On Contraction Analysis

for Nonlinear Systems

Analyzing stability differentially leads to a new perspective

on nonlinear dynamic systems

Winfried Lohmiller a and Jean-Jacques E. Slotine a, b

a Nonlinear
Systems Laboratory, Massachusetts Institute of Technology, Room
3-338, Cambridge, Massachusetts, 02139, USA
wslohmil@mit.edu, jjs@mit.edu, Fax (617) 258 5802, Phone (617) 253 0490
b Author to whom correspondence should be addressed

Abstract

This paper derives new results in nonlinear system analysis using methods in-
spired from fluid mechanics and differential geometry. Based on a differential analy-
sis of convergence, these results may be viewed as generalizing the classical Krasovskii
theorem, and, more loosely, linear eigenvalue analysis. A central feature is that con-
vergence and limit behavior are in a sense treated separately, leading to significant
conceptual simplifications. The approach is illustrated by controller and observer
designs for simple physical examples.

Key words: nonlinear dynamics, nonlinear control, observers, gain-scheduling,

contraction analysis

1 Introduction

Nonlinear system analysis has been very successfully applied to particular

classes of systems and problems, but it still lacks generality, as e.g. in the case
of feedback linearization, or explicitness, as e.g. in the case of Lyapunov theory
(Isidori, 1995; Marino and Tomei, 1995; Khalil, 1995; Vidyasagar, 1992; Slotine
and Li, 1991; Nijmeyer and Van der Schaft, 1990). In this paper, a body of
new results is derived using elementary tools from continuum mechanics and
differential geometry, leading to what we shall call contraction analysis.

Intuitively, contraction analysis is based on a slightly different view of what

stability is, inspired by fluid mechanics. Regardless of the exact technical form

NSL-961001, October, 1996. Revised, August, 1997. Final version, December 1997
in which it is defined, stability is generally viewed relative to some nominal
motion or equilibrium point. Contraction analysis is motivated by the elemen-
tary remark that talking about stability does not require to known what the
nominal motion is: intuitively, a system is stable in some region if initial con-
ditions or temporary disturbances are somehow “forgotten,” i.e., if the final
behavior of the system is independent of the initial conditions. All trajectories
then converge to the nominal motion. In turn, this shows that stability can
be analyzed differentially − do nearby trajectories converge to one another?
− rather than through finding some implicit motion integral as in Lyapunov
theory, or through some global state transformation as in feedback lineariza-
tion. Not surprisingly such differential analysis turns out to be significantly
simpler than its integral counterpart. To avoid any ambiguity, we shall call
“convergence” this form of stability.

We consider general deterministic systems of the form

ẋ = f(x, t) (1)

where f is an n×1 nonlinear vector function and x is the n×1 state vector. The
above equation may also represent the closed-loop dynamics of a controlled
system with state feedback u(x, t). In this paper, all quantities are assumed to
be real and smooth, by which is meant that any required derivative or partial
derivative exists and is continuous.

In section 2, we first recast elementary analysis tools from continuum me-

chanics in a general dynamic system context, leading to a simple sufficient
condition for system convergence. The result is then refined into a necessary
and sufficient convergence condition in section 3. The approach is illustrated
by applying it to controller and observer designs in section 4. Section 5 de-
scribes the method in the discrete-time case. Brief concluding remarks are
offered in section 6.

2 A basic convergence result

This section derives the basic convergence principle of this paper, which we
first introduced in (Lohmiller and Slotine, 1996, 1997). Considering the local
flow at a given point x leads to a convergence analysis between two neighboring
trajectories. If all neighboring trajectories converge to each other (contraction
behavior) global exponential convergence to a single trajectory can then be
concluded.

The plant equation (1) can be thought of as an n-dimensional fluid flow, where
ẋ is the n-dimensional “velocity” vector at the n-dimensional position x and

2
time t. Assuming as we do that f(x, t) is continuously differentiable, (1) yields
the exact differential relation
∂f
δ ẋ = (x, t) δx (2)
∂x

where δx is a virtual displacement − recall that a virtual displacement is

an infinitesimal displacement at fixed time. Note that virtual displacements,
pervasive in physics and in the calculus of variations, and extensively used in
this paper, are also well-defined mathematical objects. Formally, δx defines a
linear tangent differential form, and δxT δx the associated quadratic tangent
form (Arnold, 1978; Schwartz, 1993), both of which are differentiable with
respect to time.

Consider now two neighboring trajectories in the flow field ẋ = f(x, t), and
the virtual displacement δx between them (Figure 1). The squared distance
between these two trajectories can be defined as δxT δx , leading from (2) to
the rate of change

virtual displacement δ x

.
virtual velocity δx

two neighboring
trajectories

Fig. 1. Virtual dynamics of two neighboring trajectories

d ∂f
(δxT δx) = 2 δxT δ ẋ = 2 δxT δx
dt ∂x
Denoting by λmax (x, t) the largest eigenvalue of the symmetric part of the
∂f T
∂f
Jacobian ∂x (i.e., the largest eigenvalue of 12 ( ∂x
∂f
+ ∂x ) ), we thus have

d
(δxT δx) ≤ 2 λmax δxT δx
dt
and hence,
Rt
λmax (x,t)dt
kδxk ≤ kδxo k e o (3)

3
Assume now that λmax (x, t) is uniformly strictly negative (i.e., ∃ β > 0, ∀x, ∀t ≥
0, λmax (x, t) ≤ −β < 0. ). Then, from (3) any infinitesimal length kδxk con-
verges exponentially to zero. By path integration, this immediately implies
that the length of any finite path converges exponentially to zero. This moti-
vates the following definition.

Definition 1 Given the system equations ẋ = f(x, t), a region of the state
∂f
space is called a contraction region if the Jacobian ∂x is uniformly negative
definite in that region.
∂f
By ∂x
uniformly negative definite we mean that

∂f T
!
1 ∂f
∃ β > 0, ∀x, ∀t ≥ 0, + ≤ −β I < 0
2 ∂x ∂x

More generally, by convention all matrix inequalities will refer to the symmet-
ric parts of the square matrices involved − for instance, we shall write the
∂f
above as ∂x ≤ −β I < 0 . By a region we mean an open connected set. Ex-
∂f
tending the above definition, a semi-contraction region corresponds to ∂x being
∂f
negative semi-definite, and an indifferent region to ∂x being skew-symmetric.

Consider now a ball of constant radius centered about a given trajectory, such
that given this trajectory the ball remains within a contraction region at all
times (i.e., ∀t ≥ 0). Because any length within the ball decreases exponentially,
any trajectory starting in the ball remains in the ball (since by definition the
center of the ball is a particular system trajectory) and converges exponentially
to the given trajectory (Figure 2). Thus, as in stable linear time-invariant (LTI)
systems, the initial conditions are exponentially “forgotten.” This leads to the
following theorem:

Theorem 1 Given the system equations ẋ = f(x, t), any trajectory, which
starts in a ball of constant radius centered about a given trajectory and con-
tained at all times in a contraction region, remains in that ball and converges
exponentially to this trajectory.

Furthermore, global exponential convergence to the given trajectory is guaran-

teed if the whole state space is a contraction region.

This sufficient exponential convergence result may be viewed as a strength-

ened version of Krasovskii’s classical theorem on global asymptotic conver-
gence (Krasovskii, 1959, page 92; Hahn, 1967, page 270), an analogy we shall
generalize further in the next section. Note that its proof is very straight-
forward, even in the non-autonomous case, and even in the non-global case,
where it guarantees explicit regions of convergence. Also, note that the ball in
the above theorem may not be replaced by an arbitrary convex region − while
radial distances would still decrease, tangential velocities could let trajectories

4
 relative radial
other trajectory
velocity

relative
tangential
velocity

contraction
region

shrinking
given trajectory
length

Fig. 2. Convergence of two trajectories

escape the region.

Example 2.1: In the system

ẋ = −x + et

the Jacobian is uniformly negative definite and exponential convergence to a

single trajectory is guaranteed. This result is of course obvious from linear control
theory. 2
Example 2.2: Consider the system

ẋ = −t(x3 + x)

For t ≥ to > 0, the Jacobian is again uniformly negative definite and exponential
convergence to the unique equilibrium point x = 0 is guaranteed. 2

3 Generalization of the convergence analysis

Theorem 1 can be vastly extended simply by using a more general definition

of differential length. The result may be viewed as a generalization of linear
eigenvalue analysis and of the Lyapunov matrix equation. Furthermore, it
leads to a necessary and sufficient characterization of exponential convergence.

3.1 General definition of length

The line vector δx between two neighboring trajectories in Figure 1 can also
be expressed using the differential coordinate transformation

δz = Θδx (4)

5
where Θ(x, t) is a square matrix. This leads to a generalization of our earlier
definition of squared length

δzT δz = δxT M δx (5)

where M(x, t) = ΘT Θ represents a symmetric and continuously differentiable

metric − formally, equation (5) defines a Riemann space (Lovelock and Rund,
1989, page 243). Since (4) is in general not integrable, we cannot expect to
find explicit new coordinates z(x, t), but δz and δzT δz can always be defined,
which is all we need. We shall assume M to be uniformly positive definite, so
that exponential convergence of δz to 0 also implies exponential convergence
of δx to 0.

Distance between two points P1 and P2 with respect to the metric

R P2
M is defined
as the shortest path length (i.e., the smallest path integral P1 kδzk ) between
these two points. Accordingly, a ball of center c and radius R is defined as the
set of all points whose distance to c with respect to M is strictly less than R.

The two equivalent definitions of length in (5) lead to two formulations of

the rate of change of length: using local coordinates δz leads to a generaliza-
tion of linear eigenvalue analysis (section 3.2), while using the original system
coordinates x leads to a generalized Lyapunov equation (section 3.3).

3.2 Generalized eigenvalue analysis

Using (4), the time-derivative of δz = Θδx can be computed as

!
d ∂f
δz = Θ̇δx + Θδ ẋ = Θ̇ + Θ Θ−1 δz = F δz (6)
dt ∂x

Formally, the generalized Jacobian

!
∂f
F = Θ̇ + Θ Θ−1 (7)
∂x

represents the covariant derivative of f in δz coordinates (Lovelock and Rund,

1989, page 76). The rate of change of squared length can be written

d d
(δzT δz) = 2 δzT δz = 2 δzT F δz
dt dt
Similarly to the reasoning in Theorem 1, exponential convergence of δz (and
thus of δx) to 0 can be determined in regions with uniformly negative definite

6
F. This result may be regarded as an extension of eigenvalue analysis in LTI
systems, as the next example illustrates.

Example 3.1: Consider first the LTI system

ẋ = Ax

and the coordinate transformation z = Θx (where Θ is constant) into a Jordan

form
ż = ΘAΘ−1 z = Λ z
For instance, one may have
 
 λ1 ρ 0 0 0 
 
 0 λ 0 0 0 
 1 
 
Λ =  0 0 λ3 0 0
 

 
 
 0 0 0 λ
4real λ4im

 
 
0 0 0 −λ4im λ4real

where the λi ’s are the eigenvalues of the system, and ρ < 2λ1 is the normalization
factor of the Jordan form. The covariant derivative F = Λ is uniformly negative
definite if and only if the system is strictly stable, a result which obviously extends
to the general n-dimensional case.
Now, consider instead a gain-scheduled system (see (Lawrence and Rugh, 1995)
∂f
for a recent reference). Let A(x, t) = ∂x be the Jacobian of the corresponding
nonlinear, non-autonomous closed-loop system ẋ = f (x, t), and define at each
point a coordinate transformation Θ(x, t) as above. Uniform negative definiteness
of F = Λ + Θ̇Θ−1 (a condition on the “logarithmic” derivative of Θ) then
implies exponential convergence of this design.
This result also allows one to compute an explicit region of exponential conver-
gence for a controller design based on linearization about an equilibrium point,
by using the corresponding constant Θ. 2
Note that the above could not have been derived simply by using Krasovskii’s
generalized asymptotic global convergence theorem (Krasovskii, 1959, page
91; Hahn, 1967, page 270), even in the global asymptotic case and even using
a state transformation, since an explicit z does not exist in general.

3.3 Metric analysis

Equation (6) can equivalently be written in δx coordinates

!
d
T ∂f
Θ δz = Mδ ẋ + ΘT Θ̇δx = M + ΘT Θ̇ δx (8)
dt ∂x

7
using the covariant velocity differential Mδ ẋ + ΘT Θ̇δx (Lovelock and Rund,
1989). The rate of change of length is

∂f T
!
d  T  ∂f
δx M δx = δxT M + Ṁ + M δx (9)
dt ∂x ∂x

so that exponential convergence to a single trajectory can be concluded in

∂f T ∂f
regions of ( ∂x M + M ∂x + Ṁ ) ≤ −βM M (where βM is a strictly positive
constant). It is immediate to verify that these are of course exactly the re-
gions of uniformly negative definite F in (7). If we restrict the metric M to
be constant, this exponential convergence result represents a generalization
and strengthening of Krasovskii’s generalized asymptotic global convergence
theorem. It may also be regarded as an extension of the Lyapunov matrix
equation in LTI systems.

3.4 Generalized contraction analysis

The above leads to the following generalized definition, superseding Definition

1 (which corresponds to Θ = I and M = I).

Definition 2 Given the system equations ẋ = f(x, t), a region of the state
space is called a contraction region with respect to a uniformly positive definite
metric M(x, t) = ΘT Θ, if equivalently F in (7) is uniformly negative definite
∂f T ∂f
or ∂x M + M ∂x + Ṁ ≤ −βM M (with constant βM > 0) in that region.

∂f T ∂f
As earlier, regions where F or equivalently ∂x M + M ∂x + Ṁ are negative
semi-definite (skew-symmetric) are called semi-contracting (indifferent). The
generalized convergence result can be stated as:

Theorem 2 Given the system equations ẋ = f(x, t), any trajectory, which
starts in a ball of constant radius with respect to the metric M(x,t), centered
at a given trajectory and contained at all times in a contraction region with
respect to M(x,t), remains in that ball and converges exponentially to this
trajectory.

Furthermore global exponential convergence to the given trajectory is guaran-

teed if the whole state space is a contraction region with respect to the metric
M(x,t).

In the remainder of this paper we always assume this generalized form when
we discuss contraction behavior.

8
3.5 A converse theorem

Conversely, consider now an exponentially convergent system, which implies

that ∃β > 0, ∃k ≥ 1, such that along any system trajectory x(t) and ∀t ≥ 0,

δxT δx ≤ k δxTo δxo e−βt (10)

Defining a metric M(x(t), t) by the ordinary differential equation (Lyapunov

equation)

∂f ∂f T
Ṁ = −βM − M − M M(t = 0) = kI (11)
∂x ∂x

and using (9), we can write (10) as

δxT δx ≤ δxT M δx = k δxTo δxo e−βt (12)

Since this holds for any δx, the above shows that M is uniformly positive
definite, M ≥ I. Thus, any exponentially convergent system is contracting
with respect to a suitable metric.

Note from the linearity of (11) that M is always bounded for bounded t.
Furthermore, while M may become unbounded as t → +∞, this does not
create a technical difficulty, since the boundedness of δxT Mδx (from (12))
still implies that δx tends to zero exponentially and also indicates that the
metric could be renormalized by a further coordinate transformation.

Thus, Theorem 2 actually corresponds to a necessary and sufficient condi-

tion for exponential convergence of a system. In this sense it generalizes and
simplifies a number of previous results in dynamic systems theory.

For instance, note that chaos theory (Guckenheimer and Holmes, 1983; Stro-
gatz, 1994) leads at best to sufficient stability results. Lyapunov exponents,
which are computed as numerical integrals of the eigenvalues of the symmetric
∂f
part of the Jacobian ∂x , depend on the chosen coordinates x and hence do
not represent intrinsic properties.

3.6 A Note On Krasovskii’s Theorem

It should be clear to the reader familiar with the many versions of Krasovskii’s
theorem that by now we have ventured quite far from this classical result.
Indeed, Krasovskii’s theorem provides a sufficient, asymptotic convergence

9
result, corresponding to a constant metric M. Also, it does not exploit the
possibility of a pure differential coordinate change as in (4). It is also interest-
ing to notice that the type of proof used here is very significantly simpler than
that used, say, for the global non-autonomous version of Krasovskii’s theorem.
This in turn allows many further extensions, as the next sections demonstrate.

3.7 Linear properties of generalized contraction analysis

Introductions to nonlinear control generally start with the warning that the
behavior of general nonlinear non-autonomous systems is fundamentally dif-
ferent from that of linear systems. While this is unquestionably the case,
contraction analysis extends a number of desirable properties of linear system
analysis to general nonlinear non-autonomous systems.

(i) Solutions in δz(t) can be superimposed, since dtd δz = F(x, t)δz around a
specific trajectory x(t) represents a linear time-varying (LTV) system in
local δz coordinates. Note that the system needs not be contracting for
this result to hold.
(ii) Using this point of view, Theorem 2 can also be applied to other norms,
such as kδzk∞ = maxi |δzi | and kδzk1 = i |δzi |, with associated balls
P

defined accordingly. Using the same reasoning as in standard matrix mea-

sure results (Vidyasagar, 1992, page 71), the corresponding convergence
results are

d X d X
kδzk∞ ≤ max(Fii + |Fij |) kδzk∞ kδzk1 ≤ max(Fjj + |Fij |) kδzk1
dt i
j6=i dt j
i6=j

(iii) Global contraction precludes finite escape, under the very mild assump-
tion
∃ x∗ , ∃ c ≥ 0, ∀t ≥ 0 , k Θf(x∗ , t) k ≤ c
Indeed, no trajectory can diverge faster from x∗ than bounded k Θf(x∗ , t) k
and thus cannot become unbounded in finite time. The result can be ex-
tended to the case where x∗ may itself depend on time, as long as it
remains in an a priori bounded region.
(iv) A convex contraction region contains at most one equilibrium point, since
any length between two trajectories is shrinking exponentially in that
region.
(v) This further implies that, in a globally contracting autonomous system,
all trajectories converge exponentially to a unique equilibrium point. In-
deed, using V (x) = f(x)T M(x, t) f(x) as a Lyapunov-like function (an
extension of the standard proof of Krasovskii’s Theorem for autonomous

10
 systems) yields

∂f T
!
T ∂f
V̇ = f(x) Ṁ + M + M f(x) ≤ −βM V
∂x ∂x

which shows that ẋ = f(x) tends to 0 exponentially, and thus that x

tends towards a finite equilibrium point.
(vi) The output of any time-invariant contracting system driven by a periodic
input tends exponentially to a periodic signal with the same period.
Indeed, consider a time-invariant nonlinear system driven by a periodic
input ω(t) of period T > 0,

ẋ = f(x, ω(t)) (13)

Let xo (t) be the system trajectory corresponding to the initial condition

xo (0) = xI , and let xT (t) be the system trajectory corresponding to the
system being initialized instead at xT (T ) = xI . Since f is time-invariant
and ω(t) has period T , xT (t) is simply as shifted version of xo (t),

∀t ≥ T, xT (t) = xo (t − T ) (14)

Furthermore, if we now assume that the dynamics (13) is contracting,

then initial conditions are exponentially forgotten, and thus xT (t) tends
to xo (t) exponentially. Therefore, from (14), xo (t − T ) tends towards
xo (t) exponentially. By recursion, this implies that ∀t, 0 ≤ t < T , the
sequence xo (t + nT ) is a Cauchy sequence, and therefore the limiting
function limn→+∞ xo (t + nT ) exists, which completes the proof.
(vii) Consider the distance R = PP12 kδzk between two trajectories P1 and P2 ,
R

contained at all times in a contraction region characterized by maximal

eigenvalues λmax (x, t) ≤ −β < 0 of F. The relative velocity between
these trajectories verifies

Ṙ + |λmax | R ≤ 0

Assume now, instead, that P1 represents a desired system trajectory and

P2 the actual system trajectory in a disturbed flow field ẋ = f(x, t) +
d(x, t). Then

Ṙ + |λmax | R ≤ kΘdk (15)

For bounded disturbance kΘdk any trajectory remains in a boundary ball

of (15) around the desired trajectory. Since initial conditions R(t = 0) are
exponentially forgotten, we can also state that any trajectory converges
exponentially to a ball of radius R in (15) with arbitrary initial condition
R(t = 0).
(viii) The above can be used to describe a contracting dynamics at multiple
resolutions using multiscale approximation of the dynamics with bounded

11
 basis functions, as e.g. in wavelet analysis. The radius R with respect to
the metric M of the boundary ball to which all trajectories converge
exponentially becomes smaller as resolution is increased, making precise
the usual “coarse grain” to “fine grain” terminology.

3.8 Combinations of contracting systems

Combinations of contracting systems satisfy simple closure properties, a subset

of which are reminiscent of the passivity formalism (Popov, 1973).

3.8.1 Parallel combination

Consider two systems of the same dimension

ẋ1 = f1 (x1 , t)
ẋ2 = f2 (x2 , t)

with virtual dynamics

δ ż1 = F1 δz
δ ż2 = F2 δz

and connect them in a parallel combination. If both systems are contracting

in the the same metric, so is any uniformly positive superposition

α1 (t) δ ż1 + α2 (t) δ ż2 where ∃ α > 0, ∀t ≥ 0, αi (t) ≥ α (16)

Example 3.2: In the biological motor control community, there has been con-
siderable interest recently in analyzing feedback controllers for biological motor
systems as combinations of simpler elements, or motion primitives. For instance
(Bizzi, et al., 1993; Mussa-Ivaldi, et al., 1994) have experimentally studied the hy-
pothesis that stimulating a small number of areas in a frog’s spinal cord generates
corresponding force fields at the frog’s ankle, and furthermore that these force
fields simply add when different areas are stimulated at the same time. Inter-
preting each of these force fields as a contracting flow in joint-space is consistent
with experimental data, and likely candidates for the αi (t) in (16) would then
be sigmoids and pulses − so-called “tonic” and “phasic” signals (Mussa-Ivaldi,
1997; Berthoz, 1993). A simplified architecture may thus consist of weighted con-
tracting fields generated at the spinal chord level through high-bandwidth few-
synapse feedback connections, combined with the natural viscoelastic properties
of the muscles, and added open-loop terms generated by the brain, with some
time advance because of the significant nerve transmission delays. 2

12
3.8.2 Feedback Combination

Similarly, connect instead two systems of possibly different dimensions

ẋ1 = f1 (x1 , x2 , t)
ẋ2 = f2 (x1 , x2 , t)

in the feedback combination

    
d  δz1   F1 G   δz1 
=
dt δz2
  
− GT F2 δz2

The augmented system is contracting if and only if the separated plants are
contracting.

3.8.3 Hierarchical Combination

Consider a smooth virtual dynamics of the form

    
d  δz1   F11 0   δz1 
=
dt δz2
  
F21 F22 δz2

and assume that F21 is bounded. The first equation does not depend on the
second, so that exponential convergence of δz1 to zero can be concluded for
uniformly negative definite F11 . In turn, F21 δz1 represents an exponentially
decaying disturbance in the second equation. Similarly to remark (vii) in sec-
tion 3.7, a uniformly negative definite F22 implies exponential convergence of
δz2 to an exponentially decaying ball. Thus, the whole system globally expo-
nentially converges to a single trajectory.

By recursion, the result can be extended to systems similarly partitioned in

more than two equations. It may be viewed as providing a general common
framework for sliding control concepts, singular perturbations, and triangular
systems, where such hierarchical analysis can be found (see also (Simon, 1981)
in a broader context).

Consider again the system above, but now with disturbance Θ1 d1 added to
the δz1 dynamics and Θ2 d2 added to the δz2 dynamics. This means that the
relative velocities between a desired trajectory P1 and a system trajectory P2
verify

13
 ZP2 ZP2
d
kδz1 k + |λmax1 | kδz1 k ≤ kΘ1 d1 k
dt
P1 P1
ZP2 ZP2 ZP2
d
kδz2 k + |λmax2 | kδz2 k ≤ kΘ2 d2 k + F21 δz1
dt
P1 P1 P1

Bounded disturbances Θ1 d1 and Θ1 d2 thus imply exponential convergence to

a ball around the desired trajectory.

Example 3.3: Chain reactions are classical examples of hierarchical dynamics.

Consider for instance a standard polymerization process in an open stirred tank
(adapted from (Adebekun and Schork, 1989)), of the form

q Ed
I˙ = (If − I) − kd e− RT I
V
q Ep
Ṁ = (Mf − M ) − 2kp e− RT M 2 I
V
q Ep
Ṗ = (Pf − P ) + kp e− RT M 2 I
V !
q −∆H hAc
Ṫ = (Tf − T ) + kp M 2 I − (T − Tc )
V ρcp V ρcp

with I, M , and P being the initiator, monomer, and polymer concentrations,

T the temperature, Tc the coolant temperature, q(t) > 0 the feed flow rate, V
the reactor volume, kp and kd positive reaction constants, and the subscript f
corresponding to feed values. Consider now a reduced-order identity observer on
I, M , and P whose reaction rates simply reproduce the model using the measured
temperature T (t),

˙ q   Ed
Iˆ = If − Iˆ − kd e− RT Iˆ
V
˙ q   Ep
M̂ = Mf − M̂ − 2kp e− RT M̂ 2 Iˆ
V
˙ q   Ep
P̂ = Pf − P̂ + kp e− RT M̂ 2 Iˆ
V
Since this observer represents a hierarchical system, the uniform negative defi-
˙ ˙ ˙
∂ Iˆ ∂ M̂ ∂ P̂
niteness of ˆ ,
∂ I ∂ M̂
, and ∂ P̂
implies that it converges exponentially. 2

Example 3.4: Contraction analysis may be used as a more precise alternative

to zero-dynamics analysis (Isidori, 1995). Consider an n-dimensional system ẋ =
f (x, u, t) with measurement y = h(x, t). Assume that repeated differentiation of
the measurement leads to y(p) = g(x, u, t) with p ≤ n, where we can choose a
control input that leads to a contracting linear design in y, ..., y(p) .
Contracting behavior of the (n − p)-dimensional remaining states z and thus of
the whole system can then be concluded according to section 3.8.3 for uniformly
negative definite ∂∂zż . 2

14
Note that the properties above can be arbitrarily combined.

Example 3.5: Using the hierarchical property, the open-loop signal generated
by the brain in the biological motor control model of Example 3.2 may itself be
the output of a contracting dynamics. So can be the αi (t), since the corresponding
primitives are bounded. In principle, the contraction property would also enable
this term to be learned (see also (Droulez, et al., 1983; Flash, 1995; Mussa-Ivaldi,
1997)) by making the system’s behavior consistent in the presence of disturbances
or variations in initial conditions.
In this context, the remark (vi) on periodic inputs in section 3.7 may also
be relevant to the periodic phenomena pervasive in physiology. These include,
for instance, the rhythmic motor behaviors used in locomotion and driven by
central pattern generators, as in walking, swimming, or flying (Kandel, et al.,
1991; Dowling, 1992), as well as automatic mechanisms such as breathing and
heart cycles. 2

3.9 Additional Remarks

In addition to the simple properties above, we can make a few more technical
remarks and extensions on Theorem 2.

• Theorem 2 may be viewed as a more precise version of Gauss theorem in

fluid mechanics
d d
δV = div( δz) δV
dt dt
which shows that any volume element δV shrinks exponentially to zero for
uniformly negative definite div( dtd δz), implying convergence to an (n − 1)-
dimensional manifold rather than to a single trajectory. Indeed, div( dtd δz)
is just the trace of F.
• Theorem 2 still holds if the radius R of the ball is time-varying, as long
as trajectories starting in the ball can be guaranteed to remain in the ball.
Given (7) and F ≤ −β I < 0 this is the case if ∀t ≥ 0, Ṙ + β ≥ 0 .
• Assume that the metric M is only positive semi-definite, with some principal
directions pi corresponding to uniformly positive definite eigenvalues of M.
Uniformly negative definiteness of F then implies exponential convergence
to zero of the components of δx on the linear subspace spanned by the pi .
• Assume that F is not uniformly negative Rt 1
definite, but rather that ∃ κ >
1
0, ∃ to > 0, ∀t ≥ to , F ≤ − t κI. Since to τ κ dτ tends to +∞ as t → +∞, any
infinitesimal length converges asymptotically (although not necessarily ex-
ponentially) to zero, and thus asymptotic convergence to a single trajectory
can be concluded.
• Any regular Θ(x), defined in a compact set in x, yields a uniformly positive
definite metric M = ΘT Θ in this compact set.
• Since trajectories can rotate or oscillate around each other, overshoots may
occur in the elongated principal directions of the metric M(x, t).

15
• In the case that an explicit z(x, t) exists, we can alternatively compute the
∂z
virtual velocity from ż = ∂x f + ∂z
∂t
, since then

∂2z ∂z2
!
∂z ∂z ∂z ∂f d
δ ż = δ f+ = 2
δx f + δx + δx = δz
∂x ∂t ∂x ∂x ∂x ∂t∂x dt

• Contraction analysis can also be applied to differential coordinates δz whose

dimension is not the same as that of x. Of course, lower-dimensional coor-
dinates can only lead to positive semi-definite metrics M.
• Contraction regions of an arbitrary autonomous dynamic system ẋ = f(x)
may be computed by solving (in general numerically) the partial differential
equation in space
∂Θ ∂f
f +Θ = −Θ
∂x ∂x
with appropriate boundary conditions, which imposes F = − I . Consider
for instance the plant
ẋ = − sin x
dθ
and set F θ = − dx sin x − θ cos x = −θ . Integrating with θ(0) = 1 (to be
tan x
consistent with linearization) leads to θ = sin x2 . The metric is singular at
x = π + 2nπ with n ∈ Z, so that contraction regions are ]2nπ − π, 2nπ + π[.
• Theorem 2 can also be used to show exponential divergence of two neighbor-
ing trajectories. Indeed, if the minimal eigenvalue λmin (x, t) of the symmet-
ric part of the F is strictly positive, then equation (3) implies exponential
divergence of two neighboring trajectories. This may be used to impose
constraints on the flow (or, by state-augmentation, the input).

4 Applications of Contraction Analysis

This section illustrates the discussion with some immediate applications of

contraction analysis to specific control and estimation problems.

4.1 PD observers

Observer design using contraction analysis can be simplified by prior coor-

dinate transformations similar to those used in linear reduced-order observer
design (Luenberger, 1979). Consider the system

ẋ = f(x, t)
y = h(x, t)

16
where x is the state vector and y the measurement vector. Define a state
observer with

ˆ
x̄˙ = f(x̂, t) − KP (ŷ − y) − KD ẏ (17)
x̂ = x̄ + KD y
ˆ=
where ŷ = h(x̂, t) and ẏ ∂ ŷ
f(x̂, t) + ∂h . By differentiation, this leads to the
∂ x̂ ∂t
observer dynamics

x̂˙ = f(x̂, t) − KP (ŷ − y) − KD ẏ

 
ˆ − ẏ

Thus the dynamics of x̂ contains ẏ, although the actual computation is done
using equation (17) and hence ẏ is not explicitly used.

Example 4.1: Consider a simple model of underwater vehicle motion, including

thruster dynamics    
 q̈1   τ (t) − 3q̇1 |q̇1 |
 =


q̈2 −10 q̇2 |q̇2 | + q̇1 |q̇1 |
where y1 = q1 is the measured propeller angle, y2 = q2 the measured vehicle
position, q̇1 |q̇1 | the propeller thrust, q̇2 |q̇2 | the vehicle drag, and τ (t) the torque
input to the propeller (Figure 3). The system dynamics is heavily damped for
large |q̇1 | and |q̇2 |. However, this natural damping is ineffective at low velocities.
Letting ω = q̇1 , v = q̇2 , this suggests using a coordinate error feedback in the
reduced-order observer

q 1, τ (t)

q2

Fig. 3. Underwater vehicle

   
 ω̄˙   τ (t) − 3ω̂ |ω̂| − kd1 ω̂ 
 = 
v̄˙ −10 v̂ |v̂| + ω̂ |ω̂| − kd2 v̂
   
 ω̂   ω̄ + kd1 q1 
 = 
v̂ v̄ + kd2 q2

where kd1 and kd2 are strictly positive constants. This leads to the hierarchical
dynamics    
˙
ω̂ τ (t) − 3ω̂ |ω̂| − kd1 (ω̂ − ω)
 =
   

˙v̂ −10 v̂ |v̂| + ω̂ |ω̂| − kd2 (v̂ − v)

17
 ˙ ˙
The uniform negative definiteness of ∂∂ ω̂ω̂ = (−3 |ω̂| − kd1 ) and ∂∂ v̂v̂ = (−10 |v̂| − kd2 )
(which is implied by our choice of strictly positive constants kdi ) guarantees expo-
nential convergence to the actual system trajectory, which is indeed a particular
solution.
System responses to the input
 
 5 for 0 ≤ t < 1 
τ = 
−10 for 1 ≤ t < 2

with initial conditions ω(0) = 0, ω̂(0) = 4 or −4, v(0) = 5, v̂(0) = −10 or 20 and
feedback gains kd1 = kd2 = 5 are illustrated in Figure 4. The solid line represents
the actual plant and the dashed lines the observer estimates. 2
4 20

3
15
PROPELLER VELOCITY

2
VEHICLE VELOCITY 10
1

0 5

−1
0
−2
−5
−3

−4 −10
0 0.5 1 1.5 2 0 0.5 1 1.5 2
TIME TIME

Fig. 4. Underwater vehicle observer

4.2 Constrained Systems

Many physical systems, such as mechanical systems with kinematic constraints

or chemical systems in partial equilibrium, are described by an original n-
dimensional dynamics of the form

ż = f(z, t)

constrained to an explicit m-dimensional submanifold (m ≤ n)

z = z(x, t)

The constrained dynamic equations are then of the form

ż = f(z, t) + n
∂z T
where n represents a superimposed flow normal to the manifold, ∂x n= 0
− the components of n are Lagrange parameters. In a mechanical system, z
represents unconstrained positions and velocities, x generalized coordinates

18
and associated velocities, and n constraint forces. Multiplying from the left
∂z T
with ∂x results in

∂z T ∂z T ∂z T
!
∂z ∂z
ż = ẋ + = f(z) (18)
∂x ∂x ∂x ∂t ∂x

∂z T ∂z
so that a uniformly positive definite metric M = ∂x ∂x
allows one to compute,
∂z T
 
with ẋ = M−1 ∂x f(z, t) − ∂z
∂t

∂z T
!
∂z ∂z ∂z
n = ż − f = + M−1 f− −f
∂t ∂x ∂x ∂t

The variation of (18) is

T
∂z T ∂2z ∂z T ∂f
δ ż = − 2 n δx + δz
∂x ∂x ∂x ∂z
so that
T
1d T T ∂z ∂f ∂z ∂2z
(δx Mδx) = δx δx − δxT 2 n δx (19)
2 dt ∂x ∂z ∂x ∂x

Consider now a specific trajectory zd (t) of the unconstrained flow field which
naturally remains on the manifold z(x, t). Then the normal flow n around this
trajectory vanishes, so that locally the contraction behavior is determined by
the projection of the original Jacobian

∂z T ∂f ∂z
Jd = (20)
∂x ∂z ∂x

Contraction of the original unconstrained flow thus implies local contraction

of the constrained flow, and the contraction region around the trajectory can
be computed with (19).

This result can be used to study the contraction behavior of mechanical sys-
tems with linear external forces, such as PD terms or gravity − Newton’s law
in the original unconstrained state space is then linear, and z = z(x, t) are
kinematic constraints. Exponential convergence around one trajectory z(t) at
which the constraint forces vanish can then be concluded in the region where
∂f
the projection (20) of the original constant Jacobian ∂z is uniformly negative
definite. Exponential convergence of a controller or observer (see also (Marino
and Tomei, 1995; Berghuis and Nijmeyer, 1993)) can thus be achieved by
stabilizing the unconstrained dynamics with a PD part, and adding an open-
loop control input to guarantee that a desired trajectory consistent with the
kinematic constraints is indeed contained in the unconstrained flow field.

19
4.3 Linear Time-Varying Systems

As another illustration, this section shows that contraction analysis can be

used systematically to control and observe linear time-varying (LTV) systems.

4.3.1 LTV controllers

Consider a general smooth linear time-varying system

ẋ = A(t)x + b(t)u

with control input u = K(t)x+ud (t). We focus on choosing the gain K(t) so as
to achieve contraction behavior; whereas the open-loop term ud (t) then guar-
antees that the desired trajectory, if feasible, is indeed contained in the flow
field (this guarantee and a similar computation is required of any controller
design).

We need to find a smooth coordinate transformation δz = Θ(t)δx that leads

to the generalized Jacobian F
 
 0 1 0 ··· 0 
 
0 0 1 ··· 0
 
 
.. .. .. . . ..
   
−1
 
F = Θ̇ + Θ (A + bK) Θ = 
 . . . . . 
 (21)
 
0 0 0 0 1
 
 
 
 
−ao −a1 −a2 · · · −an−1

with the desired (Hurwitz) constant characteristic coefficients ai . The above

equation can be rewritten in terms of the row vectors θj (j = 1, · · · , n) of Θ
as

θ̇j + θj (A + bK) = θj+1 j = 1, · · · , n − 1 (22)

 
 θ1 
 . 
 
θ̇n + θn (A + bK) = −a  ..  (23)
 
 
θn

In order to make the coordinate transformation Θ independent of the control

input, let us impose recursively, ∀t ≥ 0, the following constraints on the θj

0 = θ1 b = θ1 Lo b

20
   d
0 = θ2 b = θ̇1 + θ1 A b − (θ1 b) = θ1 L1 b
dt
  d
0 = θ3 b = θ̇2 + θ2 A b − (θ2 b) = θ2 L1 b (24)
dt
  d  
= θ̇1 + θ1 A L1 b − θ1 L1 b = θ1 L2 b
dt
..
.
D = θn b = θ1 Ln−1 b

where the Lj b are generalized Lie derivatives (Lovelock and Rund, 1989)

L0 b = b
d j
Lj+1 b = A Lj b − Lb j = 0, ..., n − 1 (25)
dt
Choosing D = det |Lo b ... Ln−1 b| the above can always be solved algebraically
for a smooth θ1 , and from (22) the remaining smooth θj can then be computed
algebraically using the recursion

θj+1 = θ̇j + θj A j = 1, · · · , n − 1

This leads to a smooth bounded metric, and the feedback gain K(t) can then
be computed from (23)
 
 θ1 
 .
 
D K(t) = −a  ..  − θ̇n − θn A

 
 
θn

Exponential convergence of δx is then guaranteed for uniformly positive def-

inite metric M = ΘT Θ. The uniform positive definiteness of M hence repre-
sents a sufficient “contractibility” condition.

Note that the terms θ̇i and dtd Li b distinguish this derivation from the usual
pole-placement in an LTI system, as well as from related gain-scheduling tech-
niques (Wu, Packard, and Bals, 1995; Mracek, Cloutier, and D’Souza, 1996).
The method guarantees global exponential convergence, and can be extended
straightforwardly to multi-input systems.

4.3.2 LTV observers

Similarly, consider again the plant above, but assume that only the measure-
ment y = C(t)x + d(t) is available. Define the observer

x̂˙ = A(t)x̂ + E(t) (y − c(t)x̂ + d(t)) + b(t)u

21
Since by definition the actual state is contained in the flow field, no “open-
loop” term is needed, but we need to find a smooth coordinate transformation
δx̂ = Σ(t)δẑ that leads to the generalized Jacobian F
 
0 0 · · · 0 −ao 
 
1 0 · · · 0 −a1 
 
   
F = Σ−1 −Σ̇ + (A − Ec) Σ =
 
0
 1 · · · 0 −a2 
 (26)
 .. .. . . ..
 
. . .0 .


 
 
0 0 · · · 1 −an−1

with the desired (Hurwitz) constant characteristic coefficients ai . The above

equation can be rewritten in term of the column vectors σj of Σ as

−σ̇j + (A − Ec) σj = σj+1 j = 1, · · · , n − 1 (27)

 
−σ̇n + (A − Ec) σn = − σ1 · · · σn a (28)

Proceeding as before by imposing the constraints

Lj c σ1 = 0 j = 0, ..., n − 2
 
n−1
L c σ1 = det Lo c ... L n−1
c
 

where the Lj c are now defined as

Lo c = c
d j
Lj+1 c = Lj c A + Lc j = 0, ..., n − 1 (29)
dt
allows one to solve algebraically for a smooth σ1 , and then compute the re-
maining smooth σj recursively

σj+1 = −σ̇j + Aσj j = 1, · · · , n − 1

leading to a uniformly positive definite metric. The feedback gain E(t) can
then be computed from (28)
 
ED = σ1 · · · σn a − σ̇n + Aσn

Exponential convergence of δx̂ to zero and thus of x̂ to x is then guaranteed

for a bounded metric M = Σ−T Σ . The boundedness of M hence represents
a sufficient observability condition.

22
Again, the terms σ̇i and dtd Li c distinguish this derivation from pole-placement
in LTI systems, or from extended Kalman filter-like designs (Bar-Shalom and
Fortmann, 1988). The method guarantees global exponential convergence, and
can be extended straightforwardly to multi-output systems.

4.3.3 Separation principle

These LTV designs satisfy a separation principle. Indeed, let us combine the
above controller and observer (perhaps with different coefficient vectors a)

u = K(t)x̂ + ud (t)

Subtracting the plant dynamics from the observer dynamics leads with x̃ =
x̂ − x to

x̃˙ = (A(t) − E(t)C(t)) x̃

so that the Jacobian of the error-dynamics of the observer is unchanged. Since

the observer error-dynamics and the controller dynamics represent a hierar-
chical system, they can be designed separately as long as the control gain K(t)
is bounded.

Remark (vii) in section 3.7 can be used to assess the robustness of these designs
to additive modeling uncertainties.

4.3.4 Nonlinear controllers and observers

Consider now the nonlinear system

∂f ∂f ∂h
δ ẋ = δx + δu δy = δx
∂x ∂u ∂x
Let
∂f ∂f ∂h
A(t) = (xd (t), t) B(t) = (xd (t), t) c(t) = (xd (t), t)
∂x ∂u ∂x

where xd (t) is the desired trajectory. Applying the controller and observer de-
signs above to the LTV system defined by A(t), b(t), and c(t) then guarantees
contraction behavior in regions of uniformly negative definite
!!
∂f ∂f
Fcontrol = Θ̇(xd (t), t) + Θ(xd (t), t) (x, t) + (x, t)K(xd (t), t) Θ(xd (t), t)−1
∂x ∂u

23
and
! !
∂f ∂h
Fobs = Σ(xd (t)−1 −Σ̇(xd (t), t) + (x, t) − E(xd (t), t) (x, t) Σ(xd (t), t)
∂x ∂x

similarly to Example 3.1. Thus exponential convergence is guaranteed explic-

itly in a given finite region around the desired trajectory.

5 The discrete-time case

Theorem 1 can be extended to discrete-time systems

xi+1 = fi (xi , i)

The associated virtual dynamics is

∂fi
δxi+1 = δxi
∂xi
so that the virtual length dynamics is

∂fi T ∂fi
δxTi+1 δxi+1 = δxTi δxi
∂xi ∂xi
Thus, exponential convergence to a single trajectory is guaranteed for

∂fi T ∂fi
− I ≤ −βI < 0
∂xi ∂xi
This may be viewed as extending to non-autonomous systems the standard
iterated map results based on the contraction mapping theorem. The conver-
gence condition is equivalent to requiring that the largest singular value of
∂fi
the Jacobian ∂x i
remain smaller than 1 uniformly. A discrete-time version of
Theorem 2 can be derived similarly, using the generalized virtual displacement

δzi = Θi (xi , i)δxi

leading to

∂fi T T ∂fi
δzTi+1 δzi+1 = δxTi Θi+1 Θi+1 δxi = δzTi FTi Fi δzi
∂xi ∂xi
with the discrete generalized Jacobian

∂fi −1
Fi = Θi+1 Θ (30)
∂xi i

24
Note the similarity and difference with the Jacobian of an LTI system. The
above leads to the following generalized definition for discrete-time systems.

Definition 3 Given the discrete-time system equations xi+1 = fi (xi , i), a re-
gion of the state space is called a contraction region with respect to a uniformly
positive definite metric Mi (xi , i) = ΘTi Θi , if in that region

∃β > 0, FTi Fi − I ≤ −βI < 0

∂fi
where Fi = Θi+1 ∂x i
Θ−1
i .

Theorem 2 can then be immediately extended as

Theorem 3 Given the system equations xi+1 = fi (xi , i), any trajectory, which
starts in a ball of constant radius with respect to the metric Mi , centered at a
given trajectory and contained at all times in a generalized contraction region,
remains in that ball and converges exponentially to this trajectory.

Furthermore global exponential convergence to the given trajectory is guaran-

teed if the whole state space is a contraction region with respect to the metric
Mi .

Most of our earlier continuous-time results have immediate discrete-time ver-

sions, as detailed in (Lohmiller and Slotine, 1997d).

Example 5.1: As in the continuous-time case, consider first the discrete-time

LTI system

xi+1 = Axi

and the coordinate transformation zi = Θxi (where Θ is constant) into a Jordan

form

zi+1 = ΘAΘ−1 zi = Λ zi

It is straightforward to show that ΛT Λ − I is uniformly negative definite if and

only if the system is strictly stable.
Now, consider instead a discrete-time gain-scheduled system. Let Ai (xi , i) =
∂fi
∂xi be the Jacobian of the corresponding nonlinear, non-autonomous closed-loop
system xi+1 = fi (xi , i), and define at each point a coordinate transformation
Θi as above. Uniform negative definiteness of ΛT Λ − I then implies exponential
convergence of this design.
This result also allows one to compute an explicit region of exponential conver-
gence for a controller design based on linearization about an equilibrium point,
by using the corresponding constant Θi . 2

25
6 Concluding remarks

By using a differential approach, convergence analysis and limit behavior are

in a sense treated separately. Guaranteeing contraction means that after ex-
ponential transients the system’s behavior will be independent of the initial
conditions. In an observer context, one then needs only verify that the observer
equations contain the actual plant state as a particular solution to automati-
cally guarantee convergence to that state. In a control context, once contrac-
tion is guaranteed through feedback, specifying the final behavior reduces to
the problem of shaping one particular solution, i.e. specifying an adequate
open-loop control input to be added to the feedback terms, a necessary step
of any control method.

Current research includes systematically guaranteeing global exponential con-

vergence for general nonlinear systems, stable adaptation to unknown param-
eters, and further applications to mechanical and chemical systems.

Acknowledgments: This paper benefited from discussions with Ch. Bernard

and G. Niemeyer, and from reviewers’ comments and suggestions.

REFERENCES

Adebekun, D.K., and Schork, F.J. (1989). Continuous solution polymerization

reactor control. 1. Nonlinear reference control of methyl methacrylate polymeriza-
tion, Industrial Engineering and Chemical Research, 28:1308 - 1324.
Arnold, V.I. (1978). Mathematical Methods of Classical Mechanics, Springer
Verlag.
Bar-Shalom Y., and Fortmann, T. (1988). Tracking and Data Association,
Academic Press.
Berghuis, H., and Nijmeyer, H. (1993). A passivity-based approach to controller-
observer design for robots, I.E.E.E. Transactions on Robotics and Automation, 9(6).
Berthoz, A., Ed. (1993). Multisensory Control of Movement, Oxford University
Press.
Bizzi E., Giszter S.F., Loeb E., Mussa-Ivaldi F.A., Saltiel P (1995). Mod-
ular organization of motor behavior in the frogs’s spinal cord. Trends in Neuro-
sciences. Review 18:442-446
Dowling, J.E. (1992). Neurons and Networks, Belknap.
Droulez, J., et al. (1983). Motor Control, 7th International Symposium of the
International Society of Posturography, Houston, Karger, 1985.
Flash, T. (1995). Trajectory Learning and Control Models, I.F.A.C. Man-Machine
Systems Symposium, Cambridge, MA.
Guckenheimer, J., and Holmes, P. (1983). Nonlinear Oscillations, Dynamical
Systems, and Bifurcations of Vector Fields, Springer Verlag.
Hahn, W. (1967). Stability of motion, Springer Verlag.
Isidori, A. (1995). Nonlinear Control Systems, 3rd Ed., Springer Verlag.
Kandel, E.R., Schwartz, J.H., and Jessel, T.M. (1991). Principles of Neural

26
Science, Appleton and Lange.
Khalil, H. (1995). Nonlinear Systems, 2nd Ed., Prentice-Hall.
Krasovskii, N.N. (1959). Problems of the Theory of Stability of Motion, Mir,
Moskow. English translation by Stanford University Press, 1963.
Lawrence, D.A., and Rugh, W.J. (1995). Gain-scheduling dynamic linear con-
trollers for a nonlinear plant, Automatica, 31(3).
Lohmiller, W., and Slotine, J.J.E. (1996a). Metric Observers for Nonlin-
ear Systems, I.E.E.E. International Conference on Control Applications, Dearborn,
Michigan.
Lohmiller, W., and Slotine, J.J.E. (1996b). Applications of Metric Observers
for Nonlinear Systems, I.E.E.E. International Conference on Control Applications,
Dearborn, Michigan.
Lohmiller, W., and Slotine, J.J.E. (1996c). On Metric Controllers for Nonlin-
ear Systems, I.E.E.E. Conference on Decision and Control, Kobe, Japan.
Lohmiller, W., and Slotine, J.J.E. (1997a). Practical Observers for Hamilto-
nian Systems, I.E.E.E. American Control Conference, Albuquerque, New Mexico.
Lohmiller, W., and Slotine, J.J.E. (1997b). Applications of Contraction Anal-
ysis, I.E.E.E. Conference on Control Applications, Hartford, Connecticut.
Lohmiller, W., and Slotine, J.J.E. (1997c). Applications of Contraction Anal-
ysis, I.E.E.E. Conference on Decision and Control, San Diego, California.
Lohmiller, W., and Slotine, J.J.E. (1997d). Contraction Analysis: The Discrete-
Time Case, MIT-NSL Report 971101, November 1997.
Lovelock, D., and Rund, H. (1989). Tensors, differential forms, and variational
principles, Dover.
Luenberger, D. G. (1979). Introduction to Dynamic Systems, John Wiley &
Sons.
Mracek, P., Cloutier, J., D’Souza C. (1996 A new Technique for Nonlinear
Estimation, I.E.E.E. International Conference on Control Applications, Dearborn,
Michigan.
Marino, R., and Tomei, T. (1995). Nonlinear Control, Prentice-Hall.
Mussa-Ivaldi, F.A. (1997). Nonlinear Force Fields: a distributed system of con-
trol primitives for representing and learning movements, 1997 IEEE International
Symposium on Computational Intelligence in Robotics and Automation, pp. 84-90.
Mussa-Ivaldi, F.A., Giszter S.F., Bizzi E. (1994). Linear Superposition of
primitives in motor control. Proceedings and National Academy of Sciences. 91:7534-
7538.
Nijmeyer, H., and Van der Schaft, A. (1990). Nonlinear Dynamical Control
Systems, Springer Verlag.
Popov, V.M. (1973). Hyperstability of Control Systems, Springer-Verlag.
Schwartz, L. (1993). Analyse, Hermann, Paris.
Simon, H.A. (1981). The Sciences of the Artificial, 2nd edition, MIT Press.
Slotine, J.J.E., and Li, W. (1991). Applied Nonlinear Control, Prentice-Hall.
Strogatz, S. (1994). Nonlinear Dynamics and Chaos, Addison Wesley.
Vidyasagar, M. (1992). Nonlinear Systems Analysis, 2nd Edition, Prentice-Hall.
Wu, F., Packard, A., Bals G. (1995). LPV Control Design for Pitch-Axis Mis-
sile Autopilots. I.E.E.E. International Conference on Decision and Control, New
Orleans, Louisiana.

27