Preprints of the 19th World Congress

The International Federation of Automatic Control

Cape Town, South Africa. August 24-29, 2014

Pole-Placement in
Higher-Order Sliding-Mode Control
Debbie Hern
andez Fernando Casta
nos Leonid Fridman

Departamento de Control Autom

atico, Cinvestav del IPN, Mexico.
(e-mail: [email protected]).

Departamento de Ingeniera de Control y Rob

otica, Divisi
on de
Ingeniera Electrica, Facultad de Ingeniera, UNAM, Mexico

Abstract: We show that the well-known formula by Ackermann and Utkin can be generalized
to the case of higher-order sliding modes. By interpreting the eigenvalue assignment of the
sliding dynamics as a zero-placement problem, the generalization becomes straightforward and
the proof is greatly simplified. The generalized formula retains the simplicity of the original one
while allowing to construct the sliding variable of a single-input linear time-invariant system in
such a way that it has desired relative degree and desired sliding-mode dynamics. The formula
can be used as part of a higher-order sliding-mode control design methodology, achieving high
accuracy and robustness at the same time.
Keywords: Pole-Placement; Sliding-mode control; Robust stability.
1. INTRODUCTION
Sliding-Mode Control (SMC) is by now well known for
its robustness properties in the face of unmatched perturbations and uncertainties (Edwards and Spurgeon, 1998;
Utkin et al., 1999). In the SMC approach the designer
first chooses an output with well-defined relative degree
and such that the system is minimum phase. On a second
step, the designer devises a control law that drives the
output to zero. Phase minimality then ensures that the
system states go to zero along with the output. A salient
feature of SMC is that the output (sliding variable in the
SMC literature) is driven exactly to zero in finite time,
even in the presence of matched perturbations.
Conventional SMC is restricted to outputs of relative degree equal to one 1 . In contrast, modern SMC theory (i.e.,
Higher-Order Sliding-Mode Control (HOSMC)) allows for
sliding variables with relative degree higher than one (Levant, 2003).
Conventional SMC theory is fairly complete in the sense
that there exist several methods for choosing a sliding variable with desired zero dynamics (sliding-mode dynamics in
the SMC literature). One possibility is to put the system
in the so-called regular form and use part of the state
as a virtual control that will realize the desired slidingmode dynamics on a lower dimensional system (Utkin
et al., 1999, Sec. 5.1). If the system is single-input, a
sliding variable with desired sliding-mode dynamics can
be found without recourse to a coordinate transformation,
using the formula by Ackermann and Utkin (1998). A third
possibility is to use the more recent formula presented
Research supported by Conacyt, Mexico.
1 This restriction can be also found, e.g., in passivity based control:
It was shown by Byrnes and Isidori (1991) that a system is feedback
equivalent to a passive system if, and only if, it is minimum phase
and its output is of relative degree one.

Copyright 2014 IFAC

by Drazenovic et al. (2012), which works in the multi-input

case and also obviates the need to transform the system
into a regular form. Regarding the control law, it is now
well-known that a sliding variable of relative degree one
can be robustly driven to zero in finite time by means of a
simple unit control with enough gain (Utkin et al., 1999,
Sec. 3.5).
1.1 Motivation
HOSMC is under intensive development (Bartolini et al.,
2003; Laghrouche et al., 2007; Levant and Michael, 2009;
Orlov, 2009; Pisano and Usai, 2011; Moreno and Osorio,
2012). A major achievement in this area is the finding
of a complete family of sliding-mode controllers that
can robustly drive to zero a sliding variable of arbitrary
degree (Levant, 2005). While a high-order sliding variable
might appear naturally in specific cases (e.g., in the
differentiation problem or in the estimation problem),
there is at the present no general design methodology for
choosing a sliding variable with prescribed relative degree
and prescribed sliding-mode dynamics. The work reported
on this paper is motivated by the need to fill this gap.
Allow us illustrate with the simple chain of integrators
x 1 = x2 ,
x 2 = x3 ,
x 3 = u + w ,
where x R3 is the state and u, w R are the control and
the unknown perturbation at time t (we omit the time
arguments for ease of notation). Suppose that we want to
stabilize the origin.
In the conventional approach one chooses first a sliding
variable of relative degree 1 and such that the associated
2-dimensional sliding dynamics are stable. Suppose, e.g.,
we desire sliding dynamics having an eigenvalue 1 with

1386

19th IFAC World Congress

Cape Town, South Africa. August 24-29, 2014

multiplicity 2. We can use the well-known formula by

Ackermann and Utkin to obtain the sliding variable =
x1 + 2x2 + x3 . Finally, we can apply the control law
u = x2 2x3 w
sign(), where w
is a known upper
bound for |w|. It is not hard to see that the trajectories
converge globally asymptotically to zero, regardless of w.
Consider now the case = x1 . The relative degree of
is equal to the systems dimension, so there are no sliding
dynamics to worry about. It is by now a standard result
of HOSMC theory that the (substantially more complex)
controller

+ 2(||
+ ||2/3 )1/2 ( + ||2/3 sign())
, (1)
u = k0
|
| + 2(||
+ ||2/3 )1/2
with > 0 high enough, drives the state to zero in finite
time, regardless of w.
The computation of a sliding variable of relative degree
equal to the dimension of the plant was simple because
the system is in a canonical form. This suggests that,
for a general linear controllable system, we first put
it in controller canonical form and then take the state
with highest relative degree as the sliding variable. In
this way, the extreme case of relative degree equal to
the systems dimension (no sliding dynamics) can be
covered systematically. The other extreme case, that of
relative degree 1 (sliding dynamics of codimension 1),
can be covered using Ackermann and Utkins formula.
Note, however, that there is no systematic method for
constructing a sliding variable of intermediate relative
degree (in our example, of relative degree 2). To such a
sliding variable there would correspond a sliding dynamics
of dimension 1. This dynamics can be enforced with
a controller much simpler than (1), thus arriving at a
fair compromise between order reduction and controller
complexity.
1.2 Contribution
Our main contribution, Theorem 3, concerns single-input
linear time-invariant (LTI) systems. The selection of the
sliding variable is interpreted as a zero-placement problem,
which allows us to generalize the formula of Ackermann
and Utkin to the case of arbitrary relative degree. Our
proof is simpler (more insightful) than the proof of the
original problem. The formula makes it possible for the designer to construct a sliding variable with desired slidingmode dynamics of arbitrary dimension.
For the case of relative degree 2 in our motivational
example above, application of Theorem 3 to a sliding
dynamics with desired eigenvalue -1 gives the sliding
surface = x1 +x2 . The sliding dynamics can be enforced,
e.g., with the twisting controller u = x3 k0 sign()
k1 sign(),
where k0 and k1 are high enough to reject w.
1.3 Paper Structure
In the following section we give some preliminaries on
relative degree, zero dynamics and SMC. The section is
included mainly to set up the notation and to provide
some context for our main result, which is contained in
Section 3. Section 4 provides a thorough example and the
conclusions are given in Section 5.

2. PRELIMINARIES
Consider the LTI system
x = Ax + B(u + w) , x Rn , u, w R ,
(2a)
where x is the state, u the control and w the unknown
perturbation at time t (we omit the time arguments). The
pair (A, B) is assumed to be controllable. Suppose that
we want to steer x to zero despite the presence of w.
The problem can be approached in two steps: First, find a
virtual output
= Cx , R
(2b)
such that 0 implies x 0 as t . Next, design
a feedback control law that ensures that 0 either as
t or as t T , T > 0, depending on the desired
degree of smoothness and robustness of the controller.
2.1 Relative degree and zero dynamics
Recall that (2) is said to have relative degree r if
CAi1 B = 0, 1 i < r and CAr1 B 6= 0. If (2)
has relative degree r, then it is possible to take and
its successive r 1 time-derivatives as a partial set of
coordinates 1 , . . . , r . More precisely, there exists a fullrank matrix B R(nr)n
suchthat B B = 0 and

B
  C

=
x = Tx
..

.
r1
CA
is a coordinate transformation, that is, T is invertible (Isidori, 1996, Prop. 4.1.3). It is straightforward to
verify that, in the new coordinates, system (2) takes the
normal form

0
A0 + B0

1
0
2

..
.
.

(u + w) (3a)

..
..
+
. =

r1
0
r
r
r1
CA x
CA B
r
= 1 .
(3b)
The dynamics = A0 , Rnr , are the zero dynamics.
It is well known (Marino and Tomei, 1995, Ex. 4.1.3) that
the eigenvalues of A0 coincide with the zeros of the transfer
function
g(s) = C(sI A)1 B .
If the zeros of g(s) have real part strictly less than zero,
we say that the system is minimum phase. Thus, we can
reformulate our first step as: find a virtual output such
that (2) has stable zeros at desired locations.
2.2 Sliding-mode control
If |w| is majored by a known bound, then the robust
stabilization objective can be accomplished using nonsmooth control laws (solutions of differential equations with
discontinuous right-hands are taken in Fillipovs sense).
In conventional first-order SMC (Utkin et al., 1999), the
search for is confined to outputs of relative degree one
and the control takes the form 2
2

At the expense of the loss of global stability and a higher gain k0 ,

the term CAx is sometimes omitted by incorporating it into w.

1387

19th IFAC World Congress

Cape Town, South Africa. August 24-29, 2014

CAx + k0 sign(1 )
(4)
CB
with k0 > |w|. This control law guarantees that will
reach zero in a finite time T and will stay at zero for all
future time, regardless of the presence of w. The matrix
C can be set using the formula by Ackermann and Utkin,
recalled in the following theorem.
Theorem 1. (Ackermann and Utkin (1998)). Let
e1 := [0 0 0 1]
and let P be the systems controllability matrix. If C =
e1 P 1 (A) with () = n1 + n2 n2 + + 1 + 0 ,
then the roots of () are the eigenvalues of the slidingmode dynamics in the plane = 0.
u=

Restated somewhat differently, this theorem says that the

virtual output = Cx results in a zero dynamics with
characteristic polynomial (), so that the roots of ()
are precisely the eigenvalues of A0 . Thus, by choosing a
Hurwitz polynomial () we ensure that (2) is minimum
phase, which implies that all the states converge to the
origin when = 1 is constrained to zero. In view of the
previous discussion, this amount to saying that the roots
of () coincide with the zeros of g(s). This is the key
observation that will allow us to generalize the Theorem
while simplifying its proof.
Modern sliding-mode control theory considers the more
general case of relative degree r 1. Suppose, for example,
that (2) has relative degree r = 2. The second-order
twisting controller
CA2 x + k0 sign(1 ) + k1 sign(2 )
u=
CAB
with k1 > |w| and k0 > k1 + |w| will drive = 1
and = 2 to zero in finite time (again, regardless of
w). More generally, We say that an r-sliding mode occurs
whenever the successive time derivatives , ,
. . . , (r1)
are continuous functions of the closed-loop state-space
variables and = = = (r1) = 0 (i.e., = 0).
Nowadays, it is possible to construct a controller of the
form 3
CAr x + f ()
(5)
u=
CAr1 B
enforcing an r-sliding mode for arbitrary r, though it is
worth mentioning that the complexity of f () increases
rapidly as r increases (see Levant (2003) for details).
One can think of at least two circumstances that justify
the increased complexity of higher-order sliding-mode controllers: prescribed degree of smoothness and prescribed
order of accuracy in the face of unmodeled dynamics and
controller discretization.
Regarding smoothness, suppose that (2) has relative degree r and suppose that a chain of k integrators is cascaded
to the system input, uk := u(k) .
Z 2
Z t
uk d1 dk .

u=
k =0

1 =0

The relative degree of the system with new input uk and

output is r + k. Now, an (r + k)-sliding mode has to be

Actually, in the original version (Levant, 2003), the nonlinear

counterpart of CAr x is regarded as a perturbation and omitted from
the equation. Since we are dealing with simple linear systems, we
have included it to reduce the necessary gains.

enforced by uk , but the true input u is at least k 1 times

continuously differentiable.
Regarding order accuracy, it is probably best to recall the
following theorem.
Theorem 2. (Levant (2005)). Let the control value be updated at the moments ti , with ti+1 ti = = const > 0;
t [ti , ti+1 ) (the discrete sampling case). Then, controller (5) provides in finite time for keeping the inequalities




|| < 0 r , ||
< 1 r1 , . . . ,  (r1)  < r1 (6)
with some positive constants 0 , 1 , . . . , r1 .

(See Levant (2005) for the specific form of f () in (5).) It

is also shown in Levant (2010) that in the presence of an
actuator of the form z = a(z, u), v = v(z), z Rm , v R
with u the input of the actuator, v its output and the
time constant, inequalities (6) also hold under reasonable
assumptions.
3. MAIN RESULT
We have recalled in the previous section that, for arbitrary
r, it is possible to enforce an r-sliding motion despite the
presence of perturbations. Now we show that Theorem 1
holds for arbitrary relative degree, so it can be used
to select a virtual output with desired relative degree
and desired sliding-mode dynamics. The zero dynamics
interpretation allows for a simpler proof.
Theorem 3. If
C = e1 P 1 (A) ,
(7)
with () = nr + nr1 nr1 + + 1 + 0 ,
then is of relative degree r and the roots of ()
are the eigenvalues of the sliding-mode dynamics in the
intersection of the planes = = = (r1) = 0.
Proof. Let us assume that the system is given in con To
troller canonical form with system matrices A and B.
the
verify (7), we will show that for C = e1 P 1 (A),

1 B
is equal to (s).
numerator of g(s) = C(sI
A)
It is a standard result that, for a system in controller
canonical form, we have Williams and Lawrence (2007)
e1 P 1 = [1 0 0 0]
[1 0 0 0] A

= [0 1 0 0]
..
.

(8)

[1 0 0 0] An2 = [0 0 1 0]
[1 0 0 0] An1 = [0 0 0 1] .
It then follows that
C = [0 1 nr1 1 0 0] .
B
and C are in controller canonical form, the
Since A,
transfer function is simply
snr + nr1 snr1 + + 1 s + 0
g(s) =
,
sn + an1 sn1 + + a1 s + a0
which shows that the relative degree is r. Since the
numerator is equal to (s), the eigenvalues of the slidingmode dynamics are equal to the roots of (s).

1388

19th IFAC World Congress

Cape Town, South Africa. August 24-29, 2014

4
2
0
-2
-4
-6

0.1

0.01

0
-5
0.001

-10

20

0.0001

0
-20
-40
0

1e-05
0.001

Fig. 1. Simulation results for a first-order sliding mode controller. The system is perturbed by w = 0.5 sin(10t).
The control (4) is sampled and held every = 0.001
seconds.
Now, to address the general case, consider the transformation T = P P 1 , which is such that A = T 1 AT . We have
1 , that is, C = e1 P 1 (A)T
1 . Finally, from
C = CT
1
1
1
= T (A)T we recover (7).
P = P T and (A)
4. EXAMPLE
Consider the linearized model of a real inverted pendulum
on a cart (Fantoni and Lozano, 2002)

0
01 0 0
0.97
0 0 1.56 0
(u + w) ,
(9)
x+
x =
0
0 0 0 1
3.98
0 0 46.87 0
where x1 , x2 , x3 and x4 are the position and velocity of
the cart, and the angle and angular velocity of the pole,
respectively. The system is controllable and the open-loop
characteristic polynomial is 2 (+6.85)(6.85). Suppose
that we want to regulate the state to zero, in spite of any
perturbations satisfying the bound |w| 1.
4.1 First-order sliding mode control
Consider the problem of designing a first-order sliding
mode controller with sliding-mode dynamics having eigenvalues zi = 5, i = 1, 2, 3. Applying (7) with () = ( +
5)3 gives
C = [3.2002 1.9201 4.5411 0.7166] ,
which in turn yields the expected transfer function
3

(s + 5)
.
2
s (s + 6.85)(s 6.85)
To enforce a sliding motion on the surface = 0 we apply
the control (4) with k0 = 10. Fig. 1 shows the simulated
response when
w = sin(10t) and x
0 = [1 1 1 1]
g(s) = C (sI A)

0.01

10

B=

and the control law is sampled and held every = 0.001

seconds. It can be seen that, once the state reaches the
sliding surface, the state converges exponentially to the
origin, despite w.

Fig. 2. First-order sliding mode. Order of the error as a

function of the sampling period (solid) and the actuator time-constant (dashed). The error order matches
the sliding-mode order almost exactly.
To verify the order of accuracy established in (6), we take
logarithms on both sides of the inequalities (the base is
not important),
log(| (i) |) < log(i ) + (r i) log( ) ,

i = 0, . . . , r 1 .

Notice that, on a logarithmic scale, the right-hand is a

straight line with slope r i and ordinate at the origin
log(i ). To verify that the order of the error | (i) | as a
function of is indeed r i, the closed-loop system was
simulated for several values of , both for a zero order hold
with sampling period and for a (previously neglected)
actuator of the form v = v + u. We recorded the
maximum error after the transient, lim supt | (i) |. The
best linear interpolation on a least square sense was then
computed to recover an estimate of log(i ) and ri. Fig. 2
shows that the estimations agree well with (6).
4.2 Second-order sliding mode control
Suppose now that we desire a sliding-mode dynamics
with eigenvalues zi = 5, i = 1, 2. Applying (7) with
() = ( + 5)2 gives
C = [0.6400 0.2560 0.4062 0.0621]
and
2

g(s) =

(s + 5)
.
2
s (s + 6.85)(s 6.85)

To enforce a second-order sliding motion on the surface

= = 0 we apply the control (5) with f () as in Levant
(2005), that is,


1
+ ||1/2 sign()
u=
.
(10)
CA2 x + 10
CAB
||
+ ||1/2

Fig. 3 shows the simulated response for the same perturbation, initial conditions and sampling time as before. It can
be seen that, once the state reaches the sliding surface, the
state converges exponentially to the origin, again despite
w. Fig. 4 shows the system accuracy for several sampling
times and several actuator time-constants. Inequality (6)
is again verified.

1389

19th IFAC World Congress

Cape Town, South Africa. August 24-29, 2014

4
2
0
-2
-4

4
2
0
-2
-4

2
1
0
-1
-2
-3
-4
20
10
0
-10
-20
-30
-40
0

2
1
0
-1

-2
20
10
0
-10
-20
-30
0

10

Fig. 3. Simulation results for a second-order sliding

mode controller. The system is perturbed by w =
0.5 sin(10t). The control (10) is sampled and held
every = 0.001 seconds.

10

Fig. 5. Simulation results for a third-order sliding

mode controller. The system is perturbed by w =
0.5 sin(10t). The control (11) is sampled and held
every = 0.001 seconds.

0.1
0.01

0.01

0.001
0.0001
0.0001

1e-05
1e-06
1e-07

1e-06
0.001

0.01
1e-08
0.001

Fig. 4. Second-order sliding mode. Order of the errors

as functions of the sampling period (solid) and the
actuator time-constant (dashed). The error order for
matches well with the sliding-mode order.
4.3 Third-order sliding mode control
Consider the problem of designing a third-order sliding
mode controller with sliding-mode dynamics having the
eigenvalue z1 = 5. Applying (7) with () = + 5 gives
C = [0.1280 0.0256 0.0310 0.0062] ,
which in turn yields the expected transfer function
s+5
1
g(s) = C (sI A) B = 2
.
s (s + 6.85)(s 6.85)
To enforce a third-order sliding motion on the surface
= =
= 0 we apply
1
CA3 x+
u=
CA2 B

+ 2(||
+ ||2/3 )1/2 ( + ||2/3 sign())
10
|
| + 2(||
+ ||2/3 )1/2

(11)

Fig. 5 shows the simulated response for the same perturbation, initial conditions and sampling time as before. Again,
the state converges exponentially to the origin once the
state reaches the sliding surface, despite w.

0.01

Fig. 6. Third-order sliding mode. Order of the errors

as functions of the sampling period (solid) and the
actuator time-constant (dashed). The error order for
matches well with the sliding-mode order.
5. CONCLUSIONS
We have presented a generalization of the well-known
formula by Ackermann and Utkin. A complete design
cycle can now be easily carried out. Formula (7) allows
the control designer to first specify a desired slidingdynamics of any order. Then, the sliding-mode dynamics
can be enforced using the corresponding higher-order
sliding mode controller given in Levant (2005).
It is clear that there is a trade-off between complexity of
a sliding mode controller, accuracy and order reduction
of the equations of motion. By being able to choose the
relative degree of the system, the designer can now decide
on the right compromise, depending on the particular
application at hand.
We have used the notion of accuracy in the face of sample
and hold as our main motivation for using higher-order
SMC but other criteria, such as smoothness, can also
prompt the use of higher-order SMC.

1390

19th IFAC World Congress

Cape Town, South Africa. August 24-29, 2014

REFERENCES
Ackermann, J. and Utkin, V. (1998). Sliding mode control
design based on Ackermanns formula. IEEE Trans.
Autom. Control, 43, 234 237.
Bartolini, G., Pisano, A., Punta, E., and Usai, E. (2003).
A survey of applications of second-order sliding mode
control to mechanical systems. Int. J. Control, 76:9-10,
875 892.
Byrnes, C.I. and Isidori, A. (1991). Asymptotic stabilization of minimum phase nonlinear systems. IEEE Trans.
Autom. Control, 36, 11221137.
Veselic, B., and Gligic,
Drazenovic, B., Milosavljevic, C.,
V. (2012). Comprehensive approach to sliding subspace
design in linear time invariant systems. In Proc. of
the Variable Structure Systems Workshop, 473 478.
Mumbai, India.
Edwards, C. and Spurgeon, S.K. (1998). Sliding mode
control: theory and applications. CRC, Padstow, UK.
Fantoni, I. and Lozano, R. (2002). Non-Linear Control
for Underactuated Mechanical Systems. Springer-Verlag,
London.
Isidori, A. (1996). Nonlinear Control Systems. SpringerVerlag, London, U.K.
Laghrouche, S., Plestan, F., and Glumineaub, A. (2007).
Higher order sliding mode control based on integral
sliding mode. Automatica, 43, 531 537.
Levant, A. (2003). Higher-order sliding modes, differentiation and output-feedback control. Int. J. Control, 76,
924941.
Levant, A. (2005). Quasi-continuous high-order slidingmode controllers. IEEE Trans. Autom. Control, 50, 1812
1816.
Levant, A. (2010). Chattering analysis. IEEE Trans.
Autom. Control, 55, 1380 1389.
Levant, A. and Michael, A. (2009). Adjustment of highorder sliding-mode controllers. Int. J. Robust Nonlinear
Control, 19, 16571672.
Marino, R. and Tomei, P. (1995). Nonlinear Control
Design: Geometric, Adaptive, and Robust. PrenticeHall.
Moreno, J.A. and Osorio, M. (2012). Strict Lyapunov
functions for the super-twisting algorithm. IEEE Trans.
Autom. Control, 57, 1035 1040.
Orlov, Y. (2009). Discontinuous Systems, Lyapunov Analysis and Robust Synthesis under Uncertainty Conditions. Springer-Verlag, London.
Pisano, A. and Usai, E. (2011). Sliding mode control:
A survey with applications in math. Mathematics and
Computers in Simulation, 81, 954 979.
Utkin, V., Guldner, J., and Shi, J. (1999). Sliding Modes in
Electromechanical Systems. Taylor & Francis, London,
U.K.
Williams, R.L. and Lawrence, D.A. (2007). Linear statespace control systems. John Wiley & Sons, Inc., New
Jersey.

1391