Abstract
We propose a new approach to filtering under arbitrary bounded exogenous disturbances based on reducing this problem to an optimization problem. The approach has a low computational complexity since only Lyapunov equations are solved at each iteration. At the same time, it possesses advantages essential from an engineering-practical point of view, namely, the possibilities to limit the filter matrix and to construct optimal filter matrices separately for each coordinate of the system’s state vector. A gradient method for finding the filter matrix is presented. According to the examples, the proposed recurrence procedure is rather effective and yields quite satisfactory results. This paper continues the series of research works devoted to feedback control design from an optimization perspective.
Notes
It actually takes 3–4 iterations to obtain a solution with high accuracy if the starting point is not too close to the boundaries of the interval (ρ2(A – Lj + 1C), 1).
REFERENCES
Kalman, R.E., A New Approach to Linear Filtering and Prediction Problems, J. Basic Engineer., 1960, vol. 82, no. 1, pp. 35–45.
Kailath, T., Sayed, A.H., and Hassibi, B., Linear Estimation, New Jersey: Prentice Hall, 2000.
Matasov, A.I., Osnovy teorii fil’tra Kalmana (Foundations of Kalman Filter Theory), Moscow: Mosk. Gos. Univ., 2021.
Schweppe, F.C., Uncertain Dynamic Systems, New Jersey: Prentice Hall, 1973.
Kurzhanskii, A.B., Upravlenie i nablyudenie v usloviyakh neopredelennosti (Control and Observation under Uncertainty), Moscow: Nauka, 1977.
Chernous’ko, State Estimation for Dynamic Systems, Boca Raton: CRC Press, 1994.
Polyak, B.T. and Topunov, M.V., Filtering under Nonrandom Disturbances: The Method of Invariant Ellipsoids, Dokl. Math., 2008, vol. 77, no. 1, pp. 158–162.
Khlebnikov, M.V. and Polyak, B.T., Filtering under Arbitrary Bounded Exogenous Disturbances: The Technique of Linear Matrix Inequalities, The 13th Multiconference on Control Problems (MCCP 2020), Proceedings of the 32nd Conference in Memory of Nikolay Ostryakov, St. Petersburg, October 6–8, 2020, Concern CSRI Elektropribor, pp. 291–294.
Boyd, S., El Ghaoui, L., Feron, E., and Balakrishnan, V., Linear Matrix Inequalities in System and Control Theory, Philadelphia: SIAM, 1994.
Polyak, B.T., Khlebnikov, M.V., and Shcherbakov, P.S., Upravlenie lineinymi sistemami pri vneshnikh vozmu-shcheniyakh: tekhnika lineinykh matrichnykh neravenstv (Control of Linear Systems under Exogenous Disturbances: The Technique of Linear Matrix Inequalities), Moscow: LENAND, 2014.
Fazel, M., Ge, R., Kakade, S., and Mesbahi, M., Global Convergence of Policy Gradient Methods for the Linear Quadratic Regulator, Proc. 35th Int. Conf. Machine Learning, Stockholm, July 10–15, 2018, vol. 80, pp. 1467–1476.
Mohammadi, H., Zare, A., Soltanolkotabi, M., and Jovanović, M.R., Global Exponential Convergence of Gradient Methods over the Nonconvex Landscape of the Linear Quadratic Regulator, Proc. 2019 IEEE 58th Conf. Decision Control, Nice, December 11–13, 2019, pp. 7474–7479.
Zhang, K., Hu, B., and Başar, T., Policy Optimization for \({{\mathcal{H}}_{2}}\) Linear Control with \({{\mathcal{H}}_{\infty }}\) Robustness Guarantee: Implicit Regularization and Global Convergence, Proc. 2nd Conference on Learning for Dynamics and Control (2nd L4DC), Zürich, June 11–12, 2020, pp. 179–190.
Bu, J., Mesbahi, A., Fazel, M., and Mesbahi, M., LQR through the Lens of First Order Methods: Discrete-Time Case, arXiv:1907.08921, 2019.
Fatkhullin, I. and Polyak, B., Optimizing Static Linear Feedback: Gradient Method, SIAM J. Control Optim., 2021, vol. 59, no. 5, pp. 3887–3911.
Polyak, B.T. and Khlebnikov, M.V., Static Controller Synthesis for Peak-to-Peak Gain Minimization as an Optimization Problem, Autom. Remote Control, 2021, vol. 82, no. 9, pp. 1530–1553.
Polyak, B.T. and Khlebnikov, M.V., Observer-Aided Output Feedback Synthesis as an Optimization Problem, Autom. Remote Control, 2022, vol. 83, no. 3, pp. 303–324.
Polyak, B.T. and Khlebnikov, M.V., New Criteria for Tuning PID Controllers, Autom. Remote Control, 2022, vol. 83, no. 11, pp. 1724–1741.
Luenberger, D.G., Observing the State of a Linear System, IEEE Transactions on Military Electronics, 1964, vol. 8, pp. 74–80.
Luenberger, D.G., An Introduction to Observers, IEEE Trans. Autom. Control, 1971, vol. 35, pp. 596–602.
Polyak, B.T., Khlebnikov, M.V., and Shcherbakov, P.S., Linear Matrix Inequalities in Control Systems with Uncertainty, Autom. Remote Control, 2021, vol. 82, no. 1, pp. 1–40.
Nazin, S.A., Polyak, B.T., and Topunov, M.V., Rejection of Bounded Exogenous Disturbances by the Method of Invariant Ellipsoids, Autom. Remote Control, 2007, vol. 68, no. 3, pp. 467–486.
en.wikipedia.org/wiki/Kalman\_filter.
Humpherys, J., Redd, P., and West, J., A Fresh Look at the Kalman Filter, SIAM Rev., 2012, vol. 54, no. 4, pp. 801–823.
Tang, W., Zhang, Q., Wang, Z., and Shen, Y., Ellipsoid Bundle and Its Application to Set-Membership Estimation, IFAC-PapersOnLine, 2020, vol. 53, no. 2, pp. 13688–13693.
Tang, W., Zhang, Q., Wang, Z., and Shen, Y., Set-Membership Filtering with Incomplete Observations, Inform. Sci., 2020, vol. 517, pp. 37–51.
Polyak, B.T., Nazin, S.A., Durieu, C., and Walter, E., Ellipsoidal Parameter or State Estimation under Model Uncertainty, Automatica, 2004, vol. 40, no. 7, pp. 1171–1179.
Durieu, C., Walter, E., and Polyak, B., Multi-Input Multi-Output Ellipsoidal State Bounding, J. Optim. Theory Appl., 2001, vol. 111, no. 2, pp. 273–303.
Kwon, W.H., Moon, Y.S., and Ahn, S.C., Bounds in Algebraic Riccati and Lyapunov Equations: A Survey and Some New Results, Int. J. Control, 1996, vol. 64, pp. 377–389.
Funding
This work was partially financially supported by the Russian Science Foundation, project no. 21-71-30005, https://rscf.ru/en/project/21-71-30005/.
Author information
Authors and Affiliations
Corresponding author
Additional information
This paper is dedicated to the blessed memory of Boris Polyak, the author’s teacher and friend
This paper was recommended for publication by E.Ya. Rubinovich, a member of the Editorial Board
Appendices
APPENDIX A
Lemma A.1. Let X and Y be the solutions of the dual discrete Lyapunov equations with a Schur matrix A:
Then tr(XV) = tr(YW).
Proof of Lemma A.1. Indeed, direct calculations give
The proof of Lemma A.1 is complete.
Lemma A.2. The solution P of the discrete Lyapunov equation
with a Schur matrix A and Q \( \succ \) 0 satisfies the lower bounds
where ρ = \(\mathop {\max }\limits_i \left| {{{\lambda }_{i}}(A)} \right|\) and σmin(A) is the smallest singular value of the matrix A.
If Q = DDT and the pair (A, D) is controllable, then
where
i.e., u is the left eigenvector of the matrix A corresponding to the eigenvalue λ of the matrix A with the greatest magnitude. The vector u and the value λ can be complex; here, u* denotes the conjugate transpose of u.
Proof of Lemma A.2. The lower bounds (A.1) are well known; for example, see [29]. Let us prove (A.2). The explicit solution of the discrete Lyapunov equation with a Schur matrix A has the form
Multiplying this equality by u on the right and by u* on the left, due to u*Ak = λku* and (AT)ku = (λ*)ku, we obtain
where ||u*D|| > 0 by the controllability of the pair (A, D); for example, see [10, Theorem D.1.5]. The proof of Lemma A.2 is complete.
Now, we optimize the function f(α) and consider the problem
subject to the constraint
for the matrix variables P = PT ∈ \({{\mathbb{R}}^{{n \times n}}}\) and a scalar parameter 0 < α < 1.
Here we impose more stringent requirements for the problem statement: the matrix C of the system output is supposed to be square and nonsingular. This assumption could be relaxed, but the current goal is to establish the simplest and most obvious results.
Lemma A.3. Assume that A is a Schur matrix, ρ is the spectral radius of A, ρ2 < α < 1, the pair (A, D) is controllable, and the matrix C is such that CTC \( \succ \) 0. Then the function f(α) = tr CP(α)C T possesses the following properties:
(a) The function f(α) is well-defined, positive, and strongly convex on the interval ρ2 < α < 1 and its values tend to infinity at the interval endpoints. Moreover, there exists a constant c > 0 such that
(b) The function f(α) has the derivative
where P and Y are the solutions of the discrete Lyapunov equations
and
respectively.
(c) The second derivative of the function f(α) is given by
where P, Y, and X satisfy the discrete Lyapunov equations (A.4), (A.5), and
respectively. Moreover, f ''(α*) > 0 and f ''(α) is monotonically increasing on the left and right of α*.
Proof of Lemma A.3. (a) Equation (6) can be written as
according to [10, Lemma 1.2.6], there exists a unique solution if and only if \(\frac{1}{{\sqrt \alpha }}A\) is a Schur matrix: \(\left| {{{\lambda }_{i}}\left( {\frac{1}{{\sqrt \alpha }}A} \right)} \right|\) < 1, i.e., under the condition ρ2 < α < 1.
We estimate the value f(α) = tr CP(α)C T using Lemma A.2 with obvious changes:
Here, u has the same meaning as in Lemma A.2 and the value ||u*D||2 is positive by the controllability of the pair (A/\(\sqrt \alpha \), D). (It follows from the controllability of (A, D).)
Now we show that the function f(α) = tr CP(α)C T is strictly convex on the interval (ρ2, 1). According to [10, Lemma 1.2.6], the solution of Eq. (A.4) can be explicitly represented as
But Hk \( \succ \) 0 and g(α, k) > 0 for 0 < α < 1; therefore, on the interval (ρ2, 1) we have
and
Direct calculations give
(Here, differentiation is performed with respect to α.) As a result,
Thus, the second derivative of the function f(α) is positive and tends to infinity at the endpoints of the interval (ρ2, 1).
Next, with direct calculations of the fourth derivative, we obtain
so
i.e., the second derivative f ''(α) is convex and grows at the interval endpoints.
(b) Let us derive the formula for the derivative of f(α). In Eq. (A.4), the solution P is a function of α. We differentiate this equation, interpreting P ' as the derivative with respect to α:
Applying Lemma A.1 to the dual Eqs. (A.6) and (A.5) finally yields
(c) The desired expression for the second derivative of f(α) can be established by analogy. Differentiating Eq. (A.6) with respect to α, we have
Applying Lemma A.1 to this equation and Eq. (A.5) with X = P', we arrive at
The proof of Lemma A.3 is complete.
Note that the function f(α) and its two derivatives are calculated by solving three discrete Lyapunov equations.
Due to the above properties, this function can be minimized using Newton’s method. We specify an initial approximation ρ2(A) < α0 < 1, e.g., α0 = (1 + ρ2(A))/2, and apply the iterative process
The next theorem ensures the global convergence of the algorithm; it can be proved by analogy with a similar result in [16].
Theorem A.1 [16]. In the method (A.7), we have the upper bounds
where c > 0 is some constant (possibly, in explicit form).
The first bound ensures the global convergence of the method (faster than a geometric progression with a coefficient of 1/2); the second bound, the quadratic convergence in the neighborhood of the solution. In practice, it takes 3–4 iterations to obtain a solution with a high accuracy (unless the starting point is too close to the interval endpoints).
Returning to the optimization problem (4)–(5), we minimize the function
after a preliminary study of its properties.
Lemma A.4. The function f(L) is well-defined and positive on the set \(\mathcal{S}\) of admissible filter matrices.
Indeed, if (A – LC) is a Schur matrix, then ρ(A – LC) < 1 and for ρ2(A – LC) < α < 1, there exists a solution P \( \succcurlyeq \) 0 of the discrete Lyapunov equation (5). Thus, a strictly positive function f(L, α) is well-defined and f(L) > 0 due to (A.3). As in the continuous-time case, its domain \(\mathcal{S}\) may be nonconvex and disconnected and its boundaries nonsmooth.
Lemma A.5. On the set \(\mathcal{S}\) the function f(L) is coercive (i.e., it tends to infinity on the boundary of the domain). Moreover, the following lower bounds are valid:
Proof of Lemma A.5. We consider a sequence {Lj} ⊆ \(\mathcal{S}\) of admissible matrices such that Lj → L ∈ \(\partial \mathcal{S}\), i.e., ρ(A – LC) = 1. In other words, for any ε > 0 there exists a number N = N(ε) such that
for all j \( \geqslant \) N(ε).
Let Pj be the solution of Eq. (5) associated with the filter matrix Lj:
let Yj be the solution of its dual discrete Lyapunov equation
In view of Lemma A.2, we have
since ρ2(A – LjC) < αj < 1.
On the other hand,
The proof of Lemma A.5 is complete.
We introduce the level set
Obviously, Lemma A.5 implies the following result.
Corollary A.1. For any L0 ∈ \(\mathcal{S}\), the set \({{\mathcal{S}}_{0}}\) is bounded.
On the other hand, the function f(L) achieves minimum on the set \({{\mathcal{S}}_{0}}\). (This function is continuous by the properties of the solution of the discrete Lyapunov equation and is considered on a compact set.) However, the set \({{\mathcal{S}}_{0}}\) has no common points with the boundary of \(\mathcal{S}\) due to (A.8). The function f(L) is differentiable on \({{\mathcal{S}}_{0}}\); see below. Consequently, we arrive at the following result.
Corollary A.2. There exists a minimum point \({{L}_{*}}\) on the set \(\mathcal{S}\), and the gradient of the function f(L) vanishes at this point.
Let us analyze the properties of the gradient of the function f(L, α).
Lemma A.6. The function f(L, α) is defined on the set of stabilizing L for ρ2(A – LC) < α < 1.
On this admissible set, the function is differentiable, and its gradient is given by
where the matrices P and Y are the solutions of the discrete Lyapunov equations (5) and (8), respectively.
The minimum of f(L, α) is achieved at an inner point of the admissible set and is determined by the conditions
In addition, f(L, α) as a function of α is strictly convex on ρ2(A – LC) < α < 1 and achieves minimum at an inner point of this interval.
Proof of Lemma A.6. We have the constrained optimization problem
subject to the discrete Lyapunov equation (5) for the matrix P of the invariant ellipsoid.
Following Lemma A.3, differentiation with respect to a is performed using the relations (A.9), (5), and (8). To differentiate with respect to L, we add an increment ΔL and denote by ΔP the corresponding increment of P. As a result, the relation (5) takes the form
Leaving the notation ΔP for the principal terms of the increment, we obtain
Subtracting Eq. (12) from this equation yields
We calculate the increment of the functional f(L) by linearizing the corresponding values:
By Lemma B.1, from the dual Eqs. (A.11) and (8) we have
Thus, the relation (A.10) is derived and the proof of Lemma A.6 is complete.
The gradient of the function f(L) is not Lipschitz on the set \(\mathcal{S}\). But it can be shown to possess this property on the subset \({{\mathcal{S}}_{0}}\) (similar to [16]).
These properties of the minimized function and its derivatives justify the minimization method implemented as Algorithm 1.
APPENDIX B
Lemma B.1 [16]. Let X and Y be the solutions of the dual Lyapunov equations with a Hurwitz matrix A:
Then tr(XV) = tr(YW).
The properties of the function f(α) established in [16] fully apply to the case under consideration. In particular, the function f(α) is well-defined, positive, and strongly convex on the interval 0 < α < 2σ(A – LC) and its values tend to infinity at the interval endpoints. Moreover, there exists a constant c > 0 such that
The function f(α) can be effectively minimized using Newton’s method. We specify an initial approximation 0 < α0 < 2σ(A – LC), e.g., α0 = σ(A – LC), and apply the iterative process
where, according to [16],
and P, Y, and X are the solutions of the Lyapunov equations (12), (13), and (14), respectively. Theorem A.1 remains valid as well.
The following lemma is a continuous-time analog of Lemma A.4.
Lemma B.2. The function f(L) is well-defined and positive on the set \(\mathcal{S}\) of admissible filter matrices.
Indeed, if (A – LC) is a Hurwitz matrix, then σ(A – LC) > 0 and for 0 < α < 2σ(A – LC), there exists a solution P \( \succcurlyeq \) 0 of the Lyapunov equation (12). Thus, a strictly positive function f(L, α) is well-defined and f(L) > 0 due to (B.1). Its domain \(\mathcal{S}\) may be nonconvex and disconnected and its boundaries nonsmooth; see [16].
Lemma B.3. On the set \(\mathcal{S}\) of admissible matrices the function f(L) is coercive (i.e., it tends to infinity on the boundary of the domain). Moreover, the following lower bounds are valid:
Proof of Lemma B.3. We consider a sequence {Lj} ⊆ \(\mathcal{S}\) of admissible matrices such that Lj → L ∈ \(\partial \mathcal{S}\), i.e., σ(A – LC) = 0. In other words, for any ε > 0 there exists a number N = N(ε) such that
for all j \( \geqslant \) N(ε).
Let Pj be the solution of equation (12) associated with the filter matrix Lj:
let Yj be the solution of its dual Lyapunov equation
In view of [16, Lemma A.3], we have
since 0 < αj < 2σ(A – LjC) and
On the other hand,
The proof of Lemma B.3 is complete.
We introduce the level set
Obviously, Lemma B.3 implies the following result.
Corollary B.3. For any L0 ∈ \(\mathcal{S}\), the set \({{\mathcal{S}}_{0}}\) is bounded.
On the other hand, the function f(L) achieves minimum on the set \({{\mathcal{S}}_{0}}\). (This function is continuous by the properties of the solution of the Lyapunov equation and is considered on a compact set.) However, the set \({{\mathcal{S}}_{0}}\) has no common points with the boundary of \(\mathcal{S}\) due to (B.3). The function f(L) is differentiable on \({{\mathcal{S}}_{0}}\); see below. Consequently, we arrive at the following result.
Corollary B.4. There exists a minimum point \({{L}_{*}}\) on the set \(\mathcal{S}\), and the gradient of the function f(L) vanishes at this point.
Let us analyze the properties of the gradient of the function f(L, α).
Lemma B.4. The function f(L, α) is defined on the set of stabilizing L for 0 < α < 2σ(A – LC). On this admissible set, the function is differentiable, and its gradient is given by
where the matrices P and Y are the solutions of the Lyapunov equations (12) and (13), respectively.
The minimum of f(L, α) is achieved at an inner point of the admissible set and is determined by the conditions
In addition, f(L, α) as a function of α is strictly convex on 0 < α < 2σ(A – LC) and achieves minimum at an inner point of this interval.
Proof of Lemma B.4. We have the constrained optimization problem
subject to the Lyapunov equation (12) for the matrix P of the invariant ellipsoid.
Differentiation with respect to α is performed using the relations (B.2), (12), and (13). To differentiate with respect to L, we add an increment ΔL and denote by ΔP the corresponding increment of P. As a result, the relation (12) takes the form
Leaving the notation ΔP for the principal terms of the increment, we obtain
Subtracting Eq. (12) from this equation yields
We calculate the increment of the functional f(L) by linearizing the corresponding values:
By Lemma B.1, from the dual Eqs. (B.6) and (13) we have
Thus, the relation (B.4) is derived and the proof of Lemma B.4 is complete.
Rights and permissions
About this article
Cite this article
Khlebnikov, M.V. A Comparison of Guaranteeing and Kalman Filters. Autom Remote Control 84, 389–411 (2023). https://doi.org/10.1134/S0005117923040094
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1134/S0005117923040094