07-liu-shi-wangjfa

Journal of Functional Analysis 251 (2007) 573600www.elsevier.com/locate/jfa

Imperfect transcritical and pitchfork bifurcations

Ping Liu a,b, Junping Shi c,a,, Yuwen Wang aa Y.Y. Tseng Functional Analysis Research Center and School of Mathematics and Computer Sciences,

Harbin Normal University, Harbin, Heilongjiang 150025, PR Chinab School of Mathematics and Statistics, Northeast Normal University, Changchun, Jilin 130024, PR China

c Department of Mathematics, College of William and Mary, Williamsburg, VA 23187-8795, USAReceived 11 October 2006; accepted 27 June 2007

Available online 21 August 2007

Communicated by H. Brezis

Abstract

Imperfect bifurcation phenomena are formulated in framework of analytical bifurcation theory on Ba-nach spaces. In particular the perturbations of transcritical and pitchfork bifurcations at a simple eigenvalueare examined, and two-parameter unfoldings of singularities are rigorously established. Applications in-clude semilinear elliptic equations, imperfect Euler buckling beam problem and perturbed diffusive logisticequation. 2007 Elsevier Inc. All rights reserved.

Keywords: Imperfect bifurcation; Secondary bifurcation; Bifurcation in Banach spaces

1. Introduction

Nonlinear problems can often be formulated to an abstract equation

F(,u) = 0, (1.1)

Partially supported by NSFC Grants 10471032 and 10671049, Longjiang Scholarship from Department of Educationof Heilongjiang Province, China, and United States NSF grants DMS-0314736 and EF-0436318.

* Corresponding author.E-mail address: [email protected] (J. Shi).URL: http://www.math.wm.edu/~shij/ (J. Shi).0022-1236/$ see front matter 2007 Elsevier Inc. All rights reserved.doi:10.1016/j.jfa.2007.06.015

574 P. Liu et al. / Journal of Functional Analysis 251 (2007) 573600

where F : R X Y is a nonlinear differentiable mapping, and X, Y are Banach spaces. Thesolutions of nonlinear equation (1.1) and their dependence on the parameter have been thesubject of extensive studies in the last forty years. Bifurcation could occur at a solution (0, u0)if it is a degenerate solution of (1.1), i.e. the linearized operator Fu(0, u0) is not invertible.Two celebrated theorems of Crandall and Rabinowitz [5,6] are now regarded as foundation ofanalytical bifurcation theory in infinite-dimensional spaces, and both results are based on implicitfunction theorem. In both theorems, it is assumed that

(F1) dimN(Fu(0, u0)) = codimR(Fu(0, u0)) = 1, and N(Fu(0, u0)) = span{w0},

where N(Fu) and R(Fu) are the null space and the range space of linear operator Fu.

Theorem 1.1. (Saddle-node bifurcation, [6, Theorem 3.2].) Let F : R X Y be continuouslydifferentiable. F(0, u0) = 0, F satisfies (F1) and

(F2) F(0, u0) / R(Fu(0, u0)).

Then the solutions of (1.1) near (0, u0) form a continuously differentiable curve ((s), u(s)),(0) = 0, u(0) = u0, (0) = 0 and u(0) = w0. Moreover, if F is k-times continuously differ-entiable, so are (s), u(s).

Theorem 1.2. (Transcritical and pitchfork bifurcations, [5, Theorem 1.7].) Let F : R X Ybe continuously differentiable. Suppose that F(,u0) = 0 for R, the partial derivative Fuexists and is continuous. At (0, u0), F satisfies (F1) and

(F3) Fu(0, u0)[w0] / R(Fu(0, u0)).

Then the solutions of (1.1) near (0, u0) consists precisely of the curves u = u0 and ((s), u(s)),s I = (, ), where ((s), u(s)) are continuously differentiable functions such that (0) = 0,u(0) = u0, u(0) = w0.

Applications of Theorems 1.1 and 1.2 can be found in [3,11,1721,24] and many otherbooks and papers. In [24], the second author studied the perturbations of the bifurcation dia-grams appearing in Theorems 1.1 and 1.2 via an investigation of the system consisting (1.1) andFu(,u)[w] = 0. The goal of this paper is to further explore the set of solutions to (1.1) near abifurcation point (0, u0) satisfying (F1). Here our focus is the situations when the transversalitycondition (F2) is violated, that is, the opposite of (F2):

(F2) F(0, u0) R(Fu(0, u0)).

Our first result (see Theorem 2.1) describes the local solution set of F(,u) = 0 near (0, u0)when (F1) and (F2) are satisfied, which contrasts to Theorem 1.1 when (F1) and (F2) are sat-isfied. We show that the solution set is no longer a curve near the degenerate solution (0, u0)like the case in saddle-node bifurcation, but either an isolated single point, or a pair of transver-sally intersecting curves, which is similar to the solution set in Theorem 1.2. (See the remark

after Theorem 2.1 for a discussion on the connection of Theorems 2.1 and 1.2.) Our proof usesMorse lemma instead of implicit function theorem as in [5], and this idea goes back to Nirenberg

P. Liu et al. / Journal of Functional Analysis 251 (2007) 573600 575

[15,16]. Theorem 2.1 complements Theorem 1.1 better than Theorem 1.2 since (F2) and (F2)are directly opposite, and we also need a more general further transversality condition than (F3),thus Theorems 2.1 and 1.1 can be regarded as the first two parts of classification of degeneratesolutions according to the order of degeneracy.

In the second part of the paper, we continue the work of the second author [24]. We considerthe solution set of

F(,,u) = 0, (1.2)

which can be considered as a perturbation of (1.1), and the new parameter indicates the per-turbation. Following [24,25], we investigate the imperfect bifurcation of bifurcation diagram in(,u) space under small perturbations. Imperfection of the bifurcation diagrams occur when thesmall errors or noises destroy the original bifurcation structure, which occur frequently in en-gineering or other application problems. We prove several new theorems about the symmetrybreaking of transcritical and pitchfork bifurcations, see Theorems 3.13.3 for the structure of thedegenerate solutions, and Theorems 4.1 and 4.3 for the variations of solution set of (1.2). Thestatements of theorems can be found in respective sections, and here we only sketch the imperfectbifurcation diagrams they represent.

In Figs. 1 and 2, typical symmetry breaking perturbations of transcritical and pitchfork bifur-cations are shown. We use analytic bifurcation theory following [5,6] to obtain precise structureof the perturbed local bifurcation diagrams. Another not-so-typical breaking up of the transcrit-ical bifurcation is also shown as an application of the secondary bifurcation theorem which isproved in Section 2, see Figs. 3 and 4. In Section 5, we apply our abstract results to the imperfectbifurcation in classical Euler buckling beam problem and diffusive logistic equation in spatialecology.

Bifurcation problems concerning the solutions of nonlinear equation (1.1) have been studiedin abstract framework since 1970s. One important tool is LyapunovSchmidt reduction which

< 0 = 0 > 0

Fig. 1. Typical symmetry breaking of transcritical bifurcation.

< 0 = 0 > 0

Fig. 2. Typical symmetry breaking of pitchfork bifurcation.

< 0 = 0 > 0Fig. 3. Non-typical symmetry breaking of transcritical bifurcation (1).


< 0 = 0 > 0

Fig. 4. Non-typical symmetry breaking of transcritical bifurcation (2).

reduces the original problem to a finite-dimensional one, and the theory of singularities ofdifferentiable maps and catastrophe theory are useful in the qualitative studies of such finite-dimensional problems. In particular, the imperfect bifurcation of (LyapunovSchmidt) reducedmaps are investigated by Golubitsky and Schaeffer [10]. This approach is efficient but may de-pend on the reductions specific to the application problems. Another approach is to directly dealwith infinite-dimensional problems, and the results involve various partial derivatives of the non-linear maps on Banach spaces but not derivatives of reduced finite-dimensional maps. This couldbe demonstrated by the work of Crandall and Rabinowitz [5,6], which have been widely adoptedin applications. Such results are more convenient and more applicable in real world problems,since in applications one only needs to check certain linearized operators but does not need toperform LyapunovSchmidt reduction. On the other hand, some ideas of catastrophe theory havealso been generalized to infinite-dimensional setting. The fold and cusp type singularities in Ba-nach spaces have been found in various situations, and a general theory has also been developed,see the survey of Church and Timourian [4]. We also note that the two approaches are not in con-flict: for example, when establishing infinite-dimensional bifurcation theorem (see Theorem 2.1),the LyapunovSchmidt reduction is used in proof. But the abstract results in the latter approachare free of specific reductions.

In the paper, we use as the norm of Banach space X, , as the duality pair of a Banachspace X and its dual space X. For a nonlinear operator F , we use Fu as the partial derivative ofF with respect to argument u. For a linear operator L, we use N(L) as the null space of L andR(L) as the range space of L.

2. Crossing curve bifurcation

If we assume F satisfies (F1) at (0, u0), then we have decompositions of X and Y : X =N(Fu(0, u0)) Z and Y = R(Fu(0, u0)) Y1, where Z is a complement of N(Fu(0, u0))in X, and Y1 is a complement of R(Fu(0, u0)). In particular, Fu(0, u0)|Z :Z R(Fu(0, u0))is an isomorphism. Since R(Fu(0, u0)) is codimension one, then there exists l Y such thatR(Fu(0, u0)) = {v Y : l, v = 0}. Thus if F also satisfies (F2), then the equation

F(0, u0)+ Fu(0, u0)[v] = 0 (2.1)has a unique solution v1 Z. Our main result in this section is the following bifurcation theorem:

Theorem 2.1. Let F : R X Y be a C2 mapping. Suppose that F(0, u0) = 0, F satisfies(F1) and (F2). Let X = N(Fu(0, u0)) Z be a fixed splitting of X, let v1 Z be the uniquesolution of (2.1), and let l Y such that R(Fu(0, u0)) = {v Y : l, v = 0}. We assume thatthe matrix (all derivatives are evaluated at (0, u0))( )

H0 = H0(0, u0) l,F + 2Fu[v1] + Fuu[v1, v1] l,Fu[w0] + Fuu[w0, v1]l,Fu[w0] + Fuu[w0, v1] l,Fuu[w0,w0] (2.2)


is non-degenerate, i.e., det(H0) = 0.

(1) If H0 is definite, i.e. det(H0) > 0, then the solution set of F(,u) = 0 near (,u) = (0, u0)is {(0, u0)}.

(2) If H0 is indefinite, i.e. det(H0) < 0, then the solution set of F(,u) = 0 near (,u) =(0, u0) is the union of two intersecting C1 curves, and the two curves are in form of(i(s), ui(s)) = (0 + is + si(s), u0 + isw0 + syi(s)), i = 1,2, where s (, ) forsome > 0, (1, 1) and (2, 2) are non-zero linear independent solutions of the equa-tion

l,F + 2Fu[v1] + Fuu[v1, v1]

2 + 2l,Fu[w0] + Fuu[w0, v1]

+ l,Fuu[w0,w0]2 = 0, (2.3)where i(s), yi(s) are some functions defined on s (, ) which satisfy i(0) = i (0) = 0,yi(s) Z, and yi(0) = yi (0) = 0, i = 1,2.

A particularly useful special case is when F(0, u0) = 0, and immediately we have v1 = 0.

Corollary 2.2. Let F : R X Y be a C2 mapping. Suppose that F(0, u0) = 0, F satisfies(F1) and F(0, u0) = 0. We assume that the matrix

H1 ( l,F(0, u0) l,Fu(0, u0)[w0]

l,Fu(0, u0)[w0] l,Fuu(0, u0)[w0,w0])

(2.4)

is non-degenerate. Then the conclusions of Theorem 2.1 hold, and Eq. (2.3) simplifies tol,F(0, u0)

2 + 2l,Fu(0, u0)[w0] + l,Fuu(0, u0)[w0,w0]2 = 0. (2.5)

Corollary 2.3. Let F : R X Y be a C2 mapping. Suppose that F(0, u0) = 0, F satisfies(F1), (F3), F(0, u0) = 0, F(0, u0) R(Fu(0, u0)). Then the solution set of F(,u) = 0near (,u) = (0, u0) is the union of two intersecting C1 curves i , i = 1,2, 1 is tangent to axis at (0, u0) and it is in form of {(,u()), || < }; and 2 is the form of {(0 + 2s +s2(s), u0 + sw0 + sy2(s))}, where

2 = l,Fuu(0, u0)[w0,w0]2l,Fu(0, u0)[w0] .

Remark 2.4. (1) Theorem 2.1 complements CrandallRabinowitz saddle-node bifurcation theo-rem (Theorem 1.1), where (F2) is imposed. Our result is based on the opposite condition (F2)and a generic second order non-degeneracy condition det(H0) = 0.

(2) CrandallRabinowitz theorem of bifurcation from simple eigenvalue (Theorem 1.2) is aspecial case of Theorem 2.1, Corollaries 2.2 and 2.3. Indeed, in Theorem 1.2, for the curve of

constant solutions, any derivative of F in is zero, but (F3) is assumed in Theorem 1.2 so thematrix H0 is indefinite. One curve in Theorem 2.1 part 2 satisfies 1 = 0 (and we can assume


1 = 1) and only a higher order projection on w0, which corresponds to the constant solutionbranch in Theorem 1.2; the other branch satisfies 2 = 1 and

2 = l,Fuu(0, u0)[w0,w0]2l,Fu(0, u0)[w0] (2.6)

from (2.5), which determines the bifurcation direction: if 2 = 0, then a transcritical bifurcationoccurs, and on either side of = 0, (1.1) has locally exactly two solutions; and if 2 = 0 but ahigher order non-degeneracy condition is satisfied (see [24]), then a pitchfork bifurcation occurs.

(3) Theorem 2.1 is also a more general result on secondary bifurcations. The well-knownversion of secondary bifurcation is [5, p. 323, Theorem 1] (see also [7, p. 407, Theorem 29.3]),where a solution curve 1 is given, and it is shown that another curve 2 exists and intersectswith 1 transversally. In our result, no any solution curve is given, and we obtain the two curvessimultaneously from conditions on F at the bifurcation point. Later we prove that Theorem 2.1implies the CrandallRabinowitz secondary bifurcation theorem (see Theorem 2.7).

(4) Notice that v1 in (2.1) depends on the choice of the complement subspace Z, so with adifferent subspace, v1 can be in form of v1 + kw0 for some k R. However it is easy to checkthat det(H0) is independent of choice of v1 or Z, and the solutions of (2.3) are also independentof choice of v1 or Z (up to a constant scale). Note that the solutions of (2.3) are apparently notunique, but there exists two linear independent real-valued solutions since det(H0) < 0.

To prove Theorem 2.1, we first establish a result in finite-dimensional space, which is of itsown interest.

Lemma 2.5. Suppose that (x0, y0) R2 and U is a neighborhood of (x0, y0). Assume thatf :U R is a Cp function for p 2, f (x0, y0) = 0, f (x0, y0) = 0, and the HessianH = H(x0, y0) is non-degenerate. Then

(1) If H is definite, then (x0, y0) is the unique zero point of f (x, y) = 0 near (x0, y0);(2) If H is indefinite, then there exist two Cp1 curves (xi(t), yi(t)), i = 1,2, t (, ),

such that the solution set of f (x, y) = 0 consists of exactly the two curves near (x0, y0),(xi(0), yi(0)) = (x0, y0). Moreover t can be rescaled and indices can be rearranged so that(x1(0), y1(0)) and (x2(0), y2(0)) are the two linear independent solutions of

fxx(x0, y0)2 + 2fxy(x0, y0) + fyy(x0, y0) 2 = 0. (2.7)

Proof. When H is definite, then (x0, y0) is either a strict local minimum or a strict maximumpoint of f (x, y) from calculus. Thus (x0, y0) is the unique zero of f (x, y) = 0 locally. When His indefinite, consider the differential equation:x = f (x, y)y

, y = f (x, y)x

,(x(0), y(0)

) U. (2.8)


Then (2.8) is a Hamiltonian system with potential function f (x, y), and (x0, y0) is the onlyequilibrium point of (2.8) in U (if necessary we can choose smaller U ). The Jacobian of (2.8) at(x0, y0) is

J =(

fxy(x0, y0) fyy(x0, y0)fxx(x0, y0) fxy(x0, y0)

). (2.9)

Since Trace(J ) = 0 and Det(J ) = Det(H) < 0, then (x0, y0) is a saddle type equilibrium of (2.8)and J has eigenvalues k for some k > 0. From the invariant manifold theory of differentialequations, there exists a unique curve s U (the stable manifold) such that s is invariantfor (2.8) and for (x(0), y(0)) s, (x(t), y(t)) (x0, y0) as t ; and similarly the unstablemanifold is another invariant curve u for (2.8) and for (x(0), y(0)) u, (x(t), y(t)) (x0, y0)as t . Both s and u are Cp1 one-dimensional manifold by the stable and unstablemanifold theorem [22, p. 107]. f (x, y) = 0 for (x, y) s u since f (x, y) is the Hamiltonianfunction of the system and s u {(x0, y0)}. On the other hand, for any (x, y) / s u {(x0, y0)}, f (x, y) = 0. This simply follows from the Morse lemma, the Cp1 curves must beidentical to s u.

Finally we consider the tangential direction of s and u. We denote the two curves by(xi(t), yi(t)), with i = 1,2. Then

f(xi(t), yi(t)

)= 0. (2.10)Differentiating (2.10) in t twice, we obtain (we omit the subscript i for xi(t) and yi(t) in theequation)

fxx(x(t), y(t)

)(x(t)

)2 + 2fxy(x(t), y(t))x(t)y(t)+ fyy(x(t), y(t))(y(t))2+ fx

(x(t), y(t)

)x(t)+ fy

(x(t), y(t)

)y(t) = 0

evaluating at t = 0 and f (x0, y0) = 0, we obtain (2.7). We remark that Lemma 2.5 can also be deduced from a more general Morse lemma, see

Kuiper [13] and Chang [2] (Lemma 4.1 and Theorem 5.1), and a weaker result is proved inNirenberg [16, Theorem 3.2.1] in which the crossing curves are shown to be Cp2; we give analternate proof here using invariant manifold theory, and we also remark that Cp1 is the optimalregularity, see discussions in Shi and Xie [28].

Next we recall the well-known LyapunovSchmidt procedure under the condition (F1). Thefollowing version can be found in [16, pp. 3637 and 40]. We sketch a proof for the completenessof presentation.

Lemma 2.6 (LyapunovSchmidt reduction). Suppose that F : R X Y is a Cp map such thatF(0, u0) = 0, and F satisfies (F1) at (0, u0). Then F(,u) = 0 for (,u) near (0, u0) can bereduced to l,F (,u0 + tw0 + g(, t)) = 0, where t (, ), (0 ,0 + ), where isa small positive constant, l Y such that l, v = 0 if and only if v R(Fu(0, u0)), and g is aCp function into Z such that g(0,0) = 0 and Z is a complement of N(Fu(0, u0)) in X.


Proof. We denote the projection from Y into R(Fu(0, u0)) by Q. Then F(,u) = 0 is equiva-lent to

Q F(,u) = 0, and (I Q) F(,u) = 0. (2.11)

We rewrite the first equation in form

Q F(,u0 + tw0 + g) = 0 (2.12)

where t R and g Z. Since F satisfies (F1) at (0, u0), then g = g(, t) in (2.12) is uniquelysolvable from the implicit function theorem for (, t) near (0,0), and g is Cp . Hence u = u0 +tw0 + g(, t) is a solution to F(,u) = 0 if and only if (I Q) F(,u0 + tw0 + g(, t)) = 0.Since R(Fu(0, u0)) is co-dimensional one, hence it becomes the scalar equation l,F (,u0 +tw0 + g(, t)) = 0. Proof of Theorem 2.1. From the proof of Lemma 2.6, we have

f1(, t) Q F(,u0 + tw0 + g(, t)

)= 0, (2.13)for (, t) near (0,0). Differentiating f1 and evaluating at (, t) = (0,0), we obtain

0 = f1 =(Q (F + Fu[g]),Q Fu[w0 + gt ]). (2.14)

Since Fu[w0] = 0 and gt Z, and Fu(0, u0)|Z is an isomorphism, then gt (0,0) = 0. Similarlyg Z and F R(Fu(0, u0)) from (F2), hence

F(0, u0)+ Fu(0, u0)[g(0,0)

]= 0. (2.15)Hence g(0,0) = v1, where v1 is defined as in (2.1).

To prove the statement in Theorem 2.1, we apply Lemma 2.5 to

f (, t) = l,F (,u0 + tw0 + g(, t)). (2.16)From Lemma 2.6, F(,u) = 0 for (,u) near (0, u0) is equivalent to f (, t) = 0 for (, t) near(0,0). To apply Lemma 2.5, we claim that

f (0,0) = (f, ft ) = 0, and Hess(f ) is non-degenerate. (2.17)

It is easy to see that

f (0,0) =(l,F(0, u0)+ Fu(0, u0)

[g(0,0)

],l,Fu(0, u0)

[w0 + gt (0,0)

]). (2.18)

Thus f (0,0) = 0 from (2.1) and gt (0,0) = 0. For the Hessian matrix, we have

Hess(f ) =(

f ftft ftt

). (2.19)Here


ft (0,0) = ft(0,0) =l,Fu[w0 + gt ] + Fuu[w0 + gt , g] + Fu[gt ]

= l,Fu[w0] + Fuu[w0, v1], (2.20)

since gt = 0. Next we have

f(0,0) =l,F + 2Fu[g] + Fuu[g, g] + Fu[g]

= l,F + 2Fu[v1] + Fuu[v1, v1]. (2.21)

Finally,

ftt (0,0) =l,Fuu[w0 + gt ,w0 + gt ] + Fu[gtt ]

= l,Fuu[w0,w0]. (2.22)In summary, from our calculation,

Hess(f ) =( l,F + 2Fu[v1] + Fuu[v1, v1] l,Fu[w0] + Fuu[w0, v1]

l,Fu[w0] + Fuu[w0, v1] l,Fuu[w0,w0]). (2.23)

Therefore from Lemma 2.5, we conclude that the solution set of F(,u) = 0 near (,u) =(0, u0) is a pair of intersecting curves if the matrix in (2.23) is indefinite, or is a single point ifit is definite.

Now we consider only the former case of two curves. We denote the two curves by(i(s), ui(s)) = (i(s), u0 + ti (s)w0 + g(i(s), ti (s))), with i = 1,2. Then

F(i(s), u0 + ti (s)w0 + g

(i(s), ti (s)

))= 0. (2.24)From Lemma 2.5 the vectors vi = (i (0), t i (0)) are the solutions of vT Hv = 0, which are thesolutions (,) of (2.3).

Next we give an secondary bifurcation theorem generalizing a well-known one [5, Theorem 1]based on Theorem 2.1. This is not surprising considering that Theorem 2.1 is more general thanTheorem 1.2, which implies the secondary bifurcation theorem (see [5]). But here we do notassume the existence of a given solution branch as that in [5].

Theorem 2.7. Let W and Y be Banach spaces, an open subset of W and G : Y be twicedifferentiable. Suppose

(1) G(w0) = 0,(2) dimN(G(w0)) = 2, codimR(G(w0)) = 1.

Then

(1) If for any (= 0) N(G(w0)), G(w0)[,] / R(G(w0)), then the set of solutions toG(w) = 0 near w = w0 is the singleton {w0}.

(2) If there exists 1(= 0) N(G(w0)) such that G(w0)[1, 1] R(G(w0)), and there exists 2 N(G (w0)) such that G (w0)[1, 2] / R(G (w0)), then w0 is a bifurcation point ofG(w) = 0 and in some neighborhood of w0, the totality of solutions of G(w) = 0 form


two continuous curves intersecting only at w0. Moreover the solution curves are in form ofw0 + si + si(s), s (, ), i(0) = i (0) = 0, where i (i = 1,2) are the two linearindependent solutions of the equation l,G(w0)[,] = 0.

Proof. Let l Y such that l, y = 0 if and only if y R(G(w0)). Then if for any (= 0) N(G(w0)), G(w0)[,] / R(G(w0)), we must have l,G(w0)[,] > 0 (or < 0) for any(= 0) N(G(w0)). Without loss of generality, we assume > holds. We assume that W =span{1} X is a splitting of W , and we choose 2 X N(G(w0)) so that {1, 2} is a basisof N(G(w0)). Clearly X N(G(w0)) = span{2}. Define F : I X Y (I R is an openinterval containing 0)

F(,u) = G(w0 + 1 + u). (2.25)

Then F C2 and F(0,0) = 0. We check F satisfies (F1) and (F2). It is easy to calculate

F(0,0) = G(w0)[1], F(0,0) = G(w0)[1, 1],Fu(0,0)[] = G(w0)[], Fu(0,0)[] = G(w0)[1,],

Fuu(0,0)[, ] = G(w0)[, ]. (2.26)

Then N(Fu(0,0)) = span{2} and R(Fu(0,0)) = R(G(w0)), hence (F1) is satisfied. (F2) isobvious since F(0,0) = G(w0)[1] = 0. From above calculation and Corollary 2.2, we have

H1 =( l,G(w0)[1, 1] l,G(w0)[1, 2]

l,G(w0)[1, 2] l,G(w0)[2, 2]). (2.27)

Since l,G(w0)[,] > 0 for any (= 0) N(G(w0)), then det(H1) > 0 sincel,G(w0)[k11 + k22, k11 + k22]

= kH1kT > 0,which implies that H1 is positively definite, and here k = (k1, k2) R2 and kT is the trans-pose of k. We apply Theorem 2.1, part 1, then the result follows. For the second part,the calculation above remains true with 1, 2 satisfying the conditions in theorem if wechoose a complement subspace X to span{1} so that 2 X, and {1, 2} makes a basisfor N(G(w0)). But det(H1) < 0 from the assumptions, hence we can apply part 2 of The-orem 2.1 or Corollary 2.2. For the solutions (i, i) of (2.5), (1, 1) = (1,0) is one so-lution since l,G(w0)[1, 1] = 0, and the other solution is given by 2 = 1 and 2 =l,G(w0)[2, 2]/(2l,G(w0)[1, 2]). Hence the two solution branches are in form ofw0 + s1 + s1(s) and w0 + s(21 + 2) + s2(s), and one can verify that 1 = 1 and2 = 21 + 2 are the two linear independent solutions of l,G(w0)[,] = 0.

The simplest example of Theorem 2.7 is the quadratic map: f : R2 R defined by f (x, y) =x2 y2. In the + sign case, (x, y) = (0,0) is the only solution of f (x, y) = 0, and in the sign case, the solutions of f (x, y) = 0 are all points on the crossing lines x = y. To conclude

2this section, we illustrate Theorem 2.1 with an example in X = Y = R . Another example ininfinite-dimensional spaces is shown in Section 5.1.


Example. Define

F(,x1, x2) =(

5x1 + x31 + 2x1x22x2 + x32 + 2x2x21

)

(x1x2

), (2.28)

where (x1, x2) R2 and R. The equation F(,x1, x2) = 0 has a branch of zero solutions0 = {(,0,0): R}, and it also has two branch of semi-trivial solutions (with one componentzero):

1 ={(, x1,0): x1 =

5, > 5},

2 ={(,0, x2): x2 =

1, > 1}. (2.29)

We notice that both 1 and 2 are generated through pitchfork bifurcations from 0 at = 5and = 1, respectively. To further analyze the secondary bifurcations, we need the followingcalculations:

F = (

x1x2

), F = 0,

Fx =(

5 + 3x21 + 2x22 4x1x24x1x2 1 + 3x22 + 2x21

)

(1 00 1

), Fx =

(1 00 1

),

Fxx =((

6x1 4x24x2 4x1

),

(4x2 4x14x1 6x2

)). (2.30)

Along 1, Fx = diag(2 10, 9). Thus besides = 5 (primary bifurcation point), = 9is another degenerate point. We analyze the bifurcation at (, x1, x2) = (9,2,0). N(Fx) ={(0, a)T : a R}, R(Fx) = {(a,0)T : a R}, F = (2,0)T R(Fx). Hence (F1) and (F2)are satisfied. We choose Z = R(Fx), then the equation F + Fx[v1] = 0 has a unique solutionv1 = (1/4,0)T Z. From above calculation, we find that the matrix H0 in (2.2) to be

H0 =(

0 11 0

), (2.31)

which is indefinite. Thus we can apply Theorem 2.1 to this equation, and near (, x1, x2) =(9,2,0), the solution set is the union of two intersecting curves. Moreover we can choose(1, 1) = (1,0) (corresponding to 1) and (2, 2) = (0,1) (corresponding to new branch)for the directions of two curves. 2 = 0 indicates that the bifurcation is not a linear transcriticalone. Indeed it is a pitchfork bifurcation, and the new branch is

3 ={(, x1, x2): > 9, x1 =

(+ 3)/3, x2 =

( 9)/3}. (2.32)

3. Perturbed problems

We consider an equationF(,,u) = 0, (3.1)


where F C1(M,Y ), M R R X, and X, Y are Banach spaces. We define

H(,,u,w) =(

F(,,u)

Fu(,,u)[w]). (3.2)

We consider the solution (0, 0, u0,w0) of H(,,u,w) = 0. For (0, 0, u0) M and w0 X1 {x X: x = 1}, By HahnBanach theorem (see [24, Lemma 7.1]), there exists a closedsubspace X3 of X with codimension 1 such that X = L(w0) X3, where L(w0) = span{w0},and d(w0,X3) = inf{w x: x X3} > 0. Let X2 = w0 + X3 = {w0 + x: x X3}. Then X2is a closed hyperplane of X with codimension 1. Since X3 is a closed subspace of X, and X3 isalso a Banach space in the subspace topology. Hence we can regard M1 = M X2 as a Banachspace with product topology. Moreover, the tangent space of M1 is homeomorphic to M X3(see [24] for more on the setting).

In the following we will still use the conditions (Fi) on F defined in previous sections andin [24], but we will use (0, 0, u0) instead of (0, u0) in all these conditions. In addition to(F1)(F3) defined above, we also define (following [24])

(F4) Fuu(0, 0, u0)[w0,w0] / R(Fu(0, 0, u0));(F5) F(0, 0, u0) / R(Fu(0, 0, u0)).

We also use the convention that (Fi) means that the condition defined in (Fi) does not hold.We first prove a refinement of Theorem 2.6 and a generalization of Theorem 2.4 in [24]:

Theorem 3.1. Let F C2(M,Y ), T0 = (0, 0, u0,w0) M1 such that H(T0) = (0,0). Supposethat the operator F satisfies (F1), (F2), (F3), (F4) and (F5) at T0. Then there exists > 0 suchthat all the solutions of H(,,u,w) = (0,0) near T0 form a C2-curve.{

Ts =((s), (s), u(s),w(s)

), s I = (, )}, (3.3)

where (s) = 0 + (s), s I , () C2(I,R), (0) = (0) = 0, and

(s) = 0 + s + z1(s),u(s) = u0 + s(kw0 + v1)+ z2(s),w(s) = w0 + s1 + z3(s),

where s I , zi() C2, zi(0) = zi (0) = 0 (i = 1,2,3), and v1 X3 is the unique solution of

Fu(0, 0, u0)[v1] + F(0, 0, u0) = 0, (3.4)

k is the unique number such thatl,Fu(0, 0, u0)[w0]

+ l,Fuu(0, 0, u0)[v1,w0]+ kl,Fuu(0, 0, u0)[w0,w0]= 0, (3.5)

1 X3 is the unique solution of


Fu(0, 0, u0)[w0] + kFuu(0, 0, u0)[w0,w0]+ Fuu(0, 0, u0)[v1,w0] + Fu(0, 0, u0)[] = 0. (3.6)

Proof. We apply Theorem 1.1 to operator H so we need to verify all the conditions. We definea differential operator K : R X X3 Y Y ,

K[, v,] = H(,u,w)(0, 0, u0,w0)[, v,]

=(

F(0, 0, u0)+ Fu(0, 0, u0)[v]Fu(0, 0, u0)[w0] + Fuu(0, 0, u0)[v,w0] + Fu(0, 0, u0)[]

). (3.7)

(1) dimN(K) = 1. Suppose that (, v,) N(K) and (, v,) = 0. If = 0, fromFu(0, 0, u0)[v] = 0 and (F1), we have v = kw0, and

kFuu(0, 0, u0)[w0,w0] + Fu(0, 0, u0)[] = 0. (3.8)

But (F4) is satisfied, thus k = 0, and = v = 0. This is a contradiction. Next we consider = 0.Without loss of generality, we assume that = 1. Notice that F(0, 0, u0) R(Fu(0, 0, u0))from (F2), then v must be in form of kw0 + v1, where v1 is defined in (3.4). Substituting v =kw0 + v1 and applying l to it, we obtain (3.5). From (F4), k can be uniquely determined by (3.5)and 1 is also uniquely determined. Thus N(K) = {(1, kw0 + v1,1)}.

(2) codimR(K) = 1. Let (h, g) R(K), and let (, v,) R X X3 satisfy

F(0, 0, u0)+ Fu(0, 0, u0)[v] = h, (3.9)Fu(0, 0, u0)[w0] + Fuu(0, 0, u0)[v,w0] + Fu(0, 0, u0)[] = g. (3.10)

Applying l to (3.9), we get l, h = 0, hence h R(Fu(0, 0, u0)), and R(K) R(Fu)Y . Con-versely, for any (h, g) R(Fu)Y , there exists a unique v2 X3 such that Fu(0, 0, u0)[v2] = h.For any , k R, let v,k = v2 + v1 + kw0, where v1 satisfies (3.4). Then v,k solves (3.9). Sub-stituting v = v,k into (3.10), and applying l, we obtain

l,Fu[w0]

+ l,Fuu[w0, v2]+ l,Fuu[w0, v1]+ kl,Fuu[w0,w0]= l, g. (3.11)Then for fixed R, from (F4) there exists a unique k( ) so that (3.11) holds. With such choiceof k, (3.11) holds, then in (3.10) is uniquely solvable in X3. Therefore (, v,k(),) is apre-image of (h, g). This implies that R(K) = R(Fu(0, 0, u0)) Y , and codimR(K) = 1.

(3) H(0,0,u0,w0) = 0.

(4) H(0,0,u0,w0) / R(K). Since R(K) = R(Fu(0, 0, u0)) Y , we only need to showthat F(0, 0, u0,w0) / R(Fu(0, 0, u0)), but that is exactly assumed in (F5). So the statementof the theorem follows from Theorem 1.1.

In Theorem 3.1, we have (0) = 0, (0) = 0, (0) = 0 and (0) = 1. To completely de-termine the turning direction of curve of degenerate solutions, we calculate (0). Let {Ts =


((s), (s), u(s),w(s)): s (, )} be a curve of degenerate solutions which we obtain in The-orem 3.1. Differentiating H((s), (s), u(s),w(s)) = 0 with respect to s, we obtain

F(s)+ F(s)+ Fu

[u(s)

]= 0, (3.12)Fu

[w(s)

](s)+ Fu

[w(s)

](s)+ Fuu

[w(s),u(s)

]+ Fu[w(s)]= 0. (3.13)Setting s = 0 in (3.12) and (3.13), we get exactly (3.4) and (3.6). We differentiate (3.12) and(3.13) again, and we have

F[(s)

]2 + F(s)+ F[(s)]2 + F(s)+ Fuu[u(s), u(s)]+ Fu

[u(s)

]+ 2F(s)(s)+ 2Fu[u(s)](s)+ 2Fu[u(s)](s) = 0, (3.14)Fu

[w(s)

][(s)

]2 + Fu[w(s)](s)+ Fu[w(s)](s)+ Fu

[w(s)

][(s)

]2 + Fuuu[u(s), u(s),w(s)]+ Fuu[w(s),u(s)]+ Fu

[w(s)

]+ 2Fu[w(s)](s)(s)+ 2Fuu[u(s),w(s)](s)+ 2Fuu

[u(s),w(s)

](s)+ 2Fu

[w(s)

](s)+ 2Fu

[w(s)

](s)

+ 2Fuu[w(s), u(s)

]= 0. (3.15)Setting s = 0 in (3.14) and applying l to it, we obtain

l,F(0)+ l,F +l,Fuu[kw0 + v1, kw0 + v1]

+ 2l,Fu[kw0 + v1]= 0, (3.16)where k satisfies (3.5). By (3.5), we have

(0) = l,F + l,Fuu[v1, v1] + 2l,Fu[v1] k2l,Fuu[w0,w0]

l,F . (3.17)

Our next result is under the assumption (F4) instead of (F4). In this case an additionaltransversality condition (3.18) is needed to establish the saddle-node bifurcation of the degener-ate solutions.

Theorem 3.2. Let F C2(M,Y ), T0 = (0, 0, u0,w0) M1 such that H(T0) = (0,0). Supposethat the operator F satisfies (F1), (F2), (F3), (F4) and (F5) at T0. We also assume that

Fu(0, 0, u0)[w0] + Fuu(0, 0, u0)[v1,w0] / R(Fu(0, 0, u0)

), (3.18)

where v1 X3 is defined in (3.4). Then there exists > 0 such that all the solutions ofH(,,u,w) = (0,0) near T0 form a C2-curve:{

Ts =((s), (s), u(s),w(s)

), s I = (, )}, (3.19)where (s) = 0 + (s), s I , () C2(I,R), (0) = (0) = 0, and


(s) = 0 + z1(s),u(s) = u0 + sw0 + z2(s),w(s) = w0 + s2 + z3(s),

where s I , zi() C2, zi(0) = zi (0) = 0 (i = 1,2,3), 2 X3 is the unique solution ofFu(0, 0, u0)[2] + Fuu(0, 0, u0)[w0,w0] = 0. (3.20)

Proof. Recall K defined in (3.7).

(1) dimN(K) = 1. Let (, v,) N(K) and (, v,) = 0. If = 0, from Fu(0, 0, u0)[v] = 0and (F1), we have v = kw0, and we also get (3.8). From (F4), we may define 2 as in (3.20).Thus (0,w0,2) N(K). If = 0, we may assume that = 1. From (F2), The solution v1 of(3.4) exists. More generally, the solution of (3.4) is in form of v1 + kw0. Substitute v1 + kw0into (3.10) with g = 0, the solvability of is equivalent to

Fu(0, 0, u0)[w0] + Fuu(0, 0, u0)[v1,w0] R(Fu(0, 0, u0)

), (3.21)

from (F4). But we assume (3.18), hence is not solvable, and there is no (, v,) N(K)such that = 0. Hence N(K) = span{(0,w0,2)}.

(2) codimR(K) = 1. Similar to the proof of Theorem 3.1, we can show that R(K) =R(Fu(0, 0, u0)) Y. The only difference is that now in (3.11), l,Fuu[w0,w0] = 0. But with(3.18), is uniquely determined since l,Fu[w0] + l,Fuu[w0, v1] = 0. With (3.11) satis-fied, then in (3.10) is also uniquely determined. Thus R(K) = R(Fu(0, 0, u0)) Y , andcodimR(K) = 1.

(3) H(0,0,u0,w0) / R(H(,u,w)(0,0,u0,w0)). From (F5), F(0, 0, u0) /R(Fu(0, 0, u0)), hence

H(0, 0, u0,w0) =(

F(0, 0, u0)Fu(0, 0, u0)[w0]

)/ R(Fu(0, 0, u0)) Y.

(4) H(0,0,u0,w0) = 0. So the statement of the theorem follows from Theorem 1.1. In Theorem 3.2, we have (0) = (0) = 0, and we can determine (0) and (0) from

(3.14) and (3.15) by setting s = 0:

F(0)+ F(0)+ Fuu[w0,w0] + Fu

[u(0)

]= 0, (3.22)Fu[w0](0)+ Fu[w0](0)+ Fuuu[w0,w0,w0] + Fuu

[w0, u

(0)]

+ Fu[w(0)

]+ 2Fuu[2,w0] = 0. (3.23)Hence we have(0) = l,Fuu[w0,w0]l,F = 0, (3.24)


and

(0) = l,Fuuu[w0,w0,w0] + l,Fuu[w0, u(0)] + 2l,Fuu[2,w0]

l,Fu[w0] . (3.25)

Since (0) = 0, (3.22) implies that u(0) = (0)v1 +2 + k1w0 for some k1 R. Substitutingthis expression of u(0) into (3.23) and (3.25), we obtain

(0) = l,Fuuu[w0,w0,w0] + 3l,Fuu[2,w0]l,Fu[w0] + l,Fuu[w0, v1] . (3.26)

We differentiate (3.14) again,

F[(s)

]3 + F(s)+ F[(s)]3 + F(s)+ Fuuu[u(s), u(s), u(s)]+ Fu

[u(s)

]+ 3F(s)(s)+ 3F(s)(s)+ 3Fuu[u(s), u(s)]+ 3F(s)(s)+ 3F(s)(s)+ 3Fu

[u(s)

](s)+ 3Fu

[u(s)

](s)

+ 3Fu[u(s)

](s)+ 3Fu

[u(s)

](s)+ 3F(s)

((s)

)2+ 3F

((s)

)2(s)+ 3Fu

[u(s)

]((s)

)2 + Fuu[u(s), u(s)](s)+ 3Fu

[u(s)

]((s)

)2 + 3Fuu[u(s), u(s)](s)+ 6Fu[u(s)](s)(s) = 0. (3.27)Setting s = 0 in (3.27) and applying l to it, we obtain

l,F(0)+l,Fuuu[w0,w0,w0]

+ 3l,Fuu[u(0),w0]+ 3l,Fu[w0](0) = 0, (3.28)by (3.25), we have

(0) = 2 l,Fuuu[w0,w0,w0] + 3l,Fuu[2,w0]l,F . (3.29)

We recall from [24] that the bifurcation at (0, 0, u0) is a pitchfork bifurcation if

(F6) Fuuu(0, 0, u0)[w0,w0,w0] + 3Fuu(0, 0, u0)[2,w0] / R(Fu(0, 0, u0)), where 2 X3 is the unique solution of (3.20).

Thus we can conclude that if (F6) is satisfied, then (0) = 0 and (0) = 0 in Theorem 3.2.With (F2) and (F5) both satisfied, a crossing curve structure for the degenerate solutions near

a transcritical bifurcation point is possible, as we show in the next theorem.

Theorem 3.3. Let F C2(M,Y ), T0 = (0, 0, u0,w0) M1 such that H(T0) = (0,0). Supposethat the operator F satisfies (F1), (F2), (F3), (F4) and (F5) at T0. We also assume that (3.18)holds, k and v1 are defined as in (3.5) and (3.4) respectively, v2 X3 is the unique solution ofF(0, 0, u0)+ Fu(0, 0, u0)[v2] = 0, (3.30)


and the matrix

H1 ( l,D11 l,D12

l,D21 l,D22)

(3.31)

satisfies det(H1) < 0, where Dij is given by

D11 = F + 21F + 21Fu[v1] + 2Fu[v2] + 21 F + 2 21 Fu[v1]+ 21Fu[v2] + 21 Fuu[v1, v1] + 21Fuu[v1, v2] + Fuu[v2, v2],

D12 = D21 = F + kFu[w0] + Fu[v1] + 1F + 1kFu[w0] + 21Fu[v1]+ Fu[v2] + 1kFuu[w0, v1] + 1Fuu[v1, v1] + kFuu[w0, v2] + Fuu[v1, v2],

D22 = F + 2kFu[w0] + 2Fu[v1] + k2Fuu[w0,w0] + 2kFuu[w0, v1] + Fuu[v1, v1],

here 1 is determined byl,Fu[w0]

+ l,Fu[w0]+ l,Fuu[v1,w0]+ l,Fuu[v2,w0]= 0. (3.32)Then the solutions set of H(,,u,w) = (0,0) near T0 is the union of two intersecting curves,and the two curves are in form:{

Ti,s =(i(s), i(s), ui(s),wi(s)

): s (, )},

where i(s) = 0 + is + szi0(s), i(s) = 0 + is + szi1(s), ui(s) = u0 + is(kw0 + v1) +szi2(s), wi(s) = w0 + si1 + szi3(s), (1, 1) and (2, 2) are non-zero linear independentsolutions of the equation

l,D112 + 2l,D12+ l,D222 = 0, (3.33)

zij (0) = zij (0) = 0, zij (s) Z, i = 1,2; j = 0,1,2,3. If det(H1) > 0, then the solution set ofH(,,u,w) = (0,0) near T0 is the singleton {T0}.

Proof. We apply Theorem 2.1. Recall K defined in (3.7). Then the proof of Theorem 3.1shows that dimN(K) = 1 and codimR(K) = 1 since the same assumptions except (F5) holdhere but (F5) is not used in this part of proof. In particular we have N(K) = span{W0 (1, kw0 + v1,1)}, where 1 is defined by (3.6). (F5) implies that H(0, 0, u0,w0) R(H(,u,w)(0, 0, u0,w0)), thus (F2) is satisfied for H . Equation (2.1) now becomes H +H(,u,w)[, v,] = 0, which is

F + F + Fu[v] = 0, (3.34)Fu[w0] + Fu[w0] + Fuu[v,w0] + Fu[] = 0. (3.35)

We look for a solution (, v,) Z1 R X3 X3. Let v1 be defined as in (3.4), and let v2

be defined as in (3.30). Then the solution of (3.34) is given by v = v1 + v2 X3, and with thisform of v, we apply l to (3.35) and we obtain (3.32). Thus = 1 can be uniquely determined


from (3.18), and subsequently v, can be uniquely determined. We denote this solution to beV1 = (1, 1v1 + v2,3).

Next we calculate the Hessian matrix in (2.2). Since R(H(,u,w)(0, 0, u0,w0)) =R(Fu(0, 0, u0)) Y , we define l1 (Y Y) by

l1, (y1, y2)= l, y1, (3.36)

where l Y so that l, y = 0 if and only if y R(Fu(0, 0, u0)). For the simplicity of nota-tions, we also denote U = (,u,w). Then the matrix H0 in Theorem 2.1 is in form of

H1 ( l1,H + 2HU [V1] +HUU [V1,V1] l1,HU [W0] +HUU [V1,W0]

l1,HU [W0] +HUU [V1,W0] l1,HUU [W0,W0])

=(

l,F + 2FU [V1] + FUU [V1, V1] l,FU [W0] + FUU [V1, W0]l,FU [W0] + FUU [V1, W0] l,FUU [W0, W0]

), (3.37)

where V1 = (1, 1v1 + v2) and W0 = (1, kw0 + v1) are the projection of V1 and W0 from R X X3 to R X, and U = (,u). From direct calculation, we can show that H1 takes the formin (3.31). Now we can apply Theorem 2.1 to obtain the conclusions of the theorem.

An important special case of Theorem 3.3 is when we have a constant solution for all near 0, i.e. we assume that F(0, ,u0) 0 for near 0.

Corollary 3.4. Let F C2(M,Y ), T0 = (0, 0, u0,w0) M1 such that H(T0) = (0,0). Weassume that there exists a neighborhood U of (0, 0) in R2 such that F(0, ,u0) 0 for(0, ) U and F(0, 0, u0) = 0. Suppose that at T0 the operator F satisfies (F1), (F3), (F4)and

(F7) (vH2vT )/l,Fuu[w0,w0] > 0, where v = (l,Fu[w0], l,F), and

H2 ( l,F l,Fu[w0]

l,Fu[w0] l,Fuu[w0,w0]), (3.38)

then the conclusions of Theorem 3.3 hold and (1, 1) and (2, 2) are non-zero linear indepen-dent solutions of the equation(l,Fl,Fu[w0] 2l,Fl,Fu[w0])l,Fuu[w0,w0]2

+ 2l,Fl,Fu[w0]

l,Fuu[w0,w0]

(l,Fu[w0])32 = 0. (3.39)

Proof. We apply Theorem 3.3. (F2) and (F5) are satisfied and v1 = v2 = 0 sinceF(0, 0, u0) = 0, and F(0, ,u0) 0 for (0, ) U . (3.18) is also satisfied since v1 = 0and (F3). Equations (3.5) and (3.32) now become

l,Fu[w0]+ kl,Fuu[w0,w0]= 0, and l,Fu[w0]+ 1l,Fu[w0]= 0, (3.40)respectively. For the entries of matrix H1, we have l,D11 = l,F + 21F, l,D12 =l,F, and l,D22 = kl,Fu[w0] from (3.32) and Theorem 3.3. With 1 and k determined


by (3.40), then det(H1) = vH2vT /l,Fuu[w0,w0], where v and H2 are defined in (F7). Hence(F7) implies the indefiniteness of H1. So the statement of the theorem follows from Theorem 3.3and (3.39) is from calculations. Remark 3.5.

(1) We notice that (F7) is satisfied if the matrix H2 is negatively definite and l,Fuu[w0,w0] 0. (3.41)

Then the solution set of H(,,u,w) = 0 near T0 is the union of two curves of form:{Ti, =

(i(s), i(s), ui(s),wi(s)

): s (, )}, (3.42)

where i(s) = 0 + is + szi0(s), i(s) = 0 + is + szi1(s), ui(s) = u0 + iskw0 + szi2(s),wi(s) = w0 + si1 + szi3(s), (1, 1) and (2, 2) are non-zero linear independent solutionsof the equation

l,F l,Fuu[w0,w0]

2 + 2l,F

l,Fuu[w0,w0]

(l,Fu[w0])22 = 0. (3.43)

Proof. We apply Corollary 3.4. (F7) is satisfied since 1 = 0 and the expression in (F7) is nowsimplified as

l,F (l,Fu[w0])2

l,Fuu[w0,w0] (l,F)2 < 0,

because of (3.41). Therefore we can apply Corollary 3.4 to obtain the results here. Notice thatsince F = 0 and Fu[w0] = 0, then H = 0 thus we can also apply Corollary 2.2. 4. Perturbation of bifurcation diagrams

In this section we shall apply the results in Section 3 to the original equation F(,,u) = 0,and observe the variations of bifurcation diagrams in (,u) space when the parameter is per-turbed from = 0. First Theorem 3.1 is only a generalization of [24, Theorem 2.4], and the

variations of the bifurcation diagrams have been studied in [24, Theorem 2.5]. Hence we will notdiscuss that situation here. The following result illustrates the application of Theorem 3.2.


Theorem 4.1. Assume the conditions in Theorem 3.2 are satisfied, and {Ts} is defined as inTheorem 3.2. In addition we assume that F C3(M,Y ), F(0, ,u0) = 0 for | 0| < for some > 0, and F satisfies (F6). Then (0) = 0 and (0) = 0. If l,F < 0,l,Fuuu[w0,w0,w0] + 3l,Fuu[w0,2] < 0, where 2 satisfies (3.20), and l,Fu[w0] > 0,then (0) > 0 and (0) > 0, there exist 1, 1, 2 > 0 such that for N = {(,u) R X:| 0| 1, u 2}, we have

(A) for = 0,F1(0)N = {(,0): | 0| 1}0, 0 = {((t), u(t)), |t | }, (4.1)

where (0) = 0, (0) = 0, (0) > 0, and () = 0 + 1;(B) for fixed (0 1, 0) (0, 0 + 1),

F1(0)N = + , (4.2)where + = {(+(t), u+(t)), t [,+]}, where +() = 0 + 1, +(0) = 0,+(0) > 0, and (+(0), u+(0)) is the unique degenerate solution on + ; and ={(, u()): [0 1, 0 + 1]} is a monotone curve without degenerate solutions(see Fig. 2).

Proof. From Theorem 3.2 and (3.29), we have (0) = 0, (0) = 0, (0) = 0, (0) = 0. Fromthe assumptions, (3.26) (with v1 = 0) and (3.29), (0) > 0 and (0) > 0. When = 0, thereis a branch of trivial solutions, and a branch of non-trivial ones (which we denote by 0 =((t), u(t))) from Theorem 1.2. The conditions (F3) and (F4) imply that (0) = 0 and from[24, (4.6)]

(0) = l,Fuuu[w0,w0,w0] + 3l,Fuu[w0,2]3l,Fu[w0] > 0. (4.3)

Thus when = 0, the bifurcation near (0, u0) is a supercritical pitchfork one, and there isonly one degenerate solution on 0 near (0, u0). Define N = {(,u) R X: | 0| 1,u 2}, such that (4.1) holds and () = 0 + 1, for t [,+], u(t) 2/2.

We denote by the solution set of (3.1) in N for fixed . Since (0) > 0 and (0) > 0,then there exists 2 > 0 such that for (0, 0 + 2), (3.1) has a unique degenerate so-lution (+, u+), where + = (s+) > 0, s+ > 0 for s+ small. Moreover, u+ is a degener-ate solution which satisfy the condition of Theorem 1.1. In fact, we only need to check thatF(,+, u+) / R(Fu(, +, u+)). Define A(s) = l(s),F((s), (s), u(s)), where l(s) Y satisfying N(l(s)) = R(Fu((s), (s), u(s)). Then A(0) = l,Fu(0, 0, u0)[w0] > 0 andA(0) = 0, so A(s+) > 0. Define B(s) = l(s),Fuu((s), (s), u(s))[w(s),w(s)], we have

B (0) = l(0),Fuu[w0,w0]+ l,Fuuu[w0,w0,w0] + 2Fuu[2,w0]= l,Fuuu[w0,w0,w0] + 3Fuu[2,w0]< 0, (4.4)

since (0) = (0) = 0, u(0) = w0, w(0) = 2, (0) = 0. In (4.4), we obtain l(0),

Fuu[w0,w0] = l,Fuu[2,w0] by differentiating l(s),Fu((s), (s), u(s)) [w0] = 0 twiceand using (3.25). In particular B(s+) < 0. From Theorem 1.1, near (+, u+), the solutions form


a curve + = {(+(t), u+(t)), t [,+]} with +(0) = +, u+(0) = u+, +(0) > 0 and+() = 0 + 1. Here we obtain +(0) > 0 by using [24, (4.1)]

+(0) = l+,Fuu(+, u+)[w+,w+]

l+,F(+, u+)= B(s+)

A(s+). (4.5)

From the definition of N , there exists 3 > 0 such that for [0, 0 + 3], (3.1) has nosolution (,u) with u = 2 with | 0| 1. And there exists 4 > 0 such that for (0, 0 + 4], there is no degenerate solution (,u) of (3.1) with u 2 and | 0| 1/2.For = 0, (3.1) have two nontrivial solutions (0 + 1, u) N such that u 2/2. Sinceu is non-degenerate, for fixed = , by implicit function theorem, there exists 5 > 0 suchthat for (0 5, 0 +5), (3.1) has a unique solution u() near u, a unique solution u0()near the trivial solution u0 with = 0 +1, and a unique solution u0() near the trivial solutionu0 with = 0 1. Therefore from the implicit function theorem and the fact that (+, u+) isthe only degenerate solution in N , the solution set of (3.1) in N consists of + defined above,and another monotone curve which we denote by .

We claim that consists of only for (0, 0 + 1) with 1 = min{2, 3, 4, 5}.Suppose that for some (0, 0 + 1), there is another solution (,u) which is not on .Then (,u) is a non-degenerate solution since the only degenerate solution is (+, u+). So by theimplicit function theorem, (,u) is on a solution curve 1 = ((s), u(s)) which can be extendedto N . But there is no solution on u = 2, so there is a solution (0 + 1, u) on 1. Onthe other hand, has another three solutions with = 0 + 1, thus there are at least foursolutions with = 0 + 1, which contradicts with the definition of 1 and the arguments in thelast paragraph. The proof for (0 1, 0) is similar. Remark 4.2.

(1) The assumptions on the signs of l,Fuuu[w0,w0,w0] + 3l,Fuu[w0,2], l,F andl,Fu[w0] are only for the purpose of fixing an orientation, similar results for other casescan also be established.

(2) From the proof, the solutions +() are two of the three solutions u and u0, but it is notimmediately clear which two. However usually with more information, we can show thatu0 must be one of +(), i.e. the perturbation u0 of the trivial solution u0 must be on thecurve with turning point, which is consistent with observations in experiments.

Theorem 4.3. Assume the conditions in Corollary 3.6 are satisfied, and in addition we assumeF(0, 0, u0) = 0. For the purpose of fixing an orientation, we assume that

l,Fu(0, 0, u0)[w0]> 0, and

l,Fuu(0, 0, u0)[w0,w0]

< 0. (4.6)

Then there exist , 1, 2 > 0 such that for N = {(,u) R X: | 0| 1, u 2},(A) for = 0,

F1(0)N = {(,u0): | 0| 1}0,

0 =

{((t), u(t)

): t [,]}, and (t) > 0 for t [,]; (4.7)


(B) for (0 , 0) (0, + ),

F1(0)N = + , ={(

(t), u(t)): t [,]}

where (0) = 0, +(0) > 0, (0) < 0 and there is exactly one turning point on eachcomponent ; moreover +(0) > 0 > (0).

Proof. From (4.6) and (3.41), we also have l,F(0, 0, u0) < 0. When = 0, by Theo-rem 1.2, the solution set of F(0, ,u) = 0 near (0, u0) consists precisely the two curves ofform in (4.7), where

(0) = l,Fuu(0, 0, u0)[w0,w0]2l,Fu(0, 0, u0)[w0] > 0.

From the conditions in Corollary 3.6, i = 0 and i = 0 for i = 1,2, thus there are exactlytwo degenerate solutions for (0, 0 + ) as well as for (0 , 0). To be more precise,we assume i = 1 for i = 1,2. When (0, 0 + ), we denote the two degenerate solutionsby (+, u+) and (, u), where + = (s+) = 0 + 1s+ + o(|s+|) and = (s) = 0 +2s + o(|s|), s+ > 0 and s > 0 since = 0 + s+ + o(|s+|) and = 0 + s + o(|s|),u = u0 + iksw0 + o(|s|) for |s| small. From (3.43), we have

i = l,F l,Fuu[w0,w0]

l,Fu[w0] . (4.8)

To be definite we assume 1 > 0 and 2 < 0, then + > 0 and < 0.Recall the definitions of A(s) and l(s) from the proof of Theorem 4.1, and here first we

consider the branch corresponding to (1, 1). Then A is differentiable, A(0) = l,F(0,0, u0) = 0, and since F(0, 0, u0) = 0, we have

A(0) = l(0),F+1l,F + 1l,F + k1l,Fu[w0]= k1

l,Fu[w0]

= l,Fu[w0]l,Fuu[w0,w0]

l,F l,Fuu[w0,w0]

> 0.

Thus A(s+) > 0. On the other hand, sincel,Fuu(0, 0, u0)[w0,w0]

< 0,

we have l(s),Fuu

((s), (s), u(s)

)[w(s),w(s)

]< 0,

for |s| small. From Theorem 1.1, near (+, u+) the solution forms a curve + = {(+(t), u+(t)):t [,]} with +(0) = +, u+(0) = u+, +(0) = 0 and+(0) = l(s+),Fuu(+, +, u+)[w(s+),w(s+)]

l(s+),F(+, +, u+) > 0.


Similarly the solutions near (, u) the solution forms a curve = {((t), u(t)): t [,]} with (0) = , u(0) = u, (0) = 0 and (0) < 0. The proof for the case < 0is similar.

It is clear that when (0, 0 + ) (0 , 0), there exists an interval I = (0 , 0 +) such that (3.1) has no solutions near (0, u0) when I (see Fig. 3).

5. Examples

5.1. Euler bucking beam: symmetry break of pitchfork bifurcation

The Euler buckling beam problem was one of the first bifurcation problems to be studied,as early as by Euler, see for example [1,9,10,23] for more historical remarks and details onmodeling. Here we follow an example of Reiss [23]:

+ sin = 0, 0 < x < 1, (0) = (1) = 0,u = cos 1, 0 < x < 1, u(0) = 0,w = sin, 0 < x < 1, w(0) = w(1) = 0.

(5.1)

Here the length of the elastic column is normalized so that 0 x 1, the horizontal and verticaldisplacements of the buckled axis are denoted by u(x) and w(x) respectively, (x) is the anglebetween the tangent to the columns axis and the x-axis, and is a parameter proportional to thethrust. The boundary conditions imply that the ends of the column are pinned (simply supported).For this classical problem, solutions can be obtained explicitly in terms of elliptic functions, andsupercritical pitchfork bifurcations occur along the trivial solutions = u = w = 0 from =(n)2 with n = 1,2, . . . . The minimal buckling load 1 is the Euler buckling load. To accountfor imperfections in the column, we assume that the initial unstressed axis is a curve instead ofa line, and the angle between the tangent of the curve and the horizontal is I (x). Then the equation in (5.1) is modified to

+ sin = g(x), 0 < x < 1, (0) = (1) = 0, (5.2)

where g(x) = I (x) and I (0) = I (1) = 0 (thus 1

0 g(x)dx = 0). In [14,23], singular per-turbation methods matching the inner and outer expansions were used to derive the imperfectbifurcation near (,u) = (2,0). Here we will apply Theorems 3.2 and 4.1 to analytically obtainthe precise bifurcation diagrams. But first we illustrate how the secondary bifurcation theorem(Theorem 2.7) can be applied to the bifurcation at = 0:

Proposition 5.1. Suppose that g(x) C1([0,1],R), 10 g(x)dx = 0, then for any R, thesolution set of (5.2) near (,u) = (0,0) is in form of

={(0, k): k R} {(,0(x)+ o(||)): (, )}, (5.3)

where 0(x) is the solution of = g(x), 0 < x < 1, (0) = (1) = 0. (5.4)


Proof. For fixed R, define G(,) = + sin g, where R and X = {v C2([0,1]): v(0) = v(1) = 0}. We denote W = R X and Y = C0([0,1]). Then G(0,0) = 0,and

G(,)[(, )

]= (sin g) + + cos,G(,)

[(1, 1)(2, 2)

]= 0 12 + cos(12 + 21) sin12. (5.5)Thus G(0,0)[(, )]= g , which has a two-dimensional kernel N(G(0,0))= span{(0,1),(1, 0)}. If h R(G(0,0)), then there exists (, ) such that g = h. Since 1

0 g(x)dx = 0, then the solvability of (, ) is reduced to 1

0 h(x)dx = 0. Hence R(G(0,0)) ={h Y : 10 h(x)dx = 0}, which is co-dimensional one. Also G(0,0)[(0,1)(1, 0)] = 1 /R(G(0,0)), and G(0,0)[(0,1)(0,1)] = 0 R(G(0,0)). Hence Theorem 2.7 can be applied toobtain results stated since (0, k) is always a solution for any k, and we can reparameterize thenontrivial branch so that is the new parameter.

Next we apply Theorem 4.1 near the first bifurcation point 1 = 2:

Proposition 5.2. Suppose that g(x) C1([0,1],R), 10 g(x)dx = 0, and 10 g(x) cosx dx > 0.Then there exists , 1, 2 > 0 such that for (,), N = {(,u) R X: | 2| 1,u 2} such that the solution set of (5.2) in N is in form of + , where + ={(+(t), u+(t)), |t | } is a -shaped curve with +(0) > 0, and (+(0), u+(0)) is the uniquedegenerate solution on + ; and = {(, u()): | 2| 1} is a monotone curvewithout degenerate solutions. Moreover the bifurcation point satisfies

+(0) = 2 + 322

( 10

g(x) cos(x)dx

)2/32/3 + o(2/3). (5.6)

Proof. We define

F(,,) = + sin g(x), (5.7)

where , R and X = {v C2([0,1]): v(0) = v(1) = 0}. With

F(,,)[w] = w + cos w (5.8)

we can easily check that for 1 = 2 and w0 = cosx, (F1) and (F2) are satisfied, andR(F(0, 1,0)) = {v X:

10 v(x) cos(x)dx = 0}. Since F(0, 1,0)[w0] = cos(0)w0 = w0

and 1

0 w0(x) cos(x)dx = 0, then (F3) is satisfied; so is (F4) since F(0, 1,0)[w0,w0] =1 sin(0) w20 = 0. Finally (F5) is satisfied since F(0, 1,0) = 1g(x) and we assumethat

10 g(x) cosx dx = 0. Hence Theorem 3.2 can be applied, with v1 = 0 and 2 = 0,

and the degenerate solutions are on a curve {Ts = ((s), (s), u(s),w(s)): |s| < } such that (0) = (0) = (0) = w (0) = 0, and u (0) = w0. Moreover we can verify that l,F < 0,l,F[w0,w0,w0] + 3l,F[w0,2] < 0, and l,F[w0] > 0. Therefore Theorem 4.1


also holds. Equations (3.26) and (3.29) imply that

(0) = 1 1

0 w40(x) dx 1

0 w20(x) dx

= 32

4, (5.9)

and

(0) = 2 1

0 w40(x) dx 1

0 w0(x)g(x) dx> 0. (5.10)

Hence Ts = ((p/6)s3 + o(|s|3),2 + (3/8)2s2 + o(s2), s cos(x)+ o(|s|), cos(x)+ o(|s|)),where p = (0) > 0. We rescale the parameter s so (s) = (p/6)s3 then (s) = 2 +(3/8)62/32p2/32/3 + o(2/3) = 2 + (3/2)2( 10 g(x) cos(x)dx)2/32/3 + o(2/3), whichis the -coordinate of the unique degenerate solution for fixed small . Thus the turning point inTheorem 4.1 satisfies (5.6).

Now by using the perturbation analysis of degenerate solutions in Propositions 5.1 and 5.2and implicit function theorem, we can obtain the following global imperfect bifurcation picturefor 0 < < 42 with small (the proof is standard, which we omit):

Theorem 5.3. Suppose that g(x) C1([0,1],R), 10 g(x)dx = 0, and 10 g(x) cosx dx > 0.For any 1 > 0, there exists

(1) If = 0, then (5.2) has the trivial solution u = 0 for all > 0, and has exactly two othersolutions u+(x) and u(x) for (2,42) and u(x) = u+(x) = u+(1 x);

(2) If (2, 2)\{0}, then there exists such that (5.2) has exactly one solution when (0, ), has exactly two solutions when = and has exact three solutions when (,42 1). Moreover = () is given by (5.6), > 2 if 2 > > 0 and < 2 if 0 > > 2.

Note that we can show that when 1

0 g(x) cosx dx < 0, same results hold but < 2 if

2 > > 0 and > 2 if 0 > > 2. We remark that the expansion in (5.6) was also obtainedpreviously in [23], but results we have here are rigorous not just formal perturbation expansion.

5.2. Perturbed diffusive logistic equation: symmetry break of a transcritical bifurcation

Diffusive logistic equation (Fisher equation) is one of most important reactiondiffusion equa-tions with connections to biological invasion of foreign species, propagation of genetic traits[8,12]. The steady state diffusive logistic equation on a bounded domain is in form of{

u+ (u u2)= 0 in ,u = 0 on , (5.11)

where is a positive parameter, and is a smooth bounded region in Rn for n 1. The nonlin-earity f (u) can be more general as the one defined in [26,27], but here for the transparency of

2presentation, we use f (u) = uu as in the classical equation. The perturbed problem of (5.12)arises when the population is harvested:

598 P. Liu et al. / Journal of Functional Analysis 251 (2007) 573600{u+ (u u2 g(x,u))= 0 in ,u = 0 on .

(5.12)

Perturbation and bifurcation analysis have been performed in [24,26]. Here we demonstrate ournew abstract theory with some new results and also rediscovery of old results.

In the following (1, 1) is the principal eigen-pair of{ + = 0 in , = 0 on , (5.13)

and we assume 1 is normalized and 1(x) > 0 in . The bifurcation diagram for unperturbedproblem is well known (see [26,27]):

Theorem 5.4. When = 0, (5.12) has no positive solution if 1, and has exactly one posi-tive solution v if > 1. Moreover, all vs lie on a smooth curve, lim1 v = 0 and v isincreasing with respect to .

Indeed (,u) = (1,0) is a bifurcation point where a transcritical bifurcation (see Theo-rem 1.2) occurs. To show that, we define

F 0(,u) = u+ (u u2), (5.14)where R and u X = {u C2,(): u = 0 on }. Then u = 0 is a trivial solu-tion for any , N(F 0u (1,0)) = span{1}, R(F 0u (1,0)) = {v C():

v1 dx = 0}, andF 0u(1,0)[1] = 1 / R(F 0u (1,0)). Hence (F1), (F2) and (F3) are satisfied, and Theorem 1.2is applicable. Moreover, F 0uu(1,0)[1, 1] = 2121 / R(F 0u (1,0)) since

31 dx = 0, thus(F4) is satisfied, and a transcritical bifurcation occurs at (1,0). Note that all these remain truefor a perturbed operator F(,,u) if F(0, ,u) F 0(,u).

First we consider {u+ (u u2) g(x) = 0, in ,u(x) = 0, on , (5.15)

where g(x) C1() and

g(x)1(x) dx = 0. (5.16)

We define

F(,,u) = u+ (u u2) g(x). (5.17)Then (F1), (F2), (F3) and (F4) are satisfied at (, ,u) = (0, 1,0) since F(0, ,u) F 0(,u)defined above. (F5) is also satisfied, as we have F(0, 1,0) = 1g(x) / R(Fu(0, 1,0)) from(5.16). From Theorem 3.1, there exists > 0 such that all the degenerate solutions of (5.15) nearT0 = (0, 1,0, 1) form a C2-curve{

Ts =((s), 1 + s + o

(|s|), ks1 + o(|s|), 1 + o(1)), s I = (, )}, (5.18)


where

k =

21(x) dx

21

31(x) dx> 0, (5.19)

(0) = (0) = 0, and from (3.17),

(0) = k2l,Fuu[w0,w0]

l,F =2k2

31(x) dx

g(x)1(x) dx. (5.20)

We assume

g(x)1(x) dx > 0 then (0) > 0, and the variation of the bifurcation diagramsnear = 0 is shown in Fig. 1 (for more detailed proof, see [24, Theorem 2.5 and Section 6.1]).

Next we consider {u+ (u u2) 2g(x) = 0, in ,u(x) = 0, on , (5.21)

where g(x) C1(), and g satisfies

g(x)1(x) dx > 0. (5.22)

We define

F(,,u) = u+ (u u2) 2g(x). (5.23)Again (F1), (F2), (F3) and (F4) are satisfied at (, ,u) = (0, 1,0). But F(0, 1,0) = 0 hence(F5) is satisfied. Moreover we also have Fu(0, 1,0)[1] = 0, F(0, 1,0) = 21g(x), thus

m l,F l,Fuu[1, 1]

= 421

g(x)1(x) dx

31(x) dx > 0, (5.24)

from (5.22). Hence the conclusions of Corollary 3.6 hold and (1, 1) and (2, 2) are non-zerolinear independent solutions of the equation

m2 q22 = 0, (5.25)

where q = l,Fu(0, 1,0)[1] =

21 dx > 0. We can choose (1, 1) = (q,m) and

(2, 2) = (q,m). Hence for = 0, the transition in Fig. 3 occurs. However, if (5.22) ischanged to

g(x)1(x) dx < 0, for = 0, there is no degenerate solutions nearby (see Fig. 4).

AcknowledgmentsThe authors thank the anonymous referee for careful reading and several helpful suggestions.

600 P. Liu et al. / Journal of Functional Analysis 251 (2007) 573600References

[1] S.S. Antman, Nonlinear Problems of Elasticity, second ed., Appl. Math. Sci., vol. 107, Springer, New York, 2005.[2] K.-C. Chang, Infinite-dimensional Morse Theory and Multiple Solution Problems, Progr. Nonlinear Differential

Equations Appl., vol. 6, Birkhuser Boston, Boston, MA, 1993.[3] S.N. Chow, J.K. Hale, Methods of Bifurcation Theory, Springer, New York, 1982.[4] P.T. Church, J.G. Timourian, Global structure for nonlinear operators in differential and integral equations. I. Folds.

II. Cusps, in: Topological Nonlinear Analysis, II, Frascati, 1995, in: Progr. Nonlinear Differential Equations Appl.,vol. 27, Birkhuser Boston, Boston, MA, 1997, pp. 109160, 161245.

[5] M.G. Crandall, P.H. Rabinowitz, Bifurcation from simple eigenvalues, J. Funct. Anal. 8 (1971) 321340.[6] M.G. Crandall, P.H. Rabinowitz, Bifurcation, perturbation of simple eigenvalues and linearized stability, Arch.

Ration. Mech. Anal. 52 (1973) 161180.[7] K. Deimling, Nonlinear Functional Analysis, Springer, Berlin, 1985.[8] R.A. Fisher, The wave of advance of advantageous genes, Ann. Eugenics 7 (1937) 353369.[9] M. Golubitsky, An introduction to catastrophe theory and its applications, SIAM Rev. 20 (2) (1978) 352387.

[10] M. Golubitsky, D. Schaeffer, A theory for imperfect bifurcation via singularity theory, Comm. Pure Appl.Math. 32 (1) (1979) 2198.

[11] H. Kielhfer, Bifurcation Theory. An Introduction with Applications to PDEs, Appl. Math. Sci., vol. 156, Springer,New York, 2004.

[12] A. Kolmogoroff, I. Petrovsky, N. Piscounoff, Study of the diffusion equation with growth of the quantity of matterand its application to a biological problem, Moscow Univ. Bull. Math. 1 (1937) 125 (in French).

[13] N.H. Kuiper, C1-equivalence of functions near isolated critical points, in: Symposium on Infinite-DimensionalTopology, in: Ann. of Math. Stud., vol. 69, Princeton Univ. Press, Princeton, NJ, 1972, pp. 199218.

[14] B.J. Matkowsky, E.L. Reiss, Singular perturbations of bifurcations, SIAM J. Appl. Math. 33 (2) (1977) 230255.[15] L. Nirenberg, Variational and topological methods in nonlinear problems, Bull. Amer. Math. Soc. (N.S.) 4 (3) (1981)

267302.[16] L. Nirenberg, Topics in Nonlinear Functional Analysis, Courant Lect. Notes Math., vol. 6, New York Univ., Courant

Inst. Math. Sci./Amer. Math. Soc., Providence, RI, 2001.[17] P. Liu, Y.-W. Wang, A generalized saddle-node bifurcation for degenerate solutions, Commen. Math. XLV (2)

(2005) 145150.[18] J. Lpez-Gmez, Spectral Theory and Nonlinear Functional Analysis, Chapman & Hall/CRC Res. Notes Math.,

vol. 426, Chapman & Hall/CRC, Boca Raton, FL, 2001.[19] S. Oruganti, J. Shi, R. Shivaji, Diffusive logistic equation with constant effort harvesting, I: Steady states, Trans.

Amer. Math. Soc. 354 (9) (2002) 36013619.[20] T. Ouyang, J. Shi, Exact multiplicity of positive solutions for a class of semilinear problem, J. Differential Equa-

tions 146 (1) (1998) 121156.[21] T. Ouyang, J. Shi, Exact multiplicity of positive solutions for a class of semilinear problem: II, J. Differential

Equations 158 (1) (1999) 94151.[22] L. Perko, Differential Equations and Dynamical Systems, first ed., Texts Appl. Math., vol. 7, Springer, New York,

1991.[23] E.L. Reiss, Imperfect bifurcation, in: Applications of Bifurcation Theory, Proc. Advanced Sem., Univ. Wisconsin,

Madison, WI, 1976, in: Publ. Math. Res. Center Univ. Wisconsin, vol. 38, Academic Press, New York, 1977, pp. 3771.

[24] J. Shi, Persistence and bifurcation of degenerate solutions, J. Funct. Anal. 169 (2) (1999) 494531.[25] J. Shi, Multi-parameter bifurcation and applications, in: H. Brezis, K.C. Chang, S.J. Li, P. Rabinowitz (Eds.), ICM

2002 Satellite Conference on Nonlinear Functional Analysis: Topological Methods, Variational Methods and TheirApplications, World Scientific, Singapore, 2003, pp. 211222.

[26] J. Shi, R. Shivaji, Global bifurcation for concave semipositon problems, in: G.R. Goldstein, R. Nagel, S. Romanelli(Eds.), Advances in Evolution Equations: Proceedings in Honor of J.A. Goldsteins 60th Birthday, Dekker, NewYork, 2003, pp. 385398.

[27] J. Shi, R. Shivaji, Persistence in reactiondiffusion models with weak Allee effect, J. Math. Biol. 52 (6) (2006)807829.

[28] J. Shi, Z. Xie, Smoothness of crossing curves, 2007, submitted for publication.

07-liu-shi-wangjfa

Documents