Upload
others
View
9
Download
0
Embed Size (px)
Citation preview
References
Alfeld, G., & Herzberger, J. (1983). Introduction to interval computations. New York: AcademicPress.
Anderson, B. D. O., & Moore, J. B. (1979). Optimal filtering. Englewood Cliffs: Prentice-Hall.Andrews, A. (1981). Parallel processing of the Kalman filter. In IEEE Proceedings of the Interna-tional Conference on Parallel Processing (pp. 216–220).
Aoki, M. (1989). Optimization of stochastic systems: Topics in discrete-time dynamics. New York:Academic Press.
Aström, K. J., & Eykhoff, P. (1971). System identification – a survey. Automatica, 7, 123–162.Balakrishnan, A. V. (1984, 1987). Kalman filtering theory. New York: Optimization Software, Inc.Bierman, G. J. (1973). A comparison of discrete linear filtering algorithms. IEEE Transactions onAerospace and Electronic Systems, 9, 28–37.
Bierman, G. J. (1977). Factorization methods for discrete sequential estimation. New York: Acad-emic Press.
Blahut, R. E. (1985). Fast algorithms for digital signal processing. Reading: Addison-Wesley.Bozic, S. M. (1979). Digital and Kalman filtering. New York: Wiley.Brammer, K., & Sifflin, G. (1989). Kalman-Bucy filters. Boston: Artech House Inc.Brown, R. G., & Hwang, P. Y. C. (1992, 1997). Introduction to random signals and applied Kalmanfiltering. New York: Wiley.
Bucy, R. S., & Joseph, P. D. (1968). Filtering for stochastic processes with applications to guidance.New York: Wiley.
Burrus, C. S., Gopinath, R. A., & Guo, H. (1998). Introduction to wavelets and wavelet transfroms:A primer. Upper Saddle River: Prentice-Hall.
Carlson, N. A. (1973). Fast triangular formulation of the square root filter. Journal of AIAA, 11,1259–1263.
Catlin, D. E. (1989). Estimation, control, and the discrete Kalman filter. New York: Springer.Cattivelli, F. S., & Sayed, A. H. (2010). Diffusion strategies for distributed Kalman filtering andsmoothing. IEEE Transactions on Automatic Control, 55(9), 2069–2084.
Chen, G. (1992). Convergence analysis for inexact mechanization of Kalman filtering. IEEE Trans-actions on Aerospace and Electronic Systems, 28, 612–621.
Chen, G. (1993). Approximate Kalman filtering. Singapore: World Scientific.
© Springer International Publishing AG 2017C.K. Chui and G. Chen, Kalman Filtering, DOI 10.1007/978-3-319-47612-4
209
210 References
Chen, G., & Chui, C. K. (1986). Design of near-optimal linear digital tracking filters with coloredinput. Journal of Computational and Applied Mathematics, 15, 353–370.
Chen, G., Chen, G., & Hsu, S. H. (1995). Linear stochastic control systems. Boca Raton: CRCPress.
Chen, G., Wang, J., & Shieh, L. S. (1997). Interval Kalman filtering. IEEE Transactions onAerospace and Electronic Systems, 33, 250–259.
Chen, H. F. (1985). Recursive estimation and control for stochastic systems. New York: Wiley.Chui, C. K. (1984). Design and analysis of linear prediction-correction digital filters. Linear andMultilinear Algebra, 15, 47–69.
Chui, C. K. (1997). Wavelets: A mathematical tool for signal analysis. Philadelphia: SIAM.Chui, C. K., & Chen, G. (1989). Linear systems and optimal control. New York: Springer.Chui, C. K., & Chen, G. (1992, 1997). Signal processing and systems theory: Selected topics. NewYork: Springer.
Chui, C. K., Chen, G., & Chui, H. C. (1990). Modified extended Kalman filtering and a real-timeparallel algorithm for system parameter identification. IEEE Transactions on Automatic Control,35, 100–104.
Davis, M. H. A. (1977). Linear estimation and stochastic control. New York: Wiley.Davis, M. H. A., & Vinter, R. B. (1985). Stochastic modeling and control. New York: Chapmanand Hall.
Fleming, W. H., & Rishel, R. W. (1975). Deterministic and stochastic optimal control. New York:Springer.
Gaston, F.M. F.,& Irwin,G.W. (1990). SystolicKalmanfiltering:An overview. IEEProceedings-D,137, 235–244.
Goodwin, G. C., & Sin, K. S. (1984). Adaptive filtering prediction and control. Englewood Cliffs:Prentice-Hall.
Haykin, S. (1986). Adaptive filter theory. Englewood Cliffs: Prentice-Hall.Hong, L., Chen, G., & Chui, C. K. (1998). A filter-bank-based Kalman filtering technique forwavelet estimation and decomposition of random signals. IEEE Transactions on Circuits andSystems (II) (in press).
Hong, L., Chen, G., & Chui, C. K. (1998). Real-time simultaneous estimation and decompositionof random signals. Multidimensional Systems and Signal Processing (in press).
Jazwinski, A. H. (1969). Adaptive filtering. Automatica, 5, 475–485.Jazwinski, A. H. (1970). Stochastic processes and filtering theory. New York: Academic Press.Jover, J. M., & Kailath, T. (1986). A parallel architecture for Kalman filter measurement updateand parameter estimation. Automatica, 22, 43–57.
Kailath, T. (1968). An innovations approach to least-squares estimation, part I: Linear filtering inadditive white noise. IEEE Transactions on Automatic Control, 13, 646–655.
Kailath, T. (1982). Course notes on linear estimation. Stanford: Stanford University.Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Transactions ofthe ASME – Journal of Basic Engineering, 82, 35–45.
Kalman, R. E. (1963). New method in Wiener filtering theory. In Proceedings of the Symposium onEngineering Applications of Random Function Theory and Probability. New York: Wiley.
Kalman, R. E., & Bucy, R. S. (1961). New results in linear filtering and prediction theory. Trans-actions of the ASME – Journal of Basic Engineering, 83, 95–108.
Kumar, P. R., & Varaiya, P. (1986). Stochastic systems: Estimation, identification, and adaptivecontrol. Englewood Cliffs: Prentice-Hall.
Kung, H. T. (1982). Why systolic architectures? Computer, 15, 37–46.Kung, S. Y. (1985). VLSI arrays processors. IEEE ASSP Magazine, 2, 4–22.Kushner, H. (1971). Introduction to stochastic control. New York: Holt Rinehart and Winston Inc.Lewis, F. L. (1986). Optimal estimation. New York: Wiley.
References 211
Lu, M., Qiao, X., & Chen, G. (1992). A parallel square-root algorithm for the modified extendedKalman filter. IEEE Transactions on Aerospace and Electronic Systems, 28, 153–163.
Lu, M., Qiao, X., & Chen, G. (1993). Parallel computation of the modified extended Kalman filter.International Journal of Computer Mathematics, 45, 69–87.
Maybeck, P. S. (1982).Stochasticmodels, estimation, and control (Vol. 1,2,3).NewYork:Academic.Mead, C., & Conway, L. (1980). Introduction to VLSI systems. Reading: Addison-Wesley.Mehra, R. K. (1970). On the identification of variances and adaptive Kalman filtering. IEEE Trans-actions on Automatic Control, 15, 175–184.
Mehra, R. K. (1972). Approaches to adaptive filtering. IEEE Transactions on Automatic Control,17, 693–698.
Mendel, J. M. (1987). Lessons in digital estimation theory. Englewood Cliffs: Prentice-Hall.Potter, J. E. (1963). New statistical formulas. Instrumentation Lab., MIT (Space Guidance AnalysisMemo).
Probability Group (1975). Institute of Mathematics, Academia Sinica, China, ed.: MathematicalMethods of Filtering for Discrete-Time Systems (in Chinese) (Beijing)
Ruymgaart, P. A., & Soong, T. T. (1985, 1988).Mathematics of Kalman-Bucy filtering. New York:Springer.
Shiryayev, A. N. (1984). Probability. New York: Springer.Siouris, G., Chen, G., &Wang, J. (1997). Tracking an incoming ballistic missile. IEEE Transactionson Aerospace and Electronic Systems, 33, 232–240.
Sorenson, H. W. (Ed.). (1985). Kalman filtering: Theory and application. New York: IEEE Press.Stengel, R. F. (1986). Stochastic optimal control: Theory and application. New York: Wiley.Strobach, P. (1990). Linear prediction theory: A mathematical basis for adaptive systems. NewYork: Springer.
Wang, E. P. (1972). Optimal linear recursive filtering methods. Journal of Mathematics in Practiceand Theory (in Chinese), 6, 40–50.
Wonham,W.M. (1968). On the separation theorem of stochastic control. SIAM Journal on Control,6, 312–326.
Xu, J. H., Bian, G. R., Ni, C. K., & Tang, G. X. (1981). State estimation and system identification(in Chinese) (Beijing).
Yang, W., Chen, G., Wang, X. F., & Shi, L. (2014). Stochastic sensor activation for distributed stateestimation over a sensor network. Automatics, 50, 2070–2076.
Young, P. (1984). Recursive estimation and time-series analysis. New York: Springer.Yu, W. W., Chen, G., Wang, Z. D., & Yang, W. (2009). Distributed consensus filtering in sensornetworks. IEEE Transactions on Systems, Man and Cybernetics–Part B: Cybernetics, 39(6),1568–1577.
Zhang, H. S., Song, X. M., & Shi, L. (2012). Convergence and mean square stability of suboptimalestimator for systems with measurements packet dropping. IEEE Transactions on AutomaticControl, 57(5), 1248–1253.
Answers andHints to Exercises
Chapter1
1.1. Since most of the properties can be verified directly by using the definitionof the trace, we only consider trAB = trBA. Indeed,
trAB =n∑
i=1
( m∑
j=1
ai j b ji
)=
m∑
j=1
( n∑
i=1
b ji ai j
)= trBA .
1.2.
(trA)2 =( n∑
i=1
aii
)2
≤ nn∑
i=1
a2i i ≤ n(trAA�) .
1.3.
A =[3 11 2
], B =
[2 00 1
].
1.4. There exist unitary matrices P and Q such that
A = P
⎡
⎢⎣λ1
. . .
λn
⎤
⎥⎦ P� , B = Q
⎡
⎢⎣μ1
. . .
μn
⎤
⎥⎦ Q� ,
andn∑
k=1
λ2k ≥
n∑
k=1
μ2k .
Let P = [pi j ]n×n and Q = [qi j ]n×n . Then
p211 + p221 + · · · + p2n1 = 1 , p212 + p222 + · · · + p2n2 = 1 , · · · ,
© Springer International Publishing AG 2017C.K. Chui and G. Chen, Kalman Filtering, DOI 10.1007/978-3-319-47612-4
213
214 Answers and Hints to Exercises
p21n + p22n + · · · + p2nn = 1 , q211 + q221 + · · · + q2n1 = 1 ,
q212 + q222 + · · · + q2n2 = 1 , · · · , q21n + q22n + · · · + q2nn = 1 ,
and
trAA� = tr
{P
⎡
⎢⎣λ21
. . .
λ2n
⎤
⎥⎦ P�}
= tr
⎡
⎢⎢⎢⎢⎢⎢⎢⎢⎣
p211λ21 + p212λ
22 ∗
+ · · · + p21nλ2n
p221λ21 + p222λ
22
+ · · · + p22nλ2n
p2n1λ21 + p2n2λ
22
∗ + · · · + p2nnλ2n
⎤
⎥⎥⎥⎥⎥⎥⎥⎥⎦
= (p211 + p221 + · · · + p2n1)λ21 + · · · + (p21n + p22n + · · · + p2nn)λ
2n
= λ21 + λ2
2 + · · · + λ2n .
Similarly, trBB� = μ21 + μ2
2 + · · · + μ2n . Hence, trAA� ≥ trBB�.
1.5. Denote
I =∫ ∞
−∞e−y2dy .
Then, using polar coordinates, we have
I 2 =(∫ ∞
−∞e−y2dy
)(∫ ∞
−∞e−x2dx
)
=∫ ∞
−∞
∫ ∞
−∞e−(x2+y2)dxdy
=∫ 2π
0
∫ ∞
0e−r2rdrdθ = π .
1.6. Denote
I (x) =∫ ∞
−∞e−xy2dy .
Then, by Exercise1.5,
I (x) = 1√x
∫ ∞
−∞e−(
√x y)2d(
√x y) = √
π/x .
Answers and Hints to Exercises 215
Hence,∫ ∞
−∞y2e−y2dy = − d
dxI (x)
∣∣∣∣x=1
= − 1
dx
(√π/x
)∣∣∣∣x=1
= 1
2
√π .
1.7. (a) Let P be a unitary matrix so that
R = P�diag[λ1, · · · ,λn]P ,
and define
y = 1√2diag[√λ1, · · · ,
√λn]P(x − μ) .
Then
E(X) =∫ ∞
−∞x f (x)dx
=∫ ∞
−∞(μ + √
2P−1 diag[ 1/√λ1, · · · , 1/√
λn ]y) f (x)dx
=μ
∫ ∞
−∞f (x)dx
+ Const.∫ ∞
−∞· · ·
∫ ∞
−∞
⎡
⎢⎣y1...
yn
⎤
⎥⎦ e−y21 · · · e−y2n dy1 · · · dyn
= μ · 1 + 0 = μ .
(b) Using the same substitution, we have
Var(X)
=∫ ∞
−∞(x − μ)(x − μ)� f (x)dx
=∫ ∞
−∞2R1/2yy�R1/2 f (x)dx
= 2
(π)n/2 R1/2{∫ ∞
−∞· · ·
∫ ∞
−∞
⎡
⎢⎣y21 · · · y1yn...
...
yn y1 · · · y2n
⎤
⎥⎦
· e−y21 · · · e−y2n dy1 · · · dyn}R1/2
= R1/2 I R1/2 = R .
216 Answers and Hints to Exercises
1.8. All the properties can be easily verified from the definitions.1.9.Wehave alreadyproved that if X1 and X2 are independent thenCov(X1, X2) =0. Suppose now that R12 = Cov(X1, X2) = 0. Then R21 = Cov(X2, X1) = 0 sothat
f (X1, X2) = 1
(2π)n/2det R11det R22
· e− 12 (X1−μ
1)�R11(X1−μ
1)e− 1
2 (X2−μ2)�R22(X2−μ
2)
= f1(X1) · f2(X2) .
Hence, X1 and X2 are independent.1.10. Equation (1.35) can be verified by a direct computation. First, the followingformula may be easily obtained:
[I −Rxy R−1
yy0 I
] [Rxx Rxy
Ryx Ryy
] [I 0
−R−1yy R
�xy I
]
=[Rxx − Rxy R−1
yy Ryx 00 Ryy
].
This yields, by taking determinants,
det
[Rxx Rxy
Ryx Ryy
]= det
[Rxx − Rxy R
−1yy Ryx
] · detRyy
and
([xy
]−[μ
xμy
])� [Rxx Rxy
Ryx Ryy
]−1 ([xy
]−[μ
xμy
])
=(x − μ)�[Rxx − Rxy R
−1yy Ryx
]−1(x − μ) + (y − μ
y)�R−1
yy (y − μy) ,
where
μ = μx
+ Rxy R−1yy (y − μ
y) .
The remaining computational steps are straightforward.1.11. Let pk = C�
k Wkzk and σ2 = E[p�k (C�
k WkCk)−1pk
]. Then it can be easily
verified that
F(yk) = y�k (C�
k WkCk)yk − p�k yk − y�
k pk + σ2 .
FromdF(yk)dyk
= 2(C�k WkCk)yk − 2pk = 0 ,
Answers and Hints to Exercises 217
and the assumption that the matrix (C�k WkCk) is nonsingular, we have
yk = (C�k WkCk)
−1pk = (C�k WkCk)
−1C�k Wkzk .
1.12.
E xk = (C�k R−1
k Ck)−1C�
k R−1k E(vk − Dkuk)
= (C�k R−1
k Ck)−1C�
k R−1k E(Ckxk + η
k)
= Exk .
Chapter 2
2.1.
W−1k,k−1 = Var(εk,k−1) = E(εk,k−1ε
�k,k−1)
= E(vk−1 − Hk,k−1xk)(vk−1 − Hk,k−1xk)�
=⎡
⎢⎣R0
. . .
Rk−1
⎤
⎥⎦+ Var
⎡
⎢⎢⎣
C0∑k
i=1�0i�i−1ξi−1...
Ck−1�k−1,k�k−1ξk−1
⎤
⎥⎥⎦ .
2.2. For any nonzero vector x, we have x�Ax > 0 and x�Bx ≥ 0 so that
x�(A + B)x = x�Ax + x�Bx > 0 .
Hence, A + B is positive definite.2.3.
W−1k,k−1
=E(εk,k−1ε�k,k−1)
=E(εk−1,k−1 − Hk,k−1�k−1ξk−1)(εk−1,k−1 − Hk,k−1�k−1ξk−1
)�
=E(εk−1,k−1ε�k−1,k−1) + Hk,k−1�k−1E(ξ
k−1ξ�k−1
)��k−1H
�k,k−1
=W−1k−1,k−1 + Hk−1,k−1�k−1,k�k−1Qk−1�
�k−1�
�k−1,k H
�k−1,k−1 .
2.4. Apply Lemma1.2 to A11 = W−1k−1,k−1, A22 = Q−1
k−1 and
A12 = A�21 = Hk−1,k−1�k−1,k�k−1 .
218 Answers and Hints to Exercises
2.5. Using Exercise2.4 (or (2.9)), we have
H�k,k−1Wk,k−1
=��k−1,k H
�k−1,k−1Wk−1,k−1
− ��k−1,k H
�k−1,k−1Wk,k−1Hk,k−1�k−1,k�k−1
· (Q−1k−1 + ��
k−1��k−1,k H
�k−1,k−1Wk−1,k−1Hk−1,k−1�k−1,k�k−1)
−1
· ��k−1�
�k−1,k H
�k−1,k−1Wk−1,k−1
=��k−1,k{I − H�
k−1,k−1Wk−1,k−1Hk−1,k−1�k−1,k�k−1
· (Q−1k−1 + ��
k−1��k−1,k H
�k−1,k−1Wk−1,k−1Hk−1,k−1�k−1,k�k−1)
−1
· ��k−1�
�k−1,k}H�
k−1,k−1Wk−1,k−1 .
2.6. Using Exercise2.5 (or (2.10)) and the identity Hk,k−1 = Hk−1,k−1�k−1,k ,we have
(H�k,k−1Wk,k−1Hk,k−1)�k,k−1
· (H�k−1,k−1Wk−1,k−1Hk−1,k−1)
−1H�k−1,k−1Wk−1,k−1
=��k−1,k{I − H�
k−1,k−1Wk−1,k−1Hk−1,k−1�k−1,k�k−1
· (Q−1k−1 + ��
k−1��k−1,k H
�k−1,k−1Wk−1,k−1Hk−1,k−1�k−1,k�k−1)
−1
· ��k−1�
�k−1,k}H�
k−1,k−1Wk−1,k−1
=H�k,k−1Wk,k−1 .
2.7.
Pk,k−1C�k (Ck Pk,k−1C
�k + Rk)
−1
=Pk,k−1C�k (R−1
k − R−1k Ck(P
−1k,k−1 + C�
k R−1k Ck)
−1C�k R−1
k )
=(Pk,k−1 − Pk,k−1C�k R−1
k Ck(P−1k,k−1 + C�
k R−1k Ck)
−1)C�k R−1
k
=(Pk,k−1 − Pk,k−1C�k (Ck Pk,k−1C
�k + Rk)
−1
· (Ck Pk,k−1C�k + Rk)R
−1k Ck(P
−1k,k−1 + C�
k R−1k Ck)
−1)C�k R−1
k
=(Pk,k−1 − Pk,k−1C�k (Ck Pk,k−1C
�k + Rk)
−1
· (Ck Pk,k−1C�k R−1
k Ck + Ck)(P−1k,k−1 + C�
k R−1k Ck)
−1)C�k R−1
k
=(Pk,k−1 − Pk,k−1C�k (Ck Pk,k−1C
�k + Rk)
−1Ck Pk,k−1
· (C�k R−1
k Ck + P−1k,k−1)(P
−1k,k−1 + C�
k R−1k Ck)
−1)C�k R−1
k
=(Pk,k−1 − Pk,k−1C�k (Ck Pk,k−1C
�k + Rk)
−1Ck Pk,k−1)C�k R−1
k
Answers and Hints to Exercises 219
=Pk,kC�k R−1
k
=Gk .
2.8.
Pk,k−1
=(H�k,k−1Wk,k−1Hk,k−1)
−1
=(��k−1,k(H
�k−1,k−1Wk−1,k−1Hk−1,k−1
− H�k−1,k−1Wk−1,k−1Hk−1,k−1�k−1,k�k−1
· (Q−1k−1 + ��
k−1��k−1,k H
�k−1,k−1Wk−1,k−1Hk−1,k−1�k−1,k�k−1)
−1
· ��k−1�
�k−1,k H
�k−1,k−1Wk−1,k−1Hk−1,k−1)�k−1,k)
−1
=(��k−1,k P
−1k−1,k−1�k−1,k − ��
k−1,k P−1k−1,k−1�k−1,k�k−1
· (Q−1k−1 + ��
k−1��k−1,k P
−1k−1,k−1�k−1,k�k−1)
−1
· ��k−1�
�k−1,k P
−1k−1,k−1�k−1,k)
−1
=(��k−1,k P
−1k−1,k−1�k−1,k)
−1 + �k−1Qk−1��k−1
=Ak−1Pk−1,k−1A�k−1 + �k−1Qk−1�
�k−1 .
2.9.
E(xk − xk|k−1)(xk − xk|k−1)�
=E(xk − (H�k,k−1Wk,k−1Hk,k−1)
−1H�k,k−1Wk,k−1vk−1)
· (xk − (H�k,k−1Wk,k−1Hk,k−1)
−1H�k,k−1Wk,k−1vk−1)
�
=E(xk − (H�k,k−1Wk,k−1Hk,k−1)
−1H�k,k−1Wk,k−1
· (Hk,k−1xk + εk,k−1))(xk − (H�k,k−1Wk,k−1Hk,k−1)
−1
· H�k,k−1Wk,k−1(Hk,k−1xk + εk,k−1))
�
=(H�k,k−1Wk,k−1Hk,k−1)
−1H�k,k−1Wk,k−1E(εk,k−1ε
�k,k−1)Wk,k−1
· Hk,k−1(H�k,k−1Wk,k−1Hk,k−1)
−1
=(H�k,k−1Wk,k−1Hk,k−1)
−1
=Pk,k−1 .
The derivation of the second identity is similar.
220 Answers and Hints to Exercises
2.10. Since
σ2 = Var(xk) = E(axk−1 + ξk−1)2
= a2Var(xk−1) + 2aE(xk−1ξk−1) + E(ξ2k−1)
= a2σ2 + μ2 ,
we have
σ2 = μ2/(1 − a2) .
For j = 1, we have
E(xkxk+1) = E(xk(axk + ξk))
= aVar(xk) + E(xkξk)
= aσ2 .
For j = 2, we have
E(xkxk+2) = E(xk(axk+1 + ξk+1))
= aE(xkxk+1) + E(xk + ξk+1)
= aE(xkxk+1)
= a2σ2 ,
etc. If j is negative, then a similar result can be obtained. By induction, we mayconclude that E(xkxk+ j ) = a| j |σ2 for all integers j .2.11. Using the Kalman filtering equations (2.17), we have
P0,0 = Var(x0) = μ2 ,
Pk,k−1 = Pk−1,k−1 ,
Gk = Pk,k−1(Pk,k−1 + Rk)−1 = Pk−1,k−1
Pk−1,k−1 + σ2 ,
and
Pk,k = (1 − Gk)Pk,k−1 = σ2Pk−1,k−1
σ2 + Pk−1,k−1.
Answers and Hints to Exercises 221
Observe that
P1,1 = σ2μ2
μ2 + σ2 ,
P2,2 = σ2P1,1P1,1 + σ2 = σ2μ2
2μ2 + σ2 .
· · ·
Pk,k = σ2μ2
kμ2 + σ2 .
Hence,
Gk = Pk−1,k−1
Pk−1,k−1 + σ2 = μ2
kμ2 + σ2
so that
xk|k = xk|k−1 + Gk(vk − xk|k−1)
= xk−1|k−1 + μ2
σ2 + kμ2 (vk − xk−1|k−1)
with x0|0 = E(x0) = 0. It follows that
xk|k = xk−1|k−1
for large values of k.2.12.
QN = 1
N
N∑
k=1
(vkv�k )
= 1
N(vNv�
N ) + 1
N
N−1∑
k=1
(vkv�k )
= 1
N(vNv�
N ) + N − 1
NQN−1
= QN−1 + 1
N[(vNv�
N ) − QN−1]
with the initial estimation Q1 = v1v�1 .
2.13. Use superimposition.
222 Answers and Hints to Exercises
2.14. Set xk = [(x1k)� · · · (xNk )�]� for each k, k = 0, 1, · · · , with x j = 0 (andu j = 0) for j < 0, and define
x1k = B1x1k−1 + x2k−1 + (A1 + B1A0)uk−1 ,
· · · · · ·xMk = BMx1k−1 + xM+1
k−1 + (AM + BM A0)uk−1 ,
xM+1k = BM+1x1k−1 + xM+2
k−1 + BM+1A0uk−1 ,
· · · · · ·xN−1k = BN−1x1k−1 + xNk−1 + BN−1A0uk−1 ,
xNk = BNx1k−1 + BN A0uk−1 .
Then, substituting these equations into
vk = Cxk + Duk = x1k + A0uk
yields the required result. Since x j = 0 and u j = 0 for j < 0, it is also clear thatx0 = 0.
Chapter 3
3.1. Let A = BB� where B = [bi j ] �= 0. Then trA = trBB� = ∑i, j b
2i j > 0.
3.2. By Assumption2.1, η�is independent of x0, ξ
0, · · · , ξ
j−1, η
0, · · · , η
j−1,
since � ≥ j . On the other hand,
e j = C j (x j − y j−1)
= C j
(A j−1x j−1 + � j−1ξ j−1
−j−1∑
i=0
Pj−1,i (Cixi + ηi)
)
· · · · · ·
= B0x0 +j−1∑
i=0
B1iξi +j−1∑
i=0
B2iηi
for some constant matrices B0, B1i and B2i . Hence, 〈η�, e j 〉 = Oq×q for all � ≥ j .
3.3. Combining (3.8) and (3.4), we have
e j = ‖z j‖−1q z j = ‖z j‖−1
q v j −j−1∑
i=0
(‖z j‖−1
q C j Pj−1,i
)vi ;
that is, e j can be expressed in terms of v0, v1, · · · , v j . Conversely, we have
Answers and Hints to Exercises 223
v0 =z0 = ‖z0‖qe0 ,
v1 =z1 + C1y0 = z1 + C1 P0,0v0
=‖z1‖qe1 + C1 P0,0‖z0‖qe0 ,
· · · · · ·that is, v j can also be expressed in terms of e0, e1, · · · , e j . Hence, we have
Y (e0, · · · , ek) = Y (v0, · · · , vk) .
3.4. By Exercise3.3, we have
vi =i∑
�=0
L�e�
for some q × q constant matrices L�, � = 0, 1, · · · , i , so that
〈vi , zk〉 =i∑
�=0
L�〈e�, ek〉‖zk‖�q = Oq×q ,
i = 0, 1, · · · , k − 1 . Hence, for j = 0, 1, · · · , k − 1,
〈y j , zk〉 =⟨ j∑
i=0
Pj,ivi , zk
⟩
=j∑
i=0
Pj,i 〈vi , zk〉
= On×q .
3.5. Since
xk = Ak−1xk−1 + �k−1ξk−1
= Ak−1(Ak−2xk−2 + �k−2ξk−2) + �k−1ξk−1
= · · · · · ·
= B0x0 +k−1∑
i=0
B1iξi
for some constant matrices B0 and B1i and ξkis independent of x0 and ξ
i(0 ≤
i ≤ k − 1), we have 〈xk, ξk〉 = 0. The rest can be shown in a similar manner.
224 Answers and Hints to Exercises
3.6. Use superimposition.3.7. Using the formula obtained in Exercise3.6, we have
{dk|k = dk−1|k−1 + hwk−1 + Gk(vk − �dk − dk−1|k−1 − hwk−1)
d0|0 = E(d0) ,
where Gk is obtained by using the standard algorithm (3.25) with Ak = Ck =�k = 1.3.8. Let
xk =⎡
⎢⎣x1kx2kx3k
⎤
⎥⎦ , x1k =⎡
⎣�k
�k
�k
⎤
⎦ , x2k =⎡
⎣�Ak
� Ak
� Ak
⎤
⎦ , x3k =⎡
⎣�Ek
�Ek
�Ek
⎤
⎦ ,
ξk
=
⎡
⎢⎢⎣
ξ1k
ξ2k
ξ3k
⎤
⎥⎥⎦ , ηk
=⎡
⎢⎣η1kη2kη3k
⎤
⎥⎦ , vk =⎡
⎢⎣v1k
v2kv3k
⎤
⎥⎦ ,
A =⎡
⎣1 h h2/20 1 h0 0 1
⎤
⎦ , and C = [ 1 0 0 ] .
Then the system described in Exercise3.8 can be decomposed into three subsys-tems:
{xik+1 = Axik + �i
kξik
vik = Cxik + ηik ,
i = 1, 2, 3, where for each k, xk and ξkare 3-vectors, vk and ηk are scalars, Qk a
3 × 3 non-negative definite symmetric matrix, and Rk > 0 a scalar.
Chapter 4
4.1. Using (4.6), we have
L(Ax + By, v)
= E(Ax + By) + 〈Ax + By, v〉[Var(v)]−1(v − E(v))
= A{E(x) + 〈x, v〉[Var(v)]−1
(v − E(v))}
+ B{E(y) + 〈y, v〉[Var(v)]−1
(v − E(v))}
= AL(x, v) + BL(y, v) .
Answers and Hints to Exercises 225
4.2. Using (4.6) and the fact that E(a) = a so that
〈a, v〉 = E(a − E(a)) (v − E(v)) = 0 ,
we have
L(a, v) = E(a) + 〈a, v〉[Var(v)]−1(v − E(v)) = a .
4.3. By definition, for a real-valued function f and a matrix A = [ai j ], d f/d A =[∂ f/∂a ji ]. Hence,
0 = ∂
∂H
(tr‖x − y‖2n
)
= ∂
∂HE((x − E(x)) − H(v − E(v)))�((x − E(x)) − H(v − E(v)))
= E∂
∂H((x − E(x)) − H(v − E(v)))�((x − E(x)) − H(v − E(v)))
= E(−2(x − E(x)) − H(v − E(v))) (v − E(v))�
= 2(H E(v − E(v)) (v − E(v))� − E(x − E(x)) (v − E(v))�
)
= 2(H‖v‖2q − 〈x, v〉) .
This gives
H∗ = 〈x, v〉[‖v‖2q
]−1
so that
x∗ = E(x) − 〈x, v〉[‖v‖2q
]−1(E(v) − v) .
4.4. Since vk−2 is a linear combination (with constant matrix coefficients) of
x0, ξ0, · · · , ξ
k−3, η
0, · · · , η
k−2
which are all uncorrelated with ξk−1
and ηk−1
, we have
〈ξk−1
, vk−2〉 = 0 and 〈ηk−1
, vk−2〉 = 0 .
Similarly, we can verify the other formulas (where (4.6) may be used).4.5. The first identity follows from the Kalman gain equation (cf. Theorem4.1(c)or (4.19)), namely:
Gk(Ck Pk,k−1C�k + Rk) = Pk,k−1C
�k ,
226 Answers and Hints to Exercises
so that
Gk Rk = Pk,k−1C�k − GkCk Pk,k−1C
�k
= (I − GkCk)Pk,k−1C�k .
To prove the second equality, we apply (4.18) and (4.17) to obtain
〈xk−1 − xk−1|k−1, �k−1ξk−1− Kk−1ηk−1
〉= 〈xk−1 − xk−1|k−2 − 〈x#k−1, v#k−1〉
[‖v#k−1‖2
]−1v#k−1,
�k−1ξk−1− Kk−1ηk−1
〉= 〈x#k−1 − 〈x#k−1, v#k−1〉
[‖v#k−1‖2
]−1(Ck−1x#k−1 + η
k−1),
�k−1ξk−1− Kk−1ηk−1
〉= −〈x#k−1, v#k−1〉
[‖v#k−1‖2
]−1(S�
k−1��k−1 − Rk−1K
�k−1)
= On×n ,
in which since Kk−1 = �k−1Sk−1R−1k−1, we have
S�k−1�
�k−1 − Rk−1K
�k−1 = On×n .
4.6. Follow the same procedure in the derivation of Theorem4.1 with the term vkreplaced by vk − Dkuk , and with
xk|k−1 = L(Ak−1xk−1 + Bk−1uk−1 + �k−1ξk−1, vk−1)
instead of
xk|k−1 = L(xk, vk−1) = L(Ak−1xk−1 + �k−1ξk−1, vk−1) .
4.7. Let
wk = −a1vk−1 + b1uk−1 + c1ek−1 + wk−1 ,
wk−1 = −a2vk−2 + b2uk−2 + wk−2 ,
wk−2 = −a3vk−3 ,
and define xk = [ wk wk−1 wk−2 ]�. Then,{xk+1 =Axk + Buk + �ek
vk =Cxk + Duk + �ek ,
Answers and Hints to Exercises 227
where
A =⎡
⎣−a1 1 0−a2 0 1−a3 0 0
⎤
⎦ , B =⎡
⎣b1 − a1b0b2 − a2b0
−a3b0
⎤
⎦ , � =⎡
⎣c1 − a1c0−a2c0−a3b0
⎤
⎦ ,
C = [ 1 0 0 ] , D = [b0] and � = [c0] .4.8. Let
wk = −a1vk−1 + b1uk−1 + c1ek−1 + wk−1 ,
wk−1 = −a2vk−2 + b2uk−2 + c2ek−2 + wk−2 ,
· · · · · ·wk−n+1 = −anvk−n + bnuk−n + cnek−n ,
where b j = 0 for j > m and c j = 0 for j > �, and define
xk = [ wk wk−1 · · · wk−n+1 ]� .
Then{xk+1 = Axk + Buk + �ekvk = Cxk + Duk + �ek ,
where
A =
⎡
⎢⎢⎢⎢⎢⎣
−a1 1 0 · · · 0−a2 0 1 · · · 0
......
......
−an−1 0 0 · · · 1−an 0 0 · · · 0
⎤
⎥⎥⎥⎥⎥⎦,
B =
⎡
⎢⎢⎢⎢⎢⎢⎢⎢⎣
b1 − a1b0...
bm − amb0−am+1b0
...
−anb0
⎤
⎥⎥⎥⎥⎥⎥⎥⎥⎦
, � =
⎡
⎢⎢⎢⎢⎢⎢⎢⎢⎣
c1 − a1c0...
c� − a�c0−a�+1
...
−anc0
⎤
⎥⎥⎥⎥⎥⎥⎥⎥⎦
,
C = [ 1 0 · · · · · · 0 ] , D = [b0] , and � = [c0] .
228 Answers and Hints to Exercises
Chapter 5
5.1. Since vk is a linear combination (with constant matrices as coefficients) of
x0, η0, γ
0, · · · , γ
k, ξ
0, β
0, · · · , β
k−1
which are all independent of βk, we have
〈βk, vk〉 = 0 .
On the other hand, βkhas zero-mean, so that by (4.6) we have
L(βk, vk) = E(β
k) − 〈β
k, vk〉
[‖vk‖2
]−1(E(vk) − vk
) = 0 .
5.2. Using Lemma4.2 with v = vk−1, v1 = vk−2, v2 = vk−1 and
v#k−1 = vk−1 − L(vk−1, vk−2) ,
we have, for x = vk−1,
L(vk−1, vk−1)
= L(vk−1, vk−2) + 〈v#k−1, v#k−1〉[‖v#k−1‖2
]−1
v#k−1
= L(vk−1, vk−2) + vk−1 − L(vk−1, vk−2)
= vk−1 .
The equality L(γk, vk−1) = 0 can be shownby imitating the proof in Exercise5.1.
5.3. It follows from Lemma4.2 that
zk−1 − zk−1
= zk−1 − L(zk−1, vk−1)
= zk−1 − E(zk−1) + 〈zk−1, vk−1〉[‖vk−1‖2
]−1(E(vk−1) − vk−1)
=[xk−1ξk−1
]−[E(xk−1)
E(ξk−1
)
]
+[〈xk−1, vk−1〉〈ξ
k−1, vk−1〉
] [‖vk−1‖2
]−1(E(vk−1) − vk−1)
whose first n-subvector and last p-subvector are, respectively, linear combinations(with constant matrices as coefficients) of
x0, ξ0, β
0, · · · , β
k−2, η
0, γ
0, · · · , γ
k−1,
Answers and Hints to Exercises 229
which are all independent of γk. Hence, we have
B〈zk−1 − zk−1, γk〉 = 0 .
5.4. The proof is similar to that of Exercise5.3.5.5. For simplicity, denote
B = [C0Var(x0)C�0 + R0]−1 .
It follows from (5.16) that
Var(x0 − x0)
= Var(x0 − E(x0)
− [Var(x0)]C�0 [C0Var(x0)C�
0 + R0]−1(v0 − C0E(x0)))
= Var(x0 − E(x0) − [Var(x0)]C�0 B(C0(x0 − E(x0)) + η
0))
= Var((I − [Var(x0)]C�0 BC0)(x0 − E(x0)) − [Var(x0)]C�
0 Bη0)
= (I − [Var(x0)]C�0 BC0)Var(x0) (I − C�
0 BC0[Var(x0)])+ [Var(x0)]C�
0 BR0BC0[Var(x0)]= Var(x0) − [Var(x0)]C�
0 BC0[Var(x0)]− [Var(x0)]C�
0 BC0[Var(x0)]+ [Var(x0)]C�
0 BC0[Var(x0)]C�0 BC0[Var(x0)]
+ [Var(x0)]C�0 BR0BC0[Var(x0)]
= Var(x0) − [Var(x0)]C�0 BC0[Var(x0)]
− [Var(x0)]C�0 BC0[Var(x0)] + [Var(x0)]C�
0 BC0[Var(x0)]= Var(x0) − [Var(x0)]C�
0 BC0[Var(x0)] .
5.6. From ξ0
= 0, we have
x1 = A0x0 + G1(v1 − C1A0x0)
and ξ1
= 0, so that
x2 = A1x1 + G2(v2 − C2A1x1) ,
etc. In general, we have
xk = Ak−1xk−1 + Gk(vk − Ck Ak−1xk−1)
= xk|k−1 + Gk(vk − Ck xk|k−1) .
230 Answers and Hints to Exercises
Denote
P0,0 =[[Var(x0)
]−1 + C�0 R−1
0 C0
]−1
and
Pk,k−1 = Ak−1Pk−1,k−1A�k−1 + �k−1Qk−1�
�k−1 .
Then
G1 =[A0 �00 0
] [P0,0 00 Q0
] [A�0 C�
1��0 C�
1
]
·(
[ C1A0 C1�0 ][P0,0 00 Q0
] [A�0 C�
1��0 C�
1
]+ R1
)−1
=[P1,0C�
1 (C1P1,0C�1 + R1)
−1
0
],
P1 =([
A0 �00 0
]− G1[ C1A0 C1�0 ]
)[P0,0 00 Q0
][A�0 0
��0 0
]
+[0 00 Q1
]
=[[ In − P1,0C�
1 (C1P1,0C�1 + R1)
−1C1 ]P1,0 00 Q1
],
and, in general,
Gk =[Pk,k−1C�
k (Ck Pk,k−1C�k + Rk)
−1
0
],
Pk =[[ In − Pk,k−1C�
k (Ck Pk,k−1C�k + Rk)
−1Ck]Pk,k−1 00 Qk
].
Finally, if we use the unbiased estimate x0 = E(x0) of x0 instead of the somewhatmore superior initial state estimate
x0 = E(x0) − [Var(x0)]C�0 [C0Var(x0)C�
0 + R0]−1[C0E(x0) − v0] ,and consequently set
P0 =E
([x0ξ0
]−[E(x0)E(ξ
0)
])([x0ξ0
]−[E(x0)E(ξ
0)
])�
=[Var(x0) 0
0 Q0
],
then we obtain the Kalman filtering algorithm derived in Chaps. 2 and 3.
Answers and Hints to Exercises 231
5.7. Let
P0 = [ [Var(x0)]−1 + C�0 R−1
0 C0]−1
and
Hk−1 = [ Ck Ak−1 − Nk−1Ck−1 ] .Starting with (5.17b), namely:
P0 =[( [Var(x0)]−1 + C0R
−10 C0)
−1 00 Q0
]=[P0 00 Q0
],
we have
G1 =[A0 �00 0
] [P0 00 Q0
][H
�0
��0 C
�1
]
·(
[ H0 C1�0 ][P0 00 Q0
][H
�0
��0 C
�1
]+ R1
)−1
=[(
A0P0H�0 + �0Q0�
�0 C
�1
)(H0P0H
�0 + C1�0Q0�
�0 C
�1 + R1
)−1
0
]
:=[G10
]
and
P1 =([
A0 �00 0
]−[G10
][ H0 C1�0 ]
)[P0 00 Q0
][A�0 0
��0 0
]
+[0 00 Q1
]
=[(A0 − G1H0)P0A�
0 + (I − G1C1)�0Q0��0 0
0 Q1
]
:=[P1 00 Q1
].
In general, we obtain⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨
⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩
xk = Ak−1xk−1 + Gk(vk − Nk−1vk−1 − Hk−1xk−1)
x0 = E(x0) − [Var(x0)]C�0 [C0Var(x0)C�
0 + R0]−1[C0E(x0) − v0]Hk−1 = [ Ck Ak−1 − Nk−1Ck−1 ]Pk = (Ak−1 − Gk Hk−1)Pk−1A�
k−1 + (I − GkCk)�k−1Qk−1��k−1
Gk = (Ak−1 Pk−1 H�k−1 + �k−1Qk−1�
�k−1C
�k )·
(Hk−1 Pk−1 H�k−1 + Ck�k−1Qk−1�
�k−1C
�k + Rk−1)
−1
P0 = [ [Var(x0)]−1 + C�0 R−1
0 C0]−1
k = 1, 2, · · · .
232 Answers and Hints to Exercises
By omitting the “bar” on Hk , Gk , and Pk , we have (5.21).
5.8. (a){
Xk+1 = AcXk + ζk
vk = CcXk .
(b)
P0,0 =⎡
⎣Var(x0) 0 0
0 Var(ξ0) 0
0 0 Var(η0)
⎤
⎦ ,
Pk,k−1 = AcPk−1,k−1A�c +
⎡
⎢⎢⎢⎢⎣
000Qk−1
rk−1
⎤
⎥⎥⎥⎥⎦,
Gk = Pk,k−1C�c
(C�c Pk,k−1Cc
)−1,
Pk,k = (I − GkCc)Pk,k−1 ,
X0 =⎡
⎣E(x0)00
⎤
⎦ ,
X k = Ac Xk−1 + Gk(vk − CcAc Xk−1) .
(c) The matrix C�c Pk,k−1Cc may not be invertible, and the extra estimates ξ
kand
ηk in X k are needed.
Chapter 6
6.1. Since
xk−1 = Axk−2 + �ξk−2
= · · · = Anxk−n−1 + noise
and
xk−1 = An[N�CANCA]−1(C�vk−n−1 + A�C�vk−n
+ · · · + (A�)n−1C�vk−2)
= An[N�CANCA]−1(C�Cxk−n−1 + A�C�CAxk−n−1
+ · · · + (A�)n−1C�CAn−1xk−n−1 + noise)
= An[N�CANCA]−1[N�
CANCA]xk−n−1 + noise
= Anxk−n−1 + noise ,
we have E(xk−1) = E(Anxk−n−1) = E(xk−1).
Answers and Hints to Exercises 233
6.2. Sinced
ds
[A−1(s)A(s)
] = d
dsI = 0 ,
we have
A−1(s)
[d
dsA(s)
]+[d
dsA−1(s)
]A(s) = 0 .
Hence,
d
dsA−1(s) = −A−1(s)
[d
dsA(s)
]A−1(s) .
6.3. Let P = Udiag[ λ1, · · · , λn ]U−1. Then
P − λminI = Udiag[ λ1 − λmin, · · · , λn − λmin ]U−1 ≥ 0 .
6.4. Let λ1, · · · , λn be the eigenvalues of F and J be its Jordan canonical form.Then there exists a nonsingular matrix U such that
U−1FU = J =
⎡
⎢⎢⎢⎢⎢⎢⎣
λ1 ∗λ2 ∗
. . .. . .
. . . ∗λn
⎤
⎥⎥⎥⎥⎥⎥⎦
with each ∗ being 1 or 0. Hence,
Fk = U JkU−1 = U
⎡
⎢⎢⎢⎢⎢⎢⎣
λk1 ∗ · · · · · · ∗
λk2 ∗ · · · ∗
. . ....
. . . ∗λkn
⎤
⎥⎥⎥⎥⎥⎥⎦,
where each ∗ denotes a term whose magnitude is bounded by
p(k)|λmax |k
with p(k) being a polynomial of k and |λmax | = max( |λ1|, · · · , |λn| ). Since|λmax | < 1, Fk → 0 as k → ∞.
234 Answers and Hints to Exercises
6.5. Since
0 ≤ (A − B) (A − B)� = AA� − AB� − BA� + BB� ,
we have
AB� + BA� ≤ AA� + BB� .
Hence,
(A + B) (A + B)� = AA� + AB� + BA� + BB�
≤ 2(AA� + BB�) .
6.6. Since xk−1 = Axk−2+�ξk−2
is a linear combination (with constant matricesas coefficients) of x0, ξ0, · · · , ξ
k−2and
xk−1 = A xk−2 + G(vk−1 − CA xk−2)
= A xk−2 + G(CAxk−2 + C�ξk−2
+ ηk−1
) − GCA xk−2
is an analogous linear combination of x0, ξ0, · · · , ξk−2
and ηk−1
, which are uncor-related with ξ
k−1and η
k, the two identities follow immediately.
6.7. Since
Pk,k−1C�k G�
k − GkCk Pk,k−1C�k G�
k
=GkCk Pk,k−1C�k G�
k + Gk RkG�k − GkCk Pk,k−1C
�k G�
k
=Gk RkG�k ,
we have
−(I − GkC)Pk,k−1C�G�
k + Gk RG�k = 0 .
Hence,
Pk,k = (I − GkC)Pk,k−1
= (I − GkC)Pk,k−1(I − GkC)� + Gk RG�k
= (I − GkC) (APk−1,k−1A� + �Q��) (I − GkC)� + Gk RG
�k
= (I − GkC)APk−1,k−1A�(I − GkC)�
+ (I − GkC)�Q��(I − GkC)� + Gk RG�k .
6.8. Imitating the proof of Lemma6.8 and assuming that |λ| ≥ 1, where λ isan eigenvalue of (I − GC)A, we arrive at a contradiction to the controllabilitycondition.6.9. The proof is similar to that of Exercise6.6.
Answers and Hints to Exercises 235
6.10. From
0 ≤ 〈ε j − δ j , ε j − δ j 〉= 〈ε j , ε j 〉 − 〈ε j , δ j 〉 − 〈δ j , ε j 〉 + 〈δ j , δ j 〉
and Theorem6.2, we have
〈ε j , δ j 〉 + 〈δ j , ε j 〉≤ 〈ε j , ε j 〉 + 〈δ j , δ j 〉= 〈x j − x j + x j − x j , x j − x j + x j − x j 〉 + ‖x j − x j‖2n= ‖x j − x j‖2n + 〈x j − x j , x j − x j 〉
+ 〈x j − x j , x j − x j 〉 + 2‖x j − x j‖2n≤ 2‖x j − x j‖2n + 3‖x j − x j‖2n→ 5(P−1 + C�R−1C)−1
as j → ∞. Hence, Bj = 〈ε j , δ j 〉A�C� are componentwise uniformly bounded.6.11. Using Lemmas1.4, 1.6, 1.7 and 1.10 and Theorem6.1, and applyingExercise6.10, we have
tr[FBk−1−i (Gk−i − G)� + (Gk−i − G)B�k−1−i F
�]≤(n trFBk−1−i (Gk−i − G)�(Gk−i − G)B�
k−1−i F�)1/2
+ (n tr(Gk−i − G)B�k−1−i F
�FBk−1−i (Gk−i − G)�)1/2
≤(n trFF� · tr Bk−1−i B�k−1−i · tr(Gk−i − G)�(Gk−i − G))1/2
+ (n tr(Gk−i − G) (Gk−i − G)� · trB�k−1−i Bk−1−i · trF�F)1/2
=2(n tr(Gk−i − G) (Gk−i − G)� · trB�k−1−i Bk−1−i · trF�F)1/2
≤C1rk+1−i1
for some real number r1, 0 < r1 < 1, and some positive constant C independentof i and k.6.12. First, solving the Riccati equation (6.6); that is,
c2 p2 + [(1 − a2)r − c2γ2q]p − rqγ2 = 0 ,
we obtain
p = 1
2c2{c2γ2q + (a2 − 1)r +
√[(1 − a2)r − c2γ2q]2 + 4c2γ2qr } .
Then, the Kalman gain is given by
g = pc/(c2 p + r) .
236 Answers and Hints to Exercises
Chapter 7
7.1. The proof of Lemma7.1 is constructive. Let A = [ai j ]n×n and Ac = [�i j ]n×n .It follows from A = Ac(Ac)� that
aii =i∑
k=1
�2ik , i = 1, 2, · · · , n,
and
ai j =j∑
k=1
�ik� jk , j �= i ; i, j = 1, 2, · · · , n.
Hence, it can be easily verified that
�i i =(aii −
i−1∑
k=1
�2ik
)1/2, i = 1, 2, · · · , n,
�i j =(ai j −
j−1∑
k=1
�ik� jk
)/� j j , j = 1, 2, · · · , i − 1; i = 2, 3, · · · , n,
and
�i j = 0 , j = i + 1, i + 2, · · · , n; i = 1, 2, · · · , n.
This gives the lower triangular matrix Ac. This algorithm is called the Choleskydecomposition. For the general case,we can use a (standard) singular value decom-position (SVD) algorithm to find an orthogonal matrix U such that
U diag[s1, · · · , sr , 0, · · · , 0]U� = AA� ,
where 1 ≤ r ≤ n, s1, · · · , sr are singular values (which are positive numbers) ofthe non-negative definite and symmetric matrix AA�, and then set
A = U diag[√s1, · · · ,√sr , 0, · · · , 0] .
7.2.
(a) L =⎡
⎣1 0 02 2 03 −2 1
⎤
⎦ . (b) L =⎡
⎣
√2 0 0√2/2
√2.5 0√
2/2 1.5/√2.5
√2.6
⎤
⎦ .
7.3.(a)
L−1 =⎡
⎣1/�11 0 0
−�21/�11�22 1/�22 0−�31/�11�33 + �32�21/�11�22�33 −�32/�22�33 1/�33
⎤
⎦ .
Answers and Hints to Exercises 237
(b)
L−1 =
⎡
⎢⎢⎢⎣
b11 0 0 · · · 0b21 b22 0 · · · 0...
...... 0
bn1 bn2 bn3 · · · bnn
⎤
⎥⎥⎥⎦ ,
where⎧⎪⎨
⎪⎩
bii = �−1i i , i = 1, 2, · · · , n;
bi j = −�−1j j
∑ik= j+1 bik�k j ,
j = i − 1, i − 2, · · · , 1; i = 2, 3, · · · , n.
7.4. In the standard Kalman filtering process,
Pk,k �[0 00 1
],
which is a singular matrix. However, its “square-root” is
P1/2k,k =
[ε/
√1 − ε2 00 1
]�[ε 00 1
]
which is a nonsingular matrix.7.5. Analogous to Exercise7.1, let A = [ai j ]n×n and Au = [�i j ]n×n . It followsfrom A = Au(Au)� that
aii =n∑
k=i
�2ik , i = 1, 2. · · · , n,
and
ai j =n∑
k= j
�ik� jk , j �= i ; i, j = 1, 2, · · · , n.
Hence, it can be easily verified that
�i i =(aii −
n∑
k=i+1
�2ik
)1/2, i = 1, 2, · · · , n,
�i j =(ai j −
n∑
k= j+1
�ik� jk
)/� j j ,
j = i + 1, · · · , n; i = 1, 2, · · · , n.
and
�i j = 0 , j = 1, 2, · · · , i − 1; i = 2, 3, · · · , n.
This gives the upper-triangular matrix Au .
238 Answers and Hints to Exercises
7.6. The new formulation is the same as that studied in this chapter except thatevery lower triangular matrix with superscript c must be replaced by the corre-sponding upper triangular matrix with superscript u.7.7. The new formulation is the same as that given in Sect. 7.3 except that all lowertriangular matrix with superscript c must be replaced by the corresponding uppertriangular matrix with superscript u.
Chapter 8
8.1. (a) Since r2 = x2 + y2, we have
r = x
rx + y
ry ,
so that r = v sinθ and
r = v sinθ + vθ cosθ .
On the other hand, since tanθ = y/x , we have θsec2θ = (x y − x y)/x2 or
θ = x y − x y
x2 sec2θ= x y − x y
r2= v
rcosθ ,
so that
r = a sinθ + v2
rcos2θ
and
θ =(
vr − vr
r2
)cosθ − v
rθsinθ
=(ar − v2sinθ
r2
)cosθ − v2
r2sinθcosθ .
(b)
x = f(x) :=⎡
⎣v sinθ
a sinθ + v2
r cos2θ
(ar − v2sinθ)cosθ/r2 − v2sinθ cosθ/r2
⎤
⎦ .
(c)
xk+1 =
⎡
⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
xk[1] + hv sin(xk[3])
xk[2] + ha sin(xk[3]) + v2cos2(xk[3])/xk[1]
vcos(xk[3])/xk[1]
(axk[1] − v2sin(xk[3]))cos(xk[3])/xk[1]2−v2sin(xk[3])cos(xk[3])/xk[1]2
⎤
⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
+ ξk
Answers and Hints to Exercises 239
and
vk = [ 1 0 0 0 ]xk + ηk ,
where xk := [ xk[1] xk[2] xk[3] xk[4] ]�.(d) Use the formulas in (8.8).8.2. The proof is straightforward.8.3. The proof is straightforward. It can be verified that
xk|k−1 = Ak−1xk−1 + Bk−1uk−1 = fk−1(xk−1) .
8.4. Taking the variances of both sides of the modified “observation equation”
v0 − C0(θ)E(x0) = C0(θ)x0 − C0(θ)E(x0) + η0,
and using the estimate (v0 − C0(θ)E(x0))(v0 − C0(θ)E(x0))� for Var(v0 −C0(θ)E(x0)) on the left-hand side, we have
(v0 − C0(θ)E(x0))(v0 − C0(θ)E(x0))�
=C0(θ)Var(x0)C0(θ)� + R0 .
Hence, (8.13) follows immediately.8.5. Since
E(v1) = C1(θ)A0(θ)E(x0) ,
taking the variances of both sides of the modified “observation equation”
v1 − C1(θ)A0(θ)E(x0)
=C1(θ)(A0(θ)x0 − C1(θ)A0(θ)E(x0) + �(θ)ξ0) + η
1,
and using the estimate (v1 − C1(θ)A0(θ)E(x0))(v1 − C1(θ)A0(θ) ·E(x0))� forthe variance Var(v1 − C1(θ)A0(θ)E(x0)) on the left-hand side, we have
(v1 − C1(θ)A0(θ)E(x0))(v1 − C1(θ)A0(θ)E(x0))�
=C1(θ)A0(θ)Var(x0)A�0 (θ)C�
1 (θ) + C1(θ)�0(θ)Q0��0 (θ)C�
1 (θ) + R1 .
Then (8.14) follows immediately.8.6. Use the formulas in (8.8) directly.8.7. Since θ is a constant vector, we have Sk := Var(θ) = 0, so that
P0,0 = Var(xθ
) =[Var(x0) 0
0 0
].
240 Answers and Hints to Exercises
It follows from simple algebra that
Pk,k−1 =[∗ 00 0
]and Gk =
[∗0
]
where ∗ indicates a constant block in the matrix. Hence, the last equation of (8.15)yields θk|k ≡ θk−1|k−1.8.8.
⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨
⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩
[x0c0
]=[x0
c0
], P0,0 =
[p0 0
0 s0
]
For k = 1, 2, · · · ,
Pk,k−1 = Pk−1,k−1 +[qk−1 0
0 sk−1
]
Gk = Pk,k−1
[ck−1
0
][[ck−1 0]Pk,k−1
[ck−1
0
]+ rk
]−1
Pk,k = [I − Gk[ck−1 0]]Pk,k−1[xkck
]=[xk−1
ck−1
]+ Gk(vk − ck−1 xk−1) ,
where c0 is an estimate of c0 given by (8.13); that is,
v20 − 2v0x0c0 + [(x0)2 − p0](c0)2 − r0 = 0 .
Chapter 9
9.1. (a) Let xk = [xk xk]�. Then⎧⎨
⎩
xk = −(α + β − 2)xk−1 − (1 − α)xk−2 + αvk + (−α + β)vk−1
xk = −(α + β − 2)xk−1 − (1 − α)xk−2 + β
hvk − β
hvk−1 .
(b) 0 < α < 1 and 0 < β < α2
1−α .9.2. System (9.11) follows from direct algebraic manipulation.9.3. (a)
� =
⎡
⎢⎢⎣
1 − α (1 − α)h (1 − α)h2/2 −sα−β/h 1 − β h − βh/2 −sβ/h−γ/h2 1 − γ/h 1 − γ/2 −sγ/h2
−θ −θ/h −θh2/2 s(1 − θ)
⎤
⎥⎥⎦
Answers and Hints to Exercises 241
(b)
det[z I − �] =z4 + [(α − 3) + β + γ/2 − (θ − 1)s]z3+ [(3 − 2α) − β + γ/2 + (3 − α − β − γ/2 − 3θ)s]z2+ [(α − 1) − (3 − 2α − β + γ/2 − 3θ)s]z + (1 − α − θ)s .
X1 = zV (z − s)
det[z I − �] {αz2 + (γ/2 + β − 2α)z + (γ/2 − β + α)} ,
X2 = zV (z − 1)(z − s)
det[z I − �] {βz − β + γ}/h ,
X3 = zV (z − 1)2(z − s)
det[z I − �] γ/h2 ,
and
W = zV (z − 1)3
det[z I − �]θ .
(c) Let Xk = [ xk xk xk wk ]�. Then
xk = a1xk−1 + a2xk−2 + a3xk−3 + a4xk−4 + αvk
+ (−2α − sα + β + γ/2)vk−1 + [α − β + γ/2
+ (2α − β − γ/2)s]vk−2 − (α − β + γ/2)svk−3 ,
xk = a1 xk−1 + a2 xk−2 + a3 xk−3a4 xk−4 + (β/h)vk
− [(2 + s)β/h − γ/h]vk−1 + [β/h − γ/h
+ (2β − γ)s/h]vk−2 − [(β − γ)s/h]vk−3 ,
xk = a1 xk−1 + a2 xk−2 + a3 xk−3 + a4 xk−4 + (γ/h)vk
− [(2 + γ)γ/h2]vk−1 + (1 + 2s)vk−2 − svk−3 ,
wk = a1wk−1 + a2wk−2 + a3wk−3 + a4wk−4
+ (γ/h2)(vk − 3vk−1 + 3vk−2 − vk−3) ,
with the initial conditions x−1 = x−1 = x−1 = w0 = 0, where
a1 = −α − β − γ/2 + (θ − 1)s + 3 ,
a2 = 2α + β − γ/2 + (α + βh + γ/2 + 3θ − 3)s − 3 ,
a3 = −α + (−2α − β + γ/2 − 3θ + 3)s + 1 ,
and
a4 = (α + θ − 1)s .
242 Answers and Hints to Exercises
(d) The verification is straightforward.9.4. The verifications are tedious but elementary.9.5. Study (9.19) and (9.20). We must have σp,σv, σa ≥ 0, σm > 0, and P > 0.9.6. The equations can be obtained by elementary algebraic manipulation.9.7. Only algebraic manipulation is required.
Chapter 10
10.1. For (1) and (4), let ∗ ∈ {+,−, · , /}. Then
X ∗ Y = {x ∗ y|x ∈ X, y ∈ Y
}
= {y ∗ x |y ∈ Y, x ∈ X
}
= Y ∗ X .
The others can be verified in a similar manner. As to part (c) of (7), without lossof generality, we may only consider the situation where both x ≥ 0 and y ≥ 0 inX = [x, x] and Y = [y, y], and then discuss different cases of z ≥ 0, z ≤ 0, andzz < 0.10.2. It is straightforward to verify all the formulas by definition. For instance,for part (j.1), we have
AI (BC) =⎡
⎣n∑
j=1
AI (i, j)
[n∑
�=1
Bj�C�k
]⎤
⎦
⊆⎡
⎣n∑
j=1
n∑
�=1
AI (i, j)Bj�C�k
⎤
⎦
=⎡
⎣n∑
�=1
⎡
⎣n∑
j=1
AI (i, j)Bj�
⎤
⎦C�k
⎤
⎦
= (AI B)C .
10.3. See: Alefeld, G. and Herzberger, J. (1983).10.4. Similar to Exercise1.10.10.5. Observe that the filtering results for a boundary system and any of its neigh-boring system will be inter-crossing from time to time.10.6. See: Siouris, G., Chen, G. and Wang, J. (1997).
Answers and Hints to Exercises 243
Chapter 11
11.1.
φ2(t) =
⎧⎪⎪⎪⎪⎪⎪⎪⎪⎨
⎪⎪⎪⎪⎪⎪⎪⎪⎩
1
2t2 0 ≤ t < 1
−t2 + 3t − 3
21 ≤ t < 2
1
2t2 − 3 t + 9
22 ≤ t < 3
0 otherwise.
φ3(t) =
⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨
⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩
1
6t3 0 ≤ t < 1
−1
2t3 + 2t2 − 2t + 2
31 ≤ t < 2
1
2t3 − 4t2 + 10t − 22
32 ≤ t < 3
−1
6t3 + 2t2 − 8t + 32
33 ≤ t < 4
0 otherwise.
11.2.
φn(ω) =(1 − e−iω
iω
)n
= e−inω/2(sin(ω/2)
ω/2
)n
.
11.3. Simple graphs.11.4. Straightforward algebraic operations.11.5. Straightforward algebraic operations.
Chapter 12
12.1 Apply matrix analysis (see Cattivelli and Sayed 2010).12.2 Straightforward algebraic operations.
Index
AAdaptive Kalman Filtering, 202
noise-adaptive filter, 202Adaptive System Identification, 120, 122Affine model, 51Algorithm for real-time application, 110α − β tracker, 148α − β − γ tracker, 144, 145, 147α − β − γ − θ tracker, 148, 149Angular displacement, 118, 135ARMA (autoregressive moving-average)
process, 30ARMAX (auto-regressive moving-average
model with exogeneous inputs), 68Attracting point, 118Augmented matrix, 207Augmented system, 79Azimuthal angular error, 47
BBayes formula, 12Bound (lower, upper), 167
CCholesky factorization, 107Colored noise (sequence or process), 69,
78, 148Conditional probability, 11Controllability matrix, 89Controllable linear system, 89
Correlated system and measurement noiseprocesses, 51
Covariance, 13Cramer’s rule, 141
DDecoupling formulas, 139Decoupling of filtering equation, 139Descartes rule of signs, 147Determinant preliminaries, 1Deterministic input sequence, 19Digital filtering process, 22Digital prediction process, 22Digital smoothing estimate, 197Digital smoothing process, 22
EElevational angular error, 47Estimate, 16
distributed state estimate, 185least-squares optimal estimate, 17linear estimate, 17minimum trace variance estimate, 54minimum variance estimate, 17, 37, 52optimal estimate, 17operator, 53
unbiased estimate, 17, 52Event (simple), 8Expectation, 9
conditional expectation, 14Extended Kalman filter, 115, 118, 120
© Springer International Publishing AG 2017C.K. Chui and G. Chen, Kalman Filtering, DOI 10.1007/978-3-319-47612-4
245
246 Index
FFIR system, 204
GGaussian white noise sequence, 15, 121Geometric convergence, 92
IIIR system, 204Independent random variables, 13Innovations sequence, 35Inverse z-transform, 141
JJoint probability distribution (function), 10Jordan canonical form, 5, 7
KKalman–Bucy filter, 204Kalman filter, 19, 26, 33
extended Kalman filter, 115, 118, 120interval Kalman filter, 161limiting Kalman filter, 81, 82modified extended Kalman filter, 125steady-state Kalman filter, 82, 139, 189wavelet Kalman filter, 171
Kalman filtering equation (algorithm, orprocess), 26, 29, 38, 42, 57, 66,73–75, 79, 113
Kalman gain matrix, 24Kalman smoother, 197
LLeast-squares preliminaries, 15Limiting (or steady-state) Kalman filter, 81Limiting Kalman gain matrix, 82Linear deterministic/stochastic system, 19,
43, 65, 198, 204Linear regulator problem, 205Linear state-space (stochastic) system, 21,
33, 69, 81, 202, 205LU decomposition, 207
MMarginal probability density function, 10Matrix inversion lemma, 3Matrix Riccati equation, 83, 99, 139, 142
Matrix Schwarz inequality, 2, 17Minimum variance estimate, 17, 37, 52Modified extended Kalman filter, 125Moment, 10
NNonlinear model (system), 115Non-negative definite matrix, 1Normal distribution, 9Normal white noise sequence, 15
OObservability matrix, 83Observable system, 88Optimal estimate, 17
asymptotically optimal estimate, 94least-squares optimal estimate, 17optimal estimate operator, 55
Optimal prediction, 22, 202Optimal weight matrix, 17Optimality criterion, 20Outcome, 8
PParallel processing, 207Parameter identification, 123
adaptive parameter identificationalgorithm, 123
Positive definite matrix, 1Positive square root matrix, 16Prediction-correction, 22, 24, 29, 40, 82Probability density function, 9
conditional probability density function,12
Gaussian (or normal) probability densityfunction, 9, 11
joint probability density function, 12Probability distribution, 8, 10
function, 8joint probability distribution (function),
10Probability preliminaries, 8
RRadar tracking model (or system), 47, 63,
194Random sequence, 15Random signal, 177
Index 247
Random variable, 8independent random variables, 13uncorrelated random variables, 13
Random vector, 10Range, 47, 116Real-time application, 63, 76, 98, 110Real-time estimation/decomposition, 177Real-time tracking, 43, 76, 99, 142, 147
SSample space, 8Satellite orbit estimation, 118Schur complement technique, 207Schwarz inequality, 2
matrix Schwarz inequality, 2, 17vector Schwarz inequality, 2
Separation principle, 206Sequential algorithm, 101Square-root algorithm, matrix, 16, 101, 107Stabilizable system, 190Steady-state (or limiting) Kalman filter, 82Stochastic optimal control, 205Suboptimal filter, 144Systolic array, 206
implementation, 206
TTaylor approximation, 47, 126Trace, 5
UUncorrelated random variables, 13
VVariance, 10
conditional variance, 14
WWavelets, 171Weight matrix, 15
optimal weight matrix, 16White noise sequence (process), 15, 20, 69
Gaussian (or normal) white noisesequence, 15, 69
zero-mean Gaussian white noisesequence, 15
Wiener filter, 203Wireless sensor network (WSN), 185
Zz-transform, inverse, 140, 141