Chapter V: Selected Solutions

H. V. Poor
Princeton University
October 14, 2002

Exercise 3:
Since sk depends completely on Y k0 , it behaves like a constant when conditional quantities
given Y k0 are computed. The only point at which this aects the derivation of the Kalman-
Bucy lter is in the time update equations, which now become:
t+1|t = Ft X
X t|t + E{t st |Y t0 } = Ft X
t|t + t st ,

and  
t+1|t = Ft t|t FTt + Gt Qt GTt + Cov t st |Y t0 = Ft t|t FTt + Gt Qt GTt .
Note that the second of these two equations is the same as when there is no measurement
feedback.
The measurement feedback has no eect on the measurement update equations. Al-
though the presence of this feedback may cause the state to be nonGaussian, the joint
conditional statistics of X t and Y t given Y t1
0 are still Gaussian. Thus, the measurement
update is unchanged from the case of no measurement feedback since it depends only on
this joint Gaussian property and the linearity of measurement equation.

Exercise 4:
There are several ways of approaching this problem. An interesting one is to note that,
although U t and V t are dependent, the Gaussian vector U t U t Ct R1
t V t is independent
1
of V t . To take advantage of this, we may add the zero quantity Ct Rt [Y t Ht X t V t ]
to the tth state input, which yields the equivalent state equation
X t+1 = Ft X t + Gt U t + Gt Ct R1
t (Y t Ht X t ) .

So, we have an equivalent problem with independent state and measurement noises,
but
 with the measurement
 feedback term Gt Ct R1t Y t , and with the new state matrix
1
Ft Gt Ct Rt Ht . We also have a dierent correlation matrix for the state input, since

Cov(U t ) = Qt Ct R1 T
t Ct .

1
Applying the result of Exercise 3 and eliminating the measurement update equations
yields the given result.

Exercise 6:
a. This result follows by induction on t. We rst note that, for t = j, the given equality
follows by denition. From the state equation we have that, for k > j,
        
E Xj X j|k X Tk+1 = E X j X j|k (Fk X k + Gk U k )T = E X j X j|k X Tk FTk ,

j|k .
where we use the fact that U k has zero mean and is independent of both X j and X
j|k to this equation we have
Now, on applying the recursion for X
        
E j|k X Tk+1 = E
Xj X j|k1 X Tk FTk Kak E
Xj X k|k1 X Tk FTk .
Y k Hk X
We now assume that the given equality is true for t = k. From this and the denition of
Kak , we then have
     1   
j|k X T
Xj X k|k1 X Tk
k+1 = k|k1 I Hk Hk k|k1 Hk + Rk Y k Hk X
a T T
E E FTk
  1   
= ak|k1 I HTk Hk k|k1 HTk + Rk Hk E k|k1 X T
Xk X FTk
k
  1  
= ak|k1 I HTk Hk k|k1 HTk + Rk Hk k|k1 FTk = ak|k1 I HTk KTk FTk ,
where the last equality follows by deniton of Kk . Applying the recursion for ak+1|k , we
have   
E Xj X j|k X Tk+1 = ak+1|k ,
which shows that the given equation for t = k + 1. The induction principle thus gives the
desired result.
b. As with the measurement-update derivation in the Kalman-Bucy  lter, we note
here that X j and Y t are jointly Gaussian conditioned on Y 0 . Thus, E X j |Y t0 will be
t1

as given by the recursion if we have

    1
Cov X j , Y t |Y t1
0 Cov Y t |Y t1
0 = Kat .
But, since Y t = Ht X t + V t , the result from Part a. implies this equality.

Exercise 9:
This is the Kalman-Bucy problem with all dimensions equal to unity, Fk 1, Gk
n|n n , and n|n n|n
n+1|n X
Qk
0, Hk = sk , Rk 2 , m0 , 0 v 2 , X
 2
E n . The desired recursions thus follow by eliminating either set of updates
from (V.B.14) - (V.B.16). The resulting estimate is the same as that found in Example
IV.B.2.