betaFeedback

WorksMathematics

Strict rearrangement inequalities for logarithmic p-Laplacians

Abstract

We establish cases of equality for the Polya-Szego inequality of the logarithmic p-Laplacian.

Introduction

Background and main ideas

The logarithmic Laplacian LΔL_\Delta arises by differentiating the fractional Laplacian with respect to its order at zero. For u∈Cc∞(RN)u \in C^\infty_c(\mathbb{R}^N), we have

(−Δ)su=u+sLΔu+o(s)as s→0+,(-\Delta)^s u = u + sL_\Delta u + o(s) \quad\text{as } s \to0^+,

and the Fourier symbol of LΔL_\Delta is 2log⁡∣ξ∣2\log|\xi|. Chen and Weth [7] developed the variational and spectral theory of the corresponding Dirichlet problem on bounded domains. This theory was extended to the nonlinear setting by Dyda, Jarohs and Sk [9], who introduced the logarithmic pp-Laplacian for 1<p<∞1 < p < \infty.

Several aspects of the nonlinear theory have since been studied. Arora, Giacomoni, Hajaiej and Vaishnavi [1] established the pp-logarithmic Sobolev inequality and sharp embeddings into Orlicz-type spaces, and studied Dirichlet problems with critical growth. Arora, Hajaiej and Perera [2] constructed an unbounded sequence of minimax eigenvalues and obtained existence results for superlinear subcritical Dirichlet problems. For positive fractional orders, Bahrouni, Gouasmia, Hajaiej and Ouannasser [3] studied the fractional logarithmic pp-Laplacian, obtained by differentiating the fractional pp-Laplacian with respect to its order at s>0s>0.

Symmetrization for the logarithmic energy requires some care. The energy consists of a positive short-range term, a renormalized long-range interaction, and a multiple of the LpL^p-mass. Its kernel ∣x−y∣−N|x-y|^{-N} is singular at the diagonal and is not integrable at infinity. Moreover, for real-valued functions the numerator of the long-range interaction may have either sign. The cancellation in the energy must therefore be preserved, and rearrangement cannot be applied separately to terms that are only formally defined.

Chen and Weth [7 Corollary 3.6] proved that balls minimize the first Dirichlet eigenvalue among bounded Lipschitz domains of prescribed measure. Their argument uses a first-order expansion of fractional Dirichlet eigenvalues. They left open the equality case and the question of a direct symmetrization proof; see [7 Remark 3.7]. By a similar asymptotic argument, Dyda, Jarohs and Sk obtained the Faber–Krahn inequality for the logarithmic pp-Laplacian and asked whether equality characterizes balls [9 Remark 1.5]. Although fractional eigenvalue asymptotics imply the spectral comparison, the equality information need not pass to the limit. Indeed, a fractional spectral deficit that is positive for every s>0s>0 may still be o(s)o(s) as s→0+s\to0^+. Thus the equality results for fractional Pólya–Szegő inequalities, such as [10 Theorem A.1], do not by themselves resolve these questions.

In this paper, we establish a direct Pólya–Szegő inequality for the logarithmic pp-energy and determine all equality cases for real-valued functions. A nonzero function attains equality precisely when it has one sign and its modulus is a translate of its Schwarz rearrangement. The result holds on RN\mathbb{R}^N without any boundedness or finite-measure assumption on the support. Applying it to a positive first eigenfunction, we prove that balls are the unique minimizers of the first Dirichlet eigenvalue among bounded open sets of fixed measure, up to translations and null sets. In particular, no boundary regularity is needed for this conclusion.

An essential point is to identify the integrability condition under which the rearranged energy is finite. We assume that

Sp(u):=∫RN∣u∣plog⁡−∣u∣ dx<∞,log⁡−t:=max⁡{0,−log⁡t}.S_p(u):=\int_{\mathbb{R}^N}|u|^p\log^-|u|\,\mathrm{d}x<\infty,\qquad\log^-t:=\max\{0,-\log t\}.

This condition controls small values of the function on sets that may be unbounded. It is invariant under rearrangement and is weaker than the additional assumption u∈Lq(RN)u\in L^q(\mathbb{R}^N) for some 0<q<p0<q<p. We prove that, for u∈Xp(RN)u\in X^p(\mathbb{R}^N), it is necessary and sufficient for u#=(∣u∣)∗u^\#=(|u|)^* to have finite logarithmic energy. More precisely, it is equivalent both to a logarithmic spatial moment of u#u^\# and to finiteness of its long-range interaction. If the condition fails, the short-range term of u#u^\# remains finite, whereas its long-range term equals −∞-\infty.

To prove the inequality and its equality statement, we write the interaction in terms of the strictly supermodular function

Hp(a,b)=ap+bp−∣a−b∣p,a,b≥0,H_p(a,b)=a^p+b^p-\lvert a-b\rvert^p,\qquad a,b\ge0,

and regularize the kernel by Kε(z)=(∣z∣2+ε2)−N/2K_\varepsilon(z)=(\lvert z\rvert^2+\varepsilon^2)^{-N/2}. We first estimate these interactions in terms of the distribution of the function and Sp(u)S_p(u). The logarithmic energy deficit can then be expressed as the limit of the regularized rearrangement deficits. The key observation is that these deficits increase as ε↓0\varepsilon\downarrow0, because the difference of two regularized kernels is positive and strictly radially decreasing. Consequently, equality in the limiting comparison forces equality at every fixed regularization scale. By strict rearrangement at one such scale, the function is a translate of its Schwarz rearrangement. A separate comparison between uu and ∣u∣\lvert u\rvert determines the sign of a real-valued equality case.

The use of monotonicity in this argument is motivated by the work of De Luca, Novaga and Ponsiglione [8]. For characteristic functions in the linear case, they proved that balls uniquely minimize the 0-fractional perimeter at fixed volume; see [8 Theorem 7.2]. Their proof also uses monotonicity of regularized interaction differences. Here we combine this property with the strict rearrangement theorem of Burchard and Hajaiej [6] for supermodular integrands. This allows us to treat general real-valued functions, all 1<p<∞1<p<\infty, and the precise logarithmic finiteness condition. For the classical equality theory of the Riesz rearrangement inequality, we refer to Burchard [5].

Main results

Let N≥1N\ge1 and 1<p<∞1<p<\infty. Following [9], we set

CN,p:=pΓ(N/2)2πN/2=p∣SN−1∣,ρN,p:=2log⁡2−γ+p2ψ(N2).(1)C_{N,p}:=\frac{p\Gamma(N/2)}{2\pi^{N/2}}=\frac{p}{\lvert S^{N-1}\rvert},\qquad\rho_{N,p}:=2\log2-\gamma+\frac{p}{2}\psi\left(\frac{N}{2}\right). \tag*{(1)}

where γ\gamma is Euler’s constant and ψ=Γ′/Γ\psi=\Gamma'/\Gamma. We write

[u]Xpp:=∬∣x−y∣<1∣u(x)−u(y)∣p∣x−y∣N dx dy.(2)[u]_{X_p}^p:=\iint_{\lvert x-y\rvert<1}\frac{\lvert u(x)-u(y)\rvert^p}{\lvert x-y\rvert^N}\,dx\,dy. \tag*{(2)}

and define Xp(RN):={u∈Lp(RN):[u]Xpp<∞}X^p(\mathbb{R}^N):=\{u\in L^p(\mathbb{R}^N):[u]_{X_p}^p<\infty\}. For a nonnegative function uu, let u∗u^* denote its Schwarz rearrangement. For real-valued uu, set u#:=(∣u∣)∗u^\# := (|u|)^*.

We also set

Sp(u):=∫RN∣u(x)∣plog⁡−∣u(x)∣ dx,(3)S_p(u):=\int_{\mathbb{R}^N}\lvert u(x)\rvert^p\log^-\lvert u(x)\rvert\,dx, \tag*{(3)}

with the convention 0plog⁡−0=00^p\log^-0=0. This quantity is unchanged by rearrangement. Its finiteness is also preserved under multiplication by a nonzero constant, since

∣log⁡−(ct)−log⁡−t∣≤∣log⁡c∣(c,t>0).\lvert\log^-(ct)-\log^-t\rvert\le\lvert\log c\rvert\qquad(c,t>0).

For u∈Xp(RN)u \in X^{p}(\mathbb{R}^{N}) satisfying Sp(u)<∞S_{p}(u)<\infty, the logarithmic pp-energy is

EL,p(u):=CN,p2∬∣x−y∣<1∣u(x)−u(y)∣p∣x−y∣N dx dy+CN,p2∬∣x−y∣≥1∣u(x)−u(y)∣p−∣u(x)∣p−∣u(y)∣p∣x−y∣N dx dy+ρN,p∫RN∣u∣p dx.(4)\begin{aligned} \mathcal{E}_{L,p}(u) :={}& \frac{C_{N,p}}{2}\iint_{|x-y|<1}\frac{|u(x)-u(y)|^{p}}{|x-y|^{N}}\,dx\,dy \\ &+\frac{C_{N,p}}{2}\iint_{|x-y|\ge1}\frac{|u(x)-u(y)|^{p}-|u(x)|^{p}-|u(y)|^{p}}{|x-y|^{N}}\,dx\,dy \\ &+\rho_{N,p}\int_{\mathbb{R}^{N}}|u|^{p}\,dx. \tag*{(4)} \end{aligned}

The absolute convergence of the long-range integral follows from Lemma 2.2 and (14) below. With this definition, Schwarz rearrangement decreases the energy, and equality determines both the sign and the radial profile of the function.

Theorem 1.1. Let N≥1N\ge1 and 1<p<∞1<p<\infty. Suppose that u∈Xp(RN)u\in X^{p}(\mathbb{R}^{N}) is real-valued and satisfies Sp(u)<∞S_{p}(u)<\infty. Then u#∈Xp(RN)u^{\#}\in X^{p}(\mathbb{R}^{N}), Sp(u#)=Sp(u)S_{p}(u^{\#})=S_{p}(u), and

EL,p(u#)≤EL,p(u).(5)\mathcal{E}_{L,p}(u^{\#})\le\mathcal{E}_{L,p}(u). \tag*{(5)}

If u≢0u\not\equiv0, equality holds in (5) if and only if there exist x0∈RNx_{0}\in\mathbb{R}^{N} and σ∈{−1,1}\sigma\in\{-1,1\} such that

u=σu#(⋅−x0)a.e. in RN.u=\sigma u^{\#}(\cdot-x_{0})\qquad\text{a.e. in }\mathbb{R}^{N}.

The logarithmic condition is satisfied whenever ∣{u≠0}∣<∞|\{u\ne0\}|<\infty, even if this set is unbounded. It is also satisfied if u∈Lq(RN)u\in L^{q}(\mathbb{R}^{N}) for some 0<q<p0<q<p. Indeed, for 0<t<10<t<1,

tplog⁡1t≤1pe,tplog⁡1t≤tqe(p−q).(6)t^{p}\log\frac{1}{t}\le\frac{1}{pe},\qquad t^{p}\log\frac{1}{t}\le\frac{t^{q}}{e(p-q)}. \tag*{(6)}

Hence Theorem 1.1 applies to all compactly supported functions in Xp(RN)X^{p}(\mathbb{R}^{N}) and to all functions in Xp(RN)∩L1(RN)X^{p}(\mathbb{R}^{N})\cap L^{1}(\mathbb{R}^{N}). Remark 2.4 contains admissible functions that satisfy none of these stronger assumptions.

The logarithmic condition is also necessary for the rearranged energy to be finite. More precisely, for every real-valued u∈Xp(RN)u\in X^{p}(\mathbb{R}^{N}), one has u#∈Xp(RN)u^{\#}\in X^{p}(\mathbb{R}^{N}) and

EL,p(u#)∈R⟺Sp(u)<∞.(7)\mathcal{E}_{L,p}(u^{\#})\in\mathbb{R}\quad\Longleftrightarrow\quad S_{p}(u)<\infty. \tag*{(7)}

If Sp(u)=∞S_{p}(u)=\infty, the short-range term of u#u^{\#} remains finite, whereas its long-range term equals −∞-\infty. These assertions follow from Proposition 2.3 and Lemma 3.2.

We next consider the first Dirichlet eigenvalue. For a bounded nonempty open set Ω⊂RN\Omega\subset\mathbb{R}^{N}, let

X0p(Ω):={u∈Xp(RN):u=0 a.e. in RN∖Ω}.(8)X_{0}^{p}(\Omega):=\{u\in X^{p}(\mathbb{R}^{N}):u=0\ \text{a.e. in }\mathbb{R}^{N}\setminus\Omega\}. \tag*{(8)}

The first Dirichlet eigenvalue is defined by

λ1,L,p(Ω):=inf⁡u∈X0p(Ω)∖{0}EL,p(u)∥u∥Lp(RN)p.(9)\lambda_{1,L,p}(\Omega):=\inf_{u\in X_{0}^{p}(\Omega)\setminus\{0\}}\frac{\mathcal{E}_{L,p}(u)}{\|u\|_{L^{p}(\mathbb{R}^{N})}^{p}}. \tag*{(9)}

By (6), every function in X0p(Ω)X_0^p(\Omega) satisfies the logarithmic integrability condition. Theorem 1.1 therefore yields the following Faber–Krahn inequality and equality classification.

Theorem 1.2. Let N≥1N \ge1, 1<p<∞1 < p < \infty, and let Ω⊂RN\Omega\subset\mathbb{R}^N be a bounded nonempty open set. If B⊂RNB \subset\mathbb{R}^N is a ball with ∣B∣=∣Ω∣|B| = |\Omega|, then

λ1,L,p(B)≤λ1,L,p(Ω).(10)\lambda_{1,L,p}(B) \le\lambda_{1,L,p}(\Omega). \tag*{(10)}

Equality holds if and only if ∣Ω△(B+x0)∣=0|\Omega\mathbin{\triangle} (B + x_0)| = 0 for some x0∈RNx_0 \in\mathbb{R}^N. In particular, if Ω\Omega has Lipschitz boundary, equality holds if and only if Ω\Omega is a ball.

For general open sets, equality of the domains must be understood up to null sets: both the space in (8) and the quotient in (9) are unchanged by null modifications of the domain. The Lipschitz assumption in the last assertion allows us to recover equality of the domains themselves.

The equality case in Theorem 1.1 rests on a comparison of regularized interaction deficits. For u≥0u \ge0, define

Jε(u):=∬RN×RNKε(x−y)Hp(u(x),u(y)) dx dy.J_\varepsilon(u) := \iint_{\mathbb{R}^N \times\mathbb{R}^N} K_\varepsilon(x-y)H_p(u(x),u(y)) \, \mathrm{d}x \, \mathrm{d}y.

Under the hypotheses of Theorem 1.1, these interactions are finite for both uu and u∗u^*. We prove that

EL,p(u)−EL,p(u∗)=CN,p2lim⁡ε↓0(Jε(u∗)−Jε(u)).(11)\mathcal{E}_{L,p}(u) - \mathcal{E}_{L,p}(u^*) = \frac{C_{N,p}}{2}\lim_{\varepsilon\downarrow0}\bigl(J_\varepsilon(u^*) - J_\varepsilon(u)\bigr). \tag*{(11)}

The deficits on the right-hand side are nonnegative and increase as ε↓0\varepsilon\downarrow0. Hence the energy deficit bounds the Riesz deficit at each fixed scale from below. When the energy deficit vanishes, we may apply the strict rearrangement theorem at any one of these scales.

The paper is organized as follows. Section 2 establishes the long-range estimates and characterizes finiteness of the long-range term after rearrangement. Section 3 proves the regularized rearrangement inequality and preservation of Xp(RN)X^p(\mathbb{R}^N). Sections 4 and 5 prove Theorems 1.1 and 1.2, respectively. Section 6 discusses sharpness of the logarithmic condition, the equality case at p=1p = 1, and the geometric and scaling conclusions of Theorem 1.2. Appendix § 7 presents an alternative proof of Theorem 1.1 by polarization.

Long-range integrability

We first establish the long-range estimates needed to define the energy in (4). Since ∣z∣−N|z|^{-N} is not integrable at infinity, the logarithmic condition is used to control the contribution of small values of the function. Throughout this section, CC denotes a positive constant that may vary from line to line; its dependence is specified in each estimate.

For a,b≥0a,b \ge0, let

Hp(a,b):=ap+bp−∣a−b∣p,G(z):=min⁡{1,∣z∣−N},(12)H_p(a,b) := a^p + b^p - |a-b|^p,\qquad G(z) := \min\{1,|z|^{-N}\}, \tag*{(12)}

where G(0)=1G(0)=1. Then 0≤Hp(a,b)≤ap+bp0 \le H_p(a,b) \le a^p+b^p and Hp(a,0)=0H_p(a,0)=0. For 0≤b≤a0 \le b \le a, we also have

ap−1b≤Hp(a,b)≤(p+1)ap−1b.(13)a^{p-1}b \le H_p(a,b) \le(p+1)a^{p-1}b. \tag*{(13)}

The lower bound follows from (1−t)p≤1−t(1-t)^p \le1-t for 0≤t≤10 \le t \le1. The upper bound follows from the mean value theorem and bp≤ap−1bb^p \le a^{p-1}b.

To treat real-valued functions, we shall also use the following consequence of (13): for all a,b∈Ra,b \in\mathbb{R},

∣a−b∣p−∣a∣p−∣b∣p≤CpHp(∣a∣,∣b∣).(14)\lvert a-b\rvert^p-\lvert a\rvert^p-\lvert b\rvert^p \le C_p H_p(\lvert a\rvert,\lvert b\rvert). \tag*{(14)}

If ab>0ab>0, the left-hand side equals Hp(∣a∣,∣b∣)H_p(\lvert a\rvert,\lvert b\rvert). If ab<0ab<0, set A=max⁡{∣a∣,∣b∣}A=\max\{\lvert a\rvert,\lvert b\rvert\} and B=min⁡{∣a∣,∣b∣}B=\min\{\lvert a\rvert,\lvert b\rvert\}. By the mean value theorem,

0≤(A+B)p−Ap−Bp≤p2p−1Ap−1B≤p2p−1Hp(A,B).0 \le(A+B)^p-A^p-B^p \le p2^{p-1}A^{p-1}B \le p2^{p-1}H_p(A,B).

Thus (14) holds with Cp=max⁡{1,p2p−1}C_p=\max\{1,p2^{p-1}\}.

We begin with an estimate on sets of finite measure. Its logarithmic dependence on the measure will be used in the superlevel-set decomposition below.

Lemma 2.1. There exists C=C(N)>0C=C(N)>0 such that, for every measurable set E⊂RNE \subset\mathbb{R}^N of finite measure,

sup⁡x∈RN∫EG(x−y) dy≤Clog⁡(e+∣E∣).(15)\sup_{x\in\mathbb{R}^N}\int_E G(x-y)\,\mathrm{d}y \le C\log(e+\lvert E\rvert). \tag*{(15)}

Proof. Set R=max⁡{1,∣E∣1/N}R=\max\{1,\lvert E\rvert^{1/N}\}. By the definition of GG,

∫EG(x−y) dy≤∫BRG(z) dz+R−N∣E∣≤∣B1∣+∣SN−1∣log⁡R+1.\int_E G(x-y)\,\mathrm{d}y \le\int_{B_R}G(z)\,\mathrm{d}z+R^{-N}\lvert E\rvert\le\lvert B_1\rvert+\lvert S^{N-1}\rvert\log R+1.

The right-hand side is bounded by Clog⁡(e+∣E∣)C\log(e+\lvert E\rvert), independently of xx. This proves (15). □\square

We next estimate the interaction against GG in terms of Sp(v)S_p(v). Since the bound depends only on the distribution of vv, it remains valid after rearrangement.

Lemma 2.2. Let v∈Lp(RN)v \in L^p(\mathbb{R}^N) be nonnegative and satisfy Sp(v)<∞S_p(v)<\infty. Then

JG(v):=∬RN×RNG(x−y)Hp(v(x),v(y)) dx dy<∞.(16)J_G(v):=\iint_{\mathbb{R}^N\times\mathbb{R}^N}G(x-y)H_p(v(x),v(y))\,\mathrm{d}x\,\mathrm{d}y<\infty. \tag*{(16)}

Moreover, if ∥v∥Lp(RN)=1\lVert v\rVert_{L^p(\mathbb{R}^N)}=1, then

JG(v)≤C(1+Sp(v)),(17)J_G(v)\le C(1+S_p(v)), \tag*{(17)}

where C=C(N,p)>0C=C(N,p)>0.

Proof. The assertion is immediate if v=0v=0. By the homogeneity of HpH_p and the estimate following (3), we may assume that ∥v∥Lp(RN)=1\lVert v\rVert_{L^p(\mathbb{R}^N)}=1. Let μ(t):=∣{v>t}∣\mu(t):=\lvert\{v>t\}\rvert. By Chebyshev’s inequality, we have μ(t)≤t−p\mu(t)\le t^{-p} for t>0t>0. For a>0a>0 and x∈RNx\in\mathbb{R}^{N}, Tonelli’s theorem and Lemma 2.1 yield

∫{0<v≤a}G(x−y)v(y) dy≤∫0a∫{v>t}G(x−y) dy dt≤C∫0alog⁡(e+μ(t)) dt≤C∫0a(1+log⁡−t) dt≤Ca(1+log⁡−a).(18)\begin{aligned} \int_{\{0<v\le a\}}G(x-y)v(y)\,dy &\le\int_{0}^{a}\int_{\{v>t\}}G(x-y)\,dy\,dt \\ &\le C\int_{0}^{a}\log(e+\mu(t))\,dt \\ &\le C\int_{0}^{a}(1+\log^{-}t)\,dt \\ &\le Ca(1+\log^{-}a). \tag*{(18)} \end{aligned}

The last inequality follows by direct integration, treating (0,1)(0,1) and (1,a)(1,a) separately when a>1a>1.

By symmetry, (13), and (18) with a=v(x)a=v(x), we have

JG(v)≤2(p+1)∫RNv(x)p−1∫{0<v≤v(x)}G(x−y)v(y) dy dx≤C∫RNv(x)p(1+log⁡−v(x)) dx=C(1+Sp(v)).\begin{aligned} J_G(v)\le2(p+1)\int_{\mathbb{R}^{N}}v(x)^{p-1}\int_{\{0<v\le v(x)\}}G(x-y)v(y)\,dy\,dx \\ &\le C\int_{\mathbb{R}^{N}}v(x)^p(1+\log^{-}v(x))\,dx \\ &=C(1+S_p(v)). \end{aligned}

This proves (17) and completes the proof. □\square

By Lemma 2.2 and (14), the long-range integral in (4) is absolutely convergent for u∈Xp(RN)u\in X^p(\mathbb{R}^{N}) with Sp(u)<∞S_p(u)<\infty. We shall also use a translation formula for the energy. For w∈Lp(RN)w\in L^p(\mathbb{R}^{N}), set

Mw:=∥w∥Lp(RN)p,τw(z):=∫RN∣w(x+z)−w(x)∣p dx.(19)M_w:=\|w\|_{L^p(\mathbb{R}^{N})}^{p},\qquad\tau_w(z):=\int_{\mathbb{R}^{N}}|w(x+z)-w(x)|^p\,dx. \tag*{(19)}

Then 0≤τw(z)≤2pMw0\le\tau_w(z)\le2^pM_w. For nonnegative ww,

Iw(z):=∫RNHp(w(x),w(x+z)) dx=2Mw−τw(z)≥0.(20)I_w(z):=\int_{\mathbb{R}^{N}}H_p(w(x),w(x+z))\,dx=2M_w-\tau_w(z)\ge0. \tag*{(20)}

Let w∈Xp(RN)w\in X^p(\mathbb{R}^{N}) be real-valued and satisfy Sp(w)<∞S_p(w)<\infty. By the preceding integrability estimates, we may make the change of variables y=x+zy=x+z in (4) and obtain

EL,p(w)=CN,p2∫∣z∣<1τw(z)∣z∣N dz+CN,p2∫∣z∣≥1τw(z)−2Mw∣z∣N dz+ρN,pMw.(21)\mathcal{E}_{L,p}(w)=\frac{C_{N,p}}{2}\int_{|z|<1}\frac{\tau_w(z)}{|z|^N}\,dz+\frac{C_{N,p}}{2}\int_{|z|\ge1}\frac{\tau_w(z)-2M_w}{|z|^N}\,dz+\rho_{N,p}M_w. \tag*{(21)}

In particular, the energy is invariant under translations and multiplication by −1-1.

For rearranged functions, the logarithmic condition is also necessary. The following result requires only LpL^p-integrability.

Proposition 2.3. Let u∈Lp(RN)u\in L^p(\mathbb{R}^{N}) be real-valued, and put v=u#v=u^\#. Define

Tp(v):=∬∣x−y∣≥1Hp(v(x),v(y))∣x−y∣N dx dy∈[0,∞].(22)T_p(v):=\iint_{|x-y|\ge1}\frac{H_p(v(x),v(y))}{|x-y|^N}\,dx\,dy\in[0,\infty]. \tag*{(22)}

The following three conditions are equivalent:

Sp(u)<∞,∫RNv(x)plog⁡(e+∣x∣) dx<∞,Tp(v)<∞.(23)S_p(u)<\infty,\qquad\int_{\mathbb{R}^N}v(x)^p\log(e+|x|)\,dx<\infty,\qquad T_p(v)<\infty. \tag*{(23)}

Proof. By equimeasurability, Sp(v)=Sp(u)S_p(v)=S_p(u). The bound Tp(v)≤JG(v)T_p(v)\le J_G(v) and Lemma 2.2 show that the first condition implies the third.

Suppose that the third condition holds. Choose a radial nonincreasing representative of vv. For x∈RNx\in\mathbb{R}^N with r=∣x∣≥4r=|x|\ge4, let

Ax:={z∈RN:1≤∣z∣≤r2,x⋅z≤−12r∣z∣}.A_x:=\left\{z\in\mathbb{R}^N:1\le|z|\le\frac{r}{2},\quad x\cdot z\le-\frac{1}{2}r|z|\right\}.

For z∈Axz\in A_x, one has ∣x+z∣≤∣x∣|x+z|\le|x|, hence v(x+z)≥v(x)v(x+z)\ge v(x). It follows that

Hp(v(x),v(x+z))≥v(x)p.H_p(v(x),v(x+z))\ge v(x)^p.

The set of directions defining AxA_x has positive measure depending only on NN; when N=1N=1, it consists of the direction towards the origin. Consequently,

∫Axdz∣z∣N=cNlog⁡(r/2)≥cN2log⁡r.\int_{A_x}\frac{dz}{|z|^N}=c_N\log(r/2)\ge\frac{c_N}{2}\log r.

By Tonelli’s theorem, we have

Tp(v)≥c∫∣x∣≥4v(x)plog⁡∣x∣ dx,(24)T_p(v)\ge c\int_{|x|\ge4}v(x)^p\log|x|\,dx, \tag*{(24)}

where c=c(N)>0c=c(N)>0. Since v∈Lp(RN)v\in L^p(\mathbb{R}^N), the logarithmically weighted integral over ∣x∣<4|x|<4 is finite. Thus the second condition holds.

It remains to show that the second condition implies the first. For t,r≥0t,r\ge0,

tplog⁡−t≤N+1ptplog⁡(e+r)+1pe(e+r)−N−1.(25)t^p\log^{-}t\le\frac{N+1}{p}t^p\log(e+r)+\frac{1}{pe}(e+r)^{-N-1}. \tag*{(25)}

For t≥1t\ge1, this is immediate. For 0<t<10<t<1, set a=tpa=t^p and b=(e+r)−N−1b=(e+r)^{-N-1}. Since slog⁡(1/s)≤1/es\log(1/s)\le1/e for s>0s>0,

alog⁡1a=alog⁡1b+baalog⁡ba≤(N+1)alog⁡(e+r)+be.a\log\frac{1}{a}=a\log\frac{1}{b}+\frac{b}{a}a\log\frac{b}{a}\le(N+1)a\log(e+r)+\frac{b}{e}.

After division by pp, this is (25). The inequality also holds at t=0t=0 by continuity. Applying it with t=v(x)t=v(x) and r=∣x∣r=|x| and integrating over RN\mathbb{R}^N, we obtain Sp(v)<∞S_p(v)<\infty. □

The next example shows both that the logarithmic condition cannot be omitted and that it is weaker than additional LqL^q-integrability for q<pq<p.

Remark 2.4. For α>1\alpha>1, set

vα(x):=1(1+∣x∣2)N/(2p)[log⁡(e+∣x∣2)]α/p.(26)v_\alpha(x):=\frac{1}{(1+|x|^2)^{N/(2p)}[\log(e+|x|^2)]^{\alpha/p}}. \tag*{(26)}

Then vαv_\alpha is positive, smooth and radially nonincreasing. As r=∣x∣→∞r=|x|\to\infty,

vα(x)p≍r−N(log⁡r)−α,∣∇vα(x)∣p≤Cr−N−p(log⁡r)−α.v_\alpha(x)^p\asymp r^{-N}(\log r)^{-\alpha},\qquad|\nabla v_\alpha(x)|^p\le Cr^{-N-p}(\log r)^{-\alpha}.

Thus vα∈W1,p(RN)⊂Xp(RN)v_{\alpha} \in W^{1,p}(\mathbb{R}^{N}) \subset X^{p}(\mathbb{R}^{N}). The latter inclusion follows by integrating τw(z)≤∣z∣p∥∇w∥Lp(RN)p\tau_{w}(z) \leq|z|^{p}\|\nabla w\|_{L^{p}(\mathbb{R}^{N})}^{p} over B1B_{1} against ∣z∣−N|z|^{-N}. Moreover, vα∉Lq(RN)v_{\alpha} \notin L^{q}(\mathbb{R}^{N}) for every 0<q<p0 < q < p, whereas

Sp(vα)<∞⟺∫∞drr(log⁡r)α−1<∞⟺α>2.S_{p}(v_{\alpha}) < \infty\quad\Longleftrightarrow\quad\int^{\infty} \frac{\mathrm{d}r}{r(\log r)^{\alpha-1}} < \infty\quad\Longleftrightarrow\quad\alpha> 2.

Hence Theorem 1.1 applies when α>2\alpha> 2. If 1<α≤21 < \alpha\leq2, then Tp(vα)=∞T_{p}(v_{\alpha}) = \infty by Proposition 2.3, and the expression in (4) equals −∞-\infty.

A regularized rearrangement inequality

We now apply the strict rearrangement theorem for supermodular integrands to HpH_{p}. The resulting comparison will also be used with truncated kernels to prove that Schwarz rearrangement preserves Xp(RN)X^{p}(\mathbb{R}^{N}).

The interaction HpH_{p} in (12) is strictly supermodular. More precisely, for a,b≥0a,b \geq0 and h,k>0h,k > 0,

Hp(a+h,b+k)+Hp(a,b)−Hp(a+h,b)−Hp(a,b+k)=p(p−1)∫0h∫0k∣a−b+s−t∣p−2 dt ds>0.(27)\begin{aligned} &H_{p}(a+h,b+k)+H_{p}(a,b)-H_{p}(a+h,b)-H_{p}(a,b+k) \\ &=p(p-1)\int_{0}^{h}\int_{0}^{k}|a-b+s-t|^{p-2}\,\mathrm{d}t\,\mathrm{d}s>0. \tag*{(27)} \end{aligned}

Indeed, the derivative of t↦∣t∣pt \mapsto|t|^{p} is locally absolutely continuous for every p>1p>1. Two applications of the fundamental theorem of calculus imply (27). When 1<p<21<p<2, the possible singularity at a−b+s−t=0a-b+s-t=0 is locally integrable, so the argument remains valid.

For a bounded nonnegative nonincreasing function κ:[0,∞)→[0,∞)\kappa:[0,\infty)\to[0,\infty) and a nonnegative w∈Lp(RN)w\in L^{p}(\mathbb{R}^{N}), define

Jκ(w):=∬RN×RNκ(∣x−y∣)Hp(w(x),w(y)) dx dy∈[0,∞].(28)J_{\kappa}(w):=\iint_{\mathbb{R}^{N}\times\mathbb{R}^{N}}\kappa(|x-y|)H_{p}(w(x),w(y))\,\mathrm{d}x\,\mathrm{d}y\in[0,\infty]. \tag*{(28)}

The following proposition is the form of the strict rearrangement theorem of Burchard and Hajaiej [6 Theorem 2] needed here.

Proposition 3.1. Let w∈Lp(RN)w\in L^{p}(\mathbb{R}^{N}) be nonnegative, and let κ:[0,∞)→[0,∞)\kappa:[0,\infty)\to[0,\infty) be bounded and nonincreasing. Then

Jκ(w)≤Jκ(w∗).(29)J_{\kappa}(w)\leq J_{\kappa}(w^{*}). \tag*{(29)}

Suppose, in addition, that w≢0w\not\equiv0, that κ\kappa is positive and strictly decreasing, and that Jκ(w∗)<∞J_{\kappa}(w^{*})<\infty. Equality holds in (29) if and only if w=w∗(⋅−x0)w=w^{*}(\mathord{\cdot}-x_{0}) almost everywhere for some x0∈RNx_{0}\in\mathbb{R}^{N}.

Proof. Since w∈Lp(RN)w\in L^{p}(\mathbb{R}^{N}), all its positive superlevel sets have finite measure. By (27), HpH_{p} is strictly supermodular; it is also continuous, nonnegative and vanishes when either variable is zero. We may therefore apply [6 Theorem 2] with m=2m=2, u1=u2=wu_{1}=u_{2}=w, F=HpF=H_{p}, and K12=κK_{12}=\kappa to obtain (29).

Under the additional assumptions, both interactions are finite, and the nonzero LpL^{p} function ww cannot be constant almost everywhere on RN\mathbb{R}^{N}. If equality holds in (29), the equality assertion of the same theorem implies that ww is a translate of w∗w^{*}. The converse follows from translation invariance. □

Applying Proposition 3.1 to integrable truncated kernels, we obtain the next comparison of the short-range terms. This step requires no logarithmic integrability assumption.

Lemma 3.2. Let w∈Xp(RN)w \in X^{p}(\mathbb{R}^{N}) be real-valued. Then w#∈Xp(RN)w^{\#} \in X^{p}(\mathbb{R}^{N}) and

[w#]Xp≤[∣w∣]Xp≤[w]Xp.(30)[w^{\#}]_{X^{p}} \leq[|w|]_{X^{p}} \leq[w]_{X^{p}}. \tag*{(30)}

Proof. The second inequality follows from the reverse triangle inequality. We may therefore assume that w≥0w \geq0. For 0<δ<10 < \delta< 1, let

kδ(z):=1{∣z∣<1}min⁡{δ−N,∣z∣−N}.k_{\delta}(z) := \mathbf{1}_{\{|z|<1\}} \min\{\delta^{-N}, |z|^{-N}\}.

The kernel kδk_{\delta} is bounded, radially nonincreasing and belongs to L1(RN)L^{1}(\mathbb{R}^{N}). Since Hp(a,b)≤ap+bpH_{p}(a,b) \leq a^{p}+b^{p}, both interactions against kδk_{\delta} are bounded by 2Mw∥kδ∥12M_{w}\|k_{\delta}\|_{1}. By Proposition 3.1 and (20), we have

∫RNkδ(z)(2Mw−τw(z)) dz≤∫RNkδ(z)(2Mw−τw∗(z)) dz.\int_{\mathbb{R}^{N}} k_{\delta}(z)(2M_{w}-\tau_{w}(z))\,\mathrm{d}z \leq\int_{\mathbb{R}^{N}} k_{\delta}(z)(2M_{w}-\tau_{w^{*}}(z))\,\mathrm{d}z.

where Mw∗=MwM_{w^{*}}=M_{w}. The common mass term is finite, so

∫RNkδ(z)τw∗(z) dz≤∫RNkδ(z)τw(z) dz.\int_{\mathbb{R}^{N}} k_{\delta}(z)\tau_{w^{*}}(z)\,\mathrm{d}z \leq\int_{\mathbb{R}^{N}} k_{\delta}(z)\tau_{w}(z)\,\mathrm{d}z.

Letting δ↓0\delta\downarrow0 and applying the monotone convergence theorem proves (30).

For u∈Xp(RN)u \in X^{p}(\mathbb{R}^{N}), the short-range term of u#u^{\#} is finite by Lemma 3.2, whereas its long-range term is −CN,pTp(u#)/2-C_{N,p}T_{p}(u^{\#})/2. Proposition 2.3 therefore proves (7).

For ε>0\varepsilon> 0, set

κε(r):=(r2+ε2)−N/2,Kε(z):=κε(∣z∣),Jε:=Jκε.(31)\kappa_{\varepsilon}(r) := (r^{2}+\varepsilon^{2})^{-N/2}, \qquad K_{\varepsilon}(z) := \kappa_{\varepsilon}(|z|), \qquad J_{\varepsilon}:=J_{\kappa_{\varepsilon}}. \tag*{(31)}

By definition,

Kε(z)≤max⁡{1,ε−N}G(z).(32)K_{\varepsilon}(z) \leq\max\{1,\varepsilon^{-N}\}G(z). \tag*{(32)}

Lemma 2.2 and equimeasurability imply that Jε(w)J_{\varepsilon}(w) and Jε(w∗)J_{\varepsilon}(w^{*}) are finite whenever w∈Lp(RN)w \in L^{p}(\mathbb{R}^{N}) is nonnegative and Sp(w)<∞S_{p}(w)<\infty. By Tonelli’s theorem and (20), we have

Jε(w)=∫RNKε(z)Iw(z) dz=∫RNKε(z)(2Mw−τw(z)) dz.(33)J_{\varepsilon}(w)=\int_{\mathbb{R}^{N}}K_{\varepsilon}(z)I_{w}(z)\,\mathrm{d}z=\int_{\mathbb{R}^{N}}K_{\varepsilon}(z)(2M_{w}-\tau_{w}(z))\,\mathrm{d}z. \tag*{(33)}

The cancellation in 2Mw−τw(z)2M_{w}-\tau_{w}(z) is needed in this identity, since the two terms need not be integrable separately.

Proof of Theorem 1.1

We first treat nonnegative functions. Using the estimates of § 2, we express the energy difference as the limit of regularized interaction deficits. Their monotonicity will allow us to pass the equality information to the limit.

For nonnegative u∈Xp(RN)u \in X^{p}(\mathbb{R}^{N}) with Sp(u)<∞S_p(u)<\infty, define

Dε(u):=Jε(u∗)−Jε(u),ε>0.(34)D_{\varepsilon}(u):=\mathcal{J}_{\varepsilon}(u^{*})-\mathcal{J}_{\varepsilon}(u), \qquad\varepsilon>0. \tag*{(34)}

By Lemma 2.2 and (32), both interactions are finite, and Dε(u)≥0D_{\varepsilon}(u)\geq0 by Proposition 3.1. The next lemma identifies the limit of these deficits and proves their monotonicity.

Lemma 4.1. Let u∈Xp(RN)u\in X^{p}(\mathbb{R}^{N}) be nonnegative and satisfy Sp(u)<∞S_p(u)<\infty. Then

EL,p(u)−EL,p(u∗)=CN,p2∫RNτu(z)−τu∗(z)∣z∣N dz(35)\mathcal{E}_{L,p}(u)-\mathcal{E}_{L,p}(u^{*})=\frac{C_{N,p}}{2}\int_{\mathbb{R}^{N}}\frac{\tau_u(z)-\tau_{u^{*}}(z)}{|z|^{N}}\,\mathrm{d}z \tag*{(35)}
=CN,p2lim⁡ε→0Dε(u),=\frac{C_{N,p}}{2}\lim_{\varepsilon\to0}D_{\varepsilon}(u),

where the integral is absolutely convergent. Furthermore, if 0<ε1<ε20<\varepsilon_1<\varepsilon_2, then

Dε1(u)≥Dε2(u)≥0.(36)D_{\varepsilon_1}(u)\geq D_{\varepsilon_2}(u)\geq0. \tag*{(36)}

Both inequalities in (36) are strict unless uu is a translate of u∗u^{*}.

Proof. By Lemma 3.2, we have u∗∈Xp(RN)u^{*}\in X^{p}(\mathbb{R}^{N}). Since Sp(u∗)=Sp(u)S_p(u^{*})=S_p(u), both energies in (35) are finite. Near the origin,

∣τu(z)−τu∗(z)∣≤τu(z)+τu∗(z),|\tau_u(z)-\tau_{u^{*}}(z)|\leq\tau_u(z)+\tau_{u^{*}}(z),

and integrability follows from u,u∗∈Xp(RN)u,u^{*}\in X^{p}(\mathbb{R}^{N}). For ∣z∣≥1|z|\geq1, (20) and Mu=Mu∗M_u=M_{u^{*}} yield

∣τu(z)−τu∗(z)∣=∣Iu∗(z)−Iu(z)∣≤Iu(z)+Iu∗(z).|\tau_u(z)-\tau_{u^{*}}(z)|=|I_{u^{*}}(z)-I_u(z)|\leq I_u(z)+I_{u^{*}}(z).

The right-hand side is integrable against ∣z∣−N|z|^{-N} on ∣z∣≥1|z|\geq1 by Lemma 2.2. Subtracting the two identities in (21) therefore proves the first equality in (35).

By (33), we have

Dε(u)=∫RNKε(z)(τu(z)−τu∗(z)) dz.(37)D_{\varepsilon}(u)=\int_{\mathbb{R}^{N}}K_{\varepsilon}(z)\bigl(\tau_u(z)-\tau_{u^{*}}(z)\bigr)\,\mathrm{d}z. \tag*{(37)}

Since Kε(z)≤∣z∣−NK_{\varepsilon}(z)\leq|z|^{-N} for z≠0z\ne0, the dominated convergence theorem proves the second equality in (35).

To prove (36), let 0<ε1<ε20<\varepsilon_1<\varepsilon_2 and set

ℓ(r):=(r2+ε12)−N/2−(r2+ε22)−N/2.\ell(r):=(r^2+\varepsilon_1^2)^{-N/2}-(r^2+\varepsilon_2^2)^{-N/2}.

The function ℓ\ell is bounded and positive. Moreover, for r>0r>0,

ℓ′(r)=−Nr((r2+ε12)−N/2−1−(r2+ε22)−N/2−1)<0.\ell'(r)=-Nr\bigl((r^2+\varepsilon_1^2)^{-N/2-1}-(r^2+\varepsilon_2^2)^{-N/2-1}\bigr)<0.

Since 0<ℓ≤κε10<\ell\leq\kappa_{\varepsilon_1}, both interactions against ℓ\ell are finite. By Proposition 3.1, we have

Dε1(u)−Dε2(u)=Jℓ(u∗)−Jℓ(u)≥0.D_{\varepsilon_1}(u)-D_{\varepsilon_2}(u)=\mathcal{J}_{\ell}(u^{*})-\mathcal{J}_{\ell}(u)\geq0.

Applying the equality assertion of the same proposition to ℓ\ell and κε2\kappa_{\varepsilon_2} shows that both inequalities in (36) are strict unless uu is a translate of u∗u^*.

We can now prove Theorem 1.1. Lemma 4.1 settles the nonnegative case; a comparison with the modulus then handles real-valued functions and determines the sign of the equality cases.

Proof of Theorem 1.1. The assertion is immediate if u=0u=0, so assume that u≢0u\not\equiv0. Suppose first that u≥0u\geq0. Lemmas 3.2 and 4.1 yield u∗∈Xp(RN)u^*\in X^p(\mathbb{R}^N) and

EL,p(u)−EL,p(u∗)=CN,p2lim⁡ε↓0Dε(u)≥0.\mathcal{E}_{L,p}(u)-\mathcal{E}_{L,p}(u^*)=\frac{C_{N,p}}{2}\lim_{\varepsilon\downarrow0}D_{\varepsilon}(u)\geq0.

If equality holds, then by (36), for every ε0>0\varepsilon_0>0 we have

0≤Dε0(u)≤lim⁡ε↓0Dε(u)=0.0\leq D_{\varepsilon_0}(u)\leq\lim_{\varepsilon\downarrow0}D_{\varepsilon}(u)=0.

Proposition 3.1 implies that u=u∗(⋅−x0)u=u^*(\cdot-x_0) almost everywhere for some x0∈RNx_0\in\mathbb{R}^N. The converse follows from translation invariance.

Now let uu be real-valued. By the reverse triangle inequality, ∣u∣∈Xp(RN)|u|\in X^p(\mathbb{R}^N), and clearly Sp(∣u∣)=Sp(u)S_p(|u|)=S_p(u). Both energies are finite by Lemma 2.2 and (14). Subtracting their expressions in (4) yields

EL,p(u)−EL,p(∣u∣)=CN,p2∬RN×RN∣u(x)−u(y)∣p−∣∣u(x)∣−∣u(y)∣∣p∣x−y∣N dx dy≥0.(38)\begin{aligned} \mathcal{E}_{L,p}(u)-\mathcal{E}_{L,p}(|u|) &=\frac{C_{N,p}}{2}\iint_{\mathbb{R}^N\times\mathbb{R}^N} \frac{|u(x)-u(y)|^p-\bigl||u(x)|-|u(y)|\bigr|^p}{|x-y|^N}\,dx\,dy\geq0. \tag*{(38)} \end{aligned}

The integrand is nonnegative. For ∣x−y∣<1|x-y|<1, it is bounded by the short-range integrand of uu. For ∣x−y∣≥1|x-y|\geq1, integrability follows from (14) and Lemma 2.2.

If both {u>0}\{u>0\} and {u<0}\{u<0\} have positive measure, the numerator in (38) is strictly positive on their product. Thus equality in (38) holds if and only if uu does not change sign. Together with the nonnegative case, this implies

EL,p(u)≥EL,p(∣u∣)≥EL,p((∣u∣)∗)=EL,p(u#).\mathcal{E}_{L,p}(u)\geq\mathcal{E}_{L,p}(|u|)\geq\mathcal{E}_{L,p}((|u|)^*)=\mathcal{E}_{L,p}(u^\#).

If equality holds, uu has one sign and ∣u∣|u| is a translate of u#u^\#. Conversely, these properties imply equality by invariance of the energy under translations and multiplication by −1-1. □

The preceding proof also controls the deficit at each fixed regularization scale.

Remark 4.2. Let u∈Xp(RN)u\in X^p(\mathbb{R}^N) be nonnegative and satisfy Sp(u)<∞S_p(u)<\infty. By Lemma 4.1, for every ε0>0\varepsilon_0>0 we have

EL,p(u)−EL,p(u∗)≥CN,p2(Jε0(u∗)−Jε0(u)).(39)\mathcal{E}_{L,p}(u)-\mathcal{E}_{L,p}(u^*)\geq\frac{C_{N,p}}{2}\bigl(J_{\varepsilon_0}(u^*)-J_{\varepsilon_0}(u)\bigr). \tag*{(39)}

The right-hand side is positive unless uu is a translate of u∗u^*.

The assumption p>1p > 1 is used in the strict supermodularity of HpH_p in (27). At p=1p = 1, one has H1(a,b)=2min⁡{a,b}H_1(a,b) = 2\min\{a,b\}, which is not strictly supermodular, so the preceding equality argument does not apply.

Proof of Theorem 1.2

By [9 Theorem 7.1], the infimum in (9) is attained by a function ϕΩ∈X0p(Ω)\phi_\Omega\in X_0^p(\Omega) satisfying

∥ϕΩ∥Lp(RN)=1,ϕΩ>0a.e. in Ω.(40)\|\phi_\Omega\|_{L^p(\mathbb{R}^N)} = 1,\qquad\phi_\Omega> 0\quad\text{a.e. in }\Omega. \tag*{(40)}

These conclusions hold without any regularity assumption on ∂Ω\partial\Omega. Since ϕΩ\phi_\Omega vanishes almost everywhere outside the bounded set Ω\Omega, we also have Sp(ϕΩ)<∞S_p(\phi_\Omega) < \infty by (6). Thus Theorem 1.1 applies to ϕΩ\phi_\Omega.

Proof of Theorem 1.2. By translation invariance, we may assume that BB is centered at the origin. Let ϕΩ\phi_\Omega be the normalized positive first eigenfunction in (40), extended by zero outside Ω\Omega. By positivity and equimeasurability, we have

∣{ϕΩ>0}△Ω∣=0,∣{ϕΩ∗>0}△B∣=0.\lvert\{\phi_\Omega> 0\} \mathbin{\triangle} \Omega\rvert= 0,\qquad\lvert\{\phi_\Omega^* > 0\} \mathbin{\triangle} B\rvert= 0.

Therefore ϕΩ∗∈X0p(B)\phi_\Omega^* \in X_0^p(B) and ∥ϕΩ∗∥Lp(RN)=1\|\phi_\Omega^*\|_{L^p(\mathbb{R}^N)} = 1. By Theorem 1.1,

λ1,L,p(B)≤EL,p(ϕΩ∗)≤EL,p(ϕΩ)=λ1,L,p(Ω).(41)\lambda_{1,L,p}(B) \le\mathcal{E}_{L,p}(\phi_\Omega^*) \le\mathcal{E}_{L,p}(\phi_\Omega) = \lambda_{1,L,p}(\Omega). \tag*{(41)}

This proves (10).

If equality holds, it holds throughout (41), and Theorem 1.1 implies that ϕΩ=ϕΩ∗(⋅−x0)\phi_\Omega= \phi_\Omega^*(\mathord{\cdot} - x_0) almost everywhere for some x0∈RNx_0 \in\mathbb{R}^N. Taking the positivity sets and using (40), we obtain

∣Ω△(B+x0)∣=0.\lvert\Omega\mathbin{\triangle} (B + x_0)\rvert= 0.

Conversely, if this identity holds, then X0p(Ω)=X0p(B+x0)X_0^p(\Omega) = X_0^p(B + x_0) as spaces of almost-everywhere equivalence classes. Translation invariance yields λ1,L,p(Ω)=λ1,L,p(B)\lambda_{1,L,p}(\Omega) = \lambda_{1,L,p}(B).

Finally, assume that Ω\Omega has Lipschitz boundary and set D=B+x0D = B + x_0. Suppose that x∈D∖Ωx \in D \setminus\Omega. Since DD is open, a sufficiently small ball centered at xx lies in DD. If xx is an interior point of Ωc\Omega^c, then ∣D∖Ω∣>0\lvert D \setminus\Omega\rvert> 0. If x∈∂Ωx \in\partial\Omega, the same conclusion follows from the local Lipschitz representation of the boundary. Both cases contradict ∣Ω△D∣=0\lvert\Omega\mathbin{\triangle} D\rvert= 0, so D⊂ΩD \subset\Omega. By the same argument, with Ω\Omega and DD interchanged and using the smooth boundary of DD, we obtain Ω⊂D\Omega\subset D. Hence Ω=D\Omega= D. □\square

Sharpness of the main theorems

We conclude the main part of the paper with several observations on sharpness. They concern the logarithmic condition, the assumption p>1p > 1 in the equality case, and the dependence of the spectral bound on the domain and its volume.

The logarithmic finiteness condition

By Proposition 2.3 and Lemma 3.2, for every real-valued u∈Xp(RN)u \in X^p(\mathbb{R}^N),

EL,p(u#)∈R⟺Sp(u)<∞.(42)\mathcal{E}_{L,p}(u^\#) \in\mathbb{R} \quad\Longleftrightarrow\quad S_p(u) < \infty. \tag*{(42)}

Thus the logarithmic condition in Theorem 1.1 is necessary and sufficient for the rearranged energy to be finite. If it fails, the short-range term of u#u^\# remains finite, whereas its long-range term equals −∞-\infty.

The functions vαv_\alpha defined in (26) make the threshold explicit. For every α>1\alpha> 1, one has vα∈W1,p(RN)⊂Xp(RN)v_\alpha\in W^{1,p}(\mathbb{R}^N) \subset X^p(\mathbb{R}^N), while

Sp(vα)<∞⟺α>2.(43)S_p(v_\alpha) < \infty\quad\Longleftrightarrow\quad\alpha> 2. \tag*{(43)}

In particular, W1,pW^{1,p}-regularity alone does not ensure finiteness of the rearranged logarithmic energy.

The power of the logarithm cannot be decreased. Fix 0<β<10 < \beta< 1 and choose 1+β<α≤21+\beta< \alpha\le2. Since

vα(x)p≍r−N(log⁡r)−α,log⁡−vα(x)≍log⁡ras r=∣x∣→∞,v_\alpha(x)^p \asymp r^{-N}(\log r)^{-\alpha}, \qquad\log^{-} v_\alpha(x) \asymp\log r \quad\text{as } r=|x| \to\infty,

we have

∫RNvαp(log⁡−vα)β dx<∞,Sp(vα)=∞.(44)\int_{\mathbb{R}^N} v_\alpha^p(\log^{-} v_\alpha)^\beta\, dx < \infty, \qquad S_p(v_\alpha)=\infty. \tag*{(44)}

Indeed, in polar coordinates the tail of the first integral is comparable to

∫∞drr(log⁡r)α−β,\int^\infty\frac{dr}{r(\log r)^{\alpha-\beta}},

which is finite because α−β>1\alpha-\beta>1. Hence no smaller positive power of log⁡−∣u∣\log^{-}|u| can replace the logarithm in the finiteness criterion. The same family with 1<α≤21 < \alpha\le2 shows that LpL^p-integrability alone is insufficient.

For α>2\alpha> 2, the function vαv_\alpha satisfies all hypotheses of Theorem 1.1, although it is positive on RN\mathbb{R}^N and belongs to no Lq(RN)L^q(\mathbb{R}^N) with 0<q<p0 < q < p. These functions show that neither a restriction on the support nor additional LqL^q-integrability is needed in Theorem 1.1.

The equality case at p=1p=1

The assumption p>1p>1 is essential for the equality classification in Theorem 1.1. To see this, let EL,1\mathcal{E}_{L,1} denote the functional obtained by setting p=1p=1 in (1) and (4). Choose R>r>0R>r>0 and a≠0a\ne0 such that ∣a∣+r<R|a|+r<R, and set

E:=BR(0),F:=Br(a),u:=1E+1F.(45)E := B_R(0), \qquad F := B_r(a), \qquad u := \mathbf{1}_E+\mathbf{1}_F. \tag*{(45)}

Then F⊂EF \subset E and

u∗=1BR(0)+1Br(0).u^*=\mathbf{1}_{B_R(0)}+\mathbf{1}_{B_r(0)}.

Both functions have finite endpoint energy. Indeed, for a ball DD,

∫RN∣1D(x+z)−1D(x)∣ dx=∣D△(D−z)∣≤CDmin⁡{∣z∣,1}.\int_{\mathbb{R}^N} |\mathbf{1}_D(x+z)-\mathbf{1}_D(x)|\,dx = |D \mathbin{\triangle} (D-z)| \le C_D\min\{|z|,1\}.

Integrating over ∣z∣<1|z|<1 against ∣z∣−N|z|^{-N} proves short-range integrability. The corresponding term for uu is bounded by the sum of those for 1E\mathbf{1}_E and 1F\mathbf{1}_F.

The long-range term is finite because uu is bounded and compactly supported, and H1(a,b)=2min⁡{a,b}H_1(a,b)=2\min\{a,b\} for a,b≥0a,b\ge0.

For nonnegative uu, the layer-cake identity reads

∣u(x)−u(y)∣=∫0∞∣1{u>t}(x)−1{u>t}(y)∣ dt.|u(x)-u(y)|=\int_0^\infty\left|\mathbf{1}_{\{u>t\}}(x)-\mathbf{1}_{\{u>t\}}(y)\right|\,\mathrm{d}t.

In the present example, {u>t}=E\{u>t\}=E for 0<t<10<t<1 and {u>t}=F\{u>t\}=F for 1<t<21<t<2. The short-range terms therefore add, as do the mass terms and the renormalized long-range integrands. Consequently,

EL,1(u)=EL,1(1E)+EL,1(1F).(46)\mathcal{E}_{L,1}(u)=\mathcal{E}_{L,1}(\mathbf{1}_E)+\mathcal{E}_{L,1}(\mathbf{1}_F). \tag*{(46)}

By translation invariance,

EL,1(u)=EL,1(u∗).(47)\mathcal{E}_{L,1}(u)=\mathcal{E}_{L,1}(u^*). \tag*{(47)}

However, uu is not a translate of u∗u^*, since its two nontrivial superlevel sets are balls with different centers. Therefore equality does not imply a common center at p=1p=1, even for nonnegative compactly supported functions.

Equality cases and null modifications of the domain

Equality in Theorem 1.1 is attained by every admissible nonzero function σv(⋅−x0)\sigma v(\cdot-x_0), where vv is nonnegative and radially nonincreasing and σ∈{−1,1}\sigma\in\{-1,1\}. Hence the inequality cannot be improved by a positive additive constant on the whole admissible class. The sign condition is also necessary: by (38), EL,p(u)>EL,p(∣u∣)\mathcal{E}_{L,p}(u)>\mathcal{E}_{L,p}(|u|) whenever both {u>0}\{u>0\} and {u<0}\{u<0\} have positive measure.

Balls attain the bound in Theorem 1.2, so the spectral inequality admits no positive additive improvement at fixed volume. The qualification by null sets in its equality statement is necessary for general open sets. For example, let BB be a ball containing the origin and set

Ω:=B∖{0}.(48)\Omega:=B\setminus\{0\}. \tag*{(48)}

Then Ω\Omega is a bounded open set that is not a ball, but

∣Ω∣=∣B∣,Xp0(Ω)=Xp0(B),λ1,L,p(Ω)=λ1,L,p(B).|\Omega|=|B|,\qquad X_p^0(\Omega)=X_p^0(B),\qquad\lambda_{1,L,p}(\Omega)=\lambda_{1,L,p}(B).

Indeed, the zero-exterior condition in (8) is imposed almost everywhere. Removing a null set therefore leaves the admissible space unchanged. For domains with Lipschitz boundary, equality modulo null sets implies equality of the domains, as proved in § 5.

Dilation and volume dependence

To determine the dependence on volume in Theorem 1.2, we first compute the effect of dilation on the energy. For t>0t>0 and u∈Xp(RN)u\in X_p(\mathbb{R}^N) with Sp(u)<∞S_p(u)<\infty, set

ut(x):=t−N/pu(x/t).u_t(x):=t^{-N/p}u(x/t).

With the notation in (19), we have Mut=MuM_{u_t}=M_u and τut(z)=τu(z/t)\tau_{u_t}(z)=\tau_u(z/t). Dilation preserves Xp(RN)X_p(\mathbb{R}^N), since changing the cutoff radius in the defining integral adds or removes only an annulus, on which τu≤2pMu\tau_u \le2^p M_u. It also preserves the logarithmic condition, because

Sp(ut)≤Sp(u)+Np∣log⁡t∣Mu.S_p(u_t) \le S_p(u) + \frac{N}{p}|\log t|M_u.

For s>0s>0, define

As(u):=∫∣z∣<sτu(z)∣z∣N dz+∫∣z∣≥sτu(z)−2Mu∣z∣N dz.A_s(u):=\int_{|z|<s}\frac{\tau_u(z)}{|z|^N}\,dz+\int_{|z|\ge s}\frac{\tau_u(z)-2M_u}{|z|^N}\,dz.

Both integrals are finite, and comparison of the cutoff radii yields

As(u)=A1(u)+2Mu∣SN−1∣log⁡s.A_s(u)=A_1(u)+2M_u|S^{N-1}|\log s.

The change of variables z=twz=tw implies A1(ut)=A1/t(u)A_1(u_t)=A_{1/t}(u). By (21) and CN,p∣SN−1∣=pC_{N,p}|S^{N-1}|=p, we obtain

EL,p(ut)=EL,p(u)−plog⁡t ∥u∥Lp(RN)p.(49)\mathcal{E}_{L,p}(u_t)=\mathcal{E}_{L,p}(u)-p\log t\,\|u\|_{L^p(\mathbb{R}^N)}^p. \tag*{(49)}

The map u↦utu\mapsto u_t is a bijection from X0p(Ω)X_0^p(\Omega) onto X0p(tΩ)X_0^p(t\Omega) and preserves the LpL^p-norm. Taking infima of the Rayleigh quotients in (49), we conclude that

λ1,L,p(tΩ)=λ1,L,p(Ω)−plog⁡t.(50)\lambda_{1,L,p}(t\Omega)=\lambda_{1,L,p}(\Omega)-p\log t. \tag*{(50)}

Writing ωN:=∣B1∣\omega_N:=|B_1|, we may therefore express Theorem 1.2 as

λ1,L,p(Ω)≥λ1,L,p(B1)−pNlog⁡(∣Ω∣ωN).(51)\lambda_{1,L,p}(\Omega)\ge\lambda_{1,L,p}(B_1)-\frac{p}{N}\log\left(\frac{|\Omega|}{\omega_N}\right). \tag*{(51)}

Equality holds if and only if Ω\Omega agrees with a ball up to a null set. The coefficient p/Np/N is determined by dilation: for Ω=BR\Omega=B_R, formula (50) yields equality for every R>0R>0. Any other coefficient with the same unit-ball normalization would violate the bound for either R>1R>1 or 0<R<10<R<1. The fixed-scale estimate (39) bounds the energy deficit from below by a regularized Riesz deficit. A bound in terms of the distance to the equality family would require a further quantitative argument.

Remark 6.1. The argument in Section 4 and the polarization proof in Appendix A apply to the same admissible class and establish the same equality statement. In Section 4, monotonicity in Lemma 4.1 reduces equality for the logarithmic energy to equality in a rearrangement inequality at a fixed scale. The strict rearrangement theorem then identifies the center. The same argument proves (39).

For the polarization proof, the starting point is an exact deficit for a single reflection. Iterated polarizations lead to the Schwarz comparison, and equality for all affine half-spaces implies radial monotonicity about a common center. To pass to the logarithmic energy, one must also control the long-range interactions. The estimates of Section 2 and a cutoff argument ensure this control for functions of arbitrary support.

An alternative proof of Theorem 1.1

For completeness, we present a proof of Theorem 1.1 by polarization. We first compute the deficit for a single polarization and characterize its equality cases. We then justify passage to Schwarz rearrangement and show that equality forces the reflection orderings to have a common center.

Polarization and the sorting inequality

Let H⊂RNH \subset\mathbb{R}^{N} be an open affine half-space, and let σH\sigma_H denote reflection across ∂H\partial H. For a real-valued measurable function uu, define

uH(x):={max⁡{u(x),u(σHx)},x∈H,min⁡{u(x),u(σHx)},x∉H.(52)u^{H}(x) := \begin{cases} \max\{u(x),u(\sigma_H x)\}, & x \in H,\\ \min\{u(x),u(\sigma_H x)\}, & x \notin H. \end{cases} \tag*{(52)}

The value on ∂H\partial H is immaterial. Polarization permutes the two values on each reflection pair and therefore preserves their distribution. In particular,

∥uH∥Lp(RN)p=∥u∥Lp(RN)p,Sp(uH)=Sp(u).(53)\lVert u^{H}\rVert_{L^{p}(\mathbb{R}^{N})}^{p}=\lVert u\rVert_{L^{p}(\mathbb{R}^{N})}^{p}, \qquad S_p(u^{H})=S_p(u). \tag*{(53)}

For a,b,c,d∈Ra,b,c,d \in\mathbb{R}, set

a+:=max⁡{a,b}a_+ := \max\{a,b\}, a−:=min⁡{a,b}a_- := \min\{a,b\}, c+:=max⁡{c,d}c_+ := \max\{c,d\}, c−:=min⁡{c,d}c_- := \min\{c,d\},

and define

Qp(a,b,c,d):=∣a−c∣p+∣b−d∣p−∣a+−c+∣p−∣a−−c−∣p.(54)Q_p(a,b,c,d):=|a-c|^p+|b-d|^p-|a_+-c_+|^p-|a_--c_-|^p. \tag*{(54)}

For p>1p>1, we have

Qp(a,b,c,d)≥0,Qp(a,b,c,d)>0⟺(a−b)(c−d)<0.(55)Q_p(a,b,c,d)\ge0,\qquad Q_p(a,b,c,d)>0 \Longleftrightarrow(a-b)(c-d)<0. \tag*{(55)}

Indeed, if (a−b)(c−d)≥0(a-b)(c-d)\ge0, the two sums in (A.3) coincide. If a>ba>b and d>cd>c, then

Qp(a,b,c,d)=∣a−c∣p+∣b−d∣p−∣a−d∣p−∣b−c∣p=p(p−1)∫ba∫cd∣s−t∣p−2 dt ds>0.\begin{aligned} Q_p(a,b,c,d) &=|a-c|^p+|b-d|^p-|a-d|^p-|b-c|^p\\ &=p(p-1)\int_b^a\int_c^d |s-t|^{p-2}\,\mathrm{d}t\,\mathrm{d}s>0. \end{aligned}

The other case follows by interchanging the two entries in each pair. When 1<p<21<p<2, the possible singularity at s=ts=t is locally integrable, as in (27).

The inequality and equality for one polarization

We next apply the sorting inequality to the short-range form and to the full logarithmic energy. The resulting deficit identity also determines the equality cases for a single polarization.

Proposition A.1. Let 1<p<∞1<p<\infty, let u∈Xp(RN)u\in X^p(\mathbb{R}^{N}) be real-valued, and let HH be an open affine half-space. Then uH∈Xp(RN)u^{H}\in X^p(\mathbb{R}^{N}) and

[uH]Xpp≤[u]Xpp.(56)[u^{H}]_{X^p}^{p}\le[u]_{X^p}^{p}. \tag*{(56)}

If Sp(u)<∞S_p(u)<\infty, both logarithmic energies are finite and

EL,p(uH)≤EL,p(u).(57)\mathcal{E}_{L,p}(u^{H})\le\mathcal{E}_{L,p}(u). \tag*{(57)}

Equality holds in (A.6) if and only if

u(x)≥u(σHx)for a.e. x∈H,(58)u(x) \ge u(\sigma_H x) \quad\text{for a.e. } x \in H, \tag*{(58)}

or the reverse inequality holds for almost every x∈Hx \in H. Equivalently, equality holds if and only if uH=uu^H=u almost everywhere or uH=u∘σHu^H=u\circ\sigma_H almost everywhere.

Proof. For x,y∈Hx,y\in H, write

a=u(x),b=u(σHx),c=u(y),d=u(σHy).a=u(x),\quad b=u(\sigma_Hx),\quad c=u(y),\quad d=u(\sigma_Hy).

Reflection geometry implies ∣x−y∣<∣x−σHy∣|x-y|<|x-\sigma_Hy|. Let kk be a bounded, integrable, radial nonincreasing kernel. Grouping the interaction into the four pairs associated with x,y∈Hx,y\in H, we obtain

∬RN×RNk(x−y)(∣u(x)−u(y)∣p−∣uH(x)−uH(y)∣p) dx dy=∬H×H(k(x−y)−k(x−σHy))Qp(a,b,c,d) dx dy≥0.(59)\begin{aligned} &\iint_{\mathbb{R}^N\times\mathbb{R}^N} k(x-y)\left(|u(x)-u(y)|^p-|u^H(x)-u^H(y)|^p\right)\,\mathrm{d}x\,\mathrm{d}y\\ &=\iint_{H\times H}\left(k(x-y)-k(x-\sigma_Hy)\right)Q_p(a,b,c,d)\,\mathrm{d}x\,\mathrm{d}y\ge0. \tag*{(59)} \end{aligned}

This identity follows because sorting preserves the sum of the four differences. Apply (A.8) to

kδ(z):=1{∣z∣<1}min⁡{δ−N,∣z∣−N},0<δ<1.k_\delta(z):=\mathbf{1}_{\{|z|<1\}}\min\{\delta^{-N},|z|^{-N}\},\qquad0<\delta<1.

Letting δ↓0\delta\downarrow0, the monotone convergence theorem proves (A.5). By (A.2), Lemma 2.2 and (14), both energies are finite when Sp(u)<∞S_p(u)<\infty.

We may now group the logarithmic energies into the same reflection pairs. The mass terms cancel. The sum of the long-range subtraction terms on each group is

−(f1+f2)(∣a∣p+∣b∣p+∣c∣p+∣d∣p),-(f_1+f_2)(|a|^p+|b|^p+|c|^p+|d|^p),

where

f1:=1{∣x−y∣≥1}∣x−y∣N,f2:=1{∣x−σHy∣≥1}∣x−σHy∣N.f_1:=\frac{\mathbf{1}_{\{|x-y|\ge1\}}}{|x-y|^N},\qquad f_2:=\frac{\mathbf{1}_{\{|x-\sigma_Hy|\ge1\}}}{|x-\sigma_Hy|^N}.

This expression is unchanged by sorting. Consequently,

EL,p(u)−EL,p(uH)=CN,p2∬H×H(1∣x−y∣N−1∣x−σHy∣N)Qp(a,b,c,d) dx dy≥0.(60)\mathcal{E}_{L,p}(u)-\mathcal{E}_{L,p}(u^H)=\frac{C_{N,p}}{2}\iint_{H\times H}\left(\frac{1}{|x-y|^N}-\frac{1}{|x-\sigma_Hy|^N}\right)Q_p(a,b,c,d)\,\mathrm{d}x\,\mathrm{d}y\ge0. \tag*{(60)}

These regroupings are justified by finiteness of the short-range terms and absolute convergence of the long-range terms. The kernel difference in (A.9) is strictly positive almost everywhere on H×HH\times H.

Set

A:={x∈H:u(x)>u(σHx)},B:={x∈H:u(x)<u(σHx)}.A:=\{x\in H:u(x)>u(\sigma_Hx)\},\qquad B:=\{x\in H:u(x)<u(\sigma_Hx)\}.

If both sets have positive measure, (A.4) implies that the integrand in (A.9) is strictly positive on A×BA\times B. Equality therefore forces ∣A∣=0|A|=0 or ∣B∣=0|B|=0, which is precisely the asserted ordering. Conversely, either ordering makes uHu^H equal to uu or its reflection. Reflection invariance of the energy then proves equality.

Remark A.2. For real-valued uu under the hypotheses of Theorem 1.1, Proposition A.1 and (38) yield

EL,p((∣u∣)H)≤EL,p(∣u∣)≤EL,p(u).(61)\mathcal{E}_{L,p}((|u|)^{H}) \leq\mathcal{E}_{L,p}(|u|) \leq\mathcal{E}_{L,p}(u). \tag*{(61)}

If u≠0u \ne0, equality holds throughout if and only if uu has one sign and ∣u∣|u| satisfies one of the two orderings across ∂H\partial H in Proposition A.1.

Equality for signed polarization (A.1) does not require one sign, since every reflection-invariant function is fixed by that operation. Nor does equality for a single polarization imply radial symmetry. For example, every nonnegative smooth compactly supported function whose support is contained in HH satisfies uH=uu^{H}=u, regardless of the shape of its level sets.

Passage to Schwarz rearrangement

For nonnegative u∈Lp(RN)u \in L^{p}(\mathbb{R}^{N}), the approximation theorem of Brock and Solynin [4] and the constructive version of Hajaiej [11] provide finite iterated polarizations unu_n toward half-spaces containing the origin such that

un⟶u∗in Lp(RN).(62)u_n \longrightarrow u^{*}\quad\text{in }L^{p}(\mathbb{R}^{N}). \tag*{(62)}

More precisely, the constructive scheme uses a dense sequence (Hj)(H_j) of half-spaces containing the origin in their interior, with u0=uu_0=u and un+1=unH1⋯Hn+1u_{n+1}=u_n^{H_1\cdots H_{n+1}}. The repeated finite blocks are part of the convergence statement; an arbitrary sequence of half-spaces need not converge to u∗u^{*}.

To use (A.11), we first consider functions with bounded support, for which the long-range term is continuous under LpL^p-convergence. We then use the estimates of § 2 to remove the support restriction.

Proposition A.3. Let u∈Xp(RN)u \in X^p(\mathbb{R}^{N}) be real-valued and satisfy Sp(u)<∞S_p(u)<\infty. Then u#∈Xp(RN)u^{\#}\in X^p(\mathbb{R}^{N}), Sp(u#)=Sp(u)S_p(u^{\#})=S_p(u), and EL,p(u#)≤EL,p(u)\mathcal{E}_{L,p}(u^{\#})\leq\mathcal{E}_{L,p}(u).

Proof. Suppose first that u≥0u\geq0 and u=0u=0 outside BRB_R. All iterates in (A.11) also vanish outside BRB_R, since polarization toward a half-space containing the origin places the larger value at the point closer to the origin. By Proposition A.1,

EL,p(un)≤EL,p(u)for every n.(63)\mathcal{E}_{L,p}(u_n)\leq\mathcal{E}_{L,p}(u)\quad\text{for every }n. \tag*{(63)}

On functions supported in a common bounded set, the long-range term is continuous under LpL^p-convergence. Indeed, its numerator vanishes when either argument is zero, so the interaction is confined to a bounded product set. The kernel is bounded on ∣x−y∣≥1|x-y|\geq1, and the standard LpL^p estimates for differences of ppth powers imply continuity of the integral. The mass terms also converge. After passing to an almost everywhere convergent subsequence, Fatou’s lemma for the short-range term yields u∗∈Xp(RN)u^{*}\in X^p(\mathbb{R}^{N}) and

EL,p(u∗)≤lim inf⁡n→∞EL,p(un)≤EL,p(u).(64)\mathcal{E}_{L,p}(u^{*})\leq\liminf_{n\to\infty}\mathcal{E}_{L,p}(u_n)\leq\mathcal{E}_{L,p}(u). \tag*{(64)}

Now let u≥0u\geq0 have arbitrary support and satisfy Sp(u)<∞S_p(u)<\infty. Choose a smooth radial nonincreasing function χ\chi, equal to 11 on B1B_1 and to 00 outside B2B_2, with 0≤χ≤10\leq\chi\leq1. Set χR(x):=χ(x/R)\chi_R(x):=\chi(x/R) and uR:=χRuu_R:=\chi_Ru for R≥1R\geq1. Then uR↑uu_R \uparrow u almost everywhere and uR→uu_R \to u in Lp(RN)L^p(\mathbb{R}^N). To verify convergence in Xp(RN)X^p(\mathbb{R}^N), put wR:=(1−χR)uw_R := (1-\chi_R)u. By the Lipschitz bound for χR\chi_R,

[wR]Xpp≤2p−1∬∣x−y∣<1(1−χR(x))p∣u(x)−u(y)∣p∣x−y∣N dx dy+CR−p∥u∥Lp(RN)p.[w_R]_{X^p}^p \le2^{p-1}\iint_{|x-y|<1}\frac{(1-\chi_R(x))^p|u(x)-u(y)|^p}{|x-y|^N}\,dx\,dy + CR^{-p}\|u\|_{L^p(\mathbb{R}^N)}^p.

where CC is independent of RR. The first term tends to zero by dominated convergence, so uR→uu_R \to u in Xp(RN)X^p(\mathbb{R}^N).

The function HpH_p is increasing in each nonnegative argument. Since uR↑uu_R \uparrow u and uR∗↑u∗u_R^* \uparrow u^* almost everywhere, monotone convergence applies to both long-range interactions. Their limits are finite by Lemma 2.2, since Sp(u∗)=Sp(u)S_p(u^*)=S_p(u). It follows that EL,p(uR)→EL,p(u)\mathcal{E}_{L,p}(u_R)\to\mathcal{E}_{L,p}(u), while Fatou’s lemma for the rearranged short-range term yields u∗∈Xp(RN)u^*\in X^p(\mathbb{R}^N) and

EL,p(u∗)≤lim inf⁡R→∞EL,p(uR∗)≤lim⁡R→∞EL,p(uR)=EL,p(u).(65)\mathcal{E}_{L,p}(u^*) \le\liminf_{R\to\infty}\mathcal{E}_{L,p}(u_R^*) \le\lim_{R\to\infty}\mathcal{E}_{L,p}(u_R)=\mathcal{E}_{L,p}(u). \tag*{(65)}

Here the second inequality follows from the bounded-support case applied to uRu_R.

For real-valued uu, apply the nonnegative case to ∣u∣|u| and use (38). Equimeasurability implies Sp(u#)=Sp(u)S_p(u^\#)=S_p(u).

Equality and the common center

It remains to identify the equality cases. By Proposition A.1, equality for every affine half-space imposes a reflection ordering across each hyperplane. The next lemma shows that these orderings determine a common center.

Lemma A.4. Let 0≠v∈Lp(RN)0\ne v\in L^p(\mathbb{R}^N) be nonnegative. Suppose that, for every open affine half-space HH, either v(x)≥v(σHx)v(x)\ge v(\sigma_Hx) almost everywhere on HH, or the reverse inequality holds almost everywhere on HH. Then v=v∗(⋅−x0)v=v^*(\cdot-x_0) almost everywhere for some x0∈RNx_0\in\mathbb{R}^N.

Proof. Set w:=vp∈L1(RN)w:=v^p\in L^1(\mathbb{R}^N), Φ(x):=e−∣x∣2\Phi(x):=e^{-|x|^2} and F:=Φ∗wF:=\Phi*w. The function FF is positive and continuous and tends to zero at infinity. Hence it attains its maximum at some point x0∈RNx_0\in\mathbb{R}^N. The reflection orderings of vv also hold for ww.

Let HH contain x0x_0 in its interior. If w(y)≤w(σHy)w(y)\le w(\sigma_Hy) almost everywhere on HH, with strict inequality on a set of positive measure, then

F(x0)−F(σHx0)=∫H(Φ(x0−y)−Φ(x0−σHy))(w(y)−w(σHy)) dy<0.F(x_0)-F(\sigma_Hx_0)=\int_H\bigl(\Phi(x_0-y)-\Phi(x_0-\sigma_Hy)\bigr)\bigl(w(y)-w(\sigma_Hy)\bigr)\,dy<0.

Indeed, the first factor is strictly positive for y∈Hy\in H. This contradicts maximality at x0x_0. Therefore v(x)≥v(σHx)v(x)\ge v(\sigma_Hx) almost everywhere on HH for every half-space containing x0x_0 in its interior. Thus vH=vv^H=v for all such half-spaces. By the characterization of Schwarz rearrangement through inward polarizations [11], translated from the origin to x0x_0, vv is radially nonincreasing about x0x_0. Equimeasurability identifies its profile with that of v∗v^*.

We now combine this geometric fact with the polarization inequality to recover the equality statement of Theorem 1.1.

Proposition A.5. Under the hypotheses of Theorem 1.1, equality holds for u≢0u \not\equiv0 if and only if there exist x0∈RNx_0 \in\mathbb{R}^{N} and σ∈{−1,1}\sigma\in\{-1,1\} such that u=σu#(⋅−x0)u=\sigma u^{\#}(\cdot-x_0) almost everywhere.

Proof. Suppose first that u≥0u \ge0 and EL,p(u)=EL,p(u∗)\mathcal{E}_{L,p}(u)=\mathcal{E}_{L,p}(u^{*}). For every open affine half-space HH, Propositions A.1 and A.3 yield

EL,p(u∗)≤EL,p(uH)≤EL,p(u),\mathcal{E}_{L,p}(u^{*}) \le\mathcal{E}_{L,p}(u^{H}) \le\mathcal{E}_{L,p}(u),

since (uH)∗=u∗(u^{H})^{*}=u^{*}. Equality therefore holds for every polarization. By Proposition A.1, uu satisfies one of the two reflection orderings for each HH. Lemma A.4 implies that u=u∗(⋅−x0)u=u^{*}(\cdot-x_0) almost everywhere for some x0∈RNx_0 \in\mathbb{R}^{N}. The converse follows from translation invariance.

For real-valued uu, equality in

EL,p(u)≥EL,p(∣u∣)≥EL,p(u#)\mathcal{E}_{L,p}(u) \ge\mathcal{E}_{L,p}(|u|) \ge\mathcal{E}_{L,p}(u^{\#})

forces uu to have one sign by (38). The nonnegative case then shows that ∣u∣|u| is a translate of u#u^{\#}. Conversely, these properties imply equality by invariance of the energy under translations and multiplication by −1-1. □

Comparison of the two arguments

Equality for one polarization requires only an ordering across the corresponding hyperplane. It allows nonradial functions and, for signed polarization, sign-changing functions. In the proof above, radial monotonicity follows from equality for every affine half-space and Lemma A.4. The sign condition follows separately from (38).

The passage from iterated polarizations to the logarithmic energy also requires more than LpL^{p}-convergence. For nonnegative functions, the long-range contribution is negative, so Fatou’s lemma for the short-range term alone does not imply lower semicontinuity of the full energy. Bounded support ensures continuity of the long-range term in (A.13); the estimates of § 2 and the cutoff argument in (A.14) remove this restriction. Thus the polarization proof applies under the same logarithmic condition as Theorem 1.1, without additional fractional regularity or LqL^{q}-integrability for q<pq<p. In particular, it includes the functions vαv_{\alpha} with α>2\alpha>2 from Remark 2.4. In § 4, the same equality conclusion follows from monotonicity of the regularized interaction deficits, and the proof also establishes (39).

Finally, p>1p>1 is essential for strictness in (A.4). At p=1p=1, opposite orderings need not produce a positive deficit. The nested-ball example in (45) and (47) shows that the common-center equality classification fails at that endpoint.

Data availability

No data were used in the research described in this article.

Competing interests

The authors declare that they have no competing interests.

Hichem Hajaiej: Department of Mathematics, California State University, Los Angeles, USA Email address: [email protected]

Qinyang Li: School of Mathematical Sciences, Xiamen University, Xiamen 361005, P. R. China Email address: [email protected]

Weicheng Zhan: School of Mathematical Sciences, Xiamen University, Xiamen, Fujian 361005, P. R. China Email address: [email protected]

References

  1. [1]R. Arora, J. Giacomoni, H. Hajaiej and A. Vaishnavi, Sharp embeddings and existence results for logarithmic p-Laplacian equations with critical growth, NoDEA Nonlinear Differential Equations Appl. 33 (2026), Paper No. 84.
  2. [2]R. Arora, H. Hajaiej and K. Perera, Nonlocal Dirichlet problems involving the logarithmic p-Laplacian, preprint (2025), arXiv:2512.21959.arxiv.org/abs/2512.21959
  3. [3]A. Bahrouni, A. Gouasmia, H. Hajaiej and A. Ouannasser, A new class of nonlocal operators: the fractional logarithmic p-Laplacian and its properties, preprint (2026), arXiv:2605.12050v2.arxiv.org/abs/2605.12050
  4. [4]F. Brock and A. Yu. Solynin, An approach to symmetrization via polarization, Trans. Amer. Math. Soc. 352 (2000), no. 4, 1759–1796.DOI
  5. [5]A. Burchard, Cases of equality in the Riesz rearrangement inequality, Ann. of Math. (2) 143 (1996), no. 3, 499–527.DOI
  6. [6]A. Burchard and H. Hajaiej, Rearrangement inequalities for functionals with monotone integrands, J. Funct. Anal. 233 (2006), no. 2, 561–582.DOI
  7. [7]H. Chen and T. Weth, The Dirichlet problem for the logarithmic p-Laplacian, Comm. Partial Differential Equations 44 (2019), no. 11, 1100–1139.
  8. [8]L. De Luca, M. Novaga and M. Ponsiglione, The 0-fractional perimeter between fractional perimeters and Riesz potentials, Ann. Sc. Norm. Super. Pisa Cl. Sci. (5) 22 (2021), no. 4, 1559–1596.arxiv.org/abs/1906.06303
  9. [9]B. Dyda, S. Jarohs and F. Sk, The Dirichlet problem for the logarithmic p-Laplacian, Trans. Amer. Math. Soc. 379 (2026), no. 4, 2717–2779.
  10. [10]R. L. Frank and R. Seiringer, Non-linear ground state representations and sharp Hardy inequalities, J. Funct. Anal. 255 (2008), no. 12, 3407–3430.arxiv.org/abs/0803.0503
  11. [11]H. Hajaiej, Explicit constructive approximation to symmetrization via iterated polarizations, J. Convex Anal. 17 (2010), no. 2, 405–411.DOI

Paper details

vlnc:W_kag5uu49
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents