Symbolic dynamics for the

Kuramoto-Sivashinsky PDE on the line II

Daniel Wilczak and Piotr Zgliczyński1
Jagiellonian University, Institute of Computer Science,
arXiv:2405.17087v1 [math.DS] 27 May 2024

Lojasiewicza 6, 30–348 Kraków, Poland

e-mail: [email protected]

May 28, 2024

Abstract
We present a new algorithm for the rigorous integration of the vari-
ational equation (i.e. producing C 1 estimates) for a class of dissipative
PDEs on the torus. As an application for some parameter value for the
Kuramoto-Sivashinsky PDE on the line with odd and periodic bound-
ary conditions we prove the existence of infinite number of homo- and
heteroclinic orbits to two periodic orbits. The proof is computer assisted.

1 Introduction
This paper is a sequel to [WZ]. The goal of the present work is twofold: to
strengthen the results from [WZ] about the chaotic dynamics for the Kuramoto-
Sivashinsky PDE on the line and to present an algorithm for the rigorous inte-
gration of the variational equation (i.e. producing C 1 estimates) for a class of
dissipative PDEs on the torus.
We consider the one-dimensional Kuramoto-Sivashinsky PDE [KT, S] (in
the sequel we will refer to it as the KS equation), which is given by
ut = −νuxxxx − uxx + (u2 )x , ν > 0, (1)
where x ∈ R, u(t, x) ∈ R and we impose odd and periodic boundary conditions
u(t, x) = −u(t, −x), u(t, x) = u(t, x + 2π). (2)
The Kuramoto-Sivashinsky equation has been introduced by Kuramoto [KT]
in space dimension one for the study of front propagation in the Belousov-
Zhabotinsky reactions. An extension of this equation to space dimension 2 (or
more) has been given by G. Sivashinsky [S] in studying the propagation of flame
front in the case of mild combustion.
The following theorem is the main result of [WZ].
Theorem 1 [WZ, Thm. 1] The system (1)–(2) with the parameter value ν =
0.1212 is chaotic in the following sense. There exists a compact invariant set
A ⊂ L2 ((−π, π)) (A is compact in H k ((−π, π)) for any k ∈ N) which consists
of
1 Work of D.W. and P.Z. was supported by National Science Center (NCN) of Poland under

project No. UMO-2016/22/A/ST1/00077

1
 Figure 1: Two approximate time-periodic orbits u1 and u2 .

1. bounded full trajectories visiting explicitly given and disjoint vicinities of

two selected periodic solutions u1 and u2 , respectively, with any prescribed
order {u1 , u2 }Z .
2. countable infinity of periodic orbits with arbitrary large periods. In fact,
each periodic sequence of symbols {u1 , u2 }Z is realised by a periodic solu-
tions of the system (1)–(2).

The two special solutions u1 and u2 appearing in Theorem 1 are time-periodic

— see Fig.1. In the present work we extend this result include the existence of
orbits asymptotic to u1 and u2 both in backward and forward time.

Theorem 2 The invariant set A from Theorem 1 contains countable infinity

of geometrically different
1. homoclinic orbits to u1 ,
2. homoclinic orbits to u2 ,

3. heteroclinic orbits between u1 and u2 ,

4. heteroclinic orbits between u1 and u2 .
The convergence in the definition of homoclinic and heteroclinic orbits happens
in L2 ((−π, π)) and in H k ((−π, π)) for any k ∈ N.

Connecting orbits (homoclinic and heteroclinic) between periodic orbits ui and

uj are understood as non-periodic solutions defined for all t ∈ R, such that their
α-limit and ω-limit sets in L2 topology (turns out that also H k for any k) are
equal ui and uj , respectively. In fact, each non-constant sequence of symbols
(v1 , . . . , vn ) ∈ {u1 , u2 }n encodes a connecting orbit in the following way. There
is a trajectory of the system (1)–(2) with α-limit set equal to v1 and ω-limit
set equal to vn . Moreover, the solution visits disjoint neighbourhoods of u1

2
and u2 with the order prescribed by the sequence (v1 , . . . , vn ). Hence, we have
countable infinity of geometrically different connecting orbits between ui and
uj , i, j = 1, 2.
Our approach uses no special features of the Kuramoto-Sivashinsky PDE.
We just need the Fourier basis to diagonalize the linear part of the PDE under
consideration. Therefore, it should be applicable to other systems of dissipative
PDEs with periodic boundary conditions. While there is a wide literature on
computer-assisted verification of symbolic dynamics for ODEs (see for example
[AZ, GZ, MM1, MM2, Mos, RNS, T1, T2, W, WSB, WZ03, Z4]), to the best of
our knowledge, this is the first computer-assisted result of this type for PDEs.
The proofs of Theorems 1 and 2 are a mixture of geometrical methods and
rigorous numerics. The proof of Theorem 2 combines topological arguments used
to construct symbolic dynamics (as in Theorem 1) with suitable cone conditions
to obtain desired homo- and heteroclinic orbits to some periodic orbits.
In the case of the KS equation, we apply the method to certain Poincaré map.
The orbits u1 and u2 from Theorem 1 correspond to a fixed point and a period
two point. First, we construct symbolic dynamics along apparent heteroclinic
connections between these points. Then using the cone conditions we show that
these heteoroclinic indeed exist. This is the same type of construction as it
is used in the proof of the Smale-Birkhoff homoclinic theorem [GH, Thm.5.3.5]
about the existence of horseshoe dynamics for finite-dimensional diffeomorphism
with a hyperbolic fixed point whose stable and unstable manifolds intersect
transversally.
The nature of our geometric method is such that its assumptions can be
expressed as a finite set of explicit inequalities. Therefore, we can use a com-
puter to verify rigorously that these inequalities are satisfied for a given map f ,
provided we have an algorithm that computes rigorous bounds on f on compact
sets. In the case of the KS equation we have to compute rigorously bounds on
some Poincaré maps and its partial derivatives. For this purpose we propose
an algorithm, which allows to compute rigorous bounds on the trajectories of
PDEs with periodic boundary conditions (this was described in [WZ]) together
with bounds on the derivatives with the respect to initial conditions.
Let us recall from [WZ] the motivation for the choice of the KS equation for
this study. The following is known about the dynamics of KS on the line:

• The existence of a compact global attractor, the existence of a finite-

dimensional inertial manifolds for (1–2) are well established — see for
example [CEES, FT, FNST, NST] and the literature cited there. We
would like to emphasise, that we are not using these results in our work.
• There exist multiple numerical studies of the dynamics of the KS equation
(see for example [CCP, HN, JKT, JJK, SP]), where it was shown, that
the dynamics of the KS equation can be highly nontrivial for some values
of parameter ν, while being well represented by relatively small number
of modes.
• There are several papers devoted to computer-assisted proofs of periodic

3
 1.22
u2
1.20

1.18

u1
1.16

ν=0.1212
1.14

-1.40 -1.39 -1.38 -1.37 -1.36 -1.35

u2

Figure 2: Numerically observed chaotic attractor for (1–2) obtained by simula-

tion of a finite-dimensional projection of the corresponding
P∞ infinite-dimensional
ODE for the Fourier coefficients in u(t, x) = u
k=1 k (t) sin(kx). Projection
onto (u2 , u3 ) plane of the intersection of the observed attractor with the Poincaré
section u1 = 0, u′1 > 0 is shown along with an approximate location of the two
periodic points u1 , u2 appearing in Theorems 1 and 2. The point u1 is a fixed
point for the Poincaré map and u2 is of period two.

orbits for the KS equation by Zgliczyński [Z04, Z10], Arioli and Koch
[AK10] and by Figueras, Gameiro, de la Llave and Lessard [FGLL, FL17,
GL17].
While the choice of the odd periodic boundary conditions was motivated by
earlier numerical studies of KS equation [CCP, JKT], the basic mathematical
reason is the following: the equation (1) with periodic boundary conditions has
the translational symmetry. This implies, that for a fixed value of ν, all periodic
orbits are members of one-parameter families of periodic orbits. The restriction
to the invariant subspace of odd functions breaks this symmetry, and gives a
hope, that dynamically interesting objects are isolated and easier accessible for
computer-assisted proofs.
Our choice of the parameter value ν = 0.1212 is motivated by a numeri-
cal observation, that the Feigenbaum route to chaos through successive period
doubling bifurcations [F78] happens for (56) as ν decreases toward ν = 0.1212
(first observed for other values of ν in [SP]). For this parameter value a chaotic
attractor is observed — see Fig. 2.
Another result illustrating the effectiveness of our integration algorithm is
about the attracting periodic orbit for ν = 0.127.
Theorem 3 Consider system (1)–(2) with the parameter value ν = 0.127 and
φ(t, p) be the induced local semiflow.
The exists a point x0 , a compact forward invariant set A ⊂ L2 ((−π, π)),
x0 ∈ A such that

4
 • x0 is time periodic
• there exists λ > 0 and C > 0, such that for each p1 , p2 ∈ A and t > 0
holds
∥φ(t, p1 ) − φ(t, p2 )∥∗ ≤ Ce−λt ∥p1 − p2 ∥∗ , (3)
where ∥∥∗ is some norm constructed using Fourier series (see section 6.3
for details).
The existence of this periodic orbit was first proved with computer assistance
in [Z04] using the first version of the algorithm of integration of dissipative PDEs
based on self-consistent bounds. In a sense it is one of the easiest periodic
orbits to prove, however its attracting character was not proved, yet. In papers
[Z10, AK10, FL17, GL17] the existence of multiple periodic orbits for this system
has been proved for more difficult parameter values, in [AK10, GL17] the authors
were able to show that some of these orbits are unstable. It turns out that
checking the instability of an orbit is an easier task than proving that the orbit
is attracting. In proofs of instability in [GL17] it was enough to prove an
existence of one unstable direction, while the stability proof requires control
in all direction in the phase space. This is what we do in the present paper
using our C 1 -algorithm. With such an algorithm establishing the instability
or stability of apparently hyperbolic orbits is the task of the same difficulty.
Observe that in the proof of the instability of some periodic orbit in [AK10] the
whole spectrum has been estimated and a first C 1 algorithm for KS-equation
has been proposed.
The content of the paper can be described as follows. In Section 2 we present
some results about convergence solutions of Galerkin projections of dissipative
PDEs together with their derivatives with respect to initial conditions. In Sec-
tion 3 we present technical lemmas about non-autonomous linear equations. In
Section 4 we present our C 1 -algorithm for integration of dissipative PDEs. Sec-
tion 5 contains a description of some geometric tools from dynamics, which are
used in the proof of Theorem 2. In Sections 6 computer-assisted proofs of The-
orem 2 and Theorem 3. The main body of paper is followed by the appendix. It
sections of different nature. The block of Appendices A, B, C, D has a theoret-
ical character is an abstract exposition of our approach together with proofs of
convergence theorems stated in Section 2. Remaining appendices contain either
some definitions used in other part or the explicit formulas for various objects
appearing in our algorithm for KS equation.

1.1 Notation
By Z+ we denote the set of all positive integers. Given a matrix A ∈ Rn×n and
a norm ∥ · ∥ by µ(A) we denote the logarithmic norm of A induced by norm ∥ · ∥
- this notion is recalled in appendix A.
Given two normed vector spaces V, W by Lin(V, W ) we will denote the space
of all bounded linear maps from V to W .

5
2 C 1 -convergence of Galerkin projections
The goal of this section is to formulate results regarding the convergence of
solutions to Galerkin projections and associated variational equations to the
solution of underlying dissipative PDE. This, along with bounds obtained from
the algorithm, will give us bounds on partial derivatives of solutions to PDE
with respect to the initial conditions.
We will be dealing with H ⊂ RZ+ , i.e. with sequences (a1 , a2 , . . . ), with
some norm ∥ · ∥. Let ei denote the sequence such that (ei )j = 0 for j ̸= i and
(ei )i = 1. By e∗j ∈ H ∗ we denote a dual form, e∗j (ei ) = δij . If J ⊂ Z+ , then by
PJ we denote projection πi (PJ (w)) = wi ei for i ∈ J and πi (PJ (w)) = 0 if i ∈ / J.
We assume that weShave increasing family of finite subsets of Z+ J1 ⊊
J2 ⊊ Jn . . . , such that n∈Z+ Jn = Z+ . For n ∈ Z+ by Hn we will denote a
subspace spanned by {ej }j∈Jn by Pn we denote the projection on Hn and let
Qn = Id − Pn . P
Let πk be a projection onto k-th direction, i.e. πk ( wj ej ) = wk ek .
We will assume that H ⊂ RZ+ with a norm ∥ · ∥ be a Banach space, which
is gss space (see Appendix B).
We investigate the Galerkin projections of the following problem

u′ = F (u) = L(u) + N (u), (4)

where L is a linear operator and N is a nonlinear part of F. They are given by

u′ = Pn (F (u)), u(0) = Pn u0 . (5)

Denote by φn a local dynamical system on Hn induced by the n-th Galerkin

projection.
We need to introduce the notation for Galerkin projections of our vector field.
Let ιn : Hn → H be the imbedding of Hn into H. Then we define F n : Hn → Hn
by F n (z) = Pn (F (ιn (z)), i.e. this is simply a Galerkin projection.
We assume that Fi is a formal power series in variables ak and symbols
∂ k Fi
∂ai ...∂ai are defined by formal differentiation of Fi .
1 k

Definition 1 We say that F is admissible, if Pn F is defined on Hn and is C 3

as a function defined on Hn .

Definition 2 Assume that (H, ∥·∥) is gss space. Let W ⊂ H and F : dom(F ) →
H. Assume that F is admissible. We say that F satisfies conditions C1, C2,
C3, C4 on W if
C1 There exists M ≥ 1, such that Pn (W ) ⊂ W for n ≥ M
C2 W is compact.
C3 W ⊂ dom (F ), ( for each u ∈ W F (u) ∈ H) and function F : W → H is
continuous.

6
C4 for every i and z ∈ W there exists limn→∞ D(Fi|Hn )(Pn z) and we define

DF
g i (z) = lim D(Fi|H )(Pn z).
n
(6)
n→∞

g i : W → H ∗ is continuous.
We assume that DF

Observe that conditions C1, C2, C3, C4 (called later convergence conditions)
have nothing to do with splitting F = L + N and ei being eigenvectors of L.
Conditions C1, C2, C3 are needed for convergence of Galerkin projections
(C 0 -convergence), while C4 is required (plus some other conditions) for con-
vergence of solutions of variational equations for Galerkin projections (C 1 -
convergence).
Let us comment on condition C4. Instead of introducing DF g i one can be
tempted to replace it by
for every i W ⊂ dom (DFi ) and function DFi : W → H ∗ is continuous.
This approach is problematic, because it may happen that Fi is not defined
everywhere and the notation of Frechet derivative cannot be applied.
However, using DF g i we will have the following formula (if W is convex)
Z 1
Fi (z) − Fi (w) = g i (tz + (1 − t)w)dt · (z − w), z, w ∈ W.
DF (7)
0

This result is obtained in Lemma 21 in Appendix B (for f = Fi ).

Definition 3 We say that F satisfies condition D, on W ⊂ H if W is convex

and
there exists l ∈ R such that for all n = 1, 2, . . .


sup µ DPn F|Hn (x) ≤ l. (8)
x∈Pn W

The main idea behind condition D is to ensure that the logarithmic norms (see
Appendix A) for all Galerkin projections are uniformly bounded.
The following theorem is a generalization of [Z] (Thm. 13). There it is
assumed that W is a trapping region and H = l2 was used, but the main idea
proof is the same. The proof is given in appendix C.
Theorem 4 Assume (H, ∥ · ∥) is gss . Assume that W ⊂ H and F satisfy
conditions C1, C2, C3, D and W is convex. Let Z ⊂ W and T > 0 be such
that all n > M1 there holds

φn (t, x) ∈ W for x ∈ Pn (Z), t ∈ [0, T ]

Then
1. Uniform convergence and existence For a fixed u0 ∈ Z, let un : [0, T ] →
Pn (W ) be a solution of (5). Then un converges uniformly to a function u∗ :
[0, T ] → W , which is a solution of (4) and u∗ (0) = u0 . The convergence
of un is uniform with respect to u0 ∈ Z.

7
2. Uniqueness within W . There exists only one solution of the initial value
problem (4), u(0) = u0 for any u0 ∈ Z, such that u(t) ∈ W for t ∈ [0, T ].
3. Lipschitz constant. Let u : [0, T ] → W and v : [0, T ] → W be solutions of
(4) with u(0), v(0) ∈ Z. Then

|v(t) − u(t)| ≤ elt |v(0) − u(0)| for t ∈ [0, T ]. (9)

4. Semidynamical system. The partial map φ : [0, T ] × W → W , where

φ(·, u0 ) is a unique solution of equation (4), such that φ(0, u0 ) = u0 defines
a semidynamical system on W , namely
• φ is continuous
• φ(0, u) = u
• φ(t, φ(s, u)) = φ(t + s, u)
provided φ(t + s, u) exists.
Observe that the essential difficulty in application of the above theorem is to
find set W . A systematic way to construct it is based on the isolation property,
which is related to splitting F = L + N and the diagonal form of L.

2.1 Isolation property

To illustrate the isolation property let us consider a class of dissipative PDEs of
the following form
ut = Lu + N (u, Du, . . . , Dr u) , (10)
d
where u ∈ Rn , x ∈ Td = (R mod 2π) , L is a linear operator, N is a polynomial
and by Ds u we denote sth order derivative of u, i.e. the collection of all spatial
partial derivatives of u of order s. The reason to consider polynomial and
not more general functions N is technical — we need to compute the Fourier
coefficients of N (u, Du, . . . , Dr u). This can be achieved by taking suitable
convolutions of Fourier expansions of u and its spatial partial derivatives.
We require, that the operator L is diagonal in the Fourier basis {eikx }k∈Zd ,

Leikx = −λk eikx ,

with

L∗ |k|p ≤ λk ≤ L∗ |k|p , for all |k| > K and K, L∗ , L∗ ≥ 0,

p > r.

If the solutions are sufficiently smooth, the problem (10) can be written as
an infinite ladder of ordinary differential equations for the Fourier coefficients
in u(t, x) = k∈Zd uk (t)eikx , as follows
P

duk

= Fk (u) = −λk uk + Nk {uj }j∈Zd , for all k ∈ Zd . (11)
dt

8
 The crucial fact, which makes our approach to the rigorous integration of
(11) possible is the isolation property, which reads:
Let  
C
W = {uk }k∈Zd | |uk | ≤
|k|s q |k|
where q ≥ 1, C > 0, s > 0.
Then there exists K > 0, such that for |k| > K there holds
C
if u ∈ W, |uk | = , then uk · Fk (u) < 0.
|k|s q |k|
For the proof of the above fact see Theorem 3.1 and its proof in [Z10], where
the case with q = 1 was treated.
Projection of the set W onto k th mode is a closed disc (or interval) centred
at zero and of radius rk = |k|sCq|k| . Geometrically the isolation property means
that if |uk | = rk for some k > K then the k-th component of the vector field is
pointing inwards the set. As a consequence, the only way a trajectory may leave
the set W forward in time is by increasing some of leading modes |uk |, k ≤ K
above the threshold rk . However, one has to be careful with what ”inwards”
means in the above statement because W has empty interior as a compact subset
in infinite dimensional space.
This property is used in our approach to obtain a priori bounds for uk (h) for
small h > 0 and |k| > K, while the leading modes uk for |k| ≤ K are computed
using tools for rigorous integration of ODEs [CAPD, Lo, NJP] based on the
interval arithmetics [Mo], i.e. to obtain set W as in Theorem 4. Moreover,
from the point of view of topological method, the isolation property is of crucial
importance as it shows that on the tail we have the entry behaviour, which
enable us to apply the finite dimensional tools from the dynamics.
In this work we use sets of the type described above with s = 0 and q > 1.
The isolation property for KS equation on such sets is established in [WZ, Rem.
19].

2.2 Variational system

Assume that all assumptions of Theorem 4 are satisfied. From first assertion of
Theorem 4 it follows that
φn (t, Pn u) → φ(t, u), (12)
and the convergence is uniform on [0, T ] × Z.
Now let us consider the variational matrix for φn given by
∂φni
Vijn (t, u) = (t, u). (13)
∂uj
To see why we expect the convergence here for systems with diagonal L let
us remark, that V n satisfies the following differential equation
dVijn X ∂Ni
= λi Vijn + Vn (14)
dt ∂uk kj
k

9
 with initial condition V (0) = Id. The above equation has the same strong
damping as the original equation (4). Observe that the bound for Lipschitz
constant in equation (9) for equation (4) and its Galerkin projections is also a
uniform bound for the norms of matrices V n (t) for t ∈ [0, T ]. Once we have a
strong damping and a-priori bounds for V n we can use logarithmic norms to
control the convergence of V n ’s to variational matrix for (4). In [Z1] this has
been done under assumption about W being a trapping region. Moreover, one
additional condition was needed - (see condition VL formulated in Def. 4).
Since each column in Vijn evolves separately, we consider the system of vari-
ational equations
d n
x = Pn F (xn ), xn ∈ Hn , (15)
dt
d n
C = DPn F|Hn (xn )C n , C n ∈ Hn .. (16)
dt
The above system has Hn × Hn ⊂ H × H as the phase-space.
To bring the variational system to the setting discussed in this section we
proceeded as follows. On H × H we use norm ∥(x, v)∥ = max(∥x∥, ∥v∥). This
makes H × H a gss space. The Galerkin projection Pn on H × H is defined by
Pn (x, c) = (Pn x, Pn c).
In H × H we can formally write the variational system
d
x = F (x), (17)
dt
d
C = DF (x)C, (18)
dt
or
d
(x, C) = FV (x, C). (19)
dt
Observe that πi FV (x, C) for i ∈ Z+ is a formal power series in variables xk and
Ck .

Definition 4 Let F on W satisfies conditions C1,C2,C3,C4,D. We say that

(15,16) satisfies condition VL on W × Wc ⊂ H × H, if the following conditions
hold

VL -C1 there exists Mc ≥ 1, such that Pn Wc ⊂ Wc for all n ≥ Mc

VL -C2 Wc is compact.
VL -C3 W × Wc ⊂ dom (FV ),
(i.e. for each (u, C) ∈ W × Wc sequence {DF g i (u)C}i∈Z belongs to H)
+
and function FV : W × Wc → H × H given by
FV (u, C) = (F (u), {DF
g i (u)C}i∈Z ) is continuous,
+

10
VL -D there exists A ∈ R, such that for any n holds

sup µ(DPn FV |Hn ×Hn (x, v)) ≤ A. (20)

(x,v)∈Pn (W ×Wc )

The following theorem is a generalization of [Z1, Thm. 2]

Theorem 5 Same assumptions as in Theorem 4.
Assume that condition C4 is satisfied on W . Assume that Z is convex.
Assume that we have a family WCj ⊂ H of compact and convex sets, such
that condition VL is satisfied on W × WCj and for any n > j solution of the
variational problem (15–16) with initial conditions xn (0) ∈ Z and C n (0) = ej
satisfies
C n (t) ∈ WCj , t ∈ [0, T ] (21)
Then there exists a family of continuous functions Vij : [0, T ] × Z → R for
i, j = 1, . . . , ∞, such that
n
convergence: For each j the function V∗j (t, Pn z) converges to V∗j (t, z) uni-
formly on [0, T ] × Z, and there exists V (t, x) ∈ Lin(H, H), such that

∥V (t, x)∥ ≤ elt , (t, x) ∈ [0, T ] × Z,

X
V (t, x)ej = Vij (t, x)ei ,
i

and for every a ∈ H the map [0, T ] × Z ∋ (t, x) 7→ V (t, x)a is continuous.
smoothness: For any x, y ∈ Z and any t ∈ [0, T ] we have
Z 1
φ(t, x) − φ(t, y) = V (t, y + s(x − y))ds · (x − y), (22)
0

∂φ
and for every j partial derivative of the flow ∂uj (t, u) exists and

∂φ
(t, u) = V∗j (t, u). (23)
∂uj

equation for V : V (t, u) satisfies the following variational equation

dV∗j X X ∂Fi
(t, u) = ei (φ(t, u))Vkj (t, u), (24)
dt i
∂uk
k

with the initial condition V (0) = Id in the following sense: for each j the
dV
derivative dtj (t, u) exists, the series on r.h.s. of (24) converges uniformly
on [0, T ] × Z and equation (24) is satisfied.
The proof is given in appendix D.

11
3 Estimates for non-autonomous linear
equations
Theorems 4 and 5 are used in our algorithm for rigorous integration of dissi-
pative PDEs and in particular to obtain our results for Kuramoto-Sivashinsky
PDE described in Introduction to assure that the objects computed (or esti-
mated) exist. In that context the bound for the Lipschitz constant of the flow
obtained in Theorem 4 while essential in the proofs of both theorems is of no
use in rigorous numerics. To make algorithm work we need more subtle tools,
which give us precise results for particular coordinates or blocks of coordinates.
The goal of this section is to present tools, which will allow us estimate norms
of blocks of the variational matrix V (t, x).
The content of this subsection is based on the results from [KZ]. Consider a
linear equation
v ′ (t) = A(t) · v(t) + b(t), (25)
where v(t) ∈ Rk , A(t) ∈ Rk×k , b(t) ∈ Rk , A and b are bounded and measurable.
k
Assume
Ln that we have a decomposition of the phase space R of the form
Rk = i=1 Rki , ki ≥ 1 and accordingly for z ∈ Rk we will write z = (z1 , . . . , zn ),
where zi ∈ Rki . Observe that, when we have such decomposition, then equation
(25) can be written as
n
X
zi′ (t) = Aij (t)zj (t) + bi (t), i = 1, . . . , n,
j=1

where zi (t), bi (t) ∈ Rki , i = 1, . . . , n and Aij (t) ∈ L(Rki , Rkj ), i, j = 1, . . . , n

are linear maps. In this way matrix A is decomposed into blocks Aij . To each
block Aij we will assign a real number Jij and collect them in an n × n matrix
J. The quantity Jij will estimate the influence of zj on zi′ . The details will be
given in the sequel.

Definition 5 For f : U → Rn ), where U ⊂ R is open consider an ode

x′ = f (x), x ∈ Rn . (26)

Let D ⊂ R be an interval (a connected subset of R) containing t0 .

An absolutely continuous function x : D → Rn is a weak solution if for all
t ∈ D holds Z t
x(t) = x0 + f (x(s), y(s))ds. (27)
t0

The following lemma plays a crucial role in our considerations.

Ln
Lemma 6 [KZ, Lemma 4.1] Assume that z : [0, T ] → Rk = i=1 R
ki
is an
absolutely continuous map, which is a weak solution of the equation

z ′ (t) = A(t) · z(t) + b(t),

12
where b : [0, T ] → Rk and A : [0, T ] → Rk×k are bounded and measurable.
Assume that a measurable matrix function J : [0, T ] → Rn×n satisfies the
following inequalities for all t ∈ [0, T ]
(
∥Aij (t)∥ for i ̸= j,
Jij (t) ≥
µ(Aii (t)) for i = j.

Put Ci (t) = ∥bi (t)∥. Then, for i = 1, . . . , n and t ∈ [0, T ] there holds

∥zi ∥(t) ≤ yi (t),

where y : [0, T ] → Rn is a weak solution of the problem

y ′ (t) = J(t)y(t) + C(t), y(0) = (∥z1 (0)∥, ∥z2 (0)∥, . . . , ∥zn (0)∥) .

The next lemma follows immediately from Lemma 6 with b(t) ≡ 0.

Lemma 7 Assume that W : [0, T ] → Rk×k is an absolutely continuous map,
which is a weak solution of the equation

W ′ (t) = A(t) · W (t),

k×k
where A : [0, T ] → RL is bounded and measurable. Assume that we have a
n
k
decomposition R = i=1 Rki .
Let J : [0, T ] → Rn×n be a measurable matrix function satisfying the follow-
ing inequalities for all t ∈ [0, T ]
(
∥Aij (t)∥ for i ̸= j,
Jij (t) ≥
µ(Aii (t)) for i = j.

Let B : [0, T ] → Rn×n is a weak solution of the problem

B ′ (t) = J(t)B(t)

with initial conditions such that

Bij (0) ≥ ∥Wij (0)∥, i, j = 1, . . . , n

Then
∥Wij (t)∥ ≤ Bij (t), t ∈ [0, T ], i, j = 1, . . . , n.

3.1 Block decomposition

We will like to adapt Lemma 7 to the variational equation for dissipative PDEs
and its solution obtained as the limit of solutions of variational equation as in
Theorem 5. We will consider a very particular decomposition
M
H= ⟨ek ⟩ ⊕ Y, Y = (I − Pm )H
k≤m

13
where we have m 1-dimensional subspaces ⟨ei ⟩ and infinite-dimensional subspace
Y.
To distinguish between Vij , when i, j are treated as indices of particular
coordinates and the situation when they are supposed to denote blocks (as in
Lemmas 6 and 7) for the block entries involving Y we will write Vyy , Viy , Vyi .
In this way Vij for any i, j will denote always a number in the matrix V .
The block notation introduced above make sense also for F n . Below we write
∂F n
∂y meaning that the partial derivatives are meant with respect y ∈ Pn y, i.e.
this is finite dimensional matrix.

Theorem 8 Same assumptions and notation as in Theorem 5.

For some m > 0 let us consider the following decomposition of H
M
H= ⟨ek ⟩ ⊕ Y, Y = (I − Pm )H.
k≤m

Assume that matrix J ∈ R(m+1)×(m+1) satisfies

∂F n


 supn>m supw∈W ∂uij (w) , for i ̸= j, i, j ≤ m,
 n
 ∂F
supn>m supw∈W i
(w), for i = j, i ≤ m,


 ∂u
j n 
∂Fy

Jij ≥ supn>m supw∈W µ ∂y (w) , for i=j =m+1
 n
supn>m supw∈W ∂F



 ∂y (w) ,
i
for i ≤ m, j = m + 1
∂Fyn


 supn>m supw∈W ∂uj (w) , for i = m + 1, j ≤ m.

Then for any u ∈ Z and t ∈ [0, h] holds

|Vij (t, u)| ≤ eJt ij , i, j = 1, . . . , m,

(28)
∥Vyj (t, u)∥ ≤ eJt m+1,j , j = 1, . . . , m,


(29)
∥Viy (t, u)∥ ≤ eJt i,m+1 , i = 1, . . . , m,


(30)
∥Vyy (t, u)∥ ≤ eJt m+1,m+1 .


(31)

Proof: Let us take any n > m. We consider the decomposition

M
Pn H = ⟨ek ⟩ ⊕ Yn , Yn = Pn Y.
k≤m

From Lemma 7 applied to n-th Galerkin projection with matrix J we obtain

that conditions (28–31) are satisfied by V n (t, Pn u) for u ∈ Z and t ∈ [0, h]
Now we want to pass to the limit n → ∞. We know from Theorem 5 that

Vijn (t) → Vij (t), ∀(i, j) ∈ Z2+ ∀t ∈ [0, h]. (32)

The result now follows from Theorem 23.

14
4 The C 1 algorithm for integration of dissipative
PDEs.
There are several algorithms in the literature [C, Z10, Z04, WZ] that compute
bounds of the solutions on dissipative PDEs of the form (4). By a C 1 algo-
rithm we understand computation of bounds on the solutions to the following
variational equation associated to (4)

d ∂F
V (t, u0 ) = (φ(t, u0 ))V (t, u0 ), V (0, u0 ) = V0 . (33)
dt ∂u
The initial condition V0 for variational equation is chosen in our work as the
identity operator, as this needed to establish the cone condition.
Formally speaking ∂F ∂u in (33) may not exist on H, but its domain is dense in
H. Moreover, if we restrict initial conditions to sets possessing good convergence
properties (so that the assumptions of Theorem 5 are satisfied), then equation
(24) makes sense and has a unique solution.
For finite dimensional ODEs very efficient C 1 algorithms [Z02, WW] are
known along with their publicly available implementations [CAPD]. To the
best of our knowledge, there is only one C 1 algorithm [AK10] that computes
bounds of (33) in infinite dimensional case. It this section we propose a new
algorithm which simultaneously computes bounds for (4) and (33).

4.1 An outline of C 1 -integration algorithm

In the subsequent discussion we will quite often drop the dependence of V on
u0 , i.e. we will write just V (t) instead of V (t, u0 ) when u0 is known from the
context.
Denote by V∗j (t) the j-th column of V (t). It is well known that each column
V∗j (t) solves the following nonautonomous linear equation (we use notation
introduced in Section 2)

d ∂F ∂N
V∗j (t) = (φ(t, u0 ))V∗j (t) = LV∗j (t) + (φ(t, u0 ))V∗j (t), (34)
dt ∂u ∂u
but for different initial conditions – that is V∗j (0) = (V0 )∗j .
The structure of equation (34) is similar to that of (4) – the linear part L
is the same and we expect that it should dominate nonlinear part on higher
modes under the mild assumptions on the nonlinear part N . Thus, in principle
any C 0 algorithm can be applied to compute finite number of columns V∗j (t)
for j = 1, . . . , m. The remaining block of infinite number of columns will be
bounded uniformly by an operator norm.
For our further considerations it is convenient to split the phase space into
u = (x, y), where x ∈ Rm =: X, y is in the complementary subspace Y and use
the notation  
Vxx (t) Vxy (t)
V (t) = .
Vyx (t) Vyy (t)

15
The norms in X and Y is quite arbitrary. We just demand that H = X ⊕ Y is
gss and norms on X and Y are inherited from H.
In what follows we assume that the blocks Vxx (t) and Vyx (t) are computed
using our C 0 algorithm for integration of (4) (see [WZ]). Therefore for Vij (t)
for i, j = 1, . . . , m are given coordinate-wise interval bounds and for Vij with
j ≤ m and i > m we will have uniform bounds decaying with i → ∞ depending
on some constants computed by the algorithm, just like in C 0 computation of
u(t).
There remains to explain how we obtain bounds for blocks Vxy and Vyy .
Blocks Vxy and Vyy will be represented in our algorithm by estimates of their
operator norms. Let us comment, that the underlying norms in X and Y need
not to be the norms inherited from H.
In the next section we will recall some results regarding bounds for nonau-
tonomous linear ODEs. Then we will show how these estimates can be used
to obtain bounds on Vxy (t) and Vyy (t). Finally, we will present a new C 1 algo-
rithm that produces rigorous bounds on solutions of (4) and (33). Its important
part is the propagation of products in Vxx and Vxy , which helps in reducing
overestimation caused by the wrapping effect [Lo].

4.2 Input and output of one step of integration scheme.

As already mentioned, our C 1 algorithm builds on the C 0 algorithm proposed in
[WZ]. Inputs to the C 0 algorithms which computes bounds for (4) are

• [u](t0 ) – a set of initial conditions,

• h0 > 0 – a candidate for the time step.

Additional inputs to the C 1 algorithm are

• [V∗j ](t0 ), j = 1, . . . , m – leading columns of [V ](t0 ), which is an initial

condition for variational equation,
• z(t0 ) ∈ Rm+1 – vector of bounds on operator norms of rows in Vxy (t), that
is ∥ (Vxy )i∗ (t0 )∥ ≤ zi (t0 ), for i = 1, . . . , m and of ∥Vyy (t0 )∥ ≤ zm+1 (t0 ).

Successfully completed one step of integration scheme returns the following

Output:
• h ≤ h0 – accepted time step,
• [u](t0 + h) – a bound for the solutions at time t0 + h,

• [V∗j ](t0 + h), j = 1, . . . , m – a bound on leading columns of [V ](t0 + h),

• z(t0 +h) ∈ Rm+1 – such that ∥ (Vxy )i∗ (t0 +h)∥ ≤ zi (t0 +h), for i = 1, . . . , m
and ∥Vyy (t0 + h)∥ ≤ zm+1 (t0 + h),
• [E] – a rough enclosure, that is a bound for [u](t) ⊂ [E] for t ∈ [t0 , t0 + h],

16
 • [EV ] – a rough enclosure for variational equation, that is a bound for
[V ](t) ⊂ [EV ], for t ∈ [t0 , t0 + h].
Note, that usually we are interested in relatively short integration time, that
is computation of [u](T ) for T > 0 being a return to the Poincaré section. For
this purpose the algorithm is called iteratively. The enclosures [E] and [EV ]
are needed to compute intersection of a set of trajectories with surfaces, such
as in the case of finding bounds on Poincaré maps and their derivatives – see
Section 4.5.
The very preliminary step of the C 0 algorithm (see [Z02, WZ]) is computation
of the so called rough enclosure. Given a set of initial conditions [u](t0 ) and a
trial time step h0 the routine returns a pair
([E], h) ← enclosure([u](t0 ), h0 ), (35)
where h ≤ h0 is an accepted time step and [u](t) ⊂ [E] for t ∈ [t0 , t0 + h].
In particular, this routine validates that the solutions starting from [u](t0 ) are
defined for t ∈ [t0 , t0 + h]. If this step fails, the algorithm returns Failure and
we stop the computation.
In what follows we will show how Lm to compute an enclosure [EV ]. We will use
Theorem 8 for the decomposition i=1 ⟨ei ⟩ ⊕ Y , i.e. X is decomposed into 1-D
subspaces given by coordinate directions and initial condition W (0) = V (t0 ).
While B([0, h]) and B(h) give estimates for whole matrices V ([t0 , t0 + h]) and
V (t0 + h), we will be interested in the last column in matrix B as it contains the
bounds for ∥Vxi y ∥ and ∥Vyy ∥. The bounds for Vxx and Vyx obtained from matrix
B are too crude. The computation of these entries is discussed in Section 4.3.
Since the vector field is an explicit expression which satisfies on our enclo-
sures the standing assumptions C1,C2,C3,C4,D and VL , we can compute a
matrix J ∈ R(m+1)×(m+1) , such that for u ∈ [E] and A := A(u) = DF (u) there
holds (component-wise inequalities)
 
A11 |A12 | ··· |A1m | ∥[A1y ]∥
 |A21 | A22 ··· |A2m | ∥[A2y ]∥ 
 
 .. .
.. . .. .
.. ..
J ≥ . ,

 . 
 |Am1 | ··· |Am,m−1 | Amm ∥[Amy ]∥ 
∥[Ay1 ]∥ · · · ∥[Ay,m−1 ]∥ ∥[Aym ]∥ µ ([Ayy ])
where the operator norms depend on the choice of norms in X and Y .
Since ∥[Vxy ]i∗ (t0 )∥ ≤ zi (t0 ) for i = 1, . . . , m and ∥[Vyy ](t0 )∥ ≤ zm+1 (t0 ), by
Theorem 8 for t ∈ [t0 , t0 + h] there holds


∥[Vxy ]i∗ (t)∥ ≤ eJ(t−t0 ) z(t0 )
i , i = 1, . . . , m,
(36)
∥[Vyy ](t)∥ ≤ eJ(t−t0 ) z(t0 ) m+1 .

From these estimates we can obtain bounds

[EV ]xy = nM ∈ Rm × R∞ : ∥Mi∗ ∥ ≤ eJ[0,h] z(t0 ) i , io= 1, . . . , m ,


(37)
M ∈ R∞ × R∞ : ∥M ∥ ≤ eJ[0,h] z(t0 ) m+1 .


[EV ]yy =

17
4.3 Enclosure for Vxx and Vyx .
Propagation of blocks Vxx and Vyx is more subtle and before we give a routine
for computation of rough enclosure let us motivate why we need to compute
several objects.
It is well known that integration of differential equations may suffer from
the wrapping effect [Lo]. In order to reduce this overestimation Lohner [Lo]
proposes propagation of coordinate systems along the trajectory. This means,
that typical C 0 solver is almost C 1 as it keeps tract of approximate solution
to variational equations. This observation lead to the C 1 algorithm for finite
dimensional systems [Z02]. In our algorithm we take the advantage from the
linearity of the variational equation, that is its product property

V (t + h, u0 , V0 ) = V (h, u0 (t), Id) · V (t0 , u0 , V0 ),

where by V (t, u0 , V0 ) we understand a solution to variational equations after

time t and with the initial conditions u0 and V0 for the main equation and for
the variational equation, respectively.
In the infinite-dimensional case, however, this approach leads to mixing of
“near” and “far” columns in evaluation of this product, which causes huge over-
estimation. This is due to the fact, that “far” objects are roughly bounded by
their norms, while we would like to keep explicit intervals in the finite dimen-
sional block Vxx . In order to avoid this we propose to split the initial condition
for variational equations into two parts – as described in the sequel.
Now, let us denote by V (t, u, Vxx , Vxy , Vyx , Vyy ) a solution tovariational

Vxx Vxy
equation (33) with initial condition u for the C 0 part and V =
Vyx Vyy
for the C 1 part. To make the notation less cumbersome we will keep only one
argument in block component of V , that is
 
Vxx (t) Vxy (t)
:= V (t, u, Vxx , Vxy , Vyx , Vyy ).
Vyx (t) Vyy (t)

Given that the equation for variational equation is linear and non-auto-
nomous, we have

V (t, u, Vxx , Vxy , Vyx , Vyy ) = V (t, u, Vxx , 0, Vyx , 0) + V (t, u, 0, Vxy , 0, Vyy )

and Vxx (t), Vyx (t) appear in V (t, u, Vxx , 0, Vyx , 0), only. Since
 
V Vxy
V (t, u, Vxx , Vxy , Vyx , Vyy ) = V (t, u, Id, 0, 0, Id) xx ,
Vyx Vyy

we have
 
Vxx 0
V (t, u, Vxx , 0, Vyx , 0) = V (t, u, Id, 0, 0, Id)
0 0
 
0 0
+V (t, u, Id, 0, 0, Id) .
Vyx 0

18
It is easy to see that
   
V 0 V 0
V (t, u, Id, 0, 0, Id) xx = V (t, u, Id, 0, 0, 0) xx
0 0 0 0
and
 
0 0
V (t, u, Id, 0, 0, Id) = V (t, u, 0, 0, Vyx , 0).
Vyx 0
This gives
   
Vxx (t) 0 V 0
= V (t, u, Id, 0, 0, 0) xx + V (t, u, 0, 0, Vyx , 0). (38)
Vyx (t) 0 0 0

Observe that in the product of matrices in (38) we do not mix “near” and
“far” columns of V because the second matrix contains only one nonzero block
Vxx . The second component V (t, u, 0, 0, Vyx , 0) is computed independently and
we expect it should be very small due to the strong dissipation on higher modes.
Summarizing, in order to efficiently propagate bounds for m leading columns
of V we propose in each time step to integrate two different initial conditions.
Denote them by
   
Id 0 0 0
Ve = , Vb = .
0 0 [Vyx ](t0 ) 0
For each of the two initial conditions the rough enclosure for m leading
columns of the solution is computed by the same routine and for the same
vector field but for different initial conditions. Note, that in order to bound
vector field for variational equation we need already computed bound [E] for
the main equation (see (35)). We will emphasize this adding extra argument
[E] in the following sequence of calls:
 
[EeV ]∗1 , h ← enclosure([E], Ve∗1 , h),
  ...
[EV ]∗m , h
e ← enclosure([E], Ve∗m , h),
 
[EbV ]∗1 , h ← enclosure([E], Vb∗1 , h),
  ...
[EV ]∗m , h
b ← enclosure([E], Vb∗m , h).

Note, that each call to enclosure can adjust the time step. If none of the above
calls returns Failure, then we have computed a rough enclosure over time step
h for the main equation and for m leading columns of the variational equation
for two different initial conditions.
Now, let us set

[EV ]xx = [E
eV ]xx [Vxx ](t0 ) + [E
bV ]xx ,
(39)
[EV ]yx = [EV ]yx [Vxx ](t0 ) + [EV ]yx .
e b

19
Gathering (35), (37) and (39) we obtain the following partial output of our
algorithm.
Partial output:
• h – adjusted and finally accepted time step,
• [E] – rough enclosure for the main equation,
• [EV ] – rough enclosure for the variational equation.

4.4 Propagation of V (t).

Following [Z02] we represent finite block of V as the doubleton

[Vxx ](t0 ) = D + C[R] + B[R0 ],

where D, C, B are matrices and [R], [R0 ] are interval matrices containing zero.
The propagation of this block is realized as (compare (38))

[Vxx ](t0 + h) := πxx V (h, [u](t0 ), Id, 0, 0, 0) (D + C[R] + B[R0 ])

+ πxx V (h, [u](t0 ), 0, 0, [Vyx ](t0 ), 0). (40)

In order to reduce the wrapping effect the product of matrices in (40) is prop-
agated as suggested by Lohner [Lo] (see also [Z02, WW]). After that, a rather
thin part (i.e. having a small diameter) πxx V (h, [u](t0 ), 0, 0, [Vyx ](t0 ), 0) is ab-
sorbed into B[R0 ] in representation of [Vxx ](t0 + h) – see [Z02, Evaluation 4 in
Section 3.2].
Each column of Vyx is represented as a geometric series with decay q > 1
and a constant Cj = Cj (t0 ), that is

|Vij (t0 )| ≤ Cj q −i , i > m, j = 1, . . . , m.

We would like to compute new constants Cj (t0 + h), j = 1, . . . , m, so that

|Vij (t0 + h)| ≤ Cj (t0 + h)q −i , i > m, j = 1, . . . , m. (41)

From (38) we obtain

[Vyx ](t0 + h) = πyx V (h, [u](t0 ), Id, 0, 0, 0) [Vxx ](t0 )

+ πyx V (h, [u](t0 ), 0, 0, [Vyx ](t0 ), 0),

which for j = 1, . . . , m gives

m
X
Cj (t0 + h) = fk |[Vkj ](t0 )| + C
C cj , (42)
k=1

where C
fj and C
cj are computable constants satisfying

|Vij (h, [u](t0 ), Id, 0, 0, 0)| fj q −i ,

≤ C
cj q −i ,
|Vij (h, [u](t0 ), 0, 0, [Vyx ](t0 ), 0)| ≤ C

20
for i > m, j = 1, . . . , m. Indeed, these constants are computed by means of the
C 0 algorithm from [WZ] applied to variational equation with initial conditions
Ve∗j and Vb∗j , j = 1, . . . , m.
Finally, using estimates from (36) we can set z(t0 + h) := eJh z(t0 ) obtaining
bounds
zi (t0 + h) ≥ ∥[Vxy ]i∗ (t0 + h)∥ , i = 1, . . . , m,
(43)
zm+1 (t0 + h) ≥ ∥[Vyy ](t0 + h)∥ .
Gathering (40), (41), (42) and (43) we obtain remaining output of the algorithm.
Partial output:
• z(t0 + h) – a vector of bounds on norms in blocks Vxy and Vyy at time
t0 + h,
• [Vxx ](t0 + h) – finite block in the doubleton representation,
• [Vyx ](t0 + h) – tails of leading m columns represented as geometric series.

4.5 Bounds on derivatives of Poincaré map

Efficient computation of Poincaré map is performed as has been discussed in
[WZ, KaWZ]. In this section we show how to obtain bounds on derivatives of
Poincaré map from derivatives of the flow.
Let α : H → R be a smooth map such that
α(x) = α̃(x1 , x2 , . . . , xm ), (44)
so that is partial derivatives with respect to all variables xj , j > m are zero.
Let us define a Poincaré section by
Π = {x : α(x) = 0} .
Denote by T : H ⊃ dom(T ) → R the return time to the section and by P the
Poincaré map for a semiflow φ induced by x′ = f (x). In what follows we will
derive estimates for derivatives of T and P.
Differentiating the identity α(φ(T (x), x)) = 0 with respect to xj we have
 
∂T ∂φ
Dα(P(x)) (f (P(x))) · (x) + Dα(P(x)) (T (x), x) = 0. (45)
∂xj ∂xj
The quantity
m
!−1
−1
X
g := (Dα(P(x)) (f (P(x)))) = αxk (P(x))fk (P(x))
k=1

is independent from j and easily computable (finite sum) provided the vector
field is transverse to the section at P(x). Thus, due to (44), partial derivatives
of T are given by
m
∂T X ∂α ∂φk
(x) = −g (P(x)) (T (x), x). (46)
∂xj ∂xk ∂xj
k=1

21
 Differentiation of P(x) = φ(T (x), x) yields

∂Pk ∂φk ∂T
(x) = (T (x), x) + fk (P(x)) (x). (47)
∂xj ∂xj ∂xj

Let us recall, that the already proposed C 1 algorithm computes ∂φ ∂xj as ex-
k

plicit intervals, for k, j ≤ m, while three remaining blocks of Dx φ are estimated

uniformly by an operator norm – the details on evaluation of (46) and (47) will
be given in Appendix F.2.

5 Geometric tools for hyperbolicity - covering

relations and cone conditions
5.1 Covering relations.
Topological methods proved to be very useful in the context of computer-assisted
study of dynamical systems. One of the most efficient in the context of studying
chaotic dynamics is the method of covering relations introduced in [Z0, Z4] for
maps with one exit (”unstable”) direction and later extended to include many
exit directions in [ZGi] known also in the literature as the method of correctly
aligned windows [E1, E2]. In [WZ] this method has been adopted to the case of
convex and compact subsets of the real normed spaces. Here we recall definition
and the main topological result from [WZ].

Definition 6 Let X be a real normed space. A h-set, N = (|N |, cN , u(N )), is

an object consisting of the following data
• |N | ⊂ X – a compact set called the support of N,
• u(N ) – a non-negative integer,
• cN – a homeopmorphism of X such that

c−1
N (|N |) = Bu(N ) ⊕ TN =: Nc ,

where TN is a convex set.

The above definition generalizes the concept of h-sets introduced in [ZGi] for
finite-dimensional spaces, where a h-set is defined as the product of closed unit
balls B u × B s ⊂ Ru+s in a coordinate system cN . The next definition extends
the notion of covering relations from [ZGi] to compact maps acting on infinite-
dimensional real normed spaces.

Definition 7 Let N and M be h-sets in X and Y , respectively such that u =

u(N ) = u(M ). Let f : |N | → Y be a continuous map and set fc := c−1
M ◦ f ◦ cN .
f
We say that N f -covers M , denoted by N =⇒ M , if there is a linear map
L : Ru → Ru and a compact homotopy H : [0, 1] × Nc → Y , such that

22
Figure 3: An example of an h-set in three dimensions with u(N ) = 1 and
TN = D2 – a two-dimensional closed disc. Here Nc = B1 ⊕ D2 .

f
Figure 4: An example of an f −covering relation: N =⇒ M . In this case, the
homotopy joining fc (x, y) with a linear map (L(x), 0) and satisfying [CR1]–
[CR4] is simply given by H(t, x, y) = t(L(x), 0) + (1 − t)fc (x, y).

[CR1]: H(0, ·) = fc ,
[CR2]: H(1, x, y) = (L(x), 0), for all (x, y) ∈ Nc ,
[CR3]: H(t, x, y) ∈
/ Mc , for all (x, y) ∈ ∂Bu ⊕ TN , t ∈ [0, 1] and
[CR4]: H(t, x, y) ∈ Ru ⊕ TM for (x, y) ∈ Nc and t ∈ [0, 1].

A typical picture of a h-set with u(N ) = 1 is given in Figure 3. A picture

illustrating covering relation with one exit direction is given on Figure 4.
The following theorem is the main topological tools used in the proof of the
existence of symbolic dynamics in the KS equation [WZ, Theorem 1].

Theorem 9 [WZ, Theorem 4, Theorem 6] Let Xi , i ∈ Z be real normed spaces

and let Ni be h-sets in Xi , respectively. Assume
fi
Ni =⇒ Ni+1 for i ∈ Z.

23
Then there exists (ui )i∈Z such that
ui ∈ |Ni | for i ∈ Z, (48)
f (ui ) = ui+1 for i ∈ Z. (49)
Moreover, if the sequence of covering relations is periodic, that is there is K > 0
such that Ni = Ni+K and fi = fi+K then the sequence (ui )i∈Z can be chosen as
K-periodic, too.

5.2 Cone condition

The aim of this section is providing computational tool that guarantee conver-
gence of a trajectory of a discrete dynamical system to a periodic trajectory.
For this purpose we adopt the notion of cone condition from [KWZ, ZCC] and
the main results related to it to the case of real normed spaces.
We start from two definitions from [KWZ, ZCC] adopted to an infinite-di-
mensional case.
Definition 8 Let N ⊂ X be an h-set and let Q : X → R be homogenous of
degree two and continuous. The pair (N, Q) will be called an h-set with cones.
The function Q can be used to define cone fields by the condition Cx+ = {y ∈
X : Q(y − x) > 0} and Cx− = {y ∈ X : Q(y − x) < 0}.
Definition 9 [KWZ, Definition 13] Assume that (N, QN ), (M, QM ) are h-sets
f
with cones. Assume that N =⇒ M . We say that f satisfies the cone condition
with respect to the pair (N, M ), if for any x1 , x2 ∈ Nc with x1 ̸= x2 there holds
QM (fc (x1 ) − fc (x2 )) > QN (x1 − x2 ). (50)
Whenever it is convenient, we will also say that the cone conditions are satisfied
f
for the covering relation N =⇒ M , if the above condition is satisfied.
Condition (50) means that the cone field C + is preserved by the action of
f . Moreover, we have a kind of expansion and contraction in C + and C − ,
respectively.
The following theorem for finite-dimensional case directly extends to covering
relations in real normed spaces. It provides a tool for proving existence of
heteroclinic orbits between two periodic hyperbolic solutions. We would like to
emphasize, that the assumptions reduce to finite number of inequalities, and
thus in principle they can be checked by a computer.
Theorem 10 [W, Theorem 3.4] Assume that
f0 f1 f2 fn−1
N0 =⇒ N1 =⇒ N2 =⇒ · · · =⇒ Nn = N0 ,
0h 1 h2 h hm−2 hm−1
N0 =⇒ M1 =⇒ M2 =⇒ · · · =⇒ Mm−1 =⇒ K0
g0 g1 g2 gk−1
K0 =⇒ K1 =⇒ K2 =⇒ · · · =⇒ Kk = K0 ,
where

24
 • (Ni , QNi ) are h-sets with cones and fi satisfies the cone condition with
respect to (Ni , Ni+1 ), i = 0, . . . , n − 1
• (Ki , QKi ) are h-sets with cones and gi satisfies the cone condition with
respect to (Ki , Ki+1 ), i = 0, . . . , k − 1
Then there exist
1. a unique periodic orbit (u0 , . . . , un−1 ) with respect to (f0 , . . . , fn−1 ) and
the sequence (N0 , N1 , . . . , Nn−1 )
2. a unique periodic orbit (v0 , . . . , vk−1 ) with respect to (g0 , . . . , gk−1 ) and the
sequence (K0 , K1 , . . . , Kk−1 )
3. a full orbit (xi )i∈Z with respect to

(. . . , f0 , . . . , fn−1 , f0 , . . . , fn−1 ,
h0 , . . . , hm−1 ,
g0 , . . . , gk−1 , g0 , . . . , gk−1 , . . .) (51)

and the sequence

(N1 , N2 , . . . , Nn )N , M1 , M2 , . . . , Mm−1 , (K0 , K1 , . . . , Kk−1 )N (52)

such that

x0 ∈ |N0 |
lim xtk+m = v0 (53)
t→∞
lim xtn = u0 (54)
t→−∞

We omit the proof as the proof from finite-dimensional case is good also in the
present setting, as only the compactness and forward iterations are used.

6 Rigorous results about the KS equation

In this section we give computer-assisted proofs of Theorem 2 and Theorem 3.
Both proofs use C 1 algorithm described in Section 4 to obtain bounds on the
derivatives of certain Poincaré maps.

6.1 The KS equation in the Fourier basis

Consider the equation (1) with periodic and odd boundary conditions (2) and
assume that u is a classical solution to (1)–(2) given as a convergent Fourier
series
∞
X
u(t, x) = −2ak (t) sin(kx). (55)
k=1

25
 The particular form of representation of u given in (55) comes from imposing
on u(t, x) = k uk (t)eikx periodic and odd boundary conditions. Then we have
P
ak (t) = Im uk (t).
It is easy to see [CCP, ZM] that for sufficiently regular functions the system
(1)–(2) gives rise to an infinite ladder of coupled ODEs
k−1 ∞
dak X X
= Fk (a) := k 2 (1 − νk 2 )ak − k an ak−n + 2k an an+k , k ≥ 1. (56)
dt n=1 n=1

Define our space H = c0 - the space of sequence converging to zero with the
norm ∥a∥∞ = supk |ak |.
When discussing computing Poincaré
qP map of the induced semiflow
 we will use
an equivalent norm ∥a∥ = max |a | 2 , sup |a | for some m.
i≤m k k>m k
In order to apply Theorem 9 and Theorem 10 discussed in Section 5 to (56)
we need to chose the sets on which our convergence conditions C1-C4,D,VL will
be satisfied. The motivation for choosing sequences with geometrically decaying
coefficients is given in [WZ]. Here we introduce the notation.
For S > 0 and q > 1 we set

(ak )∞ −k

Wq,S = k=1 | ak ∈ R, |ak | ≤ Sq for k ≥ 1 .

Observe the for any S > 0 and q > 1 the set Wq,S ⊂ H is compact. In fact
Wq,S ⊂ H s ([−π, π]) is compact for any s, it is also compact in the space of
analytic 2π-periodic functions on the strip R × [−δ, δ] for δ < q. In fact a
coordinate-wise convergence in Wq,S is equivalent to convergence in C 0 , H s and
in the space of analytic functions. Indeed the equivalent norm in H s in terms
of the Fourier coefficient can be defined as
 1/2
X
∞ 2s 2
∥(ak )k=1 ∥H s =  |k| |ak |  , (57)
k≥1

and the convergence in such norm on Wq,S is equivalent the convergence for
each k, because the tails of elements from Wq,S are uniformly bounded in H s .
In Appendix E we show that on sets Wq,S conditions C1–C4 and condition
D are satisfied (see Theorem 25) and in Theorem 26 we show that condition
VL is satisfied on sets Wq,S × Wq,Sc . The sets with which we work with in our
proof are subsets of these sets.

6.2 Computer–assisted proof of Theorem 2.

In what follows, by a Poincaré section Θ we mean a subset of a codimension
one affine subspace Θ,
e such that the vector field is transverse to Θe on Θ. In the
applications the sets Θ will be intersections of Θ
e with the sets of the form Wq,S .
Let Θ1 , Θ2 be Poincaré sections for (56). By PΘ1 →Θ2 we denote the Poincaré

26
map between two sections, that is for a ∈ Θ1 the point PΘ1 →Θ2 (a) is the first
intersection of the trajectory of a with Θ2 , provided it exists. We also define
i i−1
PΘ 1 →Θ2
= PΘ2 →Θ2
◦ PΘ1 →Θ2 .

The following lemma has been proven with computer assistance and it was
the crucial step in the proof of Theorem 1.

Lemma 11 [WZ, Lemma 10] Consider the system (56) on l2,q with q = 32 .
There are explicitly given Poincaré sections Θ1 , Θ2 and Θi1→2 , Θi2→1 , i =
0, . . . , 10 and explicitly given pairwise disjoints h-sets N j ⊂ Θj , j = 1, 2,
i
N1→2 ⊂ Θi1→2 , N2→1
i
⊂ Θi2→1 , i = 0, . . . , 10 such that

P2 P3 P5 P5 P5 P5 P5
N 1 =⇒ N 1 =⇒ N1→2
0 1
=⇒ N1→2 2
=⇒ N1→2 10
=⇒ · · · =⇒ N1→2 =⇒ N 2 ,
P4 P4 P5 P4 P5
N 2 =⇒ N 2 =⇒ N2→1
0 1
=⇒ N2→1 2
=⇒ N2→1 =⇒ · · ·
P4 8 9P6 10 P5 P3
· · · =⇒ N2→1 =⇒ N2→1 =⇒ N2→1 =⇒ N 1 ,

where the starting and target sections in each of the above covering relations are
determined by the h-sets appearing in the relation.

The sets N j , j = 1, 2, are centred at approximate periodic points aj of

periods j, respectively, which correspond to periodic solutions uj of the original
i
system (1) with boundary conditions (2) — see Figure 1. The sets Nj→k are
i i
centred at approximate heteroclinic points a1→2 , a2→1 , i = 0, . . . , 10, found by
extensive numerical studies — see Figure 5.
In order to obtain connecting orbits between periodic points in N 1 and N 2
P2 P4
it suffices to check cone conditions for N 1 =⇒ N 1 and N 2 =⇒ N 2 and apply
Theorem 2. Preliminary simulation showed however, that in order to obtain
bounds on derivatives sharp enough to check the cone conditions on these sets
we would need to subdivide them into very large number of smaller sets (note
we have to subdivide in infinite-dimensional space). This would cause very large
time of computation.
To avoid this problem, we introduced intermediate sections and constructed
loops of covering relations satisfying cone condition, as shown in Figure 6. The
centres of new sets in these loops, that is M, K1 , K2 and K3 are located at
intersections of corresponding approximate periodic orbits with the Poincaré
section a1 = 0.
Proof of the following lemma is computer-assisted.

Lemma 12 There exist explicitly given Poincaré sections Θ11 , Θ21 , Θ22 and Θ23
• h-set with cones N 1 ⊂ Θ1 and h-set with cones M ⊂ Θ11 ,
• h-set with cones N 2 ⊂ Θ2 and h-sets with cones Ki ⊂ Θ2i , i = 1, 2, 3,

27
 a2 a91→2
-0.57
a81→2
a 1 →a 2
-0.58 a71→2 a 1

a3 -0.59

-0.60

-0.61

0.675 0.680 0.685 0.690 0.695 0.700

a2

a82→1
-0.57 a72→1

a2 a 2 →a 1
-0.58
a92→1
a1
a3 -0.59

-0.60

-0.61

0.675 0.680 0.685 0.690 0.695 0.700

a2

Figure 5: Numerically observed heteroclinic connections ai1→2 , ai2→1 i =

0, . . . , 10 between approximate periodic points a1 and a2 . All these points are
located in the section Θ = {a ∈ l2 : a1 = 0 ∧ a′1 = f1 (a) > 0} .

28
 N14→2 N15→2 N16→2
N13→2 N17→2
N12→2 N18→2
N11→2
N19→2
N10→2
N1 K1
N110→2 N2
M
K2
N20→1
N210→1 K3

N29→1
N21→1
N28→1 N22→1
N27→1 N23→1
N26→1 N25→1 N24→1

Figure 6: Graph of symbolic dynamics for the KS equation.

such that
P P
M =⇒ N 1 =⇒ M, (58)
2 P P P P 2
N =⇒ K1 =⇒ K2 =⇒ K3 =⇒ N . (59)

Moreover, the cone conditions are satisfied for these covering relations.
Proof: The trajectory of a1 , which is an approximate fixed point for P 2 ,
intersects the section a1 = 0 at a11 ≈ P(a1 ). At a11 we attached a Poincaré
section as a hyperplane almost orthogonal to the vector field at a11 . Then we
defined coordinate system in Θ11 as approximate eigenvectors of DP 2 (a11 ). In
this coordinate system we defined the set M as a parallelogram centred at a11
and with edges parallel to the axes.
Similarly, we set a2i ≈ P i (a2 ), i = 1, 2, 3. At each of these points we attached
a Poincaré section Θ2i as a hyperplane almost orthogonal to the vector field at
a2i . Then we defined coordinate systems in Θ2i as approximate eigenvectors of
DP 4 (a2i ), i = 1, 2, 3, respectively. Finally, we defined the sets Ki ⊂ Θ2i as
parallelograms centred at a2i .
Then, using algorithm for verification of covering relations [WZ, Appendix
A.4] we checked (58)–(59).
For verification of cone conditions we set homogeneous degree two continuous
functions associated with N 1 , N 2 , M , K1 , K2 and K3 as

QN 1 (a) = a21 − 0.5a22 − a23 − · · · − a217 − ∥y∥2∞ ,

QM (a) = 2a21 − a22 − a23 − · · · − a217 − ∥y∥2∞ ,
QN 2 (a) = a21 − a22 − a23 − · · · − a217 − ∥y∥2∞ ,
QK1 (a) = a21 − a22 − a23 − · · · − a217 − ∥y∥2∞ ,
QK2 (a) = a21 − a22 − a23 − · · · − a217 − ∥y∥2∞ ,
QK3 (a) = 4a21 − 2a22 − a23 − · · · − a217 − ∥y∥2∞ ,

29
 covering relation a d c ad − c2
P
N 1 =⇒ M ≥ 0.009 ≥ 0.999 ≤ 0.025 > 0.008
P
M =⇒ N 1 ≥ 0.469 ≥ 0.997 ≤ 0.042 > 0.4
2 P
N =⇒ K1 ≥ 0.626 ≥ 0.999 ≤ 0.008 > 0.6
P
K1 =⇒ K2 ≥ 0.882 ≥ 0.991 ≤ 0.103 > 0.8
P
K2 =⇒ K3 ≥ 0.041 ≥ 0.997 ≤ 0.1 > 0.03
P 2
K3 =⇒ N ≥ 0.142 ≥ 0.995 ≤ 0.063 > 0.1

Table 1: Bounds on a, d and c defined as in Lemma 27 for all covering relations

in Lemma 12.

Note, in each case the arguments of the above functions are given in section
coordinates specific for the h-set. To validate cone conditions we use Lemma 27
from the Appendix. To this end, we computed quantities a, d and c as defined
in Lemma 27. Obtained bounds are listed in Table 1. From these estimates
it is clear, that in each case a > 0 and ad > c2 and thus the assumptions of
Lemma 27 and in consequence the cone condition are satisfied.
Summarizing, the assumptions of Theorem 10 are satisfied for the set of
covering relations from Lemma 11 and (58)–(59). This completes a computer-
assisted proof of Theorem 2.

6.2.1 Implementation notes

The computation related to the existence of all covering relations from
P 2i
Lemma 11 and (58)–(59) except N i =⇒ N i , i = 1, 2, took approximately 95
minutes on a standard personal computer with AMD Ryzen 7 5800X 3.80GHz
processor and running 16 concurrent threads. This is slightly faster than the
time reported in [WZ], that is 40 minutes on 78 CPUs.
We observed, that the strategy of choice of one of the parameters in the
automatic differentiation procedure that computes Taylor coefficients of the so-
lutions to (56) has a huge impact on obtained bounds. This allowed us to enlarge
time steps of the solver and in consequence shorten the time of computation.
The vector Fk (a) when evaluated on a set a with geometric decaying tail
|ak | ≤ Cq −k for k > m and some C ≥ 0, q > 1 can be bounded by a polyloga-
e 4 q −k for some computable C
rithm function, that is |Fk (a)| ≤ Ck e ≥ 0. In [WZ,
Lemma 28] we shown that if we fix any δ ∈ (1, q), then |Fk (a)| ≤ D(q/δ)−k for
a computable D ≥ 0 and thus we obtained desired geometric decay of Fk . It
turned out that the obtained bounds are highly sensitive to the choice of δ. The
current improvement has been done by trial and choice of the strategy for δ.
This problem disappears when we use instead a larger family of sequences to
represent a and F (a), that is |ak | ≤ Ck s q −k for some C ≥ 0, q > 1 and s ∈ Z.
On the other hand, algebraic operations on such objects will be slower. Although

30
we did not made detailed analysis in this direction we believe, that overall this
is the right way to further reduce overestimation and speed up computation.
All above comments apply to computation of variational equations. First
version of the program that verified the cone condition run over 10 hours on a
computer with 448 CPUs. After these changes, and most important introduction
of intermediate sections (that is sets K, M1 , M2 , M3 ) we reduced the time of
computation related to cone condition to approximately an hour on a personal
computer running 12 concurrent threads.
The program have been written in C++ and is based on the CAPD library
[CAPD]. The source code can be found at [WSc].

6.3 Attracting periodic orbit for ν = 0.127

The goal of this section is to give a computer assisted proof of Theorem 3. The
system (56) admits a symmetry

S(a1 , a2 , a3 , a4 , . . .) = (−a1 , a2 , −a3 , a4 , . . .).

Let Θ = {(ai )∞i=1 : a1 = 0} be a Poincaré section. Since Θ = S(Θ) it follows

that S is also a symmetry for the associated Poincaré map P and the same
holds true for Poincaré maps for Galerkin projections. Since S is an involution,
we have P 2 = (S ◦ P)2 . Hence, it suffices to prove the existence of an attracting
fixed point for S ◦ P.
The existence of a fixed periodic point for S ◦ P is checked by means of the
Schauder fixed point theorem. For this we need to define a compact and convex
set K ⊂ Θ and check that K e := S ◦ P(K) ⊂ K. Then we will check that
∥D(SP)(K)∥e < 1. Observe that this implies that K is contraction on K.
Our space H = c0 - the sequences converging to zero. Let us fix m = 18.
We define norm on H by
 1,2 
m
 X
∥u∥ = max  |uj |2  , sup |uj | . (60)

j>m
j=1

It is immediate that H with above norm is gss space.

Let Pm denote Poincaré map for the flow induced by m-dimensional Galerkin
projection. The point u0 = (u1 , . . . , um ) ∈ Θm , where

u01 = 0 u02 = 1.1655388447652766

u03 = 5.7913642084409762 · 10−1 u04 = −2.768914122643753 · 10−1
u05 = −1.2582912341558827 · 10−1 u06 = 1.3015613792264165 · 10−2
u07 = 1.6757421753649995 · 10−2 u08 = 7.3178705708507339 · 10−4
u09 = −1.4755994209479948 · 10−3 u010 = −2.5601313918591901 · 10−4
u011 = 9.5427473299718919 · 10−5 u012 = 3.2627521780891864 · 10−5
u013 = −3.7164835203789059 · 10−6 u014 = −2.9884428114398931 · 10−6
u015 = −6.6059155560432874 · 10−8 u016 = 2.1596485662114442 · 10−7
u017 = 2.9759304684536878 · 10−8 u018 = −1.2262214633625861 · 10−8

31
is an approximate symmetric fixed point for S ◦ Pm (period two point for Pm ,
that is
∥S(Pm (u0 )) − u0 ∥2 ≈ 4 · 10−13
and it has been found by a simple iteration.
On the Poincaré section Θ we introduce an affine coordinate u = (u0 + A ·
Pm r, Qm r), where A is a matrix that almost diagonalizes SDPm (u) (nonrigorous
computation) and Pm , Qm are projections. In what follows we will work in this
coordinate system, that is r = 0 corresponds to (u0 , 0).
As a preliminary step we compute rigorously (using C 0 algorithm from [WZ])
a bound K 0 ⊃ S(P(0)). Clearly any reasonable choice of the set K should
contain K 0 . As a candidate for K we take

Ki = Ki0 · [−10, 10], i = 1, . . . , m

(note K10 = 0) and the tail is blown up by a factor 1.25, that is |Ki | ≤ 1.25Cq −i ,
where C is the constant bounding tail in K 0 , that is |Ki0 | ≤ Cq −i for i >
m. Using C 0 algorithm form [WZ] we compute a bound K e ⊃ SP(K) (in the
coordinate system mentioned above) and we get

i max |Ki | max |K

ei|
2 1.4037896821825919 · 10−8 8.7024122740689425 · 10−9
3 4.9756335159569671 · 10−9 8.5181999007451079 · 10−10
4 9.7156048870811418 · 10−11 6.6221478045718261 · 10−12
5 2.637578699436423 · 10−11 2.6821395044388368 · 10−12
6 8.491763367654622 · 10−12 8.6150267406921159 · 10−13
7 1.6209035540532317 · 10−11 1.7003223001285466 · 10−12
8 1.8203180667485086 · 10−11 1.8342179335531725 · 10−12
9 4.7853573165474597 · 10−11 4.8670536302275007 · 10−12
10 2.677612693142676 · 10−10 2.6635853628312471 · 10−11 . (61)
11 5.6520226737222243 · 10−11 5.6692000787577923 · 10−12
12 1.0201585471030834 · 10−10 1.013058647787179 · 10−11
13 4.341777521371891 · 10−10 4.3005552049615385 · 10−11
14 1.4398858307285336 · 10−10 1.4274077590841134 · 10−11
15 3.3157949280900669 · 10−10 3.2774891340532832 · 10−11
16 7.7478332603484965 · 10−10 7.6280987034794698 · 10−11
17 1.0424399029461924 · 10−9 1.0253441358095485 · 10−10
18 8.5396225348412251 · 10−10 8.3748052013014238 · 10−11
C 9.9201773737522156 · 10−6 7.9361550347737575 · 10−6

Because Ki is symmetric around zero it is clear from the above data that
e ⊂ K and thus the existence of a fixed point u∗ ∈ K
K e for S ◦ P is proved.
In the second step we compute a bound on the operator norm of A :=
SDP(K).
e On Poincaré section Θ = X × Y we use the norm ∥(x, y)∥ =
max {∥x∥2 , ∥y∥∞ }, where X is m−1-dimensional (the section is given by a1 = 0)

32
and Y is the tail. Using the C 1 algorithm from Section 4 we computed bounds

∥Axx ∥ ∈ [0.52569744154488263, 0.52578830885349082]

∥Axy ∥ ≤ 0.0021336347518391241
∥Ayx ∥ ∈ [5.6323483625902106 · 10−321 , 7.8313759074032856 · 10−18 ]
∥Ayy ∥ ≤ 7.8313759074032856 · 10−12 .
(62)
From the above data and the definition of the norm (see (60)) it is clear, that

∥A∥ = max {∥Axx ∥ + ∥Axy ∥, ∥Ayx ∥ + ∥Ayy ∥} ≤ 0.52792194360532996 < 1.

Remark 13 The computation regarding data in (61) and (62) took 11 minutes
and 73 minutes, respectively, in a single thread on laptop-type computer with
i7-8650U, 1.90GHz processor. The computation were performed in effective di-
mension m = 18 and with the tolerance (parameter of the algorithm specifying
the acceptable error per one time step of integration) set to tol = 10−11 .
Changing the setting to m = 15 and tol = 10−7 the validation also succeeded
within 26 seconds and 217 seconds for the C 0 and C 1 parts, respectively. Obtained
bound on the norm is much wider, that is

∥A∥ ≤ 0.92265954650065929,

although the norms in both cases are different (m = 18 and m = 15 give different
decomposition of the space). On the other hand, the stability domain is much
larger, as the set K on which we perform validation is much bigger (4 orders of
magnitude).
Appendix contains three parts. The first part consisting from Appendices
A, B, C, D has a theoretical character is an abstract exposition of our approach
together with proofs of convergence theorems stated in Section 2. Appendix E
contains a proof that for the sets used in our computations for KS equation all
conditions required for our approach are satisfied. The last part contains the
explicit formulas for various objects appearing in our algorithm.

A Logarithmic norms
Definition 10 [D58, L58] Let ∥ · ∥ denote a vector norm on Rn as well as its
subordinate operator norm on Rn×n . The logarithmic norm of operator A is

∥Id + hA∥ − 1
µ(A) = lim+ (63)
h→0 h
For the properties and the use of the logarithmic norm see [KZ] and the literature
given there.

33
 Consider the differential equation

x′ = f (x), f ∈ C 1. (64)

Let φ(t, x0 ) denote the solution of equation (64) with the initial condition x(0) =
x0 . By ∥x∥ we denote a fixed arbitrary norm in Rn .
From Lemma 6 one can easily derive the following result.
Lemma 14 Let y : [0, T ] → Rn be a piecewise C 1 function and φ(·, x0 ) be
defined for t ∈ [0, T ]. Suppose that Z is a convex set such that the following
estimates hold:

y([0, T ]), φ([0, T ], x0 ) ∈ Z

 
∂f
µ (η) ≤ l, for η ∈ Z
∂x
dy
(t) − f (y(t)) ≤ δ.
dt

Then for 0 ≤ t ≤ T there is

elt − 1
∥φ(t, x0 ) − y(t)∥ ≤ elt ∥y(0) − x0 ∥ + δ , if l ̸= 0.
l
For l = 0, there is

∥φ(t, x0 ) − y(t)∥ ≤ ∥y(0) − x0 ∥ + δt.

In the paper we frequently use norm ∥x∥∞ = maxi=1,...,n |xi | on Rn .

Lemma 15 Let A ∈ Rn×n , then the logarithmic norm of A induced by norm

∥ · ∥∞ is given by
 
n
X
µ(A) = max Ai i + |Aik | . (65)
i=1,...,n
k=1,k̸=i

B Sequence spaces with good norms

We adopt notion from the beginning of section 2.

Definition 11 We will say that H with norm ∥ · ∥ is gss (good sequence space)
if the following conditions are satisfied
S1 H is a Banach space,

S2 X
w= wi ei , ∀w ∈ H, (66)
i

34
S3 if w ∈ H, then for any α ∈ {−1, 1}Z+ wα =
P
i αi wi ei ∈ H and ∥w∥ =
∥wα ∥.
S4

∥PJ w∥ ≤ ∥w∥, ∀w ∈ H, ∀J ⊂ Z+ . (67)

Examples: l2 , l1 , c0 (sequences converging to 0) with the norm ∥ · ∥∞ .

However l∞ is not gss space, because S2 is not satisfied.

Lemma 16 If (H, ∥·∥) is gss , then for any J ⊂ Z+ projection PJ is continuous

and for any w ∈ H holds

∥w∥ = sup ∥Pn w∥, (68)

n∈Z+

∥(I − Pn )w∥ → 0, for n → ∞. (69)

Proof: Continuity of PJ follows immediately from S4.

Now we will establish (68) and (69). Let us fix w ∈ H. From S4 it follows
that for any n, M ∈ Z+ holds

∥Pn w∥ ≤ ∥w∥ ≤ ∥Pn w∥ + ∥(I − Pn )w∥, (70)

∥Pn w∥ ≤ ∥Pn+M w∥. (71)

From this it follows that condition (69) implies (68).

From S2 it follows that Pn w converges to w, therefore it satisfies the Cauchy
condition
∥Pn+M w − Pn w∥ ≤ ϵ, n ≥ N0 , M > 0.
After passing with M to ∞ we obtain

∥w − Pn w∥ ≤ ϵ, n ≥ N0 .

This proves (69) and together with (70) gives (68).

Lemma 17 Assume that H is gss . Then W ⊂ H is compact, iff it is closed,

bounded and we have uniform bound on the tail, i.e. for every ϵ > 0 there
exists N , such that for all n ≥ N for all w ∈ W holds ∥(I − Pn )w∥ < ϵ.

Proof: Implication ⇒. Assume that W is compact, then it is immediate that

it is bounded and closed. To establish the existence of uniform bound for the
tail let us fix ϵ > 0 and define an open covering {Un }n∈Z+ of W by setting
Un = {w ∈ H : ∥(I − Pn )w∥ < ϵ}. It is indeed a covering due to (69) in
Lemma 16. Observe that Un ⊂ Un+M for any n, M ∈ Z+ (this a consequence
of assumption S4). From compactness of W it follows that there exists N , such
that W ⊂ UN ⊂ UN +M for any M ∈ Z+ .
Implication ⇐. Consider a sequence ck ∈ W . We would like to prove that
there exists a convergent subsequence. From boundedness of W and the diagonal

35
argument it follows that that we can find a subsequence of ck , which we denote
by dk , such that for each j ∈ Z+ the sequence dkj is convergent to d¯k . Therefore
for each n the sequence Pn dk satisfies the Cauchy condition. We will show that
dk also satisfies the Cauchy condition.
Let us fix ϵ > 0. We have
∥dk − dk+m ∥ ≤ ∥Pn dk − Pn dk+m ∥ + ∥(I − Pn )dk ∥ + ∥(I − Pn )dk+m ∥.
From the uniform bound on tail in W we find n = N such that ∥(I − Pn )w∥ < ϵ
for w ∈ W , in particular for all dk . Let us fix this value of n. Given n from
the Cauchy condition for Pn dk it follows that there exists K, such that for all
k ≥ K and all m ∈ Z+ it follows that ∥Pn dk − Pn dk+m ∥ < ϵ.
Hence we obtain for k ≥ K and all m ∈ Z+
∥dk − dk+m ∥ ≤ 3ϵ.
Therefore sequence dk satisfies Cauchy condition. From completeness of H
(assumption S1) it follows that there exists d ∈ H, such that dk → d. Since W
is closed, so d ∈ W .
From the proof of the above lemma it is also easy to infer the following
result.
Lemma 18 Assume that H is gss and W ⊂ H is compact.
Consider a sequence ck in W . Then ck is convergent to c ∈ W , iff for all
j ∈ Z+ limk→∞ ckj = cj .
The following lemma gives a characterization of continuous linear forms on
H.
Lemma 19 Assume that H is and f : H → R be bounded linear functional.
Let fj = f (ej ). Then for every a ∈ H holds
X
|fj | · |aj | ≤ ∥f ∥ · ∥a∥ (72)
j∈Z+

and
X
f (a) = fj a j . (73)
j∈Z+

Proof: Let ā ∈ H be such that |aj | = |āj | and fj āj ≥ 0 for all j. Due to
condition S3 ā ∈ H and ∥a∥ = ∥ā∥. Since it term fj āj is nonnegative, we
obtain
X X
K∥a∥ = K∥ā∥ ≥ K∥Pn ā∥ ≥ fj · āj = |fj | · |aj |
j≤n j≤n

hence we obtained (72).

Equation (73) is a consequence of continuity of f and assumption S2.

36
Lemma 20 Assume H is gss . Assume that W ⊂ H satisfies conditions C1
and C2, and f : W → H ∗ is continuous. Then there exists L, such that

∥f|Hn (Pn z)∥ ≤ L, n ≥ M, z ∈ W. (74)

Proof: From compactness of W and continuity of f on W it follows that there

exists L ∈ R, such that
∥f (z)∥ ≤ L, z ∈ W.
From the above and condition C1 it follows that for all n ≥ M holds

∥f|Hn (Pn z)∥ ≤ ∥f (Pn z)∥ ≤ L, z ∈ W, n ≥ M.

This finishes the proof.

Lemma 21 Assume H is gss and f : H ⊃ dom (f ) → R is such that for all n

Hn ⊂ dom (f ) and f|Hn is C 1 .
Assume
• W ⊂ H satisfies conditions C1 and C2, W -convex
• f : W → H is continuous

• Assume that for all z ∈ W there exists limn→∞ D(f|Hn )(Pn z) and we
define
g(z) = lim D(F|H )(Pn z).
Df n
(75)
n→∞

g : W → H ∗ is continuous, and let us denote

We assume that Df

Df g(z)(ej ) = lim ∂f (Pn z).

g (z) = Df (76)
j
n→∞ ∂xj

Then
XZ 1
f (z) − f (w) = g (tz + (1 − t)w)dt · (zj − wj ),
Df j z, w ∈ W, (77)
j 0

Moreover, if for some z ∈ W and j there exists δ > 0, such that z + [−δ, δ]ej ∈
∂f
dom (f ), then ∂x j
(z) exists and the following equation is satisfied

g (z) = ∂f (z)
Df (78)
j
∂xj

Proof: Observe that for any v ∈ H and z ∈ H holds

Df
g(Pn z)(Pn v) = Df|H (Pn z)(Pn v).
n
(79)

37
 We have for any n ≥ M and z, w ∈ W (we use (79))
X Z 1 ∂f
f (Pn z) − f (Pn w) = (Pn (tz + (1 − t)w))dt · (zj − wj )
∂xj
j≤n 0
Z 1
= Df|Hn (Pn (tz + (1 − t)w))dt · Pn (z − w)
0
Z 1
= g(Pn (tz + (1 − t)w))dt · Pn (z − w)
Df
0

Now we pass to the limit n → ∞ in the above equation. From continuity f and
S2 it follows that f (Pn z) − f (Pn w) → f (z) − f (w) and Pn (z − w) → (z − w).
From continuity of Df
g on W follows that its uniform continuity, moreover from
Lemma 17 it follows that Pn converges to identity on W uniformly. Therefore
R1
the integral converges to 0 Dfg(tz + (1 − t)w)dt and we obtain
Z 1 
f (z) − f (w) = Df (tz + (1 − t)w)dt (z − w).
g
0
R1
g(tz + (1 − t)w)dt ∈ H ∗ , because it is bounded
Observe that 0 Df
by maxz∈W ∥Df
g(z)∥. Therefore from Lemma 19 we obtain assertion (77).
Assume that z ∈ W and z + hej ∈ W . Then from (77) we obtain
Z 1 
f (z + hej ) − f (z) = h Df
g(z + thej )dt (ej )
0
Z 1 Z 1  
=h Df
g (z + thej )dt = hDf
g (z) + h g (z + thej ) − Df
Df g (z) dt
j j j j
0 0

g it follows that the integral converges to 0 if h → 0.

From the continuity of Df j
Therefore we obtain (79).

C Convergence of Galerkin projections

We will here use the notations introduced in Section 2. The goal of this section
is to present the proof of the existence of the limit of Galerkin projections of
problem (4).
Proof of Theorem 4: Let

δn = max ∥Pn (F (x)) − Pn (F (Pn x))∥.

x∈W

Obviously δn → 0 for n → ∞, because F ◦ Pn converges uniformly to F on W

- this follows immediately from C2 and C3.
Let

ln = sup µ DPn F|Hn (x) . (80)
x∈Pn W

38
 From assumption D, it follows that ln are uniformly bounded, namely

ln ≤ l, for all n. (81)

Let us take k ≥ n. Let xn : [0, T ] → Pn W and xk : [0, T ] → Pk W be the

solutions of n- and k-dimensional projections of (4) with initial conditions in Z,
respectively.
Observe that for t ∈ [0, T ] holds

(Pn xk (t))′ = Pn F (xk (t))

= Pn F (Pn xk (t)) + (Pn F (xk (t)) − Pn F (Pn xk (t))),

hence we can treat y(t) = Pn xk (t) as ”perturbed solution” of n-th Galerkin

projection of (4) with the size of perturbation estimated by δn as follows

∥Pn F (xk (t)) − Pn F (Pn xk (t))∥ ≤ δn .

From Lemma 14 it follows immediately that

elt − 1
∥xn (t) − Pn (xk (t))∥ ≤ elt ∥xn (0) − Pn xk (0)∥ + δn (82)
l
To prove the uniform convergence of {xn } starting from the same initial
condition, observe that

∥xn (t) − xk (t)∥ ≤ ∥xn (t) − Pn (xk (t))∥ + ∥(I − Pn )xk (t)∥ ≤
elt − 1 elT − 1
δn + ∥(I − Pn )xk (t)∥ ≤ δn + ∥(I − Pn )W ∥.
l l
Since by Lemma 17 ∥(I − Pn )W ∥ → 0 for n → ∞ we see that {xn } is a Cauchy
sequence in C([0, T ], H), hence it converges uniformly to x∗ : [0, T ] → W .
Since xn (t) satisfies an integral equation
Z t
xn (t) = xn (0) + Pn F (xn (s))ds,
0

then from the uniform continuity of F on W it follows that after passing to the
limit n → ∞ that Z t
x∗ (t) = x∗ (0) + F (x∗ (s))ds,
0
which implies that x∗ is a solution of (4).
Uniqueness. Let x : [0, T ] → W be a solution of (4) with the initial
condition x(0) = x0 . We will show that xn converge to x. We apply Lemma 14
to n-dimensional projection and the function Pn x(t). We obtain

elt − 1
∥xn (t) − Pn (x(t))∥ ≤ δn . (83)
l
Since the tail ∥(I − Pn )x(t)∥ ≤ ∥(I − Pn )W ∥ is by Lemma 17 uniformly con-
verging to zero as n → ∞, we see that xn → x uniformly.

39
 Lipschitz constant on W . From Lemma 14 applied to n-dimensional
Galerkin projection for different initial conditions (we denote the functions by
xn and yn and the initial conditions x0 and y0 ), we obtain

∥xn (t) − yn (t)∥ ≤ elt ∥Pn x0 − Pn y0 ∥. (84)

Let xn → x and yn → y. Then passing to the limit in (84) gives

|x(t) − y(t)∥ ≤ elt ∥x0 − y0 ∥. (85)

Assertion 4 follows easily from the previous ones.

C.1 Galerkin projection of linear maps on gss

We will be interested in the following question: given linear maps (for example
the Jacobian matrices for the flow defined by Galerkin projections) V n : Hn →
Hn , does there exists a limit V : H → H, under some assumptions, which are
reasonable in the context of dissipative ODEs.
Consider a decomposition of H
M
H= ⟨ek ⟩ ⊕ Y, Y = (I − Pm )H
k≤m

We will adopt notation regarding blocks of V as in Section 3.1.

However care should be applied when discussing the norm of the block Vij .
Namely, we have
∥Vij ∥ = ∥ei ∥ · |Vij |/∥ej ∥. (86)
In the sequel V n : Hn → Hn will be treated as a linear map from H to H
by setting V n (u) = V n (Pn u) ∈ Hn ⊂ H. We will represent it as a matrix in
the basis e1 , e2 , . . . . For a matrix V = {Vij } ∈ RZ+ ×Z+ by V∗j we will denote
its j-th column.

Theorem 22 Assume that V n : H → H are linear maps such that,

∥V n ∥ ≤ K (87)

and for all i, j there exists Vij , such that

lim Vijn = Vij (88)

n→∞

and there exists a family {Wcj }j∈Z+ such that Wcj ⊂ H is a compact set and
for all n and j holds
n
V∗j ∈ Wcj . (89)
Then there exists a bounded linear map V : H → H, such that

πi V (ej ) = Vij ei , ∀i, j (90)

40
for all a ∈ H holds
lim V n (a) = V a, (91)
n→∞
Moreover,
∥V ∥ ≤ K (92)

Proof:
Let us fix a ∈ H. We will show that V n a is a Cauchy sequence.
We have for any n, k, m, M ∈ Z+

∥V n+k a − V n a∥ ≤ ∥V n+k (Pm a) − V n (Pm a)∥ + ∥V n+k (I − Pm )a∥

+∥V n (I − Pm )a∥
≤ ∥PM V n+k (Pm a) − PM V n (Pm a)∥ + ∥(I − PM )V n+k (Pm a)∥
+∥(I − PM )V n (Pm a)∥ + ∥V n+k (I − Pm )a∥ + ∥V n (I − Pm )a∥
M
X X
≤ ∥ei ∥ |Vijn+k − Vijn | · |aj | + ∥(I − PM )V n+k (Pm a)∥
i=1 j≤m
+∥(I − PM )V n (Pm a)∥ + 2K∥(I − Pm )a∥

Now we are ready to show the Cauchy condition. Let us take ϵ > 0. Then by
taking m big enough by Lemma 16 (condition (69)) we obtain that 2K∥(I −
Pm )a∥ ≤ ϵ. Let us fix such value of m.
Now we look at ∥(I − PM )V ℓ (Pm a)∥ for any ℓ. Observe that V ℓ (Pm a) is
linear combination of first m columns in V ℓ . Indeed,
X
V ℓ Pm a = V∗jℓ
aj .
j≤m

From this and assumption (89) it follows that any ℓ

 
X
(I − PM )V ℓ Pm a ⊂ (I − PM )  Wcj aj 
j≤m
P
The set j≤mWcj aj is compact, hence from Lemma 17 it follows there exists
P 
M , such that ∥(I − PM ) j≤m W cj aj ∥ ≤ ϵ/2. We fix such value of M .
PM
From condition (88) it follows that i=1 ∥ei ∥ j≤m |Vijn+k − Vijn | · |aj | ≤ ϵ
P
for n large enough and any k.
Therefore we obtained that ∥V n+k a − V n a∥ ≤ 3ϵ for n large enough.
V n a converge to V a, where V : H → H is linear map satisfying ∥V ∥ ≤ K.
Since V n ej = ei Vijn , we see that Vij = limn→∞ Vijn .

Theorem 23 Assume that (H, ∥ · ∥) is gss space.

Assume V n ∈ Rn×n , V ∈ RZ+ ×Z+ and B ∈ R(m+1)×(m+1) , holds

Vijn → Vij ∀i, j ∈ Z+ , (93)

41
there exists a family {Wcj }j∈Z+ such that Wcj ⊂ H is a compact set and for all
n and j holds
n
V∗j ∈ Wcj , (94)
and

∥Vijn ∥ ≤ Bij , i, j ≤ m (95)

∥Viyn ∥ ≤ Bi,m+1 , i≤m (96)
n
∥Vyj ∥ ≤ Bm+1,j , j≤m (97)
n
∥Vyy ∥≤ Bm+1,m+1 . (98)
P∞
Then for each i and all a ∈ H j=1 |Vij | · |aj | < ∞,
and {Vij } defines a bounded linear
 operator V : H → H,
P∞
such that πi V (a) = V a
j=1 ij j ei , and we have the following bounds

∥Vij ∥ ≤ Bij , i, j ≤ m (99)

∥Viy ∥ ≤ Bi,m+1 , i≤m (100)
∥Vyj ∥ ≤ Bm+1,j , j≤m (101)
∥Vyy ∥ ≤ Bm+1,m+1 . (102)

Proof: We want to apply Theorem 22. First we need to obtain K, such that

∥V n ∥ ≤ K. (103)

This is an easy consequence of (95–98). Pm P

Indeed, we have for a ∈ H, a = ax + ay = j=1 ai ei + j>m ai ei

∥V n a∥ ≤ ∥Pm V n a∥ + ∥(I − Pm )V n a∥
≤ ∥Pm V ax ∥ + ∥Pm V ay ∥ + ∥(I − Pm )V n ax ∥ + ∥(I − Pm )V n ay ∥
n n
XX X
≤ ∥Vijn ∥ · ∥ej ∥ · |aj | + ∥Viyn ∥ · ∥ay ∥
i≤m j≤m i≤m
X
n n
+ ∥Vyj ∥ · ∥ej ∥ · |aj | + ∥Vyy ∥ · ∥ay ∥
j≤m

Since by condition S4 ∥ej ∥·|aj | ≤ ∥a∥ and ∥ay ∥ ≤ ∥a∥ we obtain from conditions
(95–98) the following estimate for ∥V n a∥
   
XX X
∥V n a∥ ≤  Bij  ∥a∥ +  Bi,m+1  · ∥a∥
i≤m j≤m i≤m
   
X X
+ Bm+1,j  · ∥a∥ + Bm+1,m+1 · ∥a∥ =  Bij  ∥a∥.
j≤m i,j≤m+1

Therefore we proved (103).

42
 From Theorem 22 it follows that for each a ∈ H and sequence V n a converges
to V a and V : H → H is continuous.
Condition (99) follows immediately from (93) and (95). Observe that blocks
Viyn , Vyin and Vyy
n
can be extended to a linear maps preserving the norms given
as πi V (I − Pm )), (I − Pm )V n πi and (I − Pm )V n (I) − Pm ), respectively and
n

the same is true for blocks in V .

For each of these blocks from convergence of V n a to V a, the same follows
for the blocks. The norms for the blocks are uniformly bounded due to (96–
98). By Theorem 22 these bounds are also valid for blocks of V , i.e. conditions
(100–102) are satisfied.

D Proof of Theorem 5.
We extend V n : [0, T ] × Pn (Z) → Lin(Hn , Hn ) to V n : [0, T ] × Z → Lin(H, H)
by setting V n (t, x) = V n (t, Pn x).
We apply Theorem 4 to variational system for j-th column. From VL it
follows that all its assumptions are satisfied for system (19) with ZV = Z × {ej }
and WV = W × WCj with l = A from condition VL -D. Therefore there exists
a family of continuous functions V∗j : [0, ∞) × Z → H for j ∈ Z+ , such that for
n
each j the functions V∗j converge to V∗j uniformly on [0, T ] × Z.
Moreover, (z(t), V∗j (t, z)) satisfies variational equation (19).
Now we prove (22) and (23).
Let us fix t ∈ [0, T ].
For any n we have
Z 1
n n
φ (t, Pn x) − φ (t, Pn y) = V n (t, Pn y + s(Pn x − Pn y))ds · Pn (x − y) (104)
0

We will pass to the limit n → ∞ in above equation. The limit of l.h.s. is

settled by Theorem 4.
Now we consider the r.h.s. We will use Theorem 22 to pass to the limit
n → ∞.
From condition D and Lemma 14 it follows that for all n, x, y ∈ Z and
t ∈ [0, T ]
∥φn (t, x) − φn (t, y)∥ ≤ elt ∥x − y∥,
hence
∥V n (t, x)∥ ≤ elt , ∀n, (t, x) ∈ [0, T ] × Z. (105)
n 1
Let V (t, [x, y]) = 0 V n (t, y + s(x − y))ds. From (105) we have for x, y ∈ Z
R

n
∥V n (t, x)∥ ≤ elt , ∥V (t, [x, y])∥ ≤ elt . (106)

From convexity of WCj it follows that for all j holds

n n
V∗j (t, x), V (t, [x, y])∗j ∈ WCj . (107)

43
Since for all i, j Vijn (t, ·) are continuous on Z and converge uniformly to Vij (t, ·),
therefore
Z 1 Z 1
n
Vij (t, y + s(x − y))ds → Vij (t, y + s(x − y))ds. (108)
0 0
R1
Now we use Theorem 22 to conclude that V ∈ Lin(H, H), 0
V (t, y + s(x −
y))ds ∈ Lin(H, H) and
Z 1 Z 1
n
lim V (t, y + s(x − y))ds · (x − y) = V (t, y + s(x − y))ds · (x − y). (109)
n→∞ 0 0

Observe that from formula (22) we obtain (23). Indeed, we have

Z 1
φ(t, x + hej ) − φ(t, x) = h Vj (t, x + shej )ds · ej
0
Z 1
= hV (t, x)ej + h (V (t, x + shej )ej − V (t, x)ej ) ds
0
Z 1
= hV∗j (t, x) + h (V∗j (t, x + shej ) − V∗j (t, x)) ds
0

and the result follows from continuity of V∗j (t, ·).

It remains to prove that for every a ∈ H the map [0, T ] × Z ∋ (t, z) 7→
V (t, z)a ∈ H is continuous.
We know that for every j the map (t, x) 7→ V∗j (t, z) is continuous.
Let (t1 , z1 ), (t2 , z2 ) ∈ [0, T ] × Z. Then for any n holds

∥V (t1 , z1 )a − V (t2 , z2 )a∥ ≤ ∥V (t1 , z1 )Pn a − V (t2 , z2 )Pn a∥

+∥V (t1 , z1 )(I − Pn )a∥ + ∥V (t2 , z2 )(I − Pn )a∥ ≤
X
∥V (t1 , z1 )∗j − V (t2 , z2 )∗j ∥ · |aj | + (et1 l + et2 l )∥(I − Pn )a∥
j∈Jn

For a given ϵ > 0 we take n so large that ∥(I − Pn )a∥emax(0,l)T < ϵ/2. We fix
such n. Then from the continuity of V∗j for j ∈ Jn we can make the sum less
that ϵ/2 when (t2 , z2 ) → (t1 , z1 ). This finishes the proof.

E Convergence conditions for the KS equation

In what follows we assume that H = c0 is a space of sequences converging to
zero with norm supremum norm ∥u∥ = supn |un |.
For S > 0 and q > 1 we define

Wq,S = (uk )∞ −k

k=1 ∈ H : |uk | ≤ Sq . (110)

44
 From (56) we see that

∂Fi 2iω(i, k)z|i−k| + 2izi+k , for k ̸= i;
(z) = , (111)
∂zk i2 (1 − νi2 ) + 2izi+k , for i = k;

where ω(i, k) = −1 if i > k and ω(i, k) = 1 if i < k.

Lemma 24 For z ∈ Wq,S holds

4iSq −|i−k| ,

∂Fi for k ̸= i;
(z) ≤ (112)
∂zk |i2 (1 − νi2 )| + 2iSq −(i+k) , for i = k;

Theorem 25 Let W = Wq,S for some S ≥ 0, q > 1. Then the conditions

C1–C4 and D for the KS equation (56) are satisfied on W .

Proof:
C1: Obvious.
C2: the set is compact according to the compactness criterion from Lemma 17.
C3: For u ∈ Wq,S we have (see (56))

k−1
X ∞
X
|Fk (u)| ≤ Sq −k |k 2 (1 − νk 2 )| + k Sq −n Sq n−k + 2k Sq −n Sq −n−k
n=1 n=1
= Sq −k |k 2 (1 − νk 2 )| + k(k − 1)S 2 q −k + 2kS 2 q −k (q 2 − 1)−1
≤ Sq −k νk 4 + k 2 + Sk(k − 1) + 2k 2 S(q 2 − 1)−1 ≤ Ck 4 q −k , (113)

for some C = C(S, q).

Hence limk→ ∞ |Fk (u)| = 0 and thus F (u) ∈ H.
Observe that from the above computations follows that Fk : W → R is con-
tinuous of a limit of uniformly convergent functions F Pn , which are continuous
on W .
We will prove that F is (uniformly) continuous on W . Let us fix ε > 0. For
any u, v ∈ W using uniform bound (113) we obtain

|Fk (u) − Fk (v)| ≤ |Fk (u)| + |Fk (v)| ≤ 2Ck 4 q −k .

Hence, there exists M ∈ N such that for k > M and for all n ∈ Z+ there holds

∥(I − PM )F (u) − (I − PM )Fk (v)∥ ≤ ε, ∀u, v ∈ W. (114)

From continuity of Fk it follows that PM F is uniformly continuous on W ,

hence there exists δ > 0 such that

∥PM F (u) − PM F (v)∥ ≤ ε, if ∥u − v∥ ≤ δ. (115)

Thus F is continuous on W .

C4: Let us fix z ∈ W and define a sequence Vk = ∂F

∂zk (z). We will show that
i

∗
P
V u = k Vk uk defines a functional V ∈ H and V = limn→∞ D(Fi |Hn )(Pn z).

45
 First we will show that V ∈ H ∗ .
For z ∈ W and u ∈ H formula (112) gives

∞ ∞ ∞
!
X X ∂Fi X
|Vk uk | ≤ ∥u∥ ≤ ∥u∥ |i2 (1 − νi2 )| + 2iSq −2i + 8iS q −k
∂zk
k=1 k=1 k=1
= ∥u∥ |i2 (1 − νi2 )| + 2iSq −2i + 8iS(q − 1)−1 .

Hence V is bounded and thus V ∈ H ∗ .

Now we will show that DFgi (z) = V .
We have to find (V − D(Fi |Hn )(Pn z)) u. It is easy to see that for n ≥ i holds
!
X X
(V − D(Fi |Hn )(Pn z)) u = 2i zk−i uk + zi+k uk (116)
k>n k>n−i

Therefore for u ∈ H, ∥u∥ = 1 we have

X ∞
X
|(V − D(Fi |Hn )(Pn z)) u| ≤ 2i |zk−i | + 2i |zi+k |
k>n k>n−i
∞
X
≤ 4iS q −k = 4iSq i−n (q − 1)−1 .
k=n−i+1

Hence limn→∞ ∥(V − D(Fi |Hn )(Pn z))∥ = 0.

There remains to show that DFgi is continuous on W . Let us fix ε > 0. Let
us denote V (z) = DFi (z) for z ∈ W . Then for u ∈ H such that ∥u∥ = 1 we
g
have
∞
X ∞
X
∥(V (v) − V (z))u∥ ≤ 2i|z|i−k| − v|i−k| | + 2i |zi+k − vi+k |
k=1,k̸=i k=1
∞
X
≤ 6i |zk − vk |.
k=1
P∞
There is M ∈ N such that for all z, v ∈ W there holds k=M +1 |zk − vk | ≤ ε/2.
PM
Now if ∥v − z∥ ≤ δ, then k=1 |zk − vk | ≤ M δ ≤ ε/2 for δ small enough.
Summarizing, for sufficiently small δ there holds

∥V (v) − V (z)∥ ≤ 6iε, if ∥z − v∥ ≤ δ, z, v ∈ W

and thus DF
gi is continuous on W .
D: Let n be dimension of the Galerkin projection. Formula (65) implies that


µ DPn F|Hn (z) = max Ri,n (z), (117)
i≤n

46
where for any i ≤ n Ri,n (z) is defined and estimated using Lemma 24 for
z ∈ Pn W as follows
n
∂Fi X ∂Fi
Ri,n (z) := (z) + (z)
∂zi ∂zk
k̸=i,k=1
X
≤ i2 (1 − νi2 ) + 4iSq −|i−k| 2iSq −2i
k,k̸=i
∞
X
≤ i2 (1 − νi2 ) + 8iS q −k + 2iSq −2i ≤ i2 (1 − νi2 )
k=1
1
= i2 (1 − νi2 ) + iS( + q −2i ) =: Ri,∞
q−1
We see that Ri,∞ → −∞ if i → ∞. It is immediate that there exists l ∈ R for
all z ∈ Pn W holds


µ DPn F|Hn (z) = max Ri,n (z) ≤ sup Ri,∞ = l.
i≤n i

Theorem 26 Let W = Wq,S and Wc = Wq,Sc for some q > 1 and S, Sc ≥ 0.

Then the system (15-16) for the KS equation (56) satisfies condition VL on
W × Wc .

Proof:
VL -C1: Obvious.
VL -C2: the set Wc is compact according to the compactness criterion from
Lemma 17.
VL -C3: Let us fix (u, C) ∈ W × Wc . In order to show that {}i∈N ∈ H we
gi (u)C = 0. From (24) and using |uk | ≤ Sq −k and
have to show that limi→∞ DF
−k
|Ck | ≤ Sc q we obtain
X
gi (u)C| ≤ |i2 (1 − νi2 )Ci | + 2iSq −2i |Ci | + 4iS
|DF q −|i−k| |Ck |
k,k̸=i
X
−i −3i
2 2
≤ |i (1 − νi )|Sc q + 2iSSc q + 4iSSc q −|i−k|−k
k,k̸=i
X X
−i −3i −i
2 2
≤ |i (1 − νi )|Sc q + 2iSSc q + 4iSSc q + 4iSSc q −i−2k
k<i k>i
1
≤ |i2 (1 − νi2 )|Sc q −i + 2iSSc q −3i + 4i2 SSc q −i + 4iSSc q −i . (118)
q2 − 1

gi (u)C = 0 and thus {DF

Hence limi→∞ DF gi (u)C}i∈Z ∈ H.
+

Finally we have to show that FV (u, C) = (F (u), {DF

gi (u)C}i∈Z ) is contin-
+
uous on H × H. From C3 we know that F (u) is continuous, so there remains to

47
check continuity of the second component. Let us fix ε > 0. Using uniform and
independent on n bound (118) we obtain that there exists constant K1 ∈ R+
such that for any (u, C) ∈ W × Wc

gi (u)C| ≤ K1 i4 q −i .
|DF

Hence, there exists M ∈ N such that

∥(I − PM )DF
gi (u)C∥ ≤ ε. (119)

Using continuity of DF
gi (condition C4) and compactness of W we can find
a constant K2 such that

∥DF
gi (u)∥ ≤ K2 , u ∈ W, i ≤ M. (120)

From uniform continuity of DF gi on W (condition C4) it follows that there exists

δ > 0, such that for all u, v ∈ W such that

∥DF
gi (u) − DF
gi (v)∥ ≤ ε, i = 1, . . . , M, if ∥u − v∥ ≤ δ. (121)

Let us fix this value of δ.

Let (u, Cu ), (v, Cv ) ∈ W × Wc , such that ∥u − v∥ ≤ δ. Then for i ≤ M we
obtain (we use (120) and (121))

|DF
gi (u)Cu − DF
gi (v)Cv | ≤ |DF
gi (u)(Cu − Cv )| + |(DF
gi (u) − DF
gi (v))Cv |

≤ ∥DF
gi (u)∥ · ∥Cu − Cv ∥ + ∥DF
gi (u) − DF
gi (v)∥ · ∥Cv ∥
≤ K2 · ∥Cu − Cv ∥ + max ∥C∥ε.
C∈Wc

Summarizing, from (119) and the above estimate we obtain

n o
gi (un )C n − DF
DF gi (u)C ≤ max{2ε, K2 ∥Cu − Cv ∥ + max ∥C∥ε},
i∈Z+ C∈Wc

which can be made arbitrarily small with suitable choice of small ε and ∥Cu −
Cv ∥ < δ ′ with δ ′ small. Thus FV is continuous.
VL -D. We have to find a uniform bound on logarithmic norms of
DPn FV |Hn ×Hn , for n ∈ Z+ . Such derivative is of the following form

M n := DPn FV |Hn ×Hn (u, C)

"
DPn F |Hn (Pn u) o
#
0
= .
n
gi |H (Pn u))Pn C
D(DF DPn F |Hn (Pn u)
n
i≤n

An 0
 
n
Let us denote blocks of the above matrix by M = . Recall, that
B n An
the logarithmic norm related to the supremum norm (see (65)) is given by

48
µ(M n ) = maxi Miin + k̸=i |Mik n
P
|. It is immediate that this maximum is realized
for rows related to variables Ci .
From (111) it follows that Ci -th component of FV denoted by FV,Ci is given
by
X X
FV,Ci = 2i ω(i, j)z|i−j| Cj + 2i zi+j Cj + i2 (1 − νi2 )Ci , (122)
j,j̸=i j

hence 
∂FV,Ci  −2iCi−k + 2iCi+k , i > k;
(z, C) = 2iCi+k + 2iCk−i , i < k;
∂zk
2iC2i , i = k.


The sum related to Ci -th row for n-th Galerkin projection has the following form
(we use symbols Ri,n (z)) and Ri,∞ introduced during the proof of Theorem 25
to establish condition D )
n
X ∂Pn FV,Ci
RCi ,n (z, C) := (Pn z, Pn C) + Ri,n (z)
∂zk
k=1
∞
X ∞
X ∞
X
≤ 2i |C|i−k| | + 2i |Ci+k | + Ri,∞ ≤ 6i |Ck | + Ri,∞
k=1,k̸=i k=1 k=1
1 1
= 6iSc + i2 (1 − νi2 ) + iS( + q −2i ) =: RCi ,∞
q−1 q−1
We see that RCi ,∞ → −∞ if i → ∞. It is immediate that there exists L ∈ R

RCi ,∞ ≤ L.

This finishes the proof of condition VL -D.

F Explicit formulas
In this entire section the phase space H = c0 - the sequences converging
to zero, is decomposed as H = X ⊕ Y equipped with the norm ∥(x, y)∥ =
max {∥x∥2 , ∥y∥∞ }. For z = (x, y) we also use indexing x = (x1 , . . . , xm ) ∈
X, y = (ym+1 , . . .) ∈ Y .
Moreover, we assume that the z, P(z) ∈ Wq,S , for some constants q > 1 and
S ≥ 0.

F.1 Some formulas for norms of blocks of derivatives

A ∈ Lin(H, H). A is split into four blocks denoted by
 
Axx Axy
A= .
Ayx Ayy

49
We willalso use Aky to denotean operator Aky : Y → R given as the row matrix
Aky = Ak,m+1 Ak,m+2 · · · . The norms of the blocks of A are given by the
following explicit formulas.

X
∥Aky ∥ = |Akj | (123)
j>m
X
∥Ayy ∥ = sup ∥Aky ∥ = sup |Akj | (124)
k>m k>m j>m

∥Ayk ∥ = sup |Ajk | . (125)

j>m
 1/2
X 2
X
∥Akx ∥ =  (Akj ) ≤ |Akj | (126)
j≤m j≤m
 1/2
X 2
X
∥Ayx ∥ = sup  |Akj |  ≤ sup |Akj | . (127)
k>m j≤m k>m j≤m

F.2 Partial derivatives of Poincaré map

In this section we describe how to evaluate (46–47). Let us denote

∂T
Wkj (x) = fk (P(x)) (x) (128)
∂xj
∂φ
In the sequel we will use the following notation V (x) = ∂x (T (x), x) and V∗∗ are
blocks in V . Then
DP = V + W. (129)
Let D be a constant such that for k > m there holds

|fk (P(z)| ≤ D (130)

Let us consider four cases corresponding to four block of DP we have to

compute.
Block Wxx . For k, j ≤ m all terms in (46) are given by explicit numbers
(intervals).
Block Wyx . Put Wyx := Wk,j for j ≤ m < k. From (127) and (128) we obtain
 1/2
Xm m
X
2
∥Wyx ∥ ≤ sup  |Wkj |  ≤ sup |Wkj |
k>m j=1 k>m j=1

m m
X ∂T X ∂T
≤ sup fk (P(z)) (z) ≤ max |fk (P(z))| (z) (, 131)
k>m j=1 ∂xj k>m
j=1
∂xj

50
 ∂T
For j ≤ m derivatives of ∂xj in (46) are given by explicit numbers. From
(130) we obtain
m m
X ∂T X ∂T
max |fk (P(z))| (z) ≤ D (z) . (132)
k>m
j=1
∂xj j=1
∂xj

Gathering (131)–(132) we get a final bound

m
X ∂T
∥Wyx ∥ ≤ D (z) .
j=1
∂xj

Rows Wky . In this case we want to estimate operator norm of each row (that
is fixed k ≤ m) separately. Computation similar to (131) leads to an estimate
 1/2
X 2
X X ∂T
∥Wky ∥ ≤  |Wkj |  ≤ |Wkj | = fk (P(z)) (z)
j>m j>m j>m
∂xj
X ∂T
≤ |fk (P(z))| (z) . (133)
j>m
∂xj

From (46) we have (see (123))

X ∂T m
XX ∂φi
(z) ≤ |g| αxi (P(z)) (T (z), z)
j>m
∂x j j>m i=1
∂xj
 
Xm X ∂φi Xm
= |g| |αxi (P(z))| (T (z), z)  = |g| |αxi (P(z))| ∥Viy ∥. (134)
i=1 j>m
∂xj i=1

Gathering (133)–(134) we get

m
X
∥Wky ∥ ≤ |g| |fk (P(z))| |αxi (P(z))| ∥Viy ∥.
i=1

Block Wyy . Computation similar to (133)–(134) gives an estimate

m
!
X
∥Wyy ∥ ≤ max |g| |fk (P(z))| |αxi (P(z))| ∥Viy ∥
k>m
i=1
m
X
≤ |g|D |αxi (P(z))| ∥Viy ∥. (135)
i=1

G How to verify the cone condition for h-sets?

Recall the phase space is decomposed as H = X × Y and equipped with the
norm ∥(x, y)∥ = max {∥x∥2 , ∥y∥∞ }. Let N, M be h-sets with cones of the form
X
QK (x, y) = QK,1 (x) − ∥y∥2∞ , (136)
1≤k≤m

51
where X
QK,1 (x) = qK,k x2k . (137)
1≤k≤m

and K = N or K = M . We will write qk instead if K will be clear from the

context.
Let f be a continuous map expressed in the coordinate systems of N and
M , respectively. Assume that the map f (x, y) = (fx (x, y), fy (x, y)) has all
continuous partial derivatives and denote
 
h
∂f
i f (z) fxy (z)
Df (z) = ∂zji (z) = xx
i,j∈N fyx (z) fyy (z)

for z ∈ N . The existence of continuous partial derivatives allows us to use Euler

formula
Z 1 
f (z1 ) − f (z2 ) = Df (z1 + t(z2 − z1 ))dt (z1 − z2 ). (138)
0

We would like to derive set of computable inequalities that will guarantee

f
that the cone condition is satisfied for the covering relation N =⇒ M . To
simplify notation will use ∥ · ∥ to denote norms of x and y.
Let us fix zi = (xi , yi ) ∈ N , i = 1, 2 and set dx = x1 − x2 and dy = y1 − y2 .
We define f∗∗ as the integral average of partial derivatives on interval joining
z1 and z2 , for example
Z 1
fxx = fxx (z1 + t(z2 − z1 )dt. (139)
0

In a similar way we define fxk and fky for k = 1, . . . , m. Using this notation we
have
fx (z1 ) − fx (z2 ) = fxx dx + fxy dy,
(140)
fy (z1 ) − fy (z2 ) = fyx dx + fyy dy.
Now we derive a formula for

QM (f (z1 ) − f (z2 )) = QM,1 (fx (z1 ) − fx (z2 )) − ∥fy (z1 ) − fy (z2 )∥2 .

In what follows we Set qk = qM,k . Using (140) we have

QM,1 (fx (z1 ) − fx (z2 )) = QM,1 (fxx dx + fxy dy),

X
2 X
2
= qk fkx dx + fky dy = qk fkx dx
k≤m k≤m
X

X
2
+ 2qk fkx dx fky dy + qk fky dy .
k≤m k≤m

52
The three expressions in the above sum can be bounded as follows
X
2
= dxT (fxx )T QM,1 fxx dx,


qk fkx dx
k≤m
 
X

X
2qk fkx dx fky dy ≤ 2 |qk | ∥fkx ∥ ∥fky ∥ ∥dx∥ ∥dy∥,
k≤m k≤m
 
X
2 X
qk fky dy ≤  |qk | · ∥fky ∥2  ∥dy∥2 .
k≤m k≤m

Since

∥fy (z1 ) − fy (z2 )∥ = ∥fyx dx + fyy dy∥ ≤ ∥fyx ∥ ∥dx∥ + ∥fyy ∥ ∥dy∥

we finally obtain

QM (f (z1 ) − f (z2 )) ≥ dxT (fxx )T QM,1 fxx dx +

 
X
−2  |qk | · ∥fkx ∥ · ∥fky ∥ ∥dx∥ · ∥dy∥ +
k≤m
 
X
− |qk | · ∥fky ∥2  ∥dy∥2 − (∥fyx ∥ · ∥dx∥ + ∥fyy ∥ · ∥dy∥)2
k≤m

= dxT (fxx )T QM,1 fxx dx − ∥fyx ∥2 ∥dx∥2 +

  
X
−2  |qk | · ∥fkx ∥ · ∥fky ∥ + ∥fyx ∥ · ∥fyy ∥ ∥dx∥ · ∥dy∥ +
k≤m
  
X
−  |qk | · ∥fky ∥2  + ∥fyy ∥2  ∥dy∥2 .
k≤m

Summarizing, we obtain

QM (f (z1 ) − f (z2 )) − QN (z1 − z2 ) ≥ a∥dx∥2 − 2c∥dx∥ · ∥dy∥ + d∥dy∥2 (141)

where a, c, d are such that

dxT (fxx )T QM,1 fxx dx − ∥fyx ∥2 ∥dx∥2 − QN,1 (dx) ≥ a∥dx∥2 ,

  
X
 |qk | · ∥fkx ∥ · ∥fky ∥ + ∥fyx ∥ · ∥fyy ∥ ≤ c,
k≤m
  
X
1 −  |qk | · ∥fky ∥2  + ∥fyy ∥2  > d.
k≤m

53
Observe that a∥dx∥2 − 2c∥dx∥ · ∥dy∥ + d∥dy∥2 is positive for all dx ̸= 0 or dy ̸= 0
if and only if a > 0 and ad > c2 . Therefore we have proved the following lemma.

Lemma 27 Assume (N, QN ), (M, QM ) are h-sets with cones of the form (136).
Denote by [Dfxx (N )]I a convex hull of the set of matrices Dfxx (z), z ∈ N .
Assume that there is a > 0 such that for A ∈ [Dfxx (N )]I
 
T 2
A QM,1 A − QN,1 − sup ∥fyx (z)∥ Id − aId
z∈N

is positive definite.
Assume also that c ≥ 0, d > 0 are constants such that
  
X
sup  |qM,k | · ∥fkx (z)∥ · ∥fky (z)∥ + ∥fyx (z)∥ · ∥fyy (z)∥ ≤ c,
z∈N
k≤m
  
X
1 − sup  |qM,k | · ∥fky (z)∥2  + ∥fyy (z)∥2  > d.
z∈N
k≤m

If ad > c2 , then QM (f (z2 ) − f (z1 )) > QN (z2 − z1 ) for z1 , z2 ∈ N , z1 ̸= z2 .

H Changing coordinate system on the section.

Here we want to represent DP in a linear coordinate system A(x,
e y) = (A(x), y),
that is we want to compute
 −1  
M= A e DP A e = Mxx Mxy .
Myx Myy

Straightforward computation gives

Mxx = A−1 Pxx A

Mxy = A−1 Pxy
Myx = Pyx A
Myy = Pyy

Note, that coefficients in Pxx are given as explicit intervals, all columns in the
block Pyx are given as geometric series and we have only bounds on the operator
norms Dy Pk , k = 1, . . . , n and Pyy . Clearly Mxx can be computed by direct
multiplication of interval matrices and ∥Myy ∥ = ∥Pyy ∥. For the remaining
blocks we have the following estimates.

54
H.1 Computation of Mxy .
Put B := A−1 and denote by Mk the operator norm of k-th row of Mxy . The
range is one-dimensional and in the domain we have l∞ norm.
m
X X m
XX m
X X
Mk = Bki Pij ≤ |Bki ||Pij | = |Bki | |Pij |
j>m i=1 j>m i=1 i=1 j>m
Xm
= |Bki |∥Dy Pi ∥.
i=1

H.2 Computation of Myx .

Here we use direct bound

∥Myx ∥ ≤ ∥Pyx ∥ · ∥A∥2

References
[AK10] G. Arioli and H. Koch, Integration of dissipative PDEs: a case study,
SIAM J. of Appl. Dyn. Syst., 9:1119–1133, (2010).
[AZ] G. Arioli, P. Zgliczyński, Symbolic dynamics for the Hénon–Heiles
hamiltonian on the critical energy level, J. Diff. Eq, 171, 173–202
(2001)
[CAPD] T. Kapela, M. Mrozek, D. Wilczak, P. Zgliczyński, CAPD::DynSys:
a flexible C++ toolbox for rigorous numerical analysis of dynamical
systems, Communications in Nonlinear Science and Numerical Simu-
lation 101 (2021), 105578.
[CEES] P. Collet, J.-P. Eckmann, H. Epstein, J. Stubbe Analyticity for the
Kuramoto–Sivashinsky equation, Phys. D 67, (1993), 321–326.
[CCP] F. Christiansen, P. Cvitanovic, V. Putkaradze, Spatiotemporal chaos
in terms of unstable recurrent patterns, Nonlinearity 10, (1997), 55–70
[C] J. Cyranka, Efficient and generic algorithm for rigorous integration
forward in time of dPDEs: part I, J. Sci. Comp. 59, 28–52 (2014).
[D58] G. Dahlquist, Stability and Error Bounds in the Numerical Intgration
of Ordinary Differential Equations, Almqvist & Wiksells, Uppsala,
1958; Transactions of the Royal Institute of Technology, Stockholm,
1959.
[E1] R. Easton, Isolating Blocks and Symbolic Dynamics, J. Differential
Equations, 17 (1975), 96–118.

55
[E2] R. Easton, Homoclinic Phenomena in Hamiltonian Systems with Sev-
eral Degrees of Freedom, J. Differential Equations, 29 (1978), 241–252.
[F78] M. J. Feigenbaum, Quantitative universality for a class of nonlinear
transformations, J. Stat. Phys., 19 (1978), 25–52.

[FNST] C. Foias, B. Nicolaenko, G. Sell, R. Temam Inertial manifolds for

the Kuramoto–Sivashinsky equation and an estimate of their lowest
dimension. J. Math. Pures Appl. 67, (1988), 197–226.
[FT] C. Foias and R. Temam, Gevrey Class Regularity for the Solutions
of the Navier-Stokes Equations, Journal of Functional Analysis, Vol.
87, No. 2, 1989, 359–369,
[FGLL] J.L. Figueras, M. Gameiro, J.P. Lessard, R. de la Llave A Framework
for the Numerical Computation and a Posteriori Verification of In-
variant Objects of Evolution Equations, SIAM J. of Appl. Dyn. Syst.,
16(2), 1070–1088 (2017)

[FL17] J.L. Figueras, R. de la Llave, Numerical Computations and Computer

Assisted Proofs of Periodic Orbits of the Kuramoto-Sivashinsky Equa-
tion, SIAM Journal on Applied Dynamical Systems 16 (2), 834–852
(2017)
[GL17] M. Gameiro and J.P. Lessard, A Posteriori Verification of Invariant
Objects of Evolution Equations: Periodic Orbits in the Kuramoto-
Sivashinsky PDE, SIAM J. of Appl. Dyn. Syst., 16(1), 687–728,
(2017)
[GZ] Z. Galias and P. Zgliczyński, Computer assisted proof of chaos in the
Lorenz system, Physica D, 115, 1998,165–188

[GH] Guckenheimer J., Holmes P. (1983): Nonlinear Oscillations, Dy-

namical Systems, and Bifurcations of Vector Fields, Springer Verlag,
Berlin, Heidelberg, New York.
[H] D. Henry, Geometric Theory of Semilinear Parabolic Equations, Lect.
Notes in Math. 840, Springer 1981.

[HN] J. Hyman, B. Nicolaenko, The Kuramoto–Sivashinsky equation; A

bridge between PDEs and dynamical systems , Phys. D 18, (1986),
113–126.
[NST] B. Nicolaenko, B. Scheurer, R. Temam, Some global dynamical prop-
erties of the Kuramato-Sivashinsky equations: Nonlinear stability and
attractors, Phys. D 16, (1985), 155–183.
[HNW] E. Hairer, S.P. Nørsett and G. Wanner, Solving Ordinary Differential
Equations I, Nonstiff Problems, Springer-Verlag, Berlin Heidelberg
1987.

56
[JJK] M.E. Johnsohn, M. Jolly, I. Kevrekidis, The oseberg transition: vi-
sualization of the global bifurcations for the Kuramoto-Sivashinsky
equation, Int. J. of Bif. Chaos 11, (2001), 1–18.
[JKT] M. Jolly, I. Kevrekidis, E. Titi, Approximate inertial manifolds for the
Kuramoto–Sivashinsky equation: analysis and computations, Phys. D
44, (1990) 38–60.
[KZ] T. Kapela, P. Zgliczyński, A Lohner-type algorithm for control sys-
tems and ordinary differential inclusions, Discrete Cont. Dyn. Sys.
B, 11 (2009) 365–385.

[K] H.-O. Kreiss, Fourier expansions of the solutions of the Navier-Stokes

equations and their exponential decay rate, in Analyse mathématique
et applications, 245–262, Gauthier-Villars, Montrouge, 1988.
[KT] Y. Kuramoto, T. Tsuzuki, Persistent propagation of concentration
waves in dissipative media far from thermal equilibrium, Prog. Theor.
Phys., 55,(1976), 365
[KWZ] H. Kokubu, D.Wilczak, P. Zgliczyński, Rigorous verification of cocoon
bifurcations in the Michelson system, Nonlinearity, 20 (2007), 2147-
2174.
[KaWZ] T. Kapela, D.Wilczak, P. Zgliczyński, Recent advances in rigorous
computation of Poincaré maps, Communications in Nonlinear Science
and Numerical Simulation 110, July 2022, 106366.
[Lo] R.J. Lohner, Computation of Guaranteed Enclosures for the Solutions
of Ordinary Initial and Boundary Value Problems, in: Computational
Ordinary Differential Equations, J.R. Cash, I. Gladwell Eds., Claren-
don Press, Oxford, 1992.
[L58] S. M. Lozinskii, Error esitimates for the numerical integration of or-
dinary differential equations, part I, Izv. Vyss. Uceb. Zaved. Matem-
atica,6 (1958), 52–90 (Russian)
[MM1] K. Mischaikow and M. Mrozek, Chaos in the Lorenz Equations: a
Computer Assisted Proof , Bull. of AMS, vol. 32, pp. 66–72, 1995.
[MM2] K. Mischaikow and M. Mrozek, Isolating neighborhoods and chaos,
Jap. J. Ind. & Appl. Math., 12, pp. 205–236, 1995.
[Mo] R.E. Moore, Methods and Applications of Interval Analysis, SIAM,
Philadelphia, 1979.
[Mos] J. Moser, Stable and Random Motions in Dynamical Systems, Prince-
ton Univ. Press, 1973.

57
[NJP] Nedialkov, N. S., Jackson, K. R., Pryce, J. D., 2001. An effective
high-order interval method for validating existence and uniqueness of
the solution of an IVP for an ODE. Reliab. Comput. 7 (6), 449–465.
[RNS] T. Rage, A. Neumaier, C. Schlier, Rigorous verification of chaos in a
molecular model, Phys. Rev. E 50 (1994), 2682–2688

[S] G.I. Sivashinsky, Nonlinear analysis of hydrodynamical instability in

laminar flames – 1. Derivation of basic equations, Acta Astron. 4
(1977), no. 11-12, 1177–1206.
[SP] Y.S. Smyrlis and D.T. Papageorgiou, Predicting chaos for infinite-
dimensional dynamical systems: the Kuramoto-Sivashinsky equa-
tion, a case study. Proc. Nat. Acad. Sci. U.S.A. 88 (1991), no. 24,
11129–11132
[T1] W. Tucker, Lorenz attractor exists, C. R. Acad. Sci. Paris, t. 328,
Série I, p. 1197-1202, 1999

[T2] W. Tucker, A Rigorous ODE solver and Smale’s 14th Problem, Foun-
dations of Computational Mathematics, (2002), Vol. 2, Num. 1, 53-
117
[WW] I. Walawska, D. Wilczak, An implicit algorithm for validated enclo-
sures of the solutions to variational equations for ODEs, App. Math.
Comp., 291C, (2016) 303–322.
[W] D. Wilczak, Abundance of heteroclinic and homoclinic orbits for the
hyperchaotic Rössler system, Disc. Cont. Dyn. Sys. B, 11/4 (2009),
1039–1055.

[WSc] D. Wilczak, program https://github.com/dbwilczak/ks-chaos

[WSB] D. Wilczak, S. Serrano, R. Barrio, Coexistence and dynamical con-
nections between hyperchaos and chaos in the 4D Rössler system: a
Computer-assisted proof, SIAM J. App. Dyn. Sys., Vol. 15, No. 1
(2016), 356-390.

[WZ03] D. Wilczak and P. Zgliczyński, Heteroclinic Connections between Pe-

riodic Orbits in Planar Restricted Circular Three Body Problem - A
Computer Assisted Proof, Comm. Math. Phys. 234 (2003) 1, 37-75
[WZ] D. Wilczak, P. Zgliczyński, A geometric method for infinite-dimen-
sional chaos: Symbolic dynamics for the Kuramoto-Sivashinsky PDE
on the line, J. Diff. Eq., 269/10 (2020), 8509–8548.
[Z] P. Zgliczyński, Trapping regions and an ODE-type proof of the exis-
tence and uniqueness for Navier-Stokes equations with periodic bound-
ary conditions on the plane, Univ. Iag. Acta Math., 41 (2003) 89–113.
zz

58
[Z0] P. Zgliczyński, Fixed point index for iterations of maps, topological
horseshoe and chaos, Topol. Methods Nonlinear Anal. 8, (1996), 169–
177.
[Z1] P. Zgliczyński, On smooth dependence on initial conditions for dis-
sipative PDEs, an ODE-type approach, J. Diff. Eq., 195/2 (2003),
271–283.
[Z02] P. Zgliczyński, C 1 -Lohner algorithm, Found. Comp. Math., 2 (2002),
429–465.
[Z4] P. Zgliczyński, Computer assisted proof of chaos in the Rössler equa-
tions and in the Hénon map, Nonlinearity 10 (1), (1997) 243–252.
[ZM] P. Zgliczyński and K. Mischaikow, Rigorous Numerics for Partial
Differential Equations: the Kuramoto-Sivashinsky equation, Found.
Comp. Math., (2001) 1:255–288.
[ZGi] P. Zgliczyński and M. Gidea, Covering relations for multidimensional
dynamical systems, J. Differential Equations 202/1(2004), 33–58.
[ZCC] P. Zgliczyński, Covering relations, cone conditions and the stable
manifold theorem, J. Differential Equations, 246/5(2009), 1774–1819.
[Z04] P. Zgliczyński, Rigorous numerics for dissipative Partial Differential
Equations II. Periodic orbit for the Kuramoto-Sivashinsky PDE - a
computer assisted proof, Found. Comput. Math., 4 (2004), 157–185.
[Z10] P. Zgliczyński, Rigorous Numerics for Dissipative PDEs III. An ef-
fective algorithm for rigorous integration of dissipative PDEs, Topol.
Methods Nonlinear Anal., 36 (2010) 197-–262.

59