Topics in the Stability of Localized Patterns for some Reaction-Diffusion Systems by Ignacio Rozada B. Physics, Universidad Nacional Aut´onoma de M´exico, 2004 M. Applied Mathematics, University of New Mexico, 2006 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF Doctor of Philosophy in THE FACULTY OF GRADUATE STUDIES (Mathematics) The University Of British Columbia (Vancouver) August 2012 c Ignacio Rozada, 2012 Abstract In the first part of this thesis, we study the existence and stability of multi-spot patterns on the surface of a sphere for a singularly perturbed Brusselator and Schnakenburg reaction-diffusion model. The method of matched asymptotic expansions, tailored to problems with logarithmic gauge functions, is used to construct both symmetric and asymmetric spot patterns. There are three distinct types of instabilities of these patterns that are analyzed: self-replication instabilities, amplitude oscillations of the spots, and competition instabilities. By using a combination of spectral theory for nonlocal eigenvalue problems together with numerical computations, parameter thresholds for these three different classes of instabilities are obtained. For the Brusselator model, our results point towards the existence of cycles of creation and destruction of spots, and possibly to chaotic dynamics. For the Schnakenburg model, a differential-algebraic ODE system for the motion of the spots on the surface of the sphere is derived. In the second part of the thesis, we study the existence and stability of mesa solutions in one spatial dimension and the corresponding planar mesa stripe patterns in two spatial dimensions. An asymptotic analysis is used in the limit of a large diffusivity ratio to construct mesa patterns in one spatial dimension for a general class of two-component reaction-diffusion systems that includes the well-known Gierer Meinhardt activator-inhibitor model with saturation (GMS model), and a predatorprey model. For such one-dimensional patterns, we study oscillatory instabilities of the pattern by way of a Hopf bifurcation and from a reduction to a limiting ODEii PDE system. In addition, explicit thresholds are derived characterizing transverse instabilities of planar mesa-stripe patterns in two spatial dimensions. The results of our asymptotic theory as applied to the GMS and predator-prey systems are confirmed with full numerical results. iii Table of Contents Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv 1 2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Historical perspective . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Mathematical perspective . . . . . . . . . . . . . . . . . . . . . . 3 1.3 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 The Brusselator Model on the Surface of the Sphere . . . . . . . . . 7 2.1 2.2 The core problem and the construction of a quasi-equilibrium solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 Stability analysis of the quasi-equilibrium pattern . . . . . . . . . 18 Case I: The splitting case, m ≥ 2 . . . . . . . . . . . . . 21 Case II: m = 0 . . . . . . . . . . . . . . . . . . . . . . . 22 2.2.3 Case IIA: The competition case, m = 0 and τ = 0 . . . . 27 2.2.4 Case IIB: The Hopf bifurcation case, m = 0 and τ > 0 . . 39 Leading-order theory . . . . . . . . . . . . . . . . . . . . . . . . 46 2.3.1 46 2.2.1 2.2.2 2.3 Inner problem . . . . . . . . . . . . . . . . . . . . . . . . iv 2.3.2 Outer expansion . . . . . . . . . . . . . . . . . . . . . . 49 2.3.3 Case I: Vj0s = Vj0b , symmetric spot quasi-equilibrium . . 51 2.3.4 Case II: Asymmetric spot equilibria . . . . . . . . . . . . 51 Derivation from the S-formulation . . . . . . . . . . . . . . . . . 52 Core problem: Small S-asymptotics . . . . . . . . . . . . 53 Leading-order stability theory . . . . . . . . . . . . . . . . . . . 61 2.5.1 Stability thresholds . . . . . . . . . . . . . . . . . . . . . 70 Stability theory; Small S-Analysis from summing log formulation 73 2.6.1 Case A; τ = 0 . . . . . . . . . . . . . . . . . . . . . . . 74 2.6.2 Case B, τ = 0 . . . . . . . . . . . . . . . . . . . . . . . . 76 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . 84 The Schnakenberg Model on the Surface of the Sphere . . . . . . . 85 3.1 Localized spot patterns on the sphere . . . . . . . . . . . . . . . . 88 3.1.1 The quasi-equilibrium multi-spot pattern . . . . . . . . . 89 The spot self-replication threshold . . . . . . . . . . . . . . . . . 93 3.2.1 96 2.4 2.4.1 2.5 2.6 2.7 3 3.2 4 The competition instability threshold . . . . . . . . . . . 3.3 Slow spot dynamics on the surface of the sphere . . . . . . . . . 108 3.4 Quasi-equilibria and the cyclic matrix structure . . . . . . . . . . 116 3.5 Numerical method for reaction-diffusion patterns on the sphere . 121 3.6 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . 125 Case study: mesa patterns on the GMS system . . . . . . . . . . . . 126 4.1 Model formulation and Turing stability analysis . . . . . . . . . . 126 4.1.1 Turing stability analysis . . . . . . . . . . . . . . . . . . 127 4.2 Domain growth extension . . . . . . . . . . . . . . . . . . . . . . 131 4.3 Stability of 1-d mesa patterns for the Gierer-Meinhardt model with saturation (GMS) model when D = O(1) . . . . . . . . . . . . . . 135 4.4 4.3.1 Construction of a single mesa . . . . . . . . . . . . . . . 135 4.3.2 Bifurcation analysis . . . . . . . . . . . . . . . . . . . . 139 4.3.3 Hopf bifurcations of 1D mesa patterns . . . . . . . . . . . 144 Mesa patterns in the near-shadow limit . . . . . . . . . . . . . . . 147 4.4.1 Construction of a multi-stripe pattern . . . . . . . . . . . 148 v 4.4.2 Transverse stability in the near-shadow limit to perturbations in the y direction . . . . . . . . . . . . . . . . . . . 155 4.5 5 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . 169 The stability of mesa stripes in general reaction-diffusion systems . 170 5.1 Construction of the solution in the near-shadow limit . . . . . . . 171 5.2 Transverse stability of the K-mesa solution to perturbations along the y-axis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 5.2.1 5.3 5.4 5.5 6 The one-mesa special case . . . . . . . . . . . . . . . . . 193 Hopf bifurcation on 1-d mesa patterns in the shadow limit . . . . . 203 5.3.1 ODE-PDE system . . . . . . . . . . . . . . . . . . . . . 207 5.3.2 Stability proof for the breather case . . . . . . . . . . . . 211 Case study: the predator-prey model . . . . . . . . . . . . . . . . 216 5.4.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . 218 5.4.2 Stability in the near-shadow regime, D = O(ε −1 ) . . . . 221 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . 225 Future directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229 Appendix A Spherical coordinate tranformations . . . . . . . . . . . . 235 Appendix B Rigorous properties of NLEPs . . . . . . . . . . . . . . . 237 vi List of Tables Table 3.1 The norm of the difference between the first eigenvector and e, both normalized, for different spot configurations. . . . . . . . 120 Table 4.1 Some domain length values at which non-homogeneous solutions appear, according to Turing theory. The values were computed using the constants τ = 1, ε = 0.02, D = 1. The eigenmodes correspond to one, two, four, and eight peaks. These values can be seen overlapped in the full bifurcation diagram of Figure 4.4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 vii List of Figures Figure 2.1 Left figure: the profile of the spot solution uj (ρ) for various values of f and fixed spot strength Sj . As f increases, uj develops a volcano-shaped pattern that is commonly associated with splitting instabilities that occur in other systems such as the Gray-Scott model. Right figure: the nonlinear parameter χ(Sj ; f ) from the boundary condition in (2.6). The Boundary value problem (BVP) was solved with Matlab’s BVP5C routine. Each curve corresponds to a unique value of f , which ranges from 0.3 to 0.5. . . . . . . . . . . . . . . . . . . . . . Figure 2.2 12 The figure on the left shows the eigenvalue λ as the source strength Sj increases for the case f = 0.5. This figure is representative of what we saw for various values of f . The figure on the right tracks the values of (f, Sj ) that result in the critical case λ = 0 for m = 2. . . . . . . . . . . . . . . . . . . . . . Figure 2.3 Full numerical simulation on a sphere, u(x) for f = 0.7 and ε = 0.1. The splitting case has D = 0.35, whereas the non- Figure 2.4 22 splitting case has D = 0.45. . . . . . . . . . . . . . . . . . . 23 Competition instability threshold, i.e., stability with respect to the m = 0 mode when τ = 0. The figure on the left is for two spots located at opposite poles of the sphere, and the figure on the right computes the threshold as a function of the distance between the spots (for fixed values of f ). . . . . . . . . . . . viii 34 Figure 2.5 Full numerical simulation of the Brusselator model on a sphere (u(x)). A competition instability can be observed for the lower row (D = 1.2), which is not triggered in the top row (D = 0.8). Both scenarios have f = 0.7, ε = 0.1. Figure 2.6 Figure 2.7 . . . . . . . . . The Hopf bifurcation threshold for varying D (left), and as a function of f when D = 100. . . . . . . . . . . . . . . . . . Numerical solution to the ground-state BVP ∆ρ w − w + w2 36 44 = 0. 47 Figure 3.1 Numerical estimation of χ(Sj ) by solving the core problem (3.8) 90 Figure 3.2 Spot-splitting in the Schnakenberg model (u(x)). The same dynamics occurs in the lower hemisphere in this example, as the initial configuration consisted of two spots. The parameters were D = 1, ε = 0.1, R = 1.5, and a similar initial condition with R = 1 will does not split. . . . . . . . . . . . . . . . . . Figure 4.1 95 Solution profiles for various integration times starting close to the homogeneous solution. The figure on the left has the time evolution of the solution for u(x) up to t = 500, with time in the y-axis. The figure on the left represents four different snapshots, at t = 1, t = 50, t = 500, and t = 50, 000. We used κ = 2.5, D = 10, τ = 1, and ε = 0.01 on (4.1), with a random initial condition close to u = 0.5603, v = u2 . The numerical method we utilized was an implicit-explicit scheme. 128 Figure 4.2 Mesa profiles for various values of κ, obtained by numerically solving (4.1). The other parameters used are D = 10 and L = 1. The system was integrated using an implicit explicit scheme on a 500 point grid. . . . . . . . . . . . . . . . . . . 136 Figure 4.3 Three distinct two-mesa solutions to the GMS system. Solution I is close to the Turing instability, II is the stable mesa solution, and III is the unstable solution that develops when the domain length is increased past a critical point. The image on the right is the bifurcation diagram for the branch of two-mesa solutions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 ix Figure 4.4 Four branches of the GMS system, with an overlay of the family of stable solutions obtained by traversing it left to right . When reaching the fold point of each branch the solutions fall to the next branch, effectively doubling the number of mesas. The upper horizontal unstable line are the unstable Turing solutions, and the red points on it are the values shown on table 4.1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 Figure 4.5 The full solution curve for u(x) as the bifurcation branches in Figure 4.4 are traversed from left to right (image on the left), and from right to left (image on the right). The solutions are all plotted on a normalized domain, and the proper domain length L is represented on the y-axis. . . . . . . . . . . . . . . . . . 142 Figure 4.6 Solution curves for systems with growing domains, L(t) = eρt . Notice the delay in the bifurcation (jump between branches) as ρ gets larger. The figure on the right shows the effect on Figure 4.5 (left) when adding domain growth, with ρ = 0.002. The y-axis represents L(t), and L = 15 is reached when t = 1354. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 Figure 4.7 Stability curve of the maximum eigenvalue vs L for solutions on the 1-mesa branch. . . . . . . . . . . . . . . . . . . . . . 145 Figure 4.8 Real part of eigenvalues as τ approaches critical value. The x and y axis are the real and imaginary components of the eigenvalues, respectively. . . . . . . . . . . . . . . . . . . . . 146 Figure 4.9 Solution graphs for both u(x) (top), and v(x) (bottom), for large τ . Both numerical computations were done for τ = 380, the ones on the left with a domain length L = 1.6 and the ones on the right had L = 2.02. The horizontal axis is time, and the vertical axis is the domain length. . . . . . . . . . . . . . . . 147 Figure 4.10 A typical mesa profile in the stationary solution v(x). The left and right edges of the mesa are labelled as χl and χr respectively; and the length of the mesa section is l. . . . . . . . . . 148 Figure 4.11 A plot of the function f (w) given in (4.21). . . . . . . . . . . 149 x Figure 4.12 Eigenvalues for a two mesa solution. The parameters are D = 0.5, ε = 0.001 and κ = 3 for the Figure on the left. The λ− eigenvalues are the zigzag ones, while λ+ are the breather ones. The Figure on the right has the critical (κ, D) values for instability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 Figure 4.13 Full 2D simulation with parameters ε = 0.01, D = 0.5, κ = 1.5. The solutions were integrated using an IMEX algorithm. The solution on the left has d0 = 1.5, and the solution on the right has d0 = 2. . . . . . . . . . . . . . . . . . . . . . . . . 168 Figure 5.1 Plots of both the critical τ and λI at which a Hopf bifurcation occurs, as a function of the domain length L for the GMS model. The parameters used in the computations are D = 50, ε = 0.01. The two top figures are for κ = 1, and the bottom figures are for κ = 0.65. . . . . . . . . . . . . . . . . . . . . 206 Figure 5.2 A comparison between the ODE-PDE system (5.45) and the full numerical simulation for a system beyond the Hopf threshold. The figures in the left correspond to τ = 25, 000, and the images in the right to τ = 65, 000. The rest of the parameters are D = 50, ε = 0.01. The solution was integrated until T = 10, 000, with an IMEX scheme with 800 grid points. . . 212 Figure 5.3 The contour on which to check the Nyquist stability criterion. Figure 5.4 Turing space for the system given by 5.52, in term of the pa- 213 rameters a and b. . . . . . . . . . . . . . . . . . . . . . . . . 218 Figure 5.5 The figure on the left shows the projected half-width of a mesa versus the parameter a, for two values of b, and for L = 0.5. Notice that there is a consistency requirement on a, given that we must satisfy l < L. The second figure shows two stationary solutions for different values of the parameters that illustrates the change in mesa width. . . . . . . . . . . . . . . . . . . . 220 xi Figure 5.6 The bifurcation diagram for the one mesa solution, and solutions corresponding to various points along the branch. The bifurcation diagram was computed using AUTO [14]. The parameters used were a = 3, b = 2, τ = 1, D = 1, with the asymptotic term ε = 0.02. . . . . . . . . . . . . . . . . . . . 220 Figure 5.7 The values of u+ and V that satisfy the heteroclinic connection as a function of the parameter a (right figure), and the parameter βpp = Figure 5.8 ∞ 2 −∞ (U0 ) dy. . . . . . . . . . . . . . . . . . . . . 221 The four eigenvalues of a two mesa solution, as a function of b (left), and as a function of M (right). The parameters are D = 3, A = 1.6, ε = 0.01, L = 1,, and M = π for the figure on the left, and B = 3.5 for the figure on the right. . . . . . . 223 Figure 5.9 Full numerical simulation of the Predator-Prey model on a 2D lattice. We used ε = 0.01, D = 0.4, a = 1.6, b = 3.5, τ = 1. Both lattices were −1 < x < 1, and the lattice on the left had 0 < y < 0.8, while the lattice on the right had 0 < y < 2. The figures on the left were integrated until T = 5, 000, and the figures on the right until T = 10, 000. . . . . . . . . . . . 224 xii Glossary GMS Gierer-Meinhardt model with saturation RDE reaction-difusion equations ODE ordinary differential equations PDE partial differential equations BVAM Barrio-Varea-Arag´on-Maini model BVP Boundary value problem NAS Nonlinear algebraic system NLEP Nonlocal eigenvalue problem xiii Acknowledgments I dedicate this thesis to my father, it was a long journey and he was always there with me. I want to thank specially my advisor Michael Ward, for his support and his encouragement. The faculty, staff and graduate students at UBC have been wonderful and it’s been a great time. I appreciate the financial support of the UBC scholarship and the SEP graduate fellowship. xiv Chapter 1 Introduction 1.1 Historical perspective The formal study of pattern formation mechanisms in the life sciences owes much to Alan Turing’s seminal paper [60] on a mathematical model for a pattern-generating chemical reaction. The motivation behind the model was to shed light on the symmetry-breaking and differentiation mechanisms in biological organisms. Turing’s work presented a mechanism that can generate patterns from an initially homogeneous medium. By performing linear stability analysis, he showed it is possible to determine conditions for the existence of stable spatially-inhomogeneous solutions. In physical terms this means that under some conditions, a two-component chemical reaction can evolve into a stable non-trivial pattern. This new concept took many years to be accepted, despite early experimental evidence. In the 1950s the Russian chemist Boris Belousov [4] reported a self-oscillating chemical reaction, although he never managed to publish his findings in a peer-reviewed journal. His discovery eventually became known through one of his students, and the chemical reaction is now known as the Belousov-Zhabotinsky (BZ) reaction. 1 A mixing of two chemicals that does not lead to a dissipation in their gradients at first sight conflicts with the second law of thermodynamics; despite this seeming impossibility, similar mechanisms were soon discovered. The understanding of the physical mechanisms behind the reactions partly yielded Ilya Prigogine the 1977 Nobel prize in Chemistry [47]. In Turing’s original paper, the basic model consists of a system of two nonlinear PDEs Ut = Du ∆U + F (U, V ), (1.1) Vt = Dv ∆V + G(U, V ), with U, V representing the concentration of two chemicals, Du , Dv their diffusivities, and F (U, V ), G(U, V ) the nonlinear reaction terms. The main insight was that under the right conditions, a spatially homogeneous solution of (1.1) could be destabilized by the presence of the diffusion terms. This process is now called diffusion-driven instability. One of the so-called Turing conditions for a diffusiondriven instability is that the ratio of the diffusion coefficients Du /Dv be large. The rest of the conditions are also obtained through linear stability analysis (we work this out in detail for the Gierer-Meinhardt model with saturation (GMS model) in § 4.1). In 1972, Gierer and Meinhardt [15] extended the idea of diffusion-driven instabilities with the observation that patterning occurs through the interaction between an auto-catalytic short range activator, and a long range inhibitor. This concept became extremely popular in biological modelling, and was subsequently applied to modelling skin pigment patterns in fish [25], fingerprints [5], colouring of marine shells [34], animal coat markings [36], and many others. In a much larger scale, models for interacting populations of predators and prey instead of chemicals were studied in 1972 [53], and more recently in [62], [2]. A survey of many reaction-diffusion models can be found in [37] and [29]). While most of the early work was done in systems of one and two dimen- 2 sions, there soon started to be studies on how patterns were affected by growth and form (paraphrasing Thompson’s classic, century old book on biological patterns [57]). Early studies on 1D systems that evolved on a growing domain ([12] [13] [3]) seemed to show that domain growth increased the robustness of pattern selection. Recent studies have incorporated domain growth to modelling limb development [35], and growth on plant tips [38]. Moreover, on a fundamental level, reworking Turing analysis on a general model with domain growth (making it nonautonomous) shows that diffusion-driven instabilities can occur in more general types of kinetics, beyond the activator-inhibitor framework [28]. Besides growth, curvature has been shown to have a profound effect on reaction diffusion models. Models on spheres range from simulations of Radiolaria structure [61], spherical tumour growth [10], to modelling plant tip growth on halfhemispheres [38], and single cell models [26] where bulk diffusion within the cell was coupled to diffusion along the boundary. In more general terms, it has been shown that the geometry of the domain, and specifically changes in curvature, can stabilize localized structures to critical points of the mean curvature [59]. Without question the Turing paradigm has been extremely successful. Applications range from single cell models up to herd dynamics and probabilistic models of criminal activity ([54]), overlapping multiple fields. 1.2 Mathematical perspective As Turing models have grown more complex, incorporating domain growth and complicated topologies, there is a need for increasingly sophisticated mathematical tools capable of drawing insights from the models. The key limitation of a Turing-type analysis is that it is linear in nature, and the fact that its pattern prediction capacity is severely hampered in both higher dimensions and in large domains. By virtue of its linearity, the Turing patterns predicted 3 will be close to the homogeneous solution. However, many physical and biological systems exhibit concentration gradients and localized structures that are far from equilibrium. These cannot be studied by relying on linear analysis. Furthermore, models in two and three spatial dimensions, have the complication that the pattern modes become degenerate. In these regimes, Turing analysis cannot predict the modes that will arise for specific parameter regimes. Gjorgjieva’s work for a specific reaction-diffusion system on the surface of a sphere ([16], [17]) provides a good illustration of this phenomena. Weakly nonlinear analysis has been used to successfully study the bifurcation structure leading from the Turing regime. However, its effectiveness is constrained to a region close to the Turing instability, and as such it provides poor results in regimes far from equilibrium. In this thesis we will work with some of the classic reaction diffusion models: the Schnakenberg model [52], Brusselator [47], Gierer-Meinhardt [15], and a spatio-temporal predator-prey model [62]. The goal of our work is to develop analytical tools that are applicable to a wide variety of reaction-diffusion systems, and that provide insight into the existence and stability of solutions far from equilibrium, as well as on the dynamical processes that occur in those regimes. The models we used have been well studied and there is a large body of literature, in the case of the Brusselator going back to the seventies. Starting from dimensionless versions of the models, we will perform an asymptotic analysis in the asymptotic limit of parameters that are either very large or small. We will develop particle-like solutions from matched asymptotic expansions in the singular limit, and construct solutions both for spot-type solutions, and for mesa-type patterns. The analytical results for these solutions and their stability properties will be verified with full numerical simulations. 4 1.3 Thesis outline This thesis consists of two main parts. The first part involves the study of spot patterns for the Brusselator and Schnakenberg models on the surface of a sphere. For both models we will construct localized spot-type solutions in the singular limit of small diffusivity using the method of matched asymptotic expansions. Away from the localized spatial regions where the spots are concentrated, the approximate solution will be shown to satisfy a linear elliptic problem where the spots are replaced by effective Coulomb singularities. This leads to a particle-like solution characterization of the asymptotically reduced problem. We will use a result from the theory of point vortices, for which extensive literature already exists ([21], [39], [41], [40], [8]) regarding a Neumann Green’s function on the sphere that can be used to construct the solutions. From analyzing the stability of the full nonlinear system for the case where all the spots have a common spot strength, we will derive a DAE that couples the strengths of the spots to their position on the sphere. The resulting problem is again related to point-vortices ([18], [6]), and the possible solutions are those of the Fekete problem [56], as well as the original Thomson atom model ([58], [1]). Upon analyzing the stability properties of these solutions, we will derive analytical formulae for the thresholds of three distinct types of instabilities both by leading-order stability theory, through a related non-local eigenvalue problem [63], and by numerical calculation. These three instabilities all relate to instabilities of the amplitudes of the spots in a spot pattern, and they occur on a fast O(1) time-scale. As discussed in Chapter 3, they have no direct counterpart with any translational-type instability mechanism for fluid point-vortices on the sphere. In the second part of this thesis we will study reaction-diffusion models that admit mesa solutions, which consists of block-like patterns. Starting in one spatial dimension, we will analyze the bifurcation structure of these solutions, and study 5 the effect of a slowly growing domain on mesa-stability and mesa-splitting. We will then consider the near-shadow regime, where the ratio of diffusivities of the two reaction components is very large. In this regime we will construct solutions far from equilibrium using the method of matched asymptotic expansions, and we will study the stability properties of these solutions. Starting with the GMS model ([22],[23]), we will consider the case of multiple mesa solutions, and develop a general framework for other models. We will extend the 1-D solutions to 2-D planar stripe patterns, and then analyze transverse instabilities of these patterns. Our general framework for the analysis of transverse instabilities will also be applied to a Predator-Prey reaction-diffusion system, where we will compare our analytical predictions with results from full numerical simulations. 6 Chapter 2 The Brusselator Model on the Surface of the Sphere In this chapter we study the existence and stability of localized spot patterns on the surface of the sphere for the Brusselator reaction-diffusion model. As surveyed in the introduction, there have been many studies characterizing weakly nonlinear patterns for the Brusselator that emerge from a linearized Turing-instability type analysis. The analysis of such weakly nonlinear patterns is rather complicated owing to the degeneracy of the Laplacian eigenfunctions on the surface of the sphere. In contrast, in a singularly perturbed limit, the method of matched asymptotic expansions will be used to construct localized spot-type patterns for the Brusselator model. A precise asymptotic characterization of these patterns and the parameter ranges where they occur will be found, and the stability of these patterns analyzed. Three types of instabilities of these patterns will be discovered: a spot self-replication instability, a competition instability leading to the annihilation of spots, and a breather-type temporal instability of the spot amplitudes. Parameter ranges in terms of a phase diagram where these three instabilities occur will be determined. 7 The standard form of the Brusselator [47] model is given by ∂U = ε 2 ∇2 U + E − (B + 1)U + U 2 V, ∂t ∂V = D∇2 V + BU − U 2 V. ∂t (2.1) The variable U is the short range activator component, and V is the long range inhibitor component. We first give a formal scaling argument to determine the range of parameters with respect to ε for which spot patterns exist. We start by letting V = O(Vg ) globally, with Vg the stationary homogeneous solution. For U we have different scalings near and away from spots; Uinn = O(Uinn ) near a spot, and Uout = O(Uout ) away from spots. In the inner region, with y = ε −1 (x − x0 ), we need O(U 2 V ) = O(U ) in order to have a construct a spot profile. Hence Uin Vg = O(1), so that Uin = O(1/Vg ) in the inner region. In the outer region, from the V equation in (2.1) we need that the nonlinear term, which will be localized near a spot, be approximated by a Delta function with the correct strength. As such we have Ω 2 ) = O(DV ). U 2 V dx = O(ε 2 Vg Uin g Thus, Uin = O(ε −1 ) and consequently Vg = O(ε ). Next, from the balance in the outer region that D∆V = BU , we get that Vg = O(Uout ), which means that Uout = O(ε ). Finally, in the outer region we obtain from the U equation in (2.1) that E must balance (B + 1)U , so that E = O(Uout ) which yields E = O(ε ). This very formal scaling analysis suggests that Uin = O(1/ε ), Uout = O(ε ), Vg = O(ε ), when E = O(ε ). Next, we use a non-dimensionalization based on these scalings in order to reduce the number of parameters in (2.1) and isolate our key bifurcation parameters. 8 We define the new variables u, v, and τ by t = τT , U= µ u, ε V = εv, E = ε E0 In this way, (2.1) becomes 1 ε2 ε 2 E0 µ uτ = ∆u + −u+ u2 v, T (B + 1) B+1 B+1 µ B+1 D 1 B 1 vτ = 2 ∆v + 2 u − u2 v . 2 µ T µ ε µ (2.2) We now define ε 0 , T , and µ by ε0 = √ ε , B+1 1 , B+1 T = µ=B. Then, (2.2) transforms to E0 B −u+ u2 v, B B+1 (B + 1)2 D(B + 1) 1 vτ = ∆v + 2 u − u2 v , 2 2 B B ε0 uτ = ε 20 ∆u + ε 20 In this way, and upon replacing τ by t, we obtain the starting system for our analysis given by ut = ε 20 ∆u + ε 20 E − u + f u2 v, τ vt = D∆v + 1 (u − u2 v). ε 20 (2.3) Here we have defined f= B , B+1 τ= 1 , f2 D= D(B + 1) , B2 E= E0 . B (2.4) We remark that in our non-dimensionalization, we chose not to scale v. If we were to additionally re-scale v, then we obtain (2.3) in which E = 1 and with a slight re-definition of the other parameters D and τ . We choose instead to work with 9 (2.3) as it better isolates bifurcations due to changes in the parameter E0 . We remark that the key bifurcation parameter f is defined conveniently in the narrow interval (0, 1). As B → ∞ we get that τ → 1, f → 1, E → 0, and D = O(1/B). The system (2.3) on the surface of the sphere will be the starting point for our analysis. Without loss of generality we can let the sphere have radius one. When considering the system on the surface of the unit sphere it is understood that ∆ ≡ ∆s , where ∆ is the Laplace-Beltrami operator given by ∆s u ≡ 2.1 1 1 2 uφφ + sin θ (sin θ uθ )θ , sin θ 0 < θ < π, 0 < φ < 2π. The core problem and the construction of a quasi-equilibrium solution We now construct a multi-spot quasi-equilibrium pattern for (2.3) in the limit ε 0 → 0. For convenience we will re-label ε 0 by ε in the calculations below. We first formulate the local (or inner) problem that determines the profile of an isolated spot. We center a spot at the angular coordinates φ = φj and θ = θ j , and we define ˆ y1 = sin θ j φ, y2 = θˆ , φ − φj with φˆ = , ε θ −θj θˆ = . ε Then, in the inner region near this spot we obtain, with an O(ε ) error, that ∆s u = uy1 y1 + uy2 y2 + O(ε ). With this tangent-plane type-approximation to the sphere, we now construct a quasi-equilibrium spot pattern solution, with spots centred at (φj , θ j ) for j = 10 1, . . . , N. In the inner region near the j-th spot, we obtain to O(ε ) accuracy that (2.3) reduces to ∆y Uj − Uj + f Uj2 Vj = 0, D∆y Vj + Uj − Uj2 Vj = 0. Then, by rescaling Uj = √ −∞ < y1 , y2 < ∞, (2.5) √ Vj = vj / D , Duj , we can eliminate D and obtain the radially symmetric core problem for (uj , vj ) in terms of the sole bifurcation parameter f : ∆ρ uj − uj + f u2j vj = 0, ∆ρ vj + uj − u2j vj = 0, 0 < ρ < ∞, (2.6) uj (0) = vj (0) = 0, uj → 0 as ρ → ∞, vj ∼ Sj log ρ + χ(Sj ; f ) + 0(1), where we have defined ρ = as ρ → ∞, y12 + y22 , and ∆ρ ≡ ∂ρρ + ρ1 ∂ρ . The key feature in this problem is that we impose that vj ∼ Sj log ρ as ρ → ∞, which is appropriate for ∆ρ vj = (u2j vj −uj ) owing to the fact that uj → 0 at infin- ity. The constant Sj is a parameter at this stage, but it will eventually be determined after the asymptotic matching of the inner and outer solutions. However, in terms of Sj and the bifurcation parameter f , the key function χ(Sj ; f ) must be computed numerically from the condition that vj − Sj log ρ = O(1) as ρ → ∞. The boundary value problem (2.6) was solved for particular values of f and Sj by approximating this problem on a large but finite domain 0 ≤ ρ ≤ R, where R 1. In this way, we determined χ(Sj ; f ) by computing vj at ρ = R. We took R = 15 in our computations. This calculation of χ(Sj ; f ) is of key importance for the rest of the analysis. 11 Figure 2.1: Left figure: the profile of the spot solution uj (ρ) for various values of f and fixed spot strength Sj . As f increases, uj develops a volcano-shaped pattern that is commonly associated with splitting instabilities that occur in other systems such as the Gray-Scott model. Right figure: the nonlinear parameter χ(Sj ; f ) from the boundary condition in (2.6). The BVP was solved with Matlab’s BVP5C routine. Each curve corresponds to a unique value of f , which ranges from 0.3 to 0.5. There are a few identities that will be important later on. In the limit as R → ∞, we have R lim R ρ∆ρ vj dρ = R→∞ 0 0 (u2j vj − uj )ρdρ , and since vj ∼ Sj log ρ, with ρ(∆ρ vj ) = (ρvjρ )ρ , we obtain ∞ R lim R→∞ 0 (ρvj )ρ dρ = 0 (u2j vj − uj )ρdρ = Sj . (2.7) In a similar way, we obtain from the uj equation that ∞ R 0= 0 ∞ uj ρdρ − f ρ∆ρ uj dρ = 0 0 u2j vj ρdρ. Combining this with (2.7), we conclude that ∞ Sj = (1 − f ) 12 0 u2j vj ρdρ. (2.8) Since we do not know the sign of vj we cannot guarantee that Sj > 0. However, it is clear that as f → 1 then Sj → 0. Next, we asymptotically match the far-field behaviours of the inner solutions near each spot to a certain global solution for v, which we will construct. In doing so, we will derive a nonlinear algebraic system of equations for the unknowns Sj , referred to as the “source strengths”. Our asymptotic analysis has the key feature that it retains all of the logarithmic terms in ν ≡ −1/ log ε as ε → 0, and so our asymptotic approximation for the solution and for the source strengths has an error that is algebraic, rather than logarithmic, in ε . To determine the far-field behaviour of each inner solution we recall that u = D1/2 uj and v = D−1/2 vj , with vj ∼ Sj log |y| + χ + o(1) as |y| → ∞. We let xj = (cos φj sin θ j , sin φj sin θ j , cos θ j ) be a point on the unit sphere. Now, by Lemma A.1 in appendix A, we have that |x − xj | = ε |y| + o(1) as x → xj , with y = (φˆ sin θ j , θˆ ), and φˆ = φ−φj ε , θˆ = θ −θ j ε . Thus, we obtain the far-field behaviour and matching condition v ∼ D−1/2 Sj log |x − xj | + Sj + χ(Sj ; f ) , as x → xj ; ν ν≡− 1 . log ε This provides the singular behaviour of the outer solution for v. Next, we study the outer solution for (2.3). Meanwhile, in the outer region away from O(ε ) neighborhoods of {x1 , . . . , xn } we have that ε 2 E−u+f u2 v = 0. So the outer limit for u is u ∼ ε 2 E + O(ε 2 ). We have then that the outer and inner solutions for U are, respectively, uout = ε 20 E, and uin = D1/2 uj (ε −1 |x − xj |), whereas the inner approximation for v is v ∼ D−1/2 vj . By combining the global 13 and local parts, we get a uniformly valid approximation for u given by N u ∼ ε 2E + j=1 (D1/2 uj (ε −1 |x − xj |) + . . .). We then must estimate the term ε −2 (u − u2 v) in the v-equation of (2.3) in the sense of distributions. The evaluation of this term requires care to retain both the local contribution near each spot and the global contribution arising from the non-vanishing outer solution for u of order O(ε 2 ). In the sense of distributions we obtain ∞ 1 (u − u2 v) ∼ E + 2πD1/2 ε2 0 (uj − u2j vj )ρdρ δ (x − xj ), ∼ E − 2πD1/2 Sj δ (x − xj ). By using this result, together with the matching condition for v as given above, we obtain that the outer problem for v is 2π E =√ ∆s v + D D N j=1 Sj δ (x − xj ) v ∼ D−1/2 Sj log |x − xj | + in Ω , (2.9) Sj + χ + o(1) ν as x → xj , for j = 1, . . . , N , where Ω is the surface of the unit sphere. A key feature in this problem is that by pre-specifying the form of the non-singular O(1) term in each singularity condition, we will obtain a nonlinear algebraic system for the source strengths S1 , . . . , SN . To solve this problem we introduce the Neumann Green’s function G(x; x0 ) defined as the unique solution to ∆s G = 1 − δ (x − x0 ) , 4π 2π π G dx = G sin θ dθ dφ = 0. (2.10) 0 S 14 0 Here G is 2π periodic in φ, is smooth at the poles θ = 0, π, and the integral condition eliminates an arbitrary constant in G and thus specifies it uniquely . It is well known ([19], [20], [21]) that G=− 1 log |x − x0 | + R, 2π R= 1 (2 log 2 − 1) . 4π (2.11) We can write the solution to (2.9) in terms of G as 2π v = −√ D N v¯ Sj G(x; xj ) + √ , D j=1 (2.12) where v¯ is an arbitrary constant that must be determined as part of the analysis. To verify that (2.12) has the correct strength of the logarithmic singularity, we calculate 2π ∆s v = − √ D N 2π Sj ∆s G = − √ D j=1 N Sj j=1 1 4π 2π +√ D N j=1 Sj δ (x − xj ). This leads to the condition that the sum of the source strengths are related to E by N 2E Sj = √ . D j=1 As x → xi , the matching condition in (2.9) together with the explicit solution for v in (2.12) yields that 2π Si 2π −√ − log |x − xi | + Si R − √ 2π D D N v¯ Sj Gij + √ D j=1 j=i ∼ D−1/2 Si log |x − xi | + Si +χ , ν for i = 1, . . . , N . This results in a system of N + 1 nonlinear algebraic system 15 Nonlinear algebraic system (NAS) for Si and v¯, Si + 2π Si R + ν N Sj Gij + χ(Si ; f ) = v¯, i = 1, . . . , N, j=i (2.13) N 2E Si = √ . D i=1 Here Gij = G(xi ; xj ) is to be computed from (2.11) Before casting this system into a more convenient form, we make a few remarks. Firstly, the nonlinearity in (2.13) arises from the χ(Si ; f ) term. For a given set of spot locations x1 , . . . , xN , we can compute S1 , . . . , SN and v¯. This then determines the quasi-equilibrium pattern. For each Sj our numerical results indicate that there is a unique solution to the core problem (2.6). The outer solution for v will be given by (2.12). Due to the intractability in solving this system analytically, the solvability of this system must (typically) be explored numerically. Next, we decompose the Gij terms in the NAS to re-cast this system into a more convenient form. We write Gij = 1 2π Lij + R, where R = 1 4π (2 log 2 − 1) (see (2.11)), and where we have labelled Lij ≡ log |xi − xj |. We calculate that N Si R + j=i Sj − 1 Lij + R 2π =− 1 2π =− 1 2π N N Sj Lij + j=i Sj R, j=1 N 2RE Sj Lij + √ , D j=i so that (2.13) becomes N Si − ν νR0 E Sj Lij + √ + νχ(Si ; f ) = ν v¯, D j=i R0 ≡ 4πR = 2 log 2 − 1 , (2.14) for i = 1, . . . , N . 16 To write this system in matrix form, we introduce 1 S1 . . . . S= e= . , . , 1 SN 0 L12 · · · L1N .. .. L . . 21 G= . , .. LN 1 · · · 0 χ(S1 ; f ) .. , χ= . χ(SN ; f ) (2.15) so that the system for the source strengths becomes (I − νG)S + ν χ = νR0 E ν v¯ − √ D We multiply by eT and use the fact that eT S = 2E √ , D e. and eT e = N . This yields 2E νR E √ − eT νG S + νeT χ = ν v¯N − √ 0 N, D D which allows us to solve for v¯ as 2E R0 E eT χ eT G S v¯ = √ + √ + − . N N DN ν D By eliminating v¯ in (2.13) we obtain S+ν 1 eeT G − G S + ν χ − eeT χ N N 2E =√ e. DN We now define the matrix 1 ··· 1 T 1 .. E0 ≡ ee = N N . 1 ··· 17 1 .. . . 1 (2.16) With this we can conclude that the solution to (2.13) satisfies 2E S + ν(I − E0 )G S + ν(I − E0 )χ = √ e, DN (2.17a) and that v¯ = √ R0 E 1 2E + √ + (eT χ − eT G S). N DN ν D (2.17b) We remark that the nonlinear algebraic system in (2.17) is decoupled. One first solves for S1 , . . . , SN in (2.17a), and then the result is used to calculate v¯ in (2.17b). By using the numerical values computed from the core problem for χ(S; f ) (see Figure 2.1), one can solve (2.17a) for various spot configurations. The specific form (2.17a) of the nonlinear algebraic system is the one that is used in the analysis below. 2.2 Stability analysis of the quasi-equilibrium pattern In this section we study the stability of the N -spot quasi-equilibrium solution constructed in the previous section to O(1) time-scale instabilities. The O(1) timescale of such instabilities is fast in comparison with the expected slow dynamics of the spots with speed O(ε 2 ). Therefore, in the stability analysis we “freeze” the locations of the spots and then characterize whether the resulting quasi-equilibrium pattern is unstable to fast O(1) time-scale instabilities. There are three distinct types of instabilities that can occur and will be discussed. We remark that our stability analysis is accurate to all logarithmic orders in ν. A leading-order stability analysis is given in a later section. 18 We begin with the nondimensionalized system from (2.3) on Ω written as ut = ε 2 ∆u + ε 2 E − u + f u2 v, 1 τ vt = D∆v + 2 (u − u2 v) ε where Ω is the surface of the unit sphere. We recall that the quasi-equilibrium solution, as constructed in the previous section, satisfies N √ 2 uqe ∼ ε E + vqe ∼ j=1 Duj (ε −1 |x − xj |), √1 vj D − √2πD N j=1 Sj G(x; xj ) + √v¯ D for |x − xj | = O(ε ) for |x − xj | O(ε ) We linearize around this solution by writing u = uqe + eλt ψ, v = vqe + eλt η, to obtain that the perturbation satisfies ε 2 ∆s ψ − ψ + 2f uqe vqe ψ + f u2qe η = λψ, 1 D∆s η + 2 (ψ − 2uqe vqe ψ − u2qe η) = τ λη. ε (2.18) For the inner solution, we consider the local coordinates near the j-th spot φ − φj φˆ = , ε θ −θj θˆ = ; ε ˆ y1 = sin θ j φ, y2 = θˆ . We re-write the Laplace-Beltrami operator in the local coordinate system to get ψjy1 y1 + ψjy2 y2 − ψj + 2f uj vj ψj + f Du2j ηj = λψj , D(ηjy1 y1 + ηjy2 y2 ) + ψj − 2uj vj ψj − Du2j η = τ ληj ε 2 , 19 (2.19) since in the inner region we have uqe vqe we assume that τ λ O(ε −2 ), uj vj . We now define Nj = Dηj , and which leads to the eigenvalue problem ψjy1 y1 + ψjy2 y2 − ψj + 2f uj vj ψj + f u2j Nj = λψj , Njy1 y1 + Njy2 y2 + ψj − 2uj vj ψj − u2j Nj = 0, (2.20) on −∞ < y1 , y2 < ∞. In deriving this system we used ∆s ψ = ψy1 y1 +ψy2 y2 +O(ε ), and neglected the O(ε ) error term. In addition, we recall that uj and vj are obtained by solving the radially symmetric core problem (2.6), and that they depend on Sj and f . Finally, we note that due to the −ψj term in the ψj equation, it is consistent to impose that ψj → 0 as ρ2 = y12 + y22 → ∞, provided that λ > −1. However, the boundary conditions on Nj will depend on the type of eigenfunction that we are seeking. We now look for a separation of variable solution of the form ψj = ψˆj (ρ)eiwm , Nj = Nˆj (ρ)eiwm , where w = tan−1 (y2 /y1 ) and m = 0, 1, 2, . . . (e2πim = 1 ∀m), provides a peri- odicity condition for the perturbation in the tangent plane. Overall, this results in a local polar coordinate system on the tangent plane to the sphere at the j-th spot location. In terms of these variables, (2.20) becomes a radially symmetric problem with parameter m: 1 m2 ˆ ˆj = 0 ψˆj + ψˆj − ψj − (1 + λ)ψˆj + 2f uj vj ψˆj + f u2j N ρ ρ 2 ˆ + 1N ˆ −m N ˆj + ψˆj − 2uj vj ψˆj − u2 N ˆ N 0 < ρ < ∞, j j j j = 0, ρ ρ ˆ (0) = 0, ψˆj → 0 as ρ → ∞. ψˆ (0) = N j (2.21) j We need only consider the modes m = 0, 2, 3, 4, . . ., since m = 1 corresponds to the translation mode associated with the neutral eigenvalue λ = 0. 20 2.2.1 Case I: The splitting case, m ≥ 2 We now look for solutions to the eigenvalue problem generated by non-radially ˆj /ρ2 symmetric perturbations near the j-th spot (m = 2, 3, . . .). Due to the −m2 N ˆj equation we can impose an algebraic decay as N ˆj → ∞. Thus we term in the N ˆj → 0 as ρ → ∞. can append to (2.21) the condition N This implies that non-radially symmetric eigenfunctions are largely local instabilities, and are only coupled together through the NAS for S1 , . . . , SN . We proceed to solve the eigenvalue problem. For particular values of f , we compute uj and vj from the core problem. Despite the boundary condition being set at infinity, the solution converges exponentially fast, and it became clear that discretizing (2.21) for 0 < ρ < 14 would result in convergence to machine preˆj in the second equation, using centred differences for cision. We first solve for N the first and second derivatives, and upon substituting this into the first equation we can approximate ψˆj and the eigenvalue λ by the solution of a matrix eigenvalue problem. In Figure 2.2 we show the eigenvalues as a function of the parameter Sj for modes m = 2, 3, 4 for a specific value of f . In addition, we give the value of the pairs (f, Sj ) for which λ = 0 for the m = 2 mode. The figure agrees with the expectation that the solution becomes unstable at the range where uj attains a volcano-like profile (shown in Figure 2.1) The results indicate that if the source strength exceeds some critical value Sc , which depends on f , then the spot becomes unstable to a mode m = 2 linear instability. This linear instability is of peanut-splitting type and is the trigger for a nonlinear spot self-replication event. To illustrate this phenomena, in Figure 2.3 we show numerical results computed from the full PDE system for a one-spot initial condition for the parameter 21 λ vs S j for f = 0.5 S (F; λ = 0) for M = 2 j 0.4 12 M = 2 M = 3 M = 4 0.2 0 10 −0.2 8 −0.4 λ Sj 6 −0.6 −0.8 4 −1 2 −1.2 −1.4 0 1 2 3 4 Sj 5 6 7 0 0.2 8 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 F Figure 2.2: The figure on the left shows the eigenvalue λ as the source strength Sj increases for the case f = 0.5. This figure is representative of what we saw for various values of f . The figure on the right tracks the values of (f, Sj ) that result in the critical case λ = 0 for m = 2. values f = 0.7 and ε = 0.1. The numerical method used to obtain these results is described at the end of next chapter on the 2-D Schakenburg mode. In the first row of this figure where D = 0.35, we have that S > Sc and so there is a peanut- splitting linear instability. This instability is shown to lead to spot self-replication. Alternatively, for D = 0.45, then S < Sc and, as shown in the second row of this figure, there is no spot self-replication. 2.2.2 Case II: m = 0 The case m = 0 corresponds to a locally radially symmetric perturbation near the j-th spot. The key difference, as compared with the m = 2, 3, · · · case, is that without the −m2 Nj /ρ2 term we cannot impose that Nj → 0 as ρ → ∞. Instead, we must allow for logarithmic growth in the far-field. Returning to the eigenvalue problem (2.21), we impose the far-field behaviour Nj ∼ cj log ρ, 22 as ρ → ∞. Figure 2.3: Full numerical simulation on a sphere, u(x) for f = 0.7 and ε = 0.1. The splitting case has D = 0.35, whereas the non-splitting case has D = 0.45. The constants cj for j = 1, . . . , N will be determined later from a global matrix problem that couples all local eigenvalue problems near each of the spots. Since the eigenvalue problem is linear and homogeneous, we can write ψ = cj ψ˜j (ρ), ¯j (ρ), N = cj N so that (2.21) becomes 1 ˜j = 0, ψ˜j + ψ˜j − (1 + λ)ψ˜j + 2f uj vj ψ˜j + f u2j N 0 < ρ < ∞, ρ ˜ + 1N ˜ + ψ˜j − 2uj vj ψ˜j − u2 N ˜ N 0 < ρ < ∞, j j j = 0, ρ j ˜ (0) = 0; ˜j → log ρ + B ˜j + o(1) as ρ → ∞. ψ˜ (0) = N ψ˜j → 0 and N j j (2.22) 23 ˜j = B ˜j (Sj , f, λ) must be computed numerically From this system, the term B by solving the complex-valued BVP. Numerically solving it is not as straightforward as in the uncoupled m = 0 case; this was done carefully in § 2.2.4. ˜j does not tend to zero at infinity, the local eigenvalue In addition, since N problems near each spot will all be coupled together. This is in contrast to the m = 0 case for non-radially symmetric perturbations studied earlier, which are largely local instabilities. ˜j equation can be written as We observe that the N ˜j ρN ˜j = 0 . + ρ ψ˜j (1 − 2uj vj ) − u2j N ˜j ∼ log ρ as ρ → ∞, to get We can integrate this equation from 0 < ρ < ∞ with N ∞ 1= 0 ˜j ρdρ. ψ˜j (1 − 2uj vj ) − u2j N ˜j = Nj /cj , we obtain the identity Then, by using the fact that ψ˜j = ψj /cj , N ∞ cj = 0 ψj (1 − 2uj vj ) − u2j Nj ρdρ. (2.23) Next, we derive an equation in the outer region for η. The far-field condition of the inner solution near the j-th spot is needed. We had previously defined ηj = ˜j /D. As ρ → ∞ we have N ˜j ∼ log ρ + B ˜j + o(1), with ρ = (y 2 + Nj /D = cj N 1 y22 )1/2 . Therefore, in terms of outer variables we obtain the matching condition that ηj ∼ cj 1 ˜j , log |x − xj | + + B D ν whereν = −1/ log ε . 24 as x → xj , (2.24) Now, from (2.18) we obtain in the outer region from the ψ equation that −ψout + 2f ue ve ψout + f u2e ηout = λout ψout . Since, in the outer region, we have ue = O(ε 2 ), ve = O(1), and ηout = O(1), we obtain that ψout = O(ε 4 ) 1. From the η equation in (2.18) we obtain that the outer approximation for η satisfies D∆s η + 1 ψ + 2ue ve ψ − u2e η = τ λη. ε2 (2.25) We then must estimate the coefficient of ε −2 in (2.25). In the outer region we use ue = O(ε 2 ), ve = O(1), ηout = O(1), and ψout = O(ε 4 ) 1, to estimate ε −2 (−u2e ηout ) = O(ε 2 η) and ε −2 (ψ + 2ue ve ψ) = ε −2 O(ε 4 ) − O(ε 2 ve ψout ) O(ε 2 ) + O(ε 4 ). Therefore, both of these terms are negligible in the outer region. The estimate above shows that we need only consider the local contributions near each spot. The global contribution from the non-trivial background state of ue is negligible with regards to their contribution on the outer solution for η. To calculate the contribution from the local terms near the j-th spot we use (2.23) to obtain 1 ψ + 2ue ve ψ − u2e η → 2π ε2 N j=1 ∞ 0 ψj (1 − 2uj vj ) − u2j ηj ρdρδ (x − xj ) N = −2π j=1 cj δ (x − xj ). In this way, we obtain from (2.24) and (2.25) that the outer solution for η 25 = satisfies τλ ∆s η − η = 2π D cj η∼ D N j=1 cj δ (x − xj ), 1 ˜j , log |x − xj | + + B ν (2.26) as x → xj , for j = 1, . . . , N. Next, we will represent the solution to this problem in terms of a Green’s function, and from it we will obtain a homogeneous linear algebraic system of the form Mc = 0, where M = M(S1 , . . . , SN , λ, f, D). The condition for the existence of a non-trivial solution for c = (c1 , . . . , cN )T is that det M = 0. This condition effectively leads to a transcendental equation for any discrete eigenvalues λ. Before deriving this system, we obtain a key identity that is helpful for obtaining stability thresholds. By differentiating the core problem (2.6) with respect to Sj we get 1 (∂sj uj ) + (∂sj uj ) − (∂sj uj ) + 2f uj vj (∂sj uj ) + f u2j (∂sj vj ) = 0, ρ 1 (∂sj vj ) + (∂sj vj ) + (∂sj uj ) − 2uj vj (∂sj uj ) + u2j (∂sj vj ) = 0, ρ vj ∼ log ρ + ∂sj χ(Sj ; f ), (2.27) as ρ → ∞. By comparing this problem with (2.22) we obtain the key identity ˜j (Sj , f, λ = 0) = ∂s χ(Sj ; f ) ≡ χ (Sj ; f ). B j (2.28) In solving (2.26) we will identify two separate two cases. The case τ = 0, 26 where the solution will depend on the Neumann Green’s function for a sphere, and the case where τ = 0, for which we will identify a new Green’s function on a sphere that is related to Legendre functions. 2.2.3 Case IIA: The competition case, m = 0 and τ = 0 For this case, (2.26) reduces to N ∆s η = 2π j=1 cj η∼ D cj δ (x − xj ), (2.29) 1 ˜j , log |x − xj | + + B ν as x → xj , for j = 1, . . . , N. In terms of this Neumann Green’s function of (2.10) and (2.11), the solution to (2.29) is η=− 2π D N cj G(x; xj ) + j=1 η¯ , D (2.30) where η¯ is a constant to be found. We then calculate that 2π ∆s η = − D N j=1 2π cj ∆s G(x; xj ) = − D 1 =− 2D N cj j=1 N j=1 1 − δ (x − xj ) 4π 2π cj + D N j=1 cj δ (x − xj ). Upon comparing this with (2.29), we require that the following solvability condition be satisfied: N cj = 0. j=1 27 (2.31) Next, by letting x → xi in (2.30) and comparing it with (2.29), we obtain η=− ∼ ci D 2π ci − log |x − xi | + ci R + D 2π log |x − xi | + cj Gij + j=i η¯ D 1 ˜i . +B ν This yields a linear system of the form N −2πν ci R + ˆi . cj Gij + ν η¯ = ci 1 + η B j=i 1 Lij + R, with Lij = log |xi − xj |, to write this linear We then use Gij = − 2π system as −2πν ci R + N j=i cj R − 1 2π N ˆi . cj Lij + ν η¯ = ci 1 + η B j=i N j=1 cj Then, by using the solvability condition = 0, we can simplify the system above to an N + 1 dimensional system N ν N cj Lij + ν η¯ = ci ˆi , 1 + ηB i = 1, . . . , N ; cj = 0, j=1 j=i for the N + 1 unknowns η¯ and cj for j = 1, . . . , N . 28 (2.32) To rewrite (2.32) in matrix form, we define ˜1 B 0 1 . . , .. . B= e= . , ˜N 0 B 1 0 L12 · · · L1N .. .. L . . 21 G= . , .. LN 1 · · · 0 c1 . . c= . , cN (2.33) so that the matrix formulation of (2.32) is νGc + ν η¯e = c + νBc, eT c = 0. (2.34) Pre-multiplying by eT allows us to solve for η¯ as η¯ = 1 T e Bc − eT Gc . N Upon substituting η¯ back into (2.34) we get νGc + ν e eT Bc − eT Gc = c + νBc. N Finally, we define the matrix E0 as 1 ··· 1 T 1 .. E0 ≡ ee = N N . 1 ··· 1 .. . , 1 and then re-arrange the system above to get that c is an eigenvector of the matrix problem Mc = 0 , M≡ 1 I + (I − E0 )(B − G) ν . In summary, we conclude that c = (c1 , . . . , cN )T are eigenvectors for the homo29 ˜j (λ). We refer geneous linear system Mc = 0, where M = M(λ) since Bjj = B to this system as the globally coupled eigenvalue problem. The condition det M = 0, which yields a transcendental equation for λ, deter- mines the discrete eigenvalues for the case τ = 0. In general, solving det M = 0 ˜j = B ˜j (λ, Sj , f ) as defined in (2.22). Recall that requires the determination of B the source strengths S1 , . . . , SN terms must be computed from the nonlinear system given by 2 S + ν(I − E0 )GS + ν(I − E0 )χ = √ e. DN (2.35) Threshold calculation Although it is difficult to compute the discrete eigenvalues of the globally coupled eigenvalue problem, it is a relatively simple matter to calculate the stability threshold for eigenvalues entering Re(λ) > 0 through the origin λ = 0 by using the identity of (2.28). Near the end of section § 2.2.3 we prove that eigenvalues that cross into Re(λ) > 0 have no imaginary components when Re(λ) = 0. Recall that when λ = 0 we have ˜j (0, Sj , f ) = χ (Sj ; f ). B ˜j (0, Sj , f ) from (2.22). Therefore, when λ = 0 we do not need to compute B It is simply provided by the core problem through the numerical estimation of χ (Sj ; f ). Therefore, when λ = 0, the matrix B is known, and we need to solve det M = 0, with M ≡ 1 I + (I − E0 )(B − G), ν ν=− 1 , log ε together with (2.35). We now calculate this threshold explicitly for the special case where all the 30 spots have a common source strength, i.e., Sc ≡ S1 = · · · = SN . This will occur whenever e = (1, . . . , 1)T is an eigenvector of G, so that Ge = κ1 e, (2.36) for some κ1 . In particular, such a cyclic matrix structure will always occur for a two spot pattern, since G= 0 L12 L21 0 , and L12 = L21 . A cyclic matrix structure also occurs for other symmetric arrangements of spots, such as N equally-spaced spots lying on a ring of constant latitude on the sphere. In addition, in general the cyclic structure of the Green’s matrix will also hold when the spot locations are at the elliptic Fekete points, i.e. at the spot configuration that minimizes N − N i=1 j=1 j=i log |xi − xj | . This minimization problem is also called the Thomson problem [58], which consists of finding the optimal distribution of N equally charged particles on the surface of a sphere Assuming that (2.36) holds, we then have that S = Sc e, and from (2.35) we have that Sc = √2 . DN Similarly, for M we calculate M= 1 I + χ (Sc )(I − E0 ) − (I − E0 )G. ν We want to find conditions that guarantee that M will be a singular matrix. 31 Lemma 2.1 Consider the eigenvalue problem G b = k b, for k1 , . . . , kN , and eigenvectors b1 , . . . , bN . Assume that Ge = k1 e, so that b1 = e. Then we must have G bj = kj bj , j = 2, . . . , N, with bTj e = 0, for j = 2, . . . , N. Proof Since G is a symmetric matrix by Green’s reciprocity theorem, then the eigenspace must be orthogonal. We do need the dimension of the nullspace of (G − k1 I) to be equal to one, so that k1 is a simple eigenvalue of G. Claim 1 bj , for j = 2, . . . , N , and e are eigenvectors of M. Proof We have that Mbj = 1 bj + χ (Sc )(I − E0 )bj − (I − E0 )G bj , ν with G bj = kj bj and E0 bj = 1 T N ee bj for j = 2, . . . , N, = 0, for j = 2, . . . , N . This shows that Mbj = 1 bj + χ (Sc )bj − kj bj = ν 1 + χ (Sc ) − kj bj . ν In addition, we have that (I − E0 )e = 0, and Ge = k1 e. Therefore, the other eigenvector of M is Me = 1 1 e + χ (Sc )(I − E0 )e − (I − E0 )Ge = e. ν ν Therefore we can conclude that b2 , . . . , bN and e are the eigenvectors of M. 32 Requiring that det M = 0 then yields N − 1 algebraic equations of the form 1 + χ (Sc ) − kj = 0, ν for j = 2, . . . , N, (2.37) where kj for j = 2, . . . , N are any of the eigenvalues of G corresponding to the N − 1 dimensional subspace perpendicular to e. Stability threshold Suppose that Ge = k1 e, which implies a condition on the spot configuration x1 , . . . , xN when N > 2. The cyclic structure always holds for N = 2. Then there exists a solution to (2.35) with a common source strength Sc , given by Sc = √ 2 . DN From Figure 2.2 we know that self-replication occurs if Sc∗ > Σ2 (f ), with Σ2 the critical value where λ = 0 for the m = 2 mode. Alternatively, we know that there exists an instability with a sign-fluctuating eigenfunction whenever 1 + χ (Sc ) − kj = 0, ν for some j in j = 2, . . . , N. We now make some remarks. The eigenfunction is sign-fluctuating because c = b is one of the vectors for which bT e = 0. Recall ψ = cj ψˆj (ρ), c = (c1 , . . . , cN )T , and that ψ is a perturbation of the quasi-equilibrium solution for u. We also note that kj only depends on the number of spots N and their locations. From Figure 2.1, we know that χ (Sc ) is a negative concave down curve when Sc is small enough, with χ (Sc ) → ∞ as Sc → 0+ . We define σj = with ν = − log1 ε . For ν 1 we have σj ∼ hand, as D → ∞, we have that Sc → 0+ , 1 ν 1 ν + χ (Sc ) − jj , > 0 (to leading order). On the other and thus χ (Sc ) → −∞. Therefore, for D small enough we will eventually have σj = 0, with kj = max2≤j≤N {kj }. 33 Principal Result 2.2.1 As D is increased, i.e. as Sc is decreased, the smallest value of D for which λ = 0 is given by the root of the transcendental equation 1 − kj = −χ (Sc ), ν Sc = √ 2 , DN (2.38) T where kj = max2≤j≤N kj , and G bj = kj bj , with bj e = 0. This criterion defines a threshold Sc∗ , and since χ depends on f and N , then Sc∗ = Sc∗ (f, N ). In Figure 2.4 we compute the competition instability threshold for 3.5 6 3 5 2.5 f = 0.3 4 S Sc 2 c 3 1.5 2 1 1 0.5 f = 0.9 0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 0.2 1 f 0.3 0.4 0.5 0.6 |xi − xj| 0.7 0.8 0.9 1 Figure 2.4: Competition instability threshold, i.e., stability with respect to the m = 0 mode when τ = 0. The figure on the left is for two spots located at opposite poles of the sphere, and the figure on the right computes the threshold as a function of the distance between the spots (for fixed values of f ). two spots with common spot-strength Sc . The figure on the left considers the case of two spots located at opposite ends on a sphere, as a function of f . Since Sc = √ 2/ DN the profile of the curve f vs Dc would be monotonically increasing. If we are below any of the curves in these figures we would predict that a competition instability will occur. The figure on the right computes the competition instability threshold for two spots at varying distances from each other, for fixed values of f . The result shows that the critical spot-strength Sc increases the closer the spots are to each other, and 34 equivalently Dc will decrease as they approach each other. The key quantity determining the competition instability threshold is the largest eigenvalue of G in the subspace perpendicular to e. For equally spaced spots, the matrix G has exactly two eigenvalues which can be calculated analytically. For the more general case of N spots, this must be done numerically. We summarize our stability results so far as follows: Suppose that the spot locations are such that Ge = k1 e, so that they have a common source strength Sc . Then, • If Sc > Σ2 (f ), a peanut-splitting linear instability occurs, which numerically is shown to lead to spot self-replication. • If Sc = Sc∗ (f ; N ) then λ = 0, and in fact we predict that if Sc < Sc∗ (f ; N ), which corresponds to D too large, then there will exist a real positive eigenvalue with sign-fluctuating eigenfunction. The interaction between these two thresholds could lead to the following scenario: Consider an initial condition of N homogeneously spaced (equally spaced if N ≤ 4, or in a Fekete distribution for N > 4) spots on the surface of a sphere, and suppose that D is such that Sc∗ < Σ2 (f ). Suppose that there are N initial spots with a common source strength and that Sc exceeds the spot self-replication threshold Σ2 . Then, we predict that the N spots will split, self-replicating into 2N spots. Suppose that now, for the 2N spots, their value of Sc = Sc < Sc∗ (f ; N ). √ 2 D(2N ) now satisfies Then a sign-fluctuating instability will occur, which will annihi- late some of the 2N spots, bringing the system back to a state for which the new Sc value exceeds the self-replication value. This scenario may be too simplistic in that we are neglecting any motion of the 35 spots. More specifically, we cannot guarantee that the spots will remain equally spaced for all time. Thus, the Green’s matrix equally may not be cyclic for all time. There is also the possibility that after the self-replication event the sign-fluctuating instability annihilates a different number of spots. Figure 2.5: Full numerical simulation of the Brusselator model on a sphere (u(x)). A competition instability can be observed for the lower row (D = 1.2), which is not triggered in the top row (D = 0.8). Both scenarios have f = 0.7, ε = 0.1. In the full numerical computations of the Brusselator exhibited in Figure 2.5 we show a competition instability that occurs for a two-spot pattern when the diffusivity D is large enough. The other parameter values are given in the caption of this figure. As mentioned earlier, the analytical estimate above does not account for the motion of the spots, and this complicates the determination of a threshold. For the case τ = 0 that we have been studying, we now would like to prove for the common source-strength case Sc = S1 = S2 = · · · = SN that it is impossible for eigenvalues to enter the half-plane Re(λ) > 0 except from λ = 0. In other words, we want to rule out Hopf bifurcations. 36 ˜j (λ, Sj , f ) = B ˜c (λ, Sc , f ) for all j = 1, . . . , N . For this case, we label B ˜c (λ, sc , f )(I − E0 ) − (I − E0 )G. Recall that when Ge = k1 e, Then M = 1 + B ν then Me = ν1 e. Alternatively, for G b = kj bj , with btj e = 0, then Mbj = 1 ˜c (λ, Sc , f ) − kj bj . +B ν Thus, the eigenvalues λ are the roots of the transcendental equations 1 ˜c (λ, Sc , f ) − kj = 0, +B ν j = 2, . . . , N. (2.39) We aim to show that it is impossible for λ = iλI , with λI > 0, to be a solution to (2.39) for any j = 2, . . . , N . This would prove that eigenvalues can only enter the unstable right-half plane along the real axis. Letting λ = iλI , and separating (2.39) into real and imaginary parts we conclude that any such Hopf bifurcation must satisfy ˜c (iλI , Sc , f )] = 0, Im[B ˜c (iλI , Sc , f )] = − 1 + kj . Re[B ν (2.40) ˜c (iλI , Sc , f ) is computed from (2.22), which we write as Recall that B ˜ ˜ = iλi ψ, Lρ ψ˜ − ψ˜ + 2f uc vc ψ˜ + f u2c N ˜ − 2uc vc ) − u2 N ˜ + ψ(1 ˜ = 0, Lρ N c ψ˜ → 0, as ρ → ∞, (2.41) ˜ ∼ log ρ + B ˜c + o(1) as ρ → ∞, N with Lρ = ∂ρρ + ρ1 ∂ρ , and (uc , vc ) the solutions to the core problem. ˜ in terms of their real and imaginary components If we now separate ψ˜ and N as ψ˜ = ψ˜R + iψ˜I , ˜ =N ˜R + iN ˜I , N 37 then from (2.41) we get ˜R = −λI ψ˜I , Lρ ψ˜R − ψ˜R + 2f uc vc ψ˜R + f u2c N ψ˜R → 0, as ρ → ∞, ˜R + ψ˜R (1 − 2uc vc ) = u2 N ˜ Lρ N c R, ˜R → log ρ + Re(B ˜c )+o(1), as ρ → ∞, N ˜I = −λI ψ˜R , Lρ ψ˜I − ψ˜I + 2f uc vc ψ˜I + f u2c N ˜I + ψ˜I (1 − 2uc vc ) = u2 N ˜ Lρ N c I, ψ˜I → 0, as ρ → ∞, ˜I → Im(B ˜c )+o(1), as ρ → ∞. N ˜I = u2c N ˜I − ψ˜I (1 − 2uc vc ) for which N ˜I → 0 as There is no solution to Lρ N ˜I is bounded at infinity. In fact, in solving ρ → ∞. The best one can say is that N 1 u + u = f (ρ), ρ then u ∼ only if ∞ 0 yf (y)dy ∞ 0 yf (y)dy 0 < ρ < ∞, u (0) = 0, log ρ + o(1) at infinity. The logarithmic term vanishes = 0. Therefore, we conclude that ˜c (iλI , Sc , f )] = 0, Im[B which contradicts the first equation in (2.40). We conclude that a Hopf bifurcation is impossible for the case τ = 0 when the Green’s matrix is cyclic. Instabilities for this case can only be triggered by eigenvalues crossing into Re(λ) > 0 along the real axis in the λ−plane. 38 Case IIB: The Hopf bifurcation case, m = 0 and τ > 0 2.2.4 For the case τ > 0, we will show that Hopf bifurcations in the spot amplitudes are possible. For this case, we return to (2.26), which we write as τλ ∆S η − η = 2π D η∼ N j=1 cj δ(x − xj ), cj 1 ˜j log |x − xj | + + B D ν as x → xj , j = 1, . . . , N. We define Gλ (x; x0 ) to be the unique solution to ∆ S Gλ − τλ Gλ = −δ(x − x0 ), D 1 Gλ (x; x0 ) ∼ − log |x − x0 |+Rλ + o(1), 2π (2.42) x → x0 . Here Gλ is 2π periodic in φ, and smooth in θ = 0, π. An explicit formula for Gλ can be written in terms of Legendre functions. We notice by symmetry that Rλ is independent of x0 . The solution for η is η=− 2π D N cj Gλ (x; xj ). j=1 Now, as x → xi , we obtain for each i = 1, . . . , N that the matching condition 2π ci − log |x − xi | + ci Rλ + D 2π N j=i cj Gλij ∼ ci 1 ˜i , log |x − xi | + + B D ν must hold. By matching the O(1) terms, we conclude that N ˜i ci = 0 . cj Gλij + ν B ci + 2πν ci Rλ + j=i 39 i = 1, . . . , N. (2.43) For this case it is no longer holds that we can write for any x on sphere that Gλ (x; xj ) = − 1 log |x − xj | + constant, 2π where the constant is independent of x. In fact, there must be terms of the form |x − x0 |2 log |x − x0 |, etc. As such, we can only write (2.43) as c1 . . c= . , cN ˜ = 0, Ic+2πνGλ c+ν Bc Rλ Gλ = Gλij Gλij .. , (2.44) . Rλ where the Gλ matrix has common entries along its diagonal that are independent of the spot locations. In addition, ˜1 B 0 .. ˜= B . ˜N B 0 . Thus we conclude that λ must be such that the matrix problem ˜ = 0, Mc ˜ ≡ I + 2πνGλ + ν B ˜ M (2.45) ˜ = 0. has a non-trivial solution, i.e., that det M Next, we consider the case where the spots have a common source strength Sc , with Sc = S1 = · · · = SN . Then, ˜=B ˜c (λ; Sc , f )I, B so that ˜ = I + 2πνGλ + ν B ˜c I. M 40 In terms of the matrix spectrum of Gλ Gλ bjλ = kjλ bjλ for j = 1, . . . , N, it follows that ˜ bjλ = 1 + 2πνkjλ + ν B ˜c bjλ . M ˜ = 0 when We conclude that det M ˜c = 0, 1 + 2πνkjλ + ν B j = 1, . . . , N. (2.46) We remark that as τ increases it is impossible for eigenvalues to enter Re(λ) > 0 along the real axis by crossing through λ = 0. This is because Gλ depends only on the product τ λ, which vanishes when λ = 0. Secondly, in (2.46) we recall that ˜c (λ; Sc , f ) is obtained from the solution to B ˜ ˜ = λψ, Lρ ψ˜ − ψ˜ + 2f uc vc ψ˜ + f u2c N ψ˜ → 0, ˜ − 2uc vc ) = u2 N ˜ + ψ(1 ˜ Lρ N c , (2.47) ˜ ∼ log ρ + B ˜c (λ, Sc , f ) + o(1) as ρ → ∞. N ˜ are complex-valued when λ = λR + iλI is complex. We remark that ψ˜ and N However, given a complex-valued λ we can readily solve this BVP by separating ˜c ] and Im[B ˜c ]. the solution into real and imaginary parts and then identifying Re[B Now, we look for a Hopf bifurcation that occurs at some value τ = τHj , for j = 1, . . . , N . The stability threshold τH is defined by τh = min{τHj }. j We predict that the N-spot solution is stable to an oscillatory profile instability if 0 < τ < τH . Thus, we define λ = iλIj and τ = τHj , and we need to compute λIj 41 and τHj for which ˜ I , Sc , f )] = 0, ν −1 + 2πRe[kjλ ] + Re[B(iλ ˜ I , Sc , f )] = 0. 2πIm[kjλ ] + Im[B(iλ (2.48) In order to compute the numerical solution to (2.48), for fixed locations x1 , . . . , xN for which we have a common source strength Sc , the following steps were taken: (i) We need to calculate the Green’s function Gλ (x; xj ) that satisfies ∆s G λ − Gλ ∼ − iλIj τHj Gλ = −δ(x − xj ), D 1 log |x − xj | + Rλ , 2π as x → xj . For this we use the result in the appendix of [11], that connects this equation to the Legendre function of first kind of complex order σ, Pσ (z). We have that the solution to ∆s Gh + σ(σ + 1)Gh = −δ(x − x0 ), (2.49) with x on the sphere and where Gh is 2π periodic and smooth at φ = 0, 2π is given by Gh (x; x0 ) = − 1 Pσ (−x · x0 ), 4 sin(πσ) when σ not an integer. As z → −1+ , we have that (see the appendix of [11]) Pσ (z) ∼ sin(πσ log π 1+z 2 + 2γ + 2Φ(σ + 1) + π cot(πσ) . Here γ is Euler’s constant, Φ(z) is the Psi or digamma function Φ(z) = Γ (z)/Γ(z), with Γ(z) the gamma function. Numerically, we evaluated the complex digamma function terms on Maple 42 and the rest of the routine was done in Matlab. Now, as x → x0 , we have Gh (x; x0 ) ∼ − 1 log |x − x0 | + Rh , 2π where Rh ≡ − 1 [−2 log 2 + 2γ + 2Φ(σ + 1) + π cot(πσ)] . 4π Therefore, in order to obtain Gλ (x; xj ) and Rλ , we simply need to set σ(σ + 1) = − iλI τ , D and solve for σ. We obtain 1 σ=− + 2 1 iλI τ − , 4 D and we must choose the principal branch Re 1/4 − iλI τ /D > 0 on account of the requirement that σ = 0 when τ = 0. (ii) Once we have identified Gλ (xi ; xj ) for i = j and Rλ , we build the complexvalued matrix Rλ Gλ = Gλij Gλij .. . Rλ and calculate the complex-valued eigenvalues kλj for j = 1, . . . , N of this matrix. We remark that Gλ is symmetric, but not Hermitian when λ = iλI , since Gλij = Gλji . Thus, the kλj eigenvalues are in general complex-valued. ˜ I , Sc , f ) from the (iii) We next calculate the real and imaginary parts of B(iλ complex BVP (2.47). (iv) For each eigenvalue kλj we do a 2×2 matrix Newton update on the nonlinear system from (2.46). 43 Figure 2.6 shows the results of the full numerical simulation. We used an initial approximation obtained by using D = 100 (as outlined in the procedure described below), and we did a continuation on D to obtain the Hopf bifurcation threshold for smaller values of D. 2 7 10 6 1 10 5 τc4 τc 0 10 3 f = 0.8 2 −1 10 f = 0.7 1 f = 0.6 −2 0 0 20 40 60 80 100 10 120 0.2 Dc 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 f Figure 2.6: The Hopf bifurcation threshold for varying D (left), and as a function of f when D = 100. An important remark is that the Newton update is on the two variables (τ, λI ), and without a good starting guess it is very hard to converge to a solution. However, it is possible to find a good initial guess when considering D If we let ε = λI τ D , with ε 1. 1, we need to solve ∆s G − iεG = −δ(x − x0 ). We do an asymptotic expansion G ∼ G0 ε constant, and that G1 satisfies + G1 + · · · , and we get that G0 is a ∆s G1 − iG0 = −δ(x − x0 ). By integrating over the sphere, and using the explicit formula for the Green’s func- 44 tion in (2.11), we have G0 = − i , 4π G1 = − 1 log |x − x0 | + R1 , 2π R1 = 1 (2 log 2 − 1). 4π For the one spot case, the eigenvalue of the Green’s matrix will simply be the regular part, and therefore we have the approximation kλ = G0 /ε+R1 . This effectively decouples (2.48), and we can find the other variable by a bisection algorithm. The approximation also serves as an independent check with the full solution of the Green’s function in (2.49). With the initial condition variables, the difference between the asymptotic approximation to G and the full Legendre solution is of less than three decimal digits in both the real and imaginary parts. In particular, we used Re(kλ ) = 1 4π (2 log 2 − 1) 0.03074, and solved the first equation in (2.48). At each iteration we had to solve the complex BVP version of (2.47) on both sides of the bisection bracket. After a few iterations we found an initial estimate (λ = 0.8489380, τ = 1.621877), which we finally used as the seed to the Newton algorithm. Finding the threshold for smaller values of D is simply a matter of performing numerical continuation. In the N -spot case one of the eigenvectors will correspond to e since Gλ e = kλ1 e should hold. This eigenvalue corresponds to a synchronous oscillatory instability of the spot amplitudes. 45 2.3 Leading-order theory In the previous section, an asymptotic method based on accounting for all logarithmic terms in ν was developed in order to construct quasi-equilibrium spot patterns and to analyze their stability. However, the implementation of this theory required some numerical computations, and so the overall approach can be considered a hybrid analytical-numerical theory. In this section, we will formulate a leading-order-in-ν theory for the existence and stability of an N -spot quasi-equilibrium solution for the regime where D= D0 , ν ν=− 1 , log ε and D0 = O(1). With this approach we will obtain explicit analytical results for the profile of each spot and for the competition instability threshold. For this range of D, the equilibrium problem is ε2 ∆s u + ε2 E − u + f u2 v = 0, ν D0 ∆s v + 2 u − u2 v = 0. ε (2.50) We now construct a leading-order quasi-equilibrium solution using the method of matched asymptotic expansions. We begin with the inner problem. 2.3.1 Inner problem ˆ We let φˆ = ε−1 (φ − φj ) and θˆ = ε−1 (θ − θj ), where y1 = sin θj φˆ and y2 = θ. Then, with ρ = (y12 + y22 )1/2 , we obtain the following local problem on ρ > 0: ∆ρ Uj − Uj + f Uj2 Vj = 0, ∆ρ Vj + ν (Uj − Uj2 Vj ) = 0. D0 46 1 ∆ρ ≡ ∂ρρ + ∂ρ , ρ Now, for ν 1, we expand Uj = Uj0 + νUj1 + . . . , Vj = Vj0 + νVj1 + . . . We obtain that Vj0 is a constant, and that Uj0 satisfies 2 ∆ρ Uj0 − Uj0 + f Vj0 Uj0 = 0, so that Uj0 = 1 2 w(ρ), f Vj0 where w on ρ > 0 satisfies ∆ρ w − w + w2 = 0, w(0) > 0, ρ > 0, w → 0 as ρ → ∞. w (0) = 0, (2.51) Here w = w(ρ) is called the “ground-state” solution. Figure 2.7 shows the numerical solution for w(ρ), as computed with Matlab’s BVP5C routine. 2.5 2 1.5 w(ρ) 1 0.5 0 0 1 2 3 4 5 6 7 8 9 10 ρ Figure 2.7: Numerical solution to the ground-state BVP ∆ρ w − w + w2 = 0. 47 At next order the equation for Vj1 is ∆ρ Vj1 = 1 1 (U 2 Vj0 − Uj0 ) = D0 j0 D0 1 1 w2 − w . f 2 Vj0 f Vj0 Therefore, Vj1 ∼ Aj1 log |y| as |y| → ∞, where Aj1 ≡ 1 D0 f Vj0 1 f ∞ 0 ∞ ρw2 dρ − However, from integrating (2.51) we see that ρwdρ . 0 ∞ 0 ρwdρ = ∞ 2 0 ρw dρ. Thus, we conclude that Uj ∼ Uj0 = 1 w(ρ), f Vj0 with ∆ρ Vj1 = Vj ∼ Vj0 + νVj1 + . . . , 1 D0 f Vj0 1 2 w −w f , and Vj1 ∼ Aj1 log ρ + o(1) as ρ → ∞, where Aj1 = b D0 f Vj0 1 −1 f ∞ , b= ρw2 dρ . 0 Thus, if we set ρ = ε−1 |x − xj |, we obtain the matching condition Vj ∼ Vj0 + Aj1 ν log |x − xj | + Aj1 as x → xj . We remark that since ∆ρ Vj1 = g(ρ) for some known function g(ρ), and Vj1 = Vj1 (ρ) with Vj1 (0) = 0, we can impose that Vj1 − ∞ 0 zg(z)dz log ρ → 0 as ρ → ∞. This follows because we can always add a constant to the solution for Vj1 as it is a linear differential equation. 48 2.3.2 Outer expansion In the outer region we obtain u ∼ ε2 E, and so for |x − xj | global contribution ν (u − u2 v) ε2 O(ε) we get the νE + O(ε2 υ). In contrast, the local contribution from the x = xj region is ν (u − u2 v) ∼ 2πν ε2 ∞ 0 2 (Uj0 − Uj0 Vj0 )ρdρ δ(x − xj ), ∞ ∼ 2πν 2πν ∼ f Vj0 0 1 1 w− 2 w2 ρdρ δ(x − xj ), f Vj0 f Vj0 ∞ 1 ρw2 dρ δ(x − xj ). 1− f 0 Thus, in terms of Aj1 computed previously, the local contribution from the j-th spot is ν (u − u2 v) ∼ −2πD0 νAj1 δ(x − xj ), ε2 where Aj1 = 1 D0 f Vj0 ∞ 1 −1 f ρw2 dρ . 0 By combining the local and global contributions, we obtain the outer problem ∆s v + vE = 2πν D0 v ∼ Vj0 + Aj1 + νAj1 log |x − xj | N j=1 Aj1 δ(x − xj ), as x → xj , (2.52) j = 1, . . . , N . Our goal now is to determine a nonlinear algebraic system for Vj0 for j = 1, . . . , N . Upon integrating (2.52) over the sphere, we obtain the solvability condition E 1 = D0 2 N Aj1 . j=1 49 In terms of an as yet unknown constant v, the solution to (2.52) is N v = −2πν Aj1 G(x; xj ) + v, (2.53) j=1 where G is the Neumann Green’s function for the unit sphere. Now, equating the leading terms in v as x → xi with the non-singular terms in (2.52) as x → xi , we obtain that v ∼ v + O(ν) = Vj0 + Aj1 + O(ν), i = 1, . . . , N. Thus, our leading-order-in-ν theory yields that v = Vj0 + Aj1 , N 1 E = D0 2 with Aj1 = B , Vj0 B≡ j = 1, . . . , N, Aj1 , (2.54) j=1 1 −1 f 1 D0 f ∞ ρw2 dρ > 0. 0 Next, we solve this leading order system (2.54) for Vj0 . We have v ≡ H(Vj0 ) = Vj0 + B Vj0 , where v is a constant independent of j. A simple plot of H(Vj0 ) versus Vj0 shows that H → ∞ as Vj0 → 0+ and as √ Vj0 → +∞. Furthermore, H(Vj0 ) has a unique critical point at Vj0 = β with H > 0 for all Vj0 > 0. Therefore, by looking for intersections of H(Vj0 ) with the constant v, we conclude that there are only two possible values of Vj0 , which we label as Vj0s and Vj0b . 50 2.3.3 Case I: Vj0s = Vj0b , symmetric spot quasi-equilibrium For a collection of N spots, suppose that we take a common value for Vj0 for each spot. Then Vj0 = V0 , and we have Aj1 = V0 = D0 VN 2E B V0 , and E D0 = is the common value, and we calculate v = V0 + 1 B 2 V0 N . Hence, B V0 . Given that H(V0 ) is a concave-up curve for V0 > 0, we will be either to the left or to the right √ of the minimum at V0 = B. This construction yields to leading-order in ν an N -spot pattern with spots of equal height. For this symmetric pattern, the quasi-equilibrium solution has the form N u ∼ ε2 E + D0 BN , where V0 = 2E j=1 1 w ε−1 |x − xj | , f V0 1 B≡ D0 f 1 −1 f ∞ (2.55) ρw2 dρ. 0 Notice that B > 0 since 0 < f < 1. Thus, V0 is independent of D0 . 2.3.4 Case II: Asymmetric spot equilibria In the remainder of this chapter we will focus on symmetric spot patterns. However, we now briefly mention that one can also construct asymmetric patterns where the spots have different amplitudes. To see this suppose that we have M1 small amplitude spots with Vj0 = V0b for j = 1, . . . , M1 , and M2 large amplitude spots with Vj0 = V0s for j = M1 + 1, . . . , N , where M2 = N − M1 . We note that since Uj0 ∼ 1 f Vj0 w, a large value of Vj0 yields a small-amplitude spot, and vice versa. The small and large spots must be such that v = V0b + B B = V0s + . V0b V0s 51 Then, by Using Aj1 = B Vj0 from (2.54), we obtain 1 E = D0 2 N Aj1 = j=1 B 2 M1 M2 + V0b V0s . Therefore, we conclude that V0b , V0s satisfy the coupled nonlinear algebraic problem E B = D0 2 M1 (N − M1 ) + V0b V0s , V0b + B B . = V0s + V0b V0s (2.56) This yields two equations for the two unknowns V0b and V0s . Now, the system (2.56) is not solvable for all values of D0 . At the coalescence √ point where V0s = V0b = B, we have that 2E , D0ASY = √ BN B= b(1 − f ) . D0 f 2 At this particular value of D, solution branches of asymmetric quasi-equilibrium patterns bifurcate off of the symmetric solution branch. If D0 > D0ASY the sys- tem (2.56) should not be solvable. This construction indicates that there exists asymmetric spot equilibria to leading-order in ν when D∼ 2.4 D0 , ν and D0 < D0ASY . Derivation from the S-formulation In this section we show how to independently recover the leading-order results in the previous section by considering the limit as Sj → 0 of the S-formulation of §2.1. Near the j-th spot we recall from (2.5) that Uj = D1/2 uj , Vj = D−1/2 vj , 52 so that (uj , vj ) satisfy (see (2.6)) ∆ρ uj − uj + f u2j vj = 0, uj → 0 as ρ → ∞, ∆ρ vj + uj − u2j vj = 0, (2.57) vj ∼ Sj log ρ + χ(Sj ) as ρ → ∞. and where the Si for i = 1, . . . , N satisfy (2.14), written again as N Si − ν νR0 E + νχ(Si ; f ) = νv, Sj Lij + √ D j=i i = 1, . . . , N , (2.58) N Lij ≡ log |xi − xj |, 2E Si = √ , D i=1 R0 = 2 log 2 − 1. We want now to expand the system as Sj → 0 when D = D0 /ν 1. As such, we must consider the following side problem, which will determine the asymptotics of χ(Si ; f ) as Si → 0: 2.4.1 Core problem: Small S-asymptotics For S → 0, we now calculate u, v, and χ for ∆ρ u − u + f u2 v = 0, ∆ρ v + u − u2 v = 0, u → 0 as ρ → ∞, v ∼ S log ρ + χ as ρ → ∞. We write u = S u ˜, v = v˜/S, and χ = χ/S ˜ to obtain ∆ρ u ˜−u ˜ + fu ˜2 v˜ = 0, ∆ρ v˜ + S 2 (˜ u−u ˜2 v˜) = 0, u ˜ → 0 as ρ → ∞, v˜ ∼ S 2 log ρ + χ ˜ as ρ → ∞. This suggests that we should seek an approximate solution for S 53 1 in the form u ˜=u ˜0 + S 2 u ˜1 + S 4 u ˜2 + · · · , v˜ = v˜0 + S 2 v˜1 + S 4 v˜2 + · · · , χ ˜=χ ˜0 + S 2 χ ˜1 + S 4 χ ˜2 + . . . . Upon substituting this expansion into the problem for (˜ u, v˜) and collecting powers of S 2 we obtain the following sequence of problems: ∆ρ u ˜0 − u ˜0 + f u ˜20 v˜0 = 0, u ˜0 → 0 as v˜0 → χ ˜0 , ∆ρ v˜0 = 0, ∆ρ u ˜1 − u ˜1 + 2f u ˜0 v˜0 u ˜1 = −f u ˜20 v˜1 , ∆ρ v˜1 = −˜ u0 + u ˜20 v˜0 , ρ → ∞, as ρ → ∞ v˜1 → 0 as ρ → ∞, v˜1 → log ρ + χ ˜1 as ρ → ∞. (2.59) (2.60) In addition, we obtain that v˜2 satisfies ∆ρ v˜2 = −˜ u1 + 2˜ u0 v˜0 u ˜1 + u ˜20 v˜1 , (2.61) with v˜2 bounded as ρ → ∞. We conclude that u ˜0 = 1 w, f v˜0 v˜0 = χ ˜0 , (2.62) where w is the ground-state solution satisfying ∆ρ w − w + w2 = 0. Notice that since w → 0 at infinity, then ∞ 0 ρwdρ = solvability condition from the v˜1 equation yields that ∞ lim ρ˜ v1ρ = ρ→∞ 0 −˜ u0 + u ˜20 v˜0 ρdρ, 54 ∞ 2 0 ρw dρ. Next, the with v˜1ρ = 1/ρ as ρ → ∞. Thus, from (2.62) we obtain that 1=− Then, since ∞ 0 ρwdρ ∞ 1 f v˜0 ρwdρ + 0 ∞ 2 0 ρw dρ, = ρw2 dρ. 0 we conclude that ∞ b(1 − f ) , f2 v˜0 = χ ˜0 = ∞ 1 2 f v˜0 ρw2 dρ ≈ 4.9343. with b = 0 (2.63) The numerical value for b was obtained from the numerical computation of the ground state shown in Figure 2.7. With the leading-order terms calculated, the problem for the second-order terms u ˜1 , v˜1 can then be written as ∆ρ u ˜1 − u ˜1 + 2w˜ u1 = − w2 w − f2 f 1 ∆ρ v˜1 = v˜0 1 2 w v˜1 , f v˜02 , u ˜1 → 0 as ρ → ∞, v˜1 ∼ log ρ + χ ˜1 , as ρ → ∞. (2.64) Next, the solvability condition for the v˜2 equation (2.61) yields that ∞ 0 Upon using u ˜0 = ∞ (2˜ u0 v˜0 − 1)˜ u1 ρdρ + 1 f v˜0 w ∞ 0 0 u ˜20 v˜1 ρdρ = 0. this becomes 2w 1 −1 u ˜1 ρdρ + 2 2 f f v˜0 ∞ w2 v˜1 ρdρ = 0. (2.65) 0 We now integrate multiply the equation for u ˜1 in (2.64) by f −1 and integrate to obtain 1 f ∞ 0 ρ∆ρ u ˜1 dρ − 1 f ∞ ρ˜ u1 dρ + 0 1 f ∞ 0 55 2w˜ u1 ρdρ = − 1 f 2 v˜02 ∞ 0 w2 v˜1 ρdρ. Upon combining this with (2.65) we obtain ∞ 0 2w −1 u ˜1 ρdρ = f ∞ 0 ∞ 2w u ˜1 ρdρ − f 0 1 u ˜1 ρdρ + f ∞ 0 1 ρ∆ρ u ˜1 dρ. f Since u ˜1 (0) = 0 and u ˜1 → 0 as ρ → ∞, the last integral vanishes upon integration, and so the equation above reduces to 1 −1 f ∞ u ˜1 ρdρ = 0. 0 Since 0 < f < 1, we conclude that ∞ u ˜1 ρdρ = 0. (2.66) 0 We now show how this equation determines χ ˜1 . We return to (2.64). The solution can be written in the form v˜1 = χ ˜1 + 1 v˜1p , v˜0 f 2 where v˜1p is the unique solution on 0 < ρ < ∞ to 1 ∆ρ v˜1p = v˜1p + v˜1p = w2 − f w, ρ (2.67) v˜1p ∼ v˜0 f 2 log ρ + o(1) as ρ → ∞ ; v˜1p (0) = 0. The uniqueness of v˜1p follows from the condition that v˜1p − v˜0 f 2 log ρ → 0 as ρ → ∞. In terms of v˜1p , the equation for u ˜1 in (2.64) becomes L˜ u1 ≡ ∆ρ u ˜1 − u ˜1 + 2w˜ u1 = − χ ˜1 2 w2 w − v˜1p . f v˜02 v˜03 f 3 Now, since Lw = w2 , as seen from Lw = ∆w−w+2w2 = −w2 +2w2 = w2 , 56 we can decompose u ˜1 as u ˜1 = − χ ˜1 1 ˜1p , w − 3 3u 2 f v˜0 v˜0 f (2.68) where u ˜1p satisfies ∆ρ u ˜1p − u ˜1p − 2w˜ u1p = w2 v˜1p . (2.69) In summary, on 0 < ρ < ∞, let v˜1p and u ˜1p satisfy the coupled system ∆ρ v˜1p = w2 − f w; v˜1p ∼ v˜0 f 2 log ρ + o(1) 2 ∆˜ u1p − u ˜1p + 2w˜ u1p = w v˜1p , u ˜1p → 0 as ρ → ∞, as ρ → ∞. (2.70) Then, v˜1 and u ˜1 are given by v˜1 = χ ˜1 + 1 v˜1p , v˜0 f 2 u ˜1 = − χ ˜1 1 ˜1p . w − 3 3u 2 f v˜0 v˜0 f (2.71) Finally, we determine χ ˜1 . We substitute u ˜1 into the solvability condition (2.66) to obtain ∞ 0= 0 ρ˜ u1 dρ = − Therefore, χ ˜1 f v˜02 ∞ 0 ∞ χ ˜1 0 ρwdρ − ρwdρ = − 1 3 v˜0 f 3 ∞ ρ˜ u1p dρ. 0 ∞ 1 v˜0 f 2 ρ˜ u1p dρ. 0 However, since ∞ ∞ ρwdρ = 0 ρw2 dρ = b, v˜0 = and 0 we get χ ˜1 = − 1 2 b (1 − f ) This determines χ ˜1 in terms of u ˜1p . 57 (1 − f ) f2 ∞ ρw2 dρ, 0 ∞ ρ˜ u1p dρ. 0 (2.72) We can further simplify (2.72) by using (2.70). The goal is to obtain a formula for χ ˜1 where the parameter dependence on f appears explicitly rather than implicitly through u ˜1p . To obtain such a result, we use w2 = w − ∆w to write the problem for v˜1p as ∆ρ (˜ v1p + w) = (1 − f )w, v˜1p + w ∼ (1 − f )b log ρ + o(1) as ρ → ∞. This suggests that we can introduce v˜1Q by v˜1p + w = (1 − f )˜ v1Q , where v˜1Q satisfies the parameter-independent problem ∆ρ v˜1Q = w, ρ ≥ 0, v˜1Q (0) = 0 ; v˜1Q → b log ρ + o(1) , as ρ → ∞. Now, from the problem for u ˜1p we obtain v˜1p = −w + (1 − f )˜ v1Q , so that L˜ u1p = w2 (−w + (1 − f )˜ v1Q ). This suggests that we decompose u ˜1p as u ˜1p = −˜ u1pI + u ˜1pII (1 − f ), where L˜ u1pI = w3 , L˜ u1pII = w2 v˜1Q . Finally, from (2.72) we get χ ˜1 = 1 2 b (1 − f ) ∞ 0 ρ˜ u1pI dρ − 1 b2 ∞ ρ˜ u1pII dρ , 0 where the two integral terms are independent of f . We summarize the calculation of the asymptotics for S → 0 of the solution to the core problem as follows: 58 For S → 0, we obtain that U = D1/2 u, V = D−1/2 v, with u∼S u ˜0 + S 2 u ˜1 + . . . , v ∼ S v˜0 + S 2 v˜1 + . . . , 1 χ∼ χ ˜0 + S 2 χ ˜1 + . . . . S Here b = ∞ 2 0 ρw dρ, w2 = 0, and (2.73) w is the solution of the ground-state problem ∆ρ w − w + 1 1 χ ˜1 u1pI + (1 − f )˜ u1pII ] , w, u ˜1 = − 2 w − 3 3 [−˜ f v˜0 f v˜0 v˜0 f 1 b(1 − f ) (2.74) , v˜1 = χ ˜1 + [−w + (1 − f )˜ v1Q ] , v˜0 = 2 f v˜0 f 2 ∞ b(1 − f ) 1 1 ∞ χ ˜0 = , χ ˜ = ρ˜ u dρ − ρ˜ u1pII dρ. 1 1pI f2 b2 (1 − f ) 0 b2 0 u ˜0 = Moreover, v˜1Q is the unique solution to ∆ρ v˜1Q = w, 0 < ρ < ∞, v˜1Q (0) = 0, v˜1Q → b log ρ + o(1), as ρ → ∞, (2.75) while u ˜1pI and u ˜1pII are the unique solutions of L˜ u1pI ≡ ∆ρ u ˜1pI − u ˜1pI + 2w˜ u1pI = w3 , L˜ u1pII = w2 v˜1Q , u ˜1pII → 0 u ˜1pI → 0 as ρ → ∞, as ρ → ∞. (2.76) Remark (i) Notice that v˜1Q , u ˜1pI , u ˜1pII do not depend on any parameters (such as f ). Hence we need only compute ∞ ∞ u1pI dρ, 0 ρ˜ u1pII dρ 0 ρ˜ once in order to determine χ ˜1 . This was the motivation for introducing this decomposition from (2.72) 59 (ii) The leading order theory yields U∼ D1/2 Sf w, b(1 − f ) V ∼ D−1/2 b(1 − f ) , S f2 (2.77) and the two term expansion for χ is χ∼ 1 b(1 − f ) +χ ˜1 S + . . . S f2 (2.78) (iii) Recall now that (2.58) holds: N Si − ν νR0 E Sj Lij + √ + νχ(Si ; f ) = νv, , D j=i i = 1, . . . , N N 2E Si = √ . D i=1 N 1/2 ), i=1 Si = O(ν b(1−f ) to obtain, with f 2 Si We now assume D = D/ν where ν = −1/ log ε. Thus, which indicates that Si = O(ν 1/2 ). We then use χi ∼ Si → 0 that N νb(1 − f ) Si − O(ν) + f 2 Si 2ν 1/2 Si = √ . D0 i=1 νv, We then put Si = ν 1/2 S˜i and v = ν −1/2 v 0 , which leads to the reduced problem b(1 − f ) S˜i + = v0, f 2 S˜i N 2 S˜i = √ E. D0 i=1 (2.79) 1 b(1 − f ) , √ ˜ f2 Si D0 (2.80) Substituting into (2.77), we get U∼ w , f vi V ∼ vi ≡ 60 which holds near the i-th spot. We then solve for Si in terms of vi and substitute into (2.79) to obtain after some algebra that vi + N i=1 with U ∼ 2 f vi w b(1 − f ) D0 f 2 1 1 = √ v0, vi D0 i = 1, . . . , N, (2.81) 1 2f 2 = E, vi b(1 − f ) near the i-th spot (from (2.80)). Notice that this is precisely the same system derived in (2.54), where B= ∞ b(1 − f ) , D0 f 2 b= ρw2 dρ. 0 We conclude that we can recover the leading-order terms in the expansion in ν from the small S-asymptotics of the core problem. 2.5 Leading-order stability theory For D = D0 /ν, the time-dependent Brusselator system on the surface of the unit sphere is ut = ε2 ∆s u + ε2 E − u + f u2 v, ν ντ vt = D0 ∆s v + 2 (u − u2 v) . ε Now we linearize around the quasi-equilibrium solution by writing u = ue + eλt ψ, and v = ve + eλt η , 61 so that ε2 ∆s ψ − ψ + 2f ue ve ψ + f u2e η = λψ, ∆s η + (2.82) ν τ νλ (ψ − 2ue ve ψ − u2e η) = η. ε2 D0 D0 We will consider the regime where τ = O(1). We look for radially symmetric solutions in the inner region near the j-th spot. In this inner region we use ue ∼ 1 f vj w, ∆ρ Ψj − Ψj + 2wΨj + ve ∼ vj to obtain for ρ = |y| that 1 2 w η(xj ) = λΨj , f vj2 ρ ≥ 0. (2.83) Now, from the equation for η we have ∆y η + Assuming that τ ν −Ψj + 2ue ve Ψj + −u2e η = O D0 λτ νε2 D0 . O(ε −2 ), we use ue ve = w/f to obtain ∆y η = ν D0 −Ψj + 2 w2 wΨj + 2 2 η . f f vj (2.84) Now we expand η = ηj + νηj1 + . . ., with ηj = η(xj ). We obtain that ηj1 satisfies ∆y ηj1 = 1 D0 −Ψj + 62 2 w2 wΨj + 2 2 ηj f f vj . The far-field asymptotic behaviour of the solution is readily calculated as ηj1 ∼ 1 Aj log ρ, D0 as ρ → ∞, ∞ −Ψj + Aj = 0 ∞ Aj = 0 2 w2 wΨj + 2 2 ηj f f vj ρdρ, ∞ ηj 2 wΨj − Ψj ρdρ + 2 2 b, f f vj b= ρw2 dρ. 0 This shows that we have the far-field behaviour η ∼ ηj + Dν0 Aj log ρ as ρ → ∞. To determine the matching condition to the outer solution, we use ρ = ε−1 |x − xj |, and re-write this far-field behaviour in terms of the outer variable as η ∼ ηj + νAj Aj + log |x − xj |, D0 D0 as x → xj . (2.85) Now, in the outer region we estimate for x ≈ xj that 1 (ψ − 2ue ve ψ − u2e η) → ε2 ∞ 0 Ψj − 2 1 Ψj − 2 2 w2 ηj f f vj ρdρ δ(x − xj ). Thus, we obtain that the outer solution for η(x), valid for |x − xj | satisfies ∆s η − λτ ν 2πν η= D0 D0 N i=1 Ai δ(x − xi ), with singularity behaviour (2.85) at each xj . We now expand the solution to this problem for ν η = η + νη1 + . . . , with η = ηj + 63 O(ε), Aj , D0 1. We get j = 1, . . . , N, (2.86) where η is a constant. At next order, η1 satisfies ∆s η 1 = τ λη 2π + D0 D0 N i=1 Ai δ(x − xi ). The solvability condition for this problem, as obtained by integrating it over the sphere, is N 2π 4πτ λη + D0 D0 Ai = 0. i=1 In this way, we obtain that − ητ λ = N 1 2 Ai , η = ηj + i=1 Aj , D0 j = 1, . . . , N. (2.87) Next, we recall that ∞ Aj = 0 ηj 2 wΨj − Ψj ρdρ + 2 2 b, f f vj (2.88) so that combining it with (2.87) we get Aj 1 ηj + =− D0 2τ λ N Ai , j = 1, . . . , N. (2.89) i=1 We then write both (2.88) and (2.89) in matrix form. For (2.89) we write η+ A N = − E0 A , D0 2τ E0 = 1 T ee , N (2.90) where e = (1, . . . , 1)T . Next, for (2.88) we first define ∞ J= 0 2 wΨ − Ψ ρdρ, f Ψ1 . . Ψ= . , ΨN 64 η1 . . η= . , ηN and introduce the matrix 1/f 2 v12 H= 0 0 .. . 2 1/f 2 vN . With this notation we can rewrite (2.88) as A = J + bHη. (2.91) Then, upon combining (2.90) and (2.91), we obtain that η satisfies the matrix problem I+ N bE0 b H η=− H+ D0 2τ λ 1 N E0 J. I+ D0 2τ λ (2.92) We then write (2.83) in matrix form as ∆ρ Ψ − Ψ + 2wΨ + f w2 Hη = λΨ, (2.93) where η is given in terms of nonlocal terms via (2.92). The system (2.92)) and (2.93) is a vector nonlocal eigenvalue problem for Ψ. We will obtain an explicit nonlocal eigenvalue problem from it for the case of symmetric spot patterns where the spots have a common source strength. Recall from (2.55) that for such symmetric patterns we have vj = v ≡ D0 BN , 2 B= 1 (1 − f )b, D0 f 2 for all j, so that f 2 vj2 = f D0 BN 2 2 2 D0 f N 1 , (1 − f )b 2 D0 f 2 N 2 (1 − f )2 2 = b , for j = 1, . . . , N. 4f 2 = For this symmetric case, H = 1 I f 2 v2 (2.94) where f 2 v 2 is given in (2.94). Thus (2.92) 65 and (2.93) become I+ b Nb 1 I+ E0 η = − 2 2 D0 f v 2τ λ f 2 v 2 ∆ρ Ψ − Ψ + 2wΨ + 1 N I+ E0 J. D0 2τ λ (2.95) w2 η = λΨ. f v2 (2.96) We then write η = −AJ, where A is given by A= I+ −1 Nb b I+ E0 f 2 v 2 D0 2τ λf 2 v 2 1 1 N I+ I+ E0 , D0 D0 2τ λ Therefore, (2.96) becomes the vector Nonlocal eigenvalue problem (NLEP) ∆ρ Ψ − Ψ + 2wΨ − w2 AJ = λΨ. f v2 (2.97) Our final step in the analysis is to diagonalize this vector NLEP to obtain a scalar NLEP. To do so, we write the eigenvectors of A as Aqj = µj qj , so that in matrix form A = QΛQ−1 , | Q = q1 · · · | qN , | | µ1 0 .. Λ= 0 . . µN ˆ Letting Ψ = QΨ, we can diagonalize the vector NLEP as f w2 ˆ ˆ ˆ Q∆ρ Ψ − QΨ + 2wQΨ − 2 2 QΛQ−1 f v ∞ 0 2 ˆ ˆ ˆ wQΨ − QΨ ρdρ = λQΨ. f ˆ Multiplying by Q−1 we obtain that Ψ satisfies any one of the N scalar NLEPs f w2 ˆ ˆ ˆ ∆ρ Ψ − Ψ + 2wΨ − 2 2 Λ f v ∞ 0 66 2 ˆ ˆ ˆ wΨ − Ψ ρdρ = λΨ. f This leads to the study of the scalar NLEP ˜ −Ψ ˜ + 2wΨ ˜− ∆ρ Ψ f w2 µj f 2v2 ∞ 2 ˜ ˜ ˜ wΨ − Ψ ρdρ = λΨ, f 0 ˜ → 0, Ψ (2.98) as ρ → ∞, where µj , for j = 1, . . . , N is any eigenvalue of the matrix A defined by A= I+ Nb b I+ E0 f 2 v 2 D0 2τ λf 2 v 2 −1 1 N I+ E0 . D0 2τ λ (2.99) In contrast to the NLEP’s derived previously for other reaction-diffusion systems such as the Gierer-Meinhardt or Gray-Scott models, this NLEP involves two separate nonlocal terms. Before calculating µj we will write (2.98) in a standard form involving only one nonlocal term as ˜ −Ψ ˜ + 2wΨ ˜ − γw2 ∆ρ Ψ ∞ ˜ 0 ρw Ψdρ ∞ 2 0 ρw dρ ˜ = λΨ, 0 < ρ < ∞, (2.100) for some γ = γ(λ). This is the standard form for which Wei (e.g. [63]) has many rigorous results on the spectrum of the NLEP that are used to obtain explicit criteria for stability. To write (2.98) in standard form we define I1 and I2 by I1 = ∞ ˜ Then we can write (2.98) as I2 = 0 ρΨdρ. ˜ −Ψ ˜ + 2wΨ ˜− ∆ρ Ψ 2w2 µj w 2 µj ˜ I + I2 = λΨ. 1 f 2v2 f v2 Multiplying by ρ and integrating yields −I2 + 2I1 − 2µj b bµj I1 + 2 I2 = λI2 . 2 2 f v fv 67 ∞ ˜ 0 ρw Ψdρ and (2.101) Upon solving for I2 in terms of I1 we get I2 = ζI1 , 2(µj b − f 2 v 2 ) , µj bf − f 2 v 2 (λ + 1) ζ= so that by eliminating I1 in (2.101) we obtain 2 ˜ −Ψ ˜ + 2wΨ ˜ − 2w µj b ∆ρ Ψ f 2v2 Now, since I1 /b = f ζ 2 1− ∞ ∞ 2 ˜ 0 ρw Ψdρ/ 0 ρw dρ, I1 b ˜ = λΨ. we can now write (2.98) in the standard form (2.100), with γ= 2µj b f 2v2 1− fζ 2 , (2.102) where ζ is defined by ζ= 2(µj − f 2 v 2 /b) . µj f − (λ + 1)f 2 v 2 /b (2.103) The next step is to simplify the coefficient γ in (2.102), and calculate the eigenvalues µj . To simplify γ, we use (2.94) to define ϕ as f 2v2 N2 = 2 (1 − f )2 b, b 4f ϕ= ∞ b= ρw2 dρ . (2.104) 0 Then γ = 2µj = 2µj ϕ f (µj − f 2 v 2 /b) b 1 − , f 2v2 µj f − (λ + 1)f 2 v 2 /b ϕ(f − (λ + 1)) f − (λ + 1) = 2µj µj f − ϕ(λ + 1) µj f − ϕ(λ + 1) . Therefore, we have that the NLEP in (2.98) has multiplier γ given by γ = 2µj f − (λ + 1) µj f − ϕ(λ + 1) , 68 ϕ≡ N2 (1 − f )2 b. 4f 2 (2.105) Finally, we will calculate the eigenvalues µj of A as defined in (2.99). In terms of ϕ = f 2 v 2 /b, as given in (2.104), we can write A as A= I+ 1 N E0 I+ ϕD0 2τ λϕ −1 1 N E0 . I+ D0 2τ λ We write Aq = µq so that 1 1 N N I+ E0 q = µ I + I+ E0 q. D0 2τ λ ϕD0 2τ λϕ In this form, the eigenvectors and eigenvalues are easy to detect. It turns out that there are only two distinct eigenvalues. • Let q = e, and recall that E0 e = e(eT e)/N = e. Thus, with q = e, we have 1 N + D0 2τ λ e=µ I+ 1 N I+ ϕD0 2τ λϕ e. Thus, we obtain that for q = e, representing the synchronous mode of instability, the corresponding eigenvalue is µ= I+ 1 N D0 + 2τ λ 1 N ϕD0 I + 2τ λϕ = ϕ(N + 2τ λ/D0 ) . N + 2τ λ(ϕ + 1/D0 ) (2.106) • Let q = b, with bT e = 0. There exists N − 1 such independent vectors. These are the competition instability modes. Then 1 1 N N I+ E0 b = µ I + I+ E0 b. D0 2τ λ ϕD0 2τ λϕ However, E0 b = 0, and so D0−1 = µ (1 + 1/ϕD0 ). Therefore, with q = b and bT e = 0, then µ= ϕ . 1 + ϕD0 69 (2.107) This completes the determination of the NLEP. Next, we derive specific stability criteria from the determination of the spectrum of this NLEP. 2.5.1 Stability thresholds (I) We consider the competition instability threshold for N ≥ 2. From (2.105) and (2.107) we obtain γ= 2ϕ 1 + ϕD0 f − (λ + 1) ϕf 1+ϕD0 − ϕ(λ + 1) = 2 [f − (λ + 1)] . (2.108) f − (λ + 1)(1 + ϕD0 ) Since 0 < f < 1, it is readily seen that γ is analytic in the right half-plane Re(λ) > 0. The NLEP for this competition instability is ˜ −Ψ ˜ + 2wΨ ˜ − γw2 ∆ρ Ψ ∞ ˜ 0 ρw Ψdρ ∞ 2 0 ρw dρ ˜ = λΨ, (2.109) with γ as given in (2.108). Notice that this NLEP is not self-adjoint and that γ = γ(λ). Wei’s result (Theorem 1.4 in [63]) proves if γ is a constant, independent of λ, then Re(λ) > 0 if and only if γ < 1. Therefore, when γ is a constant the stability threshold is precisely γ = 1. However, in our case, γ = γ(λ) as given in (2.108). In Appendix B, we prove that there exists a real positive eigenvalue when γ(0) < 1. Hence, we have instability when γ(0) < 1. ˜ = w is an eigenfunction corresponding to In addition, if γ(0) = 1 then Ψ ˜ = w then λ = 0. This follows since if we set λ = 0 and Ψ ∆ρ w − w + 2w2 − γw2 ∞ 2 0 ρw dρ ∞ 2 0 ρw dρ = 0 · w. If γ(0) = 1, we use ∆ρ w −w +2w2 −γw2 = ∆w −w +w2 = 0 to establish the identity. 70 Therefore, to determine an instability threshold we set γ(0) < 1 to obtain that ϕD0 > 1 − f . Upon recalling the definition of ϕ in (2.105), we obtain our main leading-order-in-ν stability result: Principal Result 2.5.1 Let D = D0 /ν, τ = O(1). Then, to leading order in ν, an N -spot symmetric solution with N ≥ 2 is unstable to a competition instability when D0 > D0c ≡ 4f 2 , N 2 (1 − f )b ∞ b= ρw2 dρ. (2.110) 0 Remark (i) As shown below in (2.118)., this threshold agrees precisely with the leadingorder term in the stability threshold as obtained from expanding for S → 0 the results from our S-formulation involving summing logarithmic terms. (ii) Notice that as f → 1− , then D0c increases. However, as N increases, then D0c decreases. If D0 is too large, then a competition instability is triggered. This is an overcrowding type of instability. (iii) To leading-order, the stability threshold is the same for all competition modes. However, in (2.118) below we will be able to determine the correction term in this instability threshold, which yields N − 1 distinct competi- tion instability thresholds. In addition, this correction term will involve the locations of the spots on the surface of the sphere. (iv) The proof of this instability result is given in Appendix B. Although it is relatively easy to obtain an instability threshold, it is substantially more difficult to prove a stability result for the parameter range D0 < D0c . The difficulty stems from the fact that γ = γ(λ) and that one must account for complex eigenvalues. We have been unable to prove a stability result for this range of D0 . 71 (II) We now consider the synchronous instability threshold. The NLEP for synchronous instabilities is ˜ −Ψ ˜ + 2wΨ ˜ − γw2 ∆ρ Ψ ∞ ˜ 0 ρw Ψdρ ∞ 2 0 ρw dρ ˜ (0) = 0, Ψ ˜ = λΨ, ˜ →0 Ψ as ρ > 0, ρ → ∞, with γ = γ(λ) defined by f − (λ + 1) µf − ϕ(λ + 1) γ = 2µ . (2.111) Here from (2.106) µ is given by µ= ϕ(N + 2τ λ/D0 ) , N + 2τ λ(ϕ + 1/D0 ) ϕ≡ N2 (1 − f )2 b. 4f 2 Now suppose that τ = 0. Then, µ = ϕ and hence γ = 2ϕ (f − (λ + 1)) = 2. (ϕf − ϕ(λ + 1)) Hence, when τ = 0, then γ = 2 > 1 independent of λ. Wei’s result (Theorem 1.4 in [63]) then guarantees that Re(λ) < 0 and we have stability. By a perturbation argument we conclude that we must have stability when τ > 0 is sufficiently small. In addition, we notice that τ only enters through the product τ λ. If we set λ = 0, then γ = 2 > 1 independent of all the other parameters. This proves that an eigenvalue cannot enter the unstable right half-plane Re(λ) > 0 through crossing the origin λ = 0 as τ is increased. Therefore, an instability if it occurs must arise through a Hopf bifurcation as τ is increased. As shown in Appendix B, the NLEP L0 ψ˜ − γw2 ∞ ˜ 0 ρw ψdρ ∞ 2 0 ρw dρ 72 = λψ˜ , where L0 ψ˜ ≡ ∆ρ ψ˜ − ψ˜ + 2wψ˜ is equivalent to finding the roots of g(λ) = 0, where 1 g(λ) = − γ(λ) ∞ −1 2 0 ρw(L0 − λ) w dρ . ∞ 2 0 ρw dρ (2.112) As such, for the competition instability modes, we obtain from (2.108) that 1 (f − (λ + 1)(1 + ϕD0 )) = . γ 2[f − (λ + 1)] (2.113) In addition, for the synchronous instability mode we obtain that 1 µf − ϕ(λ + 1) f − (ϕ/µ)(λ + 1) = = , γ 2µ[f − (λ + 1)] 2[f − (λ + 1)] where µ= 2.6 (2.114) ϕ(N + 2τ λ/D0 ) . N + 2τ λ(ϕ + 1/D0 ) Stability theory; Small S-Analysis from summing log formulation In this section we re-derive and then improve upon our leading-order stability results of the previous section by expanding the stability formulation of § 2.2 for small source strengths. As shown previously, when D = O(1/ν), then the source strengths tend to zero. Consequently, the stability formulation of § 2.2 can be simplified in this limit. We will separate our analysis into two cases: τ = 0 for which we can obtain an explicit two-term result for the instability threshold, and τ > 0 where we can recover the NLEP problem. 73 2.6.1 Case A; τ = 0 We consider the competition instability threshold for the case where τ = 0. Recall that the leading-order result is given in (2.110). We will now derive a 2term expansion for this threshold for the case of a symmetric spot pattern where Sc = S1 = . . . = SN . We recall from the threshold calculation in § 2.2, that λ = 0 when det M = 0, where M= 1 I + (I − E0 )(B − G), ν and χ (S1 ) 0 .. , E0 = 1 B= . N 0 χ (SN ) 1 ··· . .. 1 ··· 1 0 .. . , G = 1 Lji Lij .. , . 0 with Lij = log |xi − xj | for i = j, and where S1 , . . . , SN satisfies S + ν(I − E0 )G S + ν(I − E0 )χ = √ 2 e. DN (2.115) For the symmetric case where S1 = . . . = SN = Sc , and Ge = k1 e, then √ Sc = 2/ DN is the exact solution to (2.115) for all ν, and det M = 0 reduces to (see (2.38)) 1 − kJ = −χ (Sc ), ν G bj = kj bj , T bj e = 0, (2.116) for j = 2, . . . , N. Now suppose that D = D0 /ν, with ν = −1/ log ε. Then we estimate Sc = √ 2 ν 1/2 D0 N 1. Hence, we must determine an expansion for χ (S) when S 74 1. To do so, we recall from (2.74) that d0 + d1 S + O(S 3 ) as S → 0, S ∞ b(1 − f ) 1 1 d0 ≡ , d ≡ ρ˜ u1pI dρ − 1 f2 b2 (1 − f ) 0 χ(S) ∼ where b = ∞ 2 0 ρw dρ. (2.117a) ∞ ρ˜ u1pII dρ , 0 (2.117b) Here u ˜1pI and u ˜1pII are the unique solutions of (2.76) (see § 2.3). Thus, χ (S) ∼ − Sd02 + d1 as S → 0, so that (2.116) becomes 1 d0 − kj = 2 − d1 + . . . . ν Sc We solve this relation for Sc as Sc = However, Sc2 = by 4 ν. D0 N 2 νd0 , 1 + ν(d1 − kj ) ν = −1/ log ε. Thus, the critical values D0c of D0 where λ = 0 are given D0c = 4 [1 + ν(d1 − kj )] , N 2 d0 where kj for 2 ≤ j ≤ N are any one of the eigenvalues of G with eigenvector perpendicular to e. The largest such value of kj will set the instability threshold. We summarize our result as follows. Principal Result 2.6.1 Suppose that the spot locations {x1 , . . . , xN } are such that G is a cyclic matrix so that Ge = k1 e. Consider a symmetric quasi-equilibrium solution with S1 = . . . = SN = Sc = √2 . DN Then, for D = O(− log ε), with D = D0 /ν, the smallest value of D0 , labelled by D0c , for which a competition instability occurs (i.e. λ = 0 with a sign-fluctuating eigenfunction cJ ) is at the 75 threshold D0c = 4 [1 + ν(d1 − kJ )] N 2 d0 (2.118) where kJ = max {kj }, 2≤j≤N G bj = kj bj , T bj e = 0, for j = 2, . . . , N, and d0 , d1 are given in (2.117b). Remark (i) The leading-order result is the same as was derived by NLEP theory in (2.110). However, we are able to obtain the next correction term to the leading-order result as given in (2.118). This correction term depends on a perturbation of the solution to the core problem as manifested in the parameter d1 , as well as on the locations of the spots. (ii) The sign-fluctuating eigenvector is bj ≡ cj . (iii) The limitation of this simple calculation is that it provides a threshold for λ = 0, but does not give a rigorous instability result for D > D0c 2.6.2 Case B, τ = 0 Next, we consider the case where τ = 0. Again we consider an N -spot pattern √ with spots such that Ge = k1 e. Then S1 = . . . = SN = Sc = 2/ DN . to In (2.46) we showed that the stability threshold when det M = 0 is equivalent ˜c = 0, ν −1 + 2πkjλ + B 76 for j = 1, . . . , N, (2.119) where ν = −1/ log ε, and kjλ are the eigenvalues of the matrix Gλ defined in (2.44), written as Gλ bjλ = kjλ bjλ , j = 1, . . . , N. Recall that the entries in this matrix are obtained from the Green’s function G(x; x0 ) of (2.42), written as the solution to ∆Gλ − τλ Gλ = −δ(x − x0 ), D 1 Gλ (x − x0 ) ∼ − log |x − x0 | + Rλ + o(1) as x → x0 . 2π (2.120) ˜c , In the derivation leading to (2.119) we required that λ = 0. The constant B which depends on λ and Sc , is obtained from the solution on 0 < ρ < ∞ to ˜ ˜ = λψ, Lρ ψ˜ − ψ˜ + 2f uv ψ˜ + f u2 N ˜ − 2uv) = 0, ˜ −N ˜ u2 + ψ(1 Lρ N ψ˜ → 0 ˜ → log ρ + B ˜c N (2.121) as ρ → ∞, as ρ → ∞, where u, v satisfy the core problem (2.57). Since D = O(− log ε ), then Sc ˜c for 1. As such, we must now calculate B Sc → 0. We start with the (2.121) system. As Sc → 0 we recall from (2.73) and (2.74) that u ∼ uc ∼ Sc w, f v0 vc ∼ v0 , Sc uc vc ∼ w . f Therefore (2.121) becomes S2 ˜ ˜ = λψ, Lρ ψ˜ − ψ˜ + 2wψ˜ + c2 w2 N f v0 2 ˜ + ψ˜ 1 − 2w = Sc w2 N ˜, Lρ N f f 2 v02 ˜ → log ρ + B ˜c , N ψ˜ → 0, as ρ → ∞. 77 (2.122) ˆ and B ˆc by Next, we introduce N ˜ =N ˆ /Sc2 , N ˜c = B ˆc /Sc2 , B (2.123) so that (2.122) becomes 1 ˜ ˆ = λψ, Lρ ψ˜ − ψ˜ + 2wψ˜ + 2 w2 N f v0 2 ˆ = S 2 ψ˜ 2w − 1 + w N ˆ , Lρ N c f f 2 v02 ˆ →B ˆc + Sc2 log ρ, N ψ˜ → 0, as ρ → ∞. We now expand the solution for Sc (2.124) 1 as ˆ =B ˆc + Sc2 N ˆ1 + . . . . N Then we obtain with L0 ψ˜ ≡ (Lρ − 1 + 2w)ψ˜ that 1 ˆc , (L0 − λ)ψ˜ = − 2 w2 B f v0 so that 1 ˆ −1 2 ψ˜ = − 2 B c (L0 − λ) w . f v0 (2.125) ˜ it follows that N ˆ1 from (2.124) satisfies In terms of ψ, 2 ˆ1 = ψ˜ 2w − 1 + w B ˆc , Lρ N f f 2 v02 ˆ1 ∼ log ρ N as ρ → ∞. Upon integrating this equation, we obtain the solvability condition ˆ )= lim (ρN 1 ρ→∞ ∞ 0 2w ψ˜ − 1 ρdρ + f 78 ∞ 0 2 ˆc w ρdρ, B f 2 v02 which becomes ∞ 2 f 0 ∞ ˜ 0 ψρdρ. Now we calculate ∞ 0 ∞ ˜ − ρLρ ψdρ ˜ ψwρdρ − 0 ˆc b B ˜ ψρdρ + 2 2 = 1. f v0 (2.126) We integrate the equation for ψ˜ in (2.124) to get ∞ ˜ ψρdρ +2 0 ∞ ˜ wψρdρ + 0 ˆc B f v02 ∞ ∞ w2 ρdρ = λ 0 ˜ ψρdρ. 0 Since the first integral is equal to zero, this reduces to ∞ ˜ ψρdρ = 0 2 λ+1 ∞ ˜ wψρdρ + 0 ˆc b B . f v02 (λ + 1) (2.127) ˜ wψρdρ = 1. (2.128) Then, we combine (2.126) and (2.127) to obtain ˆc b B f 2 v02 1− f λ+1 + 2 2 − f λ+1 ∞ 0 To simplify this expression we recall from (2.63) that v0 = b(1 − f )/f 2 . Therefore, we get b(1 − f )2 f 2 v02 = ≡ α. b f2 (2.129) Then, (2.128) can be written as ˆc B 1 + 2α f ∞ ˜ = wψρdρ 0 (λ + 1) . 2(λ + 1 − f ) (2.130) ˆc (L0 − λ)−1 w2 from (2.125) to get Next, we use ψ˜ = − f 1v2 B 0 ˆc ˆc B B − 2 2 2α f v0 ∞ 0 ρw(L0 − λ)−1 w2 dρ = (λ + 1) . 2(λ + 1 − f ) Now, upon multiplying the integral term top and bottom by b and then using (2.129) 79 for f 2 v02 /b, we get ˆc B ˆc B − 2α α ∞ −1 2 0 ρw(L0 − λ) w dρ ∞ 2 0 ρw dρ = (λ + 1) , 2(λ + 1 − f ) (2.131) which can be written as ˆc ˆc B (λ + 1) B − F(λ) = . 2α α 2(λ + 1 − f ) Finally, we write this problem as finding the roots of g(λ) = 0, where g(λ) = ˆc ) 1 f − (λ + 1)(1 − α/B − F(λ), 2 f − (λ + 1) F(λ) ≡ where (2.132) ∞ −1 2 0 wρ(L0 − λ) w dρ . ∞ 2 0 ρw dρ ˜c ˆc . We recall from (2.123) that B ˆc = Sc2 B This is an NLEP once we determine B √ as Sc → 0, where Sc = 2ν 1/2 / D0 N is the common source strength. Our goal is to verify that the roots g(λ) = 0 in (2.132) are the same as those obtained by the leading-order NLEP theory, as given in (2.112)-(2.114). ˆc . To do so, we first need to To establish this relationship we must determine B approximate the matrix Gλ when D = D0 /ν Gλ satisfies ∆ s Gλ − 1. For D = D0 /ν and λ = 0, then τ λν Gλ = −δ(x − x0 ), D0 (λ = 0). For ν = 0 this problem has no solution. Therefore, for λ = 0 and ν expand the solution as Gλ = a + G0 + O(ν) + . . . , ν 80 1 we must for some unknown constant a. At next order, we obtain that ∆ s G0 = aτ λ a − δ(x − x0 ), D0 G0 dx = 0. Ω We remark that the integral constraint for G0 arises from a solvability condition on the O(ν) correction term. By the divergence theorem we conclude that a = Thus, for ν D0 4πτ λ . 1, we have that Gλ (x; x0 ) ∼ D0 + G0 + O(ν), 4πτ λν where G0 is the Neumann Green’s function satisfying ∆s G0 = 1 − δ(x − x0 ), 4π G0 dx = 0 , Ω 1 with G0 ∼ − 2π log |x − x0 | + R0 as x → x0 . Then Rλ ∼ D0 4πτ λν + R0 , and we can write the Gλ matrix as Gλ = D0 N E0 + O(1), 4πτ λν 1 ··· 1 .. E0 = N . 1 ··· 1 .. . , 1 (2.133) which is valid for τ λ = 0. The eigenvalues and eigenvectors of Gλ are as follows: • Since E0 e = e, then Gλ e = k1 e, k1 = D0 N . 4πτ λν • Since E0 b = 0, whenever bT e = 0, then Gλ b = 0. There are N − 1 such independent vectors. 81 Then, we obtain from (2.119) that the threshold condition for stability is either B˜c + ν −1 = 0 whenever bT e = 0, (2.134) or, 1 B˜c + ν D0 N +1 2τ λ when Gλ e = k1 e. = 0, (2.135) The condition (2.134) represents the competition modes, whereas the condition (2.135) corresponds to the synchronous modes. We first consider the synchronous mode. We have from (2.135) that 1 B˜c + ν D0 N +1 2τ λ = 0. √ ˜c = B ˆc /Sc2 , with Sc = 2ν 1/2 / D0 N , we obtain that However, since B ˆc = − B that 2 τ λN 2 N+ 2τ λ D0 . Now, in the NLEP (2.132) we calculate with α = b(1 − f )2 /f 2 (see (2.129)) α 2τ λ b(1 − f )2 N 2 =− , 2 ˆ 4f N + 2τ λ/D0 Bc Next, we recall the definition of ϕ in (2.104) to obtain 1− α 2τ λϕ N + 2τ λ/D0 + 2τ λϕ =1+ = . ˆ N + 2τ λ/D N + 2τ λ/D0 0 Bc We then compare this with (2.114) and conclude that 1− α ϕ ≡ , ˆc µ B where µ is defined in (2.114). 82 Finally, we substitute this last expression into (2.132) to obtain g(λ) = 1 f − (λ + 1)(ϕ/µ) − F(λ), 2 f − (λ + 1) in agreement with (2.114). Next, we consider the competition modes. We have from (2.134) that B˜c = ˆc /S 2 , so that −ν −1 . Now B˜c = B c ˆc = −S 2 /ν = −4/D0 N 2 , B c √ since Sc = 2ν 1/2 / D0 N . We then calculate that α b(1 − f )2 =− ˆc f2 B D0 N 2 4 = −D0 b(1 − f )2 N 2 = −ϕD0 , 4f 2 with ϕ = N 2 (1 − f 2 )b/4f 2 . This shows that (2.132) becomes g(λ) = 1 f − (λ + 1)(1 + D0 ϕ) − F(λ), 2 f − (λ + 1) which agrees precisely with the NLEP in (2.113). We summarize the result of this section as follows. Let τ > 0 with τ = O(1). Assume that D = D0 /ν. Then, the small S asymptotics of the stability theory in the summing logs formulation of § 2.2 agrees with the NLEP as derived directly in section § 2.5. Finally, we remark that an advantage of the stability theory of § 2.2 is that it is accurate to all orders in ν. However, in order to implement the theory, numerical methods are needed and therefore it is difficult to obtain full analytical results. In contrast, for the leading-order NLEP stability theory we can obtain an explicit instability threshold in terms of D0 for the competition mode (see (2.110)), while 83 the small S-asymptotics of the stability theory of § 2.2 provides a higher order correction term (see (2.118)). However, for both stability formulations, numerical methods are needed to compute any Hopf bifurcation threshold in terms of τ . 2.7 Chapter summary In this chapter we derived a localized spot-type solution for the Brusselator model, and we analyzed its stability. By means of three different methods: a full nonlinear derivation that results in a DAE system, an asymptotic expansion in ν, and through the derivation of an NLEP for which stability results exist, we were able to derive thresholds for a competition instability, a spot-splitting instability, and an oscillatory Hopf instability. We corroborated the analytic thresholds with full numerical simulations. The DAE system obtained by uncoupling the fully nonlinear problem in a system with symmetric spot strengths was shown to have solutions that are connected to the classic Fekete/Thomson problem of distributing Coulomb charges on the surface of a sphere. In the next chapter we will apply similar techniques to the Schnakenberg model and will also derive a set of differential equations for the slow motion of the spots. 84 Chapter 3 The Schnakenberg Model on the Surface of the Sphere Reaction-diffusion systems have been previously proposed to model skin pigmentation patterns on a species of Angelfish ([25], [43]). Despite its origin as a mechanism to explain pattern formation on such two-dimensional manifolds, for analytical and computational simplicity most previous work has been done for either the case of one spatial dimension or for weakly nonlinear patterns near a spatially homogeneous equilibrium state. Our goal is to analytically characterize localized patterns for the Schnakenberg model on the surface of the sphere. The Schnakenberg model [52] on a sphere of radius L is a two-component reaction-diffusion system given in non-dimensional form by Vt − Dv ∆s V = f (U, V) ≡ b − V + UV 2 , Ut − Du ∆s U = g(U, V) ≡ a − UV 2 . (3.1) Here ∆s is the Laplace-Beltrami operator for a sphere of radius L. This system is one of the more robust pattern generators among reaction-diffusion systems, while 85 remaining amenable to analysis due to the algebraic simplicity of the kinetics [37]. The problem is connected with physical applications in which the constant terms (a, b) could represent source terms that couple a dynamical problem in the interior of the sphere with a diffusion process on the surface of the sphere ([11], [42]). The standard Turing analysis of linearizing a reaction-diffusion system around a spatially homogeneous equilibrium state is of somewhat limited use for characterizing patterns on the surface of the sphere. The difficulty is that mode predictability becomes severely hampered by the fact that the Laplacian eigenfunctions responsible for small amplitude spatially inhomogeneous patterns have a high degree of degeneracy on the surface of the sphere. This can be readily seen by linearizing around the homogeneous solution. For the Schnakenberg system, the spatially homogeneous base state is Ve = a + b and Ue = a/(a + b)2 . We linearize (3.1) around this spatially homogeneous base state to obtain the linearized problem φt = J φ + D∆s φ, D= 1 L2 Dv 0 0 Du , J = fV fU gV gU . (3.2) Ue ,Ve Here φ is a two-vector, L is the radius of the sphere, and ∆s now denotes the standard surface Laplacian on the unit sphere. Upon separating variables, it follows that the spatial eigenfunctions are the well-known spherical harmonics Y satisfying ∆s Y + k 2 Y = 0, (3.3) given explicitly in terms of Legendre polynomials as |m| Ylm (θ, φ) = cm l Pl (cos θ) exp(imφ), l = 0, 1, 2, . . . , |m| ≤ l , k 2 = l(l+1). (3.4) Here l and m are called the degree and the order of the spherical harmonic, respectively. With the exception of the simple eigenvalue k = 0, the other nonzero 86 eigenvalues k 2 = l(l + 1) with l > 0 are 2l + 1-fold degenerate, in the sense that there are 2l + 1 independent eigenfunctions, characterized by the order m in |m| ≤ l, corresponding to this eigenvalue. As l increases, the set of eigenfunctions becomes increasingly degenerate. A Turing stability analysis for the linearization of the Schnakenberg system on the surface of the sphere is a standard exercise and was done in [16]. In this 2-D case, the interval in k 2 where an instability of the spatially homogeneous base state occurs scales like O(L2 ), where L is the radius of the sphere. Thus, for large sphere radii, there will be a large number of Laplacian eigenvalues in this interval. This fact, together with the intrinsic degeneracy of the spherical harmonic eigenspace for larger eigenvalues, implies that mode prediction will only be accurate for small sphere radii (i.e. corresponding to low values of l). To illustrate this mode degeneracy, in [16] a table of values is given for the eigenvalue ranges as a function of increasing values of the radius of the sphere for the Schnakenberg system with a = 0.95 and b = 0.07. This shows that a standard linear Turing-type stability theory is not particularly well-suited for predicting small amplitude patterns on the surface of the sphere when the sphere has a large radius. The eigenpairs of the Laplacian become rather degenerate as the radius of the sphere increases, and mode determination becomes a very difficult issue. Although there have been some previous weakly nonlinear normal-form type theories for reaction-diffusion systems on a sphere near bifurcation points, showing the emergence of different solution branches from a single bifurcation point, it is in general difficult to determine which branch is the most stable and to determine its basin of attraction [30]. For the Brusselator model this has been done in [38]. Rather than adopting this weakly nonlinear viewpoint, our approach is to seek “particle-like” solutions consisting of localized spots to the fully nonlinear Schnakenberg system on the surface of the sphere. Our analysis relies on the assumption of an asymptotically large diffusion coefficient ratio. In this asymptotic limit, our goal is to characterize the existence, stability, and dynamics of such solutions. In 87 addition, we will determine a spot self-replication bifurcation that will be triggered as the radius of the sphere grows adiabatically in time. 3.1 Localized spot patterns on the sphere The specific system that we shall study in detail is the approximate system that results from (3.1) in the asymptotic regime for which Dv 1 and Du = O(Dv−1 ) 1. To derive this system, we let V = v/Dv and U = Dv u in (3.1) to obtain that vt = Dv ∆s v + bDv − v + uv 2 , Dv ut = Du Dv ∆s u + a − 1 uv 2 . (3.5) Dv We then label Dv ≡ ε 2 with ε → 0. Moreover, we define D by D = Du Dv , and we assume that D = O(1) as ε → 0. In this way, we can neglect the bDv and Dv ut terms in (3.5) and obtain the following elliptic-parabolic limit of the original Schnakenburg system (3.1): vt = ε 2 ∆s v − v + uv 2 , 0 = D∆s u + a − ε−2 uv 2 . (3.6) Morever, by a simple re-scaling of ε 2 and D by the square of the radius L of the sphere, it sufficies to consider (3.6) on the surface of the unit sphere, and so without loss of generality ∆s now denotes the usual Laplace-Beltrami operator on the unit sphere. The key bifurcation parameters in (3.6) are a > 0 and D > 0. Although this parabolic-elliptic limiting system of the original Schnakenburg model does not admit spot patterns that undergo Hopf bifurcations, there are two other instability mechanisms that occur and will be analyzed. 88 3.1.1 The quasi-equilibrium multi-spot pattern In the limit of small diffusivity ε → 0, we will first use the method of matched asymptotic expansions to construct a quasi-equilibrium solution of (3.6) with spots located at {x1 , . . . , xN } on the surface of the sphere. In the inner region near the j-th spot we introduce the local variables y = ε −1 (x − xj ) , Uj = D−1/2 u , ρ = |y| , Vj = D1/2 v , (3.7) where Vj (ρ) and Uj (ρ) are radially symmetric. Upon substituting (3.7) into (3.6) we are effectively making a tangent plane approximation to the surface of the sphere at xj ∈ Ω. We obtain that Uj and Vj satisfy the following (so-called) core problem on 0 < ρ < ∞: 1 Vj + Vj − Vj + Uj Vj2 = 0 , ρ Uj (0) = Vj (0) = 0; 1 Uj + Uj − Uj Vj2 = 0, ρ Vj → 0 and Uj ∼ Sj log ρ + χ(Sj ) + o(1) as ρ → ∞. (3.8) Upon integrating the equation for Uj on 0 < ρ < ∞ we obtain the identity that ∞ Sj = 0 ρUj Vj2 dρ . (3.9) We can solve (3.8) for a range of values of Sj , and then at each Sj output the constant χ(Sj ) defined by the limiting process limρ→∞ (Uj − Sj log ρ) = χ(Sj ). This core problem is solved numerically on a large but finite domain, and in this way we obtain an approximation to χ(Sj ), as shown in Figure 3.1. Our numerical results show that there is a unique solution to this system at least on the range 0 < Sj < 7.5. Next, we determine a nonlinear algebraic system for the source strengths S1 , . . . , SN . 89 25 20 15 χ 10 5 0 −5 −10 0 1 2 3 4 5 S 6 7 8 j Figure 3.1: Numerical estimation of χ(Sj ) by solving the core problem (3.8) This is done by asymptotically matching the inner solutions near each spot to a global outer solution for u valid away from the spot locations. The determination of the Sj for j = 1, . . . , N then specifies the inner solution near each spot. To formulate the outer problem for the inhibitor variable u, we first estimate in the sense of distributions the term proportional to ε −2 in the u-equation of (3.6). We calculate that ε−2 uv 2 → R2 1 √ Uj (DVj2 )dy D √ δ(x − xj ) = 2π DSj δ(x − xj ). Thus, the outer problem for u on the surface of the sphere is ∆s u = − a 2π +√ D D N i=1 Si δ(x − xi ), (3.10) subject to the N matching conditions that Sj 1 u∼ √ Sj log |x − xj | + + χ(Sj ) ν D as x → xj , j = 1, . . . , N , (3.11) 90 where ν = −1/ log ε. To solve this problem for u, we define the Neumann Green’s function G(x; x0 ) as the unique solution to ∆s G = 1 − δ (x − x0 ), 4π x ∈ Ω, G(x; x0 ) dx = 0 , (3.12) Ω with G being 2π periodic in φ, and smooth at θ = 0, π. The exact solution is G(x; x0 ) = − 1 log |x − x0 | + R, 2π R= 1 [2 log 2 − 1]. 4π (3.13) Then, the solution to (3.10) and (3.11) can be represented as 2π u(x) = − √ D N Si G(x; xi ) + uc , (3.14) i=1 where uc is a constant to be found. The divergence theorem, as applied to (3.10) and (3.11), yields that N 2a Si = √ . D i=1 (3.15) Then, upon expanding the solution in (3.14) as x → xj and using the matching condition (3.11) we obtain that Sj 2π − √ − log |x − xj | + Sj R + 2π D N Sj Gji + uc i=j Sj 1 ∼√ Sj log |x − xj | + + χ(Sj ) , ν D for each j = 1, . . . , N . The singular terms in these matching conditions agree automatically, while the 91 matching of the constant terms leads to the N nonlinear algebraic equations Sj + 2π Sj R + ν N i=j Si Gji + χ(Sj ) = −2πuc , j = 1, . . . N, coupled to the scalar constraint (3.15). Here Gji = G(xj ; xi ) = G(xi ; xj ). This yields an N + 1 dimensional nonlinear algebraic system for the determination of the unknowns S1 , . . . , SN and uc . In this way, we obtain that S1 , . . . , SN and uc satisfy the matrix system (I + 2πν(G + RE)) S + ν χ(S) = −2πuc νe, 0 Lij χ(S1 ) 2a .. , χ = . . . , E = 1 eeT , eT S = √ , G = . N D Lji 0 χ(SN ) (3.16) 1 log |xi − xj |. where e = (1, . . . , 1)T , and Lij = − 2π From the system we can eliminate uc and obtain a set of N nonlinear algebraic equations for the unknowns S = (S1 , . . . , SN )T . We summarize our main result for the construction of the quasi-equilibrium N -spot solution as follows: Principal Result 3.1.1 In the limit ε → 0, an N -spot quasi-equilibrium solution to (3.6) is characterized by N ve ∼ ue ∼ j=1 √ DVj (ε −1 |x − xj |), √1 Uj (ε −1 |x − xj |) D N − √2π i=1 Si G(x; xi ) + D for |x − xj | = O(ε ) uc for |x − xj | (3.17) O(ε ). Here Uj and Vj satisfy the core problem (3.8). In addition, the vector S of source 92 strengths S1 , . . . , SN satisfies the nonlinear algebraic system S + ν(I − E)(χ + 2π(G + RE)S) = 2a √ , N D E= 1 T ee , N ν= −1 . log ε (3.18) Remark (i) Suppose that {x1 , . . . , xN } are such that G is a cyclic matrix. This always occurs for two spot patterns, for spots equally spaced on a ring of constant latitude, and for other such symmetric patterns (see §3.4 below). However, this condition imposes a restriction on the spot locations for general multi-spot patterns. In the cyclic case, we have that Ge = k1 e. Therefore, we can look for a solution to (3.18) of the form S = Sc e, so that χ = χ(Sc )e. From (3.18) we obtain using (I − E)e = 0 that (I − E) χ + 2π(G + RE)S = χ(Sc )(I − E)e + 2πSc (I − E)(k1 + R)e = 0. Therefore, for the cyclic case we conclude that to all orders in ν there exists a solution to (3.18) with a common source strength Sc where Sc = 3.2 2a √ . N D (3.19) The spot self-replication threshold In this section we study a linear instability mechanism for the local deformation of a spot. This peanut-type instability is the trigger for a nonlinear spot self-replication event. Such an instability mechanism was first analyzed for the Schnakenberg system in a planar domain in [24]. Since this instability is a lo93 cal instability, the analysis of peanut-splitting instabilities for a spot on the sphere parallels that of the planar case. We now briefly outline this analysis. We linearize (3.6) around the quasi-equilibrium solution of (3.17) by writing v = ve + eλt φ , u = ue + eλt η . (3.20) By substituting (3.20) into (3.6) and linearizing, we obtain the following eigenvalue problem for φ and η: ε2 ∆s φ − φ + 2ue ve φ + ve2 η = λφ, (3.21) D∆s η − 2ε−2 ue ve φ − ε−2 ve2 η = 0. In the j-th inner region we have 1 ue = √ Uj , D ve = √ DVj , y = ε−1 (x − xj ), (3.22) where Uj and Vj satisfy the core problem (3.8). In the inner region near xj we seek an O(1) time-scale instability associated with the local angular integer mode ˆj (ρ) and Φ ˆ j (ρ) by m satisfies m ≥ 2. We introduce the new variables N η= 1 imω ˆ e Nj (ρ) , D ˆ j (ρ) , φ = eimω Φ ρ = |y| , ω = arg y . (3.23) Upon substituting (3.22) and (3.23) into (3.21), we obtain the following radially symmetric eigenvalue problem where the integer mode m ≥ 2 is a parameter: m2 ˆ 2 ˆ ˆ ˆ ∆ρ Φˆj − Φˆj − 2 Φ j + 2Uj Vj Φj + Vj Nj = λΦj , ρ m2 ˆ 2 ˆ ˆ ∆ρ Nˆj − 2 N j − 2Uj Vj Φj − Vj Nj = 0, ρ Φˆj → 0, as ρ → ∞, Nˆj → 0, as ρ → ∞. (3.24) ˆj as ρ → ∞. Since m ≥ 2 we can impose the decay condition for N 94 The eigenvalue problem (3.24) was solved numerically in §2.3 of [24]. We label λ0 to be the eigenvalue of this problem with the largest real part. Since the core solution depends on Sj from (3.8), then λ0 = λ0 (Sj , m). To determine the onset of any instabilities, the threshold value Sj = Σm where Re(λ0 (Σm , m)) = 0 was computed. In the computations of [24], only the modes m = 2, 3, 4, . . . were considered, since λ0 = 0 for any value of Sj for the translational mode m = 1. For m ≥ 2, the computations of [24] showed that λ0 (Sj , m) is real and that λ0 (Sj , m) > 0 when Sj > Σm , and that Σ2 < Σ3 < Σ4 etc.. Therefore, the smallest value of Sj where an instability is triggered occurs for the “peanut-splitting” instability m = 2. The numerical value for this threshold was found to be Σ2 ≈ 4.3. We conclude that there is a peanut-splitting instability for the j-th spot if and only if Sj > Σ2 ≈ 4.3. Figure 3.2: Spot-splitting in the Schnakenberg model (u(x)). The same dynamics occurs in the lower hemisphere in this example, as the initial configuration consisted of two spots. The parameters were D = 1, ε = 0.1, R = 1.5, and a similar initial condition with R = 1 will does not split. 95 With regards to the effect of domain growth on spot-splitting, we predict that spot self-replication will occur once the radius of a sphere becomes greater than some critical value. To show this qualitatively, suppose that we have an initial configuration of N spots for which the Green’s matrix is cyclic. Then, the common source strength Sc from (3.19) is Sc = 2a √ . N D We conclude that Sc > Σ2 ≈ 4.3, 2a N Σ2 2 when D< . Since D is inversely proportional to the square of the radius L of the sphere, we conclude that spot self-replication will occur when L exceeds a critical value LN . Further spot self-replication events will occur when L increases past a further threshold. We conclude that spot self-replication is an under-crowding type of instability. 3.2.1 The competition instability threshold In this subsection we obtain a new explicit two-term result for the competition instability threshold for an N -spot quasi-equilibrium pattern with spots at {x1 , . . . , xN } for the special case where G is a cyclic matrix. We will find that the stability threshold is D = O(1/ν) + O(1), with ν = −1/ log ε, where both terms are calculated analytically. Similarly to our analysis of the Brusselator model, in our stability analysis we will “freeze” the locations of the spots, since they evolve on a much longer timescale of order O(ε −2 ) than the O(1) time-scale needed to initiate a competition instability. The calculation of the competition instability threshold proceeds in three distinct steps. 1. For the fully coupled core problem (3.8), we determine a two-term approximation to this system for Sj → 0. In particular, we determine a two-term asymptotic expansion for χ(Sj ) as Sj → 0. Then, when G is a cyclic ma96 trix, so that Ge = k1 e, there exists a quasi-equilibrium pattern with a common source strength Sc , where Sc is given in (3.19) and is O(ν 1/2 ) when D = O(ν −1 ). This motivates the need for the small S analysis of the solution to the core problem. 2. By linearizing around the quasi-equilibrium N -spot solution we formulate a globally coupled nonlocal eigenvalue problem (NLEP). We then write the condition for λ = 0 to be an eigenvalue of this problem. The threshold condition will involve χ (Sc ) in a central way. 3. From the information obtained in step 1 and step 2, we finally eliminate Sc to determine a two-term asymptotic result for the stability threshold in terms of D. Step 1: Asymptotics as S → 0 of the solution to the core problem For the core problem (3.8), we replace Uj → U , Vj → V , and Sj → S, and write the radially symmetric core problem as ∆ρ V − V + U V 2 = 0, ∆ρ U = U V 2 , V → 0, as ρ → ∞, U ∼ S log ρ + χ(S), as ρ → ∞. (3.25) We now give a formal scaling analysis for the limit S → 0. Let U = uS −P , V = S P v, so that U V 2 = O(U V )O(V ) = O(V ). Hence, the V equation is invariant, whereas for the U equation we have ∆ρ u = S 2p uv 2 , u ∼ S 1+p log ρ + S p χ. In order to obtain a distinguished limit, we require that 2p = p + 1, and χ = O(S −p ). This yields that p = 1. Hence, for S → 0, we have U ∼ u/S, V ∼ Sv, and χ ∼ χ/S. ˆ In a systematic way, we can now expand the solution to the core problem using 97 the previous scaling. We write 1ˆ U, S U= V = S Vˆ , χ= 1 χ ˆ. S (3.26) Then, (3.25) transforms to ˆ Vˆ 2 = 0, ∆ρ Vˆ − Vˆ + U ˆ =U ˆ Vˆ 2 S 2 , ∆ρ U Vˆ → 0, ˆ ∼χ U ˆ + S 2 log ρ, as ρ → ∞, as ρ → ∞. (3.27) Since we will need a two-term asymptotic result for χ(S) as S → 0, we will expand Vˆ = Vˆ0 + S 2 Vˆ1 + S 4 Vˆ2 + · · · , ˆ =U ˆ0 + S 2 U ˆ1 + S 4 U ˆ2 + · · · , U χ ˆ=χ ˆ0 + S 2 χ ˆ1 + S 4 χ ˆ2 + · · · . We substitute this expansion into (3.27) and equate powers of S 2 to obtain ˆ0 Vˆ02 = 0, ∆ρ Vˆ0 − Vˆ0 + U Vˆ0 → 0, ˆ0 = 0, ∆ρ U ˆ0 ∼ χ U ˆ0 , as ρ → ∞, as ρ → ∞, (3.28) and ˆ0 Vˆ0 Vˆ1 = −U ˆ1 Vˆ02 , ∆ρ Vˆ1 − Vˆ1 + 2U ˆ1 = U ˆ0 Vˆ 2 , ∆ρ U 0 Vˆ1 → 0, ˆ1 ∼ χ U ˆ1 + log ρ, as ρ → ∞, as ρ → ∞. (3.29) ˆ2 is At one higher order the problem for U ˆ2 = U ˆ1 Vˆ02 + 2U ˆ Vˆ0 Vˆ1 , ∆ρ U ˆ2 U bounded as The solution to (3.28) is w , Vˆ0 = χ ˆ0 ˆ0 = χ U ˆ0 , 98 ρ → ∞. (3.30) where w is the radially symmetric ground-state solution satisfying ∆ρ w−w+w2 = 0. Then (3.29) becomes ˆ1 U L0 Vˆ1 ≡ ∆ρ Vˆ1 − Vˆ1 + 2wVˆ1 = − 2 w2 , Vˆ1 → 0, χ ˆ0 ˆ1 = 1 w 2 , ˆ1 ∼ log ρ + χ ∆ρ U U ˆ1 , χ ˆ0 as ρ → ∞, (3.31) as ρ → ∞. ˆ1 equation over 0 < ρ < ∞, we obtain that Upon integrating the U ∞ χ ˆ0 = b, b= ρw2 dρ. 0 Since we require a two-term expansion in order to obtain our stability threshold ˆ1 and Vˆ1 as below, we must calculate χ ˆ1 . To do so, we decompose U 1 ˆ ˆ1 = χ U1p , U ˆ1 + χ ˆ0 χ ˆ1 1 Vˆ1 = − 2 w + 3 Vˆ1p . χ ˆ0 χ ˆ0 (3.32) ˆ1p and Upon using the identity that L0 w = w2 , we readily derive from (3.29) that U Vˆ1p satisfy ˆ1p , L0 Vˆ1p = −w2 U ˆ1p = w2 , ∆ρ U 0 < ρ < ∞; 0 < ρ < ∞; Vˆ1p → 0, ˆ1p ∼ b log ρ + o(1), U as ρ → ∞, as ρ → ∞. (3.33) We remark that there exists a unique solution to (3.33) since we have imposed that ˆ1p − b log ρ → 0 as ρ → ∞. U ˆ2 equation in (3.30). Since U ˆ2 is Finally, to obtain χ ˆ1 we integrate the U bounded at infinity we obtain that ∞ 0 ˆ1 Vˆ 2 + 2ρU ˆ0 Vˆ0 Vˆ1 dρ = 0. ρU 0 99 (3.34) Upon using (3.32), we obtain ∞ ρ χ ˆ1 + 0 1 ˆ U1p χ ˆ0 1 2 w dρ = −2 χ ˆ20 ∞ 0 ρw − χ ˆ1 1 w + 3 Vˆ1p dρ, 2 χ ˆ0 χ ˆ0 which reduces to χ ˆ1 = 1 χ ˆ0 b ∞ ˆ1p w2 + 2wVˆ1p dρ. ρ U 0 ˆ1p w2 = −∆ρ Vˆ1p + Vˆ1p , the formula above Finally, since χ ˆ0 = b and 2wVˆ1p + U for χ ˆ1 can be written compactly as χ ˆ1 = 1 b2 ∞ ∞ ρVˆ1p dρ, with b = 0 ρw2 dρ. 0 We summarize our result as follows. In the limit S → 0, the solution to the core problem (3.25) has the following asymptotic behaviour: U∼ 1 S V ∼S Here χ ˆ0 = b = 1 ˆ U1p + · · · , χ ˆ0 χ ˆ1 1 − 2 + 3 Vˆ1p + · · · . χ ˆ0 χ ˆ0 χ ˆ0 + S 2 χ ˆ1 + w + S2 χ ˆ0 ∞ 2 0 ρw dρ, χˆ1 = 1 b2 ∞ ˆ 0 ρV1p dρ, (3.35) ˆ1p , Vˆ1p are the unique and U solutions to (3.33). The function χ(S) in (3.25) has the following two-term asymptotics for S → 0: χ∼ 1 b χ ˆ0 + S 2 χ ˆ1 + · · · = + S χ ˆ1 + · · · . S S (3.36) Therefore, for S → 0, we have χ (S) ∼ − b +χ ˆ1 + o(1) as S2 S → 0, 100 χˆ1 = 1 b2 ∞ 0 ρVˆ1p dρ. (3.37) This ends step 1. Step 2: Formulation of the globally coupled eigenvalue problem Next, we linearize (3.6) around the quasi-equilibrium solution of (3.17) to obtain on the surface of the sphere that the perturbation satisfies (3.21). In the j-th inner region we have 1 ue = √ Uj , D ve = √ DVj , y = ε−1 (x − xj ). To analyze competition instabilities, we look for a locally radially symmetric eigenfunction. In the j-th spot inner region, we let η = 1 D Nj (ρ) and φ = Φj (ρ) to obtain from (3.21) that on 0 < ρ < ∞, ∆ρ Φj − Φj + 2Uj Vj Φj + Vj2 Nj = λΦj , ∆ρ Nj − 2Uj Vj Φj − Vj2 Nj = 0. We must impose that Nj has logarithmic growth as ρ → ∞. We then set Φj = cj Φˆj , Nj = cj Nˆj , where cj is an arbitrary constant, to obtain that ˆ j + Vj2 Nˆj = λΦ ˆj, ∆ρ Φˆj − Φˆj + 2Uj Vj Φ ∆ρ Nˆj − 2Uj Vj Φˆj − Vj2 Nˆj = 0, Φˆj → 0, ˆj , Nˆj ∼ log ρ + B as ρ → ∞, as ρ → ∞. (3.38) ˆj = B ˆj (Sj , λ) must be computed numerically. By integrating the equation Here B ˆj over 0 < ρ < ∞, and recalling that Nj = cj Nˆj , we obtain the following for N identity that is needed below: ∞ cj = 0 2Uj Vj Φj + Vj2 Nj ρdρ. 101 (3.39) We note that Bˆj = Bˆj (Sj , λ). Upon differentiating the core problem (3.8) with respect to Sj , and then comparing the resulting system with (3.38), we conclude when λ = 0 that Bˆj (Sj , 0) = χ (Sj ). (3.40) Here χ(Sj ), as defined in (3.25), must be computed from the core problem. Next, we determine the matching condition and we formulate the problem for the outer solution for η. The far field of the inner solution for η, when written in outer variables, yields the following matching condition for the outer solution: η∼ 1 1 cj log |x − xj | + + Bˆj , D ν as x → xj . (3.41) In order to derive the outer problem for η we must estimate, in the sense of distributions, the terms proportional to ε−2 in the η equation in (3.21). We calculate that 2ε−2 ue ve φ → 2 ε−2 ve2 η → R2 R2 Φj Uj Vj dy δ(x − xj ), DVj2 1 Nj dy δ(x − xj ). D Upon combining these expressions, we get 2ε−2 ue ve φ + ε−2 ve2 η → 2π ∞ 0 2Φj Uj Vj + Vj2 Nj ρdρ δ(x − xj ). Therefore, upon using the identity (3.39), and the matching condition (3.41), we obtain that the outer solution for η on the surface of the unit sphere satisfies ∆η = 2π D N i=1 cj δ(x − xi ), (3.42) 1 1 η ∼ cj log |x − xj | + + Bˆj , as x → xj , D ν 102 j = 1, . . . , N. From the divergence theorem we get N ci = 0. i=1 The solution η is then represented as 2π η=− D N cj G(x; xi ) + i=1 η¯ , νD where η¯ is a constant to be determined and G is the Neumann Green’s function of (3.12). Upon expanding η as x → xj and then comparing the result with the required singularity behaviour in (3.42), we obtain that c1 , . . . , cN and η¯ satisfy the homogeneous linear system N N cj Gji + cj Bˆj ν = η¯, cj + 2πν cj R + j = 1, . . . , N ; ci = 0. i=1 i=j (3.43) To write this system more conveniently in matrix form, we introduce c1 . . c= . , cN Bˆ1 B≡ 0 0 .. . BˆN , 1 . . e= . , 1 so that (3.43) becomes c + 2πν(G + RE)c + νBc = η¯e, eT c = 0, (3.44) where G is the usual Neumann Green’s matrix of (3.16). Upon multiplying this system with eT we can then eliminate the scalar η¯. In this way, we obtain the 103 following homogeneous linear system for c: [I + 2πν(I − E)(G + RE) + ν(I − E)B] c = 0. We conclude that any discrete eigenvalues corresponding to a locally radially symmetric perturbation near each spot must satisfy det M = 0, where M = M(λ) is the N × N matrix defined by Mc = 0, M ≡ I + 2πν(I − E)(G + RE) + ν(I − E)B. (3.45) Therefore, we must find conditions that guarantee the existence of a non-trivial c. We refer to this eigenvalue problem as the globally coupled extended nonlocal eigenvalue problem (NLEP). Remark (i) In order to find the stability threshold in terms of D we will look for conditions for which det M = 0 when λ = 0. (ii) Recall that when G is cyclic, then to all orders in ν, there exists a solution with a common source strength Sc as given in (3.19). (iii) We also recall from (3.40) that when λ = 0, then Bˆj (Sj , 0) = χ (Sj ). Therefore, when the spots have a common source strength Sc , the matrix B in (3.45) is simply B = χ (Sc )I at the threshold λ = 0, where I is the identity matrix. Therefore, in the cyclic case we calculate at λ = 0 that c must be a nontrivial 104 solution to 2πν(I − E)(G + RE)c = − I + νχ (Sc )(I − E) c. (3.46) This completes step 2 of the analysis. Step 3: Calculation of the stability threshold We now look for conditions on D for which there exists a nontrivial solution c to (3.46). We label k1 and kj , for j = 2, . . . , N , to be the eigenvalues of G. The matrix spectrum of G is simply G qj = kj qj , eT qj = 0, Ge = k1 e, synchronous mode, j = 2, . . . , N, competition modes. For the synchronous mode, we replace c = e in (3.46), and use Ge = k1 e, and (I − E)e = 0. This leads to 2πν(I − E)(G + RE)e = − I + νχ (Sc )(I − E) e, which reduces to the contradictory statement that 0 = e. Therefore, as expected, we conclude that there is no instability threshold associated with the synchronous mode. For the competition modes, we let c = qj for j = 2, . . . , N where qj T e = 0. Then, we calculate G qj = kj qj , and E qj = 0. Thus, (3.46) becomes 2πν(I − E)(G + RE)qj = − I + νχ (Sc )(I − E) qj , kj 2πν(I − E)qj = −qj − νχ (Sc )qj , 2πνkj qj = −(1 + νχ (Sc ))qj . 105 We conclude that det M(0) = 0 when 2πνkj = −1 − νχ (Sc ), j = 2, . . . , N. Therefore, for the competition modes, there are N − 1 distinct thresholds where λ = 0. In terms of D they are given by the roots of the transcendental equations 1 2π χ (Sc ) + Finally, we use Sc = 2a √ , N D 1 ν = −kj , j = 2, . . . , N. (3.47) and the two-term asymptotics χ (Sc ) ∼ − Sb2 + χˆ1 c as Sc → 0 as given in (3.37) to solve (3.47) asymptotically for D. Substituting these results into (3.47) we obtain − b 1 + χˆ1 + ∼ −2πkj , 2 Sc ν so that Sc2 ∼ Then, upon recalling that Sc2 = D= ν −1 4a2 N 2D b . + χˆ1 + 2πkj from (3.19), we solve for D to obtain that 4a2 (1 + ν(2πkj + χˆ1 )) . bN 2 ν This completes the final step 3. We summarize our main result for competition instabilities as follows: Principal Result 3.2.1 Suppose that the configuration {x1 , . . . , xN } of spots are such that G is a cyclic matrix. Then, there exists an N -spot quasi-equilibrium solution with common source strengths, i.e., Sj = Sc for all j = 1, . . . N . For this solution, the globally coupled extended NLEP has a zero eigenvalue corresponding to a sign-fluctuating instability of the spot amplitudes at the critical values Dj for 106 j = 2, . . . , N of the inhibitor diffusivity D, given by Dj = 4a2 4a2 + (2πkj + χ ˆ1 ) + o(1), 2 bN ν bN 2 as ν → 0. (3.48) Here kj for j = 2, . . . , N are the eigenvalues of the Green’s matrix G in the N − 1 dimensional subspace perpendicular to e, i.e. G qj = kj qj for eT qj = 0 and j = 2, . . . , N . In addition, b = ∞ 2 0 ρw dρ, where w is the ground-state solution, ν = −1/ log ε, and χ ˆ1 is determined from a correction to the leading-order core solution as χˆ1 = 1 b2 ∞ ρVˆ1p dρ, 0 where Vˆ1p is the unique solution to (3.33). Remark (i) The instability threshold is then Dth = min2≤j≤N Dj , which involves the minimum of the kj . (ii) Our analysis was based on reducing the Schnakenberg system to the unit sphere. For a sphere of adiabatically slowly increasing radius L, our competition instability result can still be used if we identify that D and ε2 have both decreased by a factor of L2 . Therefore, competition instabilities become increasingly less prominent as the radius of the sphere increases. This is in direct contrast to the occurrence of spot self-replication instabilities, which become more prominent as the sphere radius increases. In a nutshell, a competition instability is an over-crowding instability, whereas, as discussed earlier, a spot self-replication instability is an undercrowding instability. 107 3.3 Slow spot dynamics on the surface of the sphere In this section we derive a differential algebraic system (DAE) of ODE’s characterizing the slow motion of a collection of spots. For a planar domain, such an analysis has been previously given in [24]. However, on the surface of the sphere, the analysis needed to derive the DAE system is rather more intricate in that we must retain certain new higher order terms in the inner solution near each spot related to the curvature of the sphere. As such, we must exercice care in working with the Laplacian in the spherical coordinate system. The DAE system characterizes slow spot motion in the absence of any spot self-replication or competition instability. On the unit sphere, the Laplacian in spherical coordinates is ∆θ ,φ = 1 ∂ sin θ ∂θ sin θ ∂ ∂θ + ∂2 1 1 ∂φφ . 2 ∂φ2 = ∂θ θ + cot θ ∂θ + sin θ sin2 θ (3.49) We introduce the local coordinate system near the j-th spot as s1 = ε −1 (θ − θ j (τ )), s2 = ε −1 sin θ j (φ − φj (τ )), where τ = ε 2 t. Upon substituting this into (3.49) we obtain that ∆s1 ,s2 = sin2 θ j 1 1 1 ∂ cot(θ + ε s )∂ ∂s s + + s s j 1 s 1 1 1 ε2 ε ε 2 sin2 (θ j + ε s1 ) 2 2 1 1 1 1 ∂s1 s1 + cot θ j ∂s1 + 2 ∂s s 2 ε ε ε (1 + 2ε s1 cot θ j ) 2 2 1 1 2 (∂s1 s1 + ∂s2 s2 ) + cot θ j ∂s1 − s1 cot θ j ∂s2 s2 . 2 ε ε ε (3.50) This result is used below to identify key correction terms in the inner region that are crucial for the analysis of spot motion. The Schnakenberg model (3.6) on the surface of the sphere when written in spherical coordinates is vt = ε2 ∆θ ,φ v − v + uv 2 , 0 = D∆θ ,φ u + a − ε −2 uv 2 , 108 where ∆θ ,φ is written explicitly in (3.49). In the j-th inner region, we perform the local change of variables to s1 and s2 , as written explicitly above, and we introduce U and V by u = D−1/2 U and √ v = DV . In this way, we obtain that Vt = Vs1 s1 + Vs2 s2 + ε cot θ j Vs1 − 2ε s1 cot θ j Vs2 s2 + O(ε 2 ) − V + U V 2 , 0 = D Us1 s1 + Us2 s2 + ε cot θ j Us1 − 2ε s1 cot θ j Us2 s2 + O(ε 2 ) √ − U V 2 + ε 2 Da. (3.51) Next, the time derivative Vt on the left hand side of (3.51) is calculated as ∂ ∂s1 ∂θ j ∂τ ∂ ∂s2 ∂φj ∂τ d = + = −ε Vs1 θ j − ε sin θ j Vs2 φj dt ∂s1 ∂θ j ∂τ ∂t ∂s2 ∂φj ∂τ ∂t (3.52) = −ε (Vs1 , Vs2 ) · (θ j , sin θ j φj ). Here the primes indicate derivatives with respect to the slow time variable τ given by τ = ε 2 t. The key point, which suggested the asymptotic order in ε of the slow time-scale, is that the time-derivative term must balance the O(ε ) order of the spatial correction terms on the right hand-sides of (3.51). This balance is achieved when τ = ε 2 t. As such, the explicit form of the O(ε ) correction terms suggests that we expand U = U0 + ε U1 + · · · , V = V0 + ε V1 + · · · . We substitute this expansion together with (3.52) into (3.51) and collect powers of ε. At leading order we recover the core problem V0s1 ,s1 + V0s2 ,s2 − V0 + U0 V02 = 0 , U0s1 ,s1 + U0s2 ,s2 − U0 V02 = 0. (3.53) 109 Upon collecting the O(ε ) terms we get −(V0s1 , V0s2 ) · (θ j , sin θ j φj ) = V1s1 ,s1 + V1s2 ,s2 + cot θ j (V0s1 − 2s1 V0s2 ,s2 ) − V1 + U1 V02 + 2U0 V0 V1 0 = U1s1 ,s1 + U1s2 ,s2 + cot θ j (U0s1 − 2s1 U0s2 ,s2 ) − U1 V02 − 2U0 V0 V1 . (3.54) We then rewrite (3.54) in a convenient matrix form as ∆s W1 + AW1 = f. (3.55) Here ∆s ≡ ∂s1 ,s1 + ∂s2 ,s2 , and we have defined the vectors W1 and f , and the matrix A, by W1 ≡ f≡ V1 U1 , A= −1 + 2U0 V0 −2U0 V0 V02 −V02 − cot θ j (V0s1 − 2s1 V0s2 ,s2 ) − (V0s1 , V0s2 ) · (θ j , sin θ j φj ) − cot θ j (U0s1 − 2s1 U0s2 ,s2 ) , (3.56) . The ODE system for the spot locations is obtained from imposing a solvability condition on (3.56). However, we must first determine the correct far-field behaviour for the solution W1 before invoking this condition. The required far-field condition on W1 is now determined from an asymptotic matching procedure with the outer solution. The outer solution for u was given in (3.14) as N 2π u(x) = √ Si G(x; xi ) + uc , D i=1 1 1 G(x; x0 ) = − log |x − x0 | + (2 log 2 − 1). 2π 4π 110 On the unit sphere, both |x| = |x0 | = 1, and by the law of cosines |x − x0 |2 = 2 − 2 cos ω, with ω denoting the angle between x and x0 . In terms of spherical coordinates we can write G(x; x0 ) = G(φ, θ ; φ0 , θ 0 ) explicitly as 1 1 log(1 − x · x0 ) + (log 2 − 1) 4π 4π 1 = − log(1 − sin θ sin θ 0 cos φ cos φ0 − sin θ sin θ 0 sin φ sin φ0 − cos θ cos θ 0 ) 4π 1 + (log 2 − 1). 4π (3.57) G=− From this expression we can readily compute the partial derivatives ∂G/∂φ and ∂G/∂θ at the spot location φ0 , θ 0 . We can now use this result to find a matching condition for the inner problem. We Taylor expand the outer problem as x → xj , and write the resulting expression in terms of local coordinates to obtain Sj 2π u(x) ∼ − √ − log |x − xj | + Sj R + 2π D 2πε −√ D Here we have defined N i=j and Sj Gji + uc i=j ∂Gj ∂Gj s2 s1 + ∂θ ∂φ sin θ j Si ∂Gj ∂θ N ∂Gj ∂φ (3.58) + O(ε 2 ). to be the partial derivatives of G evaluated at the j-th spot location θ = θ j and φ = φj . These terms can be calculated explicitly from (3.57). Hence, in terms of the local variables associated with the inner problem, the O(ε ) term in (3.58) gives the required far-field behaviour of the inner solution. Since the inner expansion u = D−1/2 (U0 + ε U1 + · · · ), we obtain from (3.58) that U1 must have the far-field behaviour N U1 ∼ −2π Si i=j ∂Gj ∂Gj s2 s1 + ∂θ ∂φ sin θ j 111 = α · s. (3.59) Here we have defined s = (s1 , s2 )T and α = (α1 , α2 ) as N α ≡ −2π ∂Gj ∂Gj 1 , ∂θ ∂φ sin θ j Si i=j . (3.60) Therefore, the required far-field behaviour of the inner problem for W1 , given by (3.55) is W1 ∼ 0 α·s as |s| → ∞. , (3.61) The final step in the determination of an ODE system for the dynamics of the spots is to impose a solvability condition on the solution to the inner problem (3.55) subject to the far-field behaviour (3.61). The homogeneous adjoint problem for (3.55) is ∆s P + AT P = 0, s ∈ R2 , P ≡ Φ Ψ . (3.62) Let Pˆ satisfy the radially-symmetric problem, with the far-field condition that ˆ Ψ) ˆ T → (0, 0)T as ρ → ∞. The precise asymptotic behaviour of Pˆ is readily (Φ, seen to be Pˆ ∼ (0, ρ−1 )T as ρ → ∞. Then, we seek solutions to (3.62) of the form Pc ≡ Pˆ cos Θ , or Ps ≡ Pˆ sin Θ , where Θ is the polar angle for s. In or- der to apply the solvability condition below, we need to define the inner product (u, v) = T R2 (u v)ds, where ds ≡ ds1 ds2 . Moreover, we define the operator LF ≡ ∆s F + AF and its adjoint L∗ F ≡ ∆s F + AT F . By combining the problems satisfied by W1 and Pc we get (Pc , LW1 ) − (W1 , L∗ Pc ) = R2 PcT (∆s W1 + AW1 ) − W1T (∆s Pc + AT Pc ) ds = R2 PcT ∆s W1 − W1T ∆s Pc ds1 ds2 = ∞ 112 0 2π 0 R2 PcT f ds (3.63) Pˆ T cos Θ f ρ dρdΘ . Next, we use Green’s second identity to the left hand-side of this expression to derive 2π PcT ∆s W1 − W1T ∆s Pc ds = lim σ→∞ 0 R2 ∂W1 ∂ Pˆ Pˆ T cos Θ − W1T cos Θ ∂ρ ∂ρ 2π = lim σ→∞ 0 2π PcT 0 σ dΘ ρ=σ −σ(α1 cos Θ + α2 sin Θ ) 2π σdΘ ρ=σ 1 cos Θ (α1 cos Θ + α2 sin Θ ) σ = = ∂W1 ∂Pc − W1T ∂ρ ∂ρ −1 cos Θ σ2 σ dΘ ∞ 2α1 cos2 Θ dΘ = 2πα1 . 0 (3.64) To calculate the right hand-side of the last expression in (3.63) we use the expression for f from (3.56). Then, since (θ j , sin θ j φj ) = −ε (s1 , s2 ), we get ∞ 0 2π 0 = R2 ∞ Pˆ T cos Θ f ρ dρ dΘ = 0 2π ˆ Ψ) ˆ T f cos Θ ρ dρ dΘ (Φ, 0 ˆ − cot θ j (V0s − 2s1 V0s ,s ) + ε (V0s , V0s ) · (s1 , s2 ) cos Θ ds Φ 1 2 2 1 2 + R2 ˆ − cot θ j (U0s − 2s1 U0s ,s ) cos Θ ds. Ψ 1 2 2 (3.65) Recall that both U0 and V0 are radially symmetric, as well as the two compoˆ and Ψ ˆ of the adjoint solution Pˆ . In addition, we note that ∂s = ∂ρ cos Θ , nents Φ 1 ∂s2 = ∂ρ sin Θ . Furthermore, we observe that since the terms U0s2 ,s2 and V0s2 ,s2 are even functions of s1 , so that when they are multiplied by s1 they become odd, and hence they integrate to zero over R2 . With these considerations, the left-hand 113 side of (3.65) reduces to R2 ˆ cot θ j V0ρ + ε ΦV ˆ 0ρ s − Ψ ˆ cot θ j U0ρ cos2 Θ ds = 2πα1 . (3.66) −Φ 1 Therefore, the solvability condition for Pc provides an ODE for s1 : ε s1 = −θ j = 2α1 + cot θ j ∞ ˆ ΦV0ρ + 0 ∞ ˆ 0 ΦV0ρ ρ dρ ˆ 0ρ ρ dρ ΨU = − α1 + β cot θ j γ , 2 (3.67) where γ ≡ −2/ ∞ ˆ 0 ΦV0ρ ρ dρ, and β ≡ ∞ 0 ˆ 0ρ + ΨU ˆ 0ρ ρ dρ are two funcΦV tions of the local spot strength Sj involving the core solution and the two components of the solution to the homogeneous adjoint problem. The function γ appeared in the derivation of [24] for spot dynamics in a planar domain, and it is plotted as a function of Sj in Figure 3 of [24]. For spot dynamics on the sphere, the new integral term β arises and must be computed as a function of Sj . Upon repeating the same procedure with Ps = Pˆ sin Θ , the left hand-side of the resulting expression is completely analogous and we get 2πα2 . For the right hand-side, all the ∂s1 = ∂ρ cos Θ terms integrate to zero, as well as the terms involving second derivatives by using the same symmetry argument as with Pc . In this way, we get ε s2 = −φj sin θ j = 2α2 ∞ ˆ 0 ΦV0ρ ρ dρ = −α2 γ . (3.68) We summarize our result as follows: Principal Result 3.3.1 Consider a collection of N spots on the surface of the sphere at spherical coordinates φj and θ j for j = 1, . . . , N . Then, provided that the quasi-equilibrium pattern is stable to any O(1) time-scale instability, the slow motion on the time-scale τ = ε 2 t of these collection of spots satisfies a DAE system. For this DAE system, the dynamics (θ j , sin θ j φj ) = −ε (s1 , s2 ) of the 114 spots is given by ε s1 = −θ j = 2α1 + cot θ j ε s2 = −φj sin θ j = ∞ ˆ ΦV0ρ + 0 ∞ ˆ 0 ΦV0ρ ρ dρ ˆ 0ρ ρ dρ ΨU , 2α2 , ∞ ˆ 0 ΦV0ρ ρ dρ (3.69) (3.70) for j = 1, . . . , N with primes denoting derivatives with respect to τ and N α = (α1 , α2 )T ≡ −2π Si i=j ∂Gj ∂Gj 1 , ∂θ ∂φ sin θ j . The constraints in the DAE system consist of the nonlinear algebraic system (3.18) for the source strengths S1 , . . . , SN defined in terms of the Neumann Green’s matrix, which involves the instantaneous spot locations. Remark (i) In the DAE system the spot strengths evolve slowly in time as a result of the slow motion of the collection of spots. The DAE system is valid provided that that the spots strengths are below the spot-splitting threshold, i.e. Sj < 4.3 for j = 1, . . . , N . (ii) The ODE dynamics involves two separate integral terms that must be computed as a function of the local source strength Sj . In a (practical) numerical implementation of spot dynamics these functions can be tabulated numerically in advance. It is beyond the scope of this thesis to investigate the consequences of this DAE system for spot evolution, such as the presence of stable stationary patterns or orbits. However, we remark that this system is vaguely related to, but seemingly more complicated than, the well-studied ODE systems characterizing the motion of Eulerian fluid point vortices on the surface of the sphere ([40], [9]). 115 3.4 Quasi-equilibria and the cyclic matrix structure Solving the quasi-equilibrium problem for a collection of N spots on a sphere implies finding the locations {x1 , . . . , xN } and source strengths S1 , . . . , SN that satisfy (3.18) (I + 2πν(G + RE))S + ν χ = −2πνuc e; 2a eT S = √ , D (3.71) with I the identity matrix, and S1 χ(S1 ) 1 . . . . . . S≡ χ≡ e≡ . , . , . , SN χ(SN ) 1 1 −xN | 0 ··· − log |x2π log |x2 −x1 | − 2π .. . .. .. G≡ , . . log |xN −1 −xN | − 2π N −x1 | − log |x2π ··· 0 with R = 1 4π (2 log 2 − 1) and E = 1 T N ee . The nonlinear term χ makes it impossible to solve the problem in the general case, but we can solve a reduced case where by prescribing the source strengths for all the N spots. The problem becomes that of finding the locations of the spots. Similarly, a second approach would be to prescribe the location of the spots and solve to find the source strengths. In the related problem of point vortices on a rotating sphere [18], this approach was used to categorize collections of point vortices on a sphere located at the vertices of platonic solids. We will follow the first approach and consider the problem where S is an eigenvector of the cyclic matrix G, say with corresponding eigenvector λ. We will sim- plify things further by assuming that all the spots have a common source strength 116 Sc , therefore S = Sc e and χ = χ(Sc )e. This particular choice of S, together with the assumption that G is a cyclic matrix, allows us to decouple (3.71) into √ (1 + 2πν(λ + R)2a/ D − νeT χ(Sc ) uc = − 2πνN 2a Sc = √ , N D (3.72) The problem has now been reduced to finding spot configurations that make G into a cyclic matrix. We remark that since the matrix is real symmetric, all of its eigenvalues are real. For a cyclic matrix G, it must have a constant row sum, and thus it has the eigenvector e with corresponding eigenvalue λ= −1 log 2π N j=1 j=i |xi − xj | , (3.73) for any row i. Since all the non diagonal terms in G are negative, by use of the Gershgorin circle theorem we can see that λ is the most negative eigenvalue in G. More importantly, requiring that e is an eigenvector of G is a geometrical re- striction on the position of the spots. The fact that all the rows in the matrix have to add to λ implies that the net effect of all the spots on each other is the same. The simplest way for this to happen is if the spots are the same distance apart from each other, i.e., |xi − xj | = C for all rows and i = j. This condition can be true for at most 4 spots in a sphere, at the vertices of an equilateral pyramid (or 3 equally spaced spots along a latitudinal ring, or 2 spots anywhere on the sphere. A second way to distribute the spots so that e is an eigenvector of G is by arranging them in rings. Without loss of generality, the idea is to arrange the spots in latitudinal rings, with the spots equally-space on the ring. There are results on the stability of point vortices arranged in latitudinal rings ([8], [9]), including 117 a famous result that limits the number of spots in such a configuration to ensure stability. The notion of stability in this point vortex problem is stability with respect to small perturbations of the locations of the vortices on the ring. We emphasize that in our context of spot patterns for reaction-diffusion systems, there are two different classes of eigenvalues governing the stability of the pattern. Our analysis has focused exclusively on the stability of the spot profile through our study of spot-replication, competition, and oscillatory instabilities. These spot amplitude-type instabilities, are “fast” instabilities as they result from O(1) unstable eigenvalues, and they appear to have no direct counterpart in the point-vortex fluid problem. For spot patterns, there are also weak translationaltype instabilities associated with eigenvalues that are O(ε 2 ). These eigenvalues can in principle be obtained by linearizing the DAE system in Principal Result 3.3.1 around an equilibrium configuration of spots, and determining the matrix spectrum of the associated Jacobian. It is the study of such “small” eigenvalue instabilities that has a more direct counterpart with the notion of stability of point vortices on the sphere. We have not analyzed any such small eigenvalue instability for spot patterns in this thesis, but we do expect that there should be a result similar to ([8] and [9] for the maximum number of “stable” spots that can be placed on a ring of constant latitude. In fact, for the case of a planar unit disk containing a ring of concentric equally-spaced spots, it was observed numerically for the Schnakenburg system in [24] that this system can support a maximum number of spots, beyond which some spots are pushed off the ring due to a small eigenvalue instability. With this digression, we now return to discussing configurations of spots for which G is a cyclic matrix. For more than 4 points, and not considering a ring, the points will have to be distributed in a less trivial way. We can no longer have an arrangement in which all the points are an equal distance from each other. We have to distribute the points in such a way that the sum of the contributions over each are the same. In geometrical terms, this is related to the symmetry of the configuration. Since the rows of G are all the same, there have to be rotations of the sphere that maintain the same configuration. 118 Similarly to when we were looking for equally-spaced points, there is a finite number of point distributions that exactly satisfy that condition, namely, the platonic solids. We can arrange 4, 6, 8, 12, and 20 points in a sphere such that their corresponding matrix G will also have e as an eigenvector. For homogeneous configurations with a different number of spots, e will cease to be an eigenvector. However, what we have observed is that by arranging the points in the most homogeneous possible distribution, the first eigenvector of G will be close to e (see table 3.1). A condition on a homogeneous distribution of N points on a sphere is that it maximizes the minimal distance between the points. A collection of these optimal point distributions, numerically computed, is available for up to 130 points, and can be found at Neil Sloane’s webpage [55]. Interestingly enough, maximizing the minimal distances does not necessarily correspond to platonic solids. For 8 and 20 nodes the minimizing configurations do not correspond to either a cube, or a dodecahedron, respectively. This was also empirically observed in [61]. Configurations of points that are the most homogeneously distributed on a sphere for a given number of points are called elliptic Fekete point distributions. Mathematically, Fekete points maximize the product of the distances between all pairs of points, N J= j=1 j=i N |xi − xj | , V = j=1 j=i log |xi − xj |. (3.74) Since log is monotonic, then maximizing J is equivalent to maximizing V . Notice how close that equation is to the eigenvalue equation in (3.73). As mentioned earlier, an alternative description is that of a group of equally charged particles that repel each other on the surface of the sphere. Minimizing the electrostatic potential yields the Fekete problem, and in this context is known as 119 the Thomson problem [58] The Fekete points can be numerically computed by considering V to be a spring-like potential, adding a relaxation term (friction), and using Lagrange multipliers. We solved the system with a symplectic numerical scheme, following the description in [51]. For various optimal spot locations from the numerical Fekete configurations, we computed the difference between the eigenvector corresponding to the smallest eigenvalue of G and e. Number of points n 4 6 8 10 12 20 20p 50 100 120 130 ||xn − en || 3.8459e − 16 1.8626e − 10 6.0301e − 16 0.0216 2.9374e − 16 0.0050 8.5142e − 16 0.0029 0.0029 0.0015 0.0027 Table 3.1: The norm of the difference between the first eigenvector and e, both normalized, for different spot configurations. Although, as can be seen in table 3.1, we can no longer expect to find e as an eigenvector of the distribution of points, the first eigenvector of distributions with minimal distances between points maximized is in fact close to e. Notice too, that in the 8 point case (a rotated cube), the difference in norm is within machine error, despite the fact that we no longer have a platonic solid. However, the optimal distribution in the sense of maximal minimal distances for 20 points yields a result accurate only to 3 decimals. The result 20p corresponds to a true dodecahedral distribution of 20 points, and instead we have ||xn − en || = 8.5142e − 16. For 120 comparison, a random distribution of 100 points has a norm difference of 0.2512. 3.5 Numerical method for reaction-diffusion patterns on the sphere To validate our asymptotic stability theory for the Brusselator and the Schnakenberg reaction-diffusion models, we need to compare our results with full numerical simulations of these PDE systems on the surface of the sphere. Previous work on reaction-diffusion systems on spheres can be found in [61], [10], [16], [17]. The first work that we are aware of is that of [61]. In this study, they numerically solved a generic reaction-diffusion system on a sphere and were able to obtain spotted and striped stationary solutions, depending on the the strength of the cubic and quadratic terms respectively. The numerical method they used was to write the Laplacian in spherical coordinates, and integrate in (θ, φ) space explicitly with an Euler method after an appropriate patching of the boundary. This approach was also used in [17] where they also considering growing spheres, although the method was implicit-explicit. The disadvantage of this method is that the finite difference method subdivides the (θ , φ) domain into equal subsections, and on the full sphere this corresponds to a majority of the grid points being close to the poles. Since the maximum time step in two dimensions for an explicit time integration scales as (∆x)2 , with ∆x being the smallest distance between points, the integration time will reflect the scale at the poles, making it very numerically intensive to resolve events that happen elsewhere, such as in the vicinity of localized spots. The work by [10] involved simulating a reaction-diffusion system that modelled tumour growth. The numerical method used was a method of lines with a spectral component in the reaction terms that was generalized for any reaction terms. Our previous experience with spectral methods on systems with localized 121 patterns was limited by the small number of grid points and their distribution. Although these methods typically have a high degree of accuracy and very fast convergence rates, the large number of points needed to resolve the structures made them unreliable, at least in our implementation. The approach that we chose to implement is the recently-developed Closest Point Method (see [50]). The basic idea of this method is to solve the PDE system on a grid on an embedding domain of dimension greater than the system. The method works by propagating the values of the solution on the embedded domain to the grid, iterating the solution on the grid to the next timestep, and propagating back to the embedded domain via a suitable interpolation algorithm ([7], [46]). Instead of working with the full 3-D space, it is sufficient to restrict the problem to a narrow band around the embedded domain, with the width of the band thick enough to contain enough points to make the interpolation of the same order as the differentiation scheme. This approach has the advantage that the problem is solved on the grid using basic Laplacian operators, and the geometry of the system is handled through the way the solution is propagated to and from the grid. Other than the closest point algorithm, the rest of the approach is standard explicit or implicit solvers, plus an interpolation algorithm in 3-D. For the problem of simulating spot dynamics for either the Brusselator or the Schnakenberg model on the surface of the sphere, we need to be perform very long-time integrations since the time-scale for the motion of the spots is O(1/ε 2 ). To be able to take large time steps the method has to be implicit, or at least implicit-explicit, as the CFL condition limits explicit methods to time steps of order O((∆x)2 ). For a simulation with 81 grid points in each direction, we have that ∆t = 0.0005. Currently in our full numerical simulations we typically use ε = 0.1 as a trade-off between computational difficulty and reasonable accuracy of the asymptotic theory. To fully validate the asymptotic theory we would need to use somewhat smaller values of ε . However, this has a detrimental effect on the stability of the computations, as the discretized diffusion term becomes stiffer. As a result of this computational challenge, we were not able to validate our asymptotic results for the dynamics of spots for the Schnakenberg model, nor were we 122 able to numerically realize a “chaotic attractor”, consisting of intermittent spot self-replication and annhilation events, for the Brusselator model. However, as we discussed, the stability theory that we undertook for the Brusselator model did strongly suggest that such an attractor should exist. Therefore, our numerical experiments for the full Schnakenberg and Brusselator system were undertaken only to exhibit spot self-replication and competition instabilities for moderate values of ε , since it is these instabilities that occur on a fast O(1) time-scale. In addition, the explicit version of this method that we implemented was sufficient to validate the results obtained in [17] on stationary spheres. As observed also in [61], and predicted in the asymptotic analysis, the number of spots that can stabilize on a sphere depends on the radius of the sphere, or equivalently on the value of the inhibitor diffusivity. Closest Point Method. One of the main authors of this method [49] graciously shared with us a basic explicit code that solved a linear diffusion problem on a sphere. We improved it by changing the interpolation method, which was a basic divided-differences approach, to a barycentric interpolation method (see [7]), which is about an order of magnitude more efficient. Interpolation an a 3-D lattice (N 3 points) was done by first interpolating in the z direction N 2 times, then using those values to interpolate in the y direction N times, and finally do a single interpolation on the x direction ([46]). Integrating to T = 100, on a grid of 40×40×40 takes around half an hour on a computer with an Intel T9400 processor (2.5 Ghz). We now briefly summarize the Closest Point Method [50]. The key idea of the Closest Point Method lies in the closest point extension. If the intrinsic gradient and divergence operators on the surface are denoted ∇s , and ∇s ·, by mapping the points in the grid to their closest point on the surface, we have on the grid that ∇s u(x) = ∇u(CP (x)) , ∇s · u(x) = ∇ · u(CP (x)). This is intuitively easy to see, as it all depends on the fact that u(CP (x)) 123 is constant in directions that are normal to the surface. This can be generalized to higher order operators too, and for our purposes we used that ∇s · (∇s u(x)) = ∇·(∇u(CP (x))) = ∆u(CP (x)), for the Laplace-Beltrami operator on the surface of the sphere. When the embedded domain is a unit sphere, centred at the origin, the closest point algorithm is simply cp_x = closest_point(x) d = sqrt(x(1)ˆ2 + x(2)ˆ2 + x(3)ˆ2) cp_x = x/d The following pseudo-code represents an explicit algorithm for updating w(n) to w(n+1) in a differential equation of the form wt = rhs(w): for i=1 to num_points x = domain_border + dx*i cp_x = closest_point(x) w_temp(x) = interpolate(w(n),cp_x) end for i=1 to num_points w(x,n+1) = w_temp(x) + dt*(rhs(w_temp))) end With more complicated domains, the part of the algorithm that finds the closest point can get significantly more computationally expensive. However, this computation only needs to be done once, and this is the only part of the code that has to be adjusted for different domains. In that sense, a very powerful feature of the Closest Point Method is that it can deal with general domains independently from the discretization or interpolation algorithms. 124 3.6 Chapter summary In this chapter we derived a localized spot-type solution for the Schnakenberg model, and we analyzed its stability. By employing a full nonlinear derivation that results in a DAE system similar to the one obtained for the Brusselator in the previous chapter, we obtained the threshold for a self-replication instability. We analyzed in detail the possible spot distributions that would result in solutions with equal spot strengths, and connected it with the Fekete problem, as was also the case in the Brusselator problem. We did full numerics using a specialized algorithm for solving PDEs in general surfaces called the Closest Point Method, thus verifying the analytic results. Furthermore, we considered the case of the spots motion on the surface of the sphere and derived differential equations for the dynamics of the full system. In the next chapter we will discuss 1D mesa solutions for the GMS system, derive solutions and study their stability. 125 Chapter 4 Case study: mesa patterns on the GMS system 4.1 Model formulation and Turing stability analysis One of the first models for a Turing system was introduced by Alfred Gierer and Hans Meinhardt in 1972 [15]. This model was constructed as a simplified representation of an activator-inhibitor reaction with sources, and can give rise to a large number of patterns. It was first used as a simple model for biological morphogenesis, and used to model patterns on sea shells [34]. The standard GM model has been shown to exhibit spike-type solutions [63]. On the other hand, an extension on the GM model that adds saturation to the reaction kinetics can exhibit mesa-type patterns [22], i.e., truncated spikes with wide flat tops, separated by two sharp interfaces from a quasi-constant solution. In this section we will study the dimensionless Gierer-Meinhardt model with 126 saturation, which in non-dimensional form can be written as ut = ε 2 ∆u + f (u, v) = ε 2 ∆u − u + u2 v(1 + ku2 ) (4.1) 2 τ vt = D∆v + g(u, v) = D∆v − v + u , on x ∈ [−1, 1], with homogeneous Neumann boundary conditions. Here k > 0 is the saturation parameter. We consider the singular limit where ε 1, and we will study the solutions that arise for various parameter regimes. We will consider the case of a growing domain, both dynamically L = L(t), and adiabatically, and discuss the modifications that need to be made in the dynamic growth case. In the splitting regime we will analyze the effect of varying the growth rate. 4.1.1 Turing stability analysis A spatially homogeneous stationary solution to (4.1) occurs when u2 = u, v(1 + ku2 ) u2 = v, which involves solving the cubic equation u + ku3 − 1 = 0. We will consider positive values of k only, therefore the curve h(u) = u + ku3 − 1 is monotonically increasing, thus guaranteeing that there will only be one homogeneous solution. For example, when k = 2.5, which was the used for most of the numerical calculations below , the stationary homogeneous solution is us 0.5603, vs = u2s . Starting close to the homogeneous solution, the evolution of the u(x) solution can be seen in Figure 4.1. The mesa profiles appear relatively quickly, and over a much longer time-scale they rearrange to occupy the domain equally distributed, as can be seen in the solution at t = 50, 000. According to Turing theory (see [60], [37]), when the ratio of the diffusion 127 1.0 0.8 u(x) 0.6 0.4 T=1 T=50 T=500 T=50,000 0.2 0.0 −3 −2 −1 0 x 1 2 3 Figure 4.1: Solution profiles for various integration times starting close to the homogeneous solution. The figure on the left has the time evolution of the solution for u(x) up to t = 500, with time in the y-axis. The figure on the left represents four different snapshots, at t = 1, t = 50, t = 500, and t = 50, 000. We used κ = 2.5, D = 10, τ = 1, and ε = 0.01 on (4.1), with a random initial condition close to u = 0.5603, v = u2 . The numerical method we utilized was an implicit-explicit scheme. coefficients is large (D/ε 2 1), the homogeneous solution becomes unstable and a stable heterogeneous solution develops. By linearizing around the homogeneous solution it is possible to determine the domain length L at which the new solution appears, as well as to derive general conditions on the existence of heterogeneous solutions. The linearized problem around the equilibrium solutions previously computed is ut = ε 2 ∆u + fu (ueq , veq )u + fv (ueq , veq )v (4.2) τ vt = D∆v + gu (ueq , veq )u + gv (ueq , veq )v. Using separation of variables, w(x, t) ≡ (u, v)T = w(t)W(x), the spatial component leads to the eigenvalue problem ∆W + k 2 W = 0, 128 (ˆ n · ∇)W = 0. For the 1-D domain [0, L] we have that the eigenvalues are k = nπ/L, and the corresponding eigenvectors are W ∝ cos(kx), for n ∈ N. Likewise, the purely temporal solution of the linearized system is w ∝ eλt . The full solution to the linearized problem is then given by ck eλt Wk (x) = w(x, t) = k ck eλt cos(kx), k with the constants ck defined by the initial condition. Upon substituting into (4.2), we get the system λWk = −Dk 2 Wk + AWk , D= ε2 0 0 D/τ A= fu (ueq , veq ) fv (ueq , veq ) gu (ueq , veq )/τ gv (ueq , veq )/τ (4.3) . Nontrivial Wk solutions will exist for values of λ determined by the roots of the characteristic polynomial of the matrix A − Dk 2 . Since the equilibrium solutions are constant, we have a 2 × 2 constant matrix, and thus we have that the characteristic polynomial for λ is λ2 + λ(k 2 (D/τ + ε 2 ) − fu − gv /τ ) + h(k 2 ) = 0 h(k 2 ) = k 4 ε 2 D/τ − k 2 (ε 2 gv /τ + fu D/τ ) + fu gv /τ − fv gu /τ. We are looking for solutions that are stable in the absence of diffusion, therefore if k = 0 we require fu + gv /τ < 0 (tr(A) < 0) and fu gv − fv gu > 0 (det(A) > 0). The particular solutions we are interested in become unstable when spatial effects are taken into account (k = 0). Spatially heterogeneous solutions will appear when the real part of λ(k 2 ) becomes positive. A necessary condition for this to happen is h(k 2 ) < 0, hence we 129 require the additional condition that ε 2 gv + fu D > 0. To satisfy a sufficient condition we must also require that h(k 2 ) < 0 for some 2 k 2 . By differentiation we find that the critical value is kmin = 2 h(kmin ) = det(A) − ε 2 gv +fu D , 2ε 2 D and (ε 2 gv + fu D)2 . 4ε 2 Dτ Together, we have four conditions that need to be satisfied in order for the system to have a bifurcation to non-homogeneous solutions. These are fu + gv /τ < 0, fu gv − fv gu > 0, ε 2 gv + fu D > 0, det(A) − (4.4) (ε 2 gv + fu D)2 < 0. 4ε 2 Dτ The values of L at which heterogeneous solutions appear, for various eigenmodes n, are shown in Table 4.1. Eigenmode n L 1 2 4 8 0.3729 0.7458 1.4917 2.9833 Table 4.1: Some domain length values at which non-homogeneous solutions appear, according to Turing theory. The values were computed using the constants τ = 1, ε = 0.02, D = 1. The eigenmodes correspond to one, two, four, and eight peaks. These values can be seen overlapped in the full bifurcation diagram of Figure 4.4. When considering systems with domains that grow dynamically, the standard 130 Turing equations need to be modified accordingly. When considering the extended reaction-difusion equations (RDE) system in (4.10), we considered very small growth rates r, say r 1. As a first approxima- tion, assuming t = O(1), the only thing that changes in the Turing conditions is that the eigenvalues shift to the left. Instead of (4.3), the −rW term appears in the diagonal as λr Wk = −Dk 2 Wk + AWk − rI, and the eigenvalues for this case are λr = λs −r, with respect to the λs eigenvalues of the stationary system. In § 4.2 we will discuss recent analytical results that describe in more detail the effect of growing domains on the four Turing conditions (4.4). 4.2 Domain growth extension It has long been speculated that domain growth is one of the mechanisms that influences pattern selection. In the work by Kondo and Asai [25], a domain lengthdependent term in a reaction-diffusion system was introduced, and the resulting patterns successfully mimicked the characteristics exhibited by growing fish. Further work ([12], [13], [3]) specifically addressed the effect that domain growth had on patterning, and recent research has generalized the formalisms of Turing analysis to account for domain growth, both numerically ([27],[17]), and analytically ([45], [35], [28]). Moreover, one of the main criticisms to Turing’s postulation of RDE as models for pattern formation has been that mode selection appears to not be robust enough when compared to naturally occurring patterns. Adding domain growth to the equations has shown in some cases to increase the robustness of patterns [45], and in itself works as a mechanism for pattern selection. A general system of RDE for a fixed domain Ω , in one or more dimensions, can 131 be extended to account for growing domains. The following derivation was based on Plaza et al.’s work [45]. Consider a parameter (spatial) s ∈ [0, 1], defined for each time t, which is used to parametrize the mapping ψt , such that for every time t ≥ 0, ψt : [0, 1] −→ R3 , x(s, t) ψt ≡ X(s, t) = y(s, t) z(s, t) (4.5) It is required that ψt : Ω0 ⊂ R −→ R3 be C 2 for every t ≥ 0, and continuously differentiable with respect to t. Thus, (4.5) defines a regular curve Ct embedded in R3 , with the characteristic that for every s ∈ [0, 1], and t ≥ 0 Xs (s, t) = 0. The arclength σ(s, t) of the mapping ψt , as a function of s and t, and its derivative with respect to s, are: s σ= 0 |Xs (s , t)|ds , (4.6) σs = |Xs (s, t)|. Let φ be the concentration (mols per unit length) of the chemicals φ = (φ1 , φ2 , ...)T . Assuming that the chemicals diffuse according to Fick’s law, the flux vector J of the chemicals, proportional to the concentration gradient, is: J = −D∇φ, for a particular diagonal diffusion matrix D (no cross diffusion terms are consid- ered). Given that the rate of change in the chemicals’ concentration is proportional to the flow through the boundary ∂Ω , for any outer unit normal n ˆ, the volume that 132 exits Ω through the element dS is −J · n ˆdS. d dt Ω(t) φdX = −D ∂Ω(t) J ·n ˆdS = D ∂Ω(t) ∇φ · n ˆdS. (4.7) By rewriting (4.7), and using the parametrization from (4.5), making a change of variables, using (4.6), and integrating in a segment [s1 , s2 ] of the curve Ct , defined by ψt ([si , s2 ], t), with [s1 , s2 ] ∈ [0, 1], the left hand side yields: d dt φdX = Ω(t) d dt s2 φ(X(s, t), t)σs (s, t)ds. s1 Notice that since Ω is a line, ∂Ω corresponds to the two endpoints of the interval. Doing the same change of variables for the right hand side of (4.7) results in: s2 D ∂Ω(t) ∇φ·ˆ ndS = D∇X φ s1 s2 =D s1 ∂s ∇X φ · |Xs | |Xs | s2 ds = D ∂s s1 φs σs Dropping the integral sign on both sides (as it’s valid on any interval [s1 , s2 ] ⊂ (0, 1), t ≥ 0), the generalized RDE on a growing 1-D domain is obtained: φt = D (σs )2 φss − σss φs σs − σst φ, σs or, equivalently φt = D ∂s σs φs σs − ∂t (ln σs )φ . (4.8) For the standard pattern-formation problem with two reactants, φ ≡ w = (u, v)T , (4.8) corresponds to the uncoupled portion of the equations. The coupling comes from the reaction term F (w) = (f (u, v), g(u, v))T , as seen in models such as the Schnakenberg, Barrio-Varea-Arag´on-Maini model (BVAM), Gierer-Meinhardt, etc. The full model, for two interacting chemicals w = (u, v)T , in non-dimensional 133 ds. form, with D = Du 0 0 Dv wt = , is D σs2 wss − σss ws σs − σst w + F (w) σs (4.9) In 1-D, a number of simplifying assumptions can be made that make the problem much more analytically tractable, while keeping it physically meaningful. Assuming that the system grows isotropically, it is possible to separate the growth term from the manifold parametrization, X(x, t) = ρ(t)X0 (s). Moreover, a straight line domain, characterized by X0 (s) = (s, 0, 0)T , with s ∈ [0, 1], results in σs = 1, σss = 0. The system can be further simplified by assuming slow exponential growth (ρ(t) = exp(rt)), hence σst ρt = = r. σs ρ With all of the above assumptions, the simplified version of (4.9) is, then wt = D wss − rw + F (w) ρ2 (4.10) The effect that domain growth has on the four conditions necessary for a diffusiondriven instability was recently studied [28]. Even in the simplified case of exponential growth in (4.10), the effective diffusion coefficients become non-autonomous, invalidating the standard linear approach. Asymptotic analysis showed that the conditions for patterning on growing domains are less strict, i.e., it is possible to observe patterns on systems with activators on both components, and similarly for short-range inhibition and long-range activation systems. In terms of bifurcation theory, systems with domain growth exhibit delayed self-replication events when compared to solutions obtained by adiabatically varying the domain length parameter L (integrating for each value of L until equilibrium is achieved). This behaviour was encountered in [3] for the 1-d Schnakenberg system (figure 5a), although the delayed bifurcation phenomena was not discussed. 134 4.3 Stability of 1-d mesa patterns for the Gierer-Meinhardt model with saturation (GMS) model when D = O(1) In the previous section we derived a set of Turing conditions for the existence of non-homogeneous patterns in the GMS model. We also extended the reaction diffusion system to account for growing domains. We will now use the previous results to investigate numerically the effect of domain growth on mesa patterns, as well as the difference between domains that grow adiabatically versus dynamically. We will use results from numerical bifurcation theory, and perform analytical continuation on stable solutions in order to obtain a global picture of the bifurcation diagram. Our computations are for the parameter range where D = O(1). 4.3.1 Construction of a single mesa Solving the GMS system numerically reveals that the solutions past the Turing thresholds have a mesa profile (see Figure 4.2). A single mesa is essentially a saturated pulse solution with a symmetric flat profile in the centre of the domain, separated from a quasi-zero solution by sharp interfaces located at x = ±l. Analogously to pulse systems such as the Schnakenberg system (cf. [52]), in this param- eter regime increasing the domain length will split the solution into images of the single mesa. The mesa solutions can tought of as two back-to-back heteroclinic solutions that are separated by a plateau regio. This characteristic can be exploited to obtain a first asymptotic approximation of the system. This approach was done originally in [23], and we outline the construction here. We start by looking for an asymptotic solution near the interface x = l to the 135 2.5 2.0 κ = 0.25 u(x) 1.5 κ=1 1.0 κ = 2.5 0.5 0.0 −1.0 κ=5 −0.5 0.0 0.5 x 1.0 Figure 4.2: Mesa profiles for various values of κ, obtained by numerically solving (4.1). The other parameters used are D = 10 and L = 1. The system was integrated using an implicit explicit scheme on a 500 point grid. steady state problem ut = vt = 0: u = U0 (y) + ε U1 (y) + . . . , v = V0 (y) + ε V1 (y) + . . . , y= x−l . ε To O(1), since Du = O(ε 2 ), the system is U0 = U0 − U02 V0 (1 + kU02 ) V0 = 0. From the boundary conditions we have that V0 is a constant. To determine its value, we use the Maxwell line condition [31], which states that in order to have a heteroclinic connection between x = 0 and x = L, there has to be a value vc , such 2 u that the area under the roots of a(u, vc ) = −u+ vc (1+ku 2 ) is zero. The three roots of 136 a(u, v) occur at u = 0 and at u = u± (v) = √ 1± 1−4kv 2 , 2kv The non-zero roots can be expressed as v = h(u) = with 0 < u− (v) < u+ (v). u 1+ku2 for 0 < v < vm , with v = vm the value at which both roots coalesce. Hence, the solution to the O(1) equation is V0 = vc , with vc the value that satisfies the Maxwell line condition uc a(u, vc ) = 0, 0 uc ≡ u+ (vc ), (4.11) and U0 the unique heteroclinic connection satisfying U0 (−∞) = u+ (vc ) = uc and U0 (∞) = 0. To O(ε ) we have the system L(U1 ) = U1 + au (U0 , V0 )U1 = −av (U0 , V0 )V1 V1 = 0. The solution to the second equation is V1 = V11 y + V12 . Since L(U0 ) = 0, we can derive a solvability condition to determine V12 in terms of V11 , ∞ V12 −∞ ∞ av (U0 , vc )U0 dy = V11 −∞ av (U0 , vc )yU0 dy. Furthermore, matching to the outer solution yields that V11 = v (l± ). There are two outer solutions, to the left and right of the internal layer at x = l, that when matched will determine the value for v (l± ). The problem to the left of the layer, 0 < x < l, is Dv v = b(u, h(u)) = g(u), v(l) = vc , v (0) = 0, u = u+ (v), (4.12) whereas the outer problem on the right side of the layer, l < x < L, is defined as Dv v = b(0, v), v(l) = vc , 137 v (L) = 0. (4.13) Since V11 is a constant, the matching condition is that v (l− ) = v (l+ ) Multiplying (4.12) by v = h (u)u and integrating yields the following equations for ux and vx , dv =− dx with F (u; u0 ) = u u0 2F (u; u0 ) , D du =− dx 2F (u; u0 ) , Dh (u) g(s)h (s)ds. Upon integrating between u0 = u(0) and u(l) = uc , we obtain the following relationship: l −√ = D uc u0 h (u) 2F (u; u0 ) du = 2F (uc ; u0 ) + g(uc ) uc u0 g (u) (g(u))2 2F (u; u0 )du. (4.14) The exact solution on the outer region given in (4.13) can be calculated analytically as √ cosh (L − x)/ D , v(x) = vc √ cosh (L − l)/ D √ vc v (l+ ) = − √ tanh (L − l)/ D . D (4.15) Finally, since v (l+ ) = v (l− ), we can solve for l in (4.15) and substitute it into (4.14) to obtain the critical points at which the domain will split, L √ = tanh−1 D 2F (uc ; u0 ) − vc 2F (uc ; u0 ) − g(uc ) uc u0 g (u) (g(u))2 2F (u; u0 )du. (4.16) From a practical point of view, the uc , vc values can be found using a quadrature, and a simple numerical integration will then yield the three terms in (4.16). In the next section we will calculate this threshold and compare it with numerical results of systems with growing domains. The remainder of this chapter will consist of numerical experiments based on numerical bifurcation theory. 138 4.3.2 Bifurcation analysis In the previous section we derived a solution for the D = O(1) case by treating the mesa solution as two back-to-back heteroclinic curves. Equation (4.16) determines a critical value in terms of L or D, beyond which the heteroclinic cannot exist. This provided us with an analytical estimate of the domain length at which splitting or self-replication will occur. We now want to study the splitting behaviour as the domain increases both adiabatically (when stationary solutions are recomputed at each domain length value), and dynamically (when the domain length itself is a dynamic variable). We start by focusing on how solutions change numerically as the domain length increases. The boundary conditions were homogeneous Neumann, and the values of the constants were k = 2.5, τ = 1, D = 1, and ε = 0.02. In the adiabatic growth case, numerical solutions to the full 1-D problem were computed for increasing values of L. An initial solution was iterated until convergence for some domain length L = L0 ; the stationary solution thus obtained then became the initial solution for an increased domain length L = L1 , and so forth for increasing values of L. Typical solutions for three different values of L are given in figure 4.3, while on the two-mesa regime. The three solutions all occur for varying values of the domain length. They were all obtained through numerical continuation on the L parameter. We started by obtaining a stationary solution from random initial data, and from it we used the numerical continuation package AUTO-07P [14] to follow the curve of solutions for varying L, using max(V ) as the bifurcation parameter. Solution I is essentially the leading eigenvector φ = A cos(2nπx/L), first estimated through linear stability analysis, and as expected from Turing theory, unstable. Going up on the branch beyond solution I leads to the homogeneous Turing solution u = 0.5603, v = u2 , and traversing the branch in the other direction leads to the stable mesa branch. 139 Figure 4.3: Three distinct two-mesa solutions to the GMS system. Solution I is close to the Turing instability, II is the stable mesa solution, and III is the unstable solution that develops when the domain length is increased past a critical point. The image on the right is the bifurcation diagram for the branch of two-mesa solutions. Solution II lies in the stable branch. This is a characteristic mesa structure, and it is far from the Turing equilibrium. Continuing on the branch along increasing L eventually leads to a fold point, beyond which we reach the unstable branch characterized by solution III. Going beyond the fold point causes the solution to drop to the next branch of solutions, which will have double the number of mesas. The bifurcation diagram for one, two, four, and eight mesa solutions is given in figure 4.4. The thickest line is a stable solution that is recomputed for increasing values of L. It traverses the stable branches from left to right, and at each fold point it falls to the next branch, which manifests in the solutions as a doubling in the number of mesas. Since each successive branch doubles the number of mesas, the critical value Lc at which the new set of mesas will split doubles with each iteration, hence Lc (n) = Lc (1) × 2n−1 . This exponential relationship can be readily seen in the symmetry of the bifurcation diagram of figure 4.4, which was plotted on a logarithmic scale. This numerical results show that there will always be a stable solution in the (D, τ ) = O(1) parameter regime, for all domain lengths. The asymptotic formula (4.16) provides us an estimate of the critical values Lc (n) at which an n mesa 140 0.315 max(V) 0.310 0.305 0.300 0.295 0.290 100 L 101 Figure 4.4: Four branches of the GMS system, with an overlay of the family of stable solutions obtained by traversing it left to right . When reaching the fold point of each branch the solutions fall to the next branch, effectively doubling the number of mesas. The upper horizontal unstable line are the unstable Turing solutions, and the red points on it are the values shown on table 4.1. solution splits into 2n mesas. In order to numerically compute the value, we first found the uc , vc values that satisfy the Maxwell line condition, shown in (4.11), via a quadrature. It was then straightforward to numerically integrate F (uc ; u0 ) and the third term in (4.16). The resulting value was Lc = 2.1010 for −L < x < L, or half that for 0 < x < L (as shown in Figures 4.3, 4.4). The location of the fold point in the 1-mesa branch was then calculated by solving the full system, using ε = 0.002 and 1500 grid points. The value thus obtained was Lc = 2.1325. Furthermore, the system exhibits hysteresis, traversing the bifurcation diagram left to right produces a very different picture. Traversing left to right shows the splittings occurring at the points predicted in (4.16), whereas traversing in the opposite direction results in in the solution staying in the 8-mesa branch until the left edge of the stable branch, beyond which the solution will jump either to the 4141 mesa branch, or to the 2-mesa branch. In principle, estimating those points should involve very similar analysis as was done in Section § 4.3.1. The differences between the bifurcation structure of u(x) when traversed in either direction are shown in Figure 4.5. Figure 4.5: The full solution curve for u(x) as the bifurcation branches in Figure 4.4 are traversed from left to right (image on the left), and from right to left (image on the right). The solutions are all plotted on a normalized domain, and the proper domain length L is represented on the y-axis. All of the work above was done with the domain length L fixed; when a stable solution was attained it was used as the starting point for the next computation at a slightly changed L. When considering dynamically varying domains, i.e., making L ≡ L(t), the system needs to be modified. A general framework on how to extend stationary equations into growing domains was discussed earlier in § 4.2. The extended system for isotropic exponential growth is relatively simple, ε2 u2 ∆u − ρu − u + L2 v(1 + ku2 ) D τ vt = 2 ∆v − ρv − v + u2 L Lt = ρL, ut = with ρ the rate of domain growth. 142 We solved the system by discretizing in space using centred differences, with second order boundary conditions, and Matlab’s ode15s routine was used for time integration. The solution curves for various values of ρ were obtained, and when overlapping them in the bifurcation diagram of Figure 4.4 we can see a delayed bifurcation effect in the left image of Figure 4.6. Notice the sharp transitions of the static solutions in Figure 4.5, compared to the soft dynamic transitions on the right image in Figure 4.6. The stability of the branches highlighted in the above Figure 4.6: Solution curves for systems with growing domains, L(t) = eρt . Notice the delay in the bifurcation (jump between branches) as ρ gets larger. The figure on the right shows the effect on Figure 4.5 (left) when adding domain growth, with ρ = 0.002. The y-axis represents L(t), and L = 15 is reached when t = 1354. figures had to be obtained independently from the continuation software used to generate the bifurcation the diagram. The package that we used (AUTO-07P [14]) is designed primarily for finite dimensional systems; stability can be established in a straightforward way for systems of ordinary differential equations (ODE). When extended to dealing with parabolic partial differential equations (PDE)s this capability is lost. We performed the eigenvalue analysis separately by saving the stationary solutions obtained while traversing the branches, and using them as the basis of a Taylor expansion on a perturbation to the solutions. The algorithm in AUTO utilizes non-uniform grids, a spline on the output was used to generate a uniformly spaced grid amenable with the discrete Jacobian. 143 In order to study stability we introduce the perturbation u(x) = us (x) + eλt φ(x), v(x) = vs (x) + eλt ψ(x), with us , vs the stationary solutions and φ 1, ψ (4.17) 1. Substituting this into (4.1), we get λφ = ε 2 ∆φ + au (us , vs )φ + av (us , vs )ψ λτ ψ = D∆ψ + bu (us , vs )φ + bv (us , vs )ψ. This can be written as an eigenvalue problem in matrix form, Aw = λw, with w = (φ, ψ)T , and A as A= ε 2 ∆ + au (us , vs ) av (us , vs ) bu (us , vs )/τ D∆/τ + bv (us , vs )/τ . (4.18) Plotting the eigenvalue with largest real part versus the corresponding L value for the stationary solutions reveals the stable and unstable manifolds in the branches. In Figure 4.7 we show such a curve for the full range of stationary solutions along the one-mesa branch. The labelling regarding stability on all the previous figures was based on this calculation, and due to the symmetry of the system, the curves for the different branches are essentially identical. The region in Figure 4.7 where the eigenvalues have negative values is roughly a straight line of magnitude O(ε ). This was expected from the estimates done on [24]. 4.3.3 Hopf bifurcations of 1D mesa patterns So far we have examined the instabilities that occur for increasing L, which typically result in the splitting of solutions. A second mechanism for generating instabilities is when the value of τ increases. In all previous calculations we’ve worked with τ = O(1), and the difference in difusivities marks the v(x) equation 144 0.3 0.25 max(real(λ)) 0.2 0.15 0.1 0.05 0 −0.05 0 0.5 1 1.5 2 2.5 L Figure 4.7: Stability curve of the maximum eigenvalue vs L for solutions on the 1-mesa branch. as the fast component in the system. However, if τ grows to be O(1/ε ), or even to O(1/ε 2 ), the v equation will slow down and approach the time scale of the u equation. When this happens, feedback will happen between the two equations, and phenomena characteristic of delay differential equations such as oscillatory instabilities will occur. During oscillations it is even possible that the width of the mesa might exceed the splitting threshold, which would result in a mesa splitting bifurcation. We start by doing a perturbation analysis with τ as a parameter, similar to what was done in (4.18). We start at a stationary solution on the stable part of the one-mesa branch, and we vary τ until we find a zero eigenvalue, using Newton’s method. The critical value for a stable solution originating at L = 2.02 is τ = 278. Plotting the eigenvalues with largest real part as τ increases shows how a set of complex conjugate eigenvalues becomes unstable, as depicted in Figure 4.8. The type of instability generated by large values of τ are oscillations on the 145 0.05 0.04 0.03 0.02 0.01 0 −0.01 −0.02 −0.03 −0.04 −0.05 −0.05 −0.04 −0.03 −0.02 −0.01 0 0.01 Figure 4.8: Real part of eigenvalues as τ approaches critical value. The x and y axis are the real and imaginary components of the eigenvalues, respectively. width of the mesa. We computed full numerical solutions for two different cases beyond the Hopf threshold, and these can be seen in Figure 4.8. In the first case, the domain length of the original stationary solution was small enough, that even with τ beyond the threshold the mesa doesn’t split. The complex pairs have largest real part, so stable oscillations develop. This type of instability is called a breather type instability, and we will study it in more detail when we discuss systems with D = O(1/ε ) in § 4.4. The second case involves a single mesa on a slightly larger initial domain. When τ grows, the oscillations bring it beyond the splitting threshold, and this causes the mesa to split into two mesas. The two new mesas now have domain lengths that are half the length of the original one, so they quickly stabilize. This case highlights a dynamic splitting bifurcation. 146 Figure 4.9: Solution graphs for both u(x) (top), and v(x) (bottom), for large τ . Both numerical computations were done for τ = 380, the ones on the left with a domain length L = 1.6 and the ones on the right had L = 2.02. The horizontal axis is time, and the vertical axis is the domain length. 4.4 Mesa patterns in the near-shadow limit When considering the GMS model with a diffusion coefficient D = O(1), we have shown in the previous sections numerical results highlighting the existence of both splitting and oscillatory instabilities, and an instance of an oscillatory-triggered splitting instability (Figure 4.9). We now want to extend the results obtained previously to the near-shadow limit, i.e., D = D/ε with D = O(1). We will show that this limit is more tractable analytically. We will observe that the behaviour that we observed in the numerical experiments of the previous section is amenable to asymptotic analysis when working in this regime. We want to construct a K−stripe stationary solution on x ∈ [0, 1], with L = 1/K the period of the solution, and l the length of each individual mesa, as shown in Figure (4.10). 147 χl 0 χr L l Figure 4.10: A typical mesa profile in the stationary solution v(x). The left and right edges of the mesa are labelled as χl and χr respectively; and the length of the mesa section is l. The stationary system 0 = ε 2 uxx − u + u2 , v(1 + ku2 ) D 0 = vxx − v + u2 , ε ux (0) = ux (L) = 0, (4.19) vx (0) = vx (L) = 0. Beyond the stability of 1D mesas, we will extend the mesas in the y-direction to form stripes, and derive conditions on the parameters to guarantee stability with respect to transverse perturbations. 4.4.1 Construction of a multi-stripe pattern To leading order, we have that vxx = 0. Applying the Neumann boundary condition, we have then that v ∼ V, and the value of the constant can be estimated by 148 integrating over the whole domain, V= L 1 L u2 dx. (4.20) 0 In the inner region near the left boundary of the mesa we have that v = V, and we do a change of variables for u = Vw and y = ε −1 (x − χl ). The resulting equation is −∞ < y < ∞, wyy + f (w) = 0, f (w) = −w + w2 , 1 + bw2 (4.21) with b = kV 2 . Now, we are looking for a heteroclinic connection in u as the transition mechanism that generates the mesa, one for each side of the mesa. For a heteroclinic connection to exist in (4.21), it has to satisfy the Maxwell line condition [31]. 1 f(w) 0.5 0 w− −0.5 −0.5 0 0.5 1 w+ 1.5 2 2.5 3 3.5 4 w Figure 4.11: A plot of the function f (w) given in (4.21). 149 The function f (w) = 0 has zeros at w = 0 and w± = √ 1± 1−4b , 2 with distinct real values for w± existing in the range 0 ≤ b < 1/4. The profile of the curve in that range can be seen in Figure (4.11). The Maxwell line condition states that a heteroclinic connection will exist for the value b = b0 such that w+ 0 f (w)dw = 0. Integrating f (w) we get w+ − f (w)dw = 0 =− and since we have that b0 = b0 = w+ − 1 , 2 w+ w2 w 1 + − 3/2 arctan(b1/2 w) 2 b b w+ , 0 2 w+ w+ 1 1/2 + − 3/2 arctan(b0 w+ ), 2 b0 b0 w+ −1 2 , w+ the Maxwell line condition will be satisfied if w+ − 1(w+ + 1) = 2w+ arctan( w+ − 1). (4.22) This can be solved numerically to obtain the critical values b0 = 0.211376, and w+ = 3.295209. For use later in Section §4.4, we need to compute ∞ β= w 2 dy −∞ Next, we multiply (4.21) by w to get d dy with F = − w 0 w2 2 = d F(w) dy w 2 = 2F, → f (s)ds. We then obtain ∞ β= w+ w 2 dy = −∞ 0 w2 1 dw = w w+ 2F(w)dw. (4.23) 0 This can be numerically calculated using the previously computed value for w+ , to get β ∼ 1.49882. 150 Linearizing around w = 0 for y → −∞, and for w = w+ for y → ∞, we get that for b = b0 we have a heteroclinic solution −∞ < y < ∞, w + f (w) = 0, w ∼ d− ey , −ν+ y w ∼ w+ − d+ e for ν+ = f (w) = −w + w2 , 1 + b0 w2 as y → −∞, as y → ∞, , 1 − 2/w+ . To break translation invariance we take w(0) = w+ /2, in order to guarantee uniqueness. A full mesa solution will consist of two back-to-back heteroclinic curves, and can be constructed as u ∼ V[wl + wr − w+ ], with wl ∼ w x − χl ε Integrating (4.20), we get that to first order V ∼ the width of the mesa. We then have, since b0 = 2 Vw+ L ∼ + O(ε ), l kV 2 , , wr ∼ w 1 2 2 L V w+ l, that χr − x ε . with l = χr − χr √ L k l∼ √ < L. 2 b0 w+ Therefore, a necessary condition for a K−stripe solution to exist is that √ √ k < 1. 2 b0 w+ To refine the solution, it is necessary to further expand u(x) and v(x). In the outer region we expand u and v as v ∼ V + ε v1 + ε 2 v2 + · · · . 151 Since outside of the mesa u is exponentially small, and in the plateau region u ∼ Vw+ + O(ε ), by substituting into (4.19), we get that Dv1,xx V for 0 < x < χl 2 ) = V(1 − L/l) for χ < x < χ = V(1 − Vw+ r l V for χr < x < L (4.24) with v1,x (0) = v1,x (L) = 0. In order to find the conditions on v1 at the transition layers χl and χr , we expand u as u ∼ Vw+ + ε U1 + ... on χl < x < χr . Substituting into (4.19), we get that −U1 + gu (Vw+ , V)U1 + gv (Vw+ , V)v1 = 0, with g(u, v) = u2 v(1 + ku2 ) 2 = w , the linearization terms simplify to Since 1 + b0 w+ + gu (Vw+ , V) = 2w+ 2 2 )2 = w , (1 + b0 w+ + gv (Vw+ , V) = 2 −w+ 2 = −w+ , 1 + b0 w+ (4.25) and this yields −U1 + 2 U1 − w+ v1 = 0, w+ ⇒ U1 = 2 w+ v1 . 2 − w+ We now expand u and v near the left transition point x = χl (zooming in y = ε −1 (x − χl )) as u = u0 + ε u1 + ε 2 u2 + · · · , v = V + ε V1 + ε 2 V2 + · · · , 152 with u0 = Vw+ . Substituting into (4.19), the O(ε ) system is Lu1 : ≡ u1 − u1 + gu (u0 , V)u1 = −gv (u0 , V)V1 (4.26) V1 = 0. For V1 we get that V1 = V10 + V11 y. We also have that Lu0 = 0. The solvability condition on (4.26) is that ∞ −∞ ∞ gv (u0 , V)V1 u0 dy = −∞ w2 w V1 dy = 0. 1 + bw2 (4.27) Substituting in V1 = V10 + V11 y, the condition from (4.27) implies that ∞ V10 −∞ w2 w dy + V11 1 + bw2 ∞ −∞ w2 w y dy = 0. 1 + bw2 Since V1 = O(ε ), V11 has to be zero, as otherwise for |y| 1 we would have V1 = O(1). Consequently, V10 also has to be zero, and the conclusion is that V1 = 0. We have then that Lu1 = 0, and therefore we also have u1 = 0. This result yields that v1 (χl ) = v1 (χr ) = 0, and now we have enough conditions to solve uniquely for v1 . is 2 = L/l, the full solution to (4.24) Using the fact that χr − χl = l, and that Vw+ v1 = V 2 2D (x − χ2l ) V(l−L) 2 2Dl [(x − χl ) V 2 2D [(L − x) − for 0 < x < χl − l(x − χl )] for χl < x < χr (L − χr )2 ] 153 for χr < x < L It is possible now to calculate v1,x close to the transition layers. We have Vχl , D V(L − l) v1,x (χ− , r )=− 2D V(L − l) , 2D V(L − χr ) v1,x (χ+ , r )=− D v1,x (χ− l )= v1,x (χ+ l )= (4.28) This suggests that there is a term to next order, as V2 ∼ v1,x (χl ). Expanding to next order in the inner region, as u = u0 + ε 2 u2 and v = V + ε 2 V2 , and defining g0 (w) = w2 , 1+b0 w2 we have the system Lu2 = u2 − u2 + g0 (w)u2 = g0 (w)V2 , V2 = 0. We have then that V2 = V20 + yV21 . We can derive a solvability condition from Lu2 = g0 (w)V2 , since Lw = 0. We get ∞ −∞ ∞ V2 g0 (w)w dy = (V20 + yV21 )g0 (w)w dy = 0, −∞ We can now match the inner solution to the outer solution evaluated at the interface, to determine V20 and V21 . We get that 2 V + ε 2 (V20 + yV21 ) + · · · = V + ε v1 (χl ) + ε v1,x (χ− l )(x − χl ) + ε v2 (χl ) + · · · From this we can conclude that V21 = v1,x (χ− l ), and that V20 = v2 (χl ). This last value, for both the inner and outer solutions, can be calculated from the solvability condition given that we now know V21 . Furthermore, repeating the matching procedure with x → χ+ l yields the same value, as the outer solution has no ambiguity. From this we get that v1,x (χ+ l ) = 154 v1,x (χ− l ). Repeating the procedure yet again at the right boundary χr , we get the − same result, i.e., v1,x (χ+ r ) = v1,x (χr ). Since we already knew from (4.28) that − v1,x (χ+ l ) = −v1,x (χr ), we can solve for χl and χr to get that the position of the boundaries of the mesa on [0, L] are χl = L−l , 2 χr = L+l , 2 with l the length of the plateau. A corollary from this result is that stationary mesa solutions have to be centred. We can find a second solvability condition that will be of use later on. Differentiating with respect to y, we have Lu2 = −g0 (w)u2 w + g0 (w)V2 + g0 (w)V2 w , and using the fact that Lw = 0, the solvability condition we get is ∞ −V21 ∞ g0 (w)w dy = −∞ ∞ v1,x (χr ) −∞ ∞ g0 (w)w dy = −∞ −∞ (g0 (w)V2 − g0 (w)u2 )w 2 dy, (4.29) 2 (g0 (w)V2 − g0 (w)u2 )w dy, since V2 = V21 = −v1,x (χr ). 4.4.2 Transverse stability in the near-shadow limit to perturbations in the y direction We will now extend the solutions obtained in the previous section for a 1D mesa to R2 by extending the mesas along the y-direction to form stripes. We will study the stability to perpendicular perturbations, which can give rise to buckling in the solutions. We assume that the solutions exist in a rectangular domain [0, 1] × [0, d0 ], with 155 Neumann boundary conditions on all sides. We introduce a perturbation on the equilibrium solution (ue , ve ) of the form u = ue + eλt+imy φ(x), with |φ| 1, and |ψ| v = ve + eλt+imy ψ(x); m= kπ , k = 1, 2, . . . , d0 1. Substituting into (4.1), we get the following eigenvalue problem ¯ = Lε φ + gv (ue , ve )ψ = ε 2 φxx − φ + gu (ue , ve )φ + gv (ue , ve )ψ, (4.30a) λφ ε 2ε (1 + τ λ)ψ = ψxx − m2 ψ + ue φ, (4.30b) D D ¯ = λ+ε 2 m2 , and Neumann conditions φx (0) = φx (1) = ψx (0) = ψx (1) = with λ 0. As shown in (4.25), in the plateau region we have that gu (ue , ve ) = 2/w+ , and gv (ue , ve ) = −w+ . Substituting into (4.30a), we have that, to first order and when ¯ λ 1, the asymptotic form of φ on the plateau region is φ = µψ, with µ ≡ 2 w+ , 2 − w+ χl < x < χr . Near the boundary region φ is asymptotically small, therefore, near the transition layers located at χl , χr , φ is proportional to the derivative w of the heteroclinic connection. We have then the following asymptotic form for φ −1 for x ∼ χli cli (w (ε (x − χli )) + O(ε )) φ ∼ cri (w (ε −1 (x − χri )) + O(ε )) for x ∼ χri φi = µψ for x ∈ (χli , χri ), i = 1, . . . , K, with the constants cli , cri to be found. Since φ is localized near the transition layers, we can estimate it in the sense 156 of distributions, approximating ue as ue ∼ Vw 2ε ue φ ∼ D K 2ε 2 Vcl D i=1 + 2ε 2 Vcr D ∞ −∞ wl wl dyδ (x − χli ) ∞ −∞ wr wr dyδ (x − χri ) + 2ε V w+ µψH[χli ,χri ] , D with H[χli ,χri ] = 1 on x ∈ (χli , χri ), and zero elsewhere. This then yields 2ε ue φ ∼ D K i=1 2 2 ε 2 Vcr w+ ε 2 Vcl w+ δ (x − χli ) + δ (x − χri ) D D + 2ε Vw+ µψH[χli ,χri ] D Substituting into (4.30b), we get that ψ satisfies ψxx − θ 2 ψ = − 2 ε 2 Vw+ D i (cli δ (x − χli ) + cri δ (x − χri )) , (4.31) with θ the piecewise constant function θ = θ − ≡ m2 + θ + ≡ m2 + ε D ε (1+τ λ) D 1 + τλ + 1/2 , 2w+ l(w+ −2) 1/2 for x ∈ / ∪K i=1 [χli , χri ] , for x ∈ ∪K i=1 [χli , χri ] (4.32) Since w is localized, we can define wli = w (x − χli ) and wri = w (x − χri ), and multiply it into (4.30a), to obtain the matrix eigenvalue problems ¯ cli (wli , Lε wli ) + (wli , gv (ue , ve )ψ) = cli λ(w li , wli ), ¯ cri (wri , Lε wri ) + (wri , gv (ue , ve )ψ) = cri λ(w ri , wri ), where (f, g) = 1 0 f gdx. 157 (4.33a) (4.33b) The second term in (4.33a) and (4.33b) can be readily estimated, using the fact w2 , 1+b0 w2 that w − w = g0 (w), and that gv = −g0 , with g0 (w) = K 1 (wli , gv (ue , ve )ψ) = 0 wli ψgv (ue , ve )dx = −ε K = −ε ∞ i=1 −∞ w g0 (w)dy ψ(χli ) −∞ i=1 K ∞ ψ(χli ) as w (w − w)dy = −ε ψ(χli ) i=1 2 w+ , 2 and similarly, K (wri , gh (ue , ve )ψ) = −ε ψ(χri ) i=1 2 w+ . 2 The third term can be estimated straight from the definition of β in (4.23). We get ∞ (wli , wli )) ∼ ε (w )2 dy = ε β. −∞ The first term can be estimated using some of the results previously obtained. We have that Lε wl = (wl ) − wl + gu (ue , ve )wl . We can approximate gu (ue , ve ) as gu (ue , ve ) ∼ gu (wV, V) + ε 2 (guu (wV, V) + guv (wV, V)) + · · · . 158 The derivatives can be related to g0 (w) = w2 1+b0 w2 in the following way: 2w 2u ∼ = g0 (w), v(1 + ku2 )2 (1 + b0 w2 )2 1 2 − 6ku2 1 2 − 6b0 w2 1 guu (u, v) = ∼ = g0 (w), 2 3 2 3 v (1 + ku ) V (1 + b0 w ) V 2w 1 2u ∼− = − g0 (w). guv (u, v) = − 2 v (1 + ku2 )2 V(1 + b0 w2 )2 V gu (u, v) = Substituting them in, we get gu (ue , ve ) ∼ g0 (w) + Lε wli ∼ ε2 (g (w)u2 − g0 (w)V2 ) + · · · , V 0 ε2 (g (wli )u2 − g0 (wli )V2 )wli . V 0 We can now express the first term as (wli , Lε wli ) ∼ ε2 V = 1 g0 (wli )u2 − g0 (wli )V2 wli2 dx 0 ∞ ε3 V −∞ g0 (wli )u2 − g0 (wli )V2 wli2 dy Using the solvability condition in (4.29), ∞ V2 ∞ g0 (w)w dy = −∞ −∞ (g0 (w)V2 − g0 (w)u2 )w 2 dy, we get (wli , Lε wli ) ∼ ε3 V V 2 ∞ ε3 V V 2 −∞ 2 ε 3 v1x (χli )w+ ε 3 V2 2 = w+ = 2V 2V ∞ g0 (w)w dy = 159 −∞ (w − w )w dy Similarly, on the other side of the plateau the process is identical, except for a sign change in the slope, (wri , Lε wri ) ∼ − 2 ε 3 v1x (χri )w+ 2V Putting everything together results in the following 2K × 2K system ε3 ε 2 2 cli v1x (χli )w+ − ψ(χli )w+ , 2V 2 3 ¯ ri β ∼ − ε cri v1x (χri )w2 − ε ψ(χri )w2 , ε λc + + 2V 2 ¯ li β ∼ ε λc and since from (4.28) we know that v1x (χli ) = v1x (χri ) = V(L − l) , 2D the above system can be simplified to 2 2 ¯ li ∼ ε 2 (L − l)w+ cli − w+ ψ(χli ), λβc 4D 2 2 (L − l)w w2 + ¯ ri ∼ −ε 2 λβc cri − + ψ(χri ). 4D 2 (4.34) ¯ and c = [cli , cri ]. This equation, together with (4.31) constitutes a system for λ The system given in (4.34) depends on ψ(χri ) and ψ(χli ). Solving (4.31) explicitly, we get ψ(x) = cosh(θ − x) ψ(χl1 ) cosh(θ , − χl1 ) sinh(θ + (χri −x)) sinh(θ + (x−χli )) ψ(χli ) sinh(θ + ψ(χri ) sinh(θ , + (χri −χli )) + (χri −χli )) sinh(θ − (χl(i+1) −x)) sinh(θ (x−χ − ri )) ψ(χri ) sinh(θ− (χ −χri )) + ψ(χl(i+1) sinh(θ− (χl(i+1) −χri )) , l(i+1) cosh(θ − (1−x)) ψ(χrK ) cosh(θ , − (1−χrK )) for 0 < x < χl1 for χli < x < χri for χri < x < χl(i+1) for χrK < x < 1 (4.35) 160 for i = 1, . . . , K. We have that χri − χli = l; similarly, we define d ≡ χl(i+1) − χri = and the constants cl = cosh(θ + l), sl = sinh(θ + l), cd = cosh(θ − d), sd = sinh(θ − d) 1 K − l, Additionally, the jump conditions that solution (4.35) has to satisfy are given by 2 ε 2 Vw+ cli ≡ bli , D 2 ε 2 Vw+ − cri ≡ bri , −[ψx (χ+ ) − ψ (χ )] = x ri ri D − −[ψx (χ+ li ) − ψx (χli )] = (4.36) for i = 1, . . . , K. This results in a linear system, with bli defined as c2d θ+ cd cl − ψri + ψli θ − + θ + sd sl sd sl θ− θ+ cd cl = −ψr(i−1) − ψri + ψli θ − + θ + , sd sl sd sl bli = ψr(i−1) θ − sd − and similarly, bri = −ψl(i+1) θ− θ+ cl cd − ψli + ψri θ + + θ − sd sl sl sd , for i = 2, . . . , K − 1. The values at the boundaries are slightly different and have to be derived separately. Using the identity sinh(x/2) cosh(x) − 1 = , cosh(x/2) sinh(x) 161 we finally have that θ+ cd cl θ− − ψr1 − +θ+ sd sd sl sl cd cl θ+ θ− θ− − +θ+ . − ψlK sd sd sl sl bl1 = ψl1 θ − brK = ψrK We can now write the 2K ×2K system of equations in matrix form as M ψ = b, with M the tridiagonal matrix a+c b b c a a c b . , . M = . b c a a c b b c+a and where a=− θ− , sd b=− θ+ , sl c= cd cl θ − + θ +. sd sl The eigenpairs of M can be found explicitly (appendix B in [48]), and for the reader’s convenience we will reproduce the calculation. 162 The M matrix can be simplified into M = Q + cI, with a b b 0 a a 0 b . . . Q= . b 0 a a 0 b b a We use the property that eig(M ) = c + eig(Q). We start by looking for an eigenvector q = [z, tz 2 , z 3 , tz 4 , . . . , tz 2K ]T , with t, z ∈ C, |z| = 1, and corre- sponding eigenvalue σ. From the second equation to the second to last we get the following system, atz 1−l + btz 1+l = σz l , if l is odd, bz 1−l + az 1+l = σtz l , if l is even. (4.37) Since z z¯ = 1, hence 1/z = z¯, we can solve for t in the above system to get t=± az + b¯ z , |az + b¯ z| which implies that |t| = 1. In order to satisfy the first and last equations, we look at the extended system ¯ = σ(Ah+B h), ¯ with h = [t, q, tz 2K+1 ]T . The first and last equations Q(Ah+B h) are a(At + B t¯) + b(Az + B z¯) = σ(At + B t¯), b(Atz 2K + B t¯z¯2K ) + a(Az 2K+1 + B z¯2K+1 ) = σ(Az 2K+1 + B z¯2K+1 ). 163 The first and last equations will then be satisfied if At + B t¯ = Az + B z¯, Atz 2K + B t¯z¯2K = Az 2K+1 + B z¯2K+1 . Nontrivial solutions for A and B exist if (z − t)(1 − t¯z¯)¯ z 2K = (¯ z − t¯)(1 − tz)z 2K is satisfied. Since |z| = |t| = 1, we have then that z 4K = 1. If we write z in polar form, we have that the roots of z 4K = 1 are z = ei 2π(j−1) 2K , with j = 1, 2, . . . , 2K. Substituting it into (4.37), we get σ = ±|az + b¯ z | = ± a2 + b2 + 2ab cos(θ ), with θ = 2π(j−1) 2K . Since we have both positive and negative values, we are count- ing twice when ranging j = 1, . . . , 2K; therefore the range can be restricted to obtain the following set of distinct eigenvalues σj± = ± a2 + b2 + 2ab cos θ j , with θ j = πj , K j = 1, . . . , K − 1, σK± = a ± b. Going back to (4.36), we can express the jump condition as the following system, ψ = M −1 b = 2 ε 2 Vw+ M −1 c, D and we can use it to substitute ψ in (4.34), resulting in the system ¯ = ε2 λβc 2 4 (L − l)w+ Vw+ I− M −1 c. 4D 2D (4.38) ¯ = λ + ε 2 m2 , and M −1 c = σ Since λ ˆ −1 c, with σ ˆ = c + σ, we have then that 164 in the limit when ε → 0, 2 4 (L − l)w+ Vw+ λj± 2 = −m + − σ ˆ −1 , ε →0 ε 2 4Dβ 2Dβ j± lim j = 1, . . . , K. To establish the stability of the system, we want to establish conditions that guarantee that the eigenvalues will be negative, hence stable. The largest eigenvalue corresponds to the largest σ ˆ value; since both a and b are negative, the largest σ values corresponds to j = 1; and as the number of mesas increases the largest eigenvalue tends to c + |a + b|. On the other end of the spectrum, we have that the smallest eigenvalue is always positive, c+a+b = c−|a+b| = cd θ − cl θ + θ − θ + sl θ − (cd − 1) + sd θ + (cl − 1) + − − = , sd sl sd sl sd sl since cosh(x) > 1 for x = 0. From (4.32) we have that to leading order both θ − , θ + ∼ m. Let a ˆ= −1 , sinh(md) ˆb = −1 , sinh(ml) cˆ = coth(md) + coth(ml), the stability condition is, then 2 4 (L − l)w+ Vw+ < m2 + σ ˆ −1 4Dβ 2Dβ j± = m2 + 2 Lw+ 2Dβlm for all j = 1, . . . , K − 1, and m = kπ d0 , 1 cˆ ± a ˆ2 + ˆb2 + 2ˆ aˆb cos k = 1, . . .. 165 πj K , A sufficient condition for stability is as follows: if D> 2 d2 (L − l)w+ 0 4π 2 β then the K−stripe system will be stable. For the m = 0 mode, and to first order, we approximate a, b, and c as 1 a = − + O(ε ), d 1 b = − + O(ε ), l c= 1 1 + + O(ε ), d l Since σ ˆK+ = a + b + c would be zero to first order, we can approximate c to second order for that particular case, resulting in c= 1 1 1 + + (dθ 2− + lθ 2+ ) + O(ε 2 ). d l 2 We have then that 1 1 2 1 1 + ± + + cos(θ j ), d l d2 l2 dl 1 ∼ (dθ 2− + lθ 2+ ), 2 2 ∼ . l σ ˆj± ∼ σ ˆK+ σ ˆK− j = 1, . . . , K − 1, Similarly, we have that the eigenvalues λj± for the full system, for the m = 0 166 mode, are λK+ = ε 2 = −ε λK− = ε 2 λj± = ε 2 < ε2 2 2 (L − l)w+ Lw+ − σ −1 , 4Dβ 2Dβl K+ (4.39a) 2 Lw+ 1 2w+ −1 + O(ε 2 ) < 0, − βl K 2 − w+ 2 2 (L − l)w+ lw2 Lw+ l = −ε 2 + < 0, − 4Dβ 2Dβl 2 4Dβ 2 2 Lw+ −1 (L − l)w+ − σ , j = 1, . . . , K − 1, 4Dβ 2Dβl j± 2 2 (L − l)w+ Lw+ 2 2 − + 4Dβ 2Dβl d l = ε2 (4.39b) (4.39c) −1 , 2 2 (L − l)w+ Lw+ l(L − l) − 4Dβ 2Dβl 2L = 0, with (4.39a) being negative resulting from the previous numerical estimation w+ ∼ 3.30 in (4.22). This shows that for the mode m = 0, all the eigenvalues λj , j = 1, . . . , K are negative. Hence, a 1-D K−stripe mesa pattern with D = O(ε −1 ) is a stable solution to the GMS system. For the GMS system, we computed the eigenvalues of a two-mesa solution. Of the four possible eigenvalues, their corresponding eigenvectors show that the instabilities will lead to either two breather stripes, or two zigzag stripes (using the notation from [22]). In the Figures 4.12 we have the four eigenvalues as a function of the mode m. The most unstable pair corresponds to a zigzag-type instability, and the other two correspond to a breather instability. We also found the critical value where dλ dm = 0 and λ = 0 for one of the breather and zigzag eigenvalues. Using these results we simulated the full system on a 2D domain (−1, 1) × (0, d0 ). Varying the length of the domain d0 , we observed mode-one and mode167 Figure 4.12: Eigenvalues for a two mesa solution. The parameters are D = 0.5, ε = 0.001 and κ = 3 for the Figure on the left. The λ− eigenvalues are the zigzag ones, while λ+ are the breather ones. The Figure on the right has the critical (κ, D) values for instability. two instabilities develop over time (Figures 4.13). Since the mode m = kπ d0 , it Figure 4.13: Full 2D simulation with parameters ε = 0.01, D = 0.5, κ = 1.5. The solutions were integrated using an IMEX algorithm. The solution on the left has d0 = 1.5, and the solution on the right has d0 = 2. is immediate from the formula in (4.38) that increasing the domain width will destabilize the system. 168 Both solutions were integrated until T = 10, 000 from an initial condition that had previously converged to two mesa stripes, and that had a small amount of noise added. When integrating the same system for d0 = 1 no transverse instabilities were observed up to T = 20, 000. 4.5 Chapter summary In this chapter we analyzed mesa-type solutions to the GMS system. We started by analyzing the Turing-type solutions and considering the case of domain growth. We used numerical continuation to extend the Turing solutions to the fully nonlinear regime, where we observed the solutions split as the domain length was increased. We studied the cases of both a dynamically growing domain and adiabatic growth. We next constructed a matched asymptotic solution for the fully nonlinear regime by joining two heteroclinic solutions. Studying its stability we were able to corroborate the splitting thresholds observed numerically in the splitting regime. Upon extending the mesas into stripes on the plane, we were able to derive analytic thresholds for the stability to perpendicular perturbations of multiple mesas, and verified it with full numerics. In the next chapter we will extend the theory developed for the GMS system to general mesa systems, and apply the model to a Predator-Prey model. 169 Chapter 5 The stability of mesa stripes in general reaction-diffusion systems We have previously studied the stability of mesa stripes in the GMS model in both one and two dimensions, and considered the parameter regimes D = O(1) and D = O(ε−1 ). In this chapter we will extend this previous analysis and obtain analytical results for the stability of a mesa stripe pattern for a general reactiondiffusion system when D = D/ε , with D = O(1). At the end of the chapter we will apply the results to a predator-prey RDE. The general system that we will study is the following two-component set of reaction diffusion equations: ut = ε 2 ∆u + f (u, v) τ vt = D ∆v + g(u, v), ε (5.1) with homogeneous Neumann boundary conditions on x ∈ [0, 1] × [0, d0 ]. We 170 consider the limit where ε 1, and regard all the other constants as being O(1). Some conditions on f and g are needed for the existence of a mesa pattern (see below), 5.1 Construction of the solution in the near-shadow limit The following analysis is based on work done by Kolokolnikov and McKay [32]. We will use their results to extend the analysis done earlier for the GMS model to consider the transverse stability of a general stripe pattern. We want to construct a K−stripe stationary mesa solution on x ∈ [0, 1]. A mesa structure is characterized as a function u(x) ∼ u+ on −l < x < l, and u(x) ∼ u− on l < |x| < L; with u+ > u− , and both values joined by a sharp interface. The mesa pattern will be formed by two back-to-back interfaces. We will start by constructing a solution on [0, L], with the interface centred at x = l. A full mesa solution can then be constructed by adding an even reflection, and a K−mesa solution will simply be K copies, with 2K interfaces. The stationary equation we want to solve is 0 = ε 2 uxx + f (u, v), ux (−L) = ux (L) = 0, D vxx + g(u, v), ε vx (−L) = vx (L) = 0. 0= (5.2) To first order, we have vxx = 0. Applying the Neumann boundary condition, we have then that v ∼ V, and the value of the constant can be estimated by integrating over the whole domain. The resulting equation is (5.3) with v = V constant. 171 Now, we are looking for a heteroclinic connection in u as the transition mechanism connecting u = u+ to u = u− . This imposes the algebraic constraint that f (u+ , V) ≡ f+ = 0, and f (u− , V) ≡ f− = 0, which has to be satisfied together with the Maxwell line condition [31] u+ u− f (w, V)dw = 0. For both branches to be stable we also require fu (u± , V) < 0. Solving the algebraic system determines u± , and v0 = V. In the inner region near the interface of the mesa, we have that v ∼ V, and we do a change of variables for y = ε −1 (x − l), and u(x) ∼ U0 ( x−l ε ) . Integrating (5.3) in two parts across the interface yields the following result D ε D 0= ε l 0= l vxx + 0 0 L g(u, V)dx D vx (l− ) = −lg+ , ε D vx (l+ ) = (L − l)g− . ε ⇒ L vxx + l l g(u, V)dx ⇒ Since the v solution doesn’t have sharp interfaces, we obtain to leading order that l= g− L + O(ε ). g− − g+ (5.4) Furthermore, since 0 < l < L, we require that the following consistency condition be satisfied: 0< g− < 1, g− − g+ (5.5) Here, as with f± , we have defined g± ≡ g(u± , V). We will divide the half-mesa branch into three regions: the outer part on the mesa plateau, 0 < x < l; the outer part of the mesa beyond the plateau, x > l; and the internal layer around x = l bridging the two outer regions. To zoom into the inner layer we let y = 172 x−l ε , u(ε y + l) ≡ U (y), v(ε y + l) ≡ V (y), which when substituted into (5.2) result in Uyy + f (U, V ) = 0, Vyy + ε3 g(U, V ) = 0, D ∞ < y < ∞, U → U± as y → ∓∞, ∞ < y < ∞, V → V± as y → ∓∞. (5.6) In the outer region, 0 < x < l, we have f (u, v) = 0, ε vxx + g(u, v) = 0, D ux (0) = 0, vx (0) = 0. with the boundary conditions stemming from the even symmetry imposed on the mesas. Similarly, for the region x > l we have f (u, v) = 0, ε vxx + g(u, v) = 0, D ux (L) = 0, vx (L) = 0. (5.7) ε ε u1 +· · · , v = V + D v1 +· · · , Performing an asymptotic expansion u = u− + D and substituting into a Taylor expansion of f (u, v) in (5.7), we obtain f (u− , V) + ε − (f u1 + fv− v1 ) + · · · = 0. D u − Therefore, u1 = − ffv− v1 , where fv± = fv (u± , V), and fu± = fu (u± , V). u From (5.7) we also obtain that v1xx = −g− , v1x (L) = 0, on l < x < L, v1 (l+ ) = v1− , with g− = g(u− , V), (5.8) where we have imposed the boundary condition v1 (l+ ) = v1− in terms of an unknown constant v1− to be calculated later. 173 The solution in this region is v1 (x) = −g− 1 1 (x − L)2 − (L − l)2 2 2 + v1− . (5.9) Therefore, we have that in the outer region l < x < L u ∼ u− + ε D − fv− v1 (x) , gu− (5.10) ε v ∼ V + v1 (x). D Either by taking the derivative of(5.10), or integrating (5.8) over l < x < L we get that v1x (l+ ) ≡ v1− = g− (L − l). An analogous calculation in the outer region 0 < x < l yields u ∼ u+ + v∼V+ ε D − fv+ v1 (x) , gu+ ε v1 (x), D with v1 (x) = −g+ 1 2 l2 x − 2 2 + v1+ , again with a boundary condition v1 (l− ) = v1+ in terms of an unknown constant to be found. Threfore, for both sides of the interface, we have v1x (l− ) ≡ v1+ = −g+ l, v1x (l+ ) ≡ v1− = g− (L − l). (5.11) Taylor expanding both solutions near x = l± provides matching conditions for the inner solution. The problem for the inner layer, in terms of the variable 174 y = ε −1 (x − l), is Uyy + f (U, V ) = 0, −∞ < y < ∞, ε3 g(U, V ), D −∞ < y < ∞, Vyy = − ε fv± v1± as y → ∓∞, D fu± ε V ∼ V + (V1± + ε yV1± ) as y → ∓∞. D U ∼ u± − Expanding the inner solution, U = U0 + ε D U1 we get + · · · , V = V0 + ε D V1 + ···, L(U1 ) = U1yy + fU (U0 , V0 )U1 = −fV (U0 , V0 )V1 , V1yy = 0. The matching condition for V1 = h1 y + h2 is V1 ∼ v1± as y → ∓∞. Thus, we must have that h1 = 0, and h2 = v1+ = v1− = V1 . A solvability condition can be obtained, since by translational invariance we have that L(U0 ) = 0. Hence ∞ −∞ ∞ (U0 LU1 − U1 LU0 )dy = −V1 −∞ U0 fV (U0 , V0 )dy = 0. We can conclude then that if fV = 0, then V1 = 0, thus V1± = 0. We also have then that U1 = cU0 , and without loss of generality we take c = 0. To O(ε ) we have then v(x) = u(x) = ε 2 2D g− (x − L) ε − 2D g+ (x2 − l2 ), V+ V + g− (L − l)2 , l < x < L, 0 < x < l. (5.12) ε2 u− + O( D ), l < x < L, u+ + O( εD ), 0 < x < l. 2 175 In the inner region we expand to second order, u = U0 + v=V+ ε2 D V2 ε2 D U2 + · · · , to get + · · · and L(U2 ) = U2yy + fU (U0 , V0 )U2 = −fV (U0 , V0 )V2 , V2yy = 0, with the matching condition that V2 ∼ yv1± as y → ∓∞, and v1 the O(ε /D) term for v(x) in (5.12). We must have then that V2 (y) = H20 + yH21 , and we can conclude that H21 = V1+ = V1− . Using (5.11), we can now recover the result from (5.4): g− (L − l) = −g+ l → l= g− L. g− − g+ The constant H20 can be found in terms of H21 via a solvability condition, since LU0 = 0, and LU2 = U2yy + fU (U0 , V0 )U2 = −fV (U0 , V0 )(H20 + yH21 ), We have then that ∞ −∞ (H20 + yH21 )U0 fV (U0 , V0 )dy = 0, hence H20 = ∞ −∞ yU0 fV (U0 , V0 )dy −V1± ∞ −∞ U0 fV (U0 , V0 )dy In the outer expansion, v(x) = V + ε D v1 H20 . 176 + ε2 D v2 , we require then that v2 (l) = 5.2 Transverse stability of the K-mesa solution to perturbations along the y-axis We will now use the general mesa construction outlined in the previous section to study the stability of mesa stripes to transverse perturbations. We start by considering a one-mesa steady-state solution in the domain [−L, L]× [0, d0 ]. We consider small perturbations of the form u(x, y) = ue (x) + eλt eimy φ(x), v(x, y) = ve (x) + eλt eimy ψ(x), which yield the eigenvalue problem λφ = ε 2 φxx − ε 2 m2 φ + fu (ue , ve )φ + fv (ue , ve )ψ, τ λψ = D D ψxx − m2 ψ + gu (ue , ve )φ + gv (ue , ve )ψ. ε ε (5.13) We now multiply the φ equation in (5.13) by ux , and integrate it by parts on [0, L]. Given that the equilibrium problem satisfies ε 2 uxx + f (u, v) = 0, we have then ε 2 (ux )xx + fu ux + fv vx = 0. We define the operator Lε u by Lε u ≡ ε 2 uxx + fu u, and we have that the equilibrium problem satisfies Lε ux = ε 2 (ux )xx + fu ux = −fv vx , and from (5.13) we get ¯ Lε φ + fv ψ = λφ. 177 Integrating first on −L < x < 0, we have 0 −L (ux Lε φ − φLε ux )dx = ε 2 [ux φx − φuxx ] 0 −L . The two terms on the left side of the integral each equate to 0 −L 0 −L ¯ ux Lε φdx = λ φLε ux dx = − 0 0 ux φdx − −L 0 ux fv ψdx −L φfv vx . −L Putting it all together leads to 0 ¯ λ −L 0 ux φdx − 0 ux fv ψdx + −L −L φfv vx = ε 2 [ux φx − φuxx ] 0 −L . We now make use of the following facts: ux (−L) = ux (0) = 0 from Neumann boundary conditions and even symmetry considerations, respectively. Both ψ(x) and vx (x) are approximately constant, hence ψ(x) ∼ = ψ(−l), vx (x) ∼ = vx (−l). Since ux (x) is localized near the interface, we have that 0 0 ux φdx = c− −L −L u2x dx. This reduces the equation to ¯ − λc ¯ − λc 0 −L 0 −L u2x dx ∼ = ψ(−l) 0 −L 0 ux fv dx − vx (−l)c− u2x dx ∼ = [ψ(−l) − vx (−l)c− ] 0 −L 178 −L fv ux dx − ε 2 [φuxx ] ux fv dx − ε 2 [φuxx ] 0 −L , 0 −L , We now estimate φ x=−L uxx x=−L = O(1), φ = O(ε ), uxx therefore we have that ε 2 (φuxx ) 0 −L x=0 x=0 = O(1), as well as = O(ε ), = O(ε 3 ). Changing variables to y = ε −1 (x + l), we have that 0 −L ∞ ux fv dx ∼ u+ −∞ U0 (y)fv dy = fv du, u− since u → u± when y → ±∞. Similarly, we can make the same change of variables to have 0 −L u2x dx ∼ 0 −L 1 1 (U0 )2 dx = 2 ε ε ∞ −∞ (U0 )2 dy. This yields the simplified equation ¯ − λc ∞ −∞ u+ (U0 )2 dy ∼ ε [ψ(−l) − c− vx (−l)] fv du + O(ε 4 ). u− We now define κ0 ≡ The integral in the numerator, ∞ 2 −∞ (U0 ) dy . u+ u− fv du ∞ 2 −∞ (U0 ) dy, can be further estimated by inte- grating the U equation in (5.6) ∞ −∞ 2 U0y (y)dy ∼ u+ 0 2 U0y 1 dU0 = U0y 179 u+ 2F(u)du, 0 with F(u) = − u 0 f (s)ds. Then, ¯ 0 c− ∼ ε [ψ(−l) − vx (−l)c− ] . λκ Repeating the procedure for the 0 < x < L region, we obtain the analogous equation ¯ 0 c+ ∼ ε [−ψ(l) + vx (l)c+ ] , λκ with the sign change from the fact that with a change of variables y = ε −1 (x − l) and the transition layer at x = l, in this region we have u → u± when y → ∓∞. We recall from (5.9) and (5.10) that vx (l) ∼ from (5.4) that l = g− L g− −g+ , ε D g− (L therefore vx (l) = − ε D − l), and we also know g− g+ L . g− − g+ Furthermore, since v(x) is an even function, we have that vx (−l) = −vx (l). We can collect both equations into the following linear system ¯ 0 ε −1 λκ c+ c− ∼ = −ψ(l) ∼ = −ψ(l) ψ(−l) ψ(−l) + vx (l) ε − D c+ c− g− g+ L g− − g+ (5.14) c+ c− Remark (i) The goal is to find λ; to establish the conditions under which the system is lin180 early stable or unstable it is sufficient to determine the sign of λ. To fully solve the system in (5.14) we need to find ψ(±l), and this has to be done by finding the equilibrium solution for the second equation in (5.13): ψxx − m2 ψ + ε (gu φ + gv ψ) = 0. D (5.15) (ii) At this point we want to consider solutions consisting of K−mesas. The one mesa problem in −L < x < L can be extended to the K−mesa case on −L < x < (2K − 1)L by means of Floquet theory. This can be accomplished for the ψ equation by using the following boundary conditions: ψ((2j−1)L) = z j ψ(−L), ψ ((2j−1)L) = zψ (−L), for j = 1, · · · , K. At the boundary of the whole interval [−L, (2K − 1)L], we have ψ((2K − 1)L) = z K ψ(−L). We can get standard periodic boundary conditions then by choosing z K = 1. (iii) Systems with homogeneous Neumann boundary conditions can be extended to periodic boundary conditions by adding an even reflection on one of the boundaries, yielding a system on twice the original domain. By the same token, a system with periodic boundary conditions, with even symmetry, can be folded in half into an equivalent system with Neumann boundary conditions. Applying this idea to the extended K−mesa system implies that we have to consider 2K mesas in the periodic case, and thus z = e2πik/2K , k = 0, . . . , K − 1. The case for K = 1 has to be considered separately, and will be discussed in detail in § 5.2.1. Since u(x) is approximately constant except at the interfaces, we can approximate the eigenfunction φ ∼ c± ux , and similarly ψ ∼ ψ(±l) when x ∼ ±l. 181 In the flat regions |x| < l and l < |x| < L we have that fu φ + fv ψ = λφ. We will show later that λ 1, and this can be used to approximate fv+ ψ fu+ f− φ ∼ − v− ψ fu φ∼− for |x| < l, for l < |x| < L. Substituting it into (5.15), we end up with the ODE 2 ψxx − σ± ψ = 0, (5.16) which is defined everywhere except at the interfaces, and where 2 σ± = m2 + ε τλ ε κ± + , with D D fv+ + gu fu+ f− κ− ≡ − gv− − v− gu− fu κ+ ≡ − gv+ − > 0, when |x| < l > 0, when l < |x| < L, and as before, we are using the notation fv+ ≡ (5.17) ∂f + ∂v (u , V). Remark (i) We will eventually show that for m = 0 we have that λ = O(ε 2 ), and that λ = O(ε ) for m = 0. It is tempting to disregard the τ λε term in (5.13), however, we will keep this term in order to analyze the case when τ → ∞. Since ψ is smooth, the jump across x = ±l can be estimated in the sense of distributions through the term gu φ. We start with the jump across x = l, with the 182 standard inner change of variables y = ε −1 (x − l), l+ gu φ ∼ gu c+ ux → c+ l− +∞ gu ux dxδ(x − l) ∼ c+ −∞ gU0 U0y dyδ(x − l) u− ∼ c+ u+ gU0 dU0 δ(x − l) ∼ c+ (g− − g+ )δ(x − l). (5.18) Similarly, for the jump across x = −l, with y = ε −1 (x + l), we have −l− gu φ ∼ gu c− ux → c− −l+ +∞ gu ux dxδ(x + l) ∼ c− −∞ gU0 U0y dyδ(x + l) u+ ∼ c− u− gU0 dU0 δ(x + l) ∼ c− (g+ − g− )δ(x + l). (5.19) Therefore, we have gu φ ∼ c+ (g− − g+ )δ (x − l) + c− (g+ − g− )δ (x + l), effectively taking into account the contributions from both interfaces. We can now write (5.16) defined on the whole interval −L < x < L as 2 ψxx − σ± ψ= ε (g− − g+ ) c+ δ (x − l) − c− δ (x + l) , D with the added conditions that the solution has to be continuous across+the inter− faces at x = ±l; that the jump conditions ψx −l −l+ l = c− s, and ψx l− = c+ s are satisfied; and that the Floquet boundary conditions ψ(L) = zψ(−L), and ψ (L) = zψ (−L) are satisfied as well. For the jump conditions we have s = ε D (g− 183 − g+ ) A solution to the system with prescribed continuity across the interfaces is cosh[σ− (x + L)] + AL sinh[σ− (x + l)], −L < x < −l, ψ(−l) cosh[σ− (L − l)] sinh[σ+ (x + l)] sinh[σ+ (l − x)] + ψ(l) , −l < x < l, ψ(x) = ψ(−l) cosh(2σ+ l) sinh(2σ+ l) cosh[σ− (L − x)] + AR sinh[σ− (x − l)], l < x < L, ψ(l) cosh[σ− (L − l)] (5.20) the four unknowns AL , AR , ψ(l), ψ(−l) can be determining by enforcing all four boundary and jump conditions. We start with the Floquet boundary conditions. The four relevant terms are ψ(l) + AR sinh[σ− (L − l)], cosh[σ− (L − l)] zψ(−l) zψ(−L) = − zAL sinh[σ− (L − l)], cosh[σ− (L − l)] ψ(L) = ψ (L) = AR σ− cosh[σ− (L − l)], zψ (−L) = zAL σ− cosh[σ− (L − l)]. From the condition on ψ (L) = zψ (−L), it can immediately be seen that zAL = AR . From the condition that ψ(L) = zψ(−L), we get ψ(l) − zψ(−l) = −2zAL sinh[σ− (L − l)], cosh[σ− (L − l)] hence ψ(l) − zψ(−l) = −zAL sinh[2σ− (L − l)] 184 (5.21) For the jump conditions, the four relevant terms are ψx (l+ ) = −ψ(l)σ− tanh[σ− (L − l)] + AR σ− , ψx (l− ) = ψ(l)σ+ coth(2σ+ l) − ψ(−l)σ+ csch (2σ+ l), ψx (−l+ ) = ψ(l)σ+ csch (2σ+ l) − ψ(−l)σ+ coth(2σ+ l), ψx (−l− ) = −ψ(−l)σ− tanh[σ− (L − l)] + AL σ− . The two jump conditions, ψ(l+ ) − ψ(l− ) = c+ s and ψ(−l+ ) − ψ(−l− ) = −c− s, yield − ψ(l)σ− tanh[σ− (L − l)] + AR σ− − ψ(l)σ+ coth(2σ+ l) + ψ(−l)σ+ csch (2σ+ l) = c+ s, ψ(−l)σ− tanh[σ− (L − l)] + AL σ− − ψ(l)σ+ csch (2σ+ l) + ψ(−l)σ+ coth(2σ+ l) = c− s Simplifying things slightly, and using AR = zAL , we get − ψ(l) σ− tanh[σ− (L − l)] + σ+ coth(2σ+ l) + ψ(−l)σ+ csch (2σ+ l) = c+ s − zAL σ− , − ψ(l)σ+ csch (2σ+ l) + ψ(−l) σ− tanh[σ− (L − l)] + σ+ coth(2σ+ l) = c− s − AL σ− (5.22) Putting together (5.22) and (5.21), we can express ψ(±l) as the solution to the linear system d e ψ(l) e d −ψ(−l) z 1 T ψ(l) −ψ(−l) = −s c+ c− + σ− AL z 1 = −AL z sinh[2σ− (L − l)] 185 , with d ≡ σ− tanh[σ− (L − l)] + σ+ coth(2σ+ l), e ≡ σ+ csch (2σ+ l). Solving for AL in the second equation, and substituting it into the first one yields −bT1 r zχ Gr = −sc + b0 , with G= b0 = d e e d z 1 , r= , b1 = ψ(l) −ψ(−l) 1 z c+ , c= , c− χ= , sinh[2σ− (L − l)] . σ− This yields the matrix problem (G + B)r = −sc, (5.23) where B= 1 1 b0 bT1 = zχ zχ z z2 1 z = 1 χ 1 z z¯ 1 , since z z¯ = 1. Recall that the eigenvalue system that we want to solve, from (5.14), is ¯ 0 ε −1 λκ c+ c− = −ψ(l) ψ(−l) + vx (l) c+ c− , or, using compact notation ¯ 0 c = r + vx (l)c. ε −1 λκ 186 (5.24) Solving for r in (5.23), and substituting it into (5.24) yields ¯ 0 ε −1 λκ vx (l) c = (G + B)−1 c + c. s s Now, (5.25) D vx (l) ε g− g+ L g− g+ L . =− = s D g− − g+ ε (g+ − g− ) (g+ − g− )2 And since l= g− L g− − g+ g+ L =1− , g− l → we can then calculate vx (l) l2 = s L Similarly, 1− L l = 1− L l l2 L. L2 (5.26) ¯ 0 ¯ 0 ¯ 0 l ε −1 λκ ε −1 λκ D λκ = =− 2 . s ε (g+ − g− )/D ε g− L ¯ = λ + ε 2 m2 , we obtain Substituting into (5.25), with λ (G + B)−1 c + 1 − with α = − gD− κ0 l L L l l2 α Lc − m2 α¯ c = 2 λ¯ c, 2 L ε (5.27) . The eigenpairs λ and c¯ of (5.27) are given in terms of the spectrum of the two-by-two matrix (G + B)−1 , (G + B)−1 ν± = ω± ν± , 187 they are given as c = ν+ , c = ν− , ε2 L ω+ + 1 − α l 2 L ε ω− + 1 − λ− = α l l2 L − m2 α , L2 l2 L − m2 α . L2 λ+ = and and (5.28) This will yield two eigenpairs for each value of z. Thus we only need to find the spectrum of (G + B)−1 , and this can be done by first calculating the eigenpairs of G + B, and then taking the reciprocals of the eigenvalues. Since the matrix d e G+B = + e d 1 χ 1 z z¯ 1 is a Hermitian matrix, then all the eigenvalues must be real, and it is possible to find an orthonormal basis. We let (G + B)ν = σν, and we have det d+ 1 χ e+ −σ z¯ χ e+ d+ 1 χ z χ −σ = 0. Thus 1 d+ −σ χ 2 z z¯ e+ , χ χ 1 e = e2 + 2 + (z + z¯), χ χ = e+ and we have used the fact that z z¯ = 1. 188 Thus d+ 1 1 2e − σ = ± e2 + 2 + Re(z) χ χ χ 1/2 , and we can conclude that the eigenvalues of (G + B)−1 , needed in (5.28), are simply 1 ω± = d+ 1 χ ± e2 + 1 χ2 + 2e χ Re(z) 1/2 . Now, in order to calculate the eigenvectors we first notice that d+ 1 − σ = ±|f |, χ where f = e + z , χ and |f | = (f f¯)1/2 is the length of the complex vector f . Thus, for ω+ we have d+ 1 χ e+ −σ z¯ χ e+ d+ 1 χ z χ −σ ν+ = |f | f ν+ = 0, f¯ |f | therefore ν+ = Similarly, for ω− we have −|f | f ν− = 0, ¯ f −|f | hence ν− = |f | . f¯ Notice that in a generalized dot product defined as a · b = a† b, where a1 . . a= . , aN 189 a† = (¯ a1 , · · · , a ¯N ), |f | . −f¯ then ν+ · ν− = (|f |, −f ) |f | f¯ = 0. Therefore ν+ , ν− are orthogonal with respect to this inner product. Lemma 5.1 The spectrum of (G + B)−1 ν± = ω± ν± is as follows: ω+ = ω− = 1 d+ 1 χ + |f | 1 d+ 1 χ − |f | ν+ = |f | , −f¯ ν− = |f | , f¯ (5.29) where z , e = σ+ csch (2σ+ l), χ sinh[2σ− (L − l)] χ= , d = σ− tanh[σ− (L − l)] + σ+ coth(2σ+ l). σ− f =e+ Then, we have that |f | = e2 + 1 2e + Re(z) 2 χ χ 1/2 , z = eiθ . Lemma 5.2 Consider a steady-state solution of K−mesas on an interval of length 2KL with Neumann boundary conditions. Then the linearized problem admits 2K eigenvalues. The eigenvalues are given by λ±j = ε2 L ω±j + 1 − α l l2 L − m2 α L2 190 for j = 1, · · · , K − 1, where 1 ω±j = d+ 1 χ ± e2 + 1 χ2 + 2e χ cos(πj/K) 1/2 j = 1, · · · , K − 1. , Finally, the two remaining eigenvalues are λ±K = ε2 L ω±K + 1 − α l l2 L − m2 α , L2 1 , d+e 1 . d−e with ω+K = ω−K = These are the eigenvalues that correspond to a K−mesa pattern generated by gluing together K mesas. The various quantities are: d = σ− tanh[σ− (L − l)] + σ+ coth(2σ+ l), e = σ+ csch (2σ+ l), −1 χ = σ− sinh[2σ− (L − l)], 2 σ± = m2 − ε D gv± − fv± ± gu fu± + ε τλ . D Now, we calculate with a little algebra d + e = σ− tanh[σ− (L − l)] + σ+ coth(σ+ l), d − e = σ− tanh[σ− (L − l)] + σ+ tanh(σ+ l), 1 d + = σ− (tanh[σ− (L − l)] + σ+ csch [2σ− (L − l)]) + σ+ coth(2σ+ l). χ 191 In addition, 1/2 1 2e e + 2+ cos(2πj/K) χ χ 2 1 e+ χ = 2 4e − sin2 χ πj 2K 1/2 , and 1 = σ+ csch (2σ+ l) + σ− csch [2σ− (L − l)], χ 4e = 4σ+ σ− csch (2σ+ l)csch [2σ− (L − l)]. χ e+ Thus 1 ω±j = d+ 1 χ ± e+ 1 χ 2 1/2 − 4e χ 2 sin , πj 2K j = 1, · · · , K − 1. Finally, D α = − κ0 g− l L κ0 ≡ , ∞ 2 −∞ (U0 ) dy , u+ u− fv du l g− = L g− − g+ In order to find the eigenvalues for a specific RD system, the key elements that need to be determined are u+ , u− , V, which are obtained through the heteroclinic Maxwell line condition f (u+ , V) ≡ f+ = 0, f (u− , V) ≡ f− = 0, u+ u− f (w, V)dw = 0. Regarding the stability of the heteroclinic, we also require that fu (u± , V) < 0. 192 The mesa half-width l can then be determined in terms of the domain half-width L, l= g− L, g− − g+ with g± ≡ g(u± , V), and similarly f± ≡ f (u± , V). Lastly, in order to determine κ0 we need ∞ 2 −∞ (U0 ) dy, and this can be deter- mined in the following way, ∞ −∞ (U0 )2 dy = u+ u 2F(u)du, with F(u) = − 0 f (s)ds. 0 The rest of the terms necessary to determine the breather and zigzag eigenvalues can be trivially calculated from these results. 5.2.1 The one-mesa special case The case of the stability of a one-mesa solution with Neumann boundary conditions will now be considered. This has to be done separately, as the Floquet theory used in the K− mesa case is for j = 1, · · · , K − 1, and therefore excludes the case K = 1. Since the analysis is analogous to what was previously done for K−mesas, we will start at (5.14), which we write as ¯ 0 ε −1 λκ c+ c− = with vx (l) = −ψ(l) ψ(−l) ε D 193 + vx (l) g− g+ L , g− − g+ c+ c− , (5.30) and ψ(x) satisfies c− ε c+ ε (g+ − g− )δ (x − l) − (g+ − g− )δ (x + l), D D ψ (−L) = ψ (L) = 0. ψ(x)xx − σ 2 ψ = (5.31) Analogous with (5.20), we obtain that cosh[σ− (x + L)] , −L < x < −l, ψ(−l) cosh[σ− (L − l)] sinh[σ+ (x + l)] sinh[σ+ (l − x)] + ψ(l) , −l < x < l, ψ(x) = ψ(−l) sinh(2σ l) sinh(2σ+ l) + cosh[σ− (L − x)] ψ(l) , l < x < L, cosh[σ− (L − l)] with s = ε D (g+ − g− ). −l+ The unknowns ψ(±l) are to be found by satisfying the jump conditions ψx −c− s, and ψx l+ l− −l− = c+ s. Hence, we get ψ(l) σ− tanh[σ− (L − l)] + σ+ coth(2σ+ l) − ψ(−l)σ+ csch (2σ+ l) = −c+ s, ψ(l)σ+ csch (2σ+ l) − ψ(−l) σ+ coth(2σ+ l) + σ− tanh[σ− (L − l)] = −c− s. We can write this as the linear system d e ψ(l) e d −ψ(−l) = −s c+ c− , G≡ d e e d where d ≡ σ− tanh[σ− (L − l)] + σ+ coth(2σ+ l), e ≡ σ+ csch (2σ+ l). 194 , = Inverting G and substituting into (5.31), we get ¯ 0 ε −1 λκ s c+ c− = c+ vx (l) s + G−1 c− c+ , c− with vx (l)/s as in (5.26), vx (l) D ε g− g+ L g− g+ L = = = s D g+ − g− ε (g+ − g− ) (g+ − g− )2 1− L l l L . l2 L, L2 and similarly, ¯ 0 ¯ 0 ε −1 λκ D ¯ ε −1 λκ = = − 2 λκ 0 s (ε /D)(g+ − g− ) ε g− Putting it all together, we conclude that G−1 c+ c− + 1− L l c+ l2 L 2 L c− =− D ¯ λκ0 ε 2g − l L c+ c− , (5.32) with G−1 = d2 1 − e2 d −e −e d . Thus, the stability of a single mesa is governed by (5.32). Remark (i) As a sanity check, we need to show that (5.32) evaluated for the GMS system yields an equivalent equation as the matrix problem in equation (5.15) of [22]. (ii) Similarly, we also need to show that the general K−mesa matrix problem given in (5.27) is also equivalent in the GMS case to the previously computed system in (4.38). 195 We have from (5.15) of [22] that L ¯ c, ˆ −1 cˆ = α λˆ (1 − L)I − G 2 ε2 with cˆ = cl cr , (5.33) where L is the width of the mesa, and the domain has constant length 1; and ˆ −1 = G 1 2 d − e2 d e e d , α≡ 2βLD 2 , w+ ∞ β= w 2 dy. −∞ Remark (i) Notice that in the formulation for c in (5.18) and (5.19), when compared to cˆ, we have that cl = c− , but cr = −c+ . Rewriting (5.33) in the notation we have been using, we have that L = 2l is the width of the mesa, and we have ¯ c, ˆ −1 − l(1 − 2l)I cˆ = − α λˆ G ε2 with α = (5.34) 2β(2l)D . 2 w+ Now, in our formulation, if we restrict it to the case L = 1/2 (domain length 1), we obtain G−1 − l(1 − 2l)I c = − Dκ0 (2l) ¯ λc. ε 2 g− For the left hand side, we expand (5.34) as 1 α¯ [dcl + ecr ] + l(1 − 2l)cl = 2 λc l, 2 −e ε 1 α¯ [dcr + ecl ] − l(1 − 2l)cr = − 2 λc r. d2 − e2 ε − d2 196 (5.35) Interchanging the rows, we can write this as d2 d 1 − e2 −e −e d cr − l(1 − 2l) −cl cr −cl =− α ¯ cr λ ε2 −cl , and since cr = −c+ , and cl = c− , this system is equivalent to (5.35), provided that the right hand sides match, and that the d and e values are consistent. Thus, we only need to show that α= 4βlD 2lκ0 D = , 2 g− w+ or, more succinctly, that κ0 2β = , 2 g− w+ where ∞ β= 2 w dy, −∞ U0 = Vw(y), κ0 = ∞ 2 −∞ (U0 ) dy , u+ u− fv du g− = g(u− , V). For the GMS model, we have that g(u, v) = −v + u2 , f (u, v) = −u + u2 , v(1 + κu2 ) and u− = 0. When v = V, we integrate u+ u− u+ fv du = − 0 197 u2 du. V 2 (1 + κu2 ) Now, we substitute u = Vw, u+ = Vw+ , and get u+ u− w+ fv du = −V 0 w2 dw, 1 + b0 w2 with b0 = κV 2 . From the definition of the system in (4.21), we have that w −w+ w2 = 0. 1 + b0 w2 Multiplying by w and integrating from y = −∞ to y = ∞, we get w2 2 ∞ dw w2 ∞ w2 dy = 0, + 2 dy 2 −∞ 1 + b w −∞ 0 −∞ w+ w2 w2 dw = 0. − ++ 2 1 + b0 w2 0 ∞ − Hence w+ 0 2 w+ w2 dw = , 1 + b0 w2 2 and we can conclude that u+ u− fv du = −V 2 w+ . 2 We also calculate that ∞ −∞ (U0 )2 dy = V 2 ∞ (w )2 dy, −∞ g− = g(u− , V) = −V. We conclude that ∞ V 2 −∞ (w )2 dy 2 κ0 = 2 /2) = w 2 g− −V(−Vw+ + 198 ∞ −∞ (w )2 dy. Now, the only thing remaining to ensure compatibility between our result and (5.15) of [22] is to show that the decay rates σ± and θ ± are the same. We have from (5.17) that ε (κ± + τ λ), with D f+ κ+ ≡ − gv+ − v+ gu+ > 0, fu − f κ− ≡ − gv− − v− gu− > 0, fu 2 σ± = m2 + and from (4.32) that θ ± is θ = θ − ≡ m2 + ε (1+τ λ) D θ + ≡ m2 + ε (1+τ λ) D 1/2 , 1+ 2w+ l(w+ −2) 1/2 We have then, that for the GMS system, u− = 0 u+ = Vw+ , v = V, where V = 1 2 2l . w+ Now, we have that gu− = 0, and gv− = −1, hence κ− = 1, and thus θ − = σ− . Similarly, since u+ satisfies f (u+ , V) = 0, and thus u+ = V(1 + κu2+ ), we have, gu+ = 2u+ , fu+ = −1 + gv+ = −1, fv+ = − 2u+ 2 2V = −1 + = −1 + , u+ V(1 + κu2+ )2 1 + κu2+ u2+ u+ =− . 2 2 V V (1 + κu+ ) 199 We have then κ+ = − gv+ − =1+ fv+ + gu fu+ 3 2Vw+ w+ − 2 = − −1 + =1+ u+ /V 2u+ 2V−u+ =1+ u+ 2u3+ Vu+ − 2V 2 w+ . l(w+ − 2) This shows that σ± = θ ± , hence the GMS results in (5.15) of [22] are consistent with ours general result. Thus, the GMS results constitute a particular case of our general framework. Our final result is that on a domain [−L, L], with mesa width 2l, we have G−1 c+ c− + 1− L l c+ l2 L 2 L c− D ¯ λκ0 =− ε 2g −e , − l L where G−1 = 1 d2 − e2 d −e d d = σ− tanh[σ− (L − l)] + σ+ coth(2σ+ l), e = σ+ csch (2σ+ l), ε 2 σ± = m2 + (κ± + τ λ), D f+ κ+ = − gv+ − v+ gu+ > 0, fu f− κ− = − gv− − v− gu− > 0, fu 2 2 ¯ λ=λ+ε m , g− = g(u− , V), κ0 = ∞ 2 −∞ (U0 ) dy . u+ u− fv du 200 c+ c− , (5.36) It is now convenient to define α=− D κ0 g− l L . Then, with I the identity matrix, we can rewrite (5.36) as G−1 + 1 − L l l2 α LI − m2 αI c = 2 λc 2 L ε The eigenpairs of the system are then λ± = c± = L ε2 ω± + 1− α l 1 ∓1 l2 L − m2 α , L2 (5.37) , with ω ± the eigenvalues of G−1 , given by 1 = [σ− tanh[σ− (L − l)] + σ+ tanh(lσ+ )]−1 , d−e 1 = = [σ− tanh[σ− (L − l)] + σ+ coth(lσ+ )]−1 . d+e ω+ = ω− The (λ+ , c+ ) eigenpair corresponds then to the breather mode, and the (λ− , c− ) is the zigzag mode. Two cases are worth considering, • The case m = 0, τ = O(1): We have that, irrespective of the sign of α, and for m sufficiently large, we will have that λ± < 0, i.e., the stripe is stable to short wavelength perturbations in the y-direction. We have that ω ± > 0, and in fact as L → ∞ we have ω ± ∼ 201 1 2m . Given that 1 − L/l < 0, with the ratio L/l remaining constant when only the domain length L is increased, we have that on a small enough domain the stripe will always be stable, and that there will be a critical length L beyond which the stripe will become unstable. • The case m = 0, τ = O(1): When ε → 0 we have that σ± → 0. For the zigzag mode (λ− ), we have that σ+ coth(σ+ l) = σ+ 1 1 2 ∼ ), ∼ + O(σ+ 2 l2 σ tanh(σ+ l) l l(1 − 3 ) 2 (L − l), we have and since σ− tanh[σ− (L − l)] ∼ σ− 2 ω − ∼ σ− (L − l) + 1 2 + O(σ+ ) l −1 2 2 = l + O(σ+ , σ− ). We conclude that for m = 0 we have λ− ε2 α 1− l2 ε2 2 2 L + l + O(σ , σ ) ∼ + − L2 α L l l2 L2 2 2 L + O(σ+ , σ− ) , and stability is guaranteed for this mode when α < 0. Remark (i) When τ ε τλ 1 this analysis is consistent provided that the term in ω ± , 1. Since this shows that λ = O(ε 2 ), the condition for self-consistency is that τ O(1/ε 3 ). Now, for the breather mode λ+ , we have that ω+ ∼ 2 = Since σ− ε D (κ− 1 1 2 (L − l) + σ 2 l = σ 2 L + l(σ 2 − σ 2 ) . σ− + − + − 2 − σ2 = + τ λ), and σ+ − ω 2+ ∼ ε D (κ+ − κ− ), we have that D . ε L (κ− + τ λ) + Ll (κ+ − κ− ) 202 In terms of the breather eigenvalue, this yields that λ+ = ε2 L ω+ + 1− α l l2 1 εD L ∼ . L2 Lα (κ− + τ λ) + Ll (κ+ − κ− ) Notice again that since l/L < 1 we have stability of this mode provided that α < 0. Remark (i) Notice that the breather eigenvalue λ+ = O(ε ), while the zigzag eigenvalue λ− = O(ε 2 ). (ii) The consistency condition for the breather case is that 0 < τ < O(ε −1 ), since λ+ = O(ε , τ λ). 5.3 Hopf bifurcation on 1-d mesa patterns in the shadow limit In the previous section we studied in detail the breather and zigzag instabilities that arise from transverse perturbations. We will now use the estimates on the eigenvalues for the m = 0 case to consider the possibility of a Hopf instability giving rise to oscillatory instabilities. The eigenvalues of 1-d mesa patterns are given by (5.37) when considering the case m = 0, L ε2 ω+ + 1− α l 2 ε L λ− = ω− + 1− α l λ+ = 203 l2 L , L2 l2 L , L2 (5.38) where 1 = [σ− tanh[σ− (L − l)] + σ+ tanh(lσ+ )]−1 , d−e 1 = [σ− tanh[σ− (L − l)] + σ+ coth(lσ+ )]−1 , = d+e ω+ = ω− with σ± = l D κ0 , g− L ∞ (U0 )2 dy . κ0 = −∞ u+ u− fv du ε (κ± + τ λ), D κ± = − gv± − α=− fv± ± gu , fu± (5.39) Remark (i) There are several distinguished limits in τ that are relevant. Case I: The natural distinguished limit is when λ± = O(ε 2 ). In this limit we need both ω + and ω − to satisfy ω ± = O(1), and thus we require that σ± = O(1) (we showed earlier that ω + and ω − have different limits as σ± → 0). To satisfy this condition we require that O(1). Given that λ = O(ε 2 ), ε τλ D = O(1) in order to have σ± = this means that τ = O(ε −3 ). The equations can be simplified by eliminating some constants via a suitable change of variables. Let τ0 and Λ be defined as λ=− ε2 Λ, α τ =− 1 αDτ0 . ε3 From (5.39) we obtain σ± = τ0 Λ + O(ε ), 204 and the equations in (5.38) become Λ = −√ l2 L, L2 1 L √ √ − 1− l τ0 Λ tanh( τ0 Λ(L − l)) + tanh( τ0 Λl) Λ = −√ l2 L, L2 L 1 √ √ − 1− l τ0 Λ tanh( τ0 Λ(L − l)) + coth( τ0 Λl) (5.40) for the zigzag and breather eigenvalues, respectively. In order to find the Hopf bifurcation values, we let Λ = iδ /τ0 , and for the zigzag case we get F+ = iδ 1 L +√ + 1− √ √ τ0 l iδ tanh( iδ (L − l)) + tanh( iδ l) l2 L = 0. L2 We must now find values for δ and τ such that F+ = Re[F+ ] + iIm[F+ ] = 0, hence we require both Re[F+ ] = 0 and Im[F+ ] = 0. We have then that the Hopf bifurcation values in (5.38) are the roots of Re √ 1 + 1− L √ √ l iδ tanh( iδ (L − l)) + tanh( iδ l) l2 L = 0, L2 (5.41a) Im √ 1 + δ = 0. √ √ τ0 iδ tanh( iδ (L − l)) + tanh( iδ l) (5.41b) This setup decouples the system in terms of finding δ and τ0 . To find both values one must start by solving (5.41a) in terms of δ , and then substitute the 205 values into (5.41b) in order to find τ0 τ0 = − δ Im √ 1 √ √ iδ [tanh( iδ (L−l))+tanh( iδ l)] −7 x 10 9 10 3 λ+ τ+ 2 8 10 τ− λ− 1 7 10 0 6 10 λ−1 τ 5 10 −2 4 10 −3 3 10 −4 2 10 −5 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 L 5 6 7 8 9 L −7 x 10 9 10 0 τ+ −0.5 λ+ τ− −1 λ− 8 10 7 10 −1.5 −2 6 10 λ−2.5 τ 5 10 −3 −3.5 4 10 −4 3 10 −4.5 2 10 −5 0 1 2 3 4 5 6 7 8 9 0 L 1 2 3 4 5 6 7 8 9 L Figure 5.1: Plots of both the critical τ and λI at which a Hopf bifurcation occurs, as a function of the domain length L for the GMS model. The parameters used in the computations are D = 50, ε = 0.01. The two top figures are for κ = 1, and the bottom figures are for κ = 0.65. Using a straightforward Newton method the system can be easily solved, Figure 5.1 shows the Hopf curves for both the imaginary part of λ and τ . At the point where the eigenvalues become positive, the critical value of τ becomes discontinuous. Throughout the range where λ± remains negative, the breather eigenvalue 206 will always be smaller for small values of L. Plotting the critical τ values provides the Hopf bifurcation threshold, as seen in the two bottom images in figure 5.1. We will later do a numerical simulation of an RD model when considering τ in the Hopf regime. 5.3.1 ODE-PDE system We will now compute a full time-dependent solution for the case where τ = O(ε −3 ), and λ = O(ε 2 ). We start again with the full system (5.1), ut = ε 2 ∆u + f (u, v) τ vt = D ∆v + g(u, v), ε where we look for a time-dependent mesa solution on the domain [−L, L], with the two interfaces located at x = l1 and x = l2 . Since λ = O(ε 2 ), the proper time scale of the interfaces is l1 ≡ l1 (ε 2 t); l2 ≡ l2 (ε 2 t). We also let τ = τ˜0 /ε 3 , and we define T = ε 2 t. We have then that (5.1) becomes ε 2 uT = ε 2 ∆u + f (u, v) τ˜0 D vT = ∆v + g(u, v). ε ε (5.42) We now do an asymptotic expansion near around the right layer (a similar 207 expansion will also have to be computed for the leftmost layer), with u = U0 (y1 ) + ε U1R (y1 ) + · · · , v = V0 + ε V1R + ε 2 V2R + · · · , where y = ε −1 (x − l1 (T )), and therefore uT = −ε −1 U0 l1 . Substituting into (5.42), we obtain that −ε −1 U0 l1 = U0 + ε U1R + f (U0 , V0 ) + ε fU0 U1R + ε fV0 V1R + · · · As before, we take V0 = constant, and define u+ , u− , V0 in terms of the heteroclinic Maxwell line condition u+ f (u, V0 )du = 0, f (u± , V0 ) = 0, fu (u± , V0 ) < 0. u− Thus, around the right transition layer we have U1R + fu0 U1R = −fV0 V1R − l1 U0 , −∞ < x < ∞. Similarly, for the V equation we have D V˜0 V1T = 3 (V0 + ε V1 + · · · ) + g(U0 , V0 ) + · · · ε Since V0 = constant we have that O(1) = 1 DV1R + O(1), ε2 therefore V1R = 0. We also have that V ∼ V0 on the entire interval, therefore in the internal region we cannot have V1R growing at infinity, hence V1R = V1R (T ), 208 independent of y. The inner problem on the rightmost layer is LU1 ≡ U1 + fu0 U1 = −fv0 V1R − l1 U0 . Since LU0 = 0, the solvability condition is ∞ −l1 −∞ ∞ U02 dy − V1R −∞ fv0 U0 dy = 0, and as we did before, since U0 is a heteroclinic connection, ∞ −∞ fv0 U0 dy = u− u+ fv0 dU0 = − u+ fv0 dU0 . u− Thus, the ODE for the rightmost layer is dl1 l1 = = V1R (T ) dT u+ 0 u− fv dU0 . ∞ 2 −∞ U0 dy (5.43) The same procedure on the leftmost layer at x = l2 yields that l2 = −V1L (T ) u+ 0 u− fv dU0 . ∞ 2 −∞ U0 dy (5.44) To find values for V1L and V1R we need to match with the outer solution. Now, in the outer region we expand V = V0 + εV1 + · · · , and we obtain from substituting into (5.42) that τ˜0 V1T = DV1xx + g+ if l2 < x < l1 g− if l1 < x < L or − L < x < l2 . 209 The matching condition is that V1L (T ) = v1 (l2 , T ), V1R (T ) = v1 (l1 , T ). We now define w by v1 = and recall that g+ −g− g− g− w, τ˜ = − lLe , with le the equilibrium half-length of the mesa. We can now write an ODE-PDE system that can be solved to obtain the location of the mesa interfaces as a function of time, wT = D0 wxx + 1 + L [H(x − l1 ) − H(x − l2 )] , le wx = 0 at x = ±L, (5.45) dl1 = µw(l1 , t), dT dl2 = −µw(l2 , T ), dT where D0 ≡ D , τ˜0 µ≡ 1 g− , κ0 τ˜0 H(z) = 1 if z > 0 . 0 if z < 0 Remark (i) By performing linear stability analysis on (5.45) it is possible to recover (5.31). (ii) Numerically solving (5.45) has the complication that it is necessary to discretize a Heaviside function. Doing this directly will introduce errors of order h, with h the mesh spacing. This limitation can be mitigated by using a differentiable approximation to H(z). 210 In our numerical calculation we used the following discretization, H(z) ∼ H(z) 1 2 + z 2δ + 1 2π sin πz δ if |z| ≥ δ if |z| ≤ δ , with δ = O(h). To illustrate the theory, we solved (5.45) for the GMS system, and compared it with a full numerical simulation. We chose parameters similar to those in Figure 5.1, specifically κ = 1. From the top figures we can see that for a domain halflength L = 2.5, the solution will become unstable first to a breather instability. The critical τ for the breather and zigzag instabilities is τ ∼ 10, 000 and τ ∼ 40, 000, respectively. We can see from Figure 5.2 that the full numerical solution compares well with the ODE-PDE approximation. As expected from the threshold values, the solution became unstable to a breather instability, and the ODE-PDE system matches well the period and amplitude of the full solution. When choosing a τ value that is above both the breather and zigzag instabilities, it is possible to get solutions where the mesa walls collide with each other. This is not an issue in the ODE-PDE system, however, attempting to solve the full system will result in the solution collapsing, as can be seen in Figure 5.2 on the right. A similar result was studied in more detail in [33], although they worked in the parameter regime τ = O(ε−2 ), whereas in our system we consider τ = O(ε−3 ). 5.3.2 Stability proof for the breather case In order to prove that the system is stable before the Hopf point, we will utilize the Nyquist stability criteria. The argument principle states a complex function f (z), analytic in a simply connected domain G except for at most a finite number of poles satisfies Z −P = 1 ∆arg(f )|C , 2π 211 10000 10000 9000 9000 8000 8000 7000 7000 6000 6000 T 5000 T 5000 4000 4000 3000 3000 2000 2000 1000 0 −2.5 1000 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 0 −2.5 2.5 −2 −1.5 −1 X −0.5 0 0.5 1 1.5 2 2.5 X Figure 5.2: A comparison between the ODE-PDE system (5.45) and the full numerical simulation for a system beyond the Hopf threshold. The figures in the left correspond to τ = 25, 000, and the images in the right to τ = 65, 000. The rest of the parameters are D = 50, ε = 0.01. The solution was integrated until T = 10, 000, with an IMEX scheme with 800 grid points. with Z and P the number of zeroes and poles of f (z) in G, respectively; C a closed contour in G not passing through any of the poles or zeroes; and ∆arg(f )|C the change in arg(f )|C as C is traversed counter-clockwise. The Nyquist stability criteria is an application of the argument principle on a contour that encompasses the right half-plane. If a dynamical system represented by a function f (z), with no poles on the right half plane is thus shown to have ∆arg(f )|C = 0, then the solutions to the system will be stable. 212 We will focus on equations (5.40); with Λ = z/τ0 we have that f (z) is z 1 √ √ +√ + 1− τ0 z [tanh( z(L − l)) + tanh( zl)] 1 z √ √ +√ f− (z) = + 1− τ0 z [tanh( z(L − l)) + coth( zl)] f+ (z) = L l L l l2 , L l2 , L (5.46) If we approach z = 0 along the imaginary axis, as z → 0 we have that f+ (z) f− (z) 1 , Lz l L z + + 1− τ0 z(L − l) + 1 l l2 L l2 L therefore f+ (z) has a simple pole at the origin, whereas f− (z) has a removable singularity instead. Taking this into consideration, we will check the Nyquist criterion on f+ (z) on the contour given in Figure 5.3, whereas for f− it would only be necessary to use CR and the full imaginary axis. Im(z) Im+ CR Cǫ Re(z) Im− Figure 5.3: The contour on which to check the Nyquist stability criterion. 213 The four sections on the contour are given by CR : z = Reiθ , −π/2 < θ < π/2, Cε : z = ε eiθ , Im+ : z = teiπ/2 , π/2 > θ > −π/2, R > t > ε, Im− : z = teiπ/2 , −ε > t > −R, and we consider the limit when ε → 0 and R → ∞. On CR , we expand tanh(z) as tanh(a + ib) = tanh(a) + tanh(ib) sinh(a) cos(b) + i cosh(a) sin(b) = . 1 + tanh(a) tanh(ib) cosh(a) cos(b) + i sinh(a) sin(b) (5.47) √ √ In this case we have that a + ib = Reiθ /2 (L − l), hence a = R(L − √ l) cos(θ /2), and b = R(L − l) sin(θ /2), with −π/2 < θ < π/2. Since for x → ∞ we have that sinh(x) f (z) = cosh(x), we have that on CR tanh(z) → 1, and Reiθ + O(R−1/2 ) + O(1) τ0 Reiθ . τ0 Thus, the change in argument in CR is ∆arg(f )|CR = π/2 − −π/2 = π. On Cε , we can see from (5.47) that when a, b → 0 tanh(a + ib) a + ib 1 + iab a + ib, therefore, we have that f (z) ε e−iθ eiθ L + + 1− τ0 Lε l l2 L eiθ , Lε for ε → 0. Therefore, the change in argument in Cε is also ∆arg(f )|Cε = π. 214 On Im+ , we have z = it, on R > t > ε . We have that f+ (it) = it 1 L √ √ +√ + 1− τ0 l it tanh( it(L − l)) + tanh( itl) l2 . L When t → ∞ we have Re[f+ (it)] Im[f+ (it)] 1− L l l2 < 0, L t → ∞. τ0 Therefore, since the contour is traversed counterclockwise, we have arg(f )|i∞ = −π/2. Similarly, when t → 0, we approximate tanh(x) ing f+ we get f+ (it) t2 L 2 2 3 (L − 3Ll + 3l ) − itL 4 2 t2 L2 + t 9L (L2 − 3Ll + 3l2 )2 x− x3 3 , + 1− and upon expand- L l l2 , L and upon cancelling the fourth order term in the denominator, we get that the real and imaginary components are Re[f+ (it)] Im[f+ (it)] L2 − 6Ll + 6l2 , 3L 1 − , tL (5.48a) (5.48b) with the real part independent of t. Equation (5.48a) defines a new condition on the existence of a Hopf bifurcation for the breather eigenvalue. If (5.48a) is negative, then the change in argument in both imaginary axis segments will be of −2π combined, cancelling the contributions from CR and Cε , and thus guaranteeing that solutions will be sta- 215 √ ble. However, if the domain length L satisfies that either L < (3 − 3)l, or that √ (3 + 3)l < L, then (5.48a) will be positive at one point. If it so happens that f+ is positive while crossing the real axis, then the change in argument will have the opposite sign, and by the argument theorem we will have two positive real-valued zeros, hence instability. Therefore, solutions will be stable provided that L2 − 6Ll + 6l2 < 0. 3L (5.49) Since the change from negative to positive in the real part doesn’t necessarily have to happen at the two endpoints we approximated, it is best to estimate it numerically in order to get an idea of the dependence on τ0 . It would be interesting to explore the conditoin (5.49) on a numerical simulation. 5.4 Case study: the predator-prey model We will apply the mesa theory developed in §5 to a spatio-temporal predator-prey model. We will focus on the transverse stability of a K-mesa solution in the near- shadow regime D = O(1/ε ), and we will present some numerical results in the mesa-splitting regime D = O(1). The specific model we will study is a ratio-dependent predator-prey system with a Michaelis-Menten type functional response. The spatially homogeneous model was originally posited in [64], and the full spatio-temporal model was discussed in [2], and [62]. 216 The model in question is ∂N αN P N − = D1 ∇2 N + rN 1 − , ∂t K P + αβN ∂P eαN P = D2 ∇2 P − γ P + , ∂t P + αβN (5.50) where N, P are the prey and predator densities, respectively; D1 , D2 are their diffusion coefficients; r is the maximal growth rate for the prey, K is the carrying capacity, α is the capture rate, β is the handling time, e is the conversion efficiency, and γ is the predator death rate. By nondimensionalizing time (see [2]), the system can be simplified to Ut = DU ∇2 U + U (1 − U ) − AU V U +V BU V Vt = DV ∇ V − CV + , U +V (5.51) 2 and the parameter values used in the paper were DU = 1, DV = 8, A = 1.1, B = 0.9, and C = 0.1; and with V and U the populations of predators and prey, respectively. It is possible to simplify the model a bit more. By letting x ˆ = x/L, and rescaling in the domain length, for 1D we get, after dropping the hats, ut = ε 2 uxx + u(1 − u) − τ vt = Dvxx − v + auv = ε 2 uxx + f (u, v) u+v buv = Dvxx + g(u, v). u+v (5.52) In terms of the original parameters in (5.50), we have that a = α/r, b = τ = r/γ , ε 2 = D1 , rL2 and D = D2 . γ L2 e γβ, The parameter τ , for instance, represents the ratio between the maximal prey growth rate and the predator death rate. We will discuss two distinguished regimes: the near-shadow regime where 217 D = O(1/ε ), and the splitting regime where D = O(1). 5.4.1 Preliminaries There are three possible homogeneous steady-state solutions, however, the only non-trivial one that can yield a Turing instability is given by uh = 1 − a 1 − 1 b , vh = uh (b − 1) The Turing space associated with the system, as a function of the parameters a and b, is given by Figure 5.4. Figure 5.4: Turing space for the system given by 5.52, in term of the parameters a and b. In the near-shadow limit, when D = D/ε , with D = O(1), the width of the mesa can be established in a straightforward manner. Integrating over half of the stationary v equation in 5.52, and from the Neumann boundary conditions and the 218 symmetry at the centre of the mesa we have that L 0 g(u, V)dx = 0. Furthermore, splitting the integral in two, at the location of the interface, yields l− L D D vxx + g(u, V) dx + vxx + g(u, V) dx, ε ε + l 0 D D 0 = vx (l− ) + g(u+ , V)l − vx (l+ ) + g(u− , V)(L − l), ε ε u+ + V g(u− , V) L= l= L, g(u− , V) − g(u+ , V) bu+ 0= with vx (l− ) = vx (l+ ) from the fact that v(x) is a smooth function. Plotting the half-mesa width versus the parameter a yields the result shown in Figure 5.5. Notice that for much of the parameter space which satisfies the Turing conditions the mesa width would be very close or above the domain length L. For these two specific choices in b we get the consistency conditions that a > 1.45 for b = 3, and a > 1.2 for b = 2. The stationary u(x) solution with a wide mesa depicted in Figure 5.5 corresponds to parameter values that fall squarely in the Turing regime. The parameters that give rise to the narrower mesa (a = 3, b = 2) are outside of the Turing space. The bifurcation diagram for the solution as the domain length L increases, as a function of the L2 norm of the solution, is shown in Figure 5.6. The branch of solutions connects with the Turing solution, and eventually with the homogeneous solution, at the lowest part of the branch, on one edge with the 1-mode Turing solution, and on the unstable side with the 2-mode Turing solution. Past the cusp we have the unstable solutions that lead to the splitting of the solution. As in the GMS case, there will be a family of identical branches for larger 219 Mesa width 1 0.7 0.9 0.6 a = 2, b = 1.5 a = 3, b = 2 0.8 0.5 u(x) 0.7 0.6 b = 1.5 0.5 0.4 0.3 0.2 0.4 b=2 0.1 0.3 0.2 1 1.5 2 2.5 3 3.5 0 −0.5 4 −0.4 −0.3 −0.2 −0.1 a 0 0.1 0.2 0.3 0.4 0.5 x Figure 5.5: The figure on the left shows the projected half-width of a mesa versus the parameter a, for two values of b, and for L = 0.5. Notice that there is a consistency requirement on a, given that we must satisfy l < L. The second figure shows two stationary solutions for different values of the parameters that illustrates the change in mesa width. L values representing the 2n mesa branch. 0.5 0.7 51 50 0.3 L2-NORM 49 48 0.2 0.1 47 46 44 42 0.00.0 52 53 69 70 71 72 73 68 67 66 65 56 64 57 63 0.6 58 60 62 0.5 61 59 0.4 U(1) 0.4 55 54 74 75 0.3 0.2 76 0.1 45 0.0 43 0.5 1.0 1.5 PERIOD 2.0 2.5 3.0 −0.10.0 64 62 60 63 59 65 66 67 58 57 56 55 54 53 49 47 43 61 76 42 44 46 48 50 52 74 68 69 70 72 73 75 71 45 51 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 t 0.9 Figure 5.6: The bifurcation diagram for the one mesa solution, and solutions corresponding to various points along the branch. The bifurcation diagram was computed using AUTO [14]. The parameters used were a = 3, b = 2, τ = 1, D = 1, with the asymptotic term ε = 0.02. 220 1.0 5.4.2 Stability in the near-shadow regime, D = O(ε −1 ) We now study the stability of mesa stripes to transverse perturbations. We will apply the general results from § 5 and test the theory in a different model. This type of analysis was previously computed for the GMS model in § 4.4.2. From the general mesa theory developed in § 5.2, we need to first compute the heteroclinic Maxwell line parameters u+ and V (it is straightforward to check that u− = 0). These will be the key elements necessary to determine the zigzag and breather eigenvalues. We need to find u+ and V such that f (u+ , V) = f (0, V) = 0, (5.53) u+ f (u, V)du = 0. 0 Computing the values was done using a Newton method. The results are shown in Figure 5.7, where we show the numerical values for u+ and V, as well as for βpp = ∞ 2 −∞ (U0 ) dy. This last value was constant for the GMS model, but depends on the parameter a in the Predator-Prey model. These are the building blocks that 0.7 0.06 0.6 U 0.05 + V 0.5 0.04 βpp 0.4 0.03 0.3 0.02 0.2 0.01 0.1 0 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 0 3 a 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 a Figure 5.7: The values of u+ and V that satisfy the heteroclinic connection as a function of the parameter a (right figure), and the parameter βpp = ∞ 2 −∞ (U0 ) dy. 221 are required to establish the stability of a K-mesa solution. From lemma 5.2 we have: A steady-state solution of K mesas admits 2K eigenvalues, and these are given by the following formula λ±j = ε2 L ω±j + 1 − α l l2 L − m2 α L2 for j = 1, · · · , K − 1, for the first 2K − 2 eigenvalues, and where 1 ω±j = d+ 1 χ ± e2 + 1 χ2 + 2e χ cos(πj/K) 1/2 j = 1, · · · , K − 1. , Finally, the two remaining eigenvalues are λ±K = L ε2 ω±K + 1 − α l l2 L − m2 α , L2 1 , d+e 1 . d−e with ω+K = ω−K = The various quantities are: d = σ− tanh[σ− (L − l)] + σ+ coth(2σ+ l), e = σ+ csch (2σ+ l), −1 χ = σ− sinh[2σ− (L − l)], 2 σ± = m2 − ε D gv± − fv± ± gu fu± kπ , for k = 1, . . . , d0 g− l= L + O(ε ). g− − g+ m= 222 + ε τλ , D It is worth noting that in the regime with τ = 0(1), the term discarded, as we have that λ = ε τλ D in σ± can be O(ε 2 ). Implementing this in software was very straightforward. We started by choosing a value of a, and computing the eigenvalues for the two-mesa case for a range of values in the parameter b. We wanted to find a set of parameters that enabled us to observe a transverse instability, so we next chose a b value that resulted in λ > 0, specifically for the zigzag eigenvalue λ− . With this set of values for (a, b), we next ranged on the mode M , as this would determine the type of instability (mode one, mode two, etc.) −3 −3 x 10 x 10 1 0.5 0 0 −1 −0.5 −2 λ λ −1 −3 −1.5 −4 −2 −5 −2.5 −6 −3 1 2 3 4 5 6 7 0 B 1 2 3 4 5 6 7 8 9 10 M Figure 5.8: The four eigenvalues of a two mesa solution, as a function of b (left), and as a function of M (right). The parameters are D = 3, A = 1.6, ε = 0.01, L = 1,, and M = π for the figure on the left, and B = 3.5 for the figure on the right. The results from the eigenvalue calculation are shown in Figure 5.8. With the parameter choice (a, b) = (1.6, 3.5) we expect to have a solution that becomes unstable to mode-one on a domain with width d0 = 1, while remaining stable to mode-two instabilities. If the domain width were to increase to d0 = 2, we could expect to see mode-two instabilities as well (since m = kπ/d0 , with k = 1, 2, . . .). Armed with the parameters previously computed, we proceeded to run a full 223 numerical simulation of the Predator-Prey model on a 2D lattice. We used a very similar scheme as the one used for the GMS simulation, although this model has a less diagonally-dominant matrix, necessitating higher numerical accuracy. Figure 5.9: Full numerical simulation of the Predator-Prey model on a 2D lattice. We used ε = 0.01, D = 0.4, a = 1.6, b = 3.5, τ = 1. Both lattices were −1 < x < 1, and the lattice on the left had 0 < y < 0.8, while the lattice on the right had 0 < y < 2. The figures on the left were integrated until T = 5, 000, and the figures on the right until T = 10, 000. We integrated until T = 10, 000 (Figure 5.9), and chose the domain width to allow for a mode-one instability (left), and a mode-two instability (right), and recorded an image at T = 5, 000. Both mesas started from a stationary solution with a small amount of noise added. As expected, the solution on the smaller 224 domain became unstable to a mode-one instability, and when extending the domain we saw the emergence of a mode-two instability. 5.5 Chapter summary In this chapter we extend the results obtained for the GMS system to general mesa systems. We start by constructing a solution and then derive thresholds for the transverse stability of multiple mesa stripes on the shadow regime. Furthermore, we study the stability to Hopf bifurcations and derive an ODEPDE system that reduces the problem to that of finding the location of the mesa interfaces as a function of time. The system is compared to full numerics and found to be in good agreement. We study the analytic stability for the breather case by means of the Nyquist stability criteria. The general model is then verified by applying it on a Predator-Prey model and analyzing its stability to transverse perturbations. In the splitting regime we generate a bifurcation diagram by means of numerical continuation, and in the near-shadow regime we compare the results of the theory on a two stripe system with full numerics. 225 Chapter 6 Future directions The following is a list of topics that we think would be interesting to explore in more detail. 1. The instabilities that we studied in the Brusselator model, particularly the self-replication and the competition instabilities, could interact with each other and give rise to complicated dynamics. We observed that as f → 1, the stable region where the real part Re[λ] < 0 decreases. we suspect that it is possible to find a regime where the splitting and competition instability thresholds are close enough that when one of the instabilities is triggered the system lands in the unstable regime for the other instability, and vice versa. A train of events could occur with the instabilities alternating. This is close in spirit to the work by Painter et al [44] on a chemotaxis model. The main hurdle to studying this problem is that it is computationally expensive to integrate for long enough time to observe multiple events. The time scale for the slow motion of the spots is O(ε −2 ), hence an efficient numerical solver would be key. 226 2. The problem of the slow motion of the spots was addressed for the case of the Schnakenberg model on the sphere, and a similar calculation should yield similar results for the Brusselator model. From what we can tell, the dynamical system that we obtained is new; it would be interesting to determine the existence of stationary solutions and stable orbits, and how they relate to the quasi-stationary solutions and the Fekete problem. 3. While most of the work we did was in the fully nonlinear regime, it would be interesting to connect the results we obtained with the wide body of literature that exists on Turing systems in the weakly nonlinear regime. There are two approaches where the tools of weakly nonlinear analysis would be beneficial: • To obtain a global bifurcation picture for solutions on the sphere by connecting both regimes by means of a numerical continuation. This could throw light on the existence of asymmetric solutions, as well as clarifying the picture of the degenerate solutions on the sphere. • The particle-like solutions we obtained in our analysis were useful in determining bifurcation thresholds, however the technique did not allow us to determine analytically the type of bifurcations, although from numerical experimentation it seems they were subcritical. Performing a weakly nonlinear stability analysis on the particle-like solutions would be useful to get a more complete picture of the bifurcation structure. 4. Beyond non spherical domains, the problem for a general surface has yet to be addressed. Three possible avenues for study are possible: • The case of perturbations on the sphere. • Other structures where a Green’s function can be obtained, such as spherical hemispheres, or paraboloids. • General surfaces, where the problem is that of obtaining a Green’s function. As far as we can tell this hasn’t been done yet. 227 5. Back to the Brusselator case, we established the Hopf bifurcation threshold numerically but for lack of time were not able to study it in detail. Full numerical analysis is necessary to complete the picture, and it is possible that it can interact with the other two instabilities in a complicated dynamical system. 6. As was done for the mesa case, where a general theory was developed, it should be possible to obtain a general theory for systems with spot solutions on the sphere, both for the stability of the solutions, and for the slow motion of the spots. 228 Bibliography [1] E. L. Altschuler and A. Prez-Garrido. Global minimum for thomsons problem of charges on a sphere. Physical Review E, 71(4):047703, 2005. URL http://pre.aps.org/abstract/PRE/v71/i4/e047703. → pages 5 [2] M. Banerjee. Self-replication of spatial patterns in a ratio-dependent predator-prey model. Mathematical and Computer Modelling, 51(1-2):4452, 2010. → pages 2, 216, 217 [3] I. Barrass, E. J. Crampin, and P. K. Maini. Mode transitions in a model reaction–diffusion system driven by domain growth and noise. Bulletin of Mathematical Biology, 68:981–995, 2006. → pages 3, 131, 134 [4] B. P. Belousov. A periodic reaction and its mechanism. Compilation of Abstracts on Radiation Medicine, 147:145, 1959. → pages 1 [5] D. E. Bentil and J. D. Murray. On the mechanical theory for biological pattern formation. Physica D: Nonlinear Phenomena, 63(1-2):161190, 1993. → pages 2 [6] B. Bergersen, D. Boal, and P. Palffy-Muhoray. Equilibrium configurations of particles on a sphere: the case of logarithmic interactions. Journal of Physics A: Mathematical and General, 27:2579, 1994. URL http://iopscience.iop.org/0305-4470/27/7/032. → pages 5 [7] J. P. Berrut and L. N. Trefethen. Barycentric lagrange interpolation. Siam Review, 46(3):501517, 2004. → pages 122, 123 [8] S. Boatto. Curvature perturbations and stability of a ring of vortices. Discrete and continuous dynamical systems. Series B, 10(2-3), 2008. → pages 5, 117, 118 229 [9] S. Boatto and H. E. Cabral. Nonlinear stability of a latitudinal ring of point-vortices on a nonrotating sphere. SIAM Journal on Applied Mathematics, 64(1):216230, 2003. → pages 115, 117, 118 [10] M. A. J. Chaplain, M. Ganesh, and I. G. Graham. Spatio-temporal pattern formation on spherical surfaces: numerical simulation and application to solid tumour growth. Journal of Mathematical Biology, 42(5):387423, 2001. → pages 3, 121 [11] D. Coombs, R. Straube, and M. J. Ward. Diffusion on a sphere with localized traps: Mean first passage time, eigenvalue asymptotics, and fekete points. SIAM J. Appl. Math, 70(1):302332, 2009. → pages 42, 86 [12] E. J. Crampin, E. A. Gaffney, and P. K. Maini. Reaction and diffusion on growing domains: Scenarios for robust pattern formation. Bulletin of Mathematical Biology, 61:1093–1120, 1999. → pages 3, 131 [13] E. J. Crampin, W. W. Hackborn, and P. K. Maini. Pattern formation in reaction-diffusion models with nonuniform domain growth. Bulletin of Mathematical Biology, 64:747–769, 2002. → pages 3, 131 [14] E. Doedel, A. Champneys, T. Fairgrieve, Y. Kuznetsov, B. Oldeman, R. Paffenroth, B. Sandstede, X. Wang, and C. Zhang. AUTO-07P: continuation and bifurcation software for ordinary differential equations., 2007. URL http://indy. cs. concordia. ca/auto. → pages xii, 139, 143, 220 [15] A. Gierer and H. Meinhardt. A theory of biological pattern formation. Biological Cybernetics, 12(1):30–39, 1972. → pages 2, 4, 126 [16] J. Gjorgjieva. Turing pattern dynamics for spatiotemporal models with growth and curvature. PhD thesis, PhD Thesis, 2006. → pages 4, 87, 121 [17] J. Gjorgjieva and J. Jacobsen. Turing patterns on growing spheres: The exponential case. Dynamical Systems, pages 436–445, 2007. → pages 4, 121, 123, 131 [18] M. I. Jamaloodeen and P. K. Newton. The n-vortex problem on a rotating sphere. II. heterogeneous platonic solid equilibria. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Science, 462(2075): 32773299, 2006. → pages 5, 116 [19] R. Kidambi and P. K. Newton. Motion of three point vortices on a sphere. Physica D: Nonlinear Phenomena, 116(1-2):143175, 1998. → pages 15 230 [20] R. Kidambi and P. K. Newton. Point vortex motion on a sphere with solid boundaries. Physics of Fluids, 12:581, 2000. → pages 15 [21] Y. Kimura and H. Okamoto. Vortex motion on a sphere. Physical Society of Japan, Journal, 56:42034206, 1987. → pages 5, 15 [22] T. Kolokolnikov, M. J. Ward, and J. Wei. The stability of a stripe for the gierer-meinhardt model and the effect of saturation. SIAM J. Appl. Dyn. Sys, 5:313–363. → pages 6, 126, 167, 195, 196, 199, 200 [23] T. Kolokolnikov, M. J. Ward, and J. Wei. Self-replication of mesa patterns in reaction–diffusion systems. Physica D: Nonlinear Phenomena, 236: 104–122, 2007. → pages 6, 135 [24] T. Kolokolnikov, M. J. Ward, and J. Wei. Spot Self-Replication and dynamics for the Schnakenberg model in a Two-Dimensional domain. Journal of Nonlinear Science, 19(1):1–56, 2009. → pages 93, 95, 108, 114, 118, 144 [25] S. Kondo and R. Asai. A reaction–diffusion wave on the skin of the marine angelfish pomacanthus. Nature, 376:765–768, 1995. → pages 2, 85, 131 [26] H. Levine and W. J. Rappel. Membrane-bound turing patterns. Physical Review E, 72(6):61912, 2005. → pages 3 [27] A. Madzvamuse, A. J. Wathen, and P. K. Maini. A moving grid finite element method applied to a model biological pattern generator. Journal of Computational Physics, 190(2):478–500, 2003. → pages 131 [28] A. Madzvamuse, E. A. Gaffney, and P. K. Maini. Stability analysis of non-autonomous reaction-diffusion systems: the effects of growing domains. Journal of Mathematical Biology, 2009. → pages 3, 131, 134 [29] P. K. Maini, K. J. Painter, and H. N. Chau. Spatial pattern formation in chemical and biological systems. J. Chem. Soc., Faraday Trans., 93(20): 36013610, 1997. → pages 2 [30] P. C. Matthews. Pattern formation on a sphere. Physical Review E, 67(3): 036206, 2003. → pages 87 [31] T. O. Maxwell. Periodic and connecting orbits of Hamiltonian systems. University of WisconsinMadison, 1994. → pages 136, 149, 172 231 [32] R. McKay and T. Kolokolnikov. Stability transitions and dynamics of localized patterns near the shadow limit of reaction-diffusion systems. Discrete and Continuous Dynamical Systems B, to appear, 2012. → pages 171 [33] R. C. McKay, T. Kolokolnikov, and P. Muir. Interface oscillations in reaction-diffusion systems beyond the hopf bifurcation. Journal of Dynamics and Differential Equations, 2011. → pages 211 [34] H. Meinhardt. The algorithmic beauty of sea shells. Springer Verlag, 2009. → pages 2, 126 [35] T. Miura, K. Shiota, G. Morriss-Kay, and P. K. Maini. Mixed-mode pattern in doublefoot mutant mouse limbTuring reactiondiffusion model on a growing domain during limb development. Journal of theoretical biology, 240(4):562–573, 2006. → pages 3, 131 [36] J. D. Murray. A pre-pattern formation mechanism for animal coat markings. Journal of Theoretical Biology, 88(1):161199, 1981. → pages 2 [37] J. D. Murray. Mathematical biology. Vol. 2: Spatial models and biomedical applications. Interdisciplinary Applied Mathematics, 2003. → pages 2, 86, 127 [38] W. Nagata, L. G. Harrison, and S. Wehner. Reaction-diffusion models of growing plant tips: bifurcations on hemispheres. Bulletin of mathematical biology, 65(4):571–607, 2003. → pages 3, 87 [39] J. C. Neu. Vortices in complex scalar fields. Physica D: Nonlinear Phenomena, 43(2-3):385406, 1990. URL http://www.sciencedirect.com/science/article/pii/016727899090143D. → pages 5 [40] P. K. Newton. The N-vortex problem: analytical techniques, volume 145. Springer Verlag, 2001. → pages 5, 115 [41] P. K. Newton and G. Chamoun. Vortex lattice theory: A particle interaction perspective. SIAM review, 51(3):501, 2009. URL http://www.tims.ntu.edu.tw/download/talk/20110110 1492.pdf. → pages 5 [42] I. L. Novak, F. Gao, Y. S. Choi, D. Resasco, J. C. Schaff, and B. M. Slepchenko. Diffusion on a curved surface coupled to diffusion in the volume: Application to cell biology. Journal of computational physics, 226 (2):1271–1290, 2007. → pages 86 232 [43] H. G. Othmer, K. Painter, D. Umulis, and C. Xue. The intersection of theory and application in elucidating pattern formation in developmental biology. Mathematical Modelling of Natural Phenomena, 4(4):3–82, 2009. → pages 85 [44] K. J. Painter and T. Hillen. Spatio-temporal chaos in a chemotaxis model. Physica D: Nonlinear Phenomena, 240(4):363375, 2011. URL http://www.sciencedirect.com/science/article/pii/S0167278910002617. → pages 226 [45] R. G. Plaza, F. S. Gardu˜no, P. Padilla, and R. A. Barrio. The effect of growth and curvature on pattern formation. Journal of Dynamics and Differential Equations, 16(4):1093–1121, 2004. → pages 131, 132 [46] W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling. Numerical recipes in c: the art of scientific programming. Cambridge U. Press, Cambridge, England, 1992. → pages 122, 123 [47] I. Prigogine and R. Lefever. Symmetry breaking instabilities in dissipative systems. II. The Journal of Chemical Physics, 48:1695, 1968. → pages 2, 4, 8 [48] X. Ren and J. Wei. On the spectra of three-dimensional lamellar solutions of the diblock copolymer problem. SIAM Journal on Mathematical Analysis, 35:1, 2003. → pages 162 [49] S. J. Ruuth. Private communication. 2010. → pages 123 [50] S. J. Ruuth and B. Merriman. A simple embedding method for solving partial differential equations on surfaces. Journal of Computational Physics, 2007. → pages 122, 123 [51] E. B. Saff and A. B. Kuijlaars. Distributing many points on a sphere. The Mathematical Intelligencer, 19(1):511, 1997. → pages 120 [52] J. Schnakenberg. Simple chemical reaction systems with limit cycle behaviour. Journal of Theoretical Biology, 81(3):389–400, 1979. → pages 4, 85, 135 [53] L. A. Segel and J. L. Jackson. Dissipative structure: an explanation and an ecological example. Journal of Theoretical Biology, 37(3):545559, 1972. → pages 2 233 [54] M. B. Short, M. R. DOrsogna, V. B. Pasour, G. E. Tita, P. J. Brantingham, A. L. Bertozzi, and L. B. Chayes. A statistical model of criminal behavior. Mathematical Models and Methods in Applied Sciences, 18(S1):12491267, 2008. → pages 3 [55] N. Sloane. Tables of optimal arrangements of points on a sphere, 2000. → pages 119 [56] W. J. Stortelder, J. J. de Swart, and J. D. Pintr. Finding elliptic fekete points sets: two numerical solution approaches. Journal of computational and applied mathematics, 130(1-2):205–216, 2001. → pages 5 [57] D. Thompson. On growth and form. Cambridge University Press, 1963. → pages 3 [58] J. J. Thomson, H. Geiger, E. Marsden, E. Rutherford, H. Moseley, and J. Chadwick. On the structure of the atom. Phylos Mag, 7:237265, 1904. → pages 5, 31, 120 [59] W. H. Tse, J. Wei, and M. Winter. Spikes for the Gierer-Meinhardt system with many segments of different diffusivities. Bulletin of the Institute of Mathematics, Academia Sinica, 3(3-4):525, 2008. → pages 3 [60] A. M. Turing. The chemical theory of morphogenesis. Phil. Trans. Roy. Soc, 237:32, 1952. → pages 1, 127 [61] C. Varea, J. L. Arag´on, and R. A. Barrio. Turing patterns on a sphere. Physical Review E, 60:4588–4592, 1999. → pages 3, 119, 121, 123 [62] W. Wang, Q. Liu, and Z. Jin. Spatiotemporal complexity of a ratio-dependent predator-prey system. Physical Review E, 75(5):051913, 2007. → pages 2, 4, 216 [63] J. Wei. On single interior spike solutions of the Gierer-Meinhardt system: uniqueness and spectrum estimates. European Journal of Applied Mathematics, 10(4):353378, 1999. → pages 5, 67, 70, 72, 126, 238, 243 [64] D. Xiao and S. Ruan. Global dynamics of a ratio-dependent predator-prey system. Journal of Mathematical Biology, 43(3):268ł290, 2001. → pages 216 234 Appendix A Spherical coordinate tranformations Lemma A.1 Consider a point on the unit sphere xj = (cos φj sin θ j , sin φj sin θ j , cos θ j ), with 0 < φj < 2π, 0 < θ j < π. We will now show that as x → xj then |x − xj | = |ˆ y | + small terms, ˆ φ); ˆ where yˆ = (yˆ1 , yˆ2 ) = (sin θ j φ, and where φˆ = φ − φj , and θˆ = θ − θ j , with φˆ 1, and θˆ 1. Proof Using the standard spherical coordinate transformation, we let x = cos φ sin θ , y = sin φ sin θ , 235 z = cos θ . ˆ θ = θ j + θˆ , we get Linearizing φ = φj + φ, x = xj − sin φj sin θ j φˆ + cos φj cos θ j θˆ + · · · , y = yj + cos φj sin θ j φˆ + sin φj cos θ j θˆ + · · · , z = zj − sin θ j θˆ + · · · . In matrix form we have x = xj + M φˆ , θˆ − sin φj sin θ j cos φj cos θ j 0 − sin θ j with M = cos φj sin θ j sin φj cos θ j with M a 3 × 2 matrix. We then have |x − xj |2 = (x − xj )T (x − xj ) = (M0 yˆ)T (M0 yˆ) = yˆT M0T M0 yˆ, with yˆ = yˆ1 yˆ2 = − sin φj sin θ j sin θ j φˆ , and M0 = cos φj sin θ j θˆ 0 We can now check that M0T M0 = I. Therefore, we conclude that as x → xj , we have |x − xj | = |ˆ y | + small terms. 236 cos φj cos θ j sin φj cos θ j − sin θ j Appendix B Rigorous properties of NLEPs The nonlocal eigenvalue problem in (2.100) has the general form ∆ρ ψ˜ − ψ˜ + 2wψ˜ − γw2 ∞ ˜ 0 ρw ψdρ ∞ 2 0 ρw dρ ˜ = λψ, 0 < ρ < ∞, (B.1) with ψ˜ (0) = 0 and ψ˜ → 0 as ρ → ∞. Here γ = γ(λ) is an analytic function of λ in the right half-plane Re(λ) > 0, and w is the radially symmetric ground-state solution of ∆ρ w − w + w2 = 0 where ∆ρ v ≡ v + ρ−1 v . Remark (i) Since this problem is not self-adjoint we must expect that complex eigenvalues are possible. We remark that an NLEP of the form ˜ − A(x) L0 Ψ ∞ ˜ = λψ˜ B(x)ψdx −∞ is self-adjoint if and only if the operator L0 is self-adjoint, and A(x) = cB(x) ∞ ˜ for some c independent of x. Since instead we have w2 (ρ) ρw(ρ)ψdρ, our 0 237 NLEP is not self-adjoint. A further complication arises because the multiplier γ also depends on λ. (ii) The local operator in (B.1) is defined by ˜ L0 ψ˜ ≡ ∆ρ ψ˜ − ψ˜ + 2wψ. It is well-known that the local problem L0 ψ˜ = σ ψ˜ has a unique positive eigenvalue σ0 > 0 with even eigenfunction ψ˜0 > 0 (see [63]). Since the nonlocal term vanishes identically for eigenfunctions that are odd, hence reducing the NLEP to the local problem without nonlocal term, we need only consider the even eigenfunctions of (B.1). Next, we will convert (B.1) into a transcendental equation in λ. We write (B.1) as (L0 − λ)ψ˜ − γw2 J = 0, where J ≡ ∞ ˜ 0 ρw ψdρ , ∞ 2 0 ρw dρ This shows that ψ˜ = γJ(L0 − λ)−1 w2 , and hence J= ∞ 0 ρw γJ[L0 − λ]−1 w2 dρ = γJ ∞ 2 0 ρw dρ ∞ −1 2 0 ρw(L0 − λ) w dρ . ∞ 2 0 ρw dρ We need only consider the eigenfunctions for which J = 0 (since if J = 0 the eigenfunctions of the local problem are well known). Therefore, we can cancel J in the equation above and obtain that λ is a root of g(λ) = 0, where g(λ) = C(λ) − F(λ), 1 C(λ) = , γ(λ) F(λ) = ∞ −1 2 0 ρw(L0 − λ) w dρ . ∞ 2 0 ρw dρ (B.2) In terms of the roots of g(λ) = 0, our stability criterion is as follows. We have instability if there exists λ with Re(λ) > 0 such that g(λ) = 0. We have stability 238 if for all roots of g(λ) = 0 we have Re(λ) < 0. Theorem B.1 Let λ be real. Then, the properties of F(λ) are as follows: • F(0) = 1, F → +∞ as λ → σ0− , F ≤ 0 for λ > σ0 . • F (λ) > 0 for 0 < λ < σ0 . F → 0 as λ → +∞. ˜ Here σ0 > 0 is the unique positive eigenvalue of the local problem L0 ψ˜ = σ ψ. Proof (i) • Recall that L0 w = w2 . In other words, L0 w = ∆w − w + 2w2 = w2 . Thus, F(0) = ∞ −1 2 0 ρwL0 w dρ ∞ 2 0 ρw dρ = ∞ 0 ρw(w)dρ ∞ 2 0 ρw dρ = 1. • Now, (L0 − λ)−1 does not exist at λ = σ0 the unique positive eigenvalue of L0 . Hence (L0 − λ)−1 is unbounded at λ approaches σ0 . F → +∞ as λ → σ0− ; F → −∞ as λ → σ0+ . • The proof that F < 0 for λ > σ0 is more technical and is based on the following lemma: Lemma B.1 Let ξ(ρ) be a solution to (L0 − λ)ξ = v, on 0 ≤ ρ < ∞, with ξ (0) = 0 and ξ → 0 as ρ → ∞. Assume that v is smooth, with v > 0 on 0 < ρ < ∞, and v → 0 as ρ → ∞. Then if λ > σ0 we have ξ ≤ 0 for ρ ≥ 0. 239 Proof Assume to the contrary that there exists ρ0 > 0 with ξ(ρ0 ) > 0. Then by continuity of ξ, ξ(ρ) > 0 on ρ ∈ (ρ1 , ρ2 ), with either (i) ξ(ρ1 ) = ξ(ρ2 ) = 0, ξ (ρ1 ) ≥ 0, ξ (ρ2 ) ≤ 0, ρ1 < ρ2 . Or (ii) ρ1 = 0 with ξ(0) ≥ 0, ξ (0) = 0, ξ(ρ2 ) = 0, ξ (ρ2 ) ≤ 0. Let L0 ψ˜0 = σ0 ψ˜0 with σ0 > 0 and ψ˜0 > 0 since it is the first eigenfunction. We then use Green’s identity to ξ and ψ˜0 on the subinterval ρ1 < ρ < ρ2 to get ρ2 ψ˜0 L0 ξ − ξL0 ψ˜0 ρdρ = ρ ψ˜0 ξ − ξ ψ˜0 ρ1 ρ2 ρ1 Note that −ρξ ψ˜0 ρ2 ρ1 ρ1 ρ1 ρψ˜0 [λξ + v] − ρξσ0 ψ˜0 dρ = ρψ˜0 ξ ρ2 ρ1 ρ2 ρ1 , . = 0 in either (i) or (ii). This becomes ρψ˜0 vdρ = (σ0 − λ) ρ2 ρψ˜0 ξdρ + ρψ˜0 ξ ρ1 ρ2 ρ1 . We have that the first term is positive since both ψ˜0 , v > 0, whereas the second term is negative since λ > σ0 , and the third term is ≤ 0 by (i) and (ii). This is a contradiction, hence we conclude that ξ < 0 ∀ρ. As a consequence, ξ = (L0 − λ)−1 w2 ≤ 0 when λ > σ0 , which implies that F(λ) = ∞ 0 ρw (L0 − λ)−1 w2 dρ <0 ∞ 2 0 ρw dρ for λ > σ0 . (ii) Now F → 0 as λ → ∞ is evident since (L0 − λ)−1 = O(λ−1 ) for λ 240 1. Next we use L0 w = w2 to write ∞ ∞ −1 2 −1 0 ρw(L0 − λ) w dρ 0 ρw(L0 − λ) L0 wdρ = , ∞ ∞ 2 2 0 ρw dρ 0 ρw dρ ∞ −1 [(L − λ)w + λw] dρ 0 0 ρw(L0 − λ) , ∞ 2 0 ρw dρ ∞ ∞ −1 −1 0 ρw w + λ(L0 − λ) w dρ 0 ρwλ(L0 − λ) wdρ = 1 + λ ∞ ∞ 2 2 0 ρw dρ 0 ρw dρ F(λ) = = = . From this last expression we can readily calculate F (λ) as F (λ) = ∞ −1 0 ρw(L0 − λ) wdρ ∞ 2 0 ρw dρ +λ ∞ −2 0 ρw(L0 − λ) wdρ . ∞ 2 0 ρw dρ Then, we can integrate by parts on the second integral to obtain F (λ) = ∞ −1 0 ρw(L0 − λ) wdρ ∞ 2 0 ρw dρ ∞ 0 +λ so that F (λ) = h(λ) +λ ∞ 2 0 ρw dρ where h(λ) is defined by h(λ) ≡ (L0 − λ)−1 w (L0 − ∞ 2 0 ρw dρ ∞ 0 (L0 − λ)−1 w ∞ 2 0 ρw dρ ∞ 0 ρw(L0 2 ρdρ λ)−1 w ρdρ , , (B.3) − λ)−1 wdρ. The second term in F (λ) is positive for λ > 0. Therefore, in order to prove that F (λ) > 0 on 0 < λ < σ0 it suffices to prove that h(λ) > 0 on 0 < λ < σ0 . To establish the positivity of h(λ) we will use a simple Calculus argument to show that h(0) > 0 and h (λ) > 0 on 0 < λ < σ0 . We first use the remarkable ∞ −1 0 ρwL0 wdρ that L0 w + 21 ρw 1 identity L−1 0 w = w + 2 ρw to show that h(0) = identity is readily derived by the direct verification calculate h(0) as ∞ h(0) = 0 1 ρw w + ρw 2 ∞ dρ = 0 241 ρw2 dρ + 1 2 ∞ 0 > 0. This = w. We ρ2 (ww )dρ. To determine the sign of this quantity we use integration by parts to get ∞ 1 ∞ 2 d 2 ρ (w )dρ, 4 0 dρ 0 ∞ ∞ 1 ∞ 2 1 1 ρw dρ = ρw2 dρ + ρ2 w2 − = 4 2 2 0 0 0 h(0) = ρw2 dρ + ∞ ρw2 dρ > 0. 0 This shows that h(0) > 0. Furthermore, we calculate that ∞ h (λ) = 0 ρw(L0 − λ)−2 wdρ = ∞ 0 ρ (L0 − λ)−1 w 2 dρ > 0. In addition, h(λ) → +∞ as λ → σ0− . Hence h(λ) > 0 on 0 < λ < σ0 . Thus, by (B.3), F (λ) > 0 on 0 < λ < σ0 . This concludes the proof. Next, we return to (B.2). We conclude that if C(0) > 1 and C is analytic in Re(λ) ≥ 0, then the curves C(λ) and F(λ) must cross at some λ > 0 real on the interval 0 < λ < σ0 . Theorem B.2 Suppose C(0) > 1, C(λ) is analytic in Re(λ) ≥ 0. Then there exists an unstable eigenvalue to the NLEP in Re(λ) > 0. Implication Suppose that γ(0) < 1. Then there exists an unstable eigenvalue on 0 < λ < σ0 that is real. This is precisely the criterion that was used in establishing the principal result (2.110) of §2.5.1. For instance, consider the competition instability threshold studied in §2.5.1. Then, from (2.108) γ(λ) = 2[f − (λ + 1)] . f − (λ + 1)(1 + ϕD0 ) 242 We calculate γ(0) = 2(f − 1) . f − (1 − ϕD0 ) Our rigorous result above shows that if γ(0) < 1 then we have an unstable real eigenvalue. Thus, we obtain such an instability when (see (2.110) of §(2.5.1)) D0 > D0c ≡ 4f 2 . N 2 (1 − f )b It is much more difficult to prove the converse, namely that we have stability for the competition modes when D0 < D0c . Although we anticipate that such a result is true based on our full numerical simulations of the Brusselator model, we have been unable to prove it from the NLEP. The technical difficulty with completing a rigorous proof is that one must track all the complex eigenvalues of the NLEP. We recall that if γ was a constant, independent of λ, and that γ > 1, then such a stability proof was given in [63]). 243
- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- Topics in the stability of localized patterns for some...
Open Collections
UBC Theses and Dissertations
Featured Collection
UBC Theses and Dissertations
Topics in the stability of localized patterns for some reaction-diffusion systems Rozada, Ignacio 2012
pdf
Page Metadata
Item Metadata
Title | Topics in the stability of localized patterns for some reaction-diffusion systems |
Creator |
Rozada, Ignacio |
Publisher | University of British Columbia |
Date Issued | 2012 |
Description | In the first part of this thesis, we study the existence and stability of multi-spot patterns on the surface of a sphere for a singularly perturbed Brusselator and Schnakenburg reaction-diffusion model. The method of matched asymptotic expansions, tailored to problems with logarithmic gauge functions, is used to construct both symmetric and asymmetric spot patterns. There are three distinct types of instabilities of these patterns that are analyzed: self-replication instabilities, amplitude oscillations of the spots, and competition instabilities. By using a combination of spectral theory for nonlocal eigenvalue problems together with numerical computations, parameter thresholds for these three different classes of instabilities are obtained. For the Brusselator model, our results point towards the existence of cycles of creation and destruction of spots, and possibly to chaotic dynamics. For the Schnakenburg model, a differential-algebraic ODE system for the motion of the spots on the surface of the sphere is derived. In the second part of the thesis, we study the existence and stability of mesa solutions in one spatial dimension and the corresponding planar mesa stripe patterns in two spatial dimensions. An asymptotic analysis is used in the limit of a large diffusivity ratio to construct mesa patterns in one spatial dimension for a general class of two-component reaction-diffusion systems that includes the well-known Gierer Meinhardt activator-inhibitor model with saturation (GMS model), and a predator-prey model. For such one-dimensional patterns, we study oscillatory instabilities of the pattern by way of a Hopf bifurcation and from a reduction to a limiting ODE-PDE system. In addition, explicit thresholds are derived characterizing transverse instabilities of planar mesa-stripe patterns in two spatial dimensions. The results of our asymptotic theory as applied to the GMS and predator-prey systems are confirmed with full numerical results. |
Genre |
Thesis/Dissertation |
Type |
Text |
Language | eng |
Date Available | 2012-08-22 |
Provider | Vancouver : University of British Columbia Library |
Rights | Attribution-ShareAlike 3.0 Unported |
DOI | 10.14288/1.0073037 |
URI | http://hdl.handle.net/2429/43012 |
Degree |
Doctor of Philosophy - PhD |
Program |
Mathematics |
Affiliation |
Science, Faculty of Mathematics, Department of |
Degree Grantor | University of British Columbia |
Graduation Date | 2012-11 |
Campus |
UBCV |
Scholarly Level | Graduate |
Rights URI | http://creativecommons.org/licenses/by-sa/3.0/ |
Aggregated Source Repository | DSpace |
Download
- Media
- 24-ubc_2012_fall_rozada_ignacio.pdf [ 5.99MB ]
- Metadata
- JSON: 24-1.0073037.json
- JSON-LD: 24-1.0073037-ld.json
- RDF/XML (Pretty): 24-1.0073037-rdf.xml
- RDF/JSON: 24-1.0073037-rdf.json
- Turtle: 24-1.0073037-turtle.txt
- N-Triples: 24-1.0073037-rdf-ntriples.txt
- Original Record: 24-1.0073037-source.json
- Full Text
- 24-1.0073037-fulltext.txt
- Citation
- 24-1.0073037.ris
Full Text
Cite
Citation Scheme:
Usage Statistics
Share
Embed
Customize your widget with the following options, then copy and paste the code below into the HTML
of your page to embed this item in your website.
<div id="ubcOpenCollectionsWidgetDisplay">
<script id="ubcOpenCollectionsWidget"
src="{[{embed.src}]}"
data-item="{[{embed.item}]}"
data-collection="{[{embed.collection}]}"
data-metadata="{[{embed.showMetadata}]}"
data-width="{[{embed.width}]}"
async >
</script>
</div>
Our image viewer uses the IIIF 2.0 standard.
To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0073037/manifest