This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2016.2612058, IEEE
Transactions on Automatic Control

Note on Minimax Sliding Mode Control Design for

Linear Systems
Sergiy Zhuk, Andrey Polyakov and Olexander Nakonechnyi

Abstract—The paper studies the problem of sliding mode reachability set towards the sliding surface. We show that, for
control design for linear systems with incomplete and noisy mea- the case of noisy output measurements, the classical sliding
surements of the output and additive/multiplicative exogenous mode control algorithm is just one of many possible solutions
disturbances. First, we construct a linear minimax observer to
have an estimate of the system’s state with minimal worst-case of the aforementioned optimal control problem (see Section
error. Second, we establish the optimality of the constructed IV), which, in fact, admits both continuous and discontinuous
observer in the class of all observers represented by measurable solutions in feedback form.
functionals of the output. Finally, we propose an algorithm, We rely upon the minimax state estimation framework [11],
generating continuous and discontinuous feedbacks, which steers [3], [10] and duality argument [22], [23], [24] in order to
the observer as close as possible to a given sliding hyperplane
in finite time. The optimality (sub-optimality) of the designed construct the reachability set: the minimax state estimator or
feedbacks is proven for the case of bounded noises and additive observer constructs a linear estimate of the plant’s state with
(multiplicative) disturbances of L2 -class. The efficacy of the minimal worst-case error, provided the uncertain parameters
proposed algorithm is illustrated by a numerical example. (unknown initial condition, model disturbance, observation
Index Terms—sliding mode control noise) have bounded L2 -energy. It turns out that (i) the state
of the minimax observer coincides with the minimax center of
I. I NTRODUCTION the reachability set and (ii) the observer may be realized by
a linear time-variant system taking the output of the original
Sliding mode method was introduced more that 50 years ago system as an input. The property (ii) allows us to apply linear
(see [20] and references therein). Historically it was the first separation principle and design the sliding mode control u
approach to robust control design. Indeed, it is well-known in the state space of the observer. Specifically, (i) guarantees
(see [7], [21], [18]) that sliding mode control is insensitive to that the state of the plant (with disturbances) is enclosed in
a wide class of disturbances and uncertainties. the ellipsoid, which is centered at the state of the minimax
Robust output-based feedback control algorithms are re- estimator. Hence, if the control u steers the observer’s state
quired for many practical applications. The output-based slid- towards a sliding surface, it follows that the actual state of the
ing mode control design methodology is well-developed for plant is guaranteed to be in a vicinity of the sliding surface,
linear systems (see, for example, [6], [7], [21], [18] and provided u has been applied to the plant.
references therein). However, in practise, it is quite difficult We prove that the minimax observer (in the form of a
to apply the state of the art sliding mode methods in the linear functional of outputs), used in this paper, is optimal
case of noisy measurements [19], [16] and/or mismatched among all observers represented by measurable functionals
disturbances, e.g. [8], [2], [15]. The aim of this paper is of outputs. Thus, at least theoretically, the proposed control
to propose a mathematically sound extension of the sliding design can not be further improved by using observers realized
mode control methodology allowing one to deal with the by non-linear functionals of outputs. The latter agrees with
aforementioned cases efficiently. Specifically, we consider the numerical comparisons [25] of the control laws based on
conventional (first order) sliding mode control principles and the minimax observer and the fixed-time feedbacks [13] based
study the problem of observer-based sliding mode control on the second order sliding mode (2-sm) observer [5], which
design for a linear plant with additive/multiplicative exogenous is nonlinear. We refer the reader to [18], [7], [21], [16] for
disturbances and bounded (in L2 ) deterministic measurement further information on sliding mode observers. Finally, we do
noises. Note that, in this case, it is impossible to ensure a numerical comparison of the proposed algorithm and the
the ideal sliding mode due to noisy measurements. Instead, H ∞ -observer based sliding mode control in Section V.
one can only design a control law u providing the system A preliminary version of this paper appeared in [26]. The
motion as close as possible to the selected sliding surface. key differences with respect to [26] are as follows: we show
To design such u we first construct a reachability set for the that the minimax observer in the form of a linear functional
plant and then solve the following optimal control problem: of observations is optimal for a wide class of observers
find a feedback control u steering the minimax center of the (Proposition 1), show that optimal sliding mode control solves
S. Zhuk is with IBM Research, Server 3, Damastown Ind. Park, Dublin 15, a minimax control problem (Proposition 1) and prove the exis-
Ireland, [email protected] tence of the sub-optimal control for the case of multiplicative
A. Polyakov is with Inria Lille-Nord Europe, 40, avenue Halley, 59650 disturbances (Proposition 3).
Villeneuve d’Ascq, France, [email protected]
O. Nakonechnyi is with Taras Shevchenko National University of Kyiv, The paper is organized as follows. The next section presents
Ukraine, [email protected] the problem statement and basic assumptions. The minimax

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2016.2612058, IEEE
Transactions on Automatic Control

observer for linear systems is discussed in Section III. The construct a feedback control, which will realize the motion of
problem of control design is studied in Section IV. Next the the system as close as possible to the given surface provided
numerical simulation results and conclusions are provided. The the parameters x0 , p, d, w are unknown and satisfy (3), we
proofs of all propositions are given in the appendix. consider a minimax version of the classical Mayer optimal
Throughout the paper the following p notations are used: control problem kF x(T )k → minu (see, for example, [17],
R is the set of real numbers; kxk = x21 + . . . + x2n for [1]):
> n 2 s
x = (x1 , . . . , xn ) ∈ R ; L(a,b) (R ) denotes Lebesgue space sup kF x(T )k → min
u
of squared integrable functions mapping (a, b) to Rs ; the sign
(x0 ,d,p,w)∈ΩT (4)
s.t. (1) - (2)
function is denoted by sign, i.e. sign[ρ] = 1 for ρ > 0,
sign[ρ] = −1 for ρ < 0 and sign[0] ∈ [−1, 1]. where ΩT is the ellipsoidal set composed of all x0 , d, p, w
verifying (3). The main goal of this technical note is to find
II. P ROBLEM STATEMENT (sub)optimal solution to the minimax optimization problem (4)
in the feedback form. Note that we restrict our considerations
Let us consider the following linear system:
to causal feedback laws, i.e. u(t) may depend only on mea-
ẋ(t) = Ax(t) + (B + p(t)E) u(t) + Dd(t), x(0) = x0 (1) surements y(τ ) with τ ∈ [0, t].
y(t) = Cx(t) + w(t) , t ∈ [0, T ) , (2)
III. M INIMAX O BSERVER D ESIGN
where 0 < T < +∞, x(t) ∈ Rn is the vector of system According to the classical methodology of the sliding mode
state, u(t) ∈ Rm is the vector of the control inputs, y(t) ∈ Rk control design, the precise knowledge of the so-called sliding
is the measured output, d ∈ L2(0,T ) (Rr ) and p ∈ L2(0,T ) (R) variable σ(t) := F x(t) is required in order to ensure the
are unknown external disturbances; w ∈ L2(0,T ) (Rk ) is un- motion of the system (1) on the surface F x = 0. We stress
known deterministic measurement noise, the matrices A ∈ that this information is not available as the given output y(t)
Rn×n , B, E ∈ Rn×m , D ∈ Rn×r , C ∈ Rk×n are assumed is incomplete and noisy. In this situation, the best available
to be known and time-invariant. information about the value of F x(t) can be represented by a
We study this system under the standard assumptions: minimax estimate of x(t). Following [23] we introduce
namely, the pair {A, C} is observable and the pair {A, B} is Definition 1: Assume that l ∈ Rn and U b ∈ L2 ∗ (Rk ).
(0,t )
controllable. We do not restrict matrices E and D (recall that R t∗ >
for D ∈ range(B) and E = B the exogenous disturbances A linear functional Ubl (y) := 0 U b (τ )y(τ )dτ is called a
become matched [21], [7]). minimax estimate of l> x(t∗ ) iff
σ(U b , l, t∗ ) ≤ σ(U, l, t∗ ), ∀U ∈ L2 ∗
We further assume that x0 , p, d, w are uncertain and (0,t )
deterministic, and satisfy the following inequality: where σ(U b , l, t∗ ) := sup(x ,d,p,w)∈Ω ∗ |l> x(t∗ ) − Ubl (y)| and
0 t
the set Ωt∗ is defined by (3) with T = t∗ .
ZT
2 The number σ̂(l, t∗ ) := σ(U b , l, t∗ ) is called the minimax
x>P
0 0
−1
x 0 + w>(τ )Rw(τ )+d>(τ )Qd(τ )+ p α(τ ) dτ ≤ 1, (3) estimation error.
0 Proposition 1: Let G denote a set of all continuous map-
where P0 ∈ Rn×n , R ∈ Rk×k and Q ∈ Rr×r are given pings of L2(0,t∗ ) into R and let g ∈ G. It then follows that
symmetric positive definite matrices, α > 0. b , l, t∗ ) = inf
σ(U sup |l> x(t∗ ) − g(y)| . (5)
g∈G (x0 ,d,p,w)∈Ω
t∗
The classical sliding mode control problem is (see, [21],
In other words, the worst-case estimation error of any contin-
[7]) to find a feedback control law u which steers the state
uous mapping g, defined by the sup{} in (5), cannot be less
of (1) towards a given linear hyperplane:
than the minimax error of the linear functional Ubl (y), which
F x = 0, F ∈ Rm×n , det(F B) 6= 0, has minimal worst-case estimation error σ̂. We refer the reader
to Appendix for proofs of all the propositions.
and guarantees that the state does not leave this plane. It
Proposition 2: Let x̂(t) ∈ Rn be the solution of the ODE:
is worth noting [21], [7] that the condition det(F B) 6= 0
dx̂(t)
(
is necessary for existence of a control law, which ensures = Ax̂(t) + P (t)C > R(y(t)−C x̂(t)) + Bu(t), (6)
sliding mode on the surface F x = 0. The problem of optimal dt
selection of the matrix F in order to minimize the effects x̂(0) = 0.
of disturbances is well studied in the literature (see, e.g. [2], where P (t) ∈ Rn×n is the solution to the differential Riccati
[15]). We assume that some matrix F is selected and we equation:
need to realize the reaching phase, i.e. to find the control
Ṗ (t) = AP (t)+P (t)A> +DQ−1 D> +
law u such that F x(T ) = 0. The considered problem can P (0) = P0 , (7)
be equivalently formulated as kF x(T )k → min subject to (1) αEu(t)u>(t)E>−P (t)C>RCP (t),
- (3). Indeed, obtaining a solution of this optimization problem 1
Then Ubl (y) = l> x̂(t∗ ), σ̂(l, t∗ ) = (l> P (t∗ )l) 2 .
with zero value of the cost functional guarantees the successful
A priori reachability set. Note that ∀t∗ ∈ [0, T ] the
reaching of the sliding surface. However, due to unknown
definition of the minimax error σ b implies
measurement noises and uncertain system disturbances, the
1
exact reaching F x(T ) = 0 cannot be guaranteed. In order to l> (x(t∗ ) − x̂(t∗ ) ≤ (l> P (t∗ )l) 2 , ∀l ∈ Rn . (8)

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2016.2612058, IEEE
Transactions on Automatic Control

On the other hand, if (8) holds true for any l ∈ Rn and t ∈ where σ̂(t) = F x̂(t) and x̂ satisfies (6). Moreover, the
[0, T ] then, (see, for example, [10]) it follows that equation
x(t) ∈ {z ∈ Rn : z = x̂(t) + e, e> P −1 (t)e ≤ 1} , (9) dσ̂
= F Ax̂ + F P (t)C > R(y(t) − C x̂) + F Bu(t). (10)
i.e. the state vector x(t) belongs to the ellipsoid centered at dt
x̂(t) with axes defined by the eigenvectors of P −1 (t). It is defines dynamics of the variable σ̂.
worth noting that the estimate (9) is sharp, namely, for any Theorem 1: Any control u which verifies the equality:
t ∈ [0, T ] and for any e∗ ∈ Rn such that e> ∗P
−1
(t)e∗ ≤ 1
∗ ∗ ∗ ∗
there exist (x0 , d , w , p ) ∈ ΩT such that the equality σ̂(T ) = 0 (11)
x(t) = x̂u (t) + e∗ holds for some x(t) satisfying (1)-(2). In is a solution to the minimax control problem (4) for E = 0.
fact, (9) describes the worst-case realisation of the reachability Let us consider the classical sliding mode control [21]
set of (1), i.e., it takes into account all (x0 , d, w, p) ∈ ΩT . The
estimate (9) can be further improved if one “conditions” the usm (t) = −(F B)−1 K(t)sign[σ̂(t)], K(t) > 0. (12)
reachability set on the output signal y (see [10]). We do not use
then it follows that
this result, however, as our controller design is supposed to be
implemented without prior knowledge of y. Finally, we stress dσ̂
= F Ax̂ + F P (t)C > R(y(t) − C x̂) − K(t)sign[σ̂(t)]
that the matrix P does not depend on the control parameter u, dt
provided E = 0. This suggests to design the controller u as a and taking into account x̂(0) = 0 we derive that (11) holds
function of the center of the ellipsoid, x̂ (see Section IV). true if the relay feedback gain K is selected as follows:
Stability of the observer for the case T = +∞. If
E = 0 then, for an observable (or just detectable) pair {A, C} K(t) = kF Ax̂+F P (t)C >R(y(t)−C x̂)k+µ, ∀µ > 0 . (13)
there exist a unique symmetric matrix P ∞ ≥ 0 such that Therefore, the conventional sliding mode control is the solu-
limt→∞ P (t) = P ∞ , A − P ∞ C > RC is stable, and P ∞ tion of the minimax control problem (4) for E = 0.
solves the algebraic Riccati equation:0 = AP + P A> + The right-hand side of the closed-loop sliding mode control
DQ−1 D> − P C > RCP . The proof of this well-known fact system is discontinuous with respect to the extended state
can be found in [17]. Thus, the inclusion (9) holds true for variables (x, x̂, P ). Its solution is understood in the sense of
any t ≥ 0. Moreover, if one sets P (t) := P ∞ , then, x̂ solves Filippov. Since the system is affine with respect to control
a stable system (6), and, asymptotically, all the states x(t) input, then other definitions (Utkin and Aizerman-Pyatnickii)
are contained in the ellipsoid defined by (9). In the case of are equivalent to Filippov one (see, for example, [19], [14]). In
multiplicative disturbances (E 6= 0) the separation principle particular, Utkin’s method applied to the closed-loop system
does not apply and stability of the observer depends on the (1), (7), (6), (12), (13) gives the equivalent control of the form
control law u (see Proposition 3). (14). Therefore, to fulfill the condition (11), the continuous
We stress that d ∈ L2(0,T ) “vanish on infinity”, i.e., the
control can be designed:
measure of the set {t > T : kd(t)k > 0} tends to 0
ueq (t)=−(F B)−1 F Ax̂(t)+P (t)C > R(y(t)−C x̂(t) , (14)
 
provided T → +∞. However, it is not hard to modify the
observer in order to handle the case of d 6∈ L2(0,T ) . For
instance, the external perturbation d may contain an unknown which is also optimal for the problem (4). Indeed, this feed-
absolutely continuous function v such that v(0)> Sv(0) ≤ 1 back provides F dx̂
dt = 0, and so, taking into account x̂(0) = 0,
and limt→∞ v(t) 6= 0. In this case we introduce an augmented we obtain F x̂(T ) = 0. In fact, we have proved the following
system ẋ(t) = Ax(t) + (B + p(t)E) u(t) + Dd(t) + Dv(t), Corollary 1: The control laws (12)-(13) and (14) solve the
v̇(t) = ξ(t) and v(0) = v0 , and extend (3) by adding v0> Sv0 , minimax control problem (4) for E = 0.
and placing the term ξ > (s)W ξ(s) under the integral sign to The latter suggests that one may use linear feedback (14)
account for v0 and the unknown function ξ ∈ L2(0,T ) (Rr ) instead of the discontinuous one (12) in order to guarantee the
respectively. The matrices W, S are assumed to be symmet- optimal reaching (as close as possible) of the sliding surface
ric and positive definite. Then we set xnew := (x, v)> , in the state space of the original system.
Anew = ( A0 D 0 ), Cnew = (C 0) and apply Proposition 2
for any T < +∞. Now, to ensure stability of the observer for B. The general case: E 6= 0 and D 6= 0
T = +∞ we should assume that the pair {Anew , Cnew } is Motivated by the results of the previous section, we consider
observable and use the arguments of the previous paragraph. control laws, which guarantee sliding motion on the surface
Note that this condition is sufficient for the so-called strong F x̂ = 0 starting from the initial instant of time t = 0.
observability [18] of the augmented system in the noise-free Evidently, the output-based feedback law (14) as well as the
case (w = 0). sliding mode algorithm (12), (13) are admissible sub-optimal
IV. C ONTROL D ESIGN solutions for the optimization problem (4).
A. The case of additive disturbances: E = 0 In the case of multiplicative disturbances the matrix P
depends on u and formula (7) implies that the volume of
Denote the sliding variable by σ = F x. By using (9) (recall the ellipsoid {e : e> P −1 (t)e ≤ 1} increases if ku(t)k
from Section III that (9) is sharp) we derive grows. In this case, the chattering phenomenon, introduced by
σ(t) = F x(t) = σ̂(t) + F e(t), e> (t)P −1 (t)e(t) ≤ 1 , applying the sliding mode control, could imply the significant

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2016.2612058, IEEE
Transactions on Automatic Control

1
linear feedback+H ∞ observer
degradation of the control performance compared to the linear linear feedback+minimax observer
0.8
control strategy.
We stress that the closed-loop system (1), (6), (7), (14) has
an internal dynamic loop : the control signal u(t) depends 0.4
on P (t) which in turn depends on u(t)u>(t). The following

σ
proposition proves existence and uniqueness of the control.
Proposition 3: Let F AF+ = 0. For any T > 0 there exists 0

a sufficiently small α > 0 such that the closed-loop system

(1), (6), (7), (14) has the unique bounded solution on (0, T ).
−0.4
The restriction F AF+ = 0 is introduced for simplicity. The 0 2 4 6 8 10
t
same statement can be proven for the general case. Clearly,
the control (14) exists on (0, T ), T < +∞. Fig. 2. Comparison of minimax and H ∞ observers for linear control (14)

V. N UMERICAL E XAMPLE The volume of the ellipsoid (9) defined by the matrix P (T )
has decreased by factor 105 compared to the initial one,
First, we compare the linear feedback (14) against the namely det(P (T ))/ det(P0 ) ≈ 10−5 .
sliding mode control (12) for the case of multiplicative distur- Comparison with H ∞ observer. Let us now compare the
bances. The case of additive disturbances has been studied in performance of the linear feedback ueq (14) for the minimax
[25]. Consider the system (1) with and H ∞ observers. The H ∞ observer also deals with L2(0,T ) -

0 1 0.5
 
0.25
 disturbances, so it is applicable in our case. The difference
A = −1 0 −0.7 , B = E = 0.5 , D = 0, between H ∞ observer and minimax estimator is in that
−0.5 0.7 0 1 the latter requires a priori bounds for L2(0,T ) disturbances
(the ellipsoid (3)) unlike the former. However, prior to im-
plementing H ∞ observer one needs to select a so-called

 
C = 1 0 0 ,
1 0 0
F = 1 −1 1 , P0 =  0 2 0 , T = 10. “performance parameter” γ ∈ (0, +∞), which bounds the
R = 10, Q = 0, α = 1, 0 0 4 H ∞ -norm of the transfer function from the disturbance in-
puts to the states [12]. In fact, the state equation of H ∞
3
We also √set: w(t) = 5 sin(10t)
100 , p(t) = 5 · 10− 2 and take x0 = observer coincides with (6), and the gain P solves the Riccati
( 1/6 >
2/6 2/6 ) . We used explicit Euler discretization equation (7) with an additional additive term γ12 P 2 (t) in the
with the time step 10−2 for the simulations. right hand side. This immediately implies that the ellipsoidal
Fig. 1 compares the evolution of the sliding variable σ estimate of the observation error (see, formula (9)) has ”larger”
corresponding to the linear feedback ueq (14) (red) and the shape matrix P (t). In other words, H ∞ -observers provide
sliding mode control usm (12), (13) with µ = 0.5 (blue). The more conservative but more robust estimates as opposed to the
precision of ueq appears to be better than that of usm . If µ sharp (less conservative) estimate (9) of the minimax observer.
tends to zero the precision of usm improves. The simulations In addition, H ∞ observer performance is quite sensitive to the
exhibit the chattering phenomenon which takes place due to choice of γ, as it needs to be chosen so that the difference
1 2 >
imperfections of numerical discretization of the closed-loop γ 2 P (t) − P (t)C RCP (t) does not grow too fast. The
system with the discontinuous control. As noted, we use Euler figure 2 compares the evolution of the sliding variable σ
discretization in order to model sampling effects of digital for the minimax
√ and H ∞ observers (with γ = 2.75, and
2 >
control devices. Increase of the sampling time leads to further x0 = 3 (1, 2, 2) ).
degradation of usm . In practice, non-ideal relays (with delays VI. C ONCLUSION
or/and hysteresis) may provoke additional chattering.
The paper studies a problem of sliding mode control design
for linear time invariant systems with incomplete and noisy
ueq
usm measurements of the output and additive/multiplicative exoge-
0.3
nous disturbances. The key outcomes of the presented study
for the case of additive disturbances are as follows:
0.2
• The conventional sliding mode control provides one of
many possible solutions of the problem of optimal reach-
σ

0.1 ing of the sliding surface (4).

• Sliding motion of the observer variable may be attained
0 using a linear optimal feedback (14). In fact, the latter
coincides with the equivalent control of the sliding mode
−0.1
system. Since all the parameters of the observer state
0 2 4 6 8 10 equation are known, this equivalent control can be always
t
found explicitly!
Fig. 1. Comparison of sliding mode and linear control application Despite of the fact that both solutions (sliding mode and linear
feedback) theoretically provide the same precision, the chat-

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2016.2612058, IEEE
Transactions on Automatic Control

tering phenomenon significantly decreases the performance are extreme points of Y (as ±q0 is the maximizer/minimizer
of the sliding mode control. In the case of multiplicative of ` over E0 ). Since the set Y is convex and m ∈ Y implies
disturbances, the estimation precision depends on the control −m ∈ Y , it follows that there exist a scalar s > 0 and a
magnitude implying the additional degradation of the perfor- linear functional V0 ∈ H2? such that the hyperplanes H± :=
mance. This fact is confirmed by numerical simulations. {(`(q), yg ) : s(`(q) ± `(q0 )) + V0 (yg ) = 0} are supporting for
We conclude that the linear feedback control provides the set Y , i.e., s`(q − q0 ) + V0 (yg ) ≤ 0 and s`(q + q0 ) +
an optimal solution of output-based sliding mode control V0 (yg ) ≥ 0 for all (`(q), yg ) ∈ Y . Set V (yg ) := s−1 V0 (yg ).
design problem for linear systems with L2 noises and ad- We have: |`(q) + V (yg )| ≤ `(q0 ) for all (`(q), yg ) ∈ Y so that
ditive disturbances. This fact together with practical issues sup(q,w)∈E |`(q) + V (yg )| ≤ `(q0 ) . (∗∗)
encountered while applying the discontinuous sliding mode On the other hand, we note that 0 ≤ 2`(q0 ) = `(q0 )−V (yg )+
output-based control under the same assumptions suggests that V (yg ) − `(−q0 ) ≤ |`(q0 ) − V (yg )| + |V (yg ) − `(−q0 )| ≤
the conventional sliding mode control methodology requires a 2 sup(q,w)∈E |`(q) − V (yg )|. The latter and (∗∗) prove (∗).
radical rethink for linear systems. Perhaps, the conventional This completes the proof of the proposition’s claim.
sliding mode control becomes the unique optimal solution of
(4) provided the class of noises and disturbances is restricted B. Proof of Proposition 2
to L∞ . The rigorous study of this conjecture is left for the
Let xu , xg denote the solutions of the following ODEs:
future research.
dxu (t)
= Axu (t)+Bu(t), xu (0)=0,
VII. A PPENDIX dt (15)
dxg (t)
A. Proof of Proposition 1 = Axg (t)+Dd(t)+p(t)Eu(t), xg (0)=x0 .
dt
To prove the proposition we first rewrite the minimax Then, clearly, x(t) = xu (t) + xg (t) and
estimation problem (see definition 1) in the operator form.
As noted below (see proof of proposition 2) to estimate yg (t) := y(t) − Cxu (t) = Cxg (t) + w(t). (16)
l0> x(T ) one needs to estimate just the “noisy part” of x(T ),
The function xg may be considered as a “noisy” part of x cor-
namely xg (T ) defined by the 2nd equation of (15). Let us
responding to disturbances from the ellipsoid (3). In contrast,
recall that the latter induces a linear bounded integral operator
xu represents the “mean” value of x which corresponds to the
W : H1 → H2 . Indeed, define H1 := L2(0,T ) × L2(0,T ) × Rn ,
case: x0 = 0, p = 0 and d = 0. Let us estimate xg . To do so
H2 := L2(0,T ) , q := (p, d, x0 ) and set W q := eAt x0 + we use classical results on minimax filters, namely it can be
R t A(t−s)
0
e (Dd(s) + p(s)Eu(t))ds for any q ∈ H1 . We also demonstrated (see for instance [9]) that the minimax estimate
A> T
RT >
note that l0> xg (T ) = x> 0e l + 0 d> (s)eA (T −s) D> l + of l> xg (t∗ ) is given by Ubl (yg ) = l> x̂g (t∗ ) and the minimax
1
>
p(s)(Eu(s))> eA (T −s) lds := `(q), and `(q) is a linear estimation error can be represented as σ̂(l, t∗ ) = (l> P (t∗ )l) 2
continuous functional over H1 . In other words, computing provided yg is defined by (16) and
the minimax estimate of l0> xg (T ) given yg ∈ H2 (see (16)) (
dx̂g (t)
and assuming that (3) holds true, is equivalent to computing = Ax̂g (t) + P (t)C > R(yg (t) − C x̂g (t)), (17)
dt
the minimax estimate of `(q) provided yg = CW q + w x̂g (0) = 0.
and (q, w) ∈ E where E := {(q, w) : ρ(q, w) ≤ 1} and
1 1 −1 1 Now we recall that x(t∗ ) = xu (t∗ ) + xg (t∗ ) and so x(t∗ ) =
ρ(q, w) := k(α− 2 p, Q 2 d, P0 2 x0 )k2H1 + kR 2 wk2H2 . xu (t∗ ) + x̂g (t∗ ) + e(t∗ ), where e(t∗ ) = xg (t∗ ) − x̂g (t∗ ).
Let us now prove the proposition. Our proof relies upon Then, by definition of the minimax error we get: l> e(t∗ ) ≤
Banach separation principle. Let G (Hi ) denote the set of all 1
(l> P (t∗ )l) 2 . Define x̂ := xu + x̂g . Then it is straightforward
continuous functionals f : Hi → R and let Hi? denote a to check that x̂ satisfies (6) and the error estimate holds.
subset of G (Hi ) composed of all linear functionals, i = 1, 2.
We claim that:
C. Proof of Theorem 1
inf sup |`(q)−f (yg )| = inf ? sup |`(q)−V (yg )| . Let J(u) be the value of the functional defined by (4).
f ∈G (H2 ) (q,w)∈E V ∈H2 (q,w)∈E
Having in mind (15) we write:
Indeed, the inequality inf sup ≤ inf V ∈H2? sup(q,w)∈E
f ∈G (H2 ) (q,w)∈E σ(T ) = F xu (T ) + F xg (T ) = F xu (T ) + F x̂g (T ) + F e(T )
is obvious as H2? ⊂ G (H2 ). Let us prove that for any ` ∈ H1? = F x̂(T ) + F e(T ) = σ̂(T ) + F e(T ),
there exists q0 and V ∈ H2? such that
where e(T ) ∈ E = {z ∈ Rn : z T P −1 (T )z ≤ 1}. Since
sup |`(q) − V (yg )| ≤ `(q0 ) ≤ inf sup |`(q) − f (yg )| (∗) x̂ + E is the reachability set of the system (15) it follows that
(q,w)∈E f∈G (H2) (q,w)∈E x̂ − x ∈ E if and only if x solves equation (1) provided x0 ,
p, d and v verify (3) and, hence, we obtain:
Indeed, define E0 := {(q, w) ∈ E : 0 = CW q + w}. Clearly,
E0 is a closed convex bounded set and so supE0 `(q) = `(q0 ) < J(u) = sup kF xu (T ) + F x̂g (T ) + F ek
+∞ for some (q0 , w0 ) ∈ E0 . Note that (q, w) ∈ E0 implies e∈E

−(q, w) ∈ E0 . Hence, inf E0 `(q) = −`(q0 ). Now, define Y := If E = 0 then xg and P (T ) do not depend on u, so for
{(`(q), yg ) : yg = CW q+w, (q, w) ∈ E }. Clearly, (±`(q0 ), 0) any u : σ̂(T ) = 0 we have J(u) = supe∈E kF ek. We

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAC.2016.2612058, IEEE
Transactions on Automatic Control

now note that supe∈E kη + F ek = supe∈E supl:klk=1 l> (η + is measurable with respect to t for any fixed e, P , α and
F e) = supl:klk=1 (l> η + supe∈E l> F e) = supl:klk=1 (l> η + its norm is bounded (locally) by an integrable function m:
1 1
((P (T )F l)> F l) 2 ) ≥ supl:klk=1 (((P (T )F l)> F l) 2 ) = |RHS(t, e, P )| < m(t) for t ∈ [0, T ].
supe∈E kF ek, where the inequality used in the previous line
follows from the fact that the sphere l : klk = 1 is symmetric R EFERENCES
w.r.t. 0. Hence, supe∈E kF ek ≤ supe∈E kη + F ek, ∀η ∈ Rm . [1] V. G. Boltyanski and A. Poznyak. The Robust Maximum Principle:
Theory and Applications. Birkhauser Basel, 2012.
[2] F. Castanos and L. Fridman. Analysis and design of integral sliding
D. Proof of Proposition 3 manifolds for systems with unmatched perturbations. IEEE Transactions
on Automatic Control, 51(5):853–858, 2006.
I. We prove this proposition under assumption that [3] F.L. Chernousko. State Estimation for Dynamic Systems. Boca Raton,
F AF+ = 0, where F+ = I − B(F B)−1 F is the projector FL: CRC, 1994.
onto the hyperplane F x̂ = 0. The general case can be treated [4] E. A. Coddington and N. Levinson. Theory of Ordinary Differential
Equations. New York, McGraw-Hill, 1955.
in the similar way. [5] J. Davila, L. Fridman, and Levant A. Second-order sliding-mode
Substitution of (14) into (6) gives observer for mechanical systems. IEEE Transactions on Automatic
Control, 50(11):1785–1789, 2005.
dx̂(t) [6] C. Edwards, A. Akoachere, and S. K. Spurgeon. Sliding-mode output
= F+ Ax̂(t) + P (t)C > R[Ce(t) + w(t)] , x̂(0) = 0.


dt feedback controller design using linear matrix inequalities. IEEE
Transactions on Automatic Control, 46(1):115–119, 2001.
Hence, we derive [7] C. Edwards and S. Spurgeon. Sliding Mode Control: Theory And
Applications. CRC Press., 1998.
Z t [8] C. Edwards and S. K. Spurgeon. Sliding mode stabilization of uncertain
x̂(t) = eF+ A(t−s) F+ P (s)C > R[Ce(s) + w(s)]ds. systems using only output information. Int. J. of Control, 62(5):1129–
0 1144, 1995.
[9] Arthur J. Krener. Kalman-Bucy and minimax filtering. IEEE Trans. on
and, due to F AF+ = 0, Autom. Control, 25(2):291–292, Apr 1980.
Z t [10] A. Kurzhanski and I. Valyi. Ellipsoidal calculus for estimation and
control. Birkhauser Boston Inc., Boston, MA, 1997.
F Ax̂(t) = F A eF+ A(t-s) F+ P (s)C >R[Ce(s)+w(s)]ds = 0. [11] M. Milanese and R. Tempo. Optimal algorithms theory for robust
0 estimation and prediction. IEEE Tran. Autom. Cont., 30(8):730–738,
1985.
We proved that u(t) = (F B)−1 F P (t)C > R[Ce(t) + w(t)]. [12] K. M. Nagpal and P. P. Khargonekar. Filtering and smoothing in h∞
1 1

II. Denoting L(t)=(F B)−1F P (t)C >R CP 2 (t) R−2 we setting. IEEE Transactions on Automatic Control, 36(2):151–166, 1991.
 −1   −1 > [13] A. Polyakov. Nonlinear feedback design for fixed-time stabilization of
> P 2 (t)e(t) P 2 (t)e(t)
get u(t)u (t) = L(t) 1 1 L>(t). linear control systems. IEEE Trans. Autom. Cont., 57(8):2106–2110,
R 2 w(t) R 2 w(t) 2012.
Recalling that vv > ≤ v > vIn for any v ∈ Rn and taking into [14] A. Polyakov and L. Fridman. Stability notions and lyapunov functions
for sliding mode control systems. Journal of The Frankin Institute,
account e> (t)P −1 (t)e(t) ≤ 1 we derive: 351(4):1831–1865, 2014.
 −1 >  −1  [15] A. Polyakov and A. Poznyak. Invariant ellipsoid method for min-
P 2 (t)e(t) P 2 (t)e(t) imization of unmatched disturbance effects in sliding mode control.
u(t)u>(t) ≤ L(t) 1 1 L>(t) ≤ Automatica, 47(7):1450–1454, 2011.
R 2 w(t) R 2 w(t) [16] A.S. Poznyak. Variable Structure Systems: from Principles to Imple-
mentation, volume 66 of IEE Control Series, chapter Deterministic
1+w>(t)Rw(t) L(t)L>(t)


output noise effects in sliding mode observation, pages 45–78. The
III. Let us consider the closed loop system IET, London, UK, 2004.
[17] A.S. Poznyak. Advanced Mathematical Tools for Automatic Control
Ṗ (t)=AP (t)+P (t)A>+DQ−1D>−P (t)N P (t)+αu(t)u>(t), Engineers: Deterministic Techniques. Elsevier, 2008.
[18] Y. Shtessel, C. Edwards, L. Fridman, and A. Levant. Sliding Mode
ė(t) = Ae(t)−P (t)C > R(Ce(t)+w(t))+p(t)Eu(t) + Dg(t), Control and Observation. Birkhauser, 2014.
[19] V. I. Utkin. Sliding Modes in Control and Optimization. Springer Verlag,
u(t) = (F B)−1 F P (t)C > R[Ce(t) + w(t)]. Berlin, 1992.
[20] V.I. Utkin. Variable structure systems with sliding modes. IEEE
with P (0) = P0 , e(0)T P0−1 e(0) ≤ 1. Transactions on Automatic Control, 22(2):212 – 222, 1977.
Let us denote the right-hand side of the considered system [21] V.I. Utkin, J. Guldner, and J. Shi. Sliding Mode Control in Electro-
Mechanical Systems. CRC Press., 2009.
by RHS(t, e, P, α), where e, P are components of the ex- [22] S. Zhuk. Estimation of the states of a dynamical system described
tended state vector and α considered as a constant parameter. by linear equations with unknown parameters. Ukrainian Math. J.,
In order to prove the existence and uniqueness of the 61(2):214–235, 2009.
[23] S. Zhuk. Kalman duality principle for a class of ill-posed minimax
solution of this equation for t ∈ [0, T ] and for suffi- control problems with linear differential-algebraic constraints. Applied
ciently small α > 0, we can use the theorem on con- Mathematics and Optimisation, 68(2):289–309, 2013.
tinuous dependence on the parameter (see, for example, [24] S. Zhuk and M. Petreczky. Infinite horizon control and minimax observer
design for linear daes. In Proc. 52nd IEEE Conference on Decision and
[4], Chapter 2, Theorem 4.2.). All conditions of the afore- Control, 2013.
mentioned theorem hold for the considered system. Indeed, [25] S. Zhuk and A. Polyakov. On output-based sliding mode control
for α = 0 the system has a unique solution. The right- design using minimax observer. In 13th Workshop on Variable Structure
Systems, Nantes, France, 2014.
hand side RHS is a continuous function with respect to [26] S. Zhuk and A. Polyakov. On output-based sliding mode control design
e, P , α for any fixed t. Since for any h ∈ Rn we have using minimax observer. In European Control Conference, Linz, Austria,
|h> p(t)Ξw(t)| ≤ p2 (t)h> ΞR−1 Ξh + w> (t)Rw(t), 2015.
where Ξ = E(F B)−1 F P (t)C > R, then taking into account
the estimate of u> u obtained above we derive that RHS

0018-9286 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.