Reaction-Diffusion Models in Mathematical Biology Kristian Kristiansen Kongens Lyngby 2008 Technical University of Denmark The Department of Mathematics Building 303 S, DK-2800 Kongens Lyngby, Denmark Phone +45 4525 3031, Fax +45 4588 1399 [email protected] www.mat.dtu.dk Summary In this master thesis the well-posedness of two-component reaction-diffusion equations with Neumann conditions is analysed together with a study of diffusion driven instabilities. Specifically, three models from mathematical biology is considered: The Schnakenberg model, the Gierer and Meinhardt model and the Thomas model. The techniques of upper and lower solutions described by Pao, [Pao (1992)], and abstract results on semigroups together with basic functional analysis are applied to these models in order to show well-posedness. All systems are shown to be well-posed with non-negative and bounded solutions. Linear stability is shown to imply stability of steady-states. This is exploited to analyse diffusion-driven instabilities. Sufficient conditions for the general, two-component, autonomous reaction-diffusion equation to exhibit diffusiondriven instabilities at a uniform steady-state are presented. Diffusion-driven instabilities at a uniform steady-state are analysed numerically for the Schnakenberg model. The instabilities are shown to evolve into steady-state, heterogeneous patterns. Additionally, it is observed, and proved, that there curves exist in parameter-space for which the corresponding systems are equivalent in the sense that the resulting steady-states only differ by a scaling. ii Resumé I dette eksamensprojekt studeres velstillethed og diffusionsdrevet instabilitet for to-komponente reaktion-diffusions ligninger med Neumann-betingelser. Mere specifikt betragtes tre modeller fra matematisk biology: Schnakenberg modellen, Gierer og Meinhardt modellen og Thomas modellen. Øvre- og nedre-løsningernes metode, beskrevet af Pao, [Pao (1992)], og abstrakte resultater fra semigruppe-teori sammen med simple funktional analyse benyttes til at undersøge velstilletheden af disse modeller. Alle modeller vises at være velstillede med positive og begrænsede løsninger. Lineær stabilitet vises at medføre stabilitet af stationære løsninger. Dette udnyttes til at analysere diffusionsdrevet instabiliteter. Tilstrækkelige betingelser præsenteres for at den generelle to-komponent, autonome reaktion-diffusions ligning kan fremvise diffusionsdrevet instabilitet omkring en uniform stationær løsning. Diffusionsdrevet instabiliteter for en uniform stationær løsning analyseres numerisk for Schnakenberg modellen. Instabiliteterne vises at udvikle sig til stationære, heterogene løsninger. Ydermere observeres og bevises det, at der eksisterer kurver i parameterdomænet således, at de forskellige systemer er ekvivalente i den forstand at resulterende stationære løsninger kun afviger med en skalering. iv Preface This thesis was prepared at Department of Mathematics, the Technical University of Denmark in partial fulfillment of the requirements for acquiring the Master degree in engineering. The thesis deals with reaction-diffusion models in mathematical biology. These models are coupled semilinear partial differential equations and the wellposedness of such is a complicated matter. Among other things, it is the multiple steady-states that makes these models interesting mathematically, as well as practically relevant. The aim of the thesis, is first of all to apply two techniques to show wellposedness of the Schnakenberg model, the Thomas model and the Gierer and Meinhardt model. Next, we wish to justify the use of linear analysis in this infinite-dimensional case, and apply this to the Schnakenberg model and show the existence of diffusion driven instabilities at a uniform steady-state. Finally, we shall apply numerical studies and analyse the existence of heterogeneous patterns, and study the bifurcation of such heterogeneous patterns. Lyngby, January 2008 Kristian Kristiansen vi Acknowledgements I would like to thank my supervisor Michael Pedersen for support and help during the project. Thank you for taking your time with me, and thank you for your clever guidance. Furthermore, I would like to thank my master student colleagues at the institute, Jacob Hersböll and Karen Markvard Martensen, for keeping my head high and for coping with my suddenly outbursts of celebration or confusion. Thanks to Mads Peter Sørensen, Per Grove Thomsen and Casper Skovby for MATLAB assistance. I also owe Casper Skovby and Peder Skafte a great favour for their LATEX-support. Thanks to my family, especially my mother and Anna, for always respecting my work and supporting me. Anna, also thanks for your fantastic proofreading. Thank you all viii Notation By N we shall denote the set of positive integers and by N0 the set of nonnegative integers. R denotes the real numbers, R+ and R− the positive resp. negative real numbers. By C we denote the complex numbers. Rn is the n-dimensional real Euclidean space. The open balls in Rn with centre at x and radius are denoted Br (x). d Differentiation of functions on R is indicated by ∂x = dx and similarly dif∂ n ferentiation on R is indicated by ∂xj = ∂xj . We shall also make use of the multi-index notation: When α ∈ Nn0 , α = (α1 , · · · , αn ) then ∂ α = ∂xα11 · · · ∂xαnn , and by |α| we shall denote the sum α1 + · · · + αn . We shall say that a set Ω ⊂ Rn is smooth if for each point x on the boundary ∂Ω there is a ball Br (x), r > 0, and integer i ∈ {1, . . . , n} such that ∂Ω ∩ Br (x) = {x ∈ Br (x) | xi = φ(x1 , . . . , xi−1 , xi−1 , . . . , xn )} , (1) and φ is smooth, that is infinitely many times differentiable. The set of m-times continuously differentiable functions on Ω is denoted by C m (Ω) while those with compact support is denoted by C0m (Ω). We shall denote the vector space of smooth functions by C ∞ (Ω) = ∩m∈N C m (Ω) and those with compact support C0∞ (Ω) = ∩m∈N C0m (Ω), also denoted D(Ω) and referred to as the set of test functions. P If we equip C m (Ω) with the supremum norm kukm,∞ = supΩ |α|≤m |∂ α u| we obtain a Banach space which we shall denote CLm∞ (Ω). We shall also make use of functions on product spaces Ω×(0, T ) for example. Let p, q ≥ 0 then by C p,q (Ω × (0, T )) we denote the functions on Ω × (0, T ) that x are p-times continuously differentiable for all x ∈ Ω and q-times continuously differentiable for all t ∈ (0, T ). For 1 ≤ p < ∞ we define on a measurable set S the space Lp (S) as the completion of C0∞ (S) in the norm Z 1/p p kukp = . |u| S identifying functions that agree almost everywhere. The Lp -spaces are Banach spaces; in particular L2 (S) is a Hilbert space with the inner product Z (u, v) = uv. S For 1 ≤ p < ∞ and n ∈ N the Sobolev spaces W n,p (Ω) are defined as W n,p = {u ∈ Lp (Ω)| ∂ α u ∈ Lp (Ω) for 0 ≤ |α| ≤ n where ∂ α u is the distributional partial derivative} , equipped with the Banach norm kukn,p = X 0≤|α|≤n k∂ α ukpp 1/p . (2) For p = 2 we write H n = W n,2 , and then the norm (2) stems from the inner product X (∂ α u, ∂ α v). (u, v)n = 0≤|α|≤n A definition of W s,p (Ω) for s ∈ R is more complicated, and the reader is referred to [Adams (1975)]. We define the Neumann operator ν by ψ 7→ (n · ∇)ψ, where n is the outward unit normal of Ω given by n= [−∂ α1 φ(x), · · · , −∂ αi−1 φ(x), 1, −∂ αi+1 φ(x), · · · , −∂ αn−1 φ(x)] p 1 + ∂ α1 φ(x)2 + · · · + ∂ αn−1 φ(x)2 and φ is the function from (1) and αj = (0, · · · , 1, · · · , 0) with 1 on the j’th coordinate and zero elsewhere. Working with systems of partial differential equation we shall often make use of the product topology. Let B be a Banach space, then the product space B × B is a Banach space with the product norm k(u, v)kB 2 = kukB + kvkB , for all (u, v) ∈ B 2 . Finally, tr and det denote trace resp. determinant. xi xii Contents Contents Summary i Resumé iii Preface v Acknowledgements Notation vii ix 1 Introduction 1 2 The approach of Pao 5 2.1 Upper and lower solutions . . . . . . . . . . . . . . . . . . . . . . 5 2.2 Applications of the approach of upper and lower solutions . . . . 8 2.3 Summary and remarks . . . . . . . . . . . . . . . . . . . . . . . . 13 xiv CONTENTS 3 Existence and uniqueness of two-component reaction-diffusion system 15 3.1 Local existence and uniqueness . . . . . . . . . . . . . . . . . . . 16 3.2 Application of result on local existence to specific models . . . . 27 3.3 Global existence and boundedness . . . . . . . . . . . . . . . . . 32 3.4 Summary and remarks . . . . . . . . . . . . . . . . . . . . . . . . 46 4 Linear analysis 49 4.1 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 4.2 Conditions for diffusion driven instability . . . . . . . . . . . . . 57 4.3 Summary and remarks . . . . . . . . . . . . . . . . . . . . . . . . 61 5 Numerical analysis 63 5.1 Results in Matlab . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 5.2 Summary and remarks . . . . . . . . . . . . . . . . . . . . . . . . 100 6 Conclusion 103 A Semigroups and sectorial operators 107 B Differentiation and integration in Banach Spaces 115 C Gronwall’s inequality 117 D Matlab-codes 119 D.1 main.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 D.2 cheb.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 CONTENTS xv D.3 operator.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 D.4 initcond.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 D.5 solver.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 D.6 recmovie.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 D.7 func2d.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 xvi CONTENTS Chapter 1 Introduction Reaction-diffusion equations have enjoyed a considerable amount of scientific interest. The reason for the large amount of work put into studying these equations is not only their practical relevance, but also interesting phenomena that can arise from such equations, such as multiple steady states and spatial patterns and oscillating solutions, just to mention a few. The study of these phenomena require a variety of different methods from many areas of mathematics for example bifurcation and stability theory, semigroup theory, singular perturbations, numerical analysis and many others. From a qualitatively point of view, a reaction-diffusion system is a mathematical model describing how the concentration of one or more substances vary over time and space under the influence of two terms: Reaction term or source term, in which concentration is generated or degenerated by local interaction, diffusion term which causes the substances to spread out in space. A reactiondiffusion system is therefore an equation heuristically like Change in concentration = Diffusion of concentration + Source term, or algebraically ∂t u = D∆u + f(u), in Ω, where Ω ⊂ Rn , together with some appropriate boundary conditions and initial conditions. Ω may be bounded or unbounded. 2 Introduction More specifically we shall in this work consider two-component reactiondiffusion systems on a bounded domain Ω with homogeneous Neumann conditions ∂t u − ∆u = f (t, u, v), in Ω × (0, T ), ∂t v − d∆v = g(t, u, v), u ν = 0, on ∂Ω × [0, T ) v u = u0 , v = v0 , on Ω × {t = 0} . (1.1) (1.2) (1.3) Other linear boundary conditions, such as Direchlet conditions, are also of interest in some cases, but in this thesis we shall settle with the Neumann condition. The system is called autonomous, if the reaction functions f and g do not depend upon time. Examples of autonomous reaction functions are given by Murray, [Murray (2002), Chapter 2], f (r, s) = a − r + r2 s, g(r, s) = b − r2 s, r2 , s(1 + kr2 ) ρrs f (r, s) = a − r − , 1 + r + kr2 f (r, s) = a − br + 2 g(r, s) = r2 − s, g(r, s) = a(b − s) − ρrs , 1 + r + kr2 2 for (r, s) ∈ R+ , R2+ resp. R+ . The system with the first type of reaction functions ∂t u − ∆u = a − u + u2 v in Ω × (0, ∞), ∂t v − d∆v = b − u2 v u ν = 0, on ∂Ω × [0, ∞), v u = u0 , v = v0 , on Ω × {t = 0} , (1.4) (1.5) (1.6) is known as the Schnakenberg model and it shall be of our main interest in this work. The systems with the second and third type of reaction-diffusion equations are known as the Gierer and Meinhardt model resp. Thomas model. Although these equations are non-linear and therefore troublesome to work with they all enjoy the feature of involving a nice parabolic differential operator. Furthermore the nonlinearities are on the dependent variable itself and not on its partial derivatives. A vast selection of tools are available to account for such nonlinearities. More specifically, the equations are said to be semilinear.1 1 Formally a partial differential equation of k’th order is semilinear if the equation is linear in the k’th order term. 3 In this thesis we shall be concerned with the techniques described by Pao, [Pao (1992)], on upper and lower solutions and more abstract results from semigroup theory and functional analysis utilised in [Henry (1981)] and [Hollis, S. L.; Martin, R. H.; Pierre, M (1987)]. We wish to exploit these tools to analyse the general two-component reaction-diffusion system as well as the specific models above, in terms of well-posedness. A problem is well-posed in the sense of Hadamard if the following conditions are satisfied: (W1) Existence and uniqueness; (W2) Existence for all times; (W3) Continuously dependency on initial conditions. These conditions are clearly not sufficient for a physical or biological model to be “practically well-posed.” When working with physical parameters such as concentrations, we cannot allow the solution to become negative. Furthermore the solution must be bounded. We shall therefore also be interested in the sufficient conditions for the system to satisfy: (W4) The solution is non-negative for non-negative initial data; (W5) The solution is bounded for all bounded initial data. In the following chapter we shall apply the techniques of upper and lower solutions to the specific systems. We have chosen this approach, since it applies an ad hoc approach to the subject of well-posedness of coupled, parabolic equations, and we wish to analyse to what extend this method is sufficient. The technique is, however, not adequate to analyse (W3). Next, we plan to apply the theory of semigroups and functional analysis to show the well-posedness, (W1)-(W5), for the general system (1.1)-(1.3), and then in turn apply these to the specific systems above. We shall do so by building on the work of Hollis, S. L.; Martin, R. H.; Pierre, M., [Hollis, Martin, Pierre (1987)]. Then, we wish to justify the use of a linear analysis, and apply this to analyse the ideas about diffusion-driven instabilities at a uniform steady-state. Specifically, we plan to apply these results to the Schnakenberg model. Thereafter, we aim to perform a numerical analysis of the Schnakenberg model. Our objective with the numerical analysis, is first of all to solve the system numerically and verify the well-posedness, and next to study the diffusiondriven instabilities at uniform steady-states (u∗ , v ∗ ). Moreover, by varying the size of the domain we wish to analyse the stability 4 Introduction and bifurcation of the steady-states. Finally, we want to analyse how the solutions of the Schnakenberg system are affected by the choice of (a, b, d). For example, we desire to study the possibility of obtaining similar patterns for different (a, b, d)-values. Chapter 2 The approach of Pao In this chapter the techniques described by Pao, [Pao (1992)], are presented and applied to the specific autonomous reaction-diffusion models stated in Chap. 1 in order to show existence, uniqueness, boundedness and positivity. 2.1 Upper and lower solutions In Pao, [Pao (1992)], f and g are assumed to be continuously differentiable and to satisfy either of the three, (1) ∂s f (r, s) ≥ 0, ∂r g(r, s) ≥ 0; quasimonotone nondecreasing functions; (2) ∂s f (r, s) ≤ 0, ∂r g(r, s) ≤ 0; quasimonotone nonincreasing functions; (3) ∂s f (r, s) ≥ 0, ∂r g(r, s) ≤ 0; mixed quasimonotone functions. For u0 , v0 ∈ CL0 ∞ (Ω) one can for the three type of functions define ordered upper and lower solutions of (1.1)-(1.3). Here we shall only be interested in the case of mixed quasimontone functions in which case the definition is: 6 The approach of Pao Definition 1 U = (u, v) and U = (u, v) in ordered upper resp. lower solutions if U ≥ U, 2 CL2,1 are called ∞ (Ω × [0, T )) ∂t u − ∆u ≥ f (u, v), ∂t u − ∆u ≤ f (u, v), ∂t v − d∆v ≥ g(u, v), ∂t v − d∆v ≤ g(u, v), (2.1) (2.2) νu ≥ 0 ≥ νu, νv ≥ 0 ≥ νv, (2.3) furthermore, and finally u(x, 0) ≥ u0 ≥ u(x, 0) v(x, 0) ≥ v0 ≥ v(x, 0) x ∈ Ω. (2.4) 2 The adequateness of upper and lower solutions follows from the following theorem. Theorem 1 (See [Pao (1992), Theorem 8.3.3]) Let U = (u, v) and U = (u, v) be a pair of ordered upper and lower solutions and let f and g be mixed quasimonotone. Then there exists a unique solution U = (u, v), and it is in the sector 2 2,1 hU, Ui = Z ∈ CL∞ (Ω × [0, T )) | U ≤ Z ≤ U . 2 Proof (Sketched) The result is established without the use of distributions and semigroups and even with a small amount of functional analysis. Instead (n) the result is established by considering the sequences {U }n∈N and {Un }n∈N defined through U1 = U, U1 = U and (∂t − ∆ + c1 ) un = f˜(un−1 , v n−1 ), (∂t − d∆ + c2 ) v n = g̃(un−1 , v n−1 ), (∂t − ∆ + c1 ) un = f˜(un−1 , v n−1 ), (2.5) (∂t − d∆ + c2 ) v n = g̃(un−1 , v n−1 ),(2.6) for n ≥ 2 where f˜(r, s) = f (r, s) + c1 r, g̃(r, s) = g(r, s) + c2 s, and c1 = inf{∂r f (r, s) | inf u ≤ r ≤ sup u}, Ω Ω c2 = inf{∂s g(r, s) | inf v ≤ s ≤ sup v}. Ω Ω 2.1 Upper and lower solutions 7 together with boundary and initial conditions as in (1.2) resp. (1.3). We claim that the sequences {Un }n∈N and {Un }n∈N possess the monotone property un ≤ un+1 ≤ un+1 ≤ un , v n ≤ v n+1 ≤ v n+1 ≤ v n for all x ∈ Ω for all n ∈ N. To show this, one applies the positivity lemma of parabolic operators, see [Pao (1992), Lemma 2.2.1]. Let w1 = u − u2 , w2 = v − v 2 . Then by (2.5) and (2.6) and (2.1) and (2.2): (∂t − ∆ + c1 ) w1 = ∂t u − ∆u − f (u, v) ≥ 0, (∂t − d∆ + c2 ) w2 = ∂t v − d∆v − g(u, v) ≥ 0, and w1 and w2 satisfy the boundary and initial conditions: νw1 = νu ≥ 0, w1 (x, 0) = u(x, 0) − u0 ≥ 0, νw2 = νv ≥ 0, w2 (x, 0) = v(x, 0) − u0 ≥ 0. By the positivity lemma of parabolic operators it follows that w1 and w2 ≥ 0 and hence u2 ≤ u and v 2 ≤ v. Similarly we obtain u2 ≥ u and v 2 ≥ v. The proof then goes by induction, again exploiting the positivity lemma. The existence of the pointwise limits: (u∞ , u∞ )(x) ≡ lim (un , un )(x), n→∞ and (v ∞ , v ∞ )(x) ≡ lim (v n , v n )(x), n→∞ follows from the monotone properties just proved, or at least sketched. However, it is by no means clear that these limits are solutions of the original system (1.1)(1.3) and that u∞ = u∞ and v ∞ = v ∞ . Despite this, it is shown in [Pao (1992), sec. 8.3] that a sufficient condition is the existence of k1 and k2 such that, f (r2 , s) − f (r1 , s) ≤ k1 (r2 − r1 ), for g(r, s2 ) − g(r, s1 ) ≤ k2 (s2 − s1 ), for inf u ≤ r1 ≤ r2 ≤ sup u, Ω Ω inf u ≤ r ≤ sup u, Ω Ω inf v ≤ s ≤ sup v, Ω Ω inf v ≤ s1 ≤ s2 ≤ sup v. Ω Ω This is indeed the case when (r, s) 7→ f (r, s) and (r, s) 7→ g(r, s) are differentiable for inf u ≤ r ≤ sup u, inf v ≤ s ≤ sup v, Ω Ω Ω by the mean-value theorem. Uniqueness is proved by contradiction. Ω In the following section it is the aim to prove existence and uniqueness of the reaction-diffusion systems presented in Chap. 1 by exploiting Theorem 1. 8 2.2 The approach of Pao Applications of the approach of upper and lower solutions First the Schnakenberg model (1.4)-(1.6) is considered for non-negative initial conditions. The Schnakenberg system’s reaction functions are quasimonotone: ∂s f (r, s) = r2 ≥ 0, ∂r g(r, s) = −2rs ≤ 0, for r, s ≥ 0. In order to utilise Theorem 1 we have to find a pair of ordered upper and lower solutions in accordance to Definition 1. To do so we choose v = 0. Then u has to solve the inequalities: ∂t u − ∆u ≤ f (u, 0) = a − u, νu ≤ 0, u ≤ u0 , on ∂Ω × (0, ∞). By inspection it is seen that u = m ≡ min(a, inf Ω u0 ) is a solution. We now use these lower solutions to obtain upper solutions. For v: ∂t v − ∆v ≥ g(u, v) = b − m2 v, (2.7) and boundary and initial conditions as in (2.3) resp. (2.4). We are guided to look for solutions independent of the spatial variable; the solution to the linear ordinary differential equation: ∂t y = b − m2 y, y(0) = sup v0 Ω satisfies (2.7), (2.3) and (2.4), and y is therefore an upper solution. The solution is y = mb2 + exp(−m2 t) supΩ v0 − mb2 . So y ≤ M ≡ max mb2 , supΩ v0 and then we may choose v = M . Now, u has to solve: ∂t u − ∆u ≥ f (u, v) = a − u + u2 v, (2.8) and again with boundary and initial conditions as in (2.3) resp. (2.4). Our only hope is to solve the following ordinary, nonlinear differential equation: ∂t z = a − z + z 2 M, z0 = sup u0 . Ω 2.2 Applications of the approach of upper and lower solutions 9 Three scenarios can occur. 1. M a > 41 . The solution is: 1 1 1 + α tan z(t) = αt + arctan β , 2M 2 where 1/2 α = (4M a − 1) 2M z0 − 1 β= . α , u = z satisfies (2.8) and by definition z0 ≥ u0 in Ω so it follows that u = z is an upper solution for t ∈ [0, T ), where T is given by π 1 αT + arctan β = , 2 2 T > 0, that is, T = 2 π − arctan β > 0. α 2 2. M a = 41 . The solution is: z(t) = 2a + 4a(z0 − 2a) , γ(2a − z0 )t + 4a u = z is an upper solution for t ∈ [0, T ), T > 0. If z0 > 2a then T is given by γ(2a − z0 )T + 4a = 0, T > 0. that is, 4a > 0. γ(z0 − 2a) T = If z0 < 2a then z is increasing monotonically and z → 2a as t → ∞ and hence we may take T = ∞. For z0 = 2a, z(t) = 2a for all t > 0 and hence T = ∞. 3. 0 < M a < 14 . The solution is: 1 1 z(t) = 1 − α̃ tanh α̃t + arctanh β̃ , 2M 2 1/2 α̃ = (1 − 4M a) 1 − 2M z0 . β̃ = α̃ , 10 The approach of Pao Im z(t) = 0 if and only if |β̃| < 1, or: √ √ 1 + 1 − 4M a 1 − 1 − 4M a < sup u0 < 2M 2M Ω (2.9) in which case we may take T = ∞. However, if z0 is outside this bound, then no real solution exists. Unless M a < 14 and u0 does not satisfy (2.9), then by the analysis above, Theorem 1 provides local existence and uniqueness. Furthermore, the solution is in the sector h(u, v) , (u, v)i, and therefore positive. If M a < 41 and (2.9) is satisfied, or if M = a4 and supΩ u0 ≤ 2a, then the pair of ordered upper and lower solutions exists for all t ≥ 0 and hence, again by Theorem 1, there exists a unique, global solution. Similarly the stationary equations, −∆u = a − u + u2 v, −d∆v = b − u2 v, (2.10) together with the homogeneous Neumann conditions on ∂Ω, can be analysed. Despite the fact that this problem is elliptic rather than then parabolic, the idea of theorem Theorem 1 is still applicable: If there exists (u, v) and (u, v) satisfying u ≤ u, v ≤ v and the inequalities, −∆u ≤ a − u + u2 v, −∆u ≥ a − u + u2 v, −d∆v ≤ b − u2 v, −∆v ≥ b − u2 v, (2.11) together with the boundary inequalities (2.3), then there exists a unique solution in the sector h(u, v), (u, v)i, see [Pao (1992), Theorem 5.2]. There may exist a solution outside this sector. The usefullness of this model is partly due to the existence of multiple steadystate solutions. In turn, this is also what makes these models challenging from a mathematical perspective. We shall in the following try to give sufficient conditions for the model to exhibit a unique steady-state solution. First, by solving: 0 = a − u + u2 v, 0 = b − u2 v, we realise that the Schnakenberg model has one single positive, homogeneous steady-state solution given by u∗ = a + b, v∗ = b 2, (a + b) b > 0, a + b > 0. We shall give sufficient conditions for this solution to be the unique solution of the steady-state problem (2.10). To do so we have to find a set of ordered upper 2.2 Applications of the approach of upper and lower solutions 11 and lower solutions satisfying the inequalities in (2.11). Guided by previous experience, we look for solutions independent of the spatial variable and take v = 0, u = a and v = ab2 . It is not difficult to see that these in fact solve the inequalities. For u: 0 ≥ a − u + u2 b . a2 0 = a − y + y2 b , a2 (2.12) The equation: has solutions: y± = a2 ± a2 q 1 − 4 ab 2b . Thus, for 0< 1 b ≤ a 4 y+ and y− are real and positive and we may take u = if (2.13) √ b a2 +a2 1−4 a 2b . Now, u ≥ u∗ a2 ≥ a + b, 2b or if a≥ However, 4 > √ 1+ 12 2 1+ √ 12 b. 2 and thus, as expected u > u∗ > u. Therefore: Corollary 1 For 0 ≤ ab ≤ 14 , (u∗ , v ∗ ) is the unique steady-state solution of the Schnakenberg system in the sector h(a, 0), (b/a2 , u)i. 2 The result cannot be improved since increasing v only decreases u. As mentioned other appropriate reaction functions are: r2 , g(r, s) = r2 − s, r, s > 0 s(1 + kr2 ) ρrs ρrs , g(r, s) = a(b − s) − , f (r, s) = a − r − 1 + r + kr2 1 + r + kr2 f (r, s) = a − br + r, s ≥ 0. 12 The approach of Pao Again all parameters are assumed positive. For the first model, the Gierer and Meinhardt model, the partial derivatives: ∂s f (r, s) = −r2 /(s2 (1 + kr2 )) ≤ 0, ∂r g(r, s) = 2r ≥ 0, show that the reaction functions are mixed quasimonotone on R2+ . We only study positive initial conditions. For u: ∂t y = a − by, y(0) = inf u0 . Ω a b , inf Ω The solution is y(t) ≥ m1 = min Similarly for v: u0 > 0, so we may choose u = m1 . ∂t z = m21 − z, z0 = inf v0 , Ω to conclude that z ≥ m2 = min(m21 , inf Ω v0 ), so we put v = m2 . 1 x2 Now, for u we realise that m2 (1+kx 2 ) → m k from below as x → 2 we consider: ∂t w = a − bw + ∞, and then w2 1 ≥ a − bw + , m2 k m2 (1 + kw2 ) w(0) = sup u0 . Ω It follows that w ≤ M1 = max u = M1 . Finally, for v we consider: a b + 1 bm2 k , supΩ u0 , and hence we may take ∂t q = M12 − q, q(0) = sup v0 , Ω and take v = M2 = max(M12 , supΩ v0 ). It is clear that m1 < M1 , and then m2 < M2 . So by Theorem 1 we have: Corollary 2 There exists a unique solution of the Gierer-Meinhardt system for all positive initial conditions and parameters. The solution is in the sector hU, U)i = h(m1 , M1 ), (m2 , M2 )i, and therefore it is bounded an positive. 2 The Thomas model is an example of a system with reaction functions that are not suitable for the approach of upper and lower solutions, since the reaction 2.3 Summary and remarks 13 functions are not monotone: ρr , ∂s f (r, s) = − 1 + r + ks2 ∂r g(r, s) = ρ s kr2 − 1 (1 + r + kr2 ) 2, r, s ≥ 0; clearly ∂r g changes sign. 2.3 Summary and remarks In this chapter we presented the techniques described by Pao, [Pao (1992)]. The techniques were applied to the reaction-diffusion models in order to prove existence, uniqueness, positivity and boundedness. The continuously dependency on initial data was not considered as the method is not adequate for such analysis. For the Gierer and Meinhardt system we obtained global existence and uniqueness of a positive solution for all parameters and positive initial conditions. Furthermore, we showed that the solutions were bounded. For the Schnakenberg system it was only possible to obtain local and global existence for some specific parameters and initial conditions. However, sufficient conditions were given in order for the stationary problem to display a unique steady state in a sector. It was not possible to study the Thomas model by the methods described by Pao, since the reaction-functions were not quasimonotone. For this type of reaction functions we need to rely on a different setting. Similarly, if we want to show global existence for the first Schnakenberg model for all positive parameters, we also need to rely on a different sitting. In the forthcoming chapter we shall therefore apply functional analysis and results on semigroups to the general class of two-component reaction-diffusion equations in order to prove well-posedness under certain appropriate conditions. 14 The approach of Pao Chapter 3 Existence and uniqueness of two-component reaction-diffusion system In this chapter we shall analyse the general reaction-diffusion system: ∂t u − ∆u = f (t, u, v), in Ω × (0, T ), ∂t v − d∆v = g(t, u, v), u ν = 0, on ∂Ω × [0, T ), v u = u0 , v = v0 , on Ω × {t = 0} , (3.1) (3.2) (3.3) on a smooth and bounded domain Ω, in terms of the well-posedness (W1)-(W5). To do so, we apply semigroup theory and basic functional analysis. The theory of semigroups is set forth in App. A, but for convenience we shall just recap the most important results. The Neumann realisation of the Laplacian L = −∆, D(L) = {u ∈ W 2,p | νu = 0}, is sectorial, see Definition 6 in App. A, and therefore it generates an analytical semigroup G. For α ≥ 0 we define the fractional powers of the Helmholtz operator, H = −∆ + I, and denote these Hα . The domain of the fractional powers of H, D(Hα ), equipped with graph norm k · kD(Hα ) = k · kp + kHα · kp , is continuously 16 Existence and uniqueness of two-component reaction-diffusion system embedded in CL0 ∞ (Ω) if α > n/(2p), see Lemma 9 in App. A.1 For L we obtain the following corollary to Lemma 10 in App. A. Corollary 3 Let G be the analytic semigroup generated by −L. The following properties hold for the semigroup G and the fractional powers of the Helmholtz operator, H, 1◦ G(t) : Lp (Ω) → D(Hα ) for all t > 0; 2◦ kG(t)ukD(Hα ) ≤ Cα,p t−α kukp for all t > 0, u ∈ Lp (Ω); 3◦ G(t)Hα u = Hα G(t)u for all t > 0, u ∈ D(Hα ). 3.1 2 Local existence and uniqueness The following basic hypotheses are assumed to hold: (H1) d > 0; (H2) u0 ≥ 0 and v0 ≥ 0 are continuous on Ω; u0 , v0 ∈ CL0 ∞ (Ω); 3 (H3) f and g are continuously differentiable functions from R+ into R with f (t, 0, s) ≥ 0 and g(t, r, 0) ≥ 0 for all t, r, s ≥ 0; 2 (H4) There exist m > 0 and a continuous function F : R+ → R+ such that f (t, r, s), g(t, r, s) ≤ exp(mt)F (r, s) for all t, r, s ≥ 0. Remark 1 Compared to [Hollis, Martin, Pierre (1987)] we have added (H4) to 2 archieve local existence for (u0 , v0 ) ∈ CL0 ∞ . 2 First we prove positivity. Proposition 1 Suppose that (H1)-(H4) are satisfied. Then the classical solution, when it exists, is non-negative for all t in its interval of existence, [0, T ]. I.e. the cone K + = {(u, v) ∈ C 2,1 (Ω × (0, T )) | u ≥ 0, v ≥ 0} is an invariant set: (u0 , v0 ) ∈ K + ⇒ (u(t), v(t)) ∈ K + for all t > 0. 2 1 Since 0 is in the resolvent set of ρ(Hα ) the graph-norm is equivalent to kHα · k . However, p we shall use the graph-norm throughout. 3.1 Local existence and uniqueness 17 Proof First, we assume u0 , v0 > 0 for all x ∈ Ω. Furthermore, we let u > 0 and v > 0 for every t ∈ [0, τ ) and all x ∈ Ω and either u(x, τ ) = 0 for x in non-empty set M ⊂ Ω while v(x, τ ) > 0 for all x ∈ Ω, or v(x, τ ) = 0 for x in non-empty set M ⊂ Ω while u(x, τ ) > 0 for all x ∈ Ω. If such a τ < T does not exist, then the proof is done. Now assume that it does and without loss of generality that u(x, τ ) = 0 while v(x, τ ) > 0. Let x0 ∈ M . It is now claimed that x0 ∈ / ∂Ω. This follows by the maximum principle, see [Pao (1992), Theorem 2.1.4], which states that if x0 ∈ ∂Ω is a minimum then νu(x0 ) < 0, which contradicts u being a solution. Therefore x0 ∈ / Ω. x0 is therefore an interior minimum point of u, hence ∇u(x0 , τ ) = 0, H = (∂ij u(x0 , τ )) ≥ 0, where the second inequality states that the Hessian is semi-positive definite at x0 . Thus, X −∆u(x0 , τ ) = −tr H = − λk ≤ 0 k and then by (3.1) it follows that, ∂t u(x0 , τ, x0 ) ≥ 0. ∂t u(x0 , τ ) > 0 would imply by continuity that u(x0 , t) < 0 for t ∈ (τ − ǫ, τ ) for some ǫ > 0, which contradicts the assumption made about u for t < τ . Therefore ∂t u(τ, x0 ) = 0. Similarly, if u0 (x) = 0 or v0 (x) = 0 for x ∈ M ⊂ Ω it follows ∂t u(t, x0 ) = 0, t > 0. This finishes the proof. Remark 2 The Schnakenberg and the Thomas systems’ reaction functions satisfy (H3) − (H4). On the other hand, the Gierer and Meinhardt system’s reaction functions are not continuous in (t, r, s) = (0, 0, 0). However, by the results in the previous chapter or similarly by exploiting Corollary 4, see Sec. 3.2, we have (u, v) ≥ (m1 , m2 ) > (0, 0), where m1 = min ab , inf Ω u0 > 0 and m2 = min(m21 , inf Ω v0 ) > 0, whenever inf Ω (u0 , v0 ) > 0. Therefore, we may apply the reasoning below for the Gierer and Meinhardt system if we replace (H2), (H3) and (H4) with: (H2′ ) u0 > 0 and v0 > 0 are continuous on Ω; u0 , v0 ∈ CL0 ∞ (Ω); (H3′ ) f and g are continuously differentiable functions from [0, ∞) × [m1 , ∞) × [m2 , ∞) into R with f (t, m1 , s) ≥ 0 and g(t, r, m2 ) ≥ 0 for all t ≥ 0 and (r, s) ≥ (m1 , m2 ); 18 Existence and uniqueness of two-component reaction-diffusion system (H4′ ) There exist m > 0 and a continuous function F : [m1 , ∞) × [m2 , ∞) → R+ such that f (t, r, s), g(t, r, s) ≤ exp(mt)F (r, s) for all t ≥ 0 and (r, s) ≥ (m1 , m2 ). 2 In the setting of semigroup and the abstract initial value problem it is natural 2 to define F and G on R+ × CL0 ∞ by: [F (t, u, v)] (x) = f (t, u(x), v(x)) , [G(t, u, v)] (x) = g(t, u(x), v(x)) for x ∈ Ω, t ≥ 0, u, v ∈ CL0 ∞ . Moreover, we let G1 and G2 be the analytical semigroup of L resp. d × L. The following Lemma is used frequently in what follows. Lemma 1 If u and v are continuous from [0, T ] to Lp (Ω), then the integrals: I1 (t) = I2 (t) = t Z Z 0 t 0 G1 (t − τ )F (τ, u(τ ), v(τ ))dτ G2 (t − τ )G(τ, u(τ ), v(τ ))dτ, exist and I1 (t) and I2 (t) are continuous on [0, T ) with I1 (t), I2 (t) ∈ D(L) and I1 (t), I2 (t) → 0+ in Lp for t → 0+ . 2 Proof The proof can be seen in [Henry (1981), Lemma 3.2.1]. If the reaction-diffusion system has a classical solution then u and v satisfy: Rt u(t) = G1 (t)u0 + 0 G1 (t − τ )F (τ, u(τ ), v(τ ))dτ, Rt (3.4) v(t) = G2 (t)v0 + 0 G2 (t − τ )G(τ, u(τ ), v(τ ))dτ, by variation of constants, which is shown by considering the L2 -valued functions wj (τ ) = Gj (t − τ )u(τ ), j = 1, 2. wj are differentiable since Gj is analytic and u is differentiable.2 Then by Theorem 7 in App. A: dw1 = ∆G1 (t − τ )u(τ ) + G1 (t − τ )u′ (τ ) dτ = ∆G1 (t − τ )u(τ ) − G1 (t − τ )∆u(τ ) + G1 (t − τ )F (τ, u(τ ), v(τ )) = G1 (t − τ )F (τ, u(τ ), v(τ )), and similarly for w2 . Now integration from 0 to t and the result follows. Now, on the other hand let α > n/(2p) and u and v be continuous functions from [0, T ] into D(Hα ) ֒→ CL0 ∞ satisfying (3.4). It is then claimed that u and v 2 We have 1 (wj (t+h)−w(j)) h = 1 (Gj (t+h)u(t+h)−Gj (t)u(t+h))+ h1 Gj (t)(u(t+h)−u(t)). h 3.1 Local existence and uniqueness 19 then solves the reaction diffusion system (1.1) - (1.3). The continuity of u and v implies continuity of t 7→ F (t, u(t), v(t)) and t 7→ G(t, u(t), v(t)). Then, the linear problems ∂t y − ∆y = F (t, u(t), v(t)) and ∂t z − d∆z = G(t, u(t), v(t)) and y(0) = u0 , z(0) = v0 have a unique solution, see e.g. [Henry (1981), Theorem 3.2.2], namely where y and z are given by (3.4). But then, y = u and z = v and the assertion has been shown. Now, all prerequisites have been establish to prove local existence of the reaction diffusion system (1.1) - (1.3). The result is established with L2 -theory alone. Later, when we turn to global existence we shall make use of more general Lp -theory. Proposition 2 (Local existence, see [Henry (1981), Theorem 3.3.1] and [Hollis, Martin, Pierre (1987)]) Given hypotheses (H1)-(H4) are satisfied then there exist T = T (u0 , v0 ) > 0 such that the reaction diffusion system (3.1) - (3.3) has 2 a unique solution (u, v) ∈ CL0 ∞ ((0, T ]; D(Hα )) with u(0) = u0 ∈ CL0 ∞ and v(0) = v0 ∈ CL0 ∞ . 2 Proof By the previous discussion it suffices to prove the corresponding result for (3.4). The proof is then established by utilising Banach’s Fixed Point Theorem. Choose α such that n/4 < α < 1, then the injection I : D(Hα ) → CL0 ∞ is continuous by Lemma 8 in App. A. First we establish the result for initial data in D(Hα ). For ρ > 0 and T > 0 the following closed ball is considered: n o 2 Bρ (u0 , v0 ) = (u, v) ∈ CL0 ∞ ([0, T ]; D(Hα )) | ku − u0 kα , kv − v0 kα ≤ ρ . (3.5) By the continuity of f and g the local Lipschitz properties: kF (t, u1 , v1 ) − F(t, u2 , v2 )k2 ≤ L11 ku1 − u2 kD(Hα ) + L12 kv1 − v2 kD(Hα ) kG(t, u1 , v1 ) − G(t, u2 , v2 )k2 ≤ L21 ku1 − u2 kD(Hα ) + L22 kv1 − v2 kD(Hα ) (3.6) follow for 0 ≤ t ≤ T, (u1 , v1 ), (u2 , v2 ) ∈ Bρ (u0 , v0 ) and Lipschitz-constants L11 , L12 , L21 , L22 > 0. Furthermore, for (y, z) ∈ Bρ (u0 , v0 ) define P : [0, T ] → L2 (Ω) and Q : [0, T ] → L2 (Ω) by P y(t) = G1 (t)u0 + Qz(t) = G2 (t)v0 + Z t G1 (t − τ )F (τ, y(τ ), z(τ ))dτ, 0 Z 0 t G2 (t − τ )G(τ, y(τ ), z(τ ))dτ. 20 Existence and uniqueness of two-component reaction-diffusion system Finally, set M1 = maxt∈[0,T ] kF (t, y(0), z(0))k2 , M2 = maxt∈[0,T ] kG(t, y(0), z(0))k2 and choose T so that, 3ρ k(G1 (h) − I)u0 kα , k(G2 (h) − I)v0 kα ≤ , 0 ≤ h ≤ T, 4 Z T ρ i s−α ds ≤ , i = 1, 2, Cα,2 (Mi + Li1 ρ + Li2 ρ) 4 0 (3.7) (3.8) i where Cα,2 , i = 1, 2, are the constants in 2◦ in Corollary 3 for L resp. d × L and p = 2. Remark 3 Note that such T exists since G1 (h) and G2 (h) → I as h → 0+ by 2◦ in Definition 5 in App. A and Z h 1 h1−α → 0, h → 0+ , s−α ds = 1−α 0 for α < 1. 2 The proof then goes as follows: 1◦ It is shown that (P, Q) maps Bρ (u0 , v0 ) into itself; 2◦ It is shown that (P, Q) is a strict contraction on Bρ (u0 , v0 ), allowing the use of Banach’s Fixed Point Theorem to conclude the existence of a unique fix point in Bρ (u0 , v0 ); 3◦ It is shown that the result extends to initial data in CL0 ∞ . Let (u, v) ∈ Bρ (u0 , v0 ) then, Z t kP u(t) − u0 kD(Hα ) = k(G1 (t) − I)u0 + G1 (t − τ )F (τ, u(τ ), v(τ ))dτ kD(Hα ) 0 Z t kG1 (t − τ )F (τ, u(τ ), v(τ ))kD(Hα ) dτ ≤ k(G1 (t) − I)u0 kD(Hα ) + 0 ≤ (using (3.7)) Z t 3ρ ≤ + kG1 (t − τ )k (kF (τ, u(τ ), v(τ )) − F(τ, u0 , v0 )k2 4 0 + kF (τ, u0 , v0 )k2 ) dτ ≤ (using the local Lipschitz condition and 2◦ in Corollary 3 to apply (3.8)) Z t 3ρ 1 + Cα,2 (M1 + L11 ρ + L12 ρ) (t − s)−α ds ≤ 4 0 ≤ ρ, for 0 ≤ t ≤ T, 3.1 Local existence and uniqueness 21 and similarly, kQv(t) − v0 kD(Hα ) ≤ ρ, showing that (P, Q) maps Bρ (u0 , v0 ) into itself. Furthermore, from 1◦ in Corollary 3 and Lemma 1 we compute: kP u(t + h) − P u(t)kD(Hα ) = kG1 (t)(G1 (h) − I)u0 Z t + (G1 (h) − I) G1 (t − τ )F (τ, u(τ ), v(τ ))dτ 0 + Z t t+h G1 (t + h − τ )F (τ, u(τ ), v(τ ))dτ kD(Hα ) ≤ (using Lemma 1 and the strong continuity of the semigroup) → 0 for h → 0+ , showing that P is continuous from [0, T ] into D(Hα ). A similar computation shows that Q holds the same properties and 1◦ has been shown. 2◦ follows from the estimate: Let (x, y), (z, w) ∈ Bρ (u0 , v0 ) then Z t kP x(t) − P z(t)kD(Hα ) ≤ kG1 (t − τ ) (F (t, x(t), y(t)) − F(t, z(t), w(t))) kD(Hα ) 0 ≤ (using 2◦ in Corollary 3 and (3.6)) Z t 1 ≤ Cα,2 (t − τ )−α dτ × L11 sup kx(t) − z(t)kD(Hα ) τ ∈[0,t] 0 + L12 sup ky(t) − w(t)kD(Hα ) τ ∈[0,t] ! ≤ (by (3.8)) ≤ 1 4 sup kx(t) − z(t)kD(Hα ) t∈[0,T ] ! + sup ky(t) − w(t)kD(Hα ) , τ ∈[0,t] for every t ∈ [0, T ] and thus sup kP x(t) − P z(t)kD(Hα ) ≤ t∈[0,T ] 1 4 sup kx(t) − z(t)kD(Hα ) t∈[0,T ] ! + sup ky(t) − w(t)kD(Hα ) , τ ∈[0,t] 22 Existence and uniqueness of two-component reaction-diffusion system and similarly sup kQy(t) − Qw(t)kD(Hα ) ≤ t∈[0,T ] 1 4 sup kx(t) − z(t)kD(Hα ) t∈[0,T ] ! + sup ky(t) − w(t)kD(Hα ) . τ ∈[0,t] Therefore sup k(P, Q)(x(t), y(t)) − (P, Q)(z(t), w(t))kD(Hα )2 t∈[0,T ] ≤ 1 sup k(x(t), y(t)) − (z(t), w(t))kD(Hα )2 2 t∈[0,T ] verifying that (P, Q) is a strict contraction on Bρ (u0 , v0 ). By Banach’s Fixed Point Theorem, (P, Q) therefore have a unique fix point in Bρ (u0 , v0 ). This is the solution of (3.1)-(3.3) on [0, T ] with initial value (u(0), v(0)) = (u0 , v0 ) ∈ 2 (D(Hα )) . The solution can be extended to a larger interval [0, T + ǫ], ǫ > 0, by considering w1 (t) = u(t + T ) and w2 (t) = v(t + T ) where w1 and w2 solve: ∂t w1 − ∆w1 (t) = F (t + T, w1 (t), w2 (t)), ∂t w2 − d∆w2 (t) = G(t + T, w1 (t), w2 (t)). Again the existence of w1 and w2 on an interval [0, ǫ] is assured by above. Let [0, T ∗ ) denote the maximal interval to which the solution can be extended. Before continuing with the proof let us assume that the proposition holds. Then, we state and prove the following result. Proposition 3 The system (3.1)-(3.3) has a unique solution on a maximal interval of existence [0, T ∗ ) and there are continuous functions N1 and N2 such that 0 ≤ u ≤ N1 (t), 0 ≤ v ≤ N2 (t), for in Ω × [0, T ∗ ). (3.9) If T ∗ < ∞ then lim (kuk∞ + kvk∞ ) = ∞. t→T ∗ − 2 Proof The existence of N1 and N2 directly follows by continuity. If lim (kuk∞ + kvk∞ ) < ∞, t→T ∗ − then the solution could be extended by the procedure above beyond T ∗ ; contradiction the definition of T ∗ . 3.1 Local existence and uniqueness 23 We are now ready to extend the result to initial data in CL0 ∞ (Ω). We let {(un0 , v0n )}n∈N be a sequence in D(Hα ) × D(Hα ) such that un0 , v0n ≥ 0 converge to (u0 , v0 ) in L2 : kun0 − u0 k2 , kv0n − v0 k2 → 0, for n → ∞. We show that {(un , vn )}n∈N , where (un , vn ) are the solutions corresponding to initial conditions (un0 , v0n ), is bounded in D(Hβ ), α < β < 1, and then we exploit the compactness of the injection D(Hα ) ֒→ D(Hβ ), see Proposition 5 in App. A. By the continuity of f and g, and by (H4), there exists a positive, continuous function F̃ such that, kF (t, un (t), vn (t))k∞ , kG(t, un (t), vn (t))k∞ ≤ exp(mT )F̃ (R), (3.10) for all un , vn ∈ D(Hβ ) whenever t ∈ [0, T ] and kun (t)k∞ , kvn (t)k∞ ≤ R. Let R be chosen such that kun0 k∞ , kv0n k∞ ≤ R for all n ∈ N. It is now claimed that T ∗ = inf n∈N Tn∗ > 0. To show this, we first note that φ = G1 (t)u0 solves: ∂t φ − ∆φ = 0, φ(0) = u0 , and thus by the positivity lemma if follows that kφk∞ = kG1 (t)u0 k∞ ≤ R whenever u0 ≤ R. Therefore by (3.4) and 0 ≤ t < Tn∗ : Z t n kun (t)k∞ ≤ kG1 (t)u0 k∞ + kG1 (t − τ )F (τ, un (τ ), vn (τ ))k∞ dτ 0 ≤ (using (3.10)) ≤ R + t exp(mTn∗ )B(R + 1), whenever ku(t)k∞ , ku(t)k∞ ≤ R + 1. For t ∈ [0, exp(−mTn∗ )F̃ (R + 1)−1 ], ku(t)k∞ ≤ R+1 and therefore by Proposition 3, Tn∗ > exp(−mTn∗ )F̃ (R+1)−1 for all n. This inequality defines a δ(R) > 0 such that Tn∗ ≥ δ(R) > 0 for all n, which can be seen by considering the graphs (Tn∗ , Tn∗ ) and Tn∗ , exp(−mTn∗ )F̃ (R + 1)−1 . In the following we shall suppress the dependency of R on δ. According to Corollary 3 in App. A we have: Z t kun (t)kD(Hβ ) ≤ C1,β t−β kun0 k2 + C1,β (t − τ )−β kF (τ, u(τ ), v(τ ))k2 dτ, 0 < (using (3.10)) < ∞, for all n ∈ N, and t ∈ (0, δ], and similarly for vn , showing that the sequences {un (t)}n∈N and {vn (t)}n∈N are bounded in D(Hβ ). Then by the compactness of the injection I : D(Hβ ) → 24 Existence and uniqueness of two-component reaction-diffusion system D(Hα ), see Proposition 5 in App. A, the existence of u(t) ∈ D(Hα ) and v(t) ∈ D(Hα ) such that lim kunk (t) − u(t)kD(Hα ) = 0 = lim kvnk (t) − v(t)kD(Hα ) , k→∞ k→∞ for t ∈ (0, δ], follows. By the local Lipschitz property, (3.6), and the continuity of G1 (t), G2 (t) it follows that u and v are the solutions of (1.1)-(1.3) with initial conditions u(0) = u0 and v(0) = v0 and they are in CL0 ∞ ((0, δ]; D(Hα )). By replacing [0, T ∗ ) with [δ, T ∗ ) and (u0 , v0 ) with (u(δ), v(δ)) ∈ D(Hα ) × D(Hα ) and using the results already obtained when (u0 , v0 ) ∈ D(Hα ) × D(Hα ), the proposition follows. Remark 4 It is possible to improve the regularity of the result in Proposition 2 by showing u, v ∈ CL1 ∞ ((0, T ∗ ); D(Hα )). To do so let u(t), v(t) be given from Proposition 2 and define uδ = u(δ) and vδ = v(δ) for some 0 < δ < T < T ∗ and consider: Z t G1 (t − τ )∂t F (τ, u(τ ), v(τ ))dτ, l1 (t) = G1 (t − δ)F (δ, uδ , vδ ) − ∆ G1 (t − δ)uδ + 0 Z t l2 (t) = G2 (t − δ)G(δ, uδ , vδ ) − d∆ G2 (t − δ)vδ + G2 (t − τ )∂t G(τ, u(τ ), v(τ ))dτ. δ for t ∈ [δ, T ]. Note that l1 , l2 ∈ D(Hα ) since by 1◦ and 2◦ in Corollary 3 every term is so. We may exploit the continuity of the partial derivatives of f and g to conclude that t 7→ ∂r F (t, u(t), v(t)), ∂s F (t, u(t), v(t)) and ∂r G(t, u(t), v(t)), ∂s G(t, u(t), v(t)) are continuous from [δ, T ] into the space of bounded linear operators in L2 , defined by L2 ∋ w 7→ ∂r F (t, u(t), v(t))w, ∂s F (t, u(t), v(t))w and L2 ∋ w 7→ ∂r G(t, u(t), v(t))w, ∂s G(t, u(t), v(t))w. We define: B11 (t) = k (∂r F ) (t, u(t), v(t))(x)k, B21 (t) = k (∂s F ) (t, u(t), v(t))(x)k, B12 (t) = k (∂r G) (t, u(t), v(t))(x)k, B22 (t) = k (∂s G) (t, u(t), v(t))(x)k, 3.1 Local existence and uniqueness 25 and w1 and w2 by w1 (t) = l1 (t) + Z t G1 (t − τ ) (∂r F (τ, u(τ ), v(τ ))w1 (τ ) + ∂s G(τ, u(τ ), v(τ ))w2 (τ )) dτ, δ w2 (t) = l2 (t) + Z δ t G2 (t − τ ) (∂r G(τ, u(τ ), v(τ ))w1 (τ ) + ∂s G(τ, u(τ ), v(τ ))w2 (τ )) dτ. By [Pazy (1983), Corollary 6.1.3] there exist unique w1 and w2 in CL0 ∞ ([δ, T ]; D(Hα )). We now claim that for t ∈ [δ, T ] uh (t) = u(t+h)−u(t) → w1 and vh (t) = h v(t+h)−v(t) → w as h → 0. 2 h We shall prove the claim by giving all details for u and by analogy exploit that v follows by similar arguments. First we consider h → 0+ and rewrite uh using (3.4) with uδ and vδ as initial conditions, that is at time t = δ: " Z t+h 1 G1 (t + h − δ)uδ − G1 (t − δ)uδ + G1 (t + h − τ )F (τ, u(τ ), v(τ ))dτ uh (t) = h δ # Z t − G1 (t − τ )F (τ, u(τ ), v(τ ))dτ δ " Z δ+h 1 G1 (t − δ)(G1 (h) − I)uδ + G1 (t + h − τ )F (τ, u(τ ), v(τ ))dτ = h δ # Z t Z t+h G1 (t − τ )F (τ, u(τ ), v(τ ))dτ . G1 (t + h − τ )F (τ, u(τ ), v(τ ))dτ − + δ δ+h (3.11) The third term in the brackets can be rewritten by a change of variables: Z t+h δ+h G1 (t + h − τ )F (τ, u(τ ), v(τ ))dτ = Z δ t G1 (t − τ )F (τ + h, u(τ + h), v(τ + h))dτ. Furthermore, by Taylor expansion of f and g we may write: F (τ + h, u(τ + h), v(τ + h)) = F (τ, u(τ + h), v(τ + h)) + ∂t F (τ, u(τ + h), v(τ + h)) × h + ǫ1 (h) = F (τ, u(τ ), v(τ )) + ∂r F (τ, u(τ ), v(τ ))(u(τ + h) − u(τ )) + ∂s F (τ, u(τ ), v(τ ))(v(τ + h) − v(τ )) + ǫ2 (h) + ∂t F (τ, u(τ + h), v(τ + h)) × h + ǫ1 (h), where 1 h ǫi (h) → 0 as h → 0 uniformly in L2 -norm on [δ, T ] for i = 1, 2. 26 Existence and uniqueness of two-component reaction-diffusion system Hence (3.11) becomes: Z 1 δ+h G1 (h) − I uδ + uh (t) = G1 (t − δ) G1 (t + h − τ )F (τ, u(τ ), v(τ ))dτ h h δ Z t Z t ǫ1 (h) ǫ2 (h) G1 (t − τ ) G1 (t − τ )∂t F (τ, u(τ + h), v(τ + h)) × h dτ + + dτ + h h δ δ Z t + G1 (t − τ ) [∂r F (τ, u(τ ), v(τ ))uh (τ ) + ∂s F (τ, u(τ ), v(τ ))vh (τ )] dτ. δ And then G1 (h) − I uδ − G1 (t − δ)∆uδ uh − w1 = G1 (t − δ) h ) ( Z 1 δ+h G1 (t + h − τ )F (τ, u(τ ), v(τ ))dτ − G1 (t − δ)F (δ, uδ vδ ) + h δ Z t + G1 (t − τ ) (∂t F (τ, u(τ + h), v(τ + h)) − ∂t F (τ, u(τ ), v(τ ))) × h dτ δ Z t ǫ1 (h) ǫ2 (h) + G1 (t − τ ) dτ + h h δ Z t + G1 (t − τ )∂r F (τ, u(τ ), v(τ ))(uh − w1 )dτ δ Z t + G1 (t − τ )∂s F (τ, u(τ ), v(τ ))(vh − w1 )dτ. δ By (A.3) and (B.2) in App. A resp. App. B the first four terms tends to zero in D(Hα )-norm as h → 0+ . For h → 0− we may use: G(−h) − I G1 (t + h − δ) − G1 (t − δ) uδ = G1 (t + h − δ) uδ → G1 (t − δ)∆uδ , h −h to obtain similar results. Therefore, we may write: kuh − w1 kD(Hα ) ≤ θ1 (h) + K1 Z δ t kuh − w1 kD(Hα ) + kvh − w2 kD(Hα ) dτ, (3.12) where θ1 (h) → 0 as h → 0 and K1 = max ! sup kG1 (t)kB11 (t), sup kG1 (t)kB21 (t) . t∈[δ,T ] t∈[δ,T ] 3.2 Application of result on local existence to specific models 27 In the same manner we obtain: kvh − w2 kD(Hα ) ≤ θ2 (h) + K2 Z t δ kuh − w1 kD(Hα ) + kvh − w2 kD(Hα ) dτ, (3.13) again θ2 (h) → 0 as h → 0 and: K2 = max sup t∈[δ,T ] kG2 (t)kB12 (t), sup t∈[δ,T ] ! kG2 (t)kB22 (t) . Adding (3.12) and (3.13) and using Gronwall’s inequality, see App. C and Lemma 12, we obtain: kuh − w1 kD(Hα ) + kvh − w2 kD(Hα ) ≤ 2 (θ1 (h) + θ2 (h)) exp((t − δ) max(K1 , K2 )), showing that uh → w1 and vh → w2 in D(Hα ) as h → 0, and hence w1 and w2 are the derivatives of u and v resp. Since w1 , w2 ∈ CL0 ∞ ([δ, T ]; D(Hα )), u and v are continuous differentiable on [δ, T ]. Since this holds for every 0 < δ < T < T ∗ we conclude that u, v ∈ CL1 ∞ ((0, T ∗ ); D(Hα )). 2 In the following section we shall apply the results on local existence to prove global existence of the Schnakenberg and Thomas system presented in Chap. 1. 3.2 Application of result on local existence to specific models By Proposition 2 and Remark 2 local existence and uniqueness follows for the three models described in Chap. 1. To prove global existence it suffices, according to Proposition 3, to prove that, lim (ku(t)k∞ + kv(t)k∞ ) < ∞. t→T ∗ − To do so, we shall make use of the following corollary, which is an immediate consequence of the positivity lemma for parabolic equations. Corollary 4 Let L be an elliptic operator, for example L = ∆. If u is continuous in Ω × (0, T ) and ∂t u − Lu = f (u) ≥ g(u), in Ω × (0, T ), νu = 0 ≥ h, on ∂Ω × (0, T ), u = u0 ≥ l, on Ω × {t = 0} 28 Existence and uniqueness of two-component reaction-diffusion system and y is a continuous solution of: ∂t y − Ly = g(y), in Ω × (0, T ) νy = h, on ∂Ω × (0, T ), y = l, on Ω × {t = 0}, then u ≤ y, in Ω × [0, T ). 2 Proof Define w = u − y and observe that w satisfy: ∂t w − Lw = f (u) − g(y) ≤ 0, in Ω × (0, T ) νw = −h ≤ 0, on ∂Ω × (0, T ), y = u0 − l ≤ 0, on Ω × {t = 0}. Then the positivity lemma guarantees that w ≤ 0 and thus u ≤ y and the proof is done. Again we consider the Schnakenberg model first. Assertion 1 For the Schnakenberg model there exist constants Γ1 and Γ2 such that the following estimates hold: 0 ≤v ≤ Γ1 , 0≤ kuk22 ≤ Γ2 , (3.14) (3.15) for every (x, t) ∈ Ω × (0, T ∗ ). Proof From the equation of u we readily obtain: ∂t u − ∆u = a − u + u2 v ≥ a − u and thus by Corollary 4 the solution y of: ∂t y = a − y, y = inf u0 x∈Ω satisfy 0 ≤ y ≤ u. This is completely analogous to the application of Pao in Sec. 2.2. The solution is y = a+exp(−t)(inf x∈Ω u0 −a) thus 0 ≤ min(a, inf x∈Ω u0 ) ≤ u. Let m ≡ min(a, inf x∈Ω u0 ). For v we have ∂t v − d∆v = b − u2 v ≤ b − m2 v, 3.2 Application of result on local existence to specific models 29 and thus the solution z of ∂t z = b − m2 v, z(0) = sup v0 , x∈Ω satisfy 0 ≤ v ≤ z. The solution is z(t) = mb2 + e−m hence: b 0 ≤ v ≤ max , sup v0 , m2 x∈Ω 2 t supx∈Ω v0 − b m2 and (3.16) showing (3.14). For simplicity M ≡ max mb2 , supx∈Ω v0 . Adding the equation for u and the equation for v gives: ∂t (u + v) − ∆(u + dv) = a + b − u. Multiplying by u + v and integration over Ω give after integration by parts: Z Z h i 1 2 2 ∂t (u + v)2 + (∇u) + (1 + d)∇u∇v + d (∇v) 2 Ω Ω Z (3.17) = (a + b − u)(u + v). Ω Now since, 1 (∇u + (1 + d)∇v)2 ≥ 0, 2 then by expanding we conclude that, 2 2 (∇u) + (1 + d)∇u∇v + d (∇v) ≥ 1 1 2 2 (∇u) − (1 + d2 ) (∇v) . 2 2 (3.18) Using (3.18) in (3.17) gives the inequality: Z h Z i Z 2(a + b − u)(u + v). (3.19) (∇u)2 − (1 + d2 ) (∇v)2 ≤ ∂t (u + v)2 + Ω Ω Ω 2 We estimate the (∇v) -term by use of the equation for v and (3.16): Multiply the equation for v with v and integrate over the domain Ω to obtain: Z Z Z 1 2 2 ∂t d (∇v) = (bv − u2 v) v + 2 Ω Ω ZΩ bv ≤ vol Ω bM, ≤ Ω 30 Existence and uniqueness of two-component reaction-diffusion system and thus: 2 (1 + d ) Z 2 Ω (∇v) ≤ Z 1 1 1 + d vol Ω bM − + d ∂t v2 . d 2 d Ω Insertion into (3.19): ∂t Z (u + v)2 + Ω 1 2 Z Z 1 1 2 + d ∂t + d vol Ω bM (∇u) − v2 + d d Ω Ω Z 2(a + b − u)(u + v). ≤ Ω Setting w = (u + v)2 + ∂t Z w+ Z Ω Ω 1 2 (∇u)2 ≤ 1 d Z + d v 2 we obtain: 2(a + b − u)(u + v) + Ω 1 + d vol Ω bM. d (3.20) We now claim the exist E > 0 such that 1 w≤E− 2 Ω Z ∂t To prove the claim we add (∂t + 1) Z w+ Ω Z Ω R Ω Z w. (3.21) Ω w to both sides of (3.20): 2 (∇u) ≤ + Z 1 + d vol Ω bM d Ω 2(a + b − u)(u + v) + Z w Ω ≤ (Expand the terms on the right hand side) Z 1 1 1 ≤ + d vol Ω bM + +d +1 v2 d 2 d Ω Z Z u2 . 2(a + b)(u + v) − + Ω Ω The second term on the right hand side can be estimated using (3.16): Z 1 1 1 1 +d +1 + d + 1 vol Ω M 2 v2 ≤ 2 d 2 d Ω 3.2 Application of result on local existence to specific models 31 + d vol Ω bM + 21 d1 + d + 1 vol Ω M 2 . Then Z Z Z 2 2(a + b)(u + v) − kuk22 (∇u) ≤ H + w+ (∂t + 1) Let H = 1 d Ω Ω Ω ≤ (By Cauchy-Schwarz and Minowski inequality) ≤ H + k2(a + b)k2 (kuk2 + kvk2 ) − kuk22 1 ≤ H + 8(a + b)2 vol Ω (kuk2 + kvk2 ) 2 ≤ (By Young’s inequality) 1 kuk22 + kvk22 ≤ H + 32(a + b)4 vol Ω2 + 4Z 1 4 2 w, ≤ H + 32(a + b) vol Ω + 2 Ω proving the claim, with E = H + 32(a + b)4 vol Ω2 . By Gronwall’s inequality, see App. C and Lemma 11, we deduce from (3.21) that Z Z t 2 w ≤ 2E + exp − kuk2 ≤ w0 − 2E 2 Ω Ω Z 1 1 2 + d v02 . ≤ 2E + (u0 + v0 ) + 2 d Ω R Putting Γ2 = 2E + Ω (u0 + v0 )2 + 21 d1 + d v02 (3.15) follows. We now have to turn to Lp -theory. It is clear from the definition of f and the bounds just obtained that, kF (u(t), v(t))k1 ≤ (a + Γ1 )(vol Ω + Γ2 ) and hence for p = 1 and α < 1 it follows by Corollary 3: Z t Cα,1 (t − s)−α (a + Γ1 )(vol Ω + Γ2 )ds ku(t)kD(Hα ) ≤ Cα,1 t−α ku0 k1 + 0 ≤ Cα,1 t −α ku0 k1 + Cα,1 (a + Γ1 )(vol Ω + Γ2 ) ≤ (using α < 1) ≤ Cα,1 t−α ku0 k1 + Cα,1 (a + Γ1 )(vol Ω + Γ2 ) Z 0 T∗ (t − s)−α ds T ∗ 1−α . 1−α D(Hα ) ⊂ CL0 ∞ (Ω) with continuous injection for α > n/2 for p = 1, thus, for n = 1 we have: kuk∞ < ∞. Therefore by contradiction it is concluded: 32 Existence and uniqueness of two-component reaction-diffusion system Corollary 5 For n = 1 and every u0 , v0 ∈ CL0 ∞ (Ω) the Schnakenberg system, (1.4)-(1.6), has a unique, positive, global solution, that is T ∗ = ∞. 2 In order to prove global existence for the Schnakenberg model for n = 2 and n = 3 we need some additional results, which we shall establish in the following section. But first let us consider the Thomas-model, that is the system with reaction functions: ρrs ρrs f (r, s) = a − r − , g(r, s) = α(b − s) − . 2 1 + r + kr 1 + r + kr2 The method of upper and lower solutions was insufficient for this model as the reaction functions were not quasi-monotone. But nevertheless it is possible, for this model, to utilise Proposition 2 and Proposition 3 to prove, by contradiction, global existence for space variables n = 1, 2 and 3. From: ∂t u − ∆u = f (u, v) ≤ a − u, ∂t v − d∆v = g(u, v) ≤ α(b − v), we conclude that, 0 ≤ u ≤ max(a, supΩ u0 ), 0 ≤ v ≤ max(b, supΩ v0 ). (3.22) As the bound is independent of t we therefore conclude that limt→T ∗ − (ku(t)k∞ + kv(t)k∞ ) < ∞ and therefore by contradiction: Corollary 6 For every u0 , v0 ∈ CL0 ∞ (Ω) the Thomas system has a global solution, that is T ∗ = ∞. By (3.22) it is furthermore uniformly bounded. 2 3.3 Global existence and boundedness In this section we shall present and extend the work of [Hollis, Martin, Pierre (1987)] to give sufficient conditions for the system to be global and bounded. Let us first state the theorem we aim to prove. Theorem 2 Suppose in addition to (H1)-(H4) the following hypotheses hold: (H5) N2 in Proposition 3 is bounded if T ∗ < ∞; (H6) There is an η ≥ 1 and a continuous function h : [0, ∞)2 → [0, ∞) such that |f (t, r, s)| ≤ h(t, S)(1 + r)η for all t, r, s ≥ 0 with s ≤ S; 3.3 Global existence and boundedness 33 (H7) There is an ǫ > 0 and a continuous function l : [0, ∞)2 → [0, ∞) such that ǫr + f (t, r, s) + g(t, r, s) ≤ l(t, S) for all t, r, s ≥ 0 with s ≤ S. Then the solution exists on Ω × (0, ∞) and (3.9), repeated here for convenience: 0 ≤ u ≤ N1 (t), 0 ≤ v ≤ N2 (t), for in Ω × [0, T ∗), holds with T ∗ = ∞. If furthermore N2 , h and l are bounded in t, N2 (t) ≤ N 2 , h(t, S) ≤ h(S), l(t, S) ≤ l(S), for all t ≥ 0, then there exist N 1 such that N1 (t) ≤ N 1 for all t ∈ [0, ∞) and so the solution is uniformly bounded in Ω × [0, ∞). 2 Remark 5 Notice that the Schnakenberg model (1.4)-(1.6) satisfy these hypotheses: (H5) is satisfied by (3.16) in Sec. 3.2, and the following computation, |f (t, r, s)| = |a − r + r2 s| ≤ (a + r + r2 s) ≤ (a + 1)(1 + r + r2 S) ≤ (a + 1)(S + 1)(1 + r + r2 ) ≤ (a + 1)(S + 1)(1 + r)2 , shows that we may take h = (a + 1)(S + 1) and thus that (H6) is satisfied. Simply by inspection u + f (t, r, s) + g(t, r, s) ≤ l = a + b, so (H7) follows. By Theorem 2, the solution is global and since N2 , h and l are independent of t, u and v are uniformly bounded. 2 Remark 6 Notice that by a change of variable x̃ = d−1/2 x (1.1) is transformed into: ˜ = f (t, u, v), ut − d˜∆u ˜ vt − ∆v = g(t, u, v), in Ω̃ × (0, T ), where d˜ = 1/d and hence N2 might as well be replaced with N1 and f with g in (H5) resp. (H6). 2 Remark 7 Compared to [Hollis, Martin, Pierre (1987)], we had to replace • (H7′) There is a continuous function l : [0, ∞)2 → [0, ∞) such that f (t, r, s) + g(t, r, s) ≤ l(t, S) for all t, r, s ≥ 0 with s ≤ S. 34 Existence and uniqueness of two-component reaction-diffusion system with (H7). 2 Proof To prove the theorem a couple of lemmas are necessary. For the first statement in the theorem we shall make use of four lemmas. The purpose of each lemma is described below. For 0 ≤ δ < T < ∞ we let Lp (Ω×(δ, T )) be the space of measurable functions φ : Ω × (δ, T ) → R equipped with the Banach-norm: "Z #1/p T kφkp,δ,T = δ kφkpp dt < ∞. In the four lemmas we consider the following linear auxiliary problem: ∂t φ = Hǫ φ − θ, in Ω × (δ, T ) νφ = 0, on ∂Ω × (δ, T ), φ = 0, on Ω × {t = T }, (3.23) (3.24) (3.25) where Hǫ = −∆ + ǫI. By introducing φ(x, t) = φ(x, T − t) and θ(x, t) = θ(x, T − t): ∂t φ = −Hǫ φ + θ, in Ω × (0, T − δ), νφ = 0 φ=0 on ∂Ω × (0, T − δ), on Ω × {t = 0}, it follows by the positivity lemma that φ ≥ 0, and then φ ≥ 0, whenever θ ≥ 0. Justified by similar arguments to those in Example 1, Hǫ is sectorial and therefore −Hǫ generates an analytic semigroup G̃ satisfying kG̃(t)φkp ≤ exp(−δt)kφkp by Theorem 7 in App. A. The first lemma below states that the linear mappings Lp (Ω×(δ, T )) ∋ θ 7→ φ and Lp (Ω × (δ, T )) ∋ θ 7→ ∆φ are continuous. The second and third lemma relate φ and θ to the solution (u, v), to conclude that u ∈ Lp (Ω × (0, T ∗ )) when T ∗ < ∞. Finally, the fourth lemma uses the result in the third lemma and (H5) to conclude that also F ∈ Lp (Ω × (0, T ∗ )) when T ∗ < ∞. The four lemmas are then together with the continuous embedding D(Hα ) ֒→ 0 CL∞ used to conclude that lim ku(t)k∞ < ∞, t→T ∗ − in turn, implying that the solution is global according to Proposition 3. Now for the first lemma. Lemma 2 Suppose that p ∈ (1, ∞), θ ∈ Lp (Ω × (δ, T )) and φ is the solution to (3.23)-(3.25). Then there is a constant C(p) only depending upon the norm such that kφ(δ)kp , kφkp,δ,T , k∆φkp,δ,T ≤ C(p)kθkp,δ,T . (3.26) 2 3.3 Global existence and boundedness 35 Proof By [Ladyzenskaja et al., Theorem 9.1 p. 341]: kφkp,δ,T , k∆φkp,δ,T ≤ C(p, T − δ)kθkp,δ,T , (3.27) where C may be chosen so that, C(p, T − δ) ≤ C(p, T ′ − δ), if 0 ≤ δ < T ≤ T ′ . (3.28) Applying variations of constants we get: kφ(δ)kp = kφ(T − δ)kp = k ≤ Z T −δ Z 0 T −δ G̃(T − δ − τ )θ(τ )dτ kp exp(−δ(T − δ − τ ))kθ(τ )kp dτ 0 ≤ (using Hölder’s inequality) !1/q Z T −δ ≤ exp(−qδ(T − δ − τ )dτ 0 −1/q ≤ (qδ) Z 0 T −δ kθ(τ )kpp dτ !1/p kθkp,δ,T , where the last inequality follows from the estimate: Z 0 T −δ exp(−qδ(T − δ − τ )dτ = (qδ)−1 (1 − exp(−qδ(T − δ))) ≤ (qδ)−1 . (3.29) Again, by variation of constants: kφkp,δ,T = ≤ Z T −δ 0 Z 0 T −δ k Z t 0 Z 0 G̃(t − τ )θ(τ )dτ kpp dt t exp(−δ(t − τ ))kθ(τ )kp dτ p dt ≤ (using Hölder’s inequality and p−1 + q −1 = 1) 1/p #p 1/q Z t Z T −δ "Z t 1 1 dt exp −p δ(t − τ ) kθ(τ )kpp ≤ exp −q δ(t − τ ) q p 0 0 0 # p/q Z t Z T −δ "Z t p exp(−δ(t − τ )) ≤ exp(−δ(t − τ ))kθ(τ )kp dt. 0 0 0 36 Existence and uniqueness of two-component reaction-diffusion system Then, by utilising a similar estimate as in (3.29), we obtain: kφkp,δ,T ≤ δ −p/q =δ −p/q Z T −δ 0 Z Z 0 T −δ 0 t exp(−δ(t − τ ))kθ(τ )kpp dτ dt kθ(τ )kpp ≤ δ −p/q−1 kθkpp,0,T −δ Z τ T −δ ! exp(−δ(t − τ ))dt dτ = (using p−1 + q −1 = 1) = δ −p kθkpp,δ,T . (3.30) For ∆φ, we consider ψ(t) = A(t − t0 )φ(t) for t0 ≥ 0 where A ∈ C0∞ (R) such that A = 0 in {0}, (3.31) 0 ≤A ≤ 1 in (0, 1), A = 1 in [1, 2]. (3.32) (3.33) See Fig. 3.1. 1 0 1 2 Figure 3.1: Test function satisfying (3.31)-(3.33). Then, ∂t ψ = A′ (t − t0 )φ + A(t − t0 )∂t φ = A′ (t − t0 )φ + A(t − t0 )∆φ + A(t − t0 )θ = ∆ψ + A′ (t − t0 )φ + A(t − t0 )θ, νψ = 0, on ∂Ω × {t > t0 }, ψ = 0, in Ω × {t = t0 }. in Ω × {t > t0 }, 3.3 Global existence and boundedness 37 We shall make use of the inequality: (a + b)p ≤ 2p (ap + bp ), a, b ≥ 0, (3.34) which follows from the estimate: (a + b)p ≤ (2 max(a, b))p ≤ 2p max(ap , bp ) ≤ 2p (ap + bp ). Put C2 (p) = C(p, 2), and set θ = 0 for t > T − δ. Then by (3.27) and the triangle inequality, k∆ψkpp,t0 ,t0 +2 ≤ C2 (p)p kθkp,t0 ,t0 +2 + kA′ k∞ kφkp,t0 ,t0 +2 ≤ (using (3.34)) p ≤ 2p C2 (p)p kθkpp,t0 ,t0 +2 + kA′ kp∞ kφkpp,t0 ,t0 +2 , and by the choice of A: k∆ψkpp,t0 ,t0 +2 ≥ k∆ψkpp,t0 +1,t0 +2 = k∆φkpp,t0 +1,t0 +2 . For t0 = 0, . . . , k, . . . we have: k∆φkpp,1,2 ≤ k∆ψkpp,0,2 ≤ 2p C2 (p)p kθkpp,0,2 + kA′ kp∞ kφkpp,0,2 .. . k∆φkpp,k+1,k+2 ≤ k∆ψkpp,k,k+2 ≤ 2p C2 (p)p kθkpp,k,k+2 + kA′ kp∞ kφkpp,k,k+2 .. . Since θ ∈ Lp (Ω × (0, ∞)) we have as a consequence of (3.27) that φ ∈ Lp (Ω × (0, ∞)) so we may sum over k to obtain: k∆φkpp,1,∞ ≤ 21+p C2 (p)p kθkpp,0,∞ + kA′ kp∞ kφkpp,0,∞ . And thus by (3.28), k∆φkpp,0,∞ ≤ (21+p + 1)C2 (p)p kθkpp,0,∞ + kA′ kp∞ kφkpp,0,∞ . Now applying (3.30) we finally obtain: k∆φkp,δ,T ≤ C(p)kθkp,δ,T , 1/p where C(p) = (21+p +1)1/p C2 (p) (1 + kA′ kp∞ δ −p ) the result follows. . Choosing C(q) = max(C(p), (qδ)−1/p , δ −p ) 38 Existence and uniqueness of two-component reaction-diffusion system Lemma 3 Suppose (H1)-(H4) are satisfied and let 0 ≤ δ < T < T ∗ . Then, Z Z Z φ(δ)u(δ), (3.35) φ(ǫu + f (t, u, v)) + uθ = Ω Ω×(δ,T ) Ω×(δ,T ) Z Z Z vθ ≤ φg(t, u, v) + φ(δ)v(δ) + E(T ), (3.36) Ω×(δ,T ) Ω×(δ,T ) Ω where E(δ, T ) = C(|d − 1| + ǫ)N 2 (δ, T )vol Ω1/q (T − δ)1/q kθkp , (3.37) and N 2 (δ, T ) = sup N2 (t). δ≤t<T If T ∗ < ∞, (3.35) and (3.36) also holds for T = T ∗ . 2 Proof First we integrate φ∂t u over Ω to obtain: Z φ∂t u = (using equation for u) Ω Z = (φ∆u + φf (t, u, v)) Ω = (using Green’s identity) Z Z = u∆φ + φf (t, u, v). Ω (3.38) Ω Furthermore, by integrating φut from t = δ and t = T : Z T φ∂t u = (using integration by parts) δ = φ(T )u(T ) − φ(δ)u(δ) − = (using (3.23)) Z = −φ(δ)u(δ) + δ Z T u∂t φ δ T u∆φ − ǫ Z T δ uφ + Z T uθ, (3.39) δ since φ(T ) = 0. (3.38) together with (3.39) yield: Z Z Z Z u∆φ φ(δ)u(δ) + φf (t, u, v) = − u∆φ + Ω×(δ,T ) Ω Ω×(δ,T ) Ω×(δ,T ) Z Z −ǫ uφ + uθ, Ω×(δ,T ) Ω×(δ,T ) 3.3 Global existence and boundedness 39 and thus Z uθ = Ω×(δ,T ) Z φ(δ)u(δ) + Ω Z φ(ǫu + f (t, u, v)). Ω×(δ,T ) Similarly for v we obtain: Z Z φ(δ)v(δ) + φg(t, u, v) Ω Ω×(δ,T ) Z Z vφ. v∆φ + ǫ + (d − 1) vθ = Ω×(δ,T ) Z Ω×(δ,T ) Ω×(δ,T ) Now by assumption: (d − 1) Z Ω×(δ,T ) Z vφ ≤ |d − 1|N 2 (δ, T ) Ω×(δ,T ) Z + ǫN 2 (δ, T ) |φ| v∆φ + ǫ Z Ω×(δ,T ) |∆φ| Ω×(δ,T ) ≤ (using Hölder’s inequality) ≤ |d − 1|N 2 (δ, T )vol Ω1/q (T − δ)1/q k∆φkp,δ,T + ǫN 2 (δ, T )vol Ω1/q (T − δ)1/q kφkp ≤ (using (3.27)) ≤ C(p)(|d − 1| + ǫ)N 2 (δ, T )vol Ω1/q (T − δ)1/q kθkp,δ,T , and the result follows. In the following let C̃(p, N 2 (δ, T )) = C(p)(|d − 1| + ǫ)N 2 (δ, T )vol Ω1/q . Lemma 4 Suppose 0 ≤ δ < T < T ∗ or ≤ T ∗ if T ∗ < ∞ and p ∈ (1, ∞). Then u ∈ Lp (Ω × (δ, T )). 2 40 Existence and uniqueness of two-component reaction-diffusion system Proof Adding (3.35) and (3.36) gives: Z Z Z vθ uθ + uθ ≤ Ω×(δ,T ) Ω×(δ,T ) Ω×(δ,T ) Z Z φ(δ)(u(δ) + v(δ)) + ≤ φ(ǫ + f (t, u, v) + g(t, u, v)) + E(δ, T ) Ω×(δ,T ) Ω ≤ (using (H7)) Z Z ≤ φ(δ)(u(δ) + v(δ)) + Ω φl(t, N 2 (δ, T )) + E(δ, T ) Ω×(δ,T ) ≤ (using Hölder’s inequality and (3.37)) ≤ ku(δ)kp kφ(δ)kq + N2 (s)vol Ω1/p kφ(δ)kq !1/p Z p kφkq,δ,T + C̃(q, N 2 (δ, T ))(T − δ)1/p kθkq,δ,T + l(t, N 2 (δ, T )) Ω×(δ,T ) ≤ (using (3.26)) ≤ ku(δ)kp C(q) + N 2 (δ, T )vol Ω1/p C(q) + l(t, N 2 (δ, T ))vol Ω 1/p 1/p (T − δ) 1/p C(q) + C̃(q, N 2 (δ, T ))(T − δ) where l(t, N 2 (δ, T ) = supδ≤t<T l(t, N 2 (δ, T ). Therefore Z uθ ≤ Ĉ(p, q, N 2 (δ, T )) 1 + ku(δ)kp + (T − δ)1/p kθkq,δ,T , ! kθkq,δ,T , (3.40) Ω×(δ,T ) where ( Ĉ(p, q, N 2 (δ, T )) = max C(q), N 2 (δ, T )vol Ω1/p C(q), l(t, N 2 (δ, T ))vol Ω 1/p ) C(q) + C̃(q, N 2 (δ, T )) . (3.41) Since this estimate holds for every θ ∈ Lq (Ω × (δ, T )), θ ≥ 0 we conclude by duality that u ∈ Lp (Ω × (δ, T )), u ≥ 0 and (3.42) Ĉ(p, q, N 2 (δ, T )) 1 + ku(δ)kp + (T − δ)1/p is an estimate for kukp,0,T . Remark 8 Notice that Ĉ is independent of δ and T if N 2 (δ, T ) ≤ N 2 , as will be the case in the last part of the theorem. 2 3.3 Global existence and boundedness 41 The following lemma combines the previous lemma with the hypotheses (H6). Lemma 5 Suppose the suppositions of Lemma 4 hold and that p > 1 and η is as in (H6). Then there is a constant h(T, N 2 (T )) such that Z δ T kF (t, u(t), v(t))kpp dt ≤ h(T − δ, N 2 (δ, T )) 1 + kukηp ηp,δ,T , whenever 0 ≤ δ < T < T ∗ or ≤ T ∗ if T ∗ < ∞. (3.43) 2 Proof By (H6) we have: |f (t, u, v)|p ≤ h(T, N 2 (T ))p (1 + |u(t)|) ≤ (using (3.34)) ηp ≤ 2pη h(T, N 2 (T ))p (1 + |u(t)|pη ) , and integration over Ω × (δ, T ): kF (t, u, v)kpp,δ,T ≤ 2pη h(T − δ, N 2 (δ, T ))p (T − δ)vol Ω + kukηp ηp,δ,T pη p ≤ 2 h(T − δ, N 2 (δ, T )) max(1, (T − δ)vol Ω) 1 + kukηp ηp,δ,T . Put h(T − δ, N 2 (δ, T )) = 2pη h(T − δ, N 2 (δ, T ))p max(1, (T − δ)vol Ω) and the result follows. Now we are ready to prove global existence. Suppose by contradiction that T ∗ < ∞ and thus by Proposition 3: lim ku(t)k∞ = ∞. t→T ∗ − (3.44) Select α ∈ (0, 1) and p > 1 such that n/(2p) < α and αq < 1 where q is the p → 1 from above as conjugated number of p. Clearly such p exist since q = p−1 p → ∞. Since T ∗ < ∞ we have by Lemma 4: kukpp,0,T ∗ ≤ C1 (p)p , and then by Lemma 5 there is an C2 such that kF (t, u, v)kpp,0,T ∗ ≤ C2 (p)p (1 + C1 (pη)pη ). (3.45) 42 Existence and uniqueness of two-component reaction-diffusion system Using the properties in Corollary 3, variation of constants, (3.4), gives: Z t −α Cα,p (t − τ )−α kF (τ, u(τ, v(τ ))kp dτ ku(t)kD(Hα ) ≤ Cα,p t ku0 kp + 0 ≤ (using Hölder’s inequality) Z t 1/q −α −αq ≤ Cα,p t ku0 kp + Cα,p (t − τ ) 0 × Z 0 t kF (τ, u(τ ), v(τ ))kpp ≤ Cα,p t −α ku0 kp + Cα,p 1/p Z T∗ ∗ −αq !1/q ∗ −αq !1/q (T − τ ) 0 × kF(t, u, v)kp,0,T ∗ ≤ (using (3.45)) ≤ Cα,p t −α ku0 kp + Cα,p Z 0 T∗ (T − τ ) × C2 (p)(1 + C1 (pη))1/p Cα,p T ∗ 1/q−α ≤ Cα,p t−α ku0 kp + (1 − αq)1/q × C2 (p)(1 + C1 (pη))1/p , (3.46) since αq < 1, for all t ∈ (0, T ∗ ). By the continuous injection I : D(Hα ) → CL0 ∞ limt→T ∗ − ku(t)k∞ < ∞, contradicting (3.44), and T ∗ = ∞. For the second statement in Theorem 2, regarding the boundedness of the solutions we shall make use of two lemmas. The lemmas construct a sequence {tn }n∈N , tn → ∞ as n → ∞, such that a bound on kF (t, u(t), v(t))kp,tn ,tn+1 can be established independently of n. Then the proof is completed by exploiting the continuous injection I : D(Hα ) → CL0 ∞ . When N2 (t) ≤ N 2 for all t ≥ 0, then it follows by Remark 8, that Ĉ in Lemma 4, (3.41), may be chosen independent of δ and T , so that it only depends on the norms. In the following we shall for convenience write Ĉ instead of Ĉ(p, q) and set Υ = Ĉ p (Ĉ + 2)p . ~ Lemma 6 For each p ∈ (1, ∞) there are constants C(p) and C̆(p) and a sequence {tn }n∈N such that ~ 1◦ 1 < tn+1 − tn ≤ C(p), 2◦ ku(tn )kp ≤ Ĉ + 1, 3.3 Global existence and boundedness 3◦ R tn+1 tn 43 ku(t)kpp dt ≤ C̆(p). 2 Proof We may choose Ĉ so that Ĉ + 1 ≥ ku0 kp and Ĉ ≥ 1. First we claim: If τ ≥ 0 and ku(τ )kp ≤ Ĉ + 1 then there is τ̃ ∈ (τ, τ + Υ) such that ku(τ̃ )kp ≤ Ĉ + 1. (3.47) The proof is by contradiction. So we assume that ku(t)kp > Ĉ(p) + 1 for all t ∈ (τ, τ + Υ) and so Z τ +Υ ku(t)kpp dt > Υ(Ĉ + 1)p . τ However, by (3.42) we obtain: Z τ +Υ ku(t)kpp dt ≤ Ĉ p (1 + ku(τ )kp + Υ1/p )p τ = Ĉ p (Ĉ + 2 + Ĉ(Ĉ + 2))p = Ĉ p (Ĉ + 2)p (Ĉ + 1)p = Υ(Ĉ + 1)p , and we have arrived at a contradiction. ~ We set C(p) = 2Υ and define the sequence {tn }n∈N by t1 = 0 and tn+1 = sup{tn ≤ t ≤ tn + 2Υ | ku(t)kp ≤ Ĉ + 1}, n ∈ N. tn+1 is bounded from above by tn + 2Υ. Additionally, the set Σn = {tn ≤ t ≤ tn + 2Υ | ku(t)kp ≤ Ĉ + 1} is nonempty for every n ∈ N, since t 7→ ku(t)kp is continuous and by assumption ku(t1 = 0)k ≤ Ĉ + 1. Therefore, the supremum exists and the sequence is welldefined. It is left to be shown that the sequence satisfies the lower bound in 1◦ . We claim that tn+1 − tn ≥ Υ > 1. Again by contradiction, if tn+1 − tn < Υ then by (3.47) there is a t̃ ∈ (tn+1 , tn+1 + Υ) such that ku(t̃)kp ≤ Ĉ + 1. But tn ≤ t̃ ≤ tn+1 + Υ ≤ tn + 2Υ and so t̃ ∈ Σn with t̃ > tn+1 contradicting that tn+1 = sup Σn . Thus we conclude that the sequence satisfies 1◦ . By (3.42) and Z tn+1 ku(t)kpp dt ≤ Ĉ p (1 + ku(tn )k + (tn+1 − tn )1/p )p tn ≤ Ĉ p (1 + Ĉ + 1 + (2Υ)1/p )p ≤ Ĉ p (Ĉ + 2 + (2Υ)1/p )p , we see that (3) holds with C̆ = Ĉ p (Ĉ + 2 + (2Υ)1/p )p , independent of n. 44 Existence and uniqueness of two-component reaction-diffusion system Lemma 7 Suppose that N2 , h and l are bounded in t, p ∈ (1, ∞) and η is as in (H6), and that {tn }n∈N is the sequence from Lemma 6 corresponding to ηp. Then there is a constant Ċ(p, ηp), independent of n, so that, Z tn+1 kF (t, u(t), v(t))kpp ≤ Ċ(p, ηp), for all n ∈ N. 2 tn Proof The proof is similar to the proof of Lemma 5. The boundedness on N2 , h and l applied to (3.43) imply the existence of C̊(p) independent of t and n, such that Z tn+1 Z tn+1 ηp p ku(t)kηp dt , kF (t, u(t), v(t))kp dt ≤ C̊(p) 1 + tn tn and hence by Lemma 6: Z tn+1 kF (t, u(t), v(t))kpp dt ≤ C̊(p) 1 + C̆(ηp) . tn Putting Ċ(p, ηp) = C̊(p) 1 + C̆(ηp) the result follows. We are now ready to prove the final statement in Theorem 2 about global boundedness of the solution. Again choose α ∈ (0, 1) and p > 1 such that n/(2p) < α and αq < 1 and let {tn }n∈N be the sequence from Lemma 6 corresponding to ηp. Replacing u0 by u(tn+1 ) we obtain: ku(t)kD(Hα ) ≤ Cα,p (t − tn+1 )−α ku(tn+1 )kp Z t Cα,p (t − τ )−α kF (τ, u(τ ), v(τ ))kp dτ. + (3.48) tn+1 for t ∈ (tn+2 , tn+3 ) and n ∈ N. Now since t − tn+1 ≥ 1 then (t − tn+1 )−α ≤ 1 and then by Lemma 6: Cα,p (t − tn+1 )−α ku(tn+1 )kp ≤ Cα,p ku(tn+1 )kp ≤ (using Hölder’s inequality) ≤ Cα,p vol Ω ku(tn+1 )kηp ≤ Cα,p vol Ω (Ĉ(ηp) + 1). Put R = Cα,p vol Ω (Ĉ(ηp) + 1). For the second term in (3.48) we apply similar reasoning as in (3.46) to obtain: Z t Cα,p (t − τ )−α kF (τ, u(τ ), v(τ ))kp dτ ≤ QĊ(p, ηp)1/p , tn+1 3.3 Global existence and boundedness 45 for some constant Q. Thus for t ∈ [tn+2 , tn+3 ], n ∈ N then ku(t)kD(Hα ) ≤ R + QĊ(p, ηp)1/p . Since this estimate is independent of n, and by the continuous injection I : D(Hα ) → CL0 ∞ , we conclude, in turn, that there exist N 1 such that u(t) ≤ N 1 , for all t ≥ 0, and the proof is done. We have given sufficient conditions for (3.1)-(3.3) to satisfy (W1), (W2) and (W4), (W5). It is left to be shown that the solution depends continuously on the initial data. This is shown in the following theorem. Theorem 3 The solution in Proposition 2 depends continuously on the initial data. 2 Proof Let (u, v) and (ũ, ṽ) be the solutions corresponding to initial data (u0 , v0 ) resp. (ũ0 , ṽ0 ) in CL0 ∞ . We shall show that ũ, ṽ → u, v whenever ũ0 , ṽ0 → u0 , v0 where both limits are taken in CL0 ∞ . As before, we choose α ∈ (0, 1) and p > 1 such that n/(2p) < α and αq < 1. Readily from (3.4), (3.6) and 2◦ in Corollary 3: ku(t) − ũ(t)kD(Hα ) ≤ Cα,p t−α ku0 − ũ0 kp Z t + Cα,p (t − τ )−α L11 ku(τ ) − ũ(τ )kp + L12 kv(τ ) − ṽ(τ )kp dτ, 0 kv(t) − ṽ(t)k D(Hα ) ≤ Cα,p t−α kv0 − ṽ0 kp Z t + Cα,p (t − τ )−α L21 ku(τ ) − ũ(τ )kp + L22 kv(τ ) − ṽ(τ )kp dτ, 0 for every t ∈ [δ, T ] where 0 < δ < T < ∞. And then by Hölder’s inequality and (3.34): Z t p/q p −αp p 1 1 −αq p p ku0 − ũ0 kp + Λ1 (p, L1 , L2 ) (t − τ ) dτ ku(t) − ũ(t)kD(Hα ) ≤ 2 Cα,p δ 0 × Z 0 t ku(τ ) − ũ(τ )kpp + kv(τ ) − ṽ(τ )kpp dτ ! p ≤ 2p Cα,p δ −αp ku0 − ũ0 kpp + Λ̃1 (p, L11 , L12 , T ) ! Z t p p ku(τ ) − ũ(τ )kD(Hα ) + kv(τ ) − ṽ(τ )kD(Hα ) dτ , × 0 (3.49) 46 Existence and uniqueness of two-component reaction-diffusion system and similarly for v, p kv(t) − ṽ(t)kpD(Hα ) ≤ 2p Cα,p δ −αp kv0 − ṽ0 kpp + Λ̃2 (p, L11 , L12 ) × Z t 0 ku(τ ) − ũ(τ )kpD(Hα ) + kv(τ ) − ṽ(τ )kpD(Hα ) ! dτ , (3.50) for some positive constants Λ1 , Λ2 , Λ̃1 and Λ̃2 . By adding (3.49) and (3.50) we obtain: p ku(t) − ũ(t)kpD(Hα ) + kv(t) − ṽ(t)kpD(Hα ) ≤ 2p Cα,p δ −αp ku0 − ũ0 kpp + kv0 − ṽ0 kpp p Λ̃1 (p, L11 , L12 ) + Λ̃2 (p, L11 , L12 ) + 2p Cα,p Z t × ku(τ ) − ũ(τ )kpD(Hα ) + kv(τ ) − ṽ(τ )kpD(Hα ) dτ. 0 p Λ1 (p, L11 , L12 ) + Λ2 (p, L11 , L12 ) , then Now put Π = 2 Cα,p δ and χ = 2p Cα,p by Gronwall’s inequality, see App. C and Lemma 12: p −α ku(t) − ũ(t)kpD(Hα ) + kv(t) − ṽ(t)kpD(Hα ) ≤ Π ku0 − ũ0 kpp + kv0 − ṽ0 kpp exp (tχ) →0 as ũ0 , ṽ0 → u0 , v0 in Lp , for every t ∈ [δ, T ]. Since δ and T were chosen arbitrarily this shows the continuous dependency of the initial data. By Remark 2 we therefore have: Corollary 7 The solution of the Schnakenberg and Thomas system and the Gierer and Meinhardt system depend continuously on the non-negative resp. positive initial conditions. 2 3.4 Summary and remarks In this chapter, the general two component reaction-diffusion system was considered on a smooth and bounded domain.3 First, we showed that the solution, 3 In fact, by referring to Grisvard, [Grisvard (1985)], the results in this chapter may be extended to piecewise smooth, convex domain; for example a cube in R3 . 3.4 Summary and remarks 47 when it exists and is classical, is positive whenever f (r, s) ≥ 0 and g(r, s) ≥ 0 2 for all r, s ∈ R+ . Next, we gave sufficient conditions for the solution to exist and to be continuous. Under these conditions, it was then shown that if the solution only existed on a finite time-interval, it diverged in CL0 ∞ . This result was used to prove global existence and boundedness for the Schnakenberg for n = 1 and for the Thomas model for n = 1, 2 and 3. Then, we gave sufficient conditions for the solution to be global and uniformly bounded. The reaction functions of the Schnakenberg model was shown to satisfy these conditions, and therefore the solution of the Schnakenberg system was shown to exist globally and be uniformly bounded. Finally, it was shown that under the conditions given, the solution depends continuously on the initial data. To sum up, sufficient conditions were given for the general reaction-diffusion model to be well-posed in the sense of Hadamard, (W1)-(W3), and to satisfy the “practical requirements” (W4) and (W5) in Chap. 1. In turn, this implied that (W1)-(W5) were satisfied for all three models. As a result of the analysis in this chapter, we can perform a linear analysis. This is done in next chapter, where we study the stability of uniform steadystates of the autonomous reaction-diffusion model, and give sufficient conditions for the system to exhibit diffusion driven instabilities. 48 Existence and uniqueness of two-component reaction-diffusion system Chapter 4 Linear analysis Alan Turing (1952) was the first to formulate an explanation of how the patterns of animals like leopards, jaguars and zebras develop. Turing asserted that the patterns can arise as a result of instabilities in the diffusion of morphogenetic chemicals in the animals’ skins during the embryonic stage of development.1 Mathematically he studied reaction-diffusion models and his idea was that, if in the absence of diffusion, the solution tended to a linear stable uniform state state then under certain conditions spatially inhomogeneous patterns could evolve by diffusion driven instability. In this chapter we derive the conditions for the general autonomous reactiondiffusion system (1.1)-(1.3) to exhibit diffusion driven instability, also called Turing instability. We assume that all the hypothesis of Chap. 3 are satisfied such that the solution (u, v) exists globally and is uniformly bounded for initial data in CL0 ∞ (Ω). First we need some definitions and results on stability. 1 Morphogenesis is the part of embryology that is concerned with pattern and form. 50 Linear analysis 4.1 Stability In the following let (u∗ , v ∗ ) be a steady-state solution of the autonomous equation (1.1). Definition 2 (u∗ , v ∗ ) is stable if, for any ǫ > 0, there exist δ > 0 such that any solution (u, v) with ku(0) − u∗kD(Hα ) , kv(0) − v ∗ kD(Hα ) < δ satisfies ku(t) − u∗ kD(Hα ) , kv(t) − v ∗ kD(Hα ) < ǫ for all t > 0. (u∗ , v ∗ ) is asymptotically stable if it is stable and k(u(t) − u∗ , v(t) − v ∗ )kD(Hα )2 → 0. (u∗ , v ∗ ) is unstable if it is not stable. Now suppose: f (u∗ + w1 , v ∗ + w2 ) w1 + Q(w1 , w2 ), = A w2 g(u∗ + w1 , v ∗ + w2 ) 2 (4.1) where A is a matrix in R2×2 , and kQ(w1 ) − Q(w2 )kL2 ×L2 ≤ h(ρ)kw1 − w2 kD(Hα )2 , (4.2) for kw1 kD(Hα )2 , kw2 kD(Hα )2 ≤ ρ, for some function h : R+ → R+ , continuous in 0 with h(0) = 0. By subtracting Aw on both sides we obtain an equation for the perturbed solution w, ∂t w + Lw = Q(w), (4.3) where L = −D∆ − A. If the spectrum of L lies in the right half-plane, or equivalently if the linearisation, ∂t w − D∆w = Aw, , (4.4) is asymptotically stable, then we say that (u∗ , v ∗ ) is linearly stable. On the other hand if the intersection σ(L) ∩ {λ ∈ C | Re λ < 0} is non-empty, then we say that (u∗ , v ∗ ) is linearly unstable. The following theorem is essential in the analysis to come. Theorem 4 We have the following implications: (1) (u∗ , v ∗ ) is linearly stable ⇒ (u∗ , v ∗ ) is asymptotically stable; (2) (u∗ , v ∗ ) is linearly unstable ⇒ (u∗ , v ∗ ) is unstable. 2 4.1 Stability 51 Proof (1): L is sectorial and therefore generates an analytical semigroup G. Let η > 0 be such that Re λ > η whenever λ ∈ σ(L). By Lemma 10 in App. A there exist Φ ≥ 1 such that kG(t)wkD(Hα )2 ≤ Φ exp(−ηt)kwkD(Hα )2 , Φt−α exp(−ηt)kwkL2 ×L2 . Now by (4.2) we choose ̺ > 0 and ρ > 0 so that Z ∞ 1 ̺Φ ξ −α exp(−(η − η ′ )ξ)dξ < , 2 0 2 2 α 2 kQ(w)kL ×L ≤ ̺kwkD(H ) , for kwkD(Hα )2 ≤ ρ, where 0 < η ′ < η. ρ Let kw(0)kD(Hα )2 ≤ 2Φ . Then by continuity of the solution kw(t)kD(Hα )2 ≤ ρ on some time interval and therefore by variation of constants: kw(t)kD(Hα )2 ≤ Φ exp(−ηt)kw(0)kD(Hα )2 Z t (t − τ )−α exp(−η(t − τ ))kw(τ )kD(Hα )2 dτ + ̺Φ 0 ρ ≤ + ρ̺Φ 2 Z t 0 (t − τ )−α exp(−η(t − τ ))dτ < ρ. (4.5) Again by continuity, either kw(t)kD(Hα )2 < ρ for all t > 0 or kw(t)kD(Hα )2 = ρ for some finite t. The second case contradicts the sharp inequality in (4.5), and thus kw(t)kD(Hα )2 < ρ for all t > 0. Now put Θ(t) = sup kw(ξ)kD(Hα )2 exp(η ′ ξ), t ≥ 0. 0≤ξ≤t By (4.5): kw(ξ)kD(Hα )2 exp(η ′ ξ) ≤ Φ exp(−(η − η ′ )t)kw(0)kD(Hα )2 Z ξ + Φ̺ (ξ − τ )−α exp(−(η − η ′ )(ξ − τ ))dτ × Θ(ξ) 0 ≤ (using (4.2)) ρ 1 ≤ + Θ(t), 2 2 for all 0 ≤ ξ ≤ t and thus Θ(t) ≤ ρ 1 + Θ(t), 2 2 or Θ(t) ≤ ρ. 52 Linear analysis Hence kw(t)kD(Hα )2 ≤ ρ exp(−η ′ t) showing that (u∗ , v ∗ ) is asymptotically stable. (n) (n) (2): We shall show that there exist ǫ > 0 and a sequence {(u0 , v0 )}n∈N ∈ (n) (n) L2 × L2 such that k(u0 , v0 ) − (u∗ , v ∗ )kD(Hα )2 → 0 as n → ∞, but sup k(un , vn ) − (u∗ , v ∗ )kD(Hα )2 ≥ ǫ > 0, t≥0 ∀n ∈ N, (n) (n) where un and vn are the solutions corresponding to the initial data u0 , v0 . Let us assume that 0 ∈ / σ(L). If 0 ∈ σ(L) we may apply the reasoning below by perturbing L. Therefore by assumption, there exist β > 0 such that the spectrum is disjoint from the ball in C with centre 0 and radius 2β. See Fig. 4.1. r ≤ 2β λ1 ... λJ λJ+1 ... R Figure 4.1: Eigenvalues of the operator L. The eigenvalues are assumed to be real. Let σ1 = σ(L) ∩ {λ ∈ C | Re λ < 0} and σ2 = σ(L)\σ1 . Notice that σ1 is a set of finite numbers in C. By construction L may be diagonalised such that Lw = ∞ X n=1 λn (w, wn )L2 ×L2 wn . We set X1 = span{wn | n = 1, . . . , J}, and L2 ×L2 X2 = span{wn | n ∈ N\{1, . . . , J}} , such that L2 × L2 = X1 ⊕ X2 . The projection operators onto X1 and X2 is denoted P1 resp. P2 . Now, let Li restriction of L to Xi , i = 1, 2, 4.1 Stability 53 then L1 is finite dimensional, and therefore bounded and generates an analytical semigroup G1 (t), so that there exist χ1 and χ2 , such that for t ≤ 0: kG1 (t)wkD(Hα )2 ≤ χ1 exp(2βt)kP1 wk2 , χ2 exp(2βt)kP1 wkD(Hα )2 . Similarly, L2 is sectorial and therefore by Lemma 10: kG2 (t)wkD(Hα )2 ≤ χ3 exp(−βt)kP2 wk2 , χ3 t−α exp(−βt)kP2 wkD(Hα )2 , for t > 0. Let χ = max{χ1 , χ2 , χ3 , 1}. Now we claim that Ξ given by the expression: Ξ(t) = G1 (t − ξ)σ + + Z t −∞ Z t ξ G2 (t − τ )P1 Q(Ξ(τ )) G2 (t − τ )P2 Q(Ξ(τ )), (4.6) solves (4.3) for σ ∈ X1 and t < ξ. To show the assertion we define the operator T by TΞ(t) = G1 (t − ξ)σ + + Z t −∞ Z ξ t G2 (t − τ )P1 Q(Ξ(τ ))dτ G2 (t − τ )P2 Q(Ξ(τ ))dτ, t ≤ ξ. By similar arguments to the proof of Proposition 2, it is shown that T maps Bρ (0) = ( w∈ 2 CL0 ∞ ((0, ∞); D(Hα )) ) | kwkD(Hα )2 ≤ ρ , where ρ > 0 is such that Z ∞ 1 1 −1 −α , s exp(−βs)ds ≤ χk(ρ) kP1 k β + kP2 k 2 4χ 0 (4.7) into itself. The following computation, then shows that T is a contraction on 54 Linear analysis Bρ (0): kTΨ(t) − TΞ(t)kD(Hα )2 Z t ≤ χk(ρ) exp(2β(t − τ ))kP1 kkΨ(τ ) − Ξ(τ )kD(Hα )2 dτ ξ Z t χk(ρ) exp(−β(t − τ ))(t − τ )−α kP2 k + −∞ × kΨ(τ ) − Ξ(τ )kD(Hα )2 dτ Z t ≤ χk(ρ)kP1 k exp(2β(t − τ ))dτ ξ Z t + χk(ρ)kP2 k| (t − τ )−α exp(−β(t − τ ))dτ | −∞ × sup kΨ(t) − Ξ(t)kD(Hα )2 t≤ξ Z ∞ 1 ξ −α exp(−βξ)dξ ≤ χk(ρ) kP1 k β −1 + kP2 k 2 0 × sup kΨ(t) − Ξ(t)kD(Hα )2 t≤ξ < 1 × sup kΨ(t) − Ξ(t)kD(Hα )2 , 2 t≤ξ by choice of ρ, and for Ψ, Ξ ∈ Bρ (0). Hence, by Banach’s Fixed Point Theorem there is a unique fix point in Bρ (0). To show that Ξ solves (4.3), we consider the projections of Ξ onto X1 and X2 . First, P1 Ξ(t) = G1 (t − ξ)σ + Z ξ t G1 (t − τ )P1 Q(Ξ(τ ))dτ = G1 (t)G1 (−s)σ + G1 (t) + Z t Z 0 ξ G1 (−ξ)P1 Q(Ξ(τ ))dτ G1 (t − τ )P1 Q(Ξ(τ ))dτ, 0 for 0 ≤ t ≤ ξ, and similarly P2 Ξ(t) = G2 (t) + Z 0 t Z 0 −∞ G2 (t − τ )P2 Q(Ξ(τ ))dτ G2 (t − τ )P2 Q(Ξ(τ ))dτ, for 0 ≤ t ≤ ξ. 4.1 Stability 55 Hence, Ξ(t) = P1 Ξ(t) + P2 Ξ(t) = G(t) G1 (−ξ)σ + Z ξ + Z Z G1 (−s)P1 Q(Ξ(τ ))dτ 0 −∞ + 0 G2 (t − τ )P2 Q(Ξ(τ ))dτ ! t G(t − τ )Q(Ξ(τ ))dτ Z t = G(t)Ξ(0) + G(t − τ )Q(Ξ(τ ))dτ, 0 0 for 0 ≤ t ≤ ξ, where Gw = G1 P1 w + G2 P2 w. By referring to (3.4) and the discussion in Chap. 3 we conclude that Ξ solves (4.3) with w(0) = Ξ(0). Next, we show that kΞ(t)kD(Hα )2 ≤ 2χkσkD(Hα)2 exp(2β(t − ξ)) for t ≤ ξ. By (4.6): kΞ(t)kD(Hα )2 ≤ χ exp(2β(t − ξ))kσkD(Hα )2 Z ∞ 1 −α + χk(ρ) kP1 k ξ exp(−βξ)dξ + kP2 k 2β 0 × sup kΞ(τ )kD(Hα )2 , 0≤s≤ξ ∀t ≤ ξ, but then sup kΞ(τ )kD(Hα )2 ≤ χ exp(2β(t − ξ))kσkD(Hα )2 + 0≤s≤ξ 1 sup kΞ(τ )kD(Hα )2 . 2 0≤s≤ξ Therefore, kΞ(t)kD(Hα )2 ≤ sup kΞ(τ )kD(Hα )2 ≤ 2χ exp(2β(t − ξ))kσkD(Hα )2 , 0≤s≤ξ ∀t ≤ ξ. (4.8) 56 Linear analysis D(Ãα ) Θ(t) Θ(s) − σ 0 r = 1/2kσk D(Ãα )2 σ r = 1/2kσk D(Ãα ) D(Ãα )2 Figure 4.2: By (4.9) we may conclude that kΞ(ξ)kD(Hα )2 ≥ 12 kσkD(Hα )2 . Now we are ready for the final estimate. By (4.6) we obtain: kΞ(ξ) − σkD(Hα )2 ≤ Z ξ −∞ χ(ξ − s)−α exp(−β(ξ − s))kP2 k × 2χ exp(2β(ξ − s))kσkD(Hα )2 dτ Z ξ = 2χ2 kP2 kkσkD(Hα )2 (ξ − s)−α exp(−3β(t − τ ))dτ ≤ (using (4.7)) 1 ≤ kσkD(Hα )2 . 2 −∞ (4.9) But then, kΞ(ξ)kD(Hα )2 ≥ 12 kσkD(Hα )2 , see e.g. Fig. 4.2. The theorem now follows, since if we choose kσkD(Hα )2 ≤ by (4.8), kw(0)kD(Hα )2 = kΞ(0)kD(Hα )2 ≤ ρ exp(−2βn) → 0, ρ 2χ for and ξ = n then n → ∞, while sup kw(t)kD(Hα )2 ≥ kw(n)kD(Hα )2 ≥ 0≤t≤n and therefore (u∗ , v ∗ ) is unstable. 1 kσkD(Hα )2 , 2 ∀n ∈ N Linear stability is therefore a sufficient condition for stability of the nonlinear system. This shall be utilised in the following section to present sufficient conditions for the general autonomous system to exhibit diffusion driven instability. 4.2 Conditions for diffusion driven instability 4.2 57 Conditions for diffusion driven instability In the analysis to come, it turns out that the size of the domain Ω is an important parameter. By introducing the scaling γ and x = γ 1/2 x̃ and t = γ t̃, the system is transformed into: ˜ = γf (u, v) ∂t̃ u − ∆u ˜ = γg(u, v) in Ω̃ × (0, ∞), ∂t̃ v − d∆v u ν̃ = 0, on ∂ Ω̃ × [0, ∞), v u = u0 , v = v0 , on Ω̃ × {t = 0} , where Ω̃ = x ∈ Rn | γ −1/2 x ∈ Ω , and by varying γ we can therefore study the influence of the domain size on the solutions. For convenience the tilde-notation is omitted in the following. We assume that there exists a uniform steady-state (u∗ , v ∗ ) ∈ R2+ , i.e. so that f (u∗ , v ∗ ) = 0 = g(u∗ , v ∗ ). For notational simplicity superscript ∗ is used when f and g or their partial derivatives are evaluated at (u∗ , v ∗ ). For example f ∗ = f (u∗ , v ∗ ). By Taylor expansion: ∂r f ∗ ∂s f ∗ A= . (4.10) ∂r g ∗ ∂s g ∗ Let us define what we precisely mean by diffusion driven instability. Definition 3 The autonomous reaction-diffusion model has a diffusion-driven instability at (u∗ , v ∗ ) if the following two conditions are satisfied: (1) (u∗ , v ∗ ) is asymptotically stable in the absence of diffusion; (2) (u∗ , v ∗ ) is unstable in the presence of diffusion. 2 First, the sufficient conditions for (1) to hold is determined. In the absence of spatial dependency (4.4) simplifies to ∂t w = γAw, (4.11) where w= u − u∗ v − v∗ . 58 Linear analysis The ordinary differential equation (4.11) is asymptotically stable if the eigenvalues of γA have negative real part. The eigenvalues are given by the characteristic equation, λ2 − γtr A λ + γ 2 det A = 0, with solutions: λ± = Re λ± < 0 if and only if 1/2 i 1 h . γ tr A ± tr A2 − 4det A 2 tr A = ∂r f ∗ + ∂s g ∗ < 0, det A = ∂r f ∗ ∂s g ∗ − ∂s f ∗ ∂r g ∗ > 0. (4.12) Therefore by Theorem 4, (1) in Definition 3 is satisfied if (4.12) holds. For (2) in Definition 3, we consider the linear system (4.4) and determine the conditions for linear instability and then refer to Theorem 4. It is well-known by preliminary partial differential equation theory that the system (4.4) can be solved by the method of product solutions. We write: w(x, t) = ∞ X cki exp (λki t) Xki (x) (4.13) i=1 where {cki }i∈N depend upon the initial conditions and {Xki }i∈N are the eigenfunctions of the Laplacian: −∆Xki = ki2 Xki in Ω, νXki = 0 on ∂Ω. (4.14) We shall call {ki }i∈N the wavenumbers. (u∗ , v ∗ ) is linearly unstable if and only if Re λki > 0 for some i ∈ N. It is well-known and indeed verified in Example 1 that the ki ’s are real, and they may be arranged such that 0 ≤ k12 ≤ · · · ≤ kJ2 ≤ . . . , and ki2 → ∞ for i → ∞. Inserting cki exp (λki t) Xki (x) into (4.13) we obtain an equation relating λki and ki2 : λki I − γA + ki2 D Xki = 0, ∀i ∈ N. To yield non-trivial solutions the coefficient matrix must be singular, that is, det λki I − γA + ki2 D = 0. The determinant is evaluated, and written as follows: −γ∂r g ∗ λki − γ∂r f ∗ + ki2 det −γ∂s f ∗ λki − γ∂s g ∗ + dki2 = λ2ki + ki2 (1 + d) − γ (∂r f ∗ + ∂s g ∗ ) λki + h(ki2 ) 4.2 Conditions for diffusion driven instability 59 where h(ki2 ) = dki4 + γ 2 det A − γ (d∂r f ∗ + ∂s g ∗ ) ki2 . (4.15) The solutions of (4.2) are: where 1/2 , 2λki = −B(ki2 ) ± B(ki2 )2 − 4h(ki2 ) (4.16) B(ki2 ) = ki2 (1 + d) − γtr A > 0, (4.17) by (4.12). Let us consider λ+ : R+ → C defined by (4.16): 1/2 1 . −B(ki2 ) + B(ki2 )2 − 4h(ki2 ) λ+ (k 2 ) = 2 Now, Re λ+ (k 2 ) > 0 for some k 2 > 0 if and only if h(k 2 ) < 0. In fact, at bifurcation the simple eigenvalue λ+ crosses the imaginary axis in 0. Hence Hopf’s Bifurcation Theorem is not applicable, and so we do not expect there to exist periodic solutions.2 Next, we determine the conditions for h(k 2 ) < 0 for some k 2 > 0. Since h is continuous and h → +∞ whenever k 2 → ∞, h attains a minimum on [0, ∞). Elementary differentiation of h(k 2 ) with respect to k 2 shows that, " # 2 (d∂r f ∗ + ∂s g ∗ ) 2 2 2 min h(k ) = h(k̃ ) = γ det A − , (4.18) R+ 4d where k̃ 2 = γ (d∂r f ∗ + ∂s g ∗ ) . 2d (4.19) Thus h(k 2 ) < 0 for some k 2 > 0 if and only if the following conditions are satisfied: d∂r f ∗ + ∂s g ∗ > 0 2 (d∂r f ∗ + ∂s g ∗ ) > det A. 4d The first condition is also apparent from (4.15): Cf. the requirement that det A > 0, h(k 2 ) can only attain negative values for some positive parameters if d∂r f ∗ + ∂s g ∗ > 0. From (4.12) we have tr A = ∂r f ∗ + ∂s g ∗ < 0, and therefore d∂r f ∗ + ∂s g ∗ > 0 for d > 0 implies that ∂r f ∗ > 0 while ∂s g ∗ < 0 and d > 1. Furthermore, by the condition det A > 0 it follows that ∂s f ∗ and ∂r g ∗ must be of opposite signs. Now, observe that by (4.19) and (4.18), we can choose γ such that λ+ (kJ2 ) > 0, some J ∈ N. Therefore by Theorem 4 we obtain: 2 This is not precise mathematically, we shall, however, not study this further in this thesis. 60 Linear analysis Theorem 5 If the following conditions are satisfied: (T1) tr A < 0; (T2) det A > 0; (T3) d∂r f ∗ + ∂s g ∗ > 0; (T4) (d∂r f ∗ +∂s g∗ )2 4d > det A; then, there exists a positive γc , such that for γ = γc , λ+ (k12 ) = 0. For γ < γc , (u∗ , v ∗ ) is stable, while for γ > γc , the autonomous reaction-diffusion system exhibits diffusion-driven instabilities at (u∗ , v ∗ ) when there exist wavenumbers kJ2 , J ∈ N, in the range (m2 , M 2 ), where m2 and M 2 are the zeros of h(k 2 ) = 0: h 1/2 i γ m2 = 2d (d∂r f ∗ + ∂s g ∗ ) − (d∂r f ∗ + ∂s g ∗ )2 − 4d det A (4.20) i h 1/2 γ ∗ ∗ ∗ ∗ 2 2 2 (d∂r f + ∂s g ) + (d∂r f + ∂s g ) − 4d det A M = 2d In the last section of this chapter we shall give a qualitative discription of the result of Theorem 5. The set of parameter values satisfying (T1)-(T4), are called the Turing domain and denoted T . In the numerical analysis we shall, for the Schnakenberg system, make explicit use of the Turing domain given by the inequalities (T1)-(T4). As mentioned in Chap. 2, the Schnakenberg system exhibits one single positive, uniform steady-state given by u∗ = a + b, v∗ = b 2, (a + b) b > 0, a + b > 0. The partial derivatives of f and g at this steady-states are b−a , ∂s f ∗ = (a + b)2 , a+b −2b ∂r g ∗ = , ∂s g ∗ = −(a + b)2 . a+b ∂r f ∗ = Inserting these into (T1)-(T4) we obtain: 0 < b − a < (a + b)3 , (4.21) (a + b) > 0, (4.22) d(b − a) > (a + b)3 , (4.23) 2 3 2 4 [d(b − a) − (a + b) ] > 4d(a + b) . (4.24) 4.3 Summary and remarks 61 Clearly (4.22) is redundant. Moreover the lower bound in (4.21) is redundant by the presence of (4.23) and we can therefore settle with (a + b)3 > b − a, 3 1/2 d(b − a) − (a + b) > 2d (4.25) 2 (a + b) . (4.26) Or simply: d(b − a) > (by (4.26)) > 2d1/2 (a + b)2 + (a + b)3 > (by (4.25)) > 2d1/2 (a + b)2 + b − a. Remark 9 Obviously, (4.21) is not satisfied for ab < 41 . Therefore, for ab < 41 the unique solution (u∗ , v ∗ ) in the sector h(u, v), (u, v)i, see Chap. 2, is asymptotically stable both in absence and presence of diffusion. 2 Now, let a resp. a be the real and non-negative solutions of the cubic equations: d(b − a) = 2d1/2 (a + b)2 + (a + b)3 , 3 (a + b) = b − a. (4.27) (4.28) It is not hard to see that given d > 1 and b > 0, then the following inclusion holds: Sb,d = (a, b, d) ∈ R2+ × (1, ∞) | a < a < a ⊂ T . Of course Sb,d may be empty for some d > 1 and b > 0. The expressions for the solutions of (4.27) and (4.28) are quite messy: 2/3 √ − d 2√ d3/2 + 27db + 3 6d5/2 b + 81d2 b2 − a(b, d) = d − b, (4.29) 1/3 √ 3 3/2 5/2 2 2 3 d + 27db + 3 6d b + 81d b √ 2/3 −3 27b + 3 3 + 81b2 (4.30) a(b) = √ 1/3 − b. 3 27b + 3 3 + 81b2 4.3 Summary and remarks In this chapter the linear stability analysis was shown to give sufficient conditions for the stability of the full nonlinear systems. This result was used 62 Linear analysis to analyse the diffusion-driven instabilities of an autonomous reaction-diffusion system, and the results were stated in a theorem, Theorem 5. These conditions were used to determine relevant parameters for the Schnakenberg model. By Theorem 5, it followed that by changing γ it is possible to change the stability of the uniform steady-state. Therefore, γ is a bifurcation parameter. For γ < γc , (u∗ , v ∗ ) is stable, however, for γ > γc , (u∗ , v ∗ ) may be unstable by diffusion. Now, we shall give a qualitative description of the result presented in Theorem 5. To do so, let us assume that for γ > γc , diffusion driven instabilities eventually evolves into a spatially inhomogeneous steady-state solution. Then, if the domain were to present some embryo domain and u and v represented some morphogenetic concentrations, we would expect the kinetics and diffusion coefficients to be fixed, meaning that γ would be the natural variable parameter reflecting the size of the embryo. Theorem 5, together with the assumption above then asserts that for sufficient large domains, random perturbations evolve into stable, spatial patterns. On the other hand, for sufficient small domain the uniform steady-state is asymptotically stable for small random perturbations, and no pattern evolves. This is in agreement with for example mammals such as zebras, leopards or giraffes, where the formation of coat patterns does not take place until the embryo has reached a certain size. See e.g. [Murray (2003), Chapter 3]. It is also in accordance with the fact that small mammals are often uniform in colour; rats, hedgehogs and so on. The pattern of large one-coloured mammals such as elephants, can be the result of a very fine scale pattern, so fine that essentially no patterns can be seen. See e.g. [Murray (2003), Chapter 3, p.152]. In fact, reaction-diffusion equations have been applied from a patterning point of view to a larger number of ecology situations, and it has been possible to obtain similar patterns to those seen in nature, see e.g. zebra and giraffe coat patterns in [Murray (2003), Chapter 3, Figure 3.4 and 3.5].3 3 Diffusion-driven instabilities have not only found interest in biology and ecology but also in economics. See e.g. http://www.soc.uoc.gr/ecosud/docs/Zurich%20presentation /Zurich%20presentation.ppt Chapter 5 Numerical analysis In this chapter the results from implementing the Schnakenberg system in MATLAB in the form: ∂t u − ∆u = γ(a − u + u2 v) in Ω × (0, ∞), ∂t v − d∆v = γ(b − u2 v) u ν = 0, on ∂Ω × [0, ∞), v u = u0 , v = v0 , on Ω × {t = 0} , are presented. We consider the model on the rectangle Ω = (0, 1) × 0, π2 . The rectangle is chosen since it is a simple geometry, and it is easy to handle numerically. Furthermore, it is advantageous that the wavenumbers are distinguishable from eachother on this domain. On a square domain the wavenumbers are not simple. On Ω the eigensolutions are: Xkn,m = X cos (nπx) cos (2my) , with corresponding wavenumbers: p kn,m = (nπ)2 + (2m)2 , X ∈ R2 n, m = 1, . . . , N, . . . . Chebyshev interpolation is used to approximate the Laplacian and the integration (in time) is performed by built-in MATLAB ode-solver ode15s. 64 Numerical analysis The developed code randomly perturbs the uniform steady-state solution, (u∗ , v ∗ ), and solve the system for the perturbed variables u − u∗ and v − v ∗ until the solution converge, measured by a tolerance, to a steady-state.1 The implementation was verified by considering the linearised system (4.4) for initial conditions (u0 , v0 ) = Xkn,m , and comparing the numerical solution with the analytical solution given by: u = c1 V1 exp(−σ1 t) cos(nπx) + c2 V2 exp(−σ2 t) cos(2mx), v where (σi , Vi ), i = 1, 2, are the eigensolutions of the algebraic eigenvalue problem: 2 −kn,m D + γA Vi = σi Vi , i = 1, 2, D = diag (1, d) and A is as in (4.10). The constants c1 and c2 are given by the linear system: c1 1 V1 V2 = . 1 c2 The numerical solution converged spectrally to the analytical solutions as the net was refined. Furthermore, the code evaluated the right hand side of the equation: Z f (u, v) . 0= g(u, v) Ω which follows by integration of the stationary equations for u and v, when a numerical steady-state solution was obtained for the full nonlinear system, and the equation was observed to be satisfied down to a sufficient tolerance (< 10−4 max(u∗ , v ∗ )). A convergence analysis of the numerical solution to the full nonlinear system, showed that a grid of 21 × 33 = 693 mesh points was adequate. 5.1 Results in Matlab In the numerical analysis we studied the behaviour of the system for six different (a, b, d)-values in the Turing domain. The six points in the Turing domain are seen in Fig. 5.1. 1 In practice, a tolerance is not specified. Instead, the code integrates from 0 to a large value T ≈ 106 . When |∂t u| ≤ ǫ, then the built-in ode15s solver takes steps with size of order O(ǫn ), where n > 1. So in principal, we analyse the convergence by looking at the t-values: If the solver extrapolates using small stepsizes, then the solution has not converged; if the solver extrapolates with a large stepsize, for example from t = 1 to t = 106 , then we conclude that the solution has converged. 5.1 Results in Matlab 65 0.7 Lower bound Upper bound for d =10 Upper bound for d =20 Upper bound for d =40 Upper bound for d =60 Upper bound for d =80 0.6 0.5 a 0.4 0.3 0.2 0.1 0 0 0.5 1 1.5 2 b 2.5 3 3.5 4 Figure 5.1: The parameters where diffusion driven stability occur, is, for fixed d > 1, the set confined by the the upper and the lower bound. Six points are shown. The red × is for d = 10, the black is for d = 80, while the green ◦’s are for d = 20. 66 Numerical analysis In Sec. 5.1.1 we analyse the two points (a, b, d) = (0.05, 1, 20) and (a, b, d) = (0.05, 1, 80). In Sec. 5.1.2 we consider the four remaining points (a, b, d) = (0.05, 1.5, 20), (0.1, 1.5, 20), (0.2, 0.5, 20) and (0.2, 1, 20). 5.1.1 Results for d = 10 and d = 80 In Fig. 5.2, Fig. 5.3 and Fig. 5.4 three type of developments of u are seen for a = 0.05, b = 1, d = 10 and γ = 120. Please note, that throughout it is the perturbed u and v that is visualised, i.e. u − u∗ resp. v − v ∗ . As expected the solution remains bounded and a spatial pattern develops. Furthermore, the perturbed u and v are larger than −u∗ resp. −v, so u and v are positive, as stated in Proposition 1. The final patterns in the three figures are different, so for these parameter values we have found multiple steady-states. What pattern, that ultimately develops, depends on the initial conditions. Also for v a final steady-state pattern develops. The v-pattern is similar to the u-pattern, but where u has a local maximum and minimum, v has a local minimum resp. maximum. The idea of Turing described in the previous chapter was that, u > u∗ would give rise to one colour, while u < u∗ would give rise to another colour. Therefore, the three patterns would imply stripes, spots resp. stripes. 5.1 Results in Matlab 67 * (u−u*)(0) u−u for t= 0.086461 −3 x 10 8 −3 x 10 6 1.5 1.5 6 4 4 2 2 1 1 0 y y 0 −2 −2 −4 0.5 0.5 −4 −6 −8 0 0 −6 −10 0.2 0.4 0.6 0.8 0 0 1 0.2 0.4 x 0.6 0.8 1 x * u−u* for t= 0.48595 (u−u )(t=∞) 0.5 0.4 1.5 1.5 0.4 0.3 0.3 0.2 0.2 1 0.1 1 0.1 0 −0.1 y y 0 −0.1 −0.2 −0.2 0.5 0.5 −0.3 −0.3 −0.4 −0.4 −0.5 −0.5 0 0 0.2 0.4 0.6 x 0.8 1 0 0 0.2 0.4 0.6 0.8 1 −0.6 x Figure 5.2: u at four different times for a = 0.05, b = 1, d = 10 and γ = 120. A spatial pattern develops. The final pattern is stripes. The v-pattern is similar in shape, but where u has a local maximum and minimum, v has a local minimum resp. maximum. 68 Numerical analysis * (u−u*)(0) u−u for t= 0.16821 −3 x 10 6 0.015 1.5 1.5 4 0.01 2 1 0 y y −2 0.005 1 0 −4 −0.005 0.5 0.5 −6 −0.01 −8 0 0 −10 0.2 0.4 0.6 0.8 0 0 1 −0.015 0.2 0.4 x 0.6 0.8 1 x * u−u* for t= 0.85823 (u−u )(t=∞) 0.8 0.8 1.5 1.5 0.6 0.6 0.4 0.4 1 1 0.2 y y 0.2 0 0.5 −0.2 0 0.5 −0.2 −0.4 0 0 −0.6 0.2 0.4 0.6 x 0.8 1 −0.4 0 0 −0.6 0.2 0.4 0.6 0.8 1 x Figure 5.3: u at four different times for a = 0.05, b = 1, d = 10 and γ = 120. A spatial pattern develops. The final pattern is spots. The v-pattern is similar in shape, but where u has a local maximum and minimum, v has a local minimum resp. maximum. 5.1 Results in Matlab 69 * (u−u*)(0) u−u for t= 0.10286 −3 x 10 6 0.01 1.5 1.5 0.008 4 0.006 0.004 2 1 0.002 1 0 y y 0 −2 −0.002 −0.004 −4 −0.006 0.5 0.5 −6 −0.008 −0.01 −8 −0.012 0 0 −10 0.2 0.4 0.6 0.8 0 0 1 0.2 0.4 x 0.6 0.8 1 x * u−u* for t= 0.50813 (u−u )(t=∞) 0.6 0.6 1.5 1.5 0.4 0.4 0.2 y 0 −0.2 0.5 0.2 1 0 y 1 −0.2 0.5 −0.4 −0.4 0 0 0.2 0.4 0.6 x 0.8 1 0 0 −0.6 0.2 0.4 0.6 0.8 1 x Figure 5.4: u at four different times for a = 0.05, b = 1, d = 10 and γ = 120. A spatial pattern develops. The final pattern is stripes. The v-pattern is similar in shape, but where u has a local maximum and minimum, v has a local minimum resp. maximum. 70 Numerical analysis Fig. 5.5 shows a pattern, similar to the pattern in Fig. 5.2. The difference is that local maximum and minimum in Fig. 5.2 are local minimum resp. maximum in Fig. 5.5. It was also possible to obtain similar alternatives for the patterns in Fig. 5.3 and Fig. 5.4. (u−u*)(t=∞) (v−v*)(t=∞) 0.6 0.2 1.5 1.5 0.15 0.4 0.1 0.2 1 0.05 1 −0.2 0.5 y y 0 0 −0.05 −0.1 0.5 −0.15 −0.4 0 0 0.2 0.4 0.6 x 0.8 1 −0.2 0 0 −0.25 0.2 0.4 0.6 0.8 1 x Figure 5.5: Spatial patterns for u and v, obtained for a = 0.05, b = 1, d = 10 and γ = 120. The pattern in u is similar to the pattern seen in Fig. 5.2, but local maximum and minimum in Fig. 5.2, are now local minimum resp. maximum. So here v is similar to u in Fig. 5.2. 5.1 Results in Matlab 71 In Fig. 5.6 the graph of Re λ+ as a function of k is seen for a = 0.05, b = 1, d = 10 and γ = 120. The wavenumbers kn,m , are indicated with ∗’s on the abscissa. Text-arrows describe the unstable wavenumbers. 20 k k 10 2,0 2,1 k1,3 0 −10 k Re λ+ −20 1,2 k k 0,3 k 2,2 0,4 −30 −40 −50 −60 −70 −80 0 2 4 6 k 8 10 12 Figure 5.6: The graph of Re λ+ as a function of k for a = 0.05, b = 1, d = 10 and γ = 120. The wavenumbers kn,m are indicated with ∗’s on the abscissa. 72 Numerical analysis It is not impossible to imagine, that there would exist six different steadystate patterns, each with a correspondency to the six unstable eigenfunctions. If such steady-states would exist, then it would be likely that (u0 , v0 ) = (ǫu , ǫv ) cos(πnx) cos(2my), (5.1) where 0 < ǫu ≪ u∗ and 0 < ǫv ≪ v ∗ , would be on the stable manifold of the corresponding steady-state solution. Therefore, for the first ∗ in Fig. 5.6, we let (u0 , v0 ) be as in (5.1). The solution corresponding to this initial condition converged to the steady-state solution in Fig. 5.2. Using initial conditions corresponding to the following three ∗’s in Fig. 5.6, k0,3 , k2,0 and k2,1 , the solutions converged to the steady-states in Fig. 5.4, Fig. 5.3 resp. Fig. 5.2. For the fourth ∗, k1,3 , the solution converged to the solution in Fig. 5.2, while for the fifth and sixth ∗, k2,2 resp. k0,4 , the solutions converged to the steady-state solution shown as a surface-plot in Fig. 5.7. * * (u−u )(t=∞) (v−v )(t=∞) 1 0.4 0.2 v−v* u−u* 0.5 0 0 −0.2 −0.4 2 −0.5 2 1.5 1 1 0.5 0.5 y 0 0 x 1.5 1 1 0.5 0.5 y 0 0 x Figure 5.7: u and v pattern for a = 0.05, b = 1, d = 10 and γ = 120. 5.1 Results in Matlab 73 In Fig. 5.8 six different spatial patterns are seen for a = 0.05, b = 1, d = 80 and γ = 120. Compared to the previous patterns for d = 10 it is obvious, that the increase in d is accompanied with a higher level of spatial aggregation in the stationary patterns. The steady-state solutions are shown to involve large peaks where the perturbed u is ≈ 5, such that the concentration is about six times larger than u∗ = 1.05, and areas where the concentration is constant and less than u∗ . The corresponding v-patterns are seen in Fig. 5.9. The v-patterns are also more spatial aggregated compared to d = 20, and again local maximum and minimum for u is local minimum resp. maximum for v. 74 Numerical analysis (u−u*)(t=∞) (u−u*)(t=∞) 5 1.5 5 1.5 4.5 4.5 4 4 3.5 3.5 3 3 1 2.5 2.5 2 2 y y 1 1.5 1.5 1 0.5 1 0.5 0.5 0.5 0 0 −0.5 −0.5 0 0 0.2 0.4 0.6 0.8 0 0 1 0.2 0.4 x 0.6 0.8 1 x * * (u−u )(t=∞) (u−u )(t=∞) 5 1.5 5 1.5 4.5 4.5 4 4 3.5 3.5 3 3 1 2.5 2.5 2 2 y y 1 1.5 1.5 1 0.5 1 0.5 0.5 0.5 0 0 −0.5 0 0 0.2 0.4 0.6 0.8 −0.5 0 0 1 0.2 0.4 0.6 x x (u−u*)(t=∞) (u−u*)(t=∞) 0.8 1 5 1.5 5 1.5 4.5 4.5 4 4 3.5 3.5 3 3 1 2.5 2.5 2 2 y y 1 1.5 1 0.5 1.5 1 0.5 0.5 0.5 0 0 −0.5 0 0 0.2 0.4 0.6 x 0.8 1 −0.5 0 0 0.2 0.4 0.6 0.8 1 x Figure 5.8: Six different u-patterns for a = 0.05, b = 1, d = 80 and γ = 120. 5.1 Results in Matlab 75 (v−v*)(t=∞) (v−v*)(t=∞) −0.25 1.5 1.5 −0.2 −0.3 −0.35 −0.3 1 −0.4 1 −0.45 y y −0.4 −0.5 −0.5 0.5 0.5 −0.55 −0.6 −0.6 −0.65 0 0 −0.7 0.2 0.4 0.6 0.8 0 0 1 0.2 0.4 x 0.6 0.8 1 x * * (v−v )(t=∞) (v−v )(t=∞) −0.3 −0.3 1.5 1.5 −0.35 −0.35 −0.4 1 −0.4 1 −0.5 −0.45 y y −0.45 −0.5 −0.55 −0.55 0.5 0.5 −0.6 −0.6 −0.65 0 0 0.2 0.4 0.6 0.8 −0.65 0 0 1 0.2 0.4 x 0.6 0.8 1 x (v−v*)(t=∞) (v−v*)(t=∞) −0.25 −0.3 1.5 1.5 −0.3 −0.35 −0.35 −0.4 −0.4 1 1 −0.45 −0.5 y y −0.45 −0.5 −0.55 0.5 −0.55 0.5 −0.6 −0.6 −0.65 −0.65 0 0 −0.7 0.2 0.4 0.6 x 0.8 1 0 0 0.2 0.4 0.6 0.8 1 −0.7 x Figure 5.9: v-patterns for a = 0.05, b = 1, d = 80 and γ = 120 corresponding to the u-patterns in Fig. 5.8. 76 Numerical analysis In Fig. 5.10 we consider the bifurcation of the steady-states as γ is varied. The bifurcation functional, is the Sobolev-norm H 1 of the perturbed steadystates, i.e. ku − u∗ kH 1 and kv − v ∗ kH 1 . In the topmost figure the unstable wavenumbers are made visible. The blue ∗’s indicate the wavenumber with largest corresponding λ+ , while the ▽’s and the ’s indicate when the corresponding wavenumber changes stability; from stable to unstable resp. unstable to stable. In the two buttom plots the H 1 norm of the perturbed steady-states are seen, u − u∗ resp. v − v ∗ . In the the two buttom plots several branches are seen, and in the region where the wavenumber with largest corresponding λ+ changes, indicated by the blue ∗’s in the topmost figure, it is seen that a single γ-value corresponds to two or more values of the H 1 -norm. The branches correspond to different type of solutions, and therefore in areas where we have two or more values of the H 1 -norm there exists multiple steady-states solutions. This in accordance to the plots above. It is also seen, that γc = 6.5 which is in accordance to the linear analysis. 5.1 Results in Matlab 77 10 0 0 4 20 40 60 80 100 120 140 20 40 60 80 100 120 140 20 40 60 80 100 120 140 2 0 0 1.5 1 0.5 1 H −norm of v−v * H1−norm of u−u* k 5 0 γ Figure 5.10: The bifurcation of steady-states for a = 0.05, b = 1 and d = 10. The topmost figure shows when the wavenumbers bifurcate. The topmost figure shows when the wavenumbers bifurcate. The blue ∗’s indicate the wavenumber with largest corresponding λ+ , while the ▽’s and the ’s indicate when the corresponding wavenumber changes stability; from stable to unstable resp. unstable to stable. In the two buttom figures H 1 -norms of the perturbed steady-states are shown for different γ-values. It is made clear that multiple, stable, steadystates occur near the areas where the blue ∗’s changes from one wavenumber to another. 78 Numerical analysis In Fig. 5.11 the similar bifurcation diagram is seen for d = 80. Again, multiple steady-state solutions are seen, corresponding to different branches. However, the branches are closer together, and more difficult to separate than for d = 10. This is due to the long width of the blue ∗’s in the topmost figure, showing the wavenumber with largest corresponding λ+ -value. At least for low γ-values, this results in an increased propability, compared to d = 10 where these blue ∗’s branches were shorter, of a specific type of solutions. Thus we see long branches in both of the two lower plots. We also note that the length between the ▽’s and the ’s in the topmost plot is larger for d = 80 than for d = 10. The H 1 -norms of u − u∗ for d = 80 is in general larger than those for d = 10. As seen, this was accompanied by a higher level of spatial aggregation in the patterns for d = 80. For v − v ∗ the norms are comparable, but the H 1 -norm of v − v ∗ for d = 10 is bit larger than that for d = 80. Moreover, γc = 4.5, again in accordance to the linear analysis. 5.1 Results in Matlab 79 5 0 0 15 20 40 60 80 100 120 140 20 40 60 80 100 120 140 20 40 60 80 100 120 140 10 5 0 0 1 0.5 1 H −norm of v−v * H1−norm of u−u* k 10 0 γ Figure 5.11: The bifurcation of steady-states for a = 0.05, b = 1 and d = 10. The topmost figure shows when the wavenumbers bifurcate. The blue ∗’s indicate the wavenumber with largest corresponding λ+ , while the ▽’s and the ’s indicate when the corresponding wavenumber changes stability; from stable to unstable resp. unstable to stable. In the two buttom figures H 1 -norms of the perturbed steady-states are shown for different γ-values. It is made clear that multiple, stable, steady-states occur near the areas where the blue ∗’s changes from one wavenumber to another. The H 1 -norms of u − u∗ are in general larger than the values in Fig. 5.10, while the norms for v − v ∗ are a bit smaller. 80 5.1.2 Numerical analysis Results for d = 20 In this section we consider the four different points in the Turing domain, (a, b, d) = (0.05, 1.5, 20), (0.1, 1.5, 20), (0.2, 0.5, 20) and (0.2, 1, 20). In Fig. 5.12, Fig. 5.13, Fig. 5.14 and Fig. 5.15 the corresponding bifurcation diagrams are seen. To compare the size of the H 1 -norm for the different a, b and d-values, we have introduced the relative H 1 -norms, Hu1 and Hv1 , given by 1 ku − u∗ kH 1 , u∗ resp. 1 kv − v ∗ kHv1 ≡ ∗ kv − v ∗ kH 1 . v ku − u∗ kHu1 ≡ In each of the bifurcation diagrams multiple branches are seen. They correspond to multiple, steady-state solutions. In Fig. 5.12, Fig. 5.13, and Fig. 5.15 the branches are easily distinguishable, while in Fig. 5.14 the branches are long and closer together. We also note that the length between the ▽’s and the ’s is largest for a = 0.2 and b = 0.5. In the bifurcation diagrams we can read off the γc -values. They all agree with the linear analysis. 5.1 Results in Matlab 81 10 k 5 H1u−norm 0 3 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160 20 40 60 80 γ 100 120 140 160 2 1 0 1 H1v −norm 20 0.5 0 Figure 5.12: The bifurcation of steady-states for a = 0.05, b = 1.5 and d = 20. It is shown that multiple, stable, steady-states occur near the areas where the blue ∗’s changes from one wavenumber to another. 82 Numerical analysis 10 k 5 20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160 20 40 60 80 γ 100 120 140 160 1 1 Hu−norm 0 2 0.2 v H1−norm 0 0.4 0 Figure 5.13: The bifurcation of steady-states for a = 0.1, b = 1.5 and d = 20. It is shown that multiple, stable, steady-states occur near the areas where the blue ∗’s changes from one wavenumber to another. There exist γ-values greather than γc for which the relative H 1 -norms vanish. 5.1 Results in Matlab 83 10 k 5 H1u−norm 0 6 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160 20 40 60 80 γ 100 120 140 160 4 2 0 1 H1v −norm 20 0.5 0 Figure 5.14: The bifurcation of steady-states for a = 0.2, b = 0.5 and d = 20. It is shown that multiple, stable, steady-states occur near the areas where the blue ∗’s changes from one wavenumber to another. The Hu1 and the length between the ▽’s and the ’s are seen to be larger than for the other parameters considered for d = 20, see Fig. 5.12, Fig. 5.13 and Fig. 5.15. 84 Numerical analysis k 10 5 0 1 Hu−norm 3 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160 20 40 60 80 γ 100 120 140 160 2 1 0 1 0.5 v H1−norm 20 0 Figure 5.15: The bifurcation of steady-states for a = 0.2, b = 1 and d = 20. It is shown that multiple, stable, steady-states occur near the areas where the blue ∗’s changes from one wavenumber to another. 5.1 Results in Matlab 85 In Fig. 5.13 it is observed, that there exists γ-values greather than γc such that no wavenumber are linearly unstable. In Fig. 5.16 the graph of λ+ is seen as a function of γ for the first two wavenumbers, k0,1 = 2 and k1,0 = π. For γ ≈ 15 to 20 no wavenumbers are linearly unstable. Therefore, in the bifurcation diagram Fig. 5.13, the relative H 1 -norms for these γ-values are zero and (u∗ , v ∗ ) is stable. 2 0 Re λ+ −2 −4 k0,1 = 2 −6 k 1,0 =π −8 −10 0 5 10 15 γ 20 25 30 Figure 5.16: The graph of Re λ+ as a function of γ for (a, b) = (0.1, 1.5) and for the first two wavenumbers, k0,1 = 2 and k1,0 = π. For γ ≈ 15 to 20 no wavenumbers are linearly unstable. 86 Numerical analysis In Fig. 5.17 contour-plots of steady-state u-patterns for the four parameter values are seen for γ = 72. The patterns for (a, b) = (0.05, 1.5) and (0.1, 1.5) are comparable to each other and look like the eigenfunction cos(πx) cos(4y). Equivalently, the pattern for (a, b) = (0.2, 1) looks like the eigenfunction cos(4y). The steady-state solution for (a, b) = (0.2, 0.5), where the Hu1 -norm of u − u∗ is largest, does not compare to any eigenfunction and compared to the other solutions it is more spatial aggregated. (u−u*)(t=∞) (u−u*)(t=∞) 0.8 0.4 1.5 1.5 0.6 0.3 0.4 0.2 0.1 0.2 1 1 0 y y 0 −0.1 −0.2 −0.2 −0.4 0.5 0.5 −0.3 −0.6 −0.4 −0.8 0 0 0.2 0.4 0.6 0.8 −0.5 0 0 1 0.2 0.4 x 0.6 0.8 1 x (a) a = 0.05, b = 1.5, d = 20, γ = 72 (b) a = 0.1, b = 1.5, d = 20, γ = 72 (u−u*)(t=∞) (u−u*)(t=∞) 1.4 0.4 1.5 1.5 1.2 0.3 1 0.2 0.8 1 0.1 1 0 y y 0.6 0.4 0.2 0.5 −0.1 0.5 −0.2 0 −0.3 −0.2 −0.4 0 0 0.2 0.4 0.6 0.8 1 x (c) a = 0.2, b = 0.5, d = 20, γ = 72 0 0 0.2 0.4 0.6 0.8 1 x (d) a = 0.2, b = 1, d = 20, γ = 72 Figure 5.17: u-pattern at four different points in the Turing domain. There is larger level of spatial aggregation for a = 0.2, b = 0.5. 5.1 Results in Matlab 87 Fig. 5.18 shows patterns for (a, b, d) = (0.2, 0.5, 20) and γ = 120. Again, the patterns are spatial aggregated and involve large gradients. The patterns resemble the patterns in Fig. 5.8 for d = 80 and a = 0.05 and b = 1. (u−u*)(t=∞) (u−u*)(t=∞) 1.2 1.4 1.5 1.5 1.2 1 1 0.8 0.8 1 1 0.6 y y 0.6 0.4 0.4 0.2 0.2 0.5 0.5 0 0 −0.2 −0.2 0 0 0.2 0.4 0.6 0.8 0 0 1 0.2 0.4 x 0.6 0.8 1 x (u−u*)(t=∞) 1.4 1.5 1.2 1 0.8 1 y 0.6 0.4 0.2 0.5 0 −0.2 0 0 0.2 0.4 0.6 0.8 1 x Figure 5.18: u-pattern for (a, b, d) = (0.2, 0.5, 20) and γ = 120. The patterns are spatial aggregated and involve large gradients. 88 Numerical analysis We observed above that the length between the ▽’s and the ’s in the topmost plot of the bifurcation diagrams was larger for (a, b, d) = (0.2, 0.5, 20) and (a, b, d) = (0.05, 1, 80) compared to the other parameters studied. This was accompanied by a larger Hu1 -norm and a larger level of spatial aggregation in the stationary patterns. The length between the ▽’s and the ’s refer to the length in γ for which the corresponding wavenumber is unstable. This length is given by ∆γk2 : R3 ⊃ T → R+ (2) (1) ∆γk2 (a, b, d) = γk2 (a, b, d) − γk2 (a, b, d) > 0, (2) (1) where γk2 (a, b, d) > γk2 (a, b, d), for fixed k 2 , are the zeros of: λ+ (γ) = 1/2 1 −B(γ) + B(γ)2 − 4h(γ) , 2 and B(γ) and h(γ) are given by the right hand side of (4.17) resp. (4.15). We obtain: (2) (1) γk2 (a, b, d), γk2 (a, b, d) = 1/2 k2 (d∂r f ∗ + ∂s g ∗ ) ± (d∂r f ∗ + ∂s g ∗ )2 − 4d det A 2 det A k2 b−a 2 = − (a + b) d 2(a + b)2 b+a ( )1/2 2 b−a , ± d − (a + b)2 − 4d(a + b)2 (5.2) b+a and then ∆γk2 (a, b, d) = ϕ(a, b, d)k 2 , where 1/2 1 (d∂r f ∗ + ∂s g ∗ )2 − 4d det A det A ( )1/2 2 b−a 1 d − (a + b)2 − 4d(a + b)2 , = (a + b)2 b+a ϕ(a, b, d) = (5.3) In Fig. 5.19 ϕ is shown as a contourplot for (a, b, d) ∈ T . It is seen, that for (a, b, d) = (0.2, 0.5, 20) and (a, b, d) = (0.05, 1, 80), the ϕ-value is large compared to the other parameter values considered. This is in accordance to the bifurcation diagrams and the analysis above. 5.1 Results in Matlab 89 d=10 d=20 0.25 0.3 6 15 4 0.25 0.2 10 2 0.2 0 0.15 5 0.1 a a −2 0.15 0 −4 0.1 −6 0.05 −5 0.05 −8 0 0 0.5 1 1.5 0 0 −10 0.5 1 b 1.5 2 −10 b d=80 0.6 70 60 0.5 50 0.4 a 40 0.3 30 20 0.2 10 0.1 0 0 0 1 2 b 3 4 −10 Figure 5.19: ϕ(a, b, d) for d = 10, 20 and 80. Outside the Turing domain, ϕ has been set to −10. Please note the different colourbars in the three figures. The parameter values used are indicated by ∗’s. In Fig. 5.20 a contourplot of: 1/2 γ (d∂r f ∗ + ∂s g ∗ )2 − 4d det A d )1/2 ( 2 b−a γ 2 2 d − 4d(a + b) − (a + b) = d b+a M 2 − m2 = for γ = 1 is seen for d = 10, d = 20 and d = 80. M 2 − m2 is the distance in k 2 for which λ+ > 0, see Theorem 5. It is seen that for (a, b, d) = (0.2, 0.5, 20), (0.2, 1, 20) and (0.1, 1.5, 20), M 2 − 2 m ≈ 0.2, while (a, b, d) = (0.05, 1, 10), (0.05, 1.5, 20), M 2 − m2 ≈ 0.4, and for (a, b, d) = (0.05, 1, 80), M 2 − m2 ≈ 0.9. Therefore, this measure cannot be used to explain the relative higher level of spatial aggregation of the patterns obtained for (a, b, d) = (0.05, 1, 80) and (a, b, d) = (0.2, 0.5, 20). Changing γ does not change the conclusion as M 2 − m2 ∝ γ. 90 Numerical analysis d=10 d=20 0.25 0.6 0.3 0.5 0.25 0.4 0.2 0.8 0.7 0.2 0.6 0.5 0.15 a a 0.3 0.15 0.4 0.1 0.3 0.2 0.1 0.1 0.05 0.2 0.05 0.1 0 0 0.5 1 1.5 0 0 0 0.5 1 b 1.5 2 0 b d=80 0.6 0.9 0.8 0.5 0.7 0.4 0.6 a 0.5 0.3 0.4 0.2 0.3 0.2 0.1 0.1 0 0 1 2 b 3 4 Figure 5.20: M 2 − m2 for γ = 1 and d = 10, 20 and 80. Outside the Turing domain, M 2 − m2 has been set to 0. Please note the different colourbars in the three figures. The parameter values used are indicated by ∗’s. 5.1 Results in Matlab 91 By Fig. 5.19 we choose parameters for which ϕ is relatively large and small. For example for (a, b, d) = (0.01, 1, 20), ϕ ≈ 16 and for (a, b, d) = (0.01, 3.5, 80), ϕ ≈ 2. In Fig. 5.21 patterns for these two set of parameters and γ = 100 are seen. Again we see that the relative large ϕ-value is accompanied by a relative higher level of spatial aggregation and a relative larger Hu1 -norm; 5.96 in (a) compared to 1.82 in (d). The Hv1 -norms in (c) and (d) are 1.04 resp. 0.99. (u−u*)(t=∞) (v−v*)(t=∞) 2 0 1.5 1.5 −0.1 1.5 1 0.5 0.5 −0.2 1 −0.3 y y 1 −0.4 0.5 0 −0.5 −0.5 0 0 0.2 0.4 0.6 0.8 0 0 1 0.2 0.4 x 0.6 0.8 1 −0.6 x (a) a = 0.01, b = 1, d = 20, γ = 100 (b) a = 0.01, b = 1, d = 20, γ = 100 * (v−v*)(t=∞) (u−u )(t=∞) 1 0.06 1.5 1.5 0.04 0.5 1 0.02 1 0 y y 0 −0.5 0.5 −0.02 0.5 −1 0 0 0.2 0.4 0.6 0.8 1 x (c) a = 0.01, b = 3.5, d = 80, γ = 100 −0.04 0 0 0.2 0.4 0.6 0.8 1 −0.06 x (d) a = 0.01, b = 3.5, d = 80, γ = 100 Figure 5.21: stationary solutions for (a, b, d) = (0.01, 1, 20) where ϕ(a, b, d) ≈ 16 and (0.01, 3.5, 80) where ϕ ≈ 2 for γ = 100. The relative H 1 -norm of the u- and v-pattern in (a) and (b) is 5.96 resp. 1.04, while for (c) and (d) Hu1 = 1.82 resp. Hv1 = 0.99. 92 Numerical analysis 5.1.3 Equivalent patterns The final aim of the numerical analysis is to look into the possibility of obtaining patterns for different parameters that only differ by a scaling. ˜ ∈ T , and (u, v) and (ũ, ṽ) solve: Definition 4 Let (a, b, d), (ã, b̃, d) −∆u = a − u + u2 v, −d∆v = b − u2 v, resp. −∆u = ã − u + u2 v, ˜ = b̃ − u2 v, −d∆v on domain Ω and with Neumann-conditions. (u, v) and (ũ, ṽ) are equivalent if there exists real su and sv such that u − u∗ = su (ũ − ũ∗ ), v − v ∗ = sv (ṽ − ṽ ∗ ), b̃ b where (ũ∗ , ṽ ∗ ) = (ã + b̃, ã+ ) and (u∗ , v ∗ ) = (a + b, a+b ). b̃ 2 It was observed that for fixed d, the patterns corresponding to (a, b)-values along the level curves of ϕ, involved similar spatial aggregation. But, no equivalent patterns were observed. This is due to the fact that the wavenumbers are unstable for different γvalues on the level curves of ϕ. We can ovoid this by instead fixing the position (1) (2) of the zeros γk2 and γk2 , i.e. by considering the level curves of the vector func(1) (2) tion (γk2 , γk2 ). ˜ ∈ T . Then the curve l given by Again let (ã, b̃, d) s s ˜ ˜ d+t d+t ˜ l(t) = ã, b̃, d + t , t ∈ I ⊂ R, (5.4) d˜ d˜ where I is such that l(t) ∈ T ⊂ R3 for every t ∈ I, defines a continuous level (1) (2) ˜ This can be verified by insertion.23 curve of (γk2 , γk2 ) through (ã, b̃, d). By comparing patterns for the different (a, b, d)-values on the curve l, we observed the following result, which we have formulated in a theorem. 2 The τ )2 /b̃2 ) 3l ˜ + s)2 /ã2 ) or l(τ ) = (ãτ /b̃, , b̃(ã + s)/ã, d(ã parametrisations l(s) = ( ã+s ã can also be used. ˜ see (4.20) is also a continuous level curve of (m2 , M 2 ) through (ã, b̃, d), b̃+τ b̃ ˜ b̃ + , d( 5.1 Results in Matlab 93 Theorem 6 For every steady-state solution (ũ, ṽ) of (1.4)-(1.6) for parame˜ there exists a steady-state solution (u, v) of (1.4)-(1.6) for every ters (ã, b̃, d), parameter (a, b, d) ∈ {l(t)}t∈I ⊂ T , such that (u, v) and (ũ, ṽ) are equivalent. 2 Let us first give numerical examples before √ proving the √ √ theorem. √ The (a, b, d)-values ( 2×0.2, 2×0.5, 40), ( 3×0.2, 3×0.5, 60), (0.4, 1, 80) are all on the level curve {l}t∈I through (0.2, 0.5, 20). In Fig. 5.22, Fig. 5.23 and Fig. 5.24, Fig. 5.25 similar u- and v-patterns are seen for the four (a, b, d)-values for γ = 50 resp. γ = 100. The similar patterns have the same relative H 1 -norms. For the patterns in Fig. 5.22 the relative H 1 -norms are 2.4133 and 0.3911 for u resp. v, while for the patterns in Fig. 5.24, they are 2.4536 and 0.6278 for u resp. v. Now, the main observation was that each pattern in Fig. 5.22, Fig. 5.23, Fig. 5.24 or Fig. 5.25 could be scaled to cover any of the three remaining patterns, with an accuracy of machine precision ≈ 10−16 . This is what has been claimed to hold in general in Theorem 6. 94 Numerical analysis (u−u*)(t=∞) (u−u*)(t=∞) 1.2 1.5 1.5 1.5 1 1 0.8 1 1 0.4 0.5 y y 0.6 0.2 0.5 0.5 0 0 −0.2 0 0 0.2 0.4 0.6 0.8 −0.5 0 0 1 0.2 0.4 x 0.6 0.8 1 x (a) a = 0.2, b = 0.5, d = 20, γ = 72 (b) a = √ 2 × 0.2, b = * √ 2 × 0.5, d = 40, γ = 72 * (u−u )(t=∞) (u−u )(t=∞) 2 2.5 1.5 1.5 2 1.5 1.5 1 1 1 y y 1 0.5 0.5 0.5 0.5 0 0 −0.5 −0.5 0 0 0.2 0.4 0.6 0.8 1 x (c) a = √ 3 × 0.2, b = 0 0 0.2 0.4 0.6 0.8 1 x √ 3 × 0.5, d = 60, γ = 72 (d) a = 0.4, b = 1, d = 80, γ = 72 Figure 5.22: u-pattern for (a, b, d)-values on the same level curve {l}t∈I and γ = 50. The Hu1 -norm is equal to 2.4133 for all four (a, b, d)-values. 5.1 Results in Matlab 95 * * (v−v )(t=∞) (v−v )(t=∞) 1.5 1.5 −0.1 −0.1 −0.15 1 1 −0.15 y y −0.2 −0.25 −0.2 0.5 0.5 −0.3 −0.35 −0.25 −0.4 0 0 0.2 0.4 0.6 0.8 0 0 1 0.2 0.4 x 0.6 0.8 1 x (a) a = 0.2, b = 0.5, d = 20, γ = 72 (b) a = √ 2 × 0.2, b = (v−v*)(t=∞) √ 2 × 0.5, d = 40, γ = 72 (v−v*)(t=∞) −0.02 1.5 1.5 −0.04 −0.04 −0.06 −0.06 −0.08 1 −0.08 1 −0.1 −0.1 y y −0.12 −0.12 −0.14 −0.16 0.5 −0.14 0.5 −0.18 −0.16 −0.2 −0.18 −0.22 0 0 0.2 0.4 0.6 0.8 1 x (c) a = √ 3 × 0.2, b = 0 0 −0.2 0.2 0.4 0.6 0.8 1 x √ 3 × 0.5, d = 60, γ = 72 (d) a = 0.4, b = 1, d = 80, γ = 72 Figure 5.23: v-pattern corresponding to the patterns in Fig. 5.22. The Hv1 -norm is equal to 0.3911 for all four (a, b, d)-values. 96 Numerical analysis * * (u−u )(t=∞) (u−u )(t=∞) 1.6 1.5 1.5 1.4 0.8 1.2 1 0.6 1 1 0.8 0.6 y y 0.4 0.4 0.2 0.5 0.2 0.5 0 0 −0.2 −0.2 0 0 0.2 0.4 0.6 0.8 0 0 1 −0.4 0.2 0.4 x 0.6 0.8 1 x (a) a = 0.2, b = 0.5, d = 20, γ = 100 √ √ (b) a = 2 × 0.2, b = 2 × 0.5, d = 40, γ = 100 (u−u*)(t=∞) (u−u*)(t=∞) 2 2 1.5 1.5 1.5 1.5 1 1 1 y y 1 0.5 0.5 0.5 0.5 0 0 −0.5 0 0 0.2 0.4 0.6 0.8 1 x (c) a = √ 3×0.2, b = √ −0.5 0 0 0.2 0.4 0.6 0.8 1 x 3×0.5, d = 60, γ = 100 (d) a = 0.4, b = 1, d = 80, γ = 100 Figure 5.24: u-pattern for (a, b, d)-values on the same level curve {l}t∈I and γ = 100. The Hu1 -norm is equal to 2.4536 for all four (a, b, d)-values. 5.1 Results in Matlab 97 (v−v*)(t=∞) (v−v*)(t=∞) 0.05 0.05 1.5 1.5 0 0 −0.05 −0.05 −0.1 1 −0.15 1 −0.1 −0.15 y y −0.2 −0.25 −0.2 −0.3 0.5 0.5 −0.35 −0.25 −0.4 −0.3 −0.45 0 0 0.2 0.4 0.6 0.8 1 −0.5 0 0 −0.35 0.2 0.4 x 0.6 0.8 1 x (a) a = 0.2, b = 0.5, d = 20, γ = 100 (b) a = γ = 100 √ 2 × 0.2, b = * √ 2 × 0.5, d = 40, * (v−v )(t=∞) (v−v )(t=∞) 0.05 0 1.5 1.5 0 −0.05 −0.05 1 1 −0.1 y y −0.1 −0.15 0.5 −0.15 0.5 −0.2 −0.2 −0.25 0 0 0.2 0.4 0.6 0.8 1 x (c) a = √ 3×0.2, b = √ 0 0 0.2 0.4 0.6 0.8 1 −0.25 x 3×0.5, d = 60, γ = 100 (d) a = 0.4, b = 1, d = 80, γ = 100 Figure 5.25: v-pattern corresponding to the patterns in Fig. 5.24. The Hv1 -norm is equal to 0.6278 for all four (a, b, d)-values. 98 Numerical analysis And now for the proof of Theorem 6. ˜ Proof The equations for the perturbed variables, (ũ, ṽ), for parameters (ã, b̃, d), are: −∆ũ = b b̃ − ã ũ + (ã + b̃)2 ṽ + 2(ã + b̃)ṽ ũ + ũ2 + ũ2 ṽ, ã + b̃ (ã + b̃)2 ˜ =− −d∆ṽ 2b̃ ã + b̃ ũ − (ã + b̃)2 ṽ − 2(ã + b̃)ṽ ũ − b̃ (ã + b̃)2 ũ2 − ũ2 ṽ. Let q be the q (u, v) solution of the perturbed equations with parameters (a, b, d) = d d ã, d˜b̃, d , i.e. the solution of: d̃ q −∆u = q or d b̃ d̃ − q d ã d˜ r d ã + d˜ r d b̃ d˜ !2 r d ã + d˜ r ! d b̃ vu d˜ q u+ v+2 + dd̃ b̃ q d b̃ d˜ 2 2 + q q 2 u + u v, d d ã + d˜b̃ d˜ q r !2 r ! r r 2 dd̃ b̃ d d d d q u− ã + b̃ v − 2 ã + b̃ vu −d∆v = − q ˜ ˜ ˜ d d d d d d˜ ã + b̃ ˜ ˜ d d q d b̃ d˜ 2 2 − q q 2 u − u v, d d ã + d˜b̃ d˜ d ã d̃ r 2 b̃ − ã d d −∆u = u+ ã + b̃ v + 2 ã + b̃ vu ã + b̃ d˜ d˜ s d˜ b̃ 2 2 + 2 u + u v, d ã + b̃ r 2 2 b̃ d d d ã + b̃ v − 2 ã + b̃ vu −d˜ ∆v = − u− d˜ d˜ d˜ ã + b̃ s b̃ d˜ 2 2 − 2 u + u v. d ã + b̃ (5.5) (5.6) 5.1 Results in Matlab 99 If we multiply (5.5) and (5.6) by q ˜ we obtain: d/d s ! ! s rd 2 r d ˜ ˜ d d d˜ b̃ − ã v + 2 ã + b̃ v u = u + ã + b̃ u −∆ d d d d˜ d˜ ã + b̃ s 2 s 2 r ! d b̃ d˜ d˜ + v , u u + 2 d d d˜ ã + b̃ s r ! s r ! r ! ˜ 2 d d d 2b̃ d d˜ ˜ v =− v − 2 ã + b̃ v u − ã + b̃ u −d∆ d d d˜ d˜ d˜ ã + b̃ s − b̃ ã + b̃ s 2 2 s 2 d˜ d˜ u u + d d q From this, we realise that (u, v) = q ˜ sv = dd , and then the theorem follows. d ũ, d̃ q d̃ d ṽ r ! d v . d˜ q d , and hence su = and d˜ In the examples above we saw that the relative H 1 -norms were constant along {l}t∈I . This is of course no coincidence. The relative H 1 -norms of ũ and ṽ are given by kũ − ũ∗ kHu1 1/2 Z X 1 |∂ α (ũ − ũ∗ )|2 , = ∗ ũ Ω kṽ − ṽ ∗ kHv1 = |α|≤1 1 ṽ ∗ Z X Ω |α|≤1 1/2 |∂ α (ṽ − ṽ ∗ )|2 . Now, if we move along the level curve {l}t∈I , then for every (a, b, d) ∈ {l}t∈I : ku − u∗ kHu1 1/2 Z X 1 |∂ α (u − u∗ )|2 , = ∗ u Ω |α|≤1 1/2 r Z X 1 d |∂ α (ũ − ũ∗ )|2 , =q ˜ d ∗ d Ω ũ |α|≤1 d˜ = kũ − ũ∗ kHu1 , 100 Numerical analysis since u∗ = ∗ since v = 5.2 q q d ∗ ũ , d˜ and similarly for v, kv − v ∗ kHv1 = kṽ − ṽ ∗ kHv1 , d̃ ∗ d ṽ . Summary and remarks Six different (a, b, d)-values were chosen, and for each of these the existence of multiple steady-state solutions were shown. What pattern, that ultimately developed, depended upon the initial conditions. One might have thought that the number of different steady-state patterns, at least modulus rotation and multiplication by −1 as illustrated in Fig. 5.5, for fixed (a, b, d) and γ were equal to the number of unstable wavenumbers. However, we were not able to support this claim by the numerical analysis. An example with six unstable wavenumbers was given, however, we only found four different patterns. It was observed numerically, that u had a local maximum where v had a local minimum, and u had a local minimum where v had a local maximum. Using the H 1 -norm the bifurcation of steady-states were analysed as we varied γ. The results were presented in bifurcation diagrams. γc from Theorem 5 could be read off from the bifurcation diagrams for the (a, b, d)-values considered. They all agreed with Theorem 5. Furthermore, the bifurcation diagrams showed that for some (a, b, d)-values the patterns had relative higher Hu1 -norm. The Hu1 -norm was seen to be a measure of the heterogeneity or the spatial aggregation of the steady-states. It was argued, by way of examples, that the heterogeneity was related to the size of ϕ, defined in (5.3). It was observed numerically that all the systems, (1.4)-(1.6), with (a, b, d)(2) (1) values along the level curves in T of (γk2 , γk2 ) exhibit equivalent steady-state solutions for every γ. This result was stated in a theorem, Theorem 6, and a proof was established. Moreover, it was observed numerically, and proved analytically, that the relative H 1 -norms were conserved along these level curves. This also shows, that the relative H 1 -norms are the right concepts to work with in the bifurcation analysis. Some patterns appeared to be similar to certain eigenfunctions. For example, Fig. 5.26 shows a surface-plot of the pattern in Fig. 5.2. It looks like a cosine wave in the x-direction. k2,0 = 2π is in the range of unstable wavenumbers, see Fig. 5.6. Therefore, 5.2 Summary and remarks 101 in the hope of finding an approximate solution we defined I : R2 → R by 2 I(U, V ) = kkn,m U cos(kn,m x) − f (U cos(kn,m x) + u∗ , V cos(kn,m x) + v ∗ )k2 2 +kdkn,m V cos(kn,m x) − g(U cos(kn,m x) + u∗ , V cos(kn,m x) + v ∗ )k2 , and searched for (U, V ) 6= (0, 0) satisfying ∂U I(U, V ) = 0 = ∂V I(U, V ). It is combersome to do any general, analytical work on these coupled equations, as they involve many, complicated terms. Instead we used MAPLE for the specific values a = 0.05, b = 1, d = 10 and γ = 120. No real, non-trivial (U, V ) existed. Other parameters were also tried, but the conclusion was the same. Therefore, we do not expect this method to be sufficient to predict any approximations for any parameters in the Turing domain. To obtain approximations a different approach is therefore necessary. (v−v*)(t=∞) 1 0.4 0.5 0.2 v−v* u−u* (u−u*)(t=∞) 0 −0.5 0 −0.2 −1 2 −0.4 2 1.5 1 1 0.5 0.5 y 0 0 x 1.5 1 1 0.5 0.5 y 0 0 x Figure 5.26: Surf-plot of u and v. Corresponds to the contour-plots in Fig. 5.2. 102 Numerical analysis Chapter 6 Conclusion Two different approaches were applied to the reaction-diffusion systems to prove existence and uniqueness. The first approach was the approach described by Pao, [Pao (1992)], with upper and lower solutions. Hereby, local existence was obtained for the Schnakenberg system while global existence and boundedness was achieved for the Gierer and Meinhardt system. This technique was shown to be inadequate for the Thomas system. The approach was also used to study the stationary equations for the Schnaken∗ berg system. It was shown that the uniform steady-state solution, (u∗ , v√ ), was a2 +a2 1−4 b a the unique solution in the sector of h(a, 0), (b/a2 , ui, where u = , 2b for b ≤ a/4. This is an interesting result, as it is often difficult to say anything about the number of stationary solutions. Often, there exist one single, or a “few,” stable steady-states and “many” unstable steady-states. The approach described by Pao was shown to provide an easy, ad hoc approach to show existence and uniqueness of coupled, parabolic and elliptic partial differential equations. It also provided an approach to analyse the stationary equations. A major downside, however, was the fact that, the technique only applied to quasi-monotone reaction functions. Next the functional analysis and semigroup approach was considered. The conditions: 104 Conclusion (H1) d > 0; (H2) u0 ≥ 0 and v0 ≥ 0 are continuous on Ω, that is u0 , v0 ∈ CL0 ∞ (Ω); 3 (H3) f and g are continuously differentiable functions from R+ into R with f (t, 0, s) ≥ 0 and g(t, r, 0) ≥ 0 for all t, r, s ≥ 0; 2 (H4) There exist m > 0 and a continuous function F : R+ → R+ such that f (t, r, s), g(t, r, s) ≤ exp(mt)F (r, s) for all t, r, s ≥ 0. were shown to imply positivity and local existence, i.e. existence for t < T ∗ for some 0 < T ∗ < ∞, for the general two-component reaction-diffusion equations with Neumann-conditions. Furthermore, it was shown that if the solution only existed locally, then it would diverge. By contradiction, this was used to show global existence for the Schnakenberg system, in one space dimension, and for the Thomas system. If we furthermore imposed the conditions (H5) The v-variable is bounded for all t ≤ T ∗ if the solution is local; (H6) There is an η ≥ 1 and a continuous function h : [0, ∞)2 → [0, ∞) such that |f (t, r, s)| ≤ h(t, S)(1 + r)η for all t, r, s ≥ 0 with s ≤ S; (H7) There is an ǫ > 0 and a continuous function l : [0, ∞)2 → [0, ∞) such that ǫr + f (t, r, s) + g(t, r, s) ≤ l(t, S) for all t, r, s ≥ 0 with s ≤ S. then, the solution was shown to be global. If N2 , h and l were bounded in t then the solution was additionally proved to be bounded for all t. Next, it was shown that when (H1)-(H7) were satisfied, then the solution depended continuously upon the initial data. To sum it all up, it was shown that the conditions (H1)-(H7) were sufficient conditions for the two-component reaction-diffusion system with Neumann conditions to be well-posed, in the sense of Hadamard, and to be positive and bounded. This in turn showed that the three specific systems considered satisfied (W1)-(W7). Compared to the approach with upper and lower solution, the use of semigroups and general functional analysis established a setting in which the entire analysis can be made precise and complete. However, this setting involved more work and was not in general ad hoc. Next, linear stability and instability was proved to be sufficient conditions for stability resp. instability of steady states of the general autonomous, twocomponent reaction-diffusion equations. This was used to study diffusion-driven instabilities at uniform steady state. The conditions 105 (T1) ∂r f (u∗ , v ∗ ) + ∂s g(u∗ , v ∗ ) < 0; (T2) ∂r f (u∗ , v ∗ )∂s g(u∗ , v ∗ ) − ∂s f (u∗ , v ∗ )∂r g(u∗ , v ∗ ) > 0; (T3) d∂r f (u∗ , v ∗ ) + ∂s g(u∗ , v ∗ ) > 0; (T4) (d∂r f (u∗ ,v ∗ )+∂s g(u∗ ,v ∗ ))2 4d > ∂r f (u∗ , v ∗ )∂s g(u∗ , v ∗ ) − ∂s f (u∗ , v ∗ )∂r g(u∗ , v ∗ ); were shown, for sufficient large domains, to be sufficient for the system to exhibit diffusion-driven instabilities at uniform steady-states (u∗ , v ∗ ) ∈ R2+ . Finally, the Schnakenberg model was studied numerically on a rectangle. Chebyshev polynomials were used for the Laplacian and the time integration was performed by built in MATLAB ode-solver ode15s. In accordance to the linear analysis diffusion-driven instabilities were seen to occur for sufficiently large domains. These instabilities were seen to evolve into steady-state heterogeneous patterns. The existence of multiple steady-states were shown by the use of bifurcation diagrams. The H 1 -norm was used as bifurcation functional. The patterns’ dependency upon the parameters were analysed and it was seen that in some part of the parameter domain, determined by (T1)-(T4), the spatial aggregation of the final patterns was relatively larger. By way of example, this was shown to be related to the size of s 2 1 b−a 2 − 4d(a + b)2 . ϕ(a, b, d) = − (a + b) d (a + b)2 b+a The most important observation made in the numerical analysis of the Schnakenberg model was the existence of curves in the parameter domain on which the corresponding equations exhibited equivalent steady-state solutions. This result was stated in a theorem and a proof was established. It was furthermore observed and proved that the relative H 1 -norms were conserved along these lines. This result would not have been obtained without the numerical simulations of the models. This illustrates why a numerical analysis is indispensable in almost every analysis of non-linear problems as numerical results often give us ideas about what to study analytically. 106 Conclusion Appendix A Semigroups and sectorial operators The main results on existence and uniqueness of semilinear parabolic equations are based upon the fixed point theorems from functional analysis. We are allowed to make use of these results through the setting of semigroups. In the following we introduce the concept of semigroups and state the most relevant results. It is well-known that the ordinary differential equation: ∂t u = λu, u(0) = u0 , has the solution: u(t) = exp(λt)u0 , t ≥ 0. (A.1) In the theory of semigroups we shall define similar solutions for partial differential equations like ∂t u − Au = 0, u(0) = u0 , (A.2) and some initial condition u(0) = u0 , when the linear differential operator A satisfy some suitable conditions. We shall see that the adequate conditions, for our purpose at least, are that A is sectorial, see Definition 6. The main idea is to rely on the abstract functional analysis and consider u as 108 Semigroups and sectorial operators function of t with values in Banach Space, i.e. one considers u : t 7→ X where X is some Banach space rather than u : (x, t) 7→ R or C. In this setting existence and uniqueness can often be obtained by applying fixed point theorems, for example the fundamental Banach’s Fixed Point Theorem. We are now ready for a definition of a semigroup. Definition 5 (Semigroup, see [Henry (1981), Definition 1.3.3]) An analytic semigroup is a family of bounded, linear operators on X, {G(t)}t≥0 , satisfying: 1◦ G(0) = I, G(t)G(s) = G(t + s) for t, s ≥ 0 2◦ G(t)x → x as t → 0+ , for each x ∈ 3◦ t → G(t)x is real analytic on 0 < t < ∞ for each x ∈ X The generator of this semigroup is defined by Bx = lim+ t→0 G(t)x − x , t (A.3) with domain D(B) = G(t)x − x exist in X . x ∈ X| lim t t→0+ (A.4) If a family of bounded, linear operators on X only satisfy 1◦ and 2◦ we call the family a C0 -semigroup. If a C0 -semigroup is compact for every t we call the family a compact-semigroup. 2 As already mentioned the suitable condition for our operator A in this context is that it is sectorial. Definition 6 (Sectorial operator, see [Henry (1981), Definition 1.4.1]) Let A be a linear operator in a Banach Space X and suppose A is closed and densely defined. If there exist real a, ω ∈ (0, π), M ≥ 1 such that ρ(A) ⊃ Σ = {λ ∈ C| φ ≤ arg (λ − a) ≤ π, λ 6= a} (A.5) and kRλ (A)k ≤ then we say that A is sectorial. M , |λ − a| for all λ ∈ Σ (A.6) 2 109 The adequateness of sectorial operators is implied by the following theorem. We refer to App. B for a short introduction on integration and differentiation in Banach spaces. Theorem 7 (See [Henry (1981), Theorem 1.3.4]) If A is sectorial then -A is the infinitesimal generator of an analytic semigroup, G(t). If Re λ > a, a ∈ R, whenever λ ∈ σ(A) then for any t > 0, kG(t)k ≤ C exp(−at), kAG(t)k ≤ C exp(−at). t Finally, d G(t) = −AG(t), dt t > 0. (A.7) 2 Proof See e.g. [Henry (1981), Theorem 1.3.4]. When A in (A.2) is sectorial then in the abstract setting of semigroups the solution is, according to 2. in Definition 5 and (A.7): u(t) = G(t)u0 , t ≥ 0, similar to (A.1). In our study of reaction-diffusion equations the Laplacian is of particular interest. The following example considers the Neumann realisation of −∆. Example 1 (Neumann realisation of Laplacian) L = −∆ with domain D(L) = u ∈ H 2 (Ω) | νu = 0 is a sectorial operator in L2 (Ω). Since C0∞ (Ω) ⊂ D(L) it is densely defined in L2 (Ω). Assertion 2 A is closed. Proof Let {un }n∈N ⊂ D(A) such that un → u in L2 (Ω) and Aun → v in L2 (Ω). It must be shown that u ∈ D(L) and Lu = v. Now since un → u in L2 (Ω), (Lun , φ) → (Lu, φ), for all φ ∈ D(Ω) so that v = Lu ∈ L2 (Ω) and un → u in D(L). The trace operator ν is continuous from H 2 (Ω) into H 1/2 (Ω), see [Lions and Magenes (1968)]. Hence kνu − νun k1/2 → 0. 110 Semigroups and sectorial operators But since νun = 0: kνuk1/2 = 0 ⇔ νu = 0. It is concluded that u ∈ D(L) with Lu ∈ L2 (Ω) and L is therefore closed. Assertion 3 L has a countable set of real, non-negative eigenvalues {λk }k∈N that can be arranged in a non-decreasing sequence 0 ≤ λ1 ≤ · · · ≤ λk ≤ · · · Proof L = L∗ which follows by integration by parts and extension by continuity: (Lu, v) = (u, Lv), for all u, v ∈ D(L), (A.8) and therefore all eigenvalues are real. Again by integration by parts and extension by continuity: (Lu, u) = (∇u, ∇u) = k∇uk2 ≥ 0, for all u ∈ D(L). from which the non-negativeness of the eigenvalues follows. It is now claimed that K = (I − ∆)−1 exists and is compact from L2 (Ω) into itself. To prove this consider the variational approach, where L is defined from the sesquilinear form s(u, v) = (u, v)H 1 = (∇u, ∇v) + (u, v). By Lax-Milgram I − ∆ is bijection of D(L) onto L2 (Ω). Furthermore for f ∈ L2 (Ω) kuk2H 1 = s(u, u) = (f, u) ≤ kf k2 kuk2 ≤ kf k2 kukH 1 from which it follows kKf kH 1 ≤ kf k2 and by the Rellich-Kondrachov compactness theorem K is a compact operator. Therefore the eigenvalues of K are countable and the spectrum is a pure point spectrum and the same therefore holds for L = −∆. It has been shown: Corollary 1 For any a ≤ 0 and any ω ∈ (0, π), Σ in Definition 6, is in the resolvent set. To prove that L is sectorial the existence of M in accordance to Definition 6 has to be proved. 111 Assertion 4 There exists M such that kRλ (L)k ≤ M , |λ| for all λ ∈ Σ. Proof The result follows from the following computation k(L − λI)uk2 kuk2 ≥ (Cauchy-Schwarz inequality) ≥ |((L − λI)u, u)| = (((L − Re λ)u, u)2 + Im λ2 kuk22 )1/2 |Im λ|kuk22 , for Re λ ≥ 0 ≥ 2 ((Re λ) + Im λ2 )1/2 kuk22 , for Re λ ≤ 0 |Im λ|kuk22 , for Re λ ≥ 0 = |λ|kuk22 , for Re λ ≤ 0 ≥ (writing λ = |λ| exp(iθ), θ ∈ (ω, π/2) when Re λ ≥ 0) ≥ |λ| sin(ω)kuk22 , any ω ∈ (0, π), and hence kRλ (L)k ≤ sin(ω)−1 . |λ| The fractional powers are the key issue in the existence in Sec. 3. First a definition. Definition 7 (Fractional powers, see [Pazy (1983), Section 2.6]) Suppose A is a sectorial operator with analytic semigroup G and Re λ > 0 whenever λ ∈ σ(A), then we define the fractional powers of A by Z ∞ 1 Aα = tα−1 G(t)dt, α > 0, Γ(α) 0 where sin(απ) Γ(α) = π Z ∞ ξ −α exp(−ξ)dξ, 0 and A0 = I. 2 Proposition 4 If A is sectorial in Banach space X with Re λ > 0 whenever λ ∈ σ(A), then for any α ≥ 0, A−α is a bounded, linear operator on X which is injective and satisfies A−α A−β = A−(α+β) . 2 112 Semigroups and sectorial operators Proof α = 0 is trivial. Therefore let α > 0. By assumption 0 ∈ ρ(A), and hence that A−1 exists and is injective. For every n ∈ N A−n is therefore injective. Now for n ≥ α A−n x = A−n+α A−α x = 0 ⇔ x = 0. Therefore A−α is injective. Remark 10 The name fractional powers is misleading: We shall allow α ∈ R+ and not only in Q. 2 By Example 1 H = I − ∆ is sectorial and 0 ∈ ρ(H). Therefore −H generates an analytic semigroup and the fractional powers H−α exist and are injective. Furthermore D(H) = D(L). Proposition 5 α > β > 0 implies D(Hα ) ⊂ D(Hβ ) with compact injection. Furthermore D(Hα ) is dense in D(Hβ ). 2 Proof See [Henry (1981), Theorem 1.4.8]. Lemma 8 (See [Henry (1981), Theorem 1.6.1]) D(Hα ) ⊂ CL0 ∞ (Ω) with continuous injection for α > n/4, i.e. u ∈ D(Hα ) has a version ũ ∈ CL0 ∞ (Ω) such that sup |ũ| ≤ kukD(Hα ) . Ω . 2 Proof For s > n/2 the Sobolev Theorem for smooth, bounded subsets of Rn , see e.g. [Grubb (2007), Corollary 6.12], H s (Ω) ⊂ CL0 ∞ (Ω), with sup {|u| | x ∈ Ω} ≤ Cs kukH s , or short, H s (Ω) ֒→ CL0 ∞ (Ω). Hs/2 u ∈ L2 (Ω) if and only if u ∈ H s (Ω), see e.g. [Lions and Magenes (1968)]. Therefore D(Hα ) ֒→ H s (Ω) if α = 2s > n4 . Remark 11 In the proof of local existence it shall be used that for n = 1, 2, 3 there exist an α such that n/4 < α < 1. For n ≥ 4 and more importantly when we later turn to global existence the L2 -theory is inadequate and one will have to consider general Lp -spaces, 1 ≤ p < ∞. It is therefore natural to introduce some generalisations to Lp -spaces. We do so without proof. First of all the Neumann realisation L of the laplacian, that is the the operator acting like −∆ on smooth functions and with domain D(L) = u ∈ W 2,p | νu = 0 113 is still sectorial in Lp . This is obtained by similar arguments to those used in the L2 -result in Example 1, although one has to exploit Hahn-Banach’s Theorem rather than Riesz Representation Theorem. Lemma 8 also have an Lp -counterpart, which is shown to be at the very core of the proof of global existence. Lemma 9 Let p ∈ [1, ∞). Then D(Hα ) ⊂ CL0 ∞ (Ω) with continuous injection for α > n/(2p). 2 Proof (Sketched) Again the proof goes by utilising the Sobolev Theorem, now in Lp -setting, and the interpolation results in [Lions and Magenes (1968)]. Note that in Lp the graph norm is kukD(Hα ) = kHα ukp + kukp . 2 Reminded by the previous remark we obtain the following lemma in Lp . Lemma 10 Let G be an analytic semigroup generated by a sectorial operator −A, and let δ ≥ 0 be such that −A + δ is still the generator of an analytic semigroup. The following properties then hold for the semigroup G(t) and the fractional powers of A: 1◦ G(t) : Lp (Ω) → D(Aα ) for all t > 0; 2◦ kG(t)ukD(Aα ) ≤ Cα,p t−α exp(−δt)kukp for all t > 0, u ∈ Lp (Ω); 3◦ G(t)Aα u = Aα G(t)u for all t > 0, u ∈ D(Hα ). 2 Proof (Sketched) See e.g. [Pazy (1983), Theorem 6.13]. 1◦ is a direct consequence of G(t) being analytic. 2◦ is proved by exploiting the Closed Graph Theorem to conclude that Aα G(t) is bounded and then using Aα G(t) = Aα−n An G(t), with n ∈ N chosen such that n−1 < α ≤ n, to estimate kAα G(t)ukp , u ∈ Lp (Ω). Finally 3◦ can be verified using the definition of the fractional powers and of a semigroup, see Definition 7 resp. Definition 5. 114 Semigroups and sectorial operators Appendix B Differentiation and integration in Banach Spaces The theory of semigroups is based on the study of functions with values in a Banach space X. Therefore for the sake of clarity it is appropriate to discuss differentiation and integration in Banach spaces. In the following let u(t) and v(t) be X-valued functions for all t ∈ I, i.e. u : I → X and v : I → X, where I is open in R. The definition of differentiation in a normed vector space is given in the following definition. The reader is reminded that a function ǫ : N0 → X, where N0 is an open neighbourhood of 0, is called an ǫ-function if it is continuous in 0 with ǫ(0) = 0. Definition 8 u is (norm) differentiable at t ∈ I if there exist a linear mapping L : N → X, N an open neighbourhood of t, and an ǫ-function such that u(t + h) − u(t) = L(h) + ǫ(h)|h|. for every h ∈ {d ∈ R | t + d ∈ I}. (B.1) 2 Remark 12 Since L(h) is linear (B.1) may be rewritten as u(t + h) − u(t) |h| = L(1) + ǫ(h) . h h from which it is clear that L(1) can be interpreted as the differential of u at t. It is abbreviated ∂t u(t). 2 116 Differentiation and integration in Banach Spaces And now for the definition of integration in Banach spaces. First a theorem is needed. Theorem 8 Let u : I → X be continuous with values in Banach space X. Then there exists a unique vector I(u) that satisfies: For every ǫ > 0 there exist an δ > 0, so that kI(u) − SD (u)kX < ǫ, PN for every SD (u) = i=1 u(τi )(ti − ti−1 ), where τ1 , τ2 , · · · , τN ∈ [a, b] such that τi ∈ [ti−1 , ti ] with fineness µ(D) = max {|ti − ti−1 | | 1 ≤ i ≤ n} < δ. 2 Proof (Sketched) The proof can be seen in [Hansen (1995), Theorem III.3.2]. Readily it comes down to showing, that for a sequence of subdivisions {Dn }n∈N for which µ(Dn ) → 0, the corresponding sequence of sums, {SDn }n∈N , is a Cauchy sequence in the Banach space X. And then: Definition 9 I(u) from Theorem 8 is called the integral of u over [a, b] ∈ I. 2 The following (usual) rules of integration do apply, see [Hansen (1995), Theorem III.3.4], Z Z b (αv(t) + βw(t))dt = α Z b v(t)dt + β v(t)dt + Z c v(t)dt = b a k 1 h Z b a Z c v(t)dtkX ≤ c+h Z b w(t)dt, a a a b Z c v(t)dt, a Z b a kv(t)kX , v(t)dt → v(c) in X for h → 0. (B.2) Appendix C Gronwall’s inequality In this appendix the Gronwall inequality in differential and integral form are stated and proven. Lemma 11 (Differential form) Let f , h, y and ∂t y be locally integrable functions on [0, ∞) that satisfy ∂t y ≤ f y + h, for t > 0, (C.1) y(0) = y0 , then Z t y(t) ≤ exp(F (t)) y0 + h(τ ) exp(−F (τ ))dτ , 0 where F (t) = Rt 0 f (τ )dτ . 2 Proof We multiply (C.1) with the positive valued function exp(−F (τ )) to obtain exp(−F (τ ))∂τ y − f (τ ) exp(−F (τ ))y = ∂τ (exp(−F (τ ))y) ≤ h(τ ) exp(−F (τ )). 118 Gronwall’s inequality Integration with respect to τ from 0 to t yields Z t h(τ ) exp(−F (τ ))dτ, exp(−F (t))y(t) ≤ y(0) + 0 and thus Z t y(t) ≤ exp(F (t)) y0 + h(τ ) exp(−F (τ ))dτ . 0 Lemma 12 (Integral form) Let y be locally integrable functions on [0, ∞) satisfying Z t y(t) ≤ A y(τ )dτ + B, for t > 0, (C.2) 0 for some non-negative constants A and B. Then, y(t) ≤ B exp(At). Proof Set z(t) = to Lemma 11 then Rt 0 2 y(τ )dτ ; then ∂t z ≤ Az + B with z(0) = 0 and according z(t) ≤ B exp(At)(1 − exp(−At)). A Finally, by (C.2) we get: y(t) ≤ B exp(At). Appendix D Matlab-codes D.1 main.m % num solve reaction diff eqn close all clear all global global global global gamma ab diag10 u0u0 v0v0 u0v0 N2 Nx Ny d Dx Dy a b % u’_t = gamma*(a-u-u0+(u’+u0)^2 (v’+v0)) +u’_xx % v’_t = gamma*(b-(u+u0)^2 (v+v0)) + d v’_xx % zero flux a = .05; b=1;d=80; tspan = [0 100]; Ly = pi/2; Lx = 1; Nx = 16; Ny = floor(Ly/Lx*Nx); 120 Matlab-codes if mod(Ny,2) ~= 0; Ny = Ny+1; end N2 = (Nx+1)*(Ny+1); plotN = 100; L = ((d*(b-a)-(a+b)^3)-sqrt((d*(b-a)-(a+b)^3)^2-... 4*d*(a+b)^4))/(2*d*(a+b)); M = ((d*(b-a)-(a+b)^3)+sqrt((d*(b-a)-(a+b)^3)^2-... 4*d*(a+b)^4))/(2*d*(a+b)); % Set up operator operator % Set up necessary parameters para kvec = zeros(Nx+1,Ny+1); for n= 0:Nx; for m = 0:Ny; kvec(n+1,m+1) = sqrt((n*pi)^2/Lx^2+(m*pi)^2/Ly^2); end end kvec = kvec(:) kvec = sort(kvec(:)); [X,Y] = meshgrid(x,y); Gamma = 70:1:75; h1u = []; h1v = []; kMvec = []; for gamma = Gamma; gamma tic % Initial conditions initcond evalkM [m,temp] = min(abs(kvec-kM)) out = lambdaeval(kvec(temp),a,b,d,gamma) if real(out) < 0 kMvec = [kMvec;0]; else kMvec = [kMvec;kvec(temp)]; end % Solve system D.1 main.m % solver recmovie toc dxuend = dx*uend; dyuend = dy*uend; fu = (u0+uend).^2+dxuend.^2+dyuend.^2; Fu = 1/(Lx*Ly)*trapez(fu,x,y,Nx,Ny); dxvend = dx*vend; dyvend = dy*vend; fv = (v0+vend).^2+dxvend.^2+dyvend.^2; Fv = 1/(Lx*Ly)*trapez(fv,x,y,Nx,Ny); h1u = [h1u; Fu]; h1v = [h1v;Fv]; figure surf(X,Y,reshape(uend,Ny+1,Nx+1)) title([num2str(gamma)]) drawnow end figure axes(’position’,[.1 .1 .8 .6]) plot(Gamma,h1u,’o’,Gamma,h1v,’d’); xlabel(’gamma’) ylabel(’H^1 - norm’) hold on legend(’u_s’,’v_s’) axes(’position’,[.1 .7 .8 .2]) plot(Gamma,kMvec,’.’) ylabel(’Critical wave-number’) legend(’Linear analysis prediction’) 121 122 Matlab-codes D.2 cheb.m % CHEB compute D = differentiation matrix, x = Chebyshev grid function [D,x] = cheb(N) if N==0, D=0; x=1; return, end x = cos(pi*(0:N)/N)’; c = [2; ones(N-1,1); 2].*(-1).^(0:N)’; X = repmat(x,1,N+1); dX = X-X’; D = (c*(1./c)’)./(dX+(eye(N+1))); % off-diagonal entries D = D - diag(sum(D’)); % diagonal entries D.3 operator.m D.3 operator.m % Set up local discretised operators [dx,vecx] = cheb(Nx); dx = -(2/Lx)*dx; [dy,vecy] = cheb(Ny); dy = -(2/Ly)*dy; x = -Lx*vecx’/2+Lx/2; y = -Ly*vecy’/2+Ly/2; I = eye(Ny+1); dx = kron(dx,I); I1 = [1 0;0 0]; I2 = [0 0;0 1]; Dx = sparse(kron(I1,dx)+kron(I2,dx)); I = eye(Nx+1); dy = kron(I,dy); Dy = sparse(kron(I1,dy)+kron(I2,dy)); 123 124 Matlab-codes D.4 initcond.m u0 = a+b; v0 = b/(a+b)^2; uold = zeros(Ny+1,Nx+1); for i=1:Nx for j = 1:Ny m = max(i-1,j-1); c = rand(1)*(-1)^(round(rand(1)))*m/max(m,1)*... 1/(j*i)^(2^(round(rand(1)))); uold = uold+c*cos((j-1)*pi*Y/Ly).*cos((i-1)*pi*X/Lx); end end uold = reshape(uold,N2,1); vold = zeros(size(uold)); maxu = max(abs(uold)); uold = u0/100*uold/maxu; %vold = v0/100*(-uold)/maxu/d; figure axes(’fontsize’,12) contourf(X,Y,reshape(uold,Ny+1,Nx+1)) xlabel(’x’) ylabel(’y’) colorbar title(’(u-u^*)(0)’) D.5 solver.m D.5 solver.m % Tolerance for built-in MATLAB ODE solver options = odeset(’RelTol’,1e-10,’AbsTol’,1e-10); sol = [uold;vold]; [T,sol] = ode15s(@func2d,tspan,sol,options); sol = sol’; uend = sol(1:N2,end); vend = sol(N2+1:2*N2,end); clear sol T lengthT = length(T); tint = floor(lengthT./plotN); t = zeros(plotN,1); jstart = lengthT-tint*plotN; u = zeros(N2,plotN); v = zeros(N2,plotN); for j = 1:plotN t(j) = T(jstart + j*tint); u(:,j) = sol(1:N2,jstart+j*tint); v(:,j) = sol(N2+1:2*N2,jstart+j*tint); end 125 126 D.6 Matlab-codes recmovie.m figure %axis tight xlim([0 Lx]) ylim([0 Ly]) zlim([minuv maxuv]) set(gca,’nextplot’,’replacechildren’); % Record the movie for j = 1:plotN uplay = u(:,j); uplay = reshape(uplay,Ny+1,Nx+1); vplay=v(:,j); vplay = reshape(vplay,Ny+1,Nx+1); surf(X,Y,uplay) title([’t=’ num2str(t(j))]) F(j) = getframe; end %movie2avi(F,’animation.avi’) D.7 func2d.m D.7 func2d.m function w=func(t,u) global gamma diag10 u0u0 v0v0 global N2 Nx Ny %global LL global d Dx Dy global u0v0 ab % u u uu = [u(1:N2);u(1:N2)]; % v -v vv = [u(N2+1:2*N2);-u(N2+1:2*N2)]; ux = Dx*u; ux = reshape(ux,Ny+1,2*Nx+2); ux(:,1) = 0; ux(:,Nx+1) = 0; ux(:,Nx+2) = 0; ux(:,2*Nx+2) = 0; ux = reshape(ux,2*N2,1); uy = Dy*u; uy = reshape(uy,Ny+1,2*Nx+2); uy(1,:) = 0; uy(Ny+1,:) = 0; uy = reshape(uy,2*N2,1); lapl = Dx*ux+Dy*uy; lapl(N2+1:2*N2) = d*lapl(N2+1:2*N2); w = gamma*(ab-diag10*(u+u0v0)+(uu+u0u0).^2.*(vv+v0v0))+lapl; 127 128 Matlab-codes Bibliography [1] Adams, Robert A. - Sobolev Spaces, 1995, Academic Press, Inc. [2] Grisvard, Pierre - Boundary Value Problems in Nonsmooth Domains, 1985, Pitman, London [3] Grubb, Gerd - Distributions and Operators, Lecture notes, 2007. [4] Hansen, Vagn Lundsgaard - Moderne Analyse, 4th, 1995, World Scientific. [5] Henry, Daniel - Geometric Theory of Semilinear Parabolic Equations, 1st edition, 1981, Springer-Verlag. [6] Hollis, S. L.; Martin, R. H.; Pierre, M. - Global existence and boundedness in reaction-diffusion systems, Volume 18, No. 3, 1987, Siam J. Math. Anal. [7] Lions, J.-L.; Magenes, E. - Problèmes aux Limites non Homogènes, 2nd edition, Vol 2, 1968, Springer-Verlag [8] Murray, J. D. - Mathematical Biology I: An Introduction, 3rd Edition, Volume 17, 2002, Springer-Verlag. [9] Murray, J. D. - Mathematical Biology II: Spatial Models and Biomedical Applications, 3rd Edition, Volume 18, 2003, Springer-Verlag. [10] Pao, C. V. - Nonlinear Parabolic and Elliptic Equations, 1992, Plenum Press, New York [11] Pazy, A. - Semigroups of Linear Operators and Applications to Partial Differential Equations, 2nd Edition, Volume 44, 1983, Springer-Verlag. 130 BIBLIOGRAPHY [12] Trefethen, Lloyd N. - Spectral Methods in Matlab, Volume 1, 2002, Society for Industrial & Applied
© Copyright 2026 Paperzz