06.02.2015 Views

s - Athanasios Dermanis Home Page

s - Athanasios Dermanis Home Page

s - Athanasios Dermanis Home Page

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

First International Symposium<br />

on the Mathematical and Physical Foundations of Theoretical Geodesy<br />

September 7th-9th, 1993, Stuttgart, Germany<br />

Quadratic Collocation and Robust Quadratic Collocation<br />

<strong>Athanasios</strong> <strong>Dermanis</strong><br />

Department of Geodesy and Surveying<br />

The Aristotle University of Thessaloniki<br />

1. Introduction<br />

Estimation and prediction in geodesy is typically limited<br />

to seeking optimal (minimum mean square error) estimators<br />

and/or predictors in the class of linear functions of the parameters.<br />

This can be justified in view of the well known<br />

fundamental result (see e.g. Rao, 1973) that when the random<br />

parameters of the model follow a Gaussian distribution,<br />

the linear optimal estimators and predictors cannot be improved,<br />

i.e. they are indeed the optimal non-linear ones.<br />

Since Gaussianity is rather a mathematical convenience that<br />

an absolute observational fact, it seems worthwhile to seek<br />

non-linear estimators and predictors which minimize the<br />

mean square error, without any restrictive assumptions on<br />

the distribution of the relevant random variables.<br />

The idea of using nonlinear predictors goes back to Grafarend<br />

(1972). <strong>Dermanis</strong> and Sanso (1993) have studied<br />

some basic characteristics of non linear estimators and came<br />

to the conclusion that estimation of deterministic model<br />

parameters is feasible only within a Bayesian point of view.<br />

There is no need to seek rescue in Bayesian ideas in the case<br />

of random effects models which involve no deterministic<br />

but only random parameters. Prediction of random parameters,<br />

either directly present in the model or stochastically<br />

related ones, is of much geodetic interest especially for the<br />

determination of the gravity field of the earth conceived as a<br />

stochastic process in the framework of the statistical approach<br />

of collocation (Moritz, 1980, Sanso, 1986).<br />

Among the various types of non-linear predictors the simplest<br />

and perhaps the only practically tractable ones are the<br />

quadratic ones. Quadratic estimation is widely used for the<br />

determination of variance components (e.g. Schaffrin, 1983)<br />

and rarely for combined variance component and parameter<br />

estimation (.XEiþHN <br />

<strong>Dermanis</strong> and Sanso (1993) have derived Bayesian quadratic<br />

estimators which can be directly applied to random<br />

effect models such as the models used in geodetic collocation.<br />

Here optimal quadratic predictors will be derived directly<br />

for the simplest collocation model and then extended<br />

to non-linear ones. As expected the results extend the usual<br />

(linear) collocation equations in a way which involves the<br />

effect of third and fourth order central moments which describe<br />

the deviation of the distribution of the random parameters<br />

from the Gaussian assumption.<br />

2. The simple collocation model<br />

We start with the simplest possible collocation model of<br />

direct observation of random signals s under additive zeromean<br />

observational errors v<br />

y = s+<br />

v , E { s } = m , E { v } = 0 . (1)<br />

We assume that s and v are stochastically independent, but<br />

no further assumptions are made about their probability distributions.<br />

The objective is to seek optimal predictors for any single<br />

random variable s′ , which is stochastically related to s ,<br />

within the class of quadratic functions of the observations y,<br />

i.e. of the form<br />

s ˆ ′ = γ + dT<br />

y + y<br />

T<br />

Qy (inhomogeneous) (2a)<br />

or<br />

s ˆ ′= dT<br />

y + y<br />

T<br />

Qy (homogeneous) (2b)<br />

We shall call the predictor (2a) inhomogeneous and the<br />

predictor (2b) homogeneous, not in respect to the property<br />

of homogeneity itself (of course both predictors are inhomogeneous<br />

functions of y ) but in order to establish a correspondence<br />

with the terminology used in the linear case obtained<br />

by letting Q= 0 . These two predictors can be combined<br />

with the possibility of introducing the property of unbiasedness<br />

or not, in order to produce four basic classes of<br />

best quadratic predictors:<br />

Best inhomogeneous Quadratic Prediction (inhomBQP)<br />

Best homogeneous Quadratic Prediction (homBQP)<br />

Best inhomogeneous Quadratic Unbiased Prediction<br />

(inhomBQUP)<br />

Best homogeneous Quadratic Unbiased Prediction<br />

(homBQUP)<br />

Best (optimal) prediction here refers to the minimization<br />

of the mean square prediction error :<br />

φ = E {( sˆ<br />

′−s′)<br />

2<br />

} = E{<br />

ε<br />

2}<br />

= min<br />

(3)<br />

where ε = s ˆ′−s′<br />

is the prediction error. The optional prop-


erty of unbiased prediction is<br />

β = E { sˆ<br />

′− s′<br />

} = E{<br />

ε}<br />

= 0 . (4)<br />

Since s′ ˆ is quadratic in y , ε<br />

2<br />

will be of the 4th order in<br />

y , and consequently φ will depend on up to 4th order<br />

(central) moments of s , v , and cross-moments of s′ and<br />

s . To describe these moments in matrix notation we introduce<br />

the following basic definitions :<br />

m= E{s} , δ s=<br />

s−m<br />

, δ y = y −m=<br />

δs+<br />

v (5)<br />

m ′ = E{s′<br />

} , δ s ′= s′−m′<br />

(6)<br />

C s<br />

= E{ δsδsT<br />

} , C E{ vvT<br />

v = }, (7)<br />

C = E{ δ yδy<br />

T<br />

}= C s + C v , (8)<br />

δ ′ ′ , (9)<br />

c<br />

T<br />

s s ′ = E{ sδs<br />

} = E{<br />

δyδs<br />

} = c s ′ s<br />

σ<br />

2<br />

= E{( s )<br />

2}<br />

, (10)<br />

′<br />

s ′<br />

δ<br />

⎡-1<br />

⎤<br />

- = E y y y<br />

⎢ ⎥<br />

{ vec(<br />

δ δ<br />

T<br />

) δ<br />

T<br />

} = = - s + -<br />

⎢<br />

<br />

⎥<br />

v , (11)<br />

⎢⎣<br />

- ⎥<br />

n ⎦<br />

- E{ s y y<br />

T<br />

} E{<br />

s s sT<br />

′ s = δ ′ δ δ = δ ′ δ δ } , (12)<br />

<br />

s<br />

= E{ vec(<br />

δyδy<br />

T<br />

) vec T ( δyδy<br />

T<br />

)}=<br />

⎡<br />

=<br />

⎢<br />

⎢<br />

⎢⎣<br />

3. Solution<br />

11<br />

<br />

n1<br />

<br />

<br />

<br />

1n<br />

<br />

nn<br />

⎤<br />

⎥<br />

⎥<br />

=<br />

⎥⎦<br />

+<br />

s v<br />

(13)<br />

For the determination of the optimal quadratic predictors it<br />

is more convenient to express (2a) in the equivalent form<br />

( y = m+<br />

δy<br />

)<br />

sˆ<br />

′ = γ + ( d+<br />

2Qm)<br />

T<br />

y + vecT<br />

Q vec(<br />

δyδy<br />

T<br />

−mm<br />

T<br />

) =<br />

C<br />

c<br />

z<br />

zs′<br />

⎡C<br />

-T<br />

⎤<br />

= E {( z −m<br />

z )( z −m<br />

z )<br />

T<br />

} = ⎢<br />

⎥ (17)<br />

⎣-<br />

−vecC<br />

vecT<br />

C⎦<br />

⎡ css′<br />

⎤<br />

= E {( z −m<br />

z )( s′−m′<br />

)} = ⎢ ⎥<br />

(18)<br />

⎣vec-<br />

s′<br />

s ⎦<br />

while for the homogeneous case (2b) becomes s ˆ ′ =a T z .<br />

With the above "vectorized" formulation it is easy to calculate<br />

the bias and the mean square error<br />

β γ + a m −m<br />

(19)<br />

=<br />

T<br />

z<br />

′<br />

φ a C a 2a<br />

c . (20)<br />

= β 2 + T<br />

T<br />

z + σ<br />

s<br />

2<br />

′ − zs′<br />

The problem of prediction has now the same structure as<br />

the linear problem with well known solution (Schaffrin,<br />

1985a). A simple derivation for the various types of predictors<br />

is given in appendix A, following <strong>Dermanis</strong> (1991). It<br />

has the general form<br />

s ˆ′ = α m′+<br />

c C<br />

− 1 ( z−α<br />

m )<br />

(21)<br />

T<br />

z s′<br />

z<br />

with (minimum) mean square error<br />

z<br />

φ σ<br />

2 −c C<br />

−1c<br />

+ δφ<br />

(22)<br />

=<br />

T<br />

s ′ z s ′ z zs′<br />

where α and δφ are different for different predictors:<br />

inhomBQP = inhomBQUP:<br />

α =1 , δφ = 0 , β = 0 . (23)<br />

homBQP:<br />

m C z<br />

α =<br />

, (24a)<br />

m<br />

T −1<br />

z z<br />

1+<br />

mT<br />

z C−<br />

1<br />

z<br />

T<br />

z s ′<br />

T<br />

z<br />

z<br />

( m′−c<br />

C m<br />

2<br />

z )<br />

δφ =<br />

, (24b)<br />

1+<br />

m C m<br />

T −1<br />

z s ′ z z<br />

1+<br />

mT<br />

z C−<br />

1<br />

z<br />

−1<br />

z<br />

−1<br />

z<br />

z<br />

z<br />

c C m −m′<br />

β =<br />

. (24c)<br />

m<br />

⎡ y<br />

= γ + [ dT<br />

+ 2 mT<br />

Q vecT<br />

Q]<br />

⎢<br />

⎣vec(<br />

δyδy<br />

T<br />

−mm<br />

T<br />

⎤<br />

⎥<br />

)<br />

=<br />

⎦<br />

=γ +a T z<br />

(14)<br />

homBQUP:<br />

m C z<br />

α ,<br />

T −1<br />

z z<br />

=<br />

m T<br />

z C<br />

−1<br />

z m z<br />

( m′−c<br />

C m<br />

δφ =<br />

, β = 0 . (25)<br />

m<br />

T −1 z z z ) 2<br />

s ′<br />

T<br />

z C<br />

−1<br />

z m z<br />

where a is a new unknown vector,<br />

⎡ y<br />

z = ⎢<br />

T<br />

⎣vec<br />

δyδy<br />

−mm<br />

(<br />

T<br />

⎤<br />

⎥ , (15)<br />

) ⎦<br />

⎡ m ⎤<br />

m z = E{<br />

z}<br />

= ⎢<br />

⎥ , (16)<br />

⎣vec(<br />

C−mm<br />

T<br />

) ⎦<br />

In order to obtain expicit equations we only need to replace<br />

z from (15), m from (16), C z from (17) and c z s′<br />

from (18). Also the following identity needs to be used<br />

C−<br />

1 z<br />

⎡C<br />

= ⎢<br />

⎣-<br />

<br />

-T<br />

−vecCvec<br />

T<br />

⎤<br />

⎥<br />

C⎦<br />

−1<br />

=<br />

2


where<br />

⎡C<br />

= ⎢<br />

⎣<br />

−1<br />

+ C<br />

−1-T<br />

N<br />

−N<br />

−1<br />

-&<br />

−1<br />

−1<br />

-&<br />

-&<br />

−1<br />

−C<br />

−1-T<br />

N<br />

−1<br />

N<br />

−1<br />

⎤<br />

⎥<br />

⎦<br />

(26)<br />

N= −<br />

C C−<br />

−1<br />

T<br />

. (27)<br />

vec vec T -<br />

The general solution has the form<br />

ˆ T T 1<br />

s s ′<br />

s s ′<br />

s ′ = α m′+<br />

c C<br />

−1(<br />

s−α<br />

m)<br />

+ n N<br />

−<br />

[ q(<br />

y)<br />

−α<br />

m ] (28)<br />

with<br />

n<br />

-&<br />

=<br />

1<br />

, (29)<br />

−<br />

ss′ css′<br />

−vec - s′<br />

s<br />

q = q(<br />

y)<br />

=<br />

m q<br />

-&<br />

-&<br />

−1 T T<br />

y − vec[(<br />

y − m)(<br />

y − m)<br />

q<br />

− mm<br />

] , (30)<br />

=<br />

− 1 m−vec(<br />

C−mm<br />

T ) = E{<br />

q(<br />

y )} . (31)<br />

The mean square error of the prediction is<br />

φ σ<br />

2<br />

−c C−1c<br />

−n<br />

N<br />

−1n<br />

+ δφ . (32)<br />

=<br />

T T<br />

s ′ ss<br />

′ ss<br />

′ ss<br />

′ s s′<br />

For the particular prediction types we have<br />

inhomBQP = inhomBQUP:<br />

α =1 , δφ = 0 , β = 0 . (33)<br />

homBQP:<br />

mT<br />

C<br />

−1y<br />

+ mT<br />

q N<br />

−1q<br />

α =<br />

, (34a)<br />

+ + m<br />

1 mT<br />

C<br />

−1m<br />

mT<br />

q N<br />

−1<br />

( m′−cT s ′<br />

C<br />

δφ = s<br />

1+<br />

mT<br />

C−<br />

c<br />

β =<br />

C<br />

−1<br />

m−n<br />

m+<br />

m<br />

1<br />

m−n<br />

N<br />

T<br />

s s ′<br />

T<br />

q<br />

N<br />

N<br />

T −1<br />

T −1<br />

s s ′<br />

s s ′<br />

1+<br />

mT<br />

C−<br />

1m+<br />

mT<br />

q N<br />

−1<br />

homBQUP:<br />

T<br />

q<br />

−1<br />

q<br />

−1<br />

m<br />

q)<br />

m<br />

q−m′<br />

q<br />

q<br />

q<br />

2<br />

(34b)<br />

(34c)<br />

mT<br />

C<br />

−1y<br />

+ mT<br />

q N<br />

−1q<br />

α =<br />

, (35a)<br />

mT<br />

C−<br />

1m+<br />

m N<br />

−1m<br />

( m′−<br />

−<br />

δφ =<br />

β = 0 . (35b)<br />

m<br />

cT s<br />

C−1 m nT s<br />

N<br />

−1q<br />

2<br />

s ′<br />

s ′<br />

)<br />

T<br />

C<br />

−1m<br />

+ mT<br />

q N<br />

−1m<br />

q<br />

The values from equations (33) are a quadratic extension<br />

of what is usually described as "best linear prediction" and<br />

is commonly called "collocation" in the geodetic literature.<br />

The other solutions offer some alternatives which are "robust"<br />

with respect to incorrect (prior) information about the<br />

signal mean m . The homBQUP solution in particular given<br />

by equations (35) is a quadratic extension of what has been<br />

called "robust collocation" by Schaffrin (1985b). Note that<br />

the quadratic predictors differ from the linear ones in (a) the<br />

additional third tern in equation (28) and (b) in the additional<br />

terms appearing in the definitions of α , β and δφ<br />

(terms where the matrix N<br />

−1<br />

appears).<br />

4. Vector generalization<br />

For the prediction of a vector of new signals s′ we can<br />

predict each component s′<br />

i independently and combine the<br />

solutions ŝ′<br />

i in a single matrix equation for s′ ˆ . We also<br />

need to introduce<br />

m s ′ = E {s′}<br />

, (36)<br />

C E{( s m )( s m )<br />

T<br />

′ = ′− ′ − } , (37)<br />

s s<br />

s<br />

s<br />

= E{( s′−m<br />

)( s′<br />

m )<br />

T<br />

}<br />

(38)<br />

C s′ s′<br />

− s′<br />

= E{ vec[(<br />

s−m<br />

)( s−m<br />

)<br />

T<br />

]( s′<br />

m )<br />

T<br />

}, (39)<br />

- s′ s<br />

s s − s′<br />

-&<br />

=<br />

1<br />

(40)<br />

N<br />

−<br />

s′ s Cs′<br />

s −-s′<br />

s<br />

(using the notation m s = E{s}<br />

instead of m for the sake of<br />

distinction), as well as the bias vector and the mean<br />

square error matrix M<br />

<br />

= E { sˆ<br />

′−s′}<br />

= E{<br />

0}<br />

, ( 0 =ˆ s′−s′<br />

) (41)<br />

M = E{( sˆ<br />

′−s′)(ˆ<br />

s′−s′<br />

)<br />

T<br />

} = E{<br />

00T<br />

}. (42)<br />

With the above notation the general solution becomes<br />

−1<br />

−1<br />

sˆ ′<br />

s s s<br />

s s s<br />

= α m ′ + C ′ C ( s −α<br />

m ) + N ′ N [ q(<br />

y)<br />

−α<br />

m ] (43)<br />

with α and q (y)<br />

as before and mean square error matrix<br />

M<br />

C<br />

−CT C<br />

−1C<br />

−NT<br />

N<br />

−1N<br />

+ δM<br />

. (44)<br />

= s′<br />

s′<br />

s s′<br />

s s′<br />

s s′<br />

s<br />

For the particular prediction types we have<br />

inhomBQP = inhomBQUP:<br />

α =1 , δ M = 0 , = 0 . (45)<br />

homBQP:<br />

T<br />

s<br />

T<br />

s<br />

−1<br />

m C y + m<br />

α =<br />

−1<br />

1+<br />

m C m + m<br />

s<br />

T −1<br />

qN<br />

q<br />

T −1<br />

qN<br />

m<br />

q<br />

α<br />

≡<br />

α<br />

1 −1<br />

−1<br />

= ( Cs′ sC<br />

m s + N s′<br />

s N q − m s′<br />

α D<br />

N<br />

D<br />

)<br />

q<br />

(46a)<br />

(46b)<br />

δ M = α T<br />

D <br />

(46c)<br />

homBQUP:<br />

T<br />

s<br />

T<br />

s<br />

−1<br />

s<br />

T<br />

q<br />

T<br />

q<br />

m C y + m N<br />

α =<br />

−1<br />

m C m + m N<br />

−1<br />

−1<br />

q<br />

m<br />

q<br />

α<br />

≡<br />

α<br />

N<br />

D<br />

, = 0 . (47a)<br />

3


1 T<br />

δ M = m 0m<br />

0 , α<br />

D<br />

−1<br />

−1<br />

m 0 = Cs′ sC<br />

m s + N s′<br />

sN<br />

q − m s′<br />

-&<br />

(47b)<br />

Recall that C = C s + C v and N=<br />

−vecCvec T C−<br />

−1-T<br />

where also - = - s + - v and = s + v .<br />

5. Extension to the general model<br />

A more general model involves observables a , which are<br />

non-linear functions a (s)<br />

of the signals s , for which up to<br />

fourth order central moments are known. In this case the<br />

general non-linear collocation model with additive noise is<br />

y = a( s)<br />

+ v<br />

(48)<br />

Furthermore, we want to predict a stochastic signal g<br />

which is also a nonlinear function g (s)<br />

of the same fundamental<br />

signals s .<br />

The solution is straightforward if the marginal and joint<br />

central moments (up to 4th order) are known for the random<br />

parameters a and g : we merely need to replace s and s′<br />

and their moments, by a and g and their moments, respectively.<br />

In this way the problem reduces to one of "moment<br />

propagation", i.e. to the determination of m a , C a ,<br />

- a , a , m g , c g a and - g a from the known m= m s , C s ,<br />

- s , s . If the joint distribution of s and s′ were known,<br />

the joint distribution of a and g , and therefore the required<br />

moments, could be derived in principle though tedious numerical<br />

techniques need to be used. However this requirement<br />

departs from the spirit of a "fourth order theory" and<br />

an approximate propagation law should be used instead.<br />

Linearization of a (s)<br />

and g (s)<br />

provides such an approximation,<br />

as in the case of linearized models and "second order<br />

theory". A better and hopefully sufficient approximation<br />

can be provided by propagation laws in quadratic approximations,<br />

which are derived from quadratic approximations<br />

to the nonlinear functions a (s)<br />

and g (s)<br />

. Using Taylor expansions<br />

up to the second order about the known (Taylor<br />

point) m= E{s}<br />

we obtain<br />

a= a( m)<br />

+ Aδ s+<br />

H vec(<br />

s sT<br />

a δ δ ) , (49)<br />

g = g( m)<br />

+ gT<br />

δ s+<br />

h vec(<br />

δsδsT<br />

) , (50)<br />

where δ s=<br />

s−m<br />

and<br />

T g<br />

h g = vecH g ,<br />

H<br />

g<br />

⎡<br />

T<br />

1 ∂ ⎛ ∂g<br />

⎞ ⎤<br />

= ⎢ ⎜ ⎟ ⎥<br />

⎢2<br />

∂s<br />

s<br />

⎣ ⎝ ∂ ⎠ ⎥⎦<br />

m<br />

, (53)<br />

d= −H a vecC s , c= −h vecC s . (54)<br />

Using these relations in the definitions of m a , C a , - a ,<br />

a , m g , c ag<br />

and - g a , while making use of the definitions<br />

of m , C s , - s , s , the following propagation laws are<br />

derived<br />

m<br />

C<br />

a<br />

a<br />

= a( m)+<br />

H vecC<br />

(55)<br />

a<br />

= AC AT<br />

+ A-T<br />

H<br />

T<br />

+ H - AT<br />

+<br />

s<br />

s<br />

s<br />

a<br />

a<br />

+ H ( −vecC<br />

vec T C ) H<br />

(56)<br />

a<br />

s<br />

s<br />

s<br />

s<br />

- = ⊗ + ⊗ + + ⊗<br />

s<br />

a d Ca<br />

( A A)[<br />

-s<br />

AT<br />

sH<br />

T<br />

a ] ( A H a ) vec<br />

a<br />

m<br />

c<br />

s<br />

s<br />

T<br />

a<br />

+ vec( +<br />

T T<br />

(57)<br />

C a dd ) d<br />

= ( A⊗A)<br />

( A⊗A)<br />

T<br />

−dT<br />

⊗(<br />

C ⊗d)<br />

−<br />

−(<br />

d⊗d)(<br />

d⊗d<br />

s<br />

) T<br />

+ -<br />

a<br />

s<br />

⊗d<br />

T<br />

s<br />

T g<br />

+ -<br />

a<br />

s<br />

s<br />

T<br />

a<br />

T<br />

T<br />

a C a<br />

⊗d−<br />

− vec C ( d⊗d)<br />

−(<br />

d⊗d)<br />

vec<br />

(58)<br />

= g( m)<br />

h vec<br />

(59)<br />

g +<br />

ag<br />

vec<br />

where<br />

<br />

s<br />

vec<br />

T g<br />

C s<br />

= AC + +<br />

T<br />

sg<br />

H a-<br />

sg<br />

A-s<br />

h g +<br />

+ H ( −vecC<br />

vec T C h<br />

(60)<br />

a s s s )<br />

s<br />

- ga<br />

ag<br />

( s s g<br />

= d⊗c<br />

+ A⊗A)[<br />

- g+<br />

h ] +<br />

s<br />

g<br />

+ ( A⊗H<br />

) <br />

s<br />

g+<br />

c vec(<br />

C ddT<br />

)<br />

(61)<br />

⎡vec<br />

⎢<br />

= ⎢ <br />

⎢<br />

⎣vec<br />

a vec<br />

a +<br />

s<br />

11<br />

s<br />

n1<br />

<br />

<br />

<br />

vec<br />

<br />

vec<br />

s<br />

1<br />

n<br />

<br />

s<br />

nn<br />

⎤<br />

⎥<br />

⎥<br />

⎥<br />

⎦<br />

+<br />

(62)<br />

while use of the "symmetric" Kronecker matrix product has<br />

been made, defined for any two matrices A and B by<br />

A⊗ s B=<br />

A⊗B+<br />

B⊗A<br />

(63)<br />

⎡∂a<br />

⎤<br />

A= ⎢<br />

s<br />

⎥ ,<br />

⎣ ∂ ⎦<br />

H<br />

m<br />

⎡1<br />

∂ ⎛ ∂ai<br />

⎞<br />

= ⎢ ⎜ ⎟<br />

⎢ ∂s<br />

s<br />

⎣ ⎝ ∂ ⎠<br />

ai 2<br />

T<br />

⎤<br />

⎥<br />

⎥<br />

⎦<br />

m<br />

,<br />

g<br />

T ⎡∂g<br />

⎤<br />

= ⎢<br />

s<br />

⎥<br />

⎣ ∂<br />

, (51)<br />

⎦<br />

H a<br />

m<br />

⎡vecT<br />

H<br />

⎢<br />

= ⎢ <br />

⎢ T<br />

⎣<br />

vec H<br />

a1<br />

a n<br />

⎤<br />

⎥<br />

⎥ , (52)<br />

⎥<br />

⎦<br />

The above propagation laws can be interpreted either as<br />

quadratic approximations of the true propagation laws for<br />

the original non-linear model y = a( s)<br />

+ v , g (s), or as exact<br />

propagation laws for the "quadraticized" model<br />

y = a( m)<br />

+ Aδ s+<br />

H a vec(<br />

δsδsT<br />

) + v , (64)<br />

g = g( m)<br />

+ gT<br />

δ s+<br />

h vec(<br />

δsδsT<br />

) . (65)<br />

T g<br />

4


6. Discusion<br />

The question that remains open is whether the newly derived<br />

predictors are of any practical significance. The answer<br />

depends of course on the distribution of the signals<br />

involved i.e. in whether this distribution departs from the<br />

Gaussian one. If the distribution is Gaussian linear predictors<br />

cannot be improved and the additional terms in the<br />

quadratic predictors will have to vanish. Within the framework<br />

of the "fourth order" approach followed here, the distributions<br />

themselves play no role and the question of departure<br />

from the Gaussian distribution can be confined in<br />

two specific points: (a) the existence of non-vanishing third<br />

order central moments and (b) the departure of fourth order<br />

central moments from the corresponding values in the Gaussian<br />

case which are known functions of the lower order<br />

moments.<br />

A practical way to evaluate the actual effect of such departures,<br />

when they exist, is to evaluate in each case the mean<br />

square errors φ L and φ Q from the linear predictors, respectively,<br />

and to look whether the relative improvement<br />

φ L<br />

−φ Q<br />

is significant or not.<br />

φL<br />

Turning to problems related to the gravity field, the question<br />

is directly related to the stochastic characteristics of the<br />

underlying stochastic process, namely the disturbing potential<br />

of the earth T . Sampling techniques can be used to derive<br />

the "third order central moment function" Φ T ( P,<br />

Q,<br />

R)<br />

,<br />

which is a three-point function, and the "fourth order central<br />

moment function" Ψ T ( P,<br />

Q,<br />

R,<br />

S)<br />

, which is a four-point<br />

function, in a way similar to the one used for the estimation<br />

of the covariance (second order central moment) function.<br />

Under the usual assumptions of stationarity and isotropy the<br />

above functions become functions of spherical distances ψ<br />

between the points in question.<br />

The function Φ T ( P,<br />

Q,<br />

R)<br />

=ΦT<br />

( ψ PQ , ψ PR , ψ QR ) , e.g., is<br />

approximated by fitting a function Φ T ( ψ 1,<br />

ψ 2 , ψ 3 ) to the 3-<br />

dimensional histogram (3-parameter step function) derived<br />

by taking averages of all the products<br />

[ T ( P)<br />

−mT<br />

][ T ( Q)<br />

−mT<br />

][ T ( R)<br />

−mT<br />

]<br />

where m T = E{T } , corresponding to triads of points P , Q ,<br />

R , such that the values ψ PQ , ψ PR , ψ QR fall within a prescribed<br />

cube.<br />

The results of such a sampling investigation have a local<br />

character, i.e. they cannot be taken to apply to different areas<br />

of the earth surface.<br />

In conclusion laborious numerical investigations in various<br />

parts of the earth need to be undertaken before quadratic<br />

collocation can be evaluated as a worthwhile alternative to<br />

the usual (linear) one.<br />

References<br />

<strong>Dermanis</strong>, A. (1991): A unified approach to linear estimation<br />

and prediction. IUGG General Assembly, Vienna<br />

Aug. 1991.<br />

<strong>Dermanis</strong>, A. and F. Sanso (1993): A Study of Nonlinear<br />

Estimation. General Meeting of the International<br />

Association of Geodesy, Beijing, China, August 8-<br />

14, 1993.<br />

5<br />

Grafarend, E. (1972): Nichtlineare Prädiktion. ZfV, 97, 6,<br />

245-255.<br />

.XEiþHN / /RFDOO\ EHVW TXDGUDWLF HVWLPDWRUV<br />

Math. Slovaca, vol. 35, no. 4, 393-408.<br />

Moritz, H. (1980): Advanced Physical Geodesy. H. Wichmann<br />

Verlag, Karlsruhe.<br />

Rao, C.R. (1973): Linear statistical inference and its applications.<br />

2nd edition. Wiley, New York.<br />

Sanso, F. (1986): Statistical Methods in Physical Geodesy.<br />

In: H. Sünkel (ed.) "Mathematical and Numerical<br />

Techniques in Physical Geodesy", Lecture Notes in<br />

Earth Sciences, Vol. 7, 49-155. Springer-Verlag,<br />

Berlin.<br />

Schaffrin, B. (1983): Varianz-kovarianz-komponenten-Schätzung<br />

bei der Ausgleichung heterogener Wiederholungsmessungen.<br />

DGK Reihe C, Heft Nr. 282.<br />

Schaffrin, B. (1985a): Das geodätische Datum mit stochastischer<br />

Vorinformation. DGK Reihe C, Heft<br />

Nr. 313.<br />

Schaffrin, B. (1985b): On robust collocation. In: F. Sanso<br />

(ed). Proceedings First Hotine-Marussi Symposium<br />

on Mathematical Geodesy (Roma, 1985), Milano,<br />

1986, pp. 343-361.<br />

Appendix A<br />

The various types of quadratic predictors can more easily<br />

be derived in the "vectorized" form starting with equations<br />

(14), (19) and (20)<br />

s ˆ ′ = γ +a T z (inhom) ,<br />

s ˆ ′ =a T z<br />

(hom) (A1)<br />

β<br />

= γ + aT m z −m′<br />

(inhom),<br />

β a m −m<br />

(hom) (A2)<br />

=<br />

T<br />

z<br />

′<br />

φ a C a 2a<br />

c . (A3)<br />

= β 2 + T<br />

T<br />

z + σ<br />

s<br />

2<br />

′ − zs′<br />

We note that<br />

∂β<br />

∂a =mT z<br />

∂β<br />

, = 1 . (A4)<br />

∂γ<br />

Best quadratic inhomogeneous prediction (inhomBQP):<br />

Unconditional minimization of φ follows from the set of<br />

equations<br />

1 ⎛ ∂φ<br />

⎞<br />

⎜ ⎟<br />

2 ⎝ ∂a<br />

⎠<br />

1<br />

2<br />

T<br />

T<br />

⎛ ∂β<br />

⎞<br />

= β ⎜ ⎟<br />

⎝ ∂a<br />

⎠<br />

∂φ<br />

∂β<br />

= β = β = γ + a<br />

∂γ<br />

∂γ<br />

+ Cz a−c<br />

z = β m z + C za−c<br />

z = 0 ,<br />

T<br />

m z<br />

s′ s′<br />

(A5)<br />

−m′=<br />

0 , (A6)


with solution<br />

a= C−1 z c zs′<br />

, γ = m′−c<br />

C<br />

−1<br />

z m z , (A7)<br />

s ˆ′ = m′+c C<br />

− 1 ( z −m<br />

T<br />

z s′<br />

z<br />

z<br />

)<br />

T<br />

z s ′<br />

= aT<br />

C<br />

T<br />

T −<br />

za+<br />

σ<br />

s<br />

2 ′<br />

−2a<br />

c z s ′ = σ 2<br />

s ′<br />

−c<br />

z s ′ C<br />

1<br />

z c zs′<br />

(A8)<br />

φ . (A9)<br />

Since the above prediction is already unbiased from (A6) it<br />

follows that it is identical with the Best Quadratic Unbiased<br />

inhomogeneous prediction (inhomBQUP = inhomBQP).<br />

Best quadratic homogeneous prediction (homBQP):<br />

Unconditional minimization of φ follows from<br />

1 ⎛ ∂φ<br />

⎞<br />

⎜ ⎟<br />

2 ⎝ ∂a<br />

⎠<br />

T<br />

⎛ ∂β<br />

⎞<br />

= β ⎜ ⎟<br />

⎝ ∂a<br />

⎠<br />

T<br />

+ C<br />

a−<br />

z c z s′<br />

=<br />

( mT<br />

a−m<br />

′)<br />

m + C a−c<br />

0 , (A10)<br />

= z z z zs′<br />

=<br />

with solution<br />

a= ( C + m mT<br />

)<br />

1(<br />

c ′<br />

=<br />

z<br />

z<br />

z<br />

+ m′<br />

m<br />

−<br />

zs<br />

z )<br />

m′−c<br />

C<br />

m<br />

T −1<br />

z z z<br />

C<br />

−1<br />

′<br />

z c z<br />

C<br />

−1<br />

′ + s<br />

s<br />

z<br />

1+<br />

mT<br />

z C<br />

−1<br />

z m z<br />

=<br />

m<br />

z<br />

, (A11)<br />

mT<br />

−<br />

⎛<br />

−<br />

z C<br />

1<br />

z z<br />

mT<br />

⎞<br />

′=<br />

′+<br />

−<br />

z C<br />

1<br />

z z<br />

s ˆ<br />

m<br />

T<br />

c<br />

⎜ −<br />

⎟<br />

′<br />

+<br />

−<br />

z<br />

C<br />

1<br />

z z<br />

m<br />

⎝ +<br />

− z<br />

1 m z C<br />

1<br />

z m s<br />

,<br />

T<br />

z<br />

1 mT<br />

z C<br />

1<br />

z m z ⎠<br />

(A12)<br />

c C m −m′<br />

β =<br />

, (A13)<br />

m<br />

T −1<br />

z s ′ z z<br />

1+<br />

1<br />

mT<br />

z C−<br />

z<br />

z<br />

( m′−c<br />

C m z )<br />

2<br />

φ = σ<br />

2<br />

c<br />

z<br />

C<br />

−1<br />

′<br />

−<br />

T<br />

s s ′ z c zs′<br />

+<br />

. (A14)<br />

1+<br />

m C m<br />

T<br />

z s ′<br />

T<br />

z<br />

−1<br />

z<br />

−1<br />

z<br />

Best quadratic unbiased homogeneous prediction<br />

(homBQUP):<br />

Minimization of φ under the condition β = 0 , requires the<br />

formulation of the Lagrangean function L = φ −2λβ<br />

and the<br />

solution is provided from the equations<br />

1 ⎛ ∂L<br />

⎞<br />

⎜ ⎟<br />

2 ⎝ ∂a<br />

⎠<br />

T<br />

⎛<br />

= β ⎜<br />

⎝<br />

T<br />

∂β<br />

⎞<br />

⎟<br />

∂a<br />

⎠<br />

+ C<br />

z<br />

a−c<br />

= z z zs′<br />

=<br />

zs′<br />

z<br />

⎛<br />

−λ<br />

⎜<br />

⎝<br />

T<br />

∂β<br />

⎞<br />

⎟<br />

∂a<br />

⎠<br />

( β −λ)<br />

m + C a−c<br />

0 , (A15)<br />

1 ∂L<br />

=−β<br />

= m′−a T m z = 0 , (A16)<br />

2 ∂λ<br />

Since β = 0 (A15) gives<br />

a= C λ (A17)<br />

−<br />

z<br />

1 c zs′<br />

+ C z<br />

−1 m z<br />

which can be inserted into (A16) to provide<br />

m′−c<br />

λ =<br />

m<br />

C<br />

m<br />

T −1<br />

z s ′ z z<br />

T<br />

z C<br />

−1<br />

z m z<br />

=<br />

(A18)<br />

with the above values of a and λ the prediction and its<br />

mean square error become<br />

mT<br />

−<br />

⎛ −<br />

′ =<br />

z C<br />

1<br />

z z<br />

mT<br />

′+ −<br />

⎜<br />

−<br />

′ −<br />

z C<br />

1<br />

z z<br />

s ˆ<br />

m<br />

T<br />

c<br />

z<br />

C<br />

1<br />

z z<br />

m<br />

mT<br />

−<br />

z C<br />

1<br />

z m s<br />

z<br />

⎝ mT<br />

z C<br />

1<br />

z m z<br />

T −1<br />

z s ′ z z<br />

T 1<br />

z C<br />

−<br />

z m z<br />

z<br />

⎞<br />

⎟<br />

⎠<br />

(A19)<br />

( m′−c<br />

C m )<br />

2<br />

φ = σ<br />

2<br />

c<br />

z<br />

C<br />

−1<br />

′<br />

−<br />

T<br />

s s ′ z c zs′<br />

+<br />

. (A20)<br />

m<br />

6

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!