You are on page 1of 8

# Approximation of Dynamical Systems using S-Systems ∗ Theory: Application to Biological Systems

Laurent Tournier
Laboratoire de Modelisation et Calcul, ´ 51 rue des Mathematiques, ´ B.P. 53, 38041 Grenoble cedex 9, France.

Laurent.Tournier@imag.fr

ABSTRACT
In this article we propose a new symbolic-numeric algorithm to ﬁnd positive equilibria of a n-dimensional dynamical system. This algorithm uses a symbolic manipulation of ODE in order to give a local approximation of diﬀerential equations with power-law dynamics (S-systems). A numerical calculus is then performed to converge towards an equilibrium, giving at the same time a S-system approximating the initial system around this equilibrium. This algorithm has been applied to a real biological example in 14 dimensions which is a subsystem of a metabolic pathway in Arabidopsis Thaliana. Categories and Subject Descriptors: I.1.2 [Computing Methodologies]: Symbolic and algebraic manipulation - Algorithms G.1.5 [Mathematics of Computing]: Numerical Analysis Roots of Nonlinear Equations - Iterative Methods General Terms: Algorithms. Keywords: Power-law, S-systems, Dynamical Systems, Positive Equilibria, Stability.

1.

INTRODUCTION

The modelling and study of biological or biochemical systems has become an exciting challenge in applied mathematics. The complexity of real biological dynamical systems lies essentially in the non-linearities of the dynamics as well as in the huge dimension of systems, often leading to a numerical approach. However, as the understanding of cellular mechanisms grows, it has become obvious that the modelling step strongly needs symbolic tools in order to manipulate more and more information and data, and to improve computa∗This work is part of the CALCEL project, funded by ”La R´gion Rhˆne-Alpes”, France. e o

tional tools. Therefore a new area emerged, called ”systems biology”. It involves diﬀerent ﬁelds of applied mathematics, from computer algebra (see for instance [10]) to numerical computation ([7]). In the past decades, a lot of diﬀerent frameworks have been developped to study behaviors of complex biochemical processes. Let us cite here three of them: discrete networks (see the work of R. Thomas [16]), piecewise linear systems (the so-called Glass networks [6], see also [5]) and sigmoidal switch systems ([13]). The main goal of all these approaches is to propose a (more or less) generic class of dynamical systems, either discrete or diﬀerential, that model some behaviors of complex interaction systems. Once this class is clearly deﬁned, its mathematical relevance generally allows both theoretical and numerical analysis. We use in this paper the class of S-systems (see [1], [17], [18]). The basic idea of this model is to represent interactions between biochemical species with power-law dynamics. Their mathematical expression is quite general, but suﬃciently simple to allow theoretical and practical investigations. We propose in this article a symbolic-numeric algorithm that is based upon S-systems theory. Its goal is to compute iteratively the positive equilibria of a n-dimensional system of ordinary diﬀerential equations (ODE). As it converges towards a steady state, it provides a S-system that approaches the original dynamics around this steady state. As we will see, the local approximation of some dynamics with power-laws can be made symbolically at any point of the phase space. It can also include treatment of symbolic parameters. However, iterating this process in order to converge towards equilibria needs a numerical computation, which prevents the use of pure symbolic tools to the end. In the following, we give a deﬁnition of the S-system class as it can be found in the litterature (see for instance [17]). We then propose a symbolic-numeric algorithm that computes an iteration leading to the positive equilibria of a dynamical system. We will see an application of this algorithm on a biological example in dimension 14. We ﬁnally conclude with some remarks on our algorithm and some future works.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proﬁt or commercial advantage and that copies bear this notice and the full citation on the ﬁrst page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior speciﬁc permission and/or a fee. ISSAC’05, July 24–27, 2005, Beijing, China. Copyright 2005 ACM 1-59593-095-705/0007 ...\$5.00.

2. S-SYSTEMS 2.1 Deﬁnition
In this paragraph, we give a deﬁnition of the class of Ssystems (see for instance [1]).

.j=1. As an example. . G. let us consider the one-dimensional case. H) has a unique equilibrium x in Ω (i.e. h) : x = f (x) = αxg − βxh ˙ x = b(G−H) e −1 (3) k k (1) where α. The positive equilibrium x of e (S) exists and is unique if and only if g − h = 0. .n F is C 1 and therefore locally lipschitz on the open Ω. 2.n . R∗ denotes the set of strictly positive real numbers. i = 1 . n: n αi Y gik ∂fi xk . a positive equilibrium) if and only e if the matrix (G − H) is invertible. + we deﬁne the vector xA ∈ (R∗ )n by: + (xA )i = n Y xj ij .j=1. xn ) B C . . we obtain. j = 1 . In this case. . A one-dimensional S-system is expressed by a single diﬀerential equation: S(α. S-systems are part of a broader formalism known as quasimonomial (QM) systems (see [2]). F (x) = @ A . . G. A real square matrix A of order n is said to be stable (resp. h ∈ R. We have therefore the following proposition: Proposition 2. a i = 1. βn ) ∈ (R∗ )n + and G = (gij )i. . Let us introduce the vector ﬁeld F deﬁned on Ω = (R∗ )n : + 0 1 f1 (x1 . To ﬁnd them. . An interesting result shows that QM systems can be expressed in the form of Lotka-Volterra quadratic systems (see [2] for details). fn (x1 . . αn ) ∈ .H= 4 1 3 4 5 3 xj ij = βi g j=1 j=1 n Y xj ij . . β. xn ). A S-system S(α. . we have to solve the system: αi n Y We thus obtain a formula that links the jacobian of F in x e with the matrix G − H.j=1. x Let us recall here the deﬁnition of stability of matrices: Definition 2. .e.n (2) In this paper we will use the following notation: Given a vector x ∈ (R∗ )n and a real square matrix A = (aij )i. consider the S-system:  x = 3xy 2 − 2x4 ˙ (S1) : y = 4x3 y 4 − x5 y 3 ˙ we have: „ « „ « „ « „ « 3 2 1 2 4 0 α= . G. e as we can see in the following example. n h with α = (α1 .2.1.2 Equilibrium points The study of the phase portrait of a S-system S(α.n ∈ Mn (R). .. .. β. . semi-stable) if all its eigenvalues λi . . . an obvious calculation leads to: ∂f (e) = αeg−1 (g − h) x x ∂x so the stability of x depends directly on the sign of g − h: it e is asymptotically stable if g−h < 0 and unstable if g−h > 0. i = 1 . H) begins with the search for equilibrium points in Ω. it is not trivial to link the spectrum of JF (e) with the spectrum of G − H.n j=1 With this notation. β = (β1 . non positive) real part. G...Definition 2. H) is a dynamical system deﬁned by the n diﬀerential equations: xi = αi ˙ n Y where b is the vector (β1 /α1 . Diﬀerentiating the functions fi (x1 . coeﬃcients αi and βi are sometimes called kinetic rates while gij and hij are called kinetic orders.. g. . . regardless of parameters α and β. . . .1.3 Stability analysis of the equilibrium h xj ij . Posing y = ln x. .. have a negative (resp. β. xn ) = αi n Y g xj ij (R∗ )n + − βi j=1 j=1 n Y 2.β= . β. β. . .. This deﬁnition of S-systems with power-law diﬀerential equations is strongly linked with equations of chemical kinetics. . However. However this is not true. (gij − hij ) e (4) (e) = x ∂xj xj e k=1 then the mass-action law applied to species C gives the equation: dc = k1 ab2 − k2 d3 e dt (capital letters designate species and small letters designate concentrations). x can be calculated by e the formula: xj ij − βi g j=1 j=1 n Y xj ij . we can express equation (2) as follows: xG−H = b . given any initial condition x(0) = x0 ∈ Ω. We could hope that the stability of matrix G − H was suﬃcient to deduce the stability of x. . For n = 2. . n. Taking the neperian logarithm. e As a motivating example. H = (hij )i. the stability depends on the sign of the real parts of the jacobian’s eigenvalues. CauchyLipschitz theorem ensures the existence and unicity of a maximal solution of S(α. . . + Mn (R) denotes the set of real square matrices of order n. i. βn /αn ). . A n-dimensional S-system S(α. h i = 1. xn ) with: fi (x1 .n . H) in Ω. The quesx e tion we tackle here is to ﬁnd some relationship between the stability of x and some properties of the matrix G − H. . .1. we are brought back to the resolution of a ndimensional linear system in y.G= .. this equation leads to: (G − H) ln x = ln b (the logarithm is applied to all components of vector.. . . . for i. ln x is the n-dimensional vector (ln x1 . i = 1. ln xn )). Therefore in deﬁnition 2. .. β > 0 and g. In the n-dimensional case. .. if we consider the following chemical pathway: 1 2 A + 2B → C → 3D + E The stability analysis of the equilibrium x uses the study e of the spectrum of JF (e) (the jacobian of F in x).

γn 1 xn e C A .We can calculate the two eigenvalues λ1 and e 3 9 λ2 of JF (e). . sgn(x) = : signum function: +1 0 −1 if x > 0 if x = 0 if x < 0 can thus deduce that J is semi-stable and as x is supposed e hyperbolic. sign semi-stable) if all the matrices that have the same sign pattern are stable (resp. which would make x unstable. We ﬁnd that λ1 .3. β. F (x) = @ A . The hyperbole icity of x is therefore an essential assumption in the latter e proposition. Simultaneously. the equilibrium x is asymptotically stable. we deduce that J is also invertible and does not have null eigenvalues. in matricial notation: J = ΓP D−1 where Γ and D are diagonal matrices: 0 0 1 x1 e γ1 B B C . n) in the neighborhood of y 0 at the ﬁrst order: ∀i = 1 .The matrix G − H is equal to: „ « −3 2 G−H = −2 1 Its characteristic polynomial is χ(λ) = (λ+1)2 so the matrix is stable. Since G − H is invertible.. implying that x is an x e unstable node. we can formulate the following proposition. . j = 1 . λ2 > 0. none of the eigenvalues of the e jacobian of F in x have zero real part).[11]). . 256 . in spite of the stability of matrix G − H. Let’s consider the positive vector ﬁeld F : (R∗ )n → (R∗ )n . e If the matrix G − H is sign semi-stable (i. D=@ Γ=@ A. n: ˜ ∀i = 1 . gi (y) = gi (y 0 ) + n X j=1 0 (yj − yj ) Proof. As γi > 0 and xj > 0 for all i e k=1 ek and j. fn (x1 . we propose an algorithm for approximating the equilibria of a dynamical system using S-systems. Let us note J the Jacobian of F in x and P the e matrix G − H. . We egi (y 0 ) « „ 0 ∂gi (y 0 ) exp (yj − yj ) ∂yj j=1 n X j=1 (yj − 0 ∂gi yj ) ∂yj (y ) 0 ! . . . Two real square matrices of order n. . . We assume that G − H is invertible and we note x the unique positive equilibrium of (S). . Let us remark that the latter equation gives..e. gi (y) = gi (y 0 ) + ˜ ˜ and the functions fi = exp(˜i (y)): g ˜ fi (x) def n X j=1 0 (yj − yj ) ∂gi 0 (y ) ∂yj = ˜ egi (y) = = e gi (y 0 ) exp n Y Q with γi = αi n xgik . LOCAL APPROXIMATION OF DYNAMICAL SYSTEM USING S-SYSTEMS In this part. we obtain a S-system that approximates the initial system around the equilibrium. semi-stable) in the sense of deﬁnition 2.n is sign semi-stable if and only if it satisﬁes the following three conditions: (i) ∀ i = 1 . and express the logarithm of F (x) as a function G of the new variable y: ln F (x) = ln F (ey ) = G(y) The function G is suﬃciently smooth on Rn . It could nevertheless have imaginary eigenvalues.4. ai(k−1)i(k) ai(k)i(1) = 0 (The third condition is equivalent to the fact that the directed graph associated to A admits no k-cycle for k ≥ 3) With this notion. if it veriﬁes the three conditions of theorem 2. Let consider a n-dimensional S-system S(α. . xn ) C B . G. Definition 2. n . We also ase sume that x is hyperbolic (i.n and B = (bij )i. e We need a stronger property known as sign stability (see [9]. + Let us deﬁne the following change of variables: y = ln x..j=1. it is asymptotically stable.1 Monomial approximation of a positive vector ﬁeld (see [18]. . A real square matrix A of order n is said to be sign stable (resp. . e ∂gi 0 (y ) + o( y − y 0 ) ∂yj We introduce the functions gi (y) for i = 1 . . matrices J and P have the same sign pattern. .[15]. As we have supposed that P is invertible. . . ..[17])... n .2. A = (aij )i. .. . + + 1 0 f1 (x1 .. In [9] we ﬁnd a characterization of the sign semi-stability: Theorem 2.e. e The stability of G − H is therefore insuﬃcient to deduce the stability of x.n .j=1. . The equation 4 yields: ∂fi γi (e) = x pij ∂xj xj e Proposition 2. H). the equilibrium x is unstable. . which links the stability of the equilibrium x of a Se system with the sign semi-stability of matrix G − H: 3. there is a unique equilibrium: ` ´ x = 32 . aij aji ≤ 0 (iii) for each sequence of k ≥ 3 distinct indices i1 . n . Given any arbitrary point y 0 ∈ Rn . A real square matrix A = (aij )i. Definition 2.2. .j=1. . As a result. aii ≤ 0 (ii) ∀ i = j . we have: ai(1)i(2) . xn ) We will suppose F suﬃciently smooth on (R∗ )n . sgn(aij ) = sgn(bij ) The function sgn is the classical 8 < ∀x ∈ R . . We can easily verify in the previous example that G − H is stable but not sign semi-stable (g22 − h22 = 1 > 0). let us write the Taylor expansion of gi (for i = 1 . n . . regardless of parameters α and β. . ik .1) then. . . so sgn(det(J)) = sgn(det(P )). . have the same sign pattern if: ∀ i. 3.1 (Quirk-Ruppert-Maybee).

1. we obtain: (Sx0 ) : where: xi = αi ˙ n Y j=1 x0 j xj ij − βi g j=1 j=1 n Y xj ij .n n Y 8 > g > ij < = = If the matrix G − H is invertible.. j = 1 . The main steps of this algoritm are summarized here: Algorithm 1 Search of an equilibrium point of system (S) Require: X = x0 V +.. This point xeq depends on the initial point x0 where we made our approximation. repeat Y := X for i = 1 to n do for j = 1 to n do + − Xj ∂vi Xj ∂vi gij := + . .n 8 > > α > i > < > > β > i > : + = vi (x0 ) = − vi (x0 ) j=1 n Y n Y (x0 )−gij j (7) (x0 )−hij j j=1 and: “ ” ˜ e Therefore. . For i = 1 . We propose an algorithm for ﬁnding an equilibrium point of (S) that lies in (R∗ )n . we ﬁnd a point y close to a positive equilibrium of (S) with the precision .n ! n Y g e (5) F (x) = αi x ij j j=1 i=1. Let F be a smooth n-dimensional vector ﬁeld. V − : (R∗ )n → (R∗ )n . the term vi (x) is + + − the production term of the variable xi and vi (x) the decay term of xi . Let F be a positive vector ﬁeld and F its S-approximation in x0 . . . Let us consider the n-dimensional S-system: (Sx0 ) e e x = V + (x) − V − (x) ˙ bi := βi /αi end for if det(G − H) = 0 then −1 X := b(G−H) else degenerate case: algorithm terminated → restart the algorithm with a new initial condition end if until X − Y < Result := X . we have deﬁned a vector ﬁeld F = fi i=1. we obtain the S-system (Sy ) that approximate system (S) around this equilibrium..1. Iterating this process produces a sequence of points x0 . the system (Sx0 ) admits a unique equilibrium xeq ∈ (R∗ )n : + xeq = b(G−H) −1 > > hij : x0 j + vi (x0 ) x0 j − vi (x0 ) + ∂vi (x0 ) ∂xj − ∂vi (x0 ) ∂xj (8) with: 8 > > αi (x0 ) > < = fi (x0 ) = x0 j (x0 )−gij j (6) j=1 Definition 3. The following proposition is basic for what follows: e Proposition 3. The following equalities hold: • ∀ i. .. which is equivalent: JF (x ) = JF (x0 )) e e The basic idea is to use the monomial vector ﬁeld F as an approximation of F in a neighborhood of x0 .. F : (R∗ )n → (R∗ )n and x0 any vector of (R∗ )n . We + + + e call S-approximation of F in x0 the vector ﬁeld F deﬁned by equations (5) and (6). .. we have: ˜ fi (x) = fi (x0 ) and: ∂gi (y) = ∂yj = = = ∂ (ln(fi (ey ))) ∂yj 1 ∂ (fi (ey )) fi (ey ) ∂yj 1 yj ∂fi y e (e ) fi (x) ∂xj xj ∂fi (x) fi (x) ∂xj i n Y „ xj « ∂yj (y 0 ) ∂g Using (5) and (6). x1 . h i = 1. e Given a point x0 in (R∗ )n . V − >0 ∈ (R∗ )n (initial condition) + : positive vector ﬁelds deﬁned over (R∗ )n + : precision Ensure: unless we fall in a degenerate case. . 3. we introduce the ﬁelds V + and + e − which are the S-approximations of the ﬁelds V + and V − V in x0 . Meanwhile.As y = ln x and gi (y) = ln fi (x). n. . + + V + . . that converges towards a positive equilibrium of (S). . βn /αn ). V − are positive vector ﬁelds. > > > gij (x0 ) : ∂fi 0 (x ) fi (x0 ) ∂xj with b = (β1 /α1 . hij := − vi (X) ∂Xj vi (X) ∂Xj end for n n Y Y + − αi := vi (X) (Xj )−gij . Let x1 = xeq be the new initial point where we make our next Sapproximation. Meanwhile. e • F (x0 ) = F (x0 ) ˜ ∂ fi 0 ∂fi 0 (x ) = (x ) ∂xj ∂xj 0 (or.2 Finding equilibria of a dynamical system We consider a n-dimensional dynamical system of the form: (S) x = V + (x) − V − (x) ˙ where x lies in (R∗ )n and V + . . βi := vi (X) (Xj )−hij j=1 j=1 The proof is easy and left to the reader. n . we get + a S-system that approximates the system (S) around this equilibrium. A proof of the convergence is given in section 5.

5 − y ˙ • With initial condition x0 = (0.352 y −0. we will omit the dependency in x. In other ¯ words.132 − x ˙ y = 3. (for conx x venience. 2).. n. βi (x0 ).837 x−1. We will discuss about the convergence speed in part 5. βi xj ij = αi ¯ xj ij ¯ −1 −1 (10) j=1 j=1 . The arrows represent the three experimentations described in the text. y) = 0 (dashed line) and f2 (x. As we can see on ﬁgure 1. The numerical S-system obtained is given by:  x = 1. . Let x0 ∈ (R∗ )n . Given a ﬁxed point x of Ψ.2] 0 0 1 2 Lemma 5.2).1 Correctness Let’s describe the ﬁrst iteration of our algorithm. We denote it x1 . y) = ˙ −y 3.5).266 5. n. . Let us assume that matrices G(x0 ) and H(x0 ) verify the condition: det(G − H) = 0. 1. ANALYSIS OF THE ALGORITHM 5.375 + x3 The numerical S-system obtained is given by:  x = 1. Using equation (3).500 y −1 − x ˙ y = 1.697 1. However we do not a priori know if all ﬁxed points of Ψ are indeed limits of (I). to + + each x0 ∈ (R∗ )n associates the point x1 . 4 y3 2 [2. [3]). . P 1 and P 3 are stable. Therefore: Ψ(¯) = x ⇐⇒ x ¯ ⇐⇒ ⇐⇒ ⇐⇒ b(G−H) = x ¯ b = x(G−H) ¯ n Y gij −hij βi ∀i = 1 .4. the curves f1 (x.2. . 1. Thanks to this assumption. In other words. y) = 0 (solid line).879 x−2. The numerical S-system obtained is given by:  x = 2.274 − y ˙ • With initial condition x0 = (2. y) = −x < ˙ 1 + y2 (9) > 6.647 x−0.75 : y = f2 (x. = xj ¯ αi j=1 n n Y h Y g ∀i = 1 .802 P1 ≈ P2 = P3 ≈ 1. Let’s consider the two dimensional dynamical system: 8 3 > x = f1 (x. The values of these points can be calculated: „ « „ « „ « 0.. The correctness of the algorithm (1) is a consequence of the two following lemmas: We can show that P 2 is unstable whereas P 1 and P 3 are stable (cf. we deﬁne the + quantities αi (x0 ). there exists a unique equilibrium point of the system (Sx0 ).e.n . algorithm ﬁnished in 4 iterations and found P 3 with a precision of 10−5 . the positive equilibria of (S) are the attracting ﬁxed points of Ψ. we have: x Ψ(¯) = b(G−H) x where b is the vector (βi /αi )i=1. The convergence appears to be fast since we need only 4 iterations to approach the equilibria with a precision of 10−5 . i. 2] P1 [. 0.2. this system has three equilibrium points. we found three diﬀerent initial conditions. gij (x0 ) and hij (x0 ). AN EXAMPLE WITH MULTIPLE POSITIVE EQUILIBRIA We present here the application of our algorithm for a dynamical system having multiple positive equilibrium points.535 − x ˙ y = 1. (First lemma) Let x ∈ (R∗ )n such that the de¯ + terminant of G(¯) − H(¯) is diﬀerent from zero. The equilibria of initial system (S) are the ﬁxed points of the function Ψ 4 x 5 3 Figure 1: The two curves represents the nullclines of system (9). The central equilibrium P 2 can be shown to be unstable while the two others. Applying our program in Maple.818 1. It is a system known as biological switch (see [3]). + Our algorithm computes the recurrent sequence:  0 x ∈ (R∗ )n + (I) xn+1 = Ψ(xn ) This iterative process converges towards ﬁxed points of Ψ.745 y −1. algorithm ﬁnished in 4 iterations and found P 1 with a precision of 10−5 . and we deﬁne the function Ψ : (R∗ )n → (R∗ )n that. • With initial condition x0 = (2.1. Lemma 5. and note G in ¯ place of G(¯)). there exists some ¯ initial points x0 that lead to x by the iteration (I). algorithm ﬁnished in 4 iterations and found P 2 with a precision of 10−5 .6 − y ˙ It represents the temporal evolution of two positive quantities x and y with linear decay and sigmoidal production (we Kn use here the Hill function H − (z) = n often used by K + zn biologists to model sigmoidal interactions). we must ﬁnd which ﬁxed points are attracting.2. Proof. With formulae (7) and (8). each of which tending towards one of the three equilibrium points (see ﬁgure 1 and numerical results below). .5] 1 P2 P3 [2.5 2.0 0.3.

the equilibria of (S) are the ﬁxed points of the function Ψ. We ¯ assume that det(G(¯) − H(¯)) = 0. j = 1 .. Gq . for i = 1 . αi n xj ij (resp. we produce a + sequence (xq )q∈N (with xq ∈ (R∗ )n ) that tends towards a + limit point x ∈ (R∗ )n which is an equilibrium of (11). xn Algorithm 1 ensures that. If we assume that V + and V − are at least C 1 . that it converges towards an equilibrium point of our initial system.j=1. xq = Ψ(q) (x0 ). for k = i: ! " n # n X X uj (x) ∂u−1 ∂ 2 u−1 ∂Ψi i i uj = exp − ∂xk ∂xj ∂xk xi ∂xj j=1 j=1 and # " n n X ∂ 2 u−1 ∂Ψi 1 X ∂u−1 i i × − = uj uj ∂xi ∂xj ∂xk xi j=1 ∂xj j=1 exp − j=1 n X uj (x) ∂u−1 i xi ∂xj ! As we have shown that the ﬁxed points of Ψ are the equilibria of (S). then x is the unique ﬁxed ¯ ¯ point of Ψ in W and it is attracting. Hq ) which comes from the S-approximations of functions V + and V − in xq . the iteration (I) converges towards x.n is the inverse of the matrix P = G−H. therefore: x JΨ (¯) = 0 x This implies that the spectral radius of JΨ (¯) is strictly x less than one. we get. Consider the n-dimensional dynamical system: ( dx = F (x) = V + (x) − V − (x) dt x ∈ (R∗ )n + (11) i. see e.2 Stability analysis of approximate S-system We have proven that our algorithm is correct. we can deduce that x is attracting. gij and hij are the functions deﬁned in (6). Let’s calculate pij : pij = gij „ hij − « + 1 ∂v − 1 ∂vi − − i = xj v + ∂xj vi ∂xj „ i+ « ∂ui ∂u− i = xj − ∂xj ∂xj ∂ui = xj ∂xj Therefore P −1 = ∆−1 (Ju (x))−1 = ∆−1 (Ju−1 (x)) (u−1 is the reciprocal function of u) and so: ∀i. at each step. for all x ∈ W : n Y − vj (x) + vj (x) Diﬀerentiating this (and omitting the dependency in x).n with: hq = hij (xq ) ij = 1 ∂u−1 i xi ∂xj where α. More e + precisely. it provides us with a S-system Sq (αq . . that is to say. (Second lemma) Let x be a ﬁxed point of Ψ.. i. the spectral radius of the ¯ jacobian of Ψ is strictly less than one. At the same time.Q Q g h By deﬁnition. The continuity of the x x determinant implies that there exists a neighboorhood W of x in which det(G − H) = 0. unless we fall in a degenerate case.e. β.1 im¯ plies then: Ψ(¯) = x x ¯ ⇐⇒ V + (¯) = V − (¯) x x With (10) we have. n and x ∈ W : Ψi (x) = xi „ « ∂u−1 1 exp − uj (x) i (x) xi ∂xj j=1 ! n X uj (x) ∂u−1 i = xi exp − (x) xi ∂xj j=1 n Y Thus. In order to prove the second lemma. βq . We show here that this sequence of systems is useful because it facilitates the study of the stability of the equilibrium. βi n xj ij ) is the Sj=1 ¯ j=1 ¯ approximation of V + (resp.j=1. . i. Ψi (x) = xi “ ” (−1) where pij j=1 !p(−1) (x) ij 5.n q hij i.j=1. we deduce that . V − ) in x. given any initial condition x0 in (R∗ )n . Using (7) and (8) and posing: 8 + + > U > − = log(V − ) > > U = log(V ) < „ +« V > U = U + − U − = log > > V− > : P = G−H we obtain. we deduce that ∀k = 1 .... it provides a sequence of S-systems that locally approximate the initial system. we will use the following ﬁxed point criterion (known as Ostrowski’s theorem. n. [12]): If the function Ψ is a contraction on the open set W and if x ∈ W is a ﬁxed point of Ψ. ∆=@ A .g. Proposition 3. we have: 8 > αq > < βq > Gq > : Hq = = = = α(xq ) β(xq ) ` q´ q with: gij = gij (xq ) g ` ij ´i.e.. Ψ is a contraction on an open neighboorhood of x. According to the aforementioned ﬁxed ¯ point criterion. Meanwhile. . Thus. n . ¯ Proof. This ¯ concludes the proof of the second lemma and the correctness of the algorithm. . uk (¯) = 0. Consequently. . To prove that x is attracting. for all x0 ∈ W .. ¯ ¯ we have to show that there exists an open neighbourhood of x in which Ψ is contracting. . pij (−1) in matricial notation: P = Ju (x)∆ where Ju (x) is the jacobian of the function U evaluated in x and ∆ is the diagonal matrix: 0 1 x1 B C .

e This proposition implies that the stability of x for system e (12) is the same that the stability of x for system (11). one can easily verify that the convergence of the iteration is in K n (where n is the number of iterations). we are in a case where the speed of convergence is the best possible.1. .3 Discussion about convergence The algorithm described above computes the iterations of a vectorial function Ψ on an initial point x0 ∈ (R∗ )n . we know that the convergence is very fast (up to four or ﬁve iterations in all the examples presented. in most biochemical pathways. 4]. in a biochemical pathway. for a simple reaction. Since JΨ (e) = 0. P 2 (white) and P 3 (grey).261 −2. n j=1 We want to know in which sense the system (12) approximates the system (11). 4]2 . for a precision of 10−4 or 10−5 ). each ODE represents the evolution of the concentration of a chemical species. We cannot a priori know in which of these basins the point x0 is. Then we colored the points of our grid with respect to the equilibrium towards which they lead (see ﬁg 2). in or+ der to converge towards a ﬁxed point of Ψ. An answer is given by the following proposition: Proposition 5. The dynamics of simple E substrate-product enzymatic reaction: S → P is given by the Michaelis law: kES ˙ S=− KM + S where k and KM are speciﬁc constants. e However. the algorithm’s behaviour is strongly dependent on the choice of initial point x0 . then the ﬂow generated by (12) is topologically conjugate to the ﬂow generated by (11) in a neighborhood of x. systems are obviously more complicated. As a matter of fact.e Let (S) be the following S-system: e (S) : xi = αi ˙ e j=1 n Y these sequences converge. then we can ﬁnd a neighborhood x of x wherein Ψ is K-contractant for any 0 < K < 1. we saw that the mass-action law gives. We even cannot ensure that x0 actually lies in one of them. Proof. 5. Moreover. The structure of S-systems seems therefore particularly relevant in the modeling of biochemical systems. . a dynamical system is topologically conjugate to its linearized system. e . To visualize the three basins.1 shows that systems (11) and (12) have the same linearized dynamical systems in x. We ﬁnd the equilibrium point x ≈ (1. According to proposition 5. As an example.2301.709 −2. if the function Ψ is K-contractant. The equilibrium x of (11) is an equilibrium of (12). The boundaries of such basins can be quite complicated. as equilibrium of e (Ex) is hence stable. as q tends to 8 > αq → α(e) def x = > > > def < βq → β(e) x = def > x > Gq → G(e) = > > : def Hq → H(e) = x e xj ij − βi g ˜ n Y ˜ h ∞. towards: α e e β e G e H (12) xj ij . In the example given above (equation (1)). APPLICATION TO BIOLOGY As we said in section 2.812 e e G−H = 0. the study of basins of attractions of such iterations is a complex issue. if initial system has multiple positive equilibria. S-systems are particularly appropriate to model biochemical pathways.6950) and the e e e matrix G − H: „ « −0. we launched our algorithm for the switch system (9) with initial conditions taken on a grid of ]0. Indeed.1. Proposition 3. even fractals [8]. i = 1 . [19]). Indeed. if x e e is hyperbolic.g. each of them has distinct basins of attraction. we associated a color to each equilibrium. 6. As the jacobian of Ψ is the null matrix in those ﬁxed points. Most of biochemical reactions depend on the concentration of appropriate enzymes that catalyse the reaction. The presence of such expressions implies rational functions in the equations. The ﬁrst assertion is obvious with proposition 3. even if the speed of convergence is very fast.1. monomial terms. we see that G− H is sign semi-stable. F is supposed C r (r ≥ 1). the point x. Let us prove the second assertion: it is a direct consequence of the Hartman-Grobman theorem (see e. Indeed. We obtained these graphs by applying algorithm 1 for system (9) with initial conditions taken in a regular grid of ]0. let us consider the following 2-dimensional dynamical system: 8 x x2 y 4 > x = > − < ˙ 2+y (3 + x)(4 + y 3 ) (Ex) > 5x 2xy 3 > y = : ˙ − 3+x (x + 1)(y + 2) Figure 2: Basins of attraction of points P 1 (dark).1. In fact. (11) and (12) are so topologically conjugate around x. It is composed of the diﬀerence of two positive terms (production and decay of the species). As e an exemple. According to Hartman-Grobman theorem. By transitivity of the e topological conjugation. Neverthelesss. 1.541 e e Thanks to theorem 2.1.

[7] A. [11] J. yet their simple mathematical expression allows symbolic manipulations. and B. REFERENCES [1] M. Chem. [4] G. this system exhibits a stationnary behaviour. Multistationarity.. Theor. Biol. [8] C. Page. 176:291–300. SIAM Review.W. e S.. 40:4119–4121. J. 1990 (reprint of 1970 ed. Springer Verlag. So far. [3] J.. and ii. Math. 11(1):30–51. structural conditions of multistationarity and other non-trivial behaviour. Wiggins. and J. . [19] S. 2001. Dumas. we use symbolic computation of partial derivatives).. Biochemistry. Grebogi and E. V. 26:6869–6880. We are currently working in collaboration with G. Nevertheless the technique of S-approximation (def 3. Introduction to applied nonlinear dynamical systems and chaos. Quirk. Goldbeter. i. 87:83–115. Giving realistic values of parameters and initial conditions. Thomas and M. S-approximation gives a computable and rather good approximation of ODE systems (see [18] for a comparison between power-law approximation and linearization). Math. J. Numerical Analysis: A second course. A. we have focused our study on a subsystem of 14 variables. Biochem.1) consists of symbolic manipulations (in particular.O. [15] M. [16] R. Plahte. 2002. and S. A kinetic model of the branch-point between the methionine and threonine biosynthesis pathways in arabidopsis thaliana.. 11:170–195. needs numerical estimations of symbolic parameters. 1987. Phys. Hernandez. A computer algebra approach to biological systems. N. 2000. Voit and M. Proceedings of the 2003 International Symposium on Symbolic and Algebraic Computation (ISSAC). [17] E. Laubenbacher. long-lived chaotic transients. and P. that introduce terms such as Hill functions or composition of Hill functions. 2003. Ugel. providing a practical framework of study. Phys. Mishra. HiPC 2002. Theor. Savageau. We now have to study the Sapproximation of the system near this equilibrium. Voit. Klee. a large part of research concerning the analysis of biological phenomena uses both symbolic and numerical techniques. How to make a biological switch. Tewﬁk. J. logical analysis of regulatory networks in terms of feedback circuits. [12] J. M. Ott. 50(13):935–938. The idea of S-approximation is therefore to propose a local approximation of the dynamics of a complex biochemical pathway in the shape of a S-system. the basis of cell diﬀerentiation and memory. it is not possible to give a global representation of such a system with a S-system. J. 7. 8. Qualitative simulation of genetic regulatory networks using piecewise-linear model.O. Glass. Proceedings of the International Conference on High Performance Computing. with different realistic sets of parameters. 29(2):315–326. Ortega. CONCLUSION AND PERSPECTIVES As we said in the introduction.Actually. This should provide a global approximation interpolating the system in some critical points in the phase space (see [14]). Chaos.A. Geiselmann. This work is in progress. Voit. J. Recasting nonlinear diﬀerential equations as s-systems : a canonical nonlinear form. Mestl. Combinatorial and topological methods is nonlinear chemical kinetics. as presented here. Universal canonical forms for time continuous dynamical systems. Goriely. Xs-systems : extended s-systems and algebraic diﬀerential automata for modeling cellular behavior. Adler. Fractal basin boundaries. biological enzymes may exhibit even more complex behaviours like cooperativity or allostery. and unstable-unstable pair bifurcation.. 66(2):301–340. C. Savageau and E. Maybee and J. Biol. Math. 2003.) [13] T. de Jong. Alternative designs for a genetic switch: analysis of switching times using the piecewise power-law representation.O. Biosci. Accuracy of alternative representations for integrated biochemical systems. we managed. 2002. Because of these terms. [14] M. 2004. 1976. Ravanel. J. Rev. It can be calculated in any point of the phase space and can include symbolic parameters. Van Den Driessche. Computational analysis of biochemical systems. 180:237–253. and R. Cambridge University Press. [9] C. Jeﬀries. Qualitative problems in matrix theory. Lett.M. SIAM. Biosci. [2] L. 270(23):4615–4627. 1995.. Biochemical oscillations and cellular rhythms. 1969. E. 1990. thanks to our algorithm. In vivo. Cherry and F. Antoniotti. A very interesting idea is therefore to use the context information (given for instance by biologists) of a particular system in order to create a piecewise S-approximation of this system. The goal of this work is to understand the metabolic system responsible for the synthesis of aminoacids in Arabidopsis Thaliana. Omholt. [6] L. with 9 symbolic parameters. 1975. 1996. Rev.L. 2000. Savageau. 203:117–133..R. 1989. Curien. Curien (see [4]) on a speciﬁc example. Algorithm 1. Gouz´. 1983. A mathematical framework for describing and analyzing gene regulatory networks. Brenig and A. pages 431–442. Math. The S-systems as we described represent a large class of systems. 1987. 63. A. Bull. Phys. S.A. Eur.A. Policriti. [10] R. [5] H. Biol. Cambridge University Press. [18] E..-L. When is a matrix sign stable ? Can. Kaufman. to ﬁnd this positive equilibrium.