1. Introduction
Numerous works on the fundamental gap of the first two eigenvalues have been developed in the past few decades, in particular, the fundamental gap is defined as $\lambda _2-\lambda _1$, where $\lambda _1$ and $\lambda _2$ represent the first and second eigenvalues of the given differential equation, respectively.
Most of these investigations are focused on the Schrödinger equation, which has been discussed in a variety of situations. In quantum mechanics, the size for fundamental gap is extremely crucial, if it is small enough, it will produce the well-known tunnelling effect, so the research on this problem is of great significance. Ashbaugh et al. [Reference Ashbaugh, Harrell and Svirsky4] proposed the existence and characterization for minimal and maximal fundamental gaps when the potential is constrained by various ${L^{p}}$ norm, moreover, Karaa extended this work on [Reference Karaa19]. Chern and Shen determine [Reference Chern and Shen12] the minimum fundamental gap with certain restrictions on the class of potential functions. The estimation of the upper and lower bounds of the fundamental gap has also attracted more attention of scholars. In particular, note that the fundamental gap bound related to single-well potential with transition point as midpoint is considered in [Reference Horváth16], and the symmetric potential is imposed by various constraints in [Reference Huang and Tsai17]. Yu [Reference Yu and Yang36] introduced the bound for the fundamental gap under Dirichlet and Neumann boundary condition, where the single-well potential with transition point to be not midpoint. Andrews and Clutterbuck [Reference Andrews and Clutterbuck3] solved the gap conjecture ${\lambda _2} - {\lambda _1} \ge \frac {3\pi ^{2}}{d^{2}}$ via an ingenious way, where $d$ is the diameter of domain and the potential is weakly convex. For more such topics, we may refer to [Reference Chen and Huang10, Reference Henrot15, Reference Hyun18, Reference Kikonko and Mingarelli22, Reference Lavine23].
The eigenvalue problem for degenerate elliptic equations has also attracted extensive attention. The author for [Reference Lucia and Schuricht24] considered a general class of eigenvalue problems where the leading elliptic term corresponds to a convex homogeneous energy function that is not necessarily differentiable, and they derived uniqueness of the first eigenfunction. Stuart [Reference Stuart31] developed bifurcation from the line of trivial solutions for a nonlinear eigenvalue problem on a bounded open subset $\Omega \in R^{n}$ with $n\ge 3$, and introduced a critically degenerate elliptic Dirichlet problem in [Reference Stuart32] for $n\ge 2$. For more aspects relating to degenerate elliptic equations, we can learn from other references [Reference Caldiroli and Musina7–Reference Chabrowski9, Reference Morimoto and Xu26, Reference Weber34]. Meanwhile, degenerate parabolic equation theory has flourished over the past years, interested readers can refer [Reference Alabau-Boussouira, Cannarsa and Fragnelli1, Reference Buffe and Phung6, Reference Moyano27], in particular, [Reference Buffe and Phung6] established a Lebeau–Robbiano spectral inequality for a degenerate one-dimensional elliptic operator.
With regard to sub-elliptic operator, Xu [Reference Xu35] considered the existence and the regularity for the minimum points of a certain variational problem, the properties of weak solutions are presented in detail for some sub-elliptic operator in [Reference Rodney28, Reference Sawyer and Wheeden29]. In terms of eigenvalues, Chen and Luo [Reference Chen and Luo11] studied the lower bound estimates for the $j$-th eigenvalue for some degenerate elliptic operators for $n \ge 2$.
However, the fundamental gap of sub-elliptic operator had been a largely under explored domain, too little work has been devoted to it, especially for the degenerate sub-elliptic operator even in the one-dimensional case. Inspired by the above researches, the goal of this paper is to study the minimization problem $\inf \limits _{V\in \mathcal {V} }(\lambda _2(V)-\lambda _1(V))$ about the following sub-elliptic equation:
where $h\in C^{1}[0,\,1]$ with $h(0)=0$ and $h(x)>0,\, x\in (0,\,1]$, and
where $0< m< M$ are two given constants.
And most remarkably, in some of the latest literature [Reference El Allali and Harrell2, Reference Kerner20, Reference Kerner21], Kerner [Reference Kerner20, Reference Kerner21] introduced the explicit lower bound on the fundamental gap of one-dimensional Schrödinger operators with non-negative bounded potentials with Neumann boundary conditions, which operator is a uniform elliptic operator, and the class of potential function $V(x)$ is different from ours. Allali and Harrell [Reference El Allali and Harrell2] provided lower bounds for the gap of Sturm–Liouville problem with general single-well potential $V (x)$ with a transition point $a$, without any restriction on $a$, and also for the case where the potential is convex, while $V(x)$ demands certain monotonicity or convexity, which is also different from our work.
The remaining part of the paper proceeds as follows: § 2 is devoted to the study the weak solution space of sub-elliptic equation. Under certain assumptions, we propose a more profound characterization of the weak solution space in § 3. Section 4 begins by laying out the characteristic of the eigenvalue and the corresponding eigenfunction of (1.1), and looks at the differential of eigenvalue. Section 5 presents the existence of the minimum fundamental gap problem and the performance of the optimal function.
2. Existence of weak solution
Prior to commencing the study of eigenvalue problem for (1.1), we first shall look for the proper weak solution space of the equation:
where $f\in L^{2}(0,\,1)$ are given functions. On the basis of [Reference Alabau-Boussouira, Cannarsa and Fragnelli1, Reference Buffe and Phung6, Reference Monticelli and Payne25, Reference Sawyer and Wheeden29, Reference Sawyer and Wheeden30], we provide a more comprehensive analysis of the weak solution space for this sub-elliptic operator. We remark that in the next work, the existence and uniqueness of a solution for problem (2.1) follows from the Lax–Milgram Theorem.
Throughout this paper, we let $\|\cdot \|$ denote the norm, $(\cdot,\,\cdot )$ denote the inner product and define
where
and
Theorem 2.1 Let $H^{1}(0,\,1;h),\, (\cdot,\,\cdot ),\, \|\cdot \|$ be defined as (2.2), (2.3) and (2.4), respectively. Then $(H^{1}(0,\,1;h),\,(\cdot,\,\cdot ))$ is a Hilbert space with norm $\|\cdot \|$. Moreover,
and hence $hu\in H^{1}(0,\,1)$ for all $u\in H^{1}(0,\,1;h)$.
Proof. For the first point, it is evidently that $(H^{1}(0,\,1;h),\,(\cdot,\,\cdot ))$ is an inner product space and $\|\cdot \|$ is a norm on $H^{1}(0,\,1;h)$. Besides, the expression (2.5) is deduced from
For the second point, let $\{u_k\}_{k\in \mathbb {N}}$ be a Cauchy sequence in $H^{1}(0,\,1;h)$. i.e.,
which implies that there exists $u,\,v\in L^{2}(0,\,1)$ such that
In addition,
we observe that for each $\phi \in C_c^{\infty }(0,\,1)$:
one can find that
that is $hu'=v \mbox { in } \mathcal {D}'(0,\,1)$, hence $u_k\rightarrow u$ in $H^{1}(0,\,1;h)$ is obtained.
For the third point, as $u\in H^{1}(0,\,1;h)$, we have $(hu)'\in L^{2}(0,\,1)$ by invoking (2.5), and thus
in line with the definition of $h(x)$. Overall, these results indicate that $hu\in H^{1}(0,\,1)$.
Remark 2.2 Let $h=x$, then $\sqrt {x}\in H^{1}(0,\,1;h)$, but $\sqrt {x}\notin H^{1}(0,\,1)$. So, $H^{1}(0,\,1)\subsetneq H^{1}(0,\,1;h)$.
Proposition 2.3 Let $u\in H^{1}(0,\,1;h),$ then $u\in C^{0,\frac {1}{2}}(0,\,1]$ and $hu\in C^{0,\frac {1}{2}}[0,\,1]$.
Proof. Take $\delta \in (0,\,1)$, thanks to
and employing the fact $\inf \limits _{x\in [\delta,1]}h(x)>0$, we readily check $\|u'\|_{L^{2}(\delta,1)}<\infty$, consequently, $u\in H^{1}(\delta,\,1)$. And applying the Sobolev Embedding Theorem to find that $u\in C^{0,\frac {1}{2}}[\delta,\,1]$, due to the arbitrary of $\delta >0$, we deduce that $u\in C^{0,\frac {1}{2}}(0,\,1]$.
Finally, we have $hu\in H^{1}(0,\,1)$ by theorem 2.1, then $hu\in C^{0,\frac {1}{2}}[0,\,1]$ by utilizing the Sobolev Embedding Theorem again.
For $u\in H^{1}(0,\,1;h)$, we easily check that $u(0+)=\lim \limits _{x\downarrow 0}u(x)$ exists and it is an extended real numbers, $(hu)(0)=\lim \limits _{x\downarrow 0}h(x)u(x)$ by proposition 2.3. Take into account the boundary condition, we shall have to gain further results.
Definition 2.4 Define
in the sense of
Remark 2.5 Indeed, definition 2.4 is well-defined owing to proposition 2.3. It is obvious that $H_0^{1}(0,\,1;h)$ is a subspace of $H^{1}(0,\,1;h)$ and $C_c^{\infty }(0,\,1)\subset H_0^{1}(0,\,1;h)$. Moreover, from remark 2.2 we know that $H_0^{1}(0,\,1)\subsetneq H_0^{1}(0,\,1;h)$. Note that if $u\in H_0^{1}(0,\,1;h)$, then $hu\in H_0^{1}(0,\,1)$.
Lemma 2.6 The space $H_0^{1}(0,\,1;h)$ is a Hilbert space with inner product $(\cdot,\,\cdot )$ in $H^{1}(0,\,1;h)$.
Proof. Let $\{u_n\}_{n\in \mathbb {N}}\subset H_0^{1}(0,\,1;h)$ be a Cauchy sequence, then $\{u_n\}_{n\in \mathbb {N}}$ is a Cauchy sequence in $H^{1}(0,\,1;h)$, there exists $u\in H^{1}(0,\,1;h)$ such that
from theorem 2.1. On the other side, since $\{hu_n\}_{n\in \mathbb {N}}\subset H_0^{1}(0,\,1)$ and
there exists $v\in H_0^{1}(0,\,1)$ such that
This together with (2.7), that yields $v=hu$ on $(0,\,1)$, and along with the fact that $(hu)(0)=u(1)=0$, so $u\in H_0^{1}(0,\,1;h)$ is derived.
Lemma 2.7 Generally, the space $H_0^{1}(0,\,1;h)$ cannot be embedded compactly in $L^{2}(0,\,1)$.
Proof. Assume $h=x^{2}$, for all $n\in \mathbb {N}$, let
Then
and
which implies that $\{u_n\}_{n\in \mathbb {N}}\subset H_0^{1}(0,\,1;h)$.
Suppose that there exists a subsequence $\{u_{n_k}\}_{k\in \mathbb {N}}$ of $\{u_n\}_{n\in \mathbb {N}}$, and $u\in L^{2}(0,\,1)$, such that
then $\|u\|_{L^{2}(0,1)}^{2}=\tfrac {1}{3}$, and
Furthermore, there exists a subsequence of $\{u_{n_k}\}_{k\in \mathbb {N}}$, still denoted by itself, such that
In view of the definition of $u_{n_k} (k\in \mathbb {N})$, the above conclusion implies that $u=0$ a.e. in $(0,\,1)$, but it contradicts to the fact $\|u\|_{L^{2}(0,1)}^{2}=\frac {1}{3}$.
Definition 2.8 We call $u\in H_0^{1}(0,\,1;h)$ is a weak solution of (2.1), provided
where
Now we rely on the specific bilinear form $B(u,\,v)$ for $u,\,v \in H_0^{1}(0,\,1;h)$ defined in (2.8) to test the hypothesis of Lax–Milgram Theorem.
Theorem 2.9 Let $f\in L^{2}(0,\,1)$ be arbitrarily given function, then equation (2.1) has a unique solution in $H_0^{1}(0,\,1;h)$.
Proof. For all $v\in H_0^{1}(0,\,1;h)$, we have
and hence $f: H_0^{1}(0,\,1;h)\rightarrow \mathbb {R},\, v\mapsto (f,\,v)_{L^{2}(0,1)}$ is a bounded linear function on $H_0^{1}(0,\,1;h)$.
Next, we shall explain that $B(\cdot,\,\cdot )$ satisfies the Lax–Milgram Theorem. Indeed, through applying Cauchy inequality we obtain
and
All of these show that $B(\cdot,\,\cdot )$ satisfies the Lax–Milgram Theorem. On the basis of the mentioned analysis and Lax–Milgram Theorem, we see that there exists a unique $u\in H_0^{1}(0,\,1;h)$ such that
for all $v\in H_0^{1}(0,\,1;h)$.
Lemma 2.10 Consider
where $u\in H_0^{1}(0,\,1;h)$ is the unique solution of (2.1) , then $L$ is a self-adjoint bounded linear operator.
Proof. Employing (2.8) and
for which we deduce that $\|u\|_{H_0^{1}(0,1;h)}\leq C \|f\|_{L^{2}(0,1)}$, and thereby
Evidently, the above formula provides evidence that $L: L^{2}(0,\,1)\rightarrow L^{2}(0,\,1)$ is a bounded linear operator. On other side,
for all $f,\,g\in L^{2}(0,\,1)$, where $u,\,v$ are the solutions of (2.1) with respect to $f$ and $g$ respectively. Finally, we explicitly note that $L^{*}=L$, i.e., $L$ is a self-adjoint operator.
3. Further results
In this section and later sections, we further assume that
actually, there are many functions that satisfy (3.1).
Let us consider first the boundary-value problem (2.1) with the assumption (3.1), our overall plan is first to define and then construct an appropriate weak solution $u$ of (3.1) and only later to investigate the eigenvalue problem and other properties of $u$.
Definition 3.1 Denote
and we define
Lemma 3.2 The space
Proof. As we can see the definition above has states that $\overline {C_c^{\infty }(0,\,1)}^{H^{1}(0,1;h)} \subset H_c^{1}(0,\,1;h)$, it suffices to prove that, $H_c^{1}(0,\,1;h) \subset \overline {C_c^{\infty }(0,\,1)}^{H^{1}(0,1;h)}$. Indeed, for all $u\in H_c^{1}(0,\,1;h)$, there exists a sequence $u_n \in Lip_c(0,\,1)$, where $u_n$ is a Lipschitz continuous function on $(0,\,1)$ and ${\rm supp}u_n$ is compact in $(0,\,1)$, then $u_n'$ exists a.e. on $(0,\,1)$ and $|u_n'|\leq C \mbox { a.e. on } (0,\,1)$, where $C$ is the Lipschitz constant of $u_n$, these facts illustrate that $u_n\in H_0^{1}(0,\,1)$ with compact support on $(0,\,1)$. It is well known that there exists $\{v_n\}_{n\in \mathbb {N}}\subset C_c^{\infty }(0,\,1)$ such that
and we have
by the definition of $h(x)$, finally,
so the desired result is verified.
Lemma 3.3 The space $H_c^{1}(0,\,1;h)$ is a complete subspace of $H^{1}(0,\,1;h),$ and $u(0)=u(1)=0$ for any $u\in H_c^{1}(0,\,1;h)$. Moreover, $H_c^{1}(0,\,1;h)$ can be embedded compactly in $L^{2}(0,\,1)$.
Proof. 1. Let $\{u_n\}_{n\in \mathbb {N}}\subset H_c^{1}(0,\,1;h)$ be a Cauchy sequence, then there exists $u\in H^{1}(0,\,1;h)$ such that
Hence, for arbitrary $\epsilon >0$, there exists $N\in \mathbb {N}$, such that for all $n\geq N$ we have $\frac {1}{2^{n}}<\frac {\epsilon }{2}$ and
Note that for each $n\in \mathbb {N}$, there exists $\phi _n\in C_c^{\infty }(0,\,1)$, such that
these evidences suggest that
for all $n\geq N$. i.e., $u\in H_c^{1}(0,\,1;h)$.
2. For any $u\in H_c^{1}(0,\,1;h)$, there exists $\phi _n\in C_c^{\infty }(0,\,1)$ such that
and consider that
hence using Sobolev Embedding Theorem, for any $\delta \in (0,\,1)$,
Clearly, $\lim \limits _{n\to \infty }\phi _n(1)=u(1)=0$. Besides, it should be noted here that according to the integration absolutely continuous of $h^{-1}$, for arbitrary $\epsilon >0$, there exists $\xi >0$, for all $E\subset (0,\,1),\, |E|<\xi$, we have
Then for all $\delta \in (0,\,1)$, since
which shows that $\phi _n$ is uniformly convergent, we have $\lim \limits _{n\to \infty }\lim \limits _{\delta \to 0}\phi _n(\delta )=\lim \limits _{\delta \to 0}\lim \limits _{n\to \infty }\phi _n(\delta )=0$, furthermore, we have $u(0)=\lim \limits _{\delta \to 0}u(\delta )=0$ from (3.2) and proposition 2.3.
3. Let $\{u_n\}_{n\in \mathbb {N}}\subset H_c^{1}(0,\,1;h)$ be a bounded set. On one hand, note that
for all $\delta \in (0,\,1)$, as a consequence,
by Sobolev Embedding Theorem.
Due to (3.3), for any $0< x< y<1$, we have
for all $|x-y|<\xi,\, x,\,y\in (0,\,1)$. This implies that $\{u_n\}_{n\in \mathbb {N}}\subset C[0,\,1]$ is a bounded and equicontinuous sequence according to $u\in C^{0,\frac {1}{2}}(0,\,1]$ and $u(0)=u(1)=0$ for all $u\in H_c^{1}(0,\,1;h)$. The Arzela–Asoli Theorem tells us that there exists a convergent subsequence of $\{u_n\}_{n\in \mathbb {N}}$, still denoted by itself, and $u_0\in C[0,\,1]$, such that
that is a sure sign that
Lemma 3.4 There is a relationship among the space $H_0^{1}(0,\,1),$ $H_c^{1}(0,\,1;h),$ $H_0^{1}(0,\,1;h)$ and $H^{1}(0,\,1;h)$ by
Proof. That is easily verifiable for $H_c^{1}(0,\,1;h)\subsetneq H_0^{1}(0,\,1;h)\subsetneq H^{1}(0,\,1;h)$ based on lemma 3.3 and the definition of $H_0^{1}(0,\,1;h)$.
For $H_0^{1}(0,\,1)\subsetneq H_c^{1}(0,\,1;h)$, we give an example. Let $h(x)=x^{\frac {1}{4}}$, $g_1(x)=\sqrt {x}$, $g_2(x)=(x-1)^{2}$, let $f(x)$ be infinitely differentiable satisfying $f^{(k)}(\frac {1}{2})=g_1^{(k)}(\frac {1}{2})$ and $f^{(k)}(\frac {3}{4})=g_2^{(k)}(\frac {3}{4})$, where $(\cdot )^{(k)}$ is the $k$-th derivative of $(\cdot )$, $k=0,\,1,\,2 \cdots$. Consider
then
clearly, $u\notin H_0^{1}(0,\,1)$. Now let $\xi \in C_c^{\infty }(R)$ satisfy
and let
where $m>4$. We observe that $mx>2$ if $x\in (\tfrac {2}{m},\,\tfrac {1}{2}]$ and $mx+1-m< -1$ if $x\in [\tfrac {1}{2},\,1-\tfrac {2}{m})$. According to the definition of $\xi$, we have $\xi =0$ on $(\tfrac {2}{m},\,1-\tfrac {2}{m})$, so that $u_m(x)=u(x)$ on $(\tfrac {2}{m},\,1-\tfrac {2}{m})$. See that
since $\xi (mx)$, $\xi (mx+1-m)\le 1$. Moreover, since $u_m'=u'(1-\xi )-mu\xi '$ and $\xi =\xi '=0$ on $(\tfrac {2}{m},\,1-\frac {2}{m})$, we have $u_m'(x)=u'(x)$ on $(\frac {2}{m},\,1-\frac {2}{m})$. Then
therefore $u_m(x) \to u(x)$ in $H^{1}(0,\,1;h)$. Note that $u_m=0$ on $[0,\,\tfrac {1}{m}] \cup [1-\tfrac {1}{m},\,1]$, we can mollify the $u_m$ to produce functions $\omega _m \in C_c^{\infty }(0,\,1)$ such that $\omega _m \to u$ in $H_c^{1}(0,\,1;h)$, thus $u \in H_c^{1}(0,\,1;h)$.
Lemma 3.5 The space $H_c^{1}(0,\,1;h)$ is separable and reflexive.
Proof. Define $L_2^{2}(0,\,1)=L^{2}(0,\,1) \times L^{2}(0,\,1)$, where
for $u=(u_1,\,u_2) \in L_2^{2}(0,\,1)$ as the norm of space $L_2^{2}(0,\,1)$. It is no doubt that $L_2^{2}(0,\,1)$ is a separable space in the light of that $L^{2}(0,\,1)$ is a separable. Set
evidently, $W=\{Pu|u\in H^{1}(0,\,1;h)\}$ is a subspace of $L_2^{2}(0,\,1)$. From $\|Pu\|_{L_2^{2}(0,1)}=\|u\|_{H^{1}(0,1;h)}$, we know that $P$ is an isometric isomorphism of mapping $H^{1}(0,\,1;h)$ to $W$. In view of the fact that $H^{1}(0,\,1;h)$ is complete, $W$ is a closed subspace of $L_2^{2}(0,\,1)$, furthermore, $W$ is separable. Note that $P$ is an isometric isomorphism, then $H^{1}(0,\,1;h)$ has the same properties. Lemma 3.4 implies that the space $H_c^{1}(0,\,1;h)$ is a separable Hilbert space. Similarly, reflexivity is also obtained.
Definition 3.6 Define
we call $u\in H_c^{1}(0,\,1;h)$ is a solution to equation (2.1) if
for all $v\in H_c^{1}(0,\,1;h)$.
Theorem 3.7 Let $f\in L^{2}(0,\,1)$ be arbitrarily given function, then equation (2.1) has a unique solution in $H_c^{1}(0,\,1;h)$.
Proof. Same to the proof of theorem 2.9.
So far, we have constructed a suitable weak solution space $H_c^{1}(0,\,1;h)$, especially this space can be compactly embedded into $L^{2}(0,\,1)$ space, this feature provides an vital theoretical support for the study of eigenvalues in § 4.
4. Characterization of eigenvalues
For the fundamental gap optimization problem of Schrödinger operator [Reference Ashbaugh, Harrell and Svirsky4, Reference Horváth16, Reference Karaa19], the differential of eigenvalues is a valuable tool and an important link in the process of exploration, the optimality conditions for extremum problems will be given with the help of formulae for derivatives of eigenvalues.
At the beginning this section mainly investigates the characteristics of eigenvalue and eigenfunction, so as to better serve the later exploration of the first derivative of eigenvalues. We will exploit the distinct techniques, it primarily depends on the properties of compact operators, Harnack inequality and so on [Reference Bennewitz, Brown and Weikard5, Reference Evans13, Reference Gilbarg and Trudinger14, Reference V. Egorov and Kondratiev33]. Among them, the proof for Harnack inequality is placed in the appendix.
Lemma 4.1 Define
where $u\in H_c^{1}(0,\,1;h)$ is the unique solution of (2.1). Then $L$ is a self-adjoint compact linear operator.
Proof. By the same argument of lemma 2.10 we obtain that $L$ is a self-adjoint bounded linear operator. This together with lemma 3.3 we obtain $L$ is a compact linear operator.
Remark 4.2 It is obvious that $\dim H_c^{1}(0,\,1;h)=\infty$, and hence $0\in \sigma (L),\, \sigma (L)-\{0\}=\sigma _p(L)-\{0\}$, and $\sigma (L)-\{0\}$ is a sequence tending to $0$, where $\sigma (L)$ is the spectrum of $L$ and $\sigma _p(L)$ represents the point spectrum of $L$.
Theorem 4.3 Denote
Then (i) Each eigenvalue of $S$ is real.
(ii) Furthermore, if we repeat eigenvalue according to its (finite) multiplicity, we have
where
and
(iii) Finally, there exists an orthonormal basis $\{w_k\}_{k\in \mathbb {N}}\subset L^{2}(0,\,1),$ where $w_k\in H_c^{1}(0,\,1;h)$ is an eigenfunction corresponding to $\lambda _k$:
Proof. Set $L=S^{-1}$, then $L$ is a self-adjoint compact linear operator owing to lemma 4.1. Also noteworthy,
for any given $f \in L^{2}(0,\,1)$.
Consider $L^{2}(0,1)$ is separable, by invoking theorem D7 (pp. 728) in [Reference Evans13], there exists a countable orthonormal basis of $L^2{(0,1)}$ consisting of eigenvectors of $L$. It should be pointed out that for $\eta \neq 0$, $Lw=\eta w$ if and only if $Sw=\lambda w$ for $\lambda =\frac {1}{\eta }$. Consequently, we prove theorem 4.3.
Theorem 4.4 (i) We have
(ii) Furthermore, the above minimum is attained for a function $w_1\in H_c^{1}(0,\,1;h),$ positive within $(0,\,1),$ which solves
If $u\in H_c^{1}(0,\,1;h)$ is any weak solution of
then $u$ is a multiple of $w_1$.
(iii) Let $u_k$ denote the $k$-th normalized eigenfunction for operator $S,$ then $\lambda _k=\min \{B(u,\,u)\mid u\in H_c^{1}(0,\,1;h),\, u\bot V_{k-1},\, \|u\|_{L^{2}(0,1)}=1\},$ where $V_{k-1}=span\{u_1,\,u_2\cdots,\,u_{k-1}\},$ the equality holds if and only if $u=w_k$.
(iv) The connected components of the open sets $\Omega _+=\{x\in (0,\,1),\,u_k(x)>0\}$ and $\Omega _-=\{x\in (0,\,1),\,u_k(x)<0\}$ are called the nodal domains of $u_k$, suppose the eigenvalue $\lambda _{k}$ of the operator $S$ is simple, then $u_k$ has at most $k$ nodal domains, $k\ge 2$.
Proof. We carry out this proof by several steps.
(i) From theorem 4.3, we have
Since $\{w_k\}_{k=1}^{\infty }$ is the orthogonal basis of $L^{2}(0,\,1)$, if $u\in H_c^{1}(0,\,1;h)$ and $\|u\|_{L^{2}(0,1)}=1$, then we write
the series converging in $L^{2}(0,\,1)$. By (4.1), $\frac {w_k}{\sqrt {\lambda _k}}$ is an orthonormal subset of $H_c^{1}(0,\,1;h)$, endowed with the new inner product $B(\cdot,\,\cdot )$. Indeed,
Moreover, $B(w_k,\,u)=\lambda _k(w_k,\,u)=0$ for $k=1,\,2\cdots$, which implies that $u \equiv 0$ due to $\lambda _k>0$ and $\{w_k\}_{k=1}^{\infty }$ is the orthonormal basis of $L^{2}(0,\,1)$. Hence $u=\sum \limits _{k=1}^{\infty }\mu _k \frac {w_k}{\sqrt {\lambda _k}}$ for $\mu _k=B(u,\,\frac {w_k}{\sqrt {\lambda _k}})$, the series converging in $H_c^{1}(0,\,1;h)$. Together with the equality (4.2), we know that $\mu _k= {d_k}{\sqrt {\lambda _k}}$, thus $u=\sum \limits _{k=1}^{\infty }d_kw_k$ is convergent in $H_c^{1}(0,\,1;h)$.
By employing the equalities (4.1) and (4.2), then
and the equality holds for $u=w_1$, clearly, the desired result (i) is obtained.
(ii) We next claim that if $u \in H_c^{1}(0,\,1;h)$ and $\|u\|_{L^{2}(0,1)}=1$, then $u$ is a weak solution of
if and only if
Naturally, (4.4) implies (4.5). On the other side, suppose that (4.5) is established, writing $d_k=(u,\,w_k)$, then
Therefore
Evidently, we have $d_k=(u,\,w_k)=0$ provided that $\lambda _k>\lambda _1$, then $u=\sum \limits _{k=1}^{n}(u,\,w_k)w_k$ for some $n$ according to the multiplicity finiteness of $\lambda _1$, where $Sw_k=\lambda _1w_k$. Hence
the claim is confirmed.
From (i), we see that if $u$ is a eigenfunction of $\lambda _1$, then $|u|$ is one also. By lemma A.1, we must have $|u|$ is positive in $(0,\,1)$ and hence $\lambda _1$ has a positive eigenfunction. This argument indicates that the eigenfunctions of $\lambda _1$ are either positive or negative and thereby it is impossible that two of them are orthogonal, i.e. $\lambda _1$ is simple.
(iii) We will apply the mathematical induction to prove it. Suppose the conclusion is established for $k=2,\,\cdots,\, N-1$. For $u \bot V_{N-1}$, then equality (4.1) must be satisfied, and note that
therefore
thus we must have $d_1=0,\,\cdots,\, d_{N-1}=0$ by (4.1), this leads to $\sum \limits _{k=N}^{\infty }d_k^{2}=1$. Moreover,
this inequality holds if and only if $u=w_N$.
(iv) Suppose $u_k$ has more than $k$ nodal domains, let $G_1,\,\cdots,\,G_k,\,G_{k+1}$ be the nodal domains of $u_k$ and consider
where $a_1,\,\cdots,\, a_{k+1}$ are constants to be chosen later on. It is not hard to find that $u_k{\chi _{G_n}} \in H_c^{1}(0,\,1;h)$ for every $n$, and thus $v \in H_c^{1}(0,\,1;h)$. We may choose a nontrivial $(N+1)$-tuple $(a_1,\,\cdots,\, a_{k+1})$ such that $v$ is orthogonal to the eigenfunctions $u_1,\,\cdots,\,u_k$, this is allowed since there are $N$ equations and $N+1$ coefficients. Next, multiplying by an appropriate coefficient such that $\int _{0}^{1}v^{2}{\rm d}\kern0.06em x=1$, we record the new coefficient as $(c_1,\,c_2,\,\cdots,\,c_{k+1})$. According to the assumption that the eigenvalue of $S$ is simple, by the same way as (4.7), we have
Additionally, since $-(h^{2}(u_k{\chi _{G_n}})')'+Vu_k{\chi _{G_n}}=\lambda _{k}u_k{\chi _{G_n}}$ for $n=1,\,2\cdots,\,k+1$, the equality
leads to
this contradicts inequality (4.8).
The above results provide the definition of eigenvalue and prove the characteristics of the first eigenfunction by utilizing Harnack inequality. In order to fully understand the behavior of eigenvalues and eigenfunctions of equation (1.1), we require more detailed exploration content.
Theorem 4.5 The eigenvalue $\lambda _{k}$ of equation (1.1) is simple, $k>2$.
Proof. Suppose $u$ and $v$ are two nontrivial linear independent solutions corresponding to $\lambda _{k}$, then
Let $W_h=h^{2}u'v-h^{2}v'u$, then $W_h'=(h^{2}u')'v-(h^{2}v')'u$. By invoking [Reference Evans13] (chapter 6), the interior regularity for $u,\,v \in H_{loc}^{2}(\delta,\,1)$ is obtained for any $\delta \in (0,\,1)$. Then we discover that $u,\,v\in C^{1,\frac {1}{2}}[\delta,\,1]$ take advantage of Sobolev Embedding Theorem for any $\delta \in (0,\,1)$ and $u(1)=v(1)=0$. For all $\phi \in C_c^{\infty }(0,\,1)$, we have $\int _0^{1} W_h'\phi {\rm d}\kern0.06em x= \int _0^{1} ((h^{2}u')'v-(h^{2}v')'u)\phi {\rm d}\kern0.06em x=\int _0^{1} [(V-\lambda _{k})uv-(V-\lambda _{k})uv] \phi {\rm d}\kern0.06em x =0$ by (4.10), therefore $W_h'=0$ a.e. in the sense of distribution for $x\in [0,\,1]$, which implies that $W_h$ is a constant C a.e., combined with the fact that $W_h(1)=0$ and $W_h$ is continuous on $[\delta,\,1]$ for any $\delta \in (0,\,1)$, we immediately obtain that $C=0$ a.e. on $[0,\,1]$.
If $W_h=0$ a.e. on $[0,\,1]$, that is, $u'v-v'u=0$ a.e. on $(0,\,1)$, i.e. $(\frac {u}{v})'v^{2}=0$ a.e. on $(0,\,1)$, also given that there is no positive measure subset in $(0,\,1)$ such that $v= 0$ on it by lemma 4.1.3 on [Reference Bennewitz, Brown and Weikard5], then $(\frac {u}{v})'=0$ a.e. on $(0,\,1)$. This argument shows that $\frac {u}{v}$ is a constant on $(0,\,1)$ a.e., which contradicts that $u$ and $v$ are linear independent, so that there is one and only one linearly independent solution to each $\lambda _{k}$.
Lemma 4.6 Let the functions $u$ and $v$ that do not vanish identically on $(0,\,1)$ and satisfy the equations
where $h \in C^{1}[0,\,1]$ with $h(0)=0$ and $h(x)>0$ for $x\in (0,\,1],$ $V_1(x)>V_2(x)$ in $(0,\,1)$. Suppose $x_1$ and $x_2$ be two consecutive zeros of $u,$ then there is at least one zero $x_0$ of $v$ and satisfy $x_1< x_0< x_2,$ where $x_0\in (0,\,1)$.
Proof. Let $u(x)>0$ on $(x_1,\,x_2)$, suppose $v(x)$ has no zero point on $(x_1,\,x_2)$, without loss of generality, let $v(x)>0$ on $(x_1,\,x_2)$. Multiplying equation (4.11) by $u(x)$ and equation (4.12) by $\frac {u^{2}(x)}{v(x)}$ and subtracting, that yields the equality
Integrating this equality over $(x_1,\,x_2)$, then yields
Integrating by parts and utilizing the fact that $u(x_1)=u(x_2)=0$
however, the left side of the equality is positive, this is a contradiction.
Theorem 4.7 The eigenfunction $u_k$ of equation (1.1) has exactly $k-1$ zeroes in $(0,\,1),$ $k=1,\,2\cdots$.
Proof. For $k=1$, theorem 4.4 implies that $u_1$ has no zero point on $(0,\,1)$, so it be verified. For $k=2$, we know that $u_2$ and $u_1$ are orthogonal and $u_1$ is positive on $(0,\,1)$, then $u_2$ has at least one zero. Meanwhile, $u_2$ has at most two node domains according to $\rm (iv)$ of theorem 4.4 and 4.5, hence $u_2$ has exactly one node in $(0,\,1)$.
For $k\ge 3$, consider
we observe that $V(x)-\lambda _2>V(x)-\lambda _3$, by invoking lemma 4.6, it can be obtained that $u_3$ has at least two zeros. Combined with $\rm (iv)$ of theorem 4.4, it is apparent that $u_3$ has at most three node domains, therefore $u_3$ must have two zeros in $(0,\,1)$. The rest can be deduced by this method, so that we get $u_k$ of equation (1.1) has exactly $k-1$ zeroes in $(0,\,1)$, $k=1,\,2\cdots$.
The above results have introduced the explicit feature of eigenvalues and corresponding eigenfunctions, then we will calculate the differential of eigenvalues from the perspective of partial differential equation, which is the key step to derive the nature of the optimal function.
Theorem 4.8 Consider the following problem:
with a parameter $t \in (a,\,b)$, $h \in C^{1}[0,\,1]$ with $h(0)=0$ and $h(x)>0$ for $x\in (0,\,1]$. And $V(t,\,x): (a,\,b)\times [0,\,1]\rightarrow \mathbb {R}$ and $m\leq V(t,\,x)\leq M$ a.e., suppose that $\frac {\partial V}{\partial t}(x,\,t)$ exists for a.e. $x\in [0,\,1]$ and for all $t\in (a,\,b)$.
Let $u_i(t,\,x)$ be the $i$-th normalized eigenfunction concerned with $V(t,\,x)$, then the derivative of the $i$-th eigenvalue ${\lambda _i}(t)$ in relation to $t$ is
Proof. For the sake of simplicity, let us divide the discussion into several parts.
$1^{\circ }$. Let $u_i(t):=u_i(t,\,x)$ $(i=1,\,2)$ denote the $i$-th normalized eigenfunction for problem (4.16), and $u_i^{M}\ (i=1,\,2)$ denote the $i$-th normalized eigenfunction associated with $V(t,\,x)=M$ in (4.16). Taking $t\in (a,\,b)$, by definition in theorem 4.4
likewise, we easily check that $\lambda _2$ has the same properties, that is $\lambda _i(t)\leq C,\, \forall t\in (a,\,b),\, i=1,\,2$. Furthermore, we obtain
so that
$2^{\circ }$. Taking $\{s_n\}_{n\in \mathbb {N}}$ with $s_n\rightarrow t$, there exists a subsequence of $\{s_n\}$ by $1^{\circ }$, still denoted by itself, and $\lambda _i^{*}=\liminf \limits _{n\rightarrow \infty }\lambda _i(s_n)\in \mathbb {R}$ such that $\lambda _i(s_n)\rightarrow \lambda _i^{*},\, i=1,\,2$. In addition, we can further extract a subsequence of $\{s_n\}_{n\in\mathbb{N}}$ from (4.17), still denoted by itself, and $u_i^{*}\in H_c^{1}(0,\,1;h)$ such that
by employing lemma 3.3
at the limit, the constraint
is preserved. Thanks to (4.19), there exists a subsequence of $\{u_i(s_n)\}_{n\in \mathbb{N}}\subset L^{2}(0,\,1)$, still denoted by itself, such that
Note that $u_i(s_n)$ is the solution of (4.16) with respect to $V(s_n)$, owing to the fact that $V(s_n)\rightarrow V(t)$ strongly in $L^{\infty }(0,\,1)$ as $n\rightarrow \infty$, by applying (4.18) and (4.19), elementary computations lead to
Clearly, $u_i^{*}$ is a nonzero eigenfunction related to $\lambda _i^{*}$ whence (4.20) and (4.22), $i=1,\,2$. Given that $u_1(s_n)$ is the first positive eigenfunction for $V(s_n)$, so $u_1^{*}> 0$ a.e. by (4.21), it is no doubt that $u_1^{*}$ must change sign except $\lambda _1^{*}=\lambda _1(t)$, for this reason we must have
Besides, since $u_2(s_n)$ is the second eigenfunction so it must change sign once by theorem 4.7, and consider that (4.21) and $(u_2^{*},\,u_1^{*})=0$, the eigenfunction $u_2^{*}$ exactly has one node on $(0,\,1)$, which shows that $u_2^{*}$ is the second eigenfunction for $V^{*}$ according to theorem 4.7, so that $\lambda _2^{*}=\lambda _2(t)$ and $u_2^{*}=u_2(t)$. In view of this, we conclude that $\lambda _i(s)\rightarrow \lambda _i(t)$ as $s\rightarrow t,\, i=1,\,2$.
Now, we shall show $u_i(s)\rightarrow u_i(t)$ in $L^{2}(0,\,1),\, i=1,\,2$. Indeed, if not, there exists a sequence $\{s_n\}_{n\in \mathbb {N}}\subset (a,\,b)$ and $\epsilon _0>0$ such that $s_n\rightarrow t$ and
Follow the inequality (4.17), there exists a subsequence of $\{s_n\}_{n\in \mathbb {N}}$, still denoted by itself, and $\hat u_1\in H_c^{1}(0,\,1;h)$ such that $u_1(s_n)$ weakly converges to $\hat u_1$ in $H_c^{1}(0,\,1;h)$ and $u_1(s_n)$ strongly converges to $\hat u_1$ in $L^{2}(0,\,1)$ as $\|\hat u_1\|_{L^{2}(0,1)}=1$. Hence, we can extract a subsequence of $\{u_1(s_n)\}_{n\in \mathbb{N}}\subset L^{2}(0,\,1)$, still denoted by itself, such that $u_1(s_n)\rightarrow \hat u_1 \mbox { a.e.}$, we deduce that $\|\hat u_1-u_1(t)\|_{L^{2}(0,1)}\geq \epsilon _0$ whence (4.24). However, in the same manner as above (see (4.23)), the result $\hat u_1=u_1(t)$ is obtained, which is absurd. The same argument is applied to $u_2(t)$ again, the desired result is then received.
$3^{\circ }$. We observe that
thereby
i.e.,
Multiplying both sides by $u_i(s)-u_i(t)$ and then integrating on $(0,\,1)$, we see that
utilizing the Hölder inequality, we attain further results
This suggests that if $u_i(s)\rightarrow u_i(t)$ in $L^{2}(0,\,1)$, then $u_i(s)\rightarrow u_i(t)$ in $H_c^{1}(0,\,1;h)$, $i=1,\,2$.
$4^{\circ }$. Since
naturally,
for $i=1,\,2$. Combine this result with $4^{\circ }$, in addition $V(s)\rightarrow V(t)$ as $s\rightarrow t$, this suggests that when $u_i(s)\rightarrow u_i(t)$ in $L^{2}(0,\,1)$, we have $\lambda _i(s)\rightarrow \lambda _i(t)$, $i=1,\,2$.
$5^{\circ }$. We claim that $\lambda _i(s)\rightarrow \lambda _i(t)$ as $s\rightarrow t$, then $u_i(s)\rightarrow u_i(t)$ in $L^{2}(0,\,1)$ as $s\rightarrow t$, $i=1,\,2$. If not, suppose that there exist $\epsilon >0$ and a sequence $\{u_i(s_n)\}_{n\in \mathbb {N}}$ such that $s_n\rightarrow t$ but $\|u_i(s_n)-u_i(t)\|_{L^{2}(0,1)}\geq \epsilon$. In the same way as $3^{\circ }$, there exists $\hat u_i\in H_c^{1}(0,\,1;h)$ such that $u_i(s_n)$ weakly converges to $\hat u_i$ in $H_c^{1}(0,\,1;h)$, $u_i(s_n)$ strongly converges to $\hat u_i$ in $L^{2}(0,\,1)$ with $\|\hat u_i\|_{L^{2}(0,1)}=1$ and $u_i(s_n)\to \hat {u_i}$ a.e., $i=1,\,2$. Then we obtain $\|\hat {u_i}-u_i(t)\|_{L^{2}(0,1)}\geq \epsilon$.
We see that $u_1(s_n)-u_1(t)\in H_c^{1}(0,\,1;h)$ and
Similarly, $u_2(s_n)-u_2(t) \in H_c^{1}(0,\,1;h)$ and
If $(\hat u_2-u_2(t),\,u_1(t)) \neq 0$, we must have
and if $(\hat u_2-u_2(t),\,u_1(t)) = 0$, then
For $i=1,\,2$, due to $\lambda _i(t)$ is simple, the equality of (4.25) and (4.27) holds if and only if $\hat u_i-u_i(t)=cu_i(t)$, thus $\hat u_1=(c+1)u_1(t)$, the normalization condition guarantees $c=0$ or $c=-2$, since $1=\|\hat u_i\|_{L^{2}(0,1)}=|c+1|\|u_i(t)\|_{L^{2}(0,1)}=|c+1|$. Under the condition $\|\hat u_i-u_i(t)\|_{L^{2}(0,1)}\geq \epsilon$, the possibility of $c= 0$ is excluded, $i=1,\,2$. And simultaneously we also analyse that $c\neq -2$ for $i=1$ due to $\hat u_1\geq 0$ a.e. Without loss of generality, we may choose the sign of the second eigenfunction to be positive to the left of the node and to be negative in the other side. It is easily found that $\hat {u_2}=-u_2(t)$ for $c=-2$, it means that $\hat {u_2}$ and $u_2(t)$ have the same node, so this case can not happen. All these facts have demonstrated that the inequality of (4.25) and (4.27) holds strictly.
Formula from $4^{\circ }$,
let $s_n \to t$, we check that the right side of above equation converges to $0$, but the limit on the left is nonzero by virtue of the above analysis, which causes a contradiction. So far, the desired result is proved.
$6^{\circ }$. By $3^{\circ }$,
multiplying $u_i(t)$ on the both sides of this equality and integrating on $(0,\,1)$, then yields
i.e.
Also, given that
multiplying $\frac {{u_i(s)-u_i(t)}}{s-t}$ on both sides, after integrating on $(0,\,1)$ yields
further,
whence (4.28) and (4.29). Concerning that $u_i(s)\rightarrow u_i(t)$ in $L^{2}(0,\,1)$ as $s\rightarrow t$, from the dominated convergence theorem we have
Remark 4.9 The results of the above theorem 4.8 can be extended to the case of $i>2$, the approach used is similar, which we will not repeat here.
The classical path to explore the minimum problem is to obtain the optimality condition, denote $\Gamma (V)=\lambda _2(V)-\lambda _1(V)$. Consider the set $\mathcal {V}$, recall that $P(x)$ is called admissible perturbation function associated to $V(x)$, provided $V(x)+tP(x) \in \mathcal {V}$ for any small $|t|$, where $P(x)$ is measurable bounded and real valued function. Sometimes we will say that only nonpositive $t$ or non-negative $t$ is allowed, but if there is no explicit limit, $t$ can have any sign. Now and then we consider $V(t,\,x)=V(x)+tP(x)$, theorem 4.8 guarantees that
where $u_i(x)$ be the $i$-th normalized eigenfunction for $V(x)$, $i=1,\,2$.
5. Fundamental gap
In this section we commence by exploring the existence of the minimum fundamental gap when $V(x)$ is limited to the set $\mathcal {V}$, and then we further characterize the optimal function $V^{*}(x)$ and describe its behavior.
Theorem 5.1 The fundamental gap $\lambda _2(V)-\lambda _1(V)$ reaches its minimum on the set $\mathcal {V}$.
Proof. We shall prove the existence of the minimizer $V^{*}$ such that $\Gamma ^{*}=\Gamma (V^{*})=\inf \limits _{V\in \mathcal {V}} (\lambda _2(V)-\lambda _1(V)).$
Let $\{V^{k}\}_{k\in \mathbb {N}}$ such that
In view of the compactness of the class $\mathcal {V}$, then there exists a subsequence $\{V^{k}\}_{k\in \mathbb {N}}\subset L^{\infty }(0,\,1)$ such that
and
Let $\{(\lambda _j^{k},\,u_j^{k})\}_{k\in \mathbb {N}}$ be a sequence concerned with $V^{k}$, $\lambda _j^{k}$ denotes the $j$-th eigenvalue of (1.1), $u_j^{k}$ be the normalized eigenfunction in $H_c^{1}(0,\,1;h)$, $j=1,\,2$. By the definition of eigenvalue, we realize that
so, in like manner, we obtain that $\lambda _2^{k}\le C$. Further,
we have that $u_j^{k}$ is bounded in $H_c^{1}(0,\,1;h)$, and hence by the compactness of the embedding in lemma 3.3, a subsequence, which we take as $\{u_j^{k}\}_{k\in\mathbb{N}}$ itself, such that
and
Hence, there exists a subsequence, still take as $\{u_j^{k}\}_{k\in\mathbb{N}}$ itself, such that
For all $v\in H_c^{1}(0,\,1;h)$, seeing that
let $k\rightarrow \infty$, then by (5.2) (5.3) and (5.5) (5.6)
This implies that $\lambda _j^{*}$ be the element of the spectrum for $V^{*}$, $j=1,\,2$. Consider that $u_1^{k}$ is the first eigenfunction involved with $\lambda _1^{k}$, and the ground states are characterized as the positive eigenfunctions on $(0,\,1)$, this fact determines that $u_1^{*}> 0$ on $(0,\,1)$ by (5.7), so that we must have $\lambda _1^{*}=\lambda _1(V^{*})$. In a similar fashion, we see that $u_2^{k}$ change once sign in $(0,\,1)$, so that $u_2^{*}$ change once sign by (5.7) and $(u_2^{*},\,u_1^{*})=0$, utilizing theorem 4.7, we have $u_2^{*}$ is the second eigenfunction for $V^{*}$ and $\lambda _2^{*}=\lambda _2(V^{*})$. All this suggests that $\Gamma (V^{*})=\lambda _2(V^{*})-\lambda _1(V^{*})$.
Theorem 5.2 Consider the differential equation (1.1) for $V(x) \in \mathcal {V},$ then the minimum fundamental gap $\lambda _2-\lambda _1$ is attained by
where $\omega =\{x\in (0,\,1)\mid |u_2^{*}(x)|^{2}-|u_1^{*}(x)|^{2}\ge 0\},$ $u_i^{*}(x)$ be the $i$-th normalized eigenfunction associated to $V^{*},$ $i=1,\,2$. Furthermore, $|\omega |>0$ and $|\omega ^{c}|>0$.
Proof. In order to prove (5.9), we demand display that the set $T=\{x\in (0,\,1)\mid m< V^{*}(x)< M\}$ is zero measure. If not, let $T^{k}=\{x\in [0,\,1] \mid m+\frac {1}{k}< V^{*}(x)< M-\frac {1}{k}\}$, then we may write $T=\bigcup \limits _{k = 1}^{\infty } T^{k}$, we assume that at least one of them is a positive measure for $T^{k}$. For any $x^{*}\in T^{k}$ and any measurable sequence of subsets $G_{k,j}\subset T^{k}$ including $x^{*}$, the optimal condition of the fundamental gap extremum problem determines
where $V(x,\,t)=V(x)+tP(x)$ and $P(x)=\chi _{G_{k,j}}$ is an admissible perturbation for any small $t\in (-\tfrac {1}{k},\,\tfrac {1}{k})$. We claim that $|u_2^{*}(x)|^{2} = |u_1^{*}(x)|^{2}$ on $T^{k}$ for any $k>1$, so that $|u_2^{*}(x)|^{2} = |u_1^{*}(x)|^{2}$ on the set $T$. Indeed, if there is a subset $G \subset T^{k}$ of positive measure such that $|u_2^{*}(x)|^{2} - |u_1^{*}(x)|^{2}>0$, $P(x) = {\chi _G}$ is an admissible perturbation, then we have
however, this is in contradiction with the fact that $V^{*}(x)$ is the optimal function. Likewise same procedure is carried out, we can obtain that there is no positive measure subset on $T$ with $|u_2^{*}(x)|^{2} - |u_1^{*}(x)|^{2}<0$.
Actually, the case $|u_2^{*}(x)|^{2} = |u_1^{*}(x)|^{2}$ only works on the set with zero measure. Without loss of generality, suppose $T^{+} = \{ x \in T \mid {u_2^{*}(x)} > 0\}$ with positive measure, concerning that $u_1^{*}$ has no zero point on $(0,\,1)$, as a result, $T = {T^ + } \cup {T^ - }$, where ${T^ - } = \{ x \in T \mid {u_2^{*}(x)} < 0\}$. On $T^{+}$, seeing that $u_1^{*} -u_2^{*}=0$ a.e., thus $(u_2^{*}-u_1^{*})'=0$ a.e., from equation (1.1) we can infer that
From the proof of theorem 4.5, we know that $u_1^{*},\,u_2^{*}\in H_{loc}^{2}(0,\,1)$ and $u_1^{*},\,u_2^{*}\in C^{1,\frac {1}{2}}(\delta,\,1)$ for any $\delta \in (0,\,1)$. Let $F(x)=(h^{2}(u_2^{*}-u_1^{*})')'+V(u_1^{*}-u_2^{*})-(\lambda _1u_1^{*}-\lambda _2u_2^{*})$, we obtain $F(x)=0$ a.e. on $T^{+}$ in the sense of distribution. The left side of equation (5.10) vanishes on ${T^ + }$ a.e., it directly caused $\lambda _1=\lambda _2$, a contradiction, so that $T$ is a set of zero measure. This ensures that the optimal function is $V^{*} = m{\chi _{\omega }} + M{\chi _{\omega ^{c}}}$ for some $\omega$.
In this regard, we have to study further, there is no doubt that $\chi _{\omega }$ is an admissible perturbation for any small $t\ge 0$ when $V^{*}(x)=m$, theorem 4.8 guarantees
for this we may come to a conclusion $|u_2^{*}(x)|^{2} - |u_1^{*}(x)|^{2}\ge 0$. If this is not true, there must be a set $S \subset \omega _0$ of positive measure such that $|u_2^{*}(x)|^{2} - |u_1^{*}(x)|^{2} < 0$ on $S$ and
for admissible perturbation ${\chi _S}$, $t\ge 0$, this obviously fail to meet the optimality of $V^{*}(x)$. By applying the same argument, the result $|u_2^{*}(x)|^{2} - |u_1^{*}(x)|^{2} \le 0$ as $V^{*}=M$ is straightforward.
Concerned with the normalization condition $\int _0^{1} |u_2^{*}(x)|^{2}-|u_1^{*}(x)|^{2}{\rm d}\kern0.06em x = 0$, which indicates that both the sets $\omega$ and $\omega ^{c}$ are nonempty.
Theorem 5.3 The set $\omega ^{c}$ mentioned in theorem 5.2 only contains one interval.
Proof. For any given function $V(x)$, the set $S_1$ and $S_1^{c}$ are nonempty in the light of normalization condition, here $S_1=\{x\in (0,\,1) \mid u_2(x)|^{2} - |u_1(x)|^{2}\ge 0 \}$, thereby $|u_2(x)|^{2} = |u_1(x)|^{2}$ has at least one solution. In reality, the number of solutions for this equation shall not exceed two, the following will be explained in detail.
Consider
and note that
so that
and
where $\alpha$ denotes the node of the eigenfunction $u_2(x)$. Without loss of generality, suppose $u_2(x)>0$ in $(0,\,\alpha )$ and $u_2(x)<0$ in $(\alpha,\,1)$, hence $h^{2}(x)g(x)<0$ on $(0,\,1)$ by (5.12) and (5.13), especially, $g(x)<0$ on $(0,\,1)$. Moreover, thanks to (5.11) we have
and
that is, $\frac {|u_2(x)|^{2}}{|u_1(x)|^{2}}$ is strictly monotonic decreasing on $(0,\,\alpha )$ and strictly monotonic increasing on $(\alpha,\,1)$.
Suppose there are two different points $x^{1},\, x^{2}\in (0,\,\alpha )$ such that $|u_2(x)|^{2} = |u_1(x)|^{2}$, which lead to $\frac {|u_2(x^{2})|^{2}}{|u_1(x^{2})|^{2}}=\frac {|u_2(x^{1})|^{2}}{|u_1(x^{1})|^{2}}$, this is opposite to the condition of strictly monotone for $\frac {|u_2(x)|^{2}}{|u_1(x)|^{2}}$ on $(0,\,\alpha )$, hence the equation $|u_1(x)|^{2} =|u_2(x)|^{2}$ has at most one solution on $(0,\,\alpha )$. Likewise, the same conclusion is achieved on $(\alpha,\,1)$. Consequently, there exist two points $x_0$, $x_1$, $0 \le x_0 < \alpha < x_1 \le 1$ with
and both sets are nonempty. Take into consideration that $u_1(x),\,u_2(x)$ are continuous on $(0,\,1)$, we must have $|u_2(x_0)|^{2}=|u_1(x_0)|^{2}$ and $|u_2(x_1)|^{2}=|u_1(x_1)|^{2}$.
Recognizing the facts $|u_2^{*}(x)|^{2} -|u_1^{*}(x)|^{2} \le 0$ when $V^{*}(x)=M$ according to theorem 5.2 and the inequality 5.14, we point out that the set $\omega ^{c}$ only contains one interval.
Appendix A. Harnack's inequality
Lemma A.1 Harnack's inequality
Let $0\leq f\in L^{2}(0,\,1),$ and $0\leq u\in H_c^{1}(0,\,1;h)$ be the solution of (2.1) with respect to $f$. Then there exists a constant $C>0$ such that
for any $\Omega '\subset \subset \overline {\Omega }\subset (0,\,1)$.
Proof. The proof method is similar to [Reference Gilbarg and Trudinger14], but for clarity, we show specific details.
Let $x_0\in (0,\,1),\, \overline {B(x_0,\,R)}\subset (0,\,1)$. Assume $0< r_1< r_2< R$, and $\zeta \in C_c^{\infty }(\mathbb {R}),\, 0\leq \zeta \leq 1,\, \zeta |_{B(x_0,r_1)}=1,\, \zeta |_{\mathbb {R}-B(x_0,\,r_2)}=0$, and $|\zeta '|\leq \frac {2}{r_2-r_1}$.
For each $\beta \neq 0$, taking $v=\zeta ^{2}\overline {u}^{\beta }$ with $\overline {u}=u+k,\, k>0$, then
From which we obtain that
i.e.
thus by Young inequality, we have
for any $\epsilon >0$. Choosing $\epsilon \le \frac {|\beta |}{2}$, thereby obtaining the inequality
Set
then $\theta '=\frac {\beta +1}{2}\overline {u}^{\frac {\beta -1}{2}}u'$ for $\beta \neq -1$, and
Consequently, we have
by the Sobolev inequality, where $C$ is independent of $|\beta |$. The interpolation inequality implies
hence, combine with (A.1),(A.2) and (A.3) and select the appropriate $\epsilon$, we see that
Let $\gamma =\beta +1$, note that $\zeta +|\zeta '|\leq 1+\frac {2}{r_2-r_1}\leq \frac {C}{r_2-r_1}$, then
For $\overline {B(x_0,\,R)}\subset (0,\,1)$ and $p \neq 0$, $r< R$, we introduce
then we have $\Phi (\infty,\,r)=\mathop {lim}\limits _{p \to \infty }\Phi (p,\,r)= \mathop {sup}\limits _{B(x_0,r)}{\bar u}$, $\Phi (-\infty,\,r)=\mathop {lim}\limits _{p \to -\infty }\Phi (p,\,r)= \mathop {inf}\limits _{B(x_0,r)}{\bar u}$ from [Reference Gilbarg and Trudinger14]. From (A.4), we easily check that
For $\beta >0$, then $\gamma >1$, taking $p>1$, set $\gamma =\gamma _m=3^{m-1}p$, $R_m=\rho _0+(\rho _1-\rho _0)^{m}$, where $\rho _0=r_1,\,\rho _1=\frac {r_2+r_1}{2}$, $m=1,\,2 \cdots$, so that, thanks to (A.5),
letting $m$ tend to $\infty$, then
For $\beta <0$ and $\gamma <1$, we may employ a similar method to prove it, for any $p_0$, $p$ such that $0< p_0< p<3$, $p> 1$, we have
and
where $\rho _2=r_2$. Indeed, when $0<\gamma <1$ and $p_0<1$, we may take $\gamma =p_0$, there exists $s$ such that $3^{s}\gamma \ge p>1$, by (A.5)
When $p_0\ge 1$, we may take $3\gamma =p_0$, there exists $t$ such that $3^{t}\gamma \ge p>1$, (A.7) is obtained by the same procedure. Consider the case of $\gamma <0$, let $\gamma _m=-3^{m-1}p_0$, $R_m=\rho _0+(\rho _2-\rho _0)^{m}$, $m=1,\,2,\,\cdots$, (A.8) can be obtained by utilizing equation (A.5) again.
To further explore this content, we shall further verify that
From the second of the estimates (A.1), with the aid of the Hölder inequality, the result
is obtained for any $r_1\in (0,\,r_2)$. Consequently, according to theorem 7.21 [Reference Gilbarg and Trudinger14], there exists a positive $p_0$ such that, for
we know
and hence
Then combined with the definition $\theta =\log \overline {u}$, the result (A.10) can be easily obtained. These results (A.7),(A.8) and (A.10) indicate that
i.e.
coupled with (A.6), we have $\sup \limits _{B(x_0,\rho _0)}\overline {u}\le C\inf \limits _{B(x_0,\rho _0)}\overline {u}$. Now let $\Omega ' \subset \subset \Omega$ and choose $x_1,\,x_2 \in \overline {\Omega '}$ such that $u(x_1)=\sup \limits _{\Omega '}u$ and $u(x_2)=\inf \limits _{\Omega '}$u. Take $\Gamma \subset \overline {\Omega '}$ be the line joining $x_1$ and $x_2$ and choose $R$ such that $4R< dist(\Gamma,\,\partial \Omega )$. Since $\Gamma$ can be covered by a finite number $N$ of balls of radius $R$, utilizing the above estimation on each ball and combining all inequalities, the desired result is obtained.
Acknowledgments
This work was supported by the National Natural Science Foundation of China, the Science-Technology Foundation of Hunan Province, and the Fundamental Research Funds for the Central Universities of Central South University (Grant No.2021zzts0046). The authors are grateful to editor for many useful comments on presentation. The constructive suggestions from anonymous referees are very helpful to improve the manuscript substantially.