Stable Manifold Theorem

Notes on the stable manifold theorem
University of Houston, Math 6324, Fall 2012, Vaughn Climenhaga

The stable manifold theorem is one of the most important in the theory
of non-linear ODEs and dynamical systems. Unfortunately, some of the
standard introductory texts (HirschSmale, Perko) either do not give a proof,
or do not motivate the proof, while more advanced texts (KatokHasselblatt,
BarreiraPesin) are too high-powered to be appropriate in the setting of an
introductory graduate course in ODEs, such as the one I am teaching now.
So Im taking this opportunity to turn my hastily scrawled hand-written
notes on a (hopefully) properly-motivated proof into something that will be
legible to myself and others on a more permanent basis.
The form of the theorem we will prove is this: let U Rn be an open
domain and t : U U the flow of a C 1 vector field f : U Rn . Suppose
that 0 is an equilibrium point for f and let E s Rn be the stable subspace for
Df (0) (the span of all generalised eigenvectors corresponding to eigenvalues
with negative real part). Let E cu = E c E u Rn be the centre-unstable
subspace for Df (0) (corresponding to eigenvalues with zero or positive real
part). Then there exists r > 0 and a C 1 function : B(0, r) E s E cu
such that the set W s := graph = {x + (x) | x B(0, r) E s } has the
following properties:
1. W s is positively invariant;
2. given an initial condition x W s , we have limt t (x) = 0.
(Note that if f has an equilibrium point x 6= 0, we can make the change
of coordinates y = x x and proceed as stated above. The choice x = 0
simplifies the notation.)
Proof: We present a proof using Perrons method this is essentially the
proof given in Perkos book, but with significantly more motivation and a
more abstract viewpoint. The idea is to use the same general strategy that
worked well for us in the proof of the PicardLindelof theorem on local existence and uniqueness of solutions: prove existence and uniqueness of something by obtaining as the unique fixed point of a contraction on a complete
metric space. In the PicardLindelof theorem, the space was the space of
approximate solutions, and the contraction was the Picard operator that
transformed a candidate solution into something even closer to being a solution. Here we need to consider a slightly different space and operator.
1. We will consider the space of curves that approach the fixed point 0
with a certain exponential rate, as shown in Figure 1. In particular,
for each a E s we will consider the set of such curves that begin at
a + y for some y E cu . As before, we do not assume a priori that these
curves are trajectories of the ODE.
2. We will define an integral operator on this space whose fixed points are
solutions of the ODE. Then an application of the Banach fixed point
theorem completes the proof.
Figure 1: Approximate trajectories approaching 0 exponentially.

To this end, let A = Df (0) be the linear part of f at the fixed point, and
let P = A|E s and Q = A|E cu be the restrictions of A to the stable subspace
and centre-unstable subspace, respectively. Let be the maximum value
of Re for eigenvalues of P , and fix constants 0 < < < < , as shown
in Figure 2. Then all eigenvalues of P lie strictly to the left of and all
eigenvalues of Q lie strictly to the right of . In particular, we can fix a
norm on Rn with the property that
keP t k et ,
keQt k et
(1)
for all t 0. (Note that the second of these does not give decay to 0, but
does give some control on growth.)
2
Figure 2: Bounds on eigenvalues corresponding to E s and E cu .

We work in the space of all C 1 curves approaching 0 with exponential
rate at least :

n
1
t
X := x : [0, ) R x is C , kxk := sup |x(t)|e < .
(2)
t0
This is a complete metric space with distance given by

d(x, y) = kx yk = sup |x(t) y(t)|et .
t0
(The reader familiar with Banach spaces will note that X is in fact a Banach
space.)
Fixing a E s , we consider the subspace Xa = {x X | xs (0) = a}.
Here and throughout we will write xs E s and xcu E cu for the parts of
x lying in the stable subspace and centre-unstable subspace, respectively, so
that x = xs + xcu . The goal is to find, for each sufficiently small a E s ,
some x Xa that is a trajectory of the ODE x = f (x). Then we can put
(a) = xcu (0) and conclude that has the properties claimed in the theorem.
(In fact we will not prove that is C 1 , which takes some more work, but
positive invariance and exponential stability will follow from what we show.)
3
Having defined our metric space Xa , we need to define an operator on it

whose fixed points are trajectories of the system. As in the PicardLindelof
theorem we would like to define an integral
R t operator. We cannot simply use
the one defined there ((Px)(t) = x(0)+ 0 f (x(s)) ds) for at least two reasons:
1. we need P to preserve the property x(t) 0, which this does not;
2. we will need to have the possibility that (Px)(0) 6= x(0), otherwise the
operator will preserve the subsets of Xa on which xcu (0) is constant,
and in particular each of these subsets could have a solution of the
ODE, but the whole point is that Xa should only contain one value of
xcu (0) corresponding to a solution.
Recall the idea from variation of constants: treat the ODE as a perturbation of a linear system, and write down an expression that is constant for the
linear system. Then differentiating this expression gives the effect of the perturbation at each time, and integrating yields an expression for the solution
of the full non-linear system. In other words, we write f (x) = Ax + F (x),
where F (x) = o(|x|), so that the ODE becomes
x = f (x) = Ax + F (x).
(3)
If F (x) 0 then the solution is x(t) = eAt x(0), and in particular eAt x(t) is
constant. So we differentiate this expression and use (3) to get
d At
(e x(t)) = AeAt x(t) + eAt (Ax(t) + F (x(t)))
dt
= eAt F (x(t)).
Integrating, we see that (3) is equivalent to
Z t
At
x(t) = e x(0) +
eA(ts) F (x(s)) ds,
(4)
or more generally, by integrating from T to t for some t 0,

Z t
A(tT )
x(t) = e
x(T ) +
eA(ts) F (x(s)) ds.
T
(5)
Now by decomposing everything into the part lying in E s and the part lying
in E cu , we see that x(t) is a solution of (3) if and only if we have
Z t
s
Pt s
x (t) = e x (0) +
eP (ts) F s (x(s)) ds,
0
(6)
Z t
cu
Q(tT ) cu
Q(ts) cu
x (t) = e
x (T ) +
e
F (x(s)) ds
T
for some (and hence every) T 0. The reason for using non-zero values of
T in the second equation is that we have good control on eQ when < 0
but not when > 0, and so by sending T we may observe that
|eQ(tT ) xcu (T )| keQ(tT ) k|x(T )| e(tT ) kxkeT = et kxke(+)T 0,
recalling the relationship between , illustrated in Figure 2. In particular,
we may replace (6) with
Z
cu
eQ(ts) F cu (x(s)) ds.
(7)
x (t) =
t
The conclusion is that if we write

Z t
Z
Pt s
P (ts) s
(Px)(t) = e x (0) +
e
F (x(s)) ds
0
eQ(ts) F cu (x(s)) ds, (8)
then x = Px if and only if x solves (3). One may also understand the
motivation behind sending T as follows: for a fixed T , we expect
the first term from (6) to grow exponentially as t , which makes it
difficult to verify the property (Px)(t) 0. By eliminating this term (and
the corresponding growth in the integral), we obtain an expression that is
more tractable in the limit t .
Now we show that P maps Xa to itself, and that it is a contraction,
which will complete the proof by an application of the Banach fixed point
theorem. First we observe that because the non-linear part F of the vector
field f is C 1 with DF (0) = 0, it is Lipschitz on small neighbourhoods of
0, and the Lipschitz constant can be made arbitrarily small by making the
neighbourhood small enough. More precisely, for every > 0 there exists
r > 0 such that if |x|, |y| r then |F (x) F (y)| |x y|.
Now we make a computation that shows both claims in the previous
paragraph. Fix x, y X and let r, be as above. Recall that by the definition
5
of d(x, y) we have |x(t) y(t)| d(x, y)et for all t 0, and write s =
|xs (0) y s (0)|. Then
|(Px)(t) (Py)(t)|
Z
keP (ts) k|F s (x(s)) F s (y(s))| ds

ke k|x (0) y (0)| +
0
Z
keQ(ts) k|F cu (x(s)) F cu (y(s))| ds
+
Z
Zt t
(ts)
t s
e(ts) |x(s) y(s)| ds
e
|x(s) y(s)| ds +
e +
0

Z t
Zt
(ts) s
(ts) s
t s
e
e ds +
e
e ds
e + d(x, y)
Pt

()s t
()s
t s
t e
t e
e + d(x, y) e
+e
.
s=0
s=t
Using the fact that < 0 and > 0, we obtain
t

e et
et
t s
|(Px)(t) (Py)(t)| e + d(x, y)

1
1
t s
t
+
e + d(x, y)e

( )
d(x, y) et .
= s +
( )( )
Writing L =
()()
and recalling the definition of k k in (2), this gives
kPx Pyk s (x, y) + Ld(x, y)
(9)
Given x X, by putting y = 0 we see that

kPxk |xs (0)| + Lkxk < ,
and so P maps X to itself. Moreover, if x Xa then it is apparent from
(8) that Px Xa as well. Finally, if x, y Xa for some a E s , then
s (x, y) = 0 and (9) gives
d(Px, Py) Ld(x, y).
6
By choosing r small enough we can guarantee that L < 1, and hence P is a

contraction. Thus it has a unique fixed point x Xa . This is a trajectory of
the ODE (3) which approaches 0 exponentially (with rate at least ) and has
xs (0) = a, so we put (a) = xcu (0) and conclude that has the properties
claimed in the statement of the theorem.
A few remarks are in order. The proof here follows the proof in Perkos
book (p. 107111), but includes more abstract language and more complete
motivation. We have shown that exists and its graph W s is positively
invariant with trajectories converging exponentially to 0. One can also show
that is C 1 and D(0) = 0, so that W s is tangent to E s . This proof is
not in Perkos book he refers to the text of Coddington and Levinson (p.
332333). If I feel ambitious I may try to include this part of the proof at a
later point.
The proof given above also shows that x Xa is the only trajectory
with xs (0) = a that converges exponentially to 0 with rate at least . In
particular, any such trajectory lies in the stable manifold W s . However, it
is possible that there are trajectories lying outside of W s that converge to 0
more slowly for example, there may be trajectories converging to 0 with
subexponential speed along the centre direction E c .
Ultimately, the only property of the subspace E cu that we used in the
proof was the fact that all eigenvalues of Q = Df (0)|E cu have real part
strictly greater than . In fact, the entire proof above is valid if we replace
E cu with a subspace E 2 Rn such that
1. Rn = E s E s ; and
2. there exists < 0 such that every eigenvalue of P = Df (0)|E s has
real part strictly less than , and every eigenvalue of Q = Df (0)|E s
has real part strictly greater than .
In this setting, we still get a stable manifold, which is often called a strong
stable manifold and denoted W ss , because it corresponds to the directions
in which the contraction is the strongest. There may be trajectories lying
outside of W ss that approach 0 exponentially, but the rate at which they do
so is less than .

Stable Manifold Theorem

Uploaded by

Copyright:

Available Formats

Stable Manifold Theorem

Uploaded by

Document Information

Original Description:

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

Stable Manifold Theorem

Uploaded by

Copyright:

Available Formats

Notes on the stable manifold theorem

University of Houston, Math 6324, Fall 2012, Vaughn Climenhaga

Figure 1: Approximate trajectories approaching 0 exponentially.

Figure 2: Bounds on eigenvalues corresponding to E s and E cu .

This is a complete metric space with distance given by

Having defined our metric space Xa , we need to define an operator on it

or more generally, by integrating from T to t for some t 0,

The conclusion is that if we write

eQ(ts) F cu (x(s)) ds, (8)

keP (ts) k|F s (x(s)) F s (y(s))| ds

|(Px)(t) (Py)(t)| e + d(x, y)

and recalling the definition of k k in (2), this gives

kPx Pyk s (x, y) + Ld(x, y)

Given x X, by putting y = 0 we see that

By choosing r small enough we can guarantee that L < 1, and hence P is a

You might also like