In this work, we consider the existence of solutions of the semilinear elliptic problem with a singular nonlinearity,
$ \left\{ \begin{array}{l} - {\rm{div}}\left( {\begin{array}{*{20}{c}} {\mathit{\boldsymbol{M}}\left( x \right)}&{\nabla u} \end{array}} \right) = h\left( x \right){u^{ - p}}\;\;\;{\rm{in}}\;\Omega ,\\ u > 0\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;{\rm{in}}\;\Omega ,\\ u = 0\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;{\rm{on}}\;\Omega , \end{array} \right. $ | (1) |
where Ω
$ \begin{array}{*{20}{c}} {\mathit{\boldsymbol{M}}\left( x \right)\mathit{\boldsymbol{\xi }} \cdot \mathit{\boldsymbol{\xi }} \ge \alpha {{\left| \mathit{\boldsymbol{\xi }} \right|}^2},}\\ {\left| {\mathit{\boldsymbol{M}}\left( x \right)} \right| \le \beta ,\forall x \in \Omega ,\mathit{\boldsymbol{\xi }} \in {{\bf{R}}^n},} \end{array} $ | (2) |
and h(x)>0 a.e.in Ω and -p < -1.By solutions we mean here weak solutions in H01(Ω), i.e., u∈H01(Ω) satisfying u(x)>0 in Ω and
$ \begin{array}{*{20}{c}} {\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla \nu {\rm{d}}x} - \int_\Omega {\frac{{h\left( x \right)}}{{{u^p}}} \cdot \nu {\rm{d}}x} = 0,}\\ {\forall \nu \in H_0^1\left( \Omega \right).} \end{array} $ |
Since the work by Stuart[1], people have paid much attention to the existence and multiplicity of solutions for such singular equations
$ - {\rm{div}}\left( {\begin{array}{*{20}{c}} {\mathit{\boldsymbol{M}}\left( x \right)}&{\nabla u} \end{array}} \right) = f\left( {x,u} \right), $ |
where f(x, s) is singular at s=0. See Refs.[2-5] and the rich list of references provided by these papers for a survey. Recently, Boccardo and Orsina[6] solved the problem with f(x, u)=h(x)u-p, h(x)≥0, -p < -1 and provided the existence of an Hloc1(Ω)-solution u by using approximation arguments when M(x) is a real symmetric matrix satisfying M(x)ξ·ξ≥α|ξ|2, |M(x)|≤β, ∀x∈Ω, ξ∈Rn and
We define the singular energy functional
$ \begin{array}{l} I\left( u \right) = \frac{1}{2}\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u{\rm{d}}x} + \\ \;\;\;\;\;\;\;\;\;\;\frac{1}{{p - 1}}\int_\Omega {h\left( x \right){{\left| u \right|}^{1 - p}}{\rm{d}}x} , \end{array} $ | (3) |
where -p < -1. The main difficulty is the absence of integrability of u-p for u∈H01(Ω) when -p < -1 and any inequality that relates u∈H01(Ω) will not be of much help. It should be noted also that there is a sharp contrast between the case -1 < -p < 0, for which the energy functional is continuous, and the case -p < -1. Generally, the sub-supersolution method is very effective in dealing with singularity. However, the method cannot be used for such general measurable h(x)>0. To reverse this situation, we use constrained sets to restore integrability and recast problem (1) into a variational framework in the spirit of our earlier works[9-12]. We defined constrained sets N1 and N2 as follows: N1:={u∈H01(Ω):u≥0 in Ω and ∫ΩM(x)▽u·▽u≥∫Ωh(x)|u|1-p}, N2:={u∈H01(Ω):u≥0 in Ω and ∫ΩM(x)▽u·▽u=∫Ωh(x)|u|1-p}. Here, special care must be taken to establish the validity and connection of the two constraints which simplify the existence of a minimizer for the singular functional I. It should also be noted that for -p < -1, N2 is not closed as usual (certainly not weakly closed) in H01(Ω).
In this paper we will use the notation,
C, Ci, ci, i=1, 2, …, denoting (possibly different) constants.
We denote the Dirichlet norm in H01(Ω) by ‖u‖2=∫Ω|▽u|2dx, |M(x)|=det M(x), and M(x)ξ·η:=ξTM(x)η.
1 Main resultsTheorem 1.1 Let Ω
$ - {\rm{div}}\left( {\begin{array}{*{20}{c}} {\mathit{\boldsymbol{M}}\left( x \right)}&{\nabla u} \end{array}} \right) = h\left( x \right){u^{ - p}} $ |
admits an H01-solution if there exists u0∈H01(Ω) such that
$ \int_\Omega {h\left( x \right){{\left| {{u_0}} \right|}^{1 - p}}{\rm{d}}x} < + \infty . $ | (4) |
Proof It should be noted that the topology on H01(Ω) which was generated by the norm
$ {\left( {\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u} } \right)^{\frac{1}{2}}} $ |
is equivalent to the one that was generated by the norm
$ {\left( {\int_\Omega {{{\left| {\nabla u} \right|}^2}} } \right)^{\frac{1}{2}}}, $ |
since
$ \begin{array}{*{20}{c}} {\alpha \int_\Omega {{{\left| {\nabla u} \right|}^2}} \le \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u} \le }\\ {\frac{\beta }{{{\alpha ^{n - 1}}}}\int_\Omega {{{\left| {\nabla u} \right|}^2}} .} \end{array} $ |
Hence
$ \begin{array}{*{20}{c}} {{{\left( {H_0^1\left( \Omega \right),{{\left( {\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u} } \right)}^{\frac{1}{2}}}} \right)}^ * } = }\\ {{{\left( {H_0^1\left( \Omega \right),{{\left( {\int_\Omega {{{\left| {\nabla u} \right|}^2}} } \right)}^{\frac{1}{2}}}} \right)}^ * }.} \end{array} $ |
Then, it follows that
$ {u_n} \to u\;{\rm{weakly}}\;{\rm{in}}\;\left( {H_0^1\left( \Omega \right),{{\left( {\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u} } \right)}^{\frac{1}{2}}}} \right) $ |
is equal to
$ {u_n} \to u\;{\rm{weakly}}\;{\rm{in}}\;\left( {H_0^1\left( \Omega \right),{{\left( {\int_\Omega {{{\left| {\nabla u} \right|}^2}} } \right)}^{\frac{1}{2}}}} \right). $ |
The key to prove (1) depends on a natural interpolation between the constrained sets Ni, i=1, 2. Taking u∈H01(Ω) with
$ \int_\Omega {h\left( x \right){{\left| u \right|}^{1 - p}}} < \infty , $ |
the function
$ U\left( t \right): = {t^{1 + p}}\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u{\rm{d}}x} - \int_\Omega {h\left( x \right){{\left| u \right|}^{1 - p}}} {\rm{d}}x $ |
is increasing on t>0 with
$ \begin{array}{*{20}{c}} {\frac{{{\rm{d}}I\left( {tu} \right)}}{{{\rm{d}}t}} = t\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u{\rm{d}}x} - {t^{ - p}}\int_\Omega {h\left( x \right){{\left| u \right|}^{1 - p}}} {\rm{d}}x}\\ { = {t^{ - p}}\left( {{t^{1 + p}}\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u{\rm{d}}x} - \int_\Omega {h\left( x \right){{\left| u \right|}^{1 - p}}} {\rm{d}}x} \right),} \end{array} $ |
then it follows that there exists the unique positive minimizer t(u)u such that
$ I\left( {tu} \right) \ge I\left( {t\left( u \right)u} \right),\forall t > 0. $ | (5) |
In particular, assumption (4) of Theorem 1.1 implies the existence of t(u0)>0 such that t(u0)u0∈N2 and hence N1(
$ \begin{array}{*{20}{c}} {\beta {{\left\| u \right\|}^2} \ge \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u} \ge \int_\Omega {h\left( x \right){{\left| u \right|}^{1 - p}}} }\\ { \ge {{\left( {\int_\Omega {h{{\left( x \right)}^{1 - p}}} } \right)}^p}{{\left( {\int_\Omega {\left| u \right|} } \right)}^{1 - p}}} \end{array} $ |
and Poincaré inequality
$ {\left( {\int_\Omega {\left| u \right|} } \right)^{1 - p}} \ge {C_1}{\left\| u \right\|^{1 - p}}, $ |
imply that ‖u‖≥C for all u∈N1.
It should also be noted that
$ \begin{array}{*{20}{c}} {\left| {\mathit{\boldsymbol{M}}\left( x \right)\mathit{\boldsymbol{\xi }} \cdot \mathit{\boldsymbol{\eta }}} \right| \le \frac{\beta }{{{\alpha ^{n - 1}}}}\left| \mathit{\boldsymbol{\xi }} \right| \cdot \left| \mathit{\boldsymbol{\eta }} \right|,}\\ {\forall \mathit{\boldsymbol{\xi }},\mathit{\boldsymbol{\eta }} \in {{\bf{R}}^n},x \in \Omega .} \end{array} $ | (6) |
Indeed, since M(x) is a real symmetric matrix, there exists an orthogonal matrix Q(x) such that
$ {\mathit{\boldsymbol{Q}}^{\rm{T}}}\mathit{\boldsymbol{MQ}} = \left( {\begin{array}{*{20}{c}} {{\lambda _1}}&{}&{}\\ {}& \ddots &{}\\ {}&{}&{{\lambda _n}} \end{array}} \right), $ |
where λi, i=1, …, n, the eigenvalues of M(x), satisfy λi≥α since M(x)ξ·ξ≥α|ξ|2, ∀x∈Ω, ξ∈Rn. Then it yields that
$ \mathit{\boldsymbol{M}} = \mathit{\boldsymbol{Q}}\left( {\begin{array}{*{20}{c}} {{\lambda _1}}&{}&{}\\ {}& \ddots &{}\\ {}&{}&{{\lambda _n}} \end{array}} \right){\mathit{\boldsymbol{Q}}^{\rm{T}}} $ |
and
$ \begin{array}{l} \mathit{\boldsymbol{M}}\left( x \right)\mathit{\boldsymbol{\xi }} \cdot \mathit{\boldsymbol{\eta }} = {\mathit{\boldsymbol{\xi }}^{\rm{T}}}\mathit{\boldsymbol{M}}\left( x \right)\mathit{\boldsymbol{\eta = }}{\mathit{\boldsymbol{\xi }}^{\rm{T}}}\mathit{\boldsymbol{Q}}\left( {\begin{array}{*{20}{c}} {{\lambda _1}}&{}&{}\\ {}& \ddots &{}\\ {}&{}&{{\lambda _n}} \end{array}} \right){\mathit{\boldsymbol{Q}}^{\rm{T}}}\mathit{\boldsymbol{\eta }}\\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\mathit{\boldsymbol{ = }}{\left( {{\mathit{\boldsymbol{Q}}^{\rm{T}}}\mathit{\boldsymbol{\xi }}} \right)^{\rm{T}}}\left( {\begin{array}{*{20}{c}} {{\lambda _1}}&{}&{}\\ {}& \ddots &{}\\ {}&{}&{{\lambda _n}} \end{array}} \right){\mathit{\boldsymbol{Q}}^{\rm{T}}}\mathit{\boldsymbol{\eta }}. \end{array} $ |
Hence, if one defines x:=QTξ, y:=QTη, one can obtain
$ \begin{array}{l} \mathit{\boldsymbol{M}}\left( x \right)\mathit{\boldsymbol{\xi }} \cdot \mathit{\boldsymbol{\eta }} = {\mathit{\boldsymbol{x}}^{\rm{T}}}\left( {\begin{array}{*{20}{c}} {{\lambda _1}}&{}&{}\\ {}& \ddots &{}\\ {}&{}&{{\lambda _n}} \end{array}} \right)y\\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\; = \left( {\begin{array}{*{20}{c}} {{x_1}}& \cdots &{{x_n}} \end{array}} \right)\left( {\begin{array}{*{20}{c}} {{\lambda _1}}&{}&{}\\ {}& \ddots &{}\\ {}&{}&{{\lambda _n}} \end{array}} \right)\left( {\begin{array}{*{20}{c}} {{y_1}}\\ \vdots \\ {{y_n}} \end{array}} \right)\\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\; = {\lambda _1}{x_1}{y_1} + {\lambda _2}{x_2}{y_2} + \cdots + {\lambda _n}{x_n}{y_n}, \end{array} $ |
which implies that
$ \begin{array}{l} \left| {\mathit{\boldsymbol{M}}\left( x \right)\mathit{\boldsymbol{\xi }} \cdot \mathit{\boldsymbol{\eta }}} \right| = \left| {\sum\limits_{i = 1}^n {{\lambda _i}{x_i}{y_i}} } \right|\\ \;\;\;\;\; = \left| {\prod\limits_{i = 1}^n {{\lambda _i}\left( {\frac{{{\lambda _1}}}{{\prod {{\lambda _i}} }}{x_1}{y_1} + \frac{{{\lambda _2}}}{{\prod {{\lambda _i}} }}{x_2}{y_2} + \cdots + \frac{{{\lambda _n}}}{{\prod {{\lambda _i}} }}{x_n}{y_n}} \right)} } \right.\\ \;\;\;\;\; \le \frac{\beta }{{{\alpha ^{n - 1}}}}\left| x \right| \cdot \left| y \right|. \end{array} $ |
Since
$ \begin{array}{l} {\left| x \right|^2} = {\mathit{\boldsymbol{x}}^{\rm{T}}} \cdot \mathit{\boldsymbol{x}} = {\left( {{\mathit{\boldsymbol{Q}}^{\rm{T}}}\mathit{\boldsymbol{\xi }}} \right)^{\rm{T}}} \cdot {\mathit{\boldsymbol{Q}}^{\rm{T}}}\mathit{\boldsymbol{\xi }} = {\mathit{\boldsymbol{\xi }}^{\rm{T}}}\mathit{\boldsymbol{Q}} \cdot {\mathit{\boldsymbol{Q}}^{\rm{T}}}\mathit{\boldsymbol{\xi }}\\ \;\;\;\;\; = {\mathit{\boldsymbol{\xi }}^{\rm{T}}} \cdot \mathit{\boldsymbol{\xi }} = {\left| \mathit{\boldsymbol{\xi }} \right|^2}, \end{array} $ |
$ \begin{array}{l} {\left| y \right|^2} = {\mathit{\boldsymbol{y}}^{\rm{T}}} \cdot \mathit{\boldsymbol{y}} = {\left( {{\mathit{\boldsymbol{Q}}^{\rm{T}}}\mathit{\boldsymbol{\eta }}} \right)^{\rm{T}}} \cdot {\mathit{\boldsymbol{Q}}^{\rm{T}}}\mathit{\boldsymbol{\eta }} = {\mathit{\boldsymbol{\eta }}^{\rm{T}}}\mathit{\boldsymbol{Q}} \cdot {\mathit{\boldsymbol{Q}}^{\rm{T}}}\mathit{\boldsymbol{\eta }}\\ \;\;\;\;\; = {\mathit{\boldsymbol{\eta }}^{\rm{T}}} \cdot \mathit{\boldsymbol{\eta }} = {\left| \mathit{\boldsymbol{\eta }} \right|^2}, \end{array} $ |
it follows that
$ \left| {\mathit{\boldsymbol{M}}\left( x \right)\mathit{\boldsymbol{\xi }} \cdot \mathit{\boldsymbol{\eta }}} \right| \le \frac{\beta }{{{\alpha ^{n - 1}}}}\left| \mathit{\boldsymbol{x}} \right| \cdot \left| \mathit{\boldsymbol{y}} \right| = \frac{\beta }{{{\alpha ^{n - 1}}}}\left| \mathit{\boldsymbol{\xi }} \right| \cdot \left| \mathit{\boldsymbol{\eta }} \right|. $ |
Furthermore, for u∈N1,
$ \begin{array}{l} \int {h\left( x \right){{\left| u \right|}^{1 - p}}} \le \int {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u} \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \le \frac{\beta }{{{\alpha ^{n - 1}}}}\int {{{\left| {\nabla u} \right|}^2} < \infty } . \end{array} $ | (7) |
Then there exists tu>0 such that tuu∈N2, I(u)≥I(tuu)≥
$ \mathop {\inf I}\limits_{{N_1}} \ge \mathop {\inf I}\limits_{{N_2}} . $ | (8) |
However, since N1
$ \mathop {\inf I}\limits_{{N_1}} \le \mathop {\inf I}\limits_{{N_2}} . $ | (9) |
In view of (8) and (9), it yields that
$ \mathop {\inf I}\limits_{{N_1}} = \mathop {\inf I}\limits_{{N_2}} . $ |
Now, we turn our attention to
$ \begin{array}{l} \mathop {\lim \;\rm{inf}}\limits_{n \to \infty } \int {h\left( x \right){{\left| {{u_n}} \right|}^{1 - p}}} \le \mathop {\lim \;\rm{inf}}\limits_{n \to \infty } \int {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \le \mathop {\lim \;\rm{inf}}\limits_{n \to \infty } \frac{\beta }{{{\alpha ^{n - 1}}}}\int {{{\left| {\nabla {u_n}} \right|}^2} < \infty } , \end{array} $ |
and based on Fatou' lemma, we obtain
$ \begin{array}{l} \int {h\left( x \right){{\left| u \right|}^{1 - p}}} = \int {\mathop {\lim \;\rm{inf}}\limits_{n \to \infty } h\left( x \right){{\left| {{u_n}} \right|}^{1 - p}}} \\ \le \mathop {\lim \;\rm{inf}}\limits_{n \to \infty } \int {h\left( x \right){{\left| {{u_n}} \right|}^{1 - p}}} \le \mathop {\lim \;\rm{inf}}\limits_{n \to \infty } \int {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} \\ = \int {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u} . \end{array} $ |
On the other hand, we can claim that I(u) is weakly lower semi-continuous, that is,
$ \begin{array}{*{20}{c}} {I\left( u \right) \le \mathop {\lim \;\rm{inf}}\limits_{n \to \infty } I\left( {{u_n}} \right)}\\ {{\rm{as}}\;{u_n} \to u\;{\rm{weakly}}\;{\rm{in}}\;\left( {H_0^1\left( \Omega \right),{{\left( {\int {{{\left| {\nabla {u_n}} \right|}^2}} } \right)}^{\frac{1}{2}}}} \right).} \end{array} $ |
as un→u weakly in
$ \int {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} \to \int {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u} , $ |
and by Fatou's lemma one can also obtain
$ \int {h\left( x \right){{\left| u \right|}^{1 - p}}} \le \mathop {\lim \;\rm{inf}}\limits_{n \to \infty } \int {h\left( x \right){{\left| {{u_n}} \right|}^{1 - p}}} $ |
if un→u weakly in
$ \begin{align} & \left( \text{ⅰ} \right)I\left( {{u}_{n}} \right) < \underset{{{N}_{1}}}{\mathop{\inf }}\,I+\frac{1}{n} \\ & \left( \text{ⅱ} \right)I\left( {{u}_{n}} \right)\le I\left( v \right)+\frac{1}{n}||{{u}_{n}}-v||,\forall v\in {{N}_{1}} \\ \end{align} $ |
since N1 is a closed set in H01(Ω). We may assume un≥0 as I(u)=I(|u|). Since -p < -1, I(u) is coercive on N1 and therefore (un) is bounded in H01(Ω). Indeed, by -p < -1,
$ \begin{array}{*{20}{c}} {I\left( u \right) = \frac{1}{2}\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u{\rm{d}}x} + }\\ {\frac{1}{{p - 1}}\int_\Omega {h\left( x \right){{\left| u \right|}^{1 - p}}{\rm{d}}x} \ge \frac{\alpha }{2}{{\left\| u \right\|}^2}.} \end{array} $ |
Hence, up to subsequence (still denoted by un), un
$ {u^ * } > 0\;{\rm{a}}{\rm{.}}\;{\rm{e}}{\rm{.}}\;{\rm{in}}\;\Omega $ | (10) |
as ∫Ωh(x)|u*|1-p < ∞ by Fatou's lemma. Moreover, we shall show that u*∈N2 by evaluating the best minimizing sequence (un)∈N1.
Case 1. Suppose that (un)N1\N2 for all n large. Fix φ∈H01(Ω), φ≥0 and n by now. Note that, as (un)N1\N2 and p>1, there holds that
$ \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla \left( {{u_n} + t\varphi } \right) \cdot \nabla \left( {{u_n} + t\varphi } \right)} > \int_\Omega {h\left( x \right){{\left( {{u_n} + t\varphi } \right)}^{1 - p}}{\rm{d}}x} , $ |
that is,
$ {u_n} + t\varphi \in {N_1}. $ |
In virtue of (ⅰ) and (ⅱ), we obtain that
$ \begin{array}{l} \frac{t}{n}\left\| \varphi \right\| + \frac{1}{2}\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\left( {{{\left| {\nabla \left( {{u_n} + t\varphi } \right)} \right|}^2} - {{\left| {\nabla {u_n}} \right|}^2}} \right){\rm{d}}x} \\ \ge \frac{1}{{1 - p}}\int_\Omega {h\left( x \right)\left( {{{\left| {\left( {{u_n} + t\varphi } \right)} \right|}^{1 - p}} - {{\left| {{u_n}} \right|}^{1 - p}}} \right){\rm{d}}x} . \end{array} $ |
Dividing by t>0, passing to the liminf as t→0+, we obtain
$ \begin{array}{l} \frac{{\left\| \varphi \right\|}}{n} + \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla \varphi } \\ \;\;\;\;\;\;\; \ge \int_\Omega {\mathop {\lim \;\rm{inf}}\limits_{n \to \infty } \frac{{h\left( x \right)}}{{1 - p}}\frac{{{{\left( {{u_n} + t\varphi } \right)}^{1 - p}} - u_n^{1 - p}}}{t}} \\ \;\;\;\;\;\;\; = \int_\Omega {h\left( x \right)u_n^{ - p}\varphi } . \end{array} $ |
Using Fatou's lemma again and letting n tend to infinity, we have
$ \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla \varphi } \ge \int_\Omega {h\left( x \right){u^{ * - p}}\varphi } ,\forall \varphi \ge 0. $ |
In view of (10), we obtain that u*∈N1, and by the above argument (5) there exists a unique t(u*) such that I(t(u*)u*)=
$ \begin{array}{l} \mathop {\inf }\limits_{{N_1}} I = \mathop {\lim }\limits_{n \to \infty } I\left( {{u_n}} \right)\\ \; = \mathop {\lim }\limits_{n \to \infty } \left[ {\frac{1}{2}\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} + \frac{1}{{p - 1}}\int_\Omega {h\left( x \right)u_n^{1 - p}} } \right]\\ \; \ge \mathop {\lim \;\rm{inf}}\limits_{n \to \infty } \left[ {\frac{1}{2}\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} } \right] + \\ \;\;\;\;\mathop {\lim \;\rm{inf}}\limits_{n \to \infty } \left[ {\frac{1}{{p - 1}}\int_\Omega {h\left( x \right)u_n^{1 - p}} } \right]\\ \; \ge \frac{1}{2}\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla {u^ * }} + \frac{1}{{p - 1}}\int_\Omega {h\left( x \right){u^{ * 1 - p}}} \\ \; = I\left( {{u^ * }} \right) \ge I\left( {t\left( {{u^ * }} \right){u^ * }} \right) \ge \mathop {\inf }\limits_{{N_2}} I \ge \mathop {\inf }\limits_{{N_1}} I, \end{array} $ |
and thus t(u*)=1, which means that
$ \mathop {\min }\limits_{t > 0} I\left( {t{u^ * }} \right) = I\left( {{u^ * }} \right),{u^ * } \in {N_2}. $ |
Case 2. There exists a subsequence of (un) (still denoted by un), which belongs to N2.
Let φ∈H10(Ω),φ≥0, be fixed. Since -p < -1,
$ \begin{array}{*{20}{c}} {\int_\Omega {h\left( x \right){{\left( {{u_n} + t\varphi } \right)}^{1 - p}}{\rm{d}}x} \le \int_\Omega {h\left( x \right)u_n^{1 - p}{\rm{d}}x} < \infty ,}\\ {\forall t \ge 0.} \end{array} $ |
By the previous argument (5), the function fn, φ(t):=t(un+tφ), ∀t≥0 exists, and, moreover, using the notation therein, fn, φ(0)=1 and fn, φ(t)(un+tφ)∈N2. The continuity of fn, φ(t), t>0 depends on ∫Ωh(x)|un|1-p < ∞ and dominates convergence. Indeed,
$ \begin{array}{*{20}{c}} {f_{n,\varphi }^2\left( t \right)\int {\mathit{\boldsymbol{M}}\left( x \right)\nabla \left( {{u_n} + t\varphi } \right) \cdot \nabla \left( {{u_n} + t\varphi } \right)} }\\ {f_{n,\varphi }^{1 - p}\left( t \right)\int {h\left( x \right){{\left( {{u_n} + t\varphi } \right)}^{1 - p}}} ,} \end{array} $ |
that is,
$ {f_{n,\varphi }}\left( t \right) = {\left[ {\frac{{\int {h\left( x \right){{\left( {{u_n} + t\varphi } \right)}^{1 - p}}} }}{{\int {\mathit{\boldsymbol{M}}\left( x \right)\nabla \left( {{u_n} + t\varphi } \right) \cdot \nabla \left( {{u_n} + t\varphi } \right)} }}} \right]^{\frac{1}{{p + 1}}}}. $ |
The key to showing that u*∈N2 hinges on the estimation of f′n, φ(0) defined as
$ {{f'}_{n,\varphi }}\left( 0 \right) = \mathop {\lim }\limits_{n \to \infty } \frac{{\left( {{f_{n,\varphi }}\left( t \right) - 1} \right)}}{t} \in \left[ { - \infty , + \infty } \right]. $ |
If the limit does not exist, we let tk→0 (instead of t→0) with tk>0 chosen in such a way that
$ \begin{array}{l} 0 = f_{n,\varphi }^2\left( t \right)\int {\mathit{\boldsymbol{M}}\left( x \right)\nabla \left( {{u_n} + t\varphi } \right) \cdot \nabla \left( {{u_n} + t\varphi } \right)} - \\ \;\;\;\;\;f_{n,\varphi }^{1 - p}\left( t \right)\int {h\left( x \right){{\left( {{u_n} + t\varphi } \right)}^{1 - p}}} , \end{array} $ |
$ 0 = \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} - \int_\Omega {h\left( x \right){{\left| {{u_n}} \right|}^{1 - p}}} . $ |
By the continuity of fn, φ(t), t>0, it holds
$ \begin{array}{l} 0 = \left\{ {\left( {{f_{n,\varphi }}\left( t \right) + 1} \right)\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla \left( {{u_n} + t\varphi } \right) \cdot \nabla \left( {{u_n} + t\varphi } \right)} - } \right.\\ \left. {\left( {1 - p} \right){{\left[ {{f_{n,\varphi }}\left( 0 \right) + o\left( 1 \right)} \right]}^{ - p}}\int_\Omega {h\left( x \right){{\left( {{u_n} + t\varphi } \right)}^{1 - p}}} } \right\} \cdot \\ \frac{{\left( {{f_{n,\varphi }}\left( t \right) - 1} \right)}}{t} - \frac{1}{t}\left\{ {\int_\Omega {h\left( x \right){{\left( {{u_n} + t\varphi } \right)}^{1 - p}}} + h\left( x \right)u_n^{1 - p}{\rm{d}}x - } \right.\\ \left. {\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla \left( {{u_n} + t\varphi } \right) \cdot \nabla \left( {{u_n} + t\varphi } \right) - \mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}{\rm{d}}x} } \right\}, \end{array} $ |
and by letting t→0+, then
$ \begin{array}{l} 0 \ge {{f'}_{n,\varphi }}\left( 0 \right)\left\{ {2\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} + \left( {p - 1} \right)\int_\Omega {h\left( x \right)u_n^{1 - p}} } \right\} + \\ \;\;\;\;\;2\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla \varphi } , \end{array} $ |
which implies that f′n, φ(0)≠+∞. Indeed, due to un∈N2
$ \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} \ge \alpha {\left\| {{u_n}} \right\|^2} > \alpha r_0^2 > 0. $ | (11) |
Since -p < -1, by the reversed Hölder inequality it yields that
$ \begin{array}{*{20}{c}} {\int_\Omega {h\left( x \right)u_n^{1 - p}} \ge {{\left( {\int_\Omega {h{{\left( x \right)}^{1/p}}} } \right)}^p}{{\left( {\int_\Omega {{u_n}} } \right)}^{1 - p}}}\\ { \ge {C_2}{{\left( {\int_\Omega {h{{\left( x \right)}^{1/p}}} } \right)}^p}{{\left\| {{u_n}} \right\|}^{1 - p}} > 0.} \end{array} $ |
In addition,
$ \begin{array}{*{20}{c}} {\left| {\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla \varphi } } \right| \le \frac{\beta }{{{\alpha ^{n - 1}}}} \cdot }\\ {\int_\Omega {\left| {\nabla {u_n}} \right| \cdot \left| \varphi \right| \le {C_3}\left\| {{u_n}} \right\| \cdot \left\| \varphi \right\|} .} \end{array} $ | (12) |
Furthermore, since r0 is independent of n, it follows that
$ {{f'}_{n,\varphi }}\left( 0 \right) \le {c_1}\;{\rm{uniformly}}\;{\rm{in}}\;n. $ | (13) |
On the other hand, we will show that f′n, φ(0) cannot go to -∞ as n→∞, that is, f′n, φ(0) is bounded from below uniformly for all n large. Indeed, by the fact that u∈N2, we have
$ \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u} = \int_\Omega {h\left( x \right){{\left| {{u_n}} \right|}^{1 - p}}} , $ |
which imples
$ I\left( u \right) = \left( {\frac{1}{2} + \frac{1}{{p - 1}}} \right)\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u} , $ |
and by condition (ⅱ) we have the additional condition
$ \begin{array}{l} \frac{1}{n}\left| {\frac{{1 - {f_{n,\varphi }}\left( t \right)}}{t}} \right| \cdot \left\| {{u_n}} \right\| + \frac{1}{n}{f_{n,\varphi }}\left( t \right)\left\| \varphi \right\|\\ \ge \frac{1}{n}\left\| {{u_n} \cdot {f_{n,\varphi }}\left( t \right)\left( {{u_n} + t\varphi } \right)} \right\|\frac{1}{t}\\ \ge \left[ {I\left( {{u_n}} \right) - I\left( {{f_{n,\varphi }}\left( t \right)\left( {{u_n} + t\varphi } \right)} \right.} \right]\frac{1}{t}, \end{array} $ |
that is,
$ \begin{array}{l} \frac{{\left\| \varphi \right\|}}{n}{f_{n,\varphi }}\left( t \right) \ge \frac{{{f_{n,\varphi }}\left( t \right) - 1}}{t}\left\{ { - \left( {\frac{1}{2} + \frac{1}{{p - 1}}} \right)\left[ {{f_{n,\varphi }}\left( t \right) + 1} \right] \cdot } \right.\\ \left. {\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right){{\left| {\nabla \left( {{u_n} + t\varphi } \right)} \right|}^2}} - \frac{{\left\| {{u_n}} \right\|}}{n} \cdot {\mathop{\rm sgn}} \left( {{f_{n,\varphi }}\left( t \right) - 1} \right)} \right\} - \\ \frac{1}{t}\left( {\frac{1}{2} + \frac{1}{{p - 1}}} \right)\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\left[ {{{\left| {\nabla \left( {{u_n} + t\varphi } \right)} \right|}^2} - {{\left| {\nabla {u_n}} \right|}^2}} \right]{\rm{d}}x} . \end{array} $ |
Letting t→0+, we obtain that
$ \begin{array}{l} \frac{{\left\| \varphi \right\|}}{n} \ge - {{f'}_{n,\varphi }}\left( 0 \right)\left\{ {\left( {\frac{1}{2} + \frac{1}{{p - 1}}} \right)\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} + } \right.\\ \left. {\frac{{\left\| {{u_n}} \right\|}}{n} \cdot {\mathop{\rm sgn}} {{f'}_{n,\varphi }}\left( 0 \right)} \right\} - \left( {1 + \frac{1}{{p - 1}}} \right)\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla \varphi } . \end{array} $ |
By (2) and (11) it yields that
$ \beta {\left\| u \right\|^2} \ge \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla u \cdot \nabla u} \ge \alpha {\left\| {{u_n}} \right\|^2} > \alpha r_0^2 > 0, $ |
and in view of (12) it holds that f′n, φ is bounded below. More precisely,
$ {{f'}_{n,\varphi }}\left( 0 \right) \ge {c_2}\;{\rm{uniformly}}\;{\rm{in}}\;{\rm{all}}\;n\;{\rm{large}} $ | (14) |
as r0 is independent of n.
Now, applying condition (ⅱ) again, we have that
$ \begin{array}{l} \frac{1}{n}\left[ {\frac{{\left| {{f_{n,\varphi }}\left( t \right) - 1} \right|}}{t}\left\| {{u_n}} \right\| + {f_{n,\varphi }}\left( t \right)\left\| \varphi \right\|} \right]\\ \;\;\; \ge \frac{1}{n}\left\| {{f_{n,\varphi }}\left( t \right)\left( {{u_n} + t\varphi } \right) - {u_n}} \right\|\frac{1}{t}\\ \;\;\; \ge \left[ {I\left( {{u_n}} \right) - I\left( {{f_{n,\varphi }}\left( t \right)\left( {{u_n} + t\varphi } \right)} \right.} \right]\frac{1}{t}, \end{array} $ |
that is,
$ \begin{array}{l} \frac{{\left\| {{u_n}} \right\|}}{n}\frac{{\left| {{f_{n,\varphi }}\left( t \right) - 1} \right|}}{t} + \frac{{\left\| \varphi \right\|}}{n}{f_{n,\varphi }}\left( t \right)\\ \ge \left\{ { - \frac{{\left[ {{f_{n,\varphi }}\left( t \right) + 1} \right]}}{2}\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla \left( {{u_n} + t\varphi } \right) \cdot \nabla \left( {{u_n} + t\varphi } \right) + } } \right.\\ \left. {{{\left[ {{f_{n,\varphi }}\left( 0 \right) + o\left( 1 \right)} \right]}^{ - p}}\int_\Omega {h\left( x \right){{\left( {{u_n} + t\varphi } \right)}^{1 - p}}{\rm{d}}x} } \right\} \cdot \\ \frac{{{f_{n,\varphi }}\left( t \right) - 1}}{t} + \frac{1}{{1 - p}}\int_\Omega {\frac{{h\left( x \right)\left[ {{{\left( {{u_n} + t\varphi } \right)}^{1 - p}} - u_n^{1 - p}} \right]}}{t} - } \\ \frac{{\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla \left( {{u_n} + t\varphi } \right) \cdot \nabla \left( {{u_n} + t\varphi } \right) - \mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} }}{{2t}}. \end{array} $ |
In other words,
$ \begin{array}{l} \frac{1}{{p - 1}}\int_\Omega {\frac{{h\left( x \right)\left[ {u_n^{1 - p} - {{\left( {{u_n} + t\varphi } \right)}^{1 - p}}} \right]}}{t}} \\ \le \left\{ {\frac{{\left[ {{f_{n,\varphi }}\left( t \right) + 1} \right]}}{2}\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla \left( {{u_n} + t\varphi } \right) \cdot \nabla \left( {{u_n} + t\varphi } \right) - } } \right.\\ \left. {{{\left[ {{f_{n,\varphi }}\left( 0 \right) + o\left( 1 \right)} \right]}^{ - p}}\int_\Omega {h\left( x \right){{\left( {{u_n} + t\varphi } \right)}^{1 - p}}{\rm{d}}x} } \right\} \cdot \\ \frac{{{f_v}\left( t \right) - 1}}{t} + \frac{{\left\| {{u_n}} \right\|}}{n}\frac{{\left| {{f_{n,\varphi }}\left( t \right) - 1} \right|}}{t} + \frac{{\left\| \varphi \right\|}}{n}{f_{n,\varphi }}\left( t \right) + \\ \frac{{\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla \left( {{u_n} + t\varphi } \right) \cdot \nabla \left( {{u_n} + t\varphi } \right) - \mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} }}{{2t}}. \end{array} $ |
In view of (13)and (14), it holds that
$ \begin{array}{l} \mathop {\lim \inf }\limits_{t \to {0^ + }} \frac{1}{{p - 1}}\int_\Omega {\frac{{h\left( x \right)\left[ {u_n^{1 - p} - {{\left( {{u_n} + t\varphi } \right)}^{1 - p}}} \right]}}{t}} \\ \le {{f'}_{n,\varphi }}\left( 0 \right)\left\{ {\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla {u_n}} - h\left( x \right){{\left( {{u_n} + t\varphi } \right)}^{1 - p}}{\rm{d}}x} \right\} + \\ \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla \varphi {\rm{d}}x} + \frac{1}{n}\left[ {{{f'}_{n,\varphi }}\left( 0 \right)\left\| {{u_n}} \right\| + \left\| \varphi \right\|} \right]\\ = \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla \varphi {\rm{d}}x} + \frac{1}{n}\left[ {{{f'}_{n,\varphi }}\left( 0 \right)\left\| {{u_n}} \right\| + \left\| \varphi \right\|} \right]\\ < \infty . \end{array} $ |
On the other hand, since -p < -1, φ≥0, h(x)>0, and t>0, we have
$ \frac{{h\left( x \right)\left[ {u_n^{1 - p} - {{\left( {{u_n} + t\varphi } \right)}^{1 - p}}} \right]}}{t} \ge 0, $ |
and by Fatou' lemma, we have
$ \begin{array}{l} \int_\Omega {h\left( x \right)u_u^{1 - p}\varphi {\rm{d}}x} \\ \le \mathop {\lim \inf }\limits_{t \to {0^ + }} \frac{1}{{p - 1}}\int_\Omega {\frac{{h\left( x \right)\left[ {u_n^{1 - p} - {{\left( {{u_n} + t\varphi } \right)}^{1 - p}}} \right]}}{t}} \\ \le \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u_n} \cdot \nabla \varphi } + \frac{1}{n}\left[ {{{f'}_{n,\varphi }}\left( 0 \right)\left\| {{u_n}} \right\| + \left\| \varphi \right\|} \right]. \end{array} $ |
Hence, using Fatou's lemma again and n→∞, we obtain
$ \int_\Omega {h\left( x \right){u^{ * 1 - p}}{\rm{d}}x} \le \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla \varphi ,\forall \varphi } \ge 0. $ |
In other words,
$ \int_\Omega {h\left( x \right){u^{ * 1 - p}}{\rm{d}}x} - \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot la\varphi \ge 0,\forall \varphi } \ge 0. $ |
By the same reasoning as in case 1 we derive that
$ {u^ * } \in {N_2}. $ |
Now it remains to show that u*∈H01(Ω) is a weak solution for problem (1) for all -p < -1. Letting ψ∈H01(Ω) be fixed and applying the above inequalities one finds
$ \begin{array}{l} 0 \le \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla {{\left( {{u^ * } + t\psi } \right)}^ + }} - \int_\Omega {h\left( x \right){u^{ * - p}}{{\left( {{u^ * } + t\psi } \right)}^ + }} \\ = \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla \left( {{u^ * } + t\psi } \right)} - \int_\Omega {h\left( x \right){u^{ * - p}}\left( {{u^ * } + t\psi } \right)} - \\ \;\;\;\int_{{u^ * } + t\psi < 0} {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla \left( {{u^ * } + t\psi } \right)} + \\ \;\;\;\int_\Omega {h\left( x \right){u^{ * - p}}\left( {{u^ * } + t\psi } \right)} \\ \le t\left\{ {\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla \psi } - \int_\Omega {h\left( x \right){u^{ * - p}}\psi } } \right\} - \\ \;\;\;\;\int_{{u^ * } + t\psi < 0} {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla {u^ * }} - t\int_{{u^ * } + t\psi < 0} {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla \psi } \\ \le t\left\{ {\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla \psi } - \int_\Omega {h\left( x \right){u^{ * - p}}\psi } - } \right.\\ \;\;\;\;\left. {\int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla \psi } } \right\}. \end{array} $ |
Since meas[u*+tψ < 0]→0 as t→0, we may divide the inequality by t>0 and pass to the limit as t→0, and we conclude that
$ 0 \le \int_\Omega {\mathit{\boldsymbol{M}}\left( x \right)\nabla {u^ * } \cdot \nabla \psi } + \int_\Omega {h\left( x \right){u^{ * - p}}\psi {\rm{d}}x} . $ |
By the arbitrariness of ψ∈H01(Ω), u* is indeed a H01(Ω)-solution of problem (1).
[1] | Stuart C A. Existence and approximation of solutions of non-linear elliptic equations[J]. Mathematische Zeitschrift, 1976, 147(1):53–63. DOI:10.1007/BF01214274 |
[2] | Arcoya D, Carmona J, Leonori T, et al. Existence and nonexistence of solutions for singular quadratic quasilinear equations[J]. Dierential Equations, 2009, 246(10):4006–4042. DOI:10.1016/j.jde.2009.01.016 |
[3] | Boccardo L. Dirichlet problems with singular and gradient quadratic lower order terms[J]. ESAIM:Control, Optimisation and Calculus of Variations, 2008, 14(3):411–426. DOI:10.1051/cocv:2008031 |
[4] | Giachetti D, Murat F. An elliptic problem with a lower order term having singular behaviour[J]. Bollettino Dell Unione Matematica Italiana, 2009, 2(2):349–370. |
[5] | Crandall M G, Rabinowitz P H, Tartar L. On a Dirichlet problem with a singular nonlinearity[J]. Communications in Partial Dierential Equations, 1977, 2(2):193–222. DOI:10.1080/03605307708820029 |
[6] | Boccardo L, Orsina L. Semilinear elliptic equations with singular nonlinearities[J]. Calculus of Variations and Partial Dierential Equations, 2010, 37(3):363–380. |
[7] | Boccardo L. A Dirichlet problem with singular and supercritical nonlinearities[J]. Nonli near Analysis:Theory, Methods & Applications, 2012, 75(12):4436–4440. |
[8] | Boccardo L, Casado-Dìaz J. Some properties of solutions of some semilinear elliptic singular problems and applications to the Gconvergence[J]. Asymptotic Analysis, 2014, 86(1):1–15. |
[9] | Sun Y J. Compatible phenomena in singular problems[J]. Proc Roy Soc Edinburgh, 2013, 143(A):1321–1330. |
[10] | Sun Y J, Zhang D Z. The role of the power 3 for elliptic equations with negative exponents[J]. Calculus of Variations and Partial Dierential Equations, 2014, 49(3/4):909–922. |
[11] | Sun Y J, Wu S P. An exact estimate result for a class of singular equations with critical exponents[J]. Journal of Functional Analysis, 2011, 260(5):1257–1284. DOI:10.1016/j.jfa.2010.11.018 |
[12] | Liu X, Sun Y J. Multiple positive solutions for Kirchho type problems with singularity[J]. Comm Pure Appl Anal, 2013, 12(2):721–733. |
[13] | Ekeland I. On the variational principle[J]. Math Anal Appl, 1974, 47:324–353. DOI:10.1016/0022-247X(74)90025-0 |