You are on page 1of 8

Journal of Applied Mathematics and Mechanics 81 (2017) 106–113

Contents lists available at ScienceDirect

Journal of Applied Mathematics and Mechanics


journal homepage: www.elsevier.com/locate/jappmathmech

Time dependence of the attainability regions of third order systems!


D.I. Bugrov ∗ , A.M. Formal’skii
Lomonosov State University, Moscow, Russia

a r t i c l e i n f o a b s t r a c t

Article history: A linear steady third order system with one controlling (perturbing) action that is bounded in absolute
Received 27 October 2016 magnitude is considered. The matrix specifying the system has one real and two complex conjugate
Available online 28 September 2017 eigenvalues. The behaviour of the boundary of the attainability region of this system as time increases is
studied. It is shown that the structure of the boundary of the attainability region depends on the relation
between the real eigenvalue and the real part of the complex conjugate eigenvalues.
© 2017 Elsevier Ltd. All rights reserved.

The investigation of the attainability regions (AR) of controlled (perturbed) systems when there are constraints on the controlling
(perturbing) action is of interest both in solving the problem as to which states a controlled (perturbed) system can reach after this or that
time (we recall Bulgakov’s problem on the accumulation of perturbations1 ) as well as in relation to the optimal behaviour of the participants
in games, in constructing a theory of optimal motion control. The description of AR and examples of their use and construction can be found
in many known publications from the 1950s to the 1970s (for example, see Refs 2–9). The construction and study of AR is not a simple
problem, and it has therefore been suggested in a number of investigations that these regions are estimated using ellipsoids10–12 . New
results concerning ellipsoidal estimates of AR have been discussed in comparatively recent publications.13–16 Problems in constructing the
AR for discrete systems17–19 and questions of the smoothness of the boundaries of attainability regions (BAR)20,21 were considered and
the relation between the smoothness of BAR and the observability of systems was investigated.22,23
It has been shown20 that, for a linear steady system of arbitrary order with one constrained control action having at least one real
eigenvalue, there are exactly two “conical” corner points (CCP) on the BAR at any time. A similar situation also occurs in an unsteady
system for sufficiently small time values. In a steady third order system with real eigenvalues, these two CCP are joined to one another by
two ribs.20 If a steady system (of arbitrary order) has only complex eigenvalues, then, after a certain time, the attainability set does not
have CCP.
In this paper, the “deformation” of the BAR of a steady third order system with one real and two complex conjugate eigenvalues is
investigated analytically and numerically when the instant for which the investigated AR is constructed is changed. It is shown that the
structure of the BAR depends on the difference between the real eigenvalue and the real part of the complex-conjugate eigenvalues.

1. System of equations

Consider a system described by the linear matrix differential equation with constant coefficients

(1.1)

Here, x, C and g are 3 × 1, 3 × 3 and 3 × 1 matrices respectively and ! = !(!) is a scalar control (perturbation) function of the time ! that is
assumed to be piecewise-continuous and bounded in modulus. We will assume that the initial third order system has one real and two

! Prikl. Mat. Mekh., Vol. 81, No. 2, pp. 154–164, 2017.


∗ Corresponding author.
E-mail addresses: d.bugrov@mech.math.msu.rsu (D.I. Bugrov), formal@imec.msu.ru (A.M. Formal’skii).

https://doi.org/10.1016/j.jappmathmech.2017.08.004
0021-8928/© 2017 Elsevier Ltd. All rights reserved.
D.I. Bugrov, A.M. Formal’skii / Journal of Applied Mathematics and Mechanics 81 (2017) 106–113 107

complex-conjugate eigenvalues and that Eq. (1.1) is already written in the Jordan form, that is, the matrices C and g have the form indicated
above. The quantity ! is the real eigenvalue and " and ±i# are the real and imaginary parts of the complex-conjugate values " ± i#.
The inequalities

(1.2)

that are assumed to be satisfied, are the necessary and sufficient conditions for the controllability of system (1.1) in a Kalman sense, and,
without any loss in generality, it can be assumed that # > 0.
In system (1.1), we carry out the change of variables and constant quantities

(1.3)

The possibility of such a change follows from the first inequality of (1.2).
System (1.1) becomes simpler after the change (1.3):

(1.4)

The fundamental matrix of the solutions of the homogeneous (when u(t) ≡ 0) third order system (1.4) has the form

(1.5)

2. Attainability region

Consider the set of piecewise-continuous functions u(t) satisfying inequality (1.4):

(2.1)

We will assume that functions belonging to this set are admissible controls.
Suppose, at the initial instant when t = 0, system (1.4) is located at the origin of the coordinates: x(0) = 0. Its solution at the instant t = T
is then determined by the Cauchy formula

(2.2)

We now consider the set of points in the phase space X that system (1.4), when x(0) = 0, can reach in a finite time T with an admissible
control u(t) ∈ !. This attainability region (AR) is described by the relation

(2.3)

It is well known2–9 that, in linear systems, this region is, for each finite value of T, a closed, bounded, convex set that is symmetric about
the origin of the coordinates and becomes larger as T increases, that is, Q(T2 ) ⊃ Q(T1 ) if T2 > T1 .
For purposes of illustration, the set Q(T) for a second order system is shown in Fig. 1 as well as the reference line to this set "($) that
is orthogonal to the vector $ and the distance %($) from the origin of the coordinates to the line "($) (and to the line "(−$)).
When considering the third order system (1.4), we will assume that the three-dimensional vector $ is a unit vector, that is, $ ∈ S.

We now introduce the spherical coordinates, the angles & and ' (see Fig. 2) into the treatment that determine the position of the three-
dimensional unit vector $. Then,

(2.4)

The set of all points of a unit sphere can be described by the inequalities

(2.5)

The boundary of the set Q(T) is a surface which is the envelope of its reference planes "($). Those, and only those, points x that satisfy
the set of inequalities

(2.6)
108 D.I. Bugrov, A.M. Formal’skii / Journal of Applied Mathematics and Mechanics 81 (2017) 106–113

Fig. 1.

Fig. 2.

thereby belong to the set Q(T), where !(") is the distance from the origin of the coordinates to the reference plane !(") that is orthogonal
to the vector " (" is a row matrix in relations (2.6)):

(2.7)

A permissible control u(T − #), yielding a maximum to the integral occurring in relation (2.7), has the form

(2.8)
!
It follows from conditions (1.2) and relations (1.3) that b1 =
/ 0 and ẽ = e22 + e32 / 0 and, without any loss in generality, it can also be
=
assumed that b1 > 0. Then, introducing the notation

expression (2.8) is written as:

(2.9)

Relations (2.2) and (2.9) describe the two-parameter set of boundary points of the region Q(T). The angles $ and % (see Fig. 2 and formulae
(2.4)) can be considered as the parameters which vary within ranges (2.5). The boundary points of the region Q(T) can also be found by
solving differential equation (1.4) with the control
D.I. Bugrov, A.M. Formal’skii / Journal of Applied Mathematics and Mechanics 81 (2017) 106–113 109

(2.10)

Expression (2.10) is obtained from formulae (2.9) by replacing the quantity ! with T - t.

3. Corner points of the boundary of the attainability region

When

(3.1)

" = ("1 , "2 , "3 ) = (1 0 0) (see formulae (2.4)), the first term in the expression for #(!) (see (2.9) is equal to e($−%)! and the second term
vanishes. Hence, the equality

(3.2)

holds for all values of ! and T subject to condition (3.1) in accordance with expressions (2.9). Therefore, if " = (1 0 0), the control, for which a
maximum of the integral in expression (2.7) is attained, is equal to the constant quantity (3.2). Suppose x(T) of a point of the BAR obtained
when control (3.2) is substituted into formula (2.2). The components "1 , "2 and "3 of the unit vector " are continuously dependent on
the angles & and '. It is therefore easily shown that, for every value ! ≥ 0 in the plane of the variables & and ' near the point (3.1), a
neighbourhood exists a in which the quantity #(!) (see expressions (2.9)) is positive, that is,

(3.3)

In other words, for every value ! ≥ 0 on the set S a neighbourhood of the vector " = (1 0 0) exists such that inequality (3.3) holds for all
vectors from this neighbourhood. The whole subset of unit vectors " for which inequality (3.3) holds for a specified value of ! ≥ 0 is denoted
by S(!). The subset S(!) for each value of ! contains the vector " = (1 0 0), and these subsets have a complete measure on the set S. The
left-hand side of inequality (3.3) is linearly dependent on the components of the vector " and, consequently, the subset S(!) is a convex
cone.
The set of all the vectors " for each of which inequality (3.3) holds for all values 0 ≤ ! ≤ T is subsequently denoted by S(0, T). This set S(0,
T) is the intersection of the set S(!) for all values 0 ≤ ! ≤ T

(3.4)

Being the intersection of convex cones S(!), the set S(0, T) is also a convex cone. For each vector " ∈ S(0, T), control (2.9) is the one and the
same constant quantity (see formula (3.2)) for all values 0 ≤ ! ≤ T. Consequently, all the reference planes !(") orthogonal to the vectors
" ∈ S(0, T) contain the point x(T) ∈ Q(T) obtained in the case of control (3.2). The intersection of these reference planes forms a conical
surface, the vertex of which is the point x(T), and it is naturally called the conical corner point (CCP) of the region Q(T).20 The set S(0, T)
of the unit vectors " is convex, and the intersection of the reference planes !(") orthogonal to the vectors " ∈ S(0, T) therefore forms a
convex conical surface which also follows from the convexity of the attainability set Q(T), the boundary of which is the envelope of the set
of reference planes.
To the extent that the AR is symmetric about the origin of the coordinates, it has a further CCP that is symmetric to the first CCP; it is
obtained for the control u(T − !) = −1. Similar points have been called5,7 “vertex” points. There are no other CCP in the system apart from
those obtained using the control u(T − !) = ±1.20
It follows from relation (3.4) that the set S(0,T) does not become larger as the time T increases, that is, if T2 ≥ T1 , then

(3.5)

In other words, as the time T increases, the cone S(0,T) either remains unchanged or becomes smaller. It follows from this that, as the time
T increases, the cone “encompassing” the CCP and, together with it, the whole region Q(T) either remains unchanged or becomes larger. A
similar conclusion also applies to systems of any order.

4. The dependence of the apex angle of the cone on the difference $ − % and the time T

We consider three cases: the difference $ − % is less than zero, equal to zero or greater than zero.
Initially, suppose

(4.1)

that is, the real eigenvalue $ of system (1.4) is less than the real part % of the complex-conjugate eigenvalues.
The vector " = (1 0 0) satisfies inequality (3.3) for all values of ! ≥ 0. We now show that, in the case of (4.1), the set S(0,T) of vectors
contracts to this vector " = (1 0 0) when T→ ∞.
The left-hand side of expression (3.3) is the scalar product of the two three-dimensional vectors

(4.2)

The inequality (3.3) means that the angle between the vectors (4.2) is less than (/2. The second of the vectors (4.2) is equal to the sum
of the vectors

(4.3)
110 D.I. Bugrov, A.M. Formal’skii / Journal of Applied Mathematics and Mechanics 81 (2017) 106–113

Fig. 3.

The first of vectors (4.3) is directed along the !1 axis and the second is located in the (!2 , !3 )-plane orthogonal to this axis (see Fig. 2).
Under condition (4.1), the magnitude of the first vector monotonically decreases as the value of " increases and tends to zero when "→ ∞
while the second vector rotates about the !1 axis (counter-clockwise) without changing in length and executes a complete revolution
about this axis if the value of " increases by 2#. Together with it, the second of vectors (4.2), which is equal to the sum of vectors (4.3), also
executes a complete ! revolution about the !1 axis and its magnitude decreases as the value of " becomes larger but remains greater than
the quantity ẽ = e22 + e32 . As the value of " increases, the angle between the second of vectors (4.2) and the (!2 , !3 )-plane decreases and
tends to zero when "→ ∞, that is, the vector is “pressed” towards the (!2 , !3 )-plane.
The set S(") of vectors (!1 , !2 , !3 ) satisfying inequality (3.3) is half of a unit sphere, the centre of which is the second of vectors (4.2). As
" increases, this hemisphere rotates about the !1 axis together with its “central” vector which, when "→ ∞, tends to the (!2 , !3 )-plane. It
is easy to show that the set S(0,T) of vectors (!1 , !2 , !3 ) satisfying inequality (3.3) becomes smaller as the value of T increases and, when
T→ ∞, it tends to the vector ! = (1 0 0). In this case, the set of reference planes at the corner point contracts to the single reference plane
!(!), orthogonal to the vector ! = (1 0 0).
Hence, in the case of (4.1), as the time T increases, the apex angle of the cone near the CCP (the solid angle) becomes larger. In other
words, the CCP becomes “less pointed” and disappears when T→ ∞. This smoothing of the boundary of the region Q(T) as the time T
increases occurs more rapidly, the smaller the difference $ − %. The latter is related to the fact that, in the case of (4.1), for smaller values of
the difference $ − % the exponential e($−%)" tends more rapidly to zero with an increase in the magnitude of ". The limit set Q(∞), obtained
when T→ ∞, does not have a CCP. It follows from what has been said above that, in the case of (4.1) considered, the approximation of the
region Q(T) by an ellipsoid, the boundary of which is a smooth surface, becomes more justified as the time T increases.
The set Q(∞) is bounded in the phase space X if, and only if, $ < 0 and % < 0,8 If however, $ < 0 and % ≥ 0, then the limit set is8 the strip
|x1 | ≤ |$/b1 | which, in fact, does not have corner points.
The regions Q(T) of system (1.4) are shown in Fig. 3 for the positive values $ = 0.1, % = 0.2 ($ − % < 0) and a time T = 0.8# (on the left) and
T = 2.5# (on the right). Each of the regions has two CCP. The region shown on the right is “considerably” greater in volume. This is explained
by the fact that the quantities $, % > 0 and the time T = 2.5# is greater (by roughly three times) than the time T = 0.8#. In accordance with
the statements made above, the CCP in the region shown on the right are not so pointed. The CCP in the region shown on the left are joined
to one another by two ribs that are symmetric to one another about the origin of the coordinates (see Ref. 20, for example, regarding the
existence of ribs in AR). If, $, % > 0, then, when T→ ∞, the region Q(T) encloses all points of the phase space X. In this case, the limit region
Q(∞), occupying the whole phase space, does not have a boundary and there is also, of course, no need to talk about which corner points
it has. However, we repeat that, for each finite value of T, the boundary of the region Q(T) has similar corner points.
To construct the AR, the solution of the matrix equation (1.4) with control functions (2.10) was programmed using the MATHEMATICA
software package. Expressions (2.4) are used in formulae (2.9) and (2.10) instead of the components !1 , !2 , !3 of the vector !, and the
angles & and ' appear here as parameters that are discretely varied within the ranges (2.5). The values b1 = b3 = 1, b2 = 0 were used in the
numerical investigations.
In considering Fig. 3 and the following figures, it should not be forgotten that the geometric centre of the AR shown in these figures is
located at the origin of the coordinates x1 = x2 = x3 = 0.
The domains Q (T) for the negative values $ = −0.2, % = −0.1 ($ − % < 0) are shown in Fig. 4 for a comparatively small value of the time
T = 2# (on the left) and for a “large” time T = 25# (on the right). The region shown on the left has CCP. No CCP are seen for the region shown
on the right, although they exist for any finite value of the time T. Numerical investigations show that region Q(T) when T = 25# is “close”
to the limit set Q(∞), the boundary of which does not have CCP. Hence, when T = 25#, the apex angle of each of the cones “encompassing”
the CCP is “close” to the value of #.
We now consider the case when

(4.4)
D.I. Bugrov, A.M. Formal’skii / Journal of Applied Mathematics and Mechanics 81 (2017) 106–113 111

Fig. 4.

Under condition (4.4), inequality (3.3) acquires the form

(4.5)

If the components !2 and !3 of the vector ! = (!1 , !2 , !3 ) are sufficiently small, the second term on its left-hand side will also be fairly
small. It follows from this that, if the vector (!1 , !2 , !3 ) lies in a certain neighbourhood of the vector (1, 0, 0), this inequality is satisfied for
all values of " ≥ 0.
We now analyse inequality (4.5) in greater detail to find the set of vectors S(0,T). Its left-hand side is the scalar product of the vectors

(4.6)

and inequality (4.5) means that the angle between these vectors is less than #/2. The second of the vectors (4.6) is equal to the sum of the
two vectors

(4.7)

The first of these is constant and directed along the !1 axis and the second is located in the (!2 , !3 )-plane orthogonal to the !1 axis (see
Fig. 2) and, as was stated above, as the value of " increases, it rotates around the !1 axis without changing in length and executes a complete
revolution when the value of " changes by 2#. Together with it, the second of vectors (4.6), which remains constant in magnitude and
describes a conical surface also executes a complete revolution about the !1 axis.
It follows from this that S(" + 2#) = S(") for any value of ". It therefore follows from relations (3.4) and (3.5) that, starting from the instant
T = 2#, the set of vectors S(0,T) remains unchanged as a spherical (globular) sector or, in other words, a circular cone, the axis of which is
the vector (1, 0, 0). The apex angle of the sector S(0,T) is equal to

that is, it depends on the quantities e2 and e3 but is independent of the time T. The ends of the vectors ! from this sector lie on a sphere
forming something like a “small circular cap”. Note that the set S(0,T) constructed subject to condition (4.1) belongs to the set S(0,T)
constructed for the same values of e2 and e3 but subject to condition (4.4).
The CCP together with the whole of the attainability set Q(T) are located within the circular cone conjugated to the sector S(0,T). The
apex angle of this circular cone is equal to

!
that is, the smaller the magnitude of ẽ = e22 + e32 , then the smaller this angle and the more pointed the CCP.
In the case (4.4), two CCP that are symmetric with respect to the origin of the coordinates exist for all values 0< T < ∞. Hence, unlike in
the case of (4.1), the CCP of the boundary of the region Q(T) do not disappear when the time T increases and, moreover, their pointedness
does not change. The limit set Q(∞) is bounded if, and only if, $, % < 0. It follows from what has been said above that this bounded limit set
has CCP in the case of condition (4.4).
The region Q(T) when $ = % = −0.5 and T = 16# is shown in Fig. 5. Numerical investigations show that the set Q(T) becomes “slightly”
larger with a further increase in the time T. The set constructed in Fig. 5 can therefore be considered as being close to the limit region Q(∞).
In other words, it is “almost” the limit AR that is shown in Fig. 5. It has two CCP.
In the case of (4.4) when $, % ≥ 0, the limit set Q(∞) occupies the whole phase space X and thereby does not have a boundary.
We now consider the last case when the condition

(4.8)
112 D.I. Bugrov, A.M. Formal’skii / Journal of Applied Mathematics and Mechanics 81 (2017) 106–113

Fig. 5.

is satisfied and turn to the consideration of equality (3.3) subject to this condition. Unlike in cases (4.1) and (4.4), in the case of condition
(4.8) the exponent e(!−ε)# and, together with it also the magnitude of the first of vectors (4.3), increases indefinitely as the value of !
increases. As in the two preceding cases, the second of vectors (4.2) rotates about the "1 axis when the value of ! is changed and executes
a complete revolution if the value of ! increases by 2#. Unlike in cases (4.1) and (4.4), in the case of condition (4.8) the angle between the
second of the vectors (4.2) and the ("2 , "3 ) plane becomes larger with an increase in the value of ! and tends to #/2 when !→ ∞. At the
same time, the set S(!) lies within the set S(! + 2#), that is, the inclusion

holds. With an increase in the time T starting from a certain instant T0 , the set S(0,T) remains unchanged, and this means that, starting from
this instant T0 , the apex angle (solid angle) of the conical surface “encompassing” the CCP also remains unchanged and, as in the case (4.4),
the CCP of the boundary of the region Q(T) do not disappear as T increases (unlike in the case of (4.1)). Note that the set S(0,T) constructed
under condition (4.4) belongs to the set S(0,T) constructed for the same values of e2 and e3 but under condition (4.8).
If $ < 0, the limit set Q(∞) is bounded when % < 0 and, consequently, has corner points in the case of (4.8) and, when % ≥ 0, it is a cylinder
with an axis x1 (x2 = x3 = 0)8 . The section of this cylinder with the plane x1 = 0 is the limit set constructed for a second order system consisting
of the second and third scalar equations of system (1.4). This set, constructed in the plane of the variables x2 , x3 (x1 = 0) is bounded by the
limit cycle consisting of the two integral curves, symmetric with respect to the point x2 = x3 = 0, obtained when u = 1 and u = −1 (see Refs 8
and 24) and has no corner points. The boundary of the limit set of attainability Q(∞) for system (1.4) is thereby smooth when % ≥ 0, $ < 0.
At the same time, the region Q(T) has two CCP for any finite value of T and the apex angles of the cones “encompassing” them do not tend
to # when T→ ∞.

Acknowledgement

This research was financed by the Russian Foundation for Basic Research (15-05-04503, the analytical studies) and the Russian Scientific
Foundation (14-50-00029, the numerical studies).

References

1. Bulgakov BV. Oscillations. Moscow: Gostekhizd; 1954.


2. Bellman R. Dynamic Programming. Princetion, NL: Princetion Univ. Press; 1957.
3. Bellman R, Glicksberg I, Gross O. Some aspects of the mathematical theory of control process. Santa Monica, CA: Rand Corp.; 1958.
4. Lee EB, Markus I. Foundations of Optimal Control Theory. N.Y. etc: Wiley; 1967.
5. Krasovskii NN. The Theory of Motion Control. Moscow: Nauka; 1968.
6. Boltyanskii VG. Mathematical Methods of Optimal Control. New York: Holt, Rinehart, and Winston; 1971.
7. Krasovskii NN. Game Problems on the Encounter of Motions. Moscow: Nauka; 1970.
8. Formal’skii AM. Controllability and Stability of Systems with Restricted Resources. Moscow: Nauka; 1974.
9. Kurzhanskii AB. Control and Observation under Conditions of Uncertainty. Moscow: Nauka; 1977.
10. Schweppe FC. Uncertain Dynamic Systems. Englewood Cliffs, NJ: Prentice Hall; 1973.
11. Chernous’ko FL. Estimating the Phase State of Dynamical Systems. Moscow: Nauka; 1988.
12. Chernousko FL. State Estimation of Dynamic Systems. Boca Raton, FL: SRC; 1994. p. 304.
13. Chernousko FL, Ovseevich AI. Ellipsoidal bounds of reachable sets: overview and new results. J. Optimiz. Theory Appl 2004;38(2):223–46.
14. Ovseevich AI, Taraban’ko YuV. Explicit formulas for the ellipsoids approximating attainability domains. J Comp Syst Sci Intern 2007;46(2):194–205.
15. Ovseevich AI, Chernousko FL. Methods of ellipsoidal estimation for linear control systems. In: Proc. 17th World Congress IFAC. 2008. p. 15345–8.
16. Polyak BT, Khlebnikov MV, Shcherbakov PS. The Control of Linear Systems with External Perturbations. The Technique of Linear Matrix Inequalities. Moscow: URSS; 2014.
D.I. Bugrov, A.M. Formal’skii / Journal of Applied Mathematics and Mechanics 81 (2017) 106–113 113

17. Formal’skii AM, Sirotin AN. Reachability and controllability sets of linear discrete systems. J Comp Syst Sci Intern 2002;41(4):507–18.
18. Sirotin AN, Formal’skii AM. Reachability and controllability of discrete-time systems under control actions bounded in magnitude and norm. Avtom Remote Control
2003;64(12):1844–57.
19. Formalsky AM, Sirotin AN. On the geomatric properties of reacheable and controllable sets for linear discrete systems. J. Oprimiz. Theory Appl 2004;122(2):257–84.
20. Formal’skii AM. On the corner points of regions of attainability. J Appl Math Mech 1983;47(4):446–72.
21. Ovseevich AL. Singularities attainable sets. Russ. J. Math. Phys 1998;5(3):389–98.
22. Ovseevich AI. Singularities of the boundaries of regions of attainability and problems of observability. In: Synopses of the Fifth Russian School “Mathematical Methods of
Navigation”. Moscow: Izd Mekh Mat Fak MGU; 1994. p. 30–2.
23. Bolotin YuV, Morgunova SI. Observability using angular measurements and the smoothness of the boundary of the region of attainability. Fundam Prikl Mat
2005;11(8):119–30.
24. Aleksandrov VV, Zhermolenko VN. On the absolute stability of second order systems. Vestn MGU Ser 1 Mat Mekh 1972;2:102–8.

Translated by E.L.S.

You might also like