Abstract
Given a linear controlled autonomous system, we consider the problem of including a convex compact set in the reachable set of the system in the minimum time and the problem of determining the maximum time when the reachable set can be included in a convex compact set. Additionally, the initial point and the time at which the extreme time is achieved in each problem are determined. Each problem is discretized on a grid of unit vectors and is then reduced to a linear programming problem to find an approximate solution of the original problem. Additionally, error estimates for the solution are found. The problems are united by a common ideology going back to the problem of finding the Chebyshev center.
REFERENCES
A. A. Liapounoff, “Sur les fonctions-vecteurs completement additives,” Izv. Akad. Nauk SSSR Ser. Mat. 4 (6), 465–478 (1940).
E. B. Lee and L. Markus, Foundations of Optimal Control Theory (Wiley, New York, 1967).
R. Aumann, “Integrals of set-valued functions,” J. Math. Anal. Appl. 12 (1), 1–12 (1965).
B. T. Polyak and G. Smirnov, “Large deviations for non-zero initial conditions in linear systems,” Automatica 74, 297–307 (2016).
J.-P. Aubin and A. Cellina, Differential Inclusions (Springer-Verlag, Berlin, 1984).
J.-P. Aubin, “A survey of viability theory,” SIAM J. Control Optim. 28 (4), 749–788 (1990).
H. J. Kelley, “Gradient theory of optimal flight paths,” ARS J. 30, 947–953 (1960). https://doi.org/10.2514/8.5282
A. E. Bryson and W. F. Denham, “A steepest-ascent method for solving optimum programming problems,” J. Appl. Mech. 29, 247–257 (1962). https://www.gwern.net/docs/ai/1962-bryson.pdf
Ph. Eichmeir, Th. Lauß, S. Oberpeilsteiner, K. Nachbagauer, and W. Steiner, “The adjoint method for time-optimal control problems,” J. Comput. Nonlinear Dyn. 16 (2), 021003 (2021). https://doi.org/10.1115/1.4048808
P. Cannarsa and C. Sinestrari, “Convexity properties of the minimum time function,” Calculus Var. Partial Differ. Equations 3 (3), 273–298 (1995). https://doi.org/10.1007/bf01189393
V. G. Boltyanskii, Mathematical Methods of Optimal Control (Holt, Rinehart and Winston, New York, 1971).
E. S. Polovinkin, “Strongly convex analysis,” Sb. Math. 187 (2), 259–286 (1996).
C. Le Guernic and A. Girard, “Reachability analysis of linear systems using support functions,” Nonlinear Anal. Hybrid Syst. 4, 250–262 (2010).
M. Althoff, G. Frehse, and A. Girard, “Set propagation techniques for reachability analysis,” Ann. Rev. Control Rob. Auton. Syst. 4, 369–395 (2021). https://doi.org/10.1146/annurev-control-071420-081941
M. Serry and G. Reissig, “Over-approximating reachable tubes of linear time-varying systems” IEEE Trans. Autom. Control 67 (1), 443–450 (2022). https://doi.org/10.48550/arXiv.2102.04971
A. B. Kurzhanski and P. Varaiya, Dynamics and Control of Trajectory Tubes (Birkhäuser/Springer, Basel, 2014).
M. V. Balashov, “Covering a set by a convex compactum: Error estimates and computation,” Math. Notes 112 (3), 349–359 (2022).
J. Diestel, Geometry of Banach Spaces: Selected Topics (Springer-Verlag, Berlin, 1975).
H. Frankowska and C. Olech, “R-convexity of the integral of the set-valued functions,” Contributions to Analysis and Geometry (Johns Hopkins Univ. Press, Baltimore, MD, 1981), pp. 117–129.
E. S. Polovinkin and M. V. Balashov, Elements of Convex and Strongly Convex Analysis, 2nd ed. (Fizmatlit, Moscow, 2007) [in Russian].
M. V. Balashov and E. S. Polovinkin, “M-strongly convex subsets and their generating sets,” Sb. Math. 191 (1), 25–60 (2000).
M. V. Balashov, “On polyhedral approximations in an n-dimensional space,” Comput. Math. Math. Phys. 56 (10), 1679–1685 (2016).
M. V. Balashov and D. Repovš, “Polyhedral approximations of strictly convex compacta,” J. Math. Anal. Appl. 374, 529–537 (2011).
M. V. Balashov, “Strong convexity of reachable sets of linear systems,” Sb. Math. 213 (5), 604–623 (2022).
Funding
The results in Sections 2, 3 were obtained by M.V. Balashov at the Trapeznikov Institute of Control Sciences of the Russian Academy of Sciences, and his research was supported by the Russian Science Foundation, grant no. 22-11-00042, https://rscf.ru/en/project/22-11-00042/.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
The authors declare that they have no conflicts of interest.
Additional information
Translated by I. Ruzanova
Appendices
APPENDIX 1
1.1 INTERIOR OF A MULTIVALUED INTEGRAL IN THE PLANE CASE
Lemma 5. Suppose that the eigenvalues of the matrix \(A \in {{\mathbb{R}}^{{2 \times 2}}}\) are complex conjugate numbers \(\alpha \pm i\beta \) with \(\beta > 0\) and \({{e}^{{As}}} = {{e}^{{\alpha s}}}Q(\beta s)\), where \(Q(s) = \left( {\begin{array}{*{20}{c}} {\cos s}&{ - \sin s} \\ {\sin s}&{\cos s} \end{array}} \right)\). Let \(U \subset {{\mathbb{R}}^{2}}\) be a convex compact set, \(r > 0\), \(x \in {{\mathbb{R}}^{2}}\), \(\left\| x \right\| = r\), and the conditions \({{B}_{r}}(x) \subset U\) and \(0 \in {{B}_{r}}(x) \cap \partial U\) are satisfied. Define \(C = \min \{ {{e}^{{\alpha {{t}_{1}}}}},{{e}^{{\alpha {{t}_{2}}}}}\} \). Then, for any \(0 < {{t}_{1}} < {{t}_{2}}\),
where \(\delta = Cr({{t}_{2}} - {{t}_{1}})\left( {1 - \frac{{\sin \left( {\beta ({{t}_{2}} - {{t}_{1}}){\text{/}}2} \right)}}{{\beta ({{t}_{2}} - {{t}_{1}}){\text{/}}2}}} \right)\).
Note that the conditions of the lemma imply that the boundary of the set \(U\) is smooth at the point \(0\). Moreover, the point \(0 \in \partial U\) can be touched by a ball of radius \(r\) from within \(U\).
Proof. It is true that
In view of the equality \(Q(\beta s){{B}_{r}}(0) = {{B}_{r}}(0)\), the first term on the right-hand side is estimated as
Let \(x = r{{(\cos \theta ,\sin \theta )}^{{\text{T}}}}\) for some \(\theta \in [0,2\pi )\). For the second term, we obtain
For the point \(z\), we have \(\left\| z \right\| = \frac{{2Cr\sin (\beta ({{t}_{2}} - {{t}_{1}}){\text{/}}2)}}{\beta }\), whence
Lemma 6. Suppose that the eigenvalues of the matrix \(A \in {{\mathbb{R}}^{{2 \times 2}}}\) are complex conjugate numbers \(\alpha \pm i\beta \) with \(\beta > 0\) and \({{e}^{{As}}} = {{e}^{{\alpha s}}}Q(\beta s)\), where \(Q(s)\) is defined in Lemma \(5\). Let \(U \subset {{\mathbb{R}}^{2}}\) be a convex compact set, \(r > 0\), \(\theta \in [0,2\pi )\), and \(I = {\text{co}}\{ \pm r{{(\cos \theta ,\sin \theta )}^{{\text{T}}}}\} \subset \partial U\). Define \(C = \min \{ {{e}^{{\alpha {{t}_{1}}}}},{{e}^{{\alpha {{t}_{2}}}}}\} \). Then, for any \(0 < {{t}_{1}} < {{t}_{2}}\) with \({{t}_{2}} - {{t}_{1}} < \pi {\text{/}}\beta \),
where \(\delta = \frac{2}{\beta }Cr\left( {1 - \cos \left( {\frac{1}{2}\beta ({{t}_{2}} - {{t}_{1}})} \right)} \right)\).
Proof. We have the chain of inclusions \(\int_{{{t}_{1}}}^{{{t}_{2}}} {{{e}^{{As}}}U{\kern 1pt} ds} \supset \int_{{{t}_{1}}}^{{{t}_{2}}} {{{e}^{{As}}}I{\kern 1pt} ds} \supset C\int_{{{t}_{1}}}^{{{t}_{2}}} {Q(\beta s)I{\kern 1pt} ds} \). Let \(p = (\cos \varphi ,\sin \varphi {{)}^{{\text{T}}}}\) be a unit vector. Then
where \(\psi = \theta - \varphi + \pi {\text{/}}2\). Making the substitution \(\tau = \beta s + \psi \), for \({{T}_{i}} = \beta {{t}_{i}} + \psi \), \(i = 1,2\), we obtain \(s\left( {p,\int_{{{t}_{1}}}^{{{t}_{2}}} {{{e}^{{As}}}U{\kern 1pt} ds} } \right) \geqslant \frac{1}{\beta }Cr\int_{{{T}_{1}}}^{{{T}_{2}}} {\left| {\sin \tau } \right|} {\kern 1pt} d\tau \). Since \({{T}_{2}} - {{T}_{1}} < \pi \), it follows that
Substituting this estimate into the preceding inequality yields the assertion of the lemma.
Note that Lemmas 6 and 5 give the second and third orders of the ball’s radius with respect to \({{t}_{2}} - {{t}_{1}}\), respectively.
Let \(U \subset {{\mathbb{R}}^{2}}\) be a convex compact set. Assume that there exists a number \(\rho > 0\) such that \(U \cap {{B}_{\rho }}(0) \supset H \cap {{B}_{\rho }}(0)\), where \(H = \{ x \in {{\mathbb{R}}^{2}}\,|\,(p,x) \leqslant 0\} \) is a half-space for some unit vector \(p \in {{\mathbb{R}}^{2}}\). Then \(\int_{{{t}_{1}}}^{{{t}_{2}}} {{{e}^{{As}}}U{\kern 1pt} ds} \asymp {{({{t}_{2}} - {{t}_{1}})}^{2}}\), \({{t}_{2}} - {{t}_{1}} \to + 0\).
Indeed, let \(I\) be a segment centered at the origin such that \(I \subset \partial H.\) Consider the disk \({{B}_{R}}(x)\), \(\left\| x \right\| = R\), that is tangent to \(\partial H\) at the point \(0\) and such that \(I + {{B}_{R}}(x) \supset U\). Such \(R\) and \(I\) can be chosen, since \(U\) is compact and \(\partial H\) is the supporting line of \(U\) at the point \(0\). Then
By Lemmas 5 and 6, the terms on the right-hand side of this formula have the order \({{({{t}_{2}} - {{t}_{1}})}^{2}} + {{({{t}_{2}} - {{t}_{1}})}^{3}} \asymp {{({{t}_{2}} - {{t}_{1}})}^{2}}\), \({{t}_{2}} - {{t}_{1}} \to + 0\).
Thus, the presence of a segment in U, \(0 \in \partial U,\) centered at 0 is more important than the bodiliness of \(U\) for the ball centered at 0 inscribed in the integral to be maximal.
Lemma 7. Let \({{e}^{{At}}} = \left( {\begin{array}{*{20}{c}} {{{e}^{{{{\lambda }_{1}}t}}}}&0 \\ 0&{{{e}^{{{{\lambda }_{2}}t}}}} \end{array}} \right)\), where \({{\lambda }_{1}} > {{\lambda }_{2}}\). Define \(\lambda = {{\lambda }_{1}} - {{\lambda }_{2}} > 0\) and \(C = \min \{ {{e}^{{{{\lambda }_{2}}{{t}_{1}}}}},{{e}^{{{{\lambda }_{2}}{{t}_{2}}}}}\} \). Let \(U \subset {{\mathbb{R}}^{2}}\) be a convex compact set and \(I \subset \partial U\), where \(I = {\text{co}}{\kern 1pt} {\kern 1pt} \left( { \pm r{{{(\cos \theta ,\sin \theta )}}^{{\text{T}}}}} \right)\) for \(r > 0\), \(\theta \in (0,\pi )\), \(\theta \ne \pi {\text{/}}2\). Then, for any \(0 < {{t}_{1}} < {{t}_{2}}\),
where \(\delta = Cr\frac{{\lambda {{K}^{2}}\sin \theta }}{{{{\pi }^{2}}}}{{({{t}_{2}} - {{t}_{1}})}^{2}}.\) Here, \(K = K({{t}_{1}},{{t}_{2}},\lambda ,\theta ) = \mathop {\min }\limits_{i = 1,2} \left\{ {\frac{{{{e}^{{\lambda {{t}_{i}}}}}\left| {\cot \theta } \right|}}{{1 + {{e}^{{2\lambda {{t}_{i}}}}}{{{\cot }}^{2}}\theta }}} \right\}\).
Proof. Let \(M = \int_{{{t}_{1}}}^{{{t}_{2}}} {{{e}^{{As}}}U{\kern 1pt} ds} \). Then
Let \(p = (\cos \varphi ,\sin \varphi {{)}^{{\text{T}}}}\) be an unit vector, where \(\varphi \in [0,2\pi )\). Then, for
we have \(s(p,M) \geqslant CrJ.\) Let us estimate \(J\) from below.
1. Consider the case \(\theta \in \left( {0,\frac{1}{2}\pi } \right)\). Here,
where \(\alpha (s) = {\text{arctan}}{\kern 1pt} {\kern 1pt} ({{e}^{{\lambda s}}}\theta )\). We introduce the new variable \(\tau = \alpha (s) + \varphi \), \(ds = \frac{{d\tau }}{{\alpha {\kern 1pt} '(s)}} = \frac{{1 + {{e}^{{2\lambda s}}}{{{\cot }}^{2}}\theta }}{{\lambda {{e}^{{\lambda s}}}\cot \theta }}{\kern 1pt} d\tau \). In view of the inclusion \(\alpha (s) \in \left( {0,\frac{1}{2}\pi } \right)\), it follows that
By the mean-value theorem, there exists \(\xi \in [{{t}_{1}},{{t}_{2}}]\) such that
which yields the assertion of the lemma.
2. Suppose that \(\theta \in \left( {\frac{1}{2}\pi ,\pi } \right)\). Making the substitutions \(\psi = \pi - \theta ,\) \(\alpha (s) = {\text{arctan}}{\kern 1pt} {\kern 1pt} ({{e}^{{\lambda s}}}\cot \psi ),\) and \(\tau = \alpha (s) + \varphi \) brings the integral \(J\) to the form
The rest of the proof is the same as in item 1.
Lemma 8. Let \({{e}^{{At}}} = {{e}^{{\lambda t}}}\left( {\begin{array}{*{20}{c}} 1&t \\ 0&1 \end{array}} \right)\). Define \(C = \min \{ {{e}^{{\lambda {{t}_{1}}}}},{{e}^{{\lambda {{t}_{2}}}}}\} \). Let \(U \subset {{\mathbb{R}}^{2}}\) be a convex compact set and \(I \subset \partial U,\) where \(I = {\text{co}}{\kern 1pt} {\kern 1pt} \left( { \pm r{{{(\cos \theta ,\sin \theta )}}^{{\text{T}}}}} \right)\) for \(r > 0\) and \(\theta \in (0,\pi )\). Then, for any \(0 < {{t}_{1}} < {{t}_{2}}\),
where \(\delta = Cr\frac{{{{K}^{2}}\sin \theta }}{{{{\pi }^{2}}}}{{({{t}_{2}} - {{t}_{1}})}^{2}}.\) Here, \(K = K({{t}_{1}},{{t}_{2}},\theta ) = \mathop {\min }\limits_{i = 1,2} \left\{ {\frac{1}{{1 + {{{({{t}_{i}} + \cot \theta )}}^{2}}}}} \right\}\).
Proof. Repeating the argument used in Lemma 7, for \(M = \int_{{{t}_{1}}}^{{{t}_{2}}} {{{e}^{{As}}}U{\kern 1pt} ds} ,\) \(p = (\cos \varphi ,\sin \varphi ),\) and J = \(\int_{{{t}_{1}}}^{{{t}_{2}}} {\left| {(\cos \theta + s\sin \theta )\cos \varphi + \sin \theta \sin \varphi } \right|ds} \), we have \(s(p,M) \geqslant CrJ\). For \(J,\) we obtain J = \(\sin \theta \int_{{{t}_{1}}}^{{{t}_{2}}} {\sqrt {1 + {{{(s + \cot \theta )}}^{2}}} } \left| {\sin (\alpha (s) + \varphi )} \right|ds\), where \(\alpha (s) = {\text{arctan}}{\kern 1pt} (s + \cot \theta )\). Making the substitution \(\tau = \alpha (s) + \varphi \) yields
It follows that \(J \geqslant 4\sin \theta \frac{4}{{{{\pi }^{2}}}}\frac{{{{{(\alpha ({{t}_{2}}) - \alpha ({{t}_{1}}))}}^{2}}}}{{{{4}^{2}}}}\). Applying the mean-value theorem completes the proof.
APPENDIX 2
1.1 AN EXAMPLE OF A REACHABLE SET
For a nontrivial linear system, we give an example of a reachable set that is strictly convex, but not strongly convex. Recall that \(f(s) \asymp g(s)\), \(s \to 0\) if there exists \(\delta > 0\) such that \({{C}_{1}} \leqslant \frac{{f(s)}}{{g(s)}} \leqslant {{C}_{2}}\) for all \(s \in ( - \delta ,\delta )\) and some \({{C}_{2}} \geqslant {{C}_{1}} > 0\). Let \(U = {\text{co}}{\kern 1pt} {\kern 1pt} \{ (0,0, \pm 1)\} \), \(\lambda \in \mathbb{R}\), \(A = \left[ {\begin{array}{*{20}{c}} \lambda &1&0 \\ 0&\lambda &1 \\ 0&0&\lambda \end{array}} \right]\), and \({{e}^{{As}}} = {{e}^{{\lambda s}}}\left[ {\begin{array}{*{20}{c}} 1&s&{\frac{1}{2}{{s}^{2}}} \\ 0&1&s \\ 0&0&1 \end{array}} \right]\). For \(\varepsilon \in (0,1)\), we consider the vectors \(p = \frac{1}{3}(2, - 2,1)\) and \(q = q(\varepsilon ) = \frac{{(2, - 2,1 - \varepsilon )}}{{\sqrt {9 - 2\varepsilon + {{\varepsilon }^{2}}} }}\). It is easy to see that \(\left\| {p - q} \right\| \asymp \varepsilon \), \(\varepsilon \to 0\), and, for \(t > 1 + \sqrt \varepsilon \) (\({{s}_{{1,2}}} = 1 \pm \sqrt \varepsilon \) are the roots of the equation \(({{e}^{{{{A}^{{\text{T}}}}s}}}q(\varepsilon {{),(0,0,1)}^{{\text{T}}}}) = 0\))
For any fixed \(R > 0\), the value of \(\sqrt \varepsilon {\text{/}}(R\varepsilon )\) is not bounded above for small \(\varepsilon > 0\). Therefore, \(M(t)\) is not strongly convex, since, for a strongly convex set with radius \(R > 0\), its support elements have to satisfy the Lipschitz condition on the unit sphere with constant \(R\) (see [19]; [12], Corollary 4). Note that the set \(M(t)\) is strictly convex, since for any vector \(p \in \partial {{B}_{1}}(0)\) the set \(({{e}^{{As}}}U)(p)\) is a singleton for all \(s \in [0,t]\), except for at most two values.
Rights and permissions
About this article
Cite this article
Balashov, M.V., Kamalov, R.A. Optimization of the Reachable Set of a Linear System with Respect to Another Set. Comput. Math. and Math. Phys. 63, 751–770 (2023). https://doi.org/10.1134/S0965542523050056
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1134/S0965542523050056