JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 102, NO.

B3, PAGES 5005-5017, MARCH 10, 1997

Precise point positioning for the efficient and robust

analysis of GPS data from large networks
J. F. Zumberge, M. B. Heftin, D.C. Jefferson,M. M. Watkins,
and F. H. Webb
Jet Propulsion Laboratory, California Institute of Technology,Pasadena

Abstract. Networks of dozens to hundreds of permanently operating precision

Global PositioningSystem(GPS) receiversare emergingat spatialscalesthat range
from100to 10• km. To keepthe computational
burdenassociated
with the analysis
of such data economicallyfeasible, one approach is to first determine preciseGPS
satellite positionsand clock correctionsfrom a globally distributed network of GPS
receivers. Then, data from the local network are analyzed by estimating receiver-
specific parameters with receiver-specificdata; satellite parameters are held fixed
at their values determined in the global solution. This "precisepoint positioning"
allows analysis of data from hundreds to thousandsof sites every day with 40-
Mfiop computers,with results comparable in quality to the simultaneousanalysis
of all data. The referenceframesfor the global and network solutionscan be free of
distortion imposed by erroneousfiducial constraintson any sites.

Introduction The volume of GPS data is growing rapidly, and a

means to analyze this volume in a consistent, robust,
The Global PositioningSystem(GPS) has emerged and economical manner is essential. In this article we
in the 1990s as the space geodetic technique with the
first discussthe computational burden associatedwith
simultaneousvirtues of accuracyand economy[e.g., processingGPS data in the 'contextof data volume and
Yunck, 1995]. It has been applied to a variety of number of parameters estimated. We show that the
geophysicalphenomena, including the motion of tec- computationalburden associatedwith the rigorousleast
tonic plates [Larsonand Freymueller,1995; Argusand squaresanalysisof data simultaneouslyfrom R receivers
Heftin, 1995]•plate-boundarydeformation[Feiglet al., scaleswith R3. To the extentthat globalparameters,
1993], motion associatedwith earthquakes[Blewitt et that is, orbits of GPS satellites(expressed
in an Earth-
al., 1993; Bock et al., 1993], and changingEarth ori- fixed referenceframe) and satellite clocks,can be esti-
entation [Herring et al., 1991;Lindqwisteret al., 1992]mated with a subset of the R receivers, then data from
and rotation rate. More recent uses of GPS include
the others can be analyzed one at a time. Various anal-
volcanomonitoring[Webbet al., 1995], ground-based ysesof data from R = 49 receiversare used to quantify
measurements for atmospheric[Busingeret al., 1996] the approximations involved in this technique, which
and ionospheric[e.g., Wilsonet al., 1995]applications, we call "precisepoint positioning." The validity of the
as well as atmosphereand ionospheresounding[Mel- technique is also demonstrated based on data from over
bourne,1995] usinglow-Earth orbiters equippedwith 102receiversand 104stationdays.
GPS receivers.
In heavily populated regionsof significantseismicac- Computational Burden
tivity, deployment of dense networks of precisionGPS
receiversis in progress. Already hundreds of receivers The two GPS data types,carrierphase(L) and pseu-
are in operation in Japan and dozens in southern Cali- dorange (P), measurethe receiver-to-transmitterdis-
fornia. Networks suchas these allow the on-goingmea- tance with and without, respectively,an unknown bias.
surement of the surface deformation field and are ex- L is a much lessnoisy measurementthan P, which off-
pected to be valuable both in understandingthe com- sets the fact that it requires estimation of a bias term.
plex systemof faults in the region and also in hazard Both data types exist at each of two frequencies,ap-
mitigation. Additionally, space-basedarrays of GPS- proximately 1.2 and 1.6 GHz; this allows the forma-
equipped receiversare expected over the next few years tion of the ionosphere-free linear combinationfor each,
to exploit the potential applications of GPS to climate, which to first order is independent of the ionospheric
weather, and the ionosphere. electron density. We assume in what follows that we
have formed this combination for L and P.
A phase measurement at time t between receiver r
Copyright 1997 by the American GeophysicalUnion. and transmitter x is modeled as

Paper number 96JB03860.

0148-0227/97/96JB-03860509.00
Lrxt+- Prxt + b•t+ z•tm(O•t) (1)
•,•t + C,•t - c•t +

5005
5006 ZUMBERGE ET AL.' PRECISE POINT POSITIONING OF GPS DATA

where Prxt is the true range, brxt is the phase bias or ters. It is approximately a factor of 2 slower than least
ambiguity,Zrt is the zenith troposphericdelay,m(Orxt) squaresestimation using normal equations.
is the mappingfunction for elevationangleOrxtbetween The least squaresestimate of n parameters with ra
receiver and transmitter, and OJrx
t is the phase windup measurementsrequires a number of arithmetic opera-
term to account for changesin the relative orientation tions
of the receivingand transmittingantennas[Wu et al., B (xn2m, (5)
1993]. Receiverand transmitterclockcorrectionsare
where the constant of proportionality dependson the
Crt and cxt, respectively. The term lYrxt accounts for
choice of estimation algorithm but is of order unity
data noisein the measurementof phaseto make (1)
an equality. A pseudorangemeasurementis similarly
[Bierman,1977]. We neglecttermsinvolvingn3 and
modeled as nm becausethey are small comparedto n2m in this
application.Thus (3), (4), and (5) give
Prxt - Prxt-}-Zrtm(Orxt)q-Crt - Cxtq-T]rxt, (2)
- +..., (6)
where •rxt accountsfor pseudorangedata noise. Im-
where the ellipsesdenoteterms of order R 2 and lower.
plicit in Prxt are receiver coordinates and all nonclock
transmitter parameters. The valueof k is easilyshownto equal(5 q-X)2ml,
where ral is the number of measurementsper receiver
((3) with R - 1). In (6) we assumethat the constant
Undifferenced Data
of proportionalityin (5) is 1.
Considerdata from R receiversspanninga period A; Total processing times includethree passesthrough
typically we might have A = 24 hours. Let • be the the filter module. Identification of phase-biasbreaks
time betweenmeasurements
and let f•/47r be the av- is performed after the first, and the third representsan
erage probability that a satellite is in view above an iteration, required becauseof nonlinearities in our mod-
assumedelevationcutoff (• 0.25 for a 15ø cutoff). If eling of eclipsingsatellites'yaw attitude [Bar-Severet
there are X transmitters(X = 24 for the operational al., 1996].Total processing timesalsoincludeauxiliary
GPS constellation),then the numberof measurements modulesfor input/output, data cleaning,calculations
is given by of predicted measuredvalues and their derivatives with
m = RX(f•/4•r)(A/6)d, (3) respect to estimated parameters, and so on. Shown in
Figure I are actual processingtimes as a function of
where d is the number of data types. Normally, d = 2, R, as well as the theoreticalfilter timesusing(5), with
corresponding to ionosphere-free phase and pseudor- X - 24, f•/47r - 0.25, A - 30 hours, 6 - 5 min,
ange. d- 2, a- 29, b- 10, and c- 5. From Figure 1it
The number of parameters to be estimated can be is clear that somewherein the region 50 _< R _< 100,
written
simultaneousanalysisof data from R receiversstarts to
n = aR + bX + c. (4) becomecomputationallyinfeasiblewith 40-Mfiop com-
For example, station-specificparametersmight include puting devices.
three Cartesian coordinates, a zenith tropospheric de-
Partitioning
lay parameter, a clock parameter, and one phase bias
parameter for each transmitter in view by the receiver One method to reduce the burden is to divide the
over the period of interest. Thus a - 5 + X in (4). data into J groups. For simplicity we will assumethat
Transmitter-specificparameters might include epoch- thereare J = 2 groupswith ra/2 measurements in each.
state position and velocity, two solar radiation parame- Let n be the number of parameters common to both
ters, a Y bias parameter, and a clock parameter, giving groupsdivided by the total number of parametersn.
b = 10. In the case that transmitter parameters are Then (1- n)n/2 parametersapply to the first group
fixed, we have b = 0. Polar motion values and rates are only,and another(1 -n)n/2 apply to the second group
estimablewith GPS, as is length of day, giving c = 5. only. In the contextof the GPS problemeachdata par-
Over the period A it is essential to allow stochastic tition couldcorrespondto a distinct groupof receivers;
variation of clocks because most receiver and transmit- eachreceiverbelongsto one and only one group. The
ter clocks look like white noise on timescales of minutes. common parameters are Earth orientation and satellite
Temporal variation of phase biasesis also necessaryto parameters,while the group-specificparametersare the
accountfor abrupt changeswhen the receiverloseslock parametersof receiversin the group.
or slips a cycle. Stochasticvariation for zenith tropo- The ral = ra/2 measurements in group I are usedto
spheric delays and solar radiation parameters has also estimatenl = •n q-(1 - •)n/2 = (1q-•)n/2 parameters,
been found to be effective. Becausewe use square root requiring
n•ml -- n2m(1q-•)2/8 operations.
Thesolu-
informationfilter (SRIF) sequentialestimation[Bier- tion for the secondgroupwill require an equal number
man, 1977;Lichten,1990],suchparameterscontribute of operations,so that together the two solutionscost
only onceto the parameter count; the aboveexpressions n2m(1+ n)2/4 operations.
for a and b reflect this. SRIF sequential estimation is Next, the two solutions must be combined. For
used because it is numerically stable and allows con- the combinationone can think of n parametersbe-
siderableflexibility for stochasticvariation of parame- ing determinedfrom nl + n2 measurements,
costing
 ZUMBERGE ET AL.' PRECISE POINT POSITIONING OF GPS DATA 5007

10

r-

E o
o •

E• 0.1
o

0.01

10 100
number of receivers

Figure 1. Computational burden B as a function of the number of receiversR. The open circles
are actual total processingtimes on a 40-Mfiop computer and include three passesthrough the
parameter estimation module, as well as auxiliary programs. The dotted line is the theoretical
time requiredfor three passesthrough the parameter estimationmodule only, assumingB =
2.68n2m;the 2.68is chosen
sothat Equations
(3), (4), and(5) givetheobserved
filterexecution
time for the R = 48 and 49 cases.The units of B are the number of days required by a 40-Mfiop
computer to perform the analysis.

n2(nl q-n2) --n3(1 q-n) operations.The total cost)/, Double Differencing

relativeto the n2m operationsrequiredfor the analysis For receiversq and r in view of satellites w and x at
of all data simultaneously,is thus
time t, (1) can be usedto showthat the quantity

)/- (1+ n) dLqrwxt-- Lqwt- Lrwt - Lqxtq-Lrxt (8)

is independentof all clock parameters. In practice, syn-
This can be generalizedto chronization of all measurements to GPS time is made
prior to the formation of the double difference. A sim-
x-[lq-n(J-1)]
[1q-n(J-1)
j2 n
q--- m
(7) ilar double difference can be formed for the pseudor-
ange. Least squaresestimation of nonclockparameters
can be made based on double-difference measurements.
when the m measurements are divided equally into J
groups. Note that (7) approachesunity as n -• 1, as- It is straightforwardto form a measurementcovariance
matrix that accounts for known correlations in double-
sumingn/m is small. Thus, if all parametersare com-
differenced data.
mon, partitioning is ineffective.
With X - 24, f•/4•r = 0.25,A - 24 hours,(5= 5 min, The technique has the advantage of a reduction in
d - 2, a - 29, b - 10, c - 5, and R - 40 receivers, we both the measurement and parameter counts. One
need not.include the entire set of double-difference mea-
have n - 1405 parameters of which
surements because it contains redundant information.
nn- bX + c- 245 For R receivers in view of X transmitters at time t
there are (R - 1)(X - 1) independentdouble differ-
are common(n • 0.174). There are m = 138240mea- ences that form a complete measurement set. One
surements,and the ratio of parametersto measurements suchset is dLqrwxtfor q - w - 1, 2 _<r _<R, and
is n/m • 0.01. Equation(7) then showsa 64% reduc- 2 < x < X. Thus the measurement count m is reduced
tion in burden for J = 2 and an 88% reduction for by RX - (R- 1)(X - 1) - R + X - 1 comparedto the
J = 40. Clearly, partitioning is one effectivemeansof undifferencedcase. An identical reduction in the pa-
reducing CPU when n is small and has been demon- rametercountn occursbecause(R- 1)(X- 1) is also
strated by some analysis centers in the International the number of phase bias parameters to be estimated,
GPS Servicefor Geodynamics(IGS). comparedwith RX in the undifferencedcase. Finally,
When many parameters are treated as stochastic,rig- receiverand transmitter clocksare no longer parame-
orouspartitioning is difficult. We feel that the benefits ters, so that the parameter count is further reducedby
of using stochasticsoutweigh this disadvantage. R+X.
5008 ZUMBERGE ET AL.: PRECISE POINT POSITIONING OF GPS DATA

For X = 24, then, the value of n2m is reducedby orbitsor a few tensof metersfor GPS clocks[Zumberge
about 34% for R = 10 and by 19% for R = 100. From and Bertiger•1996]. Our useof precisepoint position-
the standpoint of computational burden the double- ing differsfrom the classicaluseof the term in two ways.
difference approach has approximately this advantage First, both carrier phase and pseudorangeare used as
over the undifferencedapproach when data from R re- data. Second,the quality of transmitter parameters is
ceivers are simultaneously analyzed. a few centimetersor better. When applied to single-
To simplify the calculation, we have ignored the ex- receiver data, the resulting accuraciesare comparable
tra computational burden needed to form the double to what is obtained when data from all receivers are
differences and a measurement covariance matrix that simultaneouslyreduced and come at far lower compu-
accounts for known correlations in data noise. We have tational cost.
also ignored issuesof common visibility by assuming
that all R receivers view all X transmitters at all times.
Method
AlthoughBlewitt [1993]givesan exampleof a casein
which it is not possible to form a set of double dif- Supposethe analysis of data from R receiversis di-
ferences which contains all of the information content vided into two parts. First, a subset of S receivers is
available for estimation of nonclockparameters, perfor- usedto estimate satellite parameters,Earth orientation,
manceof IGS analysiscentersthat usedoubledifferenc- and S setsof receiverparameters. Then, data from each
ing indicates that this is not a practical limitation. of the remaining R- S receiversare analyzed,one re-
The double-differencetechnique views receiver and ceiver at a time. The computational burden B can be
transmitter clocks as nuisance parameters that can be written as the sumof two terms,onefor the globalsolu-
eliminated up front by the appropriate linear combi- tion (Bg) andonefor the point-positionedsites.For the
nation of data [Wu, 1984],exactly analogousto elim- globalsolutionwe haveng- aS q-bX q-c parametersto
inating ionosphericeffects by forming the ionosphere- estimatefrom mg= SX(f•/4•r)(A/6)d measurements.
free combinations described earlier. In the method de- Thecorresponding
burdenis Ba = kSa +..-, just asin
scribed here, however, transmitter clock parameters, to- (6) with R replacedby S.
gether with GPS orbits, are the key for the efficient The computational burden to determine nl ----5 q- X
analysisof large networksof GPS receivers. (From receiver-specific
parametersgivenrrt• ((3) with R = 1)
the point of view of information content as opposedto measurements is B• - n•m•; thisappliesto R-S sites.
computationalburden,consultKuang et al. [1997]for The total burdenis thusB = Ba + (R- S)B1 or
comparisonsof the differenced and undifferencedap-
proaches.) B = (5 q-X)2ml(S3 q-R- S) q- ..-. (9)
The Sa termtotallydominates,
sothat the receiverpa-
Precise Point Positioning rametersfrom the additional R- S sitesare essentially
Supposedata from a globally distributed network of free. Note that the computational burden includes the
R receivers are used to estimate receiver and transmit- parameter-estimation processesonly. The total point-
ter parameters. Denote the estimatesof transmitter pa- position computational burden is of the order of 1 min
rameters by P. Consider an additional receiver and the per site on a 40-Mfiop computer.
analysisof data from Rq- 1 receiversto yield transmitter Aside from computationalefficiencyan additional ad-
parameters Q. In the approximation that P = Q, that vantageof the one-receiver-at-a-timeprocessingis that
is, data from the additional receiver has negligible ef- it allows easy diagnosisof receiver-specificproblems.
fect on the values of estimated transmitter parameters, For geodeticquality receiversthe rms value of the post-
a more efficient way to estimate the parameters spe- fit residuals, that is, the rms difference between the
cific to the additional receiveris to analyze data from modeledvaluesin (1) or (2) and the measuredvalues,is
it in seclusionand fix the transmitter parameters in 5 to 10 mm for phaseand 50 to 100 cm for pseudorange.
the analysis to be P. When GPS orbits in an Earth- If measurementsfrom a particular receiverare signifi-
fixed frame and GPS clock corrections are not estimated cantly noisier, it may be an indication of a hardware
but, instead, fixed at some predetermined values and fault, a poor multipath environment,an interfering RF
the model has no other spatially correlated parameters source, or other unusual circumstance.
(e.g., troposphericdelay), then there is no needto ana- Reference Frame
lyze data from all receiverssimultaneously.
The term "point positioning" in GPS analysishas re- One way to impose a referenceframe in the analy-
ferred historically to the estimation of receiver-specific sis of GPS data is to assume that coordinates of certain
parameters by the analysisof receiverpseudorangedata, sitesare sufficientlywell known that they can be fixed or
fixing transmitter parametersto valuesbroadcastin the tightly (< 1 cm) constrainedin the analysisof data from
navigation message.The quality of results is limited in a globalnetwork[Blewitt,1993,andreferences
therein].
this casebecause(1) pseudorange
is inherentlynoisy,no For example, the analysis centers of the IGS fix the
better than about 40 cm, based on our experience with locationsof 13 globally distributed sites in their deter-
postfit residualsof geodetic-quality GPS receiversand minationsof preciseGPS ephemerides
[Kouba,1995].
(2) the transmitter parametersbroadcastin the navi- A disadvantageto the fiducial approach is that er-
gation messageare accurate to a few meters for GPS rors in assumed locations of the fiducial sites distort the
 ZUMBERGE ET AL.: PRECISE POINT POSITIONING OF GPS DATA 5009

GPS orbits in a way that is not easily undone. (This data will be required if the mistakes are significant
disadvantagedoes not apply to site coordinates pro- enough.
duced by IGS analysis centers;such coordinatesdo not To avoid the consequences of fiducial errors, a pri-
reflectstrongfiducialconstraints.)In our experienceat ori uncertaintiesof coordinatesare set fairly large at
JPL's (Jet PropulsionLaboratory)IGS analysiscenter 1 km (10 m for the site at AlgonquinPark, Canada).
[Zumberge et al., 1995],such"errors"can arisein two Although the resulting "free-network"solution will be
ways. First, the "standard" referenceframe, the Inter- in a poorly defined referenceframe, all estimated pa-
nationalTerrestrialReferenceFrame(ITRF) [Boucher rameters will be immune to fiducial errors, and ex-
et al., 1994]in the caseof the IGS, itself evolveswith pensivereprocessingof global network data is avoided.
time, so that a discontinuity in assumedlocations of Furthermore, a particular reference frame can be re-
fiducial sites will occur, typically at midnight between alized by transforming estimated receiver parameters
two calendar years. These discontinuities are not re- into that frame. A differentframe (an updatedITRF,
lated to actual motion of sitesbut indicate only a change for example)can be realizedsimplyby applyinga new
in assumed locations. transformation. Estimated parametersof sites subject
Fiducial orbits on December 31, 1994, that are based to precise point positioning with free-network trans-
on ITRF 92, for example, will exhibit a discontinuity mitter parameters will be in the same reference frame
comparedwith orbits on January 1, 1995, that are based as the original free-networksolution. A single seven-
on ITRF 93. These discontinuitiesare small, typically a parameter transformation can be applied to all sites for
few centimetersand can be reducedby applying a seven- alignment with the ITRF. The appropriate transforma-
parameter Helmert transformation but only to the ex- tion shouldbe availableto usersfrom the analysiscenter
tent that the two referenceframes are related by such that produced the free-network orbits and clocks. The
a transformation. Without reprocessingthe discontinu- free-networkapproachhas been demonstratedfor Earth
ities cannot be entirely eliminated, and a time seriesof orientation[e.g.,Herringet al., 1991],baselines [He•in
resultsderived from fixed-orbit analyseswill inherit the et al., 1992],coordinates[Blewittet al., 1992],and ve-
discontinuities. The effect can be several millimeters in locities[Feiglet al., 1993;ArgusandHeftin,1995].
positions.
Single-Day Tests
Second,mistakesin assumedantenna heightsof fidu-
cial sites, or other similar blunders, cannot be com- Transmitter parameters. Shownin Figure 2 are
pletely eliminated. Expensive reprocessingof global sites from which various networks are constructed to

..................
.?;::
....
::•:::;•::;
...... ...:!ii•.•:=•
•":•
.............
•. •i•:
.....!:?•:.i•
.........
'................
i•:.,.%::*•% ...... '....................
.............
........ :.•. .......
.....;:i•........... :;::;
.........
................ •i•
.:.•o ........
• ':':::::.•.:•:-.'i•-..43
•:..:•.... ............... :::' .........
"::::i:11'":•.:
ii:'•''..:'•'"•':::::I:
..............•:•::.•::ii'::•i• ........ .....
:.... 43 ............
'......
........ .?::i.0 ......... ?_ o o - ..•::'
............... •:..o
•i:•:.
?!....?•n:;•i:;!:.
..........
:::•
.....ß
........... ........... o• • ...... 0 :::: '............ •!.. ......
•11•:'ii•:•.:
....................
"'i
ß:" .::::"' '.-•..:.::.
o '":: •i .... : i!: ......
:'" ':' ' "::::

:•::
:::'.:iii:'•
../ .•ii!i:= ß .....................
::::::'"'"•i .. '•..•.. '...............
i! .....
i;•!:=

"::::•::'.::.
'::::"::•:.. '"•":::L ? •'"':::::'
' .:::?:
....
...:.-'::'""::'" • .:::::?

....
:.....:... "::....!;:. ...:,..:.-':"
(3 . • .......................
............................................................
...:.•f:?
.... .....................................
......................................................................
ß•:•:::::'
........................................................................................................
•:•
.......................
=(......................................................................................
?:::i:

Figure 2. Sites used for tests of point positioning. The nine large solid circles correspondto
receiversat, from left to right, Usuda, Japan; Tidbinbilla, Australia; Kokee Park, Hawaii; Casey,
Antarctica; Fairbanks,Alaska;Pietown, New Mexico; AlgonquinPark, Canada; Fortaleza, Brazil;
and Metsahovi, Finland. The 12 smaller solid circles correspondto receiversat Lhasa, China;
Perth, Australia; Guam, Pacific Ocean; Irkutsk, Russia; Chatham Island, New Zealand; Thule,
Greenland; Arequipa, Peru; St. Croix, U.S. Virgin Islands; La Plata, Argentina; Maspalomas,
Canary Islands; Noto, Italy; and Kerguelen Islands, Indian Ocean. The dotted open circles
correspondto three sitesin Australia and New Zealand, eight sites in North America, and seven
sites in Eurasia. Finally, the small open circlescorrespondto additional sites in North America
(eight) and Europe (two). Althoughthere are permanentlyoperatingreceiversin the southern
Pacific Ocean and Africa, their data on November I were either not available or of poor quality.
5010 ZUMBERGE ET AL.' PRECISE POINT POSITIONING OF GPS DATA

Table 1. Average Formal Errors of Transmitter Parameters Estimated From

Different Networks of Figure 2

S Orbit 3-D, cm Var, cm Coverage,% Clock, cm Var, cm

9 19.4 5.9 90.5 9.7 3.5

10 16.7 4.4 90.9 8.9 3.1

21 6.1 1.0 98.9 4.7 0.4

22 5.9 1.0 98.9 4.6 0.4

38 4.6 0.7 98.9 3.8 0.3

39 4.6 0.7 98.9 3.8 0.3

48 4.4 0.7 98.9 3.7 0.3

49 4.4 0.7 98.9 3.7 0.3

The nine-station network is shown as the nine large solid circles in Figure 2. The 21-
station network includes, in addition, the 12 smaller solid circles. The 39-station network
includes the 21 plus an additional 18 indicated by the dotted open circles. Finally, the
49-station network includes all sites in Figure 2. Networks with S - 10, 22, 38, 48 consist
of the network with 5' -F 1 receivers, with the addition or subtraction of the receiver in
Belgium (a dotted open circle in Figure 2). Thirty hoursof data centeredon GPS noon,
November 1, 1995, were analyzed. The coverage column indicates that for the 9- and 10-
station networksthe geometrywas insufficientto determineclock parametersabout 9%
of the time. The var columns indicate the 1-standard-deviation variability in the formal
errors among satellites and over time for the 24-hour period of November 1, 1995. To
remove contributions from reference-frame uncertainties, the locations of the nine large
solid circles in Figure 2 are assumed known for purposesof Table 1 only.

determine transmitter parameters, in the Earth-fixed pected

from1/v/•, because
(1) thecontribution
dueto
frame, based on November 1, 1995, data. The solid reference-frame uncertainty has been removed, as dis-
circles compose a network of 21 stations with good cussedabove,and (2) the additionalsitesare clustered
global distribution. This network consistsof a sparse in already populated regions.
nine-stationnetwork (large solid circles)with reason- To what extent do the data from additional receivers
able globalcoverageexceptnear 90øE longitudeand an affect the estimated transmitter parameters? Shown
additional12 sites (small solidcircles)chosento give in Figure 3 is a comparison of transmitter parameters
improved global distribution.
Given the distribution of receiversshownin Figure 2,
the uniformity of the global distributioncannot be im- 35 1

provedmuchbeyondthis 21-stationnetwork. Shownin 30

3d rms
Table 1 are the formal errors of estimated transmitter clock
parametersas a function of the network used as data, 25
assumingthat the locationsof the nine large solid cir-
clesin Figure 2 are preciselyknown. This assumptionis 20
cm
similar to one made by IGS analysiscentersin produc-
15
ing fiducial orbits. In the presentarticle it is invoked
only so that the formal errors of the transmitter param-
10
eters will not reflect reference-frame uncertainties. Free-
network estimated transmitter parameters are used in
all analyses.
Table 1 showsthat $ - 9 provides estimates of or- 0 I I I ...... ::•
bits and clocks that are, on the average, determined 0 lO 20 30 40 50
with formal errors that are factors of 4.4 and 2.6, re- number of receivers in global solution
spectively,larger than those determinedwith $ _• 38.
Furthermore,the $ - 9 and $ - 10 networkshavelarge Figure 3. Transmitter parameters determined with
fewer than 49 receivers, compared with those deter-
variations in the formal errors, as well as geometry that mined from the 49-station solution. The solid circles
is insu•cient to determine clock parameters for about give the three-dimensionalrms differencein satellite po-
10% of the time and/or satellites. sitions, while the open circles give the rms difference
On the other hand, Table 1 also indicates that for in clock solutions. The gradual improvementbeyond
S _• 21 the reduction in formal error that accompanies 20 receiversin the global solution is costly in terms of
the increase in $ is somewhat less than would be ex- computational burden.
 ZUMBERGE ET AL' PRECISE POINT POSITIONING OF GPS DATA 5011

Table 2. Formal Errors in Coordinates of Point-Positioned Sites

North, mm East, mm Vertical, mm 3-D, mm

Average 0.8 2.0 4.8 5.2

Variation 0.2 0.5 0.7 0.8

These reflect only assumeddata noise (1 m for P and 1 cm for L), the sampling
interval, and a priori uncertainties in estimated receiver parameters. The data noise
contribution dominates. By definition of point positioning, orbit and clock errors do not
contribute to the formal errors. Thus formal errors are essentiallyindependent of the
network from which global parameters were derived. (However, becauseof incomplete
coverageas indicated in Table I, formal errors of point-positioned sites based on orbits
and clocksderivedfrom the 9- and 10-stationnetworksare slightly,about 8%, larger than
for the other networks.) The average row indicates the average value over all sites in
Figure 2, and the variation row indicates i-standard-deviation variability.

estimated from different subsets of the sites shown in solution, the rms differencesin receiver coordinates due
Figure 2. To avoid contributions to the uncertainties to this changein A are 0.7 mm in the north, 2.6 mm in
that arise from imperfect knowledgeof receiver coordi- the east, and 5.4 mm in the vertical.
nates only weak, 10-m to 1-km, a priori constraints are Shown in Figure 4 are the differencesin estimated
imposed on these, as describedearlier. Becausethe re- receiver parameters as a function of $, with $ - 49
sulting transmitter parameters are in a poorly defined as the basis for comparison. To account for the refer-
Earth-fixed reference frame, the rms differences shown ence frame uncertainty of the free-network transmitter
are after a seven-parameter(three translations,three parameters, all solutions are aligned to the 39-station
rotations, and a scalefactor) transformationto align, results, using a seven-parameter Helmert transforma-
orient, and scale the reference Ëames to each other. tion to minimize the rms coordinate difference of the
From Figure 3 it can be seen that the 9- and 10- nine-station network. Clearly, the $ - 9 and $ - 10
station networks provide transmitter parameters that networks give rise to several tens of millimeters differ-
differ by tens of centimetersfrom those estimated with encesin estimated receiverparameters. Beginningwith
$ _>21. On the other hand, the differencesamong the $ - 21, however,receiverparameters are lessand less
$ >
__
21 networks is of the same order as the formal affected, and differencesare comparableto the formal
errors. Such differencesmay not be worth the signifi- errors.

cant increase in computational burden from $- 21 to

S-49.
Receiver parameters. Another measure of the 50 I I I I

technique's validity is the degree to which estimated

ß vertical
receiver parameters depend on the division of the to- 40
o horizontal -
tal R- 49 sites into $ used in the global solution and
the R- $ sites that are subsequentlypoint positioned.
For reference we show in Table 2 the formal errors of 30

point-positioned sites. By definition, uncertainties in mm

estimated transmitter parameters do not contribute to 20

formal errors in point-positioned sites and so are essen-
tially independentof $. (The exceptionhas to do with
10
$ - 9 and 10, for which insufficient coveragemeans
that transmitter clocks were not estimable about 10%
of the time.) The formal errorsare about 1 mm in the 0 I i

north component, 2 mm in the east, and 5 mm in the 0 lO 2O 3O 4O 5O

vertical.
numberof receiversin globalsolution
To improve estimatesof transmitter parameters near
the midnight boundaries, the global solution uses 30 Figure 4. Point positioningrms differencesover all
hours of data centered on GPS noon of November 1, sitesin Figure 2, showingthe effectof different divisions
1995. Receiver parametersdetermined in the global so- of R - 49 sitesinto $ usedto determinetransmitterpa-
rameters. All solutionsare alignedto the 39-stationre-
lution correspondto the same 30 hours, whereaspoint-
sults, using a seven-parameterHelmert transformation
positioned solutions correspondto only 24 hours of re- to minimize the rms coordinate difference of the nine-
ceiver data. To force all receiver parameters to corre- station network. Above 20 receiversin the global solu-
spond to the same interval, point positioningis applied tion, receiverparametersdo not dependvery strongly
to 24 hours of data from each of the 49 sites, and all on the division. The horizontal points refer to the sum
receiver parameters are reestimated. For the $ - 49 in quadratureof the north and east components.
5012 ZUMBERGE ET AL.' PRECISE POINT POSITIONING OF GPS DATA

We havealsocomputedthe rmscoordinatedifferences areas and Africa, with no receivers. That is, in Fig-
for the 11 stationsthat participatedin the $ = 49 global ure 2 the distancesfrom open circles, both dotted and
solution but not that for $ = 38. That is, if one uses small, to nearby solid circlesare small comparedto the
the $ = 38 network to determine transmitter parame- distancesfrom isolated regionsto solid circles.
ters, then usesthoseand precisepoint positioningfor Figures 3 and 4 implicitly assumethat the $ - 49
11 additionalstations,how do the resultscomparewith solution is the "truth" case to which $ < 49 results
the inclusionof the 11 stationsin the globalsolution? shouldbe compared. From Figure 3 one might conclude
The answer is 2.5-mm rms in the horizontal and 9.1- that there is a slight but steady improvementin the
mm rms in the vertical. (If the site with the largest quality of transmitter parameters for S > 21. To a
vertical difference is excluded, the rms vertical is re- lesserdegreea similar conclusionmight be drawn from
ducedto 6.8 mm.) For the verticalthis valueis slightly Figure 4 in the caseof receiverparameters.For several
higherthan that shownin Figure 4 for $ = 38. On reasons,includingmismodeling,it is not necessarilythe
the scale of the formal coordinate errors of the 11 sites, casethat $ - 49 representstruth more than, say, $ -
2.6 mm horizontal and 4.5 mm vertical, the differences 38.

are marginally significant. Even with no mismodeling, Table 1 indicates only

(Sincethe $ = 48 and $ = 49 transmitterparameters marginal improvementin the ability to determinetrans-
differby only about 1 cm or less(Figure3), the 5-mm mitter parameters with increasing $ beyond $ - 21,
rms differencefor the vertical componentfor the $ = 48 given the distribution of sites as shown in Figure 2.
solidcirclein Figure4 may seemsomewhathigh. How- In fact, for the site distributions and values of $ used
ever, this is most likely explained by the reestimation here, formal errors of transmitter parameters are ap-
of receiver parameters once transmitter parameters are proximatelylinearlyrelatedto •, with drr•/d• • 9.4 cm
determined,sothat all receiversolutionscorrespondto per 103km anddrr2/d•• 3.8 cm per 103km, where
24 hoursinsteadof 30 hours,as describedabove.) rrl denotes the three-dimensional rms orbit formal er-
Geographical distribution and truth case. To ror and rr2 denotes the clock formal error. Thus an
evaluate the uniformity of $ sites distributed on a increase in $ which is not accompanied by a decrease
sphere,we first definethe function in • is of little value in improving the overall, global
quality of transmitter parameters, even in the absence
R(O,ok)-- min[r•, r2,''- rs], of mismodeling.
Furthermore, known aspects of mismodeling include
where0 is colatitude,½ is longitude,and r• is the great
a number of simplifying assumptionsrelated to tropo-
circledistancebetween(0, ½) and siten. ThusR(O,ok)
sphericdelay,oceanloading,satellite forcemodels,and
is the distancefrom (0, ½) to the nearestof the S sites.
The rms isolation is then
variations in receiver and transmitter phase centers, to
name a few. Becauseof the geographicaldistribution
of receivers,the $ - 49 solution will suppressmanifes-

•_•fo
2'•
do)
f•dO
sin
0R2(0,
<b)(10)
47r

Shownin Figure5 is a plot of ( versus$, usingSimp-

tations of mismodeling, or systematic errors, more in
Europe and North America than in other regions. In
the presenceof systematic errors the least squaresso-
son's rule in two dimensions to numerically evaluate lution for $ - 49 may not be nearer to truth than that
(10). Clearly,( decreases negligiblyaboveS .• 21, for $- 21 or $- 38. In addition to mismodeling,un-
which is simply a consequenceof large regions,oceanic even data quality from common or additional stations
could cause the $ - 49 solution to be better or poorer,
respectively,than that for $- 38.
4OOO
Weekly reportsby the IGS analysiscoordinator(see
http://igscb.jpl.nasa.gov/IGSREPORT.html) evaluate
the quality of estimated transmitter parameters from
3000
all IGS analysis centers. The number of receiversused
varieswith center and rangesfrom fewer than 30 to over
100. There is almost no correlation, however, between
the quality of transmitter parameters and the number of
2OOO
sites. Our interpretation is that beyond $ • 30 with the
global distribution as in Figure 2, the value of data from
additional receiversin determining transmitter param-
lOOO
eters is marginal at best and may even be outweighed
by systematicerrors. As new receiversare installedin
i i i i currently isolated regionsthereby moving the • curve
downward, we of course expect that data from these
o lO 20 30 40 50
new receiverswill be valuable in determining transmit-
number of receivers
ter parameters. As of November1996, the globaldistri-
Figure 5. The rmsisolation• (Equation10)asa func- bution is such that analysisof data from approximately
tion of the number of receivers. 35-40 well-distributed sites is appropriate.
 ZUMBERGE ET AL' PRECISE POINT POSITIONING OF GPS DATA 5013

Multi-Day Tests daily deviationsare alsoshown. From Figure 6 it is clear

that there is little or no discernible difference between
The operationalGPS analysisat JPL [Zumbergeet the two classes.
al., 1995]has beenin placesince1992. Shownin Fig- Point positiontime seriesfrom 151 sitescan be viewed
ure 6 is the detrended time series of coordinates for at http://sideshow.jpl.nasa.gov/mbh/series.html. His-
the receiver at JPL spanning 2.5 years following the togramsof daily repeatabilitiesbasedon theseare shown
Northridge earthquakeof January 17, 1994. The 109 in the left of Figure 7. Repeatability is about the esti-
circlescorrespondto days when data from this receiver mated linear trend. Sites with known and uncorrected
were included in the global solution, while the 426 changesin antennaheight are not included,leaving139
shaded points correspondto coordinatesdetermined sites. The number of days for a given site rangesfrom
from precisepoint positioning.The distributionsof the a few to a few hundred; the median value is 106 days.

o-point = 4.0 mm
o-global = 4.4 mm

latitude
(cm)
-2

-4

I I
I I I I
o-point = 6.7 mm
o-global = 6.0 mm

longitude0
(cm)
-2

-4

I I I
o-point = 11.0 mm
o
o-global = 9.9 mm

height 0
(cm)
-2

-4

I I I I I
lOO
1994 1995 1996 0 20 40 60 80
days
Figure 6. Time seriesof geocentric coordinates(detrended)for the receiverat the Jet Propulsion
Laboratory(JPL), beginninglate January1994after the Northridgeearthquake.The opencircles
indicate 109 days when JPL data were includedin the daily global solution. The shadedpoints
indicate 426 days when JPL data were analyzed with precisepoint positioning. Shown also
are the distributionsof deviationsfor the global (dotted line) and point-positioned (solidline)
solutions;the histograms havea 2-mmbin width. The differences in the meansand widths(as)
of the distributions are both small and statistically insignificant.
5014 ZUMBERGE ET AL- PRECISE POINT POSITIONING OF GPS DATA

_ ,.; :: ,
latitude _
,,

,
,

lO __ ,

sites : ! :.• ::
__
i .....i -

_ --i -- '-'1 I -
_

- i-• longitude _
, :....
:

lO
sites __

_ ;..' ', __ ! t ....... : -

, -

-!"i ': F 'lq • 'F-- : '• .... -'! • •

-
_
vertical -
-

..

lO
sites - : -
_.7. . :..!i :- i .....

0
-
5
, ,

li ........
10
1
15
- 'Fi
mm
F---
0 5
i ,
i
10
".... ....
15
-
20

Figure 7. Distributionsof daily repeatabilities(dottedlines)for 139 sites(left) that havebeen

analyzedwith precisepoint positioningand 59 sites(right) that participatedin manydaily global
solutions.The solidlinesindicatesitesin the southernhemisphere(23 on the left and 15 on the
right). The medianvaluesare 3.9 mm in latitude (north), 6.3 mm in longitude(east), and 11.1
mm in the vertical for the point-positionedsites and 4.4 mm in latitude, 5.6 mm in longitude,
and 10.3 mm in vertical for global sites.

Except for a few outliers, someof which are due to re- It is well known that detailsof an estimationstrategy
ceiverswith a history of poor performance,the repeata- (the choiceof relativeweightbetweenpseudorange and
bilities do not vary too much from site to site and, in phase measurements, the a priori uncertainties in pa-
particular, do not depend strongly on northern versus rameters, and the model of tropospheretemporal vari-
southern hemisphere. The median values are 3.9 mm ability) can changethe estimatedvaluesof parameters
for the north component, 6.3 mm for the east, and 11.1 with little changein the quality of the fit, reflecting
mm for the vertical. some degree of degeneracyamong parameters. Never-
For comparison,the correspondingvaluesfor 59 sites theless, the concept of satellite clock is no less valid in
included in global analysesare 4.4 mm, 5.6 mm, and principle than that of satellite orbit. For GOA-II the
10.3 mm; their distributions are indicated on the right formal errors of orbits and clocks from Table I indi-
of Figure 7. The differencesare insignificant.What was cate that to the few-centimeter level or better there is
demonstratedin detail for the site at JPL (Figure 6) is strength in the data to determine them separately from
thus true in general. other parameters.
Comparisonsamong resultsfrom sevendifferent IGS
Mismodeling analysiscentersusingsix distinct softwaresystemshave
There are severalhigh-precisionsoftwaresystemscur- shownthat estimated parametersfrom different analy-
rently in useto analyze GPS data, of which Gipsy-Oasis sesagreeto (1) about 20 cm, three-dimensional (3-D)
II (GOA-II) usedin this work is one. Can satellitepa- rms, for daily orbits; (2) a few millimetersfor horizon-
rameters determined with one system be used in tal stationcoordinates,averagedover7 days;(3) about
other? To the extent that there is consistencyamong i cm for vertical station coordinates,averagedover 7
software systemsin how GPS observablesare modeled days;and (4) about 5 mm for daily estimatesof total
as a function of these parameters, the answer is yes. zenith troposphericdelay. Theselevelsof agreementin-
 ZUMBERGE ET AL.: PRECISE POINT POSITIONING OF GPS DATA 5015

dicate that the modeling among the different systemsis to the single-sitecoordinaterepeatabilities,depending
very similar. on component,for baselinelengthsup to about 200 km.
To further quantify the level of mismodeling,or dif- That this is not reflected in the formal errors for base-
ferent modeling, among the different systems,we have lines is the result of a compromisebetween mathemat-
applied precise point positioning to 94 sites, globally ical correctnessand computationalfeasibility.
distributed, using data from March 4, 1996. GOA-II (By examining the covariancematrix of estimated
was used for all analyses, but there were two differ- site coordinatesamongsitesdeterminedin our regular
ent sets of satellite parametersused: (1) JPL clocks global solutions,we may be able to determine approxi-
determinedevery 5 mid and JPL orbits and (2) Geo- mately correctsite-to-sitecorrelationsfor sitessubject
ForschungsZentrum (GFZ)[Gendt et al., 1995]clocks to precisepointpositioning.Forexample, pij = e-xd'J
determined every 30 mid and GFZ orbits. Parameters might modelthe correlationbetweensitesi and j sepa-
from set 2 were retrieved from the NASA Crustal Dy- rated by distancedij, with/k an empiricalparameter.)
namicsData InformationCenter(CDDIS)[Noll, 1995]. A consequenceof point 2 is that formal errors of
The 3-D rms variation in satellite positions over the point-positioned results from isolated sites will be no
day (March 4, 1996) betweenJPL and GFZ is 33 cm. different from those in dense areas. We use "isolated"
(For the presentpurpose,fiducial orbits are used. The here in the sensethat data from no nearby receivers
improvementin orbit agreementsis negligibleif a seven- were used in the global solution that produced the
parameter transformation is allowed, indicating good transmitter parameters. Transmitter parameters are
referenceframe agreementbetweenJPL and GFZ.) The probably not as well determined in such areas. This
rms agreementbetween JPL clocksand GFZ clocksis 23 effect,alreadysmallas evidencedby the lack of strong
cm. Because precise GFZ clocks are available from the northern/southern hemisphere dependence in Figure7
CDDIS only at 30-mid intervals, the data rate available (left), will becomenegligibleas the globalnetworkcon-
for analysis is 6 times less frequent than for the JPL tinues its expansion.
solutions(seediscussion on selectiveavailabilityin the A secondlimitation has to do with the availability
sectionon limitations). of precisetransmitter clockestimates. The technique
The rms differences in estimated site coordinates de- will yield resultsof the quality shownhere only when
termined in set 2 relative to set I are reasonablygood: theseare accurateto the few-centimeter(100 ps) level
9 mm in the north, 16 mm in the east, and 28 mm in at each data epoch. Becauseof selectiveavailability,
the vertical. (We have had lesssuccess in mixing or- GPS clocksvary by the orderof 103 cm (101 to l02
bits from one analysis center with clocksfrom another, ns)overtimescales
of minutes.Thuspreciseknowledge
possibly becauseof correlations in orbit and clock er- of their valuesevery, say, 5 min is of no use in pre-
rors.) The dominant contributor to these valuesmay dicting their valuesmore frequently, at least not to the
be the decreasedamount of data analyzed using GFZ few-centimeter level. As mentionedearlier,navigation-
transmitter parameters. To the question posed at the messageclockcorrections,thoughavailablein real time,
beginning of this section, we can thus say, yes, at the are essentiallylinear approximationsover • 102 rain
centimeter level. To establish this at the few-millimeter andareaccurateonlyto the 103-cmlevel.Somepartic-
level awaits further work. ipantsin the IGS [Kouba,1995]nowroutinelyproduce
GPS clockestimateswith subnanosecond precisionat
Limitations least as frequently as every 15 min and every 5 min
from somecenters. However,until preciseGPS clocks
One limitation to the precisepoint positioningtech- are producedmore frequentlythan onceevery minute
nique is that the covariancematrix of estimated receiver and preferablyonceevery10 s, data to be analyzedwith
parameters will not necessarilybe indicative of the ac- precise point positioning will have to be decimated to
tual quality of resultsbecause(1) it will showno cor- coincidewith epochsfor which precise GPS clock solu-
relation betweensitesand (2) its nonzeroelementswill tions exist.
be independentof receiverlocation. Consequently,the To achievesubcentimeterresults requires dual fre-
covariancematrix may be inappropriate for someappli- quency, geodetic-quality receiversso that ionospheric
cations. variationscan be differencedaway and data cleaning
Because of point 1, if the location of one site from can be done on a single-receiverbasis. Whenever the
point positioning is estimated as • with covarianceV• resultsfrom a givenreceiverare limited by the accuracy
and the location of another site is estimated as •b with of orbits and clocksin the broadcastmessage, however,
covariance Vb, then the baseline estimate between these useof preciseclocksand orbitswill providean improve-
sites is • - xq with covariance V• + Vb. Common error ment, even for inexpensivereceivers.
sources,from transmitter parameters, for example, do A detaileddiscussion of the integerresolutionof phase
not contribute to the individual covariances and cannot bias parameters[Blewitt, 1989] is beyondthe scope
get removedin the formation of the vector difference. of this article. Because of receiver- and transmitter-
Real measuresof performance, like daily repeatabil- specificphase delays, which are generally not known
ity, will show an improvement if common error sources and may vary with time, it is only the double-difference
are differencedaway. For example, repeatabilities of phase biasesthat will be integers. Thus resolution
baselinesformed from point-positionedresultsof south- of these parameterscannot be performed on point-
ern California sites are smaller by 20%-60% compared positionedresultsfrom a single receiver. Results from
5016 ZUMBERGE ET AL.: PRECISE POINT POSITIONING OF GPS DATA

multiple point-positioned sites in a local network can be Acknowledgments. We thank the Associate Editor
and referees for their useful comments and criticisms. The
combinedto resolvedouble differencephase ambiguity
parameters, with reduction in baselinedaily repeatabil- global network of precision GPS receiversis the result of in-
ternational collaboration among many groups. The authors
ities, especiallyfor the east and vertical components.
are grateful for the rich data set that this network provides.
For those applications that do not require accuracies This work was performed at the Jet Propulsion Laboratory,
better than a few millimeters in the horizontal dimen-
California Institute of Technology,under contract with the
sion and approximately 1 cm in the vertical dimension, National Aeronautics and Space Administration.
ambiguity resolutionis not necessary,providedthat the
observationtime is of the order of 1 day; for shorter ob- References
servingtimes, ambiguity resolutionis increasinglyvalu- Argus, D. F., and M. B. Heftin, Plate motion and crustal
able. For applicationsinvolving long baselines,for ex- deformation estimated with geodeticdata from the Global
ample, global networks,ambiguity resolutionis difficult PositioningSystem, Geophys.Res. Lett., 22 (15), 1973-
at best. For denseregional networksthe rigoroussolu- 1976, 1995.
tion of resolving double-differencephase biasesinvolv- Bar-Sever, Y. E., W. I. Bertiger, E. S. Davis, and J. A.
Anselmi, Fixing the GPS Bad Attitude: Modeling GPS
ing hundreds of sites is not computationally feasible.
Satellite Yaw During EclipseSeasons,Navigation, •3 (1),
Most approachesuse clusters of stations with overlap- 1996.
ping subnetworks;these probably suffer somedegrada- Bierman, G. J., Factorization Methods for Discrete Sequen-
tion when compared to a rigorous solution. Additional tial Estimation, Academic, San Diego, Calif., 1977.
work is necessaryto quantify the trade-off betweencom- Blewitt, G., Carrier phase ambiguity resolution for the
putational feasibilty and mathematical rigor as it relates Global Positioning System applied to geodetic baselines
to ambiguity resolution. up to 2000 km, J. Geophys. Res., 9• 10,187-10,283, 1989.
Blewitt, G., M. B. Heftin, F. Ho Webb, U• J. Lindqwister,
and R. P. Malla, Global coordinates with centimeter accu-
Conclusions
racy in the international terrestrial referenceframe using
We have shown that a robust and economical means GPS, Geophys.Res. Lett., 19 (9), 853-856, 1992.
Blewitt, G., M. B. Heftin, K. J. Hurst, D.C. Jefferson,
of analyzing data from hundredsor more GPS receivers F. H. Webb, and J. F. Zumberge, Absolute far-field dis-
involves(1) analyzinga globally distributed subsetto placements from the 28 June 1992 Landers earthquake
determinetransmitter parameters(and receiverparam- sequence, Nature, 361, 340-342, 1993.
eters for the subset)and (2) analyzingthe remaining Blewitt, G., Advances in Global Positioning System Tech-
data one receiver at a time. The quality of results for nology for Geodynamics Investigations: 1978-1992, in
station parameters is the same whether the receiver is Contributions of Space Geodesyto GeodynamicsTechnol-
ogy, Geodyn. Set., vol. 25, edited by D. E. Smith and
in the first or secondcategory. Beyond a certain num-
D. L. Turcotte, pp. 195-213, AGU, Washington, D.C.,
ber the quality of transmitter parameters does not im- 1993.
prove by including more receiversin the first category. Bock, Y., et al., Detection of crustal deformation from the
The number of receiversthat belongin category(1) Landers earthquake sequenceusing continuousgeodetic
dependson the global distribution. Although it is cur- measurements, Nature, 361, 337-340, 1993.
rently about 35-40, that number may increase as per- Boucher, C., Z. Altamimi, and L. Duhem, Results and anal-
manently operating receiverscome on-line in currently ysis of the ITRF93, IERS Tech. Note 18, Obs. de Paris,
Paris, Oct., 1994.
isolated regions. Businger, S., S. R. Chiswell, M. Bevis, J. Duan, R. A. An-
Current interest in GPS õeodesyis the densification thes, C. Rocken, R. H. Ware, M. Exner, T. VanHove, and
of the terrestrial referenceframe [Zumbergeand Liu, F. S. Solheim, The promise of GPS in atmospheric moni-
1994]. A commonview, in our opinionerroneous,is toring, Bull. Am. Meterol. Soc., 77 (1), 1996.
that the ability to processGPS data from hundredsor Feigl, Kurt L., et al., Spacegeodeticmeasurementof crustal
thousandsof receiversevery day, usingeconomicalcom- deformation in central and southern California, 1984-
1992, J. Geophys. Res., 98, 21,677-21,712, 1993.
puters, is beyond the capability of any single analysis
Gendt, G., G. Dick, and Ch. Reigber, IGS Analysis Cen-
center. With the technique described here, approxi- ter at G FZ Potsdam, in International GPS Service for
mately10• sitescanbeprocessed with a single40-Mfiop Geodynamics 1994 Annual Report, edited by J. F. Zum-
computer every day• All solutionsderived from a set of berge, R. Liu, and R. E. Neilan, JPL Publication 95-18,
fixed transmitter parameters are in the same reference Pasadena, Calif., 1995.
frame, which is the referenceframe of the global solu- Heftin, M., et al., Global geodesyusingGPS without fiducial
tion from which the transmitter parameters were deter- sites, Geophys.Res. Lett., 19 (2), 131-.134,1992.
Herring, T. A., D. Dong and R. W. King, Sub-milliarcsecond
mined. We expect that the efficacyof the techniquewill
determination of pole position using Global Positioning
be demonstrated in the denseGPS arrays in Japan and System data, Geophys.Res. Lett., 18 (10), 1893-1896,
southern California. 1991.
It has long been recognizedthat preciseGPS orbits, Kouba, J., Analysis coordinator report, in International
such as those producedby analysiscentersin the IGS, GPS Servicefor Geodynamics199• Annual Report, edited
improve the quality of GPS analysis. However, the by J. F. Zumberge, R. Liu, and R. E. Neilan, JPL Publi-
value of precise GPS clocks has not been widely rec- cation 95-18, Pasadena, Calif., 1995.
Kuang, D., B. E. Schutz, and M. M. Watkins, On the struc-
ognized. As we have shown, together with preciseor- ture of geometric positioning information in GPS mea-
bits, they allow analysis of data from one receiver at surements,J. Geodesy,71 (1), 35-43, 1997.
a time with few-millimeter daily precisionin horizontal Larson, K. M., and J. Freymueller, Relative motions of the
componentsand centimeter precisionin the vertical. Australian, Pacific and Antarctic plates estimated by the
 ZUMBERGE ET AL.: PRECISE POINT POSITIONING OF GPS DATA 5017

Global Positioning System, Geophys.Res. Lett., 22 (1), Yunck, T. P., GPS data, acquisition,environmentaleffects,
37-40, 1995. U.S. Natl. Rep. Int. Union Geod. Geophys.,1991-199•,
Lichten, S. M., Estimation and filtering for high-precision Rev. Geophys., 33, 349-352, 1995.
GPS positioning applications, Manuscr. Geod., 15, 159- Zumberge,J. F., and R. Liu (Eds), Densificationof the IERS
176, 1990. Terrestrial ReferenceFrame ThroughRegional GPS Net-
Lindqwister, U. J., A. P. Freedman, and G. Blewitt, Daily works,JPL Publication 95-11, Pasadena,Calif., 1994.
estimates of the Earth's pole position with the Global Zumberge, J. F., M. B. Heftin, D.C. Jefferson, M. M.
Positioning System, Geophys.Res. Lett., 19 (9), 845-848, Watkins, and F. H. Webb, Jet PropulsionLaboratory IGS
1992. AnalysisCenter 1994 annual report, in International GPS
Melbourne, W. G., Sounding the Earth's atmosphere and Servicefor Geodynamics 199,1Annual Report,edited by
ionospherewith GPS, Eos AGU Trans., 76 (,16), 465-466, J. F. Zumberge, R. Liu, and R. E. Neilan, JPL Publica-
1995. tion 95-18, Pasadena, Calif., 1995.
Noll, C. E., CDDIS Global Data Center report, in Interna- Zumberge,J. F., and W. I. Bettiger, Ephemerisand Clock
tional GPS Servicefor Geodynamics199• Annual Report, NavigationMessageAccuracy,in GlobalPositioningSys-
edited by J. F. Zumberge, R. Liu, and R. E. Neilan, JPL tem - Theory and Applications,vol. I, edited by B. W.
Publication 95-18, Pasadena, Calif., 1995. Parkinson and J. J. Spilker, 585-599, Am. Inst. of Aero-
Webb, F. H., M. Bursik, T. Dixon, F. Farina, G. Marshall, nautics and Astronautics,Inc., Washington,D.C., 1996.
and R. S. Stein, Inflation of Long Valley Caldera from one
year of continuousGPS observations, Geophys.Res. Lett.,
22 (3), 195-198, 1995.
Wilson, B. D., A. J. Mannucci, C. D. Edwards, Subdaily M. B. Heftin, D. C. Jefferson, M. M. Watkins,
northern hemisphereionosphericmaps using an extensive F. H. Webb, and J. F. Zumberge, Jet Propulsion Lab-
network of GPS receivers, Radio $ci., 30, 639-648, 1995. oratory, California Institute of Technology,MS 238-600,
Wu, J. T., Estimation of clock errors in a GPS based track- 4800 Oak Grove Drive, Pasadena,CA 91109-8099. (e-
ing system, paper presented at the Am. Inst. of Aero- mail: [email protected],
[email protected],
naut. and Astronaut./Am. Astronaut. Soc. Astrodynam- [email protected], [email protected],
ics Conference, AIAA, Seattle, Wash., 1984. [email protected])
Wu, J. T., S. Co Wu, G. A. Hajj, Wo I. Bertiger, and
S. M. Lichten, Effects of antenna orientation on G PS car- (ReceivedMay 3, 1996; revisedDecember2, 1996;
rier phase, Manuscr. Geod., 18, 91-98, 1993. acceptedDecember9, 1996.)