With these new expressions, the law of variance
propagation permits the expression of the
corresponding covariance matrices in equally
convenient forms. The covariance matrix of the
estimated signal and the residual noise become
respectively:
1
C=C > 6° (BR PB) La
ss ss n
+ ot (B‘PBC B'PB « o? B'pp)! (A.8)
n ss n
C= oF IP = BB ERY 5]
nn n
+ o* B(B'PBC B'PB + o B'PB) 'B (A.9)
n ss n
Moreover, taking into account (A.1), the estimated
value of the observables can be written as:
(A. 10)
Applying the law of variance propagation, its
covariance matrix becomes:
t
C** = BC -B (A. 11)
ox ss
Finally, indicating with the symbol e the error in
the estimate of the signal, i.e. the difference
between its theoretic value and its estimate: e =
s - s, the covariance matrix C becomes, taking
into account (A.8), and applying the law of
covariance propagation,
C -C -C^-
ee ss ss
c? (B'PB)'!« o* (B*PBC_ B*PB + o% B*PB)-* (A.12)
n n ss n
Consequently one has:
C~=2C -C (A. 13)
ss ss ee
and:
Qm eg pl-n Bim Pho (A. 14)
nn n ee ££
where the last matrix in expression (A.14) is the
covariance matrix of the estimate for the expected
value of the observables:
BC B
ee
C = (A.15)
££
having indicated with the error of the estimate
of the expected value of the observables, i.e. the
difference between its theoretical value and its
estimate:
388
€ =o - & 7 B(s - s) = Be
The least squares criterion,
formulation of the collocation method, can
provide, besides an estimate for a filtered
signal, also an estimate for a predicted signal
s=t: the stochastic parameters can also be
eBtimated in every point. One has to keep in mind
however, that only the properly called stochastic
parameters can be estimated.
Consequently the covariance matrix C only consists
of the properly called stochastic parámeters, and
the crosscovariance matrix between the filtered
expressed in the
and the predicted signal C ,(C ,- ct )is divided
in two parts: one contáinihg the covariance
between the predicted signal and the properly
called stochastic parameters in the filtered
signal, and one identically zero. This null matrix
is exactly the reason of the impossibility to
predict the parameters, which are strictly non
stochastic.
Given the functional:
C 0 S
ss st
1 m ta lic C gl |
2 ts tt
0 à Prn |n
+ A (Be = Ot + n - €) = min
A being a vector of Lagrange multipliers, and
taking into account expression {A.7), one has:
t zc ppc B P')'a =
ts ss n
- c B'PB (B'PBC B'PB « o? B'PB) !B'P a^ (4.16)
ts ss n
or.
t = z
with z a service vector:
z = B'PB (B'PBC .B'PB + o? B*PB) !p'P a*
s
which is to be computed once at the end of the
filtering.
Applying the law of covariance propagation, the
covariance matrix of the predicted signal becomes:
C^ - C B'PB (B'PBC B'PB « o? B'PB) !B'PBC | (A. 17)
tt ts ss n st