17.

5 Correlated Process and Observation Noise

We shall follow a development from Brown & Hwang (1997), pages 296–298 and start by
explicitly stating a somewhat changed version of the usual covariance conditions

6,k if j = k
E{ k j } =
T
0 if j 6 = k

6e,k if j = k
E{ek ejT } =
0 if j 6= k
E{ k ejT } = Ck for all j and k.

The last equation states that process noise  k and observation noise ek are correlated. We
want to study the modifications introduced by this change to the usual filter equations.

Filtering of the state vector:

x̂ k|k = x̂ k|k−1 + Kk (bk − Ak x̂ k|k−1 ).

The system equation is

x k = Fk−1 x k−1 +  k

and hence the estimation error

v k|k = x k − x̂ k|k


= x k − x̂ k|k−1 + Kk (bk − Ak x̂ k|k−1 )
= (I − Kk Ak )(x k − x̂ k|k−1 ) − Kk ek
= (I − Kk Ak )v k|k−1 − Kk ek

with covariance matrix Pk|k = E{v k|k v Tk|k }.

Now we compute the covariance between v k|k−1 and ek :

E{v k|k−1 eTk } = E{(x k − x̂ k|k−1 )eTk }

= E{(Fk−1 x k−1 +  k − Fk−1 x̂ k−1|k−1 )eTk }.

ek is neither correlated with x k−1 nor x̂ k−1|k−1 due to their whiteness. Therefore we get

E{v k|k−1 eTk } = E{ k eTk } = Ck .

We return to the main development. The covariance matrix for the estimation error is

Pk|k = E{v k|k v Tk|k }



T
= E (I − Kk Ak )v k|k−1 − Kk ek (I − Kk Ak )v k|k−1 − Kk ek
= (I − Kk Ak )Pk|k−1 (I − Kk Ak )T + Kk 6e,k KkT
− (I − Kk Ak )Ck KkT − Kk CkT (I − Kk Ak )T . (17.1)
 17.5 Correlated Process and Observation Noise 567

This is a general expression for the error covariance matrix and valid for any gain ma-
trix Kk . The two last terms are introduced due to the cross-correlation Ck . For Ck = 0 we
recover the earlier Joseph form of the variance.
We want to rewrite (17.1), and collect a factor from the first and fourth terms and
manipulate the other terms:

Pk|k = Pk|k−1 − Kk (Ak Pk|k−1 + CkT ) + Kk (Ak Pk|k−1 ATk + 6e,k + Ak Ck + CkT ATk )KkT
− (Pk|k−1 ATk + Ck )T KkT (17.2)

The Kalman matrix associated with a minimum estimation error is determined by setting
∂ tr(Pk|k )
∂Kk = 0 or

−2(Ak Pk|k−1 )T + 2Kk Ak Pk|k−1 ATk + 2Kk 6e,k − 2Ck + 2Kk Ak Ck + 2Kk CkT ATk = 0.

This yields

−1
Kk = Pk|k−1 ATk + Ck Ak Pk|k−1 ATk + 6e,k + Ak Ck + CkT ATk


. (17.3)

Let Ck → 0 and this equation reduces to the gain in the zero cross-correlation model

−1
Kk = Pk|k−1 ATk Ak Pk|k−1 ATk + 6e,k (17.4)

as it should be. Finally we insert (17.3) into (17.2) and get

Pk|k = (I − Kk Ak )Pk|k−1 − Kk CkT . (17.5)

The prediction equations are not affected by the cross-correlation between  k and ek be-
cause of the whiteness property of each.
A possible correlation between the process noise  k and the observation noise ek
leads to modifications of the Kalman gain matrix Kk and the expression for the filter co-
variance matrix Pk|k . The modifications involving terms with Ck are in (17.3) and (17.5).
568

Useful partial derivatives of traces and the determinant:

∂ tr(A)
=I
∂A  T
∂ tr(AB) B if A is asymmetric
=
∂A B + B T − diag(B), if A is symmetric
∂ tr(BAC)
= B TCT
∂A
∂ tr(ABAT )
= A(B + B T )
∂A
∂ tr(eA ) T
= eA
∂A
∂|BAC|
T
= |BAC| A−1
∂A
 REFERENCES

Brown, Robert Grover & Hwang, Patrick Y. C. (1997). Introduction to Random Signals
and Applied Kalman Filtering. John Wiley & Sons, Inc., New York, 3rd edition.

569