85

3
 Hewlett-Packard is grateful to the three authors of this application
note for sharing their expertise. Their combined knowledge offers a
resource which will undoubtedly be considered a classic reference on
the science of timekeeping for decades to come.

David W. Allan
David W. Allan was born in Mapleton, Utah on September 25, 1936. He
received the B.S. and M.S. degrees in physics from Brigham Young Univer-
sity, Provo, Utah and from the University of Colorado, respectively. From
1960 until 1992 he worked at the U.S. National Institute of Standards and
Technology (NIST), formerly the National Bureau of Standards (NBS). His
work with time and frequency research, development and generation has
been with colleagues throughout the U.S. as well as in many other coun-
tries. In cooperation with several colleagues, his principal contributions
have been: 1) development of internationally adopted methods of charac-
terizing the performance of clocks, oscillators and time and frequency
distribution systems — known as the Allan variance, the Modified Allan
variance and the Time variance; 2) development of a time-scale algorithm
technique which combines clock readings for optimum and robust perfor-
mance of the output — being better than the best clock in the ensemble
and which has been used for more than 28 years to generate official time
from NBS/NIST; 3) development of the dual-mixer time difference tech- David W. Allan
nique, which allows clock measurements at the sub-picosecond level; 4)
development of the GPS common-view time transfer technique, which is
used to transfer clock times from around the world to the International
Bureau of Weights and Measures for the generation of International Atomic
Time and UTC, and for the comparison of the frequencies of the best
primary standards in the world, and which provides a major benefit to the
timing for NASA Jet Propulsion Laboratory’s (JPL’s) Deep Space tracking
Network; 5) development of improved timing in support of the measure-
ments of millisecond pulsars and the discovery, in this regard, of a random-
walk effect of the total electron content in the galactic interstellar medium;
6) the development of a method of separating out the various error compo-
nents causing inaccuracies in the Global Positioning System (GPS) perfor-
mance; 7) the measurement of the relativistic Sagnac effect for the rotating
Earth using GPS to transport time around the world; and 8) the invention
of techniques (algorithms and measurement procedures) which allow a
clock or set of clocks to always be correct; a patent associated therewith
was later licensed by Hewlett-Packard and formed the basis of the HP
SmartClock technology.

Since retiring from NIST, Dave has developed a method of defeating the
degradation (Selective Availability) on the GPS for timing purposes.
Most recently, he led an R&D effort which has the potential of real-time
satellite orbit determination at the centimeter level.

He has published well over a hundred papers, has contributed chapters in

several books, chaired several international timing committees, organized
several tutorials and seminars, and participated on several other commit-
tees — including being the U.S. representative from NBS/NIST for two
decades on the international Consultative Committee for the Definition of
the Second. His greatest satisfaction is that his work is used and referenced
4 often — hopefully for the benefit of a better world.
Neil Ashby
Neil Ashby was born in Dalhart, Texas on March 5, 1934. He received
the B.A. degree (Summa Cum Laude) in physics from the University of
Colorado, Boulder, in 1955, and the M.S. and Ph.D. degrees from
Harvard University, Cambridge, Massachusetts in 1956 and 1961,
respectively. After spending a year in Europe as a postdoctoral fellow,
he joined the faculty of the Department of Physics at the University of
Colorado in 1962. He has been a Professor of Physics there since 1970,
and was Department Chair from 1984–1988. He consults for the Time
and Frequency Division of the National Institute of Standards and
Technology, working on relativistic effects on clocks and global time
synchronization. His work was the basis of general relativistic correc-
tion being properly included in the Global Positioning System. He has
been a member of the International Committee on General Relativity
and Gravitation from 1989–1995. He serves on several international
working groups on relativistic effects in geodesy and in metrology.
Neil Ashby
Clifford C. Hodge
Cliff Hodge received the B.Sc. degree with First Class Honours in
Physics from the Imperial College of Science, Technology and Medi-
cine in South Kensington, London, in 1985. He continued his studies at
Imperial College by joining the newly-formed Semiconductor Physics
Group. In 1990, he was awarded the Ph.D. degree for his work on
“Induced absorption of novel semiconductor superlattices” for optical
and magneto-optical studies on a range of narrow- (indium arsenide-
based) and wide- (gallium arsenide-based) band gap materials. In 1990,
Cliff worked at the NASA Jet Propulsion Laboratory in Pasadena,
California, as a Resident Research Associate in the Atmospheric and
Oceanographic Sciences Division developing high-frequency modula-
tion techniques using lead-salt semiconductor lasers for the observa-
tion in real-time of a range of gaseous species in the stratosphere, as
part of the Airborne Arctic Stratospheric Expedition (AASE II)
Programme. In 1992, he joined the National Physical Laboratory (NPL)
in Teddington, Middlesex, as Project Leader developing a cryogenic all-
sapphire Super Mirror resonator for the intercomparison of the Allan Cliff Hodge
variance of frequency standards in the microwave and optical region of
the electromagnetic spectrum. Dr. Hodge has recently advised the
European Space Research and Technology Centre (ESTEC) on “stable
clocks and time-transfer techniques for GNSS2” as part of the
Thomson-CSF-led Consortium. In March 1996, he organized an ESTEC-
sponsored GNSS2 Workshop at NPL which brought together the Time
and Frequency specialists with the GNSS experts to a one-day Work-
shop entitled: “Next-Generation GNSS and The Frontiers of Time and
Frequency Measurement.” Cliff remains actively involved in the
exciting developments throughout Europe, the United States, the
Russian Federation and Asia, towards a seamless international Global
Navigation Satellite System (GNSS).

Cliff’s interest in promoting the new science and novel applications

that could be reaped from the latest improvements in accurate clocks
and two-way time-transfer techniques has continued with the recent
submission of a collaborative proposal to the ESA for an Atomic Clock
Ensemble in Space (ACES) on board the International Space Stations
ALPHA (ISSA). 5
 2000 -1500 BC
3500 BC Mayan calendar
Egyptian obelisks
and sundials

400 BC -1600 AD
Aztec calendar
1900 -1600 BC
Stonehenge

1583
Galileo discovers
pendulum period
constancy
1094
Sung Su’s Chinese
water clock
perfected

1656
Huygens pendulum
clock

1918
Quartz crystal
oscillator
developed

1955
Essen amd Parry
start keeping time
with cesium
atomic clock

1964
HP’s First Cesium
Clock (HP 5060A)
introduced

6
 Milestones in the Progress
of Timekeeping

1736
Harrison H1
chronometer
tested at sea

1948-49
Lyons develops
first atomic clock
(ammonia)

1960
11 12 1
10 2
6 3
8 4
6 6 6

HP 105B Quartz
Frequency 1978
Standard
introduced First GPS Block 1
satellite launched

1993
GPS declared
operational
and HP 58503A
GPS Time and
Frequency
Receiver
introduced using
HP SmartClock
technology

1991
HP 5071A Primary
Reference Standard
(cesium) introduced 1997 to the Future
NAVSTAR Block 2R,
next generation of
satellites

7
 Introduction to the Science of Timekeeping
There is no such thing as a singular true time derived from natural
phenomena. Yet we live in a world in which time is nearly as important
— and taken for granted — as the cycles of nature. “What time is it?” is
a question asked almost as automatically as taking a breath. Far more
seldom the question is asked, “What is time?” Depending on the
circumstances, a degree of accuracy is sought when we ask for the
time. A remote farmer might gauge the setting sun to determine the
approach of dinnertime. Most people glance at their watch or a nearby
clock to calibrate their day’s activities. Though few individuals are
involved in the process of consulting the accuracy of atomic clocks, the
impact of their accuracy is so far-reaching in society that virtually
everyone benefits thereby. Nevertheless, no watch or clock is com-
pletely accurate. Each has its own errors due to rate imperfections and
errors in setting. Your watch will display a time that is different from
the time displayed on any other watch, so you can never really know
precisely what time it is. The correct time is simply based on an agreed
standard. Currently, Universal Time Coordinated (UTC) has been
established as the world time scale.

In the 15th century, explorers took to the high seas in search of new
worlds and exotic treasures. To navigate, seafarers could determine
their latitude by using a sextant to observe the position of the sun at
midday or bright stars at night. Unfortunately, determining longitude
was more difficult. Because the earth rotates, measuring longitude
requires both a sextant and an accurate clock. In the 15th and 16th
centuries, clocks were insufficiently accurate to navigate with any
certainty, and this all too often led to disaster.

Visibility of this burning maritime issue was heightened in 1707 when

Admiral Sir Cloudesly Shovell miscalculated his position and wrecked
his flagship and 3 other British warships off the Scillies, losing nearly
2,000 lives, including his own. Even though his main error was in
latitude estimation — exacerbated by the persistent fog so that he
could see neither night sky or daytime sun — longitude continued to
be, generally, the biggest source of error. In 1714 Sir Isaac Newton
explained “for determining the Longitude at sea, there have been
several projects, true in theory, but difficult to execute...one is by
watch,...but by reason of the motion of the ship,...variation of heat and
cold...such a watch hath not yet been made.”

Spurred to action, the British government established an official body,

The Board of Longitude, in 1714. The board offered the longitude prize
(£20,000 or about $2,000,000 US today) to the person who could
contrive a means of resolving position at sea to within 30 nautical miles
after sailing to the West Indies. This required a clock that could keep
time to within 3 seconds per day.

8
For half a century, all manner of charlatans and pseudo-scientific
crackpots tried to claim the prize, but none could solve the fundamen-
tal scientific problem of determining longitude at sea. During this time,
John Harrison (1693–1776), a woodworker and musician from
Lincolnshire, devoted his life and genius to solving the problem.
Through intuition and sheer effort he developed a clock, a maritime
chronometer, that kept time accurate to one second per day. Harrison’s
chronometer was a great advance for maritime navigation — overcom-
ing the harsh environmental conditions encountered at sea. Using a
copy of Harrison’s clock, Captain James Cook mapped the Polynesian
islands and the Pacific Ocean regions. He wrote in his log book great
praise for the new navigational instrument, “our trusty friend the
watch” and “our never failing guide.”

Having timepieces with good long-term accuracy, navigators now

needed to synchronize their chronometers to a central clock. In Britain,
ships would sail up the Thames and watch a ball drop from a tower at
the Greenwich observatory at precisely 1:00 pm. Other nations with
large navies and merchant fleets, such as Portugal and France, had
their own standard time and debate raged over which meridian should
be used for standard time. In 1884, at the International Meridian
Conference, the Greenwich observatory was accepted as source of the
world’s standard time. Thus was coined Greenwich Mean Time or
GMT, a standard that dominated world timekeeping for nearly a
century.

Over time, clocks have improved dramatically and needs have changed,
so GMT has evolved to UTC. The acronym UTC is an English-French
mixture for Coordinated Universal Time (Temps Universel Coordonné
in French). Following International Telecommunications Union
Recommendation TF.536 on Time-Scale Notations, it was internation-
ally agreed to write Universal Coordinated Time as UTC, rather than
CUT or TUC, making it language-independent. GMT was based on
mean solar time. UTC is based on a definition of the second that is
nearly a million times more accurate. Under the general umbrella of la
Convention du Mètre , this new second is based on a quantum reso-
nance within a cesium atom.

In contrast, GMT was astronomically based and had a physical clock at

the Greenwich Observatory where they used to drop the ball at the
defined time. No physical clock keeps UTC. UTC is established by the
Bureau International des Poids et Mesures (BIPM) based on an aggre-
gate of data from timing laboratories throughout the world, and from
input from the International Earth Rotation Service (IERS). In a world
of ever-increasing precise-timing needs, not to have a single reference
clock seems paradoxical. Fortunately, for everyday purposes, clocks
are sufficiently accurate to meet most needs.

9
 Precise Timing Applications Pervade Our Society
The number and variety of applications using precise timing are
astounding and increasing. Currently, the count of precise timing
components manufactured and marketed each year is in the billions.
The world has evolved into the information age, and precise timing is at
the heart of managing the flow of that information so that it is reliable,
robust and inexpensive. Thus, all of mankind may use it efficiently and
effectively.

Practical, precise timing came with the invention of the quartz-crystal

oscillator and quartz-crystal filters, which are essential elements for
radio, radar and television with their enormous, far-reaching impact on
our society. Information is literally flowing at the speed of light in
computers and communications systems. With their ever-increasing
capabilities, they wouldn’t work without precise timing of gates and
network nodes. The arrival of atomic clocks provided even more
accuracy and opened new vistas. Unprecedented navigation is now
literally at our fingertips using relatively inexpensive hand-held GPS
receivers; yet the heart of GPS is an atomic-clock synchronized system.

Since their invention, the accuracy of atomic clocks has improved, on

the average, by a factor of two every two years. It is interesting that this
geometric progression rate of improvement is the same for computer
memory density. Moving from the laboratory to the workplace, the cost
of precision clocks has dropped phenomenally, as has the size of these
devices. Now, a computer with memory, its precision timing circuitry,
and even a transmitter can be made so tiny that a single chip containing
all of them can be inserted under the skin of an animal for monitoring
and control.

The most accurate measurement known to humanity is the measure-

ment of the duration of the second. The peak of the pyramid for accu-
rate time and frequency is the international reference, UTC. The current
best accuracy for the determination of the second results in a time error
of ±0.3 nanoseconds (billionths of a second) per day. This is equivalent
to ±1 second in 10 million years.

Because time and frequency can be measured so accurately, time and

frequency devices are often used to measure other fundamental quanti-
ties, such as the volt, the ampere, the ohm, and the meter. For example,
the General Conference of Weights and Measures (CGPM, Conférence
Générale des Poids et Mesures) redefined the meter as “the length of
the path travelled by light in vacuum during a time interval of
1/299 792 458 of a second (17th CGPM, 1983, Resolution 1).” It is
expected that eventually, all base units in metrology for the support of
technological development will be traceable back to the second. Even
now, science and technology are taking more and more advantage of
the ease and accuracy with which time and frequency can be generated,
disseminated, and utilized.

10
The Global Positioning System (GPS) is a classic example of using
precise timing for accurate positioning. Many navigation systems
predating GPS also use atomic clocks, but they are ground-based. GPS
features a set of 24 orbiting satellites, each with a synchronized atomic
clock on board. It effectively puts a super-accurate clock in the sky for
everyone to see with the eyes of modern technology. At any given time
and at any point on Earth, at least four of these satellites can be seen. A
precise timing device in a GPS receiver is used by its computer to
calculate the time of flight of the signal from each of the observable
satellites. Since the signals travel at the velocity of light, and this is
known exactly, the receiver’s computer can turn the time of flight into a
very accurate estimate of the distance to each satellite — accounting for
some delays in both the neutral and ionized parts of the atmosphere.
GPS satellites also broadcast their positions. Information received from
four satellites yields four equations which can be solved for four un-
knowns: latitude, longitude, altitude, and GPS-system time. In this
process, the high accuracy of the GPS atomic clocks is transferred to
the precision quartz-crystal clock inside the receiver. Therefore, high
accuracy position and time are readily available (from which velocity
can be deduced), and a multitude of users are capitalizing on this. GPS
is like a free utility with application opportunities limited only by our
imaginations. As a result, the number of users, and the variety of uses of
GPS have literally exploded.

Location via GPS is becoming very popular for consumer applications.

Cars are already available equipped to provide navigational and direc-
tional information. Some cars even use GPS to enhance security and
mayday systems. Several GPS and time-based location systems are
currently being developed to provide precise location of cellular phones
for applications such as emergency response, location-sensitive billing,
and fraud detection. Hikers and boaters have long enjoyed the benefits
of GPS navigation and a GPS receiver is now being made for the golfing
enthusiast to measure the distance to the pin!

To understand the pervasiveness of GPS, consider a recent receiver

survey by GPS World magazine which lists over 300 different kinds of
receivers available from 50 companies. In addition, massive systems are
now being developed that largely depend on GPS. For example, most
countries are now planning to use GPS with augmentation from other
systems for navigation and control of all aircraft. In 1996, well over half
a million GPS receivers were marketed by Japan, alone, for determining
the position of ground vehicles.

Even though GPS is predominantly a navigation system, the navigation

accuracy is dependent on precise timing techniques. Precise timing is
also embedded as part of the technology in a long list of other GPS
applications. The following list of applications is not intended to be
detailed, is far from exhaustive, and continues to expand: aerial photo-
grammetry, astrometry, astronomy, atmospheric studies and measure-
ments (ionosphere and troposphere), attitude determination, aviation,
calibrating other instruments, communications, differential GPS,
dredging, earthquake monitoring and prediction, exploration, frequency
11
 comparison of world standards for the second, geodesy, geodynamics
and Earth-plate tectonics, geographic information systems, hydrogra-
phy, mapping, military, mineral exploration and mining, NASA Jet
Propulsion Laboratory’s Deep Space Network, navigation (air, land,
marine), oceanography, offshore oil exploration, position determina-
tion, precision farming, positive train control, recreational (for ex-
ample, hiking and orienteering), resource management, search and
rescue, spacecraft guidance, surveying, timing, tracking (vehicular and
nonvehicular), traffic management, and weather measurements and
prediction. GPS World magazine kindly helped generate this list.

There are a variety of novel applications using GPS — some of great

scientific interest. In 1982, the first millisecond pulsar was discovered
using the Arecibo Observatory radio telescope in Puerto Rico. The
arriving pulse stream, which left the star some 13 thousand years ago,
was exceedingly uniform with only 1.55780645169838 milliseconds
between the pulses. It challenges the mind to think of a neutron star
with about the same mass as our sun, with a radius of only about 10
km, spinning 642 revolutions per second and radiating an enormous
pulse of electromagnetic energy with each rotation. Listening to this
signal at the Arecibo Observatory is like listening to the musical note E
above C above middle C on the musical scale. Several observatories
around the world have teamed up to study these unusual celestial
objects — of which there are now more than 30 known. GPS is the
principal time and frequency transfer standard for the measurement of
these incredibly uniform timing pulses. Using GPS these signals are
compared with the best atomic clocks in the world. Some of these
millisecond pulsar timing signals have stabilities that approach the best
of atomic clocks [1, 2]. Current time-predictability uncertainty esti-
mates are on the order of a second in well over a million years. Having
a very stable galactic-pulsar clock, external to our solar system and
about 13 thousand light years away, has provided a great deal of
information about our interstellar medium, our orbit around the sun,
and the coordinate reference frame for astrometry. Along with other
things, scientists are trying to use these ultra-precise measurements to
detect gravitational waves which give rise to distortion of space and
time, as were predicted by Einstein in the early part of the century.

Usually, the largest body of users of precise and accurate timing

techniques is thought to be within the areas of navigation (including
position determination) and communication systems (including radio,
TV, video, multimedia, telephone, cellular, internet, etc.). And though
this is probably true, the trends in society show rapidly developing and
interesting applications. Optical barcoding and barcode scanners are
permeating the market place; scanners now number in the millions.
The technique uses lasers and timing signals to read the code. Laser is
an acronym for “Light Amplification by Stimulated Emission of Radia-
tion.” Lasers evolved out of the work with masers, where the “m” in
maser stands for microwave. Both are used extensively in time and
frequency metrology because they provide a unique degree of fre-
quency stability in their emission. Lasers are directly amenable to

12
barcode scanning because all of the emitted photons are in step —
providing a high-intensity and highly directed beam of light. The laser’s
high intensity spot of light bounces off the light and dark spaces of the
barcode. The timing of intensity changes of the reflected light allows
the decoding of four different widths of bars and more than 50 bits of
information in the blink of an eye. While the precise frequency and
timing requirements for this application are very relaxed, the laser
principle makes it work. In the United States, mail handling has been
facilitated greatly with the use of optical barcode scanners. Lasers are
also at the heart of the success of the popular laser printer and are
used extensively in copiers as well. CD-ROMs and CD music disks are
read with lasers at megahertz rates, and now laser dental drills can
repair cavities pain-free.

In the general body of precise time and frequency users, telecommuni-

cation and navigation systems require very high levels of precise
network synchronization. The electric power grid uses precise timing
for efficient power flow and for fault detection. These systems have
moved from milliseconds, to microseconds, to nanoseconds in their
timing requirements. Optical fiber transmissions are now working at
the sub-nanosecond level. Magnetic Resonance Imaging (MRI) is a
direct spinoff utilizing precision frequency spectroscopy techniques.
New MRI techniques now being developed will provide fine-detailed,
non-invasive scans of the brain [3]. Transportation system problems
are being solved by turning to precise timing techniques. The technol-
ogy exists today with the potential for determining the position of a
vehicle to about a centimeter using precise and accurate time and
frequency techniques. Computer networks require precise timing. The
banking and business industries are increasingly using precise timing
to time-tag transactions. New cars and appliances extensively use
microprocessors and quartz-crystal oscillators and filters. The automo-
bile of the future will have its position and course displayed on the
dashboard.

Several Nobel laureates have contributed in fundamental ways to the

success of time and frequency techniques. It would be nearly impos-
sible to list all the ways that precise time and frequency tools have
benefitted humanity. These tools and techniques are often in the
background silently serving society’s needs. The increasing communi-
cation and mobility capabilities promise to foster understanding
among the nations and cultures of the world.

13
 Clocks and Timekeeping
Almost any clock may be considered a two-part device (see Figure 1).
First, a clock will have an oscillating device for determining the length
of the second or some other desired time interval. This is usually
referred to as the clock’s frequency standard, which oscillates at some
rate determined by the laws of physics. Historically, the pendulum was
the classic source of time interval. Currently, the typical wristwatch has
as its frequency standard a quartz-crystal tuning fork with an oscillation
frequency of 32,768 Hz (one Hertz, abbreviated Hz, is a cycle per
second). This number of oscillations is convenient for the associated
digital electronic circuit, because if this number is divided by 215, which
is easy for a digital chip divider, the result is one cycle or pulse per
second.

In simple terms, atomic clocks generally provide a much more accurate

frequency than can be generated by any physical device such as a
pendulum or quartz crystal oscillator. An atomic clock uses as its
reference the oscillation of an electromagnetic signal associated with a
quantum transition between two energy levels in an atom. This bundle
of electromagnetic energy is called a photon and its energy, E, is equal
to the difference in energy between these two levels. The photon may
be either given off or absorbed by the atom.

For a given quantum transition, the photons emitted or absorbed have a

unique frequency proportional to the energy difference with little
variability around this value. The relationship between this energy
difference and the electromagnetic vibration frequency of the photons
is given by E = hν, where h is Planck’s constant and ν is the Greek
letter denoting the frequency of the photon’s electromagnetic wave.
The trick in atomic clock metrology is harnessing the frequency of
these photons while producing minimum perturbations on the natural
atomic resonance.
Figure 1. Almost any clock is a two-
part device. The first part provides
What is a Clock? An Oscillator + a Counter equally-spaced periodic events as
derived from and defined by some
oscillating device. The second part
12 adds up these events (an accumulation
of these time intervals) to provide
9 3 time from the clock.
+ 6

Oscillator Counter
(Frequency Device) (Counts Periodic Events)

+ Fast Electronic Counter

Cs - 133 Atom

14
As an example of this first part of a clock, a pendulum’s theoretical
frequency is given by
g/l
ν=
2π
where g is the gravitational acceleration at the location of the pendu-
lum and l is the length of the pendulum bob’s support wire. Specifi-
cally, an ideal pendulum that swings through its lowest point once per
second (one full cycle every two seconds, 0.5 Hz) will have a length l =
99.3621 centimeters if it is at sea level and at 45° latitude. By its nature,
a pendulum clock will depend upon several parameters, including both
its location and its environment since, for example, most materials
expand with increasing temperature. An increase in temperature would
cause the support wire to get longer and the pendulum clock to slow
down.

The current official definition of the second is much more elegant and
was agreed upon in 1967. It is based on the simple equation, E = hν.
The energy difference is specific to a particular quantum transition in
the cesium-133 atom, whose unperturbed frequency has been defined
as 9,192,631,770 Hz. When the defined number of cycles transpire for
the electromagnetic signal associated with the photon either being
given off or absorbed by this quantum transition, we have one official
second.

The second part of a clock is a counter (sometimes called an integra-

tor, adder, or accumulator) that keeps track of the number of seconds
or clock cycles that have occurred. This part of the clock is repre-
sented by the gears and clock face in a pendulum clock. It keeps track
of hours, minutes, and seconds. After being set initially, the clock can
then provide its estimate of the correct time by adding up the number
of clock cycles.

In principle, if a clock were set perfectly and if its frequency or rate

remained perfect, it would keep the correct time indefinitely. In
practice, this is impossible for several reasons: the clock cannot be set
perfectly; random and systematic variations are intrinsic to any oscilla-
tor, and when these random variations are averaged, the result is often
not well-behaved; time is a function of position and motion (relativistic
effects); and lastly and invariably, environmental changes cause the
clock’s frequency to vary from ideal. So, if a clock is measured with
sufficient precision, its reading will not agree with UTC, except at the
instant when it momentarily passes through the correct time which, of
course, is only correct by definition or convention.

The quality of a clock depends on how well it is set, how accurate and
stable its frequency is, and the degree of immunity the clock has to
environmental changes. Using modern techniques, a clock coupled
with a microprocessor and sensors can be made to compensate for
some of its timing instabilities. The quartz crystal oscillator provides a
cost-effective means of achieving reasonably good clock stability, and
properly interfaced with a computing capability and set of sensors, it
15
 can be very effective [4]. Typically, atomic clocks are much less
sensitive to environmental changes but significantly more expensive. In
addition, the intrinsic nature of atomic clocks usually yields a more
accurate estimate of correct frequency for the determination of the
second. The choice of which clock is most appropriate for a given
application should be considered from a systems point of view.

Four useful measures for describing the quality of a clock are: fre-
quency accuracy, frequency stability, time accuracy, and time stability.
These measures are not all independent. A clock’s frequency (or rate)
accuracy is how well it can realize the defined length of the second. A
commonly-used measure is the change in the error of a clock’s time
divided by the elapsed time, t, over which the change occurred. This is
often called the fractional or normalized frequency departure, y(t), and
is a time-dependent, dimensionless number. The goal of the Harrison
chronometers was to have y(t) less than three seconds per day, 3/86400
= 3.5 × 10–5. The best primary frequency standards in the world today
have y(t) values less than 1 × 10–14. Of course, the smaller the number,
the better the clock. (See Appendix A for more details regarding these
measures.)

Frequency stability, on the other hand, indicates the change in fre-

quency from one period of time to the next. A clock can have a signifi-
cant frequency error and still be very stable; in other words, the
frequency or rate error stays about the same. For example, a clock may
have a rate inaccuracy of gaining one second a day, but if that rate
remains the same, it would have perfect frequency stability. Two very
important kinds of atomic clocks, a hydrogen-maser and a cesium-
beam, are good examples of clocks having different stabilities and
accuracies. A hydrogen-maser clock typically has better frequency
stability than a cesium-beam clock from second to second or from hour
to hour, but often not from month to month and longer. On the other
hand, the typical cesium-beam clock is more accurate than the hydro-
gen-maser clock. Quartz-oscillator based clocks can be very stable for
short times, but they drift in frequency and don’t have the frequency
accuracy of atomic clocks.

Time accuracy by definition means how well a clock agrees with UTC.
There are often cases where what is needed is consistency of time at
several locations in a system. What may be important is the time
accuracy of each of the clocks with respect to the system. This is
exactly the case for the U.S. Department of Defense’s (DoD’s) Global
Positioning System (GPS) whose time differs from UTC. Specifically,
the GPS time broadcast by each of the satellites needs to be synchro-
nous with all of the other satellites in the constellation to within a
small number of nanoseconds in order for the system to work properly.
This is accomplished for GPS by the presence of atomic clocks on
board the satellites. The telecommunications industry is another
classic example, where large amounts of data are being transported
among many network nodes. These nodes need to be carefully synchro-
nized or data will be lost or transmissions will be faulty (a missing line
in a FAX, for example).
16
Time stability is usually correlated with frequency stability, but it is
often useful as a measure of changes with respect to some uniform
flow of time in time-measurement systems and/or time-distribution or
time-dissemination systems. For instance, consider the clock men-
tioned before, which gains one second a day and does this day after
day. While its time accuracy may be degrading a second a day from its
initial setting, it would have perfect frequency stability, and conse-
quently perfect time predictability. If the time or frequency errors of a
clock can be estimated, then compensating corrections can be made.
If the time and rate (frequency) of a clock with good stability are
calibrated against some better clock, then an estimate of that better
clock’s reading can be used should the better clock not be available. In
a hierarchy of calibrations, at the peak of the pyramid are the primary
reference standards. In the case of a Primary Frequency Standard,
“primary” means it can provide the length of the second indepen-
dently. In determining their accuracy, some Primary Frequency
Standards are taken through a series of measurement cycles to obtain
a best estimate of the second; this often precludes these standards
from running continuously as is needed for clock operation. In such a
case, the technique of having a secondary clock that has been cali-
brated in rate by a Primary Frequency Standard allows the secondary
clock to perpetuate an estimate of the time of the Primary Frequency
Standard, had it been operating as a clock on a continuous basis. This
is an important technique in timekeeping that is used extensively in
the generation of UTC.

Moving from astronomical time to atomic time evoked many new

concepts. One very important concept is the power of properly used
clock ensembles. Historically, astronomical time was based on the
movements of one Earth, one moon, one sun, one solar system, and
one celestial sphere of moving stars and planets. Time was deduced
from the dynamical equations of motion of these objects. Even though
there was only one of each, reliability was never an issue; the earth
was not expected to stop spinning! With the introduction of atomic
clocks, there were now many timekeepers — each of which could
generate its own estimate of time. The concept of a clock ensemble
had great merit from the standpoints of reliability and performance.
One clock can stop. From measuring the time difference between two
clocks it is impossible to tell which one is deviating. Three indepen-
dent clocks allows an independent estimate of the stability of each,
but four are needed in case one should quit. Further, it has been
shown that if the clocks are well characterized (see Appendix A), and
optimum weightings are given to them in a properly developed com-
bining algorithm, the weighted ensemble can have better performance
than the best clock in the set and even the worst clock will enhance
the ensemble output. If a clock fails or has a bad reading, it can be
detected and rejected so that it does not adversely perturb the
ensemble’s output. A very reliable and stable real-time output can be
generated by such an ensemble. The first real-time output clock
ensemble was developed at the National Institute of Standards and
Technology in Boulder, Colorado in 1968, and essentially all of the
timing centers currently use clock ensembles to generate their official
timing reference [5]. 17
 Measuring and comparing time and frequency of clocks remote from
each other can be accomplished in a wide variety of ways and with a
variety of accuracies and stabilities. Usually, the better the accuracy
and stability, the more expensive will be the system. During the last
two decades with the advent of the Global Positioning System, great
progress has been made in transporting time and frequency. With the
atomic clocks on board the satellites, GPS acts like a portable clock in
the sky, continuously available anywhere on Earth via very inexpen-
sive receivers. Accuracies better than one microsecond (one millionth
of a second) with respect to UTC are readily available by the proper
use of GPS timing receivers. Satellite techniques, in general, have
demonstrated significant advantages over terrestrial techniques, such
as improvements in accuracy, integrity, availability, continuity, cover-
age of service, and perhaps more importantly, cost.

To measure time precisely, we must account for the delay between the
clock and the user. For example, when you look at a clock, you don’t
see what time it is, but what time it was when the light that reflected
the time toward your eye left the clock. This is like the delay of sound;
the lightning is seen and then the thunder is heard. The speed of sound
is about a million times slower than the speed of light. Light travels 30
centimeters (about one foot) in a nanosecond, so the delay for a clock
in the same room as the observer is only several nanoseconds. This
delay for GPS satellite orbits is of the order of 70 milliseconds (thou-
sandths of a second). In telecommunication networks, these delays can
be very important. GPS timing receivers estimate and account for the
signal-propagation delay from the satellites to the receiver.

From the above, it is apparent that the performance one sees from a
clock is affected by signal propagation variations and delays, changes
intrinsic to the clock, and environmental perturbations on the clock
(see Figure 2). See Appendix A for a more detailed discussion of
accuracy, stability, precision, and uncertainty in specifying clock
performance.

Figure 2. What we perceive as the time

from a clock is influenced by three
Perceived Cause of Clock Deviations things: first, by the perturbations on
the time signal as it is communicated
from the clock to the receiver of the
time information; second, by the
variability of the timing intrinsic
Clock to the clock itself; and third, by
the environmental (external)
Measurement 12 Environment perturbations that cause variability
in the clock’s timing.
9 3

Perceived Performance =

Measurement Random and Systematic Environmental

Noise + Errors within Clock + Perturbations

18
The Definition of the Second and its General
Importance
In 1967, it was agreed by the 13th General Conference of Weights and
Measures (CGPM [Conférence Générale des Poids et Mesures] Resolution
1) that: “The second is the duration of 9,192,631,770 periods of the radia-
tion corresponding to the transition between the two hyperfine levels of
the ground state of the cesium-133 atom.” The current best accuracy for
the realization of the second so defined is a value “y(t),” as defined above,
of 3 × 10–15 [6]. This is equivalent to ±1 second in 10 million years.

At the time of the definition, the atomic second was made to agree as
much as possible with the ephemeris second based on astronomical
measurements. The 9,192,631,770 Hz assigned to the above cesium
transition was the result of a three-year cooperative between L. Essen
at the National Physical Laboratory (NPL) in Teddington, England and
W. Markowitz at the U.S. Naval Observatory (USNO) in Washington,
D.C. NPL supplied the cesium clock, and USNO provided the astro-
nomical measurements via a clever dual-rate moon camera which
could simultaneously compensate for the different movement of the
moon and the stars [7]. The occultation of the latter by the former gave
a very accurate estimate of ephemeris time. The commonly received
transmissions of the National Bureau of Standards time and frequency
radio transmission on WWV were used to relate the two times.

A very accurate frequency has three principle advantages over the

other base standards in metrology: first, no other standard can pres-
ently be measured so accurately; second, high levels of accuracy and
stability can be obtained relatively inexpensively; and third, it can be
communicated via electromagnetic waves (as in radio transmission,
microwave transmission and laser transmission). With the develop-
ment of these time and frequency techniques, it is not surprising that
there has been a dramatic increase in the use of precise timing over the
last few decades, and this trend is expected to continue. GPS is a
classic example.

All high-accuracy radio navigation systems employ precise and accu-

rate timing. Telecommunication systems require precise network
synchronization. The electric power grid uses precise timing for
efficient power flow and for fault detection. Computer networks
require precise timing. The banking and business industries are in-
creasingly using precise timing to time-tag transactions. Transportation
system problems are increasingly being solved by turning to precise
timing techniques. Atomic clocks are at the heart of GPS, and world-
wide avionics is moving towards adopting GPS as an important auxil-
iary navigation tool. In the long-term planning for the U.S., GPS will be
the sole means of navigation.

Historical Perspective
Historically, in an agrarian society, timing was tied to sunrise, sunset
and the seasons, and astronomical observations provided both parts of
the clock: the frequency standard (one cycle per day for the earth’s
19
 spin) and the counter (calendar). Toward the end of the 19th century,
the improved accuracy of the astronomical measurements of Simon
Newcomb demonstrated that the exact fractional number of days in a
year varied from year to year. It was later determined that the length of
the day changes at a level of about 1.5 × 10–9 for y(t) from day to day or
from year to year. In addition, there is an apparent decrease in the spin
rate of the earth of about 2 × 10–10 per year.

Because of the irregularities in the spin rate of the earth, a series of

time scales evolved [8]. The tilt of the earth with respect to its orbital
plane around the sun combined with its non-circular elliptical orbit
gives rise both to the seasons and to the “equation of time.” A visual
display of this equation is often shown on globes as a figure-eight
pattern, the Analemma, showing the path of the sun at exact 24-hour
intervals over the course of a year. If, for example, you had a perfect
clock at the Greenwich meridian and looked up at exactly noon each
day according to the time on the clock, you would observe the sun
wandering back and forth in an East-West angle about ±4 degrees (±16
minutes for the clock’s reading) during a year’s time. Correcting for the
equation of time yields UT0 or mean solar time. Astronomers and
navigators also need to correct for polar motion as the earth wobbles
with respect to its spin axis. These corrections give the time scale UT1
— the most useful time scale for knowing the earth’s angular position
with respect to the heavens (celestial sphere). Careful measurements
show an annual and a semi-annual variation in UT1; correcting for
these periodic variations yields the time scale UT2. This was done in an
effort to make a smooth time scale based on the spin rate of the earth.
The irregularities of the earth’s spin rate around the UT2 time scale are
random and cannot be predicted to better than the 1.5 × 10–9 level,
which gives rise to a time prediction error of about 60 milliseconds
over the course of a year. In contrast, UTC has a time error predictabil-
ity of about 60 nanoseconds — a million times better.

In 1960 these irregularities in the spin rate of the earth led to a new
definition of the second. Prior to this time, since there are 86,400
seconds in a day, the second was defined as “1/86400 of a mean solar
day.” The new definition used as its frequency standard the one-cycle-
per-year orbit period of the earth around the sun along with other
astrometric data, such as the orbit period of the moon around the earth.
The second was redefined as “the fraction 1/31556925.9747 of the
tropical year 1900 January 0 at 12 hours ephemeris time.” This defini-
tion, the ephemeris second, was the official length of the second from
1960 until 1967 when the atomic second, based on a hyperfine transi-
tion in cesium, was introduced. The ephemeris second was difficult to
measure and required one-year averages. In contrast, the atomic second
could be measured in timescales on the order of one-to-ten days.

The idea of an atomic clock was actually conceived in the early 1940s
by Nobel laureate I. I. Rabi. The first atomic clock, based on a micro-
wave resonance in the ammonia molecule and using the microwave
photon-absorption principle, was introduced to the world in 1949 by
Harold Lyons of the National Bureau of Standards (then based in
20 Washington D.C.). Its stability was not much different from that of the
spin rate of the earth, and it did not remain in operation as a useful
working clock, but it was an important philosophical and scientific step.
In the early 1950s, Lyons’ group researched the possibility of using a
cesium beam as an atomic frequency standard. This pioneering work
demonstrated the potential for high-accuracy atomic frequency standards.
However, this development was never turned into an operating atomic
clock. In other words, the second part of the clock — a continuous and
indefinite accumulation of atomic seconds — was not incorporated.

It was not until June of 1955 that L. Essen and J. V. L. Parry of the
National Physical Laboratory in Teddington, England introduced the
first operating atomic clock — also based on cesium. Over the next
several years, the improved accuracy and uniformity of cesium-beam
clocks became readily apparent, and the world community was ready
for a new definition. Hence, in October of 1967 at the 13th convocation
of the General Conference of Weights and Measures, it was declared
that: “The second is the duration of 9,192,631,770 periods of the
radiation corresponding to the two hyperfine levels of the ground state
of the cesium-133 atom.” This particular cesium resonance was agreed
upon under la Convention du Mètre and remains to the present time as
the official definition of the second for the world community. Fortu-
nately, the choice of cesium for the definition was a good one, and it is
evident that this definition is likely to remain with us for some time to
come. It has been well commercialized; cesium-atomic clocks now
number in the thousands.

Atomic clocks have improved dramatically since their introduction —

by a factor of about one million since being introduced in the late
1940s. As technology has moved to the foreground of society, these
new and improved clocks are being used on an ever-broadening basis.
An interesting set of problems has arisen during the move from a
society using astronomical phenomena for timing to a society where
atomic timekeeping techniques are becoming ever more important.
One problem with the current definition of the second is that it was
made to agree with the ephemeris second, which ties back to the year
1900. The spin rate of the earth is very erratic as compared to atomic
clocks, and it is very difficult to arrive at a well-defined slowing-trend
value. For example, over this century the variations are so large that it
is difficult to see any trend at all. Over the last 400 years, the decreas-
ing spin rate is about –1.1 × 10–10 per year. The long-term data for
Earth spin rate taken from S. K. Runcorn’s work using coral growth as
paleontological clocks yields a decreasing spin rate of –2.5 × 10–10 per
year . Using the definition for the ephemeris second as the earth’s rate
in 1900, the earth has slowed down about 62 seconds between 1900
and 1996 — 32 of those seconds between 1900 and 1958 when atomic
time was set synchronous with Universal Time and the remaining 30
between 1958 and 1996. Again, because of the relatively large fluctua-
tions of the earth’s spin rate, this number has little meaning. If, for
example, this 62 seconds is used to calculate a rate of decrease, we
obtain –4.3 × 10–10 per year, which makes little sense given the other
values. In a different way of looking at this problem, if the current
cesium-based definition reflected more nearly the current spin rate of
the earth, the divergence between UT1 (Earth time) and UTC would be 21
 about ten times less, but we have no guarantee that the current Earth
spin rate will continue. At the present time, the earth is running slow
about a second a year. Precision time and frequency metrologists are
satisfied with the excellent performance of atomic clocks. However,
navigators and astronomers need to know Earth angular position or
Earth time, UT1.

An Illustrative Timekeeping Example

As was shown in Figure 2, there are three things that can cause devia-
tions in the observed time of a clock: measurement noise, internal
clock deviations, and external environmental perturbations affecting
the clock. The internal clock deviations can arise either from deviations
in the clock’s frequency standard (pendulum-like device), or in its
counting mechanism which keeps track of the measured time intervals.
The following experiment was conducted to illustrate these different
deviation mechanisms. We purposely purchased three of the cheapest
stopwatches from a local department store — costing about six dollars
each. We chose watches with 1/100-second readout capability. The
readout was a liquid crystal display (LCD). The internal frequency
reference was the typical quartz tuning-fork oscillation of 32,768 Hz.
The three stopwatches were attached side-by-side to a board so that
they could be observed simultaneously.

To measure the readings of the three clocks, we mounted a camera in

front of the board and used a 1/1000-second shutter speed. Once each
day a telephone call was made to the Atomic Clock in Boulder, Colo-
rado and the camera-shutter button was pushed as near as was hu-
manly possible to align with the 14:00:05 UTC tick from the atomic
clock. The experiment was set up in the living room of one of the
authors, Neil Ashby who lives in Boulder, Colorado. This time corre-
sponded to five seconds after 7:00 a.m. Mountain Standard Time. The
five-seconds-after-the-hour delay was chosen so that the one-second
beat could be sensed to obtain the most accurate triggering of the
camera shutter. Measurements were taken once each day for 145 days.

The main environmental effect was just the seasonal temperature of the
living room. Pressure, humidity, shock and vibration effects were
probably negligible compared to the temperature effects.

Ashby also plays the piano, and we hoped his musical training would
assist in minimizing the trigger-time measurement noise as he operated
the camera. In theory, the 1/1000-second (1 millisecond) shutter speed
should easily stop the LCD readout, since it only changes every
10 milliseconds (1/100-second). However, in many cases, the LCD
readouts were caught in transition.

Appendix A explains some of the details of the tools used to analyze the
data. Once we developed the film, we extracted the readings of each of
the three clocks with a readout precision of 10 milliseconds (ms) and a
reading once per day. The theoretical standard deviation on 10 ms
rollover precision is only 2.9 ms. Contributions to the measurement
noise will also come from the precision with which the shutter is
22
 100.0 0.0

50.0 Watch #1
Watch #2
Time Error in Seconds

Time Error in Seconds

–0.5
0.0

–50.0 –1.0
Watch #3 Watch #2
–100.0
Watch #1 –1.5
–150.0 Watch #3

–200.0
0 10 20 30 40 50 60 70 80 90 100 120 130 140 0 10 20 30 40 50 60 70 80 90 100 120 130 140

Time in Days From Start of Experiment Time in Days From Start of Experiment

Figure 3. A plot of the time error of three stopwatches with Figure 4. After calibrating the clock rate (frequency) for
data taken daily for 145 days. each of the three stopwatches, these rates were subtracted
from the data and the residual time errors are shown here.

pushed plus the delay in the atomic-clock signal across the telephone
line. Since the experiment was also in Boulder and only one telephone
exchange office processed the call, this delay should be the order of a
millisecond. We therefore have a near perfect time reference as
compared with all the other time-deviation mechanisms observed in
this experiment.

The clocks were synchronized with the Boulder atomic clock at the
beginning of the experiment. Figure 3 is a plot of the daily time errors
including the initial point when the three clocks were synchronized.
The frequency accuracies of the three clocks were: –1.17, +0.48, and
–0.81 seconds per day for clocks 1, 2, and 3 respectively. If these rates
are divided by the number of seconds in a day (86,400 s/day) we obtain
the dimensionless frequency inaccuracy for each of the three clocks:
–1.36 × 10–5, +0.56 × 10–5, and –0.94 × 10–5, respectively. These values
are typical for stopwatches of this quality. A wristwatch has the
advantage of being controlled by the body’s temperature, and the
quartz crystal is usually preset to run on the correct frequency at
this temperature.

If we now use the Boulder atomic clock to calibrate the rate of each of
the three clocks and remove these rate offsets per the above numbers,
we then observe the residual errors shown in Figure 4. It will be
noticed that these errors, on a peak-to-peak basis, are nearly a hundred
times smaller than those shown in the previous figure. A high degree of
long-term correlation is also observed between the three clocks —
especially between clocks 2 and 3. This probably is due to their being
subject to the same temperature environment and having very similar
temperature dependence.

The effect of frequency drift, D, is also evident in Figure 4 as can be

seen by the parabolic shape of each of the three curves. For the kinds
of noise processes encountered in this experiment, a simple and near-
optimum estimate of the drift is given by D = 4[x(N) – 2x(N/2) + x(0)]/
T2, where x(0), x(N/2), and x(N) are the clock’s time residual errors at 23
 0.8
0.7 Watch #1 WATCH 2 - WATCH 3
0.6 Watch #2 with mean frequency and drift removed
0.5 Watch #3
Time Error in Seconds

0.4
0.3 10 –5 53 ms White PM
0.2
0.1
Measurement Noise

Mod. σy (τ)
0.0 10 –6
–0.1
–0.2 3 ms Flicker FM
–0.3 10 –7 Clock Noise
–0.4
–0.5 1 day
–0.6 10 –8
–0.7
0 10 20 30 40 50 60 70 80 90 100 120 130 140
10 4 10 5 10 6 10 7
Time in Days From Start of Experiment Averaging Time, τ (seconds)

Figure 5. After subtracting calculated frequency offsets and Figure 6. If we now analyze the frequency stability of series
frequency drifts from the data, we obtain the time error 2 minus series 3 using the methods developed in Appendix
residuals shown here. A, we obtain the plot shown here.

the beginning, in the middle and at the end of the data, respectively,
and T is the total time duration of the data set — in our case 145 days.
If the drift for each of the three time series is calculated and subtracted
from the data, we have the time error residuals shown in Figure 5.

Let’s denote the residual time error shown in Figure 5 as three time
series x1(i), x2(i), and x3(i) for each of the three clocks with the i
designating the day count from the beginning of the experiment. As
explained in Appendix A, we can calculate the day-to-day time stability
for each of these three time series and we obtain 48 ms, 52 ms, and
54 ms, respectively. These values will be the sum of the shutter-trigger
time noise, the LCD rollover noise, and the quartz tuning-fork noise
(measurement noise + counter noise + frequency standard noise). Of
course, the frequency standard’s deviations can come from internal
mechanisms or be driven by the environment. At this point, we cannot
deduce how much is coming from each of these different sources.
As will be shown in Figure 6, at an averaging time of 1 day, the mea-
surement noise is well above the crystal oscillator noise. In this case as
was implemented above, a three-point frequency drift estimator can be
shown to be a simple and efficient estimator, where the three points
are the first, middle and last time-error points. Having removed the
mean frequency to obtain the residual time error data for Figure 4
results in the first and last point being zero. Hence, the above equation
for the frequency drift can be simply obtained by D = –8 x(N/2)/T 2 ,
where x(N/2) is the middle time-error point and T is the data length.
The residuals shown in Figure 5 are probably made up of measurement
noise, correlated effects driven by a common temperature for all three
stopwatches, and long-term random variations of the quartz crystal
oscillators.

Since the data were taken each day at the same moment within 1 ms,
we can subtract one time series from the other on a day-by-day basis.
The effect of the shutter-trigger noise will cancel in these differences
since the error is identical to within 1 ms. We can write the difference
x12(i) = x1(i) – x2(i), etc. for x13(i) and x23(i). In the case of random-
24
uncorrelated deviations, the cross variances average to zero, and we
may solve for the unknown variances with the following equations:

σ 12 = 1  σ 12
2 2
+ σ 13 − σ 223  , σ 22 = 1  σ 12
2
+ σ 223 − σ 13
2 
and
2  2  

σ 23 = 1  σ 13
2
+ σ 223 − σ 12
2 
.
2 
We have here three equations with three unknowns. Solving for these
variances and taking the square root yields day-to-day time stabilities
for the LCD rollover noise of: 27 ms, 40 ms, and 36 ms, respectively,
which is much larger than the theoretical value of 2.9 ms. At this point,
we may wonder if this is clock noise. We will show later that it is not.

We may also write the above time difference equations as x12(i) =

[xL1(i) – xL2(i)] + [xC1(i) – xC2(i)], and so forth, with the first bracketed
term representing the LCD rollover error and the second the clock
error. Taking the difference between these series also has the advan-
tage of subtracting the effects due to temperature if each series is
affected the same way. If we have removed the correlated deviations
and most of the systematics, we are left with the random uncorrelated
deviations.

We can now deduce the shutter-trigger noise — how well our piano
finger follows the Boulder atomic clock. By squaring the total day-to-
day time stability given earlier for each of the three time series and
then subtracting the estimated variance for the LCD rollover noise just
deduced, we end up with an estimate of the variance for the shutter-
trigger measurement noise. Taking the square root of these calcula-
tions yields 40 ms, 33 ms, and 40 ms respectively. Taking an average of
the three variances and taking the square root yields 38 ms for how
well a person can push a button in synchronism with an accurate time
signal.

As mentioned before, the correlated effects (probably due to tempera-

ture) and the shutter-trigger noise are largely removed from this set of
time differences. In Figure 6, you will notice that the first part of the
−3 2
plot has a slope of τ ; as explained in Appendix A, this slope is
consistent with random and uncorrelated measurement noise. This
kind of noise is not observed for clock deviations for this range of
averaging times. In the long-term it will be noticed that the slope is
more like τ0. This slope is characteristic of flicker-noise frequency
modulation and is typical for clocks. If this τ0 slope is extrapolated
back to τ = 1 day, we see a level that would correspond to about 3 ms
of time stability noise. Hence, the reason for ignoring the random,
uncorrelated clock noise in the earlier calculations for the day-to-day
stability estimates of LCD and shutter-trigger noise contributions.

We see the very exciting conclusion that if these inexpensive clocks

are calibrated and the systematics are removed, their timekeeping
ability improves dramatically: from hundreds of seconds per year to
the order of one second per year. The technology exists today to
perform such calibrations and systematic error removal automatically.
When this is implemented, we will see a dramatic improvement in the 25
 common inexpensive watch. Such a watch would always be correct —
within some very small error limit — and would never need to be reset.
A patent has been awarded that explains how this could be done
(United States Patent Number 5,274,545). The HP SmartClock technol-
ogy employs some of the ideas in this patent — plus some others — to
obtain its atomic-clock-like performance from a quartz-oscillator GPS
receiver-based product.

UTC: Official Time for the World

Because most users want official time to tie to Earth time, a dilemma
arises. Precision time and frequency users want the most uniform and
accurate time possible, yet the earth speeds up and slows down and is
not useful for precision metrology. Hence, a decision was made to
make UTC a compromise time scale. The changes due to instabilities in
the earth’s spin rate would be accommodated by employing leap
seconds, while on the other hand, the UTC second would be kept as
close as possible to the definition based on the cesium atom. This
approach became official time for the world starting January of 1972.
By international agreement, a leap second may be introduced at the end
of any month. However, the preferred dates are at the end of June and
the end of December when required. They are introduced when neces-
sary to keep UTC within 0.9 seconds of Earth time (UT1), and there
have been 21 leap seconds added between January 1, 1972 and July 1,
1997 — an average of a little less than one a year. This average value
has little meaning, however, because of the random variations in the
earth’s spin and because there is a nominal slow down in the earth’s
spin rate as well. At the estimated rate of decrease, the earth would
lose about 1/2 day after 4,000 years, and about two leap seconds a
month would be needed to keep UTC in step with Earth time, UT1.

The work to generate UTC is performed at the Bureau International des

Poids et Mesures at the Pavillon de Breteuil, F-92312 SEVRES Cedex,
France, near Paris [9, 10, 11]. The staff doing the work is composed of
several international timing experts who frequently interact with, and
obtain timing data from, the rest of the world’s time and frequency
community. The leap second steps are determined by the International
Earth Rotation Service (IERS), which operates out of the Paris Obser-
vatory and which collects Earth rotation data from numerous observa-
tories and radio telescopes around the globe.

UTC was set synchronous with UT1 at 0000 hours on January 1, 1958,
and until 1972, a different technique was used to keep UTC in close
agreement with UT1. During that time, both frequency steps and 0.1-
second time steps were used to chase the instabilities in Earth time.
Table 1 lists the steering corrections introduced to keep UTC in reason-
able agreement with UT1.

UTC is generated after the fact because of the goal to have its second
be as close as possible to the definition and the goals of uniformity and
reliability. These goals are achieved by taking the times of about 230
clocks from 65 different laboratories scattered around the world and
26
Table 1. Frequency Offsets and Step Adjustments of UTC,
until 1 July 1997

Date Offsets Steps

(at 0h UTC)

1961 Jan. 1 –150 × 10–10

1961 Aug. 1 " +0.050 s

1962 Jan. 1 –130 × 10–10

1963 Nov. 1 " –0.100 s

1964 Jan. 1 –150 × 10–10

1964 Apr. 1 " –0.100 s
1964 Sep. 1 " –0.100 s
1965 Jan. 1 " –0.100 s
1965 Mar. 1 " –0.100 s
1965 Jul. 1 " –0.100 s
1965 Sep. 1 " –0.100 s

1966 Jan. 1 –300 × 10–10

1968 Feb. 1 " +0.100 s
1972 Jan. 1 0 +0.107 7580 s
1972 Jul. 1 " –1 s
1973 Jan. 1 " –1 s
1974 Jan. 1 " –1 s
1975 Jan. 1 " –1 s
1976 Jan. 1 " –1 s
1977 Jan. 1 " –1 s
1978 Jan. 1 " –1 s
1979 Jan. 1 " –1 s
1980 Jan. 1 " –1 s
1981 Jul. 1 " –1 s
1982 Jul. 1 " –1 s
1983 Jul. 1 " –1 s
1985 Jul. 1 " –1 s
1988 Jan. 1 " –1 s
1990 Jan. 1 " –1 s
1991 Jan. 1 " –1 s
1992 Jul. 1 " –1 s
1993 Jul. 1 " –1 s
1994 Jul. 1 " –1 s
1996 Jan. 1 " –1 s
1997 Jul. 1 " –1 s

combining their readings in a near optimum way. This provides unifor-

mity (stability) and reliability. Additionally, the length of the second is
currently being determined by evaluations from 11 laboratory cesium-
beam primary frequency standards. A weighted combination of these
primary frequency standards is taken according to the individual
accuracies to obtain an overall world-best estimate for the second. The
time scale generated from combining this international set of clocks
and primary frequency standards from around the world is called
International Atomic Time (TAI — Temps Atomique International).
27
 Table 2. Relationship Between TAI and UTC

Limits of Validity TAI - UTC (in seconds)

(at 0h UTC)

1961 Jan. 1 - 1961 Aug. 1 1.422 8180 + (MJD – 37300) × 0.001 296
1961 Aug. 1 - 1962 Jan. 1 1.372 8180 + " "
1962 Jan. 1 - 1963 Nov. 1 1.845 8580 + (MJD – 37665) × 0.001 1232
1963 Nov. 1 - 1964 Jan. 1 1.945 8580 + " "
1964 Jan. 1 - 1964 Apr. 1 3.240 1300 + (MJD – 38761) × 0.001 296
1964 Apr. 1 - 1964 Sep. 1 3.340 1300 + " "
1964 Sep. 1 - 1965 Jan. 1 3.440 1300 + " "
1965 Jan. 1 - 1965 Mar. 1 3.540 1300 + " "
1965 Mar. 1 - 1965 Jul. 1 3.640 1300 + " "
1965 Jul. 1 - 1965 Sep. 1 3.740 1300 + " "
1965 Sep. 1 - 1966 Jan. 1 3.840 1300 + " "
1966 Jan. 1 - 1968 Feb. 1 4.313 1700 + (MJD – 39126) × 0.002 592
1968 Feb. 1 - 1972 Jan. 1 4.213 1700 + " "
1972 Jan. 1 - 1972 Jul. 1 10 (integral number of seconds)
1972 Jul. 1 - 1973 Jan. 1 11
1973 Jan. 1 - 1974 Jan. 1 12
1974 Jan. 1 - 1975 Jan. 1 13
1975 Jan. 1 - 1976 Jan. 1 14
1976 Jan. 1 - 1977 Jan. 1 15
1977 Jan. 1 - 1978 Jan. 1 16
1978 Jan. 1 - 1979 Jan. 1 17
1979 Jan. 1 - 1980 Jan. 1 18
1980 Jan. 1 - 1981 Jul. 1 19
1981 Jul. 1 - 1982 Jul. 1 20
1982 Jul. 1 - 1983 Jul. 1 21
1983 Jul. 1 - 1985 Jul. 1 22
1985 Jul. 1 - 1988 Jan. 1 23
1988 Jan. 1 - 1990 Jan. 1 24
1990 Jan. 1 - 1991 Jan. 1 25
1991 Jan. 1 - 1992 Jul. 1 26
1992 Jul. 1 - 1993 Jul. 1 27
1993 Jul. 1 - 1994 Jul. 1 28
1994 Jul. 1 - 1996 Jan. 1 29
1996 Jan. 1 - 1997 Jul. 1 30
1997 Jul. 1 - 31

The lengths of the seconds used in the generation of TAI and UTC are
the same; that is, they are based on the world’s primary frequency
standards in the same way. Though the frequencies are the same, the
times are not with the constraint that TAI minus UTC is an exact
integer number of seconds. The integer number changes as leap
seconds are introduced into UTC. Subtracting the right number of leap
seconds from TAI yeilds UTC. This may be confusing to some because
we speak of adding leap-seconds. But, in fact, we are subtracting a
whole SI second from the uniform and monotonically increasing
reading of TAI to obtain the leap-second-adjusted UTC. The appear-
ance of adding comes because there are 61 seconds in that minute
containing one. Watching a UTC clock with a stepping seconds hand
when this happens, we note the hand spends two seconds on

28
the 60 before moving on in a regular fashion. Table 2 shows the rela-
tionship between TAI and UTC. As can be seen, TAI – UTC = 31
seconds as of 1 July 1997.

While the source of the second for International Atomic Time (TAI) is
derived from the primary frequency standards throughout the world,
the flywheel to remember the calibrations provided by these standards
is the 230 or so contributing clocks. These clocks are like the counter,
adder, or accumulator for the world’s official time UTC. For example,
if one of the 230 clocks loses 20 nanoseconds per day as averaged over
the last month, there is a certain probability that this rate will continue
over the current month. Hence, this clock can be used to predict the
time based on the best estimate of the second as given by the primary
frequency standards. Each of the 230 clocks receives a weighting
factor according to its performance. The weighted average of the best
estimate of current predicted time across all of the available clocks
yields TAI.

It takes about one month to collect all of the data and to perform the
calculations to generate TAI and UTC times. In order to obtain a real-
time estimate of UTC, there are 50 timing centers around the world
that generate their own current estimate of UTC. (See Table 4 for list
of timing centers.) These are called UTC(k), where the “k” denotes the
particular timing center. For example, UTC(NIST) and UTC(USNO
MC) are the UTC estimates generated by the National Institute of
Standards and Technology in Boulder, Colorado and by the United
States Naval Observatory’s Master Clock in Washington, D.C. NIST has
responsibility for determining the length of the second for the United
States as well as supplying this value to the BIPM. USNO has responsi-
bility for supplying time and frequency for the Department of Defense
(DoD). Both organizations, in diverse and non-overlapping ways,
supply time for the United States and for the BIPM, and their time
scales are usually synchronous to within about 20 nanoseconds [12].

When the data are reported to the BIPM, each of the 230 contributing
clocks is reported with respect to the UTC(k) available to or generated
by that timing center. Each UTC(k) — including a clock at the Paris
Observatory — is measured over a five day average against GPS time.
When these time differences are subtracted, each UTC(k) is known
against the Paris Observatory clock using GPS as a transfer standard,
which drops out in the subtraction [13]. These time difference data,
along with the UTC(k) minus the local clock time differences, are
communicated to the BIPM for the latter’s calculation of TAI and UTC.
The BIPM then sends out a monthly bulletin, which reports the time
differences of each of the UTC(k)s with respect to UTC for the previ-
ous month to the interested user community. This bulletin is like a
“report card” telling each timing center how well they have done in
predicting UTC. When it is received, it tells you “officially” what time it
was! What time it is is predicted, estimated and made available to the
world from the different UTC(k)s in the contributing nations.

29
 By international agreement all of the timing centers have been given a
goal to keep their UTC(k)s within 100 ns (nanoseconds) of UTC.
Currently, the best predictions are approximately 10 ns. A plot of some
of the UTC(k)s is shown in Figure 7. Significant progress has been
made within several countries toward improving the accuracy of the
UTC(k)’s which provide a real-time estimate of UTC.
Figure 7. An example of some of the
UTC- UTC(K) 1996 Circular-T Data best estimates of UTC as predicted
and kept by a selected set of timing
400 centers, which also contribute data
Time Difference in Nanoseconds

toward the generation of UTC. The

200
agreed-upon international goal is to
0 keep each of the UTC(k)s within 100
ns of UTC. We see significant
–200 improvement in the predicted
accuracies of the UTC(k)s as time
–400 goes on. Since the timing centers have
to predict UTC about 45 days in
–600 advance, the performance of the
UTC(k)s allows you to calculate an
–800 implicit frequency stability of UTC of
January 1996 November 1996 about 2 × 10–15.

Belgium Canada Czech Republic France Japan

Korea Poland Spain United Kingdom USA

GPS Time and UTC

GPS has become the world’s principal supplier of accurate time [13]. It
is used extensively both as a source of time and as a means of transfer-
ring time from one location to another. There are three kinds of time
available from GPS: GPS time, UTC as estimated and produced by the
United States Naval Observatory, and the times from each free-running
GPS satellite’s atomic clock. The Master Control Station (MCS) at
Falcon Air Force Base near Colorado Springs, Colorado gathers the
GPS satellites’ data from five monitor stations around the globe. A
Kalman filter software program estimates the time error, frequency
error, frequency drift and Keplerian orbit parameters for each of the
satellites and its operating clock. This information is uploaded to each
satellite so that it can be broadcasted in real time. This process pro-
vides GPS time consistency across the constellation to within a small
number of nanoseconds and accurate position determination of the
satellites to within a few meters.

Because of this process, GPS cannot tolerate the introduction of leap

seconds. Hence, in 1980, when the Department of Defense started
keeping time on the GPS satellites, its system time and frequency were
set to agree with UTC(USNO MC). At that time, TAI minus UTC was
19 seconds (see Table 2). Since then, UTC has been delayed many leap
seconds and GPS time has not. Hence, GPS time is still very close to
TAI minus 19 seconds. The specification on GPS time is that it is to be
kept within one microsecond of UTC(USNO MC) modulo one second.
In other words, as a leap second is introduced into UTC(USNO MC)

30
time, no such step occurs in GPS time. But GPS time is still steered to
agree as well as possible with UTC(USNO MC), as if no leap seconds
had occurred since 1980. In practice, the steering performance is much
better than the one-microsecond specification; typically, it is well
within 40 nanoseconds.

In order to provide an estimate of UTC time derivable from a GPS

signal, a set of UTC corrections is also provided as part of the broad-
cast signal. This broadcast message includes the time difference in
whole seconds between GPS time and UTC. During 1996 GPS time
minus UTC time was 11 seconds. Also included in this message is the
rate and time difference estimate between GPS time and UTC(USNO
MC) modulo one second. This allows a receiver, in principle, to calcu-
late an accurate estimate of UTC(USNO MC). The mission goal is 28 ns
(1σ). Outside of the purposeful current degradation of the GPS signal
(called Selective Availability, SA) by the DoD for security purposes,
this calculation may have an accuracy of about 10 nanoseconds (ns) on
an rms basis [14]. Since USNO has been successful in predicting UTC
to within about 10 ns, combining these two independent error sources
yields a real-time potential uncertainty for UTC available from GPS at
about the 14-ns level. In practice, SA prohibits achieving this accuracy
level unless special clock systems and filtering techniques are em-
ployed. The SA degradation can be filtered away, as will be discussed
later [15] (see the section below and Appendix B, see Appendix A for
the meaning of uncertainty, and see Figure 14 for an illustration of SA
filtering).

Accuracy and Stability of UTC

The accuracy and stability of UTC have improved dramatically over the
last few years. This is primarily due to the introduction of the Hewlett-
Packard HP 5071A Cesium Beam Clock. Along with having about a
factor of ten better accuracy than any other commercial frequency
standard, this clock also has more than a factor of ten less sensitivity
to the environment — giving it outstanding long-term stability —
usually well below 1 × 10–14. The work of Leonard Cutler and Robin
Giffard of Hewlett-Packard Laboratories and the design team of the HP
Santa Clara Division, in regard to solving the environmental sensitivity
problem in cesium-beam clocks, is reminiscent of what John Harrison
did for mechanical clocks. Because of their outstanding performance,
these clocks were rapidly introduced into the UTC ensemble member-
ship by the different laboratories as they became available. As a result,
the stability of UTC has improved by about a factor of ten from 1991 to
1996. With these same clocks, the timing centers are now able to do a
much better job of predicting UTC. This work could well be the basis
for eventually arriving at a highly accurate and extremely stable real-
time UTC. Then we could know precisely what time it is!

For 1995, Figure 8 shows the frequency distribution of all the clocks
contributing to UTC. Figures 9 and 10 show the distributions of the
HP 5071As and the primary frequency standards, also contributing

31
 Figure 8. A histogram showing the
Histogram of Frequency Offsets (TAI- Clock) for 1995 frequency accuracy of all the clocks
(All Clocks) contributing to TAI and UTC with
180 histogram box size of 5 × 10–13.
# of units = 288
160
Mean Offset = 0.5E–13
# of Units per Frequency Bin

140 std dev = 11.6E–13

std dev of mean = 0.7E–13
120

100

80

60

40

20

0
–30 –25 –25 –20 –15 –5 0 5 10 15 20 25
Clock Frequency Offset (TAI - Clock), parts in 10E13

Figure 9. A histogram showing the

Histogram of Frequency Offsets (TAI- Clock) for 1995 frequency accuracy of the HP 5071As
(HP 5071As) that contribute to TAI and UTC. The
120 histogram box size is 5 × 10–13. Notice
# of units = 112 that the mean offset is 24 times better
Mean Offset = 0.4E–13 than the HP accuracy specification for
# of Units per Frequency Bin

100
std dev = 1.38E–13 the HP 5071A, and the standard
std dev of mean = 0.13E–13 deviation is 7 times better.
80

60

40

20

0
–30 –25 –25 –20 –15 –5 0 5 10 15 20 25
Clock Frequency Offset (TAI - Clock), parts in 10E13

Figure 10. A histogram showing the

Histogram of Frequency Offsets (TAI- Clock) for 1995 frequency accuracy of the primary
(Lab Primary Standards) frequency standards that contribute to
5 TAI and UTC. The histogram box size
# of units = 5 is 1 × 10–13. The mean offset is mostly
Mean Offset = 0.18E–13 due to the 1996 CCDS decision to
# of Units per Frequency Bin

4 std dev = 0.47E–13 include the “black-body radiation”

std dev of mean = 0.22E–13 shift correction.
3

0
–5 –4 –3 –2 –1 0 1 2 3 4 5
Clock Frequency Offset (TAI - Clock), parts in 10E13

32
to UTC. Even though the standard deviation of the mean of the
HP 5071As’ frequencies is about the same as that of the primary
frequency standards throughout the world, one may ask the question
about the statistical independence of same-model-number devices.
The standard deviation shown for the HP 5071As’ frequencies is
seven times better than the manufacturer’s accuracy specification of
1 × 10–12, and the mean value only differs by –5 × 10–14. As can be seen
from the pie chart in Figure 11, the percentage of weight assigned to
the HP 5071As is more than 2/3 that for the entire UTC ensemble —
even though the total number is less than half of all the contributors
to UTC. This, again, is because of the excellent long-term frequency
stability and time predictability of these clocks.

Also, in the last few years, the accuracies available from cesium
primary frequency standards has improved remarkably. The emerging
technologies that have brought about these improvements in cesium
primary frequency standards are laser energy-state pumping and
detection, laser cooling using photon pressure down to near absolute
zero temperature, followed by a photon pressure pulse giving rise to a
controlled low-velocity cesium fountain. Nearly a factor of ten im-
provement in accuracy has already been achieved and the end is not in
sight. The cesium fountain work has been led by Andre Clairon at the
Laboratoire Primaire du Temps et des Fréquences [LPTF] in Paris,
France, and the laser energy-state cesium selection and detection by
Robert Drullinger at the National Institute of Standards and Technol-
ogy (NIST) in Boulder, Colorado. These have been major break-
throughs since Lyons developed the first cesium-beam device in the
early 1950s.
Figure 11. A pie chart showing the
distribution of clocks and the assigned
BIPM International Atomic Time weighting factors assigned by the BIPM
Component Clocks by Clock Weight for the clocks composing the UTC time
scale for the month of December 1996.
OSA (0.85%) FTS (0.03%)
Primary Clocks (2.04%)
Hydrogen (18.32%)
HP 5071A (72.96%)

HP 5061A/B (5.80%)

(Data as of 2/25/97)

33
 The variance used to determine the weighting factors for the TAI and
UTC computation is the frequency stability taken over an averaging
time of two months, and six values (one year’s worth of data) are used
to compute a six-sample variance. The weighting factors shown in
Figure 11 are proportional to the reciprocal of the variance except that
they are not allowed to exceed an upper limit of weight.

The primary standards are located in Canada (NRC CsV, CsVI A and
CsVI C), France (LPTF JPO and FO1), Germany (PTB CS1, CS2, and
CS3), Japan (CRL Cs1), Russia (SU MCsR 102) and USA (NIST-7). The
high accuracy and continuous operation of the PTB primary frequency
standards has made them major contributors to the accuracy of the SI
second used in TAI and UTC [16].

Table 3 lists the uncertainties of the primary frequency standards that

are used to determine the second for UTC. The current stated accuracy
given by the BIPM for the second used in the generation of UTC is
–2 × 10–14 with an uncertainty of ±1 × 10–14. As stated above, the only
clock that is correct all of the time is the one we agree to call so. Even
though there are some estimated errors in UTC, its time is correct by
definition. This small inaccuracy is due to a recent decision to include
an additional correction called the “black-body radiation shift.

The frequencies of these primary frequency standards are corrected for

the gravitational shift (of about 1 × 10–13 for an altitude of 1000 m), and
for the black-body radiation shift (of about 2 × 10–14 for a temperature
of 40°C) when available (standards tagged with an *).

Table 3. Uncertainties of the Primary Standard

The characteristics of the calibrations of the TAI frequency provided by
the different primary standards are as follows (file available via Internet:
UTAI96.AR):
Standard Unc. (1 σ) Operation Comparison Transfer
with to TAI

CRL Cs1* 1.1 × 10–13 discontinuous UTC(CRL) 60 d

LPTF JPO* 1.1 × 10–13 discontinuous UTC(OP) 10 d
LPTF FO1* 0.3 × 10–14 discontinuous H maser 5d or 10 d
NIST NIST-7* 0.5 × 10–14 discontinuous H maser 5d or 10 d
NRC CsV ≅ 1 × 10–13 continuous TAI 60 d
NRC CsVI A ≅ 1 × 10–13 continuous TAI 60 d
NRC CsVI C ≅ 1 × 10–13 continuous TAI 60 d
PTB CS1* 3 × 10–14 continuous TAI 60 d
PTB CS2* 1.5 × 10–14 continuous TAI 60 d
PTB CS3* 1.4 × 10–14 continuous TAI 60 d
SU MCsR 102* 5 × 10–14 discontinuous UTC(SU) 60 d

34
The black-body radiation shift was discovered, theoretically, by Wayne
Itano of NIST in 1982 [17]. It is due to the isotropic black-body radiation
emitted from the surroundings of the atomic beam in a cesium atomic
clock. The size of the black-body radiation shift is only –1.76 × 10–14 at
300°K (27°C). Recently, two laboratories developing the next genera-
tion of cesium fountain clocks have measured this shift, their results
confirming theory [18,19]. Because the theoretical basis of the shift was
sufficiently strong, the March 1996 CCDS (Comité Consultatif pour la
Définition de la Seconde) decided to include this effect. This resulted in
a step change in the SI second of about this amount. TAI and UTC are
being gradually steered to agree with this improved value.

Although it does not exist in nature, we may conceptualize an ideal

clock. If three or more independent clocks are compared, we may
estimate the individual instabilities of each with respect to this ideal
clock. A clock’s instabilities are directly related to its predictability.
Figure 12 shows the predictability of a wide variety of clocks, or their
timekeeping ability. The different slopes are indicative of different
kinds of random processes perturbing the timing data coming from that
particular clock.
Figure 12. A comparison of the
For instance, a horizontal slope results from being limited by measure- timekeeping ability of a wide variety of
ment noise, which is the first kind of perturbation illustrated in clocks as a function of the time since
being synchronized and syntonized.

1 000 000s
Stonehenge
Egg Timer
1 000s Earth Spin

Harrison Clocks
1 Second
ch
at Earth Orbit
W
nt
Clock Timekeeping Ability

e
em h
0.001s cap a t c Rb Gas Cell
(Millisecond) Es tz
W
u ar m
Q ulu
nd
0.000 001s Pe Millisecond Pulsar
(Microsecond) r
Super to
la
Conducting s c il UTC
Cavity a lO
0.000 000 001s y st
Cr Cs Beam
(Nanosecond) tz
ar
Qu H-Master
Hg Ion Storage
0.000 000 000 001s
(Picosecond) 6000 Years
1 1 1 100
Hr Day Yr Yrs 1 Million Years
0.000 000 000 000 001s
(Femtosecond) 10–3 1 103 106 109 1012 1015
1 Min

Time Since Synchronization (seconds)

35
 Figure 2. Data with the slopes upward are indicative of the limiting
internal random perturbations affecting that clock; these are the
second type illustrated in Figure 2. The third type, the effects of the
environment, can be enormously important, but because of their large
variety and complexity, these effects have been ignored, as much as
possible, in this figure. The range of data shown is taken from several
different experiments and is intended to suggest a nominal range of
coverage for each designated clock type. The long-term data for Earth
spin as a clock are taken from S. K. Runcorn’s work using coral growth
as paleontological clocks; in his estimates, for example, 600 million
years ago the earth had 425 days in a year. Also plotted are estimates of
the timekeeping ability of UTC. The short-term performance of UTC
will improve as better time transfer techniques are perfected. The long-
term performance will improve as the contributing clocks and comput-
ing algorithms improve.

For historical perspective, data from the revolutionary Harrison-like

chronometers is included for comparison. See Appendix B for a more
detailed analysis of the Harrison-like chronometers. Also shown is the
current estimate of the timekeeping ability of UTC as a composite of
many clocks. The low slope portion on the left is due to measurement
noise, and the low slope portion on the right is the benefit of having
primary frequency standards keeping the UTC second tied to the SI
second. As it has in the past, UTC will undoubtedly continue to im-
prove as time goes on.

In the ideal combining algorithm, the computed time of a clock en-

semble can have better stability than the best clock in the ensemble.
In order to accomplish this goal, the clocks have to be characterized
individually; then each clock can be included in the algorithm in an
optimum way. There are two practical problems with this procedure. It
takes a significant amount of time to characterize each clock; this is
one of the reasons the BIPM takes about a month to calculate UTC.
Then, there is significant measurement noise in bringing all of the
clock data into one place as is needed for the generation of UTC.
Waiting for a month has the benefit of allowing the better detection
and removal of anomalous errors from the data than if the time were
shorter. With the clocks distributed across the globe, transporting their
times to the BIPM without perturbation is a significant challenge at the
nanosecond level of accuracy. Waiting also has the advantage of
allowing time to average and filter the time transfer noise.

Einstein’s Relativity and Precise Timekeeping

The high level of accuracy now being attained with clocks requires that
relativity be included in the comparisons and computations of time and
frequency relationships. For example, relativity is an engineering
reality in the design and operation of GPS: GPS would not work
without its inclusion [20, 21].

Relativity, without going into the mathematics, deals with three

problems: the relative motion of clocks; differences of gravitational
36 potential in which clocks find themselves; and the problem of rotating
reference frames, such as the earth on which we live. Previously, we
used the term “Coordinated” — indicating an international cooperative
effort. In this section, we will be using the concept of coordinate time,
which refers to a consistent set of relativistic timing procedures. Thus,
clocks distributed in space and time can be compared and used harmo-
niously without confusion as to what a timing signal is.

A fundamental principle in relativity is the constancy of the speed of

light, which states that electromagnetic signals travel in a vacuum at a
unique speed in an inertial (non-rotating constant velocity) reference
frame (c = 299,792,458 meters per second). This principle allows one to
extend a network of self-consistently synchronized clocks throughout
an inertial reference frame. During the first part of our century,
G. Sagnac showed that in a rotating (non-inertial) reference frame, the
velocity of light is not constant. As a consequence, it is not possible to
establish a self-consistent system of synchronized clocks over the
earth’s surface simply assuming constancy of c as viewed from our
rotating earth. Instead it is useful to imagine a reference frame with
origin at the earth’s center but which is not rotating — a so-called
Earth-Centered Inertial (ECI) frame. Earth spins once per sidereal day
with respect to this ECI frame. Hypothetical clocks at rest in this frame
can be self-consistently synchronized and syntonized using constancy
of c. However, clocks in satellites or at rest on the earth’s surface are in
motion through this ECI frame and move through Earth’s gravitational
potential. Therefore corrections to atomic clock readings must be made
to compensate for motional and gravitational effects so that real clocks
read the same time as the hypothetical clocks that are at rest in the
underlying ECI frame. This results in a time called “coordinate time.”
GPS time is an example of such a coordinate time.

The reference for coordinate time for Earth has been given us by
international conventions. In 1971 the General Conference of Weights
and Measures stated that: “International Atomic Time (TAI) is the time
reference coordinate established by the Bureau International de l’Heure
on the basis of the readings of atomic clocks operating in various
establishments in accordance with the definition of the second, the unit
of time of the International System of Units.” In 1988, the responsibility
for TAI was transferred to the Time Section of the Bureau International
des Poids et Mesures, BIPM. In 1980, the Comité Consultatif pour la
Définition de la Seconde, (CCDS) declared: “TAI is a coordinate time
scale defined in a geocentric reference frame (origin of the frame at the
centre of the Earth) with the SI second as realized on the rotating geoid
as the scale unit.” Here SI stands for the International System of Units.

Earth has the shape of an oblate spheroid due to centrifugal force as it

spins. Its average equatorial radius is larger than its polar radius by
21,476 meters. The geoid mentioned above is that surface where the
total effective gravitational potential is the same, which nominally
follows this oblate shape and is approximately the same as sea level.
Ideal clocks at rest anywhere on the rotating geoid will tick at the same
rate. The SI second as currently given by the cesium atom adjusted to
the rotating geoid defines the second for TAI. Since all primary fre-
quency standards laboratories are actually above the geoid, a correc- 37
 tion has to be applied for the height at which each of these standards
rests. The primary frequency standard for the U.S., NIST-7, in Boulder,
Colorado runs fast with respect to an ideal clock at the geoid by about
1.8 × 10–13, which is very significant given its current accuracy of
0.05 × 10–13. The size of the effect is about 1 × 10–13 per kilometer of
elevation above the geoid.

For a clock fixed on Earth at distance r from the earth’s center and at
geocentric latitude θ, the fractional frequency shift is

Δ ν ν =  V r , θ – Ω E2 r 2 cos 2 θ / 2 –  V a 1 , 0 – Ω E2 a 12 / 2   / c 2
( ) ( )
 
where V(r,θ) is the earth’s gravitational potential including quadrupole
and perhaps higher multipole moment contributions, ΩE is Earth’s
angular rotation rate (one cycle per sidereal day), and a1 is Earth’s
equatorial radius. The terms proportional to Ω E2 arise because of the
spin of the earth. They can be thought of as contributions to the
effective gravitational potential in the rotating frame. Since clocks at
rest anywhere on the geoid tick at the same rate, it is convenient to
evaluate the last correction term along with the true gravitational
potential V(a1,0) at the earth’s equator. If the earth-fixed clock is at
height h above the geoid, the combination of terms gives a fractional
frequency shift of approximately gh/c2 where g is the measured value
of the acceleration of gravity.

For a clock moving with velocity v through the ECI frame the term
1 Ω E2 r 2 cos 2 θ must be replaced by 1 2 v 2 . Einstein’s second-order
2
Doppler correction states that a clock moving at a velocity “v” with
respect to an inertial frame will appear to run slow when compared
with synchronized clocks in the inertial frame by a fractional amount
2 2
“ v  2 c  ”. This is the reason for the second and fourth terms in the
above correction equation. For clocks in GPS orbits, the second-order
Doppler corrections cause the clocks to run slow by 0.823 × 10–10 with
respect to clocks at the geoid. On the other hand, the purely gravita-
tional frequency shifts cause GPS satellite clocks to run fast with
respect to clocks at the geoid by 5.289 × 10–10. The sum of these two
effects is 4.46 × 10–10. GPS satellite clocks are slowed by this amount
before launch in order for the broadcast signals to be correct when
used anywhere near the earth.

Since the GPS orbits are not perfectly circular, the eccentricity of each
satellite’s orbit is broadcast in its data message. The GPS receiver
software has to calculate the additional relativistic effects due to this
eccentricity. The magnitude of this effect can be several dozens of
nanoseconds for GPS orbits. The eccentricities of Russia’s equivalent
system to GPS, called GLONASS, tend to be much smaller than for
GPS. It appears that the very small eccentricity effect in the GLONASS
system is accounted for at the transmitters by modulation of the
transmitted clock coefficients. GLONASS was designed so that the
receivers can get by without having to calculate this relativistic effect.

38
Clocks being compared at different longitudes on Earth’s surface have
to include the relativistic Sagnac effect due to the rotation of the earth.
The size of the effect is given by 2ΩEAp/c2 = Ap × 1.6227 nanoseconds
per square megameter (Mm2) where ΩE is the angular velocity of the
earth. Ap is the total area, projected on Earth’s equatorial plane,
mapped out by the radius vector from the center of the earth to the
portable clock or to the electromagnetic signal carrying the time. The
correction is positive going eastward. In other words, if a perfect
portable clock were transported eastward around the globe on the
geoid, so slowly that its velocity relative to the ground didn’t matter,
then when it returned to its starting point, 207.4 nanoseconds would
have to be added to its reading for it to agree with its reading if it had
been left at the point of departure. The circumference of Earth is about
40 Mm — giving it a cross sectional area of 127.8 Mm2 at the equator
(127.8 × 1.6227 = 207.4 ns). For receivers at known locations such as
timing centers, when GPS time is transferred by a signal from satellite
to ground, the Sagnac effect has to be programmed into the receiver in
order to estimate the projected area of the triangle having corners at
the GPS satellite sending the signal, the receiver’s location, and the
center of the earth.

The question arises, “what about the effects due to the moon and the
sun and the fact that Earth is not in a circular orbit?” These effects are
well understood. The reference frame of choice for relativistic correc-
tions on and about the earth is a non-rotating Earth-centered frame —
an ECI frame. Fortunately, since our ECI frame is in free fall about the
sun, Einstein’s Principle of Equivalence implies that for near-earth
clocks relativistic corrections as viewed from our ECI frame remain the
same over the course of a year at very high levels of accuracy. This self
consistency holds for clocks being compared in the vicinity of the earth
[22, 23, 24]. As soon as timing measurements are made outside our
Earth-Moon system, other coordinate systems and relativistic consider-
ations enter in.

How to Access UTC

Since UTC is not directly available as a clock, real-time approximations
to it are made available from 50 timing centers around the world. A list
of the timing centers generating a UTC(k) time scale is provided in
Table 4 as taken from the BIPM annual report. Most of the UTC(k)s are
kept within one microsecond of UTC. A good percentage of them keep
within 100 nanoseconds of UTC, and a few of them are usually within
10 ns of UTC.

As UTC continues to improve in its stability, and as the different

UTC(k) time scales also become more stable, the errors in predicting
UTC will continue to get smaller. One of the biggest problems in
accessing UTC is not the accuracy of the source but rather the instabili-
ties in the time and frequency transfer techniques.

39
 Table 4. Acronyms and Locations of the Timing Centers Which Maintain
a Local Approximation of UTC, UTC(k), or/and an Independent Local
Time Scale, TA(k)

AOS Astronomiczne Obserwatorium Szerokosciowe, Borowiec, Polska

APL Applied Physics Laboratory, Laurel, MA, USA
AUS Consortium of laboratories in Australia
BEV Bundesamt für Eich - und Vermessungswesen, Wein, Oesterreich
BIRM Beijing Institute of Radio Metrology and Measurement, Beijing, P.R. China
CAO Cagliari Astronomical Observatory, Cagliari, Italia
CH Consortium of laboratories in Switzerland
CNM Centro Nacional de Metrologia, Queretaro, Mexico
CRL Communications Research Laboratory, Tokyo, Japan
CSAO Shaanxi Astronomical Observatory, Lintong, P.R. China
CSIR Council for Scientific and Industrial Research, Pretoria, South Africa
F Commission Nationale de l’Heure, Paris, France
DLR Deutsch Forschungsanstalt fuer Luft-und Raumfahrt, Oberpfaffenhofen, Germany
DTAG (Formerly FTZ) Deutsch Telecom AG, Darmstadt, Deutschland
GUM Glówny Urzad Miar, Central Office of Measures, Warszawa, Polska
IEN Istituto Elettrotecnico Nazionale Galileo Ferraris, Torino, Italia
IFAG Institut für Angewandte Geodäsie, Frankfurt am Main, Deuschland
IGMA Instituto Geografico Militar, Buenos-Aires, Argentina
INPL National Physical Laboratory, Jerusalem, Israel
IPQ Institute Portugués da Qualidade (Portuguese Institute for Quality),
Monte de Caparica, Portugal
JATC Joint Atomic Time Commission, Lintong, P.R. China
KRIS Korea Research Institute of Standards and Science, Taejon, Rep. of Korea
LDS The University of Leeds, Leeds, United Kingdom
MSL Measurement Standards Laboratory, Lower Hutt, New Zealand
NAOM National Astronomical Observatory, Misuzawa, Japan
NAOT National Astronomical Observatory, Tokyo, Japan
NIM National Institute of Metrology, Beijing, P.R. China
NIST National Institute of Standards and Technology, Boulder, CO, USA
NPL National Physical Laboratory, Teddington, United Kingdom
NPLI National Physical Laboratory, New-Delhi, India
NRC National Research Council of Canada, Ottawa, Canada
NRLM National Research Laboratory of Metrology, Tsukuba, Japan
OMH Orszagos Mérésügyi Hivatal, Budapest, Hungary
ONBA Observatorio Naval, Buenos-Aires, Argentina
ONRJ Observatorio Nacional, Rio de Janeiro, Brazil
OP Observatoire de Paris, Paris, France
ORB Observatoire Royal de Belgique, Bruxelles, Belgique
PTB Physikalisch-Technische Bundesanstalt, Braunschweig, Duetschland
RC Comité Estatal de Normalizacion, Habana, Cuba
ROA Real Instituto y Observatorio de la Armada, San Fernando, Espana
SCL Standards and Calibration Laboratory, Hong Kong
SNT* Swedish National Time and Frequency Laboratory, Stockholm, Sweden
SO Shanghai Observatory, Shanghai, P.R. China

40
Table 4. (continued)

SU Institute of Metrology for Time and Space (IMVP), NPO “VNIIFTRI”

Mendeleevo, Moscow Region, Russia
TL Telecommunication Laboratories, Chung-Li, Taiwan
TP Institute of Radio Engineering and Electronics, Academy of Sciences of
Czech Republic-Czech Republic
TUG Technische Universität, Graz, Oesterreich
UME Ulusai Metroloji Enstitüsü, Marmara Research Centre, National Metrology
Institute, Gebze-Kocaeli, Turkey
USNO U.S. Naval Observatory, Washington, D.C., USA
VSL Van Swinden Laboratorium, Delft, Nederland

* SNT ceased its time activities in May 1995.

Figure 13 shows how long an averaging time is needed to obtain a

certain level of stability and traceability via a given time and frequency
transfer technique. If the limiting transfer or distribution noise is white-
noise PM, then Mod.σy(τ) was used to show both the kind and level of
noise. In this case notice that the slope is proportional to τ − 3 2 , and
for a normal distribution of errors, the 68 percent uncertainty on
frequency transfer or distribution accuracy is given by 2 × Mod.σy(τ).
As mentioned in the beginning of this article, frequency stability is a
measure of the change of the frequency from one period of time to the
next. The length of this period of time is called the averaging time, τ. A
particular value of τ is chosen, and then the frequency stability is
ascertained over a data set for a given method of time and frequency
dissemination or transfer. Notice that essentially all techniques im-
prove with increased averaging time.

Figure 13. A plot of the frequency

Tel. Reciprocity* instabilities of traditional means of
300 Baud Modem transferring or distributing frequency.
Tel. Reciprocity* Dashed line shows experimentally
10 –6 verified potential for GPS Carrier
WWV or WWVH Phase technique. Telephone
10 –7 Reciprocity with 300 Baud Modem was
a special experiment within a calling
10 –8 area and staying “on hook” for several
Fractional Frequency Stability, σy(τ)

hours.
10 –9 TV Line-10
–10
*and Mod.σy(τ)

10
GOES
10 –11
WWVB
10 –12
Loran-C Time Transfer via
10 –13 2-way TV Line-10 or Loran-C
Satellite* GPS Carrier
Phase* GPS Direct
10 –14
GPS Common View
10 –15 and 2-way Satellite*

10 –16
10 0 10 1 10 2 10 3 10 4 10 5 10 6 10 7 10 8
Sample Time, τ (seconds)

41
 State-of-the-art techniques are used to communicate the time and
frequency data to the BIPM for the calculation of UTC. With the
frequency stability of clocks continually improving, the short-term
instabilities of the transfer techniques have increasingly become a
problem and limit the short-term instabilities achievable for UTC. The
initial flat portion of the UTC curve in Figure 12 is due to the instabili-
ties of the time transfer technique. Figure 14 gives the time stability of
some state-of-the-art time transfer techniques.

The principal operational means of communicating the times of the

contributing clocks to the BIPM is the GPS common-view technique
which may be explained by the following example. This technique has
been automatically designed into special GPS timing receivers [13].
Suppose, for example, the time of a clock in Boulder, Colorado is being
communicated to the Paris Observatory. At a predetermined time
(preset in software) the clock in Boulder measures its time difference
with respect to GPS time, B – G. At the same time the GPS receiver in
Paris, listening to the same satellite, measures its time difference with
respect to GPS time, P – G. The GPS signal for the same satellite is
averaged over the same 13-minute time window, then these measure-
ments are exchanged and subtracted, yielding B – P time difference.
With all the delay corrections properly accounted for, the accuracy of
this technique has been shown to be a few nanoseconds. Notice that
the effects of SA drop out in the subtraction to the extent that SA is
due to satellite clock dither, as do many other common-mode errors.
Even though the GPS common-view technique presently limits the
short-term stability of UTC, it improved the accuracy and stability of
time-transfer by more than a factor of 20 when it was introduced in the
early 1980s. Loran-C had provided the best operational time-transfer
technique prior to GPS.

Figure 14. A plot of the time stability of

Time Stability some of the state-of-the-art time
transfer techniques as compared with
0

Mod. σy(τ) some of the more traditional

–9

–1

–1

–1

–1

4
10

10

10

10

10

–1

µs 10 –6 techniques.
10

Loran -C
5
–1

10 –7
(seconds)

10
6
–1

10 –8
10

G PS QU
ACV RB GPS
7
–1

ns 10 –9 CS CV
10

TW
STF
T
8
–1

10 –10
10
σx(τ)

S
–1

10 –11 EGP GPS Carrier Phase

10

1 min. 1 hour 1 day 1 year

ps 10 –12
10 0 10 1 10 2 10 3 10 4 10 5 10 6 10 7 10 8

Sample Time, τ (seconds)

42
In recent years, there has been a rapid increase in the use of GPS as a
source of UTC. With minimal effort and little cost, one-microsecond
accuracies are readily achieved. There are many applications, however,
where much better timing is needed, and most of these situations need
a real-time output, such as for telecommunications or the power grid.
However, the degradation in the GPS signal caused by SA can cause
some hundreds of nanoseconds peak-to-peak variations. The current
GPS common-view technique, which avoids the SA degradation to a
large extent, only gives time differences between two clocks after the
fact.

A novel approach was developed a few years ago in which the nature of
the instabilities of SA was studied [15]. Understanding the character of
these instabilities as compared with the very different character of
the instabilities in precision clocks allowed real-time filters to be
designed for timing receivers in fixed locations. Using, for example, the
HP 5071A cesium-beam clock with an appropriate filter design essen-
tially eliminates the SA in real-time. If clocks using quartz crystal
oscillators or rubidium gas-cell frequency standards are used instead, a
good level of SA filtering is still achievable. The clock using rubidium
offers better filtering than one using a quartz-crystal oscillator, but the
latter can be built more cost-effectively and with better reliability.
Products based on these SA filtering concepts have become very
popular and useful for precision network timing. Figure 14 illustrates
the time stability achievable using these SA filtering concepts as
compared with some other precision timing techniques. They are
denoted as the Enhanced (EGPS) technique.

ACV denotes Advanced Common-View technique. The ACV technique

utilizes the degrees of freedom available from the newer GPS timing
receivers. These include the one-second data, the several satellites one
can always observe at a given site, and the convenient methods now
available for exchanging data rapidly. In contrast, the original common-
view (CV) approach only uses one satellite at a time over 13-minute
averages and is based on a one-day average, post analysis. A simple
degrees-of-freedom argument shows more than a factor of 50 advantage
of the ACV technique over the CV technique; however, the research and
development are far enough along on the ACV technique to determine
that there are correlations in the data that take away some of these
degrees of freedom. Still the actual performance obtained is very
encouraging. It has the potential to provide short-term performance in
real-time and could be made fully automatic. It cancels the SA degrada-
tion to a large extent. Hardware and software development compatible
with the experimental results is still needed.

TWSTFT denotes Two-Way Satellite Time and Frequency Transfer

technique. It exhibits excellent short-term stability — limited by only
two or three hundred picoseconds of white-noise PM taken over one-
second averages. However, the environmental effects on the transmit
and receive equipment cause the slope upward and the degradation in
the longer-term performance. Improvements are being worked on and
are expected for TWSTFT.
43
 Future Timing Techniques
EGPS denotes the Enhanced GPS technique. It is a systems approach
in which the attributes of the reference clock are best utilized to filter
the effects of the GPS SA degradation in order to obtain a real-time
estimate of UTC. Note that different levels of filtering are obtained
depending on whether the reference clock uses a quartz-crystal
oscillator (QU), a rubidium gas-cell frequency standard (RB), or
cesium-beam frequency standard (CS). The CV denotes the traditional
GPS Common-View technique, which is used to transfer the times and
frequencies of most of the standards contributing in the generation of
TAI and UTC. It removes the effects of SA to a large extent and em-
ploys one-day averages — providing an after-the-fact solution. Data
acquisition is fully automatic. Improvements are anticipated by improv-
ing GPS receiver stabilities and accuracies and by better characteriza-
tion of atmospheric delays. Both GPS CV and GPS ACV use the broad-
cast time code. Using instead the phase of the GPS carrier frequency
for transferring time and frequency information between two sites
remote to each other shows significant promise, but it is still in the
research phase. The results plotted are between hydrogen maser
clocks located in Goldstone, California and Algonquin Park, Canada
separated by four megameters (4,000 km or 2,500 miles). Loran-C has
been the principal means for navigation for several decades. Even
though it is used for timing, the signal is not self-contained in that time-
of-day information is not available and another source, such as WWV,
is needed to resolve the time ambiguity. The carrier signal is at 100 kHz
(10 microsecond period). For Figure 14, the ordinate is the time
stability, the abscissa is the amount of averaging time used in measur-
ing the change with adjacent averaging periods. Also shown are the
decade values for Mod.σy(τ) in order to have a measure on the same
graph of the uncertainty in frequency transfer accuracy.

A current International Telecommunications Union (ITU-R) handbook,

entitled “Selection and Use of Precise Frequency and Time Systems,”
contains a set of tables summarizing the various techniques for time
and frequency distribution and for comparison of clocks remotely
located from each other. By permission, these tables are included in
Appendix C.

T2L2 (Time Transfer by Laser Link) is a dedicated time-transfer

experiment under development at the Observatoire de la Côte d’Azur
and the Centre National d’Etudes Spatiales in France [25]. T2L2 is
designed to synchronize remote clocks with precision on a single
measurement of 50 picoseconds and an accuracy in the 50 picoseconds
region. Such performances have never been reached and have many
technological (navigation, telecommunication, positioning) and
scientific (gravitation, solar quadrupole momentum) applications. The
principle of this time-transfer technique is based on the propagation of
light pulses between the clocks for synchronization, as in the Laser
Synchronization from Stationary Orbit (LASSO) experiment, but
without external calibration. LASSO has been successfully tested in

44
1992 between McDonald (Texas) and Grasse (France), with a stability
of 100 picoseconds and an accuracy on the order of 1 nanosecond. The
light pulses carry the temporal information from one clock to another.
A clock on board a satellite is used as a relay between Earth clocks to
allow time transfer between remote clocks. One could also use the
technique to place a reference clock on the satellite and use the optical
time-transfer to synchronize the Earth clocks to this reference.

The usual time-transfer techniques are based on the propagation of a

wave in the microwave domain: GPS Common-View and Two-Way
Satellite Time- and Frequency-Transfer. The interaction between the
radio-frequency signal and the atmosphere presently does not allow an
accuracy much better than 1 nanosecond with an ultimate value
estimated as around 100 picoseconds. The optical wave propagation is
well controlled, and it is for this reason that the optical time-transfer
technique achieves such accuracy and precision. The transmission path
delay can be calibrated much more accurately than that for microwave
signals. Unfortunately, it is weather-dependent — clouds will block the
path.

To perform a T2L2 time-transfer, laser stations and a satellite are

needed, both equipped with a clock and time-tagging unit. The T2L2
experiment allows the monitoring of the space clock from a ground
clock and the transfer of time between the ground clocks via the
satellite clock. The laser stations emit some light pulses in the direction
of the satellite. An array of retroreflectors returns a fraction of the
received photons to the stations (the photons are returned along the
same direction). The stations record the start times of the light pulses
and the return times after reflection from the satellite. The satellite
time-tagging unit measures the onboard arrival time of the pulses.
Then, for a given light pulse emitted from station A, the offset XA
between clock A and the satellite clock can be computed. For another
light pulse emitted from station B, the offset XB can also be computed.
The time-transfer between the clocks A and B is deduced from the
difference in the offsets XA and XB. Station and satellite data are

Figure 15.
Time-transfer
stability between
40 ps a ground clock and
the satellite clock
for T2L2.
15 ps
σx(τ)

45 fs

0.1 1000 300,000

τ (seconds)

45
 regularly transmitted to the T2L2 coordination center for analysis.
The accuracy of the time-transfer between the satellite clock and the
ground clock integrated over ten days is estimated at 50 picoseconds,
considering that the ground and satellite contributions are equal. This
implies that the frequencies of the ground clock and the satellite clock
can be compared with an accuracy in the range of 6 × 10–17. However,
it must be stressed that the time origin at the satellite is arbitrary, so it
is not possible to know the phase between the ground clock signal and
the satellite clock signal.

Plans are in hand to test the next generation of atomic clocks and
time-transfer techniques in space. Scheduled for launch around 2003,
Atomic Clock Ensemble in Space (ACES) is a multi-Agency proposal
between the United States, the Russian Federation, Japan, Europe and
Canada to fly an ensemble of the next generation of atomic clocks
onboard the International Space Station Alpha. ACES would consist
of an ensemble of externally-mounted payloads. In the benign,
microgravity environment of space, the fractional frequency stability
and accuracy attainable should be in the range of 10–16 to 10–17.

None of the next generation of atomic clocks has flown in space to

date, with the exception of the hydrogen maser in 1976 on board the
Scout Rocket during the two-hour red-shift experiment for the Gravity
Probe A mission [22]. Future atomic clocks (active-cavity hydrogen
maser, cesium atomic fountain, mercury linear ion-trap) are all capable
of ground-based frequency stabilities of 10–15 (or better) above 10,000
seconds averaging time. In a microgravity environment, there is no
fundamental reason why the hydrogen maser or trapped-ion clocks
should perform better in space. But, the cesium fountain would out-
perform its counterparts through an increased interaction time be-
tween the atoms and the cavity microwave field. Local oscillators,
such as the BVA quartz crystal oscillator and the composite dielectric
resonator, would complement these new atomic clocks for short-term
averaging times and would have an important role to play.

ACES would allow the characterization and comparison of three

different ultra-stable clocks in space. The concepts of these clocks are
quite different: the maser is an active-cavity system, the fountain uses
laser-cooling techniques, and the ion-trap is based on electromagnetic
trapping of charged particles.

Two-way satellite time- and frequency-transfer through microwave

(TWSTFT) and optical (T2L2) links will provide a unique high-
performance flying timing-laboratory in space, based on the space
atomic clocks, and would be available on a global basis. Current
capabilities with the atmosphere-independent microwave link is at the
sub-nanosecond level. The optical link appears to permit clock com-
parison at the sub-picosecond level, essential to test the stability of the
space-based atomic clocks now being constructed.

46
ACES will become an essential Time and Frequency research platform.
The applications are immense and include:

1. Tests of fundamental physics with Relativity tests of unprecedented

precision (Einstein Principle of Equivalence, the Shapiro delay, the
isotropy of light, the search for the variation of the fundamental
constants)
2. Very Long Baseline Interferometry
3. High-precision geodesy and Earth spin rate dynamics
4. Study of the atmospheric propagation of light pulses and
microwave signals
5. Time and frequency comparison with ground-based clocks more
than 2 orders of magnitude over GPS
6. Establishing a global time dissemination system from space

ACES would also be the first step towards testing of atomic clocks in
space prior to their commercialization. This would be of great benefit
for:

1. Telecommunication network synchronization for digital broadcast-

ing and global mobile communications
2. Navigation and positioning in the vicinity of Earth and in
deep space

Atomic clock development with ACES could greatly benefit the devel-
opment of the next-generation of Global Navigation Satellite Systems
(GNSS), with the potential for orders of magnitude improvement over
existing GPS and GLONASS systems [26].

Global Navigation Satellite System Developments

Because so many of us fly, airline safety is a key issue. Navigation and
timing for all of avionics have taken a giant step forward with the
availability of GPS. Since aircraft navigation and timing go hand-in-
hand, this section is intended to provide an appreciation for some of
the exciting developments that are being planned. The Global Naviga-
tion Satellite System (GNSS) is a major international coordinated
initiative to provide a seamless global navigation and positioning
system by satellite that meets civilian users’ requirements in a cost-
effective way. Precise and accurate timing —including UTC — are built
into the designs, and atomic clocks are at the heart of its success. A set
of geostationary satellites will be used to augment the signals already
available from GPS and GLONASS. These satellites are being sponsored
through international cooperation by the International Maritime
Satellite Organization (INMARSAT). Developments are well in hand.

The USA plans to provide a second frequency for civilian use on the
Block 2F GPS satellites. This signal would provide civil users an
accurate measurement of the GPS signal through the ionosphere —
now one of the biggest uncertainties in precision airline navigation.

47
 Improvements in satellite position prediction and clock accuracy are
in progress. Inter-satellite links are being implemented with the GPS
Block 2R program that should already increase the time and position
accuracy available from GPS. The Block 2R satellites are scheduled for
launch starting 1997.

The Federal Aviation Administration (FAA) is in the process of devel-

oping and implementing a GPS Wide-Area Augmentation System
(WAAS) for all U.S. air traffic. The WAAS program could bring about
large fuel savings, increased inflight safety, and all-weather precision
landings. The ultimate goal of this program for precision landings is
0.8 meters (about 21/2 feet)[27].

A similar augmentation is being implemented in Europe. The European

Geostationary Navigation Overlay Service (EGNOS) will augment the
existing GPS and GLONASS, improving the accuracy, integrity, avail-
ability and continuity that is currently attainable. Sometimes termed
GNSS1, the first-generation Global Navigation Satellite System, EGNOS
will support five levels of service over different service areas. EGNOS
is designed to provide the best performance over Europe, while at the
same time guaranteeing a minimum level of service at any location
where Geostationary signals can be received. Initial operation is
expected in 1999 for levels 1, 2 and 3, with full operation scheduled for
2002 on all five levels.

The Russian authorities appear to be committed to making GLONASS

available for civil aviation and acceptable to the international commu-
nity. A study is under way to identify options for cooperation between
Western Europe and Russian industry in the field of GLONASS and
other related navigation systems.

In addition to being a provider of GNSS services at the international

level with WAAS and EGNOS, INMARSAT is also proposing to add a
dual-frequency navigation function to its Intermediate Circular Orbit
(ICO) communications satellites which are due to be launched around
the period 1999-2000. The proposed system, the International Satellite
Navigation Service (ISNS), would provide both an overlay service and
an independent dual-frequency navigation signal, which will improve
the accuracy in accounting for the timing signal’s propagation delay.
ISNS is expected to be capable of providing non-precision approach
capability worldwide without differential ground augmentation.

The Japanese Space Agency (NASDA) plans to launch two geostation-

ary satellites to provide relay of navigation, integrity, ranging and
differential messages to users in the Asia Pacific region. The service
will mainly cover the Pacific Ocean region which is at present poorly
served by the INMARSAT-3 overlay system. The satellites, which are
already in the development stage, will provide communications and
weather capabilities to users in the region. The initial Multi-functional

48
Transport Satellite (MTSAT) launch is scheduled for 1999, with a
replacement satellite in 2005. MTSAT alone will not meet requirements
for primary means of navigation, but it will probably provide a compo-
nent to the future Asian region wide-area system. The Indonesian
government is also currently planning to launch two additional Geosta-
tionary satellites with navigation payloads to complement MTSAT in
the South-Eastern Asia region. The first satellite is expected to be
launched in 2000.

Institutional and cost constraints mean that Europe is unlikely to

unilaterally launch its own satellite system to provide a service on a
global basis. Instead, European organizations are concentrating their
efforts on developing a European contribution to GNSS. The European
Space Agency is leading the developments in this area and has issued a
number of concurrent studies.

The European Space Agency (ESA) has made significant progress in its
development of the wholly-civil GNSS2. Following an initial GNSS2
Mission Analysis study of system architectures in 1995, ESA is now
concentrating on the key technical issues such as satellite orbit
configurations. The program is aimed at early demonstration of a flight
experiment by 1999, although an increase in Member State contribu-
tions to the GNSS2 program could mean seeing this time scale re-
duced. This ambitious project is expected to be funded through public-
private initiatives. Both the European Space Agency and the European
Commission are giving their full backing to the program, indicating its
importance towards the future competitiveness of European industry.
It remains that Europe is keen to see implementation of a developmen-
tal system as soon as possible. GNSS2 is expected to achieve full
operation around 2007-2010 and remain operational until at least the
year 2025.

European organizations are currently concentrating their efforts on

developing a European contribution to GNSS: the regional European
Navigation Satellite System (ENSS). The ESA is leading the develop-
ments in this area and has issued a number of concurrent studies to
examine the ENSS functional definition. One possible constellation
studied comprised twelve Inclined Geosynchronous Orbit (IGSO)
satellites and up to three Geostationary (GEO) satellites. This would
provide a vertical accuracy of better than 6 meters and provide con-
tinuous coverage for civil aviation. Following this study, the European
Commission has moved another step forward. Edith Cresson, the
European Commissioner for Science, Research and Development
(DG XII), has given her support to European moves to break into the
satellite navigation market, backing proposals for a billion-ECU
program ($1.1 billion U.S.) to develop a regional ENSS. It is hoped
that this move will allow European industry to share in a booming
market where the core technology is currently controlled by the United
States and the Russian Federation.

49
 Ultimately, the decision to fabricate the next-generation GNSS will be
taken at the political and institutional levels. Who needs it? Who will
pay for it? Only through international cooperation and coordination
can there be a truly seamless, international, next-generation GNSS. The
European Space Agency is currently engaged in discussions with the
Russians, Japanese and Americans to ensure that there is a common
goal and to minimize the duplication of effort. The cost of providing
such a system would be prohibitive without the cooperation of all
partners.

Currently, for example, GPS and GLONASS are not synchronized and
use very different coordinate reference frames [28]. One of the major
differences between GPS and GLONASS is that they use different
references for time and space. For time reference, GPS relies for its
GPS Time on UTC(USNO MC), Coordinated Universal Time (UTC) as
realized by the USNO. GLONASS relies for its GLONASS Time on
UTC(SU), UTC as realized by the Russian Federation. UTC is produced
by the BIPM and is the internationally recognized time reference for
the whole Earth.

In the past UTC(SU) and GLONASS time have been off several micro-
seconds from UTC. UTC(USNO MC) has kept very close synchroniza-
tion to UTC — within 20 ns. At the 13th Session of the CCDS (Comite
Consultatif pour la Definition de la Seconde) held on 12-13 March 1996,
it was recommended (Recommendation S4 (1996)):

“— that the reference times (modulo 1 second) of satellite navigation

systems with global coverage (such as Global Positioning System
(GPS), Global Navigation Satellite System (GLONASS), Interna-
tional Maritime Satellite Organization (INMARSAT), Global Naviga-
tion Satellite System 1 (GNSS1), Global Navigation Satellite System
2 (GNSS2)) be synchronized as closely as possible to UTC,
— that the reference frames for these systems be transformed to be in
conformity with the terrestrial reference frame maintained by the
International Earth Rotation Service (ITRF),
— that both GPS and GLONASS receivers be used at timing centers.”

Following this Recommendation the Russian Federation agreed to

improve synchronization of its time scales with UTC. This will become
an important factor for one-way time-dissemination for GNSS1 where
the timescales for the three systems will need to be coordinated. There
is also the additional problem that GLONASS includes leap seconds
and GPS does not — making the time difference, as of July 1997,
12 seconds on top of the difference between UTC(SU) and
UTC(USNO MC).

Currently, the GPS coordinate reference frame complies well within a

meter of the International Terrestrial Reference Frame (ITRF). From
the last report of the Russian authorities, getting the GLONASS refer-
ence frame also to comply poses a much bigger challenge and may not
happen for some time.

50
 0.1 ps
Optical Ion Traps 1 ps
Explosion in Timekeeping

F U T U R E
Mercury Ion Traps 10 ps
Hydrogen Masers
Performance has Provided Primary Cesium 100 ps
Major Benefit to Society
1 ns
During this Millennium

Time Stability in Seconds per Day

10 ns
Early Cesium Clocks
100 ns

1 µs
Quartz Crystal
10 µs

100 µs
Shortt Clock
1 ms
Reifler Clock
Free Pendulum Clocks 10 ms

Barometric Compensation 100 ms

Harrison’s Chronometer
Temperature Compensation 1s
Graham’s Escapement
Huygen Pendulum 10 s
Cross Beat Escapement 100 s
Chinese Hydro-mechanical
Verge & Foliot Balance 1 ks

10 ks
1000 1200 1400 1600 1800 2000

Year - A.D.

The next-generation GNSS user market will be automotive as well as

avionics. The accuracy requirements will be much more stringent —
10 centimeters is one of the long-term goals. Light travels 10 cm in
1/3 nanosecond. Methods have been proposed and are being studied
that could accomplish these incredible goals for navigation and for
clock synchronization among a set of orbiting satellites. Propagation
delay inaccuracies pose one of the biggest problems.

Finally, the issue of who controls the next-generation GNSS is complex

and beyond the scope of this paper. We merely mention some of the
challenges ahead. Encryption of the signal would need agreement from
all parties owing to the global nature of the system. Ensuring a seam-
less system will be difficult. And for safety-critical applications,
responsibility and liability for the operation of the system is currently
an active area of research [29].
51
 UTC and the Future
UTC is the common reference for all national time scales. The BIPM
timing experts have been very responsive to the needs expressed by
the nations as they have assembled under la Convention du Mètre as
part of the Consultative Committee for the Definition of the Second
(CCDS). Significant advancements have been reported by the BIPM
staff on the accuracy and stability of UTC. Essentially all of the timing
needs throughout the world are being satisfied through the current
methods of generation and distribution. The burden for providing a
real-time estimate of UTC falls on each nation. Now that GPS has
international coverage, this burden is very minimal.

Currently, GPS can be used to obtain an estimate of the UTC(USNO

MC) clock. If the time output of a good multi-channel Clear Access
(C/A) code GPS receiver is averaged for one day against a sufficiently
stable local clock, such as a cesium standard, the resulting estimate of
UTC(USNO MC) will be within 20 ns 95 percent of the time. Since
UTC(USNO MC) is steered to be within 20 ns of UTC at least
95 percent of the time, we can expect that the GPS broadcast correc-
tion will be within 30 ns of UTC 95 percent of the time. The frequency
excursions of UTC(USNO MC) via GPS are typically below 1 × 10–13
when averaged over one day. The GPS clock ensemble has a long-term
frequency stability of about 2 × 10–14. The frequency stability of
UTC(USNO MC) is about 2 × 10–15, and its rate is a predicted-forward
estimate of UTC well within the BIPM frequency uncertainty stated in
the UTC bulletins of 1 × 10–14.

As the second used in the generation of UTC is continually steered

toward the best estimate of the SI second given by the primary fre-
quency standards, we have seen, and it is anticipated we will continue
to see, significant improvement in the accuracy of the UTC second as a
result of the continued improvement of the SI second generated by the
primary standards. Primary standards are being constructed now with
anticipated accuracies of 1 × 10–16. These improved accuracies will
benefit the precision user community within the limitations of the
methods of communicating time and frequency.

As other ways to improve the usefulness of UTC are considered, there

are two areas that could help: first, to make UTC a real-time service
like GMT used to be; and second, to decrease the measurement noise
in the time and frequency transfer, dissemination and distribution
techniques.

The fact that UTC is not available in real-time means that none of the
users having real-time synchronization needs can use world official
time. The number of users needing real-time synchronization is in-
creasing rapidly; hence, they have found alternative solutions for
synchronizing their timing networks. The measurement noise associ-
ated with communicating time from one location to another is of such
a nature that the performance of state-of-the-art clocks cannot be
utilized at a distance. This measurement noise also degrades the short-
term stability of UTC.
52
The long-term performance accuracy and stability of UTC can take
advantage of the best clocks in the world and are currently doing so.
This is because the measurement noise can be averaged away over a
long enough time. The problem is that as clocks continue to improve,
the amount of time needed to average away the measurement noise is
getting longer and longer.

Thus there is a need for both real-time access to UTC and improved
methods of communicating time and frequency between clocks and
to the user community. Both problems are being addressed by the
international time and frequency metrology community, and solutions
will undoubtedly be forthcoming. Neither problem is insurmountable.
As clocks improve and as methods for communicating time and
frequency improve, the resulting available accuracies will also
improve.

Conclusions
We have seen how precise timing has provided significant benefit to
society, and we anticipate an increase in its contribution. The basic
reference for timing is UTC. The time and frequency input to UTC
comes from timing centers around the globe. Most national timing
centers generate real-time estimates of UTC that are used for con-
sumption in their respective countries. Since atomic clocks were
invented a half century ago, we have witnessed a factor of a million
improvement. These improvements have been in both the quantity and
the quality of the time and frequency signals provided to the user
community. Not surprisingly, the number of users has increased
dramatically over the last two decades. GPS has become the lead
supplier of very accurate time and frequency information, specifically
for UTC as predicted by USNO [30].

In the past, many of the precise time users were content to deduce
time and frequency information after the fact. An increasing number of
current users need precise time in real-time with very little lapse in
real-time processing of the data to deduce a current best estimate of
UTC or some timing signal. Although UTC is now only available more
than a month after the fact, there are numerous predicted estimates of
UTC, available in real-time, to satisfy most of the user community.
There is some pressure to make world official time, UTC, available in
real-time. This is under study and is possible; reasonable solutions
have been proposed.

Much of the information in this note may be found, in greater detail, in

three sources: the ITU Handbook, The Selection and Use of Precise
Frequency and Time Systems; NIST Technical Note 1337, Characteriza-
tion of Clocks and Oscillators; and the BIPM Annual Report. The
proceedings of four conferences contain a large number of papers with
relevant material to this article: IEEE International Frequency Control
Symposium; European Frequency and Time Forum; Conference on
Precision Electromagnetic Measurements; and Precise Time and Time
Interval Planning and Applications Meeting.
53
 Because time and frequency can be measured more accurately than
any other quantity, and because it can be measured with cost-effective
equipment, time and frequency techniques in various forms are perme-
ating many of society’s activities. One of the main tools used in many
of these applications is GPS; it is becoming a commonly known
system. The official provider of time for GPS is ultimately UTC. We
may expect the role of UTC, as the official provider of time for the
world, to continue to increase in its importance along with the work of
the contributing timing centers around the world. Certainly, the quality
has dramatically improved in recent time as have the number of
beneficiaries.

Acknowledgments
This note is an accumulation of the ideas and work of many notable
contributors to the far-reaching success of precise timekeeping in
modern society. We wish to take our hats off to the several who have
gone before — laying the important groundwork that allows us to be
where we are today. It would be too difficult to name names because
of the diversity of foundational activities that bring us here. Our efforts
have been to capture some of the historical perspective, highlight
where we are now, and anticipate where new discoveries in precision
metrology and timekeeping will propel us.

We have been privileged to have excellent critical feedback on the

specific contents of this note from among the outstanding clock
metrologists in the world. While the authors and HP take final respon-
sibility for the contents we are indebted to: Dr. Rob Douglas of the
Canadian National Research Council, Mr. Michael King of Motorola,
Inc., Dr. John Laverty, Head of the NPL Time & Frequency Services,
Professor Sigfrido Leschiutta of Italian Istituto Elettronico Nazionale
G. Ferrar, Dr. Wlodek Lewandowski of the Bureau International des
Poids et Mesures, Dr. Dennis McCarthy, Director of the Directorate of
Time of the USNO, Dr. Demetrios Matsakis, Head of the Time Service
Department of the USNO, Mr. William Riley of EG&G Rubidium
Products, James McAslan Steele, retired NPL timing expert, Dr. Donald
Sullivan, Chief of the Time and Frequency Division, National Institute
of Standards and Technology, Dr. Claudine Thomas, Head of Time
Section BIPM, Dr. John Vig of US Army Research Laboratory, and Dr.
Gernot M. P. Winkler, former Director of USNO Time Services. Dr.
Leonard Cutler and Robin Giffard of Hewlett-Packard Laboratories,
and Jack Kusters of Hewlett-Packard (Santa Clara Division) have
provided very valuable editorial and technical assistance. Mr. Sterling

54
Allan has also provided very valuable editorial assistance. Mrs. Sylvia
Chantler and the NPL library staff have been most helpful in research-
ing questions. The sponsorship of Hewlett-Packard (SCD) and the
assistance of their editorial staff have been the mainstay to the project.

We are also particularly indebted to Dr. Philip M. Woodward, co-author

of the book The Science of Clocks and Watches, and author of My Own
Right Time, and Jonathan Betts, Curator of Horology, National Mari-
time Museum, for providing the data for the two historic chronometers
we analyzed in detail in Appendix B: Kendall No. 1 and Mudge No. 1
Chronometers. These data were an important basis for one of the key
figures in the text, Figure 12. William J. Riley of EG&G provided the
software used in performing the statistical analysis on these chronom-
eters, and he assisted with great interest in many other ways. This is
the same statistical package used to assure that the next generation
atomic clocks on board the GPS satellites (GPS Block 2R) meet perfor-
mance specification. We wish to express appreciation to Dr. Robert F.
C. Vessot for providing most of the data for the explosion chart on
page 51.

We have received encouragement and ideas from numerous other

colleagues, to whom we would also like to express appreciation.
Clearly, the text pulls together a lot of disciplines, and without the
wide support we have received, our task would have been, perhaps,
overwhelming.

55
 Appendix A

Time and Frequency Measures Accuracy, Error,

Precision, Predictability, Stability, and Uncertainty
(See references [30] through [35] for additional details regarding the
contents of this appendix.)

Example for Illustration:

Consider the flip of a coin, for illustrative purposes, which turns out to
be very much like the random errors in an ideal atomic clock: heads
and tails are equally probable. Consider now a million people standing
in an east-west row, but all facing north — each with a coin. Each
person flips his coin 100 times. If the coin comes up heads, the person
takes one step north; if tails, one step backward (south). After a
hundred flips of the coins, we can look down the row of people and
look at the number standing at different distances from the original
line. Plotting the density of people at different steps away from the
origin will make a bell shaped curve approaching a normal or Gaussian
distribution with a maximum at the origin. In other words, more people
will have returned to where they started than to any other position. We
can compute the average distance away from the origin taken across
all one million people; this is called the mean value. If we let xi be the
distance away from the origin for the ith person, then the mean value is
computed by adding up all the xi distances and dividing that sum by a
million.

The standard deviation, σ, is obtained by subtracting from each

person’s distance the mean value; squaring the result; then adding up
all of the squared values, dividing that sum by the number of data
points minus one (999,999 values in our example); then, finally taking
the square root. The minus one is used because one degree of freedom
was removed in computing the mean value. In our example, σ will have
the value of square root of the number of flips of the coin, 100 = 10 .
For a normal distribution, 68 percent of the values will be within 1 σ of
the origin, and 95 percent (2 σ) will be within 20 steps of the origin in
our example.

The standard deviation of the mean, σm, is given by σ / N , where N is

the total number of degrees of freedom. In our example, it is a million,
since all the people flip coins that are totally independent of each
other. In our case, σ m = 10 / 1, 000 , 000 = 0.01 . In other words, if this
experiment were repeated over and over, 68 percent of the mean
values would be within 0.01 steps of the origin.

Next, carry this analogy over to an atomic clock. Suppose each flip of
the coin is the clock’s effort over one second to determine the length
of the SI second and it is able to do so with a standard deviation of
1 × 10–11. Then the second to second stability would be 1 × 10–11. The

56
uncertainty of the mean value of the SI second measurement after
100 seconds (100 flips of the coin) would be
1 × 10 –11 / 100 = 1 × 10 –12 .
This is the inaccuracy of the mean value as averaged over 100 seconds.

If we had a million of these ideal clocks or a million independent

100-second measurements, then the standard deviation of these
100-second average frequencies would be 1 × 10–12, and the standard
deviation of the mean of all one million clocks or independent measure-
–12
ments would be 1 × 10 / 1, 000 , 000 = 1 × 10 –15 .

An ideal clock, like the flip of a coin, will have an error in each fre-
quency measurement that is random and uncorrelated with any of the
past errors — the coin has no memory of any of its past flips. However,
since the time of a clock results from adding or integrating the fre-
quency or rate of a clock, the time errors get integrated or added also.
So like the people’s distance, xi, from the origin, the time error, xi, for
each clock adds up all of the independent frequency errors. The xis
follow what is called a random-walk process. And even though it is
random, the time error at any point in time correlates with the past
because it is an accumulation of all of the past errors. We can never
know the exact amount of these errors because there is no perfect
reference; hence, errors, as will be discussed in more detail below, are
often referred to in the sense of probabilities. For example, from a
normal distribution as illustrated above, 68 percent of the time the
errors would lie within one standard deviation (1 σ) of the mean value
or estimate of the measurand.

In the above example in a 1 σ sense, the time error after 1 s is

1 × 10–11 × 1 s = 10 ps (picosecond = 10–12 s). After 100 s, the clock will
be 10 ps × 100 = 100 ps away from the origin. The average deviation
time of a million such clocks would be 100 ps / 1, 000 , 000 = 0.1 ps ,
hence one can see the value of ensembling clocks.

Because of the random-walk phenomena, the standard deviation of a

clock’s time error will degrade as the square root of the running time,
and hence its time error is unbounded — even in the theoretically ideal-
clock case.

We may write the time error of a clock in terms of its frequency offset
t
as follows: x ( t ) = ∫0 y ( t ′ ) dt ′ .
Hence, if the frequency offset is positive, the time-error ramps early. If
the frequency offset, y0, is constant, we may simply write the time error
as follows: x(t)=y0 t. Early and late are sometimes confusing to people.
At the moment a clock is early its reading will be larger than a correct
clock, and if late, then smaller. Another confusion sometimes arises

57
 because the period between cycles, P, is the reciprocal of frequency,
ν=1/P. When we take the derivative of this equation and normalize it,
we have δν/ν=–δP/P=y(t) — note the minus sign in front of the δP/P. In
other words, if the frequency offset is high, the period, P, is low, etc.
The SI second is the accumulation “. . . of 9,192,631,770 periods . . .”
derived from the defined cesium resonance. So, for example, if a
primary frequency standard measures the second of TAI as being too
long, then TAI’s frequency is too low.

Definitions:
The definitions for accuracy, error, frequency instability, precision,
synchronization, syntonization, and uncertainty are given in the
Glossary and Definitions on page 81. These definitions are generally
acceptable within the time and frequency community. As much as
possible these definitions have been drawn from the work of the
International Telecommunications Union (ITU) and the International
Standards Organization (ISO). See also IEEE Standards 1139-1988 and
1193-1994 [34, 35].

Commentary on Measures:
Since the true measurand is never known in reality, the accuracy
estimates are usually given with a 1 σ (68.3%), 2 σ (95.5%), or 3 σ
(99.7%) — depending upon the confidence level desired. The percent-
ages given are for a normal distribution. To have a valid accuracy
estimate, both the random as well as the systematic errors have to be
included. For frequency accuracy the measurand is the SI second. For
time accuracy the measurand is UTC, or one may specify accuracy
with respect to a network reference clock, for example.

The definition of precision is somewhat vague. It is often used in time

and frequency metrology as defined in the dictionary. Sometimes, more
specifically per the definition given above, it is used to describe the
uncertainty of a measurement. The traditional standard deviation of
the individual measurements taken around the mean value is often
quoted as the precision of the measurement. Though this is not an
unreasonable discriptor, on the other hand, as will be shown below, it
can be very misleading. Hence, the generic use of the word is probably
safer, and other measures should be used to be more specific in time
and frequency systems specification and characterization.

The standard deviation is not recommended as a measure of frequency

instability because the long-term frequency behavior of most clocks
tends to walk off, and the standard deviation appears unbounded. Both
the IEEE and the ITU have recommended measures that are conver-
gent and well-behaved in spite of the apparent walk-off phenomena
exhibited by most clocks. (See below.)

There is sometimes a confusion between instability and stability,

inaccuracy and accuracy; though in practice it is seldom a problem. As
explained in the text of this application note, the normalized frequency

58
instability can be written as a dimensionless number; 1 × 10–12 for
example. A clock having this instability would be stable to one part in
1012. As discussed in this text, the terms are often used interchangeably
because of the lack of confusion.

The IEEE and ITU recommended measures for instability may be

described as follows: Consider three sequential time error measure-
ments of a clock, xn, xn+1, and xn+2, spaced by a measurement
interval τ. As discussed in the text, the normalized frequency departure
averaged over the n to n+1 interval is given by y n = (xn+1 – xn)/τ; or in
terms of finite difference notation this may be written: yn = Δxn/τ,
where the Δ denotes the first finite difference for the nth interval.
Similarly, the average frequency departure for the next interval may be
written yn+1 = Δxn+1/τ. The instability in this clock for its frequency
averaged over the first τ interval to the next τ interval may be repre-
sented by the change in frequency: yn+1 – yn = Δyn. Now the difference
of the difference is called the second difference and is denoted Δ2.
Hence, from the above equations we may write the following:
Δyn = Δ2xn/τ. If we compute the sum of the squares of these second
differences for n = 1 to N-2, where N is the number of time error
measurements in a series for a particular clock, and then divide by
2(N-2), we have what is called an estimate of the two-sample variance,
AVAR. We divide by N-2 because that is the number of entries in the
sum, and we divide by 2 so that AVAR is equal to the classical variance
in the case where the yns are random and uncorrelated as is the case
for classical cesium-beam and rubidium gas-cell frequency standards.
The equation for the two-sample variance may be written as follows:
2
σ 2y ( τ ) = 1  Δ2 x 
  = 1
2
( Δy ) 2 (A1)
2τ 2
where the brackets, “<>,” denote an infinite time average. As in the
example above, the average of the second difference is simply taken
over the data length. The longer the data length, the better is the
confidence on the estimate.

There are five different noise types used to model time and frequency
devices: white-noise time or phase modulation (PM), flicker or 1/f PM,
white-noise (random and uncorrelated) frequency modulation (FM),
flicker-noise or 1/f FM, and random-walk FM. The white-noise FM is
classical for atomic clocks and is like our coin toss experiment. As
explained above, in this case the time deviations are random-walk in
nature. This is because the integral of the frequency is proportional to
the time and the integral of white noise is random walk. The classical
variance is nonconvergent for the last two noise types. The two-sample
variance is not only convergent for all the noise types, but with the
observation of the dependence of the variance while the averaging time
τ is changing, the type and level of noise can be inferred except for
white-noise PM and flicker-noise PM, which have similar τ dependence.
The value of τ can be easily changed in the analysis software.

59
 As an illustration of a σy(τ) diagram, Figure A1 portrays the region of
instabilities for the most important kinds of precision clocks and
oscillators that are now in use by the Time and Frequency community.

Figure A1. A frequency stability

Stability Ranges of Various Frequency Sources diagram, σy(τ), for most of the
for 1 kHz Bandwidth precision clocks and oscillators
–9 used widely within the time and
frequency community and by an ever-
QZ increasing number of users of precision
–10 timing devices. The dashed region at
the bottom of the cesium (CS) stability
plot shows the improved long-term
–11 stability of the HP 5071A Frequency
RB Standard. QZ≡Quartz Crystal
Log (σy(τ))

–12 Oscillator, RB≡Rubidium Gas-Cell

CS Frequency Standard, CS≡Cesium-beam
Frequency Standard, HM≡Active
–13 Hydrogen-Maser Frequency Standard.

–14

HM
–15

–16
–3 –2 –1 0 1 2 3 4 5 6 7
Log (τ), seconds

The modified two-sample variance, MVAR, was developed in 1981 to

remove the ambiguity mentioned for AVAR in the previous paragraph.
MVAR removes this ambiguity by introducing, in the software statisti-
cal analysis package, a bandwidth modulation. In equation (A1) on the
previous page, there are three time-error readings making up the
second difference, Δ2xi = xi+2 – 2xi+1 + xi. If these three time-error
values are replaced by three time-error averages, each averaged over
adjacent windows of duration τ, then the desired result is obtained.
The equation for MVAR may be written as follows:
2
Mod .σ 2y ( τ ) = 1  Δ2 x  , (A2)
2τ 2
where the bar over the x denotes an average over an interval τ.

How well the time of a clock can be predicted depends upon the
prediction algorithm, the kind(s) of noise in the clock, how well the
systematics are understood and have been modeled, and the kind(s) of
measurement noise and how well they have been averaged or filtered.
The systematic effects are often driven by environmental changes —
such as may be caused by temperature effects. Hence, to perform
satisfactory prediction it may be necessary in some cases to factor in
environmental parameters.

60
Optimum predictors have been designed in many clock applications.
Here optimum means in a minimum squared error sense. It may be
shown that if the systematics and the measurement noise have been
properly dealt with, then the optimum prediction error for the random
part of the clock’s behavior is given by xrms(τp) = K(α)τpσy(τp). This
is the root-mean-square prediction error for a prediction interval τp,
and the K(α) depends upon the kind of limiting noise in the clock. For
the five noise types outlined above (α = +2, +1, 0, –1, –2), K(α) has
the following values: 2/3, ~1, 1, 1.2, and 1, respectively. The optimum
prediction algorithms are different for each of the five noise types. For
the even integer values of α, they are simple and can be deduced from
the powerful and useful statistical theorem that the optimum estimate
of a white-noise process is the simple mean. For the odd values of α,
the prediction algorithms are complex, but simple ones have been
designed that are close to optimum. In reality, most clocks are charac-
terized by more than one noise type and the prediction algorithms
become correspondingly more complex, but prediction procedures
can be designed to be very tractable and close to optimum.

During the late 1980s, the telecommunications community expressed

a need for a useful variance measure for characterizing network
performance. At the same time, the time and frequency community
had a need for a variance measure of measurement noise and of time
and frequency transfer and distribution systems. The time variance,
TVAR, was developed to satisfy these needs and has been adopted by
both communities. TVAR has a direct mathematical relationship to
MVAR, and may be written as follows:
2
σ 2x τ = τ Mod .σ 2y τ
() ()
3
2
= 1  Δ2 x  , (A3)
6
where the 6 in the denominator normalizes TVAR to be equal to the
classical variance on the time residuals in the case of classical white-
noise PM — in contrast to the above where σ 2y τ has been normalized
()
to be equal to the classical variance on the frequency residuals for
classical white-noise FM. White-noise PM is the theoretical limiting
noise for measurement systems and for networks, while white-noise
FM is the theoretical limiting noise for classical atomic clocks.

The above five values of α have proven very useful in modeling the
random fluctuations observed in precise time and frequency systems.
This α actually denotes the exponent on the Fourier frequency, f, for
the spectral density, Sy(f). The spectral density is a measure of the
power present at different Fourier frequencies. The Fourier frequency,
in contrast to the carrier frequency, is associated with the power in the
residuals (i.e., the fluctuations of the frequency y(t)) around the
nominal carrier frequency value. Suppose a quartz crystal oscillator has

61
 a carrier frequency of 5 MHz and its frequency is temperature-
dependent. If the environmental temperature for this standard goes up
and down on a daily basis, we would expect the frequency fluctuations
to have a strong Fourier frequency component at one cycle per day. A
spectral density plot of Sy(f) would show this bright-line component at
f = 1 cycle per 86400 seconds = 0.000 011574 Hz.

These five values of α are power-law spectral density models, Sy(f)~f α.

In contrast to the example above power-low spectral densities do not
exhibit bright-line components. Rather, these models represent a
performance over a band of Fourier frequencies. A classic illustration
is the previously mentioned coin-toss experiment, which is also
analogous to the ideal atomic clock. In this case, frequency fluctua-
tions are random and uncorrelated with respect to the carrier fre-
quency value. These residuals have what is called a white-noise
spectral density in which case all Fourier frequencies are equally
probable. For example, a clock with white-noise frequency modulation
has the same amount of power density causing the clock’s frequency
to fluctuate one cycle per day as that causing it to fluctuate one cycle
per second, or one cycle per hour, or at any other rate. In this case, α
is equal to zero; that is, the spectral density of the Fourier frequency
fluctuations is equal to a constant since f α = 1. The other values of α
are also useful in modeling precision clocks and timing systems. For
instance, α =+2 (white-noise time or phase modulation (PM)) is the
limiting noise model for an ideal time transfer system or time or phase
difference measurement system. The models with α = +1, –1, and –2
are experimentally useful but don’t have a strong theoretical basis at
the current time.

Since the time residual fluctuations are the integral of the frequency
fluctuations, it can be shown that for these power-law model pro-
cesses, the spectral density of the time residuals, Sx(f) ~ f β , and β = α –
2. So in the case of white-noise PM just discussed, where α = +2, then
β = 0, all Fourier frequencies of the time fluctuations have equal power
density. Both the IEEE and the ITU have recommended Sy(f) and Sx(f)
as useful measures of clock performance. These are called frequency-
domain measures of stability.

The three variances (AVAR, MVAR, and TVAR) are called time-domain
measures of stability, and have the additional virtue that the above
spectral densities can be written in terms of these variances. In fact,
the dependence of these variances has a close correspondence to

62
the Fourier frequency dependence. Specifically, if AVAR, MVAR,
µ µ′ η
and TVAR are proportional to τ , τ , and τ , respectively, then
α = – µ – 1, α = – µ ′ – 1, β = – η – 1, and η = µ ′ + 2. The correspond-
ing ranges of values over which these variances follow these three
equations are –2 ≤ µ < 2 , – 3 ≤ µ ′ < 2 , and – 1 ≤ η < 4 , respectively, and
are illustrated in Table A1. Table A2 shows the mathematical represen-
tations for the three time-domain variances. Table A3 provides the
actual relationships between the frequency-domain measures cited
above and the time-domain measures — remembering that TVAR = τ2
MVAR/3. Of course, the square root of each of these three variances are
alphabetically denoted ADEV, MDEV, and TDEV, respectively as they
were introduced above. They are written symbolically as follows: σy(τ),
Mod. σy(τ), and σx(τ), respectively.

From an electrical engineering point of view, all three of these time-

domain variances have transfer functions into the frequency-domain
that remind one of a bandpass filter. For example, the equation giving
AVAR in terms of the spectral density is given by:

2 fh sin 4 πfτ
( ) df
σy ( τ ) = 2 ∫0 ()
Sy f
,
( πfτ ) 2
where fh is the high-frequency cutoff for the applicable measurement
system. The kernel of this integral looks like a variable bandpass filter
centered at Fourier f = 1/2τ and going to zero at f = 0 and at f = 1/τ on
either side of the center frequency. Hence, the center of the effective
bandpass filter decreases as τ increases. If τ values are taken such that
τ = nτ0, where n = 2i (i = 0, 1, 2, 3 . . .) and τ0 is the initial data spacing,
then a σy(τ) diagram provides an analysis of stability over a nominally
square window in the frequency-domain — ranging from f = 1/2τmax to
1/2τ0 , where τmax is the largest τ value available from the data set. The
transfer functions for the other two variances are similar. For details
see reference [31], pp 97-108, and/or [36].

Table A1. AVAR, MVAR, and TVAR ranges of values

α Noise Type Range of Applicability

+2 White PM
Dissemination Systems
Tel. Comm.
Network
Time/Frequency

+1 Flicker PM
Quartz

MVAR α = – µ' – 1
AVAR α = – µ – 1

TVAR β = – η – 1

0 White FM
Passive Hydrogen

Active Hydrogen

–1 Flicker FM
Rubidium
Cesium

–2 Random Walk FM

63
 Table A2. AVAR, MVAR, and TVAR mathematical representations

Abbreviation Name Expression

1
AVAR Two-Sample or σ 2y (τ) = ( Δy )2
Allan Variance 2
1
= ( Δ2 x )2
2τ 2

MVAR Modified 2 1
Allan Variance
Mod. σ y ( τ) = (Δ2 x )2
2τ 2

TVAR Time Variance σ 2x (τ) = 1 (Δ2 x )2

6

Table A3. Relationships between frequency-domain measures and

time-domain measures

Noise Type Sy(f) Sx(f)

2
White PM ( 2π )τ 2 σ 2 τ  f 2 1  2
τ σx τ  f 0 ()
3f h 
y  () τ 0 fh  

2
( 2π ) τ 2 σ 2 τ  f 1
3  0 2
τ σ x τ  f
–1
()
Flicker PM
A *  y()  3.37 

12  τ –1σ 2 τ  f –2
White FM 2  τ 1σ 2y τ  f 0
() 2  x() 
  ( 2π )
1  τ 0 σ 2 τ  –1 20  τ –2 σ 2 τ  f –3
Flicker FM y f () 2  x  ()
2ln2  ( 2π ) 9ln2

6  τ –1σ 2 τ  f –2 240  τ –3 σ2 τ  f –4
Random Walk FM 2  y()  4  x ()

( 2π ) ( 2π ) 11

*
A=1.038 + 3ln(2πfhτ)

In considering an uncertainty specification, the above three time-

domain variances have proven very useful in time and frequency
metrology for characterizing the random or stochastic processes in
clocks, oscillators and sundry time and frequency systems. However,
it is very often the case that systematic effects, rather than random
effects, limit the ultimate performance of both clocks and oscillators
as well as time and frequency transfer and dissemination systems. An
uncertainty specification should reflect the one-sigma combination of
both the systematic and the random effects. They are usually combined
as the square root of the sum of the squares. If the vendor of a product
wishes to have a more conservative specification than would be given
by a one-sigma number, a two- or three-sigma number may be used, but
it is good to so designate. The distribution of errors may not always
follow a normal distribution for systematic effects. Even for random

64
effects, non-normal distributions are observed from time to time.
Following the above procedure is in essential conformity with the
BIPM’s guideline on uncertainty, and this procedure is distribution
insensitive.

One of the systematics that often occurs in precision oscillators is

frequency drift. Frequency drift occurs in all quartz-crystal oscillators
and rubidium gas-cell frequency standards. Some cesium-beam fre-
quency standards exhibit frequency drift — though the amount of drift
is usually very small. Hydrogen masers exhibit frequency drift unless
the resonant cavity used to bring about the maser oscillation has an
active servo control keeping the cavity resonance on the frequency of
the hydrogen atom. In addition, the wall coating of the cell inside the
resonant cavity containing the hydrogen atoms must remain the same
over time, or the hydrogen clock may exhibit frequency drift.

The time error caused by frequency drift is given by 1/2 Dt2, where D is
the amount of the frequency drift and t is the time since the clock was
both synchronized and syntonized. D is y(t)/t, where y(0) = 0 due to
being syntonized at that point in time. This time error is sometimes
called the time-interval error (TIE). The frequency inaccuracy caused
by frequency drift is given by Dt. Frequency drift, of course, is one
form of instability and affects the measures outlined above as follows:

σ y τ = Mod .σ y τ = Dτ
() () 2
For the time stability measure the TIE≅1.2 σx(τ) due to frequency drift,
where τ is now the time since synchronization and syntonization. The
time interval error is important in setting up a network in order to
know how often the clocks need to be calibrated to avoid exceeding
some TIE or to have an adequate “holdover” time.

Since the uncertainty needs to combine the errors from both the
systematic as well as the random parts of the timing system, it is
apparent that there will be a dependence of the uncertainty on the
averaging time since the random part is almost always τ dependent. As
an important example, consider the case where two clocks are being
compared and the limiting measurement noise is white-noise PM. It has
been shown that the uncertainty in the frequency comparison due to
this kind of measurement noise decreases as τ − 3 2 and is given by
2 × Mod.σy(τ). The frequency estimate, in this case, is deduced from
the slope of a linear regression to the time-difference measurements
taken as a time series between the two clocks. Because the white-noise
PM tends to average away so quickly, it is almost always the case that
the systematics dominate in the long-term. Hence, in specifying an
uncertainty an averaging time may also need to be given. A lot of effort
has been put into uncertainty specification documents, but that which
is most helpful is experience and common sense.

65
 Appendix B

Stability Analysis of Harrison-like Chronometers

The genius of John Harrison’s work had such far-reaching impact on
the world’s timekeeping that we elected to document his monumental
contribution. We have been fortunate to obtain some of the original
data from some early Harrison-like chronometers. The word chronom-
eter was given to a seaworthy clock of which Harrison built only five.
His great success inspired others as they patterned their designs after
his pioneering work. Specifically, we will include some data from
Kendall and Mudge chronometers. The Kendall Clock, K-1, was made
famous, as it was of great assistance to Captain James Cook in his
exploration of the Pacific Ocean regions.

Before making chronometers, Harrison’s clockmaking skills were being

developed for tower clocks, grandfather clocks and the like. As men-
tioned before, his big contribution came from removing the effects of
the environment on timekeeping uniformity. He had an insatiable desire
for learning, and was apparently largely self taught. He learned his
woodworking skills from his father, and the mechanics of motion from
a book containing a Cambridge lecture series given by Nicholas
Saunderson. This and Newton’s Principia were apparently his guide.

John Harrison finished his first pendulum clock when he was 19. The
Bocklesby Park tower clock, finished in 1722 when he was 29, is still
running and is made mostly of wooden parts. He avoided sticky oils,
which changed viscosity with temperature, by using a special hard-
wood which exuded its own natural oil. He later developed the bi-
metallic strip concept to accommodate changing temperatures. The
concept of bearings also came out of his creative mind. Some of the
land clocks made by John Harrison with the assistance of his brother
have been reported to have accuracies of one second a month, which is
astounding for that era.

Having a clock with which to compare was a serious problem for them.
He and his brother devised a sidereal clock composed of the edge of a
window in their home and a neighbor’s chimney. As this would provide
occultation of a particular star as the earth spun against the celestial
sphere, they would mark the time on their local clock. Apparently, they
could do this to better than a fraction of a second, but measurement
noise undoubtedly was a significant problem. If they achieved accura-
cies better than a second a month (4 × 10–7), this is only about forty
times worse than the annual spin-rate variations of the earth!

A pendulum clock will not work at sea because its time uniformity is
dependent upon the regular motion of a swinging bob. Newton, envi-
sioning this irregular motion of a rolling, rocking ship added on top of

66
an attempted pendulum chronometer, implied that a seaworthy clock
TIME ERROR
capable of determining accurate longitude may never be built. Harrison
Kendall 1 Chronometer
devised a counter-rotating pair of barbell-like pendulums connected by 25000
springs at the top and bottom, so that one spring was under compres-

Time Error (seconds)

20000
sion while the other was extended and vice versa — designed so that
motion would affect one half in the opposite way to the other half. The 15000
net effect was for the chronometer to be independent of the motion of 10000
its supporting platform. This clock movement came to be known as the
5000
grasshopper escapement and was used in H-1, Harrison’s first seawor-
thy chronometer. H-1 was tested on a trip to Lisbon and back to 0
0 220 440 660 880 1100
London and met the qualifications for the Crown’s £20,000 prize Number of Days
(maintaining accuracy of three seconds per day), but he told the
Figure B1. The time error of the famous
Longitude Board he could do better. His chronometer development K-1 (Kendall No. 1 Chronometer). The
continued in its sophistication from there. overall chronometer rate is 20.84 seconds
per day.
John Harrison had to build three more chronometers to partially
satisfy the critical Longitude Board. H-4 made the voyage to Jamaica — FREQUENCY DATA
losing only five seconds in 81 days at sea. It returned to London two Kendall 1 Chronometer
days after Harrison’s 69th birthday with a combined total error of two

Clock Rate (seconds per day)

50
minutes over five months. Though he was very discouraged by the 40
inordinate delays in getting the prize money, his work inspired many
30
other clockmakers.
20
We were successful in obtaining two sets of time-error data from these 10
historical jewels made more than two centuries ago [37, 38, 39]. The 0
Kendall No. 1 chronometer data was actually taken starting in July 0 220 440 660 880 1100
Number of Days
1984 — with daily readings almost every day for 1,082 days. Undoubt-
edly, this chronometer was operating far worse than at its prime — Figure B2. A plot of K-1’s clock rate
being well over 200 years old. Two hundred and three days of data (frequency). The step in frequency is
were analyzed on the Mudge No. 1 chronometer, and the data were obviously concurrent with its transfer to
a new gallery. During the data acquistion
taken in 1777, probably part of its evaluation. Both of these easily met it was not always humanly possible to be
the prize-money requirements though the environment was rather there in time to wind it before it stopped.
benign compared to conditions at sea. After stoppage (10 in the first gallery and
42 in the second), it was rewound and
reset. While in the second gallery, there
For the analysis of the data we have used two approaches: first, we was some indication that it tended to run
have employed the instability analysis procedures outlined in Appen- slightly fast for a while following a
stoppage.
dix A, and as written into a software package developed for character-
izing GPS atomic clocks [40]. Since the long-term performance of these
clocks is a key issue, we have incorporated a recent development
coming from NIST [41] which gives improved confidence on the long-
term stability estimates; it is called total sigma — using the same
statistical measure as outlined in Appendix A, but with better confi-
dence on the values. The difficulty with the first approach is that we
need a reasonably long data set of clock time errors in order to per-
form the analysis. We obtained all we could within reason. From the
data available, we were able to calculate the time prediction error per
Appendix A for inclusion in Figure 12.

67
 The second approach avoids the need for a data set taken as a time
series. For these Harrison-like clocks, such time-error series are rarely CLOCK RATE HISTOGRAM
Kendall 1 Chronometer
available. But if we have a time-prediction error over a given length of 300
time, Figure 12 is directly amenable to plotting that point as obtained
240
from experimentation. From that data point, we can effectively work

# Points Per Bin

the problem backwards and infer that the stability must be better than 180
some number in order to obtain that prediction error. The time-error 120
prediction equations given in Appendix A are under the assumption of
optimum prediction. In other words, the root-mean-square prediction 60

error cannot be better than the number given by these equations — 0

–30 –10 10 30 50 70 90
they provide a lower limit. These equations are also dependent upon Data Value
different types of noise, and so the assumed model needs to be close to
reality for validity of the calculations. We will see that these assump- Figure B3. A Histogram of the K-1 clock
rate data with each point being a one-day
tions are reasonably met. The analysis procedure will follow a similar
average. The distribution clearly does not
pattern as discussed in the illustrative example given in the text. follow the normal distribution curve also
indicated. The double distribution is a
result of the different clock rates at
Figure B1 is the time error plot for the famous Kendall K-1 Chronom-
which K-1 nominally operated in the two
eter with a beginning date of 16 July 1984. Readings were taken once a different galleries housing the clock
day as often as possible and at about the same time each day. After 230 during data acquisition. Notice that the
upper distribution is skewed to the higher
days the chronometer was moved to a different gallery. The total
clock rate values apparently as a result of
experiment length was 1,082 days. On days when readings could not be occasional stoppages.
taken, reasonable interpolations were inserted. From the knee in the
curve, it is obvious the chronometer changed frequency after being
moved to the second gallery. This could be due to temperature and/or
humidity and/or a change in the position of the chronometer with TIME ERROR minus mean frequency
respect to the earth’s gravitational vector. Kendall 1 Chronometer
150
Figure B2 is a clock rate or frequency plot of the same data. It is
Time Error (seconds)

120
apparent that the frequency stability deteriorated while it was in the
90
second gallery. There are also several large frequency steps, most of
which go positive from the nominal average value during this latter 60
segment of data. The clock suffered far fewer stoppages while in the
30
first gallery. In the second gallery, the frequency tended to go high for a
few days after the clock was rewound following a stoppage. 0
0 50 100 150 200 250
Number of Days
Figure B3 is a histogram plot of the data in Figure B2. The data values
Figure B4. A time error plot after
are in units of seconds per day. The double distribution is a result of subtracting a calibrated frequency
the different clock rates at which K-1 nominally operated in the two offset from the numbers over the first
different galleries housing the clock during data acquisition. Notice that 230 days — the time it was housed in
the first gallery. As compared to
the upper distribution is skewed to the higher clock rate values. These Figure B1, notice that this calibration
higher skewed values appeared to occur following some of the occa- reduces the peak-to-peak error by
sional weekend stoppages. about a factor of 150. Notice also the
general parabolic shape to the curve —
indicative of a negative frequency
Since K-1 is over 200 years old, we felt it necessary to give it every drift.
advantage because of the natural deterioration of all mechanical
devices; hence, we only analyzed in detail the first segment of the data.
Figure B4 is a plot of the time error for this segment after removing the
mean frequency. This would nominally be equivalent to a calibrated
clock rate correction. Notice that by this procedure the peak error
reduces from about 2,000 seconds down to 150 seconds. One sees a

68
nominal parabolic shape to this curve that could be caused by a
frequency drift. A linear regression line was subtracted from the TIME ERROR minus Drift
Kendall 1 Chronometer
frequency data for this first segment, and Figure B5 shows the remain- 75
ing time errors. In this case the peak error drops by more than a factor

Time Error (seconds)

50
of two.
25
Following the guidelines given in Appendix A and using the total
0
variance approach developed by NIST, we show the frequency stability
of both the first and second segments of the data in Figure B6. Notice –25
that the first segment is significantly lower — quantifying that which is –50
0 50 100 150 200 250
visually apparent in Figure B2. A nominally flat frequency stability Number of Days
curve like this one — proportional to τ0 — is observed in a large
variety of clocks and is characterized by what is called flicker noise, Figure B5. After subtracting a frequency
FM where S y (f) is proportional to 1/f. In a frequency stability plot, drift as estimated from K-1’s frequency
data by simple linear regression, the
this is often called the flicker floor. We see that K-1, after more than residual time errors were calculated and
200 years of operation, has a flicker floor of about 1.5 × 10–5 or about are plotted here. In this case, as
1.3 seconds per day. From Appendix A we know that the time error of compared to Figure B4, the time errors
are only reduced by a little over a factor
prediction or the timekeeping ability of such a clock is given by 1.2 of two. How much the timing errors will
τpσy(τp); hence, for one day it is 1.6 seconds. This nicely meets (~90% be reduced by frequency drift calibration
confidence interval) the requirements of 3 s/day set for the Crown is a function of the size of the drift as
compared to the random clock noise.
award money even after the instrument is over 200 years old. Of Calibrating frequency drift in quartz-
course, the environment for the experiment was very benign, and we crystal oscillators and rubidium gas-cell
have accounted for both the clock rate offset and its frequency drift. frequency standards is often very
beneficial. It is not surprising that there
is some frequency drift in a mechanical
A flicker-noise model is the reason for the upward sloping behavior of timing device.
the Harrison-like chronometers in Figure 12. Notice in Figure B6 that
for large averaging times the second segment appears to average down
toward the extrapolated values for the first segment. This would FREQUENCY STABILITY
indicate that the chronometer is still successfully hunting for its Kendall 1 Chronometer
total (τ)

designed resonance frequency in spite of the degraded operating 10 –4

conditions of the second gallery.
Total Allan Deviation, σ

Second Segment
Figure B7 is a plot of the time error of the Mudge Chronometer data 3

taken in 1777. This is a continuous time-error series taken over 203 2 First Segment
1.5
days — one measurement per day. The clock-rate offset is quite a bit
10 –5
smaller for the Mudge chronometer than for the Kendall. Since this can 10 4 10 5 10 5 10 7 10 8
be calibrated out, this should not be used as the most significant Averaging Time, τ (seconds)
criteria of clock performance. Figure B8 is the corresponding fre-
quency plot, and one can see evidence of a change in the average Figure B6. A plot of the frequency
frequency starting at about day 165. If a calibrated mean frequency is stability for the data taken in each of
subtracted from the data, the resulting timer error is shown in Figure the two galleries. K-1 was about a
factor of two less stable during its time
B9. The change in slope at day 165 clearly shows a frequency step in the second gallery than in the first.
change at that point. Notice the hump in the stability plot
for τ values of the order of one to two
weeks. This is apparently due to the
steps in frequency occurring every few
weeks. Notice also that the long-term
τ = 128 days value is almost in line with
extrapolated stability curve from the
first segment. In other words, it
appears that whatever caused K-1’s
rate to go high for a few days, this
effect tended to average out over long
enough time — a very good attribute.

69
 One wonders if something was done to the clock to induce this change.
This is clearly not approximated by a parabola, so no frequency drift TIME ERROR DATA
Mudge No. 1 Chronometer
was subtracted. The frequency stability of this clock is shown in Figure
1000
B10 — a most commendable result for that period. The lower curve is

Time Error (seconds)

Mod.σy(τ), as explained in Appendix A, and the upper curve is σy(τ). 800
Mod.σy(τ) was employed to detect any measurement noise. The steep, 600
downward slope at the shortest τ values is indicative of some. The level
400
is given by
200
( ) = 0.14
τMod .σ y τ
0
seconds at τ=1 day, which is also a most com- 0 50 100 150 200 250
3 Number of Days
mendable result for that period.
Figure B7. A plot of the Mudge No. 1
chronometer data taken in 1777 A.D.
This is the reason for the flat bottom portion — indicative of measure- over a 203 day interval — the
ment noise — for the data shown in Figure 12 for the Harrison-like chronometer’s rate being determined
over a one day interval. The time error
chronometers. The Mudge data appears to hit a flicker floor at about shown has an overall rate of 3.623
1.7 or 1.8 × 10–6 for the large τ values — of the order of weeks. This is seconds per day with very little noise
about a factor of ten better than would have been necessary for the around this slope. If observed carefully, a
slight change in slope occurs toward the
Longitude prize. end of the data — at day number 165.

The frequency change shown by the slope change in Figure B8 was

measured, and the first and second sets (before and after the frequency FREQUENCY DATA
Mudge No. 1 Chronometer
step change) were tested for the noise type of the residuals. Figure B11
Clock Rate (seconds per day)
5.5
is a plot of the frequency stability for the first 164 days. In an ideal
5.0
frequency standard, the residuals would be random and uncorrelated
— like our coin-toss experiment in Appendix A. In this case, the 4.5
frequency stability diagram, σy(τ), would behave as τ–1/2. This is also 4.0
called white noise frequency modulation (FM), and a white noise FM 3.5
line has been drawn on Figure B11 for illustrative and comparison 3.0
purposes. 2.5
0 50 100 150 200 250
Number of Days
For longer τ values the stability values rise up above a τ–1/2 slope
indicating that the noise is moving toward a flicker-noise like process. Figure B8. A plot of M-1’s daily-average
That it follows as close as it does to the theoretically ideal τ–1/2 behav- frequency values. The noise in the
chronometer and the frequency step at
ior is astounding for a clock of that vintage. day number 165 are more apparent.

TIME ERROR minus mean freq

Mudge No. 1Chronometer
5.0
Time Error (seconds)

–05.0

–10.0

–15.0
–20.0
0 50 100 150 200 250
Number of Days

Figure B9. After subtracting a calibrated

clock rate of 3.723 s/d from the data, the
timer error is plotted. By this approach
the frequency step is readily apparent.
Any frequency drift appears to be below
70 the noise.
Another simple test for non-whiteness of the noise has been developed
in which the ratio of the classical variance (standard deviation Tau Sigma Mod Sigma
8.64e+04 2.74e–06 2.74e–06
squared) to the two-sample or Allan variance is calculated. It is both a 10 –5 1.73e+05 2.05e–06 1.69e–06

Allan Deviation, σy(τ)

3.46e+05 1.76e–06 1.40e–06
necessary and a sufficient test for non-whiteness if this ratio is not 1. 6.91e+05 1.62e–06 1.32e–06
Furthermore, its departure from 1 is indicative of the kind of noise and 2
1.38e+06
2.76e+06
1.48e–06 1.26e–06
1.74e–06 1.04e–06
the degree of time dispersion that characterizes the clock [42]. For this 3
data set the ratio was 1.26. If the noise is white FM, then the uncer-
2
tainty on the optimum estimate of the frequency, which is the simple
1.5
mean, is given by the standard deviation of the mean as explained in
10 –6
Appendix A. Because the noise is close to, but not white, we used the 10 4 2 10 5 2 10 6 2 10 7
above ratio as the dispersion factor for the uncertainty of the fre- Averaging Time, τ, Seconds
quency estimates before and after the step. The statistical behavior of
the second set was very similar to the first. The clock rate (frequency) Figure B10. Frequency Stability of Mudge
No. 1 Chronometer.
step size was (0.50 ± 0.06) seconds/day.

For the electrical engineer or physicist who may be reading this, and
Tau Sigma
for any others interested in the spectral density of these clock residual 8.64e+04 2.66e–06

Total Allan Deviatiom, σtotal(τ)

fluctuations, a fast Fourier Transform was made of both clock sets. 10 –5
1.73e+05
3.46e+05
1.91e–06
1.64e–06
Nothing untoward was observed. Perhaps one small point of interest 6.91e+05 1.50e–06
1.38e+06 1.14e–06
was that the last 512 days of the Kendall No. 1 Chronometer data set 2.76e+06 8.10e–07
2
showed a slight indication of a one-cycle-per-week Fourier component
being present (see Figure B12). The amplitude of this 1/7 cycle-per-day 10 –6
component is about = 1.4 seconds per day. Not surprisingly, all the
data sets viewed in total or in segments had a negative slope to the 2
spectral density plot — indicating a non-whiteness in their residual 10 –7
10 4 2 10 5 2 10 6 2 10 7
behavior. The line fit to the Figure B12 data is f –0.83 — very close to
Average Time, τ, Seconds
flicker noise, which is f –1.
Figure B11. Frequency Stability of Mudge
One of the principle values of clock ensembles is to sense and reject No. 1 Chronometer for first 164 days.
abnormal behavior in a clock, such as the above frequency step. In this
way and because of the random uncorrelated nature of the noise, the
ensemble time can be better than the best physical clock making up
the ensemble. Even so, the reliability, ruggedness and performance of 3.00
Log PSD, Sy(f), seconds /day

these Harrison and subsequent chronometers are truly astounding and Window:
2

1 Hanning
have clearly revolutionized how we navigate. GPS is having a similar 1.50
impact now.
0.00

–1.50
Note: same component at one
cycle per week
–3.00
10 –3 10 –2 10 –1 10 0
Fourier Frequency, f, Cycles/day

Figure B12. Power Spectrum of Kendall

No. 1 Chronometer for last 512 days.

71
 Appendix C

Time and Frequency Transfer, Distribution

and Dissemination Systems
Tables C1 and C2 are taken from the International Telecommunication
Union’s handbook entitled, The Selection and Use of Precise Fre-
quency and Time Systems. These tables provide an overall summary of
the current methods that are available for both obtaining a time and
frequency calibration source and for comparing clocks located remote
to each other. In addition, where UTC is available, the time accuracy
with which it can be obtained is also given.

Figures 13 and 14 show the frequency and time stability of some of

these techniques for time and frequency distribution, dissemination or
transfer. Caution should be exercised in using any of these values
because there is often a path dependence as well as a time dependence;
in other words, the locations of the clocks can make a difference along
with the time of day or night, the time of year, and perhaps where we
are in the solar sunspot activity cycle.

For the best calibration or comparison of a clock, a systems approach

is often the best, considering the following questions: What are the
characteristics and labor intensity to use the different techniques for
calibration or comparison? What are the overall associated measure-
ment uncertainties — considering both the random and systematic
effects? What are the nominal characteristics of the clock being cali-
brated or compared and how do the measurement uncertainties impact
the calibration or comparison process? There are some extremely cost-
effective ways now available to transfer, distribute or disseminate high-
accuracy time and frequency signals. Good planning should also
include considerations for reliability and robustness.

Meanings of designations in Table C1:

1 HF (High Frequency) broadcast (3.0 MHz – 30 MHz).

2 Rec. (Recommendation) 768 of the Study Group VII of the
ITU-R.
3 LF (Low Frequency) broadcast (30 kHz – 300 kHz).
4 Loran-C, 100 kHz ground-based navigation chain of transmitters.
5 VLF (Very Low Frequency) broadcast (3 kHz – 30 kHz).
6 Omega, each station operates a different VLF carrier frequency
— coordinated globally.
7 GOES (Geostationary Operational Environmental Satellite) East
and West satellite covers the Americas.
8 INSAT (Indian Satellite)
9 DBS (Direct Broadcast Satellite)

72
Table C1. Characteristics of some potential sources and dissemination techniques for precise time-and-frequency
reference information

Type Typical time- Typical Coverage Availability Ease of Use Approximate Example Comments (1995)
transfer frequency relative system
accuracy transfer user cost
capability capability ($ US 1995)

HF broadcast1 1-10 ms 10–6 to 10–8 Global Continuous, Depends on 50 to 5,000 Many Accuracy depends on
(over 1 day) but operator accuracy services path length, time of day,
and location requirements worldwide. receiver calibrations,
dependent See Rec. 7682 etc.

LF broadcast 3 1 ms 10–10 to 10 –11 Regional Continuous Automatic 3,000 to See Rec. 768 Depends on distance
5,000 from the source and
diurnal propagation
(ionosphere height)

LF navigation 1µs 10–12 Regional Continuous Automatic 5,000 to Loran C4 Northern hemisphere
(pulsed) 12,000 coverage. Stability and
accuracy based on
ground wave reception.

VLF 10 ms 10–11 Global Continuous Automatic 4,000 Omega6 Carrier resolution can
broadcast 5 (over 1 day) provide better time
accuracy

Television 10 ns for 10–12 to 10–13 Local Dependent Automatic 5,000 Calibration required for
broadcast common view (over 1 day) upon local timing
(terrestrial broadcast
links) schedule

Navigation 20-500 ns 10–9 to 10–13 Global Continuous Automatic 3,000 to GPS and One day averaging
Satellite, (See notes 15,000 GLONASS necessary to meet
broadcast in Table C2) specified frequency
transfer capability. Best
broadcast system
available today with
commercial receivers.

Navigation 5-20 ns 10–13 to 10–15 Inter- Continuous Automatic 10,000 to GPS and Most accurate, widely
satellite, (over 1 day) continental (calculated data 20,000 GLONASS used time synchroniza-
common after the fact) acquisition. per site tion method that is
view Requires post available today (1995)
processing. with commercial
receivers for baselines
less than 8000 km.

Meteoro- 100 µs Not recom- Regional Continuous Automatic 4,000 to GOES7 May not be available
logical mended for (satellite 5,000 during satellite eclipse.
satellite, frequency footprint)
broadcast transfer

Geostationary 20 µs 5 × 10–10 Regional Continuous Automatic 4,000 INSAT8 Accuracy limited by

satellite (satellite satellite footprint. May
multipurpose footprint) not be available during
broadcast satellite eclipse.

73
 Table C1. Characteristics of some potential sources and dissemination techniques for precise time-and-frequency
reference information (continued)
Type Typical time- Typical Coverage Availability Ease of Use Approximate Example Comments (1995)
transfer frequency relative system
accuracy transfer user cost
capability capabilty ($ US 1995)

Television 0.5-10 µs 10–10 to 10–11 Regional Dependent on Automatic 7,000 DBS9 Without correction for
broadcast (satellite broadcast data Satellites satellite position
satellite footprint) schedule acquisition

10-100 ns 10–12 to 10–13 Regional Dependent on Post- 7,000 DBS With correction for
(satellite broadcast processing of Satellites satellite movement
footprint) schedule data required

Communica- 1-10 ns 10–14 to 10–15 Regional Continuous Data 50,000 North Most accurate
tion (satellite (as scheduled) acquisition per site American operational method at
satellite, footprint) can be and this time.
two-way automatic European
(depending networks
on satellite). exist.
Post
processing
required.

Telephone 1-10 ms 10–8 Telephone Continuous Automatic 100 Europe and Phone line must have
time code (over 1 day) calling North same path in both
range America directions. Assumes
computer and software
availability.

Optical fiber 10-50 ps 10–16 to 10–17 Local, less Continuous Automatic Transmitter Dedicated to Cable must be
than 50 km and receiver frequency temperature stabilized,
$30,000 per transfer (e.g. 1.5, underground).
set plus
cable and
underground
installation
costs.

100 ns 10–13 to 10–14 Long Continuous Automatic Not Synchronous Part of digital
(over 1 day) distance, applicable. Digital communication system
2000 km The Hierarchy
equipment (SDH)
is a part of network
a specific
communica-
tion system.

Microwave 1-10 ns 10–14 to 10–15 Local Continuous Automatic 50,000 to Sensitive to atmo-
link 75,000 spheric conditions and
multipath effects. Must
be 2-way to achieve
stated accuracy and
stability.

Coaxial cable 1-10 ns 10–14 to 10–15 Local Continuous Automatic 5 to 30 Sensititve to tempera-
per meter ture, VSWR, humidity,
barometric pressure.

74
Table C2. Additional information relating to the practical use of the various alternative sources of
time-and-frequency signals

System/Technique Background Information Comments on equipment and use

HF broadcasts There are approximately 13 stations worldwide broadcast- Inexpensive receivers and antennas available. Diversity receivers
ing on one or more of the allocated HF frequencies. use multiple HF frequencies to partially compensate for
Several others operate on other HF frequencies. Typical propagation effects. Simple short- or long-wire antennas are
services include standard frequencies, time signals and often usable. Other antenna design information can be found in
time intervals, time codes, voice time announcements, and amateur-radio handbooks.
UT1 time information. These services provide a convenient, Reception is generally better for the lower frequencies (<10 MHz)
easy-to-use source of UTC at modest accuracy levels. during nighttime hours and for the higher frequencies (>10 MHz)
Although HF signals can be received at large distances, during daytime hours. Reception may be intermittent due to
propagation effects can limit received accuracy and propagation disturbances and/or interference. Optimum reception
stability. Multiple stations operating on the same allocated is usually during daytime or nighttime hours when the ionosphere
frequencies may cause mutual interference in some areas. is most stable.
Reception conditions are often highly variable, depending
on factors such as season, time of day, solar activity, Voice time announcements provide a few tenths of a second
atmospheric conditions, etc. Some HF services are being accuracy. For better accuracies down to about 1 ms special
shut down in favor of other alternatives. ITU Recommenda- measurement techniques and equipment, such as oscilloscopes
tion ITU-R TF.768 contains a complete listing of HF and electronic counters, may be required. Receiver delay
services, including details of the content and format of the calibration is also necessary for highest accuracy.
broadcasts. Frequency-measurement accuracy is limited to about 1 × 10–7 by
ionospheric motion. Beat-frequency techniques are often used
along with oscilloscopes and/or counters. Frequency measure-
ments may also be inferred from daily time-difference measure-
ments.
Calculation of signal path delays is complicated by uncertainties
in the number of signal “hops” between the station and the user
and the height of the reflecting layer at any point in time. Single
hops can usually be assumed for distances of less than 1600 km.

LF broadcasts This category includes broadcasts operating in the LF band Relatively inexpensive receivers and antennas are available from
(30 - 300 kHz) that are useful sources of UTC time or commercial sources in regions served by suitable broadcasts.
frequency but excluding navigation-system broadcasts Commercial receivers are self-contained and provide a variety of
such as Loran-C. These broadcasts are of two types: outputs that can often be specified by the user. More sophisti-
(1) dedicated time-and-frequency dissemination services cated phase-tracking receivers are also available which allow
such as DCF77, HBG, WWVB, and JJF2; and (2) stations users to establish direct frequency traceability to accepted
operating in the sound-broadcasting service that have sources for UTC.
stabilized carriers and/or additional phase-or-amplitude Typical antenna types for these broadcasts include long-wire
modulations that provide coded time information. The designs (e.g., 50 - 100 meters), whip antennas (e.g., 3 meters), air-
dedicated services generally use frequencies in the loop antennas which are helpful in discriminating against
40 - 80 kHz range. interference, and small ferrite-loop antennas.
Many of these LF broadcasts provide users with very Reception conditions vary with the transmitter power, the user’s
complete time-of-year information coded form and have location, and, in some cases, the season and time of day. For
found wide acceptance in many timekeeping applications. longer paths between transmitter and user, avoid making
Time accuracies of less than 1 ms are possible. When measurements when there is sunrise or sunset anywhere along
used as a frequency standard, LF broadcasts, including the the path.
stabilized sound broadcasts, offer calibration accuracies
of less than 1 × 10–11 when averaged for about 1 day. Frequency calibrations of local oscillators may be performed by
Reliable coverage areas of the various broadcasts range continuously monitoring the phase difference between the local
from a few hundred kilometers up to 3000 km. oscillator and the received LF broadcast. Proper evaluation of the
resulting phase recordings, however, requires some operator skill
For more details on the available broadcasts for time-and- and experience in interpreting and accounting for various phase
frequency use, see ITU Recommendation ITU-R TF.768. shifts and possible “cycle slips”.
Destructive interference may occur between the first-hop
skywave and the groundwave, causing a sharp drop in received
field intensity at certain distances from the transmitter. For a
60 kHz LF broadcast, this distance is about 1200 km.

75
 Table C2. Additional information relating to the practical use of the various alternative sources of time-and-frequency
signals (continued)

System/Technique Background Information Comments on equipment and use

LF navigation Approximately 65 Loran-C stations scattered throughout Special Loran-C timing receivers and antennas are available
broadcasts (pulsed) the northern hemisphere continuously broadcast high- commercially. The more expensive models acquire and track
power navigation signals on a frequency of 100 kHz. These appropriate Loran-C signals automatically. Non-automatic
stations are arranged in chains of 4-5 stations each. Each receivers require significant operator experience and skill for
chain transmits groups of precisely controlled pulses at an optimum timing performance.
assigned unique Group Repetition Interval. Because the In order to use Loran-C to keep a local clock steered to UTC an
navigation signals are synchronized and syntonized by output 1 Hz pulse from the receiver can be synchronized to UTC
atomic standards and are carefully monitored and by using “Time of Coincidence” tables published by the US Naval
controlled, they can be very useful as time-and-frequency Observatory. These tables give the specific times when the start
references. of the Loran-C signal being received is coincident with a UTC
The Loran-C transmissions do not contain complete time- second. Some Loran-C timing receivers can perform this
of-day information and are not a direct source of UTC time. synchronization automatically.
However, if a user’s clock is initially set to UTC by some At large distances from the station the Loran-C skywave signal
other means, Loran-C can be used to keep the local clocks may sometimes be used for timing at the 50 - 100 ms level even
to within a few microseconds of UTC over long periods of when the primary groundwave signal is unusable.
time. Frequency calibrations using Loran-C can provide
1 × 10–12 accuracy when averaged over 1 day or more. Seasonal effects on Loran-C propagation may cause timing
variations of several microseconds. At this level receiver delays
Although reception from at least three different stations is also need to be considered.
necessary for navigation, time-and-frequency measure-
ments require reception from only a single station. Frequency calibrations of a local oscillator can be accomplished
by recording the phase difference between Loran-C and the local
system or by daily measurements of the phase difference using a
counter. Accuracies as good as 1 × 10–12 are possible with
24-hour averaging.
The development of very low cost Loran-C receivers (under
$1,000) for navigation creates some possibilities for their
adaptation for time-and-frequency applications, providing that the
necessary technical expertise is available.

VLF broadcast There are a number of broadcast stations operating in the Typical equipment used includes phase-tracking receivers, loop
10 - 30 kHz range that are useful for time-and-frequency antennas, and chart recorders. Receiving system delays need to
applications. These include broadcasts primarily intended be calibrated for best results.
for long-distance communications or navigation but which
are highly stabilized in frequency and time by referencing Receivers used with MSK transmissions need to reconstruct a
to multiple atomic standards. Propagation is relatively phase-coherent carrier by suitable multiplication and mixing. For
stable over very large distances (thousands of kilometers), further information on MSK signals see Note #10 to Table 2 in
which can permit phase-tracking receivers to maintain ITU Recommendation ITU-R TF.768.
phase to within a few microseconds over long periods of Omega stations are located in the United States (North Dakota
time. VLF broadcasts typically do not contain complete and Hawaii), Japan, Argentina, La Reunion, Liberia, Norway, and
UTC time information and are useful primarily as a Australia. Since each station transmits multiple frequencies in
frequency reference. sequence, use of one of the Omega navigation frequencies for
The Omega Navigation System is one VLF system that is calibration requires that a commutator be used to turn the phase-
useful for time-and-frequency applications. It features tracking receiver on and off at the proper times in order to receive
eight worldwide, 10-kW transmitters providing continuous only the particular frequency of interest.
and redundant global coverage. Each station transmits the Propagation effects often limit the useful accuracy of VLF signals,
four navigation frequencies of 10.2, 11.05, 11.33, and 13.6 especially for very long path lengths. There are, for example,
kHz sequentially in a time-shared mode. Other “unique” predominant diurnal and annual variations caused by ionospheric
frequencies in the 10 -13 kHz range are also transmitted by changes. Results may also be influenced by unpredictable sudden
each station. ionospheric disturbances (SID), which typically alter the
Several nations also operate VLF communication stations ionosphere for 20 - 30 minutes, and by polar cap absorption (PCA)
that are useful, particularly for frequency calibration. At events, which alter the polar ionosphere for up to a week.
least some of these stations operate in an MSK (minimum In addition to the diurnal and annual variations in propagation
shift keying) mode, requiring the use of special receiving delays at VLF, other variations have been observed with periods
equipment and techniques to recover a phase-stable of 27, 29.53, and 14.765 days due to various solar and lunar effects.
carrier frequency.
In recent years the use of VLF broadcasts for time-and-frequency
comparisons has declined due to the emergence of other systems
and techniques.
76
Table C2. Additional information relating to the practical use of the various alternative sources of time-and-frequency
signals (continued)

System/Technique Background Information Comments on equipment and use

Television A number of different techniques have been tried for time- Typical equipment needed includes suitable television receivers,
broadcast and-frequency dissemination and comparison that use antennas, counters, and data recorders. The television receivers
(terrestrial links) television broadcast signals. These include the insertion of must be modified to extract the particular synchronization pulse
time-and-frequency information into the television signal, from the received TV signal.
the stabilization of television carrier frequencies and At each measurement site arrange for the local clock pulse to
synchronization pulses, and the common-view reception of start the counter and the received TV signal to stop the counter.
a single television broadcast at multiple sites within a local About 10 such once-per-second measurements are usually
area. The first two techniques are still in use in limited sufficient to achieve excellent results.
geographical areas, but the common-view reception
technique is the most widely used television method. Since the measurements must be made simultaneously at each
site and the resulting data must be exchanged, active cooperation
The common-view method allows the precise time among the sites is necessary.
comparison among multiple sites within the coverage area
of a single TV station. Each site simultaneously measures By making such comparisons each day over a period of time, very
the time difference between a particular synchronization accurate frequency comparisons are possible based on the
pulse in the TV signal and its local clock. Subtracting the observed changes in the daily time differences. This assumes that
measurements from two different sites provides the the differential propagation path delay remains stable or is
difference between the local clocks plus a fixed differen- independently calibrated each time.
tial propagation delay. The local clock comparisons have a The technique is especially advantageous within a limited local
typical uncertainty of about 10 ns. region because of its simplicity, relatively low cost, and high
accuracy.

Navigation satellite There are two major satellite navigation systems in use as A variety of receivers are commercially available, especially for
broadcast of 1995 which offer outstanding time-and-frequency the GPS broadcasts. Some versions have been produced which
dissemination capabilities. These are the US Global can receive both GPS and GLONASS. Very small omnidirectional
Positioning System (GPS) and the Russian Global antennas are usually provided with the receivers. Costs have
Navigation Satellite System (GLONASS). The US Transit decreased sharply with the increasing demand and timing
system offers a third choice, but it will not be discussed receiver packages are available in early 1995 for $3,000 - $15,000.
further in view of its lower accuracy, higher cost, and Most receivers are highly automated. During initial setup they can
anticipated limited lifetime. be programmed to automatically track enough satellites to
While there are some differences between the two determine the receiver coordinates with sufficient accuracy to
systems in terms of signal structure and content, use of the support submicrosecond timing. Some care must be used in
frequency spectrum, and satellite orbits and configuration, locating the antenna to minimize multipath effects. After setup,
their similarities are much more important for time-and- receivers can continue to acquire and track all selected satellites
frequency users. Both GPS and GLONASS employ in a totally automatic mode.
redundant on-board atomic clocks, continuous global Many receivers can be easily controlled by the user to track only
coverage from 21 - 24 operational satellites, precisely certain satellites at certain times. Time differences between the
timed broadcasts which can be related to UTC(USNO) and received GPS signal and a local clock can often be stored in the
UTC(SU), respectively, to within 100 ns, and satellite- receiver’s memory for later analysis.
position information included in the broadcasts which can
be used for accurate path delay compensation by the Although the times of individual GPS and GLONASS clocks differ
user’s receiver. At least four satellites are always in view from the overall satellite system time which, in turn, differs from
from any location (required for navigation), but reception UTC, sufficient additional data are included in the satellite
from only one satellite is sufficient for time-and-frequency broadcast formats to allow a receiver to, in principle, adjust its
comparison. For accurate time the receiver’s antenna output timing signal to be with about 100 ns of UTC(USNO) or
position must also be known. UTC(SU). The actual display and output times, and their
relationship to the relevant UTC time scales, may vary from
Both GPS and GLONASS are essentially fully operational receiver to receiver, depending on the particular manufacturer
as of 1995 and provide a combined total of more than 40 and model and the effects of Selective Availability in the case of
satellites for time-and-frequency applications. Commercial the GPS signals.
development of receivers is proceeding rapidly with a
(continued) resulting sharp decrease in user costs.

77
 Table C2. Additional information relating to the practical use of the various alternative sources of time-and-frequency
signals (continued)

System/Technique Background Information Comments on equipment and use

Navigation satellite The presence of Selective Availability (SA) on the GPS Typical timing accuracies of 20 - 500 ns and frequency accuracies
broadcast signal overtly degrades the GPS timing information. A of 10–9 to 10–13 (depending on various factors) make these
(continued) Memorandum of Agreement between the US Department navigation satellite systems the best current broadcast source of
of Defense and the Department of Transportation highly accurate time-and-frequency for use with commercial
guarantees the availability of GPS, and the SA level is receivers.
expected to remain at about the current (1995) level. The Given the level and characteristics of GPS SA, a system approach
civilian use of GPS has surpassed the military use and ratio for obtaining UTC timing information has proven beneficial in
of civilian-to-military use is expect to continue increasing. inexpensive multichannel receiver designs. Using a systems
It is the intention of the U.S. to discontinue the use of GPS approach, the final time-and-frequency accuracy and stability will
Selective Availability within a decade. Beginning in 2000, depend on the receiver, the characteristics of the reference clock
the U.S. President will make an annual determination on and the processing algorithms. The performance of the output
continued use of GPS SA. improves with the performance of the reference clock. The
processing algorithms can also significantly impact the output
performance of the timing system. Levels of performance
achieved using quartz oscillators, rubidium frequency standards,
and cesium-beam frequency standards for the reference
oscillator are 10–11, 10–12, and 10–13, respectively. Timing
performance with respect to UTC can be better than 100 ns.

Navigation satellite For general background information on the GPS and Each site participating in a common-view measurement needs an
(common-view GLONASS systems see the preceding entry in this Table. appropriate GPS or GLONASS receiver and antenna, data-
mode) recording capabilities, and a communication link to other
In the common-view mode of operation with GPS or
participating sites. Accurate receiver location is also required, but
GLONASS users at two separated sites each receives a
this can often be determined automatically by the receiver itself
signal from the same satellite at the same time. Subtract-
operating in the navigation mode if the position is averaged over a
ing the (satellite – local clock) data from the two sites
few days.
provides the time difference between the local clocks. The
advantage is that, in this process, variations or errors in Care must be taken to ensure that the measurements extend over
the satellite clock are common to both paths and therefore exactly the same time period at each site. The receiver must also
cancel. If the SA degradation process for GPS is be programmed to track the proper satellite that is in common
implemented so as to cause variations in the satellite view with the other sites. Typical track lengths are 13 minutes.
clock, such changes do not affect the common-view A subcommittee of the Consultative Committee for the Definition
measurement accuracy. On the other hand, if SA causes of the Second has recommended standard data formats and other
satellite-position errors to be broadcast, such errors will procedural matters to facilitate the use of this method on a
not be totally compensated for in the common-view regular basis.
measurement because each site is receiving the signal
over a somewhat different path. The technique is usable for baselines between sites of up to
8,000 km.
The common-view method allows time comparison
accuracies of 5 - 20 ns over intercontinental distances, The results from many regular common-view time comparisons
even in the presence of SA (as it is presently implemented among national and international timing centers are published
in early 1995). Frequency comparisons can be derived from and archived by the BIPM.
such data to an accuracy of 10–13 to 10–15. To facilitate Receiver-system delays should be calibrated for the highest
common-view time comparisons among timing laborato- possible comparison accuracy and the antenna coordinates
ries throughout the world the BIPM in Paris generates and should be known to within less than 1 meter.
distributes suitable common-view tracking schedules
showing which satellites are appropriate for this method at The use of multichannel receivers in the common-view mode can
various times. provide a convenient frequency-transfer capability at the 10–14
level. The potential exists, for example, by using the GPS carrier
phase, for extending this performance down to the 10–15 region by
averaging over days.

78
Table C2. Additional information relating to the practical use of the various alternative sources of time-and-frequency
signals (continued)

System/Technique Background Information Comments on equipment and use

Meteorological Since 1974 the US Geostationary Operational Environmen- Commercial receivers with small antennas are available from
satellite broadcast tal Satellite System (GOES) has included a time code several manufacturers. Recent versions use the transmitted
referenced to the UTC(NIST) time scale. The time code is satellite position information to correct for path delay and update
disseminated continuously from two geostationary it each 1 minute. Initial setup requires the operator to enter the
satellites located normally at 75 and 135 degrees West position coordinates of the receiver location.
longitude. Satellite position data are also transmitted to The GOES time code transmissions are at 2 frequencies near
users so that suitable automatic receivers can compute 469 MHz. Because these frequencies are also allocated to the
the signal path delay and correct their 1-Hz outputs land-mobile service in the US, some interference, particularly
accordingly. Specified time code accuracy as delivered to near large metropolitan areas, can be expected. Receivers are
the user is 100 µs. The normal time code coverage area reasonably effective in “flywheeling” through such periods of
includes most of the Western hemisphere with overlapping interference.
coverage of much of North and South America.
In regions of low signal strength or frequency interference use of
The GOES time code includes information on the current simple helical or Yagi antennas may improve reception.
year, day of year, hour, minute, second, UT1 correction,
system accuracy, and indicators for Daylight-Saving Time The received time code typically shows diurnal variations with a
and leap seconds. peak-to-peak amplitude of 10 - 70 µs due primarily to imperfec-
tions in the software used to compute satellite position
predictions.
The European Meteosat system and the Japanese Geostationary
Meteorological Satellite (GMS) system are basically similar to the
GOES system but do not currently transmit a time code.
GOES satellites suffer time-code signal outages for about 2 hours/
day during Spring and Fall eclipse periods each year. Receivers
cost about $5,000 (1997).

Geostationary The Indian INSAT geostationary satellites also transmit a Commercial receivers are available (1993) at a cost of about
satellite, UTC-referenced time code as one feature of this $4,000. Antenna requirements are modest.
multipurpose multipurpose system. As in the GOES case the time code
broadcast signal also includes satellite-position information which
allows the user to compute and compensate for the signal
path delay.
The INSAT satellite footprint limits primary coverage to the
region of the Indian subcontinent. Within this region the
time accuracies of about 20 µs and frequency accuracies
of 5 × 10–10 are possible.

Television The measurement technique is the same as reported in the The equipment needed includes a small dish antenna, a
broadcast case of terrestrial links, but the signals are received in commercial satellite TV receiver, and a TV-synchronizing-pulse
(satellite links) common view from a direct-broadcast satellite (DBS), extractor.
extending the coverage area to a nearly continental
dimension. A time interval counter at each site measures the time differences
between the local clock pulse and the received TV signal from the
The main source of error in the determination of the clock satellite. Two series of at least 10 such measurements, taken 12
differences arises from the variations in the position of the hours apart, are needed daily. A data-acquisition system is also
geostationary satellite used. This drawback can be needed for data storage and exchange with the other stations for
reduced in different ways, leading to the accuracy ranges processing of the results.
reported in Table C-1.
The correction for satellite-longitude drift which degrades the
It is possible to remove the 12-hour and 24-hour periodic results can be obtained in several ways: (1) from the satellite
variations by averaging and also, most importantly, to position parameters supplied by the satellite-control station; (2)
remove the satellite longitude drift observed in the time from pseudorange measurements performed by a single station;
comparisons with various techniques. (3) from GPS satellite measurements performed by at least 3
stations; or (4) from the time measurement performed at 3 ground
stations that observe 2 satellites.

79
 Table C2. Additional information relating to the practical use of the various alternative sources of time-and-frequency
signals (continued)

System/Technique Background Information Comments on equipment and use

Communication At the current time the most precise and accurate method The earth-station equipment needed at each user site must be
satellite (two-way) for time comparisons between remote sites is the compatible with the particular satellites being used for time
simultaneous, two-way exchange of timing signals through transfer. Typical costs, including the necessary modems, may
communication-satellite channels. The high accuracy reach $50,000 per site. Operator skills needed for proper operation
achievable results from the use of a two-way exchange of may be more stringent than for most of the other techniques
signals which effectively eliminates the need for precise discussed.
knowledge of the satellite’s position, the high degree of Since the two-way technique is essentially a point-to-point
path reciprocity in the two directions, and the wide communication system, it should not be regarded as a general
bandwidth of the satellite channel which permits efficient dissemination technique.
signal design.
As typically implemented, two or more sites exchange timing
One disadvantage of the technique is the need for each signals on a regular basis, several times per week. Because of the
site to both transmit and receive signals and then to inherent time stability of the method, it is usually only necessary
exchange the data for post-processing. The earth station to perform the exchanges for a few minutes per time. The
equipment at each site tends to be rather expensive, measurement process involves measuring the difference between
especially if the system is highly automated. Participants in the satellite-signal arrival time and the local clock. Such
the time transfers must coordinate with each other and measurements are often made once per second for a period of a
with the satellite-system operator. few minutes. Subtraction of the simultaneous measurements at
Because of the potential accuracy of near 1 ns and the each site, divided by 2, provides the difference between the site
precision of 0.1 – 0.5 ns, many timing laboratories in clocks (except for corrections that may be needed to account for
various parts of the world are developing a two-way differences in equipment delays).
time-transfer capability. Special modems are being For the highest achievable accuracy of 1 – 10 ns it is important to
developed which are optimized for high accuracy and calibrate the signal delay through the ground-station equipment.
long-term stability. Suitable satellite channels appear to be This may be a difficult problem since the relevant quantity needed
available throughout the world at reasonable cost. is the difference between the delays through the transmit and
Frequency transfer is obtained most efficiently by using receive portions of the system. Several specialized techniques
continuous data.* If, for example, one hour’s worth of one- have been developed for this purpose.
second data are taken with a standard deviation of 200 ps, Depending on the particular satellite system being used and the
and the data are well modeled by white noise PM, then the locations of the stations, extensive administrative procedures may
frequency transfer uncertainty is only 3.2 × 10–15. be required in order to certify the earth-station equipment and
gain acceptance for satellite access.

Telephone time A number of timing centers in Europe and N. America have Equipment requirements to use such services are minimal. Aside
code (two-way) established services designed to disseminate coded UTC from the computer or other equipment containing the clock to be
time information over telephone lines in an automated set, only a suitable modem, access to a telephone line, and clock-
mode. Typically, computers or other automated systems setting software is needed. In order to perform the path delay
are programmed by the user to dial such services as compensation the user may also need to be able to echo the
needed, receive an ASCII time code from the timing center, received signal back to the timing center.
reset the local clock to the correct time, and, in some
Usually, a telephone connection time of only a fraction of a minute
cases, to automatically compensate for the path delay
is needed to perform a satisfactory time transfer.
through the telephone link. Depending on the particular
service, the path delay compensation can be performed Software for using such services is relatively simple to develop by
either by the time center’s equipment or at the user’s site. users or some versions of example software are often available
The compensation for delay is based on measurements of via computer bulletin boards, from the timing centers, or from
the round-trip delay time and assumes that the path is commercial sources at reasonable cost.
reciprocal. Most of the available telephone services can also be used on a
Time-transfer accuracies of 1 - 10 ms are possible, even one-way mode where there is either no compensation for path
when satellite links may be involved. In addition to the UTC delay or a fixed, average delay is used. Accuracy for this mode
time of day, most services established to date also include may be in the range of 0.1 to 0.5 seconds.
information on the year, day of year, UT1 corrections, leap By making periodic measurements of a local clock using one of
second warnings, and indicators for Daylight-Saving Time. the telephone services, an average frequency can be determined.
Accuracies of about 10–8 are possible with 1-day averages. The
most efficient frequency transfer accuracies are obtained by
staying on-hook. * (See Figure 13.)
–3
* In the case where the measurement noise is limited by white-noise PM, the frequency transfer uncertainty (1σ) is given by
12 N 2 σ / τ 0 ,
where N is the number of time difference measurements, σ is the standard deviation of those measurements, and τ0 is the data spacing.
80
Table C2. Additional information relating to the practical use of the various alternative sources of time-and-frequency
signals (continued)

System/Technique Background Information Comments on equipment and use

Optical fiber Optical fibers offer excellent potential for transferring time- In a practical implementation of a fiber-optic link for time-and-
and-frequency signals with very high accuracy over both frequency transfer at highest possible accuracy levels, it is
short (<50 km) and long distances. While dedicated UTC important to stabilize the temperature of the cable. The nominal
dissemination services using optical-fiber distribution do coefficient of delay with respect to temperature is 7 ppm/°C. In
not currently exist, the technique is included here in order to meet the stated performance in Table C-1 for links longer
recognition of its future potential. than 50 km, the cable should be put underground to a depth of at
least 1.5 m.
Two types of fibers, multimode and single mode, are in use
today. Multimode fiber is generally used to transmit digital For a dedicated optical-fiber link for time-and-frequency transfer,
data and low frequencies over a relatively short distance the cost is about $30,000 per site for transmitters and receivers
(e.g., 1 km). Single-mode fiber is best for longer distances plus the cost of the cable and its underground installation.
(e.g., 50 km) and supports wide bandwidth (e.g., 5 MHz to
Insertion loss is about 0.5 dB/km.
100 GHz). Single-mode fiber with a 1,300 nm laser is
required to meet the performance given in Table C-1 for Potential users and suppliers of UTC should maintain current
local distances. awareness of the development of regional, national, and
international digital synchronized telecommunication networks.
The accuracies stated in Table C-1 for long fiber-optic links
Such networks may provide an excellent, convenient means for
have been achieved in a digital telecommunications
distributing high-accuracy UTC time-and-frequency in the future.
system adhering to CCITT Recommendations G.707, 708,
and 709 over a distance of 2,400 km. This particular system
was designed to meet ITU-T requirements as well as to
perform time-and-frequency-transfer experiments.

Microwave link The use of microwave links to distribute time-and- Equipment is relatively expensive ($50,000 - $75,000).
frequency within local areas can provide accuracies as Results are sensitive to atmospheric conditions and multipath
high as 1 – 10 ns for timing and 10–14 to 10–15 for frequency effects.
when used in a two-way mode.
For highest accuracy two-way operation is required with a
continuously operating feedback loop for nulling out phase delay
variations.

Coaxial cable Coaxial cables offer a convenient means of transferring Cable cost is about $5 - $30 per meter.
time-and-frequency information over distances of less
than several hundred meters. To achieve the accuracy Insertion loss is dependent on cable length, type and the
performance given in Table C-1, careful attention must be frequency used.
paid to temperature environment, temperature stability, Solid-dielectric cable has a coefficient of delay of 250 ppm (or
and the type and length of cable. Good temperature even greater at 24°C). Air dielectric is 15 ppm, but must be dry-
stability can be achieved by burying the cable at least nitrogen pressurized with a dual-stage pressure regulator in an
1.5 m underground. environment controlled to with 1°C.

81
 Glossary and Definitions

Accuracy Closeness of the agreement between the result of a

measurement and a true value of the measurand.

BIPM International Bureau of Weights and Measures

(Bureau International des Poids et Mesures) located
in Sevres, France (near Paris). It is the responsible
bureau for the basic standards for international
commerce and generates the time scales: Interna-
tional Atomic Time (TAI) and Coordinated Universal
Time (UTC).

C/A Clear access channel available for civil usage with

GPS receivers. Currently, the time and position
information broadcast by GPS on this channel
(carrier frequency of 1,575 MHz) is degraded for
military security reasons. This degradation is called
Selective Availability (SA) and is not to exceed
340 ns (2σ). A U.S. presidential directive guarantees
the availability of this channel for civil use on a
world-wide basis.

CCDS Consultative Committee for the Definition of the

Second (Comité Consultatif pour la Définition de la
Seconde) is comprised of representative atomic-
time experts from the nations of the world — acting
in an advisory capacity to the International Commit-
tee of Weights and Measures (CIPM), which in turn
provides input to the CGPM.

CGPM General Conference of Weights and Measures

(Conférence Générale des Poids et Mesures) is
organized under the 1875 la Convention du Mètre.
Each signatory nation has membership. The CGPM
is the final voice for the International System (SI) of
units that provide the base standards (ampere,
candela, kelvin, kilogram, meter, mole, and the
second) for world commerce.

CIPM International Committee of Weights and Measures

(Comité International des Poids et Mesures) re-
ceives input from the sundry consultative commit-
tees, such as the CCDS, and submits input to the
CGPM for final approval of international standards
under agreement of the la Convention du Mètre.

Ephemeris A table giving the coordinates of a celestial body at

a number of specific times within a specific period.

Error Result of a measurement minus a true value.

82
Frequency The frequency change, typically averaged for an
Instability interval, τ, with respect to another frequency.
Generally one distinguishes between frequency drift
effects and stochastic frequency fluctuations.
Special variances have been developed for the
characterization of these fluctuations.

GLONASS The Russian Federation’s Global Orbiting Naviga-

tion Satellite System (GLONASS: GLobal Orbiting
NAvigation Satellite System). It features 24 satel-
lites in three orbital planes inclined 64.8° to the
equatorial plane. The GLONASS satellites transmit
the same code but at different frequencies. Both the
U.S. GPS and GLONASS are operated by their
respective defense departments and offer precise,
global and continuous position-fixing capabilities.

GMT Greenwich Mean Time is time as kept at the zero

meridian — often referred to as zulu time (referring
to the z time zone). It was the official name for
world time until 1972.

GNSS The Global Navigation Satellite System — a major

international coordinated initiative to provide a
seamless global navigation and positioning system
by satellite.

GPS The U.S. Department of Defense’s Global Position-

ing System. It features 24 satellites in six orbital
planes inclined 55° to the equatorial plane. The GPS
satellites transmit at the same carrier frequency
using codes that are orthogonal. The very accurate
atomic clocks on board the satellites allow world-
wide navigation, positioning and timing.

IEEE The Institute of Electrical and Electronics Engi-

neers, Inc.

IERS International Earth Rotation Service provides a

coordinated set of measurements of the earth’s
angular position with respect to the celestial sphere
and International Atomic Time. It is headquartered
at the Paris Observatory with primary input from
the United States Naval Observatory along with
numerous other observatories and radio telescopes
around the world.

INMARSAT International Maritime Satellite Organization.

ITU International Telecommunication Union located in

Geneva, Switzerland.

83
 Mean Solar Day A division of time equal to 24 hours and representing
the average length of the period during which the
earth makes one revolution on its axis with respect to
the sun.

Metrology The science of, or a system of, weights and measures.

MTSAT Japan’s Multi-functional Transport Satellite

Precision Random uncertainty of a measured value, expressed

by the standard deviation or by a multiple of the
standard deviation.

SA Selective Availability is an intentional degradation of

the GPS signal for civil users. This is intended to
protect high accuracy usage by the military. The
guaranteed accuracy of GPS positioning and timing
when receiving the C/A signal is 100 meters and 340
nanoseconds, respectively with a 95 percent confi-
dence interval.

SI International System of Units (Systemè International

d’Unites)(see CGPM).

Synchronization The times of clocks are in synchronization if their

readings are the same after accounting for reference
frame delays and relativistic effects. Synchronization
needs to be specified to within some level of uncer-
tainty.

Syntonization The rates or frequencies of clocks are in syntonization

if the rates are the same after accounting for refer-
ence frame corrections and relativistic effects.
Syntonization needs to be specified to within some
level of uncertainty.

Uncertainty Parameter associated with the result of a measure-

ment that characterizes the dispersion of the values
that could reasonably be attributed to the measurand.
Frequently, it is possible to distinguish two compo-
nents: Type A described by statistical analysis of a
series of observations, and Type B, those not de-
scribed as in Type A. These are often referred to as
the random and systematic components, but need not
be. See reference [32] for details.

UT? UT0, UT1, UT2 and UTC are a family of Universal

Time scales. The first three are derived from Earth
spin and orbit dynamics (see second paragraph in
“Historical Perspective” and reference [8] for more
details). The last one is official world time as ex-
plained throughout the text.
84
Bibliography

1. B. Guinot, 1989, Atomic time scales for pulsar studies and other
demanding applications, Astronomy & Astrophysics, 192, pp. 370-
373.

2. D.W. Allan, Millisecond Pulsar Rivals Best Atomic Clock Stability,

Proceedings of the 41st Annual Symposium on Frequency Control,
Philadelphia, PA, 2-11, 1987.

3. R.L. Walsworth, E.M. Mattison, E.R. Oteiza, R.E. Stoner, K.Y.N. Tan,
R.F.C. Vessot, and I.A. Yu, Precision Measurements with new
Hydrogen, Rubidium, and Noble Gas Masers, Proc. Fifth Sympo-
sium on Frequency Standards and Metrology, 1995, World Scien-
tific, pp 187-201, edited by James C. Bergquist.

4. J.R. Vig, Quartz Crystal Resonators and Oscillators for Frequency

Control and Timing Applications, A Tutorial, Available from John
R. Vig, Army research Laboratory, Fort Monmouth, NJ 07703-5601,
USA, or e-mail: [email protected]. November 1996.

5. D.W. Allan, J.E. Gray, and H.E. Machlan, The National Bureau of
Standards Atomic Time Scale: Generation, Stability, Accuracy
and Accessibility, NBS Monograph 140, Time and Frequency:
Theory and Fundamentals, 205-231, 1974.

6. A. Clairon, S. Ghezali, G. Santarelli, Ph. Laurent, E. Simon, S. Lea,

M. Bahoura, S. Weyers and K. Szymaniec, The LPTF Preliminary
Accuracy Evaluation of Cesium Fountain Frequency Standard,
Proceedings of Tenth European Frequency and Time Forum, (1996)
Brighton, UK, pp. 218-223.

7. Wm. Markowitz, R. G. Hall, L. Essen and J. V. L. Parry, 1958,

Frequency of Cesium in Terms of Ephemeris Time, Phys. Rev.
Let., vol. 1, pp. 105-106.5.

8. D.D. McCarthy, 1991, Astronomical Time, Proceedings of the

IEEE, Vol. 79, No. 7, pp. 915-920.

9. C. Thomas, and J. Azoubib, Metrologia, 1996, 33, pp 227-240.

10. P. Tavella, and C. Thomas, Metrologia, 1991, 28, pp 57-63.

11. C. Thomas, Stability and Accuracy of UTC and TAI, and Impact
of New Clock Technology, Proc. 50th 1996 IEEE International
Frequency Control Symposium, pp 1123-1130.

12. D. Bodson, R.T. Adair, M.D. Meister, Time and Frequency Infor-
mation in Telecommunications Systems Standardized by Federal
Standard, 1002A, Proc. IEEE 79, No. 7, July 1991, pp 1077-1079.

85
 13. D.W. Allan, D.D. Davis, M. Weiss, A. Clements, B. Guinot,
M. Granveaud, K. Dorenwendt, B. Fischer, P. Hetzel, S. Aoki,
M.-K. Fujimoto, L. Charron, and N. Ashby, Accuracy of Interna-
tional Time and Frequency Comparisons Via Global Positioning
System Satellites in Common-View, IEEE Transactions on
Instrumentation and Measurement, IM-34, No. 2, 118-125, 1985.

14. J.D. Crum, 1996 GPS Time Transfer Performance, Proc. 1996
Precise Time and Time Interval Planning and Applications Meeting.

15. D.W. Allan, and W. Dewey, Time-Domain Spectrum of GPS SA,

Proceedings of 1993 ION GPS Sixth International Technical
Meeting, Sept. 1993 held in Salt Lake City, Utah.

16. See Annual Report of the BIPM Time Section for Specific Calibra-
tion Values of TAI by the different primary frequency standards,
Vol. 9, 1996.

17. W. M. Itano, L. L. Lewis, and D. J. Wineland, Phys. Rev. A25, 1233,

(1982).

18. Eric Simon, P. Laurent, C. Mandache and A. Clairon, Experimental

measurement of the shifts of Cs 62 S1/2 hyperfine splittings due
to the D.C. electric field and black body radiation, Session 3:
Cesium Standards, Paper 1, 11th European Frequency and Time
Forum, Neuchâtel, Switzerland, 4-6 March 1997.

19. Rolf Augustin, A. Bauch and R. Schröder, Shift of the caesium

hyperfine transition frequency due to black-body radiation: first
experimental verification, Session 3: Cesium Standards, Paper 2,
11th European Frequency and Time Forum, Neuchâtel, Switzer-
land, 4-6 March 1997.

20. N. Ashby and D.W. Allan, Practical Implications of Relativity for

a Global Coordinate Time Scale, Radio Science, 14, No. 4, 649-669,
1979.

21. N. Ashby and M. Weiss, GPS Receivers and Relativity, NIST

Technical Note 1385, (to be published).

22. Robert F. C. Vessot, M. W. Levine, E. M. Mattison, E. L. Blomberg,

T. E. Hoffman, G. U. Nystrom, B. F. Farrel, R. Decher, P. B. Eby,
C. R. Baugher, J. W. Watts, D. L. Teuber and F. D. Wills, Test of
relativistic gravitation with a space-borne hydrogen maser,
Physical Review Letters 45, 2081-2084, 1980.

23. N. Ashby, Planetary Perturbation Equations Based on Relativis-

tic Keplerian Motion, in Relativity in Celestial Mechanics and
Astrometry, eds. J. Kovalevsky and V. A. Brumberg, Proc. of IAU
Symposium No. 114, Leningrad, May 28-31, 1985, (D. Reidel,
Dordrecht, 1986), pp 41-52.

86
24. G. Petit, P. Wolf, Relativistic Theory for Picosecond Time Transfer
in the Vicinity of the Earth, Astronomy and Astrophysics, 1994,
286, pp. 971-977.

25. P. Fridelance, E. Samain and C. Veillet, T2L2: Time transfer by

laser link: a new optical time transfer generation, accepted for
publication in Experimental Astronomy, January 1997.

26. Clifford C. Hodge, John A. Davis, John C. Gallop, David W. Allan,

Neil Ashby, Stefan Bedrich, Leonard S. Cutler, Jörg Hahn,
Hartmut Nau, Robert H. Kern, Wlodzimierz Lewandowski,
Lutfollah Maleki and Robert F. C. Vessot, Towards 10 millimetre
real-time position determination and 30 picosecond time-
transfer capability with the next-generation of global navigation
satellite systems (GNSS), Proc. ION GPS-96, Session D1: GNSS
and WAAS Architectures, Paper 3, pp. 1127-1142, Kansas City,
Missouri, United States of America, September 17-20, 1996.

27. See the Proceedings of Institute of Navigation, GPS International

Technical Meetingsheld annually for several papers on the status,
accuracy and plans for WAAS program. Contact Satellite Division
of ION for copies of proceedings.

28. Wlodzimierz Lewandowski,, J. Azoubib, M. Weiss, V. Zhang,

V. Hanns, A. Gevorkyan, P. Bogdanov, N. Tutolmin, J. Danaher,
G. de Jong, J. Hahn, M. Miranian, GLONASS time transfer and its
comparison with GPS, Session 8: Time and Frequency Compari-
sons, Paper 1, 11th European Frequency and Time Forum,
Neuchâtel, Switzerland, 4-6 March 1997.

29. B. Tiemeyer, A. Watt and R. Farnworth, Can GNSS be certified for

civil aviation?, Proc. ION GPS-96 Session E1: GPS, GLONASS and
GNSS Policy Issues, Paper 3, pp. 1483- 1491, Kansas City, Missouri,
United States of America, September 17-20, 1996.

30. L. Marton (ed.) (1977) Advances in Electronics and Electron

Physics. Volume 44, Chapter 2, Timekeeping and its Applications,
by Gernot M.R. Winkler, p.33-97. Academic Press.

31. Time and Frequency Measurement, edited by Christine Hackmand

and Donald B. Sullivan, American Association of Physics Teachers.

32. Guide to the Expression of Uncertainty in Measurement, 1993,

BIPM, IEC, IFCC, ISO, IUPAC, IUPAP, OIML.

33. D.B. Sullivan, D.W. Allan, D.A. Howe, and F.L. Walls, Characteriza-
tion of Clocks and Oscillators, NIST Tech Note 1337, 1990.

87
 H

For more information about Hewlett-

Packard test and measurement products,
34. IEEE Standard 1139-1988, Standard Terminology for Fundamen- applications, services and for a current
tal Frequency and Time Metrology, available from Secretary, sales office listing, visit our web site,
IEEE Standards Board, 445 Hoes Lane, P.O. Box 1331, Piscataway, http://www.hp.com/go/tmdir.
You can also contact one of the following
NJ 08855-1331, USA. centers and ask for a test and measure-
ment sales representative.
35. IEEE Standard 1193-1994, IEEE Guide for Measurement of
United States:
Environmental Sensitivities of Standard Frequency Generators, Hewlett-Packard Company
available from Secretary, IEEE Standards Board, 445 Hoes Lane, Test and Measurement Call Center
P.O. Box 1331, Piscataway, NJ 08855-1331, USA. P.O. Box 4026
Englewood, CO 80155-4026
1 800 452 4844
36. D. W. Allan, M. A. Weiss, J. L. Jespersen, A Frequency-Domain
View of Time-Domain Characterization of Clocks and Time and Canada:
Hewlett-Packard Canada Ltd.
Frequency Distribution Systems, Proceedings of the Forty-Fifth 5150 Spectrum Way
Annual Symposium on Frequency Control, pp. 667-678, 29-31 May Mississauga, Ontario
1991. L4W 5G1
(905) 206-4725

37. Woodward, Phillip, My Own Right Time, ISBN 0-19-856522-4. Europe:

Hewlett-Packard
European Marketing Centre
38. Sobel, Dava, Longitude, ISBN 0-14-025879-5. P.O. Box 999
1180 AZ Amstelveen
39. Betts, Jonathan, John Harrison, published by the U.K. National The Netherlands
(21 20) 547 9900
Maritime Museum, 1993.
Japan:
40. Frequency Stability Analysis by STABLE: Hamilton Technical Hewlett-Packard Japan Ltd.
Measurement Assistance Center
Services, 195 Woodbury Street, S. Hamilton, MA 01982, USA, 9-1, Takakura-Cho, Hachioji-Shi,
(508) 468-3703. Tokyo 192, Japan
Tel: (81-426) 56-7832
Fax: (81-426) 56-7840
41. D. Howe, An Extension of the Allan Variance with Increased
Confidence at Long Term, Proc. 1995 IEEE international Latin America:
Frequency Control Symposium, 95CH3446-2, 445 Hoes Lane, Hewlett-Packard
Latin American Region Headquarters
Piscataway, NJ, 08854, pp. 321-329. 5200 Blue Lagoon Drive
9th Floor
42. D.W. Allan, Should the Classical Variance Be Used as a Basic Miami, Florida 33126
U.S.A.
Measure in Standards Metrology?, IEEE Trans. on Instrumenta- (305) 267 4245/4220
tion and Measurement, IM-36, 646-654, 1987.
Australia/New Zealand:
Hewlett-Packard Australia Ltd.
31-41 Joseph Street
Blackburn, Victoria 3130
Australia
1 800 629 485

For more information: Asia Pacific:

Hewlett-Packard Asia Pacific Ltd.
17-21/F Shell Tower, Time Square,
Designing Your GPS Antenna System — Configuration Guide 1 Matheson Street, Causeway Bay,
Hong Kong
Tel: (852) 2599 7777
Application Note 1272: GPS and Precision Timing Applications Fax: (852) 2506 9285

Data Subject to Change

Printed in U.S.A. June 1997

Hewlett-Packard Company
88 Copyright © 1997
5965-7984E