Pattern formations and spatial entropy for spatially
discrete diffusion equations
Chang-Yuan Cheng, Chih-Wen Shih
∗Department of Applied Mathematics, National Chiao Tung University, 1001 Ta-Hsueh Road, Hsinchu 300, Taiwan, ROC
Received 2 March 2005; accepted 8 April 2005 Communicated by A. Pikovsky
Abstract
Formation of mosaic patterns for spatially discrete diffusion equations with cubic nonlinearity is investigated. We construct feasible basic patterns in each parameter region and combine these basic patterns into large patterns on one- and two-dimensional lattices. The basic patterns are characterized and constructed through formulating parameter conditions based on a geometrical setting. Spatial entropy associated with these patterns are computed or estimated. We also consider three typical boundary conditions and investigate their influences on pattern formations and spatial entropy. Several numerical computations are performed to illustrate such a formation of patterns.
© 2005 Elsevier B.V. All rights reserved.
PACS: 02.30.Hq; 05.45.−a; 89.90.+n
Keywords: Pattern formation; Spatial entropy; Spatial chaos; Lattice systems
1. Introduction
In this presentation, we investigate spatial patterns of the following spatially discrete diffusion equa-tions:
dui
dt = βui+ αf (ui), ui := ui+1+ ui−1− 2ui, (1.1)
∗Corresponding author. Tel.: +886 3 5722088; fax: +886 3 5724679.
E-mail address: [email protected] (C.-W. Shih).
0167-2789/$ – see front matter © 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.physd.2005.04.007
wherei ∈ Λ1⊆ Z1, or dui,j
dt = β
++u
i,j+ β××ui,j+ αf (ui,j), +ui,j:= ui+1,j+ ui−1,j+ ui,j+1+ ui,j−1− 4ui,j, ×ui,j:= ui+1,j+1+ ui+1,j−1+ ui−1,j+1+ ui−1,j−1− 4ui,j, (1.2) where (i, j) ∈ Λ2⊆ Z2, andΛ1andΛ2are connected subsets ofZ1andZ2, respectively. Herein, we consider a typical cubic nonlinearity
f (ξ) = ξ3− ξ. (1.3)
The present approach can be extended to(1.1) and (1.2)with other nonlinearity and other lattice dynamical system, continuous-time or discrete-time.
Stationary solutions (patterns) constitute fundamental structure for differential equations. This presentation at-tempts to extend previous studies on lattice dynamical systems to further generality. Moreover, it is hoped to contribute toward treating the problems of allocating the parameters with which the considered system exhibits de-sirable patterns or some specific behaviors. Such problems are a kind of inverse problems and have been attracting much scientific interests. In this work, we are especially interested in a class of stationary patterns called mosaic patterns. We shall present a methodology for constructing mosaic patterns of the above systems. These patterns are characterized and constructed through formulating parameter conditions based on a geometrical setting. Stability of these patterns can also be investigated through estimating their basins of attraction, under further parameter conditions.
Formation of mosaic patterns and their spatial entropy for systems(1.1) and (1.2)have been investigated in[1–3], with the double-obstacle nonlinearity:
f (ξ) = (−∞, −γ] if ξ = −1, γξ if|ξ| < 1, [γ, ∞) ifξ = 1, ∅ if|ξ| > 1, (1.4)
which is a set-valued function. The mosaic patterns and solutions therein take the value ui or ui,j= −1, 0, 1. Same considerations were adopted on Cahn–Hilliard equation in[4,5]. In this work, we employ the basic pattern formulation to discuss formation of mosaic patterns and spatial entropy for(1.1) and (1.2), with cubic nonlinearity
(1.3). Our treatments are motivated by numerical spirit as well as the sense from real-world pattern formations. We consider the component of the stationary solutions to lie within small ranges, instead of being some single exact value, namely
uiorui,j∈ [−1 − σ, −1 + σ] ∪ [−σ, σ] ∪ [1 − σ, 1 + σ], (1.5)
whereσ is a small number. Indeed, if a pattern in nature is represented by or is a presentation of certain quantities, these quantities are likely lying in small ranges, under a tolerance of error. The approach employed here is an extension from the work [6] on mosaic patterns of cellular neural networks. One first explores feasible basic patterns under various parameter conditions. These basic patterns are then combined through an attaching process to form patterns of larger sizes. The componentyioryi,jof mosaic patterns (output patterns) in[6]takes the value −1, 1. Herein, the attaching process needs to be modified since components of the basic patterns to be overlapped may take different values, although they lie in the same interval in one of(1.5). We propose a fixed-point argument to assure the validity of such an attaching process. The performance of this fixed-point argument is based on our geometric formulation on the parameter conditions.
IfΛ1 orΛ2 is finite, boundary conditions need to be imposed to have a well-defined system. An interested problem for systems(1.1) and (1.2)has been raised in[7]:
h = hN = hP= hD?
Herein, h denotes the spatial entropy, andhN,hPandhD, respectively, represents the spatial entropy for the same type of patterns satisfying Neumann, periodic and Dirichlet boundary conditions. Such a problem has been investigated in[8]with examples from cellular neural networks. With the present approach, the effect of boundary conditions upon pattern formations and spatial entropy for(1.1) and (1.2)can be analogously investigated. Notably, only infinite latticesZ1,Z2, and thus no boundary effects, were considered in[1–3].
Other frequently considered nonlinearities for (1.1) and (1.2)include the cubic polynomial f (ξ) = γξ + ξ3, f (ξ) = (ξ2− 1)(ξ − a), and the logarithmic nonlinearity f (ξ) = γξ + ln[(1 + ξ)/(1 − ξ)] which restricts the range of its argument to−1 < ξ < 1. Our results can be adapted to(1.1) and (1.2)with these nonlinearities. It actually can be generalized to constructing stationary states of other lattice systems with components near finite number of specific values.
Lattice dynamical systems have been attracting great scientific interests, especially in chemical reactions[9], image processing and patterns recognition[10,11], material science[12,13], and biology[14,15,19].
Asβ in(1.1)orβ+, β×in(1.2)is large, our results can be compared to the PDE case, namely the Allen–Cahn or the Nagumo equation:
∂u
∂t = dµu + f (u), (1.6)
on a one-dimensional interval domain with the Laplacianu = ∂2u/∂x2or on a two-dimensional square domain with u = ∂2u/∂x2+ ∂2u/∂y2, and with certain boundary conditions. In addition, discretization of partial differential equations and systems of partial differential equations can be regarded as lattice systems. Thus, the approach herein is also related to numerical solutions of the corresponding partial differential equations. There have been circuit implementations for simulating nonlinear PDEs via autonomous cellular neural networks[11]. Those PDEs include wave equations and reaction-diffusion equations. This study also provides a theoretical basis for pattern formation in these circuit implementations.
In the following, we write the spatially discrete diffusion equations as (sd-DE) as an abbreviation. The rest of this paper is organized as follows. In Section2, we introduce a geometric formulation to partition the parameter space. Corresponding to each partitioned parameter region, there exists a collection of basic patterns. In Section3, the basic patterns established in Section2are confirmed to be feasible basic patterns for (sd-DE), by applying a fixed-point theorem. One can then combine these basic patterns through an attaching process into mosaic patterns. We investigate stability of the mosaic patterns in Section4. In Section5, for mosaic patterns on one-dimensional lattice, transition matrices are formulated to describe the formation of patterns and compute the spatial entropy. In addition, the entropy for patterns on two-dimensional lattice is estimated. In Section6, we investigate the influence of boundary conditions upon pattern formation as well as the problem: h = hN = hP= hD? We provide some numerical illustrations for two-dimensional patterns in Section6.
2. Partitioning parameter space and basic patterns
In this section, we shall introduce the mosaic solutions and mosaic patterns for(1.1) and (1.2). The mosaic patterns are piled up through an attaching process on the so-called basic patterns. We propose a geometrical formulation to characterize the existence of basic patterns and derive the parameter conditions for such an existence. The methodology we propose is valid for systems(1.1) and (1.2)on both finite lattices and infinite lattices. The infinite lattices we consider herein is the wholeZ1orZ2. As a representative of finite lattices, we consider the rectangular
ones:
Λ1= Tk= {i ∈ Z1|1 ≤ i ≤ k}, (2.1)
Λ2= Tk= {(i, j) ∈ Z2|1 ≤ i ≤ k1, 1 ≤ j ≤ k2}, (2.2)
for casesd = 1 and d = 2, respectively, where k, k1, k2are positive integers. The results herein can be extended to other lattices and lattices of higher dimensions.
For(1.1)onTkor(1.2)onTk, boundary conditions need to be imposed so that the equations at boundary sites
are well defined. There are three typical types of boundary conditions: (i) Neumann boundary condition:
u0= u1, uk+1= uk,
ford = 1. For d = 2, 0 ≤ i ≤ k1+ 1 and 0 ≤ j ≤ k2+ 1: u0,j= u1,j, uk1+1,j = uk1,j,
ui,0= ui,1, ui,k2+1= ui,k2.
(ii) Periodic boundary condition: u0= uk, uk+1= u1,
ford = 1. For d = 2, 0 ≤ i ≤ k1+ 1 and 0 ≤ j ≤ k2+ 1: u0,j= uk1,j uk1+1,j= u1,j,
ui,0= ui,k2 ui,k2+1= ui,1.
(iii) Dirichlet boundary condition: ui= ˜ui,
for i in the exterior neighbors b of the boundary sites, where ˜uiare prescribed data and b := {0, k + 1} if d = 1
and b := {(i, 0), (0, j), (k1+ 1, j), (i, k2+ 1) | 0 ≤ i ≤ k1+ 1, 0 ≤ j ≤ k2+ 1} if d = 2.
For convenience of discussion, the prescribed boundary data ˜ui also take the values as in(1.5). Systems(1.1)
onTkor(1.2)onTkwith the Neumann, periodic, and Dirichlet boundary conditions are denoted by (sd-DE)N, (sd-DE)P, and (sd-DE)D, respectively. These systems are regular ordinary differential equations on Euclidean spaces. Notably,(1.1)on infinite latticeZ1or(1.2)onZ2is a system of differential equations on infinite-dimensional vector space. Fundamental theory on existence and uniqueness of solutions for such systems can be found in[16]. Let 0< σ < 1/11 be a fixed number. The reason for requiring σ < 1/11 will be clear later.
Definition 2.1. We say that a stationary solution u= {ui}i∈Λd of(1.1)or(1.2)is a mosaic solution if
ui ∈ [−1 − σ, −1 + σ] ∪ [−σ, σ] ∪ [1 − σ, 1 + σ],
for all i∈ Λd. We denote byMσ1(α, β) and Mσ2(α, β+, β×) the set of all mosaic solutions for(1.1)with parameters α, β and(1.2)with parametersα, β+, β×, respectively.
We employ the symbols⊕, , ⊗ to characterize such mosaic solutions. Restated, we call {si}i∈Λd the
corre-sponding mosaic pattern of a mosaic solution{ui}i∈Λd, where
si= ⊕, if 1 − σ ≤ ui≤ 1 + σ.
si= ⊗, if − σ ≤ ui≤ σ,
si= , if − 1 − σ ≤ ui≤ −1 + σ.
(2.3)
We call a 1× 3 (respectively, 3 × 3) array of ⊕, ⊗, , in the case d = 1 (respectively, d = 2), a basic pattern. There are totally 33possible basic patterns in the cased = 1 and 39possible basic patterns in the cased = 2, namely
• • •, • • • • • • • • •
, • = ⊕, ⊗, .
We denote by N1(i) = {i − 1, i, i + 1}, N2(i, j) = {(i + 1, j), (i − 1, j), (i, j + 1), (i, j − 1), (i, j), (i + 1, j + 1), (i + 1, j − 1), (i − 1, j + 1), (i − 1, j − 1)} the nearest neighbors of i and (i, j), respectively. Let u = {ui}i∈Λd
be a mosaic solution according to the above definition and let{si}i∈Λd be the corresponding mosaic pattern. We call
the projection (or restriction) of{si}i∈Λd onto the nearest neighborsN1(i) for the case of d = 1, and N2(i, j) for the
case ofd = 2, a feasible basic pattern, for any interior sites i of Λ1and (i, j) of Λ2, respectively.
A scheme for constructing mosaic patterns may go the other way around. If one can find out the feasible basic patterns for(1.1) and (1.2), then attaching these basic patterns compatibly produces patterns of larger sizes. Mosaic patterns can be obtained through such an attaching successively. This is basically the approach in[6]for constructing mosaic patterns of cellular neural networks. In cellular neural networks, a stationary solution x= {xi}
is called mosaic if the output ofxiis either exactly 1 or−1. A successful attaching yields a corresponding solution
automatically. The situation is different herein, as the component ofuiis only required to lie in a range as indicated
in(2.3). We will discuss the attaching process and justify how such a process yields a solution in Section3. We shall call those feasible basic patterns that can be confirmed by our theory in Section3“affirmatively” feasible basic
pattern.
The crucial part in the above-mentioned pattern formation scheme is allocating the parameters in(1.1)or(1.2)
to identify the existence of basic patterns. We take the cased = 1 to illustrate the idea. The stationary equation for
(1.1)is
β(ui+1+ ui−1− 2ui)+ αf (ui)= 0. (2.4)
We assumeα = 0 and set b = β/α. For a fixed i, given ui−1 andui+1,u∗i satisfies(2.4)if and only if there is an intersection (u∗i, y∗) for curves
y = b[2ui− ui−1− ui+1], (2.5)
y = f (ui), (2.6)
cf.Fig. 1. Therefore, the configurations for the graphs of these two functions determine the existence of the feasible basic patterns.
Let us use the following example to illustrate the construction of basic patterns. Given ˜ui−1, ˜ui+1∈ [−σ, σ], if there is an intersection for(2.5) and (2.6)withui−1= ˜ui−1,ui+1= ˜ui+1atu∗i ∈ [1 − σ, 1 + σ], then we have a candidate for feasible basic pattern⊗ ⊕ ⊗ corresponding to the three tuple (˜ui−1, u∗i, ˜ui+1). In order to guarantee such an intersection, we need to restrict the value of b such that the graph of f betweenL1:y = 2bx + 2bσ and L2:y = 2bx − 2bσ lies entirely in the shadow region R which is bounded by x = 1 + σ and x = 1 − σ, as indicated inFig. 1. It can be computed that such an intersection always holds if 0≤ b ≤ f (1+σ)2+4σ or−f (−1+σ)2 ≤ b ≤ 0. With
Fig. 1. Configuration of intersection for Eqs.(2.5) and (2.6).
our formulation, it will be shown in the next section that such candidates of feasible basic patterns will turn out to be real feasible basic patterns.
Through analyzing these geometrical configurations, we can characterize and classify the existence of all 27 basic patterns. The parameter spaceP1= {b : b ∈ R} can be partitioned into finitely many regions so that(1.1)has the same collection of affirmatively feasible basic patterns for parameters in each region. Through computations, it is found that some feasible basic patterns exist in groups. We thus introduce the following notations:
B• {m1,m2,...,mk} = l=m1,...,mk B• l,
whereBl•, l = 0, ±1, ±2, “ • ” = ⊕, ⊗, , are described inTable 1. The superscript bullet “•” herein means the symbol at the center of a basic pattern and the integer in the subscript indicates the states in its neighbor. Thorough computations yield the following classification for the existence of feasible basic patterns.
Theorem 2.2. Suppose that 0< σ <111 is fixed. The parameter spaceP1= {b : b ∈ R} can be partitioned so that the set of feasible basic patterns for(1.1)with(1.3)and parameters in each region contains the ones described in
Table 2.
The reason for considering 0< σ < 111 is to avoid overlap of the partitioned intervals inFig. 2. Confirmations for the feasibility of basic patterns inTheorem 2.2are in fact completed in Section3, in respecting our definition of feasible basic pattern. We remark that there may be other intersections for Eqs.(2.5) and (2.6)and thus other possibilities for the existence of feasible basic patterns for each set of parameters. Further partitioning of parameter space can be carried out to capture these possible intersections. The feasible basic patterns we list inTable 2are the ones which can be confirmed by the theory in Section3. We display, in the left half ofFig. 2, in each parameter
Table 1
Notations for collections of basic patterns,• = ⊕, ⊗,
Notation Basic patterns
B• 2 ⊕ • ⊕ B• 1 ⊕ • ⊗, ⊗ • ⊕ B• 0 ⊕ • , ⊗ • ⊗, • ⊕ B• −1 ⊗ • , • ⊗ B• −2 •
Table 2
Affirmatively feasible basic patterns corresponding to each parameter region in the cased = 1
Parameter region Affirmatively feasible basic patterns
I7= f (−1 + σ) 4σ , ∞ B⊗{0} I6= f (1 + σ) 1+ 4σ , f (−1 + σ) 4σ B⊕ {2}, B⊗{0}, B{−2} I5= f (−σ),f (1 + σ) 1+ 4σ B⊕ {2,1}, B⊗{0}, B{−1,−2} I4= f (1 + σ) 2+ 4σ , f (−σ) B⊕{2,1}, B⊗{1,0,−1}, B{−1,−2} I3= f (1 + σ) 3+ 4σ , f (1 + σ) 2+ 4σ B⊕ {2,1,0}, B⊗{1,0,−1}, B{0,−1,−2} I2= f (1 + σ) 4+ 4σ , f (1 + σ) 3+ 4σ B⊕ {2,1,0,−1,}, B⊗{1,0,−1}, B{1,0,−1,−2} I1= f (−σ) 2 , f (1 + σ) 4+ 4σ B⊕{2,1,0,−1,−2}, B{2,1,0,−1,−2}⊗ , B{1,0,−1} I0= −f (−σ) 2+ 4σ, f (−σ) 2 B⊕ {2,1,0,−1,−2}, B{2,1,0,−1,−2}⊗ , B{2,1,0,−1,−2} I−1= −f (−1 + σ) 4 , − f (−σ) 2+ 4σ B⊕ {2,1,0,−1,−2}, B{2,1,0,−1,−2}⊗ , B{1,0,−1} I−2= −f (−1 + σ)3 , −f (−1 + σ)4 B⊕{2,1,0,−1,}, B⊗{1,0,−1}, B{1,0,−1,−2} I−3= −f (−σ) 1+ 4σ, − f (−1 + σ) 3 B⊕ {2,1,0}, B⊗{1,0,−1}, B{0,−1,−2} I−4= −f (−1 + σ) 2 , − f (−σ) 1+ 4σ B⊕ {2,1,0}, B⊗{0,−1,−2}, B{0} I−5= −f (−1 + σ), −f (−1 + σ)2 B⊕{2,1}, B⊗{0}, B{−1,−2} I−6= −f (−σ) 4σ , −f (−1 + σ) B⊕ {2}, B⊗{0}, B{−2} I−7= −∞, −f (−σ)4σ B⊕ {2}, B{−2}
region, the existence of feasible basic patterns which can be confirmed by our treatment, and in the right half of
Fig. 2, with further partitioning on the parameter space, the existence of all other possible basic patterns. We will address more on that as we estimate the entropy of the system in Section5.
Let us also describe the partitioning of parameters and corresponding existence of basic patterns for the case of two-dimensional lattice, i.e., for Eq.(1.2). The formulation is analogous to the one-dimensional case. The stationary equation for(1.2)is
β++ui,j+ β××ui,j+ αf (ui,j)= 0, (2.7)
for (i, j) ∈ Λ2⊆ Z2. We assumeα = 0 and set b1= β+/α, b2= β×/α. Then, for fixed (i, j),(2.7)holds if and only if there is an intersection for curves
y = −b1+ui,j− b2×ui,j, (2.8)
Fig. 2. Partition of parameter space and feasible basic patterns. A1=f (−1+σ)4σ , A2= f (1 + σ), A3=f (1+σ)1+4σ , A4=f (−σ)1−4σ, A5=f (1+σ)2 , A6= f (−σ), A7=f (1+σ)2+4σ , A8= f (1+σ)3 , A9= f (1+σ)3+4σ , A10= f (−σ)2−4σ, A11= f (1+σ)4 , A12= f (1+σ)4+4σ , A13=f (−σ)2 , A14= −f (−σ)2+4σ, A15= −f (−1+σ)4 , A16= −f (−1+σ)4−4σ , A17= −f (−σ)2 , A18= −f (−1+σ)3 , A19= −f (−1+σ)3−4σσ , A20= −f (−σ)1+4σ, A21= −f (−1+σ)2 , A22= −f (−σ), A23= −f (−1+σ)2−4σ , A24= −f (−1 + σ), A25= − f (−1+σ) 1−4σ , A26= − f (−σ) 4σ .
The parameter space can be partitioned so that the set of feasible basic patterns for(1.2)with(1.3)and parameters in each partitioned region are the same as the cased = 1.
3. From basic patterns to mosaic patterns
Let us describe the attaching process on the basic patterns and justify that the process indeed produces corre-sponding solutions for(1.1) and (1.2), under our setting and formulations in Section2. Consider two basic patterns
sp= • p1p2and sq= q1q2•, “ • ”, p1, p2, q1, q2= ⊕, ⊗, . We say that the basic pattern sq can be attached, with two sites overlapped, to the right of basic pattern sp, ifq1= p1, q2= p2. For example, attaching sq= ⊕ to the right of sp= ⊕ with two sites overlapped, yields ⊕ . Continuing the attaching process produces mosaic patterns of any size. However, such a construction for patterns of larger sizes from patterns of smaller sizes through attaching does not automatically produce mosaic solutions to(1.1) and (1.2). Indeed, the value correspond-ing to symbolp1(respectively,p2) is only known to lie in an interval of length 2σ; thus, it is not assured a priori
whether if this value is exactly equal to the value corresponding to symbolq1 (respectively,q2) . Nevertheless, such a construction of mosaic patterns can be confirmed through a fixed-point theorem and our formulation on the existence of feasible basic patterns described in Section2.
Theorem 3.1. Assume that 0< σ < 111 is fixed. Let{si}i∈Λd be an array of symbols⊕, ⊗, (i.e., si= ⊕, ⊗, ),
obtained from the above attaching process on a collection of basic patterns corresponding to a single partitioned parameter region. Then, there exists a mosaic solution u= {ui}i∈Λdto(1.1)or(1.2). Moreover, in terms of symbols,
u is exactly represented by{si}i∈Λd so that{si}i∈Λd is indeed a mosaic pattern for(1.1)or(1.2).
Proof. We present the cased = 1. Assume that Λ1is a finite lattice. Let{si}i∈Λ1be an array of⊕, ⊗, , obtained
from the attaching process on the collection of basic patterns corresponding to a partitioned parameter region. Let {˜ui}i∈Λ1 be an array of real numbers with
˜ ui ∈ [−1 − σ, −1 + σ], if si= , ˜ ui ∈ [−σ, σ], ifsi= ⊗, ˜ ui ∈ [1 − σ, 1 + σ], ifsi= ⊕. (3.1)
According to our previous formulations, there always exists an intersection for line(2.5)and curve(2.6). Restated, yi= b[2ui− ˜ui−1− ˜ui+1],
yi= f (ui),
always have an intersection (u∗i, yi∗) for eachi ∈ Λ1with u∗ i ∈ [−1 − σ, −1 + σ], if si = , u∗ i ∈ [−σ, σ], ifsi = ⊗, u∗ i ∈ [1 − σ, 1 + σ], ifsi = ⊕. (3.2)
Notably, ifi ∈ Λ1withi + 1 or i − 1 /∈ Λ1, then ˜ui+1or ˜ui−1should be interpreted from boundary condition. Set V = {{vi}i∈Λ1 :−1 − σ ≤ vi≤ −1 + σ, if si = , −σ ≤ vi ≤ σ, if si = ⊗,
1− σ ≤ vi≤ 1 + σ, if si= ⊕}. (3.3)
Define a mappingG : V → V which maps the given {˜ui}i∈Λ1in(3.1)to{u∗i}i∈Λ1in(3.2). G is obviously continuous. It follows from the Brouwer’s fixed-point theorem that there exists a fixed point u= {ui}i∈Λ1for G. This fixed point u is exactly a stationary solution to(1.1). Moreover, u is represented by the array of symbols{si}i∈Λ1and thus{si}i∈Λ1 is exactly a mosaic pattern for(1.1). IfΛ1is an infinite lattice, for example,Λ1= Z1, then the phase space for(1.1)is
X = {u = {ui}i∈Z1, u < ∞}.
Under the circumstances, the existence of fixed point for G can be confirmed by the Schauder fixed-point theorem with a suitable topology (norm) onX.
4. Stability of mosaic patterns
In this section, we study the stability of the mosaic solutions obtained in Section3. Let u= {ui}i∈Λd ∈ Mσd,
⊕, , ⊗. We consider its neighborhood
N(u, θ, δ) = {v = {vi}i∈Λd||vi− ui| ≤ θ, if si= ⊗ and |vi− ui| ≤ δ, if si= ⊕ or }. (4.1)
We will show that the mosaic solution of system(1.1)or (1.2)with nonlinearity(1.3)is stable, by proving the positive invariance of the setN(u, θ, δ) for appropriate θ > 0 and δ > 0, under some conditions. Moreover, the asymptotic stability of u will also be established.
We introduce some notations concerning the states in the neighborhood of each i∈ Λd. For the one-dimensional case,d = 1, set
pi= card{k ∈ {i − 1, i + 1}|sk = ⊕}, ni= card{k ∈ {i − 1, i + 1}|sk = }, qi= card{k ∈ {i − 1, i + 1}|sk= ⊗}.
For the two-dimensional case, let “•” represent + (square-cross) or “×” (diagonal-cross). We denote that p•
i,j= card{(k, +) ∈ Ni,j•|sk,+= ⊕}, n•
i,j= card{(k, +) ∈ Ni,j•|sk,+= }, q•
i,j= card{(k, +) ∈ Ni,j• |sk,+= ⊗},
where Ni,j+ = {(i + 1, j), (i − 1, j), (i, j + 1), (i, j − 1)} and Ni,j× = {(i + 1, j + 1), (i + 1, j − 1), (i − 1, j + 1), (i − 1, j − 1)}. We present the following theorem for the stability of mosaic solutions on finite lattice Λd, withd = 1 in part (I), d = 2 in part (II). The case of Λd= Zd, an infinite lattice, will be remarked after the proof of the theorems.
Theorem 4.1. (I) Let u∈ Mσ1(α, β), which is represented by the patterns {si}. Then, the set N(u, θ, δ) is positively
invariant for(1.1), ifθ > 0, δ > 0 satisfy
2βδ + α[f (ui− δ) − f (ui)]− (pi+ ni)|β|δ > qi|β|θ, (4.2)
2βδ + α[f (ui)− f (ui+ δ)] − (pi+ ni)|β|δ > qi|β|θ, (4.3)
wheneversi= ⊕ or si= , and
2βθ + α[f (ui− θ) − f (ui)]− qi|β|θ > (pi+ ni)|β|δ, (4.4)
2βθ + α[f (ui)− f (ui+ θ)] − qi|β|θ > (pi+ ni)|β|δ, (4.5)
wheneversi= ⊗. (II) Let u ∈ Mσ2(α, β+, β×), which is represented by the patterns{si,j}. Then, the set N(u, θ, δ)
is positively invariant for(1.2), ifθ > 0, δ > 0 satisfy
(4β++ 4β×)δ + α[f (ui,j− δ) − f (ui,j)]− [(p+i,j+ n+i,j)|β+| + (p×i,j+ n×i,j)|β×|]δ > (q+i,j|β+| + q×i,j|β×|)θ, (4.6) (4β++ 4β×)δ + α[f (ui,j)−f (ui,j+ δ)] − [(p+i,j+ n+i,j)|β+| + (p×i,j+ n×i,j)|β×|]δ > (qi,j+|β+| + q×i,j|β×|)θ, (4.7)
wheneversi,j= ⊕ or si,j= , and
(4β++ 4β×)θ + α[f (ui,j− θ) − f (ui,j)]− (q+i,j|β+| + q×i,j|β×|)θ > [(p+i,j+ n+i,j)|β+| + (p×i,j+ n×i,j)|β×|]δ, (4.8)
(4β++ 4β×)θ + α[f (ui,j)− f (ui,j+ θ)] − (q+i,j|β+| + q×i,j|β×|)θ > [(p+i,j+ n+i,j)|β+| + (p×i,j+ n×i,j)|β×|]δ, (4.9)
wheneversi,j= ⊗.
Asymptotic stability for the mosaic solutions can further be established in the following theorem. The situations are rather different between the casesα < 0 and α > 0.
Theorem 4.2. Let u∈ Mσ1 or Mσ2, which is represented by the patterns{si}. (i) For α < 0, if(4.2) and (4.3)
(respectively,(4.6) and (4.7)) hold for i withsi= ⊕ and si = respectively, as well as(4.4) and (4.5)
(respec-tively,(4.8) and (4.9)) hold for i withui∈ [0, σ] and ui∈ [−σ, 0] respectively, then u is asymptotically stable
for (1.1) (respectively,(1.2)). (ii) Forα > 0, if si= ⊗ for all i, (4.4) and (4.5)(respectively, (4.8) and (4.9))
hold for i withui∈ [−σ, 0] and ui∈ [0, σ], respectively, then u is asymptotically stable for(1.1)(respectively,
(1.2)).
Although inequalities(4.6)–(4.9)seem complicated, for practical application, writing a computer program to examine these inequalities is straightforward. We make a few observations and arrange them in the following remarks, before we prove the theorems.
Remark 1. Notably,f (ui− δ) − f (ui) ,f (ui)− f (ui+ δ) are both negative whenever si = ⊕ or , and f (ui− θ) − f (ui),f (ui)− f (ui+ θ) are both positive whenever si = ⊗. The assumptions(4.2) and (4.3)(respectively,
(4.4) and (4.5)) are more likely to hold ifα is negative (respectively, α is positive). Similar observations are valid for(4.6)–(4.9).
Remark 2. Recall that we have takenσ < 111. With the characteristics of the nonlinearity f defined in(1.3), we can derive the following:
(a) Caseα < 0.
(i) Ifsi= ⊕, then(4.2)(respectively,(4.6)) implies(4.3)(respectively,(4.7)).
Ifsi= , then(4.3)(respectively,(4.7)) implies(4.2)(respectively,(4.6)).
Ifsi= ⊗ with ui∈ [0, σ], then(4.4)(respectively,(4.8)) implies(4.5)(respectively,(4.9)).
Ifsi= ⊗ with ui∈ [−σ, 0], then(4.5)(respectively,(4.9)) implies(4.4)(respectively,(4.8)).
(ii) Moreover, ifsi = ⊕, and(4.2)(respectively,(4.6)) holds for someθ and δ, then it also holds with θ and δ
replaced byνθ and νδ, respectively, where 0 < ν < 1. Same conclusions hold for(4.3)(respectively,(4.7)) if si= . If si = ⊗ with ui∈ [0, σ], and(4.4)(respectively,(4.8)) holds for someθ and δ, then it also holds with
θ and δ replaced by νθ and νδ, respectively, where 0 < ν < 1. Same conclusions hold for(4.5)(respectively,
(4.9)) ifsi= ⊗ with ui ∈ [−σ, 0].
(b) Caseα > 0.
(i) Ifsi= ⊕, then(4.3)(respectively,(4.7)) implies(4.2)(respectively,(4.6)).
Ifsi= , then(4.2)(respectively,(4.6)) implies(4.3)(respectively,(4.7)).
Ifsi= ⊗ with ui∈ [0, σ], then(4.5)(respectively,(4.9)) implies(4.4)(respectively,(4.8)).
Ifsi= ⊗ with ui∈ [−σ, 0], then(4.4)(respectively,(4.8)) implies(4.5)(respectively,(4.9)).
(ii) Ifsi = ⊗ with ui∈ [−σ, 0] and(4.4)(respectively,(4.8)) holds for someθ and δ, then it also holds with θ and
δ replaced by νθ and νδ, respectively, where 0 < ν < 1. Same conclusions hold for(4.5)(respectively,(4.9)) ifsi= ⊗ and ui∈ [0, σ].
Proof of Theorem 4.1. We only prove part (I), the one-dimensional case. The two-dimensional case is similar. If v= {vi} ∈ N(u, θ, δ) for some θ, δ > 0, then from the definitions of pi, ni, qiandvi, we have
vi≤ ui+1+ ui−1+ (pi+ ni)δ + qiθ − 2vi, vi≥ ui+1+ ui−1− (pi+ ni)δ − qiθ − 2vi.
Hence, we have a lower bound forβvi:
βvi≥ β(ui+1+ ui−1)− |β|(pi+ ni)δ − |β|qiθ − 2βvi
for anyβ ∈ R. Since u is an equilibrium solution of(1.1), it follows that
βvi≥ 2β(ui− vi)− αf (ui)− |β|(pi+ ni)δ − |β|qiθ. (4.10)
Similarly, we obtain a upper bound forβvias
βvi≤ 2β(ui− vi)− αf (ui)+ |β|(pi+ ni)δ + |β|qiθ. (4.11) Let v= v(t) be a solution to(1.1)lying inN(u, θ, δ),(4.10) and (4.11)imply
˙
vi(t) = βui+ αf (ui)≥ 2β(ui− vi)+ α[f (vi)− f (ui)]− |β|(pi+ ni)δ − |β|qiθ =: Li(vi, θ, δ), (4.12) and
˙
vi(t) ≤ 2β(ui− vi)+ α[f (vi)− f (ui)]+ |β|(pi+ ni)δ + |β|qiθ =: Ui(vi, θ, δ). (4.13) Now, let us prove that with v(0)∈ N(u, θ, δ), the solution v(t) to(1.1)remains in the setN(u, θ, δ) for all t ≥ 0. Notably, the inequalities(4.2) and (4.3)are equivalent toLi(ui− δ, θ, δ) > 0 and Ui(ui+ δ, θ, δ) < 0, respectively. SinceLiandUiare continuous functions of their arguments, there existµ with 0 < µ < 1, and C1> 0, C2> 0, such that Li(vi, θ, δ)≥ C1, if vi ui− δ, θ θ, δ δ ∈ (µ, µ−1), (4.14) Ui(vi, θ, δ)≤ −C2, if vi ui+ δ, θ θ, δ δ ∈ (µ, µ−1). (4.15)
On the other hand, the inequalities(4.4) and (4.5)are equivalent toLi(ui− θ, θ, δ) > 0 and Ui(ui+ θ, θ, δ) < 0, respectively. For this case, we also have
Li(vi, θ, δ)≥ C3, if vi ui− θ, θ θ, δ δ ∈ (µ, µ−1), Ui(vi, θ, δ)≤ −C4, if vi ui+ θ, θ θ, δ δ ∈ (µ, µ−1),
for someC3> 0, C4> 0. We note that µ and C1, C2, C3, C4can be chosen independent of i. For smallε > 0, there existsK > 0 such that |βvi+ f (vi)| ≤ K for all i, for all v ∈ N(u, θ + ε, δ + ε). Hence, for any solution v(t) with v(0)∈ N(u, θ, δ), we have v(t) ∈ N(u, θ + ε, δ + ε), for 0 ≤ t ≤ T := Kε. Herein, we chooseε = min{(µ1 −
1)θ, (µ1 − 1)δ}, and claim that ui− δ ≤ vi(t) ≤ ui+ δ, for all t ∈ [0, T ], whenever si = ⊕ or . Suppose, on the contrary, thatui− δ − ε ≤ vi(t) ≤ ui− δ, for some t ∈ [0, T ] and some i with si = ⊕ or . Form(4.14),
˙
vi(t) ≥ Li(vi, θ, δ) ≥ Li(vi, θ + ε, δ + ε) ≥ C1> 0.
Hence, if v(0)∈ N(u, θ, δ), vi(t) ≥ ui− δ, for all t ∈ [0, T ], for all i. In addition, by(4.15),vi(t) ≤ ui+ δ, for all t ∈ [0, T ]. Similarly, if si = ⊗, it can be shown that |vi(t) − ui| ≤ θ, for all t ∈ [0, T ]. Thus, we have that
v(t) ∈ N(u, θ, δ), for all t ∈ [0, T ]. Note that we only require v(0) ∈ N(u, θ, δ) to derive this result. Therefore, we
conclude thatN(u, θ, δ) is positively invariant.
Proof of Theorem 4.2. We only prove the one-dimensional case. Consider a solution v(t) to(1.1)with v(0)∈ N(u, θ, δ). Recall that if i is such that si = ⊕, then ˙vi(t) ≥ C1, whenever ui− δ ≤ vi(t) ≤ ui− µδ, and ˙vi(t) ≤ −C2, wheneverui+ µδ ≤ vi(t) ≤ ui+ δ. Thus, for v(0) ∈ N(u, θ, δ), |vi(t) − ui| ≤ µδ for all t ≥ (1 − µ)δ/C, whereC = min{C1, C2, C3, C4}. Similarly, we have that |vi(t) − ui| ≤ µθ, for all t ≥ (1 − µ)θ/C. Therefore, we conclude that v(t) ∈ N(u, µθ, µδ), for all t ≥ T , where T = max{(1 − µ)δ/C, (1 − µ)θ/C}. Using the observations inRemark 2(a)(ii) and (b)(ii), there is a sequence of positive timeT1< T2< T3< · · ·, which converge to infinity, such that v(t) ∈ N(u, µnθ, µnδ), for all t ≥ Tn. Notice that the choice ofTnis independent of the solution v(t). Therefore, we conclude that|v(t) − u| → 0 as t → ∞, i.e., u is asymptotically stable.
Remark 3. We can replace(4.2)–(4.9)by stronger conditions which do not depend on the exact values ofui. For example, ifα < 0, we replace(4.2) and (4.3)by
2βδ + α[f (−1 + σ − δ) − f (−1 + σ)] − (pi+ ni)|β|δ > qi|β|θ, (4.16)
2βδ + α[f (1 − σ) − f (1 − σ + δ)] − (pi+ ni)|β|δ > qi|β|θ, (4.17) ifsi= and si= ⊕, respectively, as f (−1 + σ − δ) − f (−1 + σ) > f (ui− δ) − f (ui) andf (1 − σ) − f (1 − σ + δ) > f (ui)− f (ui+ δ) for the respective case. For the case of infinite lattice Λd= Zd, one can derive similar results asTheorems 4.1 and 4.2by replacing(4.2)–(4.9)with stronger ones as(4.16) and (4.17)in the spirit mentioned herein.
A simple way to construct stable mosaic patterns is to consider the case α < 0 and the mosaic solution u represented by{si} with si= ⊕, for all i. In this situation, we only need to verify(4.2) and (4.3)for the case of d = 1 and(4.6) and (4.7)for the case ofd = 2 for the stability of u. If d = 1, and β is fixed, one can always choose negativeα with large magnitude to satisfy(4.2) and (4.3). Regarding the existence of these patterns withsi = ⊕, , we note that as|b| is small enough (b = β/α), all the basic patterns • • • with • = ⊕, , exist. More precisely, if (α, β) satisfying |b| < f (−σ)2+4σ, all mosaic patterns{si}, si = ⊕, , exist. These patterns are asymptotically stable if β is fixed, α < 0 and |α| is large. It is also straightforward to find parameters for the existence of stable pattern {si},
withsi= ⊗ for all i ∈ Λd. Notably, the patterns{si} are regarded as spatially uniform ones if si= ⊗ (or ⊕, or )
for all i∈ Λd. The existence of the above-mentioned stable patterns can be extended to the system on other lattices of higher dimension. We give a concrete example.
Example 1. Consider the cased = 1. Let Λ1= Tk, where Tk= {i ∈ Z1|1 ≤ i ≤ k}.
We impose the Neumann boundary condition onTk, i.e., uk+1= uk, u0= u1.
Fig. 3. Phase portrait for(1.1)withβ = 1, α = −100, on T2. The shadow regions depicted from the estimatesδ = θ = 0.2 indicate subsets of
the basins around the stable equilibrium points.
Since the conditions(4.2)–(4.9)inTheorems 4.1 and 4.2concern themselves with the states at each ith site and its adjacent sites, we could also examine these conditions for the boundary sitesi = 1, i = k. We illustrate the numerics by the following instance withk = 2:
˙
ui = β(ui+1+ ui−1− 2ui)+ αf (ui), i = 1, 2, (4.18)
whereu3= u2,u0= u1. If we takeβ = 1, α = −100, then δ = θ = 0.2 satisfy(4.2) and (4.3). The phase portrait for such a system is illustrated inFig. 3.
5. Spatial entropy
Let us review the notion of spatial entropy for lattice dynamical systems[2,6]. LetA be a finite set of elements (symbols) which are used to represent the patterns at each site on the lattice. In the case herein,A = {⊕, ⊗, }. LetAZd = {s|s : Zd→ A}. Consider the natural projection
πk :AZ
d
→ ATk, (5.1)
given by restricting any s∈ AZdto finite subsetTk(defined in(2.1)ford = 1,(2.2)ford = 2). Let S be a translation
invariant subset of the feasible global patterns (corresponding to stationary solutions) of(1.1) and (1.2)onZd, with certain parameters. Set
whereΓk∞denotes the number of distinct feasible mosaic patterns projected from elements inS onto Tk. The spatial
entropyh(S) of the set S is defined as h(S) := lim
k→∞
1 k1k2. . . kd
lnΓk(S). (5.3)
There are other considerations for spatial entropy; in particular, if boundary condition is taken into account, then definition(5.3)should be modified. We arrange such a consideration in Section6. According to our formulation, the partitioning of parameters in Section2allows us to discover the major portion of feasible basic patterns corresponding to each parameter region. There are some other possible basic patterns that are not included in these collections. They arise from other possible intersections for the graphs of(2.5) and (2.6). Their existence as feasible basic patterns cannot be justified from our fixed-point arguments. As a subsequence of this formulation, we further introduce the following notations. Under the same parameters forS, let S be the translation invariant subset of AZd, which is formed from attaching the affirmatively feasible basic patterns established in Section2, and letS be the one formed from attaching both the affirmatively feasible basic patterns as well as possible basic patterns. Set
h(S) := lim k→∞ 1 k1k2, . . . , kd lnΓk(S), h(S) := lim k→∞ 1 k1k2, . . . , kd lnΓk(S). (5.4)
Obviously,h(S) ≤ h(S) ≤ h(S). We recall the following definition in[2].
Definition 5.1. The system(1.1)or(1.2)is said to exhibit spatial chaos at parameters (α, β) or (α, β+, β×), if the spatial entropy is positive. The system(1.1)or (1.2)is said to exhibit pattern formation at parameters (α, β) or (α, β+, β×), if the spatial entropy is zero.
The notion of spatial entropy resembles the one of topological entropy for Markov shift[17]. In the case of one-dimensional latticed = 1, a transition matrix can be formulated to depict the attaching process of basic patterns. Accordingly, total number of mosaic patterns obtained from the attaching can be calculated and the spatial entropy can be computed exactly. Let us introduce this formulation. We employ the following identification between the indices{1, 2, 3, . . . , 9} and the nine 1 × 2 patterns {⊕⊕, ⊕⊗, ⊕, ⊗⊕, ⊗⊗, ⊗, ⊕, ⊗, } :
1←→ ⊕⊕, 2←→ ⊕⊗, 3←→ ⊕,
4←→ ⊗⊕, 5←→ ⊗⊗, 6←→ ⊗,
7←→ ⊕, 8←→ ⊗, 9←→ .
(5.5)
Consider the 9× 9 matrix M:
M = M(r) := r1 r2 r3 0 0 0 0 0 0 0 0 0 r4 r5 r6 0 0 0 0 0 0 0 0 0 r7 r8 r9 r10r11 r12 0 0 0 0 0 0 0 0 0 r13 r14 r15 0 0 0 0 0 0 0 0 0 r16 r17r18 r19r20 r21 0 0 0 0 0 0 0 0 0 r22 r23 r24 0 0 0 0 0 0 0 0 0 r25 r26r27 , (5.6)
wherer = {rj}27j=1, rj= 0 or 1, j ∈ {1, 2, . . . , 27}. The formation of feasible mosaic patterns depicted by the transition matrix can be described as follows: the (i, j)-entry of M is one if and only if the jth 1 × 2 pattern in(5.5)
can be attached, with one site overlapped, to the right of the ith 1× 2 pattern in(5.5)to form a 1× 3 feasible pattern. For example, ifb ∈ I5= [f (−σ),f (1+σ)1+4σ ], the set of affirmatively feasible basic patterns are
{⊕ ⊕ ⊕, ⊕ ⊕ ⊗, ⊕ ⊗ , ⊗ ⊕ ⊕, ⊗ ⊗ ⊗, ⊗ , ⊗ ⊕, ⊗, },
and the corresponding transition matrix is(5.6)withr1= r2= r6= r10 = r14 = r18 = r22= r26= r27= 1 and rj= 0 for all other j.
Moreover, the total number of mosaic patterns on the latticeTkobtained from such a formulation is
1≤i,j≤9 Mijk−2.
The spatial entropy can be computed from the largest eigenvalueλ1of M, namely h(S) = ln λ1.
RecallFig. 1, where we illustrate the partitioning of parameter space. Therein, we have determined a collection of basic patterns corresponding to each parameter region. These basic patterns are confirmed to be feasible later in Section3. In fact, in our geometric formulation, there may exist more possible basic patterns if the graph of f between the linesL1andL2has intersection with the shadow region R (even a point). InFig. 2, we display the existence of feasible basic patterns(left half) which can be confirmed by the treatment in Section3and the existence of all other possible basic patterns (right half) in each respective parameter region. In order to achieve this, further partitioning of the parameters needs to be performed. InTable 3, we have further partitioned regionsIiintoIi+so that all possible basic patterns are identified for parameters in subregionIi+, in addition to those feasible basic patterns already confirmed in regionIi. We summarize our computations inTable 4and the following theorem.
Theorem 5.2. System (1.1)exhibits spatial chaos in parameter regions Ii, −5 ≤ i ≤ 5, and exhibits pattern
formation in parameter regionsI±7andI±62 .
Theorem 5.2andTable 4are completely obtained from computing the eigenvalues of the transition matrix corre-sponding to each parameter region.λ1is the largest eigenvalue of the transition matrix corresponding to attaching affirmatively feasible basic patterns, while λ1is the largest eigenvalue of the transition matrix corresponding to attaching both affirmatively feasible basic patterns and possible basic patterns.
In the two dimension cased = 2, one no longer has a transition matrix to describe the formation of patterns, except some special situations[18]. Therefore, in most cases, we can only estimate the spatial entropy. By employing the methodology in[2,6], i.e., constructing adjoinable building blocks from feasible basic patterns, we can compute the lower bound of the spatial entropy. For example, if we can find three 2× 2 patterns so that any one of them can be joined (without overlapping) from the left, the right, upward, and downward directions to any one of themselves, then on a 2k × 2k square lattice, there are at least 3k2distinct mosaic patterns. It follows that a lower bound for the entropy is
lim k→∞
ln 3k2 k2 = ln 3.
If there exist only very few feasible basic patterns in a parameter region, then it can be easily seen that the spatial entropy is zero. We summarize our computations inTable 5.
Table 3
Additional possible basic patterns in further partitioned parameter regions
Parameter space Possibe basic patterns
I7= f (−1 + σ) 4σ , ∞ B⊕ {2}, B⊗{0}, B{−2} I6 I2 6= f (1 + σ),f (−1 + σ)4σ B⊕{2}, B⊗{0}, B{−2} I1 6= f (1 + σ) 1+ 4σ , f (1 + σ) B⊕ {2,1}, B⊗{0}, B{−1,−2} I5 I3 5= f (−σ) 1− 4σ, f (1 + σ) 1+ 4σ B⊕ {2,1}, B⊗{0}, B{−1,−2} I2 5= f (1 + σ) 2 , f (−σ) 1− 4σ B⊕ {2,1}, B⊗{1,0,−1}, B{−1,−2} I1 5= f (−σ),f (1 + σ)2 B⊕{2,1,0}, B⊗{1,0,−1}, B{0,−1,−2} I4= f (1 + σ) 2+ 4σ , f (−σ) B⊕ {2,1,0}, B⊗{1,0,−1}, B{0,−1,−2} I3 I2 3= f (1 + σ) 3 , f (1 + σ) 2+ 4σ B⊕ {2,1,0}, B⊗{1,0,−1}, B{0,−1,−2} I1 3= f (1 + σ) 3+ 4σ , f (1 + σ) 3 B⊕ {2,1,0,−1}, B⊗{1,0,−1}, B{1,0,−1,−2} I2 I3 2= f (−σ) 2− 4σ, f (1 + σ) 3+ 4σ B⊕ {2,1,0,−1}, B⊗{1,0,−1}, B{1,0,−1,−2} I2 2= f (1 + σ) 4 , f (−σ) 2− 4σ B⊕{2,1,0,−1}, B⊗{2,1,0,−1,−2}, B{1,0,−1,−2} I1 2= f (1 + σ) 4+ 4σ , f (1 + σ) 4 B⊕ {2,1,0,−1,−2}, B{2,1,0,−1,−2}⊗ , B{2,1,0,−1,−2} I1= f (−σ) 2 , f (1 + σ) 4+ 4σ B⊕{2,1,0,−1,−2}, B{2,1,0,−1,−2}⊗ , B{2,1,0,−1,−2} I0= −f (−σ) 2+ 4σ, f (−σ) 2 B⊕ {2,1,0,−1,−2}, B{2,1,0,−1,−2}⊗ , B{2,1,0,−1,−2} I−1= −f (−1 + σ) 4 , − f (−σ) 2+ 4σ B⊕ {2,1,0,−1,−2}, B{2,1,0,−1,−2}⊗ , B{2,1,0,−1,−2} I−2 I1 −2= −f (−1 + σ) 4− 4σ , − f (−1 + σ) 4 B⊕ {2,1,0,−1,−2}, B{2,1,0,−1,−2}⊗ , B{2,1,0,−1,−2} I2 −2= −f (−σ)2 , −f (−1 + σ)4− 4σ B⊕{2,1,0,−1}, B⊗{2,1,0,−1,−2}, B{1,0,−1,−2} I3 −2= −f (−1 + σ) 3 , − f (−σ) 2 B⊕ {2,1,0,−1}, B⊗{1,0,−1}, B{1,0,−1,−2} I−3 I1 −3= −f (−1 + σ)3− 4σσ , −f (−1 + σ)3 B⊕{2,1,0,−1}, B⊗{1,0,−1}, B{1,0,−1,−2} I2 −3= −f (−σ) 1+ 4σ, − f (−1 + σ) 3− 4σ B⊕ {2,1,0}, B⊗{1,0,−1}, B{0,−1,−2} I−4= −f (−1 + σ)2 , −f (−σ)1+ 4σ B⊕{2,1,0}, B⊗{1,0,−1}, B{0,−1,−2}
Table 3 (Continued )
Parameter space Possibe basic patterns
I−5 I1 −5= −f (−σ), −f (−1 + σ)2 B⊕{2,1,0}, B⊗{1,0,−1}, B{0,−1,−2} I2 −5= −f (−1 + σ) 2− 4σ , −f (−σ) B⊕ {2,1,0}, B⊗{0}, B{0,−1,−2} I3 −5= −f (−1 + σ), −f (−1 + σ) 2− 4σ B⊕ {2,1}, B⊗{0}, B{−1,−2} I−6 I1 −6= −f (−1 + σ) 1− 4σ , −f (−1 + σ) B⊕ {2,1}, B⊗{0}, B{−1,−2} I2 −6= −f (−σ)4σ −f (−1 + σ) 1− 4σ B⊕ {2}, B⊗{0}, B{−2} I−7= −∞, −f (−σ) 4σ B⊕ {2}, B⊗{0}, B{−2} Table 4
λ1: the largest eigenvalue of the transition matrix corresponding to attaching affirmatively feasible basic patterns, andλ1: the largest eigenvalue
of the transition matrix corresponding to attaching both affirmatively feasible basic patterns and possible basic patterns
Parameter space λ1 λ1 h I7= f (−1 + σ) 4σ , ∞ 1 1 0 I6 I2 6 = f (1 + σ),f (−1 + σ) 4σ 1 1 0 I1 6 = f (1 + σ) 1+ 4σ , f (1 + σ) 1 1.4656 0≤ h ≤ 0.3823 I5 I3 5 = f (−σ) 1− 4σ, f (1 + σ) 1+ 4σ 1.4656 1.4656 0.3823 I2 5 = f (1 + σ) 2 , f (−σ) 1− 4σ 1.4656 1.8972 0.3823 ≤ h ≤ 0.6404 I1 5 = f (−σ),f (1 + σ) 2 1.4656 2.3165 0.3823 ≤ h ≤ 0.8401 I4= f (1 + σ) 2+ 4σ , f (−σ) 1.8972 2.3165 0.6404 ≤ h ≤ 0.8401 I3 I2 3 = f (1 + σ) 3 , f (1 + σ) 2+ 4σ 2.3165 2.3165 0.8401 I1 3 = f (1 + σ) 3+ 4σ , f (1 + σ) 3 2.3165 2.5921 0.8401 ≤ h ≤ 0.9525 I2 I3 2 = f (−σ) 2− 4σ, f (1 + σ) 3+ 4σ 2.5921 2.5921 0.9525 I2 2 = f (1 + σ) 4 , f (−σ) 2− 4σ 2.5921 2.8312 0.9525 ≤ h ≤ 1.0407 I1 2 = f (1 + σ) 4+ 4σ , f (1 + σ) 4 2.5921 3 0.9525 ≤ h ≤ 1.0986
Table 4 (Continued ) Parameter space λ1 λ1 h I1= f (−σ) 2 , f (1 + σ) 4+ 4σ 2.7693 3 1.0186 ≤ h ≤ 1.0986 I0= −f (−σ) 2+ 4σ, f (−σ) 2 3 3 1.0986 I−1= −f (−1 + σ)4 , −f (−σ)2+ 4σ 2.7693 3 1.0186 ≤ h ≤ 1.0986 I−2 I1 −2= −f (−1 + σ) 4− 4σ , − f (−1 + σ) 4 2.5921 3 0.9525 ≤ h ≤ 1.0986 I2 −2= −f (−σ)2 , −f (−1 + σ)4− 4σ 2.5921 2.8312 0.9525 ≤ h ≤ 1.0407 I3 −2= −f (−1 + σ) 3 , − f (−σ) 2 2.5921 2.5921 0.9525 I−3 I1 −3= −f (−1 + σ) 3− 4σσ , − f (−1 + σ) 3 2.3165 2.5921 0.8401 ≤ h ≤ 0.9525 I2 −3= −f (−σ) 1+ 4σ, − f (−1 + σ) 3− 4σ 2.3165 2.3165 0.8401 I−4= −f (−1 + σ) 2 , − f (−σ) 1+ 4σ 1.9052 2.3165 0.6446 ≤ h ≤ 0.8401 I−5 I1 −5= −f (−σ), −f (−1 + σ) 2 1.4656 2.3165 0.3823 ≤ h ≤ 0.8401 I2 −5= −f (−1 + σ) 2− 4σ , −f (−σ) 1.4656 1.9052 0.3823 ≤ h ≤ 0.6446 I3 −5= −f (−1 + σ), −f (−1 + σ)2− 4σ 1.4656 1.4656 0.3823 I−6 I1 −6= −f (−1 + σ) 1− 4σ , −f (−1 + σ) 1 1.4656 0≤ h ≤ 0.3823 I2 −6= −f (−σ)4σ −f (−1 + σ)1− 4σ 1 1 0 I−7= −∞, −f (−σ)4σ 1 1 0
6. Effect of boundary conditions on pattern formation and spatial entropy
In Section2, three typical types of boundary conditions: Neumann (N), periodic (P), and Dirichlet (D), have been introduced. In this section, we plan to discuss the effect of these boundary conditions on pattern formation and spatial entropy. We introduce the following notations to distinguish different considerations of spatial entropy. For the definition of spatial entropy used in Section5, i.e., from counting the number of patterns projected from global patterns (patterns onZ1, Z2), we introduce the notationΓk∞to represent the number of such patterns onTk.
Since our formulation includes the situations of patterns obtained from affirmatively feasible basic patterns as well as from possible basic patterns additionally, we further denote
Table 5
Estimations for the lower bounds of spatial entropy for each parameter region, for the two-dimensional system(1.2)withb2= 0
Parameter region h Parameter region h
I13= −f (1 − σ) 8σ , ∞ 0 I−1= f (1 − σ) 8 , f (σ) 4+ 8σ ln 69 4 I12= f (1 + σ) 1+ 8σ , −f (1 − σ) 8σ 0 I−2[f (1 − σ) 7 , f (1 − σ) 8 ] ln 51 4 I11= f (σ) −1 ,f (1 + σ)1+ 8σ 0 I−3= f (σ) 3+ 8σ, f (1 − σ) 7 ln 51 4 I10= f (1 + σ) 2+ 8σ , f (σ) −1 0 I−4= f (1 − σ) 6 , f (σ) 3+ 8σ ln 49 4 I9= f (1 + σ) 3+ 8σ , f (1 + σ) 2+ 8σ ln 3 16 I−5= f (1 − σ) 5 , f (1 − σ) 6 ln 3 2 I8= f (σ) −2 , f (1 + σ) 3+ 8σ ln 3 16 I−6= f (σ) 2+ 8σ, f (1 − σ) 5 ln 3 2 I7= f (1 + σ) 4+ 8σ , f (σ) −2 ln 5 4 I−7= f (1 − σ) 4 , f (σ) 2+ 8σ ln 3 16 I6= f (1 + σ) 5+ 8σ , f (1 + σ) 4+ 8σ ln 3 2 I−8= f (1 − σ) 3 , f (1 − σ) 4 ln 3 16 I5= f (1 + σ) 6+ 8σ , f (1 + σ) 5+ 8σ ln 3 2 I−9= f (σ) 1+ 8σ, f (1 − σ) 3 ln 3 16 I4= f (σ) −3 , f (1 + σ) 6+ 8σ ln 7 2 I−10= f (1 − σ) 2 , f (σ) 1+ 8σ 0 I3= f (1 + σ) 7+ 8σ , f (σ) −3 ln 51 4 I−11= f (1 − σ),f (1 − σ) 2 0 I2= f (1 + σ) 8+ 8σ , f (1 + σ) 7+ 8σ ln 51 4 I−12= f (σ) 8σ , f (1 − σ) 0 I1= f (σ) −4 , f (1 + σ) 8+ 8σ ln 69 4 I−13= −∞,f (σ)8σ 0 I0= f (σ) 4+ 8σ, f (σ) −4 ln 3 Γ∞k = Γk(S) := card(πk(S)) Γ∞ k = Γk(S) := card(πk(S)),
whereS, S are as defined in Section5. The upper and lower bounds for the spatial entropy,h := h(S), and h := h(S) have been defined in(5.4).
On the other hand, with considerations of boundary conditions, under the same parameters, we setSBk (respec-tively,SBk) as the class of mosaic patterns onTkobtained from attaching all affirmatively feasible (respectively, all
affirmatively feasible and possible) basic patterns for (sd-DE)B, whereB = N, P, D. Moreover, let ΓkB:= Γ (SBk) (respectively,ΓBk := Γ (SBk),ΓBk := Γ (SBk)) be the number of patterns inSBk (respectively,SBk,SBk). Accordingly, we have hB= h(SBk) := lim k→∞ 1 k1· · · kd lnΓBk, hB= h(SBk) := lim k→∞ 1 k1· · · kd lnΓBk, hB= h(SB k) := lim k→∞ 1 k1· · · kd lnΓBk. (6.1)
We propose a criterion forh = hB, whereB = N, P or D, in the following proposition. For k = (k1, . . . , kd), and s ∈ R, by k − s, we mean (k1− s, . . . , kd− s).
Proposition 6.1. If there are fixed positive integerss, r such that (i) ΓBk ≥ Γ∞k−s, for all k> s, and (ii) ΓBk ≤ pc· Γ∞
k−r for some p > 0 and c = c(k) with lim
k→∞c/(k1· · · kd)= 0, then h = hB= h = hB= h = hB, where
B = N, P or D.
Proof. Let us prove the two-dimensional case. From condition (i), we have
hB= lim k→∞ 1 k1k2 lnΓBk ≥ lim k→∞ 1 k1k2 lnΓ∞k−s = lim k→∞ (k1− 2s)(k2− 2s) k1k2 lnΓ∞k−s (k1− 2s)(k2− 2s) = h(S) = h.
Condition (ii) yields that hB= lim k→∞ 1 k1k2 lnΓBk ≤ lim k→∞ 1 k1k2 ln(pc· Γ∞k−r)= lim k→∞ (k1− 2r)(k2− 2r) k1k2 c ln p + ln Γ∞ k−r (k1−2r)(k2−2r)= h(S) = h, Therefore, hB≥ h ≥ h ≥ hB.
On the other hand,hB≤ hB, from our definition. The assertion of the proposition thus follows. By applyingProposition 6.1, the following result can be derived.
Theorem 6.2. h = hN= hP = hDfor the mosaic patterns of (sd-DE) on one-dimensional latticed = 1.
The problem of whether ifh = hN= hP= hDis much more complicated for the two-dimensional cased = 2. Condition (ii) ofProposition 6.1holds for the situation herein. In several cases, we can carry out the examination for condition (i). We have not found a situation forh = hB, as there are two examples forh = hDand forh = hN in cellular neural networks[8]. One observation is that the feasible basic patterns, corresponding to each parameter region, exist in groups in which⊕ and seem to play equal roles. The observation certainly depends on the configuration for the graph of nonlinearity f.
7. Numerical Illustrations
In this section, we shall demonstrate several two-dimensional mosaic patterns for the spatially discrete diffusion equations. Herein, we employ our basic pattern formation to produce these patterns for(1.2)with cubic nonlinearity
(1.3). In the illustrations, we impose the Dirichlet boundary condition by setting ˜uij = 0, for (i, j) ∈ b (see Section
2). We first explore basic patterns needed to compose the desired patterns and locate the parameters for these basic patterns. To justify our construction, we compute the numerical solutions to system(1.2). It can be seen from the computations that each component of the solution lies within theσ-ranges centered at −1, 0, 1. We color the patterns as inFigs. 4 and 5to enhance the effect of demonstration.
Example 2. Checkerboard with horizontal interface.
Fig. 6is a 7× 7 checkerboard with horizontal interface. The 3 × 3 basic patterns needed to generate this checker-board, through attaching process, are collected inFig. 7. If we chooseb1, b2> 0, the parameters which yield these
Fig. 4. Colors corresponding to solution values.
Fig. 5. Colors corresponding to solution values. basic patterns satisfy the following conditions:
b1> 0, b2> 0, (7+ 8σ)b1+ (2 + 8σ)b2≤ f (1 + σ), (8+ 8σ)b1+ 8σb2≤ f (1 + σ).
Let us choose the parametersσ = 0.01, b1= 0.002, b2= 0.002, which satisfy these conditions, to illustrate this pattern. The computed numerical solution (using Newton’s method) is listed inFig. 8. The associated pattern (in colors) for the numerical solution{ui,j}1≤i,j≤7obviously matches the one inFig. 6.
Fig. 6. Checkerboard with horizontal interface.
Fig. 8. The numerical solution{ui,j}1≤i,j≤7to(1.2)associated with the pattern of checkerboard with horizontal interface.
Fig. 9. Vertical and horizontal stripes with vertical interface.
Example 3. Vertical and horizontal stripes with vertical interface.
We need the 3× 3 basic patterns inFig. 10to generate the vertical and horizontal stripes with vertical interface inFig. 9(a 8× 8 herein), through the attaching process. For these basic patterns to exist, the following parameter conditions are needed:
b1> 0, b2> 0, (5+ 8σ)b1+ (8 + 8σ)b2≤ f (1 + σ).
As an illustration, we choose the parametersσ = 0.01, b1= 0.002, b2= 0.001. The numerical solution is listed in
Fig. 11.
Fig. 11. The numerical solution{ui,j}1≤i,j≤8to(1.2)associated with the pattern of vertical and horizontal stripes with vertical interface.
Example 4. Checkerboard with diagonal interface.
The pattern of checkerboard with diagonal interface inFig. 12could be obtained by attaching the 3× 3 basic patterns inFig. 13with the chosen parametersσ = 0.01, b1= 0.002, b2= 0.002 which satisfy the conditions:
b1> 0, b2> 0, 8σb1− 8σb2≥ f (1 − σ), (6+ 8σ)b1+ (2 + 8σ)b2≤ f (1 + σ), (8+ 8σ)b1+ (1 + 8σ)b2≤ f (1 + σ).
Example 5. Quad junction.
The pattern of quad junction inFig. 14could be obtained by attaching the 3× 3 basic patterns inFig. 15with the chosen parametersσ = 0.01, b1= 0.002, b2= 0.001 which satisfy the conditions:
b1> 0, b2> 0, (4+ 8σ)b1+ (8 + 8σ)b2≤ f (1 + σ).
Fig. 13. Basic patterns for the checkerboard with diagonal interface.
Fig. 14. Quad junction.
