a

r
X
i
v
:
m
a
t
h
/
0
3
0
5
2
0
7
v
4


[
m
a
t
h
.
D
S
]


6

M
a
y

2
0
0
6Lipschitz Flow-box Theorem
Craig Calcaterra and Axel Boldt
Department of Mathematics, Metropolitan State University, Saint Paul, MN 55106
craig.calcaterra@metrostate.edu, phone: (651) 793-1423, fax: (651) 793-1446
February 1, 2008
Abstract
A generalization of the Flow-box Theorem is proven. The assumption
of a C
1
vector eld f is relaxed to the condition that f be locally Lipschitz.
The theorem holds in any Banach space.
Key Words: Flow-box Theorem; local linearization of a vector eld;
Straightening-out Theorem; Lipschitz continuous; Banach space
1 Introduction
The Flow-box Theorem for smooth vector elds states that the dynamic near a
non-equilibrium point is qualitatively trivial, i.e., topologically conjugate with
translation. Near a nondegenerate equilibrium point, linearizing the vector eld
by dierentiation allows a relatively simple characterization of almost all possi-
ble local dynamics. These two results characterize the local behavior of solutions
for smooth nondegenerate vector elds. A natural follow-up question is, What
dynamics are possible under nonsmooth conditions?
To be more specic, the Flow-box Theorem (also called the Straightening-
out Theorem or the Local Linearization Lemma) applies to autonomous,
rst-order dierential equations, i.e.,
x

(t) = f (x( t)). (1)

f typically is a vector eld on a manifold. For local questions such as ours,
it is enough to study the case of a map f : X X where X = R
n
or some
other Banach space. A solution to (1) with initial condition x
0
X is a curve
x : I X where I is an open subinterval of R containing 0, x(0) = x
0
, and
which satises (1) for all t I.
The traditional Flow-box Theorem asserts that if f is a C
1
vector eld and
x
0
X is not an equilibrium, i.e., f(x
0
) = 0, then there is a dieomorphism
which transfers the vector eld near x
0
to a constant vector eld. In other
words, the local ow of f is conjugate via dieomorphism to translation.
1
The Picard-Lindelof Theorem
1
, stated below, guarantees a unique solution

x
exists for every initial condition x X if f is locally Lipschitz-continuous.
The continuous dependence of solutions on initial conditions (Lemma 2, below)
is also assured when f is Lipschitz continuous. For these nonsmooth vector
elds, is the dynamic near non-equilibria still qualitatively trivial? I.e., does
the Flow-box Theorem still hold when we drop the C
1
condition on f? Yes and
no.
A transferring dieomorphism need not exist if f is merely Lipschitz (Ex-
ample 6 below). So the Flow-box Theorem does not trivially extend to the
Lipschitz case. The natural next hypothesis is that the Flow-box Theorem for
Lipschitz vector elds might work if we use a transferring lipeomorphism (a bi-
jective Lipschitz map whose inverse is also Lipschitz). Lipeomorphisms cannot
transfer vector elds, but they can still provide a conjugacy between two dy-
namics. The result of this paper, Theorem 4, is that for every non-equilibrium
of a Lipschitz vector eld there exists a local conjugacy to a constant vector
eld via lipeomorphism. Therefore the topological conjugacy with translation
still holds when a vector eld is not dierentiable.
Roughly, the trick in constructing the ow box is to track solutions to a
hyperplane transverse to the vector f (x
0
). The traditional proofs then employ
the Implicit Function Theorem or Inverse Function Theorem requiring dier-
entiability. For merely Lipschitz conditions, generalizations of those theorems
exist, but do not help when checking the transferring map is Lipschitz. We rely
on the Picard-Lindelof Theorem and Lipschitz continuous dependence on initial
conditions to nish the proof. We do not make use of Rademachers Theorem
which says a Lipschitz map is almost everywhere dierentiable.
For manifolds the Flow-box Theorem states that for any C
1
vector eld
with f(x) = 0 there is a chart around x on which f is constant. Proofs for C

Banach manifolds can be found in [2] or [8]. The results of this paper are easily
ported to this context: a vector eld is called locally Lipschitz continuous if it
is locally Lipschitz in one chart (and therefore all charts).
Thus the local qualitative characterization of dynamical systems under Lip-
schitz conditions reduces to the study of equilibria. This question has already
been broached, as dynamics with nonsmooth vector elds has enjoyed some pop-
ularity in the last few decades in control theory. Discontinuous vector elds have
been analyzed with a host of dierent approaches: see for instance [3], [4], [5],
[6], [7]. Even for the less extreme case of Lipschitz continuous vector elds, the
analysis of equilibria is ever more complicated than the smooth non-degenerate
hyperbolic case.
Interesting related results have been obtained in [1] concerning Lyapunov
exponents for systems generated by Lipschitz vector elds, and in [9] where the
Lie bracket is generalized to Lipschitz vector elds. The Flow-box Theorem
is the base case for Frobenius Theorem on the equivalence of involutive and
integrable distributions. [10] presents a generalization of Frobenius Theorem
1
Also known as The Cauchy-Lipschitz Theorem, The Fundamental Theorem of Dierential
Equations, or the Local Existence and Uniqueness Theorem. It is proven, e.g., in [2, p. 188].
2
for Lipschitz vector elds.
We nish with some examples that explore how the quality of continuity of
a vector eld (C
0
vs. Lipschitz vs. C
1
) are related to the quality of continuity
of the transferring map in the ow box setting.
2 Lipschitz Flow-box Theorem
A Banach space is a normed linear space, complete in its norm.
A map f : U V between subsets of Banach spaces is Lipschitz if there
exists K > 0 such that
f (x
1
) f (x
2
) Kx
1
x
2

for all x
1
, x
2
U. A lipeomorphism is an invertible Lipschitz map whose
inverse is also Lipschitz (i.e., slightly stronger than a homeomorphism). A
vector eld on a Banach space X is a map f : U X where U X. A
solution to a vector eld f with initial condition x is a curve
x
: I U
dened on an open interval I containing 0 such that
x
(0) = x and

x
(t) =
f (
x
(t)) for all t I.
Denote the open ball in X about x
0
X with radius r by
B(x
0
, r) := {x X : x x
0
< r} .
Theorem 1 (Picard-Lindel of ) Let X be a Banach space, x
0
X, and r > 0.
Assume f : B(x
0
, r) X is a Lipschitz vector eld, and let M be such that
f (x) M for all x B(x
0
, r). Then there exists a unique solution to f with
initial condition x
0
dened on
_

r
M
,
r
M
_
.
Proof. See, e.g., [2, p. 188] for the idea.
The following well-known result (also given in [2, p. 189]) is used in the
proof of the main theorem.
Lemma 2 (Continuous dependence on initial conditions) Let f be a Lip-
schitz vector eld with constant K dened on an open subset of a Banach space.
Let
x
and
y
be solutions to f for initial conditions x and y with interval I 0
contained in their common domains. Then

x
(t)
y
(t) x y e
K|t|
for all t I.
As a consequence of the two previous results, for any Lipschitz vector eld
f : U X on an open set U, near any point x U there exists a local ow.
Specically:
Corollary 3 Let f : U X be a Lipschitz vector eld on a Banach space X.
Let x U. There exists a neighborhood W of x in U, a number > 0, and a
map F : W (, ) X such that
3
1.
d
dt
F (x, t) = f (F (x, t))
2. F (x, 0) = x
3. F (x, s +t) = F (F (x, s) , t).
F is called a local ow of f near x.
Two vector elds f
1
: U
1
X and f
2
: U
2
X are called locally topolog-
ically conjugate near x
1
U
1
and x
2
U
2
if there exist open neighborhoods
W
1
of x
1
and W
2
of x
2
and a homeomorphism : W
1
W
2
with (x
1
) = x
2
such that a curve : I W
1
is a solution to f
1
if and only if : I W
2
is
a solution to f
2
. Less formally we require the local ows F
1
and F
2
to satisfy
(F
1
(x, t)) = F
2
((x) , t)
for all x W
1
and |t| suciently small.
Now we are ready to state and prove the main result of the paper.
Theorem 4 (Flow box) Let X be a Banach space with open subset U. Let
f : U X be a Lipschitz vector eld. Let z X be nonzero and let g : X X
be the constant vector eld g (x) = z. Then for any point x
1
U with f (x
1
) = 0,
f and g are locally topologically conjugate near x
1
and x
2
:= 0.
The homeomorphism which gives the conjugacy is a lipeomorphism.
Remark: In the usual formulation of the ow-box theorem in R
n
, the constant
vector eld g is usually chosen as g (x) = (1, 0, ..., 0).
Proof. (Outline) We may assume x
1
= 0 and f (0) = z. First specify a
hyperplane transverse to the vector f (x
1
). Then for x X near x
1
track
solutions
x
back to the plane . Dene t
x
to be the value of t such that

x
(t) and dene p
x
:=
x
(t
x
). Then the transferring lipeomorphism
we seek is (x) := p
x
+t
x
z. We check that supplies the conjugacy and nally
that is a lipeomorphism, but rst we must nd a neighborhood W
1
of x
1
on
which is well-dened. Finding W
1
is the dicult part of the proof and this
culminates at (2) below.
4
(Details) First we verify we may assume without loss of generality that x
1
=
0 and f (0) = z and z = 1. The translation to 0 and the dilation to norm one
are obvious. If z and f (0) =: y are linearly independent, then the intermediate
transferring dieomorphismA is a little more dicult to construct. Consider the
function from the subspace spanned by y and z to R given by (ay +bz) :=
b a. Extend to a continuous linear functional on X with the Hahn-Banach
Theorem. Then A : X X given by A(x) := x+(x) (y z) is its own inverse
and does the job.
We make several successive renements of a neighborhood of 0 in construct-
ing W
1
and the lipeomorphism : W
1
W
2
.
By the Hahn-Banach Theorem there exists a continuous R-linear map :
X R with (z) = 1 and |(x)| x for all x X. By the continuity of f
there is some r
1
> 0 such that
(f (x)) >
1
2
and f (x) < 2
for all x in the open ball B(0, r
1
). We can also assume r
1
is chosen so the
Lipschitz condition for f is met on all of B(0, r
1
) with constant K.
By Theorem 1 and the triangle inequality, for each x B(0, r
1
/2) a unique
solution to f with initial condition x exists,
x
: (T, T) B(0, r
1
) where
T := r
1
/4.
Denote the hyperplane in X which is the kernel of by
:= {x : (x) = 0} .
Dene
R :=
pB(0,r1/2)

p
((T, T)) .
With r
2
= min
_
r1
10
,
T
2
_
dene
W
1
:= B(0, r
2
) .
Next we show W
1
R. This will then guarantee that every point x W
1
is obtained by following a solution
p
with initial condition p . For x W
1
we know
x
((T, T)) B(0, r
1
). Then
(
x
)

(t) = (

x
(t)) = (f (
x
(t))) >
1
2
for T < t < T. Further
|
x
(0)| = |(x)| x < r
2
.
Thus there exists a unique t (2r
2
, 2r
2
) such that (
x
(t)) = 0, i.e.,
x
(t)
B(0, r
1
). Furthermore the speed of
x
is less than 2 so that the distance from
x to
x
(t) has 4r
2
as an upper bound. This follows since
x
(0)
x
(t)

_
t
0

x
(s) ds

. Thus the distance from 0 to

x
(t) is less than 5r
2

r1
2
and so
5

x
(t) B(0, r
1
/2) . Due to the uniqueness of solutions,
x(t)
(t) = x so
that x R and the claim is proven.
Thanks to the previous paragraph we can now dene . For each x W
1
since
x
is strictly increasing on (T, T), there exists a unique t
x
(T, T)
such that

x
(t
x
) B(0, r
1
/2) . (2)
Dene
p
x
:=
x
(t
x
)
and dene by
(x) := p
x
+t
x
z X
for each x W
1
.
Let F be the local ow of f near x
1
= 0 and let G be the ow of g. I.e.,
F (x, t) :=
x
(t) and G(x, t) = x + tz. To demonstrate the conjugacy we show
(F (x, t)) = G((x) , t) whenever x W
1
and |t| is suciently small. Notice
that the denitions of p
x
and t
x
above give
p
F(x,t)
= p
x
and t
F(x,t)
= t
x
+t
so that
(F (x, t)) = p
F(x,t)
+t
F(x,t)
z = p
x
+ (t
x
+t) z = G((x) , t) .
is 1-1. To see this suppose (x) = (y). Then p
x
p
y
= (t
y
t
x
) z.
Applying yields t
x
= t
y
so that p
x
= p
y
. By the uniqueness of solutions to f
we get x =
px
(t
x
) =
py
(t
y
) = y.
To show Lipschitz continuity we will use Lemma 2. Pick x, y W
1
. Since
(
x
)

(t) >
1
2
for all t (T, T),
|t
x
t
y
|
2 |(
x
) (t
x
) (
x
) (t
y
)|
2 (|(
x
) (t
x
) (
y
) (t
y
)| +|(
y
) (t
y
) (
x
) (t
y
)|)
2 (|(p
x
) (p
y
)| +
y
(t
y
)
x
(t
y
))
2
_
0 +x y e
K|ty|
_
Next, using the bound on speed f (x) < 2 gives
p
x
p
y
=
x
(t
x
)
y
(t
y
)

x
(t
x
)
x
(t
y
) +
x
(t
y
)
y
(t
y
)
2 |t
x
t
y
| +x y e
K|ty|
x y 5e
K|ty|
.
Since |t
y
| < T, dening K

:= 7e
KT
gives
(x) (y) = p
x
+t
x
z (p
y
+t
y
z) = (t
x
t
y
) z + (p
x
p
y
)
|t
x
t
y
| +p
x
p
y
K

x y .
6
Now we show
1
is Lipschitz. Pick u = p
x
+t
x
z = (x) and v = p
y
+t
y
z =
(y) then
_
_

1
(u)
1
(v)
_
_
= x y =
_
_

px
(t
x
)
py
(t
y
)
_
_

_
_

px
(t
x
)
py
(t
x
)
_
_
+
_
_

py
(t
x
)
py
(t
y
)
_
_
.
Using Lemma 2 again, we get
_
_

px
(t
x
)
py
(t
x
)
_
_
p
x
p
y
e
K|tx|
and the bound on speed f (x) < 2 gives
_
_

py
(t
x
)
py
(t
y
)
_
_
2 |t
x
t
y
| .
Dene the projection : X along z by
(q) := q (q) z.
This is a linear map and continuous since
(q) q +|(q)| z 2 q .
Then
p
x
p
y
= (u) (v) 2 u v
and
|t
x
t
y
| = |(u) (v)| u v .
Dene K

1 := 2 + 2e
KT
. Combining the above estimates gives
_
_

1
(u)
1
(v)
_
_
p
x
p
y
e
K|tx|
+ 2 |t
x
t
y
|
K

1 u v
so
1
is Lipschitz.
Finally since X might be innite dimensional, we must check W
2
:= (W
1
)
is open. Let p
x
+ t
x
z = (x) W
2
for x W
1
. Since W
1
is open there exists
s
1
> 0 such that B(x, s
1
) W
1
. Since t
x
(T, T), s
2
:= min
_
T |t
x
| ,
s1
4
_
>
0. Then pick s
3
> 0 such that B(p
x
, s
3
) B
_
0,
r1
2
_
and such that for all
p B(p
x
, s
3
) we have
p
(t
x
)
px
(t
x
) <
s1
2
(using Lemma 2). Then with
s
4
:= min
_
s
2
,
s3
2
_
> 0 we have B((x) , s
4
) W
2
. To see this notice any
member of B((x) , s
4
) may be written uniquely as p +tz for some p and
t R. Then
|t t
x
| = |([p
x
+t
x
z] [p +tz])|
[p
x
+t
x
z] [p +tz] < s
4
s
2
7
and
p p
x
= ([p
x
+t
x
z] [p +tz])
2 [p
x
+t
x
z] [p +tz] < 2s
4
s
3
.
Then

p
(t) x =
p
(t)
px
(t
x
)

p
(t)
p
(t
x
) +
p
(t
x
)
px
(t
x
) <
s1
2
+
s1
2
= s
1
so
p
(t) W
1
and therefore (
p
(t)) = p +tz (W
1
) = W
2
.
Remark 5 The Hahn-Banach theorem is essential for this proof. On a Hilbert
space or R
n
with arbitrary norm, however, an obvious modication (in the def-
inition of ) yields a proof which does not rely on the Axiom of Choice.
Example 6 The lipeomorphism constructed in the proof is not necessarily dif-
ferentiable. If a vector eld f : R
n
R
n
is smooth then the traditional Flow-box
Theorem states that near a non-equilibrium x
0
there exists a dieomorphism
from a neighborhood of x
0
to a neighborhood of the origin such that

f = g, i.e.,
d
x
(f (x)) = g ((x)), where g is the constant vector eld g (x) := (1, 0, ..., 0).
We show in this example that if f is not smooth, there may not exist such a
dieomorphism .
Consider the Lipschitz vector eld f : R
2
R
2
given by f (x, y) = (1 +|y| , 0).
If a transferring dieomorphism did exist near x
0
= (0, 0), then with :=
1
we would have
(d)
(x,y)
(g (x, y)) = f (x, y)
_

1x
(x, y)
1y
(x, y)

2x
(x, y)
2y
(x, y)
_ _
1
0
_
=
_

1x
(x, y)

2x
(x, y)
_
=
_
1 +|
2
(x, y)|
0
_
therefore there are functions u and v such that

2
(x, y) = u (y)

1x
(x, y) = 1 +|u (y)|

1
(x, y) = x +x|u (y)| +v (y)
Since (0, 0) = (0, 0) we must have u (0) = 0, and since
2x
(x, y) = 0 we cannot
have u

(y) = 0 lest the Jacobian be singular. Therefore there is no neighborhood

of the origin in which
1
is dierentiable.
Remark 7 In one dimension this whole exercise is pointless; the traditional
Flow-box Theorem applies to merely continuous vector elds. If f : R R is
continuous and f (x
1
) = 0 then there exists a dieomorphism which transfers
f near x
1
to the constant vector eld g = 1. Explicitly,
(x) =
_
x
x1
1
f (t)
dt
8
which is a dieomorphism near x
1
since f (x
1
) = 0. Therefore

(x) f (x) = 1
so (d)
x
(f (x)) = g ((x))
and so transfers f to g.
Example 8 Curiously, there exist discontinuous vector elds which can be trans-
ferred to constant vector elds by a lipeomorphism. Consider f, g : R R where
f (x) =
_
1 x < 1
2 x 1
and g (x) = 1. Then f is locally topologically conjugate to g via
(x) =
_
x x < 1
x+1
2
x 1.
References
[1] Abergel, F., Dimension Invariant Sets for Lipschitz Continuous Semigroups.
Indiana Univ. Math. J. 40 (1991), no. 3, 10291039.
[2] Abraham, R.; Marsden J; Ratiu, T., Manifolds, Tensor Analysis, and Ap-
plications. Addison-Wesley Publishing, 1983.
[3] Aubin, J.-P.; Frankowska, H., Set-Valued Analysis. Birkhauser, Boston,
Basel, Berlin, 1990.
[4] Broucke, M.; Pugh, C.; Simic, S., Structural Stability of Piecewise Smooth
Systems. Computational and Applied Mathematics, 20 (2001), no. 1-2, 51-
90.
[5] Clarke, F. H.; Ledyaev, Yu. S.; Stern, R. J.; Wolenski, P. R., Nonsmooth
Analysis and Control Theory. Springer, 1998.
[6] Deimling, K., Multivalued Dierential Equations. Walter de Gruyter, Berlin,
New York, 1992.
[7] Kunze, M., Non-Smooth Dynamical Systems. Springer 2000.
[8] Lang, S., Dierential and Riemannian Manifolds. Springer, 1995.
[9] Rampazzo, F.; Sussman, H. J., Commutators of ow maps of nonsmooth
vector elds, (2006), Preprint.
[10] Simic, S., Lipschitz Distributions and Anosov Flows. Proceedings of the
AMS, 124, no. 6 (1996), pp. 1869-1877.
9