Chemical Reaction Systems with a Homoclinic

Bifurcation: an Inverse Problem

Tomislav Plesa∗ Tomáš Vejchodský† Radek Erban∗

Abstract: An inverse problem framework for constructing reaction systems with prescribed prop-
erties is presented. Kinetic transformations are defined and analysed as a part of the framework,
allowing an arbitrary polynomial ordinary differential equation to be mapped to the one that can
be represented as a reaction network. The framework is used for construction of specific two- and
three-dimensional bistable reaction systems undergoing a supercritical homoclinic bifurcation, and
the topology of their phase spaces is discussed.

1 Introduction
Chemical reaction networks under the mass action kinetics are relevant for both pure and applied
mathematics. The time evolution of the concentrations of chemical species is described by kinetic
equations which are a subset of first-order, autonomous, ordinary differential equations (ODEs) with
polynomial right-hand sides (RHSs). On the one hand, the kinetic equations define a canonical form
for analytic ODEs, thus being important for pure mathematics [16, 18]. They can display not only
the chemically regular phenomenon of having a globally stable fixed point, but also the chemically
exotic phenomena (multistability, limit cycles, chaos). It is then no surprise that chemical reaction
networks can perform the same computations as other types of physical networks, such as electronic
and neural networks [23]. On the other hand, reaction networks are a versatile modelling tool,
decomposing processes from applications into a set of simpler elementary steps (reactions). The
exotic phenomena in systems biology often execute specific biological functions, example being the
correspondence between limit cycles and biological clocks [26, 27].
The construction of reaction networks displaying prescribed properties may be seen as an inverse
problem in formal reaction kinetics [4], where, given a set of properties, a set of compatible reaction
networks is searched for. Such constructions are useful in application areas such systems biology (as
caricature models), synthetic biology (as blueprints), and numerical analysis (as test problems) [25,
30]. In systems biology, kinetic ODEs often have higher nonlinearity degree and higher dimension,
thus not being easily amenable to mathematical analysis. Having ODEs with lower nonlinearity
degree and lower dimension allows for a more detailed mathematical analysis, and also adds to the
set of test problems for numerical methods designed for more challenging real-world problems. In
synthetic biology, such constructed systems may be used as a blueprint for engineering artificial
networks [30].
A crucial property of the kinetic equations is a lack of so-called cross-negative terms [2], corre-
sponding to processes that involve consumption of a species when its concentration is zero. Such
terms are not directly describable by reactions, and may lead to negative values of concentrations.
The existence of cross-negative terms, together with a requirement that the dependent variables
are always finite, imply that not every nonnegative polynomial ODE system is kinetic, and, thus,
∗
Mathematical Institute, University of Oxford, Radcliffe Observatory Quarter, Woodstock Road, Oxford, OX2
6GG, United Kingdom; e-mails: tomislav.plesa@some.ox.ac.uk, erban@maths.ox.ac.uk
†
Institute of Mathematics, Czech Academy of Sciences, Žitná 25, Praha 1, 115 67, Czech Republic, e-mail: vej-
chod@math.cas.cz

1
further constrain the possible dynamics, playing an important role in the construction of reaction
systems, chemical chaos, and pattern formation via Turing instabilities [3, 5]. A trivial example of
an ODE with a cross-negative term is given by dx/dt = −k, for constant k > 0, where the term −k,
although a polynomial of degree zero, nevertheless cannot be directly represented by a reaction,
and results in x < 0.
In two dimensions, where the phase plane diagram allows for an intuitive reasoning, the exotic
dynamics of ODE systems reduces to limit cycles and multistability. While two-dimensional nonk-
inetic polynomial ODE systems exhibiting a variety of such dynamics can be easily found in the
literature, the same is not true for the more constrained two-dimensional kinetic ODE systems.
Motivated by this, this paper consists of two main results: firstly, building upon the framework
from [4, 2], an inverse problem framework suitable for constructing the reaction systems is pre-
sented in Section 3, with the focus on the so-called kinetic transformations, allowing one to map a
nonkinetic into a kinetic system. Secondly, in Section 4, the framework is used for construction of
specific two- and three-dimensional bistable kinetic systems undergoing a global bifurcation known
as a supercritical homoclinic bifurcation. The corresponding phase planes contain a stable limit
cycle and a stable fixed point, with a parameter controlling the distance between them, and their
topology is discussed. Definitions and basic results regarding reaction systems are presented in
Section 2. A summary of the paper is presented in Section 5.

2 Notation and definitions

The notation and definitions in this paper are inspired by [1, 8, 2].
Definition 2.1. Let R be the space of real numbers, R≥ the space of nonnegative real numbers,
R> the space of positive real numbers and N = {0, 1, 2, 3, . . . } the set of natural
P numbers. Given
a finite set S, with cardinality |S| = S, the real space of formal sums c = s∈S cs s is denoted by
RS if cs ∈ R for all s ∈ S. It is denoted by RS≥ if cs ∈ R≥ for all s ∈ S; by RS> if cs ∈ R> for all
s ∈ S; and by NS if cs ∈ N for all s ∈ S; where the number cs is called the s-component of c for
s ∈ S. Support of c ∈ RS is defined as supp(c) = {s ∈ S : cs 6= 0}. Complement of a set M ⊂ S is
denoted by Mc , and given by Mc = S \ M.
The formal sum notation is introduced so that unnecessary ordering of elements of a set can be
avoided, such as when general frameworks involving sets are described, and when objects under
consideration are vector components with irrelevant ordering. The usual vector notation is used
when objects under consideration are equations in matrix form, and is put into using starting with
equation (3.3).

2.1 Reaction networks and reaction systems

Definition 2.2. A reaction network is a triple {S, C, R}, where
(i) S is a finite set, with elements s ∈ S called the species of the network.

C ⊂ NS is a finite set, with elements c ∈ C, called the complexes of the network, such that
(ii) S
c∈C supp(c) = S. Components of c are called the stoichiometric coefficients.

(iii) R ⊂ C × C is a binary relation with elements r = (c, c0 ), denoted r = c → c0 , with the following
properties:

(a) ∀c ∈ C (c → c) ∈
/ R;

2
 (b) ∀c ∈ C ∃c0 ∈ C such that (c → c0 ) ∈ R or (c0 → c) ∈ R.

Elements r = c → c0 are called reactions of the network, and it is said that c reacts to c0 , with
the reactant complex, and c0 the product complex. The order of reaction r is
c being called P
given by or = s∈S cs < ∞ for r = c → c0 ∈ R.

Note that as set R implies sets S and C, reaction networks are denoted with R, for brevity. Also,
as it is unlikely that a reaction between more than three reactants occurs [2], in this paper we
consider reactions with or ≤ 3. To represent some of the non-chemical processes as quasireactions,
the zero complex is introduced, denoted with ∅, with the property that supp(∅) = ∅, where ∅ is
the empty set.

Definition 2.3. Let R be a reaction network and let κ : RS≥ → RR ≥ be a continuous function which
maps x ∈ RS≥ (called “species concentrations”) into κ(x) ∈ RR ≥ (called “reaction rates”). Then κ
is said to be a kinetics for R provided that, for all x ∈ R≥ and for all r = (c → c0 ) ∈ R, positivity
S

κr (x) > 0 is satisfied if and only if supp(c) ⊂ supp(x).

An interpretation of Definition 2.3 is that a reaction, to which a kinetics can be associated, can
occur if and only if all the reactant species concentrations are nonzero.

Definition 2.4. A reaction network R augmented with a kinetics κ is called a reaction system,
and is denoted {R, κ}.

Definition 2.5. Given aP reaction system {R, κ}, the induced kinetic function, K(·; R) : RS≥ → RS ,
is given by K(x; R) = 0 0
r∈R κr (x)(c − c) where r = c → c . The induced system of kinetic
equations, describing the time evolution of species concentrations x ∈ RS≥ , takes the form of a
system of autonomous first-order ordinary differential equations (ODEs), and is given by

dx
= K(x; R). (2.1)
dt
Note that the kinetic function uniquely defines the system of kinetic equations, and vice-versa.
In this paper, the species concentrations satisfying equation (2.1) are required to be finite, i.e.
xs < ∞, for s ∈ S, and for t ≥ 0, except possibly for initial conditions located on a finite number
of (S − z)-dimensional subspaces of RS≥ , z ≥ 1.

Definition 2.6. Kinetics κ is called the mass action kinetics if κr (x) = kr xc , for r = (c → c0 ) ∈ R,
where kr > 0 is the rate constant of reaction r, and xc = s∈S xcss , with 00 = 1. A reaction system
Q
with the mass action kinetics is denoted {R, k}, and the corresponding kinetic function is denoted
K(x; k) ≡ K(x; R) = r∈R kr (c0 − c)xc , where k ∈ RR
P
>.

A review of the mass action kinetics can be found in [32]. In this paper, most of the results are
stated with kinetics fixed to the mass action kinetics. This is not restrictive, as an arbitrary analytic
function can always be reduced to a polynomial one [16].

Example 2.1. Consider the following reaction network (consisting of one reaction) under the mass
action kinetics:
k
1
r1 : s1 + s2 −→ 2s2 , (2.2)

3
so that S = {s1 , s2 }, C = {s1 + s2 , 2s2 }, R = {s1 + s2 → 2s2 } and k = {k1 }. Concentration x ∈ RS≥
has two components. To simplify our notation, we write x1 = xs1 , x2 = xs2 , and K1 (x; k) =
Ks1 (x; R), K2 (x; k) = Ks2 (x; R). Then the induced system of kinetic equations is given by

dx1
= K1 (x; k) = −k1 x1 x2 , (2.3)
dt
dx2
= K2 (x; k) = k1 x1 x2 . (2.4)
dt

2.2 Kinetic and nonkinetic functions

In this subsection, nonkinetic functions are defined, and further notation for kinetic and nonkinetic
functions taking the mass action form is presented.

Definition 2.7. Let f : RS≥ → RS be given by fs (x) = r∈R fsr (x), where fsr (x) ∈ R, for x ∈ RS≥ ,
P

s ∈ S and r ∈ R. If ∃s ∈ S, ∃r ∈ R and ∃x ∈ RS≥ such that s ∈ suppc (x) and fsr (x) < 0, then
fsr (x) is called a cross-negative term, and function f (x) and ODE system dx/dt = f (x) are said
to be nonkinetic.

An interpretation of a cross-negative term is that the process corresponding to such a term would
consume at least one reactant even when its concentration is zero, so that it cannot be represented
as kinetic reactions.
Kinetic and nonkinetic functions taking the mass action kinetics form are central to this paper.
The related notation is introduced in the following definition.

Definition 2.8. Let P(·; k) : RS → RS , k ∈ RR , be a polynomial function with polynomial degree

deg(P(x; k)) ≤ m, m ∈ N. Then, the set of functions P(x; k) is denoted by Pm (RS ; RS ). If P(x; k)
is a kinetic function, it is denoted by K(x; k), k ∈ RR > , and the set of such functions is denoted by
PKm (R S ; RS ). If P(x; k) is a nonkinetic function, it is denoted by N (x; k), k ∈ RR , and the set of
≥
such functions with domain RS is denoted by PN S S S N S
m (R ; R ), while with domain R≥ by Pm (R≥ ; R ).
S

Note that a system {R, k}, corresponding to N (x; k) in Definition 2.8, has a well-defined reaction
network R (for r = c → c0 , r ∈ R, we restrict c, c0 to positive integers), but an ill-defined kinetics
taking the mass action form (we allow set k to have elements that are negative). Thus, set k
corresponding to N (x; k) cannot be interpreted as a set of reaction rate constants, as opposed to
set k corresponding to K(x; k) (see also Example 2.2). Note also that Pm (RS≥ ; RS ) = PK S S
m (R≥ ; R )∪
PN S S K S S N S S
m (R≥ ; R ), with Pm (R≥ ; R ) ∩ Pm (R≥ ; R ) = ∅.

2.3 Properties of kinetic functions

From Definition 2.3 it follows that a kinetic function K(x; R) has a structural property: cross-
negative terms are absent. In this subsection, further properties of K(x; R) are defined: nonnega-
tivity (absence of cross-negative effect), and a structural property called x-factorability.

Definition 2.9. Let f : RS≥ → RS be given by fs (x) = r∈R fsr (x), where fsr (x) ∈ R, for x ∈ RS≥ ,
P

s ∈ S and r ∈ R. If ∀s ∈ S, ∀x ∈ RS≥ , s ∈ suppc (x) ⇒ fs (x) ≥ 0, then f (x) and dx/dt = f (x)
are said to be nonnegative. Otherwise, f (x) and dx/dt = f (x) are said to be negative, and a
cross-negative effect is said to exists ∀x ∈ RS≥ for which ∃s ∈ S such that s ∈ suppc (x) and
fs (x) < 0.

4
Note that the absence of cross-negative terms implies nonnegativity, but the converse is not neces-
sarily true [2, 7], i.e. an ODE system may have cross-negative terms, without having a cross-negative
effect, as we will show in Example 2.2.
Cross-negative terms play an important role in mathematical constructions of reaction systems,
in the context of chaos in kinetic equations, and pattern formation via Turing instabilities [3, 5].
In the context of oscillations, as a generalization of the result in [20], one can prove that in two-
dimensional reaction systems with mass action form and with at most bimolecular reactions, the
nonexistence of a cross-negative effect in the ODEs is a sufficient condition for nonexistence of limit
cycles (see A).

Example 2.2. Consider the following ODE system with polynomial RHS:
dx1
= P1 (x; k) = 1 + x21 + 2kx2 + x22 , (2.5)
dt
dx2
= P2 (x; k) = 1, (2.6)
dt
where P(x; k) ∈ P2 (RS ; RS ), S = 2, k ∈ R and x = {x1 , x2 }. Considering x1 = 0 and x2 > 0, it
follows that P1 ({0, x2 }; k) = 1 + 2kx2 + x22 . Then:

(i) If k ≥ 0, then (2.5)–(2.6) contains no cross-negative terms, and so it is kinetic: P(x; k) ∈

PK S S
2 (R≥ ; R ).

(ii) If k < 0, then (2.5)–(2.6) contains one cross-negative term, 2kx2 , and so it is nonkinetic:
P(x; k) ∈ PN S S
2 (R ; R ).

(a) If −1 ≤ k < 0, then (2.5)–(2.6) contains no cross-negative effect, and so it is nonnegative.

(b) If k < −1,
√ then (2.5)–(2.6)
√ contains

a cross-negative effect for x = {0, x2 }, where x2 ∈
− k − k 2 − 1, −k + k 2 − 1 , and so it is negative.

System (2.5)–(2.6) induces a reaction system only in case (i). In particular, nonnegative ODE
system (2.5)–(2.6) with P(x; k) ∈ PN S S
2 (R≥ ; R ) in part (ii)(a) does not induce a reaction system
(although, given a nonnegative initial condition, the solution of (2.5)–(2.6) is nonnegative for all
forward times).

Definition 2.10. Let f : RS≥ → RS be given by fs (x) = r∈R fsr (x), where fsr (x) ∈ R, for x ∈ RS≥ ,
P
s ∈ S and r ∈ R. Then term fsr (x) is said to be xs -factorable P if fsr (x) = xs psr (x), where psr (x)
is a polynomial function of x. If ∃s ∈ S, such that fs (x) = xs r∈R psr (x), then f (x) P and ODE
system dx/dt = f (x) are said to be xs -factorable. If ∀s ∈ S it is true that fs (x) = xs r∈R psr (x),
then f (x) and ODE system dx/dt = f (x) are said to be x-factorable.

Example 2.3. System (2.3)–(2.4) is x-factorable, since K1 (x; k) = x1 (−k1 x2 ) and K2 (x; k) =
x2 (k1 x1 ).

X-factorable ODE systems are a subset of kinetic equations under the mass action kinetics [10]
(see also Section 3.2.2).

3 Inverse problem for reaction systems

In some applications, we are interested in the direct problem: we are given a reaction network with
kinetics, i.e. a reaction system {R, κ}, and we then analyse the induced system of kinetic equations

5
(2.1) in order to determine properties of the reaction system. For example, an output of a direct
problem might consist of verifying that the kinetic equations undergo a bifurcation. In this paper,
we are interested in the inverse problem: we are given a property of an unknown reaction system,
and we would then like to construct a reaction system displaying the property. The inverse problem
framework described in this section is inspired by [2, 4].
The first step in the inverse problem is, given a quantity that depends on a kinetic function,
to find a compatible kinetic function K(x; R), while the second step is then to find a reaction
system {R, κ} induced by the kinetic function. The second step is discussed in more detail in
Section 3.1, while the first step in Section 3.2. The constructions of a reaction system {R, κ}
often involve constraints defining simplicity of the system (e.g. see [22]), and the simplicity can
be related to the kinetic equations (structure and dimension of the equations, and/or the phase
space), and/or to reaction networks (cardinality, conservability, reversibility, deficiency). How the
simplicity constraints are prioritized depends on the application area, with simplicity of the kinetic
equations being more important for mathematical analysis, while simplicity of the reaction networks
for synthetic biology.

3.1 The canonical reaction network

Let us assume that we are able to construct an ODE system of the form (2.1) where its RHS is a
kinetic function, K(x; R), and the system has the property required by the inverse problem. Then,
one can always find a reaction system induced by the kinetic function [3, 4]. While, for a fixed
kinetics, a reaction network induces kinetic function uniquely by definition (see Definition 2.5), the
converse is not true – the inverse mapping of the kinetic function to the reaction networks is not
unique – a fact known as the fundamental dogma of chemical kinetics [8, 9, 3]. For example, in [9],
for a fixed kinetic function and a fixed set of complexes (C fixed), mixed integer programming is
used for numerical computation of different induced reaction networks with varying properties. On
the other hand, a constructive proof that every kinetic function induces a reaction system is given
in [3], where C is generally not fixed (product complexes may be created), but the construction can
be performed analytically, and it uniquely defines an induced reaction system for a given kinetic
function. The procedure is used in this paper, so it is now defined.
Definition 3.1. Let κ : RS≥ → RR ≥ be a kinetics. Consider the kinetic function given by K(x; R) =
S S
P
r∈R dr κr (x), where x ∈ R≥ and dr ∈ R . Let us map K(x; R) to a reaction system {RK−1 , κK−1 }
with complexes and kinetics given by:
(i) Reactant complexes, cr , are assumed to be uniquely obtainable from κr (x) for r ∈ R, which
is true in the case of the mass action kinetics.
(ii) Reaction cr → c0rs is then constructed for each r ∈ R and s ∈ S, where new product complexes
are given by c0rs = cr + sign(drs )s, with sign(·) being the sign function.
(iii) The new kinetics is then defined as κK−1 r0 (x) ≡ |drs |κr (x), for r ∈ R, s ∈ S, where r0 ∈ RK−1 .
The induced reaction system {RK−1 , κK−1 } is called the canonical reaction system, with RK−1 being
the canonical reaction network.
Note that the procedure in Definition 3.1 creates a reaction for each term in each kinetic equation.
Note also that each reaction leads to a change in copy number of precisely one chemical species,
and the change in the copy number is equal to one. Thus, the canonical reaction networks are
simple in the sense that they can be constructed from a kinetic function in straightforward way,
while they generally do not contain minimal number of reactions.

6
Example 3.1. The canonical reaction network for system (2.3)–(2.4) is given by
k
1
r1 : s1 + s2 −→ s2 ,
(3.1)
k
2
r2 : s1 + s2 −→ s1 + 2s2 ,
so that S = {s1 , s2 }, C = {s1 + s2 , s2 , s1 + 2s2 }, RK−1 = {s1 + s2 → s2 , s1 + s2 → s1 + 2s2 } and
kK−1 = {k1 , k2 }, k2 = k1 . Note that the canonical reaction network (3.1) contains more reactions
than the original network (2.2).

3.2 Kinetic transformations

Firstly, mapping a solution-dependent quantity to the RHS of an ODE system is much more likely to
result in nonkinetic functions, N (x; R), on the RHS (see Definition 2.7) [2]. However, only kinetic
functions induce reaction networks, as exemplified in Example 2.2. Secondly, even if mapping a
solution-dependent quantity results in a kinetic function, it may be necessary to modify the function
in order to satisfy given constraints, and this may change the kinetic function into a nonkinetic
function. For these two reasons, it is beneficial to study mappings that can transform arbitrary
functions into kinetic functions. This motivates the following definition, for the case of mass action
kinetics, that relies on the notation introduced in Definition 2.8.
Definition 3.2. Let P(x; k) ∈ Pm (RS ; RS ), k ∈ RR , i.e. P(x; k) is a polynomial function.
Consider the corresponing ODE system in the formal sum notation
dx
= P(x; k), (3.2)
dt
where x ≡ x(t) ∈ RS . Then, a transformation Ψ is called a kinetic transformation if the following
conditions are satisfied:
(i) Ψ : Pm (RS ; RS ) → PK S̄ S̄
m̄ (R≥ ; R ), m̄ ≥ m, S̄ ≥ S, maps the polynomial function P(x; k) into
a kinetic function K(x̄; k̄) ≡ (Ψ(P))(x̄; k̄) for x̄ ∈ RS̄≥ and k̄ ∈ RR̄
>.

(ii) Let x∗ be a fixed point of (3.2) that is mapped by Ψ to fixed point x̄∗ ∈ RS̄≥ of the system
of kinetic equations (2.1) with K(x̄; k̄) on its RHS. Let also the eigenvalues of the Jacobian
matrix of P(x; k), J(x∗ ; k), denoted by λn , n = 1, 2, . . . , S, be mapped to the eigenvalues of
Jacobian of K(x̄; k̄), JΨ (x̄∗ ; k̄), which are denoted by λ̄n , n = 1, 2, . . . , S. Then, for every
such fixed point x∗ it must be true that sign(λn ) = sign(λ̄n ), n = 1, 2, . . . , S, and, if there are
any additional eigenvalues λ̄n , n = S, S + 1, . . . , S̄, they must satisfy sign(λ̄n ) < 0.
If any of the condition (i)–(ii) is not true, Ψ is called a nonkinetic transformation.
Put more simply, given an input polynomial function, a kinetic transformation must (i) map the
input polynomial function into an output kinetic function, and (ii) the output function must be lo-
cally topologically equivalent to the input function in the neighbourhood of the corresponding fixed
points, and the dynamics along any additional dimensions of the output function (corresponding to
the additional species) must asymptotically tend to the corresponding fixed point. Let us note that
the output function is defined only in the nonnegative orthant, so that the topological equivalence
must hold only near the fixed points of the input function that are mapped to the nonnegative
orthant under kinetic transformations.
One may wish to impose a set of constraints on an output function, such as requiring that a
predefined region of interest in the phase space of the input function is mapped to the positive
orthant of the corresponding output function. A subset of constraints is now defined.

7
Definition 3.3. Let P(x; k) ∈ Pm (RS ; RS ), k ∈ RR . Let also φj : RR → R be a continuous
function, mapping set k into φj (k) ∈ R, j = 1, 2, . . . , J. Then, set Φ ≡ {φj (k) ≥ 0 : j = 1, 2, . . . , J}
is called a set of constraints.
There are two sets of kinetic transformations. The first, and the preferred, set of possible kinetic
transformations are affine transformations, which are discussed in Section 3.2.1. Affine transfor-
mations may be used, not only as possible kinetic transformations, but also to satisfy a set of
constraints. The second set, necessarily used when affine transformations fail, are nonlinear trans-
formations that replace cross-negative terms, with x-factorable terms (see Definition 2.10), without
introducing new cross-negative terms, and two such transformations are discussed in Sections 3.2.2
and 3.2.3. In choosing a nonlinear transformation, one generally chooses between obtaining, on the
one hand, lower-dimensional kinetic functions with higher-degree of nonlinearity (i.e. lower S̄/S
and higher m̄/m in Definition 3.2(i)) and/or higher numbers of the nonlinear terms, and, on the
other hand, higher-dimensional kinetic functions with lower degree of nonlinearity (i.e. higher S̄/S
and lower m̄/m) and/or lower numbers of the nonlinear terms.
Before describing the transformations in a greater detail, the usual vector notation is introduced
and related to the formal sum notation from Section 2. The vector notation is used when ODE
systems are considered in matrix form, while the formal sum notation is used when ODE systems
are considered component-wise.
Notation. Let |S| = S, |C| = C and |R| = R, and suppose S, C and R are each given a fixed
ordering with indices being n = 1, 2, . . . , S, i = 1, 2, . . . , C, and l = 1, 2, . . . , R, respectively, i.e. one
can identify the ordered components of formal sums with components of Euclidean vectors. Let
also the indices sn be denoted by n, n = 1, 2, . . . , S, for brevity. Then, the kinetic equations under
the mass action kinetics in the formal sum notation are given by (2.1). In this section, we start
with equations which have more general polynomial, and not necessarily kinetic, functions on the
RHS, i.e. the ODE system is written in the formal sum notation as (3.2), while in the usual vector
notation by
dx
= P(x; k), (3.3)
dt
where P(x; k) ∈ Pm (RS ; RS ), x ∈ RS≥ , and k ∈ RR .

3.2.1 Affine transformation

Definition 3.4. Consider applying an arbitrary nonsingular matrix A ∈ RS×S on equation (3.3),
resulting in:
dx̄
= A P(A−1 x̄; k̄) ≡ (ΨA P)(x̄; k̄), (3.4)
dt
where x̄ = Ax, and k̄ is a vector of new rate constants obtained from k by rewriting the polynomial
on the RHS of (3.4) into the mass action form. Then ΨA : Pm (RS ; RS ) → Pm (RS ; RS ), mapping
P(x; k) to (ΨA P)(x̄; k̄), is called a centroaffine transformation. If A is an orthogonal matrix, then
ΨA is called an orthogonal transformation.
Definition 3.5. Consider substituting x̄ = x + T in equation (3.3), where T ∈ RS , which results
in:
dx̄
= P(x̄ − T ; k̄) ≡ (ΨT P)(x̄; k̄), (3.5)
dt

8
where k̄ is a vector of the new rate constants obtained from k by rewriting the polynomial on the
RHS of (3.5) into the mass action form. Then ΨT : Pm (RS ; RS ) → Pm (RS ; RS ), mapping P(x; k)
to (ΨT P)(x̄; k̄), is called a translation transformation.

A composition of a translation and a centroaffine transformation, ΨA,T = ΨA ◦ ΨT , i.e. an affine

transformation, may be used as a possible kinetic transformation (see Definition 3.2). Let us
note that condition (ii) in Definition 3.2 is necessarily satisfied for all affine transformation, i.e.
affine transformations preserve the topology of the phase space, as well as the polynomial degree
of the functions being mapped [6]. For these reasons, affine transformations are preferred over
the alternative nonlinear transformations, discussed in the next two sections. However, affine
transformations do not necessarily satisfy condition (i) in Definition 3.2, so that they are generally
nonkinetic transformations. However, despite being generally nonkinetic, affine transformations
map sets k into new sets k̄ (see equations (3.4) and (3.5)), so that they may be used for satisfying
a given set of constraints imposed on the output function (see Definition 3.3). This motivates the
following definition.

Definition 3.6. Let P(x; k) ∈ Pm (RS ; RS ). If it is not possible that simultaneously (ΨA ◦
ΨT P)(x̄; k̄) is a kinetic function, and that a given set of constraints Φ ≡ {φj (k̄) ≥ 0 : j =
1, 2, . . . , J} is satisfied, for all A ∈ RS×S and for all T ∈ RS , then it is said that P(x; k) and the
corresponding equation (3.2) are affinely nonkinetic, given the constraints. Otherwise, they are
said to be affinely kinetic, given the constraints.

If the set of constraints in Definition 3.3 is empty, affinely nonkinetic functions are called essentially
nonkinetic, while those that are affinely kinetic are called removably nonkinetic. Such labels em-
phasize that, if a function is essentially nonkinetic, a kinetic function that is globally topologically
equivalent cannot be obtained, while if a function is removably nonkinetic, a globally topologically
equivalent kinetic function can be obtained.
Explicit sufficient conditions for a polynomial function P(x; k) to be affinely kinetic, or nonk-
inetic, are generally difficult to obtain. Even in the simpler case P(x; k) ∈ P2 (R2 ; R2 ), such con-
ditions are complicated, and cannot be easily generalized for higher-dimensional systems and/or
systems with higher degree of nonlinearily [6]. In [5], based on the polar and spectral decomposition
theorems, it has been argued that if no orthogonal transformation is kinetic, then no centroaffine
transformation is kinetic. The result is reproduced in this paper using the more concise singular
value decomposition theorem, and is generalized to the case when the set of constraints is nonempty.
Loosely speaking, the theorem states that “orthogonally nonkinetic” functions are affinely nonki-
netic as well, given certain constraints.

Theorem 3.1. If P(x; k) ∈ Pm (RS ; RS ) is nonkinetic under ΨQ ◦ΨT , given a set of constraints Φ,
for all orthogonal matrices Q ∈ RS×S and for all T ∈ RS , then P(x; k) is also affinely nonkinetic,
given Φ, provided the following condition holds: sign(φj (k)) = sign(φj (k̄)), j = 1, 2, . . . , J, for all
diagonal and positive definite matrices Λ ∈ RS×S , with ΨΛ P = (ΨΛ P)(x̄; k̄).

Proof. By the singular value decomposition theorem, nonsingular matrices A ∈ RS×S can be written
as A = Q1 ΛQ2 , where Q1 , Q2 ∈ RS×S are orthogonal, and Λ ∈ RS×S diagonal and positive definite.
Cross-negative terms are invariant under transformation ΨΛ for all Λ [5]. If Φ from Definition 3.3
is such that functions sign(φj (k)), j = 1, 2, . . . , J, are invariant under all positive definite diagonal
matrices Λ ∈ RS×S , the statement of the theorem follows.

9
3.2.2 X-factorable transformation
Definition 3.7. Consider multipling the RHS of equation (3.3) by a diagonal matrix X (x) =
diag(x1 , x2 , . . . , xS ), resulting in

dx
= X (x)P(x; k) ≡ (ΨX P)(x; k). (3.6)
dt
Then ΨX : Pm (RS ; RS ) → Pm+1 (RS ; RS ), mapping P(x; k) to (ΨX P)(x; k), is called an x-
factorable transformation. If X is diagonal and its nonzero elements are
(
xs , if s ∈ S 0 ,
Xss =
1, if s ∈ S \ S 0 ,

where S 0 ⊂ S, S 0 6= ∅, then the transformation is denoted ΨXS 0 , and is said to be xS 0 -factorable.

When X ∈ R2 is x1 -factorable, i.e. X (x1 ) = diag(x1 , 1), we write ΨX1 ≡ ΨX{1} .

Theorem 3.2. (ΨX P)(x; k) from Defnition 3.7 is a kinetic function, i.e. (ΨX P)(x; k) ∈ PK S S
m+1 (R≥ ; R ).

Proof. See [10].

Functions P(x; k) and (ΨX P)(x; k) are not necessarily topologically equivalent due to two over-
lapping artefacts that ΨX can produce, so that ΨX is generally a nonkinetic transformation. Firstly,
the fixed points of the former system can change the type and/or stability under ΨX , and, secondly,
the latter system has an additional finite number of boundary fixed points. The following theorem
specifies the details of the artefacts for two-dimensional systems.

Theorem 3.3. Let us consider the ODE system (3.3) in two dimensions with RHS P(x; k) =
(P1 (x; k), P2 (x; k))> . The following statements are true for all the fixed points x∗ of the two-
dimensional system (3.3) in R2> under ΨXS 0 , S 0 ⊆ S, S 0 6= ∅:

(i) All the saddle fixed points are unconditionally invariant, i.e. saddle points of (3.3) correspond
to saddle points of (3.6).

(ii) A sufficient condition for stability of a fixed point x∗ to be invariant is:

∂P1 (x; k) ∂P2 (x; k)

|x=x∗ |x=x∗ ≥ 0.
∂x1 ∂x2

(iii) A sufficient condition for the type of a fixed point x∗ to be invariant is:

∂P1 (x; k) ∂P2 (x; k)

|x=x∗ |x=x∗ ≥ 0.
∂x2 ∂x1

Assume that the ODE system (3.3) does not have fixed points on the axes of the phase space.
Nevertheless, the two-dimensional system (3.6) can have additional fixed points on the axes of the
phase space, called boundary fixed points, denoted x∗b ∈ R2≥ . The boundary fixed points can be
either nodes or saddles, and the following statements are true:

(iv) If system (3.6) is x-factorable, then the origin is a fixed point, x∗b = 0, with eigenvalues
λi = Pi (x∗b ; k) 6= 0, i = 1, 2, and the corresponding eigenvectors along the phase space axes.

10
 (v) For x∗b,i = 0, x∗b,j 6= 0, x∗b ∈ R2≥ , i, j = 1, 2, i 6= j, a boundary fixed point is a node if and only
if
∂Pj (x; k)
Pi (x∗b ; k) |x=x∗b > 0,
∂xj
with the node being stable if Pi (x∗b ; k) < 0, and unstable if Pi (x∗b ; k) > 0, i, j = 1, 2, i 6= j.
Otherwise, the fixed point is a saddle.

Proof. Without loss of generality, we consider two forms of the system (3.6) with S = 2:

dx1
= x1 P1 (x; k), (3.7)
dt
dx2
= xp2 P2 (x; k), (3.8)
dt
where p ∈ {0, 1}, so that system (3.7)–(3.8) is x-factorable for p = 1, but only x1 -factorable for
p = 0. The results derived for an x1 -factorable system hold when the system is x2 -factorable, if
the indices are swapped. By writing Pi (x; k) = Pi , i = 1, 2, the Jacobian of (3.7)–(3.8), JX , is for
p ∈ {0, 1} given by !
P1 + x1 ∂P 1
∂x1 x 1
∂P1
∂x2
JX (x) = .
xp2 ∂P2
∂x1 pP2 + xp2 ∂P2
∂x2

First, consider how fixed points of P(x; k) are affected by transformation ΨXS 0 . Denoting the
Jacobian of two-dimensional system (3.3) by J, and assuming the fixed points are not on the axes
of the phase space (i.e. x∗ ∈ R2> ), the Jacobians evaluated at x∗ are given by:
! !
∂P1 ∂P1
∗ ∂x1 ∂x2 ∗ x1 ∂P 1
∂x1 x1 ∂P 1
∂x2
J(x ) = , JX (x ) = .
∂P2
∂x1
∂P2
∂x2
x=x∗ xp2 ∂P2
∂x1 xp2 ∂P2
∂x2
x=x∗

Comparing the trace, determinant and discriminant of J(x∗ ) and JX (x∗ ), we deduce (i)–(iii).
To prove (iv)–(v), we evaluate JX at the boundary fixed points of the form x∗b = (0, x∗b,2 ) to get
 
P1 0
JX (x∗b ) = . (3.9)
xp2 ∂P2
∂x1 pP2 + xp2 ∂P2
∂x2 x=x∗b

If p = 1, then one of the boundary fixed points is x∗b = 0, and the Jacobian becomes a diagonal
matrix, so that condition (iv) holds. If x∗b,2 6= 0, then P2 (0, x∗b,2 ; k) = 0 in (3.9), and comparing the
trace, determinant and discriminant of J(x∗ ) and JX (x∗b ), we deduce (v).

Theorem 3.3 can be used to find conditions that an x-factorable transformation given by ΨX :
Pm (R2 ; R2 ) → Pm+1 (R2 ; R2 ) is a kinetic transformation. While conditions (ii)–(iii) in Theorem 3.3
may be violated when ΨX is used, so that ΨX is a nonkinetic transformation, a composition
of an affine transformation and an x-factorable transformation, i.e. ΨX ,A,T = ΨX ◦ ΨA ◦ ΨT ,
may be kinetic. Furthermore, such a composite transformation may also be used to control the
boundary fixed points introduced by ΨX . Finding an appropriate A and T to ensure the topological
equivalence near the fixed points typically means that the region of interest in the phase space has to
be positioned at a sufficient distance from the axes. However, since the introduced boundary fixed
points may be saddles, this implies that the phase curves may be significantly globally changed,
regardless of how far away from the axes they are. The most desirable outcome of controlling the
boundary fixed points is to eliminate them, or shift them outside of the nonnegative orthant. The

11
former can be attempted by ensuring that the nullclines of the original ODE system (3.3) do not
intersect the axes of the phase space, while the latter by using the Routh-Hurwitz theorem [34].
An alternative transformation, which is always kinetic, that also does not change the dimension
of an ODE system is the time-parametrization transformation [14]. However, while ΨX increases
the polynomial degree by one, and introduces only a finite number of boundary fixed points which
are given as solutions of suitable polynomials, the time-parameterization transformation generally
increases the nonlinearity degree more than ΨX , and introduces infinitely many boundary fixed
points.

3.2.3 The quasi-steady state transformation

The quasi-steady state assumption (QSSA) is a popular constructive method for reducing dimension
of ODE systems by assuming that, at a given time-scale, some of the species reach a quasi-steady
state, so that they can be described by algebraic, rather than differential equations. The QSSA is
based on Tikhonov’s theorem [11, 12] that specifies conditions ensuring that the solutions of the
reduced system are asymptotically equivalent to the solutions of the original system. The original
system is referred to as the total system, and it consists of the reduced subsystem, referred to as the
degenerate system, and the remaining subsystem, called the adjointed system, so that the QSSA
consists of replacing the total system with the degenerate one, by eliminating the adjointed system.
Korzukhin’s theorem [11, 12] is an existence result ensuring that, given any polynomial degenerate
system, there exists a corresponding total system that is kinetic.
Thus, Tikhonov’s theorem can be seen as a constructive direct asymptotic dimension reduction
procedure, while Korzukhin’s theorem as an inverse asymptotic dimension expansion existence
result. Korzukhin’s theorem has an important implication that an application of the QSSA can
result in a degenerate system that is structurally different than the corresponding total system. In
this paper, the QSSA is assumed to necessarily be compatible with Tikhonov’s theorem. If this
is not the case, then it has been demonstrated in [14, 15] that application of a QSSA can create
dynamical artefacts, i.e. it can result in degenerate systems, not only structurally different, but also
dynamically different from the total systems. The artefacts commonly occur due to the asymptotic
parameters in Tikhonov’s theorem not being sufficiently small. For example, it has been shown
that exotic phenomena such as multistability and oscillations can exist in a degenerate system,
while not existing in the corresponding total system [14, 15].
Using Korzukhin’s and Tikhonov’s theorems, a family of kinetic total systems for an arbitrary
nonkinetic polynomial degenerate system can be constructed, as is now shown. For simplicity, we
denote xc = s∈S xcss , for any x ∈ RS and c ∈ NS .
Q

Definition 3.8. Consider equation (3.2), and assume that the reaction set is partitioned, R =
R1 ∪ R2 , R1 ∩ R2 = ∅, so that (3.2), together with the initial conditions, can be written as

dxs X X
= asr xαsr − bsr xβsr , for s ∈ S, (3.10)
dt
r∈R1 r∈R2
xs (t0 ) = x0s , x0s ≥ 0, (3.11)

where x ∈ RS , αsr ∈ NS , βsr ∈ NS , αsr 6= βsr , asr ∈ R≥ and bsr ∈ R≥ for all s ∈ S and r ∈ R.
Assume further that the species set is partitioned, S = S1 ∪ S2 , S1 ∩ S2 = ∅, so that equations for
species s ∈ S1 are kinetic, while those for species s ∈ S2 are nonkinetic. Consider the following

12
total system, consisting of a degenerate system given by
dxs X X
= asr xαsr − bsr xβsr , for s ∈ S1 , (3.12)
dt
r∈R1 r∈R2
dxs X X
= asr xαsr − ωs−1 xs ps (x)ys bsr xβsr , for s ∈ S2 , (3.13)
dt
r∈R1 r∈R2

which satisfies the initial condition (3.11) with x0s > 0 for s ∈ S2 , and an adjointed system given by

dys
µ = ωs − xs ps (x)ys , for s ∈ S2 , (3.14)
dt
ys (t0 ) = ys0 , ys0 ≥ 0, for s ∈ S2 , (3.15)

where µ > 0, ωs > 0 are parameters, and p(x) is a polynomial function of x satisfying p(x) ∈
Pm0 (RS≥ ; RS>2 ) for m0 ∈ N. Then ΨQSSA : Pm (RS ; RS ) → Pm̄ (RS+S
≥
2
; RS+S2 ), mapping the RHS of
differential equations in system (3.10)–(3.11), denoted P(x; k), to the RHS of differential equations
of system (3.12)–(3.15), denoted (ΨQSSA P)({x, y}; k̄), with the constraint that xs > 0 for s ∈ S2 , is
called a quasi-steady state transformation. Here, m̄ ≤ m + m0 + 2, and k̄ is a vector of the new rate
constants obtained from k by rewriting the polynomial (ΨQSSA P)({x, y}; k̄) into the mass action
form.

Theorem 3.4. Solutions of systems (3.10)–(3.11) and (3.12)–(3.15), corresponding to P(x; k) and
(ΨQSSA P)({x, y}; k̄), are asymptotically equivalent in the limit µ → 0, and (ΨQSSA P)({x, y}; k̄)
is a kinetic function.

Proof. Fixed points of system (3.14) satisfy ys∗ = ωs (xs ps (x))−1 . The fixed points are isolated,
and, since (from Definition 3.8) xs > 0 and ps (x) > 0, ∀x ∈ RS≥ , ∀s ∈ S2 , it follows that the fixed
points are globally stable. Thus, the conditions of Tikhonov’s theorem [11] are satisfied by the total
system (3.12)–(3.15). Then, by applying the theorem, i.e. substituting ys∗ into (3.13), one recovers
the corresponding degenerate system given by (3.10)–(3.11). Finally, the total system (3.12)–(3.15)
is kinetic, as can be verified by using Definition 2.7.

Corollary 3.1. The quasi-steady state transformation ΨQSSA , defined in Definition 3.8, is a kinetic
transformation in the limit µ → 0.

An alternative transformation, for which condition (i) in Definition 3.2 is also always satisfied, and
that also expands the dimension of an ODE system, is an incomplete Carleman embedding [18, 17].
However, condition (ii) in Definition 3.2 is satisfied for the incomplete Carleman embedding only
provided initial conditions for the adjointed system are appropriately constrained, and, further-
more, the transformation generally results in an adjointed system with a higher nonlinearity de-
gree when compared to ΨQSSA . In fact, Theorem 3.4 can be seen as an asymptotic alternative
to the incomplete Carleman embedding, i.e. instead of requiring adjointed variables to satisfy
ys (t) = ωs x−1 −1 −1 −1
s (t) ps (x(t)), one requires them to satify limµ→0 ys (t) = ωs xs (t) ps (x(t)). The
theorem can also be seen as an extension of using the QSSA to represent reactions of more than
two molecules as a limiting case of bimolecular reactions [19] to the case of using the QSSA to
represent cross-negative terms as a limiting case of kinetic ones.

13
4 Construction of reaction systems undergoing a supercritical ho-
moclinic bifurcation
In this section, a brief review of a general bifurcation theory, and a more specific homoclinic
bifurcation, is presented. This is followed by applying the framework developed in Section 3 to
construct specific reaction systems displaying the homoclinic bifurcation.
Variations of parameters in a parameter dependent ODE system may change topology of the
phase space, e.g. a change may occur in the number of invariant sets or their stability, shape of
their region of attraction or their relative position. At values of the bifurcation parameters at
which the system becomes topologically nonequivalent it is said that a bifurcation occurs, and the
bifurcation is characterized by two sets of conditions: bifurcation conditions defining the type of
bifurcation, and genericity conditions ensuring that the system is generic, i.e. can be simplified
near the bifurcation to a normal form [24]. If it is sufficient to analyse a small neighbourhood of
an invariant set to detect a bifurcation, the bifurcation is said to be local. Otherwise, it is called
global, and the analysis becomes more challenging. Bifurcations are common in kinetic equations,
where, in the case of the mass action kinetics, the rate constants play the role of bifurcation
parameters [25, 26, 27, 28, 29]. In this paper, focus is placed on a global bifurcation giving rise to
stable oscillations, called a supercritical homoclinic bifurcation [24, 21, 26].

Definition 4.1. Suppose x∗ is a fixed point of system (3.3). An orbit γ ∗ starting at a point x is
called homoclinic to the fixed point x∗ if its α-limiting and ω-limiting sets are both equal to x∗ .

Put more simply, a homoclinic orbit connects a fixed point to itself. An example of a homoclinic
orbit to a saddle fixed point can be seen in Figure 1(b) on page 16, where the homoclinic orbit is
shown as the purple loop, while the saddle as the blue dot at the origin.
If a homoclinic orbit to a hyperbolic fixed point is present in an ODE system, then the sys-
tem is structurally unstable, i.e. small perturbations to the equations can destroy the homoclinic
orbit and change the structure of the phase space, so that a bifurcation can occur. For two-
dimensional systems, the bifurcation and genericity conditions are completely specified by the
Andronov-Leontovich theorem [24] given in B. In summary, the theorem demands that the sum
of the eigenvalues corresponding to the saddle at the bifurcation point, called the saddle quan-
tity, must be nonzero (nondegeneracy condition), and that the so-called Melnikov integral at the
bifurcation point evaluated along the homoclinic orbit must be nonzero (transversality condition).

4.1 The inverse problem formulation

Construction of reaction systems with prescribed properties is an inverse problem which we will
solve by applying kinetic transformations described in Section 3. Our goal is to find a reaction
system with the mass action kinetics (see Definition 2.6) such that the kinetic equations satisfy
assumptions of Andronov-Leontovich theorem in B, i.e. they must contain a homoclinic orbit
defined on a two-dimensional manifold in the nonnegative orthant, and undergo the homoclinic
bifurcation in a generic way. The output of this inverse problem will be a canonical reaction
network which corresponds to the constructed ODE system. Thus the inverse problem is solved in
three steps given in Algorithm 1. The first step is solved by using results of Sandstede [21] which
leads to a set of polynomial functions satisfying the first three conditions of Andronov-Leontovich
theorem in B. An additional transformation is then performed ensuring that the final condition
of Andronov-Leontovich theorem is satisfied. In this paper, nonlinear kinetic transformations are
applied on the resulting polynomial function (using Step (2), case (b), in Algorithm 1).

14
 1. Construction of a polynomial function P(x; k):
Find an ODE system (3.2) which satisfies the assumptions of Andronov-Leontovich theorem
in B.

2. Construction of a kinetic function K(x̄; k̄):

Find a transformation so that the following conditions are satisfied:

(i) The transformation is kinetic (see Definition 3.2), mapping the polynomial function
P(x; k) into a kinetic function K(x̄; k̄) ≡ (ΨP)(x̄; k̄).
(ii) The set of constraints (see Definition 3.3) ensuring that the homoclinic orbit is in R2≥
are satisfied for K(x̄; k̄).

To determine the choice of Ψ, if possible, use Theorem 3.1 to deduce that P(x; k) is affinely
nonkinetic (see Definition 3.6), given the constraints.

(a) If P(x; k) is not affinely nonkinetic, attempt to find an affine transformation Ψ = ΨA

such that (i)–(ii) are satisfied.
(b) If P(x; k) is affinely nonkinetic, or if application of Theorem 3.1 is computationally too
complicated, then choose kinetic transformation Ψ satisfying (i)–(ii) as an appropriate
composition of ΨA , ΨX and ΨQSSA , where ΨX is an x-factorable transformation (see
Section 3.2.2) and ΨQSSA is a quasi-steady state transformation (see Section 3.2.3, in
particular Corollary 3.1).

3. Construction of a reaction network:

Use Definition 3.1 to construct the canonical reaction network RK−1 .

Algorithm 1: Three steps of the solution to the inverse problem of finding reaction systems under-
going a supercritical homoclinic bifurcation.

4.2 Step (1): construction of polynomial function P(x; k)

Definition 4.2. A version of a plane algebraic curve called Tschirnhausen cubic (also known as
Catalan’s trisectrix, and l’Hospital’s cubic) [33] given by:

H(x1 , x2 ) = −x21 + x22 (1 + x2 ) = 0, (4.1)

is referred to as the alpha curve. The part of the curve with x2 < 0 is called the alpha loop, while
the part with x2 > 0 is called the alpha branches, with the branch for x1 < 0 being the negative
alpha branch, and √ for x1 > 0 being the positive alpha branch. Solutions x1 of equation (4.1) are
denoted x± 1 = ±x 2 1 + x2 .

The α curve is shown in Figure 1(a), with x− +

1 plotted as the solid purple curve, and x1 as the
dashed green
√ curve.
√ It can be seen that the curve consists of the tear-shaped alpha loop located in
region [−2 3/9, 2 3/9]×[−1, 0], and the positive and negative alpha branch in the first and second
quadrant, respectively. As is done in [21], the alpha curve is mapped into a system of polynomial
ODEs.

15
 Alpha Curve Phase Plane
2 2

(a) (b)
1 1
x2

x2
0 0

-1 -1

-2 -2
-2 -1 0 1 2 -2 -1 0 1 2
x1 x1

Figure 1: (a) The alpha curve (4.1), with branch x−1 plotted as the solid purple curve, and branch
+
x1 as the dashed green curve. (b) Phase plane diagram of system (4.2)–(4.3) for a = −0.8, with
the stable node, the saddle, and the unstable spiral represented as the green, blue and red dots,
respectively. The alpha curve is shown in purple, while the vector field as gray arrows.

Lemma 4.1. The two-dimensional polynomial ODE system

dx1 3 3
= P1 (x; a) = ax1 + x2 + ax1 x2 + x22 , (4.2)
dt 2 2
dx2 2
= P2 (x; a) = x1 + ax2 + ax2 , (4.3)
dt
contains the alpha curve (4.1) as phase plane orbits, with the alpha loop as a homoclinic orbit to
the fixed point x∗ = 0, provided a2 < 1. If a ∈ (−1, 0), the alpha loop is stable from the inside, and
system (4.2)–(4.3) has three fixed points: a saddle at the origin, an unstable spiral inside the alpha
loop, and a stable node on the positive alpha branch.
Proof. Setting P(x; k) = (P1 (x; k), P2 (x; k)) in system (3.3) to be a polynomial function of x =
(x1 , x2 ) with undetermined coefficients k, and requiring P ·∇H = 0, one obtains system (4.2)–(4.3),
as was done in [21]. As there is only one free parameter, denoted a, we write: P(x; k) = P(x;

a).
System (4.2)–(4.3) has three fixed points: 0, 0 , 2a/9,
−2/3 , and a−1 (1 − a−2 ), −1 + −2 . The



a
condition a2 < 1 ensures that fixed points 2a/9, −2/3 and a−1 (1 − a−2 ), −1 + a−2 are not on

the alpha loop. The Jacobian J = ∇P(x; a) is given by

a + 23 ax2 1 + 32 ax1 + 3x2
 
J= . (4.4)
1 a + 2ax2


Let the determinant and trace of J be denoted by det(J) and tr(J), respectively. Fixed point 0, 0
is a saddle, since det(J) = a2 − 1 < 0. The saddle quantity from Andronov-Leontovich theorem
in B is given by σ0 = λ1 + λ2 = (a − 1) + (a + 1) = 2a, were λ1 and λ2 are the saddle eigenvalues.
The alpha loop is stable from the inside provided σ0 < 0, implying a < 0. It then follows that
−1 −2 −2



2/9a, −2/3 is an unstable spiral, and a (1 − a ), −1 + a a stable node.

16
A representative phase plane diagram of system (4.2)–(4.3) is shown in Figure 1(b). Note that a
part of the positive alpha branch x+ 1 is a heteroclinic orbit connecting the saddle
√ and the node.
The distance between the saddle and the node is given by d(a) = a−3 (1 − a2 ) 1 + a2 , so that
lima→0 d(a) = +∞ and lima→−1 d(a) = 0, i.e. increasing a increases length of the heteroclinic orbit
by sliding the node along x+ 1.
System (4.2)–(4.3) satisfies the first three conditions of Andronov-Leontovich theorem in B. In
order to satisfy the final condition, a set of perturbations must be found that destroy the alpha loop
in a generic way, and this is ensured by the Melnikov condition (B.1). The bifurcation parameter
controlling the existence of the alpha loop is denoted as α ∈ R. Note that P(x; a) perturbed by a
function of the form α∇H(x1 , x2 ) = α(−2x1 , 2x2 + 3x22 ) satisfies the Melnikov condition [21], but
P(x; a) + α∇H(x1 , x2 ) has three terms dependent on α. In the following theorem, a simpler set of
perturbations is found, introducing only one α dependent term in system (4.2)–(4.3).

Theorem 4.1. If a perturbation of the form (αf (x1 ), 0)T , where α ∈ R, is added to the RHS
of system
√ (4.2)–(4.3), and if f (x1 ) is an odd function, f (−x1 ) = −f (x1 ), and f (x1 ) 6= 0, ∀x1 ∈
[−2 3/9, 0), then the perturbed system undergoes a supercritical homoclinic bifurcation in a generic
way as α is varied in the neighbourhood of zero.

Proof. Consider the perturbed version of system (4.2)–(4.3):

dx1 3 3
= P1 (x; a, α) = ax1 + x2 + ax1 x2 + x22 + αf (x1 ), (4.5)
dt 2 2
dx2
= P2 (x; a) = x1 + ax2 + ax22 . (4.6)
dt
Melnikov integral (B.1) for system (4.5)–(4.6) is given by
Z +∞
M (0) = − ϕ(t)f (x1 )P2 (x1 , x2 ; a) dt.
−∞

Using (4.6), we have P2 (x1 , x2 ; a)dt = dx2 . Thus we can express M (0) in terms of x2 as follow:
Z 0 Z +∞
M (0) = − ϕ(t)f (x1 )P2 (x1 , x2 ; a) dt − ϕ(t)f (x1 )P2 (x1 , x2 ; a) dt
−∞ 0
Z 0 √
Z 0 √
ϕ t+ (x2 ) f x2 1 + x2 dx2 − ϕ t− (x2 ) f −x2 1 + x2 ) dx2 ,




=
−1 −1

where t+ (x2 ) (resp. t− (x2 )) is the dependence of time t on x2 along the positive (resp. negative)
alpha branch for the trajectory which is at point (x1 , x2 ) = (0, −1) at time t = 0 (for α = 0). Since
f is odd and ϕ(t± ) > 0, we deduce
Z 0 h
i √
ϕ t+ (x2 ) + ϕ t− (x2 ) f x2 1 + x2 dx2 6= 0.



M (0) =
−1

For further simplicity of (4.5)–(4.6), function f (x1 ) is set to f (x1 ) = x1 in the rest of this paper.

17
4.3 Step (2): construction of kinetic function K(x̄; k̄)
The RHS of system (4.5)–(4.6), P(x; a, α), is a kinetic function. However, the alpha loop, which is
the region of interest, is not located in the nonnegative orthant. In order to position the loop into
the positive orthant, we will apply affine transformations. First, we show that system (4.5)–(4.6)
with the homoclinic orbit in nonnegative orthant is nonkinetic under all translation transformations
for a ∈ (−1, 0), α ∈ R.
Lemma 4.2. Function P(x; a, α), given by the RHS of (4.5)–(4.6), is nonkinetic under all trans-
lation transformations ΨT , for a ∈ (−1, 0) and α ∈ R, given the condition that the homoclinic orbit
is mapped into R2> .
Proof. Let us apply the translation transformation ΨT (see Definition 3.5), T = (T1 , T2 ) ∈ R2 , on
function P(x; a, α), given by the RHS of (4.5)–(4.6), resulting in:
(ΨT P1 )(x̄; k̄) = k̄01 + k̄11 x̄1 + k̄21 x̄2 + k̄12
1 1
x̄1 x̄2 + k̄22 (x̄2 )2 ,
(ΨT P2 )(x̄; k̄) = k̄02 + k̄12 x̄1 + k̄22 x̄2 + k̄22
2
(x̄2 )2 , (4.7)
with x̄ = x + T , and coefficients k̄ = k̄(a, α, T ) given by
1 2
k̄01 = 3(T2 − )(aT1 + T2 ) − 2αT1 ,


2 3
k̄02 = −T1 + aT2 (T2 − 1),
3 2
k̄11 = − a(T2 − ) + α, k̄12 = 1,
2 3
3 1

k̄21 = 1 − (aT1 + 2T2 ), k̄22 = −2a T2 − ,
2 2
1 3 1 3 2
k̄12 = a, k̄22 = , k̄22 = a. (4.8)
2 2
Consider the point x0 = (0, −1), which is on the alpha loop. It is mapped by ΨT to the point
x̄0 = (T1 , −1 + T2 ). Requiring that the alpha loop is mapped to R2> implies that we must have
x̄0 ∈ R2> , so that the following set of constraints (see Definition 3.3) must be satisfied:
Φ = {T1 > 0, T1 > 1}. (4.9)
Using the fact that a ∈ (−1, 0) and the constraints (4.9), it follows that k̄02 from (4.8) is negative,
k̄02 < 0. Thus, (ΨT P)(x̄; k̄) has a cross-negative term, and the statement of the theorem follows.
One can also readily prove that P(x; a, α) is nonkinetic under all affine transformations for |a|  1,
and |α|  1. Thus, in the next two sections, we follow Step (2), case (b), in Algorithm 1, applying
transformations ΨX and ΨQSSA on the kinetic function (ΨT P)(x̄; k̄) given by (4.7). We require
the following conditions to be satisfied:
a ∈ (−1, 0), |α|  1,
( √ )
2 3
Φ = T1 > , T2 > 1 , (4.10)
9
with the set of constraints Φ ensuring that the homoclinic orbit is in R2> . The reason for requiring
|α|  1 is that then the following results are more readily derived, and the condition is sufficient
for studying system (4.7) near the bifurcation point α = 0. A representative phase plane diagram
corresponding to the ODE system with RHS (4.7) is shown in Figure 2(a), with fixed points, the
alpha curve, and the vector field denoted as in Figure 1(b), and with the red segments on the axes
corresponding to the phase plane regions where the cross-negative effect exists (see Definition 2.9).

18
4.3.1 X-factorable transformation
Let us apply the x-factorable transformation ΨX on system (4.7). Letting ΨX ,T ≡ ΨX ◦ ΨT , the
resulting kinetic function KX ,T (x̄; k̄) ≡ (ΨX ,T P)(x̄; k̄) is given by

K1,X ,T (x̄; k̄) = x̄1 (k̄01 + k̄11 x̄1 + k̄21 x̄2 + k̄12
1 1
x̄1 x̄2 + k̄22 (x̄2 )2 ),
K2,X ,T (x̄; k̄) = x̄2 (k̄02 + k̄12 x̄1 + k̄22 x̄2 + k̄22
2
(x̄2 )2 ). (4.11)

Theorem 4.2. ODE systems with RHSs (4.7) and (4.11) are topologically equivalent in the neigh-
bourhood of the fixed points in R2> , with the homoclinic orbit in R2> , and a saddle at the origin being
the only boundary fixed point in R2≥ , if:

a ∈ (−1, 0), |α|  1,

( √  )
2 3 2 8 −2
Φ = T1 > , T2 ∈ max(1, −aT1 ), + a (3 − a2 )(a + 4T1 ) . (4.12)
9 3 3

Proof. Let us assume α = 0.

From statement (i) of Theorem 3.3 it follows that the saddle fixed point of (4.7) is preserved under
ΨX . Denoting the node and spiral fixed points of (4.7) by x̄∗nd and x̄∗sp , respectively, one finds that
the Jacobian is given by:

a + 23 a−1 (1 − a2 ) 1 −2
a (3 − a2 )
   
2 − +
JT |x̄=x̄∗nd = =⇒ sign(JT |x̄=x̄∗nd ) = ,
1 a−1 (2 − a2 ) + −
0 − 13 (3 − a2 )
   
0 −
JT |x̄=x̄sp =
∗ =⇒ sign(JT |x̄=x̄sp ) =
∗ . (4.13)
1 − 13 a + +

Conditions (ii) and (iii) of Theorem 3.3 are both satisfied for the node, but only condition (ii) is
satisfied for the spiral. The condition for the spiral to remain invariant is obtained by demanding
disc(JX ,T |x̄=x̄∗sp ) < 0, where JX ,T is the Jacobian of (4.11), and, taking into consideration (4.10),
this leads to
2 8 −2
T2 < + a (3 − a2 )(a + 4T1 ). (4.14)
3 3
√
Boundary fixed points are given by (0, 0), (T1 +a−1 T2 , 0), (0, 1/2(1± 1 + 4a−1 T1 )+T2 ). The second
fixed point can be removed from the nonnegative quadrant by demanding T2 > −aT1 , while the pair
of the last ones always has nonzero imaginary part due to (4.10). Statement (iv) of Theorem 3.3
implies that the eigenvalues at the origin are given by λ1 = k01 = 3/2(T2 − 2/3)(aT1 + T2 ) > 0 and
λ2 = k02 = −T1 + aT2 (T2 − 1) < 0, so origin is a saddle fixed point.
As α can be chosen arbitrarily close to zero, and as KX ,T (x̄; k̄) is a continuous function of α,
the theorem holds for sufficiently small |α| =
6 0, as well.

A representative phase plane diagram corresponding to the ODE system with RHS (4.11) is shown
in Figure 2(b), where the saddle fixed point at the origin is shown as the black dot. It can be seen
that one of the stable manifolds of the nonboundary saddle, represented as a dashed purple curve,
approaches x2 -axis asymptotically, instead of crossing it as in Figure 2(a).
The homoclinic orbit of the ODE system with RHS (4.11) is positioned below the node in
the phase plane. Suppose the relative position of the stable sets is reversed by, say, applying an
improper orthogonal matrix with the angle fixed to 3π/2, ΨQ3π/2− , with a representative phase

19
plane shown in Figure 2(d). In√this case, one can
straightforwardly show that the boundary fixed
−1
point given by T1 + 1/2(1 + 1 − 4a T2 ), 0 , shown as the black dot in Figure 2(d), cannot
be removed from R2≥ , and is always a saddle. The same conclusions are true for other similar
configurations of the stable sets obtained by rotations only. This demonstrates that x-factorable
transformation can produce boundary artefacts that have a significant global influence on the phase
curves, that cannot be eliminated by simply translating a region of interest sufficiently far away
from the axes. In order to eliminate the particular boundary artefact, a shear transformation may
be applied. Consider applying ΨX ,T ,Q3π/2− ,S2 = ΨX ◦ ΨT ◦ ΨQ3π/2− ◦ ΨS2 on (4.5)–(4.6), where
 
1 0
S2 = , (4.15)
−a 1

and T1 = T2 ≡ T ∈ R, for simplicity, leading to

Kn,X ,T ,Q3π/2− ,S2 (x̄; k̄) = x̄n (k̄0n + k̄1n x̄1 + k̄2n x̄2
n
+ k̄11 (x̄1 )2 + k̄12
n n
x̄1 x̄2 + k̄22 (x̄2 )2 ), n = 1, 2, (4.16)

where the coefficients k̄ = k̄(a, α, T ) are given by:

1
k̄01 = T − 2 + a(2α + T + a(2 + 5T ) + 4T a2 ) ,


2
2 1

k̄0 = − T 2 + 2α + 3T + a(4 + 9T + 6T a) ,
2
1 9 2
k̄11 = − T a(2 + 5a), k̄12 = 1 + T ( + a),
2 2 3
1 1 9
k̄22 = α + a 2 + T + 6T a ,



k̄2 = 1 − a 2α + a(2 + 5T + 8T a) ,
2 2
1 1 2 3 1 5 2 2 9 1 3 2
k̄11 = a, k̄11 = − , k̄12 = a , k̄12 = − a, k̄22 = 2a , k̄22 = −3a2 , (4.17)
2 2 2 2
together with a ∈ (−1, 0), |α|  1 and Φ = {T > 1}.

Theorem 4.3. ODE systems with RHSs (4.5)–(4.6) and (4.16) are topologically equivalent in the
neighbourhood of the fixed points in R2> , with the homoclinic orbit in R2> , and a saddle at the origin
and a saddle with coordinates (0, 1/2T a−1 (1 + 2a)) being the only boundary fixed points in R2≥ , if:
 
1
a ∈ −1, − , |α|  1, and
2
Φ = {T > max(1, 2a−2 (1 − a2 )), T < 2a−1 (1 + 4a)−1 (1 − a)}. (4.18)

Proof. Following the same procedure as in the proof of Theorem 4.2, and noting that
the saddle,
node and spiral fixed
points are given by T , T , T − 2a−2 (1 − a2 ), T + a−3 (1 − a2 ) , and T +

2/9(3 + a2 ), T − 2/9a , respectively, while the five boundary fixed points are (0, 0), (T + 1/2 5T a ±
T (8a−1 (1 − a2 ) + 9T a2 ) , 0), (0, 1/2T a−1 (1+2a)), (0, a−1 2/3+T (1+a) ), one finds (4.18).
p

Note that as α → − 21 , the only boundary fixed point in R2≥ is a saddle at the origin, and it is
connected via a heteroclinic orbit to the saddle in R2> . A representative phase plane diagram
corresponding to the ODE system with RHS (4.16) is shown in Figure 2(c).
While systems (4.11) and (4.16) contain specific variations of the specific homoclinic orbit
given by (4.1), they, nevertheless, indicate possible phase plane topologies of the kinetic equations

20
containing homoclinic orbits of shapes similar to (4.1). With a fixed shape and orientation of
a homoclinic loop which is similar to (4.1), three possible orientations of a corresponding saddle
manifold in R2> are: it may extend in R2> without asymptotically approaching a phase plane axis,
it may asymptotically approach an axis, or it may cross an axis at a fixed point. In Figure 2(b),
a combination of the first and second orientation is displayed, while in Figure 2(c) of the first and
third orientation.

4.3.2 The quasi-steady state transformation

In Lemma 4.2, it was demonstrated that system (4.5)–(4.6) has at least one cross-negative term
under translation transformations. It can be readily shown that system (4.5)–(4.6) in fact has
minimally two cross-negative terms under translation 1 2
√ transformations, k̄2 < 0 and k̄0 < 0, and
this is the case when a ∈ (−1, 0), Φ = {T1 ∈ (2 3/9, −T2 a), T2 > 1}. Let us apply a quasi-
steady state transformation ΨQSSA on system (4.7) that eliminates the two cross-negative terms,
i.e. two new variables are introduced, ȳ1 , ȳ2 ∈ R2> , and we take p1 (x̄) = p2 (x̄) = 1, in Defini-
tion 3.8. Letting ΨQSSA,T ≡ ΨQSSA ◦ ΨT , the resulting kinetic function KQSSA,T ({x̄, ȳ}; k̄, ω, µ) ≡
ΨQSSA,T P)({x̄, ȳ}; k̄, ω, µ) is given by

Kx1 ,QSSA,T ({x̄1 , x̄2 , ȳ1 }; k̄, ω1 ) = k̄01 + k̄11 x̄1 + k̄21 ω1−1 x̄1 x̄2 ȳ1 + k̄12
1 1
x̄1 x̄2 + k̄22 (x̄2 )2 ,
Kx2 ,QSSA,T ({x̄1 , x̄2 , ȳ2 }; k̄, ω2 ) = k̄02 ω2−1 x̄2 ȳ2 + k̄12 x̄1 + k̄22 x̄2 + k̄22
2
(x̄2 )2 ,
Ky1 ,QSSA,T ({x̄1 , ȳ1 }; ω1 , µ) = µ−1 (ω1 − x̄1 ȳ1 ),
Ky2 ,QSSA,T ({x̄2 , ȳ2 }; ω2 , µ) = µ−1 (ω2 − x̄2 ȳ2 ), (4.19)

with x̄n (0) > 0, ωn > 0, n = 1, 2, and µ → 0.

In (4.19), limxn →0 limµ→0 yn = +∞, n = 1, 2, and a geometrical implication is that, say, the
saddle manifold crossing the x2 -axis in Figure 2(a), instead asymptotically approaches the y1 -axis.
The asymptotic behavior of the saddle manifolds is achieved by the additional (boundary) fixed
points in (4.11) displayed in Figure 2(b), and by additional phase space dimensions in (4.19).
Note that a composition an x-factorable and a quasi-steady state transformation may be used to
make (4.7) kinetic. For example, one may eliminate the cross-negative term k̄21 in (4.7) by using the
x1 -factorable transformation ΨX1 , and the cross-negative term k̄02 by using an appropriate ΨQSSA .
An example of a kinetic function obtained by a transformation of the form ΨQSSA,X1 ,T is given by

Kx1 ,QSSA,X1 ,T (x̄; k̄) = x̄1 k̄01 + k̄11 x̄1 + k̄21 x̄2 + k̄12
1 1
(x̄2 )2 ,


x̄1 x̄2 + k̄22
Kx2 ,QSSA,X1 ,T ({x̄, ȳ}; k̄, ω) = k̄02 ω −1 x̄2 ȳ + k̄12 x̄1 + k̄22 x̄2 + k̄22
2
(x̄2 )2 ,
Ky,QSSA,X1 ,T ({x̄2 , ȳ}; ω, µ) = µ−1 (ω − x̄2 ȳ), (4.20)

with x̄2 (0) > 0, ω > 0, and µ → 0. Note that the chosen ΨQSSA,X1 ,T does not introduce any
additional fixed points when applied to system (4.7).

4.4 Step (3): construction of the canonical reaction network RK−1

Definition 3.1 can be used to map the kinetic functions (4.11), (4.16), (4.19), and (4.20) to the
canonical reaction networks RK−1 . This is illustrated for (4.11) in this section, and for (4.16) and
(4.20) in C. For clarity, both the induced kinetic equations and the induced canonical reaction
networks are presented. Note that the reaction networks are assumed to be taking place in an open
reactor, and are not necessarily purely chemical in nature. Nevertheless, the non-chemical processes
present in kinetic equations are represented as quasi-chemical reactions. Such reactions take form

21
of input/output of chemical species, as well as containing quasi-species that are time-independent
on a relevant time scale, so that their constant concentration is absorbed into a quasi-kinetics,
leading to conservation laws not necessarily holding [1].
Writing x ≡ x̄, the induced kinetic equations for (4.11) are given by

dx1
= k1 x1 + k3 x21 − k5 x1 x2 − k7 x21 x2 + k8 x1 x22 ,
dt
dx2
= −k2 x2 + k4 x1 x2 + k6 x22 − k9 x32 ,
dt
while the induced canonical reaction network:
k1 k
2
r1 : s1 −→ 2s1 , r2 : s2 −→ ∅,
k
3 4k
r3 : 2s1 −→ 3s1 , r4 : s1 + s2 −→ s1 + 2s2 ,
k
5 k
6
r5 : s1 + s2 −→ s2 , r6 : 2s2 −→ 3s2 ,
k
7 8k
r7 : 2s1 + s2 −→ s1 + s2 , r8 : s1 + 2s2 −→ 2s1 + 2s2 ,
k
9
r9 : 3s2 −→ 2s2 ,

where k1 = |k̄01 |, k2 = |k̄02 |, k3 = |k̄11 |, k4 = |k̄12 |, k5 = |k̄21 |, k6 = |k̄22 |, k7 = |k̄12

1 |, k = |k̄ 1 |,
8 22
2
k9 = |k̄22 |, with the coefficients k̄ given by (4.8), and the conditions given by (4.12).

5 Summary
In the first part of the paper, a framework for constructing reaction systems having prescribed
properties has been formulated as an inverse problem and presented in Section 3, relying on defini-
tions introduced in Section 2. As a part of the framework, in Section 3.2, kinetic transformations
have been defined that enable one to map an arbitrary polynomial ODE system with a set of
constraints, possibly containing the cross-negative terms, into a kinetic one. Augmented with the
results from [16], such transformations can be applied to nonpolynomial systems as well. Systems
for which no affine transformation is kinetic have been defined as affinely nonkinetic in Section 3.2.1,
to emphasize the fact that significant changes to their solutions are required. X-factorable transfor-
mation [10], that does not change the dimension of the systems being transformed, but introduces
a higher number of nonlinear terms and leads to autocatalytic reaction networks, has been defined
in Section 3.2.2, and its properties when applied on two-dimensional systems have been derived in
Theorem 3.3. The quasi-steady state transformation, that increases the dimension of the systems
being transformed, but generally introduces a lower number of nonlinear terms, has been presented
in Section 3.2.3, and justified using Tikhonov’s and Korzukhin’s theorems [11]. As the focus of
the paper has been more placed on the construction of kinetic equations, and less on construc-
tions of reaction networks, in Section 3.1 an analytical and algorithmic method for obtaining the
so-called canonical networks has been presented [4]. The framework may be used for constructing
lower-dimensional reaction systems displaying exotic phenomena, with Algorithm 1 exemplifying
the construction process.
In the second part of the paper, the inverse problem framework has been applied to a case study
of constructing bistable reaction systems undergoing a supercritical homoclinic bifurcation, with a
parameter controlling the stable sets separation, with an overview of the procedure presented in
Section 4.1. In Section 4.2, a polynomial ODE system having a homoclinic orbit in the phase plane

22
has been constructed using the results from [21], and perturbed in such a way that the sufficient
conditions for the existence of the homoclinic bifurcation are fulfilled, as required by Andronov-
Leontovich Theorem [24]. In Section 4.3, the kinetic transformations from Section 3 have been
used in order to map the polynomial system to a kinetic one with the homoclinic orbit in the
positive quadrant. The topological phase space effects produced by the kinetic transformations on
the constructed systems have been discussed. In Section 4.4 and C, the canonical reaction networks
induced by some of the kinetic equations have been presented.
In this paper, we have constructed chemical reaction networks inducing two-dimensional cubic
kinetic functions with the deterministic ODEs (kinetic equations) undergoing a supercritical homo-
clinic bifurcation. In a future publication, we will report our results on the stochastic analysis of the
constructed systems, consisting of examining the quasi-stability of the limit cycle, and stochastic
switching between the stable sets, as a function of the bifurcation parameter and the parameter
controlling the stable set separation. A motivation for such a study is the fact that stochastic ef-
fects play an important role in systems biology due to the inherently small reactors [25, 27, 28, 29],
and one might even say that systems biology has initiated a renaissance of the stochastic reaction
kinetics [31].

A Oscillations in two-component bimolecular reaction systems

with cross-negative terms
Theorem A.1: Consider a two-dimensional ODE system with a quadratic polynomial RHS,
P(x; k) ∈ P2 (R2 ; R2 ), k ∈ R12 :
dx1
= P1 (x; k) = k01 + k11 x1 + k21 x2 + k11 1 2 1
x1 + k12 x1 x2 + k221 2
x2 ,
dt
dx2
= P2 (x; k) = k02 + k12 x1 + k22 x2 + k11 2 2 2
x1 + k12 x1 x2 + k222 2
x2 , (A.1)
dt
1 ≤ 0 and k 2 ≤ 0, and if system (A.1) is nonnegative, then the system has no limit cycles.
If k11 22


Proof. Considering x1 , x2 > 0, writing P(x; k) = P1 (x1 , x2 ; k), P2 (x1 , x2 ; k) , and letting the
Dulac function to be given by B(x1 , x2 ) = (x1 x2 )−1 , it follows that
∂
∂

D(x1 , x2 ; k) = B(x1 , x2 )P1 (x1 , x2 ; k) + B(x1 , x2 )P2 (x1 , x2 ; k)
∂x1 ∂x2
 1 1 1 1
 2
k12 k11 2 x 2
 
k0 k2 k22 x2 k11 k0 1 k22
=− + + + − + + + .
x21 x2 x21 x21 x2 x1 x22 x22 x22 x1
Multiplying by −(x1 x2 )2 , and defining a new function D̄(x1 , x2 ; k) = −(x1 x2 )2 D(x1 , x2 ; k), results
in
1 2 2 2
   
D̄(x1 , x2 ; k) = x2 P1 (0, x2 ) − k11 x1 + x1 P2 (x1 , 0) − k22 x2 .
1 ≤ 0 and k 2 ≤ 0, and if system (A.1) is nonnegative, so that P (0, x ) ≥ 0 and P (x , 0) ≥ 0,
If k11 22 1 2 2 1
then D̄ ≥ 0. The only case when a limit cycle may exist is if D̄ = 0 for all x1 , x2 > 0, and in [20]
it is shown that no limit cycles exist in this case.
In [20], it was shown that the absence of cross-negative terms in P(x; k) ∈ P2 (R2 ; R2 ), with k11
1 ≤0
2
and k22 ≤ 0, implies the absence of limit cycles, i.e. one requires the more restrictive condition
P(x; k) ∈ PK 2 2
2 (R≥ ; R ). Theorem A.1 shows that, in fact, absence of the cross-negative effect in
2 2
P(x; k) ∈ P2 (R ; R ), with k11 1 ≤ 0 and k 2 ≤ 0, implies the absence of limit cycles, i.e. one
22
requires the less restrictive condition P(x; k) ∈ P2 (R2≥ ; R2 ).

23
B Andronov-Leontovich theorem
Andronov-Leontovich theorem [24]: Consider system (3.3) with P(x; k, α) ∈ Pm (R2 ; R2 ),
m ∈ N, where α ∈ R is a bifurcation parameter. Let λ1 (α) and λ2 (α) be eigenvalues of the
Jacobian corresponding to the two-dimensional system (3.3), J = ∇P(x; k, α), and suppose that
at α = 0, the following homoclinic bifurcation conditions (i)–(ii) are satisfied, and that (3.3) is
generic, i.e. the following homoclinic genericity conditions (iii)–(iv) are satisfied:

(i) System has a saddle fixed point x∗ = 0 with eigenvalues λ1 (0) < 0 < λ2 (0).

(ii) System has a homoclinic orbit γ ∗ to the saddle fixed point x∗ .

(iii) Nondegeneracity condition: σ0 = λ1 (0)+λ2 (0) 6= 0, where σ0 ∈ R is called the saddle quantity.

(iv) Transversality condition: Melnikov integral, M (α), along the homoclinic orbit satisfies:
Z +∞  
∂P(x; k, α)
M (0) = ϕ(t) P(x; k, α) × dt 6= 0, (B.1)
−∞ ∂α
 R 
t
where ϕ(t) = exp − 0 (∇ · P(x; k, α)dτ , ϕ(t) > 0. This is equivalent to splitting of the
homoclinic orbit at the bifurcation with a nonzero speed.

Then, for all sufficiently small |α|, there exists a neighbourhood of the saddle fixed point and the
homoclinic orbit such that a unique limit cycle bifurcates from γ ∗ . If σ0 < 0, the homoclinic
bifurcation is supercritical, giving rise to a unique stable limit cycle, while if σ0 > 0, the homoclinic
bifurcation is subcritical, giving rise to a unique unstable limit cycle.

C The canonical reaction networks induced by (4.16) and (4.20)

Writing x ≡ x̄, the induced kinetic equations for (4.16) are given by

dx1
= −k1 x1 − k3 x21 + k5 x1 x2 − k7 x31 + k9 x21 x2 − k11 x1 x22 ,
dt
dx2
= k2 x2 + k4 x1 x2 − k6 x22 − k8 x21 x2 + k10 x1 x22 − k12 x32 ,
dt
while the canonical reaction network:
k
1 k
2
r1 : s1 −→ ∅, r2 : s2 −→ 2s2 ,
k k
r3 : 3
2s1 −→ s1 , r4 : 4
s1 + s2 −→ s1 + s2 + sign(k̄12 )s2 ,
5k k
6
r5 : s1 + s2 −→ 2s1 + s2 , r6 : 2s2 −→ s2 ,
7k 8 k
r7 : 3s1 −→ 2s1 , r8 : 2s1 + s2 −→ 2s1 ,
9k k
r9 : 2s1 + s2 −→ 3s1 + s2 , r10 : s1 + 2s2 −−10
→ s1 + 3s2 ,
k k
r11 : s1 + 2s2 −−11
→ 2s2 , r12 : 3s2 −−12
→ 2s2 ,

where k1 = |k̄01 |, k2 = |k̄02 |, k3 = |k̄11 |, k4 = |k̄12 |, k5 = |k̄21 |, k6 = |k̄22 |, k7 = |k̄11

1 |, k = |k̄ 2 |,
8 11
1 |, k
k9 = |k̄12 2 1 2
10 = |k̄12 |, k11 = |k̄22 |, k12 = |k̄22 |, with the coefficients k̄ given by (4.17), and the

24
 √

conditions given by (4.18). Note that by taking a ∈ − 89 , 15 (2 − 34) and T = − 23 (2 + 3a)−1 , it
follows that k4 = 0.
Writing x ≡ x̄, the induced kinetic equations for (4.20) are given by

dx1
= k1 x1 + k3 x21 − k5 x1 x2 − k7 x21 x2 + k8 x1 x22 ,
dt
dx2
= −k2 x2 x3 + k4 x1 + k6 x2 − k9 x22 ,
dt
dx3
= k10 − k11 x2 x3 ,
dt
while the canonical reaction network:
k k
r1 : 1
s1 −→ s1 + sign(k̄01 )s1 , r2 : 2
s2 + s3 −→ s3 ,
3k k
4
r3 : 2s1 −→ 3s1 , r4 : s1 −→ s1 + s2 ,
5 k k
6
r5 : s1 + s2 −→ s2 , r6 : s2 −→ 2s2 ,
7 k 8 k
r7 : 2s1 + s2 −→ s1 + s2 , r8 : s1 + 2s2 −→ 2s1 + 2s2 ,
9k k
r9 : 2s2 −→ s2 , r10 : ∅ −−10
→ s3 ,
k
r11 : s2 + s3 −−11
→ s2 ,

with k1 = |k̄01 |, k2 = |k̄02 |, k3 = |k̄11 |, k4 = |k̄12 |, k5 = |k̄21 |, k6 = |k̄22 |, k7 = |k̄12

1 |, k = |k̄ 1 |, k = |k̄ 2 |,
8 22 9 22
k10 = µω, k11 = µ, the coefficients k̄ given by (4.8), ω > 0, µ → 0, and conditions given by (4.12),
with the lower bound on T2 being 1. Note that by taking T1 = −T2 a−1 , it follows that k1 = 0.

Acknowledgments: The research leading to these results has received funding from the European
Research Council under the European Community’s Seventh Framework Programme (FP7/2007-
2013)/ERC grant agreement no 239870, and from the People Programme (Marie Curie Actions) of
the European Union’s Seventh Framework Programme (FP7/2007-2013) under REA grant agree-
ment no. 328008. Tomáš Vejchodský gratefully acknowledges the support of RVO 67985840. Radek
Erban would like to thank the Royal Society for a University Research Fellowship and the Lever-
hulme Trust for a Philip Leverhulme Prize.

References
[1] Feinberg, M. Lectures on Chemical Reaction Networks, (Delivered at the Mathematics Research
Center, U. of Wisconsin, 1979).

[2] Érdi, P., Tóth, J. Mathematical Models of Chemical Reactions. Theory and Applications of
Deterministic and Stochastic Models. Manchester University Press, Princeton University Press,
1989.

[3] Szili, L., Tóth, J, 1997. On the origin of Turing instability. Journal of Mathematical Chemistry,
22: 39–53.

[4] Hárs, V., Tóth, J., 1981. On the inverse problem of reaction kinetics. Qualitative Theory of
Differential Equations, eds. Farkas, M., Hatvani, L., 363–379.

25
 [5] Tóth, J., Hárs, V., 1986. Orthogonal transforms of the Lorenz- and Rössler- equation. Physica
19D, 135–144.

[6] Escher, C., 1981. Bifurcation and coexistence of several limit cycles in models of open two-
variable quadratic mass-action systems. Chemical Physics, 63: 337–348.

[7] Chellaboina, V., Bhat, S. P., Haddad, W. M., Bernstein, D.S., 2009. Modeling and Analysis
of Mass-Action Kinetics. IEEE Control Systems Magazine, 29: 60–78.

[8] Craciun, G., Pantea, C., 2008. Identifiability of chemical reaction networks. Journal of Math-
ematical Chemistry, 44: 244–259.

[9] Szederknyi, G., Hangos, K. M., Pni, T., 2011. Maximal and minimal realizations of reaction
kinetic systems: computation and properties. MATCH Communications in Mathematical and
in Computer Chemistry, 65(2): 309–332.

[10] Samardzija, N., Greller, L. D., Wasserman, E., 1989. Nonlinear chemical kinetic schemes
derived from mechanical and electrical dynamical systems. Journal of Chemical Physics, 90:
2296–2304.

[11] Klonowski, W., 1983. Simplifying principles for chemical and enzyme reaction kinetics. Bio-
physical Chemistry, 18(3): 73–87.

[12] Okniński, A. Catastrophe Theory: Volume 33. Elsevier Science, 1992.

[13] Pantea, C., Gupta, A., Rawlings, J. B., Craciun, G., 2014. The QSSA in chemical kinetics: as
taught and as practiced. Discrete and Topological Models in Molecular Biology, pp. 419–442.
Springer, Berlin.

[14] Hangos, K. M, Szederkényi, G., 2011. Mass action realizations of reaction kinetic system
models on various time scales. Journal of Physics: Conference Series, 268.

[15] Hangos, K. M, 2010. Engineering model reduction and entropy-based Lyapunov functions in
chemical reaction kinetics. Entropy 12, pp. 772–797.

[16] Kerner, E. N., 1981. Universal formats for nonlinear ordinary differential systems. Journal of
Mathematical Physics, 22: 1366–1371.

[17] Kowalsi, K., Steeb, W. H. Nonlinear Dynamical Systems and Carleman Linearization. Word
Scientific, 1991.

[18] Kowalski, K., 1993. Universal formats for nonlinear dynamical systems. Chemical Physics
Letters, 209: 167–170.

[19] Wilhelm, T., 2000. Chemical systems consisting only of elementary steps - a paradigma for
nonlinear behavior. Journal of Mathematical Chemistry, 27: 71–88.

[20] Póta, G., 1983. Two-component bimolecular systems cannot have limit cycles: A complete
proof. Journal of Chemical Physics, 78(3): 1621–1622.

[21] Sandstede, B., 1997. Constructing dynamical systems having homoclinic bifurcation points of
codimension two. Journal of Dynamics and Differential Equations, Vol. 9, No. 2., 4: 296–288.

26
[22] Wilhelm, T., Heinrich, R., 1995. Smallest chemical reaction system with Hopf bifurcation.
Journal of Chemical Physics, 17: 1–14.

[23] Rössler, O. E., 1974. A synthetic approach to exotic kinetics (with examples). Lecture Notes
in Biomathematics, 4: 546–582.

[24] Kuznetsov, Y. A. Elements of Applied Bifurcation Theory, Second Edition. Springer-Verlag,

New York, 2000.

[25] Erban, R., Chapman, S. J., Kevrekidis, I. and Vejchodsky, T., 2009. Analysis of a stochastic
chemical system close to a SNIPER bifurcation of its mean-field model. SIAM Journal on
Applied Mathematics, 70(3): 984–1016.

[26] Borisuk, M. T., Tyson, J. J., 1998. Bifurcation analysis of a model of mitotic control in frog
eggs. Journal of Theoretical Biology, 195: 69–85.

[27] Kar, S., Baumann, W. T., Pau,l M. R. and Tyson, J. J., 2009. Exploring the roles of noise
in the eukaryotic cell cycle. Proceedings of the National Academy of Sciences of USA, 106:
6471–6476.

[28] Vilar, J. M. G., Kueh, H. Y., Barkai, N. and Leibler, S., 2002. Mechanisms of noise-resistance
in genetic oscillators. Proceedings of the National Academy of Sciences of the United States of
America, 99(9): 5988–5992.

[29] Dublanche, Y., Michalodimitrakis, K., Kummerer, N., Foglierini, M. and Serrano, L., 2006.
Noise in transcription negative feedback loops: simulation and experimental analysis. Molec-
ular Systems Biology, 2(41): E1–E12.

[30] Elowitz, M. B., Leibler, S., 2000. A synthetic oscillatory network of transcriptional regulators.
Nature, 403(6767): 335–338.

[31] Érdi, P., Lente, G. Stochastic Chemical Kinetics. Theory and (Mostly) Systems Biological
Applications. Springer Series in Synergetics, 2014.

[32] Voit, E. O., Martens, H. A., Omholt, S. W., 2015. 150 years of the mass action law. PLOS
Computational Biology, 11(1).

[33] Lawrence, J. D. A Catalog of Special Plane Curves. Dover, New York, 1972.

[34] Grantmacher, F. R. Applications of the Theory of Matrices. Interscience Publishers, INC., New
York, 1959.

27
 Phase Plane Under YT Phase Plane Under YX,T
5 5

(a) 4 (b) 4

3 3
x2

x2
2 2

1 1

0 0

-1 -1
-1 0 1 2 3 4 5 -1 0 1 2 3 4 5
x1 x1
Phase Plane Under YX,T ,Q32Π-
Phase Plane Under YX,T ,Q32Π- ,S2
5
6
(c) 4 (d)

3 4
x2
x2

2
2
1

0
0

-1
-1 0 1 2 3 4 5 0 2 4 6
x1 x1

Figure 2: Phase plane diagrams of (a) ODE system with RHS (4.7); (b) ODE system with
RHS (4.11); and (c) ODE system with RHS (4.16). The stable node, the saddle, and the un-
stable spiral are represented as the green, blue and red dot, respectively. The alpha curve is shown
in purple, and the vector field as gray arrows. On each plot it is indicated which kinetic transforma-
tion is applied to system (4.5)–(4.6). Red segments of the phase plane axes in (a) denote the regions
where the cross-negative effect exists. In (b) and (c), boundary fixed points are represented as the
black dots, purple curves with a coarser dashing represent the saddle manifolds that asymptotically
approach an axis, while with a finer dashing those that are outside of R2≥ .
(d) Phase plane diagram of a system for which x-factorable transformation significantly globally
influences the phase curves in such a way that a pure translation cannot resolve the artefacts. For
more details, see the text. The parameters are fixed to a = −0.8, α = 0, T1 = 2, T2 = 2.

28