Finding Multiple Solutions to Elliptic PDE with Nonlinear Boundary Conditions An. Le∗ , Zhi-Qiang Wang∗ and Jianxin Zhou† Abstract In this paper, in order to solve an elliptic partial differential equation with a nonlinear boundary condition for multiple solutions, the authors combine a minimax approach with a boundary integral-boundary element method, and identify a subspace and its special expression so that all numerical computation and analysis can be carried out more efficiently based on information of functions only on the boundary. Some mathematical justification of the new approach is established. An efficient and reliable local minimax-boundary element method is developed to numerically search for solutions. Details on implementation of the algorithm are also addressed. The existence and multiplicity of solutions to the problem are established under certain regular assumptions. Some conditions related to convergence of the algorithm and instability of solutions found by the algorithm are verified. To illustrated the method, numerical multiple solutions to some examples on domains with different geometry are displayed with their profile and contour plots. Keywords: Multiple solutions, elliptic PDE, nonlinear boundary condition, local minimaxboundary element method AMS Mathematical Subject Classification: 35J65,35A40,35A15,58E05,58E30 1 Introduction Many studies in convection-diffusion systems, corrosion/oxidation modeling, metal-insulator or metal-oxide semiconductor systems [2,11,14,16,19,24,29] lead to solve a linear elliptic ∗ Department of Mathematics and Statistics, Utah State University, Logan, UT 84322. The second author’s research was supported in part by NSF DMS-0820327 † Department of Mathematics, Texas A&M University, College Station, TX 77843. jzhou@math.tamu.edu. This autho’s research was supported in part by NSF DMS-0713872/0820327/1115384. 1 A LOCAL MINIMAX-BEM METHOD 2 partial differential equation with a nonlinear boundary condition (BC) of the form { (−∆ + aI)u(x) = 0 in Ω, ∂u(x) = g(x, u(x)) on ∂Ω, ∂ν 2 (1.1) 2 ∂ ∂ where Ω ⊂ Rn is bounded and open with a smooth boundary Γ = ∂Ω, ∆ = ∂x 2 + · · · + ∂x2 n 1 is the Laplace operator, I is the identity, a > 0, g satisfies certain regularity and growth conditions and ν is the unit outward normal vector on Γ. Such a model also appears in the study of stationary solutions to Laplace equation with dynamical nonlinear boundary conditions [1,17,28 and references therein]. A huge literature exists on analysis and numerical computation of the case where (1.1) has a unique solution, e.g., finite element method (FEM), boundary integral equation method (BIE) and boundary element method (BEM) [6,18,26,27, etc.]. In this work we focus on the case where (1.1) has multiple unstable solutions. Multiple unstable solutions, lowly or highly, singly or multiply excited, to many nonlinear systems have been physically observed and mathematically proved to exist with a variety of solution configurations, instabilities/maneuverabilities. They used to be considered too hard to catch and therefore to apply by traditional technologies. Now scientists are able to induce, reach or control them with new advanced (synchrotronic, laser, etc.) technologies and search for NEW applications. So far, people’s understanding of such solutions is still quite limited and analytic solutions are too difficult to obtain. On the other hand, due to their strong nonlinearity, multiplicity, unstable nature, such solutions are extremely difficult to solve and very elusive to traditional numerical methods. Thus development of efficient and reliable numerical methods to solve such problems becomes very interesting to both research and applications. Most results in the literature of studying multiple solution problems focus on the case where nonlinearities appear in a differential equation. While our model equation (1.1) has nonlinearities appeared in boundary conditions. So far we have not seen any related numerical results in the literature, to the best of our knowledge. Note that in application, Ω can also be unbounded exterior domain with a bounded boundary Γ. We will keep this fact in mind while dealing with interior domains in the current work. This paper is organized as follows: in Section 2, we show how to combine a minimax approach with a boundary integralboundary element method to solve this problem for multiple solutions. Along the line, we identify a subspace in which all numerical computation and analysis can be carried out more efficiently based on the information of functions only on the boundary. Mathematical jusificationnof such approach is established. Then an efficient and reliable local minimaxBEM method is developed to find multiple solutions. More details on implementation of the algorithm are also presented in this section. With this framework, in Section 3, we establish the existence and multiplicity of solutions to (1.1) under certain regular assumptions. Some A LOCAL MINIMAX-BEM METHOD 3 conditions related to the convergence of the algorithm and instability of solutions found by the algorithm are also verified there. To illustrate the method, in Section 4, we solve some examples. Multiple numerical solutions are displayed with their profile and contour plots. 2 A Local Minimax-BEM Method It is known that solutions to (1.1) coincide with critical points of the C 2 functional ∫ ∫ 1 2 2 J(u) = (|∇u(x)| + au (x))dx − G(x, u(x))dσx , u ∈ H 1 (Ω), Ω 2 Γ where = g(x, u(x)). For any u, v ∈ H 1 (Ω), we have dJ(u + tv) (J ′ (u), v)W −1,2 ×H 1 = ∀u, v ∈ H 1 (Ω) = W 1,2 (Ω) dt t=0 ∫ ∫ = (∇u(x)·∇v(x) + au(x)v(x))dx − g(x, u(x))v(x)dσx , (2.1) ∂ G(x, u(x)) ∂u Ω (2.2) Γ where J ′ (u) ∈ W −1,2 (Ω) is the Frechet derivative of J at u. Denote an inner product ∫ ⟨u, v⟩a = (∇u(x)·∇v(x) + au(x)v(x))dx, ∀u, v ∈ H 1 (Ω), (2.3) Ω and ∥u∥2a = ⟨u, u⟩a . Thus u∗ ∈ H 1 (Ω) is a solution to (1.1) iff u∗ satisfies (J ′ (u∗ ), v) = 0 or ∫ ∗ ⟨u , v⟩a = g(x, u∗ (x))v(x)dσx ∀v ∈ H, (2.4) Γ where the left hand side is a volume integral and the right hand side is a boundary integral. Such a mixture of two types of integrals cause inconvenience and inefficiency in numerical computation. We will try to resolve this problem. Also a key issue in finding multiple critical points is whether or not such solutions can be found in certain order in their instabilities. 2.1 A Local Minimax Characterization The local minimax method (LMM) developed in [21,22,32,33,34,35] is a 2-level optimization method for finding critical points of a functional in the order of their energy levels. We briefly describe its mathematical background. Let H be a Hilbert space with norm ∥ · ∥ and J ∈ C 1 (H, R). For a closed subspace L ⊂ H, denote H = L ⊕ L⊥ and SL⊥ = {v ∈ L⊥ : ∥v∥ = 1}. For each v ∈ SL⊥ , denote [L, v] = span{L, v}. Definition 1. The peak mapping is a set-valued mapping P : SL⊥ → 2H s.t. P (v) = the set of all local maxima of J on [L, v], ∀v ∈ SL⊥ . A peak selection is a single-valued mapping p : SL⊥ → H s.t. p(v) ∈ P (v), ∀v ∈ SL⊥ . If p is locally defined, then p is called a local peak selection. A LOCAL MINIMAX-BEM METHOD 4 Recall J is said to satisfy the Palais-Smale (PS) condition in H, if any sequence {un } ⊂ H s.t. {J(un )} is bounded and J ′ (un ) → 0 has a convergent subsequence. The following theorem provides a mathematical justification for LMM and also gives an estimate for the instability of a solution approximated by LMM. Theorem 1. If p is a local peak selection of J near v0 ∈ SL⊥ s.t. (a) p is Lipschitz continuous at v0 , (b) p(v0 ) ̸∈ L and (c) v0 = arg local minv∈SL⊥ J(p(v)), then u0 = p(v0 ) is a saddle point of J. If in addition, p is differentiable at v0 and denote H 0 = ker(J ′′ (u0 )), then dim(L) + 1 = MI(u0 ) + dim(H 0 ∩ [L, v0 ]). For a given closed nonempty subspace L, let M = {p(v) : v ∈ SL⊥ }. Theorem 1 states that local-min J(u) yields a saddle point u∗ = p(v ∗ ) of J, which is unstable in H but stable u∈M on M and can be numerically approximated by, e.g., a steepest descent method. Then it leads to the following local minimax algorithm: a two-level local optimization algorithm. A solution characterized by Theorem 1 has an instability index equal to dim(L) + 1. When L = {0}, M is the wellknown Nehari manifold in the literature. 2.2 A Local Minimax Algorithm Let w1 , ..., wn−1 be n-1 previously found critical points, L = [w1 , ..., wn−1 ]. Given ε > 0, λ > 0 and v 0 ∈ SL⊥ be an ascent-descent direction at wn−1 . Step 1: Let t00 = 1, vL0 = 0 and set k = 0; Step 2: Using the initial guess w = tk0 v k + vLk , solve for wk ≡ p(v k ) = arg max J(u), u∈[L,v k ] and denote tk0 v k + vLk = w ≡ p(v ); k k Step 3: Compute the steepest descent vector dk := −J ′ (wk ); Step 4: If ∥dk ∥ ≤ ε then output wn = wk , stop; else goto Step 5; v k + sdk ∈ SL⊥ and find ∥v k + sdk ∥ { } λ tk0 λ k 2 k m k k λ k s := max : 2 > ∥d ∥, J(p(v ( m ))) − J(w ) ≤ − m+1 ∥d ∥ . m∈N 2m 2 2 Step 5: Set v k (s) := Initial guess u = tk0 v k ( 2λm ) + vLk is used to find p(v k ( 2λm )). (track a peak selection) k+1 + vLk+1 , k = k + 1, then goto Step 3. Step 6: Set v k+1 = v k (sk ), wk+1 = p(v k+1 ) = tk+1 0 v A LOCAL MINIMAX-BEM METHOD 5 Remark 1. LMM starts from n = 0 with L = {0} to find a solution w1 and then with n = 1 and L = span{w1 } to find another solution w2 . Following this way, LMM continues with L gradually expanded by previously found solutions w1 , ..., wn−1 , i.e., L = span{w1 , ..., wn−1 }. Theorem 2. [37] Let J be C 1 and satisfy the PS condition. If (a) p is locally Lipschitz continuous, (b) d(L, p(v k )) > α > 0 and (c) inf v∈SL⊥ J(p(v)) > −∞, then v k → v ∗ ∈ SL⊥ with ∇J(p(v ∗ )) = 0. 2.3 A BIE/BEM Approach For now on, we set ∥ · ∥ = ∥ · ∥a if not otherwise specified. In LMM iteration at u = p(v), a search direction d = ∇J(u) ∈ H 1 in Step 3 of LMM is the canonical identity of J ′ (u) ∈ W −1,2 (Ω) in W 1,2 (Ω) and defined by ∫ ⟨d, v⟩a = (∇d(x) · ∇v(x) + ad(x)v(x))dx Ω ∫ ∫ ∂d(x) (−∆ + aI)d(x)v(x)dx + = v(x)dσx Γ ∂ν Ω ≡ (J ′ (u), v)W −1,2 ×H 1 , ∀v ∈ H 1 (compare to (2.2). Thus d = ∇J(u) is solved from an inhomogeneous linear elliptic equation { (−∆ + aI)d(x) = (−∆ + aI)u(x), x ∈ Ω, ∂d(x) = ∂u(x) − g(x, u(x)), x ∈ Γ, ∂ν ∂ν (2.5) where the right hand side of the equation is the residue of the equation (1.1) at u and the right hand side of BC is the residue of BC in (1.1) at u. This is where FEM, BEM or other methods can be applied for numerical approximation. Here we discuss how BIE/BEM can be used more efficiently in solving such a problem. Let E be the fundamental solution defined by (−∆x + aI)E(|ξ − x|) = δ(ξ − x) where δ is the delta function. It is known { 1 − 2π ln |ξ − x| (n = 2), (4π|ξ − x|)−1 (n = 3), a = 0, √ E(|ξ − x|) = √ − a|ξ−x| 1 (n = 3), a > 0, K ( a|ξ − x|) (n = 2), e 4π|ξ−x| 2π 0 where K0 is the modified Bessel function of order 0. Then we have a simple-layer potential representation (SLPR) [15] for u satisfying (−∆ + aI)u = f , i.e., ∫ ∫ u(x) = (2.6) E(|ξ − x|)η(ξ) dσξ + E(|ξ − x|)f (ξ) dξ Γ ≡ (Lb η)(x) + (Lv f )(x), Ω ∀x ∈ Rn A LOCAL MINIMAX-BEM METHOD 6 in layer density η. For such u, by a known jump-discontinuity, we have (weakly singular) ∫ ∫ ∂u 1 ∂E(|x − ξ|) ∂E(|x − ξ|) (x) = η(x) + η(ξ) dσξ + f (ξ) dξ ∂ν 2 ∂νx ∂νx Ω Γ ≡ (∂ν Lb η)(x) + (∂ν Lv f )(x), ∀x ∈ Γ. (2.7) In the above, Lb and ∂ν Lb are two linear boundary integral operators; while Lv f and ∂ν Lv f are two volume integrals. By BIE, d has a SLPR d(x) = (Lb η)(x) + (Lv f )(x) where f (x) = (−∆ + aI)u(x) is known. BC in (2.5) leads to solve (∂ν Lb η)(x) = ∂u(x) − g(x, u(x)) − (∂ν Lv f )(x) x ∈ Γ ∂ν for layer density η. Since accurate evaluation of volume integrals Lv f, ∂ν Lv f are expensive and even impossible when Ω is unbounded, for efficiency, we try to avoid them in all computation of d and J. Next we process to identify a subspace and its expression so that all numerical computation and analysis can be carried out inside the subspace based on information of functions on the boundary Γ only. 1 From [4], for the inner product ⟨·, ·⟩a defined in (2.3), H 1 (Ω) = H 2 (Γ) ⊕ H01 (Ω) where 1 H 2 (Γ) denotes the ⟨·, ·⟩a -orthogonal complement of H01 (Ω) in H 1 (Ω) and has an ⟨·, ·⟩a orthogonal basis constructed by the Steklov eigenfunctions {ek }∞ k=1 satisfying { (−∆ + aI)ek (x) = 0, x ∈ Ω, (2.8) ∂ek (x) = λ e (x), x ∈ Γ, k k ∂ν where ∥ek ∥a = 1 and λ1 < λ2 ≤ λ3 ≤ · · · are the Steklov eigenvalues. Define a subspace Ha = {u ∈ H 1 (Ω) : (−∆ + aI)u(x) = 0}. (2.9) 1 It is clear that {ek }∞ k=1 ⊂ Ha , Ha is ⟨·, ·⟩a -orthogonal to H0 (Ω) and contains all solutions of 1 (1.1). Thus H 2 (Γ) = H̄a . So a solution to (1.1) can be approximated by using the Steklov eigenfunctions {ek }∞ k=1 ⊂ Ha . As for LMM, we show that all numerical approximation can be carried out with information of functions in Ha on the boundary Γ only. To see this, first let u ∈ Ha or f (x) = (−∆ + aI)u(x) = 0 on Ω. Then (2.5) reduces to solve d ∈ Ha from { (−∆ + aI)d(x) = 0, x ∈ Ω, (2.10) ∂u(x) ∂d(x) = − g(x, u(x)), x ∈ Γ. ∂ν ∂ν By SLPR d(x) = (Lb η)(x) and the layer density η can be solved from the linear system (∂ν Lb η)(x) = according to the following result. ∂u(x) − g(x, u(x)), ∂ν x ∈ Γ, (2.11) A LOCAL MINIMAX-BEM METHOD 7 Lemma 1. (Theorem 7.8.2. in [15]) Let a ̸= 0 and Γ be smooth. For some r ≥ − 12 , consider the interior Neumann boundary-value problem { (−∆ + aI)w(x) = 0, x ∈ Ω, (2.12) ∂w(x) r = b(x) ∈ H (Γ). ∂ν ∫ 1 ∂E(|x − ξ|) η(x) + η(ξ)dσξ = b(x), x ∈ Ω, 2 ∂ν Γ has a unique solution η ∈ H r (Γ), and the solution w of (2.12) is given by the SLPR ∫ w(x) = E(|x − ξ|)η(ξ)dσξ ∈ H r+3/2 (Ω), x ∈ Γ. (2.13) Then the BIE Γ Since in LMM, L ⊂ Ha , in the iteration, if v k ∈ Ha with ∥v k ∥a = 1, we have p(v k ) = tk0 v k + vLk ∈ Ha for some tk0 > 0 and vLk ∈ L, and then dk = ∇J(p(v k )) ∈ Ha by (2.10). k −sk dk Consequently, v k+1 = ∥vvk −s ∈ Ha . Thus LMM is invariant in Ha . k dk ∥ a We see that all the terms needed for computation and analysis in Ha can be expressed by the information of its functions on Γ only, as in (2.21)-(2.25). But the space Ha is not closed. So for analysis purpose, we define a weak form of Ha by ∫ ∂u(x) 1 H̃a = {u ∈ H (Ω) : ⟨u, v⟩a = v(x)dσx , ∀v ∈ H 1 }. (2.14) ∂ν Γ 1 It is clear that Ha ⊂ H̃a and H̃a is ⟨·, ·⟩a -orthogonal to H01 (Ω). Thus H̃a ⊂ H 2 (Γ). The map from u to its Neumann derivative is related to the Hilbert transformation. As 1 the classical Sobolev embedding from H 1 (Ω) to H 2 (Γ) is continuous, this would prompt 1 to consider the problem in an alternative way. For u ∈ H 2 (Γ) we may use the norm ∫ ||u||2Γ = Γ ( ∂u u + u2 )dσ induced by the inner product [5] ∂ν ∫ 1 ∂u(x) ⟨u, v, ⟩Γ = ( v(x) + u(x)v(x))dσx , ∀u, v ∈ H 2 (Γ). (2.15) ∂ν Γ The above process is also related to the recent work on fractional Laplacian, e.g., [5,12,13]. 1 Theorem 3. H̃a is closed and thus H 2 (Γ) = H̃a . Proof. By the Sobolev trace theorem, we see that the two norms ∥u∥2a and ∫ ∫ ∫ 2 2 2 2 2 ∥u∥a′ = ∥u∥a + u (x)dσx = (|∇u(x)| + au (x))dx + u2 (x)dσx Γ Ω (2.16) Γ are equivalent and a set is ⟨·, ·⟩a -orthogonal to H01 (Ω) iff it is ⟨·, ·⟩a′ -orthogonal to H01 (Ω) where ∫ ∫ ∫ ⟨u, v⟩a′ = ⟨u, v⟩a + u(x)v(x)dσx = (∇u(x) · ∇v(x) + au(x)v(x))dx + u(x)v(x)dσx . Γ Ω Γ (2.17) A LOCAL MINIMAX-BEM METHOD 8 1 Let uk ∈ H̃a and uk → u0 ∈ H 2 (Γ) in ∥ · ∥a -norm. We have uk → u0 in ∥ · ∥a′ -norm and Thus ⟨uk , v⟩a′ = ⟨uk , v⟩Γ , ∀v ∈ H 1 (Ω). (2.18) ⟨uk , v⟩a′ → ⟨u0 , v⟩a′ , ∀v ∈ H 1 (Ω). (2.19) 1 2 Since the classical Sobolev embedding from H 1 (Ω) to H (Γ) is continuous, we have uk → u0 1 in ∥ · ∥Γ -norm. So ⟨uk , v⟩Γ → ⟨u0 , v⟩Γ , ∀v ∈ H 2 (Γ). On the other hand, since ⟨uk , v⟩Γ = 0 and ⟨u0 , v⟩Γ = 0, ∀v ∈ H01 (Ω), it leads to ⟨uk , v⟩Γ → ⟨u0 , v⟩Γ , ∀v ∈ H 1 (Ω). Together with (2.18) and (2.19), we obtain ⟨u0 , v⟩a′ = ⟨u0 , v⟩Γ , ∀v ∈ H 1 (Ω), which implies u0 ∈ H̃a . So H̃a 1 1 1 is closed. Since Ha ⊂ H̃a ⊂ H 2 (Γ) and H̄a = H 2 (Γ), we conclude H 2 (Γ) = H̃a . Since ∇J(u) is used in our numerical computation and analysis, we need to verify Lemma 2. ∇J(u) ∈ H̃a ⇔ u ∈ H̃a . (2.20) Proof. Let u ∈ H̃a . Note Ha ⊂ H̃a and H̃a is closed. There are uk ∈ Ha s.t. uk → u. By (2.10), we have ∇J(uk ) ∈ Ha . Since J is C 1 , we obtain ∇J(uk ) → ∇J(u) ∈ H̃a . On the other hand, if d = ∇J(u) ∈ H̃a , from (2.2) and (2.3), we have ∫ ⟨d, v⟩a = ⟨u, v⟩a − g(x, u(x))v(x)dσx , ∀v ∈ H 1 (Ω). Γ But H̃a is ⟨·, ·⟩a -orthogonal to For each v ∈ H01 (Ω) we have ⟨d, v⟩a = 0 ⇒ ⟨u, v⟩a = 0, 1 or u ∈ (H01 (Ω))⊥a = H 2 (Γ) = H̃a where ⊥a means ⟨·, ·⟩a -orthogonal complement. H01 (Ω). Since all solutions of (1.1) are in H̃a and for u, v ∈ H̃a , ∫ ∂u(x) ⟨u, v⟩a = v(x)dσx , Γ ∂ν ∫ ∂u(x) 2 2 ∥u∥ := ∥u∥a = u(x)dσx ≥ 0, Γ ∂ν ] ∫ [ 1 ∂u(x) J(u) = u(x) − G(x, u(x)) dσx , Γ 2 ∂ν ∫ [ ] ∂u(x) ′ (J (u), v) = − g(x, u(x)) v(x)dσx (compare to (2.2)), ∂ν Γ u ∈ H̃a , u = 0 ⇔ ⟨u, v⟩a = 0 ∀v ∈ H̃a , (2.21) (2.22) (2.23) (2.24) ∇J(u) ∈ H̃a ⇔ u ∈ H̃a , we conclude that all numerical computation and analysis of (1.1) can be carried out efficiently in H̃a based on information of functions only on the boundary Γ. With the above notations, the condition (2.4) can be conveniently written as: u∗ ∈ Ha is a solution to (1.1) iff ∫ ∫ ∂u∗ (x) v(x)dσx = g(x, u∗ (x))v(x)dσx ∀v ∈ H̃a , (2.25) ∂ν Γ Γ or directly as in the unknown layer density η on Γ in SLPR u = Lb η and ∂b η(·) = g(·, Lb η(·)). A LOCAL MINIMAX-BEM METHOD 2.4 9 More on BEM to Solve Linear System (2.11) In LMM, SLPR in BIE leads to evaluate the gradient d by ∫ d(x) = E(|x − ξ|)η(ξ)dσξ = (Lb η)(x), x ∈ Γ, Γ 1 ∂d(x) = η(x) + ∂νx 2 ∫ Γ ∂ E(|x − ξ|)η(ξ)dσξ = (∂ν Lb η)(x), ∂νx x ∈ Γ. To solve for the unknown layer density η by BEM, we partition Γ = ∪nj=1 Γj and each Γj is a line segment centered at xj . Assume η(x) = ηj for x ∈ Γj , η = (η1 , ..., ηn ). Denote the matrices A = (aij )n×n and DA = (daij )n×n by ∫ aij = E(|xi − ξ|)dσξ , Γj { ∫ 0, i ̸= j 1 ∂ daij = δij + E(|xi − ξ|)dσξ , δij = , 2 1, i = j Γj ∂νx where for E(|x − ξ|) = √ 1 K ( a|x 2π 0 − ξ|), we have √ √ ∂ − a (x − ξ) · νx E(|x − ξ|) = K1 ( a|x − ξ|) ∂νx 2π |x − ξ| and K1 is the modified Bessel function of order 1. The integrals in aij and daij can be efficiently evaluated by, e.g., the Gaussian quadrature [15]. In particular, when i = j and xi , ξ are on the same line segment Γi , we have (xi − ξ) · νx = 0. So there will not be any singularity involved in numerical evaluation of daij . Then BC ∂d(x) ∂u(x) = b(x) ≡ − g(x, u(x)) ∂νx ∂ν leads to solve an (n × n) linear matrix system DA η = B (2.26) for η where B = (b(x1 ), ..., b(xn )). Finally we can evaluate the gradient d = Aη where d = (d(x1 ), ..., d(xn ))T . Note that in all iterations for all solutions, the matrices A, DA remain the same. So to solve the linear matrix system (2.26) repeatedly for different B, we only need to do LU decomposition of DA once for all. Thus such an algorithm is very efficient. We conclude this section by state the following result. Proposition 1. LMM-BEM is invariant in Ha and solutions to the problem (1.1) can be solved in Ha and by using information of u ∈ Ha only on Γ. A LOCAL MINIMAX-BEM METHOD 3 10 Solution Existence, Multiplicity and Properties Since most existence and multiplicity results in the literature focus on cases where nonlinearity appears in equations not BCs, in this section, we establish the existence and multiplicity of solutions to (1.1) where nonlinearity appears in Neumann type BCs. In order to check convergence of LMM-BEM and to estimate the instability index of solutions computed by LMM-BEM as stated in Theorems 1 and 2, we also verify some related conditions in the theorems. Those analysis can be carried out on the space H 1 (Ω). However, since the notations in (2.21)-(2.25) show a great advantage on the convenience and efficiency in developing our numerical method, in order to make a direct connection between our numerical method and the analysis below, we focus on H = H̃a with the notions in (2.21)-(2.25) and ∥ · ∥ = ∥ · ∥a . 3.1 Main Theoretical Results Assume the standard regularity and growth condition (p1 ) g ∈ C 1 (Γ × R, R), g(x, 0) = gξ (x, ξ) = 0; ξ=0 (p2 ) there are constants C1 , C2 > 0 s.t. |g(x, ξ)| ≤ C1 + C2 |ξ|α , ∀x ∈ Γ, n where 1 < α < p∗ ≡ n−2 for n ≥ 3 and p∗ = +∞ for n = 2; (p3 ) There exist µ > 2, M > 0 s.t. 0 ≤ µG(x, ξ) ≤ ξg(x, ξ), ∀|ξ| > M, x ∈ Γ; (p4 ) gξ (x, ξ) > g(x,ξ) , ∀x ∈ Γ and ξ ̸= 0. ξ By the Sobolev trace inequality, for s ≥ 1, it holds ∫ ∫ 1 1 2 [ Γ u(x) ∂u (x)dσ ] [ (x)dσx ] 2 u(x) ∂u x ∂ν ∂ν Γ inf ∫ ≥ inf ≥ t0 > 0 (3.1) 1 u∈H [ u∈H s+1 dσ ] s+1 C∥u∥W 1,s+1 (Ω) |u(x)| x Γ ∫ for some constant C > 0, where ∥u∥s+1 = (|∇u(x)|s+1 + |u(x)|s+1 )dx. Then following 1,s+1 W (Ω) Ω a standard approach as in [25] or [30], we have Lemma 3. Under (p1 ) and (p2 ), J is C 2 and satisfies the PS condition. Theorem 4. If g satisfies (p1 ), (p2 ) and (p3 ), then equation (1.1) has at least three nontrivial solutions. Proof. The proof of the theorem is divided into three steps. Step 1: The existence of two nontrivial solutions — a mountain pass type approach. Define for each x ∈ Γ, { g(x, t), t ≥ 0, g̃(x, t) = (3.2) 0, t < 0. 1 Consider the modified functional defined on H := H 2 (Γ) ∫ [ ] 1 ∂u(x) ˜ J(u) = u(x) − G̃(x, u(x)) dσx , Γ 2 ∂ν (3.3) A LOCAL MINIMAX-BEM METHOD 11 where G̃t (x, t) = g̃(x, t). Then following a standard approach as in [25] or [30], one may check that J˜ ∈ C 2 (H, R) and satisfies the PS condition. It is clear that under (p1 ), (p2 ), (p3 ), ˜ 1 ) → −∞ as t → ∞. We can choose r > 0 large s.t. J(re ˜ 1 ) < 0. Set we have J(te Γc = {γ ∈ C 0 ([0, 1], H) : γ(0) = 0, γ(1) = re1 } and ˜ c1 = inf sup J(γ(t)). (3.4) γ∈Γc t∈[0,1] By Lemma 4, c1 > ε > 0 and applying the Mountain Pass lemma, we obtain that c1 is a critical value of J˜ or there exists u1 ∈ H s.t. ∫ ∫ ∂u1 (x) v(x)dσx = g̃(x, u1 (x))v(x)dσx ∀v ∈ H. ∂ν Γ Γ { Denote u− 1 (x) = u1 (x), u1 (x) ≤ 0 0, u1 (x) > 0 x ∈ Γ. It is clear that u− 1 ∈ H and ∫ ∫ ∫ ∂u− ∂u1 (x) − 1 (x) − − 2 ∥u1 ∥ = u1 (x)dσx = u1 (x)dσx = g̃(x, u1 (x))u− 1 (x)dσx = 0. ∂ν ∂ν Γ Γ Γ Thus u1 ≥ 0. By the Harnack inequality, we conclude u1 > 0 in Ω. In this case, if u1 (x0 ) = 0 0) for some x0 ∈ Γ, by the strong maximum principle, we should have ∂u(x < 0, which ∂ν contradicts to g(x, 0) = 0. So u1 is strictly positive on Ω̄. It follows g̃ = g, J˜ = J and hence u1 is also a critical point of J. By a similar procedure, we can obtain the other solution u2 < 0. Without loss of generality, we assume that c1 = J(u1 ) ≤ J(u2 ) = c2 . Step 2: The local behavior of J near u1 . In following the argument developed in [30] to get another solution, the key step is to use a generalized Morse lemma to establish certain local behavior of J around the known solution u1 , see Lemma 5. Since the rest of the argument is standard, we refer to [30] for more details. So let us denote N = ker J ′′ (u1 ) and apply Lemma 5 to the function J in a neighborhood of u1 , say Bδ = {u ∈ H : ∥u − u1 ∥ < δ}, then we have 1 J ◦ Φ(u1 + z + y) = ⟨J ′′ (u1 )z, z⟩ + J(u1 + h(y) + y), 2 ∀u ∈ Bδ , (3.5) where Φ : Bδ (u1 ) → H is a homeomorphism preserving u1 and h : Bδ (θ) ∩ N → N ⊥ is a C 1 mapping with h(θ) = 0. Then by Lemma 6, the local behavior of J around u1 is given by (a) h(y) ∈ C 1 (Ω̄), for ∥y∥ < δ and (b) ∥h(y)∥C 1 → 0 as ∥y∥ → 0. A LOCAL MINIMAX-BEM METHOD 12 Step 3. This step is exactly the same as in [30] following a topological argument and we sketch it here. Lemma 3.2 of [30] can be proved here similarly to give us that for K > 0 large enough J −K ≡ S ∞ (homotopy equivalent). Now the local information near u1 and u2 from (3.5) can be used to show that the critical groups at u1 or at u2 are δk1 F (with F being the coefficients group) since they are both mountain pass critical points. Finally if we assume J has exactly three critical points 0, u1 , u2 we derive a contradiction as in the proof of Theorem 3.1 of [30] by using the above information. Thus J has another critical point besides 0, u1 and u2 . By following the proof of Theorem 9.12 in [25], if in addition to (p1 ) ∼ (p3 ), g(x, ξ) is odd in ξ, we can establish the existence of infinitely many solutions to (1.1). Next by following standard results [25, 30], if the boundary Γ is smooth, then the solutions we get are classical ones and the BIE/BEM approach described in Section 2.3 can be applied. Lemma 4. For (3.3), there exist constants ρ, ε > 0 s.t. ˜ J(u) >0 ˜ ∀0 < ∥u∥ < ρ and J(u) > ε, ∀∥u∥ = ρ. Proof. The first Steklov eigenpair (λ1 , e1 ) satisfies ∫ λ1 = min where ∥u∥2L2 (Γ) = u∈H ∫ g̃(x, t) ≤ Γ Γ ∂u(x) u(x)dσx ∂ν 2 ∥u∥L2 (Γ) > 0, (3.6) u2 (x)dσx . By (p1 ) and (p2 ), there exists C > 0 s.t. λ1 |t| + C|t|α 4 and G̃(x, u(x)) ≤ λ1 C |u(x)|2 + |u(x)|α+1 . 4 α+1 Thus with (3.6) and Sobolev embedding inequality, we have ∫ [ ∫ 1 ∂u(x) λ1 2 ] C ˜ J(u) ≥ u(x) − u (x) dσx − |u(x)|α+1 dσx 2 ∂ν 4 α + 1 Γ Γ 1 1 C 1 C0 0 2 ≥ ∥u∥2 − ∥u∥2 − ∥u∥α+1 ∥u∥α+1 Lα+1 (Ω) = ∥u∥ − Lα+1 (Ω) , 2 4 α+1 4 α+1 ∫ α+1 where ∥u∥α+1 dx. Since α + 1 > 2, the Lemma is verified. Lα+1 (Ω) = Ω |u(x)| Lemma 5. (Generalized Morse Lemma) [30] Suppose U is a neighborhood of the original θ in a Hilbert space H with inner product (·, ·) and g ∈ C 2 (H, R). Assume that θ is the only one critical point of g in U and that A = g ′′ (θ) with kernel N . If θ is at most an isolated point of the spectrum ρ(A), then there exist a ball Bδ centered at θ, an origin preserving local homeomorphism Φ and a C 1 mapping h : Bδ ∩ N → N ⊥ s.t. 1 g ◦ Φ(z + y) = (Az, z) + g(h(y) + y), ∀u ∈ Bδ , 2 where y = PN u, z = PN ⊥ u and PN is the orthogonal projection onto N . A LOCAL MINIMAX-BEM METHOD 13 Lemma 6. In (3.5), we have (a) h(y) ∈ C 1 (Ω̄) ∀∥y∥ < δ and (b) ∥h(y)∥C 1 → 0 as ∥y∥ → 0. Proof. Conclusion (a) comes directly from the generalized Morse lemma, i.e., Lemma 5. So we only need to prove (b). Recall that h(y) is the unique solution of the equation L(z) = i.e., ∂J (u1 + z + y) = 0, ∂z ∂J ≡ 0, (u + 1 + z + y) ∂z z=h(y) ∥y∥ ≤ δ, ∀∥y∥ ≤ δ, or by (2.25) we have ∫ [ ∫ ] ∂ (u1 (x) + h(y)(x) + y(x)) z(x)dσx = g(x, u1 (x) + h(y)(x) + y(x))z(x)dσx , Γ ∂ν Γ Taking J ′ (u1 ) = 0, J ′′ (u1 )y = 0 into account, we have ∫ ∫ ∂h(y)(x) z(x)dσx = f (y)(x)z(x)dσx , ∂ν Γ Γ ∀z ∈ H. ∀z ∈ H (3.7) where f (y)(x) = g(x, u1 (x) + h(y)(x) + y(x)) − g(x, u1 (x)) − gξ (x, u1 (x))y(x). Since ∫ ∫ { ∂w(x) v(x)dσx = gξ (x, u1 (x))w(x)v(x)dσx N = w∈H: ∂ν Γ Γ ∀v ∈ H (3.8) } has a finite dimension, we can assume N = span{w1 , ..., wk } ⊂ H, all wi ’s are mutually ⟨·, ·⟩a -orthogonal with ∥wi ∥ = 1. We claim that for each y ∈ N , there are scalars βi = βi (y), i = 1, ..., k s.t. ∂h(y) ∑ ∂wi + βi = f (y) on Γ, (3.9) ∂ν ∂ν ∑ where f (y) is given by (3.8). To prove the claim, since h(y) + βi wi ∈ H and by (2.25), we see that (3.9) is equivalent to ∑ ∫ ∫ ∂(h(y) + βi wi )(x) v(x)dσ = f (y)(x)v(x)dσx ∀v ∈ H. (3.10) ∂ν Γ Γ ∫ ∑ If we write v = w + z ∈ N ⊕ N ⊥ and note that h(y) ∈ N ⊥ , βi wi ∈ N and Γ u(x) v(x)dσ = ∂ν ⟨u, v⟩a , then (3.10) becomes ∫ ∫ ∫ ∑ ∫ wi (x) ∂h(y)(x) z(x)dσx + βi w(x)dσx = f (y)(x)z(x)dσx + f (y)(x)w(x)dσx . ∂ν Γ ∂ν Γ Γ Γ A LOCAL MINIMAX-BEM METHOD 14 It follows from (3.7) that βi ’s must satisfy ∫ ∑ ∫ ∂wi (x) βi w(x)dσx = f (y)(x)w(x)dσx ∂ν Γ Γ which implies for each i, ∀w ∈ N, ∫ βi = f (y)(x)wi (x)dσx . (3.11) Γ One can check that βi ’s given by (3.11) solve (3.9). By the assumption on g and the property of h, we have βi ∈ C 1 (Bδ (θ) ∩ N, R) and βi (θ) = 0. Then by (p2 ) and Lp estimate, for any p > 2, there exists constant Cp , C1 > 0 s.t. ∑ ∥h(y)∥W 2,p (Ω) ≤ Cp (∥h(y)∥ + ∥y∥ + max |βi | + 1) and ∥h(y)∥C 1 (Ω̄) ≤ C1 , ∀∥y∥ ≤ δ. From this, one may find a constant C > 0 s.t. |g(x, u1 (x) + h(y)(x) + y(x)) − g(x, u1 (x))| ≤ C(∥h(y) + ∥y∥ + ∑ max |βi |), which gives us the estimate (b) of the lemma. We remark that by arguing some more we can claim the solutions obtained contain signchanging ones except the mountain pass ones (which are one sign solutions). This requires the methods of invariant sets with gradient flows as used in [9](see [23] for more references therein). Let P± = {u ∈ H | ± u ≥ 0} and W± = {u ∈ H | d(u, P± ) ≤ ϵ}. For u ∈ H ∂v let A(u) be the solution of the equation ∂ν = g(x, u) on Γ. Then the gradient of J is of the form u − A(u) in Hilbert space H. We show that A is attractive in small neighborhoods of W± . Namely we have Proposition 2. For ϵ > 0 small enough, A(∂W± ) ⊂ int (W± ). The proof of this is by now quite standard. We sketch it here. ∫ ∫ ∂(v − w) ∂vv− 2 2 d(v, P ) = inf ||v − w|| = inf (v − w)dσ ≤ v− dσ w∈P w∈P Γ ∂ν ∂ν Γ ∫ = g(x, u− )v− dσ ≤ c||u− ||2 ||v− ||2 + C||u− ||αp∗ +1 ||v− || ∗p∗ +1 Γ p +1−α 1 ≤ ( d(u, P ) + d(u, P )α )d(v, P ). 4 By choosing ϵ > 0 small, if d(u, P ) = ϵ we have d(v, P ) ≤ 12 ϵ. With this device we may employ minimax methods to construct critical values with critical points outside W± . Then in the setting of Theorem 4 we obtain one positive, one A LOCAL MINIMAX-BEM METHOD 15 negative and one sign-changing solutions. If the nonlinearity is odd in u we also obtain infinitely many sign-changing solutions by a standard argument, see [8,9,10,20,23,31]. In the rest of this section, we verify some conditions posed in Theorems 1 and 2 for solution instability index estimate and convergence of LMM-BEM. Proposition 3. Under (p1 ) ∼ (p4 ), when L = {0}, the peak selection p(u) = tu u is uniquely defined for each u ∈ SH and is C 1 . Furthermore we have J(p(u)) > 0. Proof. By (p1 ) and (p3 ), G(x, t) is super-quadratic in t and nonnegative. For each given u ∈ SH , we have J(tu) > 0 for small t > 0, J(tu) → 0 as t → 0 and J(tu) → −∞ as t → ∞. So there exists a tu > 0 s.t. ∫ [ ∂u(x) ] d J(tu) = tu u(x) − g(x, tu u(x))u(x) dσx = 0. (3.12) dt t=tu ∂ν Γ Then d2 J(tu) = dt2 t=tu ∫ [ ∂u(x) ] u(x) − gξ (x, tu u(x))u2 (x) dσx ∂ν ∫Γ [ ∂u(x) g(x, tu u(x)) 2 ] < u(x) − u (x) dσx (by (p4 )) ∂ν tu u(x) Γ ∫ [ ∂u(x) ] 1 = tu u(x) − g(x, tu u(x))u(x) dσx = 0. tu Γ ∂ν (3.13) Thus the peak selection p(u) = tu u is uniquely defined for each u ∈ SH . Since J(tu) > 0 for d2 small t > 0, by the uniqueness and dt2 J(tu) < 0, we must have J(p(u)) = J(tu u) > 0. t=tu d By applying the implicit function theorem to the equation F (t, u) = dt J(tu) = 0, we note 2 that Ft (t, u) J(tu) < 0. Hence the peak selection p(u) = tu u is C 1 . = dtd 2 t=tu t=tu Proposition 4. Under (p1 ) ∼ (p4 ), for any nonempty closed subspace L ⊂ H, we have inf u∈L⊥ ,∥u∥=1 ∥p(u)∥ > 0 and inf u∈L⊥ ,∥u∥=1 J(p(u)) > 0. ∫ Proof. We first assume L = {0}. For each u ∈ H with ∥u∥2 = Γ ∂u(x) u(x)dσx = 1, from ∂ν (3.12), we have ∫ [ ∂u(x) ] d g(x, tu u(x)) tu 0 = J(tu) = u(x) − tu u(x) dσx dt t=tu ∂ν tu Γ ∫ [ ∂u(x) ] 1 λ1 ≥ tu u(x) − ( |tu u(x)|2 + C|tu u(x)|α+1 ) dσx (by (p2 )) ∂ν tu 4 Γ ∫ ] [ ∂u(x) λ1 u(x) − |u(x)|2 − Ctαu |u(x)|α+1 dσx = tu ∂ν 4 ∫Γ [ ∂u(x) ] 1 ∂u(x) ∂u(x) ≥ tu u(x) − u(x) − C1 tαu u(x) dσx (by (3.6) and (3.1)) ∂ν 4 ∂ν ∂ν Γ ∫ 3 = tu ( − C1 tαu ), (since ∥u∥2 = Γ ∂u(x) u(x)dσx = 1), ∂ν 4 A LOCAL MINIMAX-BEM METHOD which implies 16 ( 3 ) α1 3 − C1 tαu ≤ 0 or tu ≥ t0 = > 0. 4 4C1 Thus ∥p(u)∥ = ∥tu u∥ = tu > t0 > 0 ∀u ∈ H, ∥u∥ = 1. By H = L⊥ , we have inf u∈L⊥ ,∥u∥=1 ∥p(u)∥ > 0 and inf u∈L⊥ ,∥u∥=1 J(p(u)) > 0. (3.14) Denote N = {p(u) = tu u : u ∈ H, ∥u∥ = 1, ∇J(p(u))⊥u}. Then the two formulas in (3.14) are equivalent to inf ∥w∥ > 0 and w∈N inf J(w) > 0. w∈N (3.15) For any nonempty closed subspace L ⊂ H, denote M = {p(u) ∈ [L, u] : u ∈ L⊥ , ∥u∥ = 1, ∇J(p(u))⊥[L, u]}. (3.16) Then the two formulas in (3.14) become inf ∥w∥ > 0 and w∈M inf J(w) > 0. w∈M (3.17) Since M ⊂ N , from (3.15), the two formulas in (3.17) are also true. ∑ For L = span{u1 , ..., un−1 }, let un = p(u0 ) = t0 u0 + n−1 i=1 ti ui be a solution characterized by Theorem 1, or found by LMM-BEM. Following the argument in [35,36] or (3.16), a peak selection p at un can be computed by solving the equations Fj (t0 , t1 , ..., tn−1 , u0 ) = ⟨∇J(t0 u0 + n−1 ∑ ti ui ), uj ⟩ = 0, j = 0, 1, ..., n − 1, (3.18) i=1 for t0 , t1 , ..., tn−1 . If the n × n matrix Q=[ where ∂F (t0 , t1 , ..., tn−1 , u0 )] = [⟨J ′′ (un )ui , uj ⟩], ∂ti i, j = 0, 1..., n − 1, ∫ [ ] ∂ui (x) − gξ (x, un (x))ui (x) uj (x)dσx ⟨J (un )ui , uj ⟩ = ∂ν Γ ′′ (3.19) (3.20) is nonsingular, then by the implicit function theorem, the peak selection p is locally C 1 near un and this condition can be easily numerically checked in the algorithm. It is clear that when n = 1, (3.19) and (3.20) become (3.13), So Q is nonsingular in this case. A LOCAL MINIMAX-BEM METHOD 4 17 Numerical Examples In this last section we carry out numerical computation on some examples by our LMMBEM algorithm developed in the previous sections and display the numerical solutions. We x2 x2 set g(x, u(x)) = u3 (x) and a = 1 in (1.1) and choose domains Ω1 = {(x1 , x2 ) : 41 + 12 < 1}, Ω2 = {(x1 , x2 ) : x21 + x22 < 1}, Ω3 = (−1, 1)2 , Ω4 = a dumbbell as in Figure 1. It is a direct verification to check that assumptions (p1 ) ∼ (p4 ) are all satisfied. Next we discuss how to choose initial guesses in LMM-BEM: when Ω is a nice convex domain, we use the Steklov eigenfunctions uk = ek defined in (2.8) or use uk = Aηk where either ηk (x) = cos(kθ(x)) or ηk (x) = sin(kθ(x)) to generate a k-periodic function on Γ. On Ω1 , we use the Steklov eigenfunctions as initial guesses. Numerical solutions are shown in Figures 2 and 3. On Ω2 , this is a degenerate case. All the numerical Steklov eigenfunctions are doubled. So we have to get rid of the multiplicity and choose only one from each group as an initial guess. Note that the first solution is radial positive with 1-peak at the center of the domain. Numerical solutions are shown in Figures 4 and 5. On Ω3 , due to the corner affect, using the Steklov eigenfunctions as initial guesses will obtain only a subset of solutions. So we use sin(kπ ∗ i/n) or cos(kπ ∗ i/n) to be initial guesses and obtained much more solutions, including the first two positive solutions that are 4-rotation asymmetric and the third one is 4-rotation symmetric with 1-peak at the center of the domain. Numerical solutions are shown in Figures 6 and 7. On Ω4 , due to the complexity of the domain (not convex), we have used local symmetries to create initial guesses. The domain is shown in Fig. 1 and the numerical solutions are shown in Figures 8 and 9. In all our numerical computations, iterations are terminated when ∥∇J(uk )∥ < 10−5 and maxx∈Γ | ∂u (x) − g(x, u(x))| < 10−5 . In order to plot 2D contours and 3D profile of a solution ∂ν in one figure, we have shifted the 3D profile up or down. Although solution profiles are plotted on a 2D domain Ω, the problem is actually solved only on the boundary Γ. Due to this dimension reduction, the numerical algorithm is thus much more efficient. 1 Interior domain: counterclockwise 0.8 0.6 0.4 0.2 0 −0.2 −0.4 −0.6 −0.8 −1 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5 3 Figure 1: Boundary points start from the left endpoint (−1.5, 0) counterclockwise. A LOCAL MINIMAX-BEM METHOD 1.2 18 1.5 2 1.1 1.5 1 1 1 0.9 0.5 0.5 0.8 0.7 0 0 0.6 −0.5 −0.5 0.5 −1 0.4 −1 −1.5 0.3 0.2 0 1 2 3 4 5 6 −1.5 7 (1) 0 1 2 3 4 5 6 −2 7 (2) 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 (3) 1.5 2 2 1 1.5 1.5 1 1 0.5 0.5 0.5 0 0 0 −0.5 −0.5 −0.5 −1 −1 −1 −1.5 −1.5 −1.5 −2 −2 0 1 2 3 4 5 6 7 (4) 0 1 2 3 4 5 6 −2 7 (5) 2 (6) 2.5 2.5 2 2 1.5 1.5 1.5 1 0.5 1 1 0.5 0.5 0 −0.5 0 0 −0.5 −0.5 −1 −1 −1 −1.5 −1.5 −2 −1.5 −2 0 1 2 3 4 5 6 −2.5 7 (7) −2 0 1 2 3 4 5 6 −2.5 7 (8) (9) Figure 2: Profiles of solutions u1 ∼ u9 on Γ1 . 0 0 −0.5 0 −0.2 −1 −0.5 −0.4 −1.5 −0.6 −1 −2 −0.8 −1.5 −2.5 −1 −1.2 −2 −1.4 1 −3 −3.5 −2.5 1 2 0.5 0.5 1 0 −1 −1 −0.5 −0.5 −1 0 −2 (1) −4 −1 0 −3 −2 0 −0.5 1 2 0 −1 (2) 0.5 1 −2 −1 0 1 2 (3) 0 0 0 −0.5 −0.5 −0.5 −1 −1 −1 −1.5 −1.5 −2 −2.5 −1.5 −2 −2 −2.5 −3 −2.5 −3.5 −3 −3 −4 −3.5 −3.5 1 −4 1 2 −4 0 0 0 −4.5 −2 −1 −0.5 −1 −4.5 1 −2 (4) 2 0.5 1 0.5 1 0 0 −0.5 −1.5 −1 −0.5 0 0.5 1 1.5 2 −1 −1 −1 (5) −2 (6) 0 0 0 −0.5 −0.5 −1 −1 −1 −1.5 −1.5 −2 −2 −2 −2.5 −2.5 −3 −3 −4 −3.5 −3 −4 −3.5 −5 −4.5 −4 2 −4.5 −5 1 −5 1 (7) 0 −1 −0.5 −1 1 0 0 0.5 2 0.5 1 −2 2 0.5 1 0 −1 −1 (8) −6 1 0 −0.5 0 −0.5 −1 −1 −2 −2 (9) Figure 3: Profiles and contours of solutions u1 ∼ u9 on Ω1 with (J(u1 ), ..., J(u9 )) = (0.5750, 0.7025, 1.7940, 2.7011, 4.0194, 6.0398, 7.1387, 10.7212, 11.4944). u1 is positive. A LOCAL MINIMAX-BEM METHOD 19 0.6766 2 2 0.6766 1.5 1.5 0.6766 1 1 0.5 0.5 0.6766 0.6766 0 0 −0.5 −0.5 0.6766 0.6766 0.6766 0.6766 0.6766 0 1 2 3 4 5 6 −1 −1 −1.5 −1.5 −2 7 (1) 0 1 2 3 4 5 6 −2 7 (2) 3 3 2 2 2 1 1 1 0 0 0 −1 −1 −1 −3 −2 0 1 2 3 4 5 6 −3 7 (4) 1 2 3 4 5 6 −3 7 (5) 4 3 3 3 2 2 2 1 1 0 0 0 −1 −1 −1 −2 −2 −2 −3 −3 1 2 3 4 5 6 −4 7 (7) 3 4 5 6 7 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 (6) 4 0 2 −2 0 4 −4 1 (3) 3 −2 0 1 −3 0 1 2 3 4 5 6 −4 7 (8) (9) Figure 4: Profiles of solutions u1 ∼ u9 on Γ2 . 0 0 0.2 −1 −0.5 0.15 −1 −2 −1.5 0.1 −2 −3 −2.5 0.05 −3 0 1 −4 −3.5 0.5 1 0.5 0 −4 1 0 −0.5 −1 0.5 0 −0.5 −1 −1 (1) −5 −1 0 −0.5 0 1 −1 (2) 1 −1 −1 −1 −0.5 0 0.5 1 (3) 0 0 −1 −1 0 −2 −2 −2 −3 −3 −4 −4 −4 −5 −6 −5 −6 −6 1 −8 1 −7 1 1 0 −1 0.5 0 −0.5 −1 1 −1 (4) 0.5 0 0 −0.5 −1 0.5 0 0 1 (5) −0.5 (6) 0 0 −2 −2 −4 −4 0 −2 −4 −6 −6 −6 −8 −8 1 −8 1 0.5 0.5 0 −0.5 −1 (7) −10 1 −1 −0.5 0 0.5 0.5 0 1 −0.5 −1 (8) −1 −0.5 0 0.5 1 1 0.5 0 0 −0.5 −0.5 −1 −1 (9) Figure 5: Profiles and contours of solutions u1 ∼ u9 on Ω2 with (J(u1 ), ..., J(u9 )) = (0.3292, 1.3511, 4.3220, 9.2009, 16.1014, 25.1169, 36.3353, 49.8370, 65.6374). A LOCAL MINIMAX-BEM METHOD 20 0.8 −0.2 1.4 −0.3 1.2 0.75 −0.4 1 −0.5 0.7 0.8 −0.6 0.6 −0.7 0.65 −0.8 0.4 −0.9 0.2 0 0.6 −1 0 1 2 3 4 5 6 −1.1 7 (1) 0 1 2 3 4 5 6 0.55 7 (2) 1.5 0 1 2 3 4 5 6 7 1 2 3 4 5 6 7 (3) 1.5 2 1 1.5 1 0.5 1 0 0.5 0.5 0 −0.5 0 −1 −0.5 −0.5 −1 −1.5 −1.5 0 1 2 3 4 5 6 −2 7 (4) −1 0 1 2 3 4 5 6 −1.5 7 (5) (6) 1.5 2 1.5 0 2 1.5 1 1 1 0.5 0.5 0.5 0 0 0 −0.5 −0.5 −0.5 −1 −1 −1 −1.5 0 1 2 3 4 5 6 −1.5 7 (7) −1.5 0 2 4 6 −2 8 (8) 0 2 4 6 8 (9) Figure 6: Profiles of solutions u1 ∼ u9 on Γ3 . u1 ∼ u3 are one-sign solutions. 0 0 0 −0.1 −0.2 −0.2 −0.4 −0.5 −0.6 −1 −0.3 −0.8 −1 −1 −0.5 −1.5 −2 −1 −1.4 1 0.5 −0.5 0 0.5 1 −1.2 0 (1) 1 0 0 −0.5 0.5 0 1 1 −0.4 1 0.5 0 −0.5 −1 −1 −1 (2) −1 (3) 0 0 −0.5 −0.5 −1 0 −1 −1.5 −1 −2 −2 −1.5 −2 −3 −2.5 −3.5 1 −3 −1 −3 −2.5 −1 −4 1 0.5 0 0 −0.5 0 0.5 −1 1 1 (4) −1 −0.5 0 0.5 0 −0.5 −0.5 −1 (5) −1 (6) 0 0 0 1 0.5 0 1 −0.5 −0.5 −1 −1 −1 −1.5 −2 −1.5 −2 −2.5 −2 −3 −3 −3.5 −4 1 (7) −2.5 −1 −0.5 0.5 0 0 −0.5 0.5 −1 −1 −3 1 1 (8) 0.5 0 0 −0.5 −1 1 1 −4 −1 0 −0.5 0 0.5 1 −1 (9) Figure 7: Profiles and contours of solutions u1 ∼ u9 on Ω3 with (J(u1 ), ..., J(u9 )) = (0.2277, 0.3228, 0.3799, 0.5660, 0.8069, 0.9165, 0.9834, 1.1244, 1.3722). u1 and u2 are 4-rotation asymmetric. u3 is 4-rotation symmetric with 1-peak. A LOCAL MINIMAX-BEM METHOD 0.7 21 1 1.2 0.9 1 0.8 0.8 0.7 0.6 0.6 0.4 0.6 0.5 0.4 0.5 0.2 0.3 0.2 0.4 0 0.3 −0.2 0.2 −0.4 0.1 0.1 0 0 1 2 3 4 5 6 0 7 (1) −0.6 0 1 2 3 4 5 6 −0.8 7 (2) 1.5 0 1 2 3 4 5 6 7 (3) 2 2 1.5 1.5 1 1 1 0.5 0.5 0.5 0 0 0 −0.5 −0.5 −1 −1 −0.5 −1 −1.5 −1.5 0 1 2 3 4 5 6 −2 7 (4) −1.5 0 1 2 3 4 5 6 −2 7 (5) 0 1 2 3 4 5 6 7 (6) 2 2 2 1.5 1.5 1.5 1 1 0.5 0.5 0 0 −0.5 −0.5 −1 −1 1 0.5 0 −0.5 −1 −1.5 −1.5 −1.5 −2 0 1 2 3 4 5 6 −2 7 0 1 2 3 4 5 6 7 −2 (7) (8) 0 1 2 3 4 5 6 7 (9) Figure 8: The profiles of solutions u1 ∼ u9 on Γ4 . u1 and u2 are positive. 1 1.5 0.8 1 0.7 0.6 0.6 0.5 0.5 0.4 0 0.4 0.2 −0.5 0.3 0 0.2 −1 1 −1 0.1 1 0 0 0 0 −2 −1 0 1 2 1 −1 3 −1.5 (1) 0.5 0 −0.5 −1 1.5 1 2 −1 3 2.5 (2) −1.5 −0.5 −1 0 0.5 2 1.5 1 2.5 3 (3) 1.5 1.5 1.5 1 1 1 0.5 0.5 0.5 0 −0.5 −1 −1.5 1 0 0 −0.5 −0.5 −1 −1 −1.5 −1 −1.5 1 0 0 0 −1 −1.5 −1 −0.5 0.5 0 1 1.5 2 2.5 1 3 −1.5 −0.5 −1 1 0.5 0 1.5 2 3 2.5 −1 (4) (5) 2 0.5 0 −0.5 −1 −1.5 1 2.5 2 1.5 3 (6) 1.5 2 1 2 0.5 1.5 1.5 0 −0.5 1 1 0.5 −1 0.5 −1.5 0 0 −2 1 −0.5 −0.5 0.5 −1 1 0 −1.5 0 −0.5 −2 −1 −1.5 (7) −1 −0.5 0 0.5 1 1.5 2 2.5 3 −1.5 (8) −1 −0.5 0 0.5 1 1.5 −1 2 2.5 3 −1 −1.5 −2 −1.5 1 0 −1 −0.5 0 0.5 1 1.5 2 2.5 3 −1 (9) Figure 9: Profiles and contours of solutions u1 ∼ u9 on Ω4 . u1 and u2 are positive. (J(u1 ), ..., J(u9 )) = (0.0977, 0.3481, 0.4690, 1.0779, 1.3650, 1.4606, 2.1306, 2.4783, 3.3778). A LOCAL MINIMAX-BEM METHOD 22 Acknowledgment: We would like to thank the reviewer for making helpful comments. References [1] H. Amann and M. Fila, A Fujita-type theorem for the Laplace equation with a dynamic boundary condition, Acta Math. Univ. Comenianae, 2(1997) 321-328. [2] T. Ando, A.B. Fowleer and F. Stern, Electronic properties of two dimensional system, Reviews of Modern Physics, 54(1982) 437-621. [3] M.G. Armontano, The effect of reduced integration in the Steklov eigenvalue problem, Math. Model. Numer. Anal., 38(2004) 27-36. [4] G. Auchmuty, Steklov eigenproblems and the representation of solutions of elliptic boundary value problems, Numer. Func. Anal. Optimization, 25(2004) 321-348. [5] G. Auchmuty, Spectral characterization of the trace spaces H s (∂Ω), SIAM J. Math. Anal., 38(2006) 894-905. [6] K. T. Atkinson, The numerical solution of a nonlinear boundary integral equation on smooth surface, IMA J. Numer. Anal., 14(1994) 461-483. [7] A.K. Aziz, M.R. Dorr and R.B. Kellogg, A new approximation mehtod for the Helmholtz equation in an exterior domain, SIAM J. Num. Anal. 19(1982) 899-908. [8] T. Bartsch, K.-C. Chang, and Z.-Q. Wang, On the Morse indices of sign changing solutions of nonlinear elliptic problems, Math. Z., 233(2000) 655-677. [9] T. Bartsch, Z.L. Liu, and T. Weth, Sign changing solutions of superlinear Schrödinger equations, Comm. Partial Differential Equations, 29(2004) 25-42. [10] T. Bartsch, and Z.-Q. Wang, On the existence of sign changing solutions for semilinear Dirichlet problems, Topol. Meth. Nonl. Anal., 7(1996) 115-131. [11] R. Bialecki and A. J. Nowak, Boundary value problems in heat conduction with nonlinear material and nonlinear boundary conditions, Appl. Math. Modelling, 5(1981) 417421. [12] X. Cabré and J. Tan, Positive solutions of nonlinear problems involving the square root of the Laplacian, Adv. Math., 224 (2010) 2052-2093. [13] L. Caffarelli and L. Silvestre, An extension problem related to the fractional Laplacian, Comm. Par. Diff. Equa., 32(2007) 1245-1260. [14] J.M. Cushing, Nonlinear Steklov problems on the unit circle II—a hydro-dynamical application, JMAA, 39(1972) 267-278. [15] G. Chen and J. Zhou, Boundary Element Methods, Academic Press, London–San Diego, 1992. A LOCAL MINIMAX-BEM METHOD 23 [16] D. Fasino and G. Inglese, Recovering unknown terms in a nonlinear boundary condition for Laplace’s equation, IMA Journal of Appl. Math., 71(2006) 832-852. [17] M. Fila and P. Quittner, Global solutions of the Laplace equation with a nonlinear dynamical boundary condition, Math. Methods in Appl. Sciences, 20(1997) 1325-1333. [18] M. Ganesh, A BIE method for a nonlinear BVP, J. Comput. Appl. Math., 45(1993) 299-308. [19] O. Kavian and M. Vogelius, On the existence and ‘blow-up‘ of solutions to a twodimensional nonlinear boundary-value problem arising in corrosion modelling, Proceedings of the Royal Society of Edinburgh, 133A(2003) 119-149. [20] S.J. Li and Z.-Q. Wang, Ljusternik-Schnirelman theory in partially ordered Hilbert spaces, Trans. Amer. Math. Soc., 354(2002) 3207-3227. [21] Y. Li and J. Zhou, A minimax method for finding multiple critical points and its applications to semilinear elliptic PDEs, SIAM J. Sci. Comp., 23(2001) 840-865. [22] Y. Li and J. Zhou, Convergence results of a local minimax method for finding multiple critical points, SIAM J. Sci. Comp., 24(2002) 865-885. [23] Z. Liu and Z.-Q. Wang, Sign-changing solutions of nonlinear elliptic equations, Frontiers Math. in China, 3(2008) 1-18. [24] K. Medville and M. Vogelius, Existence and blow up of solutions to certain classes of two-dimensional nonlinear Neumann problems, Ann. I. H. Poincare, 23(2006) 499-538. [25] P.H. Rabinowitz, Minimax Method in Critical Point Theory with Applications to Differential Equations, CBMB Reginal Conference Series in Mathematics, vol. 65, Published for the Conference Board of the Mathematical Sciences, Washington, DC, 1986. [26] K. Ruotsalainen and J. Saramen, On the collocation method for a nonlinear boundary integral equation, J. Comput. Appl. Math., 28(1989) 339-348. [27] K. Ruotsalainen and W. L. Wendland, On the boundary element method for some nonlinear boundary value problems, Numer. Math., 53(1988) 299-314. [28] E. Vitillaro, On the Laplace equation with non-linear dynamic boundary conditions, Proc. London Math. Soc., 93(2006) 418-446. [29] M. Vogelius and J.-M. Xu, A nonlinear elliptic boundary value problem related to corrosion modeling, Q. Appl. Math., 56(1998) 479-505. [30] Z.-Q. Wang, On a superlinear elliptic equation, Ann. Inst. H. Poincaré Anal. Non Linéaire, 8(1991) 43-57. [31] Z.-Q. Wang, Minimax methods, invariant sets, and applications to nodal solutions of nonlinear elliptic problems, Proceedings of EquaDiff 03, International Conference on Differential Equations, Hasselt 2003, World Scientific, Singapore, (2005) 561-566. A LOCAL MINIMAX-BEM METHOD 24 [32] Z.Q. Wang and J. Zhou, An efficient and stable method for computing multiple saddle points with symmetries, SIAM J. Num. Anal., 43(2005) 891-907. [33] Z.Q. Wang and J. Zhou, A local minimax-Newton method for finding critical points with symmetries, SIAM J. Num. Anal., 42(2004) 1745-1759. [34] X. Yao and J. Zhou, A minimax method for finding multiple critical points in Banach spaces and its application to quasilinear elliptic PDE, SIAM J. Sci. Comp., 26(2005) 1796-1809. [35] J. Zhou, A local min-orthogonal method for finding multiple saddle points, JMAA, 291(2004) 66-81. [36] J. Zhou, Instability analysis of saddle points by a local minimax method, Math. Comp., 74(2005) 1391-1411. [37] J. Zhou, Global sequence convergence of a local minimax method for finding multiple solutions in Banach Spaces, Num. Funct. Anal. Optim., 32(2011) 1365-1380.