PUBLICATIONS ON GEODESY
GE,ODE,TIC
COMMISSION
NEW SERIES NUMBER 38
RADBOUD KOOP
r993
NEDERLANDSE COMMISSIE VOOR GEODESIE,THIJSSEWEG11,2629 JA DELFT, THE NETHERLANDS TEL. (31)(0) s78289, FAX (31)(0) st 823 l 1 1 48
4 ISBN906132246
Contents
Acknowledgernents Abbreviations Abstract 1 2 Introduction Satellite gradiometry: principles and applications 2.t Principle of satellite gradiometry 2.2 Applications . 2.3 Aristoteles
vl vl
vll
8 8 t2 L7
3 The gradient tensor and its series representation in different coordinate systems 2L 3.1 Potential derivativesin different coordinate systems 23 3.1.1 Algorithm 24 3.1.2 Transformation formulae for the potential derivatives. 27 3.2 Seriesexpansionof the potential and its derivatives 32 3.2.1 Sphericalharmonics 32 3.2.2 Orbital coordinates 35 3.3 Synthesisand analysis 38 4 Global gradiornetric analysis 4.1 Least squaresanalysis 4.2 Colombo's method of error analysis 4.2.1 Normal matrix : 4.2.2 Presentation 4.2.3 Ideal case 4.2.4 Band limitation . 4.2.5 Polar gaps 4.2.6 Stabilization 4.2.7 Omission,commissionand smoothing
51 52
DJ
60 64 68 80 84 90 99
III
4.2.8 Aristoteles 4.2.9 Some computational aspects 4.2.10 Conclusions 4.2.1I Other error analysismethods 4.3 Global recovery 4.3.1 Linear model 4'3'2 Iteration 4.3.3 Spacelikev. timelike . 4.3.4 Simulated data 4.3.5 Spacelikeresults 4.3.6 Timelike results 4.3.7 Some computational aspects 4.3.8 Conclusions 5 Relativistic view on gradiornetry 5.1 Some aspectsof the generaltheory of relativity 5.1.1 The spacetimeof relativity 5.L.2 Equations of motion 5.1.3 The principle of equivalence 5.1.4 The Einstein field equations 5.2 Weak field approximation 5.2.I Newtonianlimit 5.2.2 Linear approximation in relativistic terms 5.3 PostNewtonian approximation 5.4 Equations of motion revisited deviation 5.5 Equation of geodesic
105 111 113 Lt4 116 116 l2o 123 L27 L28 132 133 135 138 l4O L4L L43 L46 149 151 151 153 156 158 164 t74
B Index notation and tensor analysis B.1 Index notation 8.1.1 Kernel letters B'l'2 Indices B'1'3 Matrices 8.2 Tensor analysis B'2'1 Tensor 8.2.2 Co and contravariant B.2.3 Derivattves B.3 Geometry B.3.1 Spaces B.3.2 Metric
tv
B.3.3 Curvature C Coordinate systems C.l Definition C.2 Metric and Christoffel svmbols References
Acknowledgements
Many people, friends and colleagues,contributed in many different ways to the ac. complishment of this work. I am very.grateful for so much help. In particular I like to thank Reiner Rummel for his continuous support in practically any field and Martin van Gelderen who helped me both as a friend and as a colleague. Furthermore, I am much indebted to Wim Corbey, who wrote the timelike version of the full gradiometric analysis program and David Stelpstra for providing both the gradient synthesis program as well as the spacelike version of the analysis program. The Center for Space Research of the University of Texas at Austin is gratefully acknowledged supplying me with the simulated one month GRM gradient data. for I also like to thank Gerrit Bakker, Ernst Schrama and Prof. E. Grafarend for the discussions had with them. I
Abbreviations
Aristoteles c.p.r. CPU CSR ESA FFT GPS GRS80 GTR MIMD MSE NASA PN PPN r.m.s. SGG SIMD SISD SST STR TR Application and ResearchInvolving Space TechniquesObserving The Earth's fields from a Low Earth orbiting Satellite cycles per revolution Central ProcessingUnit Center for SpaceResearch, University of Texas at Austin European SpaceAgency Fast Fourier Transform Global Positioning System Geodetic Reference System 1980 General Theory of Relativity Multiple Instruction Multiple Data Mean Squared Error National Aeronautics and SpaceAdministration Post Newtonian ParametrizedPostNewtonian root mean square Satellite Gravity Gradiometry Single Instruction Multiple Data Single Instruction Single Data SatellitetqSatellite Tlacking Sea Surface Topography Special Theory of Relativity Tscherning/Rupp degreevariancemodel
VI
Abstract
The gravitational field of the earth can be determined globally and with high precision and resolution by means of a combined Satellite Gravity Gradiometry (SGG) and SatellitetoSatellite Tlacking (SST) mission. In such a mission a spacecraft equipped with a GPS receiverand a gradiometerwill be flown around the earth in a low and nearly polar orbit. The GPS receiveris used for the determination of the long spatial wavelengths of the earth's gravitational field and the gradiometer for the short wavelengths. As such the two techniques are complementary. This work focuses on the satellite gradiometry part only. A gradiometer delivers the second order potential derivatives relative to some local orthonormal coordinate system. In particular, the gravity gradients are determined by a technique called differential accelerometry,in which the outputs of any combination of two out of (ideally) eight accelerometersare differenced. A planar gradiometer consistingof four accelerometers and working according to this principle will be onboard the Aristoteles satellite, a mission planned by the European Space Agency (ESA). The improved knowledgeof the earth's gravitational field, resulting from such a mission, can contribute to many earth related sciences,like geodesy(levellingwith GPS), satellite orbit determination,solid earth physics (continental lithosphere,polar regions) and oceanography, latter not only for topics the like ocean circulation but also for study of climate changes. The earth's gravitational potential, together with its first and secondorder derivatives, is usually expressedas a seriesexpansion. The coefficientsof such a series (potential coefficients) describe the gravitational potential globally and are to be determined from SGG. The equations of the gradient seriesexpansions,either in geocentricpolar coordinatesor in orbital coordinates(Keplerian elements) are used , as model equations in the gradiometry analysis process. However, the measured gradients are delivered relative to a local orthonormal coordinate system connected to the instrument. We need, therefore, transformation equations for the potential derivatives between several coordinate systems. By means of a compact, general algorithm, which makes use of certain conceptsfrom tensor analysis,these equations can be derived. Using a set of known potential coefficients, the series expansions can be used to gain some insight in the signals measuredby a satellite gradiometer (spherical harmonic synthesis).Reversely, analyzing a set of gradientsto obtain the harmonic coefficientsis called spherical harmonic analysis. Both synthesis and analysis are time consumingprocesses, least if a high degreeand order seriesexpansionis used at or a large set of data points is used. Grid computation (making use of FFT routines) and the use of vector computers (for which the algorithms should be adapted) can decreasecomputation at burden. Numerical errors may enter the computations due to the use of recursive Legendre function computations and numerical quadrature formulas. For the estimation of potential coefficientsfrom a set of observed gravity gradi
vtr
ents we used an analysis technique based on least squaresadjustment, as it was, for the gradiometric case, proposed by O.L. Colombo. Under certain assumptions the normal matrix attains a blockdiagonal structure, making it easy to solve the large system of linear equations. The aposteriori error covariancematrix of the estimated potential coefficients can be computed by means of error propagation without the availability of actual measurements. We carried out such error analysis for several not only for an idealizedmission (polar orbit, no band limitation, missionscenarios, full tensor gradiometer),but also more realistic situations like polar gaps,band limitation of the gradiometer and the planar Aristoteles gradiometer measuring only three tensor components. The latter, nonideal situations have a rather large impact on the results. Apart from numerical singularities due to illconditioned subblocks of the normal matrix, some blocks may become singular too, which means that certain coefficients are not estimable any more from the observations. Of course the system can be stabilized by adding prior information, but this leads to biased estimates. Furthermore it appearsthat, even with stabilization, the results with a band limited gradiometerdid not meet the requirements.Additional GPS tracking information, as it is planned for the Aristoteles mission, is needed to resolve this problem. Sincewe had at our disposala global set of simulated gravity gradients,we could perform a global recovery of potential coefficients from this set. The procedure showed here is an iterative process, but only a first step is really implemented so far. Two strategies were applied. The spacelike method usesthe seriesexpansion in geocentricpolar coordinates, and it requiresthe set ofobservations to be transformed into a global equiangular grid ofaveraged values. Such a grid is not required for the socalled timelike method (which usesthe seriesexpansionin orbital coordinates), but this method is more time consuming. Both methods give promising results. In our error analysis computations we assumedan instrumental precision of 0.01 E l\/Hz (as foreseen for Aristoteles) . However, superconducting gradiometers are under development,aiming at a precision of 0.000L Elt/Hz. Also, orbit determination techniques are gradually improving, aiming at the centimeter level. Such improvements of precision may, in the future, require a relativistic formulation of the model. Furthermore, from a theoretical point of view, a relativistic description of satellite gradiometry emphasizesthat a gradiometer measures the curvature of fourdimensional spacetime. For those reasons,we included a relativistic view on gradiometry. The equations are derived in the weak field approximation (in particular the socalled PostNewtonian approximation), which is sufficient for earth orbiting satellites. The equations of motion of the satellite appear to be those of a spacetimegeodesicand they show the relativistic contributions to the satellite's orbit. A relativistic description of the second order potential derivatives appears to be governedby the socalled equation of geodesicdeviation. The latter shows the relativistic contribution to the observed gravity gradients. It is concluded that relativistic effects can be modelled although they are not required at present.
vill
1 l
Introduction
In spite of the fact that the concept of satellite gradiometry is more than 30 years old, no actual gradiometric mission has been undertaken yet. Several instrument and mission proposalshave been done, of which the Aristoteles mission concept of the European SpaceAgency (ESA) is the most promising one for the near future. The main purpose of a Satellite Gravity Gradiometry (SGG) mission is the global determination of the earth's gravitational field with high precisionand spatial resolution. Some central aspectsof the processto derive from the actual observations the required gravitational information will be treated in this work. This explainsthe title "Gravity Field Modelling using Satellite Gravity Gradiometry" . In this title, three topics attract our attention. Obviously, the key issue here is gravitation, which word is reflectedin the word grauily. The word gradiometryrevealsthat we have to do with rneasurements. Finally, the word satellite points out that the measurements are carried out in a spacecraft. Gravitation The tale of the falling apple marks one of the great moments in scientific history. According to it, Isaac Newton (who lived from 1642 to 1726), when sitting in his garden, watched an apple fall from a tree. The apple fell straight down, along a line which, if extended inside the earth (considereda homogeneous sphere at the moment), would go through the center of the earth. While thinking about this phenomenon,he came up with his famous law about the mutual attraction of two masses.The attractive force, called grauitation,is directed along the line connecting the two centersof the objectsand is, up to someconstant (the gravitational constant G), proportional to both massesas well as to the squared inverse of the distance between the two objects. This famous inuersesquarelaw thus tells us that on the one hand the gravitational force increases with increasingmassof one of the objects, whereason the other hand the force decreases with increasingdistancebetweenthe objects. Objects moving under the influence of a force will be acceleratedin the
1. Introduction
same direction as the force acts. Whereas the gravitational force depends on the does not. mass of the object, the grauitational acceleration Newton lived more than two and a half centuries ago, we still use Although the inversesquarelawfor all earth related applications. Not only does this law explain the falling of an apple. By using the same law, the motion of the planets around the sun and of moons and satellitesaround the planets or the trajectories of projectiles can be explained. In the beginning of the 20th century, however, measurementtechniquesbecamemore accurate,and some aspectsof the motion of the planets in the solar system which showed up in the measurements,could no longer be explained by Newton's law. Albert Einstein (18791955)came up with a new theory about gravitation. Einstein's gravitational theory is the famous general theory of relativity (GTR). The addition "general" refersto the fact that gravitation is included in the theory, where this was not the casein the socalled specialtheory of relativity (STR), which he conceivedsome years earlier. Einstein's GTR did a not state that Newton was wrong. In fact, Newtonian theory represents limiting, applications,the use of the GTR is approximate caseof the GTR. For cosmological inevitable, but for most earth related applicationswe can still confine ourselvesto Newtonian theory. However,with increasingmeasurementaccuracy,as for instance techniques,relativistic is the casewith many modern satellite related measurement have to be taken into account. effectsmight If the earth would have a perfectly spherical shape and if the mass inside the (i.e. equal mass density throughout the earth would be distributed homogeneously whole sphere) or rotationally symmetric, the line along which Newton's apple fell would indeed be a straight line, directed radially and going exactly through the earth's center. This center point would be the center of mass of the earth. The gravitational force exerted by the earth onto a proof mass (like the apple) would be exactly the sameas the force exertedby a point massof infinite small dimensionwith equal mass as the earth and located at the center of the earth. At any other point at the earth's surfacethe force on the apple would have exactly the samemagnitude and (radial) direction. Moving away from the earth's surface,the force would still with increasingaltitude. be directed radially, although its magnitudewould decrease The graaitationalfield obtainedin this way would be perfectlysphericallysymmetric. In reality, however,the situation is more complex. Although seenfrom the moon or from any other point in spacethe earth may look much like a perfect sphere,the deviations from this idealization are significant. Due to the fact that the earth is deformable and rotates about an axis going approximately through the north and south pole, it is better represented as an ellipsoid, flattened at the poles. Looking in more detail we also seemass irregularities at the surface of the earth, like oceans, mountains, plains, etc. Also the material inside the earth is not distributed homoa system,with geneously:the dynamics of the earth are more considered convective phenomenasuch as plate tecan enormous variation in temperature, resulting in radial and lateral density tonics, subduction zones,ocean ridges and a considerable differentiation. As a result, the gravitational force at different places on the earth
will not be the same: both magnitude and direction of the force will deviate from one place to the other from that of a homogeneous sphereor spheroid. The gravitational field of the earth is rather irregular, although the deviations from a spherical symmetric shape are not very large. In fact, for some applications a spherical earth may be a sufficient approximation. For other applications an ellipsoidal earth may be needed, but often higherorder approximations are required. Therefore, the gravitational field is representedby a series, of which the sphericalpart is the first term, the ellipsoidalpart the secondterm, and so on. The more detailed information or the more accurateinformation about the gravitational field we need, the more terms are required in the seriesexpansionfor an adequate representation. If we would exactly know the (irregular) mass distribution of the earth, we could compute the gravitational field from it (and thus obtain all terms in the seriesexpansion) and see how much this field deflectsfrom a perfectly spherically symmetric one. Unfortunately we still do not know the exact mass distribution. On the contrary, one of the main objectivesof gravitational field determination is to learn more about the internal mass constitution. Unfortunately an exact inference is impossible. One speaksof the gravitational inverseproblem. Measurernents If Newton would have measured the time at which the falling apple would have passed through imaginary levels of known distance, he could have computed the gravitational acceleration.Modern instruments sti[ use such free fall technique to measuregravitation. More precisely,when situated on the rotating earth, not gravitation is measuredbut grauity. Gravity is the sum of gravitation (attractive force between two masses)and centrifugalacceleration The latter is due to the rotation of the earth and its direction is perpendicularto the axis of rotation. Today's instruments are very accurate. In the caseof an absolute gravity apparatus, a proof mass is dropped in a vacuum chamber and its path is measuredinterferometrically with a laser. In this way one obtains the magnitude (or length) of the gravity vector, at the specificplace where the measurement done, to lOe. Such absolutegravity is measurements are difficult and thereforeexpensive.They are carried out only at a few points on earth. Since gravity on earth is approximately lOl m/s2 the achieved precisioncorrespondsto accelerations small as 108 ^/"2 . as In order to get an idea about gravity in between absolute stations, one measures gravity differences arbitrary measurementpoints relative to the absolute at stations. Such relotfuegravity measurements are easier to perform than absolute measurements and are carried out with spring grauirneters.Thereby, the length of a spring, mounted under a particular angle and suspendinga horizontal lever with a proof mass)is measured.With modern spring gravimetersa precisionof 5 to 10 . lOE m/s2 can be reached. The direction of the gravity vector is defined by the astronomical latitude and
1. Introduction
longitude, i.e. the direction of the plumb line as determined by astronomical observations. Thus the gravity vector in any point is determined by three quantities: astronomicallatitude and longitude (direction) and magnitude (usually referredto as grauity).In our surrounding three dimensionalworld any vector can be determined by three quantities: its componentsin all three spatial directions. Mathematically speakingthesethree componentsare the gradientof somescalarfunction, in our case the grauity potential. They describehow the scalar function changeswhen moving in all three directions. But we have seen that gravity itself also varies from place to place. These changesof gravity can also be describedby gradients,and are thus called grauity gradients.They also can be used to describethe gravitational field of the earth and are more sensitiveto small variationsof the earth's massconstitution. deAt the beginningof our century,the HungarianphysicistEcitvos(13481919) velopedan instrument to measuregravity gradients.This instrument was a socalled torsion balancer.Instead of one proof mass used in a gravimeter, a torsion balance to consistsof two proof masses,(asymmetrically)suspended the arms of a balance. The gravity gradientsproduce a torque on the beam of the balance, resulting in a by rotation of the beam. The gravity torque is counterbalanced a restoring torque exerted by the fiber with which the beam is suspended. The restoring torque is a achievedwith his instrument a precision measurefor the gravity gradients. Ecjtvcis s o f 1 E ( t E : l E c i t v c iu n i t : 1 0 * e / s 2 ) . W i t h t h e l a r g e s tg r a d i e n to n e a r t h b e i n g approximately 3000 E, this meansa precisionof 103. In generalinstruments which and the measurementtechnique measuregravity gradients are called gradiometers, is called gradiometry (measuringthe gradients of gravity), analogousto the word grauimetry (measuringgravity). In principle, Newton could have measured gravity gradients if he would have seen ttuo apples fall simultaneously from the tree and would have closely followed their adjacent trajectories. Both trajectories, if extended inside the earth, would convergeas to run through the earth's center of mass,so the distance between the paths would gradually decrease.If the change in the distance between two falling would be measured,it would be a measureof the gravity gradient, i.e. proof masses In of the variation of gravity in the direction of the line connectingthe two masses2. motion of the falling proof masses, a next step, one could think of constrainingthe
r W i t h t h e s a m e t y p e o f i n s t r u m e n t E 6 t v b s d i d a h i s t o r i c a l t e s t t o l 0  8 o f t h e s o  c a l l e dp r i n c i p l e e o f e q u i u a l e n co f i n e r t i a l a n d g r a v i t a t i o n a lm a s s ( E o t v 6 s , 1 9 5 3 ) , ( J u n g , 1 9 6 1 ) . I n r e c e n t y e a r s h i s tests were reanalyzed in relation to a search of the socalled fi,t'th /orce (Fischbachet al., 1986). tln the language of the GTR, the two proof masses,being in free fall in the gravitational field of ("shortest paths'). In a flat space,Iike a two dimensional sheet the earth, follow nearby geodesics of paper, geodesicsare straight lines, which, if being parallel initially, remain parallel (constant distance between the lines) and never cross. In the GTR, the space under consideration is not flat but curved, just like the two dimensional surface of a sphere. Fbrthermore it is four dimensional, (which is hard to visuwhere three spatial directions and time merge into one socalled spacetime alize). Geodesicsin curved spacetime are therefore not straight lines but are curves. The distance changesand this distance change is a measurefor the curvature between iniiially parallel geodesics of spacetime. So in terms of the GTR, gravity gradients describe the curvature of spacetime.
such that  during their fall  they are held in a fixed position relative to each other and their distancewould remain constant. The force neededto constrain the motion could be measuredand would again be a measureof the gravity gradients. This is gradiometry. There, a gradiometer, consisting of the principle used in spaceborne arrangedin a two or three dimensionalorthogonalsetup, two or more proof masses, is flown around the earth on board a spacecrBft. Whereas the gravity gradient is (mathematically) defined in exactly one point of infinitesimal dimension, a real life gradiometer has some definite size. Practical limitations (i.e. material properties) prevent us from reducing the size of an apparatus ad infinitum, so the distance of in two proof masses a gradiometer,although small, is not infinitesimal,but has some definite value, say 1meter3. This meanswe are actually measuring Lhe differencein gravitational accelerationbetween the locations of the two proof masses. For that reason the technique is often called differential accelerometry. ESA has planned a satellite mission called Aristoteles, which should carry a gradiometer measuring principle. gravitational gradients using the differential accelerometry Satellites The Aristoteles project is a socalled dedicated gravity field mission, its main purpose being the determination of the earth's gravitational field. The idea, however, of using earth orbiting satellitesfor such a purpose is not very new. Already from the time of the Iaunch of the first artificial earth orbiting satellite in 1957, people analyzed their orbits to determine the main characteristicsof the earth's gravitational field. As said before, the orbits of satellitesaround the earth or around any other planet, as well as the orbit of planets around the sun, are determined predominantly by the gravitational field of the earth, respectivelyof other planets or the sun. In the 17th century) even before Newton, Kepler establishedthree laws concerning the motion of planets around the sun. Actually Newton used Kepler's laws when deriving his famous inversesquarelaw. According to Kepler, planets move in elliptic orbits around a spherical body. The earth is nearly spherical, its gravitational field is nearly sphericallysymmetric, and thus the orbits of earth orbiting satellites are nearly elliptical. The main deviation from spherical symmetry, of of the earth's flattening, results in a precession the orbital plane and a precession the orbit ellipse in the orbital plane. Other departuresfrom sphericalsymmetry in the gravitational field (in generalcalled anomalies)yield other, smaller, deviations in the satellite's orbit. Precise analysis of satellite orbits thus tells us something about the gravitational field. The principle is not so very much different from Newton's falling apple. Also is satellites "fall" in the gravitational field of the earth. The difference that Newton's apple started, relative to the earth, from a rest position at the branch of the tree
3Actually, according to the principle of equivaience, infinitesirnalsmall gradiometer would not an exist. Strictly speaking this principle holds in exactly one (infinitesimal small) point only, where one could no longer discriminate between inertial and gravitational accelerations.
1. Introduction and fell straight down on the earth's surface. Would it have been thrown away horizontally (i.e. parallel to the earth's surface) with high velocity, it would have hit the ground some distance away from him. The higher above the earth's surface a proof mass is ejected ("launched") and the higher its initial velocity, the further away the point of hitting the ground. Eventually, due to the spherical shape of the earth, the proof mass never hits the earth's surface,but continuesto fall, so to say not to the earth, but around it. Obviously, as already pointed out by Newton, the same law which describesthe fall of the app.le,describesthe motion of satellites around the earth. effort has been investedinto the determination of the earth's graviConsiderable tational field from the analysisof satellite orbits. The same method will, additional to the gradiometry part, be used with ESA's Aristoteles mission. Especially the main deviations from a spherically shaped gravitational field can be determined from orbit analysis. Other, smaller deviations,i.e. the more detailed structure, represented by higherorder gravitational field parameters, are determined from the gradiometer measurements. the range of detail with which the gravitational field can be deterNevertheless, gravitation from space,is not unlimited. According to the inversesquarelaw, mined attenuation with increasingdistance between the objects. This socalled decreases effect is especiallynoticeablewith satellites,since they move at very large distances from the earth's surface. Even with Aristoteles, which will move at a "very low" altitude of "only" 200 km, many details of the gravitational field will remain undetectable. Furthermore, at 200 km there is still some atmosphericdensity left, which are extremely disturbs the motion of the satellite and due to which measurements difficult. Atmospheric drag causesthe satellite to descend,so that it needs to be kept in a 200 km orbit by maneuvers,using small rockets. These maneuvers also as disturb the measurements, does e.g. the sloshingand consumption of the fuel in the tanks neededfor the rockets. A great advantageof using satellitesfor gravity field determination is that with one single mission (nearly) the whole earth can be coveredwith measurementsin short time. Of course, launching a satellite is expensive,but travelling the earth by conventional means with an instrument and carrying out measurementswith density as a satellite, is much more expensive. Just think about corresponding high mountains, oceans and polar regions, which may constitute insurmountable hindrances for doing terrestrial measurements. Even after more than 50 years of Iarge parts of our planet remain unsurveyed. Furthermore, terrestrial measurements is a set of terrestrial measurements not Iikely to have comparableprecision all over the world, whereasthe satellite mission will deliver us a global data set of homogeneous quality.
This study in brief This study deals with the main subjects of a global determination of the earth's gravitational field from a satellite gravity gradiometry (SGG) mission, and as such and satellitesas discussed above. Beincludes aspectsof gravitation, measurements sides introductory remarks on the principles of satellite gradiometry and the applications of a preciseand highly detailed gravitational field, this study focusesmainly (gravity gradients at satellite on the analysisprocessof converting the observables altitude) into the gravitational information at z.eroaltitude. Furthermore, special attention is given to the precision of the derived gravitational information, being a function of measurementprecision,mission parametersand chosenmathematical model. Finally, as a look into the near future, a descriptionof Einstein'sGTR is to given, in as far as it may concernan SGG mission.Relativisticcorrecbions Newto ton's theory are derived,as may be necessary accountfor in future gradiometer missions.
2 l
Satellite gradiometry : principles and applications
The purpose of gradiometry is the preciseand detailed determination of the earth's gravitational field. This is done by measuring the secondorderderivatives of the thesesecondgravitational potential V of the earth. An instrument which measures order derivativesis called a gradiometer. In satellite gradiometry such a gradiometer is flown in a satellite around the earth in a low, almost circular and polar orbit. In this chapter the principle of satellite gradiometry will be illustrated by looking at the motion of test massesin space. Furthermore a short overview of the possible about instrumental applicationsof gradiometry is given, as well as a short discussion aspectsconcerningthe gradiometerand the satellite in which it is flown, with special attention to the planned Aristoteles mission.
2.L
Principle
of satellite gradiometry
situated in two nearby points A and B in space,see Consider first two proof masses figure 2.1.a. The gravitational accelerationin point A due to the attraction of the are given with respectto a local cartesiancoordiearth is U(A).The components Seeappendix A for the definition ofvarious coordinate systemsand nate system "d. appendix B for some remarks on notation. The gravitational accelerationvector is directed along the line of force going through point A, perpendicularto the equipotential surface through A. The gravitational accelerationin B is V;(B) Suppose A and B are situated on the same equipotential surface. Without any support and the earth. will fall towa.rds without any other forcesappearing,the two proof masses Since the gravitational field of the earth is almost a perfect central force field, the the earth due while falling towa.rds will decrease distance between the proof masses distance between the proof of to the convergence the Iines of force. The changein massesis a measurefor the differencein gravitational accelerationin A and B, cf. (Carroll and Savet,1959)and (Savet,1969).
2.1.
AQ
o
a
I I vi@)
O'
v;(B)
a. Proof masses on the same equipotential surface Figure 2.1 Falling proof masses.
Now consider the two points A and B situated on the same line of force but on different equipotential surfaces,see figure 2.1.b. The proof mass in B is closer to the earth as the one in ,4 and is therefore pulled harder. If the proof massesare dropped and are falling towards the earth along the same line of force, the distance between them will increase.Again, the changein this distance is a measurefor the betweenthe proof masses. differencein gravitational acceleration In a next step, consider a spacecraft(satellite) carrying out an orbital motion around the earth. The spacecraft,or more preciselythe center of mass O of the spacecraft,is in free fall in the gravitational field of the earth. In caseof a circular at orbit, the gravitational acceleration O, which is directed towards the earth, is, at resulting from the orbital any moment, compensatedby the centrifugalacceleration motion, which is directed outwards. In O no resultant forces appear' Supposethat the center of mass O of the satellite is not a material point, so that we can place one of the two proof massesfrom above at O. The proof mass will then also be in free fall around the earth, carrying out the same orbital motion as the satellite and it will remain at O (and thus at rest relative to the spacecraft)during the motion' Supposenow that we place the secondproof massat another point in the interior of the spacecraft,close to O. The second proof mass will be situated on another line of foice and/or another equipotentialsurfaceof the gravitational field, so it will
2.
Satellite gradiontetry:
start to move relatively to the first proof mass at O. This relative motion results in a distance change between the proof massesaccording to the examples above (fig.2.1). If both proof massesare placed inside the satellite at arbitrary points but not at O, again a relative motion between the proof masseswill appear as a result of the differencein gravitational accelerationat the two points, cf. (Rummel, 1989a), but furthermore they will move relatively to O. So they will start drifting inside the satellite, eventually hitting each other or the satellite skin. This skin consistsof points, which, together with all other (material) points of the satellite, form a rigid body. AII points of this body carry out the same orbital motion as O, i.e. they have the same angular velocity (disregarding any they are at any moment situated at other points in the deformations). Although gravitational field than O they remain, as a result of the material composition,in a fixed position with respect to O. Supposenow that the two proof massesare also kept in a fixed position with respect to O by means of some mechanicalor electrostaticalsuspension.Then also the proof massesare forced to carry out the same orbital motion as O. Proof masses constrainedin this way can be seenas accelerometers. The outputs of these accelerometers the forcesneededto keepthe proof masses thesefixed positions are in with respect to O (and thus with respectto each other). They are measures the for accelerationdifferences betweenthe two points and O and can therefore be used as observationsto measurethe gravitational field of the earth. These relative accelerations the two proof masses of can be expressed expandby ing the gravitational acceleration( in a Taylor serieswith respect to O. For the points ,4 and B we have:
(2.1)
If dr'(A,B), the distance betweenthe points A and B, is known, we may compute from this equation the gravitational gradientsV;i(O) sinced[ is measuredby means of the accelerometers. instrument consistingof two or more of such acceleromeAn ters and having as output the gravitational gradientsis called a gradiometer. potential derivativesin the way deThe techniqueof measuringthe secondorder see scribedaboveis known as difrerentialaccelerometry, for example (Forward, 1974)
10
or (Balmino et al., 1985). A gradiometer working according to this principle will The gradiometer will consist of a symmetrically arranged array of accelerometers. be built in the satellite in such a way that the center of the instrument coincides with the center of mass O of the satellite. The neglect of the terms proportional to the third and higherorder derivatives of the potential is of the order of 1010 E (t E : 1 EotvcisUnit : 10e s2) for a distance between A and B of about 1 m and at an altitude of the satellite of about 200 km, which is acceptablein view of the present and near future gradiometer instrument precision,cf. (Paik and Richard, 1986). So in principle it should be possibleto obtain information about the gravitational field of the earth by observing the gravitational gradients V;i in the way described above. However, up to now we consideredthe two proof massesand the satellite moving through spaceunder the influenceof gravitational and rotational forcesonly. forcesmay act on the satellite. In reality also other, external and nonconservative, example,solar radiation pressure,and, at lower altitudes, Such forcesare due to, for air drag due to the earth's atmosphere. They act on the outer skin of the satellite, the satellite housing. But if the gradiometer instrument is rigidly attached to this housing, these forceswill also act on the instrument and therefore also on the proof are due to the suspension.The measuredaccelerations in this casea mixture masses and external forces. Since we are only interested in of gravitational accelerations gravitational forces,such external forces are viewed upon as disturbing forces. One way to deal with thesedisturbing forcesis to build a socalled dragfree satellite. In such a satellite the instrument is not rigidly attached to the satellite housing. This housing, the "outer part" of the satellite, undergoesthe nongravitational disturbing forcescausinga relative motion betweenthe housing and the instrument (or the "inner part" of the satellite). In order to prevent collision of the two the motion regulated.The inner satellitethen carriesout of the outer satelliteis continuously a perfect free motion, under the influenceof gravitation (besidesthat of the earth also of the sun, moon and other planets) only. in On the other hand, since the external forces act on every accelerometer the just like the comsame manner (i.e. same direction and same magnitude), they are, mon accelerationV;(O), eliminated if we take the differencebetween two accelerometer outputs, as in eq. 2.1. This elimination,however,only works properly if the are exactly identical (in dimensionsand in orientation) and if they accelerometers are perfectly aligned relatively to one another and relatively to the instrument. The use of a non dragfree satellite therefore puts more stringent requirements to the constructionof the instrument. A final aspect which we will consider here is the rotation of the gradiometer instrument. Supposethe gradiometeris at rest with respectto the local cartesian with the center of mass O of coordinate system ri, of which the origin coincides the satellite, and thus with the center of the instrument. Supposefurthermore that, at a certain moment (e.g. the initial point of the mission), this local coordinate system is oriented such, that the zaxis is directed radially outwards, away from the
I1
earth, the caxis points in the direction of the motion of the satellite (along track) and the gaxis is perpendicular to the orbital plane (crosstrack) as to complete a righthanded coordinate system. The satellite (or the instrument) is said to have a spacefixedorientation if it orbits the earth such that it keepsthe same orientation with respectto the fixed stars, i.e. the local coordinatesystemremains parallel to the initial position. This means that the zaxis after the initial point no longer points in the radial direction, the caxis no longer along track, only the yaxis remains pointing in the crosstrack direction. On the other hand the coordinate system (and thus the instrument) does not rotate in this case. If we let the zaxis keep its radial direction and the uaxis its along track direction, the satellite is said to be earth pointing. In this casethe coordinatesystem does rotate. In particular it performs one completerotation during one orbital revolution of the satellite. Its angular velocity in that caseis 2r f T, where ? is the orbital period. In our example the rotation takes place about the yaxis. From an earth point of view, such earth pointing motion may be attractive,but it has the disadvantage occur due to the rotational motion. These accelerations that inertial accelerations all are also present in the output of the accelerometers.Since the accelerometers have different positions with respect to O, the effect of the rotation is not cancelled measurements in eq. 2.1. But if we as during the differencingof the accelerometer by can discriminate betweenthe gravitational and the rotational accelerations either some numerical method or some apriori rotational information (so that the rotaknown), cf. e.g. (Rummel, 1986),we have a meansof tional motion can be considered measuring the earth's gravitational field, called satellite gradiometry. For a comprehensivegeneraltreatment of satellite gradiometry,seefor example (Rummel, 1985a, 1985b, 1986). For related topics on the principles of satellite gradiometry see e.g. (Forward, 1981,1982),or (Moritz, 1968).
2.2
Applications
The aim of a gradiometric satellite mission is the determination of the earth's gravitational field, globally with high precision and high spectral resolution. In the caseof the first planned mission, Aristoteles,one would like to obtain the following precisions (seealso section4.2) geoidheights:o(10cm gravity anomalies: o < 5 mgal
(1 mgal : 105 m r2)
both with a spatial resolutionof between50 and 100 km (half*wavelength).This corresponds to a spherical harmonic expansion (see section 3.2) complete up to degreeand order 200 to 300. So far our knowledgeof the earth's gravity field relies on the one hand on the
12
2.2. Applications
analysis of the motion of satellites and on the other hand on terrestrial measurements. Severalgroups collect terrestrial point gravity anomaliesand processthem , t o e q u a l ' a n g u l a r e a n g r a v i t y a n o m a l i e se . g . o f s i z e 1 o x 1 o , 6 ' x 1 0 ' o r 3 ' x 5 ' m (correspondingin our Iatitudes to 100 km, 10 km or 5 km sidelength). The origithat are tied in nal point anomaliesare derived from relative gravity measurements firstorder networks,containing some absolutestations, somecountriesinto national or into the International Gravity StandardizationNetwork. For the computation of anomalies also heights are necessary. Ideally they should be levelled heights referring to a welldefined height datum. In practice one has often to sustain with it barometric heights or not welldefinedlocal levelling networks. As a consequence gravity anomaliesin some areasof the world contain must be feared that the mean systematic errors. For large areas no gravity anomaliesare available at all, either becauseof political reasonsthe data are not made available or becauseareas are e.g. polar regions,high mountain ranges. In particular ocean not easily accessible, areas, where gravity measurementsare difficult to do and very costly, large areas are not covered. A map of the current coveragewith 1' x 1o mean anomaliesis given in figure 2.2. Despite large white areas the impressionis seemingly not too bad. However,inspection of a histogram of the precisionof these valuesshows that see only a rather small portion meetspresentday requirements, figure 2.3.
Figure 2.2
One way to get gravity information in oceanareasis to convert heights,obtained from satellite altimetry, into mean gravity anomalies. There exist, however, two principal objections against doing this. First, the altimetric heights, even after
13
2.
Satellite gradiontetry:
6000
= z
sooo
4000 3000 2000
rooo o
16 20 24 28 M GALS
Figure 2.3
Histograrnshowing the precision of the 1"x 1o rnean anomalies frorn figure 2.2 (Rapp, 1977).
subtraction of our best models of the ocean topography,will not coincide with the geoid. Hencesystematicerrors will affectthe derived gravity anomalies.Thesetypes of "gravity anomalies"should certainly never be used for geoid determination in the context of ocean studies. Secondly,the conversionmethod can theoretically not be confined to ocean areas and is intrinsically unstable. Also, becauseof this, the actually applied numerical methods will introduce uncontrolledbiases.However,for some purposes,e.g. in geophysics, the gravity anomaliesdetermined in this manner are useful. Neither with terrestrial (including shipborne)nor with altimetric derived gravity anomaliesa global coverage can be achievedin the foreseeable future. Global gravity information comes from the analysis of satellite orbits and is expressedin sets of sphericalharmonic coefficients currently up to degreeand order 30 to 70. Thesesets are called geopotentialmodels. They are derived at a few computing centersin the world in a complicated estimation processconsistingof severalphases. In essence, the tracking data (laser,microwave,etc.) from a Iarge number of stations to a large number of satellitesis analyzed. Their orbits are determinedand combinedin a least squaresadjustment to yield e.g. sphericalharmonic coefficients the gravity field. of The separability of the individual coefficientsand their precision depends largely on a good distribution of the employed satellitesin terms of orbit characteristics (inclination, altitude, eccentricity). The error standard deviation per degreeof some of the recent geopotentialmodels is shown in figure 2.4. The spatial resolution is still limited. For some models a signal to noise ratio of one is reachednear degree 50. We conclude that our current knowledge of the gravity field is far from reach
14
Applications
*
h
o ci
$ s o
tr a
H C U tr Fr h q) q)
10
degree l
models. geopotential ol deuiation somerecent Figure 2.4 Error stand,ard, aimed for by gradiometry. ing the level of precision, resolution and completeness one needssuch good knowledgeof the gravity field. This leads to the question why Examples in four fields of applicationsserYeas illustration' Example One: Geodesy and levelled height Levelled heights are used in geodesyfor mapping, civil constructions, monitoring of land subsidence,control of tide gauges,etc. The processof levelling is very time consuming and therefore expensive. In recent years satellite positioning by GPS (see next section) became available. GPS measurementsdeliver meanwhile relative positions betweenpoints at a 107 to 1O8 level (dependingon the baseline AX, LY, LZ betweentwo points in the length). The cartesiancoordinatedifferences in global GPS systemcan be convertedinto differences geodeticallatitude, longitude height differencesconventional,referring in a and height. Unfortunately are the Ah, of this kind could, purely geometricalsenseto an ellipsoid. A height difference, however, be converted into a quasilevelledheight differenceAI/ were the geoid height Al[ available: L,h: AN + All . The geometric situation is sketchedin figure 2.5. The precisecomputation of AN requirespreciseknowledgeof the earth's gravity field. The gravity field as obtained by satellite gradiometry would suffice to determine the "absolute" geoid with a precision of :E 35 cm or relative with t 52 cm, t 18 cm or * 2 cm over a distance
15
2.
Satellite gradiontetry:
Figure
2.5
Relation between relerence surlace (ellipsoid), geoid and topography (ESA, 1991).
of respectively 100, 10 and 1 km. In combination with local gravity information, as available in NorthAmerica, Australia or Europe, cmprecision is feasible and levelling can in many instances be substituted by a combination of GPS anci a precise geoid.
Exarnple Two: Precise satellite orbits The main obstacleon the way to orbit determination with cmprecision is the inaccuracy in our knowledgeof the gravity field. Of coursealso nongravitational effects form a limitation, in particular for bulky spacecraft, but could be eliminated in principle by a dragfree setup or adequateparametrization. A precisegravity field in combination with accurate and densetracking  would result in orbit accuracies of a few cm. This would not only significantly increasethe value of ocean, ice and land altimetry, but be useful for geokinematicalapplicationstoo. Exarnple Three: Solid earth physics The applications of an improved gravity field in solid earth physicswere described in a number of reports. It is referred to (SESAME, 1986), (NASA, 1987) and (Lambeck, 1990). In geophysics is useful to distinguishbetweenstudiesconcerned it and lithosphere. with core or core/mantle boundary, mantle convection processes Satellite altimetry in combination with bathymetric data brought a much improved understanding of the oceanic lithosphere. Surprisingly it is the continental lithosphere,as well as the polar regions,where currently better insight is desirable. Due to the inverse character of the problem gravity alone shall never suffice for a determination of the density structure of the earth's interior. However, precise gravity and topographic data in combination with regional and global seismic tomography would drastically reduce the uncertainty range of possible solul98a) (Woodhouse and Dziewonski,1984) tions, seee.B.(Hager, 1983)(Dziewonski,
16
2.3. Aristoteles
or (Spakman, 988). 1 Exarnple Four: Oceanography and clirnate change The area where the need for an improved gravity field probably is most pressing is oceanography. Like terrain topography is defined as the deviation .Il of the topographic relief from an equipotential surface (geoid) the sea surface topography (SST) is the deviation of the actual oceansurfacefrom the geoid. The oceansurface with no external forcessuch as tides, winds, storms,etc. would.be level,coinciding with an equipotential surface. Hence any deviation from a level surfacecan directly be attributed to ocean dynamics. Satellite altimetry provides sea surface heights h relative to a chosen referencesurface (ellipsoid). If a precise gravity field were availableprecisegeoid heights N could be determined and from the relationship h : N + H the SST height 11 could be directly derived. Leaving aside the somewhat more complicated issueof the wind driven Ekman layer (Wunsch, 1992),the slope in the geoid can be translated into surface ocean circulation. In other words, altimetry in combination with a known geoid would let us see the ocean surface flow. For oceanographythis would be a milestone. However,there is more to it. To the presentday, ocean circulation is studied on basis of hydrographic measurements (salinity, temperature, pressure,depth) along selectedocean sections. In order to derive circulation from the data  employing equations of motion  an assumption on a level of motion at some depth has to be introduced (Pond and Pickard, 1983). As one knows, such a layer is purely hypothetical and consequentlysystematic errors enter the calculations. Altimetry in combination with the geoid removesthis uncertainty. It providesthe neededboundary constraint for the equationsof motion. Thus the gravity field information also permits to seethe deep ocean in the proper way. Circulation is the key to many ocean transport processes) whether it is heat transport, transport of plankton or polluted water. Heat exchangebetween ocean and atmosphereis probably the main uncertainty in a better understandingof climate changes. Oceans are considereda main buffer of atmosphereheat but the correct mechanismof heat transport in the oceansand exchangeof heat between water and air is not well understood. With theseexampleswe tried to illustrate the need for an improved gravity field knowledgethroughout earth sciences. Satellitegradiometry could have a substantial impact in this respect.
2.3
Aristoteles
At the time of writing the European SpaceAgency (ESA) is planning a solid earth mission to be launchedin the late nineties. The main purposeof this project (called
17
2.
Satellite gradiometry:
Aristoteles) will be the improvement of our knowledgeof the earth's gravitational A field through gradiometermeasurements. secondobjectivewill be a global analysis of the nearearth magnetic field since, besides gravity gradiometry, the satellite will be equipped with a magnetometer. Furthermore, our knowledge of various phenomenasuch as earth rotation and pole irregularitiescan be enriched geophysical from spaceto which extent a of precise point positioning measurements by means GPS receiverwill be on board the satellite(Dornier, 1989). The Global PositioningSystem (GPS) consistsof 24 (21 * 3 spare) satellites, orbiting the earth at an altitude of 20,240km with an orbital period of 12 h. They are evenly distributed over 6 orbit planes with an inc.linationof 55o. The satellites transmit three different kinds of pseudorandomnoise codes at two carrier frequencies. With a GPS receiveron board, tracking at any time four of the 24 GPS satelIites, the position of the receiverantenna can be determinedby pseudoranging, relative to a network of and if, in addition, differential carrier phase measurements ground stations are applied, orbit reconstitution at centimeter level is feasible. A full tensor gradiometerwould consistof eight ultra sensitivethreeaxis acceledue to the of rometersplacedon the corners a cube (Balmino et al., 1985).However, heavy affectionof the satellite by air drag in the alongtrack direction (accelerations which are more than l0 times as large as the differentialgravitational accelerations) the non dragfree Aristoteles configuration will consist of only four accelerometers mounted in the corners of a plate perpendicular to the satellite's velocity vector will (Dornier, 1989).This instrumentis calledGRADIO. The four accelerometers be very sensitivein radial and crosstrackdirection whereasthe alongtrack component will be measuredwith lower accuracy. which will be used consist of a cubic (or parallelepipedic) The accelerometers proofmass which is kept in a fixed position electrostaticallyby meansof electrodes to arranged around it (GRADIO, 1989). The force necessary maintain the proofis mass at the center of the accelerometer measuredby means of the output of the of electrodes.The differential measurement theseforcesbetweentwo accelerometers is a measurefor the gravitationalgradient(seesection2.1). Perturbing forcesresulting from air drag, solar radiation pressure,etc., are elimprocessby differencingthe accelerometer outputs (as we inated in the measurement have seenin section2.1, socalledcommon mode rejection).However,this requires a a very good linearity of the accelerometers, low coupling between the sensitivity axes and a good matching of the scale factors and of the alignments of the instrument axes (Dornier, 1989). To this extent a calibration device is provided, situated in the center of the gradiometerplate. The calibration device is furthermore needed outputs to proper gravitational units. to scale the accelerometer a The material out of which the satelliteis built up alsocauses gravitational accegradiometer(socalledself gravitation). Whereas Ieration which is measuredby the solar panels) is the signal coming from the rigid satellite parts (housing,electronics, constant and can be computed relatively simple,the part coming from the fuel in the it hydrazine tanks causesa problem, because has time varying components.Firstly,
t8
2.3. Aristoteles
there is the fuel consumption necessary maintain the satellite's to low orbit. It causes a gradual decrease the mass and thereforea changein the self gravitation. This of problem is accountedfor by a symmetricalorganizationand emptying of the fuel tanks, which furthermorehave a spherical shapeand are equippedwith elastomeric bladders to keep the fuel centered. In this casethe changein self gravitation can be computed easily. Secondly,and a more critical problem, is the sloshingof the fuel in the tanks. The tanks should be of such a size as to make sure that sloshingmode is at a frequency out of the measurement band of the GRADIO instrument (Dornier, ibid.). This measurement bandwidth, for which an 0.01 EIJH, white noise error spectrum is to be expected,is between0.005  0.125 Hz allowing only a good recoveryof the potential coefficientsabove degree 27. In (Schrama, 1990) it is shown that the onboard GPS receivercan be used in combinationwith GRADIO to obtain long wavelength(below degree27) gravrtationalinformation from the Aristoteles mission. However, combinationof gradiometric and GPS measurements not be considered will in this thesis. In table 2.1 an overviewof error sources satellitegradiometry is given. For in a detailedexplanationand discussion theseerrorsseee.g. (Touboul et al., 1991), of (Paik and Richard, 1986),(Schrama,1990)and (Rummel, 1989b),in which further references can be found. The complete Aristotelesmission will last for more than four years, of which the gravity phase,in which the gradiometermeasurements with the GRADIO instrument will be carried out, has a duration of about 6 months. This 6 months gravity mission will be performed at a low altitude of 200 km, the orbit being a dawndusk orbit with inclination95'.3. Because the relativelyhigh air density at of 200 km, drag will causean altitude decay of 400 m per revolution,so that regular orbit maintenancemaneuversare necessary keep the satellite in a band of t 3 km to around the nominal altitude of 200 km. After this 6 months period the satellite is planned to fly for another two weeksin a near polar orbit (inclination92'.3). This decreases the influenceof the relatively large polar gaps of the 6 months period.
19
Table 2.1
i E r r o r s o u r c e sn s a t e l l i t eq r a d i o m e t r u I misorientation of accelerometer
instrument i geometrical
other
scale factor mismatch between sensitivity axes of accelerometer nonorthogonality of the sensitivity axes of accelerometer in misalignment of accelerometers gradiometer frame scaleerrors of instrument axes I displacementof instrument from center of mass of the satellite nonlinearity of accelerometers incorrectcalibration nonmechanical instrument noise. due to temperature and electromagnetic fluctuations of the environment finite baselineof the instrument
unmodelled rotations orientation surface forces due to air drag, solar radiation p r e s s u r e ,e t c . environmental disturbances such as vibrations, electromagnetic and thermal irregularities
satellite
self gravitation
time varying components due to fuel consumption and sloshing resonating masses (e.g. solar arrays, antennas) reaction wheels noise
geodetic
anomalies
separationof gravitational and rotational parts integrated observables orbit errors orientationunknowns
lrnearlzatton error simplifications (truncation, symmetries)
20
3 l
The gradient tensor and its
series representation tn
The firstorder derivatives (gradient) of bhe gravitational potenbial V with respect If, for example, the system ct is a cartesian to arbitrary coordinates x' areV, : #. l,2,3togetherformtheaccec o o r d i n a t e s y s t e m , t h e t h r e e c o m p o n e n t s oT r f o r r : f leration vector or gravitational vector. The secondorder derivatives of the potential areVrr. With respect to a cartesian coordinate system (e.g.some instrument system with respect to which measurements are taken) they are the firstorder derivatives of the acceleration vector, i.e. the gradients of the gravitation. This is the reason we call the V,, the graoity gradients (or gravitational gradientsl). It also is the reason for using the word gradiornetry: "measurement of gravity gradients". In the sense of tensor analysis the quantity 7r, is a tensor. It is sometimes called gravity tensor. The two indices r,s result from a differentiation process of the scalar quantity V, and in arbitrary coordinates (not necessarily linear coordinates) this should be the process of covariant differentiation, see appendix B. Only in cartesian coordinates the covariant derivative equals the usual partial derivative, because in those coordinates the Christoffel symbols vanish. A few general remarks about V,, can be made here. First, since the gravitational potential function V is harmonic outside all masses, it fulfils Laplace's equation, which in our arbitrary coordinates r' is written as:
g"vr" : o
(3.1)
where g" is the contravariant (or associated) metric tensor. In cartesian coordinates this equation is written out in the well known formVrrlVyu lV"" : 0. We seethat
rGravity : gravitation f centrifugal acceleration.In many texts the terms gravity and gravitation are not well distinguished. If the differenceis essential,it wili be clear from the context.
21
the diagonal elements of the gradient tensor in cartesian coordinates are linearly dependent, leaving only two out of three independent components' Furthermore, since the gravitational field is irrotational, it satisfies:
e"tVg = g'
(3.2)
which implies that V* is symmetric: {%r} : {V,,}, leaving only three out of six independent offdiagonal components (e"t is the three dimensional permutation 1 symbol, which equals I if the value of the indicesconstitute an even permutation, if the permutation is odd and 0 in other cases. 0' is the nulltensor, but one usually writes O, in which case,however, the index balance no longer holds.). According to both properties (eq. 3.1 and 3.2), out of the nine components of the gradient tensor o n l y f i v e i n d e p e n d e n tc o m p o n e n t s r e m a i n . A l a s t g e n e r a l r e m a r k c o n c e r n st h e l e n s o r c h a r a c t e r o f t h e g r a d i e n t t e n s o r . B e i n g tR, u' a tensor, V* transforms to some other coordinate system, ".g.
d t ' d t 
rl YRS
ox" or"
n,
"* .
lr q Yrs
/a "\
If one knows the coordinate transformation equations t' : r'(rR) from which the i" computed these equations can easily be evaluated. transformation matrix # This will be done for severalcoordinate systems in section 3.1. We have seen in the previous chapter that, in principle, it is possible to measure the secondorder derivatives of the earth's gravitational potential using a gradiometer in an earth orbiting satellite. Actually, the gradiometer will deliver the measurements in a cartesian coordinate system connected to the instrument. The orientation of this instrument system, or satellite system, will in general differ from that of a local orbital system2 due to changes in the satellite's attitude. The local orbital system, denoted by ,' , will be orientated with the caxis along track, the yaxis cross track and the zaxis outwards. We will at this stage assume, however, that either the differences between the instrument system and the local orbital system are negligible, or that it is possible to transform the measurements from the instrument system to the local orbital system, using an equation of the type 3.3. In this particular case, such transformation will consist of a simple rotation matrix between two cartesian coordinate systems having the same origin but a different orientation, and we will assumeit to be known. Thus we consider it possible to obtain from the gradiometer bhe gravitational gradients V;1 in the local orbital system cr. Furthermore, the orientation of this local orbibal system will, due to the inclination of the satellite's orbit, in general deviate from the orientation of a (commonly used) local, northoriented coordinate system ri' with the r'axis directed north, the y'axis west and the z'axis radially outwards. However, we will also assume that it is possible to transform the components of %i from the local orbital system
2For the definition of this system and other coordinate systemswhich will be usedseeappendix A.
22
to the local, northoriented systemby meansof a known transformation, again using an equationof the type 3.3, in this casewritten as:
Ir
vitjt =
)xt
0rl
,,
ac
aivij
(3 4)
where the
is a known rotation matrix, cf. (Rummel and Colombo, 1985). If, ffi in an ideal case, all nine components of the symmetric gradienl tensor are measured, it is, in principle, possible to solve the attitude of the instrument from the measurements, at least if some additional information from star trackers is available (Rummel, 1985c). For the moment we will assume here that we are given the gravity gradients in either a local orbital system r' or a local northoriented system r' . From the available gradient measurements we like to obtain information about the earth's gravitational field. This information is usually given in terms of a set of potential coefficients, cf. section 2.2. One therefore needs a relation between these spectral coefficients, often denoted e1^ and S;, and the observabions l{i. This relationship is obtained by expanding the gravitational potential V into a seriesof spherical harmonics as function of the geocentric polar coordinates vA : (r,0,\). This will be done in section 3.2. In this section also a series expansion of the p o t e n t i a l a s f u n c t i o n o f t h e o r b i t a l c o o r d i n a t e sr o : ( r , , r r , a o ) o r r o ' : ( r , r o , 1 ) ( s e e appendix A for their definition) will be given, which appears to be better suitable for problems involving satellite observations. Given V as function of e.E. ro, one may compute the secondorder derivatives Voo (using couariant differentiation in this case because the ro are curvilinear coordinates). By means of an equation of the type 3.3, in this case
,4r
't 
a r b a r b, ,
dri Ox,j'oo
(3.5)
one obtains the desiredrelationship betweenpotential coefficients, containedin the Vo6, and measurements V;i. In order to gain some insight in the signals to be expectedfrom a gradiometer mission,the last sectionof this chapter,section3.3,dealswith the synthesis problem, i.e. how to compute from some apriori given set of potential coefficientsa set of (simulated) gradiometer data. This section furthermore shows a simple analysis strategy, i.e. the inverseproblem of recoveringthe potential coefficients again from the simulated gradient data.
3.1
Potential
derivatives
in different
coordinate
systems
In the following section3.2 we will be given two seriesexpansions the gravitational of potential, one as function ofthe geocentricpolar coordinatesrA and one as function ofthe orbital coordinates or xo' (respectively tro equations 3.14and 3.17). Both series
23
are given in terms of the same potential coefficient set er*,Srrn. Taking two times the covariant derivative of these series yields corresponding series expansions for the gravitational gradients in curvilinear coordinates, respectivelyVa6,V66 or Vot6t. They have to be related to the observed gradients, expressed in local coordinates ri or ri', respectively V;i and %,j,. So what we need are transformation equations for secondorder potential derivatives. In this section we derive these transformations, which are all of the general type
r, vRS: ,, 6rn 6rsv" 0r' )rt
'
(3.6)
where in our case we take for the ot coordinates either tA, ro or ro' and for the *e take ci o, ,i' For the evaluation of the V," we need the metrical tensor and "R the Christoffel symbols for the respective coordinate systems. They are given in appendix A. In section 3.1.1 the algorithm to derive the gradient lransformation e q u a t i o n s i s e x p l a i n e d . T h e s u b s e q u e n ts e c t i o n 3 . 1 . 2 g i v e s a l i s t i n g o f s e v e r a l e x p l i c i t transformations. Compare e.g. (Reed, 1973) or (Tscherning, 1976) were some of these transformation formulae are also derived.
3.1.1
Algorithrn
The algorithm for computing the transformation equations for the secondorder potential derivatives may serve as an example for the use of tensor analysis and index notation (seee.g. sections 8.2 and 8.3). For the definition of the coordinate systems used in this section, we refer to appendix A. In the same apPendix the elements of the metrical tensor and the Christoffel symbols for those coordinate systems are listed. We assume we have the gravitational potential l/ given as function of the coordinates co. We could equally well assume here thatV is given as a function of or rA, but we take ro as example. Since the equations are all tensor rA,ro',rA',"u equations they also hold for all other coordinate systems. Given V , we may compute by partial differentiation the firstorder derivatives tr{o, which equal the covariant  Vp : V;o. derivatives V' and which will be simply denoted by Vo, so Vo The secondorder derivatives have to be computed by covariant differentiation, since partial differentiation of a tensor with rank ) I in general does not lead to a tensor. This yields: Vo,b:Vo,6lf1,V" w h i c h w e w i l l s i m p l y d e n o t e l > ! V o u , s o V o 6 V ; a b: V o 1 . W e s e e t h a t , i n o r d e r t o e v a l u a t e V o 6 ,w e n e e d t h e C h r i s t o f f e l s y m b o l s I l i . T h e y a r e c o m p u t e d u s i n g
l'u :
;s'"(9od,b
I gud,a 9 * , d ) .
24
arl arJ
9ab: 6 r a6 r t 9 t t . Thereby g;7 equalsthe unit matrix and the transformation matrix # i, computed from the coordinate transformation s/ : rI (r"). Thus the coordinate transformation has to be known. In the present example this transformation equals equation A.6. The associatedmetric tensor g"D is computed as the inverse of 9o6. If it is hard to invert gab o\e can try to compute first the inversetransformation matrix ffi and calculate goDfrom ^ o b_ E r o 0 r o ^ r *
rtSrxc
in which g/J i" again the unit matrix. The V"6 can now be computed. Since Vo6 is a tensor, its transformation equation to another coordinate system has the form tvr J K : 0ro 6rb ,, 6 1 16 r x ' " t '
We already know ffi, so with this transformation equation we easily find the secondorder potential derivatives with respect to geocentriccartesian coordinates as function of the first (!) and secondorder potential derivatives with respect to the 1orbital coordinates ro. tansformation to a local cartesian coordinate system, for example the local orbital system r', follows analogously: tvr jk: where the transformation matrix 0r" Erb,, ail 6r.kv"b,
# Er"
art:
aC art
in which t# ," already known and fi has to be computed from the coordinate transformation zr : rI (ri). We therefore have to specify the latter transformation, or directly the matrix ffi, which in the presentexampleequalsthe matrix in equation A.3. Moreover, the transformation equationsfor the firstorder derivativesare simply
And l ,
,J:
(3.7) (3.8)
25
3.
If, during the computationsabove (either numericallyor symbolically),use is made of preprogrammedsubroutinesfor matrix manipulations,care has to be taken with the use of the transpose,inverse'andoriginal form of the matrices, especially of the transformation matric"" # etc. Strictly following the conventionsfrom section 8.1.3, this may not lead to any problems,but it is safer to convert the equations directly into nestedD0loops,the inner loops representingthe summations over the dummy indices and the outer loops representing the repetition of the computations for all the valuesof the free indices. Both methods will be illustrated below. Let us denoteall the transformation matriceswith the kernelletter X, i.e. XIo: : #, X', #, etc. Furthermore we choosethe following transition from index to matrix notation:
Vot
vo V7
V;
I a
vtt
V;j Xro X,, Xro
9ab
W U X
I L
ro rI x,'
gIJ
I
r I u
T G
The transformation equations from this section then become, in order of appearance:
9ab :
XoIgttXJt
: XTIX
g"b : x"tgJKxi
VJK : Vi* Xoj V1 vj : : : XfVo6Xb6 Xi"Vo6Xb1, X"rXti XJoVo Xiovo
TL : a
: XTg
'r
 Trg
with the transformation matrices defined through the coordinate transformation equations:
26
,I:rt(ro) ,I :
a
* +
l:Xr I  Ru uTr
,t (ri')
; / ^\
r'=x.t(ro)
As an example of how a tensor equation can easily be converted into programmable code, consider the equation for the Chrisboffel symbols l'u = * 9 b d , o 9 a b , a ') ,0""(9od,b r a n d d g ( a , b , c ) r e p r e s e n t i n ge s p e c t i v e l y I
g a m m a ( a , b , c )= O do d = 1,3 + g a m n a ( a , b , c )= g a m m a ( a , b , c ) O . 5 * 8 i ( c , d ) x (dg(a,d,b) + dg(b,d,a)  dg(a,b,d)) end do end do end do end do The algorithm sketched in this section is shown in table 3.1 as flow chart.
27
Table 3.1
input:
(1) V =V(r")
(2) (3) ,I : ,t : ,I (ro) ,I (ri)
{6"}
from (2) from (5) from (a), (5) lrom (7) from (7)
(10) qi
(11) vo:v,o (12) V1  XfVo
28
3.L.2
lYansforrnation
We assumethe gravitational potential V to be given as function of one of the foll o w i n g c o o r d i n a t es e t s : , A : ( r , 0 , \ ) , ' r " : ( r , r o , r " ) , t o ' : ( r , r o r I ) o r x , A ' : (r,Q,ro) (seeappendix A). Then we can compute the first and secondorder potential derivativeswith respectto thesecoordinates,respectivelyVe,Vo,Vo',V.a., and VABrVob,Vo,6,rV4,g,. What we need are the potential derivatives with respectto either the local northoriented coordinate system r'' or the local orbital coordinate system c', i.e. Vyit or %;. They are obtained from the derivatives with respect to curvilinear coordinatesby meansof transformationequationsof the type 3.7 and 3.6, and are given in the boxes in the remaining of this section. Notation is abbreviated For reference,also the to what is commonly used: V, means # : ffi, "t.. transformations the derivatives ellipsoidal of in and geodetic coordinates are given. V : LV,, * ' , ' r cosqt
tvr .
v
tt
vi 
0r ; 17  e Ai' v A '
, 
l t ta v Y
t / v z
 t l v r
ui: ff!fflro,",
tr / X  ! r , r r  t u nq Q r Q r, ' . t X
r r
t/
r
cosI s i n/
n
t n
t v U , , a , ,
,,
r2 cos$
V.r: 
V6.^
r cos'Q
Vu,,
r2 cos$
(3.e)
vuu:lr,*)16
V u " : T/ Tl v zz v rr
r,,
1,,
r
t V 6 +  V r o
If we put 6 :0
L,,
t
1,,
, ' Qu,,
t/ 
vrt
(3.10)
\/ z z .
Equations 3.10 will be used in chapter 4 They are obviously only uaLidon the equator of the coordinatesystemwhich, in o u r c a s e )i s t h e o r b i t of the satellite.
29
: v,1 !;Sv",o,
Vrr:
Vr, :
l,rr. J . 
V"r:Vr+ r r
rvl xu w
]V.'r,, 1
(cos.lsin @oV.,,  cos,I costioVr,,r,, sln I cos a./o  sin uoVr,,)  cosuoVr,,. l :Vr.., r r l l ,Vn l (3'12) l
Vr" :
: Vuu
;r,
F*%I
* + c o s 2I c o s u o V r . r , ,  2 c o s I c o s w o V r , , r , , c o s o r r V r , , c r , c o s l s i n c ^ . , r V r . ) Vu" = V"" : Vt sln I cos aro (r cos lVrr,,  cos IVr,,  rVrr, * Vr,,)
30
vz  V h
T/ viE
,r l
nalavn+ +MYv"&
V.l (h + U)(h + lr) cos e^ e l r , '  1 , h+ Mvhe @+ uy'v 1 1 I
r t r ' \ o
1,f
1,r viz, 
vss:
t' l/l Z  , 
tan <p
V,
(h + N) cos,p
Vnx
v. v,
vz
tl vij o2," or"1/ Az' art, ab
^,.^,,
tv,  i :
uuz,, ,  ,, Lor"rVVAB
_l
E2sin cosB,, B ,p Ln
tr
vz, =
r, lz z r rr , v
r' lu
Lru"
u
q
,, *",
,,
Ovxx
I
Lz
z 
,,
^ r u A
Lu'cosp
^ r A
Lcosp
rr : Vzz
31
3,
The gradient tensor and its series representation in different coordinate systertls
v,   Lva
vu'  j'v^
V"' : V,
v;,i, ##r"u
Vr,",:V, lVoo r r '
,v x t u t : cosd rl I ,, ^ rr r'r"0'0^ 1 2s . t n z ' 0
(3.13)
3.2 3.2.L
The gravitational potential V, being a solution of Laplace's equation, is expanded into a series of spherical harmonics as function of the geocentric polar coordinates ,A : (r,d,)) as: GM2 v ( r , 0 , ) ) + t ( : ) tt a where GM R er*,Srr,I,m Plr (cos0) r,0,\ gravitational constant times mass of the earth r e f e r e n c er a d i u s n o r m a l i z e d p o t e n t i a l c o e f f i c i e n t so f d e g r e e / a n d o r d e r m degree and order normalized Legendre functions g e o c e n t r i cp o l a r c o o r d i n a t e s( d i s c o  l a t i t u d e ) . /P\l.IL. () f ^ L l e , * r o . m ) , * S 1  ^ sn ) P 1 ^ ( c o s d3 . 1 4 ) si , \r /
32
V (r,0,I) : t
tn=O
(3.r5)
I D rcf^Q,e) I t
Kf,(r,0)
l=m
(3.!.6)
Kfl(r,0)
KB^(r,0)
t:
vLm
H6(r,0)
pLm
e,
Hr,n(r,0): ut(r)Pt(cos d)
GM / R \ ' * ' / \ ut\r)= R \;/ We saw that, in order to compute the potential derivativesin the local northorientedsystem,;',*eneed the derivatives VaandVaB ofthe potentialwith respect to the polar coordinatesrA. Thesecan be derived from equation 3.14 or 3.15 by means of simple partial differentiation. The derivatives,computed in this way, can also be expressedin seriesanalogousto 3.15, with appropriate expressions for the factors Kh, KE^ and H1^. In table 3.2 one finds these factors for all the first and secondorder derivatives V with respectto r,d,.\. In this table we abbreviated of P!*: APt,"lA0 and Pll:02Pt^002. The local derivatives V;,i, are expressed in the sameway, the expressions Kfi, Kf^ and H1^now to be taken from table 3.3. for They are found by using equations3.13. Already at this point one might find out, by taking a glanceat this table, why satellite gradiometry is especiallysuitable for the determination of higher degrees and orders of the gravitationalpotential. On the one hand, one noticesthe factor ( R l 4 ' * t , w h i c h i s p r e s e n ti n a l l s i x g r a d i e n t s( s i n c el t : ut/r2 : (GMlRlrz). (Rlr)'*t).This factor causes attenuationeffectwith height for the gravitational an potential and related quantities. It means that, since Rf r is smaller than l, the power contents of the potential signal decreases with increasingaltitude above R. Due to the power of / * I this decrease stronger for higher degrees making them is /, harder to detect if one goesfurther away from the earth. The attenuation effect is inherent to all spaceborne gravity determinationtechniques.
33
3. The gradient tensor and its series representation in different coordinate systems
Table 3.2
H[)
Kh
H
KP,, H[2sh"
HI},SM
JPu,P,* u,pl^
mutPt^
9e,,"
IA\ _
@p"tPt,. @u,p!^
;utrtn
zn(l+t) 6
H i,,i,Cu.
e H,i'],),^
H,
Hl
(")) e.
n eafL
",pi,*
mulP!^
m2u1P1*
f!e,*
Hr'^t Sr*
On the other hand, in the case of the secondorderderivatives, this attenuation by effectis to someextent compensated multiplicationfactorslike (/+ l)(l +2),(l+ the power contentsfor higher degrees.This m, which tend to increase 2),(l+ 1) and makessatellite gradiometry (the techniqueof observinggravity gradientsat satellite altitude) an attractive candidate for gravitational recoverywith high precisionand resolution compared to other techniqueslike orbit determination.
Table 3.3 Potential deriuatiues witlt, respect to local northoriented nates (rt ,y', z'). lt : 3 coordi
n( )
"lm
Kh
Hii,.''cm
. tU I mY ) q u l f T a t V
KBt*
Hli.' 'sm
 , t\ _
 nl].''),^ e
' ' I m 9 I ' m
( + z)r,Pi*
0 11(cot Pl^  (/ + t + m2 sinz0)P,*) m ( + 2 ) s i n  10 1 1 P 1 ^ ( / + t ) ( /+ 2 ) r t P h
Hli''cm Hli,,''ct^
ts ( v ' t lm
q('' '') q,
s,^ ull)u')
v'r'
t l
t') 4.
ulm
 n[I')e,^
"
ni*'ulm
tz z  ^
g(z'z') 4, vlffl
Ln
34
3.2.2
Orbitalcoordinates
For the description of the motion of planets around the sun or satellitesaround the earth the orbital elementsa,e,I,dl,w,M can be employed,which are respectively the semimajor axis of the elliptical orbit, eccentricity,inclination, right ascension of the ascendingnode, argument of perigeeand mean anomaly (seefigure A.1). It is convenient to have also an expressionof the gravitational potential V in terms of these orbital elements. Upon keeping the harmonic coefficientse1* and 31* from eq. 3.14 as coefficients the potential expansion, in but changingfrom r,0,A to r,I,u,Q,M (we assume the orbit is (nearly)circular,so e E 0) we obtain:
vu"'\lal R ?v \ , ) t=
t^.f @
zprl+l
,l I,.: :
e t*
^
vrm
m=O p=0
t t F'*,(I)'
I
cos!,t1^o I
'
I st*
sint!1^o\ : "..
(3.17)
normalized inclination functions (l  Zp) eo I rnut, subscript tto" referring to "orbit" I I see (Schrama, 1989) J
u)o : e * M ee: 06
A derivation of eq. 3.17 can be found in (Kaula, 1966). In (Sneeuw, 1991a) the same expression is derived by means of a grouptheoretical approach which makes use of the socalled representation coefficients (also known as Wigner coefficients) . In such an approach it becomes obvious that it is, in fact, more fundamental, and for satellite purpbses also more convenient, to use in eq.3.17 the index k instead of p, with k: l2p. The index khas astepsize2and rangesfrom l to l. So switching from p to ft and changing the notation accordingly, we make the following replacementsin eq. 3.17:
al LP=o
Fn r(I) ,bt^p
F,r*(t)
,bk*: kuo I mu,
From ft : I 2p we seethat k and I alwayshavethe sameparity, a property we will uselater on. It appears convenient abbreviate 3.17(includingthe replacements to eq.
35
above) as
o o l l
v:
with
B lAn*(r,I)cos$p^* u*(r,/)sinr/;l
(3'18)
l c m l _  fi  Jr,oaa L
 ^ 1lzn:even
1lrn:even
Furthermore, we mention the possibility to rearrangethe summation order l,m,k to k,m,l, as a result of which we may rewrite eq. 3.18 (truncated at some maximum degree.L) in the following manner, cf. (Schrama,1989):
"
: f
t
* B1*1,sin{p,n) lA6lacosrb**
lAtnkcosrbr,^*81*psinr!1,*)
t
L L
:
where
k= L m=O
Dllr^cos$1,**
\ 
B**sintl'1,*l
(3.19)
Ar,*I l :
Bo* )
,=,7,np11 "t*r
.1 z
\l A , * o
36
3.2.
and H1^p  utFtk^. In the expression abovelmin: mar(lkl,m) + 5 where 6: O if & and rnar(lkl, rn) have the same parity and 6 : I if they have opposite parity, cf. (Schrama, ibid.). The Ap,n and Bp* coefficientsare the socalled lumped coefficients, (Schrama, see ibid.) or (Wagner,1983). From eq. 3.18 we may compute the partial derivativesof V with respectto the coordinatesr,u)o)u)e,1, resultingin differentexpressions the quantitiesHunk,Auo* for and B1^p. Those quantities are listed in table 3.4 for all the first and secondorder partial derivatives V with respectto r,0)o,or,1. FollowingappendixA.l we either of considerthe set (r,ro,.I) or the set (r, uro,ar) so the mixed derivativewith respect : to.I and c,.r, not listed in table 3.4. In this table we abbreviated Fro,o' AFf*lAI is and F[n" : A2Ftk^/A12. tansformation to loca] derivatives is, fot ii,ir situaiion, discussed section4.2. in Table 3.4 Potential deriaatiues w.r.t. r,o)o,ti",I differentiation
w r f
H[)r
At^*
Bt^*
T"tFf*
utFf^
kulFrk^
mul,f !'n
r
u)o
QJe
g'^H[2r g,*H!2r
ahH[:""k)
B,*Hli."o)
,,(rr I almntmi
B,*n!1::)",*n[Tl
g,*Hlil.) p,^n[i]")
rr II
0)oQo @eQe
",*u!l,l) g'*n!!,!)
"t^nl'is "t^u[i.f) B,^ullii') o,*H[li) g,*Hli,!.) g,,Hllif) o,*H[li[)
rI
TQo TQe QoQe
Tu,F!^'
ltculFl*
_t9!"tF,1,, p,^H[:;)
mku1F!,o
Iuo
kulF!^'
37
3.3
In the previous sections we have seen that gravity gradients are measurable, in principle, in an earth orbiting satellite by means of differential accelerometry(section 2.1). We have seen how this technique will be realized in practice and what kinds of problems thereby will occur (section 2.3). We have seen how to express the observedgradients in severalcoordinate systems(section3.1), how they are related to the desired potential coefficients(section 3.2) and for which applications as we need these potential coefficients, they are derived from satellite gradiometry in (section2.2). But what should we expect from gradiometric measurements terms of of magnitude and spatial or spectral characterisbics the signal? the easiestway to get at least a first, rough estimate of the size of the Probably gravity gradients is to take for the gravitational potential V the central term W spherical mass distribution. ln only, representingthe potential of a homogeneous of a sphericalharmonic expansion(eq. 3.14) this meansthat the only nonzero terms potential coefficientir Coo : 1. Then we can easily compute the gravity gradientsin Sincein this caseV is only a function a local north*oriented coordinate system "i'. partial derivativeswith respect to 0 and ,\ are of r all the first and secondorder of for zero. Using eq. 3.13 this leaves the elements the gradienttensor:
1
GM
a
0 1
'l
0 l
, )
v;,i,
0 rr v
0 0
0 V,,
At a satellite altitude of 200 km we find in this case for the diagonal elements Vr,r,  Vr,o,: 1V",", av1400 E. spherical mass. In terms of In reality the earth is not a perfectly homogeneous exists. the seriesexpansion3.14 an infinite number of nonzero potential coefficients The potential V and the gravity gradientsare not only a function of r but also of d and ). A global computation of the gravity gradient signal should thereforeinclude a full evaluation of the series expansion. Whereas in this equation (and also in eq. 3.17) the summation over / rangesfrom zero to infinity, one has to choosesome finite maximum degree L ) 0 when doing practical computations. Such evaluation can be done in arbitrary points of which the geocentricpolar coordinates (r,0, )) are specified. If we carry out such a computation in individual points, it is called single point computation. Very often, however,we like to compute the gravity gradients in the nodes of an equiangular world wide grid located on a sphere with radius r. Such a computation, based on the evaluation of eq. 3.14, is called sphericalharmonic synthesis (Colombo, 1981). Besidesthe polar or orbital coordinatesof the (grid) points, grathe availabilityof someknown set of (usingeq.3.14 or 3.17)requires dient synrhesis potential coefficients.Nowadaysseveralpotential coefficientsets exist, for example GEMT2, completeup to degreeand order L : 36 (Marsh et al., 1989),OSU180
38
3.3. SyntJresis ana]vsis and with .L : 180 (R.pp, 1981) and OSU86F with tr : 360 (R.pp and Cruz, 1986). The first one is derived purely from orbit analysisof a variety of satellites,the second included altimeter data from Seasat and terrestrial data (gravity anomalies), whereasthe third model was computed from altimeter data and land measurements including geophysicallypredicted gravity anomalies. A maximum degree tr of 180 wavelengthof 2o thus giving the possibility of correspondsto a smallest recoverable gravitational features in a grid of 1o x 1' blocks (half of the smallest waveshowing length or the resolution).In the sameway data in 0.5ox 0.5'blocks corresponds to ,L : 360. is So if some known set of potential coefficients available,the synthesisproblem may give us some insight into the signalsto be expectedfrom a gradiometermission by computinga set of gravity gradients. But the reversereasoningis also of interest to us: given the values of the gradients in a set of points, regular or irregular distributed over the whole earth, compute from them the potential coefficients. If one uses an equation of the type 3.14 this problern is called sphericalharmonic analysis(Colombo, 1981). The set of gradientscould be thought of as obseroed during a real gradiometer mission,from which we like to recover unknown potential coefficients. In generalany functional of the gravitational potential may be generatedand/or analyzed by means of, respectively,the synthesisand analysis problems discussed above. Such potential functionals could include for instance the gravitational potential itself, its first and secondorderderivativesin any coordinate system, geoid deflections the vertical, etc. Let us write /(d,)) for of heights,gravity anomalies, such arbitrary function, defined on the surface of a sphere, and let us expand this function in a spherical harmonic seriesas
L I
f (e,\ : t
(3.21)
l=O rn=O
where the a1r, and bpn are the normalized spherical harmonic coefficients. Comparison with eq. 3.14 revealsthe same structure,so that, if we identify / with the (GMIR)(Rlrf+rey* gravitational potential V and ala and 61 with respectively a n d ( G M I R ) ( R l r ) t + r 3 1  , t h e t w o e x p r e s s i o n s c o i n c iI n a n a l o g y w i t h e q . 3 . 1 5 w e de. could write 3.21 as
L
f (0, ^) : t
rn=0
(3.22)
with
::l:r^
41... a; {
atm
I a'^
(3.23)
The harmonic coefficientso1, Dlm constitute the spectrumof / and are therelore bherebyreferring to the bheoryof Fourier analsometimescalled spectral coefficients, ysis, where one also usesthe term synthesfs the Pourier seriesor inverseFourier for
39
in representation differentcoordinatesystems 3. The gradienttensorand its series transform of respectivelyperiodic or aperiodic signalsand the term analgslsfor the resp. the Fourier transform, e.g. (Oppenheim et al., 1983). Fourier seriescoefficients Note, however, that eq. 3.21 does not correspond exactly to a twodimensional Fourier transform in the plane, cf. (Schwarz,1985). Flom the spectral coefficients q p t o n e c o m p u t e s h e ( q u a d r a t i c ) o w e rs p e c t r u m , ( l : 0 , . . . , . L ) a s
(3.24)
The c1 are also called degreevariances. The degreeorder uarianceschn are computed from the power spectrum as ct (3.25) ctrn:
2l+t'
the square root of which is referred to as r.m.s. value per coefficient per degree. Given the /(d, )) in a grid the spectral coefficientsa6n and 63 themselvescan be computed using the following numerical quadrature formula:
a,* I
r 'r
tkt
(3.26)
are where A; : A)A0sin?; and where "^" indicatesthe fact th at these coefficients valuesat the grid points. estimatesof the a6y and b1*. The I (0;, Ai) are the function Eq. 3.26 is a discrete approximation of
"r',:\:
I I
d^ d0
where o representsthe unit sphere. This formula is based on the orthogonality property of sphericalharmonics,seefor example (Heiskanenand Moritz, 1967). For the cosm) and sin rn) terms the orthogonality also appliesin the caseof a discrete summation, like the summation overy in eq. 3.26. For the fsummation over the P1 terms the orthogonality property holds only approximately so the d6n and 61will contain an error due to this approximation. The computation of potential functionalsin a grid, and the estimation of spectral coefficientsfrom such grid, involve some equivalent numerical techniques (duality betweensynthesisand analysis(Colombo,1981)). This can be seenby looking at the transform pair 3.22 and 3.26. Firstly, both problemsrequire the computation half of the problem of all Legendrefunctions up to degreeand order tr. Secondly, can be solvedby applying preprogrammedFFT routines. For sphericalharmonic osynthesis one first computesthe coefficients and b using eq.3.23. Afterwards the grid values are obtained by applying an inverseFFT with those coefficients (eq. 3.22). When evaluatingthe sphericalharmonic analysisequation 3.26, the jsummation can also be done by an FFT, after which the rsummation is carried out separately.
40
The value one choosesfor tr depends mainly on the specific application, but it may have a rather large impact on numerical computations. Typical numerical problems like underflow, numerical stability, but also problems with computation time and data storage may occur when simulating gravity gradients (or in general any potential functional) up to high maximum degree,for example L : 360, and at a large number of points. Therefore,we will pay here some attention to a few numerical aspectsof gradient synthesis,in particular to the recursivecomputation of Legendrefunctions, single point versusgrid computation and vectorization. Legendre functions In the processof evaluating a seriesexpansionof the potential and its derivatives up to high degree and order the computation of the Legendrefunctions plays an functions by meansof explicit important role. Sincethe computation of the Legendre formulas (for example (Heiskanenand Moritz, 1967), eq. l60 or 162) is too time consuming,one usesrecurrent relations. For the normalized Legendrefunctions we use the relations: Pt,t: h,t_r: Pt,*: with the starting values Po,0: 1 Pr,r: 16sind where
t _ tl 
fisin0P11,11 f2cos0P1_1,1_1 f s ( f e c o s 0 P t  r , ^ I s P t  2 , * )
f2: J2t + 
fs:
(tm)(t+rn) 14: t/2t  
l ,, r _ V( t  ^  l ) ( r + m  I ) _ l 2 _B
We may visualize these recurrent relations by means of arrows in an /, mscheme with degreeJ on the horizontal axis and order rn on the vertical axis (seefigure 3.1). Equation 3.27 is the diagonal recurrenceupwards on the main diagonal,eq. 3.28 the first step in horizontal direction for fixed order rn, arriving at the first suball diagonal and eq. 3.29 expresses the further steps in horizontal direction for increasingdegreel. There also exists a vertical downward recurrent relation (for fixed
41
eq.3.29 o r d
e
degree /
Figure 3.1
h ,^ z
with
t _ J 6 
fa
lrr^')#n,,
tl
a fzh,^l
m:2 rn*2
if
fz=
(l+rn)(lmi
This recurrence,however, is not used here since it becomesnumerically singular Furthermore, a downward recurrence is not for small 0 due to the factor fr. recommendablewhen, during the computational process,an underflow occurs on the main diagonal. An overview of all possiblerecurrencerelations can be found in (Ilk, 1983). Comparealso (Gerstl, 1980). with Poo= I the value of LhePa for increasing/ Starting the diagonal recurrence will decrease rapidly. The degreeat which a numericalunderflow will occur depends on the value of the maximum exponent of the used computer and on the latitude 0 (seefig. 3.2). The underflow results in a zero value for the P11.In that case all subsequent fi will also be zero. If a downward recurrence would have been used a much larger part of all the fl would be zero due to the underflow on the main cf. diagonal as in the caseof horizontal recurrences, (Koop and Stelpstra, 1989).
42
o o o
$
o ' o
o q) h (7)
;' o o ol o o o
P o
d (coIatitude)
Figure 3.2
DegreeI at which the P]1 (during the d.iagonalrecurrence)become srnallerthan 10s08 (and is thereforeset to zero on the CONVEX) as function of 0.
Whereas the value of the fi1 for increasingJ on the main diagonal decreases, the value of the Ptrn for fixed rn and increasing I during a horizontal recurrence slightly increases.An underflow may therefore causeinformation to be lost for high degrees.This is even more the casefor the secondorderderivativesof the Legendre functions P/l which can reach values up to 106. This problem of underflows can be dealt with by scaling techniques,as described in (Koop and Stelpstra, ibid.) for recursive Legendre function computations or in (Sneeuw,1991a) for recursive inclination function computations. derivativesof the Legendre The recurrent relationsfor the first and secondorder be derived from equations3.27 3.29 by meansof differentiationwith functions may respect to d, obtaining: 0 P!,r: /1(cos kt,tt + sindP/r,,r) Pl.tr: /2(sin 0Ptt,tr+ cosdP/t,,t) Pl'n: fs( fqsin0Pl1,,o* /n cos0P!r,*  fsP!r,*)
: P6,o o
Pl"' = rA cosd
43
P l ' , r  r : / 2 (  c o s0 P t  t , t  t 2 s i n0 P l  r , ,  r + c o s 0 P / 1 1 , 1  1 ) P{,* = f e( I + cos P1 1,*  2f a sin0P! r,,, * /a cos0P!\,^  f s P/r,,*) 0 with
P61o: o
Pt: r,/$sind.
Single point versus grid cornputation Gradient computation, using high degree potential models, is a time consuming process. Especially if one wishes to compute the gradients in a large number of points. Dependingon further applications,one may proceedin differentways. Either one evaluatesthe full spherical harmonic expansion3.14 in each point separately (single point computation) or one applies spherical harmonic synthesisusing FFT techniquesas describedabove,resulting in a grid of gradient values,having constant step size in d and ,\ direction (grid computation). Grid computation is especiallyuseful if one, for example,has to compute world wide representationsof gravity gradients using different coefficientsets as input, aiming at, for instance,comparisonof potential modelsor comparisonwith gradients obtained from other techniques.The main advantageof this method is of coursethe use of the FFT, decreasing computation time drastically. The step sizein ) direction dependson the degreeof the FFT, which in turn dependson the maximum degree ,L of the potential model which is used as input. For example, using a potential coefficientset with .L : 180 allowsto compute independentgradient valuesin points with A) : 1" (Nyquist frequency). The degreeof the (onedimensional)FFT in that case will be 360. As for the d direction, one may arbitrarily choose the value of d (Colombo, 1981) with the limitation that the smallest possiblewavelength to be present in the grid is determined by I (the maximum degree of the potential model), even if one chooses more d's than .L. For e.g. satellite applicationsthe grid computation may not be well suited, since the points along a satellite orbit where one wishesto compute the gradientsusually do not coincidewith the nodes of a grid. Using, however,a seriesexpansionof the type 3.1.9,we may still apply an FFT, in this case not along a parallel but along the satellite orbit. Another possibility is to compute a grid at satellite altitude and interpolate between the grid points. Both latter methods, however,fail if the satellite orbit is elliptical so that r is no longer constant. The interpolation method could be extended to three dimensjons(Schrama,1984)or (Wichiencharoen,1985), for which grids at different altitudes have to be computed first. A final possibility is to compute gradients at each point of the orbit individually, bringing us back to single point computation. The choice between single point and grid computation depends, however, not only on the specific application, but also on the availablecomputer hardwa^re and software. As for the machine characteristics.we mention the maximum value of
44
the exponent, optimization properties (Koop and Stelpstra, 1989), CPU speed and available memory. Concerning the software, accurate and fast interpolation and FFT routines are nowadayswidely available. Other aspectsare handling of underflows, speed, portability, memory usage, handling of various types of coordinates and potential models, accuracy of the input (coordinatesand coefficients) use of , single or double precision numbers, vectorization and parallelization capabilities, for etc. In (Balmino et al., 1990) severalsoftware packages computation of potential functionals are comparedon (someof) these aspects.The influenceof machine characteristics and softwa.recapabilities on computation time and accuracy of the with increasingmaximum degreeL. In general,single point comresults increases putation will be the most time consuming method, but it does not require much memory. Interpolation methods introduce some loss of accuracy and they require more memory capacity. Methods using FFT are very fast, but have the restrictton of computing the gradient valuesat regularly distributed points. Due to the wide availability of ever larger and faster computers the machine it influence will become less and Iess important in the future. Nevertheless, is to at this moment that for future gradiometermissionsthe use of vector be expected Therefore,someremarks computersor evenparallel computerswill be indispensable. about vectorizationwill be made in the following.
Vectorization Vectorization is, in fact, only one aspectof what is called more generally optirnization. When talking about optimization we have to discriminate between time and space optimization. In particular time optimization plays an important role when dealing with very large computational problems like gradient synthesisup to high degree. First of all it must be stated that optimization is not only a matter of hardware and system software. If one likes to fully benefit the optimization capabilities of the hardware configuration, also the userssoftwareshould be optimally adapted of to it. In spite of the fact that there are somegeneraltechniques optimization, the determinesthe way in which the software specificcomputer which is used eventually should be set up. Already in the case of "conventional" computers (scalar processoror so called "single Instruction SingleData" (SISD) machines)the impact of the algorithm setup can be significant. Just look at the differencein executionspeedbetween linear and binary search. Compilersrunning on machinesequippedwith a vector processor (so called "single Instruction Multiple Data" (SIMD) machines)modify the code in order to fully benefit the use of the vectorizationcapabilities. But the programmer still has to design the algorithm in such a way as to fit optimally the compiler. CPU time, parallelizationtries to reduce Whereas vectorizationintends to decrease by spreading work across multiple CPU's. The development "time to solution" of parallel computers ("Multiple Instruction Multiple Data" (MIMD) machines), is equipped with a large number of parallel connectedprocessors, still going on, and
45
they are not yet very common in use. how the algorithmsfor the recursive In (Koop and Stelpstra, 1989)it is described computation of the Legendrefunctions as well as the final evaluation of the potential functionalshave to be organizedas to perform optimally on a CONVEX C240 vector computer. It appeared that, in the case of singlepoint potential synthesis,the best way to go through the l, rnschemewas diagonally,i.e. first the main diagonal ( m : l , Y m 1 . L ) , t h e n t h e f i r s t s u b  d i a g o n a(lm : J  1 , V m I L  l ) , t h e n t h e secondsubdiagonal (rn  I  2, V m 1 L  2), etc. For grid computation, the same order is used, evaluating in each call the Legendre functions or potential values for all 0's. Nurnerical tests We did some numerical tests concerningsphericalharmonic synthesisand analysis. The tests were carried out on the CONVEX C240 vector computer. We used the howeverGRS80 referOSU86F (Rupp and Cruz, 1986) set of potential coefficients, were subtracted and Czr and ,S21 were encevaluesfor the first four zonal coefficients assigned specialvalues,seetable 3.5. Furthermore,a maximum degreeL :240 wa.s chosenfor the computations. tests Table 3.5 Specifications the numerical of
OSU86F with J2,J4,Ja,J3from GRS80 subtracted Coo:o er, : 0.f 0. loe . : ,Sr, 0.102108
First we examinedthe approximationerror introduced by the numerical quadrature formula 3.26. We have seen that this error is caused by the lsummation only, so we inserted eq. 3.21 into 3.26. In the resulting equation, from which the jrsummation is eliminated due to the orthogonality of Fourier series,we inserted on the right hand side the OSU86F coefficients.After numerical evaluation of this equation,we obtain the estimates dpn and 61, which can then be comparedwith the original OSU86F set. Figure 3.3 showsthe logarithm of the degreevariancesof the for betweenthe estimatedand the original coefficients two cases: relative differences a step size in 0 direction of Ad : Oo.5(i.e. 360 parallels)and a step size L0 : 0".25 (720 parallels). Since the fsummation is in fact nothing more than a numerical with decreasing step size, as indeed can be integration, the error should decrease with increasing seenin the figure. Furthermorewe seethat the error slowly increases
46
degree /, reaching about 106 at I : 240 in the case of 360 parallels. We have to bear in mind, however, that part of this error may be caused by underflows in the Legendre recursions, compare figure 3.2 and the paragraph on Legendre functions.
120
degree
quadrature error mad,e usingthe numerical by Figure 3.3 Relatiae formula Secondly,a program was made for grid computation of potential functionals. The program is an adapted versionof the single point program EVLPOT described in (Koop and Stelpstra, 1989),which is especiallytailored for use on the CONVEX C240 vector computer. It computes a grid of nine potential functionals in one run (potential and first and secondorderderivatives),both in geocentriccartesianand local northoriented coordinates.Due to the vectorizedsetup is does not make use of an FFT routine. Comparisonwith a sphericalharmonic synthesisversion (which in usesFFT) did not show any significantdifferences computation time (!), at least up to ,L : 24Oon the CONVEX. With the program we computed a world wide grid of gravity gradients in a local northoriented coordinate system at an altitude of 200 km. This grid was then analyzed for each of the six gradients using Colombo's HARMIN program for spherical harmonic analysis (Colombo, 1981). We slightly adapted the program in order to perform better on the vector computer. HARMIN of makes use of the numerical quadratureformula 3.26 so it yields estimatesd.pn,b6o etm,hrn (eq,.3.21)for some function /, in our case gravity the spectral coefficients gradients. From these estimateswe computed power spectra using 3.24, which are
47
shown in figure 3.4. From this figure (and confirmed by degreewisecomparison of the power spectra) we seethat the power of the various gradientsapproximatelyobey the relations (cf. (Rummel and v. Gelderen))
o! (r'r') x
o! (r' z') x
Note that the spectral coefficients etm,b1^ are in general not in a simple way r e l a t e d t o t h e p o t e n t i a l c o e f f i c i e n t se r * , S r ^ . O n l y f o r V " , r , s y g l 1r e l a t i o n i s e a s y :
48
c, I
c)
tr1
ti : a t r l
t{ bI)
n o
{
lO
120 degree
o o
a, I
tl
t;
: + t rl
lr{ h0
tO I
Fl
F0
60
720 degree
49
q
o
o
N N
b
E O
a) F.o o : f d
{)
o I
120
degree
and co' original potential coefficients Figure 3.5 Relatiuedifferencebetween uia and analysis of aV",",grid. efficients computed. synth.esis
50
4 l
Global gradiometric analysis
In chapter 2 we haveseenthat our main interestin satellitegradiometry is the preclse and detailed determinationof the earth's gravitational field. The gravitational information derived from the observedgravity gradientsis presentedby meansof a set of e potential coefficients Un,St^ up to somemaximum degreeand order, together with and the an indication of their precision. The relations between the measurements or unknown parameters (potential coefficients potential coefficientcorrections)are the presentchapter attention shall be paid to the method of derived in chapter 3. In (analysis) We restrict our. from the measurements solving the potential coefficients which means each sphericalharmonic coefficientreflects selvesto a globalanalysis, global field, correspondingto some spatial wavelength. Investigaa feature of the can be found in e.g. (Tscherninget al', 1990), tions into regional (local) approaches 1990), (Ilk, 1987) or (IIk et al., 1990). (Robbins,1985), (Arabelosand Tscherning, error analysis,which means that we study Furthermore, emphasiswill be laid on the error propagation during the adjustment process.This will be done for several kinds of possiblemission scenarios. The method we use here is basedon leastsquaresadjustmentand was proposed of by O.L. Colombo (1987, 1989). One of the main characteristics this method is (see section 4.2), the normal matrix in the least that, under certain assumptions squaresestimation processattains a blockdiagonalstructure which can be inverted without much (computational) effort. Without the blockdiagonal structure it will be a tough job to invert the normal matrix, especiallyfor high degree solutions (Balmino and Barriot, 1990). After a description of the Ieast squares adjustment method in the first section of this chapter, the secondsection deals with the error analysis. Finally the third sectiondealswith methodsof global recovery(i.e. solving the potential coefficients) again basedon a least squaresadjustment. ,
51
4.
Global gradiontetric
analysis
4.L
from gradiWe will show here how the problem of solving the potential coefficients is ometer measurements set up in terms of a least squaresadjustment model. For we the measurements take a set of observedsecondorderradial derivatives. This is just an example. Equivalent derivations hold for other gradients, as well as for combinationsof gradients. For the moment, orbit and orientation uncertaintiesare neglected. Observations are given along a satellite orbit and it is therefore appropriate to use an expansionof the gravitational potential in inclination functions, as was are given in section 3.2. We assumethe measurements given with respect to a local orbital coordinate system ci. In appendix A this local orbital system is defined. Using equation3.18 and table 3.4 from From section3.1.2we find thatV"":Vrr. section 3.2.2 we have for the secondorderradial derivative:
v". :t
with
t=on=o k=_tl2l
^ fdfff)r"rr/tr a!ff)sin,1,p*l
(4.1)
f Qlrn= 
' 1l  m : e v e n 
l c t * l
&
l,,ooo
Plm
a.
lul
let*1,,ooo
,lr*rn: kwo * mw" For practical computations we have to truncate the infinite summation over I at e7^ and ,S1are some maximum degree tr. In equation 4.1 the potential coefficien's
52
4.1.
the unknowns. Now consider a large set of measurementsV", given for all kinds of values of the coordinates@o>@e assumethe inclination 1 of the orbit remains (we constant during the mission). For each measurementwe have an equation of the type above, each as a function of the same potential coefficients Qrr., S1. Now let us arrange all these measurements V"" in one vector, denoted I (the tt" denotes stochasticquantities). We may also order the unknown coefficients, each value for of degree/ and order rn successively, one vector, denoted c, merging the double in summation over / and rn into one. In this way we obtain the following system of equations:
i:Ac+6
(4.2)
in which 6 are the residualsbetweenthe observations and the model. This model is representedby the designmatrix A of which each element is of the type
I s)_ ,.(zz) nink cosl4km
(4.3)
k=tlz
I \
,(zz)
(4.4)
depending on whether we are dealing with a Cpn or S1* coefficientand whether I  rn is even or odd. The apriori variancesand covariances the observations of are collected in the variancecovariance matrix Qrr. Solving the model 4.2 in a least squaressense(i.e. minimizing {Q;"t4, we find for the estimates6 of the potential coefficients: r 6 : 1J 4,r e;"1 t (4.5) with the normal matrix N
u: ere;)A.
The inverse of the normal matrix, .y'r, representsthe aposteriori variancesand covariancesof the estimated potential coefficients. It can be calculated without having to perform a complete adjustment. This offersthe possibility of carrying out an error analysis without the availability of a real set of observations. This fact will be used in the next section. Since a gradiometer not only measuresthe secondorderradial derivative but also, as in the case of Aristoteles, the gradients Vuu and Vy", and in the future hopefully all six gradients, we will need observationequations for these gradients too. The relation between the local gradientsV;i and the potential derivativeswith respect to rrl,eotu)e are given in section 3.1.2. There we see that we have two : (r,uo,u.'r)(equations possibilities.Either we choose the set {fla:1,2,3} 3.12) o r t h e s e t { r o ' l a ' : L , 2 , 9 } : ( r , r o , / ) ( e q u a t i o n s . 1 1 ) . I t c a n b e s h o w nt h a t b o t h 3 give the same result if used for gradient synthesis. Just for the sets of expressions
53
4.
sake of convenience we choose equations 3.11. For all six gradients an expression of the type 4.1 can be derived, using equations 3.11 and table 3.4, resulting in again different expressions for the quantities Atrok, Bt^k and H1*p. For all six gradients, those quantities are listed in table 4.1. Table 4.1 Etpressionsfor local gradientsin terrns of deriuatiuesol the potential with respectto (rrwor l).
H[)o
tg
Alro*
Bl**
on"Hliii
( \
0mHi#i
(ko,*  cotwoBp,.)H!7)
( \
sy tz
(cotwoc,1* nP,in[7)
vv
yz zz
( + z)rpF,k* ((t+t)2kz)r{!^
 ( t + 2 ) s i n  1w " l 1 F [ , .
B,*n[i,])
o,^H[#)
o,*u{ll)
( t + t ) ( t+ z ) t t F t k ^
",*n[ii)
In this table we computed the expression for V* using the Laplace equation: Vcy : (Vrr*V""), thereby avoidingthe secondorderderivativesof the inclination functions and obtaining a simpler equation. As an example we derive here the expression for Vr, From equation 3.ll we have
1 Vr":lVrr^ ,, ,vu,,
Ht*kfor the roroderivative From table 3.4 we have for tn" qru.rtitl u" or,oo','"r*,* and
: AIXP F,^Hliif)
: BIX;) o,*H[#t) ulkF!* HlxP: (F1)
and for the aroderivative
54
) ,L,rlolX)
( / + 1 1 l, )\ r r/
I F a6nku1 f*sinrltlrr,l J
: t
l,m,k
: t IAIT)cosrh*^t Bffisinrbr*l ,
l,n,k
JJ
4.
between 3.10 and 3.11 for these two gradients is easy if we use equation C.l from appendixC:
v ru : coso.ro I r , T ;v t l vI Juo n r sln o.,o r' sln (,o
I a /
ro d(to
.;1.
r ,,\ ,ti
/
\sln@o
_ 1 v, d,w " q
and 1 Vur: * sln (,o Vll 1 Vrt . r sln @o
FV,
V,Q 1. .
In terms of the crosstrack inclination functions Frk^ we now find for the gradients Vro andVr" the expressions listed in table 4.2. The other gradients are the same as in table 4.1. Table 4.2 andVr" in terms of the deriaatiuesof New erpressions forV, t h e p o t e n t i a lw i t h r e s p e c t o ( r r $ r u o ) .
H[)r
ry yz
Al*k
Bt**
I .,\
kt1F!^
omHi#i
gt*u!7)
are In table 4.2the A1*p and 81,,,1, now the coefficientsof respectively the cosine where now k : I  2p  1 (compare and sine of the argument ,lrkn: kwo* rnrr.r" appendix C). The summation over ,t runs for thesegradientsfrom (,  1) to I  1. from table 4.2 for the gradients 7r, and In the sequelwe will use the expressions from table 4.1. Vy" and for the other gradientsthe expressions
4.2
Colombo's
method
of error analysis
The idea of a gradiometric error analysis as it is presentedhere originated from (Colombo, 1987). As describedin the previous section,it consistsof the computation of the inverseof the normal matrix in the senseof a least squaresadjustment, for thereby using an expression the gradientsin terms of inclination functions. The
56
of diagonal elements this inverserepresentthe aposteriori variances the estimated of potential coefficients. As we have seen,they can be computed without the use of actual observations.The latter implies that we are not restricted to some specified orbit and instrument characteristics.Given some mission goals in terms of resolution and accuracy requirements of derived gravity anomalies or geoid heights (as were given in section 2.2) one can searchfor the specificorbit and instrument characteristicsmatching these requirements. In turn, given some orbit and instrument characteristics(e.g. mission duration limitations, instrument accuracy limitations one can derive the resulting etc. imposed by technical or practical considerations) or accuracy and resolution of the estimated potential coefficients of certain derived quantities like gravity anomaliesor geoidheights. Of course,the value of such computations is limited by the underlying assumptionsof the method of analysis. For example, as already stated before,the presentmethod does not include orbit or orientation uncertainties. We simply assumeto be given a global, regularly distributed relative to someknown local coordinatesystem. For the set of gradient observations sake of simplicity, however,we make severalother assumptions.Thesewill be listed below, seealso (Colombo, 1987) and (Colombo, 1989). 1. We assumethe orbit of the satellite to be circular. In reality the orbit will have an eccentricitydifferent from zero, although very small. Furthermore, by assumingthe orbit to be circular we neglectthe orbital decay as a result of air drag. For noncircularorbits ("lO), we would have to include in the model the socalled eccentricitg functions GUok) (Kaula, 1966) which in generalare to be summed for oo < g < oo. However,it appearsthat for nearly circular orbits the index g can be restricted, with sufficient accuracy,to three values: 1 < I 1 l, cf. (Schrama,1989) or (Wagner, 1989). In spite of the fact that inclusion of these eccentricityfunctions would not influenceour analysis method fundamentally,we leavethem out for simplicity. If required, the error analysis can be carried out at different values of r to reveal the influence of varying satellite altitude. 2. We consider the data to be distributed regularly along the orbit. During the measurementperiods this assumption is very well met, but gaps occur maneuvers, sincein those during e.g.instrument failuresand orbit maintenance periods data is not usefulor not even present. Also in a real mission gaps near the poles and, to a smaller extent, at the equator may occur as a result of drag variations (Touboul et al., 1991). However, if the mission is excessive long enough the overall data set may yield a very regular data distribution. 3. Regarding the mission length we assumethat it consistsof an integer number of complete revolutions and an integer number of nodal days. Furthermore we will assumethat the mission duration equalsthat of one repeat period ?r, so that no groundtrack repeat will occur during the mission. To this extent, the number of orbit revolutions N, during the mission period and the number of (nodal) days Na contained in it have to be relative prime integers. This
ct
can be seen as follows. Consider equation 3.18, which constitutes a Fourier series as function of the argument ry'1. This series can be regarded a time seriesif successive measurementpoints along the orbit are considered. The time parameter t enters the argument 4t*ro if we write wo : u2 + 6"(t  ts) and o" : u2 + ,;"(t  t6) where the dot denotesdifferentiationwith respectto t and where ar! and u! are evaluationsof respectivelyuo andu" at t6. Thus:
,bko: kwo I rnw"
: t' (r? + b"( t  to) )+ m( w! + 6"( t  to) ) : krf; t rnu! + (kd)" mb")(t  ts) + = ,b2*+ ,lt*^(t ro).
(4.6)
Whereas the term ,/tlr^ only representsa phase shift, the term tftlr* represents the actual measurementfrequency.However,the frequencyin terms of.cycles per reuolutfon (c.p.r.) is denoted 0**(: rb*r.lr") and may be written as (cf. (Schrama,1989)or (Schrama,1990)):
9 k * : t '+ ^ ? (ro
so that tltk^: t!'r*+ 0*^6o(t  ts). Now one nodal day is 2rf 6" secondsand one revolutionis2rf wo seconds.If the mission (?,. seconds)takes exactly lfa nodal days and N" orbit revolutions,there exists an integer ratio betweenur, and aro,i.e.
Yt:Y!
Qo Nt'
This meansthat after Na nodal days and N" orbit revolutionsthe ground track will exactly repeat. However,if Na and N, would have some common divisor d, this repeat would already occur after Nafd days and Nrf d revolutions. Requiring Na and N" to be relative primes therefore ensures no orbit repeat during the missionperiod ?r. This also meansthat during this mission period the finest possible coverageof the earth's surface is obtained. More repeat periods, being nothing more than a "repetition of experiments",would, on the other hand, only result in a rescaling of the variances. which the instrument delivers are in generalthe result of 4. The measurements an averagingprocessover some time interval At', which has to be lessor equal to the sampleinterval At. This averagingcan be accountedfor in the following manner. Consider equation 4.6. Let us, for the sake of simplicity, assumeto : 0 and the initial value rllrrn to be zero. We then have $1"^: ,lr*^t. Being a function of time t, the expression for the gradients can be integrated in order to account for the timeaveraging mentioned above. The integration only applies
58
to the cosineand sine terms and takes place over the time interval Atf . Upon introducing the integration variable r, this integration becomes: _ I t f + + cosrlprrrdr : At'lt{ : 1 1 f . ; ,, At'. i t,  At'rl ptrz(t + ) srnrlp,"lt 2 )) z At, il;fstn 2
ir^,
N J,+
St+{ sin$p*rdr:
in may be incorporated the inclinaThe factor Jk,n=zsin(ff)l\!**L/) tion function to obtain a kind of smoothedfunction J*^Ff^. In the sequelwe will assumeall inclination functions to be multiplied by J**. We furthermore take At': At. 5 . Nongravitational effectswill not be included in the present error analysis. We assumethat, to first order, these effectsare eliminated through common mode rejection (seechapter 2). Only secondordereffects,due to e.g. nonsymmetry in the instrument, remain. Following (Colombo, 1987) we assume a large part of these effects,as well as some instrument errors Iike thermal noise, to be present in a low frequency band, mainly below /nin : 4 c.p.r. (Schrama, 1990). Removing from the analysis the low frequenciesl9**l < gnin may be considered appropriate to account for these nongravitational effects. Other highpass filters may also be used, but are not consideredhere. 6 . For the moment the covariance matrix of the measurements,Q, will be consideredto be a scaledunit matrix,i.e.Q: o2I. This meansthat consecutive are consideredto be uncorrelated and of equal variance. The measurements variance factor o2 rnay, however, have arbitrary values for different gradients. in Spatial differences the precisionof the data due to phenomenalike e.g. drag variations are thereforenot included in the present analysis. of One of the most important consequences the assumptionsabove is that the normal matrix (which is to be inverted) attains a blockdiagonal structure. This will be explained in the next section. Each block of the normal matrix can be inverted separately,which means an enormousreduction of computation time. In this way it becomespossibleto carry out the error analysisseveraltimes with only limited (computational) effort. Each time we may changethe orbit and instrument characteristics(like the inclination or instrument noise level) thereby allowing the investigation of severaldifferent mission scenarios.These mission scenariosdepend on the value of the following mission parameters:
59
4.
h 1 Lt T, o 9*;n 9^o"
radius R, i.e. r : R + h satellite height above reference inclination of the orbit sample interval, i.e. time betweentwo successive measurements mission duration (one repeat period) measurement precision lowest measurement frequency included in the analysis frequencyincluded in the analysis highest measurement
It will be shown later (section 4.2.3) that only for a change in some of these parametersa new run of the error analysisprogram is needed(e.g.for the inclination 1), but that a changein the other parameters(e.g. the height h) can be accounted for by simply scaling the inuertednormal matrix. 4.2.1 Norrnal rnatrix
The normal matrix N of the least squares adjustment problem 4.2 is given by ArQrl. In this section we will derive the expressions for the elements of this given in the last section. matrix, under the assumptions Inserting for the covariancematrix Q the expression o21 leavesfor the normal matrix N = \Ar A. We see that the elementsof the normal matrix are computed by taking the inner product of columns of the design matrix A. Each column of A contains, for one specific combination of I and m, terms of the form 4.3 or 4.4, the argument **ro (being a function of time t) indicating consecutive observationpoints (epochs). If we, for simplicity,assumeto:0 and r/.,f;*:0, we have from equation4.6 r!** k6ot + mw"t f.orall , : 0,...,7,. where ?" is one repeat period, in our caseequal to the mission duration. Now we only have measurements discrete points along in the orbit with interval At, so we may write ,lr**: kuoj\t* mb"jAt with jr 0,.'.,NrlandNr:Trf L,t,thetotalnumberofmeasurementsduringthemission. In the previous section we have seen that Na is the number of nodal days during the mission period ?, and N" is the number of orbit revolutions. So we have:
q .  t r 2 t tr tf d;
,:,e w 
!AL2r, '(tT
Nt1 2r Lt Np
, n tr
trri)o
nr?r
2 w o  t tt rr T r I ?N , 2 r l &
60
:2f,@r^N,) ,l,r^
j:0,...,Nol
with 81,,o k +ff. Not" that, since L < k I L, \r^may take negative values too. Taking the inner product of two columns 1 and 2 of A means a summation over the indexJ over all measurement pointsj :0,. . . , No I for differentvalues11,12,m1 and m2. Let us take as an example the gradientV"" and the caseI  m even. For a e1ro coeffi.cient(indicated by an upper index cc) the element of the normal matrix now becomes(leavingout, for the moment,the factor (/+ t)(/+ 2)f;):
r N p  l l
,  ,
\ /
,  :
According to the orthogonality properties of trigonometric series, the summation overy in this equation reduces to:
& ,
N.
if if
if
In case .ly'ois even the result of the 7summation may also be equal to Npl2 it l 1 * r * r l : l g t r * r l + + a n d e q u a lt o N , i f g n r * , : g k " * r : + , b u t i n a r e a l i s t i c mission with millions of measurementpoints this situation will never occur. The frequency Bp equals k + rn# where k may take values between  L and,L and m may take values from 0 to .L. This means that there may be several ft, rn combinations corresponding to the same frequency Bpr". This also implies that we may have (4.7) l9*r^rl: l9*r*"1 for h * kz and rnt * rnz cf. (Schrama, 1990). This situation will not lead to a blockdiagonal structure of the normal matrix, as will be explained later. Now the situation /kr*, : 9*",o" implies: No__lqkz N,. rmt  rnz where (as explained before) the numbers N4 and N" are relative primes (i.e. they have no common divisor, so their fraction cannot be simplified any more). The
61
4.
denominator on the right hand side will never be greater than .L. If we now ensure that N, is always greater than .L there is no possible combination of the numbers kt,kz,fftL,trr2 for which 4.7 holds. Furthermore, we have to avoid the situation where for certain k, rn combinations Asimila'r phase,i.e9*rror:9k"^". f r e q u e n c i e s B l r l n o c c u rw h i c h a r e l 8 0 o o u t o f N" be greater than reasoning as above will, for this case, lead to the requirement that : kz = k, 2L. If now N' > 2L we only have gkr^t: 9k"^" if m1 : m2 : m and k1 at least for m I 0. Furthermore, in this case we never have Fr* : O. If rn1 : rn2 = 0*"oVkr  ft2. Also, m : O w e h a v e a s p e c i a l s i t u a t i o n s i n c e i n t h a t c a s eg k r o : k Z : k : O . T h e e x p r e s s i o nf o r a n e l e m e n t o f we have Blrro:0 for kt: for rn:0 the normal matrix can now be simplified to:
, (zz)cc
hi,ti* :
T,
2o2Ll
rnin(l 1, . 2 ) l a = m i n ( l
F+_
t Jz)lz\
w h e r ef . o rm l O
F1  F!,*F!"*
h.r", 'oll2m 
lt,lhT,
2oz A,t
(4.8)
hiir*:
: htirr*: h't"rt"* O
with 211116taken from table 4.3 (with factors like (t + 1)(, + 2)ft restored).
62
4.2.
Table 4.3
Zlrl"^k
tr ry
(h + 2)(tz 2)k2 +F +
vv
yz
ri : Ff,,iF,r,,i
and for rn: O
z,'1 ,'" '!i ;'" { [(/r + 1 ) '  k' ]l( tr + t) 2  /c2l+ ': (/r + 1 )(/r 2) ( t2 1) ( 12 2) ) r ' +  + + + + ( h + 2 ) ( t z + 2 ) F ; 'If some combination of gradientsis used as a singleobservation,for example 2Vr, * Vr", the expressionfor the elementsof the normal matrix is found by using equation 4.8 with
(.,.
63
4. GIobaIgradiometric analysis
choice of the ordering of the unknown potential coefficients,the normal matrix will attain a blockdiagonal structure (Colombo, 1987) which is much more easy to invert. First of all, we have seen that the normal matrix elementsconnecting an arbitrary e m and an arbitrary $ coefficientare all zero. This means that if we order the coefficientsin such a way that ail e6 coefficientscome first, the normal matrix is divided into four blocks, the upper right and lower left one (the offdiagonal blocks) contain only zero's. Furthermore the two diagonal blocks are exactly the same. Also, since only elements for equal rn differ from zero (at least if If" > 2.L), ordering the unknowns accordingto m causesthe two main diagonal blocks to be divided into smaller blocks, one for each rn. The size of these smaller blocks decreases from,L* 1 for m : O to I for m : L. Finally, as a consequence of the vanishing of the summation terms for which h * kz (again only if N, > 2L) only elementsfor which k is the same remain, which in turn implies that 11and /2 have to have the same parity. So, the rnblocks are again divided into four smaller blocks if for each m the / are ordered accordingto their parity, first all even / and then all odd l. Of those four blocks only the diagonal blocks differ from zero, see figure 4.1. The size of these blocks is half the size of the mblocks. The inversion problem is now reduced from that of a matrix of size L2 to one of 2.L times a matrix with a sizeof at most 2, which meansan enormousreduction of computation time. The blockdiagonal structure also implies that the error analysiscan be carried out for certain groups of coefficientsindependently. 4.2.2 Presentation
With the expressions the elementsof the normal matrix N which are derived in for the last section,we are now in a position to compute in an easyway the aposteriori variancesof the estimated potential coefficients. The assumptionsstated in the beginning of section4.2 may, to someextent, put limitations on the usefulness the of method. Nevertheless, method is still very useful to gain a better understanding the of the gradiometric analysis problem and at the same time shows possible problem areas. The required aposteriori variancesare found to be the diagonal elements of the inverse normal matrix N1. A very common way of representationof these variancesis by meansof degreevariances.The error degree uariance(of the formally propagatederror) is
a
Loirn,
(4.e)
where use is made of the equality of the error variances for the epn and S; coefficients and where or2^ it a diagonal element of the inverse normal matrix .l[l
64
et*
I
sun
I I
tt.
m: O
),
m: I
lrn:2 )
m:l
J 
t ,m : 2
J
ll,.,l
o
r.+r
iL + L t{ "l r
' L
tl,_,
"))
I I I I I I _1 I I L _ _
I I
Figure 4.1 oe" n B l o c k  d i a g o n as t r u c t u r eo l t h e n o r m a l m a t r i z . T h , e o t a t i o n l eu n ' L e o , n s e nI a n d " o " o d d l .
65
4.
Global gradiometric
analysis
for a specific value of J and rn. From equation 4.9 we compute the error r.m.s. per coefficient per degree as
(4.10)
the square of which is called error degreeorder uariance. Degree variances and degreeorder variances can also be computed for the potential coefficients themselves, in which case they are called respectively signal degree uariances and signal degreeorder uariances. Ttey are defined as
I
(4.11)
and
ct*:2r+r'
cl
(4.r2)
For computational purposes, the potential coefficientsC1* and ,S1 in eq. 4.11 may be taken from some apriori known potential coefficientmodel, e.g. OSU86F (R.pp and Cruz, 1986). Another possibility is to compute the c1directly using some degree variance model, like Kaula's rule (Kaula, 1966) or the Tscherning/Rapp model (Tscherningand Rapp, 1974). Note that, since the potential coefficients so et^rSun are dimensionless, are o!, oun, c1 and c1^. Whereasthe degreevarianceof somefunction (either signal or error) represents the power of the function per degree,one is also often interestedin the total power of the function over all degrees. For example, the total signal power for potential coefficientsi" DEocl with the c1 taken from equation 4.11. This total power also representsthe averagesquare value (or norm) of the signal over the unit sphere, cf. (Heiskanen and Moritz, 1967). The globalr.m.s. is the squareroot of the total power. In the same way we may compute from the error degree variances(equaby tion 4.9) the global error r.m.s. for potential coefficients summing the of over all in degrees. The mission goals for gradiometry are expressed terms of global error r.m.s. values for gravity anomaliesor geoid undulations. They can easily be computed from the global error r.m.s. for potential coefficients multiplying the latter by )1 with the eigenvalues of the linear operator connecting the respectivequantity (Rummel, 1991). The global error r.m.s. becomes with the potential coefficients
D, t7"7
ln
S": ff(l  1) and for geoidundulations)3: R. The with for gravity anomalies global r.m.s. of each arbitrary other gravitational quantity can be computed in a similar manner, as long as it is linearly related to the potential. Here we will only show the two mentioned quantities, namely the surface gravity anomaliesand the geoid heights.
66
from zero to infinity. The We seethat the summation aboveincludesall degrees reason for this is that the gravitational potential is consideredan element of an sphere infinite dimensionalHilbert spaceof functions defined outside a convergence which is approximately the surface of the earth. The spectrum of such functions from zero up to infinity. In our error analysis,based on the includes all frequencies availability of a finite number of measuredsamples,we can only estimate a limited the part of this spectrum, e.B. up to some maximum degree L. As a consequence, even never be perfectly reproducedfrom the measurements, gravitational field can in case of perfect measurements. There always remains a part of the spectrum which is neglectedin the analysis. This neglectedsignal is called omission error. In case of regular sampling, it contains primarily the high frequency signal part I for degrees > L. As a result, the global error r.m.s. should consistof two parts: the propagated error from the analysis for degreesI < L (called the commission error) and the omission error (neglectedsignal for degrees/ > .L). Since we do not know the true spectrum of the gravitational field, and since it is impossiblein practice to really include all degreesup to infinity, the omission part is computed only up to some limited maximum degree/rncs using a model representingthe true spectrum. Thus, the global error r.m.s. valuesfor gravity anomaliesor geoid heights are computed as
re
(4.13)
where the first summation indicatesthe commissionerror and the secondthe omtssion error. and where
9t
F**
#lnt(cosr/.)
 Pr+r(.orry'")] smoothingoperator
,h"
radius of a sphericalcap with equal area as an equiangularblock of size d" at the equator
I ^ 
,\1
I W(t
I
 t)
[
Lt
ft
degree of truncation of the commission error which is less or equal to the maximum degree .L of the performed error analysis truncation degree for the omission error
lmrs
The reasonfor including the smoothing operator 81, seee.g. (Meissl, 1971),will be explained in section 4.2.7. A truncation degree for the commission error L1 smaller than ,L may be chosen,for example,if the error in the estimated coefficients the signal to I for degrees > Lt exceedsIOO% of the signal, i.e. if for those degrees o?l"t > 1, cf. (Colombo, 1989)or (Rapp, 1989). Choosing .L1in this way noise ratio is in fact comparableto "ideal" lowpass filtering. However,a more optimal way for
67
4. Globalgradiometricanalysis
computing the global r.m.s. is using a minimum variancefilter, like a Wiener filter, becomes: e.g. (Papoulis,1965).In that casethe expression
( 4.r 4)
The Wiener filter is derived as to ensure a minimum total error (cornmission* where the signal power largely exceedsthe noise power, the omission). For degrees filter will approximately attain the value l. In the estimation procedurethis would imply that the data at these degreesare almost completely used. For the global error computation it thereforemeans that the error for those degreescomes almost exclusively from the propagated error ol from the gradiometric analysis. If the noise power, however,is larger than the signal power, the filter tends to zero. For the global error this means that for those degreesthe contribution comes largely from apriori availablegravity information, in our casethe signal degreevariances c1,which can be seen from equation 4.L4. In this equation, the term including o! the commissionpart and both other terms the omissionpart. This means represents that, if one usesa Wiener filter, one may also have an omissionerror part for degrees below ,L, which in fact results from the imperfectionof the measurements.We will, use the latter equation4.14. in our computations, exclusively The above derivations shall now be used to study the error behaviour of gradiometric experiments. Thereby we shall start with the most ideal situation. It is all characterized a gradiometerinstrument which measures six componentsof the by gradient tensor (a socalled full tensor gradiometer). The full measurementsignal (the complete spectrum of measurementfrequencies, here indicated by the index ft) is used with equal variancesfor all frequencies.Consideringthe orbit, an ideal situation is a polar orbit (1 : 90') in order to obtain complete global coverage.Afterwards we shall considera number of restrictions (band limitation and nonpolar orbits) and finally we discussa scenariothat resemblesas closely as possible the current plans of the Aristoteles mission. 4.2.3 Ideal case
We consider here a full tensor gradiometer,which measuresall six componentsof we the inclination of the gradient tensor. To obtain completeglobal coverage assume the orbit to be 90'. Finally we considerthe instrument not to be bandlimited. The frequencies, low and high enough to latter implies that we include all measurement maximum degree,in the analysis,and that recoverthe potential up to somespecified on we assumethem to have equal accuracy.A discussion a bandlimited instrument will be given in the next section. For the moment we simply take B*;n to be zero and
68
4.2.
up Frro, ) 254, which is high enoughto recoverpotential coefficients to ,L : 240, see section 4.2.4. These, and other characteristics the ideal casegradiometermission of are listed in table 4.4. Table 4.4 Characteristics the id,eal oJ casegradiorneter rnission. inclination /:90o
components v", ,vru rvr" ,vgg,vg rv"" " no band Iimitation 9 * ; n : O satellite height sampling interval mission duration h:200km Al:4s T, :6 months
In figure 4.2 the (dimensionless) error r.m.s. valuesfor potential coefficients computed with equation 4.10 are shown for 10 different cases:for each of the six tensor componentsindividually, for severalcombinationsof components(namely all six together, {Vsy,Vy,,V""} and {Vsy,V""}) and for the quantity 2Vuu 1V"". The latter quantity is chosen becauseit will be used in the potential coefficientrecovery in section 4.3. The ordinates in these figures have a logarithmic scale. For all situations the overall pattern of the graphs is the same: after an initial decrease of the r.m.s. value (i.e. an increaseof the aposteriori accuracy) the lowest point is reached around degree I = 60, after which a steady increaseoccurs. This overall pattern can be explained by remembering the expressions the elementsof the for normal matrix from table 4.3. As the result of two differentiationsall the gradients are multiplied by a factor proportional to at least /2. For the signal this means a higher contribution for higher degrees, to be expectedfor secondorderderivatias ves. For the aposteriori error (inverseof the normal matrix) this means a decrease with increasingdegreeJ. However,as already indicated in section 3.2.1,the natural attenuation effect of the gravitational potential with height is also present. For the gradients this factor is (Rlr)t+3. For the elementsof the normal matrix this factor is again squared. For higher degrees, the signal will decrease, and the error r.m.s. will thereforeincrease.On a logarithmic scalethis attenuation factor will result in a straight line, starting left below and reachingfor the upper right side (seefigure 4.3). Eventually this effect will be dominant, so the r.m.s. plots are likely to approximate this straight line for higher degrees, can be seenin the figure. as Combination of the effectsdescribed above results in the lowest error for the part of the spectrum somewhere betweendegrees and 100,slightly dependenton 30 the observation type, making gradiometry especiallysuitable for determination of
69
4.
t{ O) O t
qj
R o
A =
a l
AI
I I
h o ) O l +i li
o
A a
ol I
@
 w
k O )
o r l
c..;
P o
t o l 0 e '
I:90
Fi,,:O
ol
I0
60
120
degree I
180
240
Figure 4.2
error frorn gradiometryfor uarious situations. VertiPropagated For all casesthe satellite'saltitude was is cal scale logarith.mical. km, the rnissionduration T, 6 months, the sample interual 200 was and L,t I second.s a 0.Ot EftHz white noiseerror spectrurrL assurned,.
70
4.2.
term term
attenuation term
(r73;2(t+s)
 o
k
o e.i o) t o m
\)
l ltr
I
=.1
oil
iu
i+ I
120 degree I
Figure 4.3
Infl.uence f attenuation actor on error r.m.s. plots. The figo f u r e s h o w se r r o r r . t n . s .f r o m V " " . k f t a e r t i c a ls c a l ei s f o r r . m . s aalues, right aertical scalelor attenuation factor. Both scales logarithmical.
71
4.
Globa gradiontetric
anilysis
thesedegrees. From figure 4.2 we see that the largest error is obtained for the V* and Vuu components, followedby that from respectively 2V* IV""rVrrrV"" and Vrr. Out of the seven individual observationsshown in the figure, V"" givesthe best result. Redundancy,however,ensureseven better results for combinationsof observations, the combination {Vor,Vrr} being slightly better than V"" alone, followed by the combination {Vuu,Vs",V"") and the best results for the full tensor combination {Vrr,Vry,Vr",Vvv,Vy",Vr"). Note that for the six individual gradients,the level of the aposteriori error relative to one another is not necessarily the inverseof the signal power level, cf. figure 3.4. Once we have computed the r.m.s. valuesfor a specificinclination and 0*;n, it is very easy to derive from them equivalentvaluesfor various altitudes, with various mission durations, sampling intervals and measurement precisions.We do not need to invert a new normal matrix in these cases. This can be understood as follows. Columns of the designmatrix A are multiplied by the downward continuation factor ( R l i ' * " , c f . e q . 4 . 1 , w h e r er : R * h . S o A c a n b e w r i t t e n a s A : A ' D , w h e r eD is a diagonalmatrix containingonly the factors (Rlr)t*" and Atequals A without thesefactors. Then, sinceN  ATQA, we have N : DTA'rer A' D : DNt D
with N' :{tTgr4'. Computing N'1 first, it can be scaledby pre and post multiplication with Dr to obtain the results at satellite altitude, since y'fl : DL N'r D1. Furthermore, from eq. 4.8 we seethat eachelementof l{ is multiplied by a factor Trf o2Lt. The square root of this factor can be included in D, and can thus be left out of the matrix inversion. If we now like to compute error r.m.s. values for other, arbitrary values of h,Lt,T, or o2, we do not need to carry out a new normal matrix inversion. The inverse y'y''1, computed only once, is simply scaled with an appropriate diagonal matrix D. Note that for combinations of gradients, like {7rr,V""}, the scaling with o2 does not alwayswork. If different gradients are given different o's, a new inversion has to be established. For the graphs of figure 4.2 scaling with ?r, oz or L,t means a uniform shift of the complete curve. A different altitude means that the straight line representing the attenuation factor (seeabove) obtains another slope with subsequenteffect on the r.m.s. curve. In figure 4.4 the effects mentioned above are illustrated for the error r.m.s. from V".. As describedin the previous section the error degreevariancescan be used to compute global error r.m.s. valuesfor gravity anomaliesand geoid undulations. The influenceof the individual mission parametersTr,A't,h and o on this global r.m.s. is not so obvious any more. For a full tensor gradiometer,measuring all six tensor components, and for the case that all these six componentsare used in the error analysiswith equal weights, this global r.m.s. is shown in figures4.5 and 4.6. Figure 4.5 shows the global r.m.s. for three different mission durations as a function of height. (Note that the straight lines are just added for clearness.They
72
4.2.
o = o.o7t/Jttz
t{
o
}j
R o
6
ri
{)
a
ol I
60
120 degree I
180
240
Figure 4.4
altitude, mission duration and measurement Effect of a d,ifferent precisiorlotu the o,posteriorierror r.n'I.s.aaluesfromV"".
73
4.
IN MCAL
tb 0a {
c
o t p
c'i
140
160
180
200
220
240
260
N 'V Cil
*
AI
o O
E o k
140
160
180
200
220
240
260
satellite height in km
Figure 4.5
Global r.rn.s. (commission* omission error) for surfacegravity anomaliesL,g and geoid undulations N for three different mission durations as function of the satellite's altitude. Matimurn d e g r e eo r o r n i s s i o ne r r o r w a s 1 0 0 0 . N u r n b e r sr e p r e s e n1 " x 7 " t f aaeraqes. blocle
74
Ag IN MGAL
u) trro l <
m
l. t
.3+
lO
zoo
*
tr
o.ooo1 n/Jnz
0.01 n/Juz
o
d c l o L d O
o.ozn/JHz
140
160
180
200
220
240
260
satellite height in km
Figure 4.6
Global r.rn.s. (comrnission* omission error) for surfacegraaity undulationsN lor three different rnea' anomaliesLg and geoid, surement precisionsas lunction of the satellite's altitude. Max'irnurn degree t'or omission error was 7000. Nurnbers represent j " x 7 " b l o c ka a e r a q e s .
la
4. Global gradiometricanalysis
do not representfunctional valuessince only the nodes (5 different altitudes) were really computed.) The numbersin this and the following figuresof this type were all computed using equation 4.14, including the Wiener filter, where for the omission error (computed using the Tscherning/Rupp degree variance model) a maximum summation value /mcs : 1000 was taken. The numbers furthermore represent 1o x 1o block averagessince the smoothing operator B1 was included with 0, : 1o. The influenceof lmrs and d" on the resultswill be discussed section4.2.7. in The most remarkable fact from this figure is that at the lowest altitude (160 km) tripling the mission duration gives no substantial improvement of the total r.m.s., cf. (Rapp, 1989). Figure 4.6 is of the same kind as the previous one, but now results are shown for three different measurement precisions.The caseo:0.0001 E/JH, refersto the anticipated measurementprecision of the superconductinggradiometer under developmentat NASA (Paik and Richard, 1986). What can be seenfrom this figure is that with this kind of high measurementaccuracy,satellite altitude becomesof lessimportance, whereasfor the lower accuracies still plays an important role. At it 160 km, however,all o's give nearly the same result, cf. (Rapp, 1989). Apparently there is some lower limit for the total r.m.s. To understand this, we have to bear in mind that the r.m.s. values representthe total error) commission and omission error (seethe previous section). Increasingthe measurement precisionchangesthe commission error part. For example, the commissionerror for gravity anomalies for the three increasingmeasurement precisions(0.02, O.Ol and 0.0001nlt/nz) at 160 km altitude is respectively 0.28 mgal,0.14 mgal and 0.0014mgal. As for the omission error) only the contribution for degreesbelow tr is changed (due to the Wiener filter) when increasingthe measurementprecision,whereasthe part above .L is not changed at all. The total error therefore has some constant base level determined by the omission error, especiallythe part above Z. This part has to be decreasedin order to fully benefit the improved measurementaccuracy,which implies that we have to increase.L. The maximum value for .L is determined by the sampling density of the measurements, which is represented our caseby the in combination of mission duration ?" and sampling interval At. So we see that in the end improving the measurementaccuracywill not give better results if at the same time not also the sampling density is increased.A samereasoningexplainsthe limited influenceof the mission duration at 160 km altitude in figure 4.5. In any case,both figuresshow that in the presentideal casethe goalsdescribed in chapter 2 (5 mgal accuracyfor gravity anomaliesand 10 cm for geoid undulations with a resolution of better than 100 km) can easily be met, at least for a full tensor gradiometer. This conclusionis, however,a theoretical one since in reality the situation may and probably will not be so ideal. Nevertheless, the previous discussiongives more insight in the way in which various parametersinfluence the processand it shows the maximum obtainable result from gradiometry, of course under the present assumptions.
76
4.2.
subblocks
When trying to invert the subblocks (mblocks) of the normal matrix, it appears (and it will appear for other casesin subsequentsections) that some blocks are singular and can therefore not be inverted. The reason for this might be either a physical one or a numerical one. An example of a physical cause underlying such singularitiescan be found for the componentsVruand Vr" (both of which contain one differentiationin the crosstrackdirection). For a perfect polar orbit (1: 90") zonal coefficientscannot be estimated from these components,the crosstrack direction being always orthogonal to the direction of variation of the zonal coefficients (along the meridians). This is reflectedin the model by means of vanishing of the crosstrack inclinationfunctions,Ql for m:0. Consequently, m: O blocksfor these the componentsare singular. We may also encounternumerical singularities.Especiallyfor low orders (which constitute the largest subblocks,sincethe degrees containedin them run from rn / to .L) the differencebetween the largest and smallesteigenvaluemay become very large,so the block will be illconditioned(seealsosection4.2.6). In either case, we do not obtain propagated error variancesfor the potential coefficients belongingto those blocks. Or stated otherwise,the potential coefficients belonging to the singular blocks apparently cannot be estimated from the observations. However, when computing error degree variancesusing equation 4.9, a summation is done over all orders m. Leaving out certain orders, due to the singularities describedabove, would thereforeresult in incorrect valuesfor the degree variances.We would have, so to say,assumedzero error for those orders,where perhaps an infinite error would have been more appropriate. Therefore,the error r.m.s. plots included in this work have to be interpreted very carefully, always bearing in mind which rnblocks are left out. To this extent, table 4.5 shows which blocks appearedto be singular and were omitted from the analysis. For later reference, we included in this table all the casesyet to follow in subsequent sections. What can we do about this? Probably the best (but certainly not the most pleasant) answer would be not to compute degree variancesat all. The problems above first arise when computing the degreevariances,whereasthere is, of course, nothing wrong with the propagated error variancesfor each individual coefficient themselves,at least as far as they canbe estimated (i.e. the nonsingular blocks). There are, however,some reasonsfor using degreevariances. First of all, they are generally found to be very illustrative, giving a good picture of the gravitational field and its main spectral characteristics.Secondly,if we do not compute degree variances,we are left with a very large amount of individual error variancesto be comparedfor severalcases.Whereasthe number of degreevariancesttonly" amounts to.L (the maximum degreeof the analysis),the number of individual coefficients is (L + l)(L * 2) minus the nonestimable ones. It is hard to make good pictorial or graphical representations for so much coefficients, at least for high maximum degrees,like .L : 240 in our case.
77
analysis 4. GIobaIgradiontetric An alternative would be to compute order uariances.Then orders for which the subblocks of the normal matrix are singular appear in the graphs simply as gaps, but at least they would not interfere with other orders. But from table 4.5 we see for that often only the even or odd degrees a certain rn appear to be singular. Then we are faced again with a similar problem as above. Furthermore, one is not used to interpret order variances,so we will not show this alternative here. We already suggestedabove that instead of zero error, an infinite error for the coefficientsof the singular blocks would perhaps be realistic. So we could try, in someway, to fill in the shortagesoccurring due to the singular blocks. For example, as one could regard the nonestimablecoefficients contributing to the omissionerror variances from somedegreevariance and thus insert for them the signal degreeorder potential coefficientmodel. But since from a model or the value of the coefficients for large parts of the spectrum the signal is much higher (partly by severalorders) than the error) the results would become extremely bad (very high error r.m.s. from someexisting . values) Another possibility would be to insert the error variances potential coefficientmodel. But especiallyfor the Iow degrees(where the influence of the singular blocks is relatively large) the error of the existing models is much smaller than that from gradiometry. Our results would therefore hardly change. Even a multiplication of the model error varianceswith a factor 3 before adding them to the degreevariancesdid not have much influence. Furthermore, both methods, inserting model signal or error variancesfor the singular blocks, lead to the problem of how to interpret the correspondingobservational model belonging to such a strange combination of (error) variances. Concerning the additional error variancesfrom some model, one could view upon them as apriori information added to the estimation procedure. But those model error variancesthemselvesare the result of another, former, estimation process,leading to a strange mixture of data, models and methods. we not pay any special attention In the remaining of this chapber, will bherefore the to the phenomenonof the singular subblocks. For reference, singular blocks are the influenceof the singular blocks on the computed listed in table 4.5. In any case, degreevariancesis not so large, as might be illustrated by the following test. We for computedthe error r.m.s.values the case1:90o and B^;n:0 for the six tensor componentsleaving out exactly the same orders for all components. From table 4.5 . one can seethat this means leaving out m : 0 (both even and odd degrees) Apart r.m.s. curves, the ratios between the error from a slight change in the individual r.m.s. curves,as comparedto figure.4.2, did not alter. The error f.romV* andV* remains largest, followed by that from Vr, , lhen V* and Vy" and finally Vr, . 4.2.4 Band lirnitation
a for In the previoussectionwe assumed the gradiometricmeasurements O.OlEl\/Hz < lgrr"l ( oo. In bandwidth 0 white noise error spectrum over the full measurement in the time domain will have equal precision,i.e. that case all measurements %1(t)
78
Table 4.5
Singular subbloclcs th,enorrnal rnatrir. ne" rneanseuen deof grees, oott means od,d, degrees, * means both eaen and od,d d e g r e e s . F o r t h e c o m b i n a t i o n t{ V r r , V " ! r V " " r V s y r V y " r V r " } , m { V y s , V y " , V r " } a n d .{ V r u , V r r } n e u e r s i n g u l a r  b l o c k s a p p e a r . order rn
I 9U900 920.3
tl;U 4 0
7
V,,
95".3
920.3 YU900
0
4
vxu
U 4
r r
e e
e e
e e
e e
0 0 4
U 4
V
o e
o
0 0 4
U
vu, 4 0 0 4
U 4 0 0 4
U 4
f
+
Vu"
95'.3 92'.3
9U" 90" 92".3 95".3 92".3
9U" 900
+ + +
+ e e e e e e e e e e
Vt.
0 0 4
U 4 0 0 4
+
e
zv,u + V,,
o e
79
4. GIobaI gradiontetric analysis the covariancematrix Q is a scaledunit matrix. As a result of Parseval'srelation this measurementprecisionwill in our casebe 0.01 E. In reality the error spectrum of a gradiometerwill only be white in a limited band, betweensomelower band limit Brn;n and an upper lirrrit p^o, The measurementfrequencieslB6 l below B^;n and above Brno, wlll either not be present in the signal at all due to the sampling rate or will be distorted too much due to instrumental and environmentalerror sources (coloured noise). An upper band limit 9*o, is causedby the sampling rate with which the measurementsare taken along the satellite'sorbit. For Aristoteles this sampling rate is frequencypresentis 0.125Hz (twice at foreseen 4 s, so that the highest measurement the sampling rate or Nyquist rate). At an altitude of 200 km, the orbital period of the satelliteis approximately5310 s, so that 0*or:5310 x 0.125a: 664 c.p.r. We limit our error analysisto .L  240, so also l,tl and m are limited to.L. As a result, c.p.r. frequencyincluded in the analysisi" l0**l:254 the highest measurement g^or. An upper band limit as discussed here will therefore not which is well below above causeany problems. Note that the choiceof L :240 means that frequencies although presentin the signal,are not usedin the analysis. 254c.p.r., A lower band limit is the result of instrumental and environmental influences. thermal stability) can only be mainFor Aristoteles instrument stability (especially a maximum period of 200 s. This limits the white noise error band to tained over 200 s, i.e. the lowest measurewith a period not exceeding frequencies measurement with aperiod exceeding mentfrequencyis 0.005Hz, or 0^;nx27 c.p.r. Frequencies A l 2 O O s ( 1 8 1 ,<.2 7 ) w i l l b e p r e s e n t i n t h e s i g n a l , b u t w i t h d e g r a d e d a c c u r a c yl . B , 1 can be assumedfor such coloured noise situaerror behaviour for thesefrequencies lower band limit of 27 c.p.r. (with all an tion (Schrama,1990). Therefore, absolute here. below this Iimit removedfrom the analysis)will not be discussed frequencies Finally, nongravitational orbital effectsdisturb the spectrum, mainly below 4 c.p.r. from the analysis can be considered (seesection 4.2). Removing these frequencies appropriate to account for these effects. One could think of this as assuming an infinite error for these frequencies. A 0^;n of 4 c.p.r. should be consideredas a minimum, also for the colourednoise situation. of In this section we will investigatethe consequences a lower band limit of 4 white noisefor the full error spectrum abovethis limit. A c.p.r., thereby assuming LIB error behaviourbetween4 and 27 c.p.r. and colourednoise situation, with " in white noiseabove27, will be discussed section4.2.8. bandlimitation describedaboveis that the covariof One of the consequences lhe is ancematrix of the measurements no longer a scaledunit matrix. In particular, the matrix will be a full matrix, however,with a certain favourablestructure. covariance It will have constant diagonals(socalled Toeplitzrnatrir) and it can be shown that, with such covariancematrix, orthogonality properties of the trigonometric funcso tions are preserved, that the normal matrix will still be blockdiagonal (compare 4 section .2.1). I of Another consequence (lower) band limitation is that somedegrees below the
60
value of Bp;n cannot be estimated from the measurements any more, since for those degrees the total power content mainly (or solely) comes from the measurement frequencies below Brn;n. This can be seen as follows. Removing from the analysis the frequencies below Bro;n means that only the Bp* for which l0**l 2 Ba;n are included. Let us first consider the case Bpro ) 0, so that we have Bp^ ) Br";n. With gk^ : k+ ma (where a : NalN,) and k  I  2p, o I p < I we find that only those degrees / are present for which I ) 0*;r. The other case, Bp,o < 0 leads to / > (1 a)10^;n. If we take N" :727 and N4 : 45 (Schrama, 1990),whichfulfil the conditions stated in section 4.2.1 and which refer to the Aristoteles mission, we find that e.g. for /min:4 d e g r e e sI < 4 a r e n o t p r e s e n t a n d f o r / m i n : 2 7 degrees
I <26. Figure4.7 shows aposteriori the error r.m.s.values the casethat 0^;n  4. for
Frequencies below 4 c.p.r. were removed from the analysis and for the remaining frequenciesa white noise error spectrum was assumed. A remarkable difference with figure 4.2 is the zigzag pattern which occurs for certain components and combinations. This has to do with the step sizeof 2 for the index /c,cf. equation4.8. Given a certain value of m a BIimitation also implies a klimitation, since 0k^ : k+ma. This is especially clearfor low ordersm (sinceo is smali),where a relatively large part of the total signal content is concentrated.The consequences limiting of ft are different for even and odd degrees,due to the step size of 2. For example, consider the casem: O. This is the easiest casesincefor m = 0 we have0r*: k. If we choose0*in: 1 this meansall kcontributions for k : 0 are removed. But only the even degrees include a k :0 part, so that the odd degrees remain unaffected. For other values of /rn;n a same reasoningcan be applied. For higher valuesof m, limiting the 9n^ results in a difficult pattern of dropped &, rn contributions. It is not so clear any more how this limits the /csummationfor a certain degreeI and how this is related to the parity of /. But since the contribution to the total signal content for higher orders is relatively small, the main effect of Blimitation is to be expected to come from the low orders. It can be seen from the figure that, for certain observationtypes, the zigzag pattern damps out with increasingdegree. Two explanationscan be given for this. As we have seenabove,the main effectof a Iower band limitation comesfrom the low orders. Low degrees are thereforemore affectedthan high degrees since a relatively larger part of the total signal content comesfrom the low orders. Furthermore, as the r.m.s. curves approximate explained in the previous section,for higher degrees more and more a straight line induced by the attenuation factor (Rlr)'n". It should be noted that (part of) the zigzag pattern could be caused by the occurrence of singular rnblocks, as discussedin the previous section. Compared to the case 1min : O significantly more rnblocks are singular, howeveroften only the even or the odd degrees, table 4.5. see The singular mblocks make it furthermore difficult to compare the overall level of the r.m.s. curves in the caseB*;n : 0 with that of 0*in : 4, especiallyif the causeof the singularitiesis numerical instability. That the situation becomesworse
81
4.
Global
gradiontetric
analysis
 w
xK,xy,xz,
w'w'zz
. q) l
w,yz,zz
o tr o o
A H
AI I
xy t{ A @ . l c) lr o o A i
xz
xx
l
ol
li
zz
Aco
q) t o o A i a l
 zw+zz I:90
w.zz
Fio:4
ol I
Figure 4.7
Propagateil error frorn gradiometry. Specificationsas in fig' ure 1.2 etcePt now B*;n : f .
E2
in the latter case (i.e. larger propagated error) is not very surprising, but only a comparisonfor each coefficientindividually would give a clear picture of the exact try to compare figures 4.2 and 4.7 we amount of distortion. If we, nevertheless, see that indeed mainly the lower degreesare affected. R.m.s. values reach up to the two orders of magnitude higher valuesin the caseB^;n: 4. As a consequence, minima of the curves are moved in the direction of increasingdegree,around I : 60 for Brn;n: 0 but around I  l2o for B^;n : 4. A last remark concernsthe discontinuousbehaviour for some components for of low degrees(l < 20, e.g.Vs"). This is a direct consequence the singular rnblocks. As discussedin the previous section, the singular rnblocks are removed from the analysisso no contribution for those orders is included in the r.m.s. values,resulting in too optimistic error estimates. From table 4.5 we see that singular mblocks are relatively more affectedby Iow mainly appear for low orders. Sincelow degrees the bumpinessfor low degrees. orders, this explains Figures 4.8 and 4.9 present global r.m.s. values for surface gravity anomalies and geoid undulations for the present caseB*;n: 4, figure 4.8 for severalmission durations and figure 4.9 for severalmeasurementprecisions. Both figures give the results for a full tensor gradiometer (i.e. the combination of all six gradients), the r.m.s. being computed with a maximum degreefor the omissionerror of 1000 and a Comparisonof thesefigures smoothing factor B1representing1' x 1' block averages. is with figures4.5 and 4.6 revealsthat the r.m.s. for gravity anomalies hardly affected value of Bro;n,whereasthe geoid undulations show clear differences.These by this differencesnot only concern the overall higher level of r.m.s. Also the influence of both the mission duration and the measurementprecision is noticeable now, even for lower altitudes, where it was not in the previous section (seethe differencesat h : 160 km). Geoid undulations constitute a smoother signal than gravity anomalies. The latter are firstorder derivativesof the gravitational potential, resulting in multiplication factors (/ + t) etc., seechapter 3. It is thereforeto be expected that the geoid undulations are more affectedby changesin the lower degrees(as is the case if we choose9^;n * 0) than gravity anomalies. Note that the high precision gradiometer(o :0.0001 EIJH, ) no* showsits value,the r.m.s. of the geoid undulations being for this o nearly the same for both gmin :0 as well as Brn;n: 4, for all altitudes. We have to bear in mind, however,that the figurespresent the results for a full tensor gradiometer. For the combinations {Vyy,Vy",V".} or {Vyy,V""} (as for Arisfor toteles) the results are somewhatworse,especially the geoid undulations. Global r.m.s. valuesfrom those combinationsare respectively15.6cm and 97.0 cm for geoid the valundulations(at 200 km,6 months missionand o:0.01 EIJH, ), whereas ues for the gravity anomaliesare hardly changing. Individual componentsgive even worse results. This shows the importance of combinations of components,i.e. the in need for additional observations, the band Iimited case. If other componentsare not available,Iike for Aristoteles (planar gradiometer,band limited, degradedlower techniques(like GPS) is spectrum), additional information from other measurement
83
4.
Global gradiontetric
analysis
see necessary, section4.2.8. Furthermore,in section4.2.6we will discussthe possibility of improving the solution by adding prior information, i.e. stabilizing the solutton collocation(Rummel et al., 1979). of by adding constraintsin the sense least squares 4.2.5 Polar gaps
A realistic gradiometric mission not only differs from the ideal case sketched in section 4.2.3 through band limitation. In reality the satellite is also not likely to fly in a perfect polar orbit. For Aristoteles two different inclinations are considered: 92".3 and 95o.3. Aristoteles is planned to fly for approximately six months in an orbit with inclination 95'.3 and another two weeksin an orbit with inclination 92".3. are available. A nonpolar orbit givesrise to polar regionswhere no measurements the inclination of the orbit is .I, polar caps,of size2x (190)', will arisewhich are If an not coveredby ground tracks of the satellite'sorbit. As a consequence incomplete global coverageis obtained which is expectedto influencea global recoverymethod as it is performed here. At first sight, one expects a polar gap, with size larger than half of the smallest wavelength to be recovered,to disturb the solution, at least for those short wavelengths(i.e. high degrees).For recovery,on a global scale, of gravitational features with very short wavelengthsthere are simply not enough (well distributed) data points. On the other hand, it is very difficult to exactly translate the missing of data in certain spatial regions to a possibledistortion of recoveredpotential coefficients.This may becomeclear from figure 4.10, in which the propagatederror for the case/:92o.3 is shown. Remarkably no substantial deterioration occurs compared to the case 1 90o, except for Voo. Whereas the result for some components becomesonly slightly worse, other componentsseemto becomeeven better. In spite of the fact that this outcome is confirmed by other investigations(Schrama,1990),it is difficult to give a clear explanation for it. Several aspects may play a role here. First consider the data distribution. In both cases,.I : 90o and I : 92o.3 we have the same amount of observations(namely a total ol Trf A,t observations). Although for a 92o.3 inclination the polar gaps occur, the data density in the remaining part (the may distort the band 2'.3 < 0 < 177".7) becomeshigher. Whereasthe polar SaPS improve the recovery of the higher data density may recoveryof some coefficients, others. On the whole, a small improvementmay occur if the inclination differs not too much from 90". If the deviation from 90" becomestoo large, the influence of the polar gaps will ultimately surpassthat of the higher data density' A secondaspect which may play a role is the behaviour of the inclination functions. Besideson factors like f1, (/ + t), k, etc. the total signal power contained in a specificdegreedependson the inclination functions Ftk* (or the crosstrack incliF n a t i o n f u n c t i o n s f * * f o r V r o a n d V y " ) . T h e y m a k e p a r t o f t h e s o  c a l l e ds e n s i t i u i t y r.\ H!,1, from tables 4.1 and 4.2. Let us focus here on one specificdegree coeffi.cients and order (i.e. fixed I and rn). What remains is a summation over k of inclination functions and a possiblefactor k or lcz (cf. tables 4.1 and 4.2). If we compute the
64
Ag IN MGAL
? s ib .+ xx,xy,xz,yy,yz,zz
o
r
I
140
160
180
200
220
240
260
N IN CM
* o O
@ F I
(\I
140
160
180
200
220
240
260
satellite height in km
Figure 4.8
85
A,g IN MCAL
r0 trtr) tr3
m L t
.9+
rO
zoo zzo
z4o z'60
*
o
GI
o o
q l ( )
t{
P
o
o
140
160
180
200
220
240
260
satellite height in km
Figure 4.9
precisions.SpecifiGtobatr.rn.s. ualues lor seueraltneasuretnent : f. cations as in figure 1.6 ercept now B*;n
E6
4.2.
tI
 w
xx,xy,xz, w,yz'zz w,yz'zz
+i o' q, I . o i l H a l a a ) E a
&
l
I
AI
as I
xy xz
xx
si qr o l
6 i
o r l U ) a
> H E l
F I
 w
zz I : 92.3 Zyy+zz w,zz Fio : 0
r r l c) i q o l ,
i1 H
92* E *
C\l I
120 degree I
180
Figure 4.10
87
analysis 4. Global gradiontetric inclination functions for a specified I, rn and /c as function of 1, not necessarily a maximum value is obtained for the F!r, at 1 : 90o. The value of some inclination for functions will increaseif ,I > 9Oo,somewill perhaps decrease higher inclinations. this behaviour. If a// inclination We did some test computations which confirmed functions would obtain a maximum value at 1 : 90o, one would expect the signal for amount to decrease I > 90", so that the propagatederror would increase.Since this is not the case,the signal amount may also increasefor higher inclinations (so . that the propagatederror decreases)Carrying out the ksummation for one specific l,m overall correspondingF,f, (someof which have a larger value, others a smaller) the result for / : 92".3 might either be Iarger or smaller than the same sum for (i.e. on the gradient). sensitivitycoefficient /:90o, dependingon the specific A third aspect applies only to the V"u and Vy" components. As mentioned in section4.2.3 the crosstrackinclination functions are all zero for I : 90o and rn : 0.  92".3 this does not This resulted in two singular mblocks, cf. table 4.5. For I happen any more, i.e. the potential coefficientsfor m: 0 are now estimable from This meansthat the m: O contributionis includedin the r.m.s. the observations. valuesshown in figure 4.10, whereasit was not in figure 4.2. This troubles a direct comparisonof the two figures. An inclination of 95'.3 has a very strong impact on the results, as can be seen in figure 4.11. Not only does the level of the r.m.s. increase,also the shape of the curves changesdrastically. Whereas it is expected that for this case the exchange between positive and negative influenceof respectivelydata density and polar gaps will result in the negative, inclination function behaviour and singular rnblocks may play an important role here. In any way, without any stabilization method or without any additional information to accountfor the data gaps at the poles, these results are not acceptable.
4.2.6
Stabilization
The problem of solving gravitational information at the surface of the earth from observeddata at satellite altitude is known to be improperly posed (also called illposed or illconditioned). A problem is called improperly posed if it does not meet at least one of the following three requirements: 1.) existenceof the solution, 2.) uniquenessof the solution and 3.) stability of the solution (Rummel et al., 1979), (Moritz, 1980)or (Neyman, 1985). In our case,where we try to derive detailed gravitational information at the surfaceof the earth from a relatively smooth (attenuated) signal at satellite altitude (downward continuation), we encounter problems with the uniquenessand with the stability of the solution. The former case applies of when some eigenvalues a normal matrix subblock becomezero so that the block singular, as we have seen e.g. for the V", and Vv, gradients for rn : 0 at becomes may becomeextremely small, 1:90o (cf. table 4.5). In the latter case,eigenvalues and errors in the data, but also roundoff errors so that the matrix is illconditioned greatly amplified during the inversionprocess.The block may of the computer, are
88
 w
i F t < l 0)
*.: P
0 r l
k O ) o l A a ^
> =
I (o I
l r l 0) A
s i 9 o l o k o ) q ) a ^
>{=
(o I
i F
r { l
0) A
q , I
I : 95.3
a o ) o l o.
@
Fio : 0
> :l E
I
Figure 4.11
89
even becomenumerically singular in this case. We have already seenclear examples of such improperly posed problems in the foregoingsections,e.g. the caseB^;n = { or the casef: 95o.3. For both casesthe instability is again illustrated by means of the condition number of the individual blocks of the normal matrix. As we know, the normal matrix consistsof blocks on the diagonal, decreasing size. There is one block for in eachorder rn and degrees of equal parity. Figures4.12 and 4.13 show the condition / number for the even degree blocksfor the case Bro;n:4 and 1:95o.3 respectively. The odd degreeblocks give similar results. The condition number rc is defined here as the largest eigenvalue )o, divided by the smallesteigenvalue ,\;,r:
. ^ \*o,
^^r^'
The blocks which are expectedto be numerically unstablecan be traced by comparing the inverseof the condition number rc1 for eachblock with the internal accuracy of the computer. In our case,all the computationsare done using FORTRAN REAL*8 (doubleprecision)numbers,i.e. some 15 or 16 significantdigits. Relativeerrors of the order 1015or 1016are likely to occur. Blocksfor which the condition number approaches (".g.nt > 1012) or exceeds this value are very illconditioned. Inspection of figure 4.13 leads to the conclusionthat for almost all componentsand combination of componentsthe blocks belonging to orders approximately below 20 tend to be illconditioned. Hence, also degreesbelow 20 are likely to give bad results. Also higher degreesare influencedby those illconditionedlow orders. For the case1nin:4, condition numbers for loworder blocks (approximately m < l0) reach even higher values,up to 1018or higher, as can be seenfrom figure 4.12. One of the possibilitiesto stabilizethe solution is to make use of additional prior inforrnation about the unknowns. In terms of least squarescollocation one adds the prior expectationsof the unknowns together with their covariancematrix. The solution then becomes(compareequation 4.5); 6: N1(A'Qt tf Prc') with the matrix N now defined as N: ATQrA+ Pl (4.15)
where c' are the prior expectationsof the potential coefficientsand P is their covariance matrix. In our casec' could contain the coefficients one of the existing of geopotentialmodels,such as OSU86F, with P the corresponding covariance matrix. For the purpose of a global error analysis,as discussed here, a more common choice for c' is zero with P one of the existing signal degreevariancemodels, e.g. Kaula's rule (Kaula, 1966). Here we take the degree variancesfrom a Tscherning/R.pp (TR) model (Tscherning and Rupp, 1974). Such a choice implies that we consider the prior value of the potential coefficients be zero and their variancesto follow to
90
I:90
F^;^:4
even degrees
Figure 4.12
Condition nurnberlor blocksol tlt,enorrnal matrit for euen d.e' as grees.Specifications in figure 1.2 ezceptp*in : 4'
91
rO
;
4
xx,xy,xz,
w'w'zz
w,yz'zz
trs
{lo
xx
_xz
lO
;
trgr o
_zz
w,zz Zyy+zz
C)
url)
I : 95.3
Fi. : 0
even degrees
Figure 4.13
Condition number for blocksof the normal matrix for eaen degrees.Specifications in figure 1.2 ercept I : 95".3. as
92
the TR model. The solution is in this case again given by equation 4.5, but now with the normal matrix as in equation 4.15. This means that the solution is a kind of weighted mean of apriori information and gradiometry results. where the solution was Figure 4.14 shows the error r.m.s. for the caseB*;n:4 stabilized by adding to the normal matrix a diagonal matrix containing the TR degreevariances.This figure should be comparedwith figure 4.7. For completeness the curve for the TR model is added. We see that the zigzag pattern is damped for most solutions (clearly visible for Vy"). On the other hand, the curves tend to approximate the TR curve. Questionscan be asked about the value of the, in especiallyfor higher degrees,since they obviously this way, estimated coefficients, do not add much information to the apriori known model degreevariances.Above degree240 all curveswill follow almost exactly the TR model curve. This is a reason why we choosea maximum degreeof 240 for our computations. Above this degree, potential recovery is not likely to give any new information, at least not for more realistic (and thus illconditioned) situations, like the one in figure 4.14. The stabilizedsolution for,I:95o.3 is shown in figure 4.15. Compared with the figure 4.11the stabilization certainly has much influence,sincefor lower degrees error r.m.s. curves now show a behaviour comparableto figure 4.2, though a little the curves tend to follow that of the TR model. disturbed. For higher degrees Following (Xu and Rummel, 1991) one may view upon stabilization as biased estimation. As a purely mathematical technique, one of the purposes of biased estimation is to control instabilitv. It is defined as:
6 b: (f
e  rA + K)  r nr q t ,
where the index D means "biasedestimate" and K is somearbitrary but positive definite matrix. This way of looking upon stabilization becomesimportant if the prior information added to the problem is not correct (Xu, 1991). In fact, this is exactly the situation in our stabilization method (where we take for Kr the covariance matrix P of the prior expectationsof the coefficients)since the degreevariancesof , but are the TR model do not describeerror variancesof the potential coefficients, possiblemagnitudes of the coefficients themselves(signal variances). Following (Xu and Rummel, 1991) the magnitude of the expected bias in this case can be computed as B;a(6) : (Ar Qt,q + Pr)r Pr c where we should insert for c, in principle, the true potential coefficients.In general one does not have the real coefficients,so the bias is approximated by replacing the real coefficientswith either the biased estimates themselves,or some model coefficients.In (Xu and Rummel, l99l) it appearedthat the former choice (biased leads to too optimistic conclusionsabout the bias. Choosing estimates themselves) coefficientscomputed from the model which was used for the K matrix appeared to give more conservative results.
93
CO
F.
{)
xx,xy,xz,
A(D . l o
ti
w'w'zz
W'Yz'zz
o o
A H
ol I (o I
k q,
ac)
9i o tr o o A i a l
H
I
Rt
(o I
L q)
zz
ACo +i o, o o
k r D O A H
I:90
a 0l
Figure 4.14
Propagated. error frorn gradiornetry. Specificationsas in figure 4.2, exceptFmin : l, and nou the solution is stabilizedby n'Leans a Tscherning/Rappmodel addedto the norrnal matrix. ol
94
@ I
F{ O
O) l
".; o Pd X
L q) A a
C\l I @
xy
t{ O) O l +i
xz
xx
$ o
, l
l ^ t . **
q) A a
ol I
(D
r< o) (l)

il
P o
tr
I : 95.3
o U ) a
Fio : 0
stabilized solution
Al
Figure 4.15
Propagated error fronx gradiometry. Specificationsas in f'g' ure 4,2, emept I  95".3, and' nou the solution is stabilizedby model addedto the normal rnatria. rneansof a Tsch.erning/Rapp
95
4.
Global gradiontetric
analysis
 w
o s
U O d q) k
9rB i .
ts
ry
x2 xx
o s aoo
C' 0, k 9 f FIi r$
o s
ffJ,
 w
_zz
o $ { O
6
o o k 9 f s $
Figure 4.16
of Percentages the biaseswith respectto the OSU86F mod.el for the casep*in : 4.
96
o ro
(7J
xy xz
Y:(
o a l a00 d
c Q'
k
9 ( o
Ir. <l
o
@ c/)
zz
 w,zz Zyy+zz
I : 95.3 Fio : O
Figure 4.17
97
4.
Global gradiometric
analysis
For both stabilized solutions presentedabove (9^;n : 4 and 1 : 95".3) we computed the bias using the equation above. We choseas coefficientsc the same valuesfrom the TR model which were used in the P matrix, but with the sign taken from the OSU86F potential coefficientset. Since we) at this moment, do not have tttruett coefficients estimated from gradiometer measurements,we cannot comPare may give the most the biaseswith such coefficients.But since using TR coefficients conservativeresult, cf. (Xu and Rummel, 1991),we may compare them to e.g. the of of OSU86F model. To this extent, the percentages the degreevariances the biases with respect to those of the OSU86F model are computed and shown in figures4.16 and 4.17. Depending on which component or combination of components is used for the analysis,large parts of the spectrum show biaseswhich exceed20 To, even However, the case1:95o.3 this only happensfor for rcO% of the model variances. the results for all six componentstogether, for V"" and higher degrees. Note that for the combinations {Vor,Vy",V""} and {Vrr,V""} are not so bad at all, the degree variancesof the bias in thesecasesreachingapproximately 25 Toof the model values only for very high degrees (l > 22a). The accuracyof biasedestimatesis no longer given by the inverseof the normal matrix only. The bias term needs to be added to the aposteriori error. The socalled mean squarederror of 66, MSE(66), obtained in this way is given by:
(4.16)
with N from eq.4.15, see(Xu and Rummel, l99l). Figures4.18 and 4.19 show the percentages this MSE with respectto the propagatederror (as coming from the of inverseof the normal matrix only) from figures4.14 and 4.15, for the casesp^in: 4 and f : 95o.3 respectively. In these figures the straight horizontal line indicates the 100 Vo Iine. If the bias would not have any effect, all percentageswould have smaller than 100 indicate degreesfor which to lie on this 1OO line. Percentages Vo (no bias term included) is larger than the MSE. It means the propagated error the previously shown results (without the bias) were in fact that, for those degrees, too pessimistic. In this casewe don't have to worry that stabilization leads to too indicate degreesfor which inclusion optimistic error estimates. Higher percentages of the bias causesa higher aposteriori error, so for those degreesthe results from figures 4.14 and 4.15 were too optimistic. reach high values of nearly 160, they Whereas in figure 4.18 some percentages show no substantial degree dependent systematic deviation from 100 %. On the average,the componentsVy"rV* and V* seemto give lessfavourableresults due to the stabilization (if the bias is included in the error), whereasthe other components seem to become slightly better. For .I : 95o.3, on the average,for most components the percentageis smaller than 100 (at least for large parts of the spectrum), indicating that if one accountsfor the bias, estimateswill becomebetter. So, although the estimatesbecomebiased,this bias stayslimited, the aposteriori error (which has to include the bias, so it is rhe MSE) being positively affectedby
98
AT (D
_  @ !u_ o
P q)
t t D *
a rrl
o AI @
_  @ !u o
d q)
k O < . Al)
zz
 w,zz Zyy*zz
I=90 Fio:4
ol Figure 4.1.8 Percentages the MSE with respectto the aposteriori enor (NL onlY) for 9*;n : 4'
99
()*
k
aii
o
GI @
_
 . o
o o
_xz
Fe
k Y $ R U )
xx
zz
w,zz ?yy+zz
I : 95.3
Fio : 0
Figure 4.19
100
method of error analysis 4.2. Colombo's it on the average,especiallyfor the caseI : 95o.3. Thus, stabilization seemsto be a satisfactory means of obtaining a stable and meaningful solution. Ornission, cornrnission and smoothing
4.2.7
In sections4.2.3 and 4.2.4 global r.m.s. valuesfor gravity anomaliesand geoid undulations were presented. They were computed using equation 4.14 and therefore the included commissionerror and omissionerror. The commissionerror represents that part of the spectrum for which error. It only concerns propagatedmeasurement i.e. can be estimated from the measurements, for degrees the potential coefficients < .L (seesection4.2.2). The omissionerror includesthe neglectedpart of the spec/ trum above ,L, in principle up to infinity, but in practical computationstruncated at some maximum degreelmrs. Furthermore, due to the Wiener filter, the omission error includes that contribution from degreesbelow .L for which the signal to noise ratio is smaller than one. This part, in fact, stems from the imperfection of the measurementprocess.For the omissionpart we usedthe TR degreevariancemodel. In the results presentedabovewe truncated the summation for the omissionpart at Jrncs : 1000. The magnitudesof the degreevariancescomputed with the TR model gradually with increasingdegree. The signal power contained in the upper parts of decrease the spectrum therefore will be small. In order to see whether a maximum degree of 1000 is reasonableor not (i.e. the power above degree 1000 is negligible) we compute here the total error once again, but choosea maximum degreeof 10,000. Thereby one should keepin mind that such a model is only basedupon observational between300 and information up to a certain maximum degree,typically somewhere 1000, above which it is speculation. This may become evident by comparing the high spectral part of variousexisting models,such as Kaula's rule (Kaula, 1966) TR , Jekeli/Moritz (Moritz, 1980)and (Jekeli,1978). (Tscherningand Rapp, 1974)or Table 4.6 shows the global (commission,omission and total) error for surface gravity anomaliesand geoid undulations in three different situations for the cases e l m r s : 1 0 0 0 a n d l m r s : 1 0 , 0 0 0 . F o r r e f e r e n c a l s ot h e c a s el m r s : 2 4 0 i s a d d e d error only due to Wiener filtering). It can be seenthat the error for (i.e. omission geoid undulations does not change if the omission error is summed up to degree 10,000compared to the casewhere it is summed "only" up to degree1000. Only if could the numbers would have been presentedwith more decimal digits, differences be detected,but then the numbers would pretend to be very accurate,which is not the case. Since the geoid signal is relatively smooth, this result is not surprising. slightly (0.0t mgal), the gravity field The error for gravity anomaliesonly increases being a little rougher than the geoid. In view of the present accuracy, this test confirms our procedure to carry out the omission error summation only up to a maximum degree/rncs: 1000. below 240 is relatively small, at least The part of the omissionerror for degrees : 9Ooand B^;n : 4,1 : 9Oo.For the other case(1 : 95".3) for the case Brr;n: O,/
101
Table 4.6
Comrnission,om.issionand,total error t'or the obseraationcomb i n a t i o n{ V " r r V " o r V r " , V y y , V u " r V ? ) a n d h : 2 0 0 k r n , A t : o 4 t , 7 , : 6 m o n t h s , : 0 . 0 1E / t / H z , L : 2 1 0 a n d 0 " : 1 " . gravity anomalies (mgal)
0.40
8.39
.I : 95o.3 and B*;n = 0 lmxs :240 lrncs : 1000 lmrs : 10,000 5.90 5.90 5.90 5.55 6.63 6.64 8.10 8.87 8.88 29.78 29.78 29.78 24.49 38.56 25.86 39.44 25.87 39.44
102
4.2.
Table 4.7
Cornrnission,omission anil total error for the obseruationcornand h : 200 krn, L,t : bination {vrr,vrr,vr",vyyrvu"rv":) :2f0,0":1" and" , 4 s , T , : 6 r n o n t h so : 0  0 1 E / 1 / H z , L solution' a stabilized,
0.38
8.38
103
the propagated error is so high that a large part of the omission error comesfrom degrees below 240. Obviously,for this case,the gradiometercontributes little to the outcome. The large errors in the case 1:95o.3 are due to the instability of the solution, as pointed out in the previous sections. For comparisonwe therefore also give a similar table (table 4.7), now computed from the stabilized solutions. The cases grnin : 0, f : 90' and 0*in : 4,1  90o are hardly affected. Compare this to the figures 4.14 and 4.15. For all cases,the differencebetween lrnss : 1000 and l m x s : 1 0 , 0 0 0 i s s t i l l v e r y s m a l l . O n t h e o t h e r h a n d , f o r t h e c a s e1 : 9 5 " . 3 , t h e contribution to the omission error from degreesbelow 240 decreasessubstantially, as does of coursethe commissionpart. Another point which is important in this context, is the smoothing of the total error. If we look at equation 4.14, we seea smoothing operator B1is included. This is done, because a global recovery of the gravitational potential always has some limited resolution as a result of the finite sampling distance of the measurements. This sampling distancelimits the maximum obtainable degree,thereby introducing an error to the potential solution. In terms of equal angular blocks,the resolution to be obtained from a gradiometric analysisup to somemaximum degree.L is typically I8O"f L. If no smoothing is carried out, the global r.m.s. for gravity anomaliesor any other quantity computed from a potential coefficientset which is derived from are a global analysis,representsthe error in discretepoints in which all frequencies present. This error would be very large, mainly due to the relatively poor determined in higher frequenciesand the frequencies the omission part. This is not realistic, as we know that the analysis aims for, and has, a limited resolution. For certain applications (especially global gravity field investigationsfor the purpose of e.g. oceanography),the r.m.s. values should therefore be consideredas blockaverages over blocks with size induced by the maximum degree of the analysis. For other the high frequencypart should investigations, applications,e.g. regional geophysical In be obtained from a densenetwork of local gravity measurements. order to obtain blockaverages, smoothing operator B1 is used which in fact damps out the high a of aiming at a decrease the sampling error. frequencies, In our analysis,we used a maximum degreeL : 240 for the potential recovery, correspondingto a resolution of 0o.75x 0'.75. But as we have seenthat the higher degrees only poorly determinedand sincethe goal for gradiometry is a resolution are to of 100 km (corresponding 1o x 1') we introduceda blocksizeof 0": Lo for the in smoothing operator. Nevertheless, table 4.8 the influence of the smoothing is shown. There we computed the global r.m.s. for three cases: 0" : 0o (i.e. no s m o o t h i n g ) , 0 " : 0 " . 7 5 a n d 0 " : l o , c f . ( R a p p , 1 9 8 9 ) .F r o m t h i s t a b l e w e s e et h a t the influenceof smoothing is much larger than that of the maximum degreeof the omission error. No smoothing gives unsatisfactorilyerror estimates. Again, geoid undulations are much less affectedby both aspectsthan gravity anomalies. It can alsobe seenthat the two aspects(inclusionof omissionerror above,L on the one hand and smoothing on the other) have opposite effectson the global r.m.s. Inclusion of
104
Table 4.8
0c: O"
0.29
r.46
0.81
4.40
0.55
3.L4
lmrs:10,000
0": I"
lmrs :240 lrncs : 1000 lmrs:10,000
the global r.m.s. becausewe have to account for omission error above tr increases the the neglectof the higher spectrum parts. But smoothing decreases global r.m.s. of these higher degrees,which, nevertheless, by gradually removing the influence remain present. 4.2.8 Aristoteles
In previous sectionsdeviations from an ideal casewere investigatedby considering either a band limitation or a nonpolar orbit. A combination of both effectswill be investigatedin this section,aiming at the situation of ESA's Aristotelesmission (see section 2.3). We do, however,not include in our analysisthe possiblecontribution we from GPS observations, only give "gradiometry only" results. An error analysis for the combination "GPS * gradiometry" can be found in (Schrama,1990) or (Visser,1992). Concerningthe band limitation we refer to section 4.2.4. First of all we choose B*;n equal to 4 in order to account for the nongravitational effects. Secondly, approximately to Aristoteleswill have a lower band limit of 0.005Hz, corresponding
105
4. GIobaIgradiometricanalysis 9 : 27. Below 27, the frequencieswill be available but with degraded accuracy. The 0.01 E/y'Hz white noiseerror spectrum appliesto the frequencyband between0.005 and 0.125 Hz (the latter approximately correspondingto 0*o, : 664). Following (Schrama, 1990) a LIB behaviour is assumedfor the frequenciesbetween 4 and 27. Between 4 and 27 the precision per frequency then becomes (2719*^). o6 with in oo:0.01 El\/Hz. Such a colourednoisesituation will be assumed the present section. As for the inclination, we choosehere I : 92".3, despitethe fact that Aristoteles will fly for 6 months in an orbit with inclination I : 95o.3. The reasonis that in the two weeks after this 6 months phase, in which the orbit of the satellite will have an inclination of 92o.3,a so densecoverageof the polar regions is achieved(of course a small cap of 4o.6 excluded) that the complete mission may safely be assumed to be flown in this 92o.3 inclination orbit. The important mission parametersand specificationsare listed in table 4.9. Table 4.9 rnission. Characteristics the Aristoteles of oradiometer inclination
I :92".3
components Vyy,V". band limitation satellite height sampling interval mission duration 0^in: 4 h:200km At:4 s T, :6 months
error spectrum O.Ol El\/Hz white noiseabove27 c.p.r. lf B behaviour between4 and 27 c.p.r.
Figure 4.20 shows the error r.m.s. valuesfor the present specifications.In this figure again the results for all l0 componentsand combinations are given. Since Aristoteles only measures Vro and V"" with sufficient accuracy the lowest of the and for three graphs is of specialinterest here. The others are included for reference preserving a uniform presentation. Comparing with figure 4.7 (in which p*in : 4 and .I : 90") it appearsthat the latter is little worse, despite the fact of complete global coverage when I : 9Oo. This fact was already observed and discussed in section 4.2.5. As a result of the lower band limit and the coloured noise the results for low degrees(approximately below 30) are not satisfactory. For this reason also a stabilized solution was computed for the present case (shown in figure 4.21), which, however,appears to have mainly effect on high degrees.Comparing the unstabilizedwith the stabilizedsolution one can seethat some
106
F{ c) A @ . l ID
li
o o
A d
C\I
I (o I
tr 0) A @ ID
xz
xx
o
k
o o
A H
NI
I
@
lr q) A @ . l 0)
zz
 zw+zz I : 92.3
w.zz
Fi: 4
lr
o o
A H
ol I
Figure 4.20
Propagotederrcr rorn gradiornetrywith the follouing specifi.ca' g p o . 3 ,9 _  , " : 4 , 7 , : 6 m o n t h s , , t : 4 s , h : 2 0 0 L t i o n s :I : km, o : 0.01 E/t/Hz aboue27 c.p.r. o,rldlf B behauiourbetueen j a n d .2 7 c . p . r .
107
4.
F{ o)
A(o . l c, o tr
oc)
A H
=
ol I
k A C o . l q)
c r o
O . H
=
u ) l
ol I (o
li ID A @
. o
tr
I : 92.3
o o
A e
v ) l
F::
ol
 0
180
240
Figure 4.21
108
Table 4.10
unstabilizedsolution
vuu
ZVyy *Vzz
V"" {Vuu,Vu}
stabilized solution
componentsbecomeworse for lower degreesin the latter case (".g.Vur).However, this conclusionmay be misleading,since contributions from singular blocks are excluded in the unstabilizedcase(comparetable 4.5) and sincewe saw in section4.2.6 that for a proper descriptionof the error of a stabilizedsolution we should compute the MSE, equation 4.16, which includesthe bias term. Finally, table 4.10 gives global r.m.s. values for surface gravity anomalies and geoid undulations for the observation types Voo,2Vau1V"", V"" and {VrurV""). Also the results of the stabilized solution are included. The numbers in this table (i.e. smoothing operator included). The maximum represent1o x 1" block averages degreefor the omissionerror is 1000. The most important conclusionfrom this table is that if geoid undulations have to be derived from the presentmission scenario,additional information is necessary since the results do not meet the goalsfrom section2.2. Even the stabilizedsolution givestoo large errors for geoidundulations. The causeof this is the lower band limit of 4 c.p.r. and the coloured noise of the lower part of the measurementspectrum, which influencethe geoid undulations rather heavily due to the relative smoothness on of the signal. Gravity anomalies, the other hand, already give satisfactoryresults, at least for the stabilized solution (although of coursebiasedin this case). They are than geoid undulations. Obviously,the higher part more sensitiveto higher degrees
109
4. Global gradiontetricanalysis of the spectrum is determined well enough by the gradiometer. In (Schrama, 1990) or (Visser, 1992) it is shown that additional GPS tracking information can be used to obtain long wavelengthgravitational information, in which casealso geoid undulations will be derived with the desiredprecision. The two techniques,GPS for the lower part of the spectrum and gradiometry for the higher part, are thus complementary, and the combination of the two on Aristoteles will meet the requirements of section2.2. Finally, is can also be seenfrom table 4.10 that stabilization in the presentcase mainly influences the omission part of the total error. This can be explained as follows. In the unstabilized case, the propagatederror is relatively large, but the its Wiener filter decreases contribution to the commissionpart of the total error' thereby at the same time increasingthe omissionpart. Sincewe have seenthat the above .L, is small, the Iarge value of the omissionpart, due to neglectingthe degrees be mainly due to the mentionedfilter. In omissionpart in the unstabilizedcasemust the stabilized situation, the propagatederror givessignificantlybetter results (compare figures 4.20 and 4.21), causingthe filter to letting passmore commissionerror substantially. (howeverwith lesspower). As a result the omissionpart decreases aspects
4.2.9
Sorne cornputational
The software for the error analysisruns on a CONVEX vector computer. This is equipped with 4 CPU's a socalled SIMD machine (see section 3.3), nevertheless the fact that the optimalwhich offer some possibilitiesof parallelization. Despite ization capabilitiesof this computer (and especiallythe vectorizationtechniques)in general greatly reduce execution time, a single run of the error analysis program takes as much as 500 CPU seconds(for all 10 observationtypes). This, of course, dependsmainly on the maximum degreeL (and minimum degree)of the analysis, the mentioned time referring to an analysisup to degree240. The number of degrees up to tr determines the size of the blocks of the normal matrix, the range of the orders rn between mmin and mmar determines the number of blocks in the normal matrix, though one usually takes mmin equal to 0 mmat equal to .L. Whereas one often expects matrix inversion to account for the major part of CPU time, it appearedhere that setting up the normal matrix blocks is more costly. This is mainly due to the computation of the inclination functio"" 4f. and F,l, which is responsiblefor approximately 46 Yo of the total CPU time. One could suggestto compute the inclination functions only once and store them for future reference. This costs, however, much memory space. Storing all inclination functions and crosstrack inclination functions up to degree .L requires the storage of 3) numbers (Sneeuw,1991a). If representedby t tt, + \ (L + 2) (4 L + for double precision numbers this corresponds, L:240, to over 71 Mbyte, for one the memory costs may be too much. inclination only. For many inclinations, We tried two methods for the computation of these inclination functions. The first one is describedin appendix C and is based on an algorithm used by Wagner
110
(1983). It evaluatesa kind of unit potential function at equidistant points along a great circle (which representsthe satellite's orbit). With an FFT routine one obtains the Fourier coefficientsof this function from which the inclination functions can be derived. For the evaluation of the unit potential we need to compute the Legendre functions. Even if one makes optimal use of the symmetry properties of these functions, cf. (Schrama,1989), the largest part of CPU time goes to the recursivecomputation of the Legendrefunctions. The FFT is relatively fast (as it should be, consideringits name). The secondmethod for computation of the inclination functions is by using reprogrammedby N. Sneeuw(Sneeuw,1991a) current relations. A subroutine package based on an algorithm by Emeljanov and Kanter (1989) was used for this purand pose, with a supplementary routine for the crosstrack inclination functions F,lfor (Sneeuw,1991b). Sincesomeof the recurrences the inclination functions are unto occur, extensiveuse is made of scaling techniques, stable and overflowsare likely which becomeespeciallyimportant if the inclination functions are required for high degrees. betweenthe two methods are to be mentioned. First of all, the A few differences second(recurrent) method appearedto be slower. One of the main reasonsfor this is that recurrent relations are (almost) not suitable for vectorization. Furthermore, the method of scalingmakes extensivelyuse of conditional statements,which, especially if used inside inner D0loops,prevent vectorization. For the FFT in the "unit potential" method a FORTRAN programmed subroutine from a CONVEX mathematical library was used, which is especiallytailored for vectorization. A more conventionalFFT routine, made for use on scalar computers,proved to be up to 6 times slower for FFT's of the order 1024if run on the CONVEX. Whereas the two methods give identical results in nearly all cases,small differencesmay occur in special situations. The latter are mainly the situations when the inclinations functions ought to be zero. The first method will almost always but the secondmethod may give small valuesof the give zero valuesin these cases, order l015 or l0ro due to numerical roundoff errors during the recursions. For example, if I : 9Ooand n1 : O, all Ff* are identically zero, resulting in singular blocks of the normal matrix for the componentsVruandVs". The recursivemethod, however,would result in F!* valuesof the order of 1015, so that not all elements of the rn : O blocks will be equal to zero. If none of the diagonal elementsof such block is exactly zero, the block may not be singular, so that, after inversion, very large values for the aposteriori variancesappear. Therefore,care has to be taken with the recursivemethod. Sincewe do not need the completeinversenormal matrix, but only the diagonal elements,in reality the inverseis not computed at all. A Cholesky factorization is carried out on the normal matrix N, resulting in an upper triangular matrix R, such that .lf = Rr R. The rth diagonal element of the inversenormal matrix, N;1, is then computed in the following way:
111
4.
rzl lYii
'1";
on in which the expression the right hand side is the inner product of the vector sr' with itself. This vector s; is computed by solving the triangular system Rr s1: e; where e; is a vector containing only zero'sexcept the rth element,which equals 1. The computation of the diagonal elementsof the inversenormal matrix in this way costs about 7 % of the total CPU time of the error analysis. 4.2.LO Conclusions
Due to the assumptionsstated in section4.2 the error analysiscarried out here refers given there (circular orbit, to a somewhatidealizedsituation. Of the six assumptions no groundtrack repeat during mission period, measureregular data distribution, over sampling period, nongravitational effects ments are consideredto be averaged are uncorrelatedand of equal variance),the secondone is excluded, measurements probably the most critical one. As already mentioned,data gaps are likely to occur. A noncontinuous data stream destroys the orthogonality properties on which in fact the whole analysisis based. Without orthogonality,the normal matrix becomes a full matrix, leading to an enormousamount of extra work for the inversion. In an idealizedsituation (polar orbit and no band limitation) the mission goals in terms of r.m.s. values for derived gravimetric quantities like gravity anomalies and geoid undulations can be met in terms of accuracylevel as well as resolution, figure 4.2 suggeststhat also a at least for a full tensor gradiometer. Nevertheless, (of the type as will be used in the Aristotelesmission) may give planar gradiometer satisfactory results in such an idealizedsituation, as numerical verification showed (although these results are not given here). Whereassatellite altitudes lower than 200 km seem to improve the results not very much, higher altitudes are certainly not desirable. The same appliesfor the mission duration. Finally, the measurement precision has somewhat more influence,although mainly for higher altitudes than 200 km. We consideredtwo major deviations from an idealizedsituation: (lower) band limitation and nonpolar orbits. Of thesetwo, lower band limitation has the largest impact on the results. The influenceof a nonpolar orbit can be greatly reducedby of of stabilization. A negative consequence stabilization, biasedness the estimates, of coursepresent and appearedto have not so much influence,although it remains one has to bear that in mind. Whereas stabilization may help for nonpolar orbits, it is no remedy against band limitation, as can be concluded from section 4.2.6. Lower band limitation of the lower degrees the spectrum,whereasstabilization seems significantlydegrades
112
4.2.
to improve primarily the higher degrees.As a result too large biasesoccur for the lower degrees. For this problem, only the addition of supplementaryobservations, such as GPS tracking, can be of any real help (Schrama, 1990), (Visser, 1992). For Aristoteles we assumedthe total 6 months mission to be flown in a 92".3 inclination orbit. Whereas this is not actually the case (only for an additional two weeks) the conclusionsabove indicate this assumptionnot to be very critical. As , of for the lower band limitation, this may really jeopardizethe success the mission, if we take a look to the r.m.s. values of geoid undulations (table 4.10). especially really becomevery important to obtain the long waveAdditional GPS observations length part of the spectrum with high enough precision. Two complementarytechGPS for the low degreesand gradiometry for higher degrees. niques are necessary, Indeed, a GPS receiveris planned to be on board the Aristoteles satellite. Inclusion of omission error and smoothing have a major influence on the presented global r.m.s. results. As for the omissionerror) neglectingthe degreesabove omission but, especiallyin nonideal cases, 1000doesnot make much of a difference, parts for degreesbelow .L do have some influence. In fact, the omission error part somelower limit for the global r.m.s.)seesection4.2.3. determines Much more influencethan the inclusionof the omissionerror has the introduction of a smoothing operator into the global r.m.s. computations. Without such an operator, r.m.s. valuesbecomeextremely high. However,for our purposes,representation is of the results in terms of blockaverages justified by taking into considerationthe (although presentin the true signal) globally from difficulty of deriving high degrees techniquedue to the inevitable limited sampling distanceand misany measurement the goal of a gradiometric mission is a resolution of approximately sion altitude. If for 1o x 1o blocks are appropriate. 100 km, blockaverages 4.2.LL Other error analysis methods
so The error analysispresented far was first usedby O.L. Colombo (Colombo, 1987). as uses the actual gradiometric measurements observations,r6lated to the unIt by known potential coefficients a linear model of the type 4.2. Another possibility Ak*, Bk,n (in eq. 3.19) is describedin (Schrama,1990). There, Iumped coefficients Equations of the type 3.20 are used as model equation, are used as observations. for H1^p. with for each gradient a different expression are lumped coefficients determinedfrom the observedgradientsby meansof The for a Fourier transformation. They servesotosay as pseudo observables the next are derived from them using equation 3.20 step, in which the potential coefficients as observationalmodel. Both methods of error analysis, the one in which the observed gradients or the one in which lumped coefficientsaie taken as observations,agree in the sense given in subsequentpoints along that they both treat the gradient measurements, a satellite orbit, as a time series. The methods will therefore be called timelike (Koop et al., 1989). Furthermore, since the method used in this work takes the
113
gradient measurements themselvesas observations, it is referred to as tirneIike in the time d,ornain. The method which uses the lumped coefficients as observations may then be referred to as timelike in the frequencyd,omain. In (Koop et al., ibid.) also a different approach is mentioned. Instead of considering the measurementsas being functions of time, one may also consider them as functions of space. One then assumesthe measurementsto be given on a sphere or in a spherical shell at satellite altitude. Such an approach is referred to as spaceliker. It is equivalent to solving a geodeticboundary value problem. Depending on assumptionsabout the spatial distribution of the measurementsand the apriori variances,different strategiesare possible. First one may assume to have a complete, continuous global coverageof the earth with measurements. In this case an analytical solution for the inverse of the normal matrix exists (Rummel et al., 1989), at least for V"", {V"",Vu"} and is {2V"u,VrrVau}.In reality, of courseno continuouscoverage available.This fact is taken into account by weighting the apriori covariance matrix in such a way that some assumedfinite spatial resolution enters the solution. This method is referred to as spacelikecontinuous. The idea can be modified by assuming we are given a finite number of measurements randomly distributed inside some global spherical shell containing the satellite's orbit. This spherical shell is divided into equalangular cells, inside of which the measurementsare averagedper cell. Observationsare now assumedto be regularly distributed point values(or block mean values),the observationalnoise being adapted accordingly,see e.g. (Rapp, 1989). Using a spherical harmonic expansion for the gradients, the normal matrix takes on a blockdiagonal structure, due to the orthogonality of the sine and cosineseries,and can be easily inverted. This method is referred to as spacelikediscrete.In (Koop et al., 1989) it is shown that all four mentioned methods (tirnelike in the frequencgdomain, timelike in the time domain, spacelikecontinuousand spacelikediscrete)yield the same results.
4.3
Global
recovery
Whereasthe previoussection (section4.2) "only" discussed quality of the potenthe tial coefficients be derived from spaceborne to gradiometry (in terms of aposteriori error r.m.s. values) we are n fact primarily interestedin those estimated potential , coefficients(em,Sm) themselves. As already explained in section 4.1 we try to solve potential coefficients from globally distributed gradiometric data. This processis indicated with the term globalrecouerg. There are severalmethods of global recovery,one of them will be discussedin the present section. In order to carry out a global recoveryof potential coefficients, need a data set of measuredgrawe gradients. Such a set is not yet available,the first set has to be generatedby vity
rThe terms spacelikeand timelike used in this context have nothing to do with the same terms known from the theory of relativity.
114
4.3. Globalrecovery
the Aristoteles mission. Therefore,in this section, a simulatedset of gradient data is used, computed at the Center for Space Researchof the University of Texas at Austin, (Schutz et al., 1987) and (Schutz et al., 1988). A descriptionof thissimuladenote in the sequelthe tion will be given in section4.3.4. We will, for convenience, p o t e n t i a lc o e f f i c i e n tb y { e u n o l o : 0 , 1 } , m e a n i n ge 6 i f o : 0 a n d , S t  i f o : 1 . s 4.3.1 Linear rnodel
The observed gravity gradients are functions of the location of measurementsand of the gravity field. In general,the location (e.g. the coordinatesof the measurement by point in somecoordinatesystem) as well as the gravity field (e.g.expressed means of a series expansion in terms of potential coefficients)are unknown. These are To exactly the unknownswhich we like to derivefrom the measurements. this extent, are linked to a mathematical model, describing (approximately) the measurements the physical reality. In section4.1 such a mathematical model (e.g. equation 4.1 for secondorderradial derivatives)was the starting point for our discussionon (least squares) error analysis. In that section, we simply assumedto be given a global, regularly distributed set of gradient observationswith some known measurement error. The given model equation was assumedto properly describe the physical were assumedto be obtainable from reality and the unknown potential coefficients the measurementsby means of a least squares adjustment. All this meant that we implicitly assumedthat the location of the measurementpoints was known, i.e. along the orbit to circular orbit with known radius and equal spacedmeasurements invoke orthogonality of the trigonometric series. Also the orientation of the local are coordinate system in which the measurements given was assumedto be known. In reality, the coordinates of the measurementpoints are not exactly known, only approximately. Then the model equation which we used becomesnonlinear. In this sectionwe will show one possibility of dealingwith the nonlinear problem, a linearization procedurefollowing (Rummel and Colombo, 1985) for gradiometry or in generalfollowing (Rummel, 1985a). Consider an observationpoint P on the true (or actual) satellite orbit. Initially, the gradiometer output (i.e. the gradient observations)is given in an instrument or satellite coordinate system which has its origin in P. But as already mentioned in chapter 3 we assumethat the gradientscan be transformed to a local orbital or Iocal northoriented coordinate system (with origin also in P) by meansof a known rotation matrix. However,sincethe actual orbit is unknown, the orientation of such local system in P is also unknown. Supposenow that we are given the coordinates of a nearby point P' on a known approximate (or nominal) orbit. Then we may take for the orientation of the local system the orientation of such a system in P', see figure 4.22. We will denote the local coordinate system situated in P but with the orientation of the system in P', with ci, and assumewe are given the componentsof the gradient tensor, V;i@), in this system (the coordinatesof P of courseremaining unknown).
115
4. Global gradiometricanalysis
If we were given an initial state vector, the nominal orbit could be computed by meansof integrating the dynamic equationsof motion of the satellite, using as force model a gravitational field representedby some known set of potential coefficients (up to some maximum degree and order) , and using the measurementsfrom some satellite tracking system. The initial state vector is usually given in terms of the three location coordinatesin a geocentriccartesiancoordinatesystem (X,Y,Z) and the three componentsof the velocity vector (X,Y,2) at ttte initial time ts. The between the true orbit and the nominal orbit are the orbit errors. The differences orbit errors contain a geometricallyinduced part (due to errors in the initial state vector) and a gravitationally inducedpart (due to the imperfect choiceof a potential are coefficient for the orbit integration). This meansthat the observations doubly set dependenton the gravitational parameters,directly through the measurementtype itself and indirectly by meansof the dynamics of the satellite orbit, the latter being reflectedin the orbit errors (Betti and Sansb,1989). Furthermore, the orbit errors of contain contributions from various other sources, which atmosphericdrag will be the most important. The latter error parts will, however, at the moment not be considered.
al orbit
.t.
\<
nominal orbit
Figure 4.22
What follows is the introduction of a mathematical model to which the observations are linked, e.g. equationslike the onesderived in section 4.1. In generalwe
116
4.3. Globalrecovery
For our purpose, however,we assumethat the stochasticbehaviour of the residuals d after linearization is properly describedby their first moment (expectation) to be zero,i.e. E{;} :0 (zero mean) so that E{7} : I (unbiasedness), its second and moment (error variance)tobe E{{}: with Q the error covariance matrix of Q, the observations.In general,the mathematical model t(P,V) will be nonlinear in P and V. The linearizationprocedurethereforeconsistsof two parts, a linearization with respectto the gravitational potential T and one with respectto the coordinates ,; (P) of the point P. The former is done by splitting the gravitational potential V into a known normal (or reference)part U and an unknown disturbing part 7:
v(P):u(P)+r(P).
Hence for the secondorder derivatives we have
(4.r7)
The secondstep of the linearization processis the linearization with respect to the Iocation P, for which we expand U;i@) into a Taylor serieswith respect to the coordinatesof the known approximate point P', truncated after the linear term: ( p U ; i P ) x U ; 1 P ' ) + U ; 1 ( P ' )L , r k  ri(P')  (Ar, L,y,a,z) rhe coordinatecorwhere (J;j*: ffi L.i = ""a "'(P) rections between P and P' , i.e. the orbit corrections between the actual and the nominal orbit. It is convenient (although not necessary)to compute the nominal orbit using the same potential [/ as is used above for the potential linearization. Since the disturbing potential ? is a small quantity of first order we assume T;i(P') xT;i(P). Insertedinto equation4.I7 this yields:
(4.18)
with Al;; the gradient anomaly. In the simplestcasethe normal potential [/ is the potential of a spherically symmetric gravitational field (i.e. U : Y), but it may also be the potential of an ellipsoidalfield or any other higher order approximation of the actual gravitational field. In general,the normal potential [/ is expandedinto
117
4. Globalgradiometricanilysis
by a sphericalharmonic series(or a seriesof inclination functions) represented some ej^". With the actual potential V expanded set of harmonic coefficients (limited) into a similar series with coefficients C1^o we find for the disturbing potential ? an expansionin terms of the coefficientcorrectionsLe t^o, where e kno : e't*o * Aeuoo. In equation 4.18 the unknowns are the potential coefficient corrections Le uro in Ti1(P'), and the coordinate correctionsArt. It is known from the theory of adjustment that, although the choiceof approximate values does not influencethe results (as long as they are inside the bounds of so convergence that the iterative estimation processwill converge) it is important , to compute these approximate values (in our case the normal potential U in the approximate point P') accurately,in order to obtain small anomaliesand thus accelerating the iteration process.On the other hand, the coordinate correctionsAcd will, in general,also be small so that the elementsof the coefficientmatrix U;ix(P') can be computed with less accuracy. Whereas one usually takes for [/ a series expansion up to some (high) maximum degreeif it concernsthe computation of the anomalies(i.e. for U;i(P')), a spherical or ellipsoidal [/ sufficesfor computation of matrix (i.e. for Urir(P')). the coefficient sphericalmassdistribuLet us take, for example,the potential of a homogeneous tion as simplified normal potential for the computation of the coefficientmatrix [1i3. This matrix can then be computed very easily,and if inserted into equation 4.18, it yields (for all six gradient tensor componentsseparately):
Af*
0 0
U U
7,,
e,,
Er"
3GM
0 0
I
^"1
o'l*
Lrl
T,,
Tr"
,b,
(4.1e)
0 0 0 l
eoo
cuz
Li
""
0 0 ,
6""
An equation like 4.19 can be set up in each observationpoint. If we have N, observationpoints, we have 3No coordinateunknowns and 6No unknown ?;i's. But the gravitational potential is common to all measurementpoints, so, expressing the potential by a seriesexpansionup to some maximum degree L, the potential coefficientcorrections Aeuno are the samefor all points, leavingus with (r+ 1)2 3 potential unknowns (which is the total number of unknown coefficientsin a series (l expansionup to degree,L minus the three firstdegreecoefficients : 1) which are : O of our geocentriccoordinate system is the center identically zero if the origin r of gravity of the earth). A system of observationequations arisesif one joins the model equation 4.19 for all observationpoints together as function of the same set of potential coefficients.
116
4.3. Globalrecoverv
Now one may proceedin differentways. Either one solvesthe systemof equations for all unknowns simultaneously(i.e. position as well as potential unknowns). This see method is used in the socalled theory of integratedgeodesy, e.g. (Hein, 1986) or (Moritz, 1980). Or one eliminates certain coordinate unknowns by forming appropriate linear combinations of anomalies, a method often used in the theory of the geodeticboundary value problem. One is then left with an adjustment problem with only potential coefficient corrections as unknowns. After they have been estimated, back substitution into the model equation givesestimatesfor the coordinate unknowns. Iteration of this procedure should convergeto the right results. The next section will discusssuch a procedure.
4.3.2
fteration
In this section a possibleiteration procedurewill be sketchedwhich can be used to solve potential coefficients(and coordinate unknowns) from gradiometric observain tions. The procedureusedhere was first presented (Rummel and Colombo, 1985). method is to carry out a preadjustment of the diagonal comThe first step in this ponents of the gradient tensor using as model equation the Laplace equation (zero trace of the gradient tensor). In the caseof Aristoteles this preadjustment cannot be carried out since not all diagonal tensor componentsare measured. As starting point for the adjustment to follow we use the linearizedmodel 4.19 derived in the previous section. The idea is to separatethe solution of the potential from that of the coordinateunknowns. If we take a look at equation 4.19 coefficients we see that, in the presentsphericalapproximation, only Af', depends on the coordinate correction L,r, Lls" only on Ay but that the three diagonal components all depend on Az. These three can therefore be used to eliminate the radial orbit and Ai', to be correction Az from the model equation. If we assumeonly Ai* available (like with Aristoteles) then the linear combination
z\lua * Ll
""
(4.20)
eliminates A,z. If we take this linear combination in each observationpoint as new there remains a systemof linear equationswith only the potential coefobservables, ficient correctionsas unknowns. Note that this is the reasonwe took this particular combination of gradients 2Vo, lV"" as one of the combinationsin the previous section on error analysis,section4.2. Now we have separatedthe determination of the gravitational field from that of the orbit. This is only true up to the order of the earth's flattening, sincewe took the coefficientmatrix of the coordinate corrections in equation 4.19 in sphericalapproximation. On the left hand side of the resulting model equation we now have the linear combination andon the right hand side the corresponding combination 2Airr+A i ", of the disturbing potential 2Tuu+ ?"r. This equation can be used to estimate the potential coefficientcorrectionsLet*o.
119
4.
Global gradiontetric
analysis
Also, from equation 4.19 we seethat the tensor componentAi* does (in spherical approximation) not dependon any coordinatecorrectionat all. This equationcan therefore be used for the estimation of the potential coefficient corrections together with the linear combination above (at least if this tensor component is measured, which is not the casein the Aristoteles mission). The linear model is now of the sameform as equation 4.2,and can be solved by means of a least squaresestimation procedure,as describedin section 4.1. This results in a set of potential coefficient correctionestimatesLe t*o, from which we may compute an updated set of potential coefficients adding them to the coefficients by : e't*o * Le mo. e'rrooof the normal potential U: e'rl" What should follow is the computation of the coordinate corrections A,z by means of backward substitution of the solved potential coefficientcorrectionsinto the linear model eq.4.19. We saw alreadythat thesecoordinatecorrections are in fact the radial orbit errors, the differences betweenthe true and the nominal orbit in radial direction, and that they contain gravitationally and geometricallyinduced parts. The orbit errors reflect the unmodelled part of the motion of the satellite. This motion (being the time evolution of the state vector (Betti and Sansd,1989)) can be describedby an initial state vector (ri and ii at the initial time t6) and some system of dynamical equations(equations motion), e.g.: i;(t) : V;(rt(t)) + P;, of where P; are perturbing terms, which are neglectedhere for convenience, and t is time. In order to include the satellite motion (through the dynamical equation) into the estimation procedure, the determination of the coordinate corrections is divided into two parts. Assuming that the corrected set of potential coefficients Cjf* constitutes a better gravitational field representationthan the set e'1^o, *. first compute an updated nominal orbit with these new coefficients Cfl, but with the old initial state vector. In this way we obtain a new set of updated approximate points P" in which we computenew anomalies: Aili: iii(P) U;i(P").Inserting these new, smaller, anomalies and the estimated potential coefficientcorrections x Le uoo into equation4.19 and assumingT;1(Pt') T;1(Pt) x T;1(P), we obtain 6rli  T;i(P"): U;j*(P")Lrk
(4.2r)
where the U;ip is still in spherical approximation and where the (also smaller) coordinate corrections Axk : ,ke)  *o(P") are now the orbit errors between the true orbit and the updated nominal orbit. They are thereforefree from the influence of the A6'so and are only due to uncertaintiesin the initial state vector, cf. (Rummel and Colombo, 1985). From this equation coordinate correction estimates L,2 can be obtained. With the resulting set of coordinate corrections A2 we may compute an updated initial state vector using some linear model. An example of such a linear model is the solution to the homogenousHill equations (ibid.). The Hill equations are derived from Iinearizingthe equationsof motion of the satellite relative to some referencemotion (u.g.u circular motion in a spherical gravitational field). For an
120
4.3. Globalrecovery
exact derivation of the Hill equations and their solution see e.g. (Kaplan, 1976) or solution (Colombo, 1986). For example, for the radial direction, the homogeneous is given by: of the Hill equations
(4.22)
and n : at where the initial conditions Azs,L,is,Aig are considered t:0 is the mean motion of the satellite. For each measurement epoch t an equation like 4.22 can be set up. All theseequationsare function of the sameinitial conditions Azs,L,2s,Ai6. Merging theseequationsinto one system,estimatesof Lzs, A2s, Ais can be obtained by solving the system of equationsby least squaresadjustment. Now we have improved the two error parts contributing to the orbit errors: the set gravitational part by improving the potential coefficient and the geometricalpart we by improving the initial state vector. As a consequence may again compute an orbit with the coefficients e'1'*o and the improved initial state updated nominal vector. In fact, from here on, the whole processdescribedaboverepeatsitself, since with this new nominal orbit, updated gradient anomaliesare derived, the linear correctionsare estimated,etc. Assuming combination is formed, potential coefficient to to the process converge the right solution (cf. (Rummel and Colombo, 1985)),one repeats the above procedureuntil satisfactoryaccuracyis obtained. In the caseof Aristoteles the procedureends here. If also the gradientsVr" and Vs" aremeasuredthe proceduremay be continuedby estimating the other two coordinate correctionsAs and Ay. The coordinatecorrection Ar can be obtained from Ai,, (which is computed along the final nominal orbit) by inserting the final potential coefficientset resulting from the repetition processabove into equation 4.19. In the same way Ay is obtained from Alyr. The correspondingtwo solutions (crossHill equations are: track and alongtrack directions) of the homogeneous nAio 2L'io  (6na,zs+ SAio)t + at(t) ) sin nl ' n cosnt f (6Azo + n 2\io ^ Aron Ai(t) : Ayo cosnt + Aio ,in ,,t n
(4.23) (4.24)
which, togetherwith the radial solution, may be usedto estimate a last improvement of the initial conditions and to compute a last nominal orbit. This best estimate of the orbit could be used to carry out a final potential coefficientestimation. The whole estimation procedureis listed in table 4'11. 4.3.3 Spacelike v. tirnelike
Whereas the iteration method describedabove may seem a nice example of gradiometric data analysis (since it separatesthe determination of the orbit from that
121
4. Global gradiontetricanalysis
of the gravity parameterse m.) there remains one big problem, namely that of the enormous amount of data to be handled by the software. Two aspects have to be consideredin this context. First the number of observations.During a six months mission and with a data sampling rate of 4 seconds, almost 4 million measurement epochsbecomeavailable. If we confine ourselves the caseof Aristoteles this means that at each epoch two to observations given,Vru andV"",fromwhich the linear combinatior,2Lluu+Llzz are is derived. In terms of the model equation4.2 this thereforemeansthat the vector of observationsI has over 4 million elements.In the leastsquares estimationprocedure, eq. 4.5, we have to carry out the multiplication 6 : ,+r t (if.we assumethe covariance matrix Q to be a scaledunit matrix) which involvesinner products of the columns of A with the observation vector l. One such inner product thus consists of a summation over 4 million elements. Each column of A belongs to a certain l, rncombination and each row of A belongsto one individual observation.Each element of A is therefore the type 4.3 or 4.4for a uniqueset of (ro,rr,r). For eachof the 4 of million inner product elementsthe trigonometric functions cosrfilrr, sin rftprrhave er to be evaluated for all necessary valuesof the index fr. Even when using recursive type methods for the computation of thesetrigonometric functions, this part of the estimation process,i.e. computing 6, remains very time consuming. Another aspect concernsthe number of unknowns. Given some specifiedmaximum degree ,L up to which one likes to estimate the potential coefficients, total a number of (I + 1)(, + 2) unknown e1*o coeffi,cients appear. E.g. if L : 240 (like it is usually the case in this thesis) we have over 58,000unknowns. This means that the design matrix A has more than 58,000 columns and, as we have seen, almost 4 million rows. The normal matrix N has a size of (sa,ooo)2,but it is symmetric. Solving such a linear system is still an enormouseffort, which under certain assumptions, however,can be drastically reduced by invoking a blockdiagonal structure, as was done in section 4.2 with the error analysis. If we are willing to accept some assumptions and approximations as far as it concernsthe observationdistribution, somethingcan be done to reducethe computational efforts. Two methods are mentioned here: the spacelike and the tirneIike method. The terms spacelikeand timelike are those used in section 4.2.11. The differencebetweenthe two is the way in which we look upon the data distribution. Both methods will be illustrated below.
Spacelike In the spacelikeapproachwe view upon the data set as being a set of observations distributed in some way (regular or irregular) in three dimensionalspace. The mutual relationship between the observations(they are all taken along one common satellite orbit) is not taken into consideration.The latter point is only reflectedby the fact that during the iteration process(table 4.11) the initial state vector components are updated using some dynamical model for the satellite motion (i.e. Hill
122
4.3. GIobaIrecoverv
to are equations),and by the fact that all rneasurements considered be located inside a spherical shell around the earth. This shell, concentricwith the earth's center of mass, is just thick enough to contain all measurements.An observation is considered the gradients a function of its position coordinatesr,0, .\. An easyway of expressing in this case is by means of a spherical harmonic expansion,equation 3.16. When using this formulation, a blockdiagonal structure of the normal matrix is achieved if the data, assumedto be situated on a sphere(i.e. constant radius approximation), is distributed regularly in ddirection (observations along parallels) as well as in ,\direction (observations at equal intervals At). Along each parallel the observations constitute a set of equally spacedsamples. On such a set, the trigonometric the elementsof the normal functions sin rnl and cos rn\ arc orthogonal. This causes e6^ro, to be only different from etr*ro, and matrix belonging to the coefficients zero if dr : a2 and rn1 : trt2 (Rummel and Colombo, 1985). Compare this with the orthogonality in section 4.2.1. Thus a regular, global grid of observationsis needed in this case. But we do not have such a grid. If the actual data distribution shows some particular pattern at all, it will certainly not be of the type described above. However, in that such a grid structure can be ob(Rummel and Colombo, 1985) it is suggested tained in the following way. The spherical shell in which all observationsare contained, is divided into equal angular cells of thicknessAr (the thicknessof the shell) and with sizeL,0  Al, the latter chosenat most half of the smallest gravitational inside each cell are averwavelength to be recovered.AII the gradient observations aged. All equiangular cells with their averagegradient valuesformed in this way, together constitute a global grid of gradient observations. A blockdiagonal structure of the normal matrix is now obtained which means an enormousreduction of computation time. But also the number of "observations" (i.e. cell averages) has becomeless. If we, for example,like to recoverthe potential coefficients up to a maximum degree and order L : 240, the grid size L,0 = A) has to be at most 0'.75. The number of cells contained in a grid of this size is 115,200,which is much less than 4 million. So also the amount of time neededfor the computation of D has been decreased. The drastic decreaseof computation time is a big advantage of gridding the data. On the other hand, averagingmeans a loss of precisiondue to the smoothing of the signal. Furthermore, we may put some questions about how to carry out the averaging process. A first problem is that in reality, we do not know the exact coordinates of the measurement points P. The decision in which cell an observation lies and where it is located inside the cell, can only be made on the basis of the coordinates of the approximate points P' along the nominal orbit. Due to orbit uncertainties in crosstrack or alongtrack direction, an observation may therefore contribute to the cell averageof the wrong cell (the radial orbit error will not lead to such problems). Since we have seenthat the secondorderderivativeshave a large power content in higher degrees,these kinds of mistakes may have considerable influence. Furthermore, since the error variance of cell averagesis computed from
123
4.
Global gradiometric
analysis
the resulting set the measurementerror propagationthrough the averagingprocess, of grid values will not have equal variances any more. Unfortunately, as long as we do not have the true coordinates, nothing can be done about the former problem, unlesswe could determine the true orbit with enough precision independently,e.g. by using GPS, which would imply that we are back at the earlier situation, where we don't have to estimate the orbit from the gradient measurements. A related question is how to interpret the averagevalues. One possibility is to first average the original gradient observations, and to attribute the averagevalues to the center points of the equiangularcells. The coordinatesof those center points are known. The problem is reducedto that of just estimating potential coefficients, compare (Vermeer,1991). No linearization since the "orbit problem" is suppressed, with respect to the coordinate unknowns is possibleany more. Another possibility is to attribute the averagegradient value to some imaginary averagelocation inside the cell, not coinciding with the cell center. This averagemeasurementpoint could be thought of as being computed by means of averaging the (unknown) coordinates of all the actual measurement points inside the cell. The cell center could then be chosen as approximate point P'. In this case there remains the problem of which kind of relation there is between on the one hand the averagedmeasurement points and the true orbit and on the other hand the cell centersand the nominal orbit. Probably the best choiceis to first linearizethe problem and compute the gradient anomalies(eq. 4.18) and the linear combinations(eq. 4.20) f.or all observations individually. AII observationsinside the same cell are linearizedwith respectto the same approximate point P', namely the cell center, and for the normal gradient U;i the value at the cell center P' is taken. The resulting set of linear combinations of gradient anomalies are now averagedon the basis of their approximate locations P'. In the estimation procedure from table 4.11 this means that the cell averagesare formed betweenstep 6 and 7, and that they only apply to step 7. Back substitution (step 10) is done with the original systemof equations(one equationper observation, only for computing . not the averages) Furthermore, one could use the cell averages : and inverting the normal matrix w #A'A, whereasthe multiplication AT lis carried out for all observationsindividually. In that case it is not even necessary to actually compute the cell averages.One simply choosesa global grid at a certain mean radius and with a certain step size in ddirection. The ,\dependence is cancelled due to the trigonometric orthogonality relations. In this way however, we have to compute D again for almost 4 million observations. Note furthermore that, if we want to use the inverse of the present normal matrix (where cell averages are assumed) to describe the aposteriori precision of the estimated potential coefficients, we have to adapt the scalefactor o2 of the variance matrix of the observations Q to represent the variance of cell averages. In fact, in general the variance of a cell average becomes a function of the number of observations inside the cell and of their location relative to the cell center. This results also in a ddependence the variances. there remains the question if Nevertheless, one likes to compute cell averages,
124
4.3. GIobaIrecovery
of the validity of the averaging process. Inside each cell the observations might be distributed rather irregularly. It should therefore be preferred to reduce the observations to the cell center using somehigher order model, at least one which includes a dependencyon the relative location of the measurementpoints with respect to the cell center. This becomesextremely important if the actual observations V;y are averaged,since these numbers are relatively large and vary considerably throughout a cell. E.g. at 200 km altitude differences the value of the gradientsinside a cell in with size Ar: 10 km and A,0 :0o.5 mav be of the order of 1010s2 due to the C2s contribution only. It should be mentioned that a possible negative influence on the results of the averagingprocessmay be overcomeduring the iteration procedure (seetable 4.11, step 14), at least ifthe cell averages formed betweenstep 6 and 7 (as suggested are above). Timelike As pointed out in section 4.2.LL, the timelike approach considersthe data set a time series,the observationsgiven in subsequent points along a satellite orbit. The usual way of expressingthe gradientsin this caseis by means of an expansioninto a seriesof inclinationfunctions,like eq.3.18. The time dependency expressed is by the argumentth**: !tl* * rb**t, compareeq. 4.6. In order to achieve a blockdiagonal structure for the normal matrix in this case,we do not need an equiangular global grid of observations. In section 4.2.1 we saw that such a structure is already obtained if the data is distributed regularly only along the orbit, under the condition that the number of nodal days .l[a and the number of orbit revolutions N" containedin the mission are relative primes and that .l/" > 2L. For .L  240 the latter condition is fulfilled in a 6 months mission, like with Aristoteles. Also, with the foreseen constant sampling rate (4 secondsfor Aristoteles) with which the measurements are taken, the requirement of a regular data distribution along the orbit is very well met. What remains as an assumption is the absenceof data gaps. Unfortunately, data gaps will occur in a real mission (due to orbit maintenancemanoeuvresand excessive drag variations at the poles). Note that in the spacelike situation alongtrack data gaps play no role as far as it concernsthe orthogonality requirements. Another problem that remains is the fact that not one referenceorbit is computed for the full mission period. Instead, reference trajectoriesare integratedfor somelimited amount of time, e.g. every three or six days. In the timelike approachthere are no problemsconcerningthe true orbit, nominal orbit and averagevalues,becauseno averagingof the data is neededto invoke orthogonality. Except for the data gaps assumption,the timelike approach therefore seems more suitable for the gradiometric analysis (in the way described in section 4.3.2 above) than the spacelikeapproach.
125
4.
Global gradiometric
analysis
4.3,4
Simulated data
Whereasan error analysiscan be carried out without the availability of a set of real measuredquantities (like it was done in section4.2), potential coefficientestimates can only be obtained if one has at ones disposal some set of observations. Up to now, however,no gradiometricsatellitemissionhas been carried out, so no true data is availableyet. In order to be able to investigatethe gradiometric analysisprocess and to carry out test computations one dependson simulated data. The Center for Space Research(CSR) of the University of Texas at Austin computed a set of gravity gradient data along a simulated satellite orbit. Detailed information about this simulation can be found in (Schutz et al., 1987) and (Schutz et al., 1988). In of this section we will only briefly discussthe main characteristics this simulation. the CSR simulation spans 32 sidereal In order to obtain a good global coverage, days after which the ground track of the satellite'sorbit will repeat to within 10 km , d a y s ) . F o r a s a t e l l i t e m o v i n ga l o n ga p o l a r o r b i t ( / : 9 0 " ) a t a m e a n (i."."":32 radius of approximately 6527 km, this results in 525 orbital revolutions. The force model used for the orbit computation consistedof the complete 360 x 360 OSU86F potential coefficient field (Rapp and Cruz, 1986) and the following values for the radius .B: gravitational parameter GM and the reference G M : 3 . 9 8 6 0 0 4 4 0 4 '1 0 1 4 , n 3 "2 : 6378137m . .R the were assigned valuesezt : 0.10'109 coefficients Furthermore,the C21and ,521 . 108. The force model did not include nongravitational forces, : and ,S21 0.102 lunisolar effects or temporal variations in the gravity field (Schutz et aL.,1983). of The computed data set containsthe ephemeris a loworbiting GRM (Geopotential ResearchMission) satellite at 4 secondintervals. This resulted in a total amount of 691,210measurementepochs. At eachepoch the elementsof the gravity gradient tensor V;i are evaluated. Despite the fact that only five elementsof the gradient tensor are independent (V;; is symmetric and traceless) all six elementscomprisingthe upper (or lower) triangular , part of Vi are provided. The gradientswere computed as the secondorderpartial derivativesof the gravitational potential in a geocentriccartesiancoordinate system using a spherical harmonic expansionup to degreeand order 360. The potential coefficientsfor this expansion were also taken from the OSU86F field, but the conAll was not included (i.e. Coo :0). tribution from the zeroth order term (ff) purpose of the commeasurementerrors were excluded from the simulation. For the putations the observationswere assumedto be uncorrelated,having a 0.01 F l\/Hz white noise error spectrum. At this point already some remarks can be made about this simulation, as far as they concernour intended use. Referringto section4.2.lit is reminded here that for from satellite data up to a maximum degree a good recoveryof potential coefficients by represented somefr, rn ,L (avoiding a situation where different orbital frequencies,
126
4.3. Globalrecoverv
combination, are projectedonto the samegravitational frequency,represented an by l,m combination), the number of orbital revolutions N, in one repeat period has to be larger than2L. This in turn implies that for the presentsimulation (525 orbital revolutions) we can expect only a good recoveryup to approximately L:262. The choice for repeating ground tracks after 32 days ensuresa finest possible global coveragewithin this period. As a result the ground track nodes (cross over points between descendingand ascendingtracks) for the 525 orbital revolutions included in this repeat period have, at the equator, a spacing in )direction of approximately 0".68. Simple reasoningtells us not to expect recoveryof potential coefficientsabove degree I : 262 in this case, despite the fact that, due to the 4 second sampling interval in alongtrack direction, the spacing in ddirection will be approximately 0".27. Figure 4.23 shows for a small area (50o S ) < 60" and Oo I g < 10') the ground tracks (continuouslines) of the simulated orbit, as well as the grid lines of a 0o.5 x 0".5 grid (broken lines). It can be seenfrom this figure that several blocks of the grid are not coveredby ground tracks. An attempt to derive from the simulation potential coefficients to degree 360 with the spaceup like method (for which such a 0o.5 x 0'.5 grid is needed)is therefore not likely to succeed, least if the empty blocks are not accountedfor in some manner. at
Longitude
Figure 4.23
Grounil tracks (continuous lines) of the sirnulated.CSR orbit for a small area. The brolcen lines are the grid lines of o 0" .5x 0" .5 grid..
127
4. GIobaIgradiometric analysis
4.3.5
Spacelikeresults
A computer program was written by D. Stelpstra (Stelpstra, 1990) to estimate potential coefficients from gradiometer measurementsfollowing the linearization procedure in the spacelike approach, which was described in previous sections. As input data the satellite ephemerisand gravity gradientsfrom the CSR simulation, described in the last section, was used. Test computations were done to demonstrate the possibilitiesof the present recoverymethod (Koop and Stelpstra, 1991). To this extent, a global 1o x 1o grid, being a commonly used and easily manageablegrid type, was establishedfrom the gradient data set. Such a grid consists of 64,800 equiangular blocks. For each measurementepoch, the given geocentric cartesiancoordinateswere transformed into curvilinear coordinates0, ). The latter coordinates determined in which block the measurementwas assumedto be situated. Since there were 691,210measurementepochs in the simulation, the result per block. To simulate the elimwas an averagenumber of N,  l0 measurements ination of the coordinate unknowns (i.e. the orbit errors), as is done in the linear epoch the linear combination of observations model approach,at each measurement The resulting numbers were then averagedinside z\f * I Al", was computed. to we each grid block. For convenience assumedall those block averages have the with o the precisionof one gradient), samemeasurementprecision(namely "Jffi which of course is not really true, since not each block contains the same number of measurements.At this stage we are only interestedin estimatesfor the potential in coefficients.Their aposteriori error behaviourwas already discussed section 4.2. variancesis the asThis means that the one important item concerningthe error sumption of equal variance for block averages. This leads to a simple variance matrix, namely a scaled unit matrix. The value which we choosefor the variance factor a2 is not important here. As can be seenfrom eq. 4.5 it will, in casethat Q is a scaledunit matrix, drop out when computing 6. up From the 1o x 1" grid one may estimate potential coefficients to degree and from the simulated order 180. We expecteda good recoveryof potential coefficients 525 orbit revolutions only up to degree 262, so with .L : 180 we are on the safe were used to compute signal degreevariside. The estimated potential coefficients ances. They are shown in figure 4.24, together with the degree variances from the original OSU86F model. Up to degree 140 the signal power seemsto be recovered very well. Above degree 140 the estimated power is higher than the original one. This fact may be due to an aliasing effect, where the power of higher degrees,contained in the measurementsignal but not solved for, is reflectedin the lower ones (Jenkins and Watts, 1968). In time seriesanalysis,this effect always appearswhen the chosen sampling interval is not small enough to detect the highest frequency which is present in the signal. The highest frequencywhich can be detected given some sampling interval A is the socalled NyquistfrequencyIx : Il(2L). If fv is smaller than the highest frequencypresentin the signal, aliasing occurs. When usin ing spherical harmonics the aliasingeffect cannot be expressed a simple manner,
128
4.3.
Global recoverv
but it will be present. In our case,the OSU86F field contains potential coefficients up to degreeand order 360, which were all used to simulate the gradient data. But we only solve for coefficients to degree180, so the power content for the degrees up 181 < I < 360 will be reflectedin the spectrum between0 and 180. Since in reality the gravitational field contains degreesup to infinity, an aliasing effect will always be present, even for higher degreesolutions, although blockaveragingof the data will decrease influenceto a large extent. its
osu86f 
solved
9 O rI E
a) o0 '6 co b o l
a.a/t11^.^\
.__\
OU
90
degree
' tn
r Rn
180
potential Figure 4.24 Degree aariances solaed coefi.cients sirnulated from. frorn gradient as datausingthespacelike approach, wellas those frorn OSU86F. Nevertheless, the solution seemsnot bad at all. Especially if we remember that this estimation is in fact only one step of an iteration processin which one simultaneously solvesfor the gravitational field as well as the orbit. Furthermore, since we did not remove from the simulated data any other potential contribution apart from the central (zeroth order) term, the anomaliesare relatively large. The degree variances, however, show the spectral behaviour of the signal in terms of the power per degree. If this signal power per degreeis well solved for, it does not necessarily mean that the distribution of the power over all orders inside this degree is properly solved for. One has to look into the individual potential coefficients themselves.To this extent the absolutevaluesof the relative differences betweenthe solvedcoefficients 1*o and the original OSU86F coefficients 1*o *"r. e e : _ computed,i.e.l(Cp""Cu.")lCtol. value is, for example,10u: 1 If the resulting this meansthe relativeerror is IOO%. A valueof 102 means1 % relativeerror. We counted how much of these numbers were presentin each 100lclass between 105
129
4.
and 105. A cumulative graph of these numbers, in percentages the total amount of of solvedcoefficients, shown in figure 4.25. From this figure we seethat over 80 % is of the coefficients solvedwith an error lessthan 100 %, howeveronly about 20 % is with an error lessthan l0 To. In order to seewhether someparticular pattern exists in the error or not, a greyscale map of the same numbers in an f , rnscheme was created. This map showed no particular pattern, except for a slight increasewith increasingdegree.
sr_
o
d
0
log abs rel dif
Figure 4.25
Cumulatiue graph of the number of errors betweensohed and. OSU86F coefficients percentages the total arnount of solaed in of coefficients,Horizontal atis: ord.erol rnagnitudeof the relatiae differ ence Iog s, arithmical.
It should be mentioned that the differences between the solved and original coefficients, as they are computed here, do not contain measurement errors, since we adopted a scaled unit matrix for the covariance matrix of the observations. The remaining error therefore consists of model errors and numerical errors. The latter are inevitable in each computational process. 4.3.6 Tirnelikeresults
In the timelike approach the observations are treated (as they are in reality) as measurements done in successive points along the satellite orbit. The parameters to indicate those measurement points are the orbital elements a,e,I,u)o,tt)". For the CSR simulation we have 1 : 90o and for our convenience we assume e : 0. Furthermore, we assume all measurements to be situated on a sphere with radius equal to the mean radius of the simulated orbit. This leaves us with c.roand <,r"as
130
4.3. Globalrecovery
measurements.They are to be computed from variables to indicate the successive the given geocentriccartesian coordinatesrI and the time t of the epoch. Care especiallywith the has to be taken with this transformation from ,I ,t to u)o,u)e, identification of the correct quadrants. has to be done, so we are left with a data No averagingof the measurements set of 691,210records,each containing the measurement(or the linear combination and at z\laz+ Afrr) as well as the two coordinates r^,ro <.,r, the measurement time t, in double precisiona storagerequirementof a little more than 16 Mbyte. A computer following program was written to estimate from this data set potential coefficients the method describedin sections4.3.1 and 4.3.2. This estimation program is in fact an extended version of the error analysisprogram from section 4.2, now including the computation of b : Ar I and the solution of the normal equations. Due to the enormous CPU time requirementsfor bhis program, we decided for the first thereby halving the number of measurements, run to averageevery two successive to observations be used in the estimation process.Sincewe only attempt to estimate up coefficients to a maximum degree.L : 180 (like in the spacelikeapproach) the data interval in ddirection after this averaging(approximately 0'.54) will still be sufficient. Whereas in a real mission like Aristoteles the lower part of the measurement spectrum will be too much distorted or not availablewith enough precision (band limitation, see section 4.2.4) we included in the present estimation the complete measurementspectrum (aIl B6for 0 ( m 4 L and L < fr < I). No stabilization techniquewas used (seesection4.2.6). solved with the preDegreevariances,computed from the potential coefficients togetherwith the original OSU86F sent timelike program, are shown in figure 4.26, degreevariances.Above degree120 the solvedspectrum divergesfrom the original are larger than in the spacelikeapproach (compare one. However, the differences figure 4.24). Since, in both cases)we did not remove the contribution from the the differencesare due to an aliasing degreesabove 180 from the measurements, this averspacelikeapproach,where we first computed cell averages, effect. In the decreasing power content of the high degrees. the aging acts as a smoothing process, Therefore the aliasing effect is expectedto be lessthan here. Also for the present solution we show (figure 4.27) the cumulative percentages of the relative errors in each coefficient.Somewhatlessthan 80 % of the coefficients is solvednow with an error lessthan IOO%, but still about 20Yowith an error less than 10 %. that, for both the spacelike as well as the timeIt should be reemphasized like approach, the results shown here are "only" a first step of what should be an iterative processin which at the same time the orbit is improved and the coefficients are estimated. A simplified simulation, based on zonal coefficientsonly for (Rummel and Colombo, 1985),indicated that only afew steps are necessary the iteration processto converge. A final judgement of the results and the estimation method can thus only be given if a complete iterative solution is computed. Nev
131
osuB6f solved
9D r a l
6 ol) o ^ o,0 
en
OU
on
120
150
180
degree Figure 4.26 potential coefficients Degreeaariances frorn sirnulated, lrom solued grad.ientdata using tlt,etirnelike approach, as well as thosefrom OSU86F. ertheless, the present results are promising, especially considering the fact that in both cases no stabilization technique was used.
4.3.7
Sorne cornputational
aspects
From a computational point of view, the programs for the spacelikeand the timelike methods are very much alike, at least in broad outlines. Both programs do, in principle, nothing more than solving a system of the type I  Ac by means of a least squaresadjustment for the samevector of unknowns c (which are the potential coefficients). The differencesbetween the two approachesare found in the vector of for the elementsof the design matrix A. In case observationsI and the expression of the spacelike method, the vector of observationsI consistsof block averages, a total of 64,800 elementsin the caseof a 1" x 1' grid. In the timelike approach this vector consistsof the original observationsalong the satellite orbit which are 691,210elementsin caseof the one month CSR simulation. betweenthe two methods is that Concerningthe designmatrix A, the difference the spacelike method usesspherical harmonics as base functions (for which we need to compute the Legendrefunctions Pt* and their derivatives), whereasthe timelike method is set up in the rotated orbital system and is thereforebased upon the inclinations functions F!* and Frk*. as Both differenceshave farreaching consequences far as it concerns computation time. Before we illustrate this, we have to remind ourselvesthat the required
132
Figure 4.27
and Cumulatiue graph ol the number of errors betweensolued' in oJ OSU86F coefi,cients percentages the total amount of sohted ol coefficients.Horizontal aris: order of rnagnitud,e the relatiae diff er ences, I ogarithmic al.
computational effort strongly dependson the maximum degree ,L and the available hardware. All CPU results given in this section refer to software running on the CONVEX C240. First of all, computing the inclination functions for all valid /, rn,k wp to some maximum degree.L is much more time consumingthan computation of the Legendre functions for all l, m up to the same degree. If the spacelike program is used only for error analysis (i.e. setting up and inverting the normal matrix) it requires only 40 CPU seconds (for one observation type), against the timelike program and types), both up to L:180 approximately800 seconds(for 10 observations both on the same computer. Secondly, and what is more important, the product D : A" I puts a heavy burden on the computations. Compared to the computationsinvolved in b the error analysiscontributesonly little to the total CPU time for the adjustment process.For the spacelike program the coefficientestimation up to L  180 requires 370 CPU we seconds(Stelpstra, 1990). For each of the 64,800cellaverages have to evaluate of rnA, multiply them with the Legendrefunctions, some other the sine and cosine factors and with the observation. But since the observations are regularly spaced on a grid, the amount of extra effort can be kept limited. For the timelike program the discrepancybetween "only" error analysis and adjustment is much larger. This is due to the fact that we have much more ob
133
4. GIobaIgradiometricanalysis servationsthan in the spacelike method. For each observation an element of A consistsof a summation over k. For all l,m combinationswe have to evaluate the sine and cosine of the argument ,lr*^: kro + rn<r", multiply them with the inclination functions and some other factors, add them for all k and multiply them with up the observation.As an example,considerthe estimation of the zonal coefficients of to L :100, which requiresTI7 CPIJ seconds, which 95 % (681 sec.) is neededfor D. Bearing in mind the total runtime for the error analysisfor a// orders setting up see up and degrees to ,L : 240 (500 CPU seconds, section4.2.9),one can imagine the computational effort to be deliveredfor a complete estimationup to such maximum up degree. Indeed, the estimation of all potential coefficients to degree and order 180 from the halved data set of 345,605observations(as describedin the previous section) required approximately 69 hours CPU time on the CONVEX. The time neededfor setting up and inverting the normal matrix in this casebecomesalmost negligible. A last remark concernsthe system of normal equations. The solution of the system of normal equations, N c : 6 is obtained in an equivalentmanner as were the diagonal elements of the inverse normal matrix in the error analysis. With the upper triangular matrix .R, resulting from a Cholesky factorization of N, the triangular system RTw:b is solved for tl by forward substitution. Afterwards, backward substitution of the system Rc: w gives the required vector c. 4.3.8 Conclusions
On basis of the presentedresults we cannot yet decidein favour of either the spacelike or the timelike approach to estimate potential coefficientsfrom gradiometer observations. However, the results shown in figures 4.24 to 4.27, make us believe that both methods bear the potential of convergingto the right solution, if the iteration procedure is continued. At this moment we have to keep in mind that we only carried out the first step of this iteration process. Furthermore, in an actual of mission one shall be faced with the presence data gaps and nonpolar orbits. The spacelike method is very attractive becauseit reducescomputation time drastically, not only becauseof the blockdiagonal structure of the normal matrix, but much more becauseof limitation of the length of the vector b : AT l. On the acts other hand, the required averagingprocessin this case(to obtain cell averages) degreesare as a kind of smoothing operator, which may not be attractive if higher Another point is that the spacelikemethod to be obtained from the measurements. the possibility of including the band limitation of the gradiometerinto doesnot offer the procedure in a proper manner, whereasthis band limitation is very likely to be present in a real mission.
134
4.3. Globalrecoverv
due to its setup in The timelike method does not have those disadvantages terms of measurementfrequenciesin c.p.r. (by using inclination functions). Furhas to be done in this method to create thermore, no averagingof the observations a blockdiagonal structure of the normal matrix. On the other hand, the influence of alongtrack data gaps may have more influence than in the spacelike approach. of A big disadvantage this method is furthermore the required CPU time, which, in caseof a real 6 months mission,will be large. Both methods suffer from aliasing effects. Since in our simulation we know where and with which potential coefficientsthe observationswere calculated, we could remove the contribution for degreeswhich we do not solve for from the observations. In reality, however, this can never be exactly done since we do not know these contributions, especiallynot the high degrees.A better test might be to set up and invert the mblocks of the normal matrix (the step in the computations where the coefficientsbecome correlated and aliasing finds its origin) for as much degreesas possible,but compute the vector D and solvethe systemonly for the requireddegrees. For example,in our casewe could have constructedand inverted the normal matrix for all degreesup to 360 (the maximum degree of the simulation) and solve only for degreesup to 180. In that case, at least theoretically,the aliasing effect must vanish.
135
Table 4.11
1 , ,
Preadjustment of the diagonal tensor components (if available) using Laplace' equation, including error propagation. Choice of the normal potential field I/. Given (/, tracking data and some first estimate of the initial state vector, compute a first nominal orbit. Compute at approximate points P' the normal gradients U;i, with which the gradient anomaliesAi;, are derived. Evaluation of the coefficientmatrix of the coordinate corrections U;ip at the approximate points P' in sphericalapproximation. Forming a linear combinatioD,.8. eq. 4.20 which meanselimination of the coordinate corrections. The apriori variance of the linear combination should be computed by error propagation. Least squaresadjustment to determineestimatesof the potential coefficient corrections LCt^o, based on the linear combination 4.20 and Air, (if available). Update of the nominal orbit with the old initial state vector and the updated potential coefficients e'r;". Compute at new approximate points P" the normal gradients 4;, with which updated gradient anomaliesAi;i are derived.
4 5 6
8 9
1 0 Back substitution to obtain coordinatecorrectionsA,2, eq. 4.21. 1 1 Improvementof the initial state vector elements L,zs,A,2g,L,is,eq. 4.22. L2 Update the nominal orbit again with improved initial state vector and
improved potential coefficientse!;".
1 3 Again update the gradient anomalieswith new nominal orbit. T4 Repeat the processfrom step 5 on until it converges. I f t re gradients Vr" and Vs" are available: t5
16 Back substitution into eq.4.19 to obtain coordinatecorrectionsAi Ly. and
Final improvement of the initial state vector with equations 4.22, 4.23 and 4.24. Compute final best estimateof the nominal orbit with this improved initial state vector and the final potential coefficientset.
t7
5 l
Relativistic vrew on gradiometry
the secondorderpartial derivativesof the As we have seen,a gradiometermeasures gravitational potential. The nine elementt {%rli, j : L,2,3} (of which only 5 are independent)together built up the gradient tensor, also calledtidal tensor or Etitvcis tensor. The latter names point at a geometricalinterpretation of the gradients in terms of curvature. Indeed, the elements of V;i are related to the curvature of the equipotential surfaces and the lines of force in the following manner:
K1 t
ft fz
H*
{v;i}: g
t r c z ft fz
cf. (Hotine, 1969), (Marussi, 1982) or (Rummel, 1986), where /{1 and rc2 are the normal curvatures of the equipotential surface in 11, respectively12 direction, t is the geodetic torsion, /1 and f2 are the components of the curvature vector of the line of force in cl, respectively c2 direction and I{* is the mean curvature of the equipotential surface. In this sense,gradiometry gives us insight in the geometrical structure of the space around the earth, being the result of the earth's gravitational field. However, these curvatures refer to either a two dimensional space (an equipotential surface) or a one dimensional space (the line of force). The embedding three dimensional space is just flat. This geometrical picture is appropriate in the Newtonian theory of gravitation. In Einstein's theory of gravitation, better known as the general theory of relativity, the viewpoint is different. The embedding space is no longer a three dimensional flat space (which of course is simply the space surrounding us, where we live in) but a four dimensional space,called spacetime, built from the usual three dimensional spatial part and the time as fourth dimension. This spacetime can hardly be visualised any more, but it has proven to be a useful mathematical concept in order to accurately describe physical features.
137
There are several reasonswhy a relativistic description of gradiometry may be of interest. We will see later, that Einstein's general theory of relativity is in fact a geometrical theory of gravitation. Geometry and gravitation become two indistinguishable concepts. Geometry plays an essentialpart in the theory of relativity and, since gradiometry is so closely related to Newtonian geometry, it may be interesting to find out how gradiometry looks like in relativistic terms. One of Einstein's conditions in the developmentof his gravitational theory was that, in the limit, it should reduce to Newton's theory. Indeed it can be shown that, after specifyingthis limit, Einstein's equationsyield the Newtonian equations as an approximation. Conversely,we may expect that Newtonian concepts have a generalization in relativistic terms which, at least conceptually, must be very similar. So we may expect that in some way a relativistic description of gradiometry will include the geometrical structure, especiallythe curvature, of spacetime. This indeed will appear to be the case. The general relativistic counterpart of the gradiometric model equation is the socalled equation of geodesicdeviation, in which relative accelerations are related to the curvature of spacetime. This curvature is fully described by the elements of the RiemannChristoffel tensor and it appears that the nonvanishing elements of this Riemann tensor in the Newtonian limit, appear to be elementsof the tidal tensor, the one shown above. In this way relativity contributes to the gradiometric measurementsthemselves since the Vi are part of the Riemann tensor. A second aspect which should be included in a relativistic description of satellite gradiometry, is the orbit of the satellite. One may be familiar with the ideas from special relativity (valid in the absence of gravitation) that features like time dilation or length contraction may play a role, at least if we consider particles moving with high velocities, such as earth orbiting satellites. However, satellites are in free fall in the gravitational field of the earth. In the general theory of relativity (with gravitation included) such motion is governed by the geodesicequation. Gravitation is not looked upon as some external force, but it is part of the geometry of spacetime. In fact, gravitation "curves" spacetime. Hence a satellite falling around the earth is regarded as a free particle, on which no forces act (disregarding disturbing forces like air drag etc.). Its path through spacetime is the "straightest" possibleline, a geodesic,but spacetime, and thus the geodesic,are curved due to gravitation. So, the equations of motion of a particle moving under the influence of gravitation, constitute a second object of study in which relativistic influences may play a role. A third point worth mentioning concerns the local referencesystem with respect to which the gradiometer measurementsare taken. In the Newtonian case, as described in the foregoing chapters, we use a local cartesian (or orbital) coordinate system with its origin in the center of mass of the gradiometer and one axis always pointing radially outwards. Another possibility would be to keep the gradiometer "inertially" fixed (with respect to far stars), so that the coordinate system would not rotate. In both cases,the formulation implies the possibility of using some preferential coordinate system (in the latter case an inertial system). Newtonian theory
138
is based upon the existence of such inertial system(s). But in the general theory of relativity, no (global) inertial systems exist, only local ones, valid over some small spacetime region. Furthermore, each coordinate system should be equally well suitable for the formulation of physical features, there is no system to be preferred. It is therefore common to work with four dimensional local inertial systems, defined in the framework of relativity as moving with the particle along its worldline (its trajectory in spacetime). Their motion is determined by parallel transport of the system axes along the trajectory. Due to the curvature of the space,the local system may therefore move somewhat different than it would do in Newtonian physics. In this chapter, we will pay some attention to the relativistic description of gradiometry. The first two aspects, (the relativistic contribution to the gradiometric measurements in terms of the Riemann tensor and the satellite's equations of motion) will be illustrated, the third aspect (parallel transport of the local inertial system) will not be treated separately.The reader is not supposedto be acquainted with the general theory of relativity. To this end, the chapter contains introductory sectionson the theory of relativity, as far as needed for our purposes. The intention is to sketch the outlines of general relativity and the way in which it may play a role in gradiometry. However, this thesis is certainly not a textbook on relativity, and we do not try to give a complete treatment of all aspects of the theory. Therefore, only a few of them are treated, some of which only briefly. Many textbooks are written on the theory of relativity. As for the general theory of relativity, we mention (Bergmann, 1942), (Misner et al., 1973), (Ohanian, 1976) or (Foster and Nightingale, 1979). Extensive referencecan be made everywhere to the literature mentioned above, as well as to others. Furthermore, recently several authors have published on the relativistic description of gradiometry. As for the three aspectsmentioned above, we refer to (Thei8, 1984), (Soffelet al., 1987), (Soffel,1989), (Ries et al., 1990), (Kopejkin, 1991) and (Gill et al., 1992). Furthermore, as gradiometry is expected to be a very sensitive technique (especially the superconductinggradiometerof Paik and Richard (1986)),gradiometricmeasurements could be used to test certain effects predicted by the general theory of relativity. See e.g. (Paik, 1989) or (Mashhoon et al., 1989).
5.1
theory
of relativity
It is very customary to distinguish between the specialand the generaltheory of relativity, abbreviated respectively by STR and GTR. The essentialdifferencebetween the two is that the latter is capable of incorporating gravitation into the theory, whereas the former is only valid in the absenceof gravitation. Despite this difference, and despite the fact that the special theory was developed first, the indication "general" and "special" suggest that the latter can be derived from the former by considering a special case (namely no gravitation), which indeed is true. We will therefore, in this chapter, focus on the GTR. Furthermore, since gravitation is the
139
major subject of this thesis, we will not specializeto the STR. Moreover, only those aspects of the GTR, which are either of special interest to us or are indispensable to a good understanding of the theory, are treated. 5.1.1 The spacetirne of relativity
In Newtonian theory, any event (or happening) can be described by stating where (space) and when (time) it took place. "Where" is usually indicated by three coordinates, {cilr : 1,2,3} (for example r,y,z). The chosencoordinate system thereby spans a three dimensional space. Analogously, we may look upon the time t as a coordinate spanning a one dimensional space,indicating the "when". In the theory of relativity, the three dimensional space and the one dimensional time are merged into a new, four dimensionalspace,called spacetime (Minkowski, 1952). Plotting the position of an event in an, in general,four dimensionaldiagram, gives its trajectory in spacetime, called the worldline. Usually a two dimensional subspaceis plotted, one axis of which is always the time coordinate. In order to put a geometry onto the space, we introduce a quadratic form ds2, calledthe line element(seeappendix B.1.3). This quadratic form is set up in terms of dr,dy,dz and dt, describingin a way the infinitesimal distancebetweentwo events. In the Newtonian casewe have in fact two such "distances" (Ohanian, 1976): dP  d* * dyz + dzz and dt
but in relativistic spacetimethere is only one: ds2 : "2 dt2  ds,2 dyz  dz2
(5.1)
where the scalar c is the velocity of light in vacuum (multiplication of dt with c matches the dimensions of the separate terms; with c included, the unit of time is chosensuch, that the velocity of light in vacuum is equal to unity). The "distance" between two events, expressedby the line element ds is an invariant, i.e. it is independent of the coordinate system used to express it. The specific form of equation 5.1 can be understood by recalling one of the fundamental postulates of relativity, which says that the speedof light c is the same in all inertial coordinate frames. Consider the motion of a light photon. If we have two coordinate systems cd a n d c r ( w i t h { r d l r : L , 2 , 3 } : ( r , y , z ) a n d{ r I l I = 1 , 2 , 3 } = ( X , Y , Z ) ) , i t s v e l o c i t y is c : dr/dt in the first coordinate system, and c : dR/ilT in the other, where dr2 : drz + dy2 + dzz and dR2 : dX2 + dY2 + dzz. The constantnessof c can then be expressedas: c 2 d t z d x z  d y ' dzz: "2dr2  dx2  dy2  dz2 :o
which is consistent with the assumption of the invariance of the line element ds in eq. 5.1, cf. (Foster and Nightingale, 1979). Eq. 5.1 shows that, for objects, other
140
than photons, the line element (also called sometimes spacetime interaat) can be positive or negative. lf ds2 is positive, the line element is called tirnelike, if it is negative spacelike,and if it is zero (like for a photon) it is called lightlike. It is convenient, and very customary in relativity, to introduce the following notation:
:i::;:::"]'t"i";'
where we adopt the convention to use Greek indices to denote the four relativistic spacetime coordinates and Latin indices to denote the three "usual" space coordinates. It is, furthermore, convenient to write
ds2 : eprdrPd.su
(5.2)
(5.3)
where
0 0 1 0
0 0 0 1
{nu"):
0 1 0 0 0 0
is the socalled Minkowski tensor. Comparing equation 5.3 with equation B.12 from appendix 8.3.1 we seethat the Minkowski tensor corresponds the metrical tensor, to describingthe metrical propertiesof the space. Furthermore,sinceit is diagonal and has constant elements, learn from this appendix that it is the metrical tensor for we flat space expressedin "cartesian" coordinates. This space is the spacetime from STR. The Minkowski metric replaces the three dimensional cartesian metric if we switch to four dimensions. In the Newtonian formalism, in three dimensions, it is possible to introduce a cartesian coordinate system covering the entire space. Such a coordinate system is called (globally) inertial. All other coordinate systems, connected to such an inertial system by means of linear coordinate transformations (not involving time), socalled Galilean transformations, are also inertial systems. In relativistic spacetime, the coordinate system (ct,x,y,z)rfor which the line element 5.1 holds, is the generalization of a Newtonian inertial system. Different inertial systems are connected via the wellknown Lorentz transformations, which are still linear coordinate transformations, but now the time t is transformed too. Accelerated coordinate systemsare not inertial. In the presenceof a gravitational field, all coordinate systemsare accelerated,so there is no privileged system (inertial
141
5. Relativistic view on gradiometry system) in which the metrical tensor takes on the simple Minkowski form. Therefore, gravitation cannot be included in the STR. Einstein solvedthis problem when setting up his GTR. We will see later that, in the GTR, we are no longer restricted to linear Lorentz coordinate transformations, but more general nonlinear coordinate transformations are involved. The metrical tensor is no longer diagonal and its elements are in general functions of the coordinates, instead of constants. We will, furthermore, see that in this case the space is no longer flat, but curved, so that it is no longer possible to introduce into the space a cartesian (inertial) coordinate system which is "global in the sensethat its coordinate neighbourhood is the whole of spacetime" (Foster and Nightingale, 1979). In a curved space such a coordinate system is only valid in an infinitesimal region of the space ("locally"). 5.L.2 Equations of rnotion
Considera particle freely moving in space.Sinceno forcesact on the particle, it will of is move along a geodesic.A geodesic the generalization the concept of a straight line in three dimensional Euclidean space (Es) to more general curved spaces of arbitrary dimension. This generalization is done by requiring that there exists a parametrization u of the geodesic,such, that the tangent vectors 4rttf du along the curve constitute a parallel vector field. The concept of parallelism is generalizedby Dup(: ap.rdr', seeappendix B'2'3)' The parameter deriuatiue meansof the absolute u is called an affine parameter. In the space under consideration the geodesicgives the shortest path between two points. Newton already stated (in his first law) that each particle at rest or in uniform motion along a straight line preservesits state if no forces are exerted on it. Considering a free particle to move along a geodesicis therefore nothing more than a generalization of this Newtonian law to arbitrary spaces. The equations of motion of the particle may thus be representedby the equation of a geodesic:
i l P+ l f ; o i " i " = o
(5.4)
(seeappendix 8.2.3). In equation 5.4 the dot denotesdifferentiationwith respectto the affine parameter u, i.e. i" : dr" ldu, and thetfi are the Christoffel symbols (of the second kind), which may be computed from the metrical tensor I pv as r3 : Lrtee (g,p,ot 9op,, 9ro,p)
(5.5)
(compare eq. B.15). Note that, if we would like to insert for the parameter u time f, we should bear in mind that, since time is in four dimensional spacetime one of the coordinates, it is no longer universal for the whole of space, so it is not an affine parameter. Therefore, we introduce the socalled proper tirne r, which is measured by a clock, comoving with the particle, and not by some kind of for "inertial ticking" clock. This has also consequences the velocity of the particle. In three dimensional nonrelativistic space, the components of the velocity vector
142
5.1. Someaspects the general of theory of relativity are ud : doildt. In four dimensional relativistic spacetime, we have to take the derivative with respect to the proper time r. We obtain the socalled 4velocity ur' : d* /dr, also called world velocity. We may of course also differentiate the coordinates with respect to the coordinate time t. Then we obtain the coordinate : dt/dr velocity utt  4suf dt  upll. The quantity "y is given by I : Ll ,/T=777 (with u2  uiui), which appears in the four dimensional Lorentz transformation of STR. With eq. 5.2 we may write the coordinate velocity as ur: (",rt). Thus, if the elements of the metrical tensor are given we may compute the equations of motion of the particle by means of equation 5.4. Often, however, the metric has a complicated form, and one would have a hard time computing the equations of motion in this way. One may therefore proceed in another manner (Foster and Nightingale, 1979). The equations of motion can also be obtained by using the EulerLagrange equations:
!(e!\_?L:o
du \)ir') 1ru
(5.6)
which can be found in many textbooks on classical mechanics, tensor analysis or relativity, e.g. (Goldstein, 1980), (Sokolnikoff,1951) or (Ohanian, 1976). In this equation "L is the Lagrangian, which is a function of the independent variables ip and rp. Let us consider the following Lagrangian: [,2 : grriryi'
(5.7)
As we will show now, with this .L the EulerLagrange equations reduce to the geodesicequation 5.4. (Note the resemblanceof this .t with the infinitesimal length (or line element)ds (equation B.12) from appendix B.3.1, which will appear to be more than just a coincidence.)From equation 5.7 we obtain:
a L
U*=
t a
*t*(g'oi"i")
:
:
+ fi(n,,uii" s,,i"6fl)
L
. v
79P"x'
and on multiplying with the inverse (or associate) metrical tensor gop and using eq. 5.5 together with the identity
l gpr,oiv id = g ur,oi" i" 1 +'rg uo,ri' io
we obtain the geodesic equation 5.4. This means that the EulerLagrange equations 5.6 are equivalent to the geodesic equations 5.4, at least if one uses a Lagrangian ^Lof the form eq. 5.7. Evaluating equations 5.6 for deriving the equations of motion is often more convenient than straightforward application of the geodesic equation. The variational principle and the equations of rnotion
Geodesicsare the "shortest" paths between two points in the spaceunder consideration (e.g. in Euclideanspacea straight line). Considera particle moving along such a geodesicand supposewe have parametrized the curve with the arclength s, instead of some arbitrary parameter u. It can be shown (Foster and Nightingale, 1979) that the arclength s is an affine parameter. The parameter s can be used as a measure of the length of the curve. It appears in the expressionof the invariant line element ds: ds2 : gprdrpdr, With the help of this line elementds, the distanceS betweentwo points on the curve as: and P2, with coordinatesrp(s2), can be expressed P1, with coordinates "p(sr),
S:
I ds y,S t
fsz
Iu','{%rt",'d*
J,,
g urittiv ds
_ ["' 
(5.s)
we where tt' : drpf ds. Sincewe are dealing with a geodesic look for the "shortest" distance between the points P1 and P2. The integral in equation 5.8 therefore has to have a stationary (in our caseminimal) value. This value can be found with the help of the socalled calculusof oariations. If the geodesicS has a minimal value, it will grow if we replace rp(s) with an arbitrary function
cP(s) * 6cP(s)
(5.e)
in which the 6cp(s) are small in the interval sl < s ( s2 (we call 6cp(s) the aariation of rP(s)). So eq. 5.9 would representanother curve through the end points P1 and P2 in the neighbourhood of the geodesic.Since this other curve also passesthrough P1 and P2 it is :6 (5.10) 6zP(s1) rp(sz):0.
144
If we now put Jg;EW: f (rP,it') we may write the change in distance between P1 and P2, being the result of replacing cp with xt'+ 6rp, as
ut :
1",'f(ru
(5.11)
is the equation determining the geodesic. This equation has exactly the same form as the EulerLagrange equations5.6 if we put f : L, so that the Lagrangian from eq. 5.7 is indeed identified as the line element. Already at this point we may suspect an important concept which is used in the GTR. In classical mechanics, the Lagrangian L is related to the energy of the particle. If a Lagrangian of the form as in equation 5.7 properly describesthe motion of the particle, then the expressionfor the line element ds must contain in some way the particle's energy. As the drp are nothing but infinitesimal coordinate differences, the energy must be present in the elementsof the metrical tensor! This notion is very important and we will come back to it in the next section. 5.1.3 The principle of equivalence
According to the principle of equiualence is not possibleto distinguish between it gravitational and inertial accelerations. Consider for example an observer (astronaut) in a freely falling spacecraft. Supposethe spacecraft has no windours and the astronaut awakens after a long sleep. He does not know whether he is still in orbit
145
around the earth or has drifted away from the earth and from all other attracting masses. In order to find out he releasesan object (proofmass) and observes its motion. In both cases,the proofmass will stay floating in front of him. If no forces act on either the astronaut or the proofmass, this can be understood very easily. Newton already stated it in his first law (law of inertia). The spacecraft, astronaut and the proofmass are in rest or in straightline, uniform motion and will persist in doing so, as long as no forces are exerted on them. If ad is the acceleration of the body with respect to some inertial coordinatesystem ud, the equationsof motion of the body can be describedby:
od:O
i.e. the motion is unaccelerated.If we would describethe motion of the spacecraft with respect to some other coordinate system tr', which rotates with a constant a' angular velocity with respectto x,i, the acceleration is no longer zero. Multiplying of by theseaccelerations the masses the bodies,rna',we obtain socalledapparent or inertial forces, which seernto act on the bodies (centrifugal and/or Coriolis forces). These are in fact no real forces, since with a linear coordinate transformation c' : ,i(r') we can always "transform them away". Let us call the mass rn mentioned above the inertial mass. If there exists a coordinate system with respect to which all bodies not subjected to (real) forces are in rest or in uniform motion, it is called an inertial system. The cd would be such inertial system, the c' not, becauseof its rotation. However, all coordinate systems which are at rest or in uniform motion with respect to each other, are inertial systems. If now the astronaut would find himself in free fall in a gravitational field, a force does act on him and the proofmass. This gravitational force is proportional to the mass of the object it acts on: fi : Mgi with M the graaitational massand gi the gravitational acceleration. So if the motion of the astronaut is due to a gravitational force, it undergoesan acceleration , M , o.:ig Currently we know that for all objects the ratio Mf mis equal to I (accuratelyto lott), i.e. the gravitational force results in the same accelerationfor all objects, independent of their mass. This is exactly what the principle of equivalencesays: there is no difference between inertial and gravitational mass. So also in this case the proofmass will remain floating in front of the astronaut. Obviously, there is a coordinate system, comoving with the spacecraft,relative to which the gravitational force can be "transformed away", as if it was an inertial force. However, there is a solution to the astronaut's problem. Due to the convergence of the lines of force of the gravitational field of a spherical body, the proofmass will very slowly move relative to the astronaut (often referred to as moving under the influence of tidal forces). Note that this is exactly the reason why satellite
146
gradiometry is possible. The relative motion as a result of the radial force field provided we limit the validity does not really violate the principle of equivalence, of the principle to infinitesimal small regions of space,where the tidal forces can be considerednegligible. The comoving coordinate system, extended over this small region (often referred to as local coordinate system) looks very much like a real inertial system, of which the metric is describedby the Minkowski tensor. In the STR (valid in the absenceof gravitation) the laws of physics apply with respect to inertial coordinate systems. Noninertial systemscan always be transformed to inertial systems by suitable Lorentz transformations. According to the principle of equivalence, all frames of reference, including those in which a gravitational field is present, are equally well suited for formulating the laws of nature, since we have seen that gravitational fields can (at least locally) be transformed away by a suitable coordinate transformation. In this sense,the difference between the STR and the GTR is that in the latter all continuous,differentiablecoordinate transformations are considered,whereas the STR "only" involves linear Lorentz transformations. In the literature one finds many different formulations of the princoordinate systems,coordinate in ciple of equivalence, terms of mass, accelerations, transformations, STR or GTR, validity of the laws of nature or physics, etc. etc. Seee.g. (Bergmann,1942),(Fosterand Nightingale,1979),(Ohanian,1976). We have seen that a free particle moves along a geodesicof which the equations are given by 5.4. This equation is a tensor equation,that is, it has the sameform in all possible coordinate systems. If a gravitational field is present, the particle will no longer be a free particle since it moves under the influence of gravitation. However, according to the principle of equivalence,the gravitational field may be "transformed away" (at least locally) by a suitable coordinatetransformation. Indeed,an observer moving along with the particle and carrying a coordinate frame with him, will considerthe particle to be a free particle with respectto his coordinate frame, just like the astronaut in the example above. For this observer,equation 5.4 still holds for the motion of the particle. Then this equation should also hold in all other coordinate systems, including those not comoving with the particle (in which the particle d,oes accelerate).In that caseit must be so that the gravitational effects are already present in equation 5.4. They must be included in the Christoffel symbols th. The Christoffel symbols in turn, are related to the metrical tensor gp, (see equation 5.6), so that we conclude that the gravitational field must be present in the elementsof gpv, or as Bergmann (toaZ) writes: "t'he g* are the potentials of the gravitational field". Since the gravitational potential is in generala function of the coordinates,the elements of the metrical tensor will also be functions of the coordinates. They describe the geometricalstructure of the space. In flat space, in global cartesian coordinatesrgprhas a diagonal form with constant elements. In such a coordinate system the I;[ are zero and equation 5.4 reduces to
'iF :0 .
147
If a gravitational field would be present it would appear as if we could transform away the effects of gravitation globally by a coordinate transformation to global cartesian coordinates. We know, however, that the principle of equivalence only holds locally (in an infinitesimal region of space). Apparently we are not able to introduce a global cartesian coordinate system. In such a situation we infer from geometry (see appendix 8.1.3) that the space is curved. The conclusionmust be that gravitation curves the space. Note that mathematically the correct formulation for a curved space is a nonvanishing RiemannChristoffel tensor. of The aboveconclusionimplies that, in the GTR (i.e. in the presence gravitation) gravitation is no longer considered a force, but a part of the geometry of the space. A free particle is now meant to move under the influence of gravitation alone, and it will still move along a geodesic,the geodesic,however, now being some curved trajectory, but still the "shortest" distance between two points, but in a curved space. This notion is the most important concept in the GTR. Gravitation and geometry are no longer two different concepts. The theory of gravitation becomes a geometricaltheory in which the usual geometricalformalism can be applied. This is also one of the reasonsthat index notation and tensor analysis are much used in the GTR, since they are both especiallysuitable to geometricalproblems. Note that in the above reasoning the crucial point was the validity of the geodesicequation in all coordinate systems. Since this is a natural result of the tensor formalism, one can understand Einstein's choice to use tensor analysis as mathematical tool in the developmentof the GTR. 5.1.4 The Einstein field equations
In the foregoing,we have seenthat, in the GTR, based on the principle of equivalence,the gravitational potential is presentin the elementsof the metrical tensor. In Newtonian theory, the gravitational force field, being the gradient of a gravitational of potential, finds its origin in the presence matter. Poisson'sequation TP'Vrr: 4rPG
(orin index free notation Y2V = atrpG) where p is the mass density,V the gravitational potential and G the gravitational constant, describesthe relation between this gravitational field and the matter distribution. This equation is referred to as field equation. When developinghis GTR, Einstein was searchingfor an equivalent relation, with on one side the gravitational potential (in terms of (functions of) the metrical tensor g pr) and on the other side some quantity expressingthe matter (or in general: energy) distribution. Naturally, the equation had to be a tensor equation (since it must be invariant with respect to coordinate transformations), so the quantity expressing the energydistribution had alsoto be a symmetric, secondorder tensor, tensor. It is called the energy tensor (also called: energymomentumstress and it is denotedby Tuu. [t may contain stresstensor, stressenergytensor, etc.), the matter density, momentum, as well as other kinds of energy,like electromagnetic
146
(5.12)
(i.e. vanishing divergence) which merely expressesconservation of energy. A last condition for the relativistic field equations is, that they have to reduce to Poisson's equation in the Newtonian limit. After a long search,Einstein came up with equations fulfilling all conditions, the famous Einstein field equations:
I Rttu_lRopr:KTFU
"
(5.13)
where Rp' is the contravariant Ricci tensor, R the curvature scalar, equal to a contraction, R,u, of the Ricci tensor (the trace), and rc some constant, equal to 8rGf ca. Equivalently one may write this equation in covariant form, with lower indices, or, after some manipulations, in the following form: Rtu : n(71r,  lTnrr) , (5.14)
with 7  Tuu. The lefthand side of equation 5.13 is called the Einstein tensor Gpv. Since the Ricci tensor is a contraction of the RiemannChristoffel curvature tensor, i.e. Rpr: Ropvotand since the latter contains the secondorderderivatives of the metrical tensor, the field equation is indeed very analogousto the Newtonian situation, where we have also on the left hand side somefunction of the secondorder derivativesof the gravitational potential, and on the right hand side the sourceterm. ln general, the Einstein field equations are nonlinear. Furthermore, since the Ricci tensor has 10 independentcomponents,equation 5.14 is in fact a system of 10 nonlinear equations. Solving these equations for gp, if some information or assumptionson the energytensor are given, is thereforea very tediousj ob. Whereas the field equation is a tensor equation and thereforevalid in all coordinate systems, one coordinate system may appear more convenient for finding a solution of the field equations than another. In order to simplify the searchfor a solution, one may, under certain assumptions, linearize the field equations and put some constraints on sections. the matter distribution. This will be done in subsequent
5.2
'Weak
field approximation
(5.15)
where 4p, is the Minkowski metric and the hu, are small quantities, such that products of them may be neglected.Sincethe elements the metrical tensor include of
149
5.
Relativistic
view on gradiometry
the gravitational potential, the latter must now be presentin the hpr. Scalingwith the velocity of light c to obtain dimensionlessquantities, we have the condition
(5.16)
If this condition is fulfilled (which, in many applications,e.B. computations in our solar system, is usually the case)one often calls the linearization 5.15 the weakfield approrimatioz. But it is in fact not the above condition that makes the field weak. A weak field is obtained only if we assume the derivatives of h* to be small too! The value of the gravitational potential itself does not tell us anything about the weaknessor strongnessof the field. We have to look at the amount in which the potential varies throughout space,since it is the potential differencewhich produce the gravitational force, i.e. the potential derivatives. In general the metrical tensor can be used to raise and lower indices (see apwe pendix B.3.2). In the linearized caseof eq.5.15,however, shall,sincethe h* are small, raise and lower indices using 4rr. 5.2.1 Newtonian lirnit
In the developmentof the GTR one of the conditions was, that, in the limit, the by theory reducesto the Newtonian case. This limit is partly characterized the weak field approximation, as described above. But, to arrive at Newton, some additional assumptions have to be taken. One of them is the assumption of a quasistatic gravitational field. For a quasistatic field, the time derivatives of lhe hp, are small compared to the spacederivatives,i.e.
#n,,:!*n*u*ou,
(5.17)
and are therefore neglected. Furthermore, we only consider particles which move Ll\nuzJA ry 1 so that with low velocity u, i.e. v 11 c. For such a particle l: equation 5.4 and neglectingterms utsx r.tP. Using theseassumptionsin the geodesic quadratic in u, we obtain: d'ru , nrdro dro +LdoEE:o'
dr,
150
5.2. Weak field approximatton where we used the stationarity condition 5.17. The equation of motion for the p : O component will in this case read
p:o ,
"#
: o,
1, and furthermore implies our assumption'y : dtldr: which exactly expresses that i0 : c. Therefore, switching from r to t, the spatial part of the equations of motion become . )2a : Lrr'q'i hoo,i F:t :
ifr
Using wellknown symbolic notation, the right hand side of this equation can be delta) and 6ij # V. written.. f,Vnos, where we used nii : 6ii (Kronecker From Newtonian theorv we know d2r;
orz
: 6'1vi
with V the Newtonian gravitational potential, so that, in the presentapproximation, the relativistic theory will reduce to Newton if we identify: hoo: '* with which the 966component becomes:
.rl f
(5.18)
2V 9oo:1 ,
(5.le)
Note that this expressionfor 9oo (Newtonian limit) is derived by comparing the relativistic equations of motion in the weak field, slow motion approximation, with the Newtonian equations. Nothing can be said at this point about the other elements of the metrical tensor (gfi and 9;l), cf..(Misner et al., 1973). In fact, it does not matter how those other componentslook like, because,to the present level of approximation, they will not contribute to the equations of motion, because they are multiplied with small quantities u or v2. In the nonrelativistic limit, the dominant component of the energy tensor is Too: pcz (Ohanian, 1976) where p is the mass density. For this OOcomponent,the Einstein field equations5.14 are:
&o:"(t*
trn*)
^, Roo #ot
5. Relativisbicview on gradiometry
(5.20)
which is the familia,r Poisson equation. We see that, indeed, with the present assumptions and level of approximation, the relativistic theory reducesto Newtonian theory. 5.2.2 Linear approximation in relativistic terms
When working in terms of relativity, we no longer assumethat we have a quasistatic field. Furthermore, we no longer deal with small velocities. We proceed, in this case,by inserting the metric 5.15 into the field equations5.13. It can be shown, e.g. (Foster and Nightingale, 1979) or (Ohanian, 1976), that the field equations reduce to (5.21) o h p r , o o + ( r 1 u r h " p , o  h o r , u o h 1 n , v o ) : 2 n T p , where
(5.22)
If we would carry out a small coordinate transformation, rt'' "+rp * (e, a straightforward calculation shows (Foster and Nightingale, 1979) that
htt'r' *htt" 4ro p,o npo r,o * rlp, o,o
(5.23)
so that
lr''o',o, ' h4o,o  u,oo .
(5.24)
This is called the gauge condition. The quantilies hp, are the potentials of the field. The field equations will not determine these potentials uniquely. We are from eq.5.23, becauseit will leave the field free to replace h* with the expression equations intact. Since we are free to choose the coordinates we like (because all equations are tensor equations), it is always possible to transform to a coordinate
152
system such that the gauge condition is fulfilled. This particular choice for (p, will simplify eq. 5.21 a great deal, becauseall three terms between brackets are zero. We obtain (we drop the primes) :
hpr,oo :ZnTp, One often introduces the d,'Alembert operator !, ' defined by
(5.25)
J: e"'
a
_a7Z_
v 
, , a t ,,
e O
which is the extension of the Laplace operator to four dimensional spacetime. With this operator, the simplified field equations 5.25 (in contravariant form) may be written as (5.26) aht'"  2rcTp' which holds provided the quantities hr'" satisfy the gaugecondition (cf. eq. 5.2a):
hr',: o .
Equation 5.26 is the extension to four dimensional relativistic spacetime of the Poisson equation in nonrelativistic three dimensional space. The generalization to relativity of the Laplace equation would be 'htt' : O
which is valid in empty space. The d'Alembertian is known from the theory of electromagnetism. It is sometimes called the "wave operator", and an equation like 5.26 is referred to as a wave equation. Analogously to electromagnetism, one therefore assumesthe existence of gravitational waves, generated by a source term 2nTt"', and propagating through space like radiation. We will solve the field equation 5.26 for some special (simplified) situations. To do this, consider an isolated gravitating mass. We will (as we did in the Newtonian limit) assumethis mass to be stationary, so that both ht" and TY' are only functions of the spacecoordinatescd. With this assumptionthe d'Alembertian reducesto the Laplacian V2 so that we have to solve the equations: Vzhp" (x) : ZnTpu (x) where x denotesthe position vector in spatial coordinates.The generalsolution to this equation is
2rJx lxx'l cf. (Ohanian,1976),where the integral is taken over the region E containingthe mass, and the prime denotesthe integration variable. Expanding tll*  x'l into
153
a Taylor series and truncating after the second term (i.e. we neglect higherorder multipoles of the mass), we obtain:
. lr,'rr,, t*\dE
t
(5.27)
With a few manipulations, one can show that, in this expression,the integrals
I rira> Jx,
vanish. The vanishing of the first and last of these three integrals is a consequence of the conservation of energy (since the energy tensor satisfies the conservation law eq. 5.12), whereas the vanishing of the second integral expressesthe choice for the origin of the coordinates to be in the center of mass. In the following we shall give, for a few simple masses,the components of the energy tensor and derive, by means of the linearized field equations, the expression for the metrical tensor. The field of a spherical, stationary mass The elementsof the energy tensor Tp' for a spherical,stationary and nonrotating mass, are: Too : pc2 ? 0 d: 0 Tii:O where p is the mass density, which can be obtained by comparing the energy content of spacetimewith that of a perfect fluid without internal stressand pressure,of which in the particles are motionless. Using theseexpressions equation 5.27,we obtain for the hr'":
while all other components vanish. From hp' = ht"  ihnu" (which follows from eq. 5.22) we obtain
ht": '4#uu' ' rcWith these disturbances. the elements of the metric tensor become: goo:1
2U
o
(5.28)
,fr ,l + ;)
154
5.3. PostNewtonianapproximation
stationary
rnass
Consider now a rotating mass,howeverstill stationary (i.e. negligibletime derivatives). The mass particles,out of which the body is built up, have, in this case,some spatial velocity, ,d. The elementsof the energy tensor are: Too : pc2 Toi  pcvi Tii : pviuj Inserting these valuesinto eq. 5.27 yields:
,a
roo _ 
4GM
rc'
,
: ho' #rj
hij:o where
#r,'or,ro lror'1u;ar:
9 o oI : r y
;^k ooi: fiei;*xr s* 2G
(5.29)
,ll
Assume that the mass rotates around the c3axis with constant angular velocity or. Then the elements96 become:
goi:
>
(5.30)
5.3
PostNewtonian
approximation
Let us compare the metrical elementsfor the Newtonian limit, a spherically symmetric, stationary mass and a rotating mass, respectivelyeq. 5.19, 5.28 and 5.29. We seethat the term proportional to c2 in the goocomponent is not a relativistic one. The factor c2 arises from the fact that the gs6 corresponds to the time coordinate,
155
which is c0: ct. All other terms in the three metrics that contain c, ate relativistic contributions. These metrics were obtained by considering simplified situations. In general we have to do with more complicated forms of the energy tensor ?p/, with more types of energy included (e.g. internal stress and pressure)and with a nonspherical and nonhomogeneousmass distribution. This may make it difficult to obtain an analytical solution of the field equations. Following the line of construction of the metrics above, it seemsa good approximation to expand the elementsof the metrical tensor in terms of the small parameter tJf c2. The smallness of this quantity was one of the conditions for the weak field approximation, equation 5.16. Starting from the Newtonian limit, equation 5.19,the first step would b" 9ooup to 0("n),9ii up to 0(c2), and 96 up to 0(c3). Following steps would include terms of respectivelyO(c6), 0("o) and 0(c5). this way of approximating the metric is called the PostIfewtonian approrirnatioa (PN approximation), and the metric is called PNmetric, see e.g. (Misner et al., 1973). Including, for all elements, only the first relativistic contribution, we obtain the socalled 1PNmetric, given by :
4 e o o :1 * T . o ( "  u )
s o i :+ + o ( c  s )
(2 : e ; 1 _ . 6 ; i+ t c4 '1 + O ( c  a ) t
cf. (Soffel,1990), (Mashhoon et al., 1989),or (Gill et al., 1992),where [/ is the gravitational potential of the mass and V; is a vector potentiall describing the field contributions of magnetic type, arising from moving masses(compare the go in metric 5.29). Note that, whereas the metrics 5.28 and 5.29 the potential U :GM/r in comes from a spherical mass distribution, the potential [/ and I in 5.31 (and also in 5.19) are generated by arbitrary massesand are determined respectively by the equations: (5.31)
v2u  4rGp
YzV:4rGprt; cf. e.g. (Chandrasekhar,1965),where ui are the spatial velocity componentsof the mass points. So in general
rA note on notation: The use of the notation U for the Newtonian (scalar) gravitational potential and %' for the vector potential is customary in PN theory. For this reason we adopt this notation here, in apite of the fact that we, in previous chapters, used V for the Newtonian potential and % for the first derivative (gradient) of the potential.
156
Expanding tll*x'l into a Taylor series,we obtain, for a nearly spherical body, metric 5.29 as a first approximation, i.e. U x GMlr and V; x ffieprisk (cf. (Soffel,1990)). It must be noted that the expansionof the metrical elementsin this way does in generalnot constitute an exact analytical solution of the Iinearizedfield equations. The PN formalism simply assumes the existenceof a coordinate system where the metriial tensor has the form 5.31, cf. (Soffel,1989).
5.4
Equations
of motion
revisited
If we are given some expressionfor the elementsof the metrical tensor, we may compute the equationsof motion, either by meansof a direct evaluation of the geodesic equation 5.4 or using the EulerLagrange equations 5.6. In relativistic terms, the gravitational potential is included in the metric, so theseequationsdescribethe motion of afree particle (i.e. moving under the influenceof gravitation alone), e.g. an earth orbiting satellite (if we neglectdisturbing forceslike drag or solar pressure). The geodesic describedby equation 5.4 is parametrizedwith an affineparameter u, in our casethe proper time r. For practical computations,one likes to obtain the equations of motion with respect to the coordinate time t. These can be obtained in the following way. We replace the differentiation f, UV #jf . fnen we obtain: drP dr dzrq dzrq / dt\z
dx,P dt
d,r:lrr\a) *a,r\a) E
rh#ff:o
dzt dro
d t,
ar' at
(#)'
:
+ dzt ( dt\' drz\dr)
d2ct
d,tr:o
so that, if we use this in the equationsof motion for the spatial componentr (p : r), we obtain:
4!"_r:g
157
or
r:"#" #.(,,
t l
(5.32)
as the spatial equations of motion in coordinate time. For the computation of the equationsof motion one needsthe Christoffel symbols. They are computed using equation 5.5. In the following we will neglect the time derivatives of the metrical elements, since we assume they will be small compared to the spatial derivatives. Furthermore, we will put c = 1. Since c is constant, it acts as a kind of scaling parameter, so that putting it equal to 1 merely means a change of units. When numerical values have to be computed, one inserts in every term as many c's as needed,in order to match the dimensionsof all terms, cf. (Ohanian, 1976). The PNmetric for which we will give the equationsof motion is
(5.33)
where the parameters B and 1 are added for convenience. If B :1 and ? : 1, we obtain eq. 5.29, the metric of a we obtain the lPNmetric 5.31. If. P :0 rotating, stationary mass, and if Vi : O, we obtain 5.28, the metric of a nonrotating mass. In the socalled ParametrizedPost*Newtonian (PPN) theory, valid for more generalmass (or in generalenergy) distributions, other types of potential functions are included in the metric, eachwith its own parameter as factor, seee.g. (Misner et al., 1973) or (Will, 1981). Nowadays it is common to include 10 such parametersin the PPN metric, two of which are B and'y, as above. Depending on the value of these parameters, different relativistic theories are obtained (also called metrical theories of gravitation). Einstein's theory of gravitation does not include all these extra PPN terms (i.e. they do not appear in the solution of the Einstein field equations). However,if we chooseall PPN parametersto be zero except for B and 7 which we chooseto be 1, Einstein's theory is included in the PPN formalism. The latter is what we will do here, so we insert I : 1. The parameter B is retained and will be propagatedinto the results so that the influenceof the term containing it will be clear. For the lPNmetric 5.33 the nonvanishing Christoffel symbols become:
[b:u,;(rz(B+1)u) Ib?:u,i(rz(B1)u)
l,!  O;eU,t 6i*U,; 6;iU,* * rii  z(vi,j  Vi,t) (5.34)
\l :2(v;,,+v1,;)
158
where the comma, as usual, denotes partial differentiation. In order to agree with much used (index free) notation, we introduce: (r,y, particle's spatial position vector particle's spatial velocity vector (the dot means differentiationwith respect to t)
")
@,!,t)
VU
v
u2
With this notation, the equations of motion will become (cf. (Soffel,1989), (Mashhoonet al., 1989)or (Ries et al., 1990)):
d  : o tu r l r * A [  r t B + t u v u  4 ( v u . v ) v *, z v u  4 v x ( v x v ) ] , ( b . 3 5 )
where clearly the term betweensquarebracketsis the 1PN relativistic contribution to the equations of motion. The dominant relativistic contribution will come from a spherical,nonrotating mass, eq. 5.28. If we, therefore,insert V :0, I : L and U : Y (so that YU : #*), equation5.35will reduceto dv GM
13
x+
GMI
GM
,I
r%, Ln;**a(x'v)vt'"x.J
(5.36)
These equations may be solved for x to reveal the relativistic effects in the particle's motion, see e.g. (Soffel,1989). Here, we are interested in the relativistic contributions to the motion of an earth orbiting satellite, in particular the Aristoteles satellite. So we take for M the massof the earth and for {ri1;:1,2,3}: (r,y,z) a geocentriccoordinate system. It should be noted that this coordinate system, despite the commonly used notation (c, y, z), is not a rectilinear (cartesian)coordinate system. In fact, the three coordinatesare the spatial part of the four dimensional coordinatesystem (t,r,y,z) which labelspoints (events)in spacetime.And we have seen that spacetimeis curved, compare the metric 5.33, so that no rectilinear coordinate system can be introduced on a global scale, only locally. In particular, for a spherical symmetric mass distribution, the curvature of spacetimemanifests itself in the radial direction and in the time coordinate. This feature was already used by K. Schwarzschild 1916 who found the first exact solution of the (nonin linearized) Einstein field equations. His famous solution for a static spherically symmetric gravitational field (nowadays known as The Schwarzschild solution) was
159
5. Relativistic view on gradiometry put in terms of polar coordinates(t,r,0,,\). In this casespacetimeis curved in the r two coordinates (t,r) . As a consequence does not measurethe usual radial distance. However, the geometry for those parts of space which have fixed t and r is just that of a usual sphere. It can be shown that, by a proper choice of coordinates, the Schwarzschild solution gives the lPN metric as a first approximation, at least the lPN for a static (i.e. nonrotating) spherical symmetric mass. This means for the coordinates{rdlr : 1,2,3} : (r, y,z), if one would relate them to some kind : (r,0,)) by meansof a coordinatetransforof polar coordinates{r"la:1,2,3}  'i(r") with c: rsin0cos), !: rsindsin) and z: rcoso that r is mation ti not the usual radial distance. In order to deal with this problem, one sometimes where c acts as a introduces a new radial coordinate r', defined by r' : r + aff the choiceof a definite coordinate system (Soffel,1989). As a gauge parameter for all consequence orbital parametersused to describethe satellite motion depend in this senseon the choice of the coordinatesystem. Nevertheless,we will solve the equations of motion 5.36 by interpreting the satellite motion to be the solution of a perturbed Keplerian problem, where the secondterm on the right hand side (the one with the squarebrackets)is regardedas the perturbing force. If, in our simulation, one subtracts from the solution obtained in this way, that of a simple Keplerian problem with the sameinitial state vector but without the relativistic term, one obtains the relativistic contribution to the orbit of to by the satellite. We used a computer program developed E. Schrama (1SSZ) carry out such computations. In reality, when working with real data, one solves,among other parameters, the elements of the initial state vector from the observations in the orbit determination process. In that case the relativistic effect cannot be determined in the way described above, since both orbits (with and without the relativistic perturbation) would have different initial state vectors and can thus not simply be subtracted. Furthermore we have, in that case, to bear in mind that existing values of parameters,which are used in the computations, could originally have been determinedwithout relativistic models,so that possiblerelativistic effects are implied in those parameters. It is well known that the dominant relativistic contribution in the motion of a particle in orbit around a spherical symmetric mass (whether it is an artificial satellite around the earth or a planet around the sun) is an advanceof the periapse. This can be seen in figure 5.1. In this figure the relativistic orbit contributions for in the Aristotelessatellite are shown,expressed terms of radial, crosstrack and along This figure was computed in the way describedabove, namely track components. by solving equation 5.36 for x and subtracting the solution of the same equation without the relativistic term. The figure revealsthe dominant relativistic orbit contribution as a linearly Srowing along track effect to be interpreted as the perigeum advancementioned above. It reachesover 38 m after two weeks. The figure also shows a periodic effect with a frequencyof once per orbital revolution. This can be clearly seenif one removesthe linear trend and computes the power density spectrum, as is done in figure 5.2 for
160
0 !
744
t9?
time (h)
Figurc 5.1
orbit contribu,tions tlte Aristoteles Relatiuisiic satellite where for only tlte cen"tral terrn of the potentialof the earthwas included, (socalled Schwarzschild effect).Satellite heightwastakento be 200km and orbitinclination95".3. Thealongtrackperturbotion is plottedin m, the otherperturbations cm. in
the along track effect. For the radial perturbation this periodic effect, with a linearly growing envelope,can be seenin figure 5.1. Nevertheless the radial perturbation is much smaller then the along track effect, being of the order of a few cm after two weeks. The cross track contribution remains below the level of lOa m. The along track relativistic effect can not be separatedfrom a change in the value of the gravitational parameter G M , which relatesthe satellite'smean motion to the semimajor axis a and therefore merely acts as a scaling factor. Estimating the value of G M in the orbit determination processwill therefore yield a value in which the dominant relativistic orbit effect is absorbed,cf. (Tapley, 1989)or (Schrama,1992). One of the most remarkable differencesbetween the classicalNewtonian gravitational theory and Einstein's relativistic theory of gravitation is the fact that in relativity not only a static massdistribution causesa gravitational field (generatedby a scalar potential function) , but also the motion of a mass causesa gravitational field (generatedby a vector potential function). In analogy with electromagnetic theory the former gravitational field is said to be of electric type (sometimescalled gravitoelectric field), whereasthe latter is of magnetic type (gravitomagnetic field). The vector potential for moving masses (for example rotating objects) appears in the metric in the ps; components,the ones describingthe interaction betweentime and space. In the lPN metric it was included as the vector potential function % (or V in index free notation). The equations of motion 5.35 in their general form (with V; included) allow us an easy check of the relativistic contribution to the satellite
161
Figure 5.2
orbit for ttteatonstrack retatiaistic efiect of after rernoual a lineat rern'aining satellite, the .tlristotele:; for trcn.LI.
p":,":;,'*",,',i,;';;,"
motion caused by moving masses.This gravitomagnetic effect on satellite orbits is also called the LenseThirring effect,first describedby Lense and Thirring in 1918 (Thirring and Lense, 1918). From eq. 5.35 we see that, leaving the vector potential V in, an additional relativistic term 4 x (V x V) "v enters the disturbing acceleration.Let us evaluatethis term for our sphericalsymmetric mass,which we now assumeto rotate around the zaxis with angular velocity <r". Comparing eq. 5.30 with the 1PN metric 5.33 (with 'y : 1) we find for V:
n::ry
or in vector notation
(:)' w"ei;sri
v
Y : ;
rGM /n\2 u  [  l
D r \r,/
162
With this expression for V the LenseThirring part in the relativistic acceleration becomes
I i@' + v'  2r')t3vz21 " 4GM 4)' rf :% '   'i\@ , '*r y z 2 " ) ) t , " i I + Y 'z r  o ' 5;
\;i 
s"@i vn)
cf. (Soffel,1989). The effectof this term on the orbit of the earth orbiting Aristoteles satellite is shown in figure 5.3. Ib is compubedin the same manner as figure 5.1 but now for the LenseThirring terrn only.
cross track (cm) along track (rnm) radial (cm)
t 6
o
k
time (h)
Figure 5.3
'lhe
jilitl;i::ii,i!;!,'i{,':i:,,:':;:!"'i'1" {:'x::::",'
The figure mainly shows a once per revolution periodic effect due to this term, being largest in the cro'sstrack direction with a linearly growing envelope. In the along track direction the effect is smaller showing a secular trend. The radial effect remains below the level of 10a m.The crosstrack effect is causedby a drift of the right ascensionof the ascendingnode O at a rate of approximately 3.8 . t0tl /s. Such a drift rate is well within the observablelimits of modern tracking systems (Schrama,1992).
5.5
model can be As mentionedin the beginningof this chapter,the gradiometric interpretedin relativisticterms by considering socalledequationof geod,esic the
163
5. Relativisticview on gradiometry deuiation. This equation describesthe relative motion of two neighbouring particles moving on geodesics.We will derive and discussit in this section. Consider both affinely parametrized, e.g. with the Proper to this extent the two geodesics, geodesics close,and the particles move along them close together, are time r. The such that, at any value of the parameter r the coordinates of the first particle are rp(r) and of the second ru(r) * '("), where p is a small vector of coordinate differences. We will assume these coordinate differencesto be infinitesimal, as well as their derivatives d(P/dr, which means that not only are the particles close together, but they remain so for a long time. Higherorder terms in, or products of these quantities will therefore be neglectedin the derivations. Note that, in general, the coordinates rP do not constitute a vector (rank 1 tensor). A tensor Ap should transform to another coordinate system rp' with the transformation matrix transform via the coordinate Xur: 6rt'' 1Arr,i.e. AP' : Xp;H". But coordinates  ,u'7ru1 from which the transformation matrix is transformation equations rP' , derived by means of partial differentiation. However, if differencesbetween coordinates are small (like we assumefor the {p) they may be considered the elements of a vector (Foster and Nightingale, 1979),such that they transform as qu' : yy' g'' The equationsof motion for the two particles become:
 r [  l , ' r 'vo\*t ln
dzr\
drz
dx'dxo
4, dr
rfl(r+)xrft(r)+rft,'@)(
so that, if we take the difference between the two geodesic equations, we obtain (neglecting terms quadratic in (r' 3n4 ir')'
(5.37)
where the dot denotes differentiation with respect to r. The first absolute derivative (seeappendix B.2.3, eq. B.10) of gr' 1. Dt' dr : dP dr +lu Fod" ' 'vo\ dr
D2 u
drz
164
v,r,ip) "i"iP
where we made use of the geodesic equation and relabelleddummy indiceswhenever convenient. Inserting this result in eq. 5.37, yields
(5.38)
with Rp,,1othe RiemannChristoffel curvature tensor. This is the equation of geodesic deviation, sometimescalled Jacobi equation.It shows how two neighbouring geodesics deviate relative to eachother. It is interesting,though not surprisingly, that the Riemann tensor enters this equation. The Riemann tensor describesthe curvature of the space under consideration. If a global (i.e. covering the whole of space) coordinate system can be introduced, in which all componentsof the Riemann tensor vanish, the spaceis flat. In a flat space,in cartesian coordinates,the absolute derivative D{P/dr reduces to the ordinary total derivative d{pf dr. So, with Rqouo: 0, the equation of geodesic deviation would reduce to dzt,
77:o
of which the solution for (e It '("):cqraPu where Cp and DP are constants. We seethat in this case,the separation between the geodesics grows linearly, as it should, since in a flat spacewe know that geodesics are straight lines. In a curved space,some componentsof the Riemann tensor will always be different from zero, whatever coordinatesystemwe choose.The separationvector between the two geodesics will not changelinearly as function of r, but accelerated (the quantity D2Eufdrz is the "relative acceleration" vector), depending on the curvature of the space. So the curvature of a space determines the relative accelerationbetween two free particles. On the other hand, measurementof this relative acceleration,tells us about the curvature of the space. Since in satellite gradiometry we measure such relative accelerations,this technique should give us the curvature of spacetime, which is curved due to the earth's gravitational field. If this is so, then the equation of geodesic deviation should reduce to the basic gradiometric model equation 2.1in the Newtonian limit. This will be shown in the sequel.
165
5. Relativistic view on gradiometry neededto arrive at Newton. The Newtonian First we will repeat the assumptions the linearized (weak field) approximation. In this approximalimit is derived from tion we take for the metrical tensor gp, the expression\py * hpv, with small hp, and small derivatives hu,,o. The latter also implies that the Christoffel symbols Tf,, are small. Another assumptionwe made in order to arrive at Newton was u << c, so the Newtonian situation of universal that 7 = 1, thus dldr x dldt. This expresses time t, i.e. all clocks read the same time. Furthermore, derivativeswith respect to time are assumedsmall in comparison to spatial derivatives. [n section 5.2.1 we with the Newtoequation 5.4, under the aboveassumptions, compared the geodesic equationsof motion, in order to show that the former will reduce to the latter. nian This appearedto be so if the only nonvanishing Christoffel symbols are
lob: 16t'noo,r'
( 5.3e)
Sinceall clocksmeasurethe sameuniversaltime t, so will the two clocksmoving along so with the two particles on the neighbouring geodesics, that the first component will be zero, i.". o : 0. This simply means that the of the separation vector p particle's accelerationsare compared at equal times (Ohanian, 1976). With this result, and with eq. 5.39, the equation for the absolutederivative of the separation vector will become: Dt'
d'
In the same way, we find for the secondabsolutederivative D2 pf dr2 : 6zEuf drz . The latter is equal b d2pfd* since,y:1. Thus, in the Newtonianlimit, the left deviation 5.38 reducesto d2{'fdt2. In order hand side of the equation of geodesic for find the Newtonian limit of the right hand side,we look at the expression the to are neglected, RiemannChristoffeltensor B.16. Products of the Christoffelsymbols and we are left wittr : RPooF *B,o  t!",8 ' With eq. 5.39, the only nonvanishing components of this tensor are r?oro : Btooo : Iiio,r = deviation becomes l6ii hss,ip.For P: O the equation of geodesic
fi:o
which indeed agrees with (0 : O. The spatial components yield
d2fo
## ###:R'o*o* R'oo*o
166
5.5.
Equation
of geodesic deviation
: :
(5.40)
: o ; ia2trf i ; k . f
In the latter equation we inserted the hss from eq. 5.18. Comparing the above equation with 2.1, we see that, indeed, the equation of geodesicdeviation is the relativistic generalization of the gradiometric equation, at least if we take (l as the coordinate differencesbetween the two proof massesand d2i I dtz as their acceleration difference. In relativistic terms, a gradiometer measuresthe curvature of the space,cf. (Misner et al., 1973),(Ohanian,1976)or (Gill et al., 1992). This curvature is completelydeterminedby the elementsof the RiemannChristoffel tensor. Therefore, we will give those elements (or rather those of the covariant form of the curvature tensor, Rr,.og)for the 1PN metric 5.33. Inserting the Christoffel symbolsfrom eq. 5.34 into the expression the Riemann tensor B.16, we obtain for for the nonvanishing elementsof Rroog:
167
5.
Relativistic
view on gradiontetry
worldline with the tangent vectorsto the to coincidein every point of the observer's coordinate lines of some global coordinate grid. Such a frame is sometimes called a coordinate induced frame (Soffel,1989). In this casewe have, denoting the base vectors by e1a) (sometimese6) e(o) :
a
6ts"
(5.42)
the index (o) indicates which base vector it is and the inIn the notation e1a1u dex pt indicates the component of this base vector with respect to the coordinate system xp. In the sequel,the "indices between parentheses"(which in fact are labels to distinguish the base vectors) are treated as normal indices, which can be . raised or lowered and summed over, cf. (Synge, 1960) This means we may define "contravariant" base vectors in the following way:
" (" )
 n@o) "( O .
Note that for the componentsof these basevectorswith respectto some coordinate system the usual relations for raising and lowering of indices applies,i.e.
e ( o ) p : g u r e ( o 1. If the observer moves along a geodesic (i.e. he is not accelerated due to some external force or rotation) this proper reference frame is a local Lorentz frame, in which
g@o: n@p)
and
l!"/l
: o
(5.43)
the basevectorsof such orthonormal tetrad form a parallel vector Along a geodesic field, i.e.
':'tq
au
: o'
(5.44)
any Constructing an orthonormal tetrad in the way describedabove,we may resolve vector or tensor into componentsalong this tetrad, e.g.
,(")  rr"!i)vre@)u u(o): uPe61u: vuelol  t*e@)ur(A), "(") "(9) ep1ue61,: trr"(q"(B) t(oB): tPY 1@0) y"
168
and the inverse relations : up : ,u@)e(o)u u@) e7o, 0t, : u@)ef)  u@) rroru
"+:++rfie('vff:o
*:_r,*"Gt#
(5.45) vector (p betweenthe two neighbouringgeodesics For the infinitesimal displacement from eq. 5.38 we have Dt' du : +lt" f,4x" ' vo> du du du e(,, this yields and if we write qr : q@)
o#:
D du
(#):#'("r.,#)
169
,t.r#
t Ru,opoi'iP:0.
"lf)"l,t*du2
azr@) #
If we now insert for the vectors i' , ic and (" their transformations to the invariant components relative to the orthonormal tetrad (respectivelyi@)"(il, i(d)r[o) u.d (")t["]) we obtain s2c@)
R(o u) : Rp,ope(ol"(n"?,l"bl . B^, As a final step we define K(".,)= RpB.,6p@) i(6) equation becomes, (Synge, cf. 1960): with which the invariantdeviation
t*
du2
:o #+KpB16(e)
of which the spatial part is
:o #*K6i16(i)
170
(5.47)
which closely resembles the Newtonian form of the gradiometric equation. The quantity K1;i1 takes the place of the gradient tensor and is therefore sometimes referred to as tidal matrix. Its components are found as K(ri) : which approximately R6p161iB i6
(5.48)
The dominant relativistic contributions to the gradients are again expected to result from a static spherical symmetric mass distribution. For such a mass distribu: polar coordinates to tion it is convenient usethe 4dimensional {ru'ltt' : 0, 1, 2,3} the non(ct,r,0,,\) with the primes dropped in the sequel. In these coordinates, vanishing componentsof the Riemann tensor, computed from eq. 5.41 with c's restored, are (up to lPN order in c2)
(5.4e)
rc'
.1
Rzszs:2rsinz09+
where we have put p : 1. Finally we need,in order to evaluateeq. 5.48, expressions for the components of the base vectors of the orthonormal tetrad. The tangent vectors to the coordinate lines of the coordinate grid (ct,r,0,,\) are respectively 0f c)t,0f 0r, 0100 and AIA^. In order to fulfil eq. 5.42 (so that the base vectors are orthonormal) we have (up to first order in U f cz)
e(0) : e(r) : e(2) : e(3) :
GM\
;[' *)u
I
t 1 _ _ t _
GM\A
rcz / 0),
rsind \
171
This system of base vectors can be constructed in every point of the observer's worldline. But each of those frames is at rest with respect to the spatial part of spacetime. The proper reference frame of an observer comoving with a satellite is moving through space, and if it is earth pointing, it also undergoesa spatial rotation. Without any loss of generality (as a result of the spherical symmetry of the mass) we will, for convenience, consider here an equatorial satellite orbit, i.e. = rl2. The motion of the proper referenceframe then takes place in the e1s; 0 direction, and we Eursume to be circular. In order to obtain the comoving tetrad it e(d), we apply a Lorentz transformation to the e(a) frame where we take for the velocity the satellite's mean orbital motion ns. Thus u(i)  nork$) where k(d) is a spatial unit vector pointing in the along track direction, which in our caseis the e1s1 direction,so {k(i)}: (0,0,1). Such a transformationis also called a Lorentz boost in elsydirection. So (nl e(d,)= 41")i'e1B; where ,f1"!f) ir the Lorentz transformation matrix (e.g. (Misner et al., 1973) or given (up to 1PN order) by (SoffeI,1989))
l +'+ t9 c 4 t
o o 1 0 0 1
tLrr
c
{^("!f)}:
0 0
n(\f
0 0
L rc'
0 o r++44
0 no0
With this matrix the base vectors of the comoving frame e1a,;become
e ( 0 , ): e ( 1 , ):
ei/9r\ :
(/
I /
(t*t;F),at+;a^
GM\ A
G M \ A
3cM\
\t;t)u
; l . t * ) m
  :   r  1 1 _ _  l 
: e(s,)
nor0 l/ 'r\cc)t
lGM\ A 2rc2/0),
Now we are in a position to compute the components of the tidal matrix Klii; (omitting the primes). We obtain from eq. 5.48:
:29+t'1#' Klrr)
ry Kpz):
: ry+u%Klss)
172
(5.50)
of which the terms with c2 are obviously the first relativistic contributions to the gradients. The first terms on the right hand side coincide with the Newtonian expressionsfor the gradients of a spherical symmetric mass distribution. For example, sincethe s(o)=r componentpoints in the radial direction, Klrr; is equivalenttoV* in the local orbital (cartesian)coordinatesystemwhich we used in the nonrelativistic approach. In the sameway Klss; coincides with V* and K1zz1 with Vrr. Estimating the sizeof the relativistic contributions,we seethat they are of the order of 3.10o E for a satelliteat 200 km altitude, cf. (Soffelet al., 1987)or (Gill et al., 1992). For the present Aristoteles mission scenario (ZOO km, O.Ol El\/Hz measurementprecision) such small contribution can be neglected.Other relativistic effectson the gradients due to e.g. the gravitomagnetic field or higherorder multipole expansionsare not discussed here. The reader is referred to (Thei8, 1984) and (Gill et al., 1992).
173
Conclusions
By means of a technique called diferential accelerometryit is possible to measure the secondorderpartial derivativesof the earth's gravitational potential. An instrument which measuresthese socalled gravity gradientsis called a gradiometer. If a gradiometer is on board an earth orbiting satellite, a global set of gravity gradients can be obtained in a few months time. Flom this set of observationsthe can be derived by gravitational potential (represented a set of potential coefficients) squaresadjustment, with which at the same time the by means of an iterative least satellite's trajectory can be estimated. The linear model used for the adjustment is usually obtained from partial differentiation of a seriesexpansion of the gravitational potential. If, in the future, more accurategradiometersand/or satellite tracking to systemsbecome available,it might be necessary include relativistic effectsin the model both for the satellite orbit and the gradients. It is customary, and for our purposes also appropriate, to expand the gravitational potential function, using geocentricpolar coordinates,^  (r,0, )), into a series of sphericalharmonic functions. In such a seriesthe Legendre functions appear. For satellite applications,however,it is more suitable to use a coordinatesystem related to the satellite orbit. In that case,the gravitational potential is written as function of orbital coordinatesro  (r,wo,w") or 7@': (r,wo, 1) . Due to the rotation of an equatorial coordinate system to an orbital system, socalled inclination functions appear in the series. Expressionsof the first and secondorder derivatives of the potential can be easily derived by differentiating the series expansion with respect to the coordinates. In the case of satellite gradiometry, expressionsfor the secondorder derivatives are neededwith respect to a local cartesian coordinate system. Therefore,the partial derivativesof the potential seriesexpansionhave to be transformed from any of the geocentriccurvilinear coordinate systemsrArto or ro' to this local cartesian system. With the help of inder notation and tensor analgsisa compact, general algorithm can be formulated for the transformation of potential derivativesbetween arbitrary coordinatesystems. for The transformation equations,in which seriesexpressions the potential deriequationsin a gradiometric analysis vatives are inserted, then serve as observation procedure. By meansof a least squaresadjustmentapproach,the unknown parameters (in our case the potential coefficients)can be estimated. An advantage of such least squaresestimation is that an error analysiscan be carried out prior to an actual
174
ConcJusions
experiment. The inverseof the normal matrix providesthe formal variancesand covariancesof the unknown parameters.Thus, given some mission scenario(in terms of satellite altitude, orbital inclination, missionduration, sampling interval and measured tensor components) and some apriori error model for the observations,the precisionof the estimatedpotential coefficients can be computed and comparedwith given such requirements, the error analysiscan be used the requirements.Revqrsely, to put demands on the mission scenario as to meet these requirements. The requirements for a gradiometer mission typically are lessthan 100 km spatial resolution (halfwavelength) and a precision of lessthan 10 cm global r.m.s. for geoid heights or 5 mgal for gravity anomalies. The mentioned resolution corresponds to a spherical harmonic expansion up to at least 180 or more (somewherebetween 200 or 300). From the error analysisappearedthat, for most mission scenarios, signal to a noiseratio of 1 was reachednear degree240, which will be sufficientfor the required resolution. If globalr.n1.s.valuesfor geoidheightsor gravity anomalies have to be computed from the aposteriori error estimates of the potential coefficients(as they result from the error analysis),we have to bear in mind that in reality the gravitational potential is a continuous function and should be representedby an infinite series (including all frequenciesfrom zero to infinity). The error analysis, however, is limited to some maximum degree and order, in our case 240. The error resulting from this part is called comrnissionerror. The neglectedpart of the spectrum, above degree 240, also contributes to the global mean error, and should in fact be taken into account. This neglectedpart, is called omission error. Since we do not know the true gravitational spectrum, the omission part can only be approximated by using some prior knowledgeof the averagebehaviourof the gravity field. The global error r.m.s. values presented in this work always represent the sum of commission and omission error. The omission part is taken into account up to a maximum degree of 1000. Above this degree no substantial contribution was found. Furthermore, we always show srzoothed global error values, representing 1" x 1o block averages. Due to the mission altitude and the sampling distance it is difficult to estimate higher degrees that smoothing is justified for our purposes.Dependingon further so applications, unsmoothed values (representingpoint values) can be used too. A spherical harmonic expansion up to a maximum degree and order L of. 24O contains more than 58,0m potential coefficients.If these are to be estimated in a least squaresadjustment, the normal matrix will have a sizeof (58,000)2. Even with large computers, the solution of a system of linear equations of this size will be a and by ordering the unknowns in a specific major task. Under certain assumptions, manner, the normal matrix attains a blockdiagonal strtcture. The largest block to be inverted will have a size of (24012+ 1)', so that computation time is drastically decreased. The assumptions to arrive at the blockdiagonal stnrcture are: 1.) circular orbit; 2.) regularly distributed data along the orbit (i.e. no data gaps); and 3.) the number of orbital revolutions and the number of nodal days contained in one
175
Conciusions
repeat period are relative prime integers, where the number of revolutions should be larger than 2L. The second of these assumptions seemsto be the most critical one. Doto gaps are likely to occur, not only due to instrumental failures, but also due to orbit maintenance maneuvers and as a result of excessivedrag variations near the poles and the equator. A noncontinuous data stream destroys the orthogonality properties on which the blockdiagonal structure is based. In an ideal situation, an SGG mission will be flown with a full tensor gradiometer, all measuringat a sampling rate of 4 seconds nine gradientswith a 0.01 E//Hz white The satellite will fly in a circular, polar orbit at e.g. 200 km noise error spectrum. altitude for at least 6 months. Global error r.m.s. valuesderived from such an ideal mission will be approximately 8.6 cm for geoid undulations and 3.6 mgal for gravity anomalies at a resolution of 0."75 halfwavelength. These results easily fulfil the requirementsgiven above, even without having additional information, either from a GPS receiver on board the satellite or in terms of apriori information in a kind of least squares collocation setup. However, a more realistic mission like Aristoteles differs from the ideal situation are: 1.) band limitation on severalaspects. The three most important differences error spectrum; 2.) nonpolar orbit; and 3.) planar and colourednoisemeasurement gradiometer measuring only the out of plane tensor componentswith high enough precision. All three limitations have a rather large effect on the results. Apart from numerical singularities (resulting from illconditioned normal matrix subblocks) other normal matrix subblocks will be singular too as a result of the limitations. The occurrence of such singularities shows that certain unknown parameters are not estimable from the observations any more. The planar gradiometeris a penalty resulting from the non drag free concept of due to air drag are too large Aristoteles. The along track disturbanceaccelerations axes in this direction. to allow for very sensitiveaccelerometer Aristoteles is planned to fly for approximatelysix months in an orbit with inclination 95."3. During additional two weeksthe orbit will have an inclination of 92."3. In both cases,polar gaps occur) i.e. polar areas which are not coveredby ground tracks of the satellite. If a global recovery of potential coefficientsis intended, problems might occur due to these gaps. Small deviations from a 90o inclination do not deteriorate the results significantly. Remarkably, the results slightly improve for a 92."3 inclination. It is very likely that this phenomenonresults from a higher data density in the remaining part of the earth. Furthermore, the behaviour of the inclination functions may play a role, as well as the fact that with a polar orbit, zonal coefficients are not estimable from gradients involving one crosstrack derivative. For a 95.o3inclination large polar gaps occur and the problem to derive a complete global set of potential coefficients from the data becomes improperly posed. The by solution can be stabilized adding prior information (e.g. prior expectationsof the potential coefficientsand their covariancematrix) and solve the system in the sense of least squares collocation. The solution becomes stable, but the estimates will be biased. However, investigations showed that, for large parts of the spectrum, a
176
Conclusions
stabilized solution does not lead to too optimistic error estimates. More than the polar gaps, the band limitation of the Aristoteles gradiometer will be a problem. Only in a frequency band between 0.005 Hz and 0.125 Hz a 0.01 Dl\/Hz error level can be achieved. The upper limit (resulting from the sampling rate of 4 seconds)will not cause any problems if the maximum degree and order of the analysis is 240. However, a lower limit (resulting from instrumental instability and nongravitational orbital effects), has influence on all degreesin the solution. Whereas a frequency of 0.005 Hz would correspond to approximately 27 cycles per revolution (".p.t.), we took an absolute lower band limit of 4 c.p.r. (to account for nongravitational. orbital effects). Furthermore, we investigated, apart from a flat error spectrum above 4 c.p.r., a l/B coloured noise enor behaviour between 4 and 27 c.p.r. With a band limited gradiometer, certain potential coefficientsare no longer estimable from the measurementsdue to a lack of information, resulting in singular normal matrix subblocks. Especially the lower part of the spectrum, where the contribution of the low orbital frequenciesis relatively large, will cause problems. The requirements for an SGG mission are no longer met. Additional information is necessary. Given the present Aristotelesgradiometry specifications, a combined SGG/SST mission seemsto be the best alternative. For an SGGonly mission one has to use a gradiometerwith either a much better precision (e.g.0.0001 E l\/Hr) or one which measuresmore components of the gradient tensor. The results of the error analysis are presentedin terms of error degreeuariances. The degree variance is the sum over all orders of a degree of the individual error estimates per coefficient. In some of the above situations, however,we do not obtain error estimates for all coefficients, viz. those cases where singular normal matrix subblocks appear. The degree variances computed in such casesthus do not represent the whole spectrum. Adding prior information in the sense of a stabilized solution overcomesthis problem, but it is difficult to interpret the resulting "mixed" degree variances. In fact, the problem of singularities on certain degreesor orders is inherent to satellite methods. Degree variances are not the appropriate representation of the error situation in such cases. "Order variances" might be an alternative. Perhaps it gives more insight to plot all individual error estimates per coefficient in a perspectively plotted l, rnscheme, and find ways to identify the contribution of the prior information. Whereas an error analysis learns us much about the influence of mission scenarios on the precision of the estimated unknowns, the final goal of a SGG mission is a set of estimated potential coefficients,together with the precision of all coefficients. The recooer! of potential coefficientscan be done in an iterative least squaresadjustment scheme, which at the sametime the satellite'strajectory is determined. Depending in on the way in which the data is handled there are, in the present setup, two possible methods for the recovery. In the spacelikemethod one transforms the data along the orbit into a global equiangular grid of mean values. The observation equations are based on a seriesexpansion of the gravitational potential in geocentric
177
Conclusions polar coordinates. Equal step size in .\direction and orthogonality of trigonometric series give rise to the blockdiagonal structure of the normal matrix. The number of "observations" in the least squares estimation process is limited (it equals the number of blocks in the grid) so computation time is relatively short. The averaging process,required to obtain the gridded data, on the one hand acts as a smoothing are operator so that high frequencies lost, but on the other hand aliasingis reduced. of A disadvantage this method is the impossibility to include band limitation of the gradiometer in an easy manner. In the timetike method the data does not need to be transformed into an equiangular grid. A series expansion of the potential in orbital coordinates is used and with equal data step size along the orbit a blockdiagonal normal matrix is as are so obtained. Data is not averaged, higher degrees preserved, a result of which aliasingwill play a larger role. Due to the very large number of observations however the computation time will be rather large. Since possibledata gaps may destroy orthogonality they have to be accountedfor. With this method it is easy to include the effectof band limitation of the gradiometer.A test was done with both methods. We used a set of simulatedgradients,computed along a circular, polar orbit for a 32 days mission. Only a first step of the iteration processwas implemented, without were estimatedup to degreeand order 180 updating the orbit. Potential coefficients (approx. 33,000 unknowns). Although the.test should be consideredpreliminary, the results look promising. The requirementsfor an SGG mission are very high. Already a mission like Aristoteles puts high demands on the technologyof the instrument and the spacewith ever improving technologicaldevelopmentsgradiometers craft. Nevertheless, with much higher precision may become available,e.g. superconductinggradiometers. Also satellite orbit determination techniquesare still improving. In order to to it fully benefit from such future perspectives shall be necessary improve the mathematical models as to include relativistic effects. Even better, gradiometry might become a technique such accurate, that aspectsof the theory of relativity can be tested. Finally, from a theoretical (geometrical)point of view, gradiometry perfectly lends itself to be formulated in terms of curvature of spacetime. We mention three aspectsof an SGG mission where relatiuistic theory may play a role: f.) the satellite orbit; 2.) the observedgradient tensor; and 3.) the moving local referenceframe. The third one is important for the orientation of the are taken, but this aspect instrumental frame with respect to which measurements is not treated here. In the generaltheory of relativity a satellite can be considered a freely falling particle, moving along a four dimensionalspacetimegeodesic.Gravitation is not consideredan external force, but is translated into the geometry of the space. Due to gravitation, four dimensionalspacetimeis curuedand so are the thereforeconstitute the equationsof motion geodesics. The equation of the geodesic of the satellite. Extracting the spatial part of theseequationsrevealsthe relativistic contribution to the satellite motion. Viewing upon gradiometry as the relative movementof two or more proof masses,
178
ConcJusions
we find the relativistic correspondence the socalled equation of geodesicdeuioin tion. This equation is governedby the Riemann tensor, in which the secondorder derivatives of the gravitational potential appear. Thus relativistic effects are also present in the observed gradients, and in this sense, a gradiometer measures the curvature of four dimensionalspacetime. In relativistic terms, the gravitational field of the earth can be considereda weak field. The equationsof the satellite motion and the gradients are therefore derived in the socalled weakfield approrimation. ln particular, we use the socalled first PostNewtonian (1PN) approximation of the spacetime metric, keeping only the firstorder relativistic contributions. The dominant relativistic orbit contribution is an effect in along track direction, which, however,in reality cannot be separated from a change in the value of the gravitational pararneter GM. For a spherical, nonrotating earth and for a satellite like Aristoteles,other relativistic orbit effects are, at the moment, small enough to be neglected. Effects due to the rotation of the earth are even smaller. Furthermore, in the case of a spherical, nonrotating km are of the order of 106 earth, relativistic contributions to the gradientsat 2OO E. Although this is much smaller than the Aristoteles measurementprecision, the questionwhether the relativistic effectsin the gradientscan be determinedfrom such precision satellite missionsis a difficult one, dependingnot only on the measurement but also on the mission duration and the principle character of the instrument.
179
appendix A
Coordinate systems
Severalcoordinate systemsare used in this thesis,curvilinear as well as rectilinear, with different orientations and located with the origin in different points. Furthermore, the derivativesof the gravitational potential function with respectto most of these coordinate systems are needed at different stagesduring the derivations' The choicefor a particular coordinate system often dependson the specificapplication, either becausethe coordinate system is particularly well suitable for the geometry of the problem, or becausethe practical or instrumental implementation of the problem prescribesthe use of a specificcoordinate system. In principle one could choose any coordinate system one likes, in practice only a few coordinate systems will have favourable characteristics for that specific problem. On the other hand, the introduction of a coordinate system is only artificial, it has in fact nothing to do with the physical reality of the problem itself. . In this appendix we introduce some special, much used, coordinate systems' Some of them are applied in this thesis, others are just given for illustration. Referin ring to section 8.3.1, we will only considermetrical spaces, most casesalso linear, En, in particular for n : 3' i.e. Euclidean spaces In the first section the coordinate systems will be defined, in most casesrelative to a cartesian geocentriccoordinate system, by giving the coordinate transformation equations between the geocentric cartesian and the new coordinates. Each coordinate system will be addressedby a special set of indices which indicate the particular system. Note that in appendix B the choice of indices was arbitrary, since at that stage no connection to some particular coordinate systemswas given. In the second section the metric tensor and the Christoffel symbols for some of these coordinate systemswill be listed. They are neededwhen deriving expressions for the transformation of potential derivativesby using tensor analysis, as is done in section3.1.
180
4.1.
Definition
A.1
Definition
Possibly the most familiar coordinate system is the cartesian coordinate system. This is a rectilinear coordinate system of which the coordinate axes intersect in one point (the origin), are mutually orthogonal and along which the scaleis the same in all directions. If the origin of such a cartesiancoordinatesystem is in the geocenter, the coordinatesare denotedwith cr, the indicescbming from the set {1, J,K,L}. This system is called the geocentric cartesiancoordinate sgstem.As we will do for all coordinatesystems,the separatecoordinatesare given specialnames(kernelletters), which, in this case,are {rIlI : 1,2,8}: (rlt,rI=2,r1=3) = (x,y, z) . The Xaxis points to the Greenwich meridian, the Zaxis to the NorthPole and the Yaxis completesthe set to a right handed coordinate system, seefigure A.1.
Figure A.1
In cartesian coordinates, the metrical tensor takes on the diagonal form with all
1E1
{crt}
:l
0 1 sinl
1 0 0 0 (A.1)
0 0 1
system(with In this casewe also have {gIJ} : {Stil.Also in everyother cartesian a different origin and/or a different orientation) the metrical tensor has this form. For example, if we rotate the rr system in such a way that the XYplane coincides with the orbital plane of a satellite in orbit around the Earth, we obtain a new geocentric cartesiancoordinatesystem,denotedby {rt' l1': 1,2,3): (X',Y',2'), which has the same origin but a different orientation, see figure A.1. The X'axis of this system is directed towards the ascendingnode. This system is called the system. The coordinatetransformation betweenthe two orbital cartesiancoordinate : ,I(rI' ), is as follows: systems,,I
x
Y :
 .or r,
sin 4.,,sin /
fx'
I L
rotations about the angles u.r,and .I. The which is obtained through successive matrix in this equation is the transformation matrix ffi. A third cartesian coordinate system we will frequently use is a local cartesian system, denoted {ril; : 1,2,3) = (r,y, z), which has its origin in a point on a satellite'sorbit and which is oriented with the zaxis radially outwards, the caxis directed along track and the yaxis crosstrack such that it is a righthanded system (seealso figure A.1). This system is called the local orbital coordinatesystern.The cartesiansystemand the local orbital system, transformation betweenthe geocentric rotations about the anglesw",I and ,I : ,I (ri), can be obtained through successive aroand a translation in radial direction by r. This will yield:
1",
(A.2)
x
Y
 cosoc sin arocos cos sin c,.r, a,ro f  sin c,.r, oro* sin (re cos cOS cosc.ro 1 cosar,sin 1
 cosu,r, 1 sin a.r, u,ro* cos sin cos &.re cOS sin c,.ro 1 cos1 sin oo sin 1
v
z*r
(A.3)
Another cartesiancoordinate system to be used is a local northoriented cartes i a n s y s t e m ,d e n o t e db y { " t ' l f ' : 1 , 2 , 3 } : ( r ' , y ' , 2 ' ) , w h i c h h a s i t s o r i g i n i n s o m e terrestrial or spacepoint and is oriented with the z'axis radially outwards, the ctaxis directed north and the y'axis directed west. This system is called local northoriented.coordinatesystem. The transformation ,I : ,I (ri') is obtained through
1E2
A.1. Definition
successlve rotations about the angles d and ,\ and a translation in radial direction (see figure A . 2 ) :
X Y
sin ) cos) 0
tr'
v'
z'+r
(A.4)
Figure A.2
Whereas in the transformation equations above the quantities Irw",worrr0, \ act as parameters to fix the orientation of the ,I' ,si' and cd coordinate systems with respect to the r/ system, they themselves can also be used as coordinates. From the mentioned quantities we select the following coordinate sets:
{"olA: {r"la:
g e o c e n t r ip o l a r c o o r d i n a t e s c Iorbital coordinates a r r  o r b i t ac o o r d i n a t e s l .
{r"'lo' :1,2,3}:
183
X:rsindcos) Y:rsin0sin) Z:rcosd. The c" and co' systems both have the same transformation equations: X : Y : Z : r(cos (te coso)o  sin or, sin alo cos 1) r(sin u)ecos"io f cosrr.r, sin<r.ro 1) cos r sin o,ro 1 sin
(A.5)
(A.6)
where in the caseof the ro system<,.r, the third coordinate (r"=3) and / acts as a is parameter fixing the orientation of cd and vice versa for the ro' system. Note that equation A.5 can be obtained from equation A.4 if we put r' : y' : zt : 0 and that equation A.6 can be obtained from equation A.3 by putting r : y : z : 0. Note furthermore that the ra system is not an orthogonal coordinate system so we expect some offdiagonal componentsof the metrical tensor in these coordinatesto be unequal to zero. All of the above coordinate systems,except one, can be used to label all points of the 3dimensional spaceunder consideration,regardless the arbitrary values of of the possibleparameters. The exception is the ca system. This system has the inclination I as parameter. If I + 90" parts of the space are not "covered" by the coordinate system. Theseparts are coneswith the top in the geocenterand with the Zaxis as symmetry axis and with the top anglewith respectto the Zaxis equal to .f. However,this system will only be used to describepoints along a satellite orbit with inclination .I and for that purpose the system can be used very well. A fourth curvilinear coordinate system which will be used is also a polar coordinate system but not relative to the geocentriccartesian coordinate system cr (like ra), but relative to the orbital cartesiansystem ,I'. It is called orbital polar (r,6,ro). We havethe s a c o o r d i n a t e y s t e m n d i s d e n o t e db y { " o ' l A ' : 1 , 2 , 3 } : following relations (seefigure A.1):
X' : r cos/cosaro Y' : r cos/sinc.ro Z':rsin/.
(A.7)
The transformation rI : ,I (rA') .un be found by inserting equation A.7 into equation A.2. The final two coordinate systems to be introduced are two ellipsoidal systems (seefigure A.3): {"AIA: 1,2,3} : (h,9,\) geodetic oordinates c coordinates ellipsoidal ,
184
Y : u cosBsin) Z : usin9
rI : rl(ru) :
(A.8)
(A.e)
with u : \/u\ (o, b2)fa2, a E i 2 , E 2: is the semimajor axis of the ellipsoid and 6 the semiminor axis. For problems expressedin ellipsoidal coordinates it is often convenient to have a local cartesian coordinatesystemdirected alongthe normal to the ellipsoidalsurface. This cartesian system is called local ellipsoidal coordinatesystem and it is denoted by {rtll : (i,y,z). The Zaxis is directed outwards, normal to the ellipsoid,the 1,2,3\: faxis directed north, tangent to the ellipsoidalsurface and the graxis is directed west (seefigure A.3). We will not actually use these ellipsoidalcoordinate systems in this work, but they are given here just for reference. In table A.1 all mentioned coordinate systemsare listed.
A.2
Metric
1
{grt} : {grt,) : {g;i} : {e;,i,}:
0
I
0
The Christoffel symbols are therefore zero in all these coordinates. This is not true in curvilinear coordinates. From equations A.5  A.9 we may compute the transformationmatrices 6rtd,etc. Thesecan be usedin equation B.13 to compute the metrical tensor in the systemsrA,xorro'rrA',rA and sa from the one given above. For each of the coordinate systems the Christoffel symbols are then computed using equation B.15. In the sequelall the metrical tensors (table A.2) and the Christoffel symbols (tables A.3 to A.8) are listed.
185
rl
rA ,I'
,A' ,i' tr' ro xo' rA rd xii
(x,Y, z)
(r,0,\) (xt ,Y' ,z')
(r,6,ro) (r',yt,"') (',Y, ") (rrworu") (r,wor I) (h,p, \) (", F, \) (n,!,2)
geocentriccartesiancoordinates geocentricpolar coordinates orbital cartesiancoordinates orbital polar coordinates Iocal northoriented coordinates local orbital coordinates ^Iorbit al coordinates orbital coordinates c,.r, geodeticcoordinates ellipsoidalcoordinates local ellipsoidalcoordinates
186
Table A.2
geocentric polar ( r , 0 ,\ )
1 0
0
r
0
0 r2 sinz 0
t l
0
orbital polar ( r ,d , r o )
I
0 0
, r'
0 0
0
rz cos26 0 rz cosI
 r 2 ( s i n 2 l s i n 2 a r o 1 )
0
n
.Iorbital
(rrworw")
0 0
rr2 cos I
t.rrorbital ( r r w o rl )
n
r'
0 0
12 sinz wo
U
0 0
0 0
geodetic (h,p,\)
0 0
( n +u ) 2
0 0
0 (n + lr)2 cosz 9 0
0 v2 coszB
ellipsoidal
L2
(u,9\) ,
trz
0
1E7
Table A.3
0
r
0 0
r sinz0
r)."
0 0 0
0
I
?
0 0
sindcosd I
f
r],"
; 0
0
0
0 0
cot 0
!i'
0
I
r
cot 0
Table A.4
0
r
0 0 r coszS
ry,8'0
0 0
r
7
0
0 sin/cos/
I
ry,8'
0 0 0
tan Q
0
nA ' 3 L
B'
;
tan /
0
t
r
186
4.2.
Table A.5
0
r r'cos 1
I r
0 r cos.r
r(sin2 o.ro sin2 I  l)
ql
0 0 0
0
cos .I tan c,ro
l"'u I
f
qt, 0
t
7
Table ,4'.6
0
r
U
t
0 0
r sin2 aro
I)10, 0 0 0
0 0
 sln u,ro cos(ro
tlu'
I
r
0 0
U
0 0 I)90, 0
I
f
T
r
0
cotwo
cot uo
189
Table A.7
0 0
(h + N) cos2,p
!iu
0 0 0
(h + M)
0
i+Ff
ri,a
I h+M
rp pcos "ffiAsin 0
0 0
0
U
ffi si"pcosp
1 h+N
riu
0
I
ffi
n
tun'P
[+lv
ffi tune
Table A.8
'F="oao"z 0
fisinBcosB
 2u,u 2
n
L2
0
t l
qi
fisinBcosB
0
S.or2 B
U
ShsinBcosB
tl"i
F
fisinBcosB
0 0 0 tanB
0
n
fisinBcosB
u
+1
t l
tanB
U
190
appendix B
A large part of the time spent on the solution to problems is often devoted to, on the one hand, the languagein which the problem and its solution are written (i.e. notation) and on the other hand the mathematical formalism or tools used for solving the problem. It is said that without the use of index notation and tensor analysisEinstein could not have made his generaltheory of relativity to becomesuch a success, even that this theory would never have been developedat all without or its use. This may illustrate the influencewhich the choiceof a proper notation and formalismmay have on the theory and it thereforejustifies the amount of effort put into the question of notation and mathematical formalism. This appendix will therefore contain a short treatise on the two mentioned subjects: notation and mathematicaltools. However,not all possibilitieswill be treated here. For the notation we choose here the (kernel) index notation and as mathematical tool the tensor analysis. Of course there are other possibilities,which are perhaps more modern or more commonly used. There are, however, two reasons that we choosehere for index notation and tensor analysis. At first index notation and tensor analysis are very elegant tools and easy to work with. Of course it requires, as always, some experienceto work efficiently with them, but as soon as one is getting acquainted with it, one discovers that they are very suitable for almost all our problems and that they therefore create a single general framework in which all parts of the subject fit, that formulas become very short and easy to read and write and (last but not least) that these formulas can be very easy converted into programmable code for computational purposes. The secondreason is that this thesis contains a chapter on relativistic aspects.At present index notation and tensor analysisare still very commonly used for relativistic purposes. Nevertheless, the use of vectors and matrices and their correspondingnotation is perhaps even more widespread. AIso in this thesis it is used, since it is often convenient and it sufficesour needs. However, sometimes people like to switch
191
appendixB. Index notation and tensor analysis between index notation and matrix notation. To this extent, the present appendix will also contain a section on the relation between them.
E}.1
Index
notation
The terms index notation and tensor analysis are often mentioned in one breath. One has to bear in mind, however, that they are two distinct topics. As already illustrated above, index notation is only a kind of language for writing something down. It can be used for all kinds of quantities, including the ones which are not tensors (Moon and Spencer,1986). Tensor analysis is a kind of mathematical using index tool which can be used to solve certain problems. It can be expressed notation but also other notations may be used. The two topics will therefore be treated separately. In this section some fundamental concepts of index notation will be pointed out, in the next section someconceptsof tensor analysis. 8.f.1 Kernel letters
In index notation an object (or entity) is a set (or system or array) of numbers which are in some senserelated to each other. The relationship between the numbers may be based on geometrical or physical properties or may be a purely mathematical one. In the latter casewe may also use the word holor (Moon and Spencer,1986), in the former case the word quantfty is often used. Since in this work almost all objectshave somephysicalor geometricalbackground,we will use the word quantity or throughout the work. The numbersmay also be called elements merates. Usually the numbersout of which a quantity is built up are taken relative to somecoordinate We system. Then they may be called coordinatesor cornponents. will use the word componentused in this componentsthroughout this work. The meaningof the word sensemust not be confusedwith the meaning of the same word customary in vector analysis,although the two may in specialcasescoincide(Moon and Spencer,1986)' A quantity is represented by a letter. For example, the gravitational potential is denoted by V, the metric of a space by g, the Christoffel symbols are f , a base vector is e, coordinates are x, etc. As one can see,letters to indicate quantities may be roman or greek, small or capital, even boldfaced. It is not recommendable to add other symbols or characters, like bars, accents,numbers or tildes, to the kernel letter. This is to avoid confusion with indices and to keep formulas visually as clear as possible. One should, furthermore, never assumethat the reader is familiar with what the writer thinks is a commonly acceptednotation. It is better to state always explicitly the meaning of all kernel Ietters and the quantities they represent. In the kernelindex notation of Schouten(Schouten,1954) a quantity is always denoted by the same letter, independentof the coordinate system with respect to which the components of the quantity are expressed.In this casewe call the letter the kernel letter. Not all authors using index notation follow this convention, so one often
192
8.1. Index notation seesthat if the samequantity is expressed with respectto another coordinatesystem, another kernel letter is used,or that a bar or someother symbol is addedto the kernel letter (".g.g + or V V'). In the kernelindex notation a changeof coordinates t is indicated by a change of the type of indices (see next section). So the word "kernel letter" not only has the visual meaning of the central character representing a quantity but it also has the fundamental implication that the introduction of an arbitrary coordinate system is only artificial and does not change anything of the physical or geometrical nature of the quantity. The numerical value of the componentsof the quantity may changewhen transforming to another coordinate system, but the physical or geometricalcharacteristics remain, of course,the same. Furthermore, we will try to leave the kernel Ietter intact as much as possible, also when performing other operationson it then coordinate transformations. So if the effectofany operation can be indicated by a changein the type, number or place of the indices,the kernel letter will remain the same. Thesekinds of operations are, for example, raising and lowering of indices, contraction, covariant differentiation, transpose. 8.1.2 Indices
A quantity which consistsof more than one component (so it is not a scalar) is indicated by a kernel letter with one or more subscriptsor superscripts,also called (lower or upper) indices, attached to it. For example, the componentsof the gravitational accelerationvector are indicated by Va, the componentsof the metrical tensor by g;i, the componentsof the christoffel symbols by Ifi, etc. Indices may be roman or greek letters and small letters or capitals. The index is in fact a short hand notation for the separatecomponents,for example in Sdimensionalspacewe have . { V e l A : 1 , 2 , 3 } : ( 7 a = 1V A = z , V , c = s ) , If this is simply abbreviated by Ve we have to state explicitly the range (dimension) and the meaning of the index 1., for example: the index A representspolar coordinatesin 3dimensionalspace: {rolA: 1 , 2 , 3 }: ( z A = r r A = ' , r A = 3 ), ,
the index I representscartesiancoordinates { c d l r: L , 2 , 9 } : ( " i = t , , i = 2 , r i = s ) . lt is often convenientto give the various componentsspecialsymbols like
{ x i 1 :i L , 2 , 2 } : ( r , y , z ),
193
appendix B.
{ V n l A : t , 2 , 8 }= ( % , V e , V t )
and
. {Vil; : 1,2,3}= (V",Vy,V") The latter notation is ofcoursenot accordingto the conventions the last paragraph in but in most cases this kind of notation will not lead to any confusionand the meaning of the symbols will always have to be evident from the context. The following notation, however,is ambiguous:
(v1,v2,vs)
becausethis notation does not show with respect to which coordinate system the componentsare given. From now on we will adopt a much used convention which is to confuse a quantity with its components(Foster and Nightingale, lgzg). For example,we refer to a quantity Va rather than a quantity V with components7a. Note that this convention becomesimportant if we follow the rule given in the last section to leave the kernel letter intact as much as possible,even under operationschangingthe number of indices such as contraction or covariantdifferentiation. When referring explicitly to the quantity and not to its components, the kernel letter is often printed boldface, like in the example above: quantity V with componentsV4. So the type of the indices indicates the coordinate system. In the kernelindex notation we address to each coordinate system a special (limited) set of successive letters from the alphabet to be used as indices. For example, the set of indices{r,i,k,/} can be usedfor alocal cartesian coordinate system,{A,B,c,D} f.or curvilinear polar geocentriccoordinates,etc. We allow for indices the use of other s y m b o l s t t a c h e d o t h e l e t t e r s , i k e { r , ,j t , k , , l , } , { A , B , e , D } . I n c o m b i n a t i o n w i t h a t l the kernel letter we then have for exampleVa the componentsof the gravitational accelerationvector with respect to the polar coordinates,A , Vi the same quantity but expressed with respect to local cartesiancoordinatesx,i, etc. As one can see, indices may be placed high or low. Upper indices represent contravariant components and lower indices representcovariant components, at least if we are dealing with tensors. In section 8.2.2 the meaning of these terms will be discussed. Any index appearing twice (one upper and one lower index) in a quantity or in a product (in general in a term of an expression)is, according to Einstein's summation convention, summed over all values the index can take (i.e. over its range or dimension),for example R " p T B:
, D R"uTu
9=L
where n is the dimension. Such indices are called dummy or summation indices because after evaluating the expression they are cancelled. Summing over equal
194
8.1.
Index notation
indices in a quantity or in a product is also called contraction Indices which are not dummy indices are called /ree indices. After summation over dummy indices in an equation the remaining (free) indiceson the left hand side and on the right hand side should be the same and in the same place (upper or lower), for example ai rbi  ci a i b J 11  c k : d k , are correct expressions, but oi ibk : 'i is not. We will call this rule here (weak) inder balance. Since dummy indices are summed over, we may always replace them by other letters, as long as they belong to the same coordinate system and as long as they are not already in use as free indices. The number of indices attached to a kernel letter indicates the rank (or valence or order) of the quantity. If a quantity is of secondrank it is represented a kernel by letter with 2 indices, for example o;i. If the iindex is raised we could write the result as a'i, ai, or a;.r. When working only with index notation, the first form is always clear. The indices may always be placed as near as possibleto the kernel letter. If one likes to convert certain equationsat the end of a derivation to matrix notation, the second(or third) form is to be preferred. In this casenot only the place of an index upper or lower matters bub also the place right or left: when raising or lowering an index it must keep its place in horizontal direction relative to the kernel letter (seenext section). It may help to place a dot under each upper index if other indices follow (thira form), but this makes the expression visually unattractive. 8.1.3 Matrices
Despite the fact that the languageof index notation is sufficient for writing down all kinds of problems, many people still like to work with matrices becausematrices and matrix notation (including vector notation, sometimescalled symbolic or abstract notation) are found to be more illustrative than index notation. Now the conversion between the two is not always clear. Problems occur with the ordering of the quantities in a term, with the transposeand with the inverse of a matrix. In this section we will discussthese problems. Furthermore, additional conventions will be given, which, if strictly followed,will prevent these problems to occur. One of the most important differences between the two languages is that in matrix notation only quantities up to rank 2 can be expressed adequately,whereas in index notation there is no limitation on the rank of quantities. So in this section we will restrict ourselves scalars(rank 0), vectors (rank 1) and matrices (rank 2). to In index notation the ordering of the quantities in a term is of no importance:
o' jbt r
195
is the same as bj ro'i (i.e. multiplication is commutative). But in matrix notation it is known that AB I BA (where A and B are matrices). So if we derive in index notation an expression and we like to convert it later to matrix notation, we have to find a way to discover the ordering of the quantities in a term. Furthermore we have to find out if we are dealing with the original quantity, its transposeor its inverse. In index notation the latter two play no explicit role. An expressioncan always be evaluated correctly if one strictly follows the rules for dummy and free indices, without having to know anything about the transpose or the inverse. It is not even necessary to define these forms (although this is possible). Since in matrix notation the two are very important, we have to find a way to seewhether the quantity is a transpose, inverse or the original one. The problems mentioned can be dealt with by following some additional conventions than the ones mentioned already earlier. 1. First of all we connect to the first index (when reading from left to right) the rows of the matrix and to the secondindex the columns, for example in A;i, r indicates rows and 3 columns. AIso ci is a one column ("standing") vector, the index i indicating the rows. A "lying" vector has, in this sense,to be written as r'd, r denoting the columns now, the dot (in the place of the rows) being added for clearness.This implies that indices should always maintain their place in horizontal direction. This is especiallyimportant during the processes raising and lowering the indices. A quantity with one upper and of one lower index belongingto the same set, should thereforealways be written with shifted indices: Art and nor A!r. In an equation the free indices on the left and right hand side should now also be in the same place left or right, not only upper and lower. For example
ro : Rio tri
should be written as
ro : Roiri
We call this strong index balance.In terms of a matrix product this is easy to understand: ri and ro are "standing" vectors (the indices indicating rows) of which the lengths have to agreewith respectivelythe number of columns and the number of rows of the matrix ,R. This also means that in a product equal indices always appear twice: once as row and once as column index. 2. Secondly, when reading the indices from left to right, the alphabetical order indicates which form of the quantity we have: alphabetical order indicates the original form and reverseorder the transpose,for example A;i is the original
196
8.1.
Index notation
quantity and Aii the transpose. Also: Drftoriginal, bt, transpose or di origr inal, df transpose. This rule only applies if the indices belong to the same coordinate system. In the case of coordinate transformation matrices we are dealing with quantities of which the indices belong to different coordinate systems, for example 8do. In that case we have to look into the definition of n. tr r? (o, arry other kernel letter used for the coordinate transformation) is defined (or introduced) by means of an equation like
ri : Ri .'x,o
the order is {a,r}. So the indices indicating the "new" coordinates(left hand side of the transformation equation) come first. Strictly speakingthe transposeshould be expressed a different kernel letter. by But since the two are very strongly related (they are built up out of the same (numerical) values only in a different ordering), since the difference between the transposeand the original quantity can be shown without ambiguity by means of the indices and since it is commonly accepted,we will use the same kernel letter for the original quantity and the transpose.This is also common in matrix notation. 3. In matrix notation the inverse Ar of a matrix A is define as Ar A : I d, where.I is the unit matrix. In index notation the inversea,, of a quantity Are is defined through A'" ail  6', . where 611is the Kronecker delta whose value is 1 if r : I and zero if r I t. We seeanother kernel letter has to be usedfor the inverse,since the inverseis in fact another quantity. The definition equation relates the inversewith the original quantity. We try to pronouncethis relationship a little bit more by using for the original quantity and the inversethe capital and small version of the same letter (or vice versa) but this is not strictly necessary. Using a different kernel letter becomesextremely important if we are dealing with mixed quantities (quantities having both upper and lower indices) like Rdr.  6'o. If we would have Its inverse rd, would be defined by Ri used r/ r the same kernel letter here the notation would have been very ambiguously: R'iBir  6'r. one has to customizeoneselfto take for the inverse always a different kernel letter and state explicitly the definition equation. So if rBi; is in matrix notation .B and rd, is its inverse r?r we have the following four possibilities:
197
appendix B.
index notation ' matrix notation original quantity transpose inverse Rri R;
,'j
There are however two situations in which we may use the same kernel letter for the inverse and the original quantity. That is if the quantity has either two upper or two lower indices or if the indices do not belong to the same coordinate system: A'r , B;i , C t o t A Aii , Bi; , c o i ' A T A,i , B'i , Coi t
Al
,B ,BT , Bl
,C ,CT , gL
AT ,3T
,gT
It is repeated here once again that for mixed quantities the latter does not apply so a different kernel letter lras to be used. 4. The last new convention concernsthe ordering of the quantities in a product. From convention 1 it follows that this ordering is such that equal indices (dummy indices) are as closeas possibletogether. So if a derivation in index notation leads to a Product like
attb*
it has to be written as
brra"t ,
the dummy index s appearingin both terms closetogether..In the same way: and 9;ir'tyt as t'tg;iyl ' to Vo6R;"Rb1has be written as RroVo6R0, Example we of To illustrate the use and consequences the new conventions show an example. Imagine two coordinate systemsof which the base vectors are respectivelyei and e". we assumethe dimensionsof i and a are equal. The componentsof a vector x in the two systemsare respectivelycd and ro:
x:t'igi:troQa
198
Inspection of the last equation shows that the transformation of the components is x,o : Roixi ,
where we have interchanged the indices r and a of R based on convention 1 and reversed the order of c and R based on convention 4. The inverse of the last equation is rj:foro with
R;"rJ  6rt So if, in matrix notation, we substitute.R for ,Rrawe see that if the base vectors transform with R, the coordinatestransform with RT, basedon convention3. If a vector uo transforms like v; : A;auo, a matrix I/16will transform like v;j : A;ovotAb j (see section B.2.1). In matrix notation this becomes(if we substitute Vj Vot  V): I: AVAT l,
a familiar result. There remainsone thing to be explained. Sometimes have the situation where we the numerical values of two sets of components (of the same rank and dimension) are the same, but the two sets are not representedby kernel letters with the same indices in the same places. According to the new conventions above we may only equate quantities if free indices are of the same type and in the same place (upper and lower as well as right and left). For example, in "ordinary" index notation, the symmetry of a matrix is expressedas adt : aii. This would not be correct based on the strong index balance. Another situation where this occurs is with orthogonal rotation matrices, where the transposeequals the inverse. Expressing this as Rxi: ril is not correct. These kinds of equalitieswill be expressed with bracketsl:
{att} : {att}
notation is chosenfollowing (Moon and Spencer,1986). In this book a very beautiful and extensive treatment on index notation and tensor theory is given, which, however, does not fully coincide with the conventions presented here.
rThis
199
and {Rt,} = {"t'} . The number and type of the indices on the left and right hand side are still the same, but the place may be different. This indicates numerical equality for all values the indicescanassume,so{oii} { " i t } m e a n s a r r : a r r r a r 2 : a 2 r r a L S : a 3 r , Q . 2 3 :Q 3 2 , etc.
Example Supposethat, in the previous example,the matrix V is symmetric: {Vo6} : {Vaol,.. Now let us derive the transposeof l, which is in index notation Vii: 1V1;\ : {Ri"V"bRbi \
{R,bV6oR"r} {RjbV"bR";} {R"rV"rR1b} {R:V"bRb j} {V;i} (dummyindices) (symmetry of V) (convention 4) (convention 1) (expression for V;1) ,
so we see I is also symmetric. As already said, the additional conventionsgiven in this sectionare only needed if one likes to convert from index notation to matrix notation. If this is not the case (what is to be recommended) "ordinary" index notation can be used (Moon and Spencer,1986).
8.2
Tensor analysis
We use tensor analysisin some chaptersof this work becauseit is a suitable mathematical tool, as stated in section B.l. Another reasonwe use it, follows from the fact that this work deals with some aspects of gravitation. And since Einstein we know that gravitation is closelyconnectedto geometry,in particular the geometry of curved spaces. Especially in the latter case, the tensor analysis, as it is based on the absolute differential calculus of Ricci and LeviCivita, is found to be a very suitable mathematical concept. In the same senseas index notation and tensor analysis are two different topics, though strongly connectedin practice,also geometry and tensor analysisare not the same. If we, in this work, talk about geometry,we are dealing with metrical spaces, curvature, surfaces,metric, curves, etc., all in a more or less concrete sense. The geometrical objects are the objects of study themselves. If we look upon geometry in this way, tensor analysis is again (only) a mathematical tool to help us solve our geometrical problems. The two topics, being distinct, are therefore treated
200
8.2. Tensoranalvsis here separately. First we deal, in the present section, with tensor analysis. We discuss the tensor concept, the terms contravariant and covariant and the processof differentiation. In the next section, some geometrical aspects are discussed:spaces, metric and curvature. The division of the topics, treated in this appendix, into three parts (i.e. index notation, tensor analysis and geometry) based on the considerationsgiven above and in the last section, may seem artificial to some readers, but it is just a way of ordering the broad range of topics related to the theory and it suits the present work. 8.2.1 Tensor
A quantity is defined as a set of elementswhich are in somesensemutual connected. This connection depends in its turn on the definition of the quantity and it will in most casesbe a physical or a geometricalone. Such a relationship between the elementsof a quantity is one of the characteristics tensors.But a more important of aspect of a tensor is that its elements are the components of the quantity with respect to some coordinate system. This does not mean that all sets of components are tensors. To be a tensor, the set needsto havesomeadditional, specialproperties, which will be explained in this section. The important point here is that there is a distinction between on the one hand the quantity itself, which has some physical meaning, and on the other hand its representation a set of numberswith which we by can carry out computations. Thesenumbers maybe the componentsof the quantity relative to a coordinate system and this hasto be so as one of the conditions for the set to be a tensor. One can imagine that the value of the numberschanges another if coordinate system is chosen to represent the same quantity. Consider for example the componentsof a displacementvector in E3 in two different cartesiancoordinate systemshaving the sameorigin but a different orientation. Sincethe introduction of a coordinate system is only artificial and has nothing to do with the quantity itself, we like to have a mathematical framework underlying our computations which is independent of this arbitrary choice of coordinates. This is exactly the essential characteristic of tensor analysis and it is captured in the word "absolute" in the absolute differential calculus of Ricci and LeviCivita: "The tensor calculus is said to be absolute becauseit is independentof the details of the choiceof coordinates, that is, the equationshave the sameform in all coordinate frames" (Ohanian, 1976, p.22L). To see what this means mathematically we discuss the transformation between two coordinate systems. Let us denote the "old" coordinatesystem with zo and the "new" with cd and let the two be connectedby the relationship
rt : r'(ro) ,
(8.1)
expressing the fact that the new coordinates are some functions of the old. The coordinate transformation B.1 may be chosen at will, but for conveniencewe will
201
appendixB. Index notation and tensor analysis impose some limitations on it, namely that the transformation function is singlebetween r' and ro), valued (which means that there is a onetoone correspondence (i.e. its derivatives up to a sufficient order exist and are continuous) that it is analytic and the Jacobian of the transformation is different from zero in every point, cf. (Moon and Spencer, 1986). In this casealso the inverseof the transformation exists: r" : x"(xi) . For the rest the transformation may be completely arbitrary, nonlinear as well as linear. Manipulations of equation B.1 tend to be complicatedif the transformation is nonlinear. But if we are willing to deal with infinitesimal coordinate changes we (which is the casein tensor analysis), may linearizethe transformation B.1. Differential changesin the coordinates are then transformed as
4,;  9!d,o
or,'
or lnverse
: d,xo
#O} ,
where the partial derivatives ffi ur" computed from 8.1. These partial derivatives are in general functions of the coordinates. Only in the case of linear coordinate transformations they are constants. Now we definea contravarianttensor (under generalcoordinatetransformations) as a quantity which transforms according to _ 6ri )ri fij...p6rd ErF '
0rB'
(B.2)
a covariant tensor as a quantity which transforms according to A ; i . " p: 0ro Tso ' ' ' xo A"p o ai aJ aro
(B.3)
. '. . o r l 'A. r . r n " p : o r ' t ; ..r " a"" 5"0ffi wherethe partial derivatives ffi to another according
(8.4)
a*ail"j'
Tensors of rank 0 are called scalars, tensors of rank 1 are called vectors. We will not eachset of numbers which in generalnot call tensorsof rank 2 matrices,because (a matrix) has to be a tensor. A can be written as a system with rows and columns
202
8.2. Tensor analysis quantity is a tensoronly if it fulfils one of the transformationlaws B.2, 8.3 or 8.4. A tensor of rank 2, however, can always be representedas a matrix, but that does not make the two identical. As a concessionto matrix notation we will, in the sequel, refer to # u" the transformation matrix and to # u" the inverse transformation matrix, thereby keeping in mind the remark about secondrank tensors made above. We seethat the tensor character of a quantity is only of interest if we are dealing with different coordinate systems or with coordinate transformations. This does not always necessarily have to be the case. But f different coordinate systems are involved and a quantity is a tensor, its representationin each coordinate system is the same except for the type of the indices,for example Ad versus Ao, Bob"versus B;,it1r,, etc. tansforming from one coordinate system to another may never cause the tensor representationto changefrom, for example, Ai to ao, A; to Ao6 or Ad to Ao + Bo, etc. This means that, if an equation consistsonly of tensors, its form is the same in any coordinate system, except for the type of the indices. This fact is one of the most important and most powerful conceptsof tensor analysis. We have for example
u ij : wi *zkj + Y;r'j S : g;iriai
in one coordinate system, which is written uob : as
S :
in another. But also if all componentsof a tensor A;i are zero in one coordinate system,.e. A;i:0 i ( o r A ; j : 0 ; i , t h e r i g h t h a n d s i d eb e i n gt h e " n u l l t e n s o r " ) t h e n they are zero in any other coordinate system. In the examples above, we see that tensors may be added or multiplied with other tensors. Addition of tensors in only possiblewith tensors of the same rank and the same type (remember: dummy indices are to be consideredcancelledafter evaluation). The result is again an tensor of the same rank and the same type. Multiplication of tensorswill result in a new tensor of which (c) the rank is the sum of the rank of the multiplied tensors minus the number of dummy indices, and (6) the type of the indices is the type of the remaining free indices in the product. It can be shown (cf. (Hotine, 1969), (Moon and Spencer,1986)) that these operations indeed result in new tensors. Reversely, each tensor can be written as the product or sum of other tensors,with or without contraction, as long as they obey the rules above. 8.2.2 Co and contravariant
In the previous section covariant and contravariant tensors were defined as quantities of which the componentstransform in a certain manner. This way of introducing
203
tensors is based on the classicalapproachto tensor algebra and tensor analysis (cf. (Ohanian, 1976), (Sokolnikoff,1951), (McConnell, 1957), (Hotine, 1969) and many other textbooks). This classicalway of introducing tensors is a geometrical one and it in fact goes back to Euclides. H", for the first time, gave a meaning to the concept of a geometricalspace when trying to describe the world (or space) surrounding us. Later the connection to some specific geometrical structure was abandoned and a space was considered just a set of points with some arbitrary structure. Whereas the spaceconsideredby Euclideswas a linear, 3dimensional, geometricalspace (Euclidean spaceor Et), u spacein a purely mathematical sense may be of any dimension, linear or nonlinear and with or without any special geometricalstructure. to Later, when Descartessuggested project the collection of realnumbers onto concept of coordinateswas introduced into the space. The elea straight line, the by ments, out of which a spacewas built up (i.e. points) could now be represented a set of numbers({cdl; : 1,..., n},n being the dimensionof the space).The points are said to be labelled by n real coordinates. If the structure of the spaceis reduced to zero, we are merely considering a set of labelled points and the space is said to be arithmetfc, denoted by X'. WhereasDescartesused linear coordinatesto label the points of E3, in general also nonlinear coordinatesmay be introduced. A linear space (a spacein which linear coordinatescan be introduced) without metrical structure (seesection B.3.1) is called an affine space: A'. In such a space,various coordinate systemsare connectedthrough linear (or affine) coordinate transformations. A next step is the introduction of an orthonormal coordinate system into an affine space (i.e. a system of which the coordinate axes have the same origin, are mutual orthogonal (independent)and have the same scale). In doing so, we endow the spacewith a specificmetric and the spaceis called Euclidean. As already mentioned before, the tensor character is in fact only of importance if different coordinate systems and the transformations between them are concerned. A tensor is a quantity which behaves as an invariant under such coordinate transfortnations. In this sense,a tensor is an extensionof a uector,a concept originally coming from physics. Gibbs introduced a geometrical visualization of a vector, namely an arrow, characterizedby its direction and its length. A point, having a fixed location, is the most elementary (geometrical)invariant under coordinate transformations. A vector, fixed by means of its begin and end points, is also an invariant. Of course the numerical values of the componentsof the vector may change if we carry out a coordinate transformation, but the vector itself, having some definite physical or geometrical meaning, remains unchanged. Its components express, however, the connection between the invariant vector and the arbitra.rily introduced coordinate system. The components of a vector I are the scalars )o relative to a basis {e"lo : a 1,. .., n) such that ) : ,\oea. The set of vectorss c&n be considered basisof space if it spans the space (i.e. every vector in the space may be written as a the linear combination of the members of the set eg) and if its members are linearly
204
8.2.
Tensor analysis
a=l ;c=l
Figure B.1
Contravariant components.
independent. The base vectors may also be visualizedby arrows, however of unit length, lying along the coordinate axes,pointing in the positive coordinatedirection and with the begin points in the origin of the coordinate system. In a linear space (affine space) the componentsof an arbitrary vector with begin point not in the origin of the coordinate system, may be obtained by parallel transport of the basis from the origin to the begin point of the vector. If we now consider a vector field, defined throughout the whole space)such transport of the basis will result in a basisfield,i.e. in each point of the spacewe have a set of basevectors. If we use curvilinear coordinatesin a linear space,or if the space is not linear itself, we generalizethe concept of a basis to the set of tangent vectors to the coordinate curves in each point. The tangent vectors now span in each point the (local) tangent space. The basis now dependson the coordinates, just as the vectors of a vector field may dependon the coordinates.(In tensor algebra deal with problems we either restricted to the local tangent spacein a point or concerning the whole space in which case the space has to be linear (so that all tangent spacescoincide with it). Tensor analysisstudies problems either dealing with the connection of different tangent sPaces a nonlinear spaceor dealingwith curvilinear coordinatesin linear in space,in both casesforcing us to linearizeand to considerinfinitesimal coordinate changes.) The componentsof a vector relative to a basisin the sensedescribedabove (i.e. in general the tangent vectors to the coordinate curves in a point) are now called contrauariant components, denoted with upper indices as in l : )oea. This definition is, at this point, completelyarbitrary, but we have to make a start somewhere. The geometrical interpretation of these contravariant components is probably well known to everybody, at least in Euclidean 2 or 3dimensional space. There, those componentsare the parallel projectionsof the end point of the vector onto the coordinate lines, see figure B.1. In the case of nonlinear coordinatesthe picture is essentially the same being only differential now. Sets of coordinate lines in an ndimensional space, along each of which one particular coordinatevariesand the other n 1 coordinatesare constant,are not the only way in which a coordinate system reflects the geometrical structure of a space.
205
appendix
B.
lndex
notation
A coordinate system provides us also with a dual structure of sets of coordinates planes on each of which one particular coordinate is constant and the other n  I coordinates vary. In the former structure (sets of coordinate lines) we may find the base vector in the rDdirection by taking the derivativesof all the coordinates x 4r a : 1 , . . . , n w i t h r e s p e ctto r b : 0ro
o{"u)
for o : 1,.. . , n and fixed D. This givesa unit vector (tangent vector) in cbdirection. The complete basis is found by repeating this for all values b can take (i.e. b
lr'" rn)'
In the case of sets of coordinate planes (dual structure) we take the derivative rb of one particular c" with respectto all coordinates ,b : lr. . . ,n:
o{r"}
0rb for b : 1, . . . , n and fixed a. This also gives a unit vector which is to be regarded as the gradient vector of the scalar field co. Repeating this for all values of a we obtain ee. the socalled dual Dosr's Gradient vectors of arbitrary scalar fields rp 9(r") may now be written with their componentsrelative to this dual basis. So if the vector p is the gradient of gr then F : poea. The componentswith respectto the dual basis and they are written with lower indices2. In the are called couariantcomponents samesenseas the gradient of a scalarfield reflectsthe rate of changeof it in a certain direction (i.e. the density of the g : constant surfaces)the covariant components are related to the density of the coordinateplanes in the direction perpendicular to those planes. Only in an Euclidean space (with a metric defined) we may visualize these components as the orthogonal projections of the end point of a vector onto the coordinate lines, seefigure B.2 (Hotine, 1969). Note that in this figure the vectors are not of the same type as in figure B.1, which fact we tried to indicate by using dashedlines in figure 8.2. In other spacesno satisfactoryvisualization is possible. So we have at this moment two types of vectors. At first we have ordinary vectors (contravariant vectors) with (contravariant) componentsrelative to the basis. Examvectors,velocity vectors,etc. These are ples of this type of vectors are displacement with the vectors everybody is customary to work with. Secondlywe have coaectors covariant components relative to the dual basis. These are, for example, gradient vectors. Only if in the space a metric is defined the two sets of components are related by meansof the metrical tensor (seesectionB.3). In this caseeachvector may be written with covariant or contravariant components. The distinction "vector" and "covector" looses its usefulnessand we will simply talk about vectors. For the components,however,the distinction betweencovariant and contravariant remains.
2Sometimesquantities which can be written with covariant componentsrelative to the dual basis are called lforms, cf. (Misner et al., 1973).
206
8.2.
Tensor analysis
Pa=2
/:
"a=l
Figure B.2
Coaariont components.
Tensors,being an extension of vectors, may also have covariant or contravariant components. For tensorsof rank 2 or higher thesecomponentsare, however,difficult to visualizegeometrically,the extensionbeing a purely mathematical one. Now considera generalcoordinate transformation ct : r;(r").From this transformation we may compute the transformation matrix ffi and the inversetransformation matrix ffi such that u*Uffi: 6r{. In a point P we now have two sets of coordinate lines, one set for the cdcoordinates and one for the cocoordinates. The tangent vectorsalong all the coordinatelines u" (togetherforming the basise") were : 1,. . . , n for eachb. In the same manner we may find found abovethrough ffi,a  1, . . . n for a set of tangent vectors along the cd coordinate lines through , ffi, " each r. This set forms a new basis ei in P, spanning the same tangent space as does ea. The quantities # ^", be interpreted as the components of the base vectors ei relative to the basis eg, just as the quantities ffi: 6Do the componentsof the are base vectors e" relative to this basis itself:
: "" ff",
": ffi." i
207
appendixB. lndex notation and tensor analysis In order to seehow the contravariant components of a vector transform, we proceed as follows: Ani
so the contravariant components transform with the transformation matrix ffi. In a similar manner we may show that covariant components transform with the Wtureas we already defined the contravariant inverse transformation matrix ff. and covariant components at the moment of introduction of the two sets of base vectors(basisand dual basis)we could haveequallywell definedthem at this stageas the componentstransforming under a generalcoordinatetransformation ci : ri(*) the inversetransformation matrix with respectively the transformation matrix # ", to quantities of higher rank forms the background ffi. n*t"nsion of this reasoning of the tensor definition given in section 8.2.1. Besides the classical theory of tensors described above (which is strongly connected with geometry due to historical developments)there is also a more modern approach for introducing tensors, which uses the abstractalgebraicconcepts of space, vector space, sets, linear and multiIinear functionals, mappings, manifolds, groups, charts, atlases, etc. Examples of this approach can be found in (Foster and Nightingale, 1979), (Bishop and Goldberg, 1968) and others. whereas in the classicalapproach the first step was a special metrical 3space (Euclidean of to spaceEg) which was later generalized nonmetrical spaces any dimension,the of starts with generalspaces arbitrary dimension and may end up modern approach (after imposing more and more conditions onto the space)with the ES as a very special case. In spite of the EB being in fact the traditional "working space" of some appeal to them if interpreted geodesists, modern approachmust also.have the as a framework in which we work from "large" (most general) to "small" (specific case). 8.2.3 Derivatives
In the last sectionwe saw that tensor analysisdealseither with problemsconcerning in nonlinear spacesor problems in linear spacesexpressed curvilinear coordinates. in a nonlinear space is in(It is obvious that the use of curvilinear coordinates evitable.) Consider now one of thesetwo possiblesituations. In the space,a tensor field is definedwhich, in general,will be a function of the coordinates.In a point of this spacewe may construct a basisspanning the local tangent spacein that point. In curvilinear coordinatesthis basis will also depend on the coordinatesso we are in fact consideringa field of coordinate dependentbase vectors and tangent spaces (O'Neill, 1966). A changeof coordinateswill in generalimply a changeof the local basis with respect to which the tensor field is defined. Study of the tensor field therefore requires knowledgeabout the way in which neighbouringtangent spaces
206
8.2. Tensor analysis are connected. The situation of continuously changing base vectors (which is the casewhen using curvilinear coordinates) forces us to consider differential coordinate changes.This brings us to examining the processof differentiationof tensors. If we like to continue our computations using tensor analysis,we require that this difterentiation processyields again a tensor. Partial differentiationof a tensor. as known from ttordinary" analysis,does, however,in generalnot lead to a tensor, as can be shown easily. Consider an arbitrary contravariantvector u" which, under a general coordinate transformation rd : ri(r") will transform as ui ffiu". The partial derivative of this equation with respectto ct yields:
or.,i a (a"r\
_ t _ l  .  I
6rj
 ' Ed \Er" ) Ex" Eri o2xi orb ^ ar; arb aoo  ' 6ra6"b Qbi flso flri Qab
oriou"
(B.5)
The partial derivative is denoted by a comma precedingthe differentiation index: ,i.,j : # lt the term containing the secondorderpartial derivative was absent, u',, would transform as a mixed tensor. We conclude that obviously the ordinary partial derivative is not a tensor. One exception to this conclusionis the partial derivative of a scalar field ,S,which, accordingto the chain rule, transforms as AS Arr: )xa 63 ' A"t A""
So for each scalar field ,S, S.o is a tensor. Partial differentiation obviously is not the right method to account for the connection of neighbouringtangent spaces.In order to find this connection,we look at the way in which the basise" changes under a differentialchangeof coordinatesdro in each of the coordinate directions. Expressingthis changeof the basevectors ffi relative to the basis e" itself, we may write de" : Litec 6rt (8.6)
where the Ili are called the coeffi.cients the linear connectionor for short "linear of connection", ttconnection coefficients",ttaffineconnectiont',etc. The way of introducing the connectioncoefficients in equation B.6 may be interpreted as imposing as a certain structure on an arithmetic space(although this structure is not a metric) . This structure consistsof nothing more then requiring the basevectorsto changein a linear way under a differential coordinate changeand not completely arbitrarily. The spaceis now said to be a linearly connectedspace (Moon and Spencer,1986). The Ili may in general be functions of the coordinates. Sometimes they are introduced just by giving their transformation equation (ibid.), which is:
tlh:" 6xt }xJ 0x" n1
ar" aroarr\i 
0r"
02ri
flri flroflrb
(B.7)
209
appendixB. Index notation and tensor anilysis This equation (which we give without derivation) shows that the connection coefficients do not form a tensor. It furthermore shows that the linear connection is : We will in the sequel symmetric in its two subscripts, at least if ## ##. the case. If this is not true, the basis field is called assumethat the latter is always and the spaceis a socalled torsionalspacein which differentiation is anholonornfc possible,but integration is not. With the linear connection we try to find a process of differentiation of a tenthrough sor which gives again a tensor. From equation 8.7 we solve b, ffi multiplication with $$:
02 ri 0ro Erb
0r' 0r"
ry,##,*
into equation 8.5 givesafter somemanipulations (cf. Substitution of this expression (Moon and Spencer,1986)): oxi orb (y: or, *' r t'i* n , l _ . u *'n o' " \ i r Aro Ail \Arb 5Ai ) This equation shows that the quantity on the left hand side transforms as a mixed tensor. It is called the couariant deriuative of a contravariant vector and is denoted by a semicolon precedingthe differentiationindex: ui i (or Viui):
,i ,j = ui ,, + t1rrk .
(8.8)
We seethe covariant derivative consistsof two parts. One part is the familiar partial derivative. The additional term, containing the connection coefficients,accounts for the change of the basis in the differentiation process. Sometimes it is no longer important that a tensor was originally created through the processof covariant differentiation. In that casewe simply write u].. This alsomeansthat everycovariant index of a tensor of rank r can be viewed upon as originating from a process of covariant differentiation of another tensor of rank r  1, namely one covariant index V;rs ot 9o : 9;a : 9,a' less. For example: Ai i1, : A' irn Ai.i1,or Vr, : Vr;, '': uorto. Since for a scalar field the partial Consider now the scalar'field S derivative transforms as a tensor, it is S,o S,o. So we have:
(uow") p (oow");t
ua,b uo + vo wo,6 : :
210
8.3. Geometry which has to be true for arbitrary t a, so that we find for the covariant derivative of a covariant vector:
0oib : rto,6  lf6 v"
(B.e)
The covariant derivative of a contravariant tensor of rank 2, tab can be found in a similar manner by writing 1ob ,a*b. The extension to tensors of arbitrary rank or type follows analogously. For each superscript one adds to the partial derivative a term with I, and for each subscript one subtractsa term with f (like the ones in equations B.8 and B.9). The covariant derivative is the tensor analogy ofthe partial derivative ofclassical analysis. For the total derivative duo : ffidxb the analogy in tensor analysis is the absolutederiaatiue, defined as:
Doo: uo.6d* : duo tl6f;u"drb
(8.10)
Again only for a scalarfield the absolutederivativeof tensor analysisequalsthe total derivative of ordinary analysis. The absolutederivative can be used to extend the notion of a field of "parallel" vectorsalong a straight line in Es (which is ff : g) 66 a parallel field ofvectors along a generalcurve in a generalspace(seesection B.3.1). For the latter we have
D ! ": o
du
(B.11)
along the curve parametrizedby the parameter u. If we take for uo in this equation the tangent vector dx"f du along the curve, equation 8.11 becomes d2ro , nndrb dr"
du'+tui^E:o
which is the equation of a geodesic.So geodesics are those curves in space along which the tangent vectors in all points form a parallel vector field. Note that the meaning of the word parallel in this senseis much more general and abstract than the visual meaning it has in Euclidean space.
B.3
Geometry
In the previous sections some aspects of index notation and tensor analysis were treated. As explained in section B.2 purely geometricalsubjects were left out. In the present section attention will be paid to geometry. First the notion and classification of severalspaces will be treated, which was already initiated in section 8.2.2. Furthermore, we saw in section B.2.3 that the introduction of the linear connection I already imposes some kind of structure to a space. The step from such nonRiemannian (nonmetrical) spacesto Riemannian (metrical) spacesconsists of the introduction of a metrical tensor. This tensor will be treated in section B.3.2. Special attention will be drawn to curvature and curvature related quantities in the last part of this section. Many general textbooks exist on geometry. As an example see (O'Neill, 1966).
211
8.3.1
Spaces
In section 8.2.2 a spacewas considereda set of points with some structure. Such a definition still remains vague becauseno explanation of the words "point" and "structure" is given (Moon and Spencer,1986). The reader, probably familia,rwith these concepts,will have an intuitive understandingof these ideas. When proceeding, the meaning will becomeclearer through the description of some specialcases. Instead of "space", the word manifold is also used sometimes,but in the literature one encountersvarious definitions of a manifold, strongly dependent on further applications. Most frequently it is usedwhen introducing the tensor conceptin a modern, abstractalgebraicmanner (seesection 8.2.2). In that case,the structure imposed on the set of points (which has to have a finite dimension) has to be such that coordinate transformations involve differentiablefunctions. Or simply put: the space has to have certain smoothnessproperties. The classical,geometricaltreatment of tensor analysisusually usesthe word space.The introduction of a coordinatesystem (viewed upon as a mapping of the elementsof the space(the points) onto the set of real numbers) allows labelling of the points in a unique manner. If the structure of denoted by Xt. the spaceis reduced to zero, we call it an arithmetic space, can be imposed on the arithmetic space by giving A very primitive structure t'out", ttneart', etc. The space is then a certain meaning to concepts like "in", called topological.As we already saw, another way of giving some kind of structure to a space is the introduction of the connection coefficientsIi!. In general these are functions of the coordinates,i.e. Qi(r'). The spaceis now a linearly connected denotedby Lt. If, in such a space,linear coordinatesmay be innonmetrical space, troduced, we have a linearly connected,nonmetrical linear space,which we already denotedby At. Calling the latter spaceaffine before as an affine space, encountered may tempt us to call the former "curved", but sincecurvature is a concept or linear this may be misleading. As a comprowhich only has a meaning in metrical spaces, mise a linearly connected,nonmetrical and nonlinear space is sometimescalled "pseudocurved". Different coordinate systems labelling a pseudocurved space are connected via nonlinear transformations. The "point" is one of the few invariants of such a space. In affine space(in which coordinatesystemsare connectedthrough linear transformations) there are more invariants (like proportionalities of lengths and surfacesor all kinds of intersectionproperties) giving already the possibility of solving some elementarygeometricalproblems. What remains are conceptslike distance,angle, orthogonality and others, which first obtain a meaning if a metric is introduced in the space. This idea originated from Riemann, who proposed to introduce the concept of distance by means of a quadratic differential form (8.12) dsz : go6dr"dxb , the infinitesimal distance ds being an invariant and the quantities 9o6the metrical coefficients(or metrical tensor),which, in general,are functions of the coordinates. in The metrical tensor will be discussed the next section. A spaceendowedwith a
212
8.3. Geometry
denoted by Ro. metric by means of equation B.12 is called a Riemannian spaces, In general a Riemannian spaceis curved. We may use this word here, since in a metrical space we can give a meaning to the word curvature, as we will see in section 8.3.3. Only if it is possibleto introduce in the space globally (i.e. covering the whole space) a linear orthogonal coordinate system, the space is called linear, flat or Euclidean and is denoted by Eo. The complete classificationof spacesas it is presentedhere is given schematicallyin figure B.3. arithmetic: Xn
r
Iin. coord. nonmetrical: Iinearlv connected:Ln affine: An
o o
metrical:
Riemannian:
Rn
Euclidean: En
Figure B.3
Spaces
8.3.2
Metric
The metric, imposed on a spaceby meansof the introduction of the metrical tensor g, allows us to give a meaning to geometricalconceptslike length, angle, volume, etc. From the definition of the invariant ds (equationB.f 2) it followsthat 9o6is symmetric. In general9o6is a full matrix of which the componentsare functions of the coordinates.Since9o6is a tensor (which we will not prove here), its transformation
3In fact, Riemann introduced the quadratic form directly into an arithmetic space. The connection coefficients I were afterwards expressed in terms of the metrical tensor. as we shall see Iater.
213
appendix B.
Index notation
equation is
9ab: 6ri }sj jro artg;i '
(B.13)
This equation is extremely usefirl for computation of the components of 9o6in vartous curvilinear coordinate systems from g;; in cartesian coordinates, at least if the transformation equations cd : ,'("") are known. The metric tensor provides the spacewith an inner product. So analogouslyto equation B.12, which gives the infinitesimal length ds belongingto a differentialchangeof coordinatesdr, the length u of a vector u" is given by
'2  go6'o'b , and the inner product of two vectors uo and tlo is gobuo b w If the latter quantity is zero, the two vectors are orthogonal. The angle 0 between the two vectors is now defined as cosd:
g obvo tttb uID
If we have an orthogonal coordinate system ub with u : 1/lo6u" and tu : \/ g "bw;fi. and we take for uo and tu" differential changes of the coordinates in each of the coordinate directions, we can show that the metrical tensor 9o6takes on the diagonal form, cf. (Moon and Spencer,1986). If, furthermore, the coordinatesare linear, the diagonal componentsof 9o6 are constants. In a linear, nonorthogonal coordinate system, 9o6has not the diagonal form, .but its componentsare still constants. or The contrauariantmetric tensor goo (also called associated conjugate metric tensor) is defined through gobgb" 6o, the inverseof goo.Thecontravariantmetric tensor provides and it can be considered the dual spacewith an inner product: g"bvowt . Now let us define the components of a covariant vector ;r as the tensor product
Po: gobAb
where )D are the components of an arbitrary contravariant vector l. Using the definition of the contravariant metric tensor we find the inverse relation as
\o : goblrb .
r,* ,: {n"u
214
8.3.
Geometry
and
gobl"Ab
from which we conclude that p.2 : )2. We therefore identify the two vectors and write for the components
)o : 9o6AD
or
)o : goD)o .
These operations (which involve the metrical tensor) are called respectively the Iowering and raising of indices. They give us the possibility to compute from a vector either the covariant or the contravariant components. In a space where a metric is defined,the distinction betweencovariantand contravariantthereforeonly appliesto components,not any more to the vectors(or in generaltensors)themselves (comparesection8.2.2). Whereasthe introduction of a metric offers us the possibility of solving a large number of geometricalproblems tnorethan without a metric, it is in fact a limitation on the generality of the space. As we shall see,the introduction of the metric limits the choice for the connection coefficientsf . These have now a specified form to be computed from the metrical tensor. This relationshipwill be derived now. Consider two arbitrary parallel vector fields )o and po along a curve parametrized by a parameter u. According to equation B.11 we have for the two vector fields
D ! ": o
qu
and Dp" :0 du .
In E3 the inner product of two parallel vector fields is constant along the curve, so we require analogously that in curved space of any dimension d(g"bl" ttb)_ D(so,)"pb) _ o. du du (For scalars the total derivative equals the absolute derivative, see section B.2.3.) This yields:
:(T)^",'
215
appendix
B.
Index notation
:9ab;c
0.
This must hold for arbitrary vectorfields,\o andp.oand for arbitrary tangentvectors # "" we have
9ob:": O
(which is a tensor equation, so it holds in any coordinate system) or if we use equation B.9 : )9ob,c t!" suo *" s"a . Cyclic permutation of the indices gives lca,b:*g"o**s"o lbc,a: tA g"o t t!, soo' Adding the secondand the third equationand subtracting the first and also dividing by 2 yields tlu Sra: ; (9o"1f gb",o 1ot,c) or rfo: 1 ,o'o t (goa,u 9bd,a 9ob,d) (B'15)
. l
(B'14)
Christoffel already derived this quantity (equation B.15) some time before the development of the tensor analysis, and it is therefore called Christofel syrnbol of the secondkind. The Christoffel symbol of the first kind is equation B.14: l o a .= 9 " a l f . The explicit form of the components of the metric tensor and the Christoffel symbols are given for several coordinate systems in section A.2' We should mention that, with the use of the metric tensor, we may also give a definite meaning to the concepts of divergenceand the Laplacian. Consider a scalar field gr. The gradient of this scalar field is g;o: got which forms a covariant vector. The contravariant componentsare rpo: gobpb.Covariant differentiation of the latter and contraction of the indices givesthe divergenceof go:
go,o = 9oo: g u b4 o b ,
216
8.3.
Geometry
8.3.3
Curvature
In a curved spaceno linear coordinates can be introduced,at least not globally. Take for example the 2dimensional surfaceof a sphere. It may be described by curvilinear polar coordinates (0,4) but not with'cartesian coordinates(r,y). In a flat space, however, linear coordinates as well as curvilinear coordinates may be introduced. Take for example in E3 cartesiancoordinates(r, y, z) and polar coordinates(r,g, \). In general, the components of the metrical tensor in curvilinear coordinates are functions of the coordinates. Only in linear coordinates, those components are constants. Suppose now, that we are given a metrical tensorof which the components are functions of the coordinates. How can we find out if this metric belongs to a curved space, so that it by no means can be transformed to a coordinate system in which the componentsare constants,or that it is just the metrical tensor of flat space but expressedin curvilinear coordinates? Or stated otherwise: what is the condition for a space to be flat?. In order to get an answer to this question we look at the way in which the componentsof the metrical tensor changeunder a differentialchangeof coordinates, i.e. we compute the derivatives of the metrical tensor. In linear coordinates the componentsof the metrical tensor are constants,so their partial derivativeswill be zero. But the partial derivative is not a tensor, so we may not concludethat in that casethe componentsof the metrical tensorwill be constant in every other coordinate system. We have to look to the covariant derivative. However,this does not help us since we already saw in the last section that the first covariant derivative of the metrical tensor is zero in all coordinatesystems,independentof the structure of the space. We may try to look to the Christoffel symbols. They are some function of the first partial derivativesof the metrical tensor. But they also are no tensors,so if they are zero in linear coordinates,this does not mean that they are zero in other coordinate systems. The answer to the question above has to be found by consideringthe second partial derivatives of the metrical tensor. Now the first partial derivatives of the metrical tensor appear, via the Christoffel symbols, in the expressionfor the first covariant derivative of an arbitrary vector uo (in generalof arbitrary tensors). For the secondpartial derivativeswe thereforecompute the secondcovariant derivative of uo. We find: : ua;bc (uo;t)," *" ro,u \1v";a : (rop  t"do ,"  ljr(ro,o * ro) \'a r) \1@",0 lfa rr) : la,bc t1,. ,o  t&va,, \7va,a *  tf"vo,it* tifior" . I:"lbit," We also have, on interchangingthe indices 6 and c: : za;cb ua,cb t!"p ra  *"ro,u  *tra," + +l:b\iu"  lbd.uo,,t ruo"tior". *
217
appendix B.
Index notation
Subtraction yields: uo;bc ua;cb: ua(*r,a  *," + L'" ql  qi 4:) : and the symmetry of the Christoffel symbols and we also where we used uo,bc u4,c0 changeddummy indices. The left hand side of this equation is obviously a tensor, so is u4, which implies that the term betweenbracketsis also a tensor. This is the also called the cutuature tensor: RiernannChristoffel tensor Rdob",
(B.16)
Now, in flat spaceand in linear coordinates,the componentsof the metrical tensor are constant. The componentsof the Christoffel symbols are in this casezero, and so are all the components of the RiemannChristoffel tensor, as follows from equation 8.16. Since the latter is a tensor, its componentswill in this casealso be zero in all other coordinate systems. The vanishing of the RiemannChristoffel tensor is therefore a condition for the space to be flat, and vice versa. It can be shown and sufficientcondition (seefor example (Ohanian, 1976)). that this is a necessary The curvature tensor contains all the information about the curvature of the space . in its na components (where n is the dimension of the space) However, it posa sesses number of symmetries and its componentssatisfy certain identities, which reduce the number of independent componentsto Snz(n2  L) (see for example it (Moon and Spencer,1986)). Nevertheless, is sometimesuseful to consider contractions of the curvature tensor. The Ricci tensor is defined as the contraction of the curvature tensor on its first and last index:
Rob: R"ob" and the curvature scalar is defined as R: g o bR o u_ R o o .
We mention one last contracted form of the curvature tensor, especially valuable in the general theory of relativity. It is lhe Einstein tensor, defined as _ Gob: Rob rr*n* .
It can be shown that this tensor is symmetric and has zero divergence, i.e.
G " ' . u :o .
2lE
appendix C
Inclination functions
In chapter 4 we used, as observationequations,expressions the local cartesian for secondorderpotential derivativesin terms of derivativeswith respectto the orbital polar coordinates (r,6,ro). The latter coordinate system has the satellite's orbit as equator. The expressions are derived in section 3.1. We see that in this casewe have to know the expressionfor V6, the potential derivative with respect to /, for example for the computation of the gradientsV", andVs". However,in section 3.2 we had the potential given as function of either (r,0,\) or (r,wo,ur,I) and not of the coordinates (r,6,uo).In order to find V6 we proceedas follows. Let us compare the expressions V, in terms of ro' and in terms of rA' (cf. for (Betti and Sansd,1989)):
v,:
*%vr
vu : \vr " f
From theseexpressions seethat obviously we
(c.1)
For V7 a seriesexpansionexists (seesection 3.2) so a similar expansionhas to exist f.or V6 in which the term sin1a,, has already been included. This expansioncan be found in an analogousmanner as the expansionof the potential in inclination functions. The latter expansionis derived in (Kaula, 1966) and we will only briefly show here the derivation of the seriesexpansion for V6. We start with formula (3.58)from (ibid.) which we write in the following manner
sunsin * mo,)
219
" L
.,
ms
t  m  2 t * . cI
j where : '/i,.:)"'
w
R \;i
2t)cor21ssin2 l]
where  m f (I) : rinlnr'2t1 Icos"1 I [(/ g(u") : (C,,o  jSho) si^"' h(wo) : alnlzn2t*sL '')o cos^t '!o ' Equation C.2 can be compared with formula (3.59) from (Kaula, 1966). Proceeding in the same way as (ibid.), we arrive at
sho 0r^,r gy (!\'*' : i J  . ' ' " p { , t L o_ . J , _..o , o o o n;,(1)\ / , sttmp* R \r/ ao
I  ",* 1trn:even sin,1,,*ol +  ; s'Jr*,oaa L )
where $kop:
tftu@r nL
( f
1lrn:even
(C'3)
"H,n\
220
"
/\or")'
f (I) k trnax : :
crnar : cmin :
Methods of computing of the inclination functions and their derivatives with respectto .[ can be found in e.g. (wagner, 1983), (Schrama,19g9)and (Goad, 19gz). The crosstrack inclination functions f,io(1) as derived above can be computed in a similar manner as will be shown hereafter. For the description of a recursiae algorithm for computation of the inclination functions or the crosstrack inclination functions, we refer to respectively(sneeuw, 199ra) and (sneeuw, 1991b). Following (Wagner, 1983) the method for computation of the inclination functions describedin (Schrama, 1989) makes use of a "unit potential function', developed along a great circle with inclination 1. Along this circular unit orbit (defined : b y . R : 1 r _ 9 0 c = e : 0 ( s o a l s oc , . r , ) a n d M + u : :0 uo j u) we evaluatehe t function ffi' which can be obtained by applying the chain rule of differentiation
(c.4)
where we use a unit potential function definedby GM : R =, : et^: Sho: L. Now we have to find the partial derivatives and tr'rom appendix A we have ffi ff. on the unit sphere X: sindcos,\ : sin0sin) Y Z : cos? and X' : cos Scoswo Y ' = c o s$ s i n . , . o
L
Dl
(C.b)
(C.6)
:slne.
221
appendix
C,
Inclination
functions
The transformation between the coordinate systems c/ and cI' consists only of a rotation about the Xaxis by an angle 1(since @, :0), or:
(c.7)
( c.8)
(c.e)
( c.10)
A6:
 cosI "i"0
sinl cos) ' ri"0
n' o v t mI t ; r : 4 ,
ao ?o
f tl
t lI
l s i n ( /z p  r ) u l
where the bar over F indicates that the crosstrack inclination functions are now = normalized. This seriescan be regarded as a Fouriertype series. Introducing k I2plyields
k ) p ) os ku*
sin ku ) )
222
for odd/ "! : Ff^1t_t+*112 1"0r u0o11z /  n1 eve\ for + : Fi,,"1t_t_*112 1"1 o0r11z I  rn even for : Ff*1t_r+*y1z pfr  u!112 for /  rn odd : Fi*1t_r_*112 @f + of11z for J  rn odd.
The Fourier coefficientsof and bf, are derived by computing the unit potential at discrete points along a great circle. With an FFT routine these potential values (time domain) are transformed to the coefficients and 6f (frequencydomain). afo
Ff*1,r.11, :
223
References
Arabelos, D. and C.C. Tscherning (1990), Simulation of Regional Gravity Field Recoveryfrom Satellite Gravity Gradiometer Data using Collocation and FFT, 64, Bulletin Gdodsique, 363382 Balmino, G., F. Barlier, A. Bernard, c. Bouzat, R. Rummel and P. Touboul (1985), proposal for a Satellite Gravity Gradiometer Experiment for the Geosciences, Bureau Gravim6trique International, Toulouse Balmino, G. and J.P. Barriot (1990), Methods of Global Recovery of Harmonic 520, Study on Precise Coefficientsfrom SGG in the General case,Workpackage Gravity Field Determination Methods and Mission Requirements,Final Report (Phase2), ESA Contract no. 8153/88F IFL Balmino, G., J. Barriot, R. Koop, B. Middel, N.c. Thong and M. vermeer (1991), Simulation of Gravity Gradients: a ComparisonStudy, Bulletin G{oddsique,65, 218229 Betti, B. and F. Sansd (1989), The Integrated Approach to Satellite Geodesy,in: 25, Theory of Satellite Geodesyand, Graoity Lecture Notesin Earth Sciences, R. Rummel (Eds.), Springerverlag, Berlin HeiDetermination, F. sansd, Field, delberg New York Bergmann, P.G. (1942),Introduction to the Theory of Relatiuity,Dover Publications, New York Bishop, R.L. and S.I. Goldberg (rsoa), Tensor Analysison Manifolds, Macmilla,n, New York carroll, J.J. and P.H. Savet (1959), SpaceNavigation and Exploration by Gravity Difference Detection, Aero/Space Engineering, 4447 Chandrasekhar, S. (1965), The PostNewtonian Equations of Hydrodynamics in Journal, L42, 14991512 General Relativity, The Astrophysical Colombo, O.L. (1981), Numerical Methods for Harmonic Analysis on the Sphere, report 310, Ohio State University, Columbus Department of Geodetic Science, Colombo, O.L. (1986), Not"r on the Mapping of the Gravity Field using Satellite Mathematicaland Nurnerical TechData, in: Lecture Notesin Earth Sciences,T, (Ed.), SpringerVerlag,Berlin Heidelberg H. niquesin PhysicalGeodesy, Siinkel New York
224
References
Colombo, O.L. (1987), The Global Mapping of the Gravity Field with an Orbiting FullTensor Gradiometer: an Error Analysis, presented at XIX Assembly of the IUGG, Vancouver,Canada Colombo, O.L. (1989), Advanced Techniques for HighResolution Mapping of the Gravitational Field, in: Lecture Notes in Earth Sciences,25, Theory of SatelIite Geodesyand Grauity Field Determination, F. Sansb, R. Rummel (Eds.), SpringerVerlag,Berlin HeidelbergNeu' York Dornier (1990),AristotelesAdditional Study, ExecutiveSummary Report 5, ESTEC Contract no. 8355/89/NL/JS Dziewonski, A.M. (1984), Mapping the Lower Mantle: Determination of Lateral Heterogeneityin P velocity up to Degree and order 6, Journal of Geophysical Research, 89, 87, 59295952 Emeljanov, N.V. and A.A. Kanter (1989),A Method to Compute Inclination Functions and their Derivatives,Manuscripta Geodaetica, L4,7783 Eiitviis, R. (1953), Gesarnmelte Arbeiten,Akaddmiai Kiad6, Budapest ESA (1991), Report of the Earth Observation User Consultation Meeting, ESA sP1143 Fischbach,E., D. Sudarsky,A. Szafer,C. Talmadgeand S. Aronson (1986), Reanalysis of the Ecitvcis Experiment, PhysicalReuiewLetters,56, 1, 36 Forward, R.L. (1974),Review of Artificial Satellite Gravity Gradiometer Techniques for Geodesy,in: Useof Artificial Satellites and Geodynarnics, Veis G. for Geodesy (Ed.), L577e2 Forward, R.L. (1981), Gravity Sensorsand the Principle of Equivalence, IEEE Transactions Aerospace on and ElectronicSystems, L7, 4,511519 Forward, R.L. (1982), Flattening Spacetimenear the Earth, PhysicalReuiewD,26, 4,735744 Foster, J. and J.D. Nightingale (1979), A Short Coursein GeneralRelatiuity,Longman Group, Harlow, Essex Gerstl, M. (1980),On the RecursiveComputation of the Integralsof the Associated LegendreFunctions, Manuscripta Geodaetico, 181199 5, Gill, 8., M. Soffel,H. Ruder and M. Schneider(1992), Relativistic Motion of Gyroscopesand SpaceGradiometry, DGK A107, Miinchen Goad, C.C. (1987), An EfHcient Algorithm for the Evaluation of Inclination and Eccentricity Functions, Manuscripta Geodoetica, L2, lI15 Goldstein, H. (1980), ClassicalMechanics,2nd edition, AddisonWesley,Reading, Massachusetts GRADIO (1989), AccelerometerPredevelopment, Final Report, ESTEC Contract 7631/88/NL/JS no.
225
References
Hager, B.H. (1983),Global Isostatic Geoid for Plate and Boundary Layer Models of the Lithosphere, Earth and,Planetary ScienceLetters,63, 97109 Hein, G.W. (1986), Integrated Geodesy StateoftheArt 1986 referencetext, in: in T, Lecture Notes in Earth Sciences, Mathernaticaland Numerical Techniques H. Physical Geodesy, Siinkel (Ed.), SpringerVerlag,Berlin HeidelbergNew York Freeman and Co., San Heiskanen,W.A. and H. Moritz (1967), Physical Geodesy, Francisco ESSA Monograph 2, U.S. Department of Hotine, M. (1969), Mathematical Geodesy, Commerce, Washington, D.C. Kiirper  GravitationswechIlk, K.H. (1983),Ein Beitrag zur Dynamik ausgedehnter selwirkung , DGK C288, Miinchen Ilk, K.H. (1987), Regional Gravity Field Mapping: Satellite Gravity Gradiometry versus Satellitetosatellite Tlacking Techniques, presented at XIX Assembly for the IUGG, Vancouver,Canada Ilk, K.H., R. Sigl and M. Thalhammer (IOOO),Regional Gravity Field Recovery 440, Study on PreciseGravity From Gradiometer Measurements,Workpackage Field Determination Methods and Mission Requirements,Final Report (Phase 2 ) , E S A C o n t r a c tn o . 8 1 5 3 / 8 8 l F l F L Jekeli, C. (1978),An Investigationof two Models for the DegreeVariancesof Global report 275, Ohio State CovarianceFunctions, Department of Geodetic Science, University, Columbus Jenkins, G.M. and D.G. Watts (1OOA),Spectral Analysis and, its Applications, HoldenDay, Inc., Oakland, California Akademische in Jung, K. (1961), Schwerkraftuerlahren der AngewandtenGeophysffr, Verlaggesellschaft Geest & Portig K.G, Leipzig Dynarnics E Control, John Wiley & Sons, Kaplan, M.H. (1976), Modern Spacecraft New York Blaisdell Publishing Company, Kaula, W.M. (1966), Theory of Satellite Geodesy, Waltham, Massachusetts Koop, R. and D. Stelpstra (fOaO),On the Computation of the Gravitational Potential and its First and Second Order Derivatives, Manuscripta Geodaetica,14, 373382 Koop, R., E.J.O. Schrama,R. Rummel and M. van Gelderen (1989), Gravity Field Recovery Performance, Aristoteles AddOn Study, Part 1, Dornier W.P. no.
24rolr
Koop, R. and D. Stelpstra (tOOf), Potential CoefficientRecoveryfrom the CSR Set ResearchLetters, of Simulated Satellite Gradiometry Observations,Geophysical 18, 10, 18971900
226
References
Kopejkin, S.M. (1991), Relativistic Manifestations of Gravitational Fields in Gravimetry and Geodesy,Manuscripta Geodaetica, 301312 L6, Lambeck, K. (1990), Aristoteles: An ESA Mission to study the Earth's Gravity Field  Impact of a HighResolution GravityField Mission for Planet Earth, ESA Journal, L, L4, I2L Marsh, J.G., F.J. Lerch,B.H. Putney,T.L. Felsentreger, B.V. Sanchez, S.M. Klosko, G.B. Patel, J.W. Robbins, R.G. Williamson, T.E. Engelis, W.F. Eddy, N.L. Chandler, D.S. Chinn, S. Kapoor, K.E. Rachlin, L.E. Braatz and E.C. Pavlis (1989), The GEMT2 Gravitational Model, NASA Technical Memorandum 100746 Marussi, A. (1982), On the Structure of the Tidal Field, in: Geodesyand Global Geodynamics,Lectures of the 3rd Int. Summer School on Geodesyand Global Geodynamics,H. Moritz, H. Siinkel (Eds.), Institut fiir TheoretischeGeoddsie, Graz, Austria Mashhoon, 8., H.J. Paik and C.M.will (1989), Detection of the Gravitomagnetic Field using an Orbiting Superconducting Gravity Gradiometer.TheoreticalPrinciples.,Physical Review D, Particles and Fields, 39, 10, 28252888 McConnell, A.J. (1957), Applications of Tensor Analysis, Dover Publications, New York Meissl, P. (1971),A Study of CovarianceFunctionsrelated to the Earth's Disturbing Potential, Department of Geodetic Science,report 151, Ohio State University, Columbus Minkowski, H. (1952), Spaceand Time, in: The Principle ol Relatiaity, by A. Einstein, H.A. Lorentz, H.Weyl and H. Minkowski, Dover Publications, New York Misner, C.W., K.S. Thorne and J.A. Wheeler(fOZS),Grauitation,W.H. Freeman, New York Moritz, H. (1968), Kinematical Geodesy, DGK A59, Miinchen Moritz, H. (1980), AdaancedPhysical Geod,esy, Wichman Verlag, Karlsruhe H. Moon, P. and D.E. Spencer(tOA0), Theory of Holors. A Generalizationof Tensors, Cambridge University Press, Cambridge NASA (1987), Geophysicaland Geodetic Requirementsfor Global Gravity Field Measurements19872000,Report of a Gravity workshop, Colorado Springs Neyman, Y.M. (1985),Improperly posedproblems in geodesyand methods of their solution, in: Local Grauity Field Approximation,proceedings the Beijing Interof national Summer School,Ed. K.P..Schwarz,Publication 60003,Calgary, Canada Ohanian, H. (1976), Grauitation and Spacetime,W.W. Norton & Company, New York O'Neill, B. (1966), ElernentaryDifferential Geometry,Academic Press,New York
References Oppenheim, A.V., A.S. Willsky and I.T. Young (1983), Signals and Sgstems, PrenticeHall International, London Paik, H.J. and JP. Richard (1986), Developmentof a SensitiveSuperconducting Gravity Gradiometer for Geologicaland Navigational Application, NASA Contractor Report 4011 Paik, H.J. (1989), Tests of GeneralRelativity in Earth Orbit using a SuperconductRes.,9, no. 9, 4150 ing Gravity Gradiometer,Adu. Space Papoulis, A. (1965), Probability, Random Variables and Stochastic Processes, McGrawHill Kogakusha,Tokyo ed., Pond, S. and G.L. Pickard (1983),IntroductoryDynamical Oceanography,2nd Oxford Pergamon Press, Rupp, R.H. (1977), Potential Coefficient Determinations from 5o Terrestrial Grareport 251, Ohio State University, vity Data, Department of Geodetic Science, Columbus Rapp, R.H. (1981),The Earth's Gravity Field to Degreeand Order 180 using Seasat Altimeter Data, TerrestrialGravity Data, and other Data, Department of Geodetic Science,report 322, Ohio State University, Columbus Rapp, R.H. and J.Y. Cruz (1986), SphericalHarmonic Expansionof the Earth's Gravitational Potential to Degree 360 using 30' Mean Anomalies, Department report 376, Ohio State University, Columbus of Geodetic Science, to Rapp, R.H. (1989), Signalsand accuracies be expectedfrom a satellite gradiomeL4, ter mission, Manuscripta Geodaetica, 3642 Reed, G.B. (1973), Application of Kinematical Geodesyfor Determining the Short Wave Length Components of the Gravity Field by Satellite Gradiometry, Dereport 201, Ohio State University, Columbus partment of Geodetic Science, Ries, J.C., C. Huang, M.M. Watkins and B.D. Tapley (1990),The Effectsof General Relativity on NearEarth Satellites, in: Astrodynamics1989; Proceedings held August 710, 1989, Stowe, of the AAS/AIAA AstrodynamicsConference, 71, Sciences, 8593 in Vermont, Aduances the Astronautical Robbins, J.W. (1935),Least SquaresCollocation applied to Local Gravimetric Solutions from Satellite Gravity GradiometerData, Department of Geodetic Science, report 368, Ohio State University, Columbus and M. Gerstl (1979), Least SquaresCollocation and Rummel, R., K.P. Schwa.rz Regularization,Bulletin G6odisique,53, 343361 Rummel, R. (1985a),From the ObservationalModel to Gravity Parameter Estimaof tion, in: Local Grauity Field Approzimation, proceedings the Beijing International Summer School,Ed. K.P. Schwarz,Publication 60003,Calgary, Canada Rummel, R. (1985b), Satellite Gradiometry  a Promising Concept since 25 Years, presentedat Spring Meeting of AGU, Baltimore, Maryland
22E
References
Rummel, R. (1985c),Satellitengradiometrie, ZFV, G, 242257 Rummel, R. and O.L. Colombo (1OSS), Gravity Field Determination from Satellite Gradiometry, Bulletin G{odisique, 59, 2Bg_24O Rummel, R. (1986), Satellite Gradiometry,in Lecture Notesin Earth Sciences,T, Mathernaticaland,Numerical Techniques PhysicalGeodesg, Stinkel (Ed.), in H. SpringerVerlag,Berlin HeidelbergNew York Rummel, R. (1989a), Aristoteles: surface Gravity from space Geometry, in: proceedingsof the Italian Workshop on the European SolidEarth Mission Aristoteles, Aeritalia, Tlevi Rummel, R. (1989b),SGG Principlesand state of the Art, workpackage130, study on Precise Gravity Field Determination Methods and Mission Requirements, Final Report, ESA Contract no.752If 87lF IFL Rummel, R., P. Teunissenand M. van Gelderen (1989), Uniquely and overdetermined Geodetic Boundary Value Problems by Least Squares,Bulletin G6odisiquer 63, l33 Rummel, R. (1991), Physical Geodesy,Lecture Notes, Delft University of Technology, Faculty of Geodetic Engineering,Delft Rummel, R. and M. van Gelderen (1992), Spectral Analysis of the Full Gravity Tensor, Geophysical Journal International, LLL, lbg169 savet, P.H. (1969), Gravity Field Exploration by a new Gradient Technique, J. Spacecraft,6, 710716 Schouten, J.A. (1954),Ricci Calculus,2nded., SpringerVerlag, Berlin Schrama, E.J.O. (1984), Orbit Integration based upon Interpolated Gravitational Gradients, graduate thesis, Delft University of Technology,Faculty of Geodetic Engineering,Delft Schrama,E.J.O. (1989),The Role of Orbit Errors in Processing SatelliteAltimeter of Data, NetherlandsGeodetic Commission,Publications on Geodesy,New Series, 33 Schrama,E.J.O. (1990),Gravity Field Error Analysis: Application of GPS Receivers and Gradiometers on Low Orbiting Platforms, NASA TechnicalMemorandum 100769 Schrama, E.J.O. (1992), Relativistic effectson motion of nearEarth satellites,internal report schutz, B.8., B.D. Tapley, J.B. Lundberg and P. Halamek (1982), simulation of a Geopotential Mission for Gravity studies, Manuscripta Geodaetica, L2, s168 schutz, B.E., J.B. Lundberg, L.K. white and P.G. Antreasian (1988), Developments in the Simulation of a GeopotentialResearch Mission, proceedings the of AIAA/AAS Astrodynamics Conference, Minneapolis,Minnesota
229
References Schwarz, K.P. (1985), Data types and their spectral properties, in: Local Gravity of Approrirnation, proceedings the Beijing International summer school, Fietd. Ed. K.P. Schwarz,Publication 60003,Calgary, Canada of SESAME (1986), Proceedings an ESA Special Workshop on Solid Earth Science for Europe, ESA SP1080 & Application Mission sneeuw, N.J. (199la), Inclination Functions. Group theoretical Background and a Recursive Algoriihm, Faculty of Geodetic Engineering lDept. of Mathematical report 91.2, Delft University of Technology and Physical Geodesy, Sneeuw,N. (1gglb), Nonsingular Crosstrack Derivativesof the Gravitational Potential using Rotated spherical Harmonics, proceedingsof the XX General Assembly of the IUGG, IAG Symposium G3, Vienna Soffel, M., E. Gill, H. Ruder and M. Schneider (1987), Relativistic Gradiometry, proceedingsof the International Associationof GeodesySymposia, Vancouver, Canada, August lO22, 1987 Soffel, M.H. (1989), Relativity in Astrometry, Celestial Mechanics and Geodesy, SpringerVerlag,Berlin New York Heidelberg soffel, M.H. (1990),Lectureson PostNewtonianTheory, Editorial de la universidad de Costa Rica, San Jos6 sokolnikoff, LS. (1951), Tensor Analysis.Theory and Applications.,John wiley & Sons, New York Spakman, w. (1988), Upper Mantle Delay Time Tomography, Geologica ultraiectina,53, Utrecht Stelpstra, D. (1gg0), Gravity Recoveryfrom Satellite Gradiometry, graduate thesis, Delft university of Technology,Faculty of Geodetic Engineering, Delft Tostewart, R., L.L. Fu and M. Lefebvre (roao), Scienceopportunities from the california pex/ PoseidonMission, JPL Publication 8618, NASA, Pasadena, synge, J.L. (1960), Relativity: The General Theory, NorthHolland Publ. comp., Amsterdam in Tapley, B.D. (1989), Fundamentalsof orbit Determination, in: Lecture Notes and Grauity Field Determination, Earth sciences,25,Theory of satettiteGeodesy F. Sansb,R. Rummel (Eds.), springerverlag, Berlin HeidelbergNew York fiir Theig, D.S. (1984), Neue gravitative Effekte rotierenderMassen.Moglichkeiten InauguralDissertation,Uniweitere Tests der Allgemeinen Relativititstheorie, versitlt Koln, Ktiln Thirring, H. and J. Lense (1918), Uber den Einfluss der Eigenrotation der zentralkcirper auf die Bewegungder Planeten und Monde nach der Einsteinschen Gravitationstheorie,Phys. 2., 19, 156163 John Tikhonov, A.N. and v.Y. Arsenin (1963), Solutionsof lllposed Problerns, Wiley & Sons, New York
230
References Touboul, P., A.Bernard, F. Barlier and C. Berger (lg9l), Air Drag Effect on Gradiometer Measurements,Manuscripta Geodaetica, 7391 16, Tscherning,C.C.and R.H. Rapp (t97a), Closed Covariance Expressions Grafor vity Anomalies, Geoid Undulations and Deflectionsof the Vertical implied by Anomaly Degree Variances,Department of Geodetic Science, report 208, Ohio State University, Columbus Tscherning, C.C. (1976), Comparison of the SecondOrderderivativesof the Normal Potential Based on the Representation by a Legendre series, Manuscripta Geodaetica, 2,7192 Tscherning, C.C., R. Forsbergand M. Vermeer (1990), Methods for Regional Gravity Field Modelling from SST and SGG Data, Report of the Finnish Geodetic Institute 9O:2, Helsinki Vermeer,M. (1991), Geoid Recoveryat 0.5 DegreeResolutionfrom Global Satellite Gradiometry Data Sets, in: Determination of the Geoid. Present and Future, IAG Symposium No. 106,Eds. R.H. Rapp, F. Sansd,SpringerVerlag,New York Berlin Heidelberg Visser, P.N.A.M. (1992), The Use of Satellitesin Gravity Field Determination and Model Adjustment, PhD Thesis, Delft University Press,Delft Wagner, C.A. (1983), Direct Determination of Gravitational Harmonics from lowlow GRAVSAT Data, Journal of Geophysical Researcl,88, 309321 Wagner, C.A. (1989), Summer School Lectures on Satellite Altimetry, in: Lecture Notes in Earth Sciences,25, Theory ol Satellite Geodesy and Grauity Field Determination,F. Sansd,R. Rummel (Eds.), SpringerVerlag, Berlin Heidelberg Wichiencharoen, (1985) Recoveryof l"mean Anomaliesin a Local Region from C. , a LowLow Satellite to Satellite tacking Mission, Department of Geodetic Science,report 363, Ohio State University, Columbus Will, C. (1981), Theorg and Erperiment in Grauitational Physics,Cambridge University Press, Cambridge Woodhouse,J.H. and A.M. Dziewonski(1984),Mapping the Upper Mantle: ThreeDimensional Modelling of Earth Structure by Inversion of Seismic Waveforms, Journal of Geophysical Researci., 89, 87, 59535986 Wunsch, C. (1992), Physicsof the Ocean Circulation, Lecture Notes of the International Summerschool Theoretical Geodesyon "satellite Altimetry in Geodesy of and Oceanography",Tlieste Xu, P.L. (1991), Least SquaresCollocation with Incorrect Prior Information, ZFV, LL6,266273 Xu, P.L. and R. Rummel (1992),The Value of Minimum Norm Estimation of Geopotential Fields, Geophysical Journal International, lII, 170178
231
Viel mehr als nur Dokumente.
Entdecken, was Scribd alles zu bieten hat, inklusive Bücher und Hörbücher von großen Verlagen.
Jederzeit kündbar.