1908 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO.

11, NOVEMBER 2003

Backstepping Control of Linear Time-Varying

Systems With Known and Unknown Parameters
Youping Zhang, Member, IEEE, Barış Fidan, Student Member, IEEE, and Petros A. Ioannou, Fellow, IEEE

Abstract—The backstepping control design procedure has been Matrix Frobenius norm of .
used to develop stabilizing controllers for time invariant plants Estimate of scalar or vector signal .
that are linear or belong to some class of nonlinear systems. The Estimate error .
use of such a procedure to design stabilizing controllers for plants
with time varying parameters has been an open problem. In this shifted truncated -norm( -norm),
paper we consider the backstepping design procedure for linear .
time varying (LTV) plants with known and unknown parameters.
We first show that a backstepping controller can be designed Set ,
for an LTV plant by following the same steps as in the linear for a
time-invariant (LTI) case and treating the plant parameters as given constant , where are some
constants at each time . Its stability properties however cannot be finite constants, and is independent of .
established by using the same Lyapunov function and techniques
as in the LTI case. We then introduce a new parametrization
and filter structure that takes into account the plant parameter
variations leading to a new backstepping controller. The new
control design guarantees exponential convergence of the tracking
error to zero if the plant parameters are exactly known. If the
parameters are not precisely known but the time variations of
the parameters associated with the system zeros are known,
the appropriate choice of certain design parameters, without
using any adaptive law, leads to closed-loop stability and perfect for a given function ,
regulation. This new control design is modified and supplemented where are some finite constants.
with an update law to be applicable to LTV plants with unknown
parameters. In the adaptive control design, the notion of struc- Mean-square-error (MSE) norm of function ,
tured parameter variations is used in order to include possible .
a priori information about the plant parameter variations. With Any exponentially decaying to zero signal.
this formulation, only the unstructured plant parameters are Any positive constant.
estimated and are required to be slowly time varying, and the
structured plant parameters are allowed to have any finite speed
( ) identity matrix.
of variation. The adaptive controller is shown to be robust with
respect to the unknown but slow parameter variations in the I. INTRODUCTION
global sense. We derive performance bounds which can be used to
ESEARCH on adaptive nonlinear control has been accel-
select certain design parameters for performance improvement.
The properties of the proposed control scheme are demonstrated
using simulation results.
R erated during the last decade, after introduction of a class
of controllers for a set of general classes of nonlinear systems
Index Terms—Adaptive control, backstepping, certainty equiv- [1]–[7]. These controllers are based on integrator backstepping
alence, parametric robustness, structured parameter variations, together with other nonlinear design tools such as nonlinear
time varying systems. damping [1], [7], [8], tuning functions [7], [9], and and MT
filters [4], [7], [10], [11]. In the absence of modeling uncer-
NOMENCLATURE tainties, these controllers achieve global boundedness, asymp-
totic tracking, passivity of the adaptation loop irrespective of
The following notation is used throughout this paper, unless the relative degree, and most importantly, systematic improve-
otherwise stated. ment of transient performance [7], [12]. These controllers have
th element of vector . later on been modified so that they can tolerate a class of mod-
th coordinate column vector in . eling uncertainties, especially high frequency unmodeled dy-
th row of matrix . namics, in the global sense [13]–[16]. The set of systems which
can be controlled by these controllers includes linear time-in-
Manuscript received October 25, 2002; revised May 5, 2003. Recommended variant (LTI) systems. Moreover, for LTI systems, these con-
by Associate Editor P. A. Iglesias. This work was supported by the National trollers bear strong parametric robustness in the sense that global
Science Foundation under grant ECS 9877193.
Y. Zhang is with Synopsys, Inc., Mountain View, CA 94043 USA (email: stability can be achieved by choosing appropriate design param-
[email protected]). eters without the precise knowledge of the values of the plant pa-
B. Fidan and P. A. Ioannou are with the Department of Electrical Engi- rameters [6], [7], [17]. The corresponding adaptive controllers
neering, University of Southern California, Los Angeles, CA 90089 USA
(email: [email protected]; [email protected]). which deal with unknown but constant parameters [9], [7] can
Digital Object Identifier 10.1109/TAC.2003.819074 achieve arbitrarily improved transient performance [7], [12].
0018-9286/03$17.00 © 2003 IEEE
ZHANG et al.: BACKSTEPPING CONTROL OF LINEAR TIME-VARYING SYSTEMS 1909

The stability properties of this class of controllers are based performance. Furthermore, we show that the proposed adaptive
on the assumption that the plant parameters are time invariant controller is robust with respect to unknown but slow parameter
(TI). In most applications, however, plant parameters may vary variations. Finally, we demonstrate the properties of the devel-
with time and therefore the properties of the controllers that are oped controllers using simulations.
designed for LTI plants need to be evaluated in a time varying
(TV) environment. The early attempts to design adaptive con- II. PROBLEM STATEMENT
trollers for linear time-varying (LTV) systems are based on the
A single-input–single-output (SISO) linear plant with param-
use of the certainty equivalence approach that combines a con-
eters that are smooth and bounded and have bounded deriva-
troller structure with a robust adaptive law [18]–[21]. These con-
tives which is strongly controllable and observable is topolog-
trollers use the notion that slow time variations of the plant pa-
ically equivalent to the following observable canonical form
rameters act as a perturbation to the system in the same manner
[21], [26], [27]:
as unmodeled dynamics. Based on this notion, robust adaptive
control schemes for LTI systems are used to guarantee signal (2.1)
boundedness and small tracking error of the order of the time (2.2)
variations of the plant parameters. Later, consideration of the
TV nature of the plant and some a priori information about the where
parameter variations led to new adaptive model reference and
pole placement control designs that allow the system to be fast .. ..
TV [22], [23], [21]. These controllers bear the strong stability . .
and robustness properties of their traditional counterparts for
LTI systems. However, they can not guarantee good transient
behavior [24], [25], and generally can not be extended to non- ..
.
linear time varying systems. In this paper, we fill this gap using
..
the backstepping control design procedure. .
We first consider the use of the backstepping controllers pro- ..
posed in [6], [17] based on TI models to control LTV systems .
with known parameters by treating the time varying parameters
as constant at each time . We demonstrate that the quadratic The state–space model (2.1),(2.2) can also be represented in the
Lyapunov function-based analysis used in [6], [17] to show sta- input–output form
bility and asymptotic tracking for LTI systems does not work
for LTV systems in general, even when the plant parameters (2.3)
are known exactly at each time . In addition, we establish that where
signal boundedness can only be guaranteed if the plant param-
eters associated with the plant zeros vary slowly with time.
We, then, propose a new controller that guarantees stability
and convergence of the tracking error to zero independent of
the speed of variation of the plant parameters. The new con- are the right polynomial differential operators (PDOs) [19], [21]
troller uses integrator backstepping and nonlinear damping and for the plant. Equivalently, (2.3) can also be represented using
exploits the fact that the TV plant parameters and their varia- the right polynomial integral operator (PIO) as
tions are known exactly. The stability and performance of the
proposed controller is examined in the presence of parametric
uncertainty. The controller guarantees signal boundedness pro- We make the following assumptions about the plant.
vided that the time variations of the parametric uncertainty as- Assumption 1 The PIO is exponentially
sociated with the plant zeros are small. In particular, if we know stable with a rate at least ,
the time variations of these parameters exactly, then exponential i.e., the transition matrix cor-
regulation can be achieved for zero reference input. responding to satisfies
The new controller developed for the known parameter case for some constant
based on the LTV plant model is modified and combined with .
an adaptive law to deal with the case of unknown plant param- Assumption 2 The PDO’s , are strongly co-
eters. The notion of structured parameter variations is used to prime with known orders , , re-
incorporate any available a priori information about the modes spectively, and .
of variation of the plant parameters into the parameter estimates. Assumption 3 The plant parameters , are time
The resulting adaptive backstepping controller has the following functions which are bounded and have
advantages as applied to LTV plants. First, only the unstructured bounded derivatives.
plant parameters are estimated and are required to be slowly Assumption 4 The sign of the high frequency gain
TV. The structured plant parameters are allowed to have any fi- is known and constant, and there ex-
nite speed of variation. Second, the performance bounds derived ists a known constant such that
can be used to choose certain design parameters for improved .
1910 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 11, NOVEMBER 2003

The control objective is to design an output feedback control For LTI systems, is a constant vector. In our case, however,
law so that the closed-loop system is uniformly stable, and the is a vector function of time. The TV nature of does not affect
plant output tracks as close as possible a bounded, continu- the form of the observer equation given by (3.2) which is the
ously differentiable reference signal with measured bounded same as that with constant. The observation error equation,
derivatives up to order . however, is given by

III. BACKSTEPPING CONTROL: POINTWISE DESIGN

Let us assume that the plant parameters are known at each where . It is clear that in the LTI case, where is a
time and use the backstepping approach to design a control constant vector, as . Since is TV, i.e., ,
law that could meet the control objective if the plant parameters can no longer be guaranteed to go to zero in general.
were frozen in time at each point in time. We refer to this design Let . Then, a plant parameterization to be used for con-
as pointwise in time. In other words, we use the backstepping trol design is obtained by differentiating and using (3.2) as
design approach developed for LTI plants to a plant that is con- follows:
sidered for design purposes to be an LTI plant at each frozen
time in the parameter space. Then we examine whether such
a design approach can lead to a controller that can handle the
parameter time variations. where
We consider the state dynamics (2.1) to construct a state es-
timator. Selecting a design vector
such that the matrix (or the polynomial Introducing the notation
) is Hurwitz, we can rewrite (2.1) in the
Laplace domain (assuming frozen parameters at each time ) as

Hence, we can use the following equation to derive a state we can write
estimator:
(3.1)
(3.5)
Assuming that we have no a priori information about the state
vector, we can set to zero and expand (3.1) as The time variations affect the plant parametric (3.5) through
the signal that also depends on the filtered values of , .
Since is not considered to be known, it can only be treated
Treating the plant as LTI (the plant coefficients as constant), we
as a modeling error. The backstepping control design based on
obtain
(3.5) with is given as follows:
(3.6)
(3.7)
(3.8)

Denoting , ( (3.9)
), and ( ) by , , and ,
respectively, we obtain a state estimation scheme which is very (3.10)
similar to [6], [7], and [17]
(3.11)
(3.2)

where

(3.12)
Noticing that for and ,
where , are positive design constants and
we can generate ( ) and ( ) using the
following filters:

The control law is

(3.3) if
(3.13)
(3.4) if
ZHANG et al.: BACKSTEPPING CONTROL OF LINEAR TIME-VARYING SYSTEMS 1911

Let us analyze the stability properties of the controller (3.13) which follows from (4.1) by applying the linear swapping
designed for when applied to the TV plant with by lemma [21], [28]. Combining (4.3) with (4.2), and denoting
considering the following Lyapunov function: and by and , respec-
tively, we obtain

(4.4)
which has been used in the LTI case. Its derivative can be com-
puted using (3.5)–(3.11) as The signals , , and can be generated using the filters

(4.5)
(4.6)
(4.7)

The number of the filters can be reduced by defining

As we can see, if , then , and will decay to , i.e., combining (4.5) and (4.6) as follows:
zero exponentially fast, noting that is the generic notation
for exponentially decaying to zero signals and (4.8)
for any constant , which can be chosen
arbitrarily small. However when , due to the presence of The filter (4.7) is used for backstepping design purposes. It is
which depends on and , signal boundedness is not guaran- easy to verify that the estimation error satisfies
teed let alone asymptotic tracking unless the time variations
(4.9)
are sufficiently small or decay to zero with time.
The above analysis demonstrates that, in the presence of plant which indicates that , and therefore exponentially.
parameter variations, we can not prove stability with the control Using (4.7), (4.8), and (4.4), the following plant parameteriza-
law based on the backstepping approach for LTI plants using tion is obtained:
the quadratic Lyapunov function based analysis of [6] and [17].
In the following section, we modify the backstepping control (4.10)
design to take into account the time varying nature of the plant
parameters. Note that the observer (4.4) incorporates the TV parameters
in the filter design which gives us the desired observation error
IV. BACKSTEPPING CONTROL: TIME VARYING DESIGN (4.9). In addition, only two filters are used, hence, this observer
scheme has the potential of reducing the mathematical com-
In this section, we use the backstepping procedure for control
plexity of the control law. However, we also note that the number
design by taking into account the fact that the plant is TV. As
of th order filters required for observer (4.4) is three, which is
before, we assume that the plant parameters are known at each
one more than that in the LTI case. This is for compensating for
time .
the time variations of the plant parameters and achieve perfect
A. Observer Design for the Time Varying Plant tracking.
In the following section, we apply the backstepping proce-
The reason that the controller (3.13) can not guarantee per- dure to design a controller for (2.1) and (2.2) based on observer
fect tracking or even global stability is due to the term in the (4.4) and parameterization (4.10) that are more suitable for LTV
parameterization (3.5). The signal , which acts as a perturba- plants.
tion to the closed-loop system, is due to the time variation
of the parameter vector and depends on the closed-loop sig- B. Backstepping Controller Design
nals , , and is therefore not guaranteed to be vanishing or
Let us apply the backstepping controller design steps to the
even bounded. However, can be constructed as follows if is
LTV plant given by (4.10).
known. Consider the filter
Step 1) We treat as the first virtual control. We define
(4.1)
and define (4.11)

(4.2) and choose

It can be easily verified that with defined in (4.2) (4.12)
satisfies
Step ) In each subsequent step, we individually
treat as the virtual control and, therefore, the associ-
ated error signals and stabilizing functions are recursively
and, therefore, converges to the true state as . If is
defined as
not known then in (4.2) can be generated from
(4.3) (4.13)
1912 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 11, NOVEMBER 2003

complexity of the stabilizing functions. This is another advan-

tage of the new controller.
The results of Theorem 4.1 are based on the assumption that
the TV plant parameters are known for all . In many ap-
plications, this assumption may not hold. Consequently, it is of
interest to examine the robustness of the proposed controller
when only some nominal (TV) values of the plant parameters
are known. In the following section, we address the robustness
of the controller with respect to parametric uncertainties.

V. PARAMETRIC ROBUSTNESS OF THE PROPOSED CONTROLLER

In the previous sections, we assumed that the plant parameters
(4.14) were known exactly. A natural question one may ask is: What
if the plant parameters are not precisely known? That is, what if
In the final step, when we differentiate , the control the actual plant dynamics are described by
appears in the form of . Therefore, we can design the control
law as (5.1)
if
(4.15) instead of (2.1), and there exist errors ,
if
between the actual parameters
where is the th stabilizing function bearing the same defi- , and the pa-
nition as (4.14), which completes the design. rameters , used in the control design? This section an-
The stability of the control law (4.15) can be established by swers this question.
using the Lyapunov function Due to the parameter errors and , the observer described
in (4.2) or (4.4) is no longer a true state observer. In fact, if we
substitute and in (4.4), we obtain
(4.16)

whose derivative is given by

Applying the swapping lemma, we have

(4.17) where

for any constant , which can be chosen arbitrarily small.

From (4.16) and (4.17), it follows that as
exponentially fast. Hence, the tracking error converges to
zero exponentially, and the signals and are uniformly
bounded. To establish the boundedness of , we first see that The corresponding plant parameterization is
is bounded due to the exponential stability of .
Using the boundedness of , we can recursively establish that
and finally are all bounded.
We summarize our results using the following theorem. (5.2)
Theorem 4.1: For the LTV plant (2.1) and (2.2) with As-
sumptions 1–4, the controller given by (4.15) guarantees that
the closed-loop system is internally stable, and the tracking error A lengthy analysis based on the plant parameterization (5.2)
converges to zero exponentially fast. results in the following theorem, which establishes the robust-
We note that the traditional polynomial based model refer- ness properties of the controller (4.15) with respect to the para-
ence controller scheme cannot guarantee perfect tracking when metric uncertainties.
the TV plant parameters are completely known [19], and that Theorem 5.1: Assume that the parameter error
a different filter structure has been proposed in [22] to resolve remains small for all time in the sense that
this problem for the model reference control case. In our case, such that .
perfect tracking is achieved using two th order filters. We also Furthermore, select the design parameters , to satisfy
notice that by using only the signals and instead of a series
of ’s and ’s, we have significantly reduced the mathematical (5.3)
ZHANG et al.: BACKSTEPPING CONTROL OF LINEAR TIME-VARYING SYSTEMS 1913

VI. ADAPTIVE BACKSTEPPING CONTROL

In the previous sections, it is assumed that the plant parame-
ters are known precisely or with some small error. In this sec-
(5.4) tion, we consider these parameters as unknown functions of
time which satisfy Assumptions 1–4. In order to incorporate any
available a priori information about the modes of variation of
and
the plant parameters, we use the structured parameter variations
representation [21], i.e., we assume that the plant parameters
(5.5) , have the following known structure:

for , where is a state-

space realization of the equation shown at the bottom of the
page. is the positive–definite symmetric solution (6.1)
to the Lyapunov equation
, and is a constant positive–definite matrix where , form the decomposition
satisfying . Then, there exists a of which is a matrix of known time functions,
such that , if the derivatives of the parameter errors is the unstructured parameter vector that is un-
satisfy known; is a known parameter vector which can
be decomposed to . Note that the leading
rows of , are zeros. Furthermore, we assume
the following about the leading nonzero term of and the un-
structured parameters.
for some , , then the closed loop system (5.1), Assumption 5: The sign of is the same as the sign of
(2.2), (4.15) is uniformly stable, and the tracking error is of the for all . Moreover, the unstructured parameter vector
order of in the mean square sense. is differentiable with respect to time and satisfies
Proof: The proof is long and technical, and is presented in
Appendix A. , i.e., the signals are bounded and
Remark 5.1: Theorem 5.1 indicates that the uncertainty in
the parameters can be counteracted by increasing the values
of the design parameters, , , , in particular. Note
that for sufficiently small, we can find design parameters ,
such that (5.3)–(5.5) hold. As for the parameters, only the and for some and , a ”small” scalar.
derivatives of the parameter errors have to be small, not nec- Assumption 5 requires the mean square value of the time vari-
essarily the parameter errors themselves. This suggests that the ations to be of order , where will be required to be small.
backstepping controller has strong parametric robustness as op- Next, we exploit the TV model based filter design of Section IV
posed to the traditional ones. A special situation is the LTI case, to construct a state estimator for the unknown parameter case.
where the time variations of the plant parameters or parameter Using (6.1), we can rewrite (4.7) and (4.8) as
errors are zero. Then, we reach the same conclusion as in [6]. In
this case, if the reference input is zero, then exponential regula-
tion is achieved.
In this section, we assume that the nominal values of the TV
plant parameters are known. For stability, we require that the
parametric uncertainty is small in the sense that the time varia-
tion (first time derivative) of the parametric uncertainty is small
in the average sense, i.e., small most of the time. In the fol- Applying the linear swapping lemma, we get
lowing section, we combine the proposed controller designed
for LTV plants with known parameters with an appropriate pa-
rameter estimation scheme to deal with the case of unknown
plant parameters.
1914 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 11, NOVEMBER 2003

where , , linearly independent time functions, then the filter can be

, and . Hence, obtained using th order filters. For example, suppose
constructing the filters

(6.2)
(6.3)
where , ’s are constant matrices,
(6.4)
are linearly independent scalar time functions, then it suffices to
(6.5) implement

where , is
such that is exponentially stable with stability
margin , i.e., has all eigenvalues nonpositive, and
considering the “virtual observer” where . The matrix , can be realized using
as
(6.6)

it is straightforward to verify that the observation error

satisfies

When , as . Hence, (6.6) is a true state Moreover, if , can be linearly represented by ,

observer for (2.1), (2.2) when the parameter vector is known , respectively, say
and constant. If is not constant, then the observation error
is nonvanishing and is represented in the following transfer
function form:

(6.7) where , are constant vectors of length , then the filter

signals and can be obtained through , as follows:
Using (6.7), we can obtain the following plant parameterization:

(6.8)

where
In this case the filters are of total order . In partic-
ular, when the parameters are not structured, and
, then the total filter order is , which is similar
to the LTI case [6], [7], [17].

A. Certainty Equivalence Control Law

The controller design follows the same procedure as in the
Similar to the pointwise design of Section III, the parameter- known parameter case. The idea is to recursively treat
ization (6.8) appears to be in the same form as the LTI case [7], as a virtual control signal, and apply the backstepping procedure
[17] except for the term in , and is suitable for applying using the certainty equivalence, i.e., replacing the unknown pa-
the adaptive backstepping design. rameter vector with its on-line estimate . The design steps
Remark 6.1: Note that (6.1) covers the general case in- are as follows.
cluding the fully structured, unstructured, and known parameter Step 1)
cases. If the parameters are unstructured, then we simply have
. If the parameters are fully structured, then (6.9)
is constant but unknown. The case where corresponds
to the known parameter case. (6.10)
Remark 6.2: Even though the filters (6.2) and (6.3) appear
to be of high order since both , are matrices, the Step )
actual implementation of these two filters can be of lower order,
depending on the elements of , . In general, if contains (6.11)
ZHANG et al.: BACKSTEPPING CONTROL OF LINEAR TIME-VARYING SYSTEMS 1915

suitable for synthesizing an adaptive law based on a Lyapunov

function. We define the auxiliary filter

(6.16)

and the auxiliary error signal

Then, the error signal satisfies the equation

(6.12) Since is not guaranteed to be bounded, we introduce the fol-

lowing normalizing signal:

(6.13) where and are design constants.

Some important properties of are given by the following
lemma.
In Step ), the control appears in the form of Lemma 6.1: We have
, therefore the control law can be
chosen as (6.17)

if (6.18)

if (6.19)
(6.14)
where is either 0 or 1, the latter corresponding to the case where is a positive constant.
where appears explicitly in the control law. Note that for the Proof: The state (2.1) can be rewritten as
control law (6.14) to exist, the adaptive law must assure that
.
With the control law (6.14), the corresponding error system from which we obtain
is given by
(6.20)

(6.15) Observing (6.2)–(6.5) and (6.20), we see that , , , , and

where , can be represented as outputs of stable filters with inputs and
. Hence, the result (6.17) follows immediately. Similarly, (6.7)
implies that , which together with
.. .. .. (6.17) leads to (6.18). Finally, using the inequality
. . .
, we obtain
.. ..
. .
.. .. .. ..
. . . .

Now, define the normalized estimation error

Then, satisfies

(6.21)
B. Adaptive Law With an Auxiliary Filter
The adaptive law for generating the parameter estimates used By considering (6.21), (6.19), and the following Lyapunov-like
in the control law (6.14) is based on the idea of introducing function:
an auxiliary filter to counteract the effect of term in the
equation, therefore ending up with a new error system that is (6.22)
1916 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 11, NOVEMBER 2003

the following robust adaptive law can be chosen: we obtain

(6.23)

where is the nominal value of , i.e., is expected to be (6.26)

close to the constant vector , is a positive definite symmetric
gain matrix, is the projection operator to project along the In view of (6.18), if ,
boundary , which is defined hence, (6.26) implies that . In addition, integrating
as (6.25), we get that

if and
otherwise
and is the leakage coefficient [28] defined as
Using and , it follows that
if
if (6.24) , and, consequently, using i) and ii),
it follows that .
Due to the linearity of the stabilizing functions, depends
is a known upper bound for , and is a small only on the parameter estimates and is linear in , , ,
constant. , , thus , and follows from
The stability properties of the adaptive law are described by (6.23) and . Using (6.16), satisfies
the following lemma.
Lemma 6.2: Assume that , then the (6.27)
adaptive law (6.23) guarantees the following.
i) , , from which follows immediately.
. If , then and all the
ii) . properties become properties, i.e., . Fi-
iii) , , , , , . In particular, nally, using (6.21) and (6.27) we see that
, which together with implies that
if , then and as .
as . Having established the stability properties of the adaptive law,
Proof: The properties in i) are direct consequences of the we analyze the closed-loop stability properties of the adaptive
projection and switching -modification, see [28]. To prove ii) control scheme based on the error system (6.15) next. The fol-
and iii), let us consider the Lyapunov-like function (6.22). For lowing theorem summarizes the results of this analysis, which
simplicity and without loss of generality, we assume . is presented in Appendix B in details.
The derivative of along the solution of (6.21), (6.23) can be Theorem 6.1: The adaptive controller described by (6.14)
computed as and (6.23), when applied to the LTV plant (2.1),(2.2), guarantees
the existence of a constant such that , all
the closed-loop signals are uniformly bounded, and the tracking
error is of the order in the mean square sense, i.e.,

(6.28)

where and are finite positive constants. Moreover, can

be expressed as
(6.25)
where
where and is a finite positive constant indepen-
dent of , , .
Proof: The proof is similar to that of Theorem 5.1 and is
Using the inequality presented in Appendix B.
Theorem 6.1 indicates that, using the adaptive controller
(6.14) and (6.23), only the time variations of the unstructured
plant parameters are required to be small in the mean square
sense to guarantee closed-loop stability and tracking with small
ZHANG et al.: BACKSTEPPING CONTROL OF LINEAR TIME-VARYING SYSTEMS 1917

MSE. The overall system is not necessarily restricted to be

slowly TV. The requirement about the time variations of the
unstructured plant parameters is necessary since it is not pos-
sible to estimate unknown and arbitrarily fast TV parameters
using a general adaptive law with finite speed of adaptation
[22], [28]. Once this requirement is satisfied, the mean square
tracking error is guaranteed to be of the order of the speed of
the unstructured plant parameter variations.
Besides establishing stability and tracking properties, the
theorem provides guidelines for performance improvement as
well. It shows that the MSE performance can be improved by
amplifying , , and possibly for small enough to satisfy
the stability conditions. Unlike [12], arbitrary performance
improvement is only assured in terms of the normalized
tracking error . The bound on depends
on . However, although might increase by (a)
increasing , , , this can be counteracted by reducing
the normalization coefficient . Hence, the performance of the
adaptive backstepping controller can be improved by adjusting
the design parameters , , , and . We demonstrate
this fact via simulations in Section VII.

C. Fully Structured Parameter Variations

The case of fully structured parameter variations corresponds
to being constant. We generalize it to the situation that
. For this special class of LTV
plant, the proposed adaptive controller (6.23) and (6.14) has
the following properties.
Corollary 6.1: If the speed of parameter variations satisfy
, then the adaptive controller
(b)
(6.14) and (6.23) guarantees that all the closed-loop signals are
uniformly bounded, and the tracking error converges to zero Fig. 1. Response using the pointwise design and exact knowledge of g , g ,
g . (a) c = d = c = d = 1. (b) c = c = 5, d = d = 1.
asymptotically.
Proof: This is a direct consequence of Theorem 6.1 and
Lemma 6.2 iii). where , , and . It
Due to the transformation (6.1), the parameter vector may is required to design a controller so that the output tracks the
not reflect the plant parameters themselves, and can contain reference signal .
more or less than elements, which corresponds to Let us first apply the pointwise design, i.e., the control
the overparameterized and the underparameterized case, respec- scheme (3.6)–(3.13) together with the estimation filters (3.3)
tively. Corollary 6.1 indicates that when full knowledge of the and (3.4), assuming that , , are all known exactly. Noting
parameter variations is available, then regardless of the speed of that , , for the plant, the filter parameters are
the parameter variations of the plant, global stability is guaran- chosen as , and the design parameters are chosen
teed, and asymptotic tracking is achieved. as . Fig. 1(a) shows the result. Although
In addition, in the case of fully structured parameter varia- the output signal is bounded, tracking performance is not that
tions, is exponentially vanishing. Therefore, in this case, we successful. Next, we increase the values of the design constants
can apply the tuning design given in [6], [17] instead of the cer- and to 5. Tracking performance is enhanced as shown in
tainty equivalence approach using parameterization (6.8). The Fig. 1(b). However, asymptotic tracking is not achieved.
advantage is a guaranteed performance improvement, as in the Then, we repeat the same simulations with the LTV design
TI case [7], [12]. i.e., the control scheme (4.11)–(4.15) together with the estima-
tion filters (4.5)–(4.7). Tracking is perfect as shown in Fig. 2.
VII. SIMULATION RESULTS Next, we consider some parametric uncertainty. We assume that
Let us consider a simple unstable second-order LTV plant our plant model is a little bit erroneous, e.g., models of the ac-
whose state-space representation is tual functions , , of the plant are ,
. Choosing the design parameters as ,
(7.1) , we can see from Fig. 3 that the system is
stabilized, and a relatively small tracking error (smaller than
(7.2) that of controller (3.13) with known parameters) is obtained.
1918 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 11, NOVEMBER 2003

(a) (a)

(b) (b)
Fig. 2. Response using the LTV design and exact knowledge of g , g , g . Fig. 4. Adaptive control with different choices for c , c , 0 (d = d = 1,
(a) c = d = c = d = 1. (b) c = c = 5, d = d = 1.  = 2,  = 1). (a) Tracking. (b) Parameter estimation.

Finally, we consider the unknown parameter case assuming

that the plant structure (7.1)–(7.2) is known but the functions ,
, are unknown. In order to build up an adaptive controller,
we first write the plant parameter in structured parameter varia-
tions form as follows:

where

Fig. 3. Response using the LTV design in presence of parametric uncertainty

(g (t) = 1, g (t) = g (t) = 2, c = c = 5, d = d = 1).

Note that the parametric uncertainty has amplitude 1, however,

its derivative has a much smaller amplitude 0.1. This demon-
strates that the time variation of the uncertain parameter, not
the size of uncertainty itself, determines the system stability and Noting that and hence are zero, the estimation filters
performance. are implemented using (6.2), (6.4), and (6.5). The following
ZHANG et al.: BACKSTEPPING CONTROL OF LINEAR TIME-VARYING SYSTEMS 1919

where , , , , , , are design constants, is the

switching- coefficient defined in (6.24).
Fig. 4 shows the tracking error and the parameter
estimate for simulations with different choices of the design
parameters and the adaptive gain parameters. The switching
and normalization parameters are set as , ,
, in all of these simulations. The response for
, , is redrawn in Fig. 5(a)
in order to make it comparable with the results for the cases
with known and unknown parameters. As can be seen in these
figures, the system is stabilized, and the tracking error remains
in a neighborhood of 0, for all design parameter choices.
As can be seen in Fig. 4, increasing the value of the design
parameters improves the tracking performance as in the
(a) known parameter case. By increasing the adaptive gain, not only
parameter estimation gets faster but the tracking performance is
improved further as well.
Later, fixing , , , ,
, , the effect of the normalization coefficient
is tested. The results in terms of the tracking errors are shown
in Fig. 5(b). As seen in this figure decreasing has a similar
effect with increasing on enhancement of tracking.
In the aforementioned simulations, we see that the param-
eter estimates adapt to the parameter changes. We have also
observed that the control effort remains within a reasonable
bound. Since the only unknown TV parameter is slowly
time varying, stability is guaranteed.

VIII. CONCLUSION
(b)
In this paper, we introduced a new backstepping controller
Fig. 5. (a) Response using the adaptive controller (c = c = 5, d = d =
1, 0 = 10,  = 2,  = 1). (b) Adaptive tracking with different choices for for LTV systems with known and unknown parameters. The
 (c = c = 5, d = d = 1, 0 = 10,  = 2). controller guarantees exponential tracking when the plant pa-
rameters are known exactly. When the plant parameters are not
control law is designed based on the steps in Section VI se- known exactly but their time variations are small enough, re-
lecting : gardless of the size of the parameter errors (except for the high
frequency gain), global stability can be guaranteed by choosing
certain design parameters properly. Hence, the proposed con-
troller has strong parametric robustness properties which most
of the traditional model reference controllers do not have.
When the plant parameters are unknown, the proposed con-
troller is combined with an online parameter estimator to form a
new adaptive controller. This new adaptive controller guarantees
the following. All the closed-loop signals are globally uniformly
bounded. The tracking error remains small and of the order of
the speed of the unstructured plant parameter variations, which
is required to be small in the mean square sense. If the plant
parameter variations are fully structured, the tracking error con-
The adaptive law and the associated auxiliary signal are de- verges asymptotically to zero. The performance bounds for the
fined as tracking error developed can be used to select certain design
parameters for performance improvement. The expected perfor-
mance of the proposed controller and the effects of design pa-
rameter selections on the transient performance are illustrated
by simulation results.
The proposed controller is suitable for use in many applica-
tions where the plant parameters are time varying. An example
of such application is the control of aerospace systems where the
parameters of the system vary with time and/or flight conditions
1920 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 11, NOVEMBER 2003

[29]–[31]. Application of the proposed backstepping scheme to where . Suppose that

flight control of high performance aircrafts and hypersonic air is a state-space realization of
breathing vehicles is currently under investigation. , then
(A.3)
APPENDIX I (A.4)
PROOF OF THEOREM 5.1
Note that and are independent of the design
Using (5.2), (4.11), and (4.14), we obtain the following error parameters . Since is an exponentially stable matrix,
equation: there exists a positive-definite matrix such that
(A.1)
Define , then we substitute (A.4) into the
where
error equation of and get

(A.5)
.. .. ..
. . . Next, we consider the term , which we write in the following
.. .. state-space form:
. .
.. .. .. ..
. . . .

(A.6)
Augmenting (A.6) with (3.3), we get

(A.7)
We first consider the term . We have
Finally, we augment the error systems (A.1) and (A.3) with
(A.7) using (A.5) and get

.. .. (A.8)
. .

where

(A.2)

Let the th order monic polynomial and the th order

monic polynomial be a decomposition of , i.e.,
. Then, the operator in of (A.2) can .. ..
be written as . .
.. ..
. .

.. ..
. .
.. .. ..
. . .
.. .. .. ..
which is the sum of two proper and exponentially stable I/O . . . .
.. .. ..
operators. Hence, the operator in of (A.2), which we denote . . .
as , is a proper and exponentially stable I/O operator. .. .. .. ..
. . . .
Using (A.2), (4.11), and the definition of , we can write
ZHANG et al.: BACKSTEPPING CONTROL OF LINEAR TIME-VARYING SYSTEMS 1921

We first analyze the homogeneous part of (A.8) by individu- we get

ally considering the following two partial Lyapunov functions:

where are constants to be chosen and

is a constant matrix satisfying . The
derivatives of , along the solution of (A.8) are computed (A.9)
as

(A.10)
Let us choose

where is an arbitrary constant, and consider the

Lyapunov function
Using the inequalities
where is another constant to be chosen. Using (A.9) and
(A.10), we have the following:

We first pick

where , are arbitrary constants. With these choices,

if

where are arbitrary constants, then

where . Since , , , ,
are arbitrary, the existence of , , is guaranteed pro-
vided that (5.3)–(5.5) are satisfied. Hence, if , , and for
1922 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 11, NOVEMBER 2003

are chosen to satisfy (5.3)–(5.5), then the homoge- where , , are chosen as before. Using (A.13) and conti-
neous part of (A.8) is exponentially stable. Now, let us suppose nuity of
, , and for are chosen as such, and go back to
(A.8). Since , the corresponding output is also
bounded. Therefore it suffices to consider the subsystem where (A.16)
, , are all zero. Using (A.8), (A.16), and the fact that
Next, we define a fictitious normalizing signal

(A.11)
for an arbitrary function , the derivative of can be computed
where is a constant such as
that , are exponentially
stable. The normalizing property of can be described by
the following lemma.
Lemma 8.1: Regardless of the boundedness of any closed-
loop signal, we have

(A.12)

for all and some . Therefore, , all the normalized signals are
Proof: Using the inequality bounded and small in the order of in the mean square sense. In
(A.13) particular, . On the other hand, using (3.6)–(3.12) and
(4.11)–(4.14), we can represent the control law (3.13) or (4.15)
which follows directly from the definition of given by in the form
(A.11), we have that for all for which
is exponentially stable

(A.17)

for some bounded functions , , , .

Substituting

in (A.17), we get

where , is a Hurwitz polynomial of

(A.14) order , ,
where is the impulse response of and . Hence, applying [30, Th.
. Note that is finite since is (22), p. 113], we obtain
exponentially stable. Using Holder’s inequality, we have that
and

(A.15) where and are impulse responses of

and , respectively. Since
The result follows directly from (A.14) and (A.15). and are both exponentially
Now, define the following normalized errors: stable and , , , are bounded smooth functions of
which is a bounded function of time, the supremum terms
mentioned before are all finite. Hence, using , we get

and consider the Lyapunov function

ZHANG et al.: BACKSTEPPING CONTROL OF LINEAR TIME-VARYING SYSTEMS 1923

or where is a constant such that

is an exponentially stable operator. Applying the arguments we
have used in the proof of Theorem 5.1 again, we derive that

where is some constant. Applying the Bellman–

Gronwall Lemma, we have where is some finite positive constant. Since

Now, if
and

we have
where , then and .
Once is bounded, is bounded. Then, we apply the same
argument as before, and conclude that all the closed-loop signals
are uniformly bounded, and the closed-loop system is internally where , i.e., such
stable. In addition, all the error signals satisfy that
. That is, all the error signals, including the tracking error
are of the order of in the mean square sense, i.e.,

Applying the Bellman–Gronwall lemma, we obtain

for any and some constants independent of

.
Let , then , we have .
APPENDIX II
Since bounds which bounds all the closed-loop signals,
PROOF OF THEOREM 6.1 it follows that all signals are uniformly bounded. In addition, the
Following equations (6.9)–(6.13), it can be easily shown that tracking error satisfies .
the control law (6.14) can be represented as In order to get some quantitative results, we first derive some
performance bounds for the estimation error . We start by cal-
culating the bound of the Lyapunov function . Noting that
is assumed to be zero for simplicity and without loss of gener-
ality, (6.26) yields
(B.1)

for some continuous functions , , , . On the other

hand, from (2.3), we get (B.3)
From (B.3), we notice that is a function which decreases
(B.2) with increasing , , . Therefore, the bound on ,
is decreasing with the increase of , , . Using the fact
Substituting (B.2) into (B.1), we obtain that , and , we
integrate (6.25) and get

where so that is some Hurwitz (B.4)

polynomial of order , ,
and
. Let us define a fictitious normalizing (B.5)
signal
where is some positive constant independent of
.
1924 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 48, NO. 11, NOVEMBER 2003

Next, we derive some performance bounds for the auxiliary [13] Y. Zhang and P. Ioannou, “A new class of nonlinear robust adaptive con-
signal . Considering the quadratic function trollers,” Int J. Control, vol. 65, no. 5, pp. 745–769, 1996.
[14] Y. Zhang and P. A. Ioannou, “Robustness and performance of a modi-
fied adaptive backstepping controller,” Int. J. Adapt. Control Signal Pro-
cessing, vol. 12, no. 3, pp. 247–265, 1998.
[15] F. Ikhouane and M. Krstić, “Robustness of the tuning functions adaptive
backstepping design for linear systems,” IEEE Trans. Automat. Contr.,
vol. 43, pp. 431–437, 1998.
it follows directly from (6.27) and the definition of that [16] M. Arcak, M. Seron, J. Braslavsky, and P. Kokotović, “Robustification
of backstepping against input unmodeled dynamics,” IEEE Trans. Au-
tomat. Contr., vol. 45, pp. 1358–1363, July 2000.
[17] M. Krstić, I. Kanellakopoulos, and P. V. Kokotović, “Nonlinear design
of adaptive controllers for linear systems,” IEEE Trans. Automat. Contr.,
vol. 39, pp. 739–752, 1994.
Therefore [18] G. Kreisselmeier, “Adaptive control of a class of slowly time varying
plants,” Syst. Control Lett., vol. 8, no. 2, pp. 97–103, 1986.
[19] K. S. Tsakalis and P. A. Ioannou, “Adaptive control of linear
time-varying plants,” Automatica, vol. 23, no. 4, pp. 459–468, 1987.
[20] R. H. Middleton and G. C. Goodwin, “Adaptive control of time-varying
linear systems,” IEEE Trans. Automat. Contr., vol. 33, pp. 150–155,
1988.
[21] K. S. Tsakalis and P. A. Ioannou, Linear Time Varying Systems: Control
and Adaptation. Upper Saddle River, NJ: Prentice-Hall, 1993.
Since , , , depend only on , , , , and is non- [22] , “Adaptive control of linear time-varying plants: A new model ref-
erence controller structure,” IEEE Trans. Automat. Contr., vol. 34, pp.
increasing with the increase of , using (6.23) and
1038–1047, Oct. 1989.
(B.5), we obtain [23] , “A new indirect adaptive control scheme for time-varying plants,”
IEEE Trans. Automat. Contr., vol. 35, pp. 697–705, June 1990.
[24] Z. Zang and R. R. Bitmead, “Transient bounds for adaptive control sys-
tems,” IEEE Trans. Automat. Contr., vol. 39, pp. 171–175, Jan. 1994.
[25] I. M. Y. Mareels and R. R. Bitmead, “Nonlinear dynamics in adaptive
control: Chaotic and periodic stabilization,” Automatica, vol. 22, pp.
(B.6) 641–655, 1986.
[26] L. M. Silverman, “Transformation of time-variable systems to canonical
(phase-variable) form,” IEEE Trans. Automat. Contr., vol. AC-11, pp.
where is some positive constant independent of , , . Fi- 300–303, Mar. 1966.
nally, combining (B.4) and (B.6) we obtain (6.28). [27] L. M. Silverman and B. D. O. Anderson, “Conrollability, observability,
and stability of linear systems,” J. SIAM Control, vol. 6, no. 1, 1968.
[28] P. A. Ioannou and J. Sun, Robust Adaptive Ccontrol. Upper Saddle
REFERENCES River, NJ: Prentice-Hall, 1996.
[29] R. Wiśniewski, “Linear time-varying approach to staellite attitude
[1] I. Kanellakopoulos, P. V. Kokotović, and A. S. Morse, “Systematic de-
control using only electromagnetic actuation,” AIAA J. Guid., Control,
sign of adaptive controllers for feedback linearizable systems,” IEEE
Dyna., vol. 23, no. 4, July-Aug. 2000.
Trans. Automat. Contr., vol. 36, pp. 1241–1253, 1991.
[30] J. J. Zhu and M. C. Mickle, “Missile autopilot design using a new linear
[2] R. Marino, I. Kanellakopoulos, and P. V. Kokotović, “Adaptive tracking
time-varying control technique,” AIAA J. Guid., Control, Dyna., vol. 20,
for feedback linearizable SISO systems,” in Proc. 28th Conf. Decision
no. 1, Jan.-Feb. 1997.
Control, Dec. 1989, pp. 1002–1007.
[31] J. M. Biannic, P. Apkarian, and W. L. Garrard, “Parameter varying con-
[3] R. Marino, P. Tomei, I. Kanellakopoulos, and P. V. Kokotović, “Adap-
trol of a high-performance aircraft,” AIAA J. Guid., Control, Dyna., vol.
tive tracking for a class of feedback linearizable systems,” IEEE Trans.
20, no. 2, Mar.-Apr. 1997.
Automat. Contr., vol. 39, pp. 1314–1319, June 1994.
[32] C. A. Desoer and M. Vidyasagar, Feedback Systems: Input–Output Prop-
[4] R. Marino and P. Tomei, “Global adaptive output-feedback control of
erties. New York: Academic, 1975.
nonlinear systems, Part I: Linear parameterization,” IEEE Trans. Au-
tomat. Contr., vol. 38, pp. 17–32, Jan. 1993.
[5] M. Krstić and P. V. Kokotović, “Adaptive nonlinear design with
controller-identifier separation and swapping,” IEEE Trans. Automat.
Contr., vol. 40, pp. 426–440, 1995.
[6] M. Krstić, I. Kanellakopoulos, and P. V. Kokotović, “Passivity and para-
metric robustness of a new class of adaptive systems,” Automatica, vol. Youping Zhang (M’02) received the B.S. degree
30, pp. 1703–1716, 1994. from the University of Science and Technology of
[7] , Nonlinear and Adaptive Control Design. New York: Wiley, China, Hefei, Anhui, R.O.C. in 1992, and the M.S.
1995. and Ph.D. degrees in electrical engineering from the
[8] I. Kanellakopoulos, P. V. Kokotović, and A. S. Morse, “A toolkit for University of Southern California, Los Angeles, CA
nonlinear feedback design,” Syst. Control Lett., vol. 18, pp. 83–92, 1992. in 1994 and 1996, respectively.
[9] M. Krstić, I. Kanellakopoulos, and P. V. Kokotović, “Adaptive nonlinear From 1996 to 1999, he was a Research Engineer
with the United Technologies Research Center, East
control without overparameterization,” Syst. Control Lett., vol. 19, pp.
Hartford, CT. He joined Numerical Technologies,
177–185, 1992.
Inc., San Jose, CA, in July 1999 and stayed for
[10] G. Kreisselmeier, “Adaptive observers with exponential rate of conver- nearly four years until it was acquired by Synopsys
gence,” IEEE Trans. Automat. Contr., vol. AC-22, pp. 2–8, Feb. 1977. in February 2003. At Numerical Technologies, he was in several different
[11] I. Kanellakopoulos, “Passive adaptive control of nonlinear systems,” Int. positions including software product development, technology research, and
J. Adapt. Control Signal Processing, vol. 7, pp. 339–352, 1993. technical marketing for resolution enhancement technologies. He is currently
[12] M. Krstić, P. V. Kokotović, and I. Kanellakopoulos, “Transient perfor- a Technical Marketing Manager for Mask Synthesis at Synopsys. His research
mance improvement with a new class of adaptive controllers,” Syst. Con- interests are in the areas of optimizations, numerical methods, computer aided
trol Lett., vol. 21, pp. 451–461, 1993. designs, and intelligent systems.
ZHANG et al.: BACKSTEPPING CONTROL OF LINEAR TIME-VARYING SYSTEMS 1925

Barış Fidan (S’02) received the B.S. degrees in Petros A. Ioannou (S’80–M’83–SM’89–F’94)
electrical engineering and mathematics from Middle received the B.Sc. degree (first class honors) from
East Technical University, Ankara, Turkey in 1996, University College, London, U.K., and the M.S.
and the M.S. degree in electrical engineering from and Ph.D. degrees from the University of Illinois,
Bilkent University, Ankara, Turkey in 1998. He is Urbana, in 1978, 1980, and 1982, respectively.
currently working on the Ph.D. degree in Electrical In 1982, he joined the Department of Electrical
Engineering-Systems at the University of Southern Engineering-Systems, University of Southern
California, Los Angeles. California, Los Angeles, California, where he is
His research interests include adaptive and currently a Professor and the Director of the Center
nonlinear control, switching and hybrid systems, of Advanced Transportation Technologies. His
robotics, high performance and hypersonic flight research interests are in the areas of adaptive control,
control, semiconductor manufacturing process control, and disk-drive servo neural networks, nonlinear systems, vehicle dynamics and control, intelligent
systems. transportation systems, and marine transportation. He was a Visiting Professor
at the University of Newcastle, NSW, Australia and the Australian National
University, Canberra, in fall 1988, the Technical University of Crete in summer
1992 and fall 2001, and served as the Dean of the School of Pure and Applied
Science at the University of Cyprus in 1995. He is the author/coauthor of five
books and over 150 research papers in the area of controls, neural networks,
nonlinear dynamical systems, and intelligent transportation systems.
Dr. Ioannou was a recipient of the Outstanding Transactions Paper Award in
1984, and the recipient of a 1985 Presidential Young Investigator Award.