Sieber, J. (2006) - Dynamics of Delayed Relay Systems.: University of Bristol - Explore Bristol Research

Sieber, J. (2006). Dynamics of delayed relay systems.
https://doi.org/10.1088/0951-7715/19/11/001
Early version, also known as pre-print
Link to published version (if available):

10.1088/0951-7715/19/11/001
Link to publication record in Explore Bristol Research

PDF-document
University of Bristol - Explore Bristol Research

General rights
This document is made available in accordance with publisher policies. Please cite only the
published version using the reference above. Full terms of use are available:
http://www.bristol.ac.uk/pure/user-guides/explore-bristol-research/ebr-terms/
Dynamics of delayed relay systems
J Sieber
Bristol Centre of Applied Nonlinear Mathematics, Dept. of Engineering
Mathematics, University of Bristol, U.K.,
E-mail: [email protected]
Abstract. The paper studies the dynamics near periodic orbits in dynamical
systems with relays (switches) that switch only after a fixed delay. As a motivating
application, we study the problem of stabilizing an unstable equilibrium by
feedback control in the presence of a delay in the control loop. We show that
saddle-type equilibria can be stabilized to a periodic orbit by a switch even if this
switch is subject to an arbitrarily large delay. This is in contrast to linear static
feedback control, which fails when the delay is larger than a problem-dependent
critical value. Our analysis is based on the reduction of the return map near a
generic periodic orbit to a finite-dimensional map. This map is smooth if the
periodic orbit satisfies two genericity conditions. A violation of any of these two
conditions causes a discontinuity-induced bifurcation of the periodic orbit. We
derive asymptotic expressions for the piecewise smooth return map for each of
these two codimension-one bifurcations. This analysis shows that the introduction
of a small delay into the switching decision can induce chaos in a relay system
that had a single stable periodic orbit without delay. This small-delay behaviour
is fundamentally different from smooth dynamical systems.
AMS classification scheme numbers: 37G15, 34K18, 34K35
1. Introduction
This paper is concerned with dynamical systems with delayed relays. Relay systems
follow two different smooth vector fields in two different regions of their physical space.
Specifically, we consider the effects of a time delay in the decision when to switch from
one vector field to the other. As an initial motivation let us consider the problem of
stabilizing an unstable equilibrium by feedback in the presence of delay in the feedback
loop, which is a typical situation in applications. For example, a controlled inverted
(massless and frictionless) pendulum on a cart, as shown in figure 1, is governed by
the equation
θ̈ = sin θ − F cos θ. (1)
In (1), the dependent variable θ is the inclination angle of the pendulum. The force
F is applied as a feedback to the cart with the goal of stabilizing thep
unstable upright
position θ = 0; see figure 1. Time has been rescaled to units of 2L/(3g) in (1)
where L is the length of the pendulum and g describes the gravitational acceleration.
This implies that a fixed reaction time in the application of the feedback force F (θ, θ̇)
gives a delay τ in the arguments of F which increases for decreasing L. The inverted
pendulum is a prototype for balancing tasks in robotics and biomechanics [1, 2], and
Dynamics of delayed relay systems 2
Figure 1. Sketch of the setup for the controlled inverted pendulum on a cart.
a textbook example in control theory [3] and the study of delay effects [4]. Let us
consider the following question:
Problem 1 (Balancing) Let τ > 0 be a given, potentially large, delay. Find a
function F : R2 → R such that the feedback law F (θ(t − τ ), θ̇(t − τ )) inserted into (1)
is able to stabilize the upright position θ = 0.
For linear F this is impossible
√ as soon as the delay τ exceeds a certain critical value τc .
The critical delay τc = 2 is derived in the textbook [4] for the classical PD control
law F (θ(t − τ ), θ̇(t − τ )) = aθ(t − τ ) + bθ̇(t − τ ). The works [5, 6] have found critical
delays also for other specific linear control laws. Reference [7] presents a complete
stabilizability analysis for two-dimensional linear systems with static feedback subject
to time-delay, giving the critical delay in dependence of all relevant system parameters.
The references [8, 9] include small oscillations and other nonlinear phenomena, which
occur for delays close the critical value, into their study. A conclusion of [6] is that,
even if one accepts small stable oscillations around the upright position as successful
balancing, the restriction on the delay cannot be relaxed substantially beyond the
critical value obtained from the linear theory.
In order to overcome this fundamental restriction, we consider a relay switch in
(1) of the form
F = εsgn[g(θ(t − τ ), θ̇(t − τ ))], (2)
2 2
where g : R → R is a smooth or piecewise affine function dividing R into two
simple domains G1 = {g < 0} and G2 = {g ≥ 0}. A feedback of the form (2) can
never stabilize the equilibrium θ = 0 perfectly but will, at best, admit small stable
oscillations that switch back and forth between F = ε and F = −ε [10]. If we accept
small oscillations as successful balancing then a relay switch F of the form (2) can
achieve stabilization [11]. Surprisingly, one can even construct a stabilizing feedback
of form (2) for any given delay τ , thus, removing any restriction on the delay. In
section 5 we will give a geometric illustration how to construct the switching function
g for a given delay τ for the inverted pendulum (1) and prove the following general
result:
Theorem 2 (Existence of stable periodic orbits) Let f : Rn × R → Rn be
smooth and let ẋ = f (x, 0) have a saddle equilibrium x0 . Let τ > 0 be arbitrarily
large and ε > 0 be sufficiently small. If the pair (∂x f (x0 , 0), ∂u f (x0 , 0)) is controllable
then there exists a smooth function g : Rn → R such that
ẋ = f (x, εsgn[g(x(t − τ ))])
has a stable periodic orbit.

The only condition on (∂x f (x0 , 0), ∂u f (x0 , 0)), apart from the saddle type of x0 , is
controllability (which is a genericity condition), which is obviously necessary. The key
point of Theorem 2 is that controllability is also sufficient. The stable periodic orbit
is in a neighbourhood of x0 of order ε. In the formulation of Theorem 2 the switching
law g depends on ε and τ . In particular, a function g that guarantees a stable periodic
orbit for a certain large delay τ does not necessarily also guarantee a stable periodic
orbit for all smaller τ .
A crucial ingredient in the proof of Theorem 2 is a precise description of
the dynamics near periodic orbits of a general n-dimensional system of differential
equations of the form
(
f1 (x(t)) if g(x(t − τ )) < 0
ẋ(t) = (3)
f2 (x(t)) if g(x(t − τ )) ≥ 0
where τ > 0 is the delay in the switching decision. The presence of the delay in (3)
gives rise to an infinite-dimensional phase space, the space C([−τ, 0]; Rn ) of continuous
functions on the history interval [−τ, 0]. However, if a periodic orbit x̃(·) of (3) switches
only finitely often per period and satisfies two genericity conditions then the dynamics
of (3) near x̃(·) is described by a smooth finite-dimensional local return map. In short,
the genericity conditions are that
(i) all intersections of the periodic orbit x̃(·) with the switching manifold {g = 0}
are transversal, and
(ii) along the periodic orbit x̃(·) none of the delayed switching events coincides with
a crossing of the switching manifold {g = 0} (that is, if g(x̃(s)) = 0 then
g(x̃(s + τ )) 6= 0 for all s ∈ R).
We will derive a precise relation between the dimension of the image of the return map
and the location of the switching times of the orbit x̃(·). In particular, this dimension is
n−1 (where n is the dimension of the physical space of (3)) if all switching times along
x̃(·) are separated by more than the delay time τ . This kind of the periodic orbits is
called slowly oscillating. The finite-dimensionality of the local return maps of periodic
orbits is in contrast to the situation in smooth delay differential equations (DDEs)
where periodic orbits typically have infinitely many non-zero Floquet multipliers [12].
The second main result of the paper gives a complete description of possible
discontinuity-induced bifurcations of codimension one for a slowly oscillating periodic
orbit x̃(·). Each of these bifurcations corresponds to a violation of one of the genericity
conditions (i) and (ii). Violation of condition (i) implies generically that x̃ grazes
(touches) the switching manifold {g = 0} quadratically. This induces a return map
for x̃ that is asymptotically linear on one side of the grazing manifold and square-root
like on the other side. This square-root asymptotics implies that the introduction of
a small delay into the switching decision of a relay system can change the dynamics
drastically. In particular, it can introduce chaos into a system that, without delay,
has a stable periodic orbit as its only attractor. This small-delay limit behaviour is
fundamentally different from the case of smooth DDEs and will be illustrated with a
more detailed example in section 6.3. The violation of condition (ii) corresponds to
a corner collision and gives rise to a piecewise asymptotically linear return map near
the colliding periodic orbit x̃. This reduction to piecewise smooth finite-dimensional
maps links the local bifurcation theory of periodic orbits in delayed relay systems
to the well-established results of the bifurcation theory for piecewise smooth maps
[13, 14, 15, 16].
The paper is organized as follows. Section 2 outlines how the results of this paper
relate to previous and recent studies on the dynamics of piecewise smooth ordinary and
delay differential equations, and how the result of Theorem 2 relates to common delay
compensation techniques in control theory and engineering. Section 3 revisits some
common notation for the definition of the forward evolution of DDEs, also pointing out
the differences to the case of smooth DDEs. Section 4 shows under which conditions
the local return map of a periodic orbit reduces to a finite dimensional smooth map.
Section 5 first shows how one can construct a switching law g in (2) that gives rise to
a stable periodic orbit for the inverted pendulum in the presence of an arbitrary delay.
This construction reveals already the main ideas of the proof for the general result
in Theorem 2. The section also lists the main differences between the illustrating
example and the general n-dimensional case. The detailed proof of Theorem 2 can be
found in Appendix C. Section 6 studies the two codimension-one bifurcations of slowly
oscillating periodic orbits, stating secondary non-degeneracy conditions and deriving
asymptotic expressions for the return maps. The sections 4, 5 and 6, which contain
technical material and general theoretical results, include also simple but instructive
examples illustrating the main concepts and ideas. More technical parts of the proofs
for statements in the sections 4, 5 and 6 are given in separate appendices.
2. Background
Piecewise smooth dynamical systems model many problems in control engineering

[17, 18], in mechanics (for example in systems with dry friction [19] or impacts [20]),
in electrical systems with switches [21], or in biological systems with threshold effects
[22]. In these situations one observes an evolution that is governed by different smooth
vector fields in different regions of the phase space, which are separated by switching
manifolds. These hybrid systems are an attractive subject of study as they can
generate complex dynamics even if all of the vector fields and switching manifolds
are simple enough (for example linear) to study them analytically. Moreover, they
show phenomena such as chaotic attractors robustly, which are often non-hyperbolic
and, thus, extremely subtle, in smooth maps and vector fields. This feature allows one
to ‘engineer’ particular dynamics such as chaos [16]. In control theory piecewise linear
systems are used to approximate nonlinear systems to understand the global dynamics
and guarantee global stability [17]. See [14] for a survey on the active development of
general bifurcation theory for piecewise smooth dynamical systems.
Whenever the non-smoothness of the dynamical system is induced by the
implementation of a switch one can expect that the actual switch is subject to a delay,
giving rise to delayed relay models such as (3). In applications this delay is often
artificially increased (or hysteresis is introduced) since otherwise so-called ‘sliding’
along the switching manifold can occur, which would involve are large number of
switchings in a short time interval [18].
The works [10, 23] studied one-dimensional prototype examples of the form
ẋ = κ x − sgnx(t − τ ), (κ > 0), (4)
and found that this type of system typically admits periodic orbits that switch back
and forth between the two vector fields. Moreover, they have classified all possible
dynamics of system (4) completely and also studied its behaviour with respect to
perturbations, including periodic forcing. The references [24, 25, 26] have studied other
simple piecewise linear systems (typically with a two-dimensional physical space). In
contrast to the studies of (4) these investigations have found a huge variety of different
dynamics such as chaos [26] or a complex network of periodic orbits [24, 25]. The
different regimes are connected by grazing or collision events that show similarities
to those in impacting or dry-friction systems [14]. However, even the behaviour of
simple prototype systems such as studied in [24, 25, 26] is far from being classified
completely.
In this paper we adopt a different approach. We consider a general system of
form (3) and assume that it has a periodic orbit x̃(t) (t ∈ [−p, 0]) that has a finite
number of switchings between the vector fields. Then we study the dynamics near
this periodic orbit and its bifurcations. In this way the results of our paper will be
more general than studies of specific classes of examples such as [10, 24, 25, 26] but all
statements are valid only locally. The consideration of only two vector fields in (3) is
not really a restriction when one studies the local dynamics near a particular periodic
orbit.
A further motivation for the study of the general system (3) is its connection with
smooth delay differential equations (DDEs) with steep nonlinearities. Often one can
start from (3) as a limiting case where the existence of stable periodic orbits is easy
to prove and then deduce the persistence of these orbits for smooth DDEs close to (3)
[27]. Reference [24] also continued periodic orbits of (3) approximately by standard
numerical software for smooth DDEs after ‘smoothing’ the discontinuity in (3). The
limit turns out to be well-behaved if the periodic orbit is not close to one of the
bifurcations discussed in section 6.
Finally, let us put Problem 1 and Theorem 2 into perspective compared to
classical delay compensation techniques in control theory and engineering. The studies
[4, 5, 6, 7] and Theorem 2 are restricted to static feedback. That is, the feedback
law (for example F in (1)) is only a function of a single instance of the delayed
state. Classical delay compensation techniques that can cope with an arbitrarily
large delay rely on dynamic feedback where the feedback depends on a predictor,
obtained by a real-time solution of a functional equation (see, for example, [28, 29]).
The fact that the basin of attraction of the periodic orbit in Theorem 2 will, in
general, be exponentially small for large delay τ is only formally a difference to classical
dynamic feedback schemes. Even though methods based on functional predictors can
be globally asymptotically stable on the linear level, they have exponentially large
transients if the initial condition is not already exponentially close to the equilibrium.
See also [30] for a survey on implementation problems of functional predictors and
how to overcome them. In the case of small delays polynomial forward prediction,
such as used in substructuring [31, 32] in civil and mechanical engineering, is often
successful and easier to implement in real-time.
3. Fundamental properties of delayed relay systems — definition of

forward evolution
We define a delayed relay system as a dynamical system governed by a differential

equation of the form (3) where τ > 0, and f1 , f2 : Rn → Rn are Lipschitz continuous.
We assume that the switching function g : Rn → R is a piecewise smooth Lipschitz
continuous function. Furthermore, we assume that the gradient g ′ (x) is non-zero
whenever it exists and g(x) is zero. These assumptions on g imply that the set
{x : g(x) = 0} constitutes a piecewise smooth (n − 1)-dimensional submanifold of Rn ,

which we call the switching manifold.
We denote the n-dimensional flow corresponding to fj (j = 1, 2) by Φj . That is,
the time-t map generated by ẋ = fj (x) is Φj (t; ·) : Rn → Rn .
When solving differential equations where the right-hand-side depends also on
the state in the past one typically has to keep track of the solution history along the
trajectory [12]. Thus, the natural phase space, also for a system of the form (3), is the
space C([−τ, 0]; Rn ) of continuous functions on the closed interval [−τ, 0]. This section
is concerned with the definition of the forward evolution E(t; ·) for (3), which maps
an initial value x0 ∈ C([−τ, 0]; Rn ) to its time-T image E(T ; x0 ) ∈ C([−τ, 0]; Rn ).
In the case of a single delayed argument with a fixed delay τ as in (3) an intuitive
way to define E(T ; x0 ) is the method of steps [12, 33]: treat the past x0 (t) (t ∈ [−τ, 0])
as an inhomogeneity, solve the ensuing ordinary differential equation (ODE) for all
times up to τ , and then shift the history and repeat the process.
For example, consider an initial history segment x0 ∈ C([−τ, 0]; Rn ) that
intersects the switching manifold only finitely many times, that is, g(x0 (t − τ )) = 0
only for t = t1 , . . . , tµ ∈ [0, τ ]. This divides the interval [0, τ ] into µ + 1 subintervals
Ik :
I0 = (0, t1 ], Ik = (tk , tk+1 ] for k = 1, . . . , µ − 1, Iµ = (tµ , τ ].
The forward evolution will follow one of the flows Φjk (jk = 1 or 2 for k = 0, . . . , µ)
in each subinterval Ik . Thus, we can define the curve x(t) for t ∈ [0, τ ] recursively by
x(t) = Φj0 (t; x0 (0)) for t ∈ I0
x(t) = Φjk (t − tk ; x(tk )) for t ∈ Ik , k = 1 . . . , µ − 1, (5)
x(t) = Φjµ (t − tµ ; x(tµ )) for t ∈ Iµ .
For any t in the interior of any of the intervals Ik the point x(t) satisfies the
differential equation (3) with the history x0 . Thus, the forward evolution E(T ; x0 ) ∈
C([−τ, 0]; Rn ) for T ∈ [0, τ ] is defined by
(
x(t + T ) if t ∈ (−T, 0]
E(T ; x0 )(t) = (6)
x0 (t + T ) if t ∈ [−τ, −T ].
For times T > τ we define E(T ; ·) as a concatenation of time steps smaller than τ , for
example E(T ; ·) := E(T /(k + 1); ·) ◦ . . . ◦ E(T /(k + 1); ·) when T ∈ [kτ, (k + 1)τ ). This
definition is independent of the particular partition of the interval (0, T ).
Recursion (5) reveals that the evolution E(·, x0 ) does not depend on the complete
shape of x0 ∈ C([−τ, 0]; Rn ) but only on the position of x0 (0) ∈ Rn (the headpoint
of x0 ) and the finitely many switching times t1 − τ ,. . . ,tµ − τ in the interval [−τ, 0].
This suggests that the dynamics of delayed relay systems such as (3) is governed by
only finitely many coordinates despite the infinite-dimensionality of the underlying
phase space. This is generically the case near periodic orbits with only finitely
many intersections of the switching manifold, which are discussed in section 4. The
construction also shows that delayed relays cannot induce ‘sliding’, which is common
in non-delayed systems of the form (3) (that is, if τ = 0 in (3)).
The above construction of E(·; x0 ) assumes that x0 intersects the switching
manifold only finitely many times within [−τ, 0]. For many elements of C([−τ, 0]; Rn )
this is not the case. For general x0 ∈ C([−τ, 0]; Rn ) we define the curve x(t) as the
solution x ∈ C([0, T ]; Rn ) of the variation-of-constants formula corresponding to DDE

(3)
1 t
Z
x(t) = x0 (0) + f1 (x(s)) [1 − sgn g(x0 (s − τ )))] +
2 0 (7)
+ f2 (x(s)) [1 + sgn g(x0 (s − τ ))] ds.
In the integral equation (7) we use the convention that sgn 0 = 1. Equation (7)
has a unique solution x ∈ C([0, T ]; Rn ) satisfying x(0) = x0 (0) due to the Lipschitz
continuity of f1 and f2 and the measurability of sgn g(x0 (·)). In general, the points
x(t) satisfy the differential equation (3) for t in the open and dense subset of (0, T )
{t ∈ (0, T ) : g(x0 (t − τ )) < 0} ∪ int {t ∈ (0, T ) : g(x0 (t − τ )) ≥ 0}
where the notation int refers to the interior of a set. We remark that this set may
not necessarily have full Lebesgue measure on the interval (0, T ). For general x0 we
use the solution x of (7) instead of the simple recursion (5) in the definition (6) of
E(T ; x0 ).
We observe that the evolution E(T ; x0 ) depends continuously on T but, in
general, it does not depend continuously on x0 . In fact, arbitrarily close to any
x0 ∈ C([−τ, 0]; Rn ) that intersects {g = 0} at least once (say, in s1 ∈ (−τ, 0)) we find
a xε ∈ C([−τ, 0]; Rn ) which has g(xε (s)) = 0 for all s ∈ (s1 − ε, s1 + ε). In general,
we cannot expect that E(T ; xε ) is continuous in its second argument in xε . Thus, E
is not a semiflow in the classical sense of [34].
4. Behaviour near generic relay periodic orbits
Although equation (3) does not define a semiflow we can often understand the
dynamics generated by (3) near periodic orbits by studying smooth finite-dimensional
maps. This section will explain in detail how this reduction near periodic orbits works
in the simplest (but generic) case.
4.1. Illustration — linearized inverted pendulum

Let us consider the example of the inverted pendulum from the introduction to
illustrate how the infinite-dimensional semiflow simplifies to a low-dimensional map
close to a periodic orbit. Inserting the relay feedback (2) into the differential equation
governing the controlled inverted pendulum leads to a system of form (3). In the
consideration of small periodic orbits close to the upright position the nonlinearities
in (1) can be regarded as small perturbations. If ε ≪ 1 and after rescaling
(θ, θ̇) = (εx1 , εx2 ) the nonlinear equation (1) with (2) is a perturbation of order
O(ε2 ) of
ẋ1 (t) = x2 (t)
(8)
ẋ2 (t) = x1 (t) − sgn g(x1 (t − τ ), x2 (t − τ )).
Any structurally stable periodic orbit found in (8) will persist under small
perturbations, and, thus, after rescaling, also exist in the nonlinear system (1) for
sufficiently small ε. This reduction of a piecewise smooth system to the piecewise
linear system (8) is an expression of the general fact that many key features of piecewise
smooth dynamical systems can already be found in piecewise linear systems [14] where
they simply persist under the perturbation caused by a small nonlinearity. The two
x2
(0, 1)
Gτ1 = Φ1 (τ ; {g = 0}) l2s
{g = 0} C1
l1u
(1, 0)
(−1, 0) x1
l2u
C2
l1s
Gτ2 = Φ2 (τ ; {g = 0})
(0, −1)
Figure 2. Sketch of the dynamics near a periodic orbit of the linearized inverted
pendulum. The flows Φ1 (dashed) and Φ2 (dotted) are superimposed in R2 . The
periodic orbit is the bold closed curve with corners the C1 and C1 . The return
map to this orbit is a 1D map from Gτ1 back to itself.
flows Φ1 and Φ2 can be computed analytically for (8), giving rise to the affine maps
Φ1 (t; v) = A(t)v − v0 (t), and Φ2 (t; v) = A(t)v + v0 (t)
where

cosh(t) sinh(t) 1 − cosh(t)
A(t) = , v0 (t) = .
sinh(t) cosh(t) − sinh(t)
Let us choose, for illustration, a linear switching function g of slope α, namely
g(x1 , x2 ) = x1 cos α + x2 sin α, α ∈ (0, π/2),
and consider a delay τ < log(1 + tan α). Figure 2 shows a sketch of the situation.
The flows Φ1 (dashed) and Φ2 (dotted) are superimposed in the plane R2 . The flow
Φ1 has a saddle equilibrium at (−1, 0), the flow Φ2 has a saddle at (1, 0). The stable
s u
(l1,2 ) and unstable (l1,2 ) subspaces of both flows form a square, which is sketched in
figure 2. We denote the Φ1 (τ ; ·)-image of {g = 0} by Gτ1 and its intersection point
with the axis x1 = 0 by C1 . Correspondingly, the Φ2 (τ ; ·)-image of {g = 0} is denoted
by Gτ2 and its intersection point with the axis x1 = 0 by C2 . The points C1 and C2 are
mirror images of each other (C2 = −C1 ). If τ < log(1 + tan α) they are also mapped
onto each other by the flows. That is, C2 = Φ2 (p/2; C1 ), C1 = Φ1 (p/2; C2 ) where
τ
e tan α + 1 − eτ

p = 2 τ + log . (9)
tan α + 1 − eτ
This implies that the closed curve W = Φ1 ([0, p/2); C2 ) ∪ Φ2 ([0, p/2); C1 ) is the graph
of a periodic orbit of (8). Moreover, the dynamics near W are given by the return map
to the line Gτ1 , which is a one-dimensional map. Any initial value x ∈ C([−τ, 0]; R2 )
that is sufficiently close to Φ1 ([−τ, 0]; C1 ) will, after time τ , follow Φ2 . Thus, the next
switching to Φ1 will invariably be located on the time-τ image of {g = 0} under Φ2 ,
which is Gτ2 . From now on the trajectory will always follow Φ1 to Gτ1 and Φ2 to Gτ2 ,
reducing the evolution of (8) to a smooth one-dimensional map from Gτ1 back to itself.
This map is nonlinear if α 6= π/4 even though both flows and g are linear.
The facts that make this reduction possible are that
(i) the switchings of W (C1 and C2 ) have a positive distance from {g = 0},
(ii) the intersections of W with {g = 0} are transversal,
(iii) the time between successive crossings of the switching manifold is larger than the
delay τ .
The first two conditions are genericity conditions. Their violations correspond to
discontinuity induced bifurcations, which are discussed in section 6. Periodic orbits
that satisfy the last condition are called slowly oscillating.
We observe that the same curve W is the graph of a periodic orbit also if the
delay τ in (8) is replaced by a delay of size τ + kp where k is a positive integer and p
is the period of W given in (9). Then all time differences between successive crossings
of the switching manifold {g = 0} are smaller than the switching delay. Thus, W for
delay τ + kp with k ≥ 1 would be a rapidly oscillating periodic orbit. A general lemma
expressing the return map will be developed in the following section. It also applies
to rapidly oscillating orbits and gives for the pendulum case a dimension of 1 + 2k for
the return map.
4.2. Return map for periodic orbits in the general case

Suppose that the evolution of (3) has a periodic orbit W of period p. We can assume
that p ≥ τ without loss of generality because we do not require p to be the minimal
period. We denote the elements of W by x̃t where x̃t = E(t; x̃0 ) ∈ C([−τ, 0]; Rn ) for
t ∈ (0, p] and x̃p = x̃0 , and denote the corresponding trajectory of headpoints x̃t (0)
by x̃(t). The closed curve x̃([−p, 0]) ⊂ Rn comprises the graph of the periodic orbit in
the physical space Rn . The function x̃(·) can be extended to the whole real axis due
to the periodicity of W . In the following paragraphs we formulate three fundamental
assumptions on f1 , f2 , g and x̃. If f1 , f2 and g are at least piecewise smooth functions
then these conditions (3, 5, and 6) are genericity conditions on the periodic orbit.
Condition 3 (finitely many intersections with switching manifold)
We assume that the graph x̃([−p, 0]) ⊂ Rn of the periodic orbit W intersects the
switching manifold in at most finitely many points. That is, g(x̃(s̃k )) = 0 for at most
finitely many times s̃k (k = 1, . . . , m) in (−p, 0).
Then, x̃([−p, 0]) is composed of m + 1 curves following either Φ1 or Φ2 with switching
(or touching) times t̃k = [(s̃k +τ ) mod p]−p. We can assume without loss of generality
that the intersection times s̃k and the switching times t̃k lie in the open interval (−p, 0)
for k = 1, . . . , m. The trajectory t → x̃(t) is differentiable for all t ∈ [−p, 0] except
possibly in t̃k (k = 1, . . . , m).
The following lemma states that the evolution E is continuous with respect
to initial conditions in all points of the periodic orbit W in the topology of the
Banach space C([−τ, 0]; Rn ). As pointed out in section 3 this continuity statement
is rather subtle. If x̃(·), the graph of W , has at least one intersection with the
switching manifold, and if f1 6= f2 in this intersection, we cannot find a whole open
neighbourhood of W where E(t; x) is continuous with respect to x.
Lemma 4 (Continuity of evolution in periodic orbits)

Let t0 ∈ R, T > 0 be arbitrary, and x̃t0 be an element of a periodic orbit with finitely
many intersections of the switching manifold. Then, E(T ; x) is continuous with respect
to x ∈ C([−τ, 0]; Rn ) in the point x = x̃t0 . Moreover, this continuity is uniform in T
for T in any finite interval [0, T0 ].
Lemma 4 exploits the fact all x close to x̃t0 in C([−τ, 0]; Rn )-topology follow the
same flow as x̃ outside of small neighbourhoods of the finitely many time points
t̃1 , . . . , t̃m , t̃1 + p, . . . , t̃m + p, . . . Thus, the proof of continuity for all T > 0 relies
strongly on the periodicity of x̃(·). The complete proof is included as Appendix A.
Due to the continuity stated in Lemma 4, it makes sense to define a local return
map, or Poincaré map, for the periodic orbit W . It is defined as the map induced
by the first return to a hyperplane in the phase space transversal to the periodic
orbit. For simplicity of notation we restrict ourselves in all further considerations to
return maps to hyperplanes defined by a condition on the headpoint z(0) of a function
z ∈ C([−τ, 0]; Rn ). Let l0 ∈ Rn be a vector of length 1 such that l0T x̃(0) ˙ > 0. The
curve x̃(·) is differentiable in t = 0 because all switching times t̃k are different from 0.
The Poincaré map is defined as the local return map from the hyperplane
H := {z ∈ C([−τ, 0]; Rn ) : l0T [z(0) − x̃(0)] = 0} (10)
to itself. Locally, this is a well-defined map in the following sense.
There exist small open neighbourhoods U0 ⊂ U1 ⊂ C([−τ, 0]; Rn ) of x̃0 (x̃(0) is
the headpoint of x̃0 ) and a neighbourhood (T1 , T2 ) ⊂ R of p such that
• E(p; U0 ) ⊂ U1 , and
• for all z ∈ H ∩ U0 , there exists a unique time T (z) of first return to the set U1 ∩ H
within (T1 , T2 ).
That is, T (z) is defined as the minimal time in the interval (T1 , T2 ) satisfying
l0T [E(T (z); z)(0) − x̃(0)] = 0. This first return time T (z) is a well defined function
of z for all z ∈ U0 . The return time T (z) is continuous in the point z = x̃0 but,
in general, not in U0 because not necessarily all trajectories starting in U0 intersect
H transversally. Hence, also the Poincaré map P : U0 ∩ H → U1 ∩ H defined by
P z := E(T (z); z) is well defined in H ∩ U0 , is continuous in z = x̃0 , but not necessarily
continuous in H ∩ U0 . Let us denote by S the domain of definition of P , the local
Poincaré section U0 ∩ H.
Let L ∈ Rn×(n−1) be such that the augmented matrix [l0 L] ∈ Rn×n is orthogonal.
Thus, l0T Lv = 0 for all v ∈ Rn−1 . The headpoints of elements of the local Poincaré
section S all have the form x̃(0) + Lv where v ∈ Rn−1 is small.
The following two assumptions on the periodic orbit x̃ and the switching function
g will allow us to reduce the Poincaré map P to a smooth finite-dimensional map.
Condition 5 (Smoothness in intersection points)
We assume that g(x̃(t̃k )) 6= 0 for all k = 1, . . . , m.
Condition 5 implies that the two sets of points {x̃(s̃1 ), . . . , x̃(s̃m )} (where x̃(·)
intersects the switching manifold) and {x̃(t̃1 ), . . . x̃(t̃m )} (the corners of x̃, where x̃(·)
actually switches) are disjoint. Furthermore, it implies that x̃ is differentiable in its
intersections with the switching manifold at the times s̃k (k = 1, . . . , m). Thus, x̃
follows either Φ1 or Φ2 in s̃k .
Condition 6 (Transversality of all intersections)

For all k = 1, . . . , m holds: The function g is differentiable in the vicinity of x̃(s̃k ), and
˙ k ) 6= 0. More precisely, if x̃ follows Φj in s̃k then g ′ (x̃(s̃k ))fj (x̃(s̃k )) 6= 0.
g ′ (x̃(s̃k ))x̃(s̃
Condition 6 asserts that the switching manifold {g = 0} is differentiable whenever it
intersects the periodic orbit x̃. Moreover, Condition 6 asserts that the orbit x̃ intersects
the switching manifold transversally in all its intersection points x̃(s̃k ) (k = 1, . . . , m).
Consequently, the number of switching times, m, must be even. We can assume that
x̃ follows Φ1 in t = 0 without loss of generality.
To fix notation, we number the intersection and switching times so that, for some
µ ∈ {0, . . . , m},
−τ < t̃1 < . . . < t̃µ < 0, and − p < t̃µ+1 < . . . < t̃m < −τ (11)
with, correspondingly, s̃k = [(t̃k − τ ) mod p] − p. Thus, x̃(·) has the form


 Φ1 (t; x̃(0)) if t ∈ [t̃µ , 0],

 Φ2 (t − t̃k ; x̃(t̃k ))

if t ∈ [t̃k−1 , t̃k ) and µ − k is even,
x̃(t) = (12)
 Φ1 (t − t̃k ; x̃(t̃k ))

 if t ∈ [t̃k−1 , t̃k ) and µ − k is odd,

Φ1 (t + t̃µ+1 ; x̃(t̃µ+1 )) if t ∈ [−p, t̃µ+1 )

for k ∈ {1, . . . , m}. Lemma 7 below states that the dynamics of the local Poincaré
map P : S → S is attracted by a finite-dimensional local invariant manifold M
after finite time. Moreover, the local manifold M can be parametrized by tuples
(v, t1 , . . . , tµ ) ∈ Rn−1+µ where each of the µ numbers tk is close to t̃k and the vector
v ∈ Rn−1 is small. The number µ equals the number of switchings of the periodic orbit
in the interval [−τ, 0] (see (11)). In the formulation of the lemma we use the notation
that a set M is ‘invariant under P relative to a set N ’ if any trajectory starting in
M ∩ N stays in M ∩ N under iterations of P as long as it stays in N .
Lemma 7 There exists an open neighbourhood N of x̃0 in the local Poincaré section
S that is mapped by P 2 = P ◦ P into a local manifold M ⊂ S of dimension n − 1 + µ
where µ is defined by (11). The local manifold M is invariant under P relative to
N . Moreover, M can be parametrized by a small open ball B ⊂ Rn−1+µ around
(0, t̃1 , . . . t̃µ ) ∈ Rn−1 × Rµ . The parametrization of M
IM : (v, t1 , . . . , tµ ) ∈ B → z ∈ C([−τ, 0]; Rn ) (13)
is defined recursively by

 Φ1 (t; x̃(0) + Lv)
 if t ∈ [tµ , 0],
z(t) = Φ2 (t − tk ; z(tk )) if t ∈ [tk−1 , tk ) and µ − k is even, (14)

Φ1 (t − tk ; z(tk )) if t ∈ [tk−1 , tk ) and µ − k is odd

where we use the notation t0 := −τ .

We note that the manifold M and even the number µ defining the dimension of M may
depend on the choice of the hyperplane of the Poincaré section H. The proof, which
appears in full in Appendix B, is based on the fact that any initial condition sufficiently
close to x̃0 will also always intersect {g = 0} transversally after one iteration of P .
This gives rise to parametrization (14) in the second iteration of P . Description (14)
of M expresses that, for elements z of M, we have to store only the location v of
the headpoint (z(0) = x̃(0) + Lv) and the switching times within (−τ, 0), of which we
have exactly µ if we are sufficiently close to the periodic orbit x̃.
Furthermore, IM induces a map P0 on B defined by IM P0 y = P IM (y). If B is

sufficiently small all intersections of the headpoint trajectory E(t; IM (y))(0) with the
switching manifold are still transversal for all y ∈ B. Thus, P0 is differentiable and
the smoothness of P0 is only limited by the smoothness of the switching function g
and the flows Φj . Hence, if fj and g are differentiable to a higher degree then P0 is
as well.
We remark that the decay to zero in finite time for certain components of
the infinite-dimensional initial condition is a common feature also in smooth delay
differential equations. This phenomenon manifests itself in the existence of ‘small
solutions’; see [33] for a comprehensive discussion in a textbook.
4.3. Poincaré map P0 for slowly oscillating orbits

For slowly oscillating periodic orbits the manifold M simplifies to the local manifold
in Rn
G0 := {z ∈ Rn : l0T (z − x̃(0)) = 0} ∩ U (x̃(0)),
where we denote by U (ξ) a sufficiently small neighbourhood of a point ξ ∈ Rn . The
intersection times s̃k are separated by more than the delay time τ . This means that,
without loss of generality, we can order the switching times as
−p < s̃1 < t̃1 = s̃1 + τ < s̃2 < t̃2 = s̃2 + τ < . . . < s̃m < t̃m = s̃m + τ < 0. (15)
Introducing the local switching manifolds (that is, local neighbourhoods of x̃(s̃j ) within
the switching manifold {g = 0}) and their time-τ images (in the same way as for the
pendulum in section 4.1 and figure 2)
Gj := {g = 0} ∩ U (x̃(sj )) for j = 1, . . . , m
Gτj := Φ1 (τ ; Gj ) for j = 1, 3, . . . , m − 1 (odd),
Gτj := Φ2 (τ ; Gj ) for j = 2, . . . , m (even),
we can express the map P0 corresponding to the Poincaré map as a map from G0 back
to itself by the concatenation of maps
Φ Φ Φ Φ Φ Φ Φ
P0 : x ∈ G0 →1 G1 →1 Gτ1 →2 G2 →2 Gτ2 →1 . . . →2 Gτm →1 G0 . (16)
Φj
The symbol Gτk → Gk+1 is defined as the map from a submanifold Gτk to a submanifold
Gk+1 obtained by following the flow Φj . All maps in (16) are well defined and smooth
because the intersection of the flow Φj with the target manifold is always transversal
due to Condition 6. For slowly oscillating orbits the reduction of the Poincaré map to
P0 is a well-established fact that has been used extensively in many studies of delayed
relay systems, for example, in [24, 25, 26].
We remark that the map P0 will, in general, be nonlinear, even if the local
switching manifolds Gj and the flows Φ1 and Φ2 are affine, because the maps
Gτj 7→ Gj+1 are nonlinear.
Furthermore, we remark that rapidly oscillating solutions (periodic orbits with
µ > 0 for all choices of Poincaré sections) can also occur as stable periodic orbits of a
delayed relay system. Some of the periodic orbits found in [24, 25] have this structure.
The rapidly oscillating orbits in the linearized pendulum discussed in section 4.1 are,
however, all dynamically unstable [35].
(0, 1)
Gτ1
C1
l1u
G1 = Φ1 (−τ ; Gτ1 ) l2s
(−1, 0) (1, 0)
G∗
l1s
l2u G2 = Φ2 (−τ ; Gτ2 )
C2
Gτ2
(0, −1)
Figure 3. Construction of the switching line G = G1 ∪ G∗ ∪ G2 and the stable

periodic orbit C1 → C2 → C1 . For large delays τ , the lines G1 and G2 will be
close to l1s and l2s , respectively. Trajectories of Φ1 are dashed, trajectories of Φ2
are dotted.
5. Existence of stable periodic orbits for arbitrarily large delay
Let us come back to Problem 1 formulated in the introduction and the resulting
general Theorem 2 about the existence of stable periodic orbits. By choosing a suitable
switching law g, we can create a periodic orbit resembling any closed curve in Rn that,
in alternating fashion, follows Φ1 and Φ2 , always for a time longer than the delay τ .
Thus, this periodic orbit will be slowly oscillating, and its dynamical stability will be
determined by the concatenation of (n − 1)-dimensional maps of the form (16). Hence,
we can achieve the dynamic stability of the periodic orbit by ‘tilting’ the local switching
manifolds such that their time-τ images are tangential to desired hyperplanes.
5.1. Stabilization of the inverted pendulum

We first illustrate the main idea behind our construction of the desired switching
function g for the inverted pendulum example. As mentioned in section 4.1, it is
sufficient to find a function g, dividing R2 into two simple domains, such that the
piecewise affine equation (8) has a stable periodic orbit for a given τ > 0. More
precisely, it is sufficient to construct the two domains, D1 for the flow Φ1 and D2 for
the flow Φ2 , and a smooth boundary G separating them. A smooth function g can then
always be chosen such that clos D2 = {g(x1 , x2 ) ≤ 0} and clos D2 = {g(x1 , x2 ) ≥ 0}
(the notation clos refers to the closure of a set).
Figure 3 illustrates the following construction. First, we find a closed curve that
consists of two segments, one following Φ1 , one following Φ2 , both for a time longer
than τ . Such a curve exists: the periodic orbit W found in section 4.1 is of this type
if the points C1 and C2 are sufficiently close (0, ±1), respectively. Let h ∈ (0, 1/2) be
such that
eτ ∈ h−1 − 1, h−1 .

(17)
If we choose C1 = (0, 2eτ h−1)T and C2 = (0, 1−2eτ h) then Φ1 (2τ −log(1−h−1 ); C2 ) =
C1 and Φ2 (2τ − log(1 − h−1 ); C1 ) = C2 . The traveling time 2τ − log(1 − h−1 ) is larger
than τ by construction of h. Next, we find the boundary G such that this curve
W = Φ1 ([0, 2τ − log(1 − h−1 )]; C2 ) ∪ Φ2 ([0, 2τ − log(1 − h−1 )]; C1 ) (18)
is a stable periodic orbit of (8). The local delayed switching manifolds have to
contain the corners: C1 ∈ Gτ1 and C2 ∈ Gτ2 . If Gτ1 = C1 + s∂1 Φ2 (0; C1 ) where
s ∈ (−δ, δ) then Gτ1 is tangent to the outgoing flow Φ2 in C1 . At the same time Gτ1
is transversal to the incoming flow Φ1 in C1 . Thus, the image of Gτ1 under Φ1 (−τ ; ·)
is an affine line segment G1 intersecting W transversally within the (dashed) segment
Φ1 ([0, 2τ − log(1 − h−1 )]; C2 ) of the curve W . The corresponding local manifolds for
C2 are Gτ2 = −Gτ1 and G2 = −G1 . Since W does not self-intersect we can connect
G1 and G2 by a segment G∗ and extend G1 ∪ G∗ ∪ G2 to a global piecewise affine
manifold G0 which generates the periodic orbit W . This piecewise affine manifold can
subsequently be smoothed at its corners (which have a positive distance to the curve
W ) to obtain a smooth switching manifold G; see Figure 3.
Lemma 8 The periodic orbit W defined by (18) is stable.
Proof: As demonstrated in section 4.1, the Poincaré map P for the periodic orbit
W can be reduced to a one-dimensional return map P0 from Gτ1 to itself, defined by
following Φ2 to Gτ2 and then Φ1 back to Gτ1 . Let p0 = C1 + s∂1 Φ2 (0; C1 ) be a point in
Gτ1 close to C1 (that is, s ∈ R is small). Thus, p0 = Φ2 (s; C1 ) + O(s2 ). The traveling
time t(s) from p0 to Gτ2 is 2τ − log(1 − h−1 ) − s + O(s2 ) for small s. Thus, the image
of p0 under the flow Φ2 to Gτ2 is
p′0 = Φ2 (t(s); Φ2 (s; C1 )) = C2 + O(s2 ).
Since the map defined by following Φ1 from Gτ2 to Gτ1 is smooth this implies that
P0 (p0 ) = C1 + O(s2 )).
We observe that the orbit W is even quadratically stable. That is, the
linearization of P0 in C1 is zero. The periodic orbit W is also structurally stable.
That is, it is robust with respect to small nonlinearities or small perturbations of the
parameters, for example, of τ , or the location of the switching manifold. However,
this tolerance is exponentially small for large τ since (17) gives effectively a condition
on τ once h is chosen. Similarly, the basin of attraction of the periodic orbit W is
exponentially small with respect to τ .
Remark 1: Apart from the fact that the relay stabilizes to a periodic orbit instead of
the equilibrium, the exponential smallness of the basin of attraction is a difference to
classical methods for delay compensation, such as finite spectrum assignment [29].
Finite spectrum assignment is a linear dynamic control law based on an explicit
predictor. However, even though methods, such as finite spectrum assignment,
are globally asymptotically stable on the linear level, they have exponentially large
transients if the initial condition is not exponentially close to the equilibrium.
Remark 2: The study in [36] discusses systems of the form αẍ(t) = −ẋ(t) + kx(t) −
sgnx(t−1), finding conditions under which there are no bounded non-trivial oscillatory
solutions. The crucial difference between [36] and the above construction, which gives
a stable slowly oscillating solution for all α > 0, k > 0 is that we allow the relay to
depend not only on x(t − 1) but also on ẋ(t − 1). That is, the relay is of the form
sgng(x(t−1), ẋ(t−1)), giving rise to a switching curve in figure 3 which is not vertical.
5.2. Stable periodic orbits in n-dimensional systems

As we have stated in Theorem 2 in the introduction, the construction of a relay switch
for the simple two-dimensional inverted pendulum can be generalized to n-dimensional
systems with saddle equilibria. The general piecewise smooth nonlinear system
ẏ = f (y, ε sgn g̃(y(t − τ ))) (19)
close to an equilibrium y0 of ẏ = f (y, 0) and for small ε can be rewritten as a
perturbation of order ε of a piecewise affine system. Introducing the rescaled variable
x = ε−1 (y − y0 ), the rescaled representation of the switching law g(x) = g̃(y0 + εx),
and the partial derivatives A = ∂1 f (y0 , 0), v = −∂2 f (y0 , 0), the rescaled nonlinear
system reads as
ẋ = Ax − v sgn g(x(t − τ )) + εh(x, sgn g(x(t − τ )), ε) (20)
where the function h is uniformly smooth for all ε near zero. Any stable periodic orbit
x̃(·) found in the truncated system (20) with ε = 0 persists under small perturbations
to the two vector fields (that is, to ε > 0). This implies the existence of a corresponding
stable periodic orbit close to x̃(·) in (20) for sufficiently small ε > 0. Thus, (19) has a
stable periodic orbit with an amplitude of order ε.
The pair (A, v) is called controllable if (v, Av, . . . , An−1 v) has full rank n. An
unbounded domain is called simple if its closure is homeomorphic to a half-space, say
{z ∈ Rn : z1 ≥ 0}. We call a periodic orbit W quadratically stable if
(i) it has a Poincaré map P which has a fixed point corresponding to W and is two
times differentiable on the image of P 2 in a neighbourhood of its fixed point, and
(ii) the linearization of the Poincaré map in this fixed point is zero.
With these notations and arguments we can formulate a “linearized version” of
Theorem 2, which, due to the above arguments, implies Theorem 2.
Theorem 9 Let A ∈ Rn,n be a matrix that has eigenvalues with positive real part
and eigenvalues with negative real part but no eigenvalues on the imaginary axis. Let
v ∈ Rn be such that the pair (A, v) is controllable. Let τ > 0 be arbitrary. Then
there exists a smooth function g : Rn → R such that {z ∈ Rn : g(z) = 0} is a smooth
manifold that splits Rn into two simple domains and such that the differential equation
ẋ(t) = A [x(t) − v sgn g(x(t − τ ))] (21)
has a quadratically stable periodic orbit.
The spectral properties of A imply that, without the relay (v = 0), the origin is an
equilibrium of saddle type. Moreover, they imply that the statement of Theorem 9 is
equivalent to the corresponding statement for the truncated system (20) with ε = 0
because (A, v) is controllable if and only if (A, A−1 v) is controllable.
The main difference between the n-dimensional case of Theorem 9 and the
construction of W and G in section 5.1 is that the choice of Gτj tangential to
the outgoing flow eliminates only one dimension of the linearization. Thus, each
switching at one of the delayed switching manifolds Gτj acts as a projection with a
one-dimensional kernel on the linearization of P0 . This means that we have to find
a closed curve Ψ consisting of an even number m > n of segments Ψj , alternating
between the two flows and always following each of the flows for a time θj greater
than the delay τ . Subsequently, we have to verify that
(i) we can find a switching manifold that intersects each segment Ψj transversally
exactly once in a point of our choice (namely, in Ψ(θj − τ )), and
(ii) we can tilt the switching manifold locally in the intersection points with Ψ in a
manner such *that the concatenation of the projections induced by the switchings
cancels out all components of the linearization.
The existence of an appropriate closed curve and point (i) follow from the saddle
property of A, which implies that all trajectories that spend a long time near the
equilibrium approximately follow first the stable and then the unstable subspace of
A. The second point is implied by the controllability required in Theorem 9. The
detailed proof of Theorem 9 is given in Appendix C.
6. Discontinuity-induced bifurcations
This section discusses what happens generically to the dynamics near relay periodic
orbits that violate one of the transversality requirements, either Condition 5 or
Condition 6. To simplify our presentation we restrict ourselves in this section to the
practically most relevant case of slowly oscillating periodic orbits. We assume that
the general delayed relay system (3) depends on a parameter λ where the dependence
of f1 , f2 , g on x and λ and the dependence of τ on λ are smooth:
(
f1 (x(t), λ) if g(x(t − τ (λ)), λ) < 0,
ẋ(t) = (22)
f2 (x(t), λ) if g(x(t − τ (λ)), λ) ≥ 0.
Moreover, we assume that, for λ < 0, (22) has a slowly oscillating periodic orbit x̃(·, λ)
of uniformly bounded period p(λ), which satisfies the transversality conditions 5 and
6. Section 6.1 investigates the case of x̃(·, 0) violating Condition 5, section 6.2 studies
the case of x̃(·, 0) violating Condition 6. This study treats (22) and the periodic orbit
only at the parameter λ = 0. Thus, we can drop the parameter λ, which is always 0,
from our notation in the remainder of the section.
6.1. Corner collision

The graph x̃([−p, 0)) of the periodic orbit is a continuous piecewise smooth curve in
Rn . The violation of Condition 5 means that one of the corners (switching points) of
the curve x̃([−p, 0)) lies in the switching manifold {g = 0}. That is, at λ = 0 there
are two times s̃1 and s̃2 such that
s̃1 + τ = s̃2 , g(x̃(s̃1 )) = 0, g(x̃(s̃2 )) = 0, (23)
which violates Condition 5. The fact that x̃(·) is slowly oscillating for parameters
λ < 0 implies that s̃1 is the only intersection point of x̃(·) with {g = 0} in the interval
[s̃1 − τ, s̃2 ). Thus, x̃(·) follows exactly one flow in the time interval [s̃1 , s̃2 ] (say, Φ1
without loss of generality).
We assume that x̃ satisfies the following two secondary non-degeneracy conditions:
Condition 10 (secondary genericity conditions for corner collision)

(a) Both vector fields intersect the switching manifold transversally in x̃(s̃2 ). That
is, g ′ (x̃(s̃2 ))f1 (x̃(s̃2 )) 6= 0, and g ′ (x̃(s̃2 ))f2 (x̃(s̃2 )) 6= 0
(b) The local delayed switching manifold Gτ1 = Φ1 (τ ; {g = 0})∩U (x̃(s̃2 )) and the local
switching manifold G2 = {g = 0} ∩ U (x̃(s̃2 ))) are not tangent to each other in
x̃(s̃2 ). That is, ∂2 Φ1 (τ ; x̃(s̃1 ))g ′ (x̃(s̃1 ))T and g ′ (x̃(s̃2 ))T are linearly independent.
Let us choose as Poincaré section S the set of all z ∈ C([−τ, 0]; Rn ) with headpoint
z(0) ∈ Gτ1 . This is an admissible choice since Gτ1 intersects the incoming flow Φ1
transversally.
Condition 10(b) implies that Gτ1 and G2 intersect each other in the smooth local
manifold Gτ1 ∩ G2 of codimension 2, which contains x̃(s̃2 ). This intersection divides
Gτ1 into two parts, F− := Gτ1 ∩ {g < 0} and F+ = Gτ1 ∩ {g ≥ 0}.
The following lemma states that the return map for x̃ can still be expressed as an
(n − 1)-dimensional return map P0 to Gτ1 but that P0 is only piecewise smooth with,
in general, different derivatives in F− and F+ .
Lemma 11 (Return map for corner collision) The image of the local return
map P of the periodic orbit x̃(·) is contained in a (n − 1)-dimensional manifold that
can be parametrized by the elements of Gτ1 . On Gτ1 , P is described by a piecewise
smooth (n − 1)-dimensional map P0 which is smooth in F+ and F− . More precisely,
there exist linear maps A1 , A2 ∈ Rn×n such that the local return map P0 has the form
(
A1 x + O(kxk2 ) if x̃(s̃2 ) + x ∈ F− ,
P0 (x̃(s̃2 ) + x) = x̃(s̃2 ) + 2
(24)
A2 x + O(kxk ) if x̃(s̃2 ) + x ∈ F+
for all sufficiently small x ∈ Gτ1 − x̃(s̃2 ).
The first statement of Lemma 11 follows from the fact that all elements of S have
an image under P of the form Φ1 ([−τ, 0]; z0 ) where z0 ∈ Gτ1 . The piecewise linear
asymptotics of P0 comes, roughly speaking, from the fact that a trajectory through
x̃(s̃2 ) + x ∈ F+ spends a different time in {g ≥ 0} than a trajectory through F− . This
time difference is asymptotically linear in x.
The precise dependence of A1 and A2 on the right-hand-side is described in
Appendix D. There are three distinct cases (shown in a piecewise affine example
in Figure 4), giving rise to different expressions for A1 and A2 :
(a) g ′ (x̃(s̃2 ))f1 (x̃(s̃2 )) · g ′ (x̃(s̃2 ))f2 (x̃(s̃2 )) > 0, shown in figure 4(a). This case
corresponds to the situation where the periodic orbit x̃ intersects the switching
manifold {g = 0} transversally in x̃(s̃2 ) in the sense that all convex combinations
fc = cf1 (x̃(s̃2 )) + (1 − c)f2 (x̃(s̃2 )) (0 ≤ c ≤ 1) of the two tangent vectors at the
corner satisfy g ′ (x̃(s̃2 ))fc 6= 0. That is, fc points through the switching manifold
for all c ∈ [0, 1]. Figure 4(a) illustrates this configuration where the periodic orbit
intersects {g = 0} at x̃(s̃2 ).
(b),(c) g ′ (x̃(s̃2 ))f1 (x̃(s̃2 )) · g ′ (x̃(s̃2 ))f2 (x̃(s̃2 )) < 0. In both cases the periodic orbit lies
(locally near x̃(s̃2 )) entirely on one side of the switching manifold. This implies
that x̃(·) intersects (touches) the switching manifold an odd number of times. The
form of A1 and A2 depends also on the existence of another intersection x̃(s̃3 ) of
the periodic orbit x̃(·) with the switching manifold {g = 0} within in the interval
(s̃2 , t̃2 ) where t̃2 = s̃2 + τ . The difference between the cases (b) and (c), shown in
figure 4(b) and (c), is the order of x̃(s̃3 ) and x̃(t̃2 ) along the orbit x̃. The effect of
(a) {g < 0} (b) x̃(s̃3 ) (c) x̃(s̃3 )

{g ≥ 0}
{g ≥ 0}
x̃(t̃2 ) x̃(t̃2 )
x̃(s̃2 ) Gτ1
x̃(t̃2 ) x̃(t̃3 ) x̃(t̃3 ) Gτ1
x̃(s̃2 ) x̃(s̃2 )
Gτ1
x̃(s̃1 )
{g = 0} x̃(s̃1 )
x̃(s̃1 )
{g = 0}
{g ≥ 0} {g < 0} {g < 0} {g = 0}
Figure 4. Illustration of a periodic orbit x̃ undergoing a corner collision at x̃(s̃2 ),

showing the three different possible configurations. The times s̃j (j = 1, 2, 3) are
the moments when x̃ intersects the switching manifold {g = 0}. The times t̃j are
the corresponding switching times (t̃j = s̃j + τ , t̃1 = s̃2 ). The manifold Gτ1 is the
time-τ image of {g = 0} under Φ1 . The return map P0 , discussed in Lemma 11,
maps the local manifold Gτ1 back to itself. The dashed part of x̃ follows Φ1 , the
dotted part follows Φ2 .
this difference on the dynamics becomes clear if both flows are linearly dependent
in the delayed switching point x̃(t̃2 ) (that is, f1 (x̃(t̃2 )) and f2 (x̃(t̃2 )) are linearly
dependent). Then the linearization of the return map will be continuous for case
(b) but, in general, it will still be discontinuous for case (c). Case (c) is the most
complex scenario because the discontinuity is affected by the configuration at four
different points along the orbit: at x̃(s̃2 ), x̃(s̃3 ), x̃(t̃2 ), and x̃(t̃3 ).
The dynamics of piecewise asymptotically linear maps have been studied in [13, 14],
also classifying possible bifurcations when the parameter λ unfolds the degeneracy
transversally. Thus, Lemma 11 links the study of the infinite-dimensional delayed
relay system to the bifurcation theory of piecewise smooth asymptotically linear finite-
dimensional maps.
6.2. Tangential grazing

The violation of Condition 6 means that there exists a time s∗ when the periodic orbit
grazes (touches) the switching manifold {g = 0} tangentially, that is, g(x̃(s∗ )) = 0
˙ ∗ ) = 0. Let us denote the transversal switching times along x̃ by s̃j
and g ′ (x̃(s∗ ))x̃(s
(j = 1, . . . , m where m is even). We assume that x̃(·) satisfies the following secondary
non-degeneracy conditions:
Condition 12 (secondary genericity conditions for tangential grazing)
(a) The orbit x̃ is quadratically tangent to the switching manifold {g = 0} in s∗ , and

not to a higher order. That is,
1 d2 1h ′ ˙ ∗ ) 2 6= 0.
¨ ∗ ) + g ′′ (x̃(s∗ )) x̃(s
i
q := g(x̃(t)) = g (x̃(s∗ ))x̃(s
2 [dt]2 t=s∗ 2
(b) The time when the tangency is noticed along the orbit x̃ does not coincide
with another crossing of the switching manifold. That is, g(x̃(t∗ )) 6= 0 where
t∗ = s∗ + τ .
(c) The grazing does not coincide with a simultaneous violation of Condition 5 (a
corner collision). That is, g(x̃(s∗ − τ )) 6= 0. Hence, s∗ 6= t̃j (j = 1, . . . , m) where
t̃j = s̃j + τ .
The periodic orbit x̃ is slowly oscillating for parameter λ < 0. Thus, s∗ lies in an
interval [a, b] which is longer than the delay τ (that is, b − a > τ ) where x̃ follows one
flow. Without loss of generality, let us assume that x̃([a, b]) follows Φ1 . We choose as
Poincaré section S the set of all z ∈ C([−τ, 0]; Rn ) with headpoint z(0) ∈ G where G
is a hyperplane intersecting x̃ transversally at time t̃0 = (a + b + τ )/2. The following
lemma describes the local return map P to the Poincaré section S to leading order.
Lemma 13 (Return map for tangential grazing) The image of the local return
map P to the Poincaré section S is contained in a (n − 1)-dimensional manifold that
can be parametrized by the elements of the affine hyperplane
˙ ∗ )T [x − x̃(s∗ )] = 0}.
F0 := {x : x̃(s
On F0 , P is described by a piecewise smooth (n − 1)-dimensional map P0 : F0 7→ F0 .
There exists a smooth function m : U (x̃(s∗ )) → R such that the map P0 is smooth in
F+ = F0 ∩ {x : m(x) > 0} and F− = F0 ∩ {x : m(x) < 0}. For small x ∈ F0 − x̃(s∗ )
the map P0 has the form
Ax + O(kxk2 )
(
if x̃(s∗ ) + x ∈ F+ ,
P0 (x̃(s∗ ) + x) = x̃(s∗ ) + p (25)
v −m(x̃(s∗ ) + x) + O(kxk) if x̃(s∗ ) + x ∈ F−
where A ∈ Rn×n and v ∈ F0 − x̃(s∗ ) ⊂ Rn .
The expansion of the function m in x̃(s∗ ) is
m(x̃(s∗ ) + x) = q −1 g ′ (x̃(s∗ ))x + O(kxk2 ).
This implies that the return map of all trajectories near x̃(·) that intersect F− expands
to lowest order like a square root. The first statement of Lemma 13 follows from
the fact that all elements of S will have an image under P which has the form
Φ1 ([−τ, 0]; z0 ) where z0 ∈ G. This reduces the Poincaré map P to a return map
to the hyperplane G ⊂ Rn . Since both hyperplanes F and G are transversal to x̃,
return maps to G and to F are conjugate to each other under the local diffeomorphism
obtained by following the flow Φ1 from F to G.
The function m(x) used in Lemma 13 is defined as the local minimum of the
parabola-shaped function q −1 g(Φ1 (·; x)) near 0. This local minimum is uniquely
defined and depends smoothly on x. The square-root asymptotics of P0 arises, roughly
speaking, from the fact that the time which a trajectory through x̃(s̃2 )+x ∈ F− spends
in {x : m(x) < 0} depends asymptotically linearly on the square root of −m(x).
The precise dependence of A and v on the right-hand-side is described in detail in
Appendix E. Figure 5 illustrates the two different cases that can arise. The difference
between the two cases is that in case (a) x̃ does not cross the switching manifold
{g = 0} between s∗ and t∗ , whereas in case (b) there is an intermediate crossing at
x̃(s̃2 ). Both cases have two sub-cases depending on the existence of the intermediate
switching at t̃1 between s∗ and t∗ , but those cause only minor differences. Case
(b) is more complex because the discontinuity of the return map is affected by the
(a) {g ≥ 0} (b) {g < 0}
x̃(t∗ )
x̃(t̃2 ) x̃(t̃2 )
F0 F0 {g = 0}
x̃(s̃1 )
x̃(s∗ ) x̃(s∗ )
x̃(s̃2 ) x̃(s̃1 ) x̃(s̃2 )
{g = 0}
x̃(t̃1 ) x̃(t̃1 )
x̃(t∗ )
{g < 0} {g ≥ 0}
Figure 5. Illustration of a periodic orbit x̃ undergoing a tangential grazing at

x̃(s∗ ), showing the two different possible configurations. The times s̃j (j = 1, 2)
are the moments when x̃ intersects the switching manifold {g = 0} transversally.
The times t̃j are the corresponding switching times (t̃j = s̃j + τ ). At the point
x̃(t∗ ) the system notices the grazing at s∗ . The return map P0 , discussed in
Lemma 13, maps the local manifold F0 (defined as being orthogonal to x̃(s ˙ ∗ ))
back to itself. The dashed part of x̃ follows Φ1 , the dotted part follows Φ2 .
configurations near four points along the periodic orbit: x̃(s∗ ), x̃(s̃2 ), x̃(t∗ ) and x̃(t̃2 ).
A special case of type (a) is a periodic orbit of period larger than the delay τ that has
no transversal intersections with the switching manifold {g = 0}.
Lemma 13 allows one to link phenomena occurring close to a grazing periodic
orbit in a delayed relay system to the bifurcation theory of piecewise smooth maps
with square-root asymptotics on one side of the discontinuity. The general results in
[15, 16] classify the dynamics for maps of this type.
6.3. The dynamics near grazing bifurcations in the small-delay limit — illustrating
example
The occurrence of square-root terms in return maps as in Lemma 13 is typical
for impacting systems in the vicinity of periodic orbits with slow-velocity impacts
rather than ordinary differential equations with discontinuous right-hand-side (that is,
systems such as (3) with τ = 0, so-called Filippov systems [14]). A consequence of this
fact is that the dynamics of system (3) can change dramatically by changing τ from 0 to
a small positive value. The reason behind this change is that codimension-one grazing
events of periodic orbits generically induce C 1 -smooth or piecewise asymptotically
linear return maps for Filippov systems, in contrast to impacting systems, or the case
of (3) with a positive delay. As an illustrative example we consider the system in R2

 0 −1 x − a · (kxk − λ)

if x1 (t − τ ) − 1 < 0,
ẋ = 1 0 (26)

 f (x)
2 if x1 (t − τ ) − 1 ≥ 0
where a > 0, λ > 0, k · k is the Euclidean norm in R2 , and f2 (x0 ) = (−b, 0) at
x0 = (1, 0) with b > 0. The switching function g is g(x) = x1 − 1. This system
(a) (b)
τ =0
0
P0 (y)
x2
0 y
x1 x0
(c)
0<τ ≪1
0
O(τ )
P0 (y)
{x1 = 1}
0 y
Figure 6. Illustration of the configuration for the periodic orbit of (26). Panel (a)
shows the phase portrait for the grazing periodic orbit. The dashed trajectories
correspond to flow Φ1 with its stable limit cycle, the dotted arrow shows f2 (x0 ).
Panels (b) and (c) show the asymptotics of the local return map P0 to {x2 = 0}
for y = x1 − 1 for small y, and delay τ = 0 (b) or small delay (c).
has a stable limit cycle (x̃1 (t), x̃2 (t)) = (λ cos(t), λ sin(t)) if λ < λ0 = 1. At the
parameter value λ = 1 the periodic orbit x̃(·) grazes tangentially the switching line
{x1 = 1} in x0 . Figure 6 illustrates this situation in panel (a). For τ = 0 the
orbit continues to exist also for λ ≥ 1 (λ ≈ 1), changes its shape continuously and
√ 1 < λ ≪ 2 the orbit slides along the line {x1 = 1} from
remains√stable. In fact, for
x2 = − λ2 − 1 to x2 = λ2 − 1 (due to f2 (x0 ) pointing toward the sliding line in
the grazing point). Thus, at the grazing its only non-trivial Floquet multiplier jumps
from c = exp(−4πa) ∈ (0, 1) for λ < 1 to 0 for λ > 1.
If, however, τ is small but positive the return map P0 to the line segment x2 = 0,
x1 ∈ [1 − δ, 1 + δ] has the form described in Lemma 13 for λ = 1. Specifically,
introducing the variable y = x1 − 1,
(
cy if y < 0
P0 (y) = √ (27)
−d y + O(|y|) if 0 ≤ y ≪ 1
where d is a positive factor close to 2bc. The Poincaré map of the grazing periodic
orbit is depicted in figure 6(c), comparing it to the return map without delay in panel
(b). The expression in (27) captures only the first square root branch of the return
map of height O(τ ). The dynamics of maps with square-root asymptotics has been
studied by [15, 16]. A consequence of the results of [16] is that, if exp(−4πa) > 2/3, for
any given τ > 0 the system has a chaotic attractor for all λ in an interval (1, λmax (τ )).
This sudden transition to chaos by an introduction of an arbitrarily small delay is
fundamentally different from the behaviour of smooth systems. If one introduces a
small delay in one of the arguments of a smooth system of ODEs the delay acts as
a regular perturbation parameter, preserving, for example, hyperbolic equilibria or
periodic orbits without changing their stability [37].
7. Conclusion
The paper considered the dynamics of dynamical systems with delayed relays in the
vicinity of periodic orbits. First, we found that the dynamics can be described
generically by low-dimensional local return maps, even though the phase space of
the original system is infinite-dimensional. Generically these return maps are smooth.
Specifically, we provided two sufficient genericity conditions on the periodic orbit that
guarantee the smoothness and finite-dimensionality of the local return map.
We exploited the existence and form of these local return maps to show that relays
can be used to design simple static feedback laws that are able to stabilize saddle-type
equilibria to nearby periodic orbits even in the presence of arbitrarily large delays.
Finally, we studied the two most common bifurcations that occur when one of
the genericity conditions is violated: the corner collision and the tangential grazing.
They give rise to piecewise smooth local return maps. These return maps are either
piecewise asymptotically linear (corner collision) or have square-root asymptotics on
one side (grazing). The reduction to piecewise smooth maps provides a link to well-
established results of the bifurcation theory for these types of maps [13, 14, 15, 16].
It also shows that the small-delay limit for relay systems is more subtle than the
corresponding limit for smooth DDEs.
The main open problem concerning the bifurcation theoretic part of our studies
is that the secondary non-degeneracy conditions, even though they are genericity
conditions, are often not fulfilled in practice. Typically, symmetric periodic orbits of
piecewise linear systems of the form ẋ = Ax − vsgn[bT x(t − τ )] violate the secondary
non-degeneracy conditions formulated in the sections 6.1 and 6.2 whenever they violate
the primary conditions 5 or 6. This gives rise to much more degenerate bifurcation
scenarios in the systems studied in [24, 25].
A caveat of the stabilizability result in Theorem 9 is that the basin of attraction
of the quadratically stable periodic orbit shrinks not only for increasing τ but also for
decreasing amplitude of the orbit (which is related to the size of ε). A possible solution
to this problem are the more general hybrid feedback control law. For example, a
hybrid feedback for the inverted pendulum would give rise to a dynamical system of
the form

 1
 if x(t − τ ) ∈ D+ ,
ẍ = x − α where α = 0 if x(t − τ ) ∈ D0 ,

−1 if x(t − τ ) ∈ D− ,

where D+ ∪ D0 ∪ D− is a partition of the physical space Rn . This type of hybrid

control could potentially allow one to decrease the amplitude of the periodic motion
without shrinking its basin of attraction. Moreover, one can choose this partition such
that the time spent in D0 by the periodic orbit is arbitrarily close to p, the period of
the orbit. This means that the relay control could be switched off most of the time.
Another limitation of the stabilizability result in Theorem 9 is its restriction to
matrices of saddle-type. A modification of the arguments given in the illustration in
section 5.1 and the general proof in Appendix C enables one to extend the result to
completely unstable matrices in certain cases (for example, the case of a matrix with
two different real positive eigenvectors).
Acknowledgments
The research of J.S. is supported by EPSRC grant GR/R72020/01. The author

thanks his colleagues Bernd Krauskopf, Piotr Kowalczyk and David Barton for
fruitful discussions and, in particular, Bernd Krauskopf for his help on improving
the presentation of the material.
[1] M. Garcia, A. Chatterjee, and A. Ruina. Efficiency, speed, and scaling of passive dynamical
bipedal walking. Dynamics and Stability of Systems, 15(2):75–99, 2000.
[2] F. Moss and J.G. Milton. Balancing the unbalanced. Nature, 425:911–912, 2003.
[3] H. Kwakernaak and R. Sivan. Linear Optimal Control Systems. Wiley Interscience, New York,
1972.
[4] G. Stépán. Retarded Dynamical Systems: Stability and Characteristic Functions. Longman
Scientific and Technical, 1989.
[5] F.M. Atay. Balancing the inverted pendulum using position feedback. Appl. Math. Letters,
12:51–56, 1999.
[6] J. Sieber and B. Krauskopf. Extending the permissible control loop latency for the controlled
inverted pendulum. Dynamical Systems: An International Journal, 20(2):189–199, 2005.
[7] W. Michiels and D. Roose. Limitations of delayed state feedback: a numerical study.
International Journal of Bifurcation and Chaos, 12(6):1309–1320, 2002.
[8] M. Landry, S.A. Campbell, K. Morris, and C. Aguilar. Dynamics of an inverted pendulum with
delayed feedback control. Preprint, University of Waterloo, 2003. submitted.
[9] J. Sieber and B. Krauskopf. Bifurcation analysis of an inverted pendulum with delayed feedback
control near a triple-zero eigenvalue. Nonlinearity, 17(1):85–104, 2004.
[10] L. Fridman, E. Fridman, and E. Shustin. Steady modes and sliding modes in relay control
systems with delay. In J. P. Barbot and W. Perruguetti, editors, Sliding Mode Control in
Engineering, pages 264–295. New York, 2002.
[11] L. Fridman, V. Strygin, and A. Polyakov. Stabilization of amplitude of oscillations via relay
delay control. Int. J. Control, 76(8):770–780, 2003.
[12] O. Diekmann, S. van Gils, S.M. Verduyn Lunel, and H.-O. Walther. Delay Equations, volume
110 of Applied Mathematical Sciences. Springer-Verlag, 1995.
[13] S. Banerjee and C. Grebogi. Border collision bifurcations in twodimensional piecewise smooth
maps. Phys. Rev. E, 59:4052–4061, 1999.
[14] M. di Bernardo, M.I. Feigin, S.J. Hogan, and M.E. Homer. Local analysis of C-bifurcations in
n-dimensional piecewise smooth dynamical systems. Chaos, Solitons and Fractals, 10:1881–
1908, 1999.
[15] A.B. Nordmark. Universal limit mapping in grazing bifurcations. Phys. Rev. E, 55:266–270,
1997.
[16] H. Nusse, E. Ott, and J. Yorke. Border collision bifurcations: an explanation for observed
bifurcation phenomena. Phys. Rev. E, 49:1073–1076, 1994.
[17] J. Gonçalves, A. Megretzki, and M.A. Dahleh. Global stability of relay feedback systems. IEEE
Transactions on Automatic Control, 46(4):550–562, 2001.
[18] K.H. Johansson, A.E. Barabanov, and K.J. Åström. Limit cycles with chattering in relay
feedback systems. IEEE Transaction on Automatic Control, 47(9):1414–1423, 2002.
[19] K. Popp and P. Stelter. Stick-slip vibrations and chaos. Philosophical Transactions: Physical
Sciences and Engineering, 332(1624):89–105, 2000.
[20] B. Brogliato. Impacts in mechanical systems — analysis and modelling, volume 551 of Lecture
notes in physics. Springer-Verlag, 2000.
[21] S. Banerjee and G.C. Verghese. Nonlinear phenomena in power electronics: attractors,
bifurcations, chaos, and nonlinear control. IEEE Press, New York, 2001.
[22] J.-L. Gouzé and T. Sari. A class of piecewise linear differential equations arising in biological
models. Dynamical systems, 17:299–316, 2003.
[23] D.A.W. Barton, B. Krauskopf, and R.E. Wilson. Explicit periodic solutions in a model of a
relay controller with delay and forcing. Nonlinearity, 18(6):2637–2656, 2005.
[24] D.A.W. Barton, B. Krauskopf, and R.E. Wilson. Periodic solutions and their bifurc-
ations in a non-smooth second-order delay differential equation. Preprint 2005.15,
University of Bristol, Bristol Centre for Applied Nonlinear Mathematics, 2005.
http://www.enm.bris.ac.uk/anm/preprints/2005r15.html.
[25] W. Bayer and U. an der Heyden. Oscillation types and bifurcations of a nonlinear second-order
differential-difference equation. J. Dynam. Diff. Eq., 10(2):303–326, 1998.
[26] U. Holmberg. Relay feedback of simple systems. PhD thesis, Lund Institute of Technology,
1991.
[27] H.-O. Walther. Contracting return maps for monotone delayed feedback. Discrete and
Continuous Dynamical Systems, 7(2):259–274, 2001.
[28] Y.-H. Roh and J.-H. Oh. Robust stabilization of uncertain input-delay systems by sliding mode
control with delay compensation. Automatica, 35:1861–1865, 1999.
[29] Q.G. Wang, T.H. Lee, and K.K. Tan. Finite spectrum assignment for time-delay systems. In
Lecture notes in control and information sciences, volume 239. Springer Verlag, New York,
1999.
[30] S. Mondié and W. Michiels. Finite spectrum assignment of unstable time-delay systems with a
safe implementation. IEEE Transactions on Automatic Control, 48(12):2207–2212, 2003.
[31] T. Horiuchi and T. Konno. A new method for compensating actuator delay in real-time hybrid
experiments. Philosophical Transactions of the Royal Society A, 359:1893–1909, 2001.
[32] M.I. Wallace, D.J. Wagg, and S.A. Neild. An adaptive polynomial based forward prediction
algorithm for multi-actuator real-time dynamic substructuring. Accepted for publication in
Proceedings of the Royal Society of London A.
[33] J.K. Hale and S.M. Verduyn Lunel. Introduction to Functional Differential Equations, volume 99
of Applied Mathematical Sciences. Springer-Verlag, 1993.
[34] A. Pazy. Semigroups of Linear Operators and Applications to Partial Differential Equations.
Applied mathematical Sciences. Springer Verlag, New York, 1983.
[35] J. Sieber. Dynamics of delayed relay control systems with large delays. In Proceedings of the
5th IFAC Workshop on Time-Delay Systems, Leuven, Belgium, 2004.
[36] E. Shustin, E. Fridman, and L. Fridman. Stable oscillations in a discontinuous delay system
of the second order. Discrete Contin. Dynam. Systems, (Added Volume II):209–233, 1998.
Dynamical systems and differential equations, Vol. II (Springfield, MO, 1996).
[37] C. Chicone. Inertial and slow manifolds for delay equations with small delays. Journal of
Differential Equations, 190:364–406, 2003.
Appendix A. Proof of Lemma 4
It is sufficient to prove the continuity of E(T ; ·) in x̃t0 for T ∈ (0, τ ) since x̃t0 lies on
a periodic orbit. The set of all roots of g(x̃t0 (θ)) within [−τ, −τ + T ] is a subset of
{t0 + s̃1 , . . . , t0 + s̃m }. Let us denote these time points by rj (j = 1, . . . , q, q ≤ m) in
ascending order: −τ ≤ r1 < . . . < rq ≤ T − τ . We define the constant
C = max [kf1 (x̃(θ))k + kf2 (x̃(θ))k]
θ∈R
which is bounded since x̃ is periodic and the functions fj are Lipschitz continuous.
Let ε > 0 be small enough such that exp(4εL) < 2 where L is a Lipschitz constant for
f1 and f2 . In order to prove continuity it is sufficient to find a δ > 0 such that
q+1
kE(T ; ξ) − E(T ; x̃t0 )k < (8C + 1) 2eLT ε (A.1)
for all ξ ∈ C([−τ, 0]; Rn ) satisfying kξ − x̃t0 k < δ in the (maximum) norm of
C([−τ, 0]; Rn ). We choose δ ∈ (0, ε) such that all ξ ∈ C([−τ, 0]; Rn ) with kξ − x̃t0 k < δ
meet the following condition:
Sq g(ξ(θ)) is nonzero and has the same sign as g(x̃t0 (θ))
for all θ ∈ [−τ, T − τ ] \ j=1 (rj − ε, rj + ε). That is, for all ξ in the δ-neighbourhood
of x̃t0 , g(ξ(·)) can have zeroes only in the vicinity of the zeroes of g(x̃t0 (·)).
Let ξ ∈ C([−τ, 0]; Rn ) be such that kξ− x̃t0 k < δ. Since E(T ; ·)(θ) = E(0; ·)(θ+T )
for θ ∈ [−τ, −T ], we have
kE(T ; ξ)(θ) − E(T ; x̃t0 )(θ)k < δ < ε for θ ∈ [−τ, −T ]. (A.2)
If θ ∈ (−T, 0] then E(T ; ·)(θ) = E(θ + T ; ·)(0). Consequently, we have to focus on the
evolution of the difference between the headpoints, x(t) := E(t; ξ)(0) and x̃(t0 + t),
∆(t) := kx(t)
Sq − x̃(t0 + t)k, for t ∈ [0, T ]. Inequality (A.2) implies that |∆(0)| < ε. If t
is not in j=1 (τ + rj − ε, τ + rj + ε) then both headpoints follow the same flow (either
Φ1 or Φ2 ). Thus, we have the following set of recursive inequalities for the evolution
of ∆(t) in the intervals (τ + rj + ε, τ + rj+1 − ε):
∆(t) < eLt ∆(0) < eLT ε if 0 ≤ t ≤ τ + r1 − ε,
∆(t) < eL(t−(τ +rj +ε)) ∆(τ + rj + ε)
< eLT ∆(τ + rj + ε) if τ + rj + ε ≤ t ≤ τ + rj+1 − ε (j = 1, . . . , q − 1), (A.3)
L(t−(τ +rq +ε))
∆(t) < e ∆(τ + rq + ε)
LT
<e ∆(τ + rq + ε) if τ + rq + ε ≤ t ≤ T .
The variation-of-constants formula (7) implies an estimate on how ∆(t) evolves in the
intervals (τ + rj − ε, τ + rj + ε). Let t1 , t2 be in [0, T ] ∩ (τ + rj − ε, τ + rj + ε) for
some j and t1 < t2 :
∆(t2 ) ≤ ∆(t1 ) + kx(t2 ) − x(t1 )k + kxt0 (t2 ) − xt0 (t1 )k
Z t2
≤ ∆(t1 ) + kf1 (x(s))k + kf2 (x(s))kds+
t1
Z t2
+ kf1 (xt0 (s))k + kf2 (xt0 (s))kds
t1
t2 t2
(A.4)
Z Z
≤ ∆(t1 ) + 2 kf1 (xt0 (s))k + kf2 (xt0 (s))kds + 2L ∆(s)ds
t1 t1
Z t2
≤ ∆(t1 ) + 2(t2 − t1 )C + 2L ∆(s)ds
t1
2L(t2 −t1 )
≤ [∆(t1 ) + 2C(t2 − t1 )] e
≤ 2∆(t1 ) + 4C(t2 − t1 )
The recursion of inequalities (A.3) and estimate (A.4) (where always t2 − t1 < 2ε)
allow for a global estimate of ∆(t) for t ∈ [0, T ]:
∆(t) ≤ eLT 8C + 8C 2eLT + . . . + 8C(2eLT )q−1 + (2eLT )q ε

q+1 (A.5)
≤ (8C + 1) 2eLT ε
The inequalities (A.2) and (A.5) combined imply the validity of the estimate (A.1) for
the whole maximum norm of the function E(T ; x) − E(T ; xt0 ).
Appendix B. Proof of Lemma 7
Let δ > 0 be such that all intervals (−p, −p + δ), (s̃k − δ, s̃k + δ), (t̃k − δ, t̃k + δ) and
(−δ, 0) are disjoint (k = 1, . . . , m). This is possible due to Condition 5 on x̃.
Let s̃k (k ∈ {1, . . . , m}) be one of the zeroes of g(x̃(·)) in (−p, 0). Due to
Condition 5 the periodic orbit x̃ follows one of the flows in s̃k , say Φj . Because
of the transversality of Φj with {g = 0} in x̃(s̃k ) (Condition 6) there exists a εk > 0
such that the function t → g(Φj (t; z)) changes its sign and has exactly one zero in
(−δ, δ) for all z ∈ Rn with kz − x̃(s̃k )k < εk . Furthermore, for sufficiently small δ there
exists a ε0 such that the function t → l0T [Φ1 (t; z) − x̃(0)] has exactly one regular zero
in (−δ, δ) for all z ∈ Rn with kz − x̃(0)k < ε0 . We define ε := min{εk : k = 0 . . . , m},
which is larger than zero.
Let the open neighbourhood N ⊂ S of x̃0 be sufficiently small such that for all
ξ ∈ N the following three Conditions B0–B2 are satisfied:
B0 on the history interval,

Sm g(ξ(t)) is non-zero and has the same sign as g(x̃(t)) for
all t ∈ [−τ, 0] \ [ k=1 (s̃k − δ, s̃k + δ)],
and the headpoint trajectory x(·) = E(·; ξ)(0) ⊂ Rn satisfies
B1 g(x(t)) is non-zero and has the same sign as g(x̃(t)) for all t ∈ / (s̃k +lp−δ, s̃k +lp+δ)
(l = 1, 2, k = 1, . . . , m),
B2 kx(s̃k + lp) − x̃(s̃k )k < ε for all l = 1, 2, k = 1, . . . , m, and kx(2p) − x̃(0)k < ε.
A neighbourhood N satisfying the conditions B0–B2 above exists due to Condition 5,
because S E is continuous in x̃, and because g(x̃(·)) has no zeroes in the compact set
m
[−p, 0] \ k=1 (s̃k − δ, s̃k + δ).
Let ξ ∈ N be arbitrary, and x(·) = E(·; ξ)(0) be its headpoint trajectory. Let
s̃k (k ∈ {1, . . . , µ}) be one of the zeroes of g(x̃(·)) in [−p, 0] corresponding to the
switching time t̃k and denote by Φj the flow that x̃ is following in s̃k . Due to the
conditions B0 and B1, x(·) also follows Φj in the intervals (s̃k + p − δ, s̃k + p + δ) and
(s̃k + 2p − δ, s̃k + 2p + δ). Since kx(s̃k + p) − x̃(s̃k )k < ε and kx(s̃k + 2p) − x̃(s̃k )k < ε,
g(x(·)) changes its sign and has exactly one zero in each of the intervals (s̃k + p −
δ, s̃k + p + δ) and (s̃k + 2p − δ, s̃k + 2p + δ) due to condition B2.
Consequently, g(x(·)) has exactly µ zeroes s1 , . . . , sµ ∈ [2p − 2τ, 2p − τ ] (the same
number as g(x̃(·))), and x(·) follows Φ1 in the time interval (2p − δ, 2p + δ) (as does
x̃(·)). Thus, due to condition B2, x(·) also has a unique transversal intersection with
the plane {x̃(0) + Lv : v ∈ Rn } at a time t0 in a point x̃(0) + Lv. Then this v gives
rise to the form (14) for P 2 x where tk = (sk + τ ) − t0 (k = 1, . . . , µ).
Appendix C. Proof of Theorem 9 in section 5
The proof of Theorem 9 requires several steps which we will follow through in the form
of several lemmas. Let us denote by Φ1 the flow corresponding to ẋ = A(x + v) and by
Φ2 the flow corresponding to ẋ = A(x − v) (following the notation of the section 3).
We observe that Φ1 and Φ2 are symmetric to each other with respect to rotation by
π in the origin, that is,
Φ2 (t; z) = −Φ1 (t; −z) (C.1)
n
for all z ∈ R . The flow Φ1 can be expressed as an affine map
Φ1 (t; z) = exp(At)z + [exp(At) − I]v (C.2)
n
for z ∈ R and t ∈ R. The equilibrium of the flow Φ1 is at −v and is of saddle type.
There exist nonzero invariant projections P+ and P− corresponding to the stable (P− )
and unstable (P+ ) eigenspaces of A such that P− + P+ = I. Let us assume (without
loss of generality) that the basis of Rn is chosen such that kP± k = 1 and, for certain
constants K2 > K1 > 0, the dichotomy inequalities
exp(K2 t)kP+ zk ≥ kP+ exp(At)zk ≥ exp(K1 t)kP+ zk
(C.3)
exp(−K1 t)kP− zk ≥ kP− exp(At)zk ≥ exp(−K2 t)kP− zk
hold for all t ∈ R and z ∈ Rn in the original Euclidean norm of Rn .
Let m be an even number greater than n + 1. We now construct a g that gives
rise to a slowly oscillating periodic orbit intersecting the switching manifold {g = 0}
transversally m times. More precisely, the periodic orbit switches m times between
the two flows Φ1 and Φ2 and the time between successive switches is always greater
than the delay τ of the switch.
Lemma 14
Let δ > 0 be sufficiently small and denote by Bδ := {z ∈ Rn : kP− z − P− vk <
δ and kP+ z + P+ vk < δ}. Then there exist m-tuples (θ1 . . . , θm ) ∈ Rm and
(x1 , . . . , xm ) ∈ (Rn )m such that
(i) θj > τ for all j = 1, . . . , m,
(ii) xj ∈ Bδ for all j = 1, . . . , m,
(iii) xj+1 = −Φ1 (θj ; xj ) for j = 1, . . . , m − 1 and x1 = −Φ1 (θm ; xm ),
(iv) the curves Φ1 ([0, θj ]; xj ) for j = 1, . . . , m are mutually disjoint, and,
(v) using the notation rj = θj + . . . + θm (j = 1, . . . , m), the n vectors x1 , v,
exp(rm A)v,. . . , exp(rm−n+3 A)v are linearly independent.
Remark: We often use sets indexed by j = m − n + 3, . . . , m throughout this section
(for example in point (v) above). For n = 2 these sets are meant to be empty.
Proof: Let θ1 , . . . , θm be larger than τ . A tuple (x1 , . . . , xm ) ∈ (Rn )m has
property (iii) if and only if it satisfies the linear system of equations
xj+1 = − [exp(Aθj )xj + (exp(Aθj ) − I)v] for j = 1 . . . , m − 1,
(C.4)
x1 = − [exp(Aθm )xm + (exp(Aθm ) − I)v] .
Using the invariant projections P+ and P− we can split (C.4) into an equivalent pair
of systems of equations for P− xj and P+ xj (j = 1, . . . , m):
P− xj+1 = P− v − exp(Aθj )(P− xj + P− v) for j = 1 . . . , m − 1,
P− x1 = P− v − exp(Aθm )(P− xm + P− v),
(C.5)
P+ xj = −P+ v − exp(−Aθj )(P+ xj+1 − P+ v) for j = 1 . . . , m − 1,
P+ xm = −P+ v − exp(−Aθm )(P+ x1 − P+ v)
where the recursion for P+ xj follows from (C.4) after premultiplication with
exp(−θj A). If the θ1 , . . . , θm are sufficiently large, the linear system (C.5) is a small
perturbation of the regular system
P− xj = P− v for j = 1 . . . , m,
(C.6)
P+ xj = −P+ v for j = 1 . . . , m,
due to the dichotomy inequalities (C.3). Consequently, we can find a θ0 > τ such that,
for any tuple (θj )m
j=1 of numbers greater than θ0 , the perturbed system (C.5) (and,
hence, (C.4)) is uniquely solvable and such that its solution has a distance less than
δ from the solution of the unperturbed system (C.6). Thus, for any tuple (θj )m j=1 of
numbers greater than θ0 we find a unique tuple (xj )m j=1 that meets the properties (ii)
and (iii) in the lemma. For a sufficiently small δ let the time Tδ be bigger than
sup{t ≥ 0 : Φ1 (t; Bδ ) ∩ Bδ 6= ∅} + sup{t ≥ 0 : Φ1 (t; −Bδ ) ∩ −Bδ 6= ∅}.
Both summands are finite since −v (the equilibrium of Φ1 ) is neither in Bδ nor in
−Bδ due to the controllability of the pair (A, v). In fact, Tδ becomes smaller when δ
gets smaller. If the tuple (θj )mj=1 is chosen such that each two members of the tuple
differ by more than Tδ then all curves Φ1 ([0, θ1 ]; x1 ),. . . , Φ1 ([0, θm ]; xm ) are mutually
disjoint. If, in addition, all θj (j = 1, . . . , m) are greater than θ0 then the assertions
(ii)–(iv) of the lemma are satisfied simultaneously.
Let us finally adapt the tuple (θj )m m
j=1 (and, thus, simultaneously (xj )j=1 defined
by (C.4)) further to achieve property (v). Due to the controllability of the pair (A, v)
we can find, for any (m − 1)-tuple (θj )m j=2 , a (m − 1)-tuple nearby such the set
Σ = {v, exp(rm A)v, . . . , exp(rm−n+3 A)v}
of n − 1 vectors spans a (n − 1)-dimensional space (where rj = θj + . . . + θm ). The

elements of Σ appear in assertion (v) in addition to x1 . We solve system (C.4) to
obtain the representation
x1 = [I − exp(r1 A)]−1 [exp(r1 A)v − 2 exp(r2 A)v +−+. . .− 2 exp(rm A)v + v] (C.7)
for the point x1 . The matrix I − exp(r1 A) is regular because A has no eigenvalues
on the imaginary axis. It remains to be shown that we can find an arbitrarily small
adjustment of the m-tuple (θ1 , . . . θm ) (or, equivalently, (r1 , . . . , rm )) that makes x1
independent of the span of Σ.
We observe that r2 can be varied without changing Σ. The vector x1 has the
form x1 = M [w − 2 exp(r2 A)v] where w ∈ Rn and M = [I − exp(r1 A)]−1 is a regular
matrix. The controllability of the pair (A, v) implies the following: we can find a
n-tuple of times (t1 , . . . , tn ) such that all tj are close to r2 and such that the set of
vectors {w − 2 exp(t1 A)v, . . . , w − 2 exp(tn A)v} spans the whole Rn . Consequently,
M [w−2 exp(tj A)v] must be independent of the span of Σ (which is (n−1)-dimensional)
for one tj ≈ r2 . Thus, the small modification of the tuple (θ1 , . . . , θm ) such that r2
changes to tj and r1 , r3 ,. . . rm (and, thus, Σ and M ) remain unchanged guarantees that
x1 is independent of the span of Σ. Hence, with this small modification (θ1 , . . . , θm ),
along with the tuple (x1 , . . . , xm ) given by (C.4), also satisfies property (v).
Corollary 15 For tuples (θ1 , . . . θm ) and (x1 , . . . , xm ) having the properties (iii)
and (v) from Lemma 14 the sets {exp(rj A)xj , v, exp(rm A)v, . . . , exp(rm−n+3 A)v} are
linearly independent (using the notation rj = θj + . . . + θm for j = m − n + 3, . . . , m).
The statement of Corollary 15 follows recursively for j = m, m − 1, . . . , m − n + 3 from
system (C.4).
Due to the symmetry (C.1), the m-tuple (x1 , . . . , xm ) constructed in Lemma 14
defines a closed curve Ψ in Rn that follows trajectories of Φ1 and Φ2 in alternating
fashion and switches at xj (for odd j) from Φ2 to Φ1 and at −xj (for even j) from
Φ1 to Φ2 . We denote the smooth segments of the closed curve Ψ by Ψj for odd j and
−Ψj for even j:
Ψ = Ψ1 ∪ (−Ψ2 ) ∪ . . . Ψm−1 ∪ (−Ψm )
= Φ1 ([0, θ1 ]; x1 ) ∪ (−Φ1 ([0, θ2 ]; x2 )) ∪ . . .
∪ Φ1 [0, θm−1 ]; xm−1 ) ∪ (−Φ1 ([0, θm ]; xm )) (C.8)
= Φ1 ([0, θ1 ]; x1 ) ∪ Φ2 ([0, θ2 ]; −x2 ) ∪ . . .
∪ Φ1 [0, θm−1 ]; xm−1 ) ∪ Φ2 ([0, θm ]; −xm ).
The curve Ψ is continuous (by construction of (x1 , . . . , xm ); see Equation (C.4)) and
piecewise smooth. It can only be non-differentiable at its joints x1 , −x2 , x3 , . . . ,
xm−1 , −xm .
In the next step we show how to find, for sufficiently small δ, an appropriate
switching function g such that Ψ is a slowly oscillating periodic orbit of the differential
equation (21).
Lemma 16 Let δ > 0 be sufficiently small. Let the tuples (θ1 , . . . , θm ) and
(x1 . . . , xm ) be as constructed in Lemma 14 and the closed curve Ψ ⊂ Rn be as defined
in (C.8). Define the m points x̃j = Φ1 (θj − τ ; xj ) and let the vectors β̃j (j = 1, . . . , m)
be such that
β̃jT A[x̃j + v] 6= 0. (C.9)
Then there exists a smooth function g such that g(−v) < 0, g(v) > 0 and such
that {g = 0} is a manifold which partitions Rn into two simple domains and which
intersects Ψ transversally exactly once in each of its smooth segments. Moreover,
• for odd j, {g = 0} intersects the segment Ψj in x̃j and its tangential hyperplane
has the normal vector β̃j , and
• for even j, {g = 0} intersects the segment −Ψj in −x̃j and its tangential
hyperplane has the normal vector −β̃j
Property (i) in the construction of Lemma 14 guarantees that x̃j lies on Ψj for
j = 1, . . . , m). Condition (C.9) guarantees that the affine hyperplane attached to
x̃j with normal vector β̃j in x̃j is indeed transversal to the segment Ψj in x̃j .
Proof: It is sufficient to construct two simple domains, G1 for the flow Φ1
and G2 for the flow Φ2 , and a smooth boundary b separating them. Then, a smooth
function g can always be chosen such that the interior of G1 is {g < 0} and the closure
of G2 is {g ≥ 0}. For any sufficiently small ε > 0 we can choose a manifold b0 such
that (using the notation Bε (z) for the open ball of radius ε around z ∈ Rn )
• b0 has the form {z : β0T (z − x̃0 ) = 0} in the ball Bε (x̃0 ) where x̃0 = Φ1 (−τ ; P+ v −
P− v) and β0 is chosen such that b0 intersects all trajectories of Φ1 transversally
in Bε (x̃0 );
• b0 has the form {z : −β0T (z − x̃0 ) = 0} in the ball −Bε (x̃0 ), thus, b0 intersects all
trajectories of Φ2 transversally in −Bε (x̃0 );
• b0 has a positive distance from the ε-neighbourhoods of the affine subspaces
V1 = {z : P− (z + v) = 0} and V2 = {z : P− (z − v) = 0}
• the subspace V1 (including the equilibrium −v of Φ1 ) lies on one side of b0 and
the subspace V2 (including the equilibrium v of Φ2 ) lies on the other side of b0 .
The balls Bε (x̃0 ) and −Bε (x̃0 ) do not intersect with the ε-neighbourhoods of V1
and V2 for sufficiently small ε. The transversality of the intersection of b0 with
Φ1 ([−∞, 0]; P+ v − P− v) in x̃0 implies that β0T A[x̃0 + v] 6= 0.
Next we choose δ sufficiently small and the times of the tuple (θj )m
j=1 sufficiently
large such that
• each of the segments Ψj intersects b0 exactly once;
• all intersections of Ψj with b0 occur in Bε (x̃0 ) (let us denote these intersections
by x̃0j , which are given by x̃0j = Φ1 (θ̃j ; xj ) for some θ̃j ∈ (0, θj ));
• Φ1 (−τ ; −Bδ ) ⊂ Bε (x̃0 ), and, thus, x̃j = Φ1 (θj −τ ; xj ) = Φ1 (−τ ; −xj+1 ) ∈ Bε (x̃0 )
for j = 1, . . . , m.
1
This choice of δ and (θj )m j=1 is possible due the C -closeness of the segments Ψ1 ,. . . ,
Ψm to the curve Φ1 ((0, ∞]; −P+ v + P− v) ∪ Φ1 ([−∞, 0]; P+ v − P− v) outside of a
small neighbourhood of −v (the equilibrium of Φ1 ) for large times θ1 ,. . . , θm . The
construction of b0 implies that, for even j, the segments −Ψj intersect b0 in −Bε (x̃0 ).
Furthermore, all intersections of b0 with Ψj (j odd) and −Ψj (j even) are transversal
as they occur in Bε (x̃0 ) and −Bε (x̃0 ), respectively.
Finally, we modify b0 in the interior of Bε (x̃0 ) such that the modification b
intersects Ψj in x̃j with normal vector β̃j (instead of intersecting Ψj in x̃0j with
normal vector β0 ) for odd j. In exactly the same manner we also modify b0 in the
interior of −Bε (x̃0 ) such that the modification b intersects −Ψj in −x̃j with normal
−β̃j (instead of intersecting −Ψj in −x̃0j with normal vector −β0 ) for even j. An
additional restriction on the modification from b0 to b is the requirement that b has

to be transversal to all trajectories of Φ1 in Bε (x̃0 ) (and to all trajectories of Φ2 in
−Bε (x̃0 )). We demonstrate that this modification can be achieved by following the
flow Φ1 from each point z ∈ b0 ∩ Bε (x̃0 ) for an appropriate time h(z) (and then apply
the same procedure to −Bε (x̃0 )).
Let z ∈ U ⊂ Rn−1 7→ x̃0 + M z ∈ Rn be a parametrization of b0 ∩ Bε (x̃0 ) =
{ζ ∈ Bε (x̃0 ) : β0T (ζ − x̃0 ) = 0} where the matrix M ∈ Rn×(n−1) has full rank n − 1
and satisfies β0T M = 0. Then the intersection points x̃0j of Ψj with b0 have the form
x̃0j = x̃0 + M zj for a certain zj ∈ U . We choose h : C ∞ (clos U ; R) (clos U is the
closure of U ) such that
(i) h(zj ) = θj − τ − θ̃j and, for j = 1, 3, . . . , m − 1,
h i
−β̃jT exp A θj − τ − θ̃j M
h′ (zj ) = , (C.10)
β̃jT A[x̃j + v]
(ii) h(z) = 0 and h(k) (z) = 0 for all z ∈ ∂U (the boundary of U ) and k ∈ N;
(iii) Φ1 (h(z); M z + x̃0 ) ∈ Bε (x̃0 ) for all z ∈ clos U .
We choose the modified manifold b as b = {Φ1 (h(z); M z + x̃0 ) : z ∈ U } ∪ (b0 \ Bε (x̃0 )),
which is, first, a smooth manifold due to the points (ii) and (iii) of the construction
of h. Second, due to the smoothness of h, b is transversal to all trajectories of Φ1 in
Bε (x̃0 ). Third, the manifold b intersects Ψj in x̃j and has the normal vector β̃j due
to point (i) of the construction of h. Thus, after applying the analogous modification
in −Bε (x̃0 ) for even j, the manifold b is suitable as a switching manifold separating
the domains of Φ1 and Φ2 .
For the g constructed in the proof of Lemma 16 the closed curve Ψ is a slowly
oscillating periodic orbit that satisfies all genericity conditions postulated in section 4.
The intersection points x̃j (for j = 1, 3, . . . , m − 1) and −x̃j (for j = 2, 4, m) are
immediately followed by the corresponding switch at xj+1 and −xj+1 , respectively,
without any intermediate crossing of the switching manifold {g = 0}. Thus, Poincaré
maps along Ψ are, after symmetry reduction due to (C.1), concatenations of maps
generated by following Φ1 between the delayed switching manifolds and rotations by
π. The choice of the cross-section for a Poincaré map does not affect the linearized
stability of its fixed point corresponding to Ψ.
In the next step we choose the normal vectors β̃j (which are arbitrary in Lemma 16
apart from the transversality condition (C.9) in the construction of g) such that the
linearization of a Poincaré map along Ψ becomes identically zero in its fixed point
corresponding to Ψ. Equivalently, we can choose the normal vectors
T
βj+1 = β̃jT exp(−Aτ ) for j = 1, . . . , m − 1 and
(C.11)
β1T = T
β̃m exp(−Aτ )
to the local delayed switching manifolds Φ1 (τ ; {g = 0}) ∩ U (−xj ) in −xj for
j = 2, 4, . . . , m and (after symmetry reduction) −Φ1 (τ ; −{g = 0}) ∩ U (xj ) in xj
for j = 1, 3, . . . , m − 1. The transversality condition (C.9) on β̃j translates into the
condition
βjT A[v − xj ] 6= 0 (C.12)
for βj and j = 1, . . . , m. For any tuple (βj )m
satisfying (C.12), we find the
j=1
corresponding tuple of normal vectors to the switching manifold at (x̃j )m j=1 , which
is needed in the construction of g, by using the relation (C.11).
Let us choose a Poincaré section G0 for Ψ through a point p on the segment −Ψ2
(transversally to the curve −Ψ2 ). Let the intersection time t0 be such that t0 is in
the open interval (0, θ2 ). The periodic orbit Ψ corresponds to a fixed point p of the
return map. The return map along Ψ from G0 to back G0 has the simple form of a
concatenation of maps between (n − 1)-dimensional local hyper-surfaces of Rn . The
linearization of this concatenation in p contains the product of matrices
Π2 exp(θ1 A)(−I)Π1 exp(θm A)(−I)Πm exp(θm−1 A) · . . . · Πm−n+3 (C.13)
where the maps Πj are the discontinuity maps at the switching points xj (after
symmetry reduction). They are projections of the form
A[v − xj ]βjT
Πj = I − ,
βjT A[v − xj ]
which are well defined if the βj satisfy the transversality condition (C.12). Thus, the
kernel of Πj is spanned by A[v − xj ] and its image is {z ∈ Rn : βjT z = 0}. Let us
define the following recursion of matrices for j from m downward to m − n + 3:
P1 := −Π2 exp(θ1 A)Π1 , Pm = −P1 exp(θm A)Πm , Pj := −Pj+1 exp(θj A)Πj
for j < m. The product (C.13) coincides with the final iterate Pm−n+3 of this
recursion. We now prove inductively that we can choose the vectors βj (j =
2, 1, m, m − 1, . . . , m − n + 3) defining Πj such that
ker P1 = AL (x1 , v) ,
ker Pm = exp(−rm A)AL (exp(rm A)xm , v, exp(rm A)v)
(C.14)
= AL (xm , v, exp(−θm A)v) ,
ker Pj = exp(−rj A)AL (exp(rj A)xj , v, exp(rm A)v, . . . , exp(rj A)v)
for j = m − 1, . . . , m − n + 3 where the notation L(w1 , . . . , wk ) refers to the subspace
spanned by the vectors w1 , . . . , wk . Lemma 14 and Corollary 15 imply that all sets
on the right-hand-side of (C.14) are linearly independent. Thus, (C.14) implies that
dim ker Pm−j = j + 3, and, hence, Pm−n+3 = 0.
Initial step of induction (j = 1): Let β2 be arbitrary but satisfying the transversality
condition (C.12). The kernel of Π2 is spanned by A(v − x2 ), which is non-zero. Thus,
the kernel of Π2 exp(θ1 A) is spanned by exp(−θ1 A)A(v − x2 ) = A(v + x1 ) (due to
(C.4)). Because x1 and v are linearly independent, so are A(v+x1 ) and A(v−x1 ) (since
A is regular). Thus, we can choose β1 such that β1T A(v + x1 ) = 0 but β1T A(v − x1 ) 6= 0
(thus, β1 satisfies transversality condition (C.12)). The condition β1T A(v + x1 ) = 0
implies that ker[Π2 exp(θ1 A)] ⊂ ImΠ1 . Since ker Π1 = L(A(x1 − v)), this implies
ker P1 = L(A[x1 − v], A[x1 + v]) = AL(x1 , v). For n = 2 the product (C.13) is already
identically zero.
Inductive step from j + 1 to j (for n ≥ 3)

Assumption. Assume that
ker Pj+1 = exp(−rj+1 A)AL (exp(rj+1 A)xj+1 , v, exp(rm A)v, . . . , exp(rj+1 A)v) . (C.15)
Thus,
ker[−Pj+1 exp(θj A)] = exp(−rj A)AL (exp(rj+1 A)xj+1 , v, exp(rm A)v,
(C.16)
. . . , exp(rj+1 A)v) .
Step. The space spanned by the m − j + 3 vectors

V = L(exp(rj A)xj , v, exp(rm A)v, . . . , exp(rj+1 A)v, exp(rj A)v) (C.17)
has dimension m − j + 3 as stated in Corollary 15. Thus, the same holds for
exp(−rj A)AV , which can be split into the direct sum of the two components
(separating the first and the last two vectors in the right-hand-side of (C.17) from
the remaining m − j vectors)
AL(xj , v, exp(−θj A)v) ⊕ exp(−rj A)AL(v, exp(rm A)v, . . . , exp(rj+2 A)v). (C.18)
The first (three-dimensional) summand can be rewritten due to the identity
exp(−Aθj )xj+1 = exp(−Aθj )v − xj − v following from (C.4):
AL(xj , v, exp(−θj A)v) =
= AL(xj − v, exp(−θj A)xj+1 , exp(−θj A)v)
= L(A[xj − v]) ⊕ exp(−rj A)AL (exp(rj+1 A)xj+1 , exp(rj+1 A)v) .
Recombining this with the second summand of the direct sum (C.18) of exp(−rj A)AV
gives the direct sum
exp(−rj A)AV = L(A[xj − v]) ⊕ exp(−rj A)AL (v, exp(rm A)v,
. . . , exp(rj+2 A)v, exp(rj+1 A)v, exp(rj+1 A)xj+1 ) .
The second summand is by the assumption of the inductive step (C.16) the kernel of
Pj+1 exp(θj A). The first summand is the kernel of Πj . This implies that we can choose
βj such that βjT A(xj − v) 6= 0 (thus, satisfying the transversality (C.12)) but βjT z = 0
for all z ∈ [ker Pj+1 exp(θj A)], defining the map Πj such that ker[Pj+1 exp(θj A)] ⊂
ImΠj . With this choice of βj we have ker Pj = ker[Pj+1 exp(θj A)] ⊕ ker Πj =
exp(−rj A)AV , which has dimension m − j + 3 and assumes the form of (C.14) for j,
thus, proving the inductive step.
Consequently, if we choose the switching manifold as in the construction of
Lemma 16 with β̃j given by (C.11) and the βj as defined inductively above, the matrix
product (C.13) has a kernel of dimension n and is, thus, zero. Hence, the linearization
of the Poincaré map in G0 for the periodic orbit Ψ has also vanishes. This implies
that Ψ is quadratically stable, which proves Theorem 9.
Appendix D. Proof of Lemma 11
This section explains how the piecewise linearizations A1 and A2 in the statement
of Lemma 11 depend on the right-hand-side and the concrete configuration of the
periodic orbit x̃. Assume (without loss of generality) that x̃ follows Φ1 for times
smaller than s̃2 and then switches to Φ2 at s̃2 . Furthermore, we order the intersection
times −p < s̃1 < . . . < s̃m < 0 and denote the corresponding switching times by
t̃1 , . . . , t̃m .
Case (a)
Let us denote fj0 = fj (x̃(s̃2 )), fj1 = fj (x̃(t̃2 )) where j = 1, 2 and t̃2 = s̃2 + τ ,
and g ′ = g ′ (x̃(s̃2 )). Furthermore, let F1 be the hyperplane intersecting x̃ in x̃(t̃2 )
orthogonal to the outgoing flow f11 . Let R be the return map along x̃(·) from F1 to
Gτ1 , which is a concatenation of smooth maps. We denote its derivative ∂x R|x=x̃(t̃2 ) by
R′ . Case (a) is defined in section 6.1 by g ′ f10 · g ′ f20 > 0, which means that the periodic
U (x̃(s̃2 )) U (x̃(t̃2 ))
R1
x̃(·)
Gτ1 x̃(·)
f11
z0 {g < 0} Gτ2
f20
G2 z3 y3
x̃(s̃2 ) y1 x̃(t̃2 ) F1
{g ≥ 0}
z4
f10 f21
z1 z2 y0 y2
x̃(·) x̃(·)
R1 Φ2 (τ ; ·)
Φ1 Φ2
Figure D1. Sketch of the neighbourhoods U (x̃(s̃2 )) and U (x̃(t̃2 )) when x̃(·)
undergoes a corner collision of type (a). Dashed trajectories follow flow Φ1 ,
dotted trajectories follow flow Φ2 . The return map P0 is a concatenation of a
non-smooth map from Gτ1 to F1 and a smooth map R1 from F1 back to Gτ1 . The
non-smooth map from Gτ1 to F1 maps x̃(s̃2 ) to x̃(t̃2 ), y0 ∈ F− to y3 ∈ F1 , and
z0 ∈ F+ to z4 ∈ F1 .
orbit intersects the switching manifold {g = 0} transversally. Figure D1 shows the

configuration of the neighbourhood of x̃(s̃2 ) in the left panel. In this case there can
be no other intersection or switching point on x̃ between s̃2 and t̃2 . The maps A1 (for
x + x̃(s̃2 ) ∈ F− ) and A2 (for x + x̃(s̃2 ) ∈ F+ ) have the form
" #
1 1T
f20 g ′

′ f1 1f
A1 = R I − T ∂2 Φ2 (τ ; x̃(s̃2 )) I − ′ 0
f11 f11 g f2
" T
# (D.1)
f11 f11 f20 g ′

′
A2 = R I − T ∂2 Φ2 (τ ; x̃(s̃2 )) I − ′ 0 .
f11 f11 g f1
Notice that A1 and A2 differ only in the last factor. Let us first consider the case
of a trajectory through a point y0 = x̃(s̃2 ) + x ∈ F− (see figure D1). It follows Φ2
until it reaches G2 in y1 . It continues to follow Φ2 for time τ until it reaches Gτ2 , the
time-τ image of G2 in y2 near x̃(s̃2 ) (see right panel of figure D1). The point y3 is the
projection of y2 onto F1 under Φ1 . The points y1 , y2 , and y3 have the expansions
f 0 g′

y1 − x̃(s̃2 ) = I − 2′ 0 x + O(kxk2 ),
g f2
y2 − x̃(t̃2 ) = ∂2 Φ2 (τ ; x̃(s̃2 ))(y1 − x̃(s̃2 )) + O(kxk2 ),
" T
#
f11 f11
y3 − x̃(t̃2 ) = I − T (y2 − x̃(s̃2 )) + O(kxk2 )
f11 f11
giving the expression for A1 in (D.1). A trajectory through a point z0 = x̃(s̃2 )+x ∈ F+
has crossed G2 in z1 at time δ, following Φ1 . Thus, the trajectory follows Φ2 from z0
U (x̃(s̃2 )) U (x̃(t̃2 ))
R2
Gτ1
{g < 0} f11
z0
z5 x̃(·) y1
z2 G2
x̃(s̃2 ) {g ≥ 0} f21
z1 f20 z3
F2
f10 x̃(·)
y2
y0 x̃(t̃2 )
z6
x̃(·)
z4
x̃(·)
R2 Φ2 (τ ; ·)
Φ1 Φ2
Figure D2. Sketch of the neighbourhoods U (x̃(s̃2 )) and U (x̃(t̃2 )) when x̃(·)
undergoes a corner collision of type (b). The return map is a concatenation of a
non-smooth map Gτ1 7→ F2 and the smooth map R : F2 7→ Gτ1 . The non-smooth
map maps x̃(s̃2 ) to x̃(t̃2 ), y0 ∈ F− to y2 ∈ F2 and z0 ∈ F+ to z6 ∈ F2 .
up to z3 = Φ2 (τ ; z2 ) near x̃(s̃2 ) (see right panel of figure D1) where z2 = Φ2 (δ; z0 ).

The point z4 is the projection of z3 onto F1 under Φ1 . The expansions of δ, z2 to z4
are
δ= − g ′ x/(g ′ f10 ) + O(kxk2 ),
f 0 g′

z2 − x̃(s̃2 ) = I − 2′ 0 x + O(kxk2 ),
g f1
z3 − x̃(t̃2 ) = ∂2 Φ2 (τ ; x̃(s̃2 ))(z2 − x̃(s̃2 )) + O(kxk2 ),
" T
#
f11 f11
z4 − x̃(t̃2 ) = I − T (z3 − x̃(s̃2 )) + O(kxk2 ),
f11 f11
which implies the expression for A2 in (D.1).
Case (b)
Let us denote fj0 = fj (x̃(s̃2 )), fj1 = fj (x̃(t̃2 )) where j = 1, 2 and t̃2 = s̃2 + τ ,
and g ′ = g ′ (x̃(s̃2 )). Furthermore, let F2 be the hyperplane intersecting x̃ in x̃(t̃2 )
orthogonal to the flow f21 . Let R be the return map along x̃(·) from F2 to Gτ1 , which
is a concatenation of smooth maps. We denote its derivative ∂x R|x=x̃(t̃2 ) by R′ . Case
(b) is defined in section 6.1 by g ′ f10 · g ′ f20 < 0, which means that the periodic orbit
lies entirely on one side of the switching manifold {g = 0} near x̃(s̃2 ), touching it in
x̃(s̃2 ). The left panel of figure D2 shows the configuration of the manifolds G2 , Gτ1
and the periodic orbit x̃ in the neighbourhood of x̃(s̃2 ). In addition, case (b) requires
that the orbit x̃ does not intersect {g = 0} between s̃2 and t̃2 . The maps A1 (for
x + x̃(s̃2 ) ∈ F− ) and A2 (for x + x̃(s̃2 ) ∈ F+ ) have the form

" T
#
′ f21 f21
A1 = R I − T ∂2 Φ2 (τ ; x̃(s̃2 )),
f21 f21
" #
T
(D.2)
f21 f21 f11 g ′ f11 g ′

′
A2 = R I − T ∂2 Φ2 (τ ; x̃(s̃2 )) + ′ 0 − ′ 0 .
f21 f21 g f1 g f2
Let us first consider a trajectory going through a point y0 = x̃(s̃2 ) + x ∈ F− . It follows
Φ2 until it reaches F2 in y2 . The point y2 is the projection of y1 = Φ2 (τ ; y0 ) onto F2
under Φ2 . The expansion for y1 and y2 is
y1 − x̃(t̃2 ) = ∂2 Φ2 (τ ; x̃(s̃2 ))x + O(kxk2 ),
" T
#
f21 f21
y2 − x̃(t̃2 ) = I − T (y1 − x̃(t̃2 )) + O(kxk2 ),
f21 f21
which, in concatenation with R′ : F2 7→ Gτ1 , gives the expression for A1 in (D.2). For a
trajectory through a point z0 = x̃(s̃2 ) + x ∈ F+ we have to compute the time that this
trajectory spent in {g ≥ 0} ∩ U (x̃(s̃2 )), which is the traveling time −δ1 from z1 to z0
under Φ1 plus the traveling time δ2 from z0 to z2 under Φ2 (see figure D2). The point
z3 is the Φ2 (τ ; ·)-image of z0 . In the point z4 = Φ2 (δ1 ; z3 ) the trajectory switches to
the flow Φ1 for time δ2 − δ1 , reaching z5 where it switches back to Φ2 . The point z6
is the projection of z5 = Φ1 (δ2 − δ1 ; z4 ) onto F2 under Φ2 . The expansions of δ1 , δ2 ,
and z3 to z6 are
δ1 = −g ′ x/(g ′ f10 ) + O(kxk2 ), (δ1 < 0),
δ2 = −g ′
x/(g ′ f20 ) 2
+ O(kxk ), (δ2 > 0),
z3 = x̃(t̃2 ) + ∂2 Φ2 (τ ; x̃(s̃2 ))x + O(kxk2 ),
z4 = z3 + δ1 f21 + O(kxk2 ),
z5 = z4 + (δ2 − δ1 )f11 + O(kxk2 ),
" T
#
f21 f21
z6 = x̃(t̃2 ) + I − T (z5 − x̃(t̃2 )) + O(kxk2 ),
f21 f21
which implies the expression for A2 in (D.2) because the difference between z3 and z4
is in the kernel of the projection onto the hyperplane F2 .
Case (c)
In this case four locations in the physical space are involved in determining the
discontinuity in the linearization of the Poincaré map P0 : Gτ1 7→ Gτ1 0: U (x̃(s̃2 )),
U (x̃(s̃3 )), U (x̃(t̃2 )) and U (x̃(t̃3 )). A characteristic feature of this case is that the orbit
x̃ crosses the switching manifold between s̃2 and t̃2 = s̃2 + τ (in s̃3 ; see figure 4(c)).
Locally near x̃(s̃2 ), the orbit x̃ lies entirely on one side of {g = 0}, say, {g ≥ 0},
switching from Φ1 to Φ2 in s̃2 . Figure D3 shows the four neighbourhoods.
Let us denote fj0 = fj (x̃(s̃2 )), fj1 = fj (x̃(s̃3 )), fj2 = fj (x̃(t̃2 )), and fj3 = fj (x̃(t̃3 ))
where j = 1, 2 and t̃k = s̃k + τ (k = 2, 3). Furthermore, let g0′ = g ′ (x̃(s̃2 ))
and g1′ = g ′ (x̃(s̃3 )), and F3 be the hyperplane intersecting x̃ in t̃3 orthogonal to
the outgoing flow Φ1 . The intersection and switching manifolds are called Gτ1 =
Φ1 (τ ; {g = 0}) ∩ U (x̃(s̃2 )), G2 = {g = 0} ∩ U (x̃(s̃2 )), G3 = {g = 0} ∩ U (x̃(s̃3 )) and
Gτ3 = Φ2 (τ ; {g = 0}) ∩ U (x̃(t̃3 )), respectively. We denote by Π1 the projection along
U (x̃(s̃2 )) G2 U (x̃(s̃3 ))
{g < 0} {g ≥ 0}
{g ≥ 0}
Φ2
z2 f20 x̃(·) x̃(s̃3 )
Gτ1 z3 {g < 0}
z0
y1
x̃(s̃2 ) f21
y0 G3
f10
x̃(·)
R z1
x̃(·)
U (x̃(t̃3 )) U (x̃(t̃2 ))
Φ2
x̃(·)
Gτ3
y2
f13
x̃(t̃2 )
y3 Φ2
f22
z7 x̃(t̃3 ) z4
F3 f12
z5
z6
Φ1 Φ2
Figure D3. Sketch of the neighbourhoods U (x̃(s̃2 )), U (x̃(s̃3 )), U (x̃(t̃2 )) and
U (x̃(t̃3 )) when x̃(·) undergoes a corner collision of type (c). The return map is a
concatenation of the non-smooth map Gτ1 7→ F3 , mapping y0 to y3 and z0 to z7 ,
and the smooth map R : F3 7→ Gτ1 .
Φ2 onto G3 , linearized in x̃(s̃3 ), and by Π3 the projection along Φ1 onto F3 , linearized

in x̃(t̃3 ). The projections Π1 and Π3 read
T
f21 g1′ f 3f 3
Π1 = I − ′ 1 , Π3 = I − 1 T1 .
g1 f2 f13 f13
Let R be the return map along x̃(·) from F3 back to Gτ1 , which is a concatenation
of smooth maps. We denote its derivative ∂x R|x=x̃(t̃3 ) by R′ . We denote by
B1 = ∂2 Φ2 (s̃3 − s̃2 ; x̃(s̃2 )), B2 = ∂2 Φ2 (t̃2 − s̃3 ; x̃(s̃3 )) and B3 = ∂2 Φ2 (t̃3 − t̃2 ; x̃(t̃2 )) the
linearizations of the flow Φ2 in x̃ between the different neighbourhoods. The vectors
f2k satisfy the relations f21 = B1 f20 , f22 = B2 B1 f20 , and f23 = B3 B2 B1 f20 . Using these
notations, the piecewise linearizations in the statement of Lemma 11 are
A1 = R′ Π3 B3 B2 Π1 B1 ,
f 0 g ′ B1 f 0 g′ f 0 g′ (D.3)
2 ′
f 2 g′

f g
A2 = R′ Π3 B3 B2 B1 I − 2′ 1 0 + 2′ 00 − 2′ 00 + 1′ 00 − 1′ 00 .
g1 B1 f2 g0 f2 g0 f1 g0 f1 g0 f2
Let us first consider a trajectory going through a point y0 = x̃(s̃2 ) + x ∈ F− . It
follows Φ2 for time τ + s̃3 − s̃2 . The point y1 is the intersection point of the trajectory
with G3 , y2 is Φ2 (τ ; y1 ), y3 is the projection of y2 onto F3 under the flow Φ1 . This

point y3 is then mapped back to Gτ1 by R. The expansions of y1 , y2 and y3 read
y1 − x̃(s̃3 ) = Π1 B1 x + O(kxk2 ),
y2 − x̃(t̃3 ) = B3 B2 (y1 − x̃(s̃3 )) + O(kxk2 ),
y3 − x̃(t̃3 ) = Π3 (y2 − x̃(t̃3 )) + O(kxk2 ),
which, in concatenation with R′ : F3 7→ Gτ1 , gives the expression for A1 in (D.3).
For a trajectory through a point z0 = x̃(s̃2 ) + x ∈ F+ we have to compute the
time that this trajectory spends in {g < 0} ∩ U (x̃(s̃2 )), which is the traveling time −δ1
from z1 to z0 under Φ1 plus the traveling time δ2 from z0 to z2 under Φ2 (the same
as in case (b)). The point z3 is the intersection point of the trajectory with G3 . The
intersection time of the trajectory with G3 , starting from z0 , is s̃3 − s̃2 + δ3 for a small
δ3 . In the point z4 = Φ2 (τ + δ1 ; z0 ) the trajectory switches to the flow Φ1 , follows it
for time δ2 − δ1 to z5 . From z5 it continues to follow Φ2 for time s̃3 − s̃2 + δ3 − δ2 to
z6 . The point z7 is the projection of z6 onto F3 under the flow Φ1 . This point z7 is
then mapped back to Gτ1 by R. The expansions of δ1 , δ2 , z4 and z5 are the same as
in case (b), the expansion of δ3 , and z4 to z7 read
δ3 = −g1′ B1 x/(g1′ f21 ) + O(kxk2 ),
z4 = x̃(t̃2 ) + B2 B1 x + δ1 f22 + O(kxk2 ),
= x̃(t̃2 ) + B2 B1 (x + δ1 f20 ) + O(kxk2 ),
z5 = z4 + (δ2 − δ1 )f12 + O(kxk2 ),
z6 = x̃(t̃3 ) + B3 (z5 + (δ3 − δ2 )f22 − x̃(t̃2 )) + O(kxk2 ),
= x̃(t̃3 ) + B3 B2 B1 x + (δ3 − δ2 + δ1 )f20 + B3 (δ2 − δ1 )f12 + O(kxk2 ),

z7 = x̃(t̃3 ) + Π3 (z6 − x̃(t̃3 )) + O(kxk2 ),

which, in concatenation with R′ : F3 7→ Gτ1 , gives the expression for A2 in (D.3).
Appendix E. Proof of Lemma 13
Case (a)
The characteristic feature of case (a) is that the orbit x̃ does not cross the switching
manifold between the grazing time s∗ and t∗ = s∗ + τ . Furthermore, let us assume
that the configuration is such that the orbit x̃ does not switch from Φ1 to Φ2 between
˙ ∗ ) = f1 (x̃(s∗ ))
s∗ and t∗ , either; see figure E1 where we use the abbreviations f 0 = x̃(s
T
and fj1 = fj (x̃(t∗ )). The hyperplane F1 = {x : f11 [x − x̃(t∗ )] = 0} intersects x̃
orthogonally in x̃(t∗ ). The linearized projection along Φ1 onto F1 defined by
T
f11 f11
Π=I− T
f11 f11
is orthogonal. We express the return map to F0 as a concatenation of a piecewise
smooth map from F0 to F1 and a smooth map R along x̃ from F1 back to F0 , which
is a concatenation of smooth maps. Let us denote the derivative ∂x R(x̃(t∗ )) by R′ .
Using these notations the matrix A and the vector v in the statement of Lemma 13
have the form
A = R′ Π∂2 Φ1 (τ ; x̃(s∗ ))
(E.1)
v = 2R′ Πf21 .
U (x̃(s∗ )) U (x̃(t∗ ))
{x : m(x) < 0}
F0
xm (z0 ) z5
0} f21
R z0 z2 {g = x̃(·) R
y1
f0
x̃(s∗ ) z3
z1 f11
x̃(·)
z6
x̃(t∗ ) y2 F1
x̃(·) y0 z4
Φ1 (τ ; ·)
x̃(·)
{x : m(x) > 0}
Φ1 Φ2
Figure E1. Sketch of the neighbourhoods U (x̃(s∗ )) and U (x̃(t∗ )) when x̃(·)
undergoes a grazing bifurcation of type (a). The return map to F0 is a
concatenation of a non-smooth map F0 7→ F1 and a smooth map R. The non-
smooth map maps y0 ∈ F+ to y2 ∈ F1 and z0 ∈ F− to z6 ∈ F1 .
Figure E1 illustrates how points of F0 near x̃(s∗ ) are mapped to F1 . A trajectory

through a point y0 = x̃(s∗ ) + x ∈ F+ never crosses {g = 0} in U (x̃(s∗ )). Thus, it is
mapped to y1 ∈ U (x̃(t∗ )) by Φ1 (τ ; ·). The point y2 is the projection of y1 onto F1
under Φ1 . Thus, the expansion of y2 with respect to y0 is
y1 − x̃(t∗ ) = ∂2 Φ1 (τ ; x̃(s∗ ))x + O(kxk2 ),
y2 − x̃(t∗ ) = Π(y1 − x̃(t∗ )) + O(kxk2 ),
which implies the expression for A in (E.1) in the case x̃(s̃) + x ∈ F+ .
The function m : U (x̃(s∗ )) → R, used to define F− and F+ in Lemma 13, and
defined by
m(x) = min q −1 g(Φ1 (δ; x))
δ∈[−δ0 ,δ0 ]
is uniquely defined and smooth for a sufficiently small δ0 > 0 due to Condition 12 (see
page 18) stating the non-degeneracy of the grazing event. Moreover, the function
δm : U (x̃(s∗ )) 7→ [−δ0 , δ0 ], defined by q −1 g(Φ1 (δm (x); x)) = m(x),
and the map
xm (x) : U (x̃(s∗ )) 7→ U (x̃(s∗ )), defined by xm (x) = Φ1 (δm (x); x),
are also well-defined and smooth in U (x̃(s∗ )). The function δm describes the traveling
time to the minimum in the definition of m. The map xm describes the position
in Rn where this minimum is attained. Thus, δm (x̃(s∗ )) = 0, which implies
δm (x̃(s∗ ) + x) = O(kxk), and xm (x̃(s∗ )) = x̃(s∗ ).
A trajectory through z0 = x̃(s∗ ) + x ∈ F− has two intersections z1 and z2 with
{g = 0}. The traveling time from z0 to xm (z0 ) is δm (z0 ). The traveling times −δ1 from
z1 to xm (z0 ) and δ2 from xm (z0 ) to z2 are solutions of h(δ) := q −1 g(Φ1 (δ, xm (z0 ))) = 0,
which expands as
0 = h(δ) = m(z0 ) + δ 2 + O(kxk2 ) + O(δ 3 ).
Thus, δ1 and δ2 have the expansions (keeping in mind that m(z0 ) = O(kxk))
p p
δ1 = − −m(z0 ) + O(kxk), δ2 = −m(z0 ) + O(kxk). (E.2)
This implies that both, thep traveling time from z1 to z0 and the traveling time from
z0 to z2 , are of the order −m(z0 ) + O(kxk) (because δm (z0 ) = O(kxk)).
The trajectory through z0 switches to the flow Φ2 time −δ1 before it reaches the
point
z3 = Φ1 (τ ; z0 ) = x̃(t∗ ) + O(kxk)
(see figure E1). This happens in point
z4 = z3 + δ1 f11 + O(kxk) = x̃(t∗ ) + δ1 f11 + O(kxk).
Subsequently the trajectory follows Φ2 for time δ2 − δ1 up to
p p
z5 = z4 + 2 −m(z0 )f21 + O(kxk) = x̃(t∗ ) + δ1 f11 + 2 −m(z0 )f21 .
The point z6 is the projection of z5 onto F1 under Π, which projects f11 to 0. Thus,
the expansion of z6 is
z6 = x̃(t∗ ) + Π(z5 − x̃(t∗ )) + O(kxk2 )
p
= x̃(t∗ ) + Π[2f21 ] −m(z0 ) + O(kxk),
which implies the expression for v in (E.1).
If the orbit x̃ switches from Φ1 to Φ2 between s∗ and t∗ (at some time t̃1 ∈ (s∗ , t∗ )
a modification of (E.1) applies. Since x̃ follows Φ2 in t∗ instead of Φ1 the role of f11
and f21 is interchanged in the definition of Π and v. Furthermore, the time τ -map from
U (x̃(s∗ )) to U (x̃(t∗ )) is no longer Φ1 (τ, ·) but R0 (x) = Φ2 (τ − t(x); Φ1 (t(x); x)) where
t(x) is the traveling time from x to the delayed switching manifold Gτ1 = Φ1 (τ ; {g =
0}) ∩ U (x̃(t̃1 )). This traveling time depends smoothly on x, which implies that R0 is
smooth as well. With these modifications the arguments given above lead to
" T
# " T
#
f21 f21 f21 f21
A=R I− ′
T
′
∂x R0 (x̃(s∗ )), v = 2R I − T
f11 . (E.3)
f21 f21 f21 f21
Case (b)
The characteristic feature of this case is that the orbit x̃ intersects the switching
manifold {g = 0} between s∗ and t∗ at some time s̃2 . Four locations in the physical
space are involved in determining the discontinuity in the linearization of the Poincaré
map P0 from F0 back to F0 : U (x̃(s∗ )), U (x̃(s̃2 )), U (x̃(t∗ )) and U (x̃(t̃2 )). Let us first
assume that the orbit x̃ does not switch from flow Φ1 to Φ2 between s∗ and s̃2 .
Figure E2 shows this configuration. It uses the abbreviations f 0 = x̃(s ˙ ∗) =
f1 (x̃(s∗ )), fj1 = fj (x̃(s̃2 )), fj2 = fj (x̃(t∗ )) and fj3 = fj (x̃(t̃2 )) for j = 1, 2. The
T
hyperplane F3 = {x : f23 [x − x̃(t̃2 )] = 0} intersects x̃ orthogonal to the outgoing flow
Φ2 in x̃(t̃2 ). We denote by Π1 the projection along Φ1 onto G2 = {g = 0} ∩ U (x̃(s̃2 )),
linearized in x̃(s̃2 ), and by Π3 the projection along Φ2 onto F3 , linearized in x̃(t̃2 ).
The projections Π1 and Π3 read
T
f11 g ′ (x̃(s̃2 )) f 3f 3
Π1 = I − ′ 1 , Π3 = I − 2 T2 .
g (x̃(s̃2 ))f1 f23 f23
U (x̃(s∗ )) {g = 0} U (x̃(s̃2 ))
{x : m(x) < 0}
z2 {g < 0}
Φ1 z3
xm (z0 ) x̃(s̃2 )
x̃(·) {g ≥ 0}
f0
z0 y1
{x : m(x) > 0} G2 f11
x̃(s∗ )
y0 x̃(·)
z1
F0
x̃(·)
U (x̃(t̃2 )) U (x̃(t∗ )) Φ1
x̃(·)
R Gτ2
Φ2 y2 x̃(t∗ )
f23
y3 Φ1
f12 z4
z7 x̃(t̃2 )
F3 f22
z5
z6
Φ1 Φ2
Figure E2. Sketch of the neighbourhoods U (x̃(s∗ )), U (x̃(s̃2 )), U (x̃(t∗ )) and
U (x̃(s̃t )) when x̃(·) undergoes a grazing bifurcation of type (b). The return map
to F0 is a concatenation of a non-smooth map F0 7→ F3 and a smooth map R.
The non-smooth map maps y0 ∈ F+ to y3 ∈ F3 and z0 ∈ F− to z7 ∈ F3 .
We express the return map to F0 as a concatenation of a piecewise smooth map from

F0 to F3 and a smooth map R from F3 back to F0 . The return map R along x̃(·)
from F3 back to F0 is a concatenation of smooth maps. Let us denote its derivative
∂x R(x̃(t̃2 )) by R′ . We also make use of the function m defined in section 6.2 and
discussed in more detail in the treatment of case (a).
Using these notations the matrix A and the vector v in the statement of Lemma 13
have the form
A = R′ Π3 ∂2 Φ1 (τ ; x̃(s2 ))Π1 ∂2 Φ1 (s̃2 − s∗ ; x̃(s∗ ))
(E.4)
v = 2R′ Π3 ∂2 Φ1 (s̃2 − s∗ ; x̃(t∗ ))[f22 − f12 ].
A trajectory through a point y0 = x̃(s∗ ) + x ∈ F+ does not cross {g = 0} in U (x̃(s∗ ))
(see figure E2). It is mapped to y1 ∈ G2 by Φ1 . The trajectory continues to follow
Φ1 for time τ from y1 to y2 . The point y3 is the projection of y2 onto F3 under Φ2 .
Thus, the expansion of y1 , y2 , y3 with respect to y0 is

y1 − x̃(s̃2 ) = Π1 ∂2 Φ1 (s̃2 − s∗ ; x̃(s∗ ))x + O(kxk2 ),
y2 − x̃(t̃2 ) = ∂2 Φ1 (τ ; x̃(s̃2 ))(y1 − x̃(s̃2 )),
y3 − x̃(t̃2 ) = Π3 (y2 − x̃(t̃2 )) + O(kxk2 ),
which implies the expression for A in (E.4).
A trajectory through a point z0 = x̃(s∗ ) + x ∈ F− has two intersection points
with {g = 0}, z1 and z2 (see figure E2). The traveling times −δ1 from z1 to z0 and δ2
from z0 to z2 have been computed to leading order already in (E.2) in the treatment
of the grazing case (a). The intersection of the trajectory with G2 is named z3 . The
difference δ3 between the traveling time from z0 to z3 and s̃2 − s∗ is of order O(kxk).
At z4 = Φ1 (τ +δ1 ; z0 ) the trajectory switches to Φ2 for time δ2 −δ1 , reaching z5 . From
z5 it continues to follow Φ1 for time t̃2 − t∗ + δ3 − δ2 = s̃2 − s∗ − δ2 + O(kxk) reaching
z6 . The point z7 ∈ F3 is the projection of z6 onto F3 following the outgoing flow Φ2 ,
and is then mapped back to F0 by R. The expansion of z7 in z0 is to leading order
δ3 = O(kxk),
z4 = x̃(t∗ ) + δ1 f12 + O(kxk),
z5 = z4 + (δ2 − δ1 )f22 + O(kxk)),
z6 = x̃(t̃2 ) + ∂2 Φ1 (s̃2 − s∗ ; x̃(t∗ )) z5 − δ2 f12 − x̃(t∗ ) + O(kxk) + O(kxk),

= x̃(t̃2 ) + ∂2 Φ1 (s̃2 − s∗ ; x̃(t∗ ))(δ2 − δ1 ) f22 − f12 + O(kxk),

p
= x̃(t̃2 ) + 2∂2 Φ1 (s̃2 − s∗ ; x̃(t∗ )) f22 − f12

−m(z0 ) + O(kxk),
z7 = x̃(t̃2 ) + Π3 [z6 − x̃(t̃2 )] + O(kxk),
which implies the expression for v in (E.4).
If the orbit x̃ switches from Φ1 to Φ2 between s∗ and s̃2 (at some time t̃1 ∈ (s∗ , s̃2 ))
a modification of (E.4) applies. Since x̃ follows Φ2 in s̃2 and t∗ instead of Φ1 , and
switches from Φ2 to Φ1 in t̃2 , the role of f11 and f21 is interchanged in the definition
of Π1 , Π3 and v. Furthermore, the time-(s∗ − s̃2 ) map from U (x̃(s∗ )) to U (x̃(s̃2 ))
is no longer Φ1 (s̃2 − s∗ , ·) but R1 (x) = Φ2 (s̃2 − s∗ − t(x); Φ1 (t(x); x)) where t(x)
is the traveling time from x to the delayed switching manifold Gτ1 = Φ1 (τ ; {g =
0}) ∩ U (x̃(t̃1 )). This traveling time depends smoothly on x, which implies that R1 is
smooth as well. With these modifications the derivation given above leads to
" T
#
f13 f13 f21 g ′ (x̃(s̃2 ))

′
A=R I− ∂ Φ
2 2 (τ ; x̃(s2 )) I − ∂x R1 (x̃(s∗ )),
T
f13 f13 g ′ (x̃(s̃2 ))f21
" T
# (E.5)
′ f13 f13 2 2
v = 2R I − T
∂2 Φ2 (s̃2 − s∗ ; x̃(t∗ ))[f1 − f2 ].
f13 f13
This completes the proof of Lemma 13.

Sieber, J. (2006) - Dynamics of Delayed Relay Systems.: University of Bristol - Explore Bristol Research

Uploaded by

Copyright:

Available Formats

Sieber, J. (2006) - Dynamics of Delayed Relay Systems.: University of Bristol - Explore Bristol Research

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sieber, J. (2006) - Dynamics of Delayed Relay Systems.: University of Bristol - Explore Bristol Research

Uploaded by

Copyright:

Available Formats

Sieber, J. (2006). Dynamics of delayed relay systems.

Early version, also known as pre-print

Link to published version (if available):

Link to publication record in Explore Bristol Research

University of Bristol - Explore Bristol Research

AMS classification scheme numbers: 37G15, 34K18, 34K35

has a stable periodic orbit.

Piecewise smooth dynamical systems model many problems in control engineering

3. Fundamental properties of delayed relay systems — definition of

We define a delayed relay system as a dynamical system governed by a differential

{x : g(x) = 0} constitutes a piecewise smooth (n − 1)-dimensional submanifold of Rn ,

solution x ∈ C([0, T ]; Rn ) of the variation-of-constants formula corresponding to DDE

4. Behaviour near generic relay periodic orbits

4.1. Illustration — linearized inverted pendulum

4.2. Return map for periodic orbits in the general case

Lemma 4 (Continuity of evolution in periodic orbits)

Condition 6 (Transversality of all intersections)

where we use the notation t0 := −τ .

Furthermore, IM induces a map P0 on B defined by IM P0 y = P IM (y). If B is

4.3. Poincaré map P0 for slowly oscillating orbits

Figure 3. Construction of the switching line G = G1 ∪ G∗ ∪ G2 and the stable

5. Existence of stable periodic orbits for arbitrarily large delay

5.1. Stabilization of the inverted pendulum

5.2. Stable periodic orbits in n-dimensional systems

6.1. Corner collision

Condition 10 (secondary genericity conditions for corner collision)

(a) {g < 0} (b) x̃(s̃3 ) (c) x̃(s̃3 )

Figure 4. Illustration of a periodic orbit x̃ undergoing a corner collision at x̃(s̃2 ),

6.2. Tangential grazing

(a) The orbit x̃ is quadratically tangent to the switching manifold {g = 0} in s∗ , and

(a) {g ≥ 0} (b) {g < 0}

Figure 5. Illustration of a periodic orbit x̃ undergoing a tangential grazing at

where D+ ∪ D0 ∪ D− is a partition of the physical space Rn . This type of hybrid

The research of J.S. is supported by EPSRC grant GR/R72020/01. The author

Appendix A. Proof of Lemma 4

Appendix B. Proof of Lemma 7

B0 on the history interval,

Appendix C. Proof of Theorem 9 in section 5

of n − 1 vectors spans a (n − 1)-dimensional space (where rj = θj + . . . + θm ). The

additional restriction on the modification from b0 to b is the requirement that b has

Inductive step from j + 1 to j (for n ≥ 3)

Step. The space spanned by the m − j + 3 vectors

Appendix D. Proof of Lemma 11

orbit intersects the switching manifold {g = 0} transversally. Figure D1 shows the

up to z3 = Φ2 (τ ; z2 ) near x̃(s̃2 ) (see right panel of figure D1) where z2 = Φ2 (δ; z0 ).

x + x̃(s̃2 ) ∈ F− ) and A2 (for x + x̃(s̃2 ) ∈ F+ ) have the form

Φ2 onto G3 , linearized in x̃(s̃3 ), and by Π3 the projection along Φ1 onto F3 , linearized

with G3 , y2 is Φ2 (τ ; y1 ), y3 is the projection of y2 onto F3 under the flow Φ1 . This

z7 = x̃(t̃3 ) + Π3 (z6 − x̃(t̃3 )) + O(kxk2 ),

Appendix E. Proof of Lemma 13

Figure E1 illustrates how points of F0 near x̃(s∗ ) are mapped to F1 . A trajectory

We express the return map to F0 as a concatenation of a piecewise smooth map from

Thus, the expansion of y1 , y2 , y3 with respect to y0 is

= x̃(t̃2 ) + ∂2 Φ1 (s̃2 − s∗ ; x̃(t∗ ))(δ2 − δ1 ) f22 − f12 + O(kxk),

You might also like