POLONICI MATHEMATICI LXXIII (2000)
Topological conjugacy of cascades generated by gradient flows on the two-dimensional sphere
by Andrzej Bielecki (Krak´ow)
Abstract. This article presents a theorem about the topological conjugacy of a gra- dient dynamical system with a constant time step and the cascade generated by its Euler method. It is shown that on the two-dimensional sphere S
2the gradient dynamical flow is, under some natural assumptions, correctly reproduced by the Euler method for a suffi- ciently small time step. This means that the time-map of the induced dynamical system is globally topologically conjugate to the discrete dynamical system obtained via the Euler method.
1. Introduction. In recent years several papers have been devoted to studying the qualitative properties of discrete-time dynamical systems ob- tained via discretization methods. The basic question is whether the quali- tative properties of continuous-time systems are preserved under discretiza- tion. Various concepts of differentiable dynamics were investigated. Re- sults on stability and attraction properties ([KL]), the saddle-point structure about equilibria ([AD], [Bey1], [Bey2]), invariant manifolds ([BL], [Fec1]), averagings ([Fec2]) and algebraic-topological invariants ([MR]) can be men- tioned as examples. A number of applications have been studied as well ([Gar4]). The investigations are concerned with both local (see, for instance, [Gar1], [Fec3]) and global conjugacy ([Gar2], [Gar3], [Gar5]).
This paper is devoted to the problem of topological conjugacy between the discretization of a gradient dynamical system and the cascade generated by its Euler method. Similar problems have been solved in recent years for numerical methods of order greater than one (see [Gar2], [Li]).
2. Topological conjugacy of gradient cascades.As mentioned above we consider a gradient differential equation and its Euler method. The time- map of the induced solution is compared to the cascade obtained via the
2000 Mathematics Subject Classification: 34C35, 34A50.
Key words and phrases: topological conjugacy, gradient dynamical system, Euler method.
[37]
Euler method. We show that on the two-dimensional sphere S
2a gradient dynamical system is, under some natural assumptions, correctly reproduced by the Euler method for a sufficiently small time step. This means that the time-map of the induced dynamical system is globally topologically conju- gate to the discrete dynamical system obtained via the Euler method. This fact can be expressed as follows.
Theorem 2.1. Let S
2be the unit sphere in R
3and let φ : S
2× R → S
2be the dynamical system generated by a differential gradient equation
(1) x = − grad E(x), ˙
where E ∈ C
2(S
2, R), having a finite number of singularities, all hyperbolic.
Let , furthermore, the dynamical system φ have no saddle-saddle connec- tions. Moreover , let φ
h: S
2→ S
2be the discretization of φ, i.e. φ
h(x) :=
φ(x, h), and let ψ
h: S
2→ S
2be generated by the Euler method for (1).
Then, for sufficiently small h > 0, there exists a homeomorphism α = α
h: S
2→ S
2globally conjugating the cascades generated by φ
hand ψ
h, i.e.
(2) φ
h◦ α = α ◦ ψ
h.
Remarks. 1. Axiom A and the strong transversality condition are known to be equivalent to the structural stability of a dynamical system (see [PM], p. 171, and [Man]). On the other hand, for gradient dynamical systems, Axiom A implies that the system has only a finite number of singularities, all hyperbolic, whereas the strong transversality condition implies that the system has no saddle-saddle connections. Thus, the structural stability of a dynamical system (S
2, φ) implies the assumptions of Theorem 2.1. Moreover, the set of structurally stable systems is open and dense in the space of gradient dynamical systems (see [PM], p. 116).
2. A dynamical system generated by the equation (1), having only a finite number of singularities, all hyperbolic, without saddle-saddle connections is called a gradient Morse–Smale system.
3. Estimation of the Euler method on S
n. Let n
0∈ N and let a = n
0h denote the length of the time interval on which the error e
n:=
%
M(φ
nh0(x), ψ
hn0(x)) is estimated (%
Mdenotes the Riemannian metric on M). We will show that on the n-dimensional sphere
(3) e
n< ξ(a)h,
where, for a given problem, ξ(a) is a constant value which only depends on a.
Consider the problem
(4) x = f (x), ˙ x(0) = x
0, 0 ≤ t ≤ a,
on a compact manifold M, where f is a vector field on M. The Euler iterative rule for the problem (4) is of the form
(5) x
n= exp
xn(−hf (x
n−1)), where f (x
n−1) is a vector of the tangent space T
xn−1M.
The Euler method in R
nis a first order numerical method ([Kru], p. 31).
For x, y ∈ M in the domain of a chart ϑ of a manifold M, we have (6) m
1%(ϑ(x), ϑ(y)) ≤ %
M(x, y) ≤ m
2%(ϑ(x), ϑ(y)),
where m
1, m
2are constant for a given chart ϑ (see [Rob], p. 453, formula (2.2)); % is the euclidian metric on R
n. By (6), the Euler method on a compact manifold is also a first order method.
The error in a single step of a numerical method is defined by r(x, h) := %
M(ψ
h(x), φ
h(x))
and is a continuous function of x for a constant h. Therefore, on a compact manifold, it reaches its maximum. Set
r(h) := max
x∈M
r(x, h).
Let us estimate the error of the Euler method for a gradient equation on the sphere S
nin R
n+1. By the assumptions of Theorem 2.1 a gradient dynamical system on a compact manifold has at least one attracting singu- larity. Change coordinates in R
n+1so that this fixed point is the south pole of the sphere. The sphere can be covered by two charts ϑ
1, ϑ
2with
(7) ϑ
1: S
n\ p
north→ R
n, ϑ
2: S
n\ p
south→ R
n, where p
north, p
southare the poles of the sphere.
If the south pole is the starting point of the Euler method then
%
M(φ
nh(x), ψ
nh(x)) = 0
for each n ∈ N as the gradient is zero at the south pole. In this case the error is zero.
In the other case, if the starting point x is different from p
souththen
φ
nh(x) 6= p
southfor each n ∈ N, because ψ
his invertible for h sufficiently
small. Thus, for each n ≤ n
0, both φ
nh(x) and ψ
nh(x) lie in the domain
of the chart ϑ
2. Therefore, the point x and the systems φ
hand ψ
hcan
be transformed into R
nin order to perform the iterations e ψ
nh0(ϑ
2(x)) and
φ e
nh0(ϑ
2(x)). Afterwards, we return to the sphere via ϑ
−12. The dynamical
systems e ψ
hand e φ
hare the systems ψ
hand φ
htransformed from the sphere
to R
nby the maps of the atlas. Notice that if ψ
his generated by the Euler
method of the gradient equation generating φ
hon the sphere then e ψ
his
generated by the Euler method of the equation generating e φ
hin R
nbecause
the atlas preserves the differential structure.
The error estimate with step h for the equation
˙
x = f (x),
where f is lipschitzian with a constant L in R
n, is given by the formula (see [Kru], p. 32, formula 1.22)
(8) e
n≤
e
0+ rh L
e
La,
where a = n
0h is the time interval on which the solution is considered, rh
2is the maximal error in a single step and e
0is the initial error. Since the right side of (1), being a differentiable map on a compact space, is a lipschitzian map, so is the right side of the equation transformed to R
n, by (6). Therefore, the estimate (8) can be applied to the systems e ψ
hand e φ
h. The flow and its Euler method start from the same point, thus e
0= 0 and
%( e ψ
hn0(ϑ
2(x)), e φ
nh0(ϑ
2(x))) ≤ rh L e
La. By (6) we get the following estimate on S
n:
(9) %
M(ψ
hn0(x), φ
nh0(x)) ≤ m
2rh L e
La,
which is of the same form as in R
n. Thus we have obtained (3).
4. Lemmas. One of the key points in the proof of Theorem 2.1 is a construction of proper homeomorphisms conjugating the cascades φ
hand ψ
hin a neighbourhood of an attracting singularity (see Lemma 4.6). This construction is based on the following geometric lemma.
Lemma 4.1. Let γ
1, γ
2, δ
1, δ
2be curves in R
2, each parametrized by τ ∈ [0, 1], homeomorphic to a line segment and such that γ
1(0) = δ
1(0), γ
1(1) = δ
2(0), γ
2(0) = δ
1(1) and γ
2(1) = δ
2(1). Assume that their union is the boundary of a simply connected domain D. Then there exists a homeo- morphism Λ : D → [0, 1]
2such that Λ(γ
1) = [0, 1] × {1}, Λ(γ
2) = [0, 1] × {0}, Λ(δ
1) = {1} × [0, 1], Λ(δ
2) = {0} × [0, 1].
The proof of this well known fact is omitted. It can be found in [Bie].
A great number of theorems concerning the topological conjugacy near hyperbolic singularities are known. We will need the following theorem (see [Bey], [Gar2], [Gar3]).
Theorem 4.2. For each equilibrium point x
0of the cascade φ
hand for
sufficiently small h there exists an equilibrium point x
hof the cascade ψ
h,
a neighbourhood V
x0of x
0, and a homeomorphism α
h: V
x0→ α
h(V
x0) such
that α
h(x
0) = x
hand (α
h◦ φ
h)(x) = (ψ
h◦ α
h)(x), whenever x and φ
h(x)
are in V
x0.
In our case, since x
his an equilibrium point of ψ
hiff gradE(x
h) = 0, we have x
h= x
0for every h > 0.
Throughout this section M denotes a compact differentiable manifold of dimension greater than one, φ denotes a Morse–Smale gradient system on M, and φ
hand ψ
hare the cascades defined as in Theorem 2.1. Let, furthermore,
g
ci: V
ci→ V
c∗ibe a homeomorphism locally conjugating φ
hand ψ
hin the neighbourhood V
ciof the singularity c
iand V
c∗i= g(V
ci).
Lemma 4.3. Let c
ibe an attracting or saddle singularity. Then there exists r
i> E(c
i) such that for every x 6= c
iin W
φs(c
i) ∩ V
cithe following implication holds: if E(x) ≤ r
ithen there exists t > 0 such that
E(φ(x, −t)) > r
iand φ(x, −t) ∈ W
φs(c
i) ∩ V
i. Furthermore the sets
K
ci,0: = {x ∈ W
φs(c
i) : E(x) ≤ r
i},
K
ci,1: = {x ∈ W
φs(c
i) : E(x) ≤ r
i}, where E(c
i) < r
i< r
i, are nonempty.
P r o o f. The point c
iis attracting on the stable manifold W
φs(c
i). Choose R < min{%
M(c
i, c
j) : i 6= j, c
jis a fixed point}.
Let B
rel(c
i, R) be the closed ball and S
rel(c
i, R) the sphere in W
φs(c
i). By compactness there exists x
0∈ S
rel(c
i, R) such that
E(x
0) = inf{E(x) : x ∈ S
rel(c
i, R)} =: e
i.
The following proposition is necessary to complete the proof of Lemma 4.3. Its proof follows simply from the Morse lemma and therefore is omitted (it can be found in [Bie]).
Proposition 4.4. Let e
i> r
∗> E(c
i). Then the set B
lev(c
i, r
∗) := {x ∈ W
φ,cs i: E(x) ≤ r
∗} and the ball B
rel(c
i, R) are homeomorphic.
Since the potential E decreases along each orbit, every r
i< r
∗(where r
∗is defined in Corollary 4.4) satisfies the implication in Lemma 4.3. That implies that the sets K
ci,m, m = 0, 1, are nonempty.
Lemma 4.5. The sets K
ci,m, m = 1, 2, have the following properties:
(i) c
i∈ int
relK
ci,m,
(ii) ∂
relK
ci,0= {x ∈ W
φ,cs i: E(x) = r} and ∂
relK
ci,1= {x ∈ W
φ,cs i: E(x) = r},
(iii) {φ(x, t) : t > 0} ⊂ K
ci,mfor every x ∈ K
ci,m,
(iv) K
ci,mis arcwise connected ,
(v) inf{%
M(x, y) : x ∈ K
ci,1, y ∈ K
ci,0} > 0.
The boundaries and interiors are considered in the relative topology of W
φ,cs iand therefore are marked by the subscript “rel”.
P r o o f. (i) This is clear since K
ci,mis homeomorphic to a ball B
rel(c
i, r).
(ii) Let x
0∈ ∂
relK
ci,0. Then, for every neighbourhood V
x0of x
0, V
x0∩ {x ∈ W
φ,cs i: E(x) ≤ r} 6= ∅ and V
x0∩ {x ∈ W
φ,cs i: E(x) > r} 6= ∅.
Since the map E is continuous, E(x
0) is equal to r.
Let x
0∈ W
φ,cs i. The definition of K
ci,mimplies that x
0∈ K
ci,0. Assume, by contradiction, that x
0∈ int
relK
ci,0. Then there exists t
0> 0 such that φ(x
0, −t) ∈ K
ci,0for every t ≤ t
0. However, for those t,
E(φ(x
0, −t)) > E(φ(x
0, 0)) = E(x
0) = r.
This means that φ(x
0, t) 6∈ K
ci,0, a contradiction.
The property of K
ci,1can be proved in the same way.
(iii) Let x
0∈ ∂
relK
ci,0. Since the potential decreases along orbits, E(φ(x
0, t)) < E(x
0) for every t > 0. Hence φ(x
0, t) ∈ K
ci,0by the defi- nition of K
ci,m.
If x
0∈ int
relK
ci,0, then E(x
0) < r. Since the set K
ci,0= B
lev(c
i, r) is homeomorphic to a ball B
rel(c
i, R) such that K
ci,0⊂ B
rel(c
i, R) (see Corollary 4.4), the set {φ(x
0, t) : t > 0} would intersect the boundary
∂
relK
ci,0if the semiorbit φ(x
0, t), t > 0 were not included in K
ci,0. However this is impossible because the potential on the boundary of the set K
ci,0is greater than in the interior.
The property of K
ci,1can be shown in the same way.
(iv) Define τ : R 3 t 7→ τ (t) ∈ [−1, 1] by τ (t) =
( −1 if t = −∞,
t/(1 + |t|) if t ∈ R,
1 if t = ∞.
Let x
1, x
2∈ K
ci,m, where m equals 1 or 2. Then the formulas γ
12τ (t) = φ(x
1, t) and γ 1 −
12τ (t) = φ(x
2, t)
define an arc from x
1to x
2; the arc is included in K
ci,mby Lemma 4.5(ii).
(v) The assertion follows from the compactness of M. Assume, by con- tradiction, that inf{%
M(x, y) : x ∈ K
ci,1, y ∈ K
ci,0} = 0. Then there exists a sequence {x
n} ⊂ K
ci,1which converges to a point x
0∈ ∂
relK
ci,0. But E(x
0) = r and E(x
n) ≤ r, which is a contradiction because r < r and E is continuous.
The following lemma is necessary for the construction of a global hom-
eomorphism conjugating φ
hand ψ
h.
Lemma 4.6. Let a be an attracting fixed point of the flow φ on a two- dimensional manifold M. Let , furthermore, s
k, k ∈ {1, . . . , K}, be saddle points whose stable manifolds intersect the stable manifold of a. Then there exists a neighbourhood V
aof a and a homeomorphism g
adefined on this neighbourhood locally conjugating the flows (M, φ
h) and (M, ψ
h) such that for every k ∈ {1, . . . , K},
(10) g
a(W
φuh(s
k) ∩ V
a) ⊂ W
ψuh(s
k).
P r o o f. We first record some facts and introduce a few definitions. As M is two-dimensional, the components of the unstable manifolds W
φuh
(s
k) of the saddle points are curves with a common end at a. Choose a level set of E on which E > E(a) and such that its connected component is included both in W
φsh(a
i) and in a sufficiently small neighbourhood of a. Denote this component by F
1. Since near an attracting singularity a level is homeo- morphic to a circle (see Corollary 4.4), this level can be parametrized as F
1= F
1(τ ) with F
1(0) = F
1(1), τ ∈ [0, 1], where the map F
1is continuous.
The set W
φu(s
k) \ {s
k, a} has two connected components, each being an orbit of φ. As already mentioned, the curve F
1is homeomorphic to a circle and the value of the potential decreases along an orbit. Thus, every connected component of S
Kk=1
(W
φuh
(s
k) \ {s
k, a}) intersects F
1in exactly one point.
Let there be M such components and let they be numbered from 0 to M − 1 according to the increasing τ, say W
φuh,m, m = 0, . . . , M − 1.
Set F
2:= φ
h(F
1). The curve F
2does not intersect F
1and is homeomor- phic to F
1. The point a lies in the interior of the domain bounded by F
2. Therefore the set
P
F1,F2:= {φ(x, t) : x ∈ F
1, t ∈ [0, h]}
is homeomorphic to a closed annulus B(a, R) \ int B(a, r), where r < R.
Define
δ
m:= W
φuh,m∩ P
F1,F2, m = 0, . . . , M − 1, (11) Λ : [0, 1] × [0, h] 3 {τ, t} 7→ Λ(τ, t) = φ(F
1(τ ), t) ∈ P
F1,F2.
The mapping Λ is a homeomorphism because it is a superposition of homeo- morphisms and there exists a finite sequence τ
0= 0, τ
1, . . . , τ
M −1such that Λ(τ
m, [0, h]) = δ
m.
Now, we begin the proof of Lemma 4.6.
Step 1: Construction on fragments of W
φuh(s
k). Let g
skdenote a home- omorphism locally conjugating φ
hand ψ
hnear s
kand let φ
−hbe the inverse of φ
h. The formula
(12) g
sk,a(x) := (ψ
hn0(x)◦ g
sk◦ φ
n−h0(x))(x)
defines a homeomorphism conjugating φ
hand ψ
hon δ
m. The natural number
n
0(x) is chosen in such a way that φ
n−h0(x)(x) ∈ K
sk,1and φ
n−h0(x)−1(x) ∈ int
rel(K
sk,0\K
sk,1). Note that the parameter m uniquely identifies the saddle point s
k.
The mapping
g
a:= g
sk,a, x ∈ δ
m,
defines a homeomorphism conjugating φ
hand ψ
hon S
M −1m=0
δ
m. The map- pings ψ
h, g
skand g
aare continuous in h and each is the identity for h = 0.
Step 2: Construction on a fragment of the annulus. Set
γ
1,m:= {x = F
1(τ ) : τ ∈ [τ
m, τ
m+1]}, γ
2,m:= {x = F
2(τ ) : τ ∈ [τ
m, τ
m+1]}, where τ
mis the value of τ for which
F
j(τ
m) = F
j∩ W
φuh,m=: x
j(m), j ∈ {1, 2}.
Note that γ
j,m, j = 1, 2, m = 0, . . . , M − 1, is a curve with end points x
j(m), x
j(m + 1) whereas δ
mis a curve with end points x
1(m), x
2(m). All these curves are homeomorphic to line segments. This implies that γ
1,m∪ γ
2,m∪ δ
m∪ δ
m+1bounds a simply connected domain in R
2. By the Riemann Theorem this domain is homeomorphic to a ball and hence to a rectangle.
By Lemma 4.1 there exists a homeomorphism Λ
m:= Λ|
[τm,τm+1]×[0,h]from [τ
m, τ
m+1] × [0, h] onto the closed domain Sq
γ1,m,γ2,m,δm,δm+1bounded by the curves γ
1,m, γ
2,m, δ
mand δ
m+1. The mapping Λ has the following properties:
Λ
m(τ
m, [0, h]) = δ
m, Λ
m([τ
m, τ
m+1], 0) = γ
1,m, Λ
m([τ
m, τ
m+1], h) = γ
2,m.
Perform a similar construction for the cascade ψ
h. Set x
∗j(m) := g
a(x
j(m)).
Let a curve F
1∗:= g
a(F
1) parametrized by τ be homeomorphic to a circle and such that a is in the interior of the simply connected domain bounded by F
1∗. Let, furthermore, the points x
∗1(m), m ∈ {0, . . . , M − 1}, lie in F
1∗and let the parametrization have the property
F
1∗(τ
m) = x
∗1(m), m = 0, . . . , M − 1.
Let F
2∗:= ψ
h(F
1∗) and suppose ψ
h(F
1∗(τ )) = F
2∗(τ ) for all τ. Set δ
∗m:= g
a(δ
m),
and let δ
m∗be parametrized in such a way that
δ
m∗(h
0) := g
a(δ
m(h
0)), h
0∈ [0, h].
The homeomorphism g
ais a continuous function of h and is the identity for h = 0, so for sufficiently small h, the curve δ
∗mlies near δ
mand does not intersect any other δ
m0, m
06= m. Thus, if m increases then so does the value of τ which parametrizes F
1∗. Therefore, F
1∗can be reparametrized in such a way that the values of τ at x
j(m) equal those at x
∗j(m), m = 0, . . . , M − 1.
Define
γ
1,m∗:= {x ∈ F
1∗(τ ) : τ ∈ [τ
m, τ
m+1]}, γ
2,m∗:= {x ∈ F
2∗(τ ) : τ ∈ [τ
m, τ
m+1]}.
Arguing as for φ
h, we can show that the domain bounded by F
1∗and F
2∗is homeomorphic to an annulus whereas the domain Sq
γ∗1,m,γ2,m∗ ,δm∗,δ∗m+1
bounded by γ
1,m∗, γ
2,m∗, δ
m∗and δ
m+1∗is homeomorphic to a rectangle. By Lemma 4.1 the homeomorphism can be chosen to map the vertices of the rectangle to the points x
∗1(m), x
∗2(m), x
∗1(m + 1) and x
∗2(m + 1). Denote it by
ξ
m: Sq
γ∗1,m,γ2,m∗ ,δ∗m,δ∗m+1
→ [τ
m, τ
m+1] × [0, h].
It has the following properties:
ξ
m(γ
1,m∗) = [τ
m, τ
m+1] × {0}, ξ
m(γ
2,m∗) = [τ
m, τ
m+1] × {h},
ξ
m(δ
∗m) = {τ
m} × [0, h].
Furthermore, ξ
mcan be constructed in such a way that on the curves δ
m∗it is consistent with the mapping g
m,a(the parameter m determines the saddle point s
kuniquely):
(13) g
m,a(φ(x
m, t)) = ξ
m(τ
m, t), t ∈ [0, h], and
(14) g
m+1,a(φ(x
m+ 1, t)) = ξ
m(τ
m+1, t), t ∈ [0, h].
This can be shown in the following way. Equations (13) and (14) imply that on the vertical sides τ
m×[0, h] and τ
m+1×[0, h] the parametrization is settled by increasing homeomorphisms f
1: [0, h] → [0, h] and f
2: [0, h] → [0, h].
We will show that the mapping g : [τ
m, τ
m+1] × [0, h] 3 {τ, t} 7→
g(τ, t) = (g
1(τ, t), g
2(τ, t)) ∈ [τ
m, τ
m+1] × [0, h]
defined by
g
1(τ, t) = τ, g
2(τ, t) = τ
m+1− τ
τ
m+1− τ
m· f
1(t) + τ − τ
mτ
m+1− τ
m· f
2(t)
is a homeomorphism with the required properties. We will prove that it
transforms the rectangle [τ
m, τ
m+1] × [0, h] onto itself. Fix τ
0∈ [τ
m, τ
m+1]
and t
0∈ [0, h]. Without losing generality we can assume that f
1(t
0) ≤ f
2(t
0).
Then
g
2(τ
0, t
0) ≥ τ
m+1− τ
0τ
m+1− τ
m· f
1(t
0) + τ
0− τ
mτ
m+1− τ
m· f
1(t
0) = f
1(t
0) ∈ [0, h]
and
g
2(τ
0, t
0) ≤ τ
m+1− τ
0τ
m+1− τ
m· f
2(t
0) + τ
0− τ
mτ
m+1− τ
m· f
2(t
0) = f
2(t
0) ∈ [0, h].
Furthermore, g
2(τ
m, ·) = f
1(·) and g
2(τ
m+1, ·) = f
2(·). The jacobian jac g = ∂g
1/∂τ ∂g
1/∂t
∂g
2/∂τ ∂g
2/∂t
= τ
m+1− τ τ
m+1− τ
m· df
1(t)
dt + τ − τ
mτ
m+1− τ
m· df
2(t) dt is positive at each point because both components are nonnegative (f
1and f
2are increasing) and do not equal zero simultaneously.
The superposition
Υ
m= Λ
m◦ g ◦ ξ
mis a homeomorphism transforming the closure of Sq
γ∗1,m,γ∗2,m,δ∗m,δm+1∗
onto the closure of Sq
γ1,m,γ2,m,δm,δm+1in such a way that for each x ∈ δ
m∪ δ
m+1,
(15) Υ
m(x) = g
m,a(x).
Step 3: Construction on the annulus. By (13) and (14), Υ
m(x)=Υ
m+1(x) for each x ∈ δ
m+1and Υ
0(x) = Υ
M −1(x) for x ∈ δ
0. Therefore the mapping defined as
Υ (x) = Υ
m(x) for x ∈ Sq
γ∗1,m,γ∗2,m,δ∗m,δm+1∗
and m ∈ {0, 1, . . . , M − 1}
is a homeomorphism transforming the “annulus” P
F1∗,F2∗onto the “annulus”
P
F1,F2.
Step 4: Construction on the neighbourhood of a. Extend Υ to the whole neighbourhood of a. Let y ∈ P
F1∗,F2∗, and define x ∈ V
aby
x = ψ
k(x)h(y), k ∈ N.
Set
(16) Υ e
a(x) :=
(ψ
−k(x)h◦ Υ ◦ φ
k(x)h)(x)
if ∃y ∈ P
F1∗,F2∗: x = ψ
hk(x)(y), k ∈ N,
a for x = a.
For points of the curve F
2∗the mapping e Υ
ais defined in two ways. First,
F
2∗⊂ P
F1∗,F2∗so we can take zero as the value of k. On the other hand we
can take k = 1 because every point of F
2∗is the image of a point of F
1∗.
However, both φ
hand ψ
hpreserve values of τ so for each x ∈ F
2∗both the
ways give the same image. This also implies that e Υ
ais continuous on F
2∗.
To prove the continuity of e Υ
aat a consider a sequence {x
n} converging to a. For each n there exists a natural number k
nsuch that ψ
h−kn(x
n) ∈ P
F1∗,F2∗. Furthermore, k
n→ ∞. Thus e Υ
a(x
n) → a. Hence, the function
Υ e
a: P
F1∗,F2∗→ P
F1,F2is a homeomorphism.
Since on the sets δ
m, m = 0, . . . , M − 1 (see (11)) the mapping Υ is defined by the local conjugating homeomorphism g
sk(see (12) and (15)) which transforms W
φuh(s
k) onto W
ψuh(s
k), the inclusion (10) is satisfied.
Introduce the following notations:
D
φh— the set of all repelling points,
S
φh— the set of all saddle points and the attracting points which are not contained in the closure of the unstable manifold of any saddle point, P
φh— the set of attracting points which are contained in the closure of the
unstable manifold of a saddle point.
Define
W
φsh(S
φh) := [
c∈Sφh
W
φsh(c), W
φsh(P
φh) := [
c∈Pφh
W
φsh(c),
and Θ : W
φsh(S
φh) → M by
(17) Θ(x) = (ψ
−hn0◦ g
c◦ φ
nh0)(x) for x ∈ W
φsh(c) \ {c},
x for x = c,
where c ∈ S
φh, g
cis a local homeomorphism conjugating the flows φ
hand ψ
hon a neighbourhood of c and ψ
−his the inverse of ψ
h. If x is in one of the sets K
c,1, then the natural number n
0= n
0(x) is zero. In the other case it is chosen in such a way that φ
nh0(x) ∈ K
c,1and φ
nh0−1(x) ∈ int
rel(K
c,0\ K
c,1).
Define α : M → M as follows:
(18) α(x) =
Θ(x) for x ∈ W
φsh(S
φh),
Υ e
a(x) for x ∈ W
φsh(P
φh), a ∈ P
φh, x for x ∈ D
φh.
Every nonrepelling point on a compact manifold is contained in a stable manifold of an attracting or saddle singularity. Thus (18) defines α on the whole M. In the next section it is shown that α is a homeomorphism globally conjugating the flows (M, φ
h) and (M, ψ
h).
Lemma 4.7. For all positive constants ε, there exists a positive constant h
0such that for all x ∈ M, t ∈ R and 0 < h < h
0,
%
M(φ(x, t), φ(x, t + h)) < ε.
The proof of this simple lemma is omitted (it can be found in [Bie]).
Lemma 4.8. Let a be an attracting fixed point of the flow φ, and p
1, p
2repelling fixed points, p
16= p
2. Assume that p
1, p
2∈ cl W
φs(a). Then there exist saddle points q
1, q
2, not necessarily different , such that p
1∈ cl W
φs(q
1) and p
2∈ cl W
φs(q
2).
P r o o f. Since p
1, p
2∈ ∂W
φs(a), there exist y
1and y
2such that
t→∞
lim φ(y
1, t) = a, lim
t→−∞
φ(y
1, t) = p
1,
t→∞
lim φ(y
2, t) = a, lim
t→−∞
φ(y
2, t) = p
2.
Let z
1∈ orb(y
1) ∩ V
aand z
2∈ orb(y
2) ∩ V
a, where orb(y) denotes the orbit of y, and let V
a⊂ W
φ,asbe a neighbourhood of a. The manifold M is locally arcwise connected, so let
l : [0, 1] 3 τ 7→ l(τ ) ∈ X
be a closed arc included in V
a, avoiding a and such that l(0) = z
1, l(1) = z
2. Since z
1and z
2lie in the unstable manifolds of p
1and p
2respectively, the boundaries of the unstable manifolds intersect l in points w
1and w
2, not necessarily different. As the unstable manifolds of repelling points are open and M is compact, the points w
1, w
2are contained in the unstable manifolds of saddle points q
1, q
2, not necessarily different. From the λ-lemma it follows that for every neighbourhood V
qiof q
ithere exists t > 0 such that
φ(l ∩ W
φu(p
1), −t) ∩ V
q16= 0.
We have
(W
φs(q
1) \ {q
1}) ∩ V
q16= 0.
Thus, for every ε > 0 there exists t > 0 such that φ(l∩W
φu(p
1), −t) intersects the ε-envelope of W
φs(q
1). This implies that there exists u ∈ W
φs(q
1) such that
t→−∞
lim φ(u, t) = p
1.
Therefore, the point p
1∈ cl W
φu(q
1). The same can be said about the points q
2and p
2.
Corollary 4.9. Let p be a repelling point which is not contained in the closure of the stable manifold of any saddle point. Then there exists only one attracting point a such that p ∈ cl W
φs(a).
This follows easily from Lemma 4.8.
Corollary 4.10. Let M = S
n, n > 1. Then, for every h > 0, the cascade (S
n, φ
h) has no saddle-saddle connections and , for sufficiently small h, the cascade (S
n, ψ
h) has no such connections either.
P r o o f. By the assumptions of Theorem 2.1, the flow (S
n, φ) has no
saddle-saddle connections. Thus each saddle point q
i, i = 1, . . . , I, has a
neighbourhood U
qisuch that every point x
0∈ U
qilies in W
φs(q
i) or in W
φs(a
j) for an attracting point a
j. For every x ∈ S
nand h > 0,
t→∞
lim φ(x, t) = lim
n→∞
φ
nh(x), hence x
0∈ W
φsh
(q
i) or x
0∈ W
φsh
(a
j).
Define
P
i:= {x ∈ U
qi: ∀
n∈{1,2,...}ψ
hn(x) 6∈ U
qi}
and let U
qi⊂ int V
qi, where V
qiis a neighbourhood on which the cascades φ
hand ψ
hare locally conjugate. Choose real numbers R and h
isuch that B(q
i, R) ⊂ int U
qiand P
i∩ B(q
i, R/2) 6= ∅ for every h ∈ [0, h
i].
Step 1: We will show that P
i∩ W
φs(q
i) = ∅ for sufficiently small h
i. Define
t
i:= sup
x∈Uqi∩Wφs(qi)
{t : φ(x, t) ∈ B(q
i, R/4) \ B(q
i, R/8)}.
It is obvious that t
i∈ (0, ∞). Suppose h
iis such that there exists a natural number n
1with t
i= n
1h
iand
ξh
i< R/4,
where ξ is the constant from (3). Since there are only a finite number of singularities, we can take ξ as the maximum value of the constants of all saddle points.
Let x ∈ U
qi∩ W
φs(q
i). Then
φ
nhi1(x) ∈ B(q
i, R/4) and by (3) and the choice of h
i,
%
M(φ
nhi1(x), ψ
hni1(x)) < R/4, or equivalently
ψ
nhi1(x) ∈ B(q
i, R/2).
Since P
i∩ B(q
i, R/2) = ∅, the point x is not in P
i.
Step 2: Proof of lemma. Let x ∈ P
i. The first step implies that x 6∈ W
φs(q
i). Since there are no saddle-saddle connections, the point x is in W
φs(a
j) for an attracting point a
j. Let V
ajbe the neighbourhood of a
jon which φ
hand ψ
hare topologically conjugate. Decompose the set P
iinto disjoint components in the following way:
P
ij:= {x ∈ P
i: x ∈ W
φs(a
j)}.
Let, furthermore,
t
ij:= sup{t : φ(x, t) ∈ B(a
j, r/2) \ B(a
j, r/4), x ∈ P
ij},
where r is chosen in such a way that B(a
j, r) ⊂ V
aj. It is obvious that
t
ij∈ (0, ∞). Denote by e
nthe error after the nth step of the Euler method.
Choose h
ijso small that φ
hand ψ
hare topologically conjugate on V
ajand ξh
ij< r/2. Moreover, let t
ij= n
2h
ij, n
2∈ {1, 2, . . .}. Then φ(x, t
ij) ∈ B(a
j, r/2) for every x ∈ P
ij. By (3) and the choice of h
ij,
e
n:= %
M(φ
nh2ij(x), ψ
hnij2(x)) < r/2.
Hence
ψ
hn2ij
(x) ∈ B(a
j, r).
Assuming that there are I saddle points and J attracting points, set
(19) h
0:= min
i∈{1,...,I}, j∈{1,...,J }
{h
ij}.
As V
aj⊂ W
φsh0
(a
j) and the cascades are locally topologically conjugate, we have V
aj⊂ W
ψsh
(a
j) for every h ∈ [0, h
0]. Hence ψ
nh2(x) ∈ W
ψsh(a
j).
Therefore, every x ∈ U
qiis either in W
ψsh(q
i) or in W
ψsh(a
j), where a
jis an attracting point. This implies that ψ
hhas no saddle-saddle connections.
Remark. By Lemma 4.10 and Theorem 4.2, if (S
n, φ) is a gradient Morse–Smale system, then so are the cascade (S
n, φ
h) (for every positive h) and (S
n, ψ
h) (for sufficiently small h).
Corollary 4.11. Let U
qibe a neighbourhood of saddle point q
ion which the cascades φ
hand ψ
hare conjugate by a homeomorphism g
qi: U
qi→ g
qi(U
qi). Then there exists a constant h
0> 0 such that for every h ∈ (0, h
0) and x ∈ U
qi∩ W
φuh
(q
i) if x ∈ W
φsh(a
j), where a
jis an attracting fixed point , then g
i(x) ∈ W
ψsh(a
j).
P r o o f. Step 1. Define
(20) H
ij:= W
φuh(q
i) ∩ (B(q
i, r
i) \ int B(q
i, r
i/2)) ∩ W
φsh(a
j).
First, it will be shown that there exists h
0> 0 so small that for every h ∈ (0, h
0) and every x ∈ H
ijwe have g
i(x) ∈ W
ψsh
(a
j). The radius r
iis chosen such that B(q
i, r
i) ⊂ U
qi.
The definition (20) implies that H
ijis closed in M, hence compact.
The stable manifold W
φsh(a
j) is open and H
ij⊂ W
φsh
(a
j). Thus, for every x ∈ H
ijthere exists r
x> 0 such that B(x, r
x) ⊂ W
φsh
(a
j). The set K := {B(x, r
x) ∩ W
φuh(q
i)}
is a covering of H
ij. Thus, we can choose a finite subcovering K
∗. Let r
ijbe the smallest radius of the balls B(x, r
x) used to construct K
∗. Take h
0ij> 0 such that g
qi(x) ∈ B(x, r
ij) for every point x ∈ H
ij. Then g
qi(x) ∈ W
φsh(a
j).
It can be shown (in the same way as in the second step of the proof of
Corollary 4.10) that there exists h
1ijsuch that g(H
ij) ⊂ W
ψsh(a
j) for every
h ∈ (0, h
1ij). Set h
ij:= min{h
0ij, h
1ij}.
Step 2. Suppose, by contradiction, that there exists y ∈ W
φuh(q
i) ∩ int B(q
i, r
i/2) such that y
1∈ W
φsh
(a
j) and z
1:= g
i(y
1) ∈ W
φsh(a
k), where a
kis an attracting fixed point and k 6= j. Corollary 4.7 implies that, for sufficiently small h, there exists a natural m
0such that
φ
mh0(y
1) ∈ B(q
i, r
i) \ int B(q
i, r
i/2).
As y
1∈ W
φsh
(a
j), the point y
2:= φ
mh0(y
1) is in W
φsh(a
j). Therefore, by Step 1, z
2:= g
i(y
2) ∈ W
ψsh(a
j) as y
2∈ P
ij. However
z
2:= g(y
2) = g(φ
mh0(y
1)) = (g ◦ φ
h)(φ
mh0−1(y
1))
= (ψ
h◦ g ◦ φ
h)(φ
mh0−2(y
1))
= (ψ
h◦ ψ
h◦ g)(φ
mh0−2(y
1)) = . . . = ψ
hm0(g(y
1)) = ψ
hm0(z
1), and ψ
mh0(z
1) ∈ W
ψsh(a
k) as z
1∈ W
ψsh
(a
k), which leads to a contradiction.
5. Proof of Theorem. Firstly, we prove that the map α defined in (18) is a bijection. Then we prove that it is continuous on M. This implies that α is a homeomorphism. The map α conjugates the cascades φ
hand ψ
h, which follows directly from its definition (see (16)–(18)).
Injectivity. Let x
1, x
2∈ S
2. If at least one of them is a repelling singu- larity of the system φ, then α(x
1) 6= α(x
2) by the definition of α. Otherwise the following two cases have to be considered:
Case 1: The points lie in different orbits. Then y
1:= φ
nh0(x1)(x
1) 6= φ
nh0(x2)(x
2) =: y
2.
The system φ has only a finite number of singularities and every singularity c
iis a fixed point of g
i. The sphere S
2is compact so every point is contained in the stable manifold of a stable point: x
1∈ W
φs(c
i), x
2∈ W
φs(c
j).
If i = j, then z
1:= g
i(y
1) 6= g
i(y
2) = z
2as g
iis a bijection. Since g
iis a local conjugating homeomorphism, the images of different orbits do not intersect. Thus ψ
n−h0(x1)(z
1) 6= ψ
n−h0(x2)(z
2), which implies that α(x
1) 6= α(x
2).
If i 6= j, then we can choose the domains V
iof the homeomorphisms g
iin such a way that V
i∩ V
j= ∅ and g
i(V
i) ∩ g
j(V
j) = ∅ for i 6= j. Thus z
1:= g
i(y
1) 6= g
j(y
2) =: z
2. As x
1∈ W
φsh,ai
and x
2∈ W
φsh,aj
, the points z
1and z
2lie in the disjoint manifolds W
ψsh,aiand W
ψsh,ajrespectively. This implies that α(x
1) 6= α(x
2).
Case 2: The points lie in the same orbit. We can assume that E(x
2) >
E(x
1). Then x
1= φ
mh(x
2) for some positive natural number m = m(x
1, x
2).
Thus
α(x
2) = (ψ
−hn0(x1)+m◦ g
ci◦ φ
nh0(x1)+m)(x
2)
= (ψ
−hm◦ ψ
−hn0(x1)◦ g
ci◦ φ
nh0(x1))(x
1) = ψ
m−h(α(x
1)) 6= α(x
1).
Surjectivity. If y ∈ S
2is a repelling fixed point, then y = α(y) (see definition of α). Otherwise y ∈ W
ψsh,cifor some fixed point c
i(attracting or saddle). Let n
0= n
0(y) be such that
y
n0−1:= ψ
hn0−1(y) ∈ int
rel
(g(K
ci,0) \ g(K
ci,1)) and y
n0:= ψ
hn0(y) ∈ g(K
ci,0).
As g is a conjugating homeomorphism,
(φ
−1h◦ g
−1◦ ψ
h)(y
n0−1) = g
−1(y
n0−1).
Furthermore, since ψ
h(y
n0−1) = y
n0, we have
(φ
−1h◦ g
−1)(y
n0) = g
−1(y
n0−1).
Set
x
n0:= g
−1(y
n0) and x
n0−1:= g
−1(y
n0−1) = φ
−1h(x
n0).
Since g
ciis a homeomorphism, x
n0−1∈ (int
relK
ci,0\ K
ci,1) and x
n0∈ K
ci,1. This means that for x
∗= φ
n−h0(x
n0) the natural number n
0is the same as in the definition of α, which implies that y = (ψ
−nh 0◦ g ◦ φ
nh0)(x
∗) = α(x
∗).
Continuity
Case 1: Continuity on the stable manifold of an attracting point. Let x lie on the stable manifold of an attracting point a
i. There exists n
0= n
0(x) such that
φ
nh0(x) ∈ K
ai,1and φ
nh0−1(x) ∈ int(K
ai,0\ K
ai,1).
If φ
nh0(x) ∈ int K
ai,1, then as int K
ai,1and int(K
ai,0\ K
ai,1) are open and the map φ(·, t) is continuous, there exists a neighbourhood U
xof x such that
φ
nh0(U
x) ⊂ int K
ai,1and φ
nh0−1(U
x) ⊂ int(K
ai,o\ K
ai,1).
As g
i, ψ
−h(·, t) and φ
h(·, t) are continuous, the map α = ψ
n−h0◦ g
i◦ φ
nh0is continuous at x.
If y := φ
nh0(x) ∈ ∂K
ai,1, then every neighbourhood U
yof y intersects int(K
ai,0\ K
ai,1). Let (y
n)
∞n=1⊂ int(K
ai,0\ K
ai,1) converge to y. Then there exists N such that for every natural n greater than N, φ
h(y
n) ∈ int K
ai,1. Indeed, suppose otherwise. Then there exists a subsequence (y
nk) such that φ
h((y
nk)
∞k=1) 6⊂ int K
ai,1. However, φ
h(y) ∈ int K
ai,1because the potential E is constant on ∂K
ai,1and decreases along a trajectory. This means that for every k, %
M(φ
h(y), φ
h(y
nk)) > inf{%
M(φ
h(y), w) : w ∈ ∂K
ai,1)} > 0.
This is a contradiction because φ is continuous.
We have demonstrated that for every y ∈ ∂K
ai,1there exists a neigh- bourhood U
y?with φ
h(U
y?) ⊂ K
ai,1. The neighbourhood U
y?has two disjoint components: the first one, U
y?1, is included in int K
a1,1whereas the second, U
y?2, is not. The point x is transformed by the map α in such a way that U
y?1is transformed by g
i, and U
y?2by ψ
−h◦ g
i◦ φ
h. Afterwards U
y?1∪ U
y?2is transformed by the map ψ
−hn0. But g
iis a conjugating homeomorphism so
(ψ
−h◦ g
i◦ φ
h)(U
y?2) = (ψ
−h◦ ψ
h◦ g
i)(U
y?2) = g
i(U
y?2)
and therefore we can say that the whole neighbourhood U
y?is transformed by g
i. The continuity of φ
h, g
iand ψ
−himplies that α is continuous at x.
Case 2: Continuity at repelling fixed points. There are two possibilities in this case: either
(i) the repelling point, say p, is in the closure of the stable manifold of a saddle point q, or
(ii) p is in the closures of the stable manifolds of attracting points.
Suppose first that (i) holds.
Step 1: Restriction to the stable manifold. Let {x
k}
k∈N⊂ W
φ,qs, p ∈ cl W
φs(q) and lim
k→∞x
k= p.
Let V
qbe a neighbourhood of q such that the homeomorphism g
h,qconju- gating φ
hand ψ
his defined on V
q. For sufficiently small h and every natural N almost all elements of the sequence {x
k}
k∈Nhave the following property:
φ
nhk(x
k) ∈ int K
q,1, φ
nhk−1(x
k) ∈ int(K
q,0\ K
q,1), n
k> N, which follows from Lemma 4.7 (for the definition of the sets K
q,0and K
q,1, see Lemma 4.3). The same lemma also implies that there exists r > 0 such that y
k= φ(x
k, n
kh) 6∈ B(q, r)∩W
φs(q) = K
q,2for all k. In other words, since the step on the manifold M is small, all the y
klie near ∂K
q,1. Since q is a fixed point of g
h,q, g
h,q(V
q) is a neighbourhood of q and g
h,q(V
q) ∩ W
φsh(q) ⊂ W
ψsh,qas g
h,qlocally conjugates φ
hand ψ
h. Since p is also a repelling point of ψ
h, it lies in W
ψsh(q) (by Corollary 4.11). As g
h,qis a homeomorphism, g
q(y
k) 6∈ g
h,q(K
q,2) for every k. Let W
ψp,sh(q) be the maximal connected component of W
ψsh(q) \ {q} containing p. Then, for every ε > 0, there exists N such that
ψ
−hn(W
ψp,sh
(q) \ K
q,2) ⊂ B(p, ε) for all n > N.
Since almost all n
kare greater than N and every y
kis in W
ψp,sh(q) \ K
q,2, almost all ψ
−hnk(y
k) lie in B(p, ε). This means that
k→∞
lim ψ
−hnk(g
q(y
k)) = p.
Thus, we have shown that the map α restricted to W
φsh(q)∪{p} is continuous
at p.
Step 2: Continuity at the repelling point. Let ε > 0. Since α is continuous on the stable manifold of every saddle point (see case 3), for every x
∗in such a manifold and ε
1> 0, there exists δ
1> 0 such that if x ∈ B(x
∗, δ
1) then α(x) ∈ B(x
∗, ε
1). On the other hand, p is a fixed point of α, and α|W
φs(q) is continuous at p (see step 1). Thus, for every ε
2> 0 there exists δ
2> 0 such that if x
∗∈ B(p, δ
2), then α(x
∗) ∈ B(p, ε
2). Let ε
1and ε
2be such that ε
1+ ε
2= ε. Then the triangle inequality implies that if x ∈ B(p, δ
1+ δ
2) then α(x) ∈ B(p, ε
1+ ε
2).
Suppose now that (ii) holds. Corollary 4.9 implies that there exists only one attracting singularity a such that p ∈ W
φs(a), whereas Lemma 4.8 implies that only one repelling point can lie in cl W
φs(a). This means that there exists a neighbourhood V
pof p such that
t→∞
lim φ(x, t) = a
for all x ∈ V
p\ {p} and there exists a neighbourhood V
aof a such that
t→∞
lim ψ(y, −t) = p for all y ∈ V
a\ {a}.
We can now repeat the first step of subcase (i) to show that the map α is continuous at p.
Case 3: Continuity on the stable manifold of a saddle point. Let q
ibe a saddle point. If the restriction φ
h|W
φsh
(q
i) is considered, then q
iis an attracting singularity and, repeating the argument from case 1, we have continuity of α in the relative topology on the stable manifold.
Thus, let x
0∈ W
φsh
(q
i) and x ∈ U
x0\ W
φsh
(q
i), where U
x0is a neighbour- hood of x
0. Since there are no saddle-saddle connections, there exist attract- ing points a
j, a
k, not necessarily different, such that W
φuh(q
i)
1⊂ W
φsh
(a
j) and W
φuh(q
i)
2⊂ W
φsh