On the Nazarov–Shcheglova Conjecture for Sharp Sobolev Inequalities: The Case $(n,p)=(3,2)$

Abstract.

For integers $n>k\geq 0$ and $1\leq p,q\leq\infty$, let $\lambda_{3}(n,k,p,q)$ denote the optimal constant in the one-dimensional Sobolev inequality

$$\|u^{(k)}\|_{L^{q}(0,1)}\leq\lambda_{3}(n,k,p,q)\|u^{(n)}\|_{L^{p}(0,1)},\qquad u\in\mathring{W}_{p}^{n}(0,1).$$

Nazarov and Shcheglova [16] conjectured that

$$\lambda_{3}(n,1,p,1)=2\lambda_{3}(n,0,p,\infty),\qquad n\geq 2,\quad 1\leq p\leq\infty,$$

and that the corresponding extremal functions coincide and are symmetric about the midpoint of the interval. We prove this conjecture for $(n,p)=(3,2)$ and, in particular, obtain

$$\lambda_{3}(3,1,2,1)=\frac{1}{32\sqrt{5}}.$$

The proof is based on a reduction to an operator norm problem with an additional moment constraint. After rescaling to $(-1,1)$, this constraint leads to orthogonality with respect to quadratic polynomials. We extend the inverse of the second derivative through the corresponding orthogonal projection and analyze the adjoint operator by an odd–even decomposition. The odd component is controlled by a positive Gram kernel, while the even component is treated by a weak-$L^{2}$ estimate. The equality cases yield the characterization of all extremal functions.

1. Introduction

The determination of sharp constants in Sobolev inequalities is a classical problem in analysis, closely connected with the identification and qualitative properties of the corresponding extremal functions. In the Euclidean setting, the fundamental results of Aubin and Talenti [1, 20] established the optimal constants in the classical Sobolev inequality and identified the corresponding extremal functions. For a broader treatment of Sobolev-type inequalities and their geometric and analytic aspects, we refer to [18]. Another important development was the mass-transport approach of Cordero-Erausquin, Nazaret, and Villani [5], which yields sharp Sobolev and Gagliardo–Nirenberg inequalities together with a characterization of the equality cases.

In the one-dimensional setting, sharp Sobolev-type inequalities have a long history and have been investigated from several points of view; see, for instance, [17, 10, 15, 3, 11, 12]. Beyond the determination of optimal constants, a central question concerns the structure and symmetry of extremal functions; see, for example, [2, 6, 21, 19, 4]. Classical examples include the Poincaré, Wirtinger, and Steklov-type inequalities, while higher-order Sobolev embeddings give rise to a considerably richer class of extremal problems; see, for instance, [7, 8, 9, 13]. For surveys and further historical references, we refer to [14, 16].

Adopting the notation of Nazarov and Shcheglova [16], for integers $n>k\geq 0$ and $1\leq p,q\leq\infty$, let $\lambda_{3}(n,k,p,q)$ denote the best constant in the higher-order Sobolev inequality

(1.1) $$\|u^{(k)}\|_{L^{q}(0,1)}\leq\lambda_{3}(n,k,p,q)\,\|u^{(n)}\|_{L^{p}(0,1)},\qquad u\in\mathring{W}_{p}^{n}(0,1).$$

Here

$$\mathring{W}_{p}^{n}(0,1)=\left\{u\in W_{p}^{n}(0,1):u^{(j)}(0)=u^{(j)}(1)=0,\quad j=0,\ldots,n-1\right\}.$$

Exact values of $\lambda_{3}(n,k,p,q)$ and the structure of the corresponding extremal functions are known only for restricted ranges of the parameters; see [16] and the references therein.

Among the known results, the endpoint regimes $q=\infty$ and $q=1$ are particularly relevant to the present work. For $q=\infty$, sharp constants and the corresponding extremal functions have been studied in a number of cases; see, for example, [7, 8, 12, 13]. In particular, Watanabe, Kametaka, Nagai, Yamagishi, and Takemura [21] determined $\lambda_{3}(3,0,p,\infty)$ for $1<p<\infty$ and established the symmetry of the corresponding extremal functions with respect to $x=\frac{1}{2}$. Together with the endpoint cases $p=1$ and $p=\infty$, this gives the determination of $\lambda_{3}(3,0,p,\infty)$ for every $1\leq p\leq\infty$; see [16] and the references therein. At the endpoint $p=1$, the right-hand side of (1.1) is understood in the measure-valued sense.

For the endpoint $q=1$, a recent higher-order result was obtained by Hindov, Nitzan, Olsen, and Rydhe [9]. They proved that

$$\lambda_{3}(n,0,2,1)=\frac{n!}{(2n)!\sqrt{2n+1}},$$

and showed that the corresponding extremal functions are, up to multiplication by a nonzero constant,

$$x^{n}(1-x)^{n}.$$

In [16], Nazarov and Shcheglova conjectured that

$$\lambda_{3}(n,1,p,1)=2\lambda_{3}(n,0,p,\infty),\qquad n\geq 2,\quad 1\leq p\leq\infty.$$

Moreover, the corresponding extremal functions coincide and are symmetric with respect to $x=\frac{1}{2}$. The conjecture is known to hold for $n=2$; see [16, Remark 4.15].

In the present paper, we prove the Nazarov–Shcheglova conjecture in the case $(n,p)=(3,2)$. For $p=2$, the result of Watanabe et al. [21] gives

(1.2) $$\lambda_{3}(3,0,2,\infty)=\frac{1}{64\sqrt{5}}.$$

Thus, in this case, the conjecture reduces to proving that

$$\lambda_{3}(3,1,2,1)=\frac{1}{32\sqrt{5}},$$

together with the asserted coincidence and symmetry of the corresponding extremal functions.

Our main result is the following.

Theorem 1.1.

The sharp constant in (1.1) for

$$(n,k,p,q)=(3,1,2,1)$$

is given by

(1.3) $$\lambda_{3}(3,1,2,1)=\frac{1}{32\sqrt{5}}.$$

Moreover, the families of extremal functions for $\lambda_{3}(3,1,2,1)$ and $\lambda_{3}(3,0,2,\infty)$ coincide up to multiplication by a nonzero constant, and every such extremal function is symmetric about $x=\frac{1}{2}$.

Our proof is inspired in part by the left-inverse approach developed by Hindov, Nitzan, Olsen, and Rydhe [9] for the sharp embedding of $W_{0}^{k,2}(-1,1)$ into $L^{1}(-1,1)$. In the present problem, however, the reduction $v=u^{\prime}$ introduces the additional moment condition

$$\int_{0}^{1}v(x)\,dx=0.$$

After rescaling to $(-1,1)$ by

$$V(t)=v\left(\frac{t+1}{2}\right),\qquad-1<t<1,$$

the boundary and moment conditions imply

$$V^{\prime\prime}\perp\mathcal{P}_{2},\qquad\mathcal{P}_{2}=\operatorname{span}\{1,t,t^{2}\}.$$

Indeed, the boundary conditions yield the orthogonality of $V^{\prime\prime}$ to $1$ and $t$, while the zero-moment condition yields the orthogonality to $t^{2}$. This additional constraint prevents the fixed-sign kernel argument used in [9] from applying directly.

To overcome this difficulty, we extend the inverse of the second derivative from $\mathcal{P}_{2}^{\perp}$ to $L^{2}(-1,1)$ by composing it with the orthogonal projection onto $\mathcal{P}_{2}^{\perp}$. We then analyze the adjoint of the resulting operator through an odd–even decomposition. The odd component is controlled by an explicit positive kernel, whereas the even component requires a weak-$L^{2}$ estimate. This yields both the sharp constant and the characterization of the extremal functions.

The rest of the paper is organized as follows. In Section 2, we develop the reductions, construct the relevant inverse operator, and establish the kernel estimates needed in the sequel. In Section 3, we prove Theorem 1.1.

2. Preliminaries

2.1. Notation

We first fix some notation. Throughout the paper, $\operatorname{sgn}$ denotes the sign function,

$$\operatorname{sgn}(x)=\begin{cases}1,&x>0,\\ 0,&x=0,\\ -1,&x<0.\end{cases}$$

For $s\in\mathbb{R}$, let

$$s_{+}=\max\{s,0\}.$$

For each $x\in(-1,1)$, define

(2.1) $$\phi_{x}(t)=(x-t)_{+},\qquad-1<t<1.$$

We denote by

$$\langle f,h\rangle_{L^{2}(-1,1)}=\int_{-1}^{1}f(t)h(t)\,dt$$

the inner product in $L^{2}(-1,1)$.

We introduce the space

(2.2) $$X=\left\{V\in H_{0}^{2}(-1,1):\int_{-1}^{1}V(t)\,dt=0\right\},$$

and set

(2.3) $$C_{*}=\sup_{V\in X\setminus\{0\}}\frac{\|V\|_{L^{1}(-1,1)}}{\|V^{\prime\prime}\|_{L^{2}(-1,1)}}.$$

We set

$$\mathcal{P}_{2}=\operatorname{span}\{1,t,t^{2}\},$$

the space of real polynomials of degree at most two, viewed as a subspace of $L^{2}(-1,1)$. We denote by

$$\Pi_{2}:L^{2}(-1,1)\longrightarrow\mathcal{P}_{2}$$

the orthogonal projection onto $\mathcal{P}_{2}$, and by $I$ the identity operator on $L^{2}(-1,1)$.

For each $x\in(-1,1)$, define

(2.4) $$r_{x}=(I-\Pi_{2})\phi_{x}\in\mathcal{P}_{2}^{\perp},$$

where $\mathcal{P}_{2}^{\perp}$ denotes the orthogonal complement of $\mathcal{P}_{2}$ in $L^{2}(-1,1)$.

For $f\in L^{2}(-1,1)$, set

(2.5) $$(Af)(x)=\int_{-1}^{1}r_{x}(t)f(t)\,dt,\qquad-1<x<1.$$

Lemma 2.1.

The operator

$$A:L^{2}(-1,1)\longrightarrow L^{1}(-1,1)$$

defined by (2.5) is linear and bounded.

Proof.

Linearity is immediate. Since $I-\Pi_{2}$ is the orthogonal projection onto $\mathcal{P}_{2}^{\perp}$,

(2.6) $$\|r_{x}\|_{L^{2}(-1,1)}\leq\|\phi_{x}\|_{L^{2}(-1,1)}=\left(\frac{(x+1)^{3}}{3}\right)^{1/2}\leq\sqrt{\frac{8}{3}},\qquad-1<x<1.$$

Hence, by the Cauchy–Schwarz inequality,

$$|(Af)(x)|\leq\sqrt{\frac{8}{3}}\,\|f\|_{L^{2}(-1,1)},\qquad-1<x<1.$$

Therefore,

$$\|Af\|_{L^{1}(-1,1)}\leq 2\sqrt{\frac{8}{3}}\,\|f\|_{L^{2}(-1,1)},$$

and the result follows. ∎

We denote by

$$A^{*}:L^{\infty}(-1,1)\longrightarrow L^{2}(-1,1)$$

the Banach adjoint of $A$, where $L^{2}(-1,1)$ is identified isometrically with its dual in the usual way. By (2.6) and Fubini’s theorem, for every $g\in L^{\infty}(-1,1)$,

(2.7) $$(A^{*}g)(y)=\int_{-1}^{1}g(x)r_{x}(y)\,dx$$

for almost every $y\in(-1,1)$.

We denote the corresponding operator norms by

(2.8) $$\|A\|_{2\to 1} =\sup_{f\in L^{2}(-1,1)\setminus\{0\}}\frac{\|Af\|_{L^{1}(-1,1)}}{\|f\|_{L^{2}(-1,1)}},$$

$$\|A^{*}\|_{\infty\to 2} =\sup_{g\in L^{\infty}(-1,1)\setminus\{0\}}\frac{\|A^{*}g\|_{L^{2}(-1,1)}}{\|g\|_{L^{\infty}(-1,1)}}.$$

We introduce the reflection operators

(2.9) $$R_{2}:L^{2}(-1,1)\longrightarrow L^{2}(-1,1),\qquad(R_{2}f)(t)=f(-t),$$

and

(2.10) $$R_{\infty}:L^{\infty}(-1,1)\longrightarrow L^{\infty}(-1,1),\qquad(R_{\infty}g)(t)=g(-t),$$

with the identities understood almost everywhere on $(-1,1)$.

Finally, define

(2.11) $$D^{2}:X\longrightarrow L^{2}(-1,1),\qquad D^{2}V=V^{\prime\prime}.$$

2.2. Reduction to an operator norm problem

We first reduce the computation of $\lambda_{3}(3,1,2,1)$ to the determination of the operator norm of $A$.

Lemma 2.2.

The following assertions hold.

1. The operator $D^{2}$ defined by (2.11) maps $X$ bijectively onto $\mathcal{P}_{2}^{\perp}$, and

   $$\left.A\right|_{\mathcal{P}_{2}^{\perp}}=(D^{2})^{-1}.$$

2. The constant $C_{*}$ defined by (2.3) satisfies

   $$C_{*}=\|A\|_{2\to 1}=\|A^{*}\|_{\infty\to 2}.$$

3. The map

   $$T:\mathring{W}_{2}^{3}(0,1)\longrightarrow X,\qquad(Tu)(t)=u^{\prime}\left(\frac{t+1}{2}\right),$$

   is a linear bijection, and

   (2.12) $$\frac{\|u^{\prime}\|_{L^{1}(0,1)}}{\|u^{\prime\prime\prime}\|_{L^{2}(0,1)}}=\frac{1}{4\sqrt{2}}\frac{\|Tu\|_{L^{1}(-1,1)}}{\|(Tu)^{\prime\prime}\|_{L^{2}(-1,1)}},\qquad u\in\mathring{W}_{2}^{3}(0,1)\setminus\{0\}.$$

4. Consequently,

   $$\lambda_{3}(3,1,2,1)=\frac{C_{*}}{4\sqrt{2}},$$

   and $u$ is an extremal function for $\lambda_{3}(3,1,2,1)$ if and only if $Tu$ is an extremal function for (2.3).

Proof.

(i) Let $V\in X$. By (2.2),

$$V(-1)=V(1)=V^{\prime}(-1)=V^{\prime}(1)=0$$

and

$$\int_{-1}^{1}V(t)\,dt=0.$$

Integration by parts gives

$$\int_{-1}^{1}V^{\prime\prime}(t)\,dt=0,\qquad\int_{-1}^{1}tV^{\prime\prime}(t)\,dt=0,\qquad\int_{-1}^{1}t^{2}V^{\prime\prime}(t)\,dt=2\int_{-1}^{1}V(t)\,dt=0.$$

Hence

$$V^{\prime\prime}\in\mathcal{P}_{2}^{\perp},$$

and therefore

$$D^{2}(X)\subset\mathcal{P}_{2}^{\perp}.$$

Conversely, let $f\in\mathcal{P}_{2}^{\perp}$ and define

$$V(t)=\int_{-1}^{t}(t-s)f(s)\,ds,\qquad-1\leq t\leq 1.$$

Then $V\in H^{2}(-1,1)$,

$$V^{\prime}(t)=\int_{-1}^{t}f(s)\,ds,\qquad V^{\prime\prime}(t)=f(t)\quad\text{a.e. in }(-1,1),$$

and

$$V(-1)=V^{\prime}(-1)=0.$$

Since $f\perp\mathcal{P}_{2}$,

$$V^{\prime}(1)=\int_{-1}^{1}f(s)\,ds=0,\qquad V(1)=\int_{-1}^{1}(1-s)f(s)\,ds=0.$$

Thus

$$V\in H_{0}^{2}(-1,1),\qquad D^{2}V=f.$$

Moreover, by Fubini’s theorem,

$$\int_{-1}^{1}V(t)\,dt =\int_{-1}^{1}\int_{-1}^{t}(t-s)f(s)\,ds\,dt$$

$$=\frac{1}{2}\int_{-1}^{1}(1-s)^{2}f(s)\,ds$$

$$=0,$$

since $(1-s)^{2}\in\mathcal{P}_{2}$. Hence $V\in X$.

If $V\in X$ and $D^{2}V=0$, then $V$ is affine, and the boundary conditions in (2.2) imply $V\equiv 0$. Thus $D^{2}$ is injective. Consequently,

$$D^{2}:X\longrightarrow\mathcal{P}_{2}^{\perp}$$

is bijective.

Finally, let $f\in\mathcal{P}_{2}^{\perp}$. By (2.4) and (2.5),

$$(Af)(x) =\int_{-1}^{1}r_{x}(t)f(t)\,dt$$

$$=\int_{-1}^{1}\phi_{x}(t)f(t)\,dt$$

$$=\int_{-1}^{x}(x-t)f(t)\,dt,$$

since $\Pi_{2}\phi_{x}\in\mathcal{P}_{2}$ and $f\perp\mathcal{P}_{2}$. Comparing with the preceding construction, we obtain

$$Af=(D^{2})^{-1}f.$$

Therefore,

$$\left.A\right|_{\mathcal{P}_{2}^{\perp}}=(D^{2})^{-1}.$$

(ii) By (2.3) and (i),

$$C_{*}=\sup_{f\in\mathcal{P}_{2}^{\perp}\setminus\{0\}}\frac{\|Af\|_{L^{1}(-1,1)}}{\|f\|_{L^{2}(-1,1)}}.$$

Moreover, by (2.4) and (2.5),

$$Af=A(I-\Pi_{2})f,\qquad f\in L^{2}(-1,1).$$

Hence

$$\|Af\|_{L^{1}(-1,1)}\leq C_{*}\|(I-\Pi_{2})f\|_{L^{2}(-1,1)}\leq C_{*}\|f\|_{L^{2}(-1,1)},$$

and therefore

$$\|A\|_{2\to 1}\leq C_{*}.$$

Conversely, since $\mathcal{P}_{2}^{\perp}\subset L^{2}(-1,1)$,

$$C_{*}\leq\sup_{f\in L^{2}(-1,1)\setminus\{0\}}\frac{\|Af\|_{L^{1}(-1,1)}}{\|f\|_{L^{2}(-1,1)}}=\|A\|_{2\to 1}.$$

Thus

$$C_{*}=\|A\|_{2\to 1}.$$

Finally, the equality of the norms of a bounded operator and its Banach adjoint gives

$$\|A\|_{2\to 1}=\|A^{*}\|_{\infty\to 2}.$$

This proves (ii).

(iii) Let $u\in\mathring{W}_{2}^{3}(0,1)$ and set

$$V=Tu,\qquad V(t)=u^{\prime}\left(\frac{t+1}{2}\right).$$

Then $V\in H^{2}(-1,1)$ and

$$V^{\prime}(t)=\frac{1}{2}u^{\prime\prime}\left(\frac{t+1}{2}\right).$$

Since

$$u^{\prime}(0)=u^{\prime}(1)=u^{\prime\prime}(0)=u^{\prime\prime}(1)=0,$$

we have

$$V(\pm 1)=V^{\prime}(\pm 1)=0,$$

and hence

$$V\in H_{0}^{2}(-1,1).$$

Moreover, by the change of variables $x=(t+1)/2$,

$$\int_{-1}^{1}V(t)\,dt=2\int_{0}^{1}u^{\prime}(x)\,dx=2\bigl(u(1)-u(0)\bigr)=0.$$

Therefore,

$$Tu=V\in X.$$

Conversely, let $V\in X$ and define

$$u(x)=\int_{0}^{x}V(2s-1)\,ds,\qquad 0\leq x\leq 1.$$

Then

$$u^{\prime}(x)=V(2x-1),\qquad u^{\prime\prime}(x)=2V^{\prime}(2x-1),\qquad u^{\prime\prime\prime}(x)=4V^{\prime\prime}(2x-1),$$

and hence $u\in W_{2}^{3}(0,1)$. Since $V\in H_{0}^{2}(-1,1)$,

$$u^{\prime}(0)=u^{\prime}(1)=u^{\prime\prime}(0)=u^{\prime\prime}(1)=0.$$

Moreover,

$$u(0)=0$$

and, by the defining moment condition of $X$,

$$u(1)=\int_{0}^{1}V(2s-1)\,ds=\frac{1}{2}\int_{-1}^{1}V(t)\,dt=0.$$

Hence

$$u\in\mathring{W}_{2}^{3}(0,1),$$

and

$$(Tu)(t)=u^{\prime}\left(\frac{t+1}{2}\right)=V(t).$$

Thus $T$ is surjective.

Suppose that $Tu=0$ for some $u\in\mathring{W}_{2}^{3}(0,1)$. Since $u^{\prime}$ is continuous,

$$u^{\prime}(x)=0,\qquad 0\leq x\leq 1.$$

Thus $u$ is constant, and since $u(0)=0$, we obtain $u=0$. Hence

$$\ker T=\{0\}.$$

Since $T$ is linear, it is injective. Consequently, $T$ is a bijection from $\mathring{W}_{2}^{3}(0,1)$ onto $X$.

Finally, for $u\in\mathring{W}_{2}^{3}(0,1)\setminus\{0\}$, a change of variables gives

$$\|Tu\|_{L^{1}(-1,1)}=2\|u^{\prime}\|_{L^{1}(0,1)}$$

and

$$\|(Tu)^{\prime\prime}\|_{L^{2}(-1,1)}=\frac{1}{2\sqrt{2}}\,\|u^{\prime\prime\prime}\|_{L^{2}(0,1)}.$$

This yields (2.12) and completes the proof of (iii).

(iv) By the definition of the optimal constant in (1.1),

$$\lambda_{3}(3,1,2,1)=\sup_{u\in\mathring{W}_{2}^{3}(0,1)\setminus\{0\}}\frac{\|u^{\prime}\|_{L^{1}(0,1)}}{\|u^{\prime\prime\prime}\|_{L^{2}(0,1)}}.$$

Since $T$ is a bijection from $\mathring{W}_{2}^{3}(0,1)$ onto $X$, taking the supremum in (2.12) gives

$$\lambda_{3}(3,1,2,1)=\frac{C_{*}}{4\sqrt{2}}.$$

Moreover, by (2.12) and the bijectivity of $T$, the left-hand quotient attains its supremum at $u$ if and only if the right-hand quotient attains its supremum at $Tu$. Hence $u$ is an extremal function for $\lambda_{3}(3,1,2,1)$ if and only if $Tu$ is an extremal function for (2.3). This proves (iv). ∎

2.3. Reflection symmetry and decomposition of the adjoint

We next exploit the reflection symmetry of the family $\{r_{x}\}_{x\in(-1,1)}$ and use it to decompose the adjoint operator $A^{*}$ into its odd and even components.

Lemma 2.3.

The reflection operator $R_{2}$ defined by (2.9) has the following properties.

1. The operator $R_{2}$ is an orthogonal involution on $L^{2}(-1,1)$. In particular,

   $$R_{2}^{*}=R_{2}^{-1}=R_{2},\qquad R_{2}^{2}=I,$$

   and

   $$\|R_{2}f\|_{L^{2}(-1,1)}=\|f\|_{L^{2}(-1,1)},\qquad f\in L^{2}(-1,1).$$

2. The subspaces $\mathcal{P}_{2}$ and $\mathcal{P}_{2}^{\perp}$ are invariant under $R_{2}$, and

   $$\Pi_{2}R_{2}=R_{2}\Pi_{2}.$$

3. The family $\{r_{x}\}_{x\in(-1,1)}$ defined by (2.4) satisfies

   $$r_{-x}=R_{2}r_{x},\qquad x\in(-1,1).$$

Proof.

(i) By (2.9),

$$R_{2}^{2}=I\qquad\text{and}\qquad\|R_{2}f\|_{L^{2}(-1,1)}=\|f\|_{L^{2}(-1,1)}$$

for every $f\in L^{2}(-1,1)$. Hence $R_{2}$ is an orthogonal involution and

$$R_{2}^{-1}=R_{2}.$$

Since an orthogonal operator satisfies $R_{2}^{*}=R_{2}^{-1}$, we obtain

$$R_{2}^{*}=R_{2}^{-1}=R_{2}.$$

(ii) Let $g\in\mathcal{P}_{2}$. Then

$$g(t)=a+bt+ct^{2}$$

for some $a,b,c\in\mathbb{R}$. By (2.9),

$$(R_{2}g)(t)=g(-t)=a-bt+ct^{2}\in\mathcal{P}_{2}.$$

Hence $\mathcal{P}_{2}$ is invariant under $R_{2}$.

Let now $h\in\mathcal{P}_{2}^{\perp}$ and $g\in\mathcal{P}_{2}$. By (i) and the invariance of $\mathcal{P}_{2}$,

$$\langle R_{2}h,g\rangle_{L^{2}(-1,1)}=\langle h,R_{2}g\rangle_{L^{2}(-1,1)}=0.$$

Thus

$$R_{2}h\in\mathcal{P}_{2}^{\perp},$$

so $\mathcal{P}_{2}^{\perp}$ is invariant under $R_{2}$.

Finally, for $f\in L^{2}(-1,1)$, write

$$f=\Pi_{2}f+(I-\Pi_{2})f.$$

Applying $R_{2}$, we obtain

$$R_{2}f=R_{2}\Pi_{2}f+R_{2}(I-\Pi_{2})f.$$

By the invariance of $\mathcal{P}_{2}$ and $\mathcal{P}_{2}^{\perp}$ under $R_{2}$,

$$R_{2}\Pi_{2}f\in\mathcal{P}_{2},\qquad R_{2}(I-\Pi_{2})f\in\mathcal{P}_{2}^{\perp}.$$

Hence, by uniqueness of the orthogonal decomposition,

$$\Pi_{2}(R_{2}f)=R_{2}(\Pi_{2}f),$$

and therefore

$$\Pi_{2}R_{2}=R_{2}\Pi_{2}.$$

(iii) Let $x\in(-1,1)$. By (2.4) and (ii),

$$R_{2}r_{x}=R_{2}(I-\Pi_{2})\phi_{x}=(I-\Pi_{2})R_{2}\phi_{x}.$$

By (2.9), for almost every $t\in(-1,1)$,

$$(R_{2}\phi_{x})(t)=\phi_{x}(-t)=(x+t)_{+}.$$

Moreover,

$$(R_{2}\phi_{x})(t)-\phi_{-x}(t)=(x+t)_{+}-(-x-t)_{+}=x+t.$$

Hence

$$R_{2}\phi_{x}-\phi_{-x}\in\mathcal{P}_{2}.$$

Applying $I-\Pi_{2}$, we obtain

$$(I-\Pi_{2})R_{2}\phi_{x}=(I-\Pi_{2})\phi_{-x}.$$

Therefore, by (2.4),

$$R_{2}r_{x}=r_{-x}.$$

This proves (iii). ∎

We next analyze the adjoint operator $A^{*}$ defined by (2.7).

For $g\in L^{\infty}(-1,1)$ and almost every $x\in(0,1)$, set

(2.13) $$a_{g}(x)=\frac{g(x)-g(-x)}{2},\qquad b_{g}(x)=\frac{g(x)+g(-x)}{2}.$$

Then $a_{g},b_{g}\in L^{\infty}(0,1)$.

Lemma 2.4.

The following assertions hold.

1. The adjoint operator $A^{*}$ intertwines the reflection operators $R_{\infty}$ and $R_{2}$ defined by (2.10) and (2.9), respectively:

   $$A^{*}R_{\infty}=R_{2}A^{*}.$$

2. For every $g\in L^{\infty}(-1,1)$,

   $$\|A^{*}g\|_{L^{2}(-1,1)}^{2}=\left\|\int_{0}^{1}a_{g}(x)\bigl(r_{x}-r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2}+\left\|\int_{0}^{1}b_{g}(x)\bigl(r_{x}+r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2},$$

   where the integrals are understood as Bochner integrals in $L^{2}(-1,1)$.

Proof.

(i) Let $g\in L^{\infty}(-1,1)$. By (2.7) and (2.10),

$$A^{*}R_{\infty}g=\int_{-1}^{1}g(-x)r_{x}\,dx=\int_{-1}^{1}g(x)r_{-x}\,dx,$$

where the second equality follows from the change of variable $x\mapsto-x$. By Lemma 2.3(iii),

$$r_{-x}=R_{2}r_{x},$$

and therefore

$$A^{*}R_{\infty}g=\int_{-1}^{1}g(x)R_{2}r_{x}\,dx=R_{2}\int_{-1}^{1}g(x)r_{x}\,dx=R_{2}A^{*}g.$$

Hence

$$A^{*}R_{\infty}=R_{2}A^{*}.$$

(ii) Let $g\in L^{\infty}(-1,1)$. By (2.13), for almost every $x\in(0,1)$,

$$g(x)=a_{g}(x)+b_{g}(x),\qquad g(-x)=b_{g}(x)-a_{g}(x).$$

Splitting the integral in (2.7) over $(-1,0)$ and $(0,1)$, and using the change of variable $x\mapsto-x$ on $(-1,0)$, we obtain, in $L^{2}(-1,1)$,

(2.14) $$A^{*}g=\int_{0}^{1}a_{g}(x)\bigl(r_{x}-r_{-x}\bigr)\,dx+\int_{0}^{1}b_{g}(x)\bigl(r_{x}+r_{-x}\bigr)\,dx.$$

By Lemma 2.3(i) and (iii),

$$R_{2}(r_{x}-r_{-x})=-(r_{x}-r_{-x}),\qquad R_{2}(r_{x}+r_{-x})=r_{x}+r_{-x}.$$

Since $R_{2}$ is bounded and linear, it commutes with the above Bochner integrals. Hence

$$\int_{0}^{1}a_{g}(x)\bigl(r_{x}-r_{-x}\bigr)\,dx$$

is odd, whereas

$$\int_{0}^{1}b_{g}(x)\bigl(r_{x}+r_{-x}\bigr)\,dx$$

is even. The two terms are therefore orthogonal in $L^{2}(-1,1)$. Using (2.14), we obtain

$$\|A^{*}g\|_{L^{2}(-1,1)}^{2}=\left\|\int_{0}^{1}a_{g}(x)\bigl(r_{x}-r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2}+\left\|\int_{0}^{1}b_{g}(x)\bigl(r_{x}+r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2}.$$

This proves (ii) and completes the proof. ∎

2.4. Kernel identities and estimates

For $x,y\in(-1,1)$, define the Gram kernel

(2.15) $$K(x,y)=\langle r_{x},r_{y}\rangle_{L^{2}(-1,1)}.$$

For $x,y\in(0,1)$, set

(2.16) $$K_{-}(x,y) =K(x,y)-K(x,-y),$$

$$K_{+}(x,y) =K(x,y)+K(x,-y).$$

By (2.6) and the Cauchy–Schwarz inequality,

$$K\in L^{\infty}((-1,1)^{2}),$$

and hence

$$K_{-},K_{+}\in L^{\infty}((0,1)^{2}).$$

We also define

(2.17) $$\begin{aligned} \kappa_{-}(x)&=\int_{0}^{1}K_{-}(x,y)\,dy,\\ \kappa_{+}(x)&=\frac{1}{\sqrt{2}}\|r_{x}+r_{-x}\|_{L^{2}(-1,1)},\end{aligned}\qquad 0<x<1.$$

Lemma 2.5.

The following assertions hold.

1. For all $x,y\in(-1,1)$,

   $$K(-x,-y)=K(x,y),\qquad K(-x,y)=K(x,-y).$$

   Consequently,

   $$K_{\pm}(x,y)=K_{\pm}(y,x),\qquad 0<x,y<1.$$

2. For $0<x\leq y<1$,

   $$K_{-}(x,y)=\frac{x(1-y)^{2}}{12}\left(3y-x^{2}(y+2)\right).$$

   In particular,

   $$K_{-}(x,y)>0,\qquad 0<x,y<1.$$

3. For $0<x<1$,

   $$\kappa_{-}(x)=\frac{x(1-x)^{2}(2x+1)}{48}.$$

4. For $0<x<1$,

   $$\kappa_{+}(x)^{2}=K_{+}(x,x)=\frac{(1-x)^{3}}{192}P(x),$$

   where

   (2.18) $$P(x)=15x^{5}+45x^{4}+30x^{3}-30x^{2}+3x+1.$$

Proof.

(i) Let $x,y\in(-1,1)$. By Lemma 2.3(i) and (iii),

$$K(-x,-y) =\langle r_{-x},r_{-y}\rangle_{L^{2}(-1,1)}$$

$$=\langle R_{2}r_{x},R_{2}r_{y}\rangle_{L^{2}(-1,1)}$$

$$=\langle r_{x},r_{y}\rangle_{L^{2}(-1,1)}$$

$$=K(x,y),$$

and

$$K(-x,y) =\langle r_{-x},r_{y}\rangle_{L^{2}(-1,1)}$$

$$=\langle R_{2}r_{x},r_{y}\rangle_{L^{2}(-1,1)}$$

$$=\langle r_{x},R_{2}r_{y}\rangle_{L^{2}(-1,1)}$$

$$=\langle r_{x},r_{-y}\rangle_{L^{2}(-1,1)}$$

$$=K(x,-y).$$

Hence, by the symmetry of the inner product and (2.16), for $x,y\in(0,1)$,

$$K_{\pm}(y,x) =K(y,x)\pm K(y,-x)$$

$$=K(x,y)\pm K(-x,y)$$

$$=K(x,y)\pm K(x,-y)$$

$$=K_{\pm}(x,y).$$

This proves (i).

(ii) For $z\in(0,1)$, set

$$d_{z}=r_{z}-r_{-z}.$$

By Lemma 2.3(iii),

(2.19) $$R_{2}d_{z}=R_{2}r_{z}-R_{2}r_{-z}=r_{-z}-r_{z}=-d_{z},\qquad 0<z<1.$$

Using also Lemma 2.3(i), for all $x,y\in(0,1)$,

$$\langle r_{-x},d_{y}\rangle_{L^{2}(-1,1)} =\langle R_{2}r_{x},d_{y}\rangle_{L^{2}(-1,1)}$$

$$=\langle r_{x},R_{2}d_{y}\rangle_{L^{2}(-1,1)}$$

$$=-\langle r_{x},d_{y}\rangle_{L^{2}(-1,1)}.$$

Hence

$$\langle d_{x},d_{y}\rangle_{L^{2}(-1,1)} =\langle r_{x}-r_{-x},d_{y}\rangle_{L^{2}(-1,1)}$$

$$=2\langle r_{x},d_{y}\rangle_{L^{2}(-1,1)}.$$

Consequently, by (2.16),

(2.20) $$K_{-}(x,y) =\langle r_{x},r_{y}-r_{-y}\rangle_{L^{2}(-1,1)}$$

$$=\langle r_{x},d_{y}\rangle_{L^{2}(-1,1)}$$

$$=\frac{1}{2}\langle d_{x},d_{y}\rangle_{L^{2}(-1,1)},\qquad 0<x,y<1.$$

We next compute $d_{z}$. Recalling (2.1), for $z\in(0,1)$ set

$$\psi_{z}(t)=\phi_{z}(t)-\phi_{-z}(t)-z=(z-t)_{+}-(-z-t)_{+}-z,\qquad-1<t<1.$$

Since the constant function $t\mapsto z$ belongs to $\mathcal{P}_{2}$, by (2.4),

(2.21) $$d_{z}=(I-\Pi_{2})\psi_{z}.$$

The function $\psi_{z}$ is odd and satisfies

(2.22) $$\psi_{z}(t)=\begin{cases}z,&-1<t<-z,\\ -t,&-z\leq t\leq z,\\ -z,&z<t<1.\end{cases}$$

By Lemma 2.3(ii),

$$R_{2}\Pi_{2}\psi_{z}=\Pi_{2}R_{2}\psi_{z}=-\Pi_{2}\psi_{z}.$$

Hence $\Pi_{2}\psi_{z}$ is odd. Since $\Pi_{2}\psi_{z}\in\mathcal{P}_{2}$, it follows that

$$\Pi_{2}\psi_{z}\in\operatorname{span}\{t\}.$$

Since

$$\|t\|_{L^{2}(-1,1)}^{2}=\frac{2}{3}$$

and, by (2.22),

$$\langle\psi_{z},t\rangle_{L^{2}(-1,1)} =-2\left(\int_{0}^{z}t^{2}\,dt+z\int_{z}^{1}t\,dt\right)$$

$$=\frac{z^{3}-3z}{3},$$

we obtain

$$\Pi_{2}\psi_{z}(t)=\frac{z^{3}-3z}{2}\,t,\qquad-1<t<1.$$

Therefore, by (2.21) and (2.22), setting

(2.23) $$\beta_{z}=\frac{z(3-z^{2})}{2},\qquad 0<z<1,$$

we obtain, for $0<t<1$,

(2.24) $$d_{z}(t)=\begin{cases}(\beta_{z}-1)t,&0<t\leq z,\\ \beta_{z}t-z,&z<t<1.\end{cases}$$

Let $0<x\leq y<1$. By (2.19) and (2.20),

$$K_{-}(x,y)=\int_{0}^{1}d_{x}(t)d_{y}(t)\,dt.$$

Using (2.24),

$$K_{-}(x,y)=\int_{0}^{x}(\beta_{x}-1)(\beta_{y}-1)t^{2}\,dt+\int_{x}^{y}(\beta_{x}t-x)(\beta_{y}-1)t\,dt+\int_{y}^{1}(\beta_{x}t-x)(\beta_{y}t-y)\,dt.$$

Using (2.23) and simplifying, we obtain

$$K_{-}(x,y)=\frac{x(1-y)^{2}}{12}\left(3y-x^{2}(y+2)\right).$$

Finally, since $0<x\leq y<1$,

$$3y-x^{2}(y+2)\geq 3y-y^{2}(y+2)=y(1-y)(y+3)>0.$$

Thus

$$K_{-}(x,y)>0,\qquad 0<x\leq y<1.$$

By (i), $K_{-}$ is symmetric, and hence

$$K_{-}(x,y)>0,\qquad 0<x,y<1.$$

This proves (ii).

(iii) Let $0<x<1$. By (2.17) and the symmetry of $K_{-}$ established in (i),

$$\kappa_{-}(x)=\int_{0}^{x}K_{-}(y,x)\,dy+\int_{x}^{1}K_{-}(x,y)\,dy.$$

Using (ii), we obtain

$$\kappa_{-}(x)=\frac{(1-x)^{2}}{12}\int_{0}^{x}y\left(3x-y^{2}(x+2)\right)\,dy+\frac{x}{12}\int_{x}^{1}(1-y)^{2}\left(3y-x^{2}(y+2)\right)\,dy.$$

A direct computation gives

$$\kappa_{-}(x)=\frac{x(1-x)^{2}(2x+1)}{48}.$$

This proves (iii).

(iv) Let $0<x<1$. By (2.17),

$$\kappa_{+}(x)^{2} =\frac{1}{2}\|r_{x}+r_{-x}\|_{L^{2}(-1,1)}^{2}$$

$$=\frac{1}{2}\left(\|r_{x}\|_{L^{2}(-1,1)}^{2}+\|r_{-x}\|_{L^{2}(-1,1)}^{2}+2\langle r_{x},r_{-x}\rangle_{L^{2}(-1,1)}\right).$$

By Lemma 2.3(i) and (iii),

$$\|r_{-x}\|_{L^{2}(-1,1)}=\|r_{x}\|_{L^{2}(-1,1)}.$$

Therefore, by (2.15),

$$\kappa_{+}(x)^{2} =\|r_{x}\|_{L^{2}(-1,1)}^{2}+\langle r_{x},r_{-x}\rangle_{L^{2}(-1,1)}$$

$$=K(x,x)+K(x,-x).$$

Thus, by (2.16),

(2.25) $$\kappa_{+}(x)^{2}=K_{+}(x,x).$$

Set

$$\eta_{x}(t)=(|t|-x)_{+},\qquad-1<t<1.$$

By (2.4) and Lemma 2.3(ii)–(iii),

$$r_{x}+r_{-x} =(I-\Pi_{2})\phi_{x}+(I-\Pi_{2})R_{2}\phi_{x}$$

$$=(I-\Pi_{2})(\phi_{x}+R_{2}\phi_{x}).$$

Since

$$\phi_{x}(t)+(R_{2}\phi_{x})(t)=2x+\eta_{x}(t),\qquad-1<t<1,$$

and the constant function $t\mapsto 2x$ belongs to $\mathcal{P}_{2}$, we obtain

(2.26) $$r_{x}+r_{-x}=(I-\Pi_{2})\eta_{x}.$$

The function $\eta_{x}$ is even. Since $\Pi_{2}$ commutes with $R_{2}$ by Lemma 2.3(ii), $\Pi_{2}\eta_{x}$ is also even. Hence

$$\Pi_{2}\eta_{x}\in\operatorname{span}\left\{1,t^{2}-\frac{1}{3}\right\},$$

where the displayed basis is orthogonal in $L^{2}(-1,1)$. Moreover,

$$\|1\|_{L^{2}(-1,1)}^{2}=2,\qquad\left\|t^{2}-\frac{1}{3}\right\|_{L^{2}(-1,1)}^{2}=\frac{8}{45},$$

and a direct computation gives

$$\|\eta_{x}\|_{L^{2}(-1,1)}^{2}=\frac{2(1-x)^{3}}{3},\qquad\langle\eta_{x},1\rangle_{L^{2}(-1,1)}=(1-x)^{2},$$

and

$$\left\langle\eta_{x},t^{2}-\frac{1}{3}\right\rangle_{L^{2}(-1,1)}=\frac{(1-x^{2})^{2}}{6}.$$

Therefore,

$$\|\Pi_{2}\eta_{x}\|_{L^{2}(-1,1)}^{2} =\frac{\langle\eta_{x},1\rangle_{L^{2}(-1,1)}^{2}}{\|1\|_{L^{2}(-1,1)}^{2}}+\frac{\left\langle\eta_{x},t^{2}-\frac{1}{3}\right\rangle_{L^{2}(-1,1)}^{2}}{\left\|t^{2}-\frac{1}{3}\right\|_{L^{2}(-1,1)}^{2}}$$

$$=\frac{(1-x)^{4}}{2}+\frac{5(1-x^{2})^{4}}{32}.$$

By (2.17) and (2.26),

$$2\kappa_{+}(x)^{2} =\|(I-\Pi_{2})\eta_{x}\|_{L^{2}(-1,1)}^{2}$$

$$=\|\eta_{x}\|_{L^{2}(-1,1)}^{2}-\|\Pi_{2}\eta_{x}\|_{L^{2}(-1,1)}^{2}$$

$$=\frac{2(1-x)^{3}}{3}-\frac{(1-x)^{4}}{2}-\frac{5(1-x^{2})^{4}}{32}.$$

Using (2.18) and simplifying, we obtain

$$\kappa_{+}(x)^{2}=\frac{(1-x)^{3}}{192}P(x).$$

Together with (2.25), this proves (iv) and completes the proof. ∎

2.5. Estimates for the odd and even components

We now use the kernel identities established above to estimate separately the odd and even components in the decomposition (2.14) of $A^{*}g$.

We first estimate the odd component.

Lemma 2.6.

For every $a\in L^{\infty}(0,1)$,

$$\left\|\int_{0}^{1}a(x)\bigl(r_{x}-r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2}\leq 2\int_{0}^{1}\kappa_{-}(x)a(x)^{2}\,dx.$$

Moreover, equality holds if and only if $a$ is constant almost everywhere on $(0,1)$.

Proof.

Let $a\in L^{\infty}(0,1)$. By (2.20) and Fubini’s theorem,

$$\left\|\int_{0}^{1}a(x)\bigl(r_{x}-r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2} =\int_{0}^{1}\int_{0}^{1}a(x)a(y)\langle d_{x},d_{y}\rangle_{L^{2}(-1,1)}\,dx\,dy$$

$$=2\int_{0}^{1}\int_{0}^{1}a(x)a(y)K_{-}(x,y)\,dx\,dy.$$

By Lemma 2.5(ii),

$$K_{-}(x,y)>0,\qquad 0<x,y<1.$$

Hence, using

$$2a(x)a(y)\leq a(x)^{2}+a(y)^{2},$$

together with (2.17) and the symmetry of $K_{-}$ established in Lemma 2.5(i), we obtain

$$\left\|\int_{0}^{1}a(x)\bigl(r_{x}-r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2} \leq\int_{0}^{1}\int_{0}^{1}\bigl(a(x)^{2}+a(y)^{2}\bigr)K_{-}(x,y)\,dx\,dy$$

$$=2\int_{0}^{1}a(x)^{2}\left(\int_{0}^{1}K_{-}(x,y)\,dy\right)\,dx$$

$$=2\int_{0}^{1}\kappa_{-}(x)a(x)^{2}\,dx.$$

If equality holds, then

$$\int_{0}^{1}\int_{0}^{1}(a(x)-a(y))^{2}K_{-}(x,y)\,dx\,dy=0.$$

Since $K_{-}(x,y)>0$ on $(0,1)^{2}$, it follows that

$$a(x)=a(y)$$

for almost every $(x,y)\in(0,1)^{2}$. By Fubini’s theorem, $a$ is constant almost everywhere on $(0,1)$.

Conversely, if $a$ is constant almost everywhere, then

$$2a(x)a(y)=a(x)^{2}+a(y)^{2}$$

for almost every $(x,y)\in(0,1)^{2}$, and equality holds. ∎

Lemma 2.7.

Let $h\in L^{\infty}(0,1)$ satisfy

$$0\leq h(x)\leq 1\qquad\text{for almost every }x\in(0,1),$$

and

$$\|h\|_{L^{\infty}(0,1)}>0.$$

Then

(2.27) $$\left(\int_{0}^{1}\kappa_{+}(x)h(x)\,dx\right)^{2}<\int_{0}^{1}\kappa_{-}(x)h(x)\,dx.$$

Proof.

By Lemma 2.5(iii),

$$\kappa_{-}(x)>0,\qquad 0<x<1.$$

Define the finite measure $\mu$ on $(0,1)$ by

$$d\mu(x)=\kappa_{-}(x)\,dx,$$

and set

$$\rho(x)=\frac{\kappa_{+}(x)}{\kappa_{-}(x)},\qquad 0<x<1.$$

By (2.17),

$$\rho(x)\geq 0,\qquad 0<x<1.$$

Moreover,

(2.28) $$\int_{0}^{1}\kappa_{+}(x)h(x)\,dx=\int_{0}^{1}\rho(x)h(x)\,d\mu(x),\qquad\int_{0}^{1}\kappa_{-}(x)h(x)\,dx=\int_{0}^{1}h(x)\,d\mu(x).$$

We first establish a weak-$L^{2}(\mu)$ estimate for $\rho$. By Lemma 2.5(iii) and (iv),

(2.29) $$\rho(x)^{2}=\frac{12P(x)}{x^{2}(1-x)(2x+1)^{2}},\qquad 0<x<1.$$

Set

$$W(x)=\mu((0,x))=\int_{0}^{x}\kappa_{-}(s)\,ds,\qquad 0\leq x\leq 1.$$

Using Lemma 2.5(iii), a direct integration gives

(2.30) $$W(x)=\frac{x^{2}(8x^{3}-15x^{2}+10)}{960},\qquad 0\leq x\leq 1,$$

and, in particular,

(2.31) $$W(1)=\frac{1}{320}.$$

We claim that

(2.32) $$\rho(x)^{2}W(x)<\frac{1}{8},\qquad 0<x\leq\frac{1}{2},$$

and

(2.33) $$\rho(x)^{2}\bigl(W(1)-W(x)\bigr)<\frac{1}{8},\qquad\frac{1}{2}\leq x<1.$$

Indeed, by (2.18), (2.29), and (2.30),

(2.34) $$\frac{1}{8}-\rho(x)^{2}W(x)=\frac{3x^{2}R_{-}(x)}{80(x-1)(2x+1)^{2}},\qquad 0<x\leq\frac{1}{2},$$

where

$$R_{-}(x)=40x^{6}+45x^{5}-145x^{4}-180x^{3}+308x^{2}+101x-105.$$

For integers $0\leq i\leq j$, let

$$B_{i,j}(s)=\binom{j}{i}s^{i}(1-s)^{j-i},\qquad 0\leq s\leq 1.$$

A direct expansion in the Bernstein basis gives

$$-R_{-}\left(\frac{s}{2}\right)=\sum_{i=0}^{6}c_{i}B_{i,6}(s),\qquad 0\leq s\leq 1,$$

where

$$(c_{0},\ldots,c_{6})=\left(105,\frac{1159}{12},\frac{2491}{30},\frac{2619}{40},\frac{3651}{80},\frac{4919}{192},\frac{225}{32}\right).$$

All these coefficients are positive, and hence

$$R_{-}(x)<0,\qquad 0\leq x\leq\frac{1}{2}.$$

Together with (2.34), this yields (2.32).

Similarly, using (2.29), (2.30), and (2.31),

(2.35) $$\frac{1}{8}-\rho(x)^{2}\bigl(W(1)-W(x)\bigr)=-\frac{R_{+}(x)}{80x^{2}(2x+1)^{2}},\qquad\frac{1}{2}\leq x<1,$$

where

$$R_{+}(x)=120x^{9}+255x^{8}-180x^{7}-720x^{6}+204x^{5}+382x^{4}-68x^{3}-98x^{2}+12x+3.$$

A direct expansion in the Bernstein basis gives

$$-R_{+}\left(\frac{1+s}{2}\right)=\sum_{i=0}^{9}d_{i}B_{i,9}(s),\qquad 0\leq s\leq 1,$$

where

$$(d_{0},\ldots,d_{9})=\Bigg(\frac{1325}{256},\frac{4691}{768},\frac{8269}{1152},\frac{6019}{672},\frac{1577}{126},\frac{2453}{126},\frac{1775}{56},\frac{3625}{72},\frac{220}{3},90\Bigg).$$

Again all coefficients are positive. Thus

$$R_{+}(x)<0,\qquad\frac{1}{2}\leq x\leq 1.$$

Together with (2.35), this yields (2.33).

We claim that

(2.36) $$\mu\bigl(\{x\in(0,1):\rho(x)>\lambda\}\bigr)<\frac{1}{4\lambda^{2}},\qquad\lambda>0.$$

For $\lambda>0$, set

$$c_{\lambda}=\frac{1}{8\lambda^{2}},$$

and let

$$E_{\lambda}^{-}=\left\{0<x\leq\frac{1}{2}:\rho(x)>\lambda\right\}.$$

By (2.32),

$$W(x)<c_{\lambda},\qquad x\in E_{\lambda}^{-},$$

and hence

(2.37) $$E_{\lambda}^{-}\subset\left\{0<x\leq\frac{1}{2}:W(x)<c_{\lambda}\right\}.$$

If

$$c_{\lambda}>W\left(\frac{1}{2}\right),$$

then, since $W$ is strictly increasing,

$$\left\{0<x\leq\frac{1}{2}:W(x)<c_{\lambda}\right\}=\left(0,\frac{1}{2}\right],$$

and therefore

$$\mu(E_{\lambda}^{-})\leq W\left(\frac{1}{2}\right)<c_{\lambda}.$$

Suppose now that

$$0<c_{\lambda}\leq W\left(\frac{1}{2}\right).$$

Since $W$ is continuous and strictly increasing, with $W(0)=0$, there exists a unique $\xi\in(0,\frac{1}{2}]$ such that

$$W(\xi)=c_{\lambda}.$$

Thus, by (2.37),

$$E_{\lambda}^{-}\subset(0,\xi).$$

Moreover, (2.32) gives

$$\rho(\xi)^{2}W(\xi)<\frac{1}{8},$$

and hence

$$\rho(\xi)<\lambda.$$

Since $\rho$ is continuous on $(0,1)$, there exists $0<\delta<\xi$ such that

$$\rho(x)<\lambda,\qquad\xi-\delta\leq x\leq\xi.$$

Consequently,

$$E_{\lambda}^{-}\subset(0,\xi-\delta),$$

and therefore

$$\mu(E_{\lambda}^{-})\leq W(\xi-\delta)<W(\xi)=c_{\lambda}.$$

Thus, in either case,

(2.38) $$\mu(E_{\lambda}^{-})<c_{\lambda}.$$

Likewise, let

$$E_{\lambda}^{+}=\left\{\frac{1}{2}<x<1:\rho(x)>\lambda\right\}.$$

By (2.33),

$$W(1)-W(x)<c_{\lambda},\qquad x\in E_{\lambda}^{+},$$

and hence

$$E_{\lambda}^{+}\subset\left\{\frac{1}{2}<x<1:W(1)-W(x)<c_{\lambda}\right\}.$$

Since $x\mapsto W(1)-W(x)$ is continuous and strictly decreasing on $[\frac{1}{2},1]$, there exists $\xi\in[\frac{1}{2},1)$ such that

$$\left\{\frac{1}{2}<x<1:W(1)-W(x)<c_{\lambda}\right\}=(\xi,1),$$

with

$$W(1)-W(\xi)\leq c_{\lambda}.$$

Therefore,

(2.39) $$\mu(E_{\lambda}^{+}) \leq\mu((\xi,1))$$

$$=W(1)-W(\xi)$$

$$\leq c_{\lambda}.$$

Since

$$\{x\in(0,1):\rho(x)>\lambda\}=E_{\lambda}^{-}\cup E_{\lambda}^{+},$$

and the union is disjoint, (2.38) and (2.39) give

$$\mu\bigl(\{x\in(0,1):\rho(x)>\lambda\}\bigr) =\mu(E_{\lambda}^{-})+\mu(E_{\lambda}^{+})$$

$$<2c_{\lambda}$$

$$=\frac{1}{4\lambda^{2}}.$$

This proves (2.36).

Now let $E\subset(0,1)$ be $\mu$-measurable with

$$m=\mu(E)>0.$$

Set

$$\lambda_{0}=\frac{1}{2\sqrt{m}}.$$

By (2.36),

$$\mu\bigl(E\cap\{\rho>\lambda\}\bigr)<\frac{1}{4\lambda^{2}},\qquad\lambda>\lambda_{0},$$

and hence

$$\int_{\lambda_{0}}^{+\infty}\mu\bigl(E\cap\{\rho>\lambda\}\bigr)\,d\lambda<\frac{1}{4\lambda_{0}}.$$

On the other hand,

$$\mu\bigl(E\cap\{\rho>\lambda\}\bigr)\leq m,\qquad 0<\lambda\leq\lambda_{0},$$

so

$$\int_{0}^{\lambda_{0}}\mu\bigl(E\cap\{\rho>\lambda\}\bigr)\,d\lambda\leq m\lambda_{0}.$$

Therefore, by the layer-cake representation,

$$\int_{E}\rho\,d\mu =\int_{0}^{+\infty}\mu\bigl(E\cap\{\rho>\lambda\}\bigr)\,d\lambda$$

$$<m\lambda_{0}+\frac{1}{4\lambda_{0}}$$

$$=\sqrt{m}.$$

Thus

(2.40) $$\left(\int_{E}\rho\,d\mu\right)^{2}<\mu(E)$$

for every $\mu$-measurable $E\subset(0,1)$ with $\mu(E)>0$.

Finally, for $0<t<1$, set

$$E_{t}=\{x\in(0,1):h(x)>t\}.$$

Since $\|h\|_{L^{\infty}(0,1)}>0$, there exists

$$0<\varepsilon<\|h\|_{L^{\infty}(0,1)}$$

such that

$$|E_{\varepsilon}|>0.$$

Since $\kappa_{-}>0$ on $(0,1)$,

$$\mu(E_{\varepsilon})>0.$$

Moreover,

$$E_{\varepsilon}\subset E_{t},\qquad 0<t<\varepsilon,$$

and hence

$$\mu(E_{t})>0,\qquad 0<t<\varepsilon.$$

By (2.40),

$$\int_{E_{t}}\rho\,d\mu<\mu(E_{t})^{1/2},\qquad 0<t<\varepsilon,$$

whereas, for every $0<t<1$,

$$\int_{E_{t}}\rho\,d\mu\leq\mu(E_{t})^{1/2}.$$

Consequently, by the layer-cake representation and the Cauchy–Schwarz inequality,

$$\int_{0}^{1}\rho(x)h(x)\,d\mu(x) =\int_{0}^{1}\left(\int_{E_{t}}\rho(x)\,d\mu(x)\right)dt$$

$$<\int_{0}^{1}\mu(E_{t})^{1/2}\,dt$$

$$\leq\left(\int_{0}^{1}\mu(E_{t})\,dt\right)^{1/2}$$

$$=\left(\int_{0}^{1}h(x)\,d\mu(x)\right)^{1/2}.$$

Hence

$$\left(\int_{0}^{1}\rho(x)h(x)\,d\mu(x)\right)^{2}<\int_{0}^{1}h(x)\,d\mu(x).$$

Together with (2.28), this yields (2.27). ∎

Corollary 2.8.

Let $b\in L^{\infty}(0,1)$ satisfy

$$\|b\|_{L^{\infty}(0,1)}>0.$$

Then

$$\left\|\int_{0}^{1}b(x)\bigl(r_{x}+r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2}<2\|b\|_{L^{\infty}(0,1)}\int_{0}^{1}\kappa_{-}(x)|b(x)|\,dx.$$

Proof.

Set

$$M=\|b\|_{L^{\infty}(0,1)},\qquad h(x)=\frac{|b(x)|}{M}.$$

Then

$$0\leq h(x)\leq 1\qquad\text{for almost every }x\in(0,1),$$

and

$$\|h\|_{L^{\infty}(0,1)}=1.$$

By Lemma 2.7,

$$\left(\int_{0}^{1}\kappa_{+}(x)|b(x)|\,dx\right)^{2}<M\int_{0}^{1}\kappa_{-}(x)|b(x)|\,dx.$$

On the other hand, by the triangle inequality for Bochner integrals and (2.17),

$$\left\|\int_{0}^{1}b(x)\bigl(r_{x}+r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)} \leq\int_{0}^{1}|b(x)|\,\|r_{x}+r_{-x}\|_{L^{2}(-1,1)}\,dx$$

$$=\sqrt{2}\int_{0}^{1}\kappa_{+}(x)|b(x)|\,dx.$$

Consequently,

$$\left\|\int_{0}^{1}b(x)\bigl(r_{x}+r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2} \leq 2\left(\int_{0}^{1}\kappa_{+}(x)|b(x)|\,dx\right)^{2}$$

$$<2M\int_{0}^{1}\kappa_{-}(x)|b(x)|\,dx.$$

This proves the corollary. ∎

3. Proof of Theorem 1.1

Proof.

Step 1: The sharp constant.

Let $g\in L^{\infty}(-1,1)$ satisfy

$$\|g\|_{L^{\infty}(-1,1)}\leq 1.$$

By (2.13), for almost every $x\in(0,1)$,

(3.1) $$|a_{g}(x)|+|b_{g}(x)|=\max\{|g(x)|,|g(-x)|\}\leq 1.$$

By Lemma 2.4(ii), Lemma 2.6, and Corollary 2.8, with the case $b_{g}=0$ being trivial,

(3.2) $$\|A^{*}g\|_{L^{2}(-1,1)}^{2}=\left\|\int_{0}^{1}a_{g}(x)\bigl(r_{x}-r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2}+\left\|\int_{0}^{1}b_{g}(x)\bigl(r_{x}+r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2},$$

(3.3) $$\left\|\int_{0}^{1}a_{g}(x)\bigl(r_{x}-r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2}\leq 2\int_{0}^{1}\kappa_{-}(x)a_{g}(x)^{2}\,dx,$$

(3.4) $$\left\|\int_{0}^{1}b_{g}(x)\bigl(r_{x}+r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2}\leq 2\|b_{g}\|_{L^{\infty}(0,1)}\int_{0}^{1}\kappa_{-}(x)|b_{g}(x)|\,dx.$$

Together with (3.1), this yields

(3.5) $$\|A^{*}g\|_{L^{2}(-1,1)}^{2} \leq 2\int_{0}^{1}\kappa_{-}(x)a_{g}(x)^{2}\,dx+2\|b_{g}\|_{L^{\infty}(0,1)}\int_{0}^{1}\kappa_{-}(x)|b_{g}(x)|\,dx$$

$$\leq 2\int_{0}^{1}\kappa_{-}(x)\left(a_{g}(x)^{2}+|b_{g}(x)|\right)\,dx$$

$$\leq 2\int_{0}^{1}\kappa_{-}(x)\left(|a_{g}(x)|+|b_{g}(x)|\right)\,dx$$

$$\leq 2\int_{0}^{1}\kappa_{-}(x)\,dx.$$

Using Lemma 2.5(iii), a direct integration gives

(3.6) $$\int_{0}^{1}\kappa_{-}(x)\,dx=\frac{1}{320}.$$

Therefore,

$$\|A^{*}g\|_{L^{2}(-1,1)}^{2}\leq\frac{1}{160},$$

and hence, by (2.8),

(3.7) $$\|A^{*}\|_{\infty\to 2}\leq\frac{1}{4\sqrt{10}}.$$

Now let

(3.8) $$g_{0}(t)=\operatorname{sgn}(t),\qquad-1<t<1.$$

Then

$$\|g_{0}\|_{L^{\infty}(-1,1)}=1,\qquad a_{g_{0}}(x)=1,\qquad b_{g_{0}}(x)=0,\qquad 0<x<1.$$

Since $a_{g_{0}}$ is constant, the equality case in Lemma 2.6, together with Lemma 2.4(ii) and (3.6), gives

$$\|A^{*}g_{0}\|_{L^{2}(-1,1)}^{2}=2\int_{0}^{1}\kappa_{-}(x)\,dx=\frac{1}{160}.$$

Thus equality is attained in (3.7), and

$$\|A^{*}\|_{\infty\to 2}=\frac{1}{4\sqrt{10}}.$$

By Lemma 2.2(ii),

$$C_{*}=\frac{1}{4\sqrt{10}},$$

and Lemma 2.2(iv) therefore yields

$$\lambda_{3}(3,1,2,1)=\frac{C_{*}}{4\sqrt{2}}=\frac{1}{32\sqrt{5}}.$$

This proves (1.3).

Step 2: Extremal functions for $\lambda_{3}(3,1,2,1)$.

We first determine all functions in the closed unit ball of $L^{\infty}(-1,1)$ at which the norm $\|A^{*}\|_{\infty\to 2}$ is attained.

Let $g\in L^{\infty}(-1,1)$ satisfy

$$\|g\|_{L^{\infty}(-1,1)}\leq 1$$

and

(3.9) $$\|A^{*}g\|_{L^{2}(-1,1)}=\|A^{*}\|_{\infty\to 2}=\frac{1}{4\sqrt{10}}.$$

Then equality holds throughout (3.5). In view of (3.2), (3.3), and (3.4), this yields

(3.10) $$\left\|\int_{0}^{1}a_{g}(x)\bigl(r_{x}-r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2}=2\int_{0}^{1}\kappa_{-}(x)a_{g}(x)^{2}\,dx$$

and

(3.11) $$\left\|\int_{0}^{1}b_{g}(x)\bigl(r_{x}+r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2}=2\|b_{g}\|_{L^{\infty}(0,1)}\int_{0}^{1}\kappa_{-}(x)|b_{g}(x)|\,dx.$$

By (3.11) and Corollary 2.8,

(3.12) $$b_{g}(x)=0\qquad\text{for almost every }x\in(0,1).$$

Hence, by Lemma 2.4(ii),

(3.13) $$\|A^{*}g\|_{L^{2}(-1,1)}^{2}=\left\|\int_{0}^{1}a_{g}(x)\bigl(r_{x}-r_{-x}\bigr)\,dx\right\|_{L^{2}(-1,1)}^{2}.$$

Moreover, by (3.10) and Lemma 2.6, $a_{g}$ is constant almost everywhere on $(0,1)$. Thus, for some $c\in\mathbb{R}$,

(3.14) $$a_{g}(x)=c\qquad\text{for almost every }x\in(0,1).$$

Together with (3.6), (3.9), (3.10), and (3.13), this gives

$$\|A^{*}g\|_{L^{2}(-1,1)}^{2}=2c^{2}\int_{0}^{1}\kappa_{-}(x)\,dx=\frac{c^{2}}{160}=\frac{1}{160}.$$

It follows that

(3.15) $$c=\pm 1.$$

By (2.13), (3.12), (3.14), and (3.15), we conclude that

$$g(t)=\pm g_{0}(t)\qquad\text{for almost every }t\in(-1,1),$$

where $g_{0}$ is defined by (3.8).

Combining this characterization with Step 1, we conclude that the norm of $A^{*}$ on the closed unit ball of $L^{\infty}(-1,1)$ is attained exactly at $\pm g_{0}$. In particular, by Lemma 2.2(ii),

(3.16) $$\|A^{*}\|_{\infty\to 2}=\|A^{*}g_{0}\|_{L^{2}(-1,1)}=C_{*}.$$

We next compute $A^{*}g_{0}$ explicitly. By (2.4), (2.7), and (3.8),

(3.17) $$A^{*}g_{0}=(I-\Pi_{2})F,$$

where

$$F(t)=\int_{-1}^{1}\operatorname{sgn}(x)(x-t)_{+}\,dx,\qquad-1<t<1.$$

A direct computation gives

(3.18) $$F(t)=\frac{1}{2}-t+\frac{1}{2}t|t|,\qquad-1<t<1.$$

Since the even part of $F$ is the constant $1/2$, while the odd subspace of $\mathcal{P}_{2}$ is spanned by $t$, we have

$$\Pi_{2}F(t)=\frac{1}{2}+\frac{\langle-t+\frac{1}{2}t|t|,t\rangle_{L^{2}(-1,1)}}{\|t\|_{L^{2}(-1,1)}^{2}}\,t.$$

Moreover,

$$\|t\|_{L^{2}(-1,1)}^{2}=\frac{2}{3},\qquad\left\langle\frac{1}{2}t|t|,t\right\rangle_{L^{2}(-1,1)}=\frac{1}{2}\int_{-1}^{1}t^{2}|t|\,dt=\frac{1}{4}.$$

Hence

$$\Pi_{2}F(t)=\frac{1}{2}-\frac{5}{8}t,\qquad-1<t<1.$$

Together with (3.17) and (3.18), this yields

(3.19) $$f_{0}(t):=(A^{*}g_{0})(t)=\frac{1}{2}t|t|-\frac{3}{8}t=\frac{t(4|t|-3)}{8},\qquad-1<t<1.$$

Define

(3.20) $$V_{0}(t)=\frac{t(1-|t|)^{2}(1+2|t|)}{48},\qquad-1\leq t\leq 1.$$

The function $V_{0}$ is odd,

$$V_{0}(\pm 1)=V_{0}^{\prime}(\pm 1)=0,$$

and a direct differentiation gives

(3.21) $$V_{0}^{\prime\prime}=f_{0},$$

where $f_{0}$ is defined by (3.19). Hence $V_{0}\in H_{0}^{2}(-1,1)$, and its oddness yields

$$\int_{-1}^{1}V_{0}(t)\,dt=0.$$

Therefore,

$$V_{0}\in X.$$

By Lemma 2.2(i),

(3.22) $$f_{0}\in\mathcal{P}_{2}^{\perp}\qquad\text{and}\qquad V_{0}=Af_{0}.$$

Moreover, by (3.8) and (3.20),

(3.23) $$\operatorname{sgn}(V_{0}(t))=\operatorname{sgn}(t)=g_{0}(t),\qquad-1<t<1.$$

Together with (3.19) and (3.22), this yields

(3.24) $$\|V_{0}\|_{L^{1}(-1,1)} =\int_{-1}^{1}g_{0}(t)V_{0}(t)\,dt$$

$$=\int_{-1}^{1}g_{0}(t)(Af_{0})(t)\,dt$$

$$=\langle A^{*}g_{0},f_{0}\rangle_{L^{2}(-1,1)}$$

$$=\|f_{0}\|_{L^{2}(-1,1)}^{2}.$$

Hence, by (3.16), (3.19), (3.21), and (3.24),

$$\frac{\|V_{0}\|_{L^{1}(-1,1)}}{\|V_{0}^{\prime\prime}\|_{L^{2}(-1,1)}}=\|f_{0}\|_{L^{2}(-1,1)}=\|A^{*}g_{0}\|_{L^{2}(-1,1)}=C_{*}.$$

Thus $V_{0}$ is an extremal function for (2.3).

We now prove that the extremal functions for (2.3) are unique up to multiplication by a nonzero constant.

Let $V\in X\setminus\{0\}$ be an arbitrary extremal function for (2.3), and set

$$f=V^{\prime\prime}.$$

By Lemma 2.2(i),

$$f\in\mathcal{P}_{2}^{\perp},\qquad V=Af.$$

Set

$$g(t)=\operatorname{sgn}(V(t)),\qquad-1<t<1.$$

Then

$$\|g\|_{L^{\infty}(-1,1)}=1$$

and

$$\|V\|_{L^{1}(-1,1)}=\int_{-1}^{1}g(t)V(t)\,dt.$$

Since $V$ is extremal, by (2.3), (2.8), and Lemma 2.2(ii),

$$C_{*}\|f\|_{L^{2}(-1,1)} =\|V\|_{L^{1}(-1,1)}$$

$$=\int_{-1}^{1}g(t)(Af)(t)\,dt$$

$$=\langle A^{*}g,f\rangle_{L^{2}(-1,1)}$$

$$\leq\|A^{*}g\|_{L^{2}(-1,1)}\|f\|_{L^{2}(-1,1)}$$

$$\leq\|A^{*}\|_{\infty\to 2}\|g\|_{L^{\infty}(-1,1)}\|f\|_{L^{2}(-1,1)}$$

$$=C_{*}\|f\|_{L^{2}(-1,1)}.$$

Since $f\neq 0$, all the inequalities above are equalities. Consequently,

(3.25) $$\langle A^{*}g,f\rangle_{L^{2}(-1,1)}=\|A^{*}g\|_{L^{2}(-1,1)}\|f\|_{L^{2}(-1,1)}$$

and

(3.26) $$\|A^{*}g\|_{L^{2}(-1,1)}=C_{*}.$$

By (3.26) and the characterization above,

$$g=\pm g_{0},$$

and therefore, by (3.19),

$$A^{*}g=\pm f_{0}.$$

Moreover, by (3.25), $f$ and $A^{*}g$ are linearly dependent. Hence

$$f=cf_{0}$$

for some $c\neq 0$. Since $V=Af$ and, by (3.22), $V_{0}=Af_{0}$, the linearity of $A$ gives

$$V=cV_{0}.$$

Thus the extremal functions for (2.3) are precisely the nonzero scalar multiples of $V_{0}$.

We now return to the original interval $(0,1)$. By Lemma 2.2(iii), the inverse image of $V_{0}$ under $T$ is given by

$$x\longmapsto\int_{0}^{x}V_{0}(2s-1)\,ds.$$

For convenience, we choose the normalization

(3.27) $$U_{*}(x)=-120\int_{0}^{x}V_{0}(2s-1)\,ds,\qquad 0\leq x\leq 1.$$

Then

$$(TU_{*})(t)=U_{*}^{\prime}\left(\frac{t+1}{2}\right)=-120V_{0}(t),\qquad-1\leq t\leq 1.$$

Using (3.20), a direct integration gives

(3.28) $$U_{*}(x)=\begin{cases}x^{3}\bigl(16x^{2}-25x+10\bigr),&0\leq x\leq\frac{1}{2},\\[2.84526pt] (1-x)^{3}\bigl(16(1-x)^{2}-25(1-x)+10\bigr),&\frac{1}{2}\leq x\leq 1.\end{cases}$$

By Lemma 2.2(iv) and the above characterization of the extremal functions for (2.3), the extremal functions for $\lambda_{3}(3,1,2,1)$ are precisely the nonzero scalar multiples of $U_{*}$. Moreover, (3.28) gives

$$U_{*}(1-x)=U_{*}(x),\qquad 0\leq x\leq 1.$$

Thus every extremal function for $\lambda_{3}(3,1,2,1)$ is symmetric about $x=\frac{1}{2}$.

Step 3: Extremal functions for $\lambda_{3}(3,0,2,\infty)$.

By (1.1),

$$\lambda_{3}(3,0,2,\infty)=\sup_{U\in\mathring{W}_{2}^{3}(0,1)\setminus\{0\}}\frac{\|U\|_{L^{\infty}(0,1)}}{\|U^{\prime\prime\prime}\|_{L^{2}(0,1)}}.$$

We first show that $U_{*}$ is an extremal function for $\lambda_{3}(3,0,2,\infty)$.

Indeed, by (3.23) and (3.27),

$$U_{*}^{\prime}(x)=-120V_{0}(2x-1),\qquad 0\leq x\leq 1,$$

and

$$\operatorname{sgn}(V_{0}(t))=\operatorname{sgn}(t),\qquad-1<t<1.$$

Hence $U_{*}$ is increasing on $(0,\frac{1}{2})$ and decreasing on $(\frac{1}{2},1)$. Together with

$$U_{*}(0)=U_{*}(1)=0,$$

this yields

$$\|U_{*}\|_{L^{\infty}(0,1)}=U_{*}\left(\frac{1}{2}\right)$$

and

(3.29) $$\|U_{*}^{\prime}\|_{L^{1}(0,1)} =\int_{0}^{\frac{1}{2}}U_{*}^{\prime}(x)\,dx-\int_{\frac{1}{2}}^{1}U_{*}^{\prime}(x)\,dx$$

$$=2U_{*}\left(\frac{1}{2}\right)$$

$$=2\|U_{*}\|_{L^{\infty}(0,1)}.$$

Since $U_{*}$ is an extremal function for $\lambda_{3}(3,1,2,1)$, (1.3) gives

$$\frac{\|U_{*}^{\prime}\|_{L^{1}(0,1)}}{\|U_{*}^{\prime\prime\prime}\|_{L^{2}(0,1)}}=\frac{1}{32\sqrt{5}}.$$

Therefore, by (1.2) and (3.29),

$$\frac{\|U_{*}\|_{L^{\infty}(0,1)}}{\|U_{*}^{\prime\prime\prime}\|_{L^{2}(0,1)}}=\frac{1}{64\sqrt{5}}=\lambda_{3}(3,0,2,\infty).$$

Thus $U_{*}$ is an extremal function for $\lambda_{3}(3,0,2,\infty)$.

Let

$$\widetilde{\mathcal{P}}_{2}=\operatorname{span}\{1,s,s^{2}\}\subset L^{2}(0,1),$$

and denote by

$$\widetilde{\Pi}_{2}:L^{2}(0,1)\longrightarrow\widetilde{\mathcal{P}}_{2}$$

the orthogonal projection onto $\widetilde{\mathcal{P}}_{2}$. For $x\in[0,1]$, set

$$\sigma_{x}=(I-\widetilde{\Pi}_{2})\left(\frac{(x-\cdot)_{+}^{2}}{2}\right).$$

A direct computation gives

$$\|\sigma_{x}\|_{L^{2}(0,1)}^{2}=\frac{x^{5}(1-x)^{5}}{20},\qquad 0\leq x\leq 1.$$

Consequently,

(3.30) $$\|\sigma_{x}\|_{L^{2}(0,1)}\leq\frac{1}{64\sqrt{5}},$$

with equality if and only if $x=\frac{1}{2}$.

Let $U\in\mathring{W}_{2}^{3}(0,1)\setminus\{0\}$ be an arbitrary extremal function for $\lambda_{3}(3,0,2,\infty)$. Choose $x_{0}\in(0,1)$ such that

$$|U(x_{0})|=\|U\|_{L^{\infty}(0,1)},$$

and set

$$h=U^{\prime\prime\prime}.$$

Then $h\neq 0$ and

$$\frac{|U(x_{0})|}{\|h\|_{L^{2}(0,1)}}=\lambda_{3}(3,0,2,\infty)=\frac{1}{64\sqrt{5}}.$$

The boundary conditions

$$U(0)=U(1)=U^{\prime}(0)=U^{\prime}(1)=U^{\prime\prime}(0)=U^{\prime\prime}(1)=0$$

and integration by parts give

$$\int_{0}^{1}h(s)\,ds=\int_{0}^{1}sh(s)\,ds=\int_{0}^{1}s^{2}h(s)\,ds=0.$$

Hence

$$h\perp\widetilde{\mathcal{P}}_{2}$$

and

$$U(x)=\frac{1}{2}\int_{0}^{1}(x-s)_{+}^{2}h(s)\,ds=\langle\sigma_{x},h\rangle_{L^{2}(0,1)},\qquad 0\leq x\leq 1.$$

Using (3.30), we obtain

$$\frac{1}{64\sqrt{5}}\|h\|_{L^{2}(0,1)} =|U(x_{0})|$$

$$=|\langle\sigma_{x_{0}},h\rangle_{L^{2}(0,1)}|$$

$$\leq\|\sigma_{x_{0}}\|_{L^{2}(0,1)}\|h\|_{L^{2}(0,1)}$$

$$\leq\frac{1}{64\sqrt{5}}\|h\|_{L^{2}(0,1)}.$$

Since $h\neq 0$, equality holds throughout. Therefore,

$$\|\sigma_{x_{0}}\|_{L^{2}(0,1)}=\frac{1}{64\sqrt{5}},\qquad|\langle\sigma_{x_{0}},h\rangle_{L^{2}(0,1)}|=\|\sigma_{x_{0}}\|_{L^{2}(0,1)}\|h\|_{L^{2}(0,1)}.$$

Consequently, $x_{0}=\frac{1}{2}$ and

$$h=c\,\sigma_{1/2}$$

for some $c\neq 0$.

Applying the same argument to the extremal function $U_{*}$ gives

$$U_{*}^{\prime\prime\prime}=c_{*}\,\sigma_{1/2}$$

for some $c_{*}\neq 0$. Hence

$$U^{\prime\prime\prime}=c_{0}U_{*}^{\prime\prime\prime},\qquad c_{0}=\frac{c}{c_{*}}.$$

Since

$$U-c_{0}U_{*}\in\mathring{W}_{2}^{3}(0,1)$$

and

$$(U-c_{0}U_{*})^{\prime\prime\prime}=0,$$

we obtain

$$U=c_{0}U_{*}.$$

Thus the extremal functions for $\lambda_{3}(3,0,2,\infty)$ are precisely the nonzero scalar multiples of $U_{*}$. Combining this with Step 2, the families of extremal functions for $\lambda_{3}(3,1,2,1)$ and $\lambda_{3}(3,0,2,\infty)$ coincide up to multiplication by a nonzero constant. Moreover, every such extremal function is symmetric about $x=\frac{1}{2}$. This completes the proof.

∎