betaFeedback

WorksMathematics

Gagliardo–Nirenberg extremals on manifolds with small holes

Abstract

We study the sharp homogeneous Dirichlet Gagliardo–Nirenberg quotient on a closed Riemannian manifold with holes of small capacity and connected complements. For each fixed exponent 2<q<2∗2<q<2^*, the nonnegative L2L^2-normalized extremal exists and is unique when the first Dirichlet eigenvalue λDλ_D is sufficiently small. If the ambient volume is VV and β=n(q−2)/2β=n(q-2)/2, the sharp constant satisfies CD=V1−q/2λD−β/2(1+O(λD))C_D=V^{1-q/2}λ_D^{-β/2}(1+O(λ_D)). For the normalized first eigenfunction φDφ_D, we obtain ∥uD−φD∥H1=O(λD3/2)\|u_D-φ_D\|_{H^1}=O(λ_D^{3/2}) and CD/WMD(φD)=1+O(λD2)C_D/W_{M_D}(φ_D)=1+O(λ_D^2), together with a quantitative stability estimate in terms of the relative deficit. The constants depend only on the ambient manifold and the exponent. On general compact manifolds with Dirichlet boundary conditions, we prove compactness above the Euclidean sharp constant and describe concentration at equality. Positive scalar curvature at one interior point implies attainment. For the L2L^2-critical nonlinear Schrödinger energy, the mass threshold for negative energy tends to zero as the capacity of the holes tends to zero, whereas the threshold for unboundedness below is unchanged.

Introduction

The sharp homogeneous Dirichlet Gagliardo–Nirenberg constant on a bounded Euclidean domain equals its whole-space value and is not attained. On a closed manifold with a smooth hole of sufficiently small capacity and connected complement, we prove that the supremum is attained by a unique nonnegative L2L^2-normalized extremal. We determine the asymptotics of the sharp constant, estimate the difference between the extremal and the first Dirichlet eigenfunction, and prove a stability inequality (Theorem A).

Let (Mn,g)(M^n,g) be a smooth compact connected Riemannian manifold with smooth nonempty boundary. Throughout the paper,

n≥2,2<q<2∗,2∗={2n/(n−2),n≥3,∞,n=2.(1)n \ge2,\qquad2 < q < 2^*,\qquad2^* = \begin{cases} 2n/(n-2), & n \ge3,\\ \infty, & n=2. \end{cases} \tag*{(1)}

and

β=n(q−2)2,α=q−β>0.(2)\beta= \frac{n(q-2)}{2},\qquad\alpha=q-\beta>0. \tag*{(2)}

For u∈H01(M)∖{0}u\in H^1_0(M)\setminus\{0\}, define

WM(u)=∥u∥Lq(M)q∥u∥L2(M)α∥∇u∥L2(M)β,CM=sup⁡u∈H01(M)∖{0}WM(u).(3)W_M(u)=\frac{\|u\|_{L^q(M)}^q}{\|u\|_{L^2(M)}^\alpha\|\nabla u\|_{L^2(M)}^\beta},\qquad C_M=\sup_{u\in H^1_0(M)\setminus\{0\}}W_M(u). \tag*{(3)}

We write C∗=CRnC_* = C_{\mathbb{R}^n} for the Euclidean sharp constant. Functions are real-valued unless stated otherwise. Norms and integrals on a manifold use its Riemannian volume unless another measure is displayed.

Extremals for small holes

Fix a smooth closed connected manifold (Nn,g)(N^n,g) of volume VV. A hole D⊂ND\subset N is a nonempty smooth open set with proper closure, and MD=N∖D‾M_D=N\setminus\overline{D} carries the Dirichlet condition on ∂D\partial D. Put CD=CMDC_D=C_{M_D} and λD=λ1(MD)\lambda_D=\lambda_1(M_D). When MDM_D is connected,

2020 Mathematics Subject Classification. 35A23, 35J61, 35Q55, 53C21, 58I50.

Key words and phrases. Gagliardo–Nirenberg inequality, small holes, uniqueness, quantitative stability, nonlinear Schrödinger equation. let φD>0\varphi_D > 0 be its first Dirichlet eigenfunction, normalized by ∥φD∥2=1\|\varphi_D\|_2 = 1. A tilde denotes zero extension to NN. We use the mean-zero capacity

cap⁡N(D)=inf⁡{∫N∣∇v∣2:v∈H1(N), v=1 a.e. on D, ∫Nv=0}.(4)\operatorname{cap}_N(D) = \inf\left\{\int_N |\nabla v|^2 : v \in H^1(N),\ v = 1\ \text{a.e. on }D,\ \int_N v = 0\right\}. \tag*{(4)}

The mean-zero condition excludes nonzero constant functions.

For f∈H01(MD)f \in H^1_0(M_D) with ∥f∥2=1\|f\|_2 = 1, define the relative deficit by

δD(f)=1−WMD(f)CD.(5)\delta_D(f) = 1 - \frac{W_{M_D}(f)}{C_D}. \tag*{(5)}

Theorem A (Uniqueness, stability, and spectral asymptotics). There are λ0>0\lambda_0 > 0, δ0∈(0,1/2)\delta_0 \in(0,1/2), and K<∞K < \infty, depending only on NN and qq, such that the following holds for every smooth hole DD with connected complement and 0<λD<λ00 < \lambda_D < \lambda_0.

One has CD>C∗,C_D > C_{*,}, and there is a unique nonnegative maximizer uDu_D with ∥uD∥2=1\|u_D\|_2 = 1. As λD→0\lambda_D \to0,

CD=V1−q/2λD−β/2(1+O(λD)),(6)C_D = V^{1-q/2}\lambda_D^{-\beta/2}\left(1 + O(\lambda_D)\right), \tag*{(6)}

and

∥u~D−φ~D∥H1(N)=O(λD3/2),(7)\|\widetilde{u}_D-\widetilde{\varphi}_D\|_{H^1(N)} = O(\lambda_D^{3/2}), \tag*{(7)}
CDWMD(φD)=1+O(λD2),(8)\frac{C_D}{W_{M_D}(\varphi_D)} = 1 + O(\lambda_D^2), \tag*{(8)}
∥∇uD∥22=λD+O(λD3),(9)\|\nabla u_D\|_2^2 = \lambda_D + O(\lambda_D^3), \tag*{(9)}
∥u~D−V−1/2∥H1(N)=O(λD1/2).(10)\|\widetilde{u}_D - V^{-1/2}\|_{H^1(N)} = O(\lambda_D^{1/2}). \tag*{(10)}

Every nonnegative f∈H01(MD)f \in H^1_0(M_D) with ∥f∥2=1\|f\|_2 = 1 and δD(f)≤δ0\delta_D(f) \le\delta_0 satisfies

∥f~−u~D∥H1(N)≤KλDδD(f).(11)\|\widetilde{f}-\widetilde{u}_D\|_{H^1(N)} \le K\lambda_D\delta_D(f). \tag*{(11)}

For a family of such holes, λD→0\lambda_D \to0 if and only if cap⁡N(D)→0\operatorname{cap}_N(D) \to0, and

CD=V1−α/2cap⁡N(D)−β/2(1+O(cap⁡N(D))).(12)C_D = V^{1-\alpha/2}\operatorname{cap}_N(D)^{-\beta/2}\left(1 + O(\operatorname{cap}_N(D))\right). \tag*{(12)}

All error bounds are uniform over the holes.

The constants in Theorem A are independent of curvature and collar bounds for ∂D\partial D.

Compactness at the Euclidean threshold

Let Bx,ℓ,RB_{x,\ell,R} be the cutoff rescaling defined in (34) from a Euclidean optimizer satisfying ∥Q0∥2=∥∇Q0∥2=1\|Q_0\|_2 = \|\nabla Q_0\|_2 = 1.

Theorem B (Threshold compactness). Under (1), one has C∗≤CM<∞C_* \le C_M < \infty. If CM>C∗C_M > C_*, every maximizing sequence normalized by ∥uk∥2=1\|u_k\|_2 = 1 is precompact in H01(M)H^1_0(M), and CMC_M is attained by a function positive in M∘M^\circ.

If CM=C∗C_M = C_*, every such maximizing sequence has a subsequence that either converges strongly in H01(M)H^1_0(M) to an extremal, or satisfies

ℓk=∥∇uk∥2−1⟶0.(13)\ell_k = \|\nabla u_k\|_2^{-1} \longrightarrow0. \tag*{(13)}

In the latter case there are xk∈M∘x_k \in M^\circ, Rk→∞R_k \to\infty, and signs sk∈{−1,1}s_k \in\{-1,1\} such that ℓkRk→0\ell_kR_k \to0, the functions Bxk,ℓk,RkB_{x_k,\ell_k,R_k} are supported in M∘M^\circ, and

∥uk−skBxk,ℓk,Rk∥2+ℓk∥∇(uk−skBxk,ℓk,Rk)∥2⟶0.(14)\|u_k-s_kB_{x_k,\ell_k,R_k}\|_2+\ell_k\|\nabla(u_k-s_kB_{x_k,\ell_k,R_k})\|_2 \longrightarrow0. \tag*{(14)}

Their centers satisfy

dg(xk,∂M)ℓk⟶∞.(15)\frac{d_g(x_k,\partial M)}{\ell_k} \longrightarrow\infty. \tag*{(15)}

For every smooth bounded domain Ω⊂Rn\Omega\subset\mathbb{R}^n, one has CΩ=C∗C_\Omega= C_*, and the supremum is not attained.

In the noncompact alternative, the centers may approach ∂M\partial M, but their distance from the boundary divided by ℓk\ell_k tends to infinity. The limiting profile is therefore defined on Rn\mathbb{R}^n.

Scalar curvature and attainment

Let Q∈H1(Rn)Q \in H^1(\mathbb{R}^n) be the positive radial solution, centered at the origin, of

−ΔQ+Q=Qq−1in Rn.(16)-\Delta Q + Q = Q^{q-1}\quad\text{in }\mathbb{R}^n. \tag*{(16)}

Exponential decay follows from the radial differential equation as in [24 p. 155]; interior elliptic estimates give the same decay for the gradient. Thus Q(y)+∣∇Q(y)∣≤Ce−c∣y∣Q(y) + |\nabla Q(y)| \leq C e^{-c|y|} for some C,c>0C,c > 0. The test functions Ux,εU_{x,\varepsilon} are defined in (43).

Theorem C (Positive scalar curvature). If Scal⁡g(x)>0\operatorname{Scal}_g(x) > 0 at one interior point, then CM>C∗C_M > C_* *and the supremum is attained. More precisely, uniformly for x∈K⋐M∘x \in K \Subset M^\circ,

WM(Ux,ε)=C∗(1+An,qScal⁡g(x)ε2+OK(ε3)),(17)W_M(U_{x,\varepsilon}) = C_* \left(1 + A_{n,q}\operatorname{Scal}_g(x)\varepsilon^2 + O_K(\varepsilon^3)\right), \tag*{(17)}

where

An,q=α12+q−212n∫Rn∣y∣2Qq dy∫RnQq dy>0.(18)A_{n,q} = \frac{\alpha}{12} + \frac{q-2}{12n}\frac{\int_{\mathbb{R}^n}|y|^2Q^q\,dy}{\int_{\mathbb{R}^n}Q^q\,dy} > 0. \tag*{(18)}

Consequently, CM=C∗C_M = C_* implies Scal⁡g≤0\operatorname{Scal}_g \leq0 in M∘M^\circ.

Positive scalar curvature is sufficient but not necessary for attainment. Theorem A applies, for example, to flat tori with small geodesic balls removed.

Constrained nonlinear Schrödinger energy

At q=2+4/nq = 2 + 4/n, define

EM(u)=12∥∇u∥22−1q∥u∥qq,IM(m)=inf⁡u∈H01(M)∥u∥22=mEM(u),(19)\mathcal{E}_M(u) = \frac{1}{2}\|\nabla u\|_2^2 - \frac{1}{q}\|u\|_q^q,\qquad I_M(m) = \inf_{\substack{u\in H_0^1(M)\\ \|u\|_2^2=m}}\mathcal{E}_M(u), \tag*{(19)}

and

m0(M)=(q2CM)n/2,m∗=(q2C∗)n/2.(20)m_0(M) = \left(\frac{q}{2C_M}\right)^{n/2},\qquad m_* = \left(\frac{q}{2C_*}\right)^{n/2}. \tag*{(20)}

The thresholds for negative energy and for unboundedness below need not coincide; see [2] for metric graphs. The following theorem gives these thresholds on compact manifolds with Dirichlet boundary conditions and their asymptotics as the capacity of the holes tends to zero.

Theorem D (Constrained energy minimization). There is a function of mass mm with negative energy if and only if m>m0(M)m > m_0(M). For 0<m<m∗0 < m < m_*, the fixed-mass energy is coercive and IM(m)I_M(m) is attained. For m>m∗m > m_*, one has IM(m)=−∞I_M(m) = -\infty.

If CM>C∗C_M > C_*, every mass in (m0(M),m∗)(m_0(M),m_*) admits a negative-energy constrained minimizer. For holes in the setting of Theorem A, as λD→0\lambda_D \to0,

m0(MD)=(q2)n/21VλDn/2(1+O(λD))⟶0,(21)m_0(M_D) = \left(\frac{q}{2}\right)^{n/2} \frac{1}{V\lambda_D^{n/2}}(1+O(\lambda_D)) \longrightarrow0, \tag*{(21)}

whereas m∗m_* depends only on the dimension. In particular, m0(MD)/m∗→0m_0(M_D)/m_* \to0, and every fixed mass m∈(0,m∗)m \in(0,m_*) eventually admits a negative-energy constrained minimizer.

Related work and motivation

The interpolation inequalities originate in the work of Gagliardo [17] and Nirenberg [23]. Weinstein [Wei83] characterized the Euclidean sharp constant by a variational problem, and Kwong [18] proved uniqueness of the positive radial solution of (16). Non-attainment on bounded Euclidean domains is proved in [14 §4.3], which also observed that removing a small ball from a closed manifold can make the sharp constant exceed the Euclidean value and asked whether the supremum is attained on compact Riemannian manifolds with boundary. Banica [3 Remark 6.1, p. 1668] asked whether the sharp constant on hyperbolic space equals the Euclidean constant. Mukherjee [22 Theorem 1.3] proved equality and non-attainment on hyperbolic space; Remark 2.1 of that paper records divergence of the supremum for punctured spheres in dimensions n≥3n \geq3. In this paper, we address the compact Riemannian manifold question. Compactness of the underlying manifold does not prevent maximizing sequences from concentrating at arbitrarily small scales. To exclude such concentration, we prove that the sharp constant strictly exceeds its Euclidean value for complements of holes of sufficiently small capacity. When the complement is connected, we also establish uniqueness and quantitative stability of the nonnegative normalized extremal.

Del Pino and Dolbeault [16] determined sharp constants and optimizers for a family of Gagliardo–Nirenberg inequalities and related them to nonlinear diffusion. Cordero–Erausquin, Nazaret, and Villani [CENV04] gave mass-transportation proofs of sharp Sobolev and Gagliardo–Nirenberg inequalities. Bianchi and Egnell [5] proved quantitative stability for the critical Sobolev inequality. Carlen and Figalli [11] prove stability for a planar Gagliardo–Nirenberg–Sobolev inequality involving the L4L^{4} and L6L^{6} norms and use fast diffusion to obtain stability for the logarithmic Hardy–Littlewood–Sobolev inequality and quantitative convergence for the critical-mass Keller–Segel equation. Bonforte, Dolbeault, Nazaret, and Simonov [4] use fast diffusion and entropy estimates to prove constructive stability for the Lp+1L^{p+1}–L2pL^{2p} family and improved convergence rates toward self-similar solutions under suitable tail assumptions.

Riemannian inequalities with lower-order terms are treated in [7, 8, 9]; the last paper relates the second sharp constant to scalar curvature through weighted integrals of Euclidean optimizers. Cheng [13] obtains curvature restrictions for interpolation families with exponent pairs 2a2a and a+1a+1. Theorem C concerns the homogeneous Dirichlet quotient (3), which has an L2L^{2} factor, throughout the range (1). The relation between eigenvalue perturbations and capacity is studied in [15, 1].

The concentration–compactness method is due to Lions [19, 20]. Coordinate estimates for manifolds with boundary are recorded in [21].

Localization and Euclidean comparison

Euclidean inequalities

The Euclidean sharp inequality is

∫Rn∣f∣q dy≤C∗(∫Rn∣f∣2 dy)α/2(∫Rn∣∇f∣2 dy)β/2.(22)\int_{\mathbb{R}^{n}} |f|^{q}\,dy \le C_{*}\left(\int_{\mathbb{R}^{n}} |f|^{2}\,dy\right)^{\alpha/2}\left(\int_{\mathbb{R}^{n}} |\nabla f|^{2}\,dy\right)^{\beta/2}. \tag*{(22)}

The nonzero real equality cases are aQ(b(⋅−y0))aQ(b(\cdot-y_{0})), with a≠0a\ne0 and b>0b>0. To recall the classification, existence follows from [Wei83]. If ff is an optimizer, then so is v=∣f∣v=|f|, and the Euler equation is

−Δv+Av=Bvq−1,A=α∥∇v∥22β∥v∥22>0,B=q∥∇v∥22β∥v∥qq>0.-\Delta v+Av=Bv^{q-1},\qquad A=\frac{\alpha\|\nabla v\|_{2}^{2}}{\beta\|v\|_{2}^{2}}>0,\qquad B=\frac{q\|\nabla v\|_{2}^{2}}{\beta\|v\|_{q}^{q}}>0.

Elliptic regularity and the maximum principle give v>0v>0; local elliptic estimates and v∈H1(Rn)v\in H^{1}(\mathbb{R}^{n}) give v(y)→0v(y)\to0 as ∣y∣→∞|y|\to\infty. The symmetry theorem recalled in [18 pp. 243–244] makes vv radial about some point. After rescaling to (16), Kwong’s uniqueness theorem identifies it with QQ. Since v>0v>0, the function f/v∈Hloc1(Rn)f/v\in H^{1}_{\mathrm{loc}}(\mathbb{R}^{n}) takes only the values ±1\pm1 and is constant. Conversely, multiplication, translation and dilation preserve equality. In particular, every real optimizer has constant sign and is nonzero everywhere.

The dilation f(y)↦ℓ−n/2f(y/ℓ)f(y)\mapsto\ell^{-n/2}f(y/\ell) preserves the L2L^{2} norm, multiplies the squared gradient norm by ℓ−2\ell^{-2} and the nonlinear integral by ℓ−β\ell^{-\beta}, and leaves the quotient unchanged.

Let Q0Q_{0} be the centered positive Euclidean optimizer normalized by

∥Q0∥2=∥∇Q0∥2=1.(23)\|Q_{0}\|_{2}=\|\nabla Q_{0}\|_{2}=1. \tag*{(23)}

It is obtained from QQ by multiplication by a constant and a dilation. Fix a smooth radial cutoff χ\chi, equal to one on B1B_{1}, zero outside B2B_{2}, and satisfying 0≤χ≤10\le\chi\le1.

Lemma 2.1 (Dirichlet half-space and domains). The Dirichlet sharp constant on R+n\mathbb{R}^{n}_{+} equals C∗C_{*} and is not attained. The same conclusions hold on every smooth bounded Euclidean domain.

Proof. Zero extension gives the upper bound C∗C_{*}. A compactly supported Euclidean near-optimizer can be translated and dilated into any interior ball, giving the reverse inequality. Attainment would make its zero extension a whole-space optimizer vanishing on an open set, contrary to the equality classification.

Local coordinates

We use normal coordinates in the interior and Fermi coordinates at the boundary. The boundary normal points inward, and II(X,Y)=−g(∇Xν,Y)\mathrm{II}(X,Y)=-g(\nabla_X\nu,Y), as in [21 §1.1]. Put Kg=tr⁡IIK_g=\operatorname{tr}\mathrm{II}. At a boundary center,

∣g(y)∣=1−Kgyn+O(∣y∣2),gab(y)=δab+2IIabyn+O(∣y∣2),(24)\sqrt{|g(y)|}=1-K_g y_n+O(|y|^2),\qquad g^{ab}(y)=\delta^{ab}+2\mathrm{II}^{ab}y_n+O(|y|^2), \tag*{(24)}

with gan=0g^{an}=0 and gnn=1g^{nn}=1; see [21 Lemma 3.7]. The coordinate bounds are uniform on fixed smaller neighborhoods by [21 Lemmas 3.2 and 3.6]. The same geometric construction applies in dimension two.

Lemma 2.2 (Local comparison). Let f∈H1f\in H^1 be supported in B2RB_{2R} in an interior chart, or in B2R+B^+_{2R} with zero trace on the flat boundary in a Fermi chart. Write

Px,εf(z)=ε−n/2f(Φx−1(z)/ε),P_{x,\varepsilon}f(z)=\varepsilon^{-n/2}f(\Phi_x^{-1}(z)/\varepsilon),

extended by zero. For fixed RR and sufficiently small ε\varepsilon,

∥Px,εf∥22=(1+OR(εγ))∥f∥22,(25)\|P_{x,\varepsilon}f\|_2^2=(1+O_R(\varepsilon^\gamma))\|f\|_2^2, \tag*{(25)}
ε2∥∇Px,εf∥22=(1+OR(εγ))∥∇f∥22,(26)\varepsilon^2\|\nabla P_{x,\varepsilon}f\|_2^2=(1+O_R(\varepsilon^\gamma))\|\nabla f\|_2^2, \tag*{(26)}
εβ∥Px,εf∥qq=(1+OR(εγ))∥f∥qq,(27)\varepsilon^\beta\|P_{x,\varepsilon}f\|_q^q=(1+O_R(\varepsilon^\gamma))\|f\|_q^q, \tag*{(27)}

where γ=2\gamma=2 for an interior normal chart and γ=1\gamma=1 for a Fermi chart. Consequently,

WM(Px,εf)=Wflat(f)(1+OR(εγ)).(28)W_M(P_{x,\varepsilon}f)=W_{\mathrm{flat}}(f)(1+O_R(\varepsilon^\gamma)). \tag*{(28)}

The estimates are uniform when the charts and their supports lie in the indicated coordinate neighborhoods.

Proof. Since Px,ε=ε−1Tx,εP_{x,\varepsilon}=\varepsilon^{-1}T_{x,\varepsilon} in the notation of [21], its Dirichlet-energy identity (3.11) gives (26). For the norms, use

∫M∣Px,εf∣r dVg=εn−nr/2∫∣f(y)∣r∣g(εy)∣ dy.\int_M |P_{x,\varepsilon}f|^r\,dV_g =\varepsilon^{n-nr/2}\int|f(y)|^r\sqrt{|g(\varepsilon y)|}\,dy.

In normal charts both metric and density errors are OR(ε2)O_R(\varepsilon^2); in Fermi charts they are OR(ε)O_R(\varepsilon). The powers of ε\varepsilon cancel in the quotient. □

Proposition 2.3 (Almost-sharp compact inequality). For every η>0\eta>0 there exists Bη<∞B_\eta<\infty such that

∥u∥qq≤(C∗+η)∥u∥2α∥∇u∥2β+Bη∥u∥2q(u∈H01(M)).(29)\|u\|_q^q\leq(C_*+\eta)\|u\|_2^\alpha\|\nabla u\|_2^\beta+B_\eta\|u\|_2^q\qquad(u\in H_0^1(M)). \tag*{(29)}

On a fixed smooth closed manifold NN, the same inequality holds for all u∈H1(N)u\in H^1(N).

Proof. Given δ>0\delta>0, choose a finite normal/Fermi cover sufficiently fine that the local sharp inequality on each chart has constant at most C∗+δC_*+\delta, by Lemmas 2.1 and 2.2. Take real smooth functions θi\theta_i subordinate to the cover with ∑iθi2=1\sum_i\theta_i^2=1. Put Mi=∥θiu∥22M_i=\|\theta_i u\|_2^2 and Ei=∥∇(θiu)∥22E_i=\|\nabla(\theta_i u)\|_2^2. Minkowski’s inequality in Lq/2L^{q/2} and then Hölder’s inequality for the finite sum give

∥u∥q2=∥∑i∣θiu∣2∥q/2≤∑i∥θiu∥q2\|u\|_q^2=\left\|\sum_i|\theta_i u|^2\right\|_{q/2}\leq\sum_i\|\theta_i u\|_q^2
≤(C∗+δ)2/q∑iMiα/qEiβ/q≤(C∗+δ)2/q(∑iMi)α/q(∑iEi)β/q.\leq(C_*+\delta)^{2/q}\sum_i M_i^{\alpha/q}E_i^{\beta/q} \leq(C_*+\delta)^{2/q}\left(\sum_iM_i\right)^{\alpha/q}\left(\sum_iE_i\right)^{\beta/q}.

Here α/q+β/q=1\alpha/q+\beta/q=1. The identity ∑iθi2=1\sum_i\theta_i^2=1 gives

∑iMi=∥u∥22,∑iEi=∥∇u∥22+∫M(∑i∣∇θi∣2)u2≤∥∇u∥22+Kδ∥u∥22.\sum_iM_i=\|u\|_2^2,\qquad \sum_iE_i=\|\nabla u\|_2^2+\int_M\left(\sum_i|\nabla\theta_i|^2\right)u^2 \leq\|\nabla u\|_2^2+K_\delta\|u\|_2^2.

Raising to the power q/2q/2 yields

∥u∥qq≤(C∗+δ)∥u∥2α(∥∇u∥22+Kδ∥u∥22)β/2.(30)\|u\|_q^q\leq(C_*+\delta)\|u\|_2^\alpha\left(\|\nabla u\|_2^2+K_\delta\|u\|_2^2\right)^{\beta/2}. \tag*{(30)}

For b>0b>0 and ρ>0\rho>0, (X+Y)b≤(1+ρ)Xb+Cb,ρYb(X+Y)^b\leq(1+\rho)X^b+C_{b,\rho}Y^b for X,Y≥0X,Y\geq0. Choose δ\delta and ρ\rho with (C∗+δ)(1+ρ)≤C∗+η(C_*+\delta)(1+\rho)\leq C_*+\eta and apply this inequality with b=β/2b=\beta/2. This proves (29). On a closed manifold the proof uses only interior charts. □

Corollary 2.4 (Spectral bounds). Let VM=Vol⁡g(M)V_M=\operatorname{Vol}_g(M) and λ1(M)\lambda_1(M) be the first Dirichlet eigenvalue. Then

VM1−q/2λ1(M)−β/2≤CM≤C∗+η+Bηλ1(M)−β/2.(31)V_M^{1-q/2}\lambda_1(M)^{-\beta/2}\leq C_M\leq C_*+\eta+B_\eta\lambda_1(M)^{-\beta/2}. \tag*{(31)}

In particular CM<∞C_M<\infty.

Proof. For the L2L^2-normalized first eigenfunction φ1\varphi_1, one has ∥∇φ1∥22=λ1(M)\lVert\nabla\varphi_1\rVert_2^2=\lambda_1(M). Hölder’s inequality gives ∥φ1∥qq≥VM1−q/2\lVert\varphi_1\rVert_q^q\geq V_M^{1-q/2}, proving the lower bound. Divide (29) by ∥u∥2α∥∇u∥2β\lVert u\rVert_2^\alpha\lVert\nabla u\rVert_2^\beta and use the Dirichlet Poincaré inequality for the upper bound. □

Remark 2.5 (Dirichlet Cheeger constant). Let

hD(M)=inf⁡∅≠E⋐M∘E smoothPer⁡g(E)Vol⁡g(E).h_D(M)=\inf_{\varnothing\ne E\Subset M^\circ\atop E\ \mathrm{smooth}}\frac{\operatorname{Per}_g(E)}{\operatorname{Vol}_g(E)}.

Cheeger’s inequality [12] gives λ1(M)≥hD(M)2/4\lambda_1(M)\geq h_D(M)^2/4. Indeed, the coarea formula applied to u2u^2, for u∈Cc∞(M∘)u\in C_c^\infty(M^\circ), gives

hD(M)∥u∥22≤∫M∣∇(u2)∣≤2∥u∥2∥∇u∥2.h_D(M)\lVert u\rVert_2^2\leq\int_M|\nabla(u^2)|\leq2\lVert u\rVert_2\lVert\nabla u\rVert_2.

Consequently, (31) implies

CM≤C∗+η+2βBηhD(M)−β.(32)C_M\leq C_*+\eta+2^\beta B_\eta h_D(M)^{-\beta}. \tag*{(32)}

If a family has fixed positive volume, uniform BηB_\eta for one fixed η>0\eta>0, and λ1(Mj)≲hD(Mj)θ→0\lambda_1(M_j)\lesssim h_D(M_j)^\theta\to0 for some θ>0\theta>0, then (31) gives CMj≲hD(Mj)−θβ/2C_{M_j}\lesssim h_D(M_j)^{-\theta\beta/2}.

Compactness of maximizing sequences

Compactness above the Euclidean sharp constant

Proof of the strict part of Theorem B. Insert compactly supported Euclidean near-optimizers to obtain CM≥C∗C_M\geq C_*. Let ∥uk∥2=1\lVert u_k\rVert_2=1 and WM(uk)→CMW_M(u_k)\to C_M. If the gradients are unbounded, a subsequence has ∥∇uk∥2→∞\lVert\nabla u_k\rVert_2\to\infty, and (29) gives lim sup⁡kWM(uk)≤C∗+η\limsup_k W_M(u_k)\leq C_*+\eta for every η>0\eta>0. Thus CM>C∗C_M>C_* forces boundedness in H01(M)H_0^1(M).

After extraction, uk⇀uu_k\rightharpoonup u in H01(M)H_0^1(M) and uk→uu_k\to u in L2(M)∩Lq(M)L^2(M)\cap L^q(M), with ∥u∥2=1\lVert u\rVert_2=1. Lower semicontinuity gives WM(u)≥CMW_M(u)\geq C_M, hence equality, and

∥∇uk∥2β=∥uk∥qαWM(uk)⟶∥u∥qαCM=∥∇u∥2β.\lVert\nabla u_k\rVert_2^\beta=\frac{\lVert u_k\rVert_q^\alpha}{W_M(u_k)}\longrightarrow\frac{\lVert u\rVert_q^\alpha}{C_M}=\lVert\nabla u\rVert_2^\beta.

Convergence is therefore strong in H01(M)H_0^1(M). The argument applies to every subsequence. Taking the absolute value gives a nonnegative extremal. □

Proposition 3.1 (Equation for an extremal). Every nonnegative extremal uu satisfies

−Δgu+α∥∇u∥22β∥u∥22u=q∥∇u∥2ββ∥u∥qquq−1in M,u∣∂M=0.(33)-\Delta_g u+\frac{\alpha\lVert\nabla u\rVert_2^2}{\beta\lVert u\rVert_2^2}u=\frac{q\lVert\nabla u\rVert_2^\beta}{\beta\lVert u\rVert_q^q}u^{q-1}\quad\text{in }M,\qquad u|_{\partial M}=0. \tag*{(33)}

It is positive in M∘M^\circ and belongs to C2(M‾)C^2(\overline{M}).

Proof. Differentiate log⁡WM(u)\log W_M(u) against ϕ∈H01(M)\phi\in H_0^1(M) and multiply by ∥∇u∥22/β\lVert\nabla u\rVert_2^2/\beta. This gives (33) weakly. Subcritical elliptic regularity gives boundedness and then C2C^2 regularity up to the smooth boundary. Both coefficients in (33) are positive, and the strong maximum principle gives interior positivity. □

Rescaled maximizing sequences

In an interior normal chart Ψx\Psi_x, put

Bx,ℓ,R(z)=ℓ−n/2χ(Ψx−1(z)ℓR)Q0(Ψx−1(z)ℓ),(34)B_{x,\ell,R}(z)=\ell^{-n/2}\chi\left(\frac{\Psi_x^{-1}(z)}{\ell R}\right)Q_0\left(\frac{\Psi_x^{-1}(z)}{\ell}\right), \tag*{(34)}

extended by zero. This notation is used only when the support lies in M∘M^\circ.

We next consider a sequence with

∥uk∥2=1,ℓk=∥∇uk∥2−1→0,WM(uk)→C∗.(35)\|u_k\|_2=1,\qquad\ell_k=\|\nabla u_k\|_2^{-1}\to0,\qquad W_M(u_k)\to C_*. \tag*{(35)}

Set gk=ℓk−2gg_k=\ell_k^{-2}g and vk=ℓkn/2ukv_k=\ell_k^{n/2}u_k. Then

∫Mvk2 dVgk=1,∫M∣∇vk∣gk2 dVgk=1,∫M∣vk∣q dVgk→C∗.(36)\int_M v_k^2\,dV_{g_k}=1,\qquad\int_M|\nabla v_k|_{g_k}^2\,dV_{g_k}=1,\qquad\int_M|v_k|^q\,dV_{g_k}\to C_*. \tag*{(36)}

Lemma 3.2 (A nonzero rescaled profile). There are points zk∈Mz_k\in M and a subsequence such that one of the following holds.

If dg(zk,∂M)/ℓk→∞d_g(z_k,\partial M)/\ell_k\to\infty, the rescalings in normal coordinates at zkz_k converge weakly in Hloc1(Rn)H^1_{\mathrm{loc}}(\mathbb{R}^n) and strongly in Lloc2∩LlocqL^2_{\mathrm{loc}}\cap L^q_{\mathrm{loc}} to a nonzero w∈H1(Rn)w\in H^1(\mathbb{R}^n).

Otherwise, after choosing nearest boundary points aka_k and Fermi coordinates there, the rescalings converge in the same local senses on R+n\mathbb{R}^n_+ to a nonzero w∈H01(R+n)w\in H^1_0(\mathbb{R}^n_+). In either case,

0<m:=∥w∥22≤1,0<e:=∥∇w∥22≤1.(37)0<m:=\|w\|_2^2\le1,\qquad0<e:=\|\nabla w\|_2^2\le1. \tag*{(37)}

Proof. In the metrics gkg_k, the normal/Fermi atlas at scale ℓk\ell_k gives coordinate neighborhoods of bounded radius and overlap, with uniformly bounded cutoff gradients. Apply the nonvanishing argument of [19, 20] on this cover.

Set r0=2+4/nr_0=2+4/n. The Euclidean GN inequality applied to a cutoff of vkv_k in each neighborhood, followed by summation, gives

∥vk∥Lr0(gk)≤C(sup⁡z∈M∫Bgk(z,ρ)vk2 dVgk)2/n∫M(∣∇vk∣gk2+vk2) dVgk,(38)\|v_k\|_{L^{r_0}(g_k)}\le C\left(\sup_{z\in M}\int_{B_{g_k}(z,\rho)}v_k^2\,dV_{g_k}\right)^{2/n}\int_M\left(|\nabla v_k|_{g_k}^2+v_k^2\right)\,dV_{g_k}, \tag*{(38)}

for fixed ρ>0\rho>0. The exponent r0r_0 makes the gradient power two; cutoff derivatives give the L2L^2 term. Boundary cutoffs are extended by zero.

For every fixed 2<s<2∗2<s<2^*, put βs=n(s−2)/2\beta_s=n(s-2)/2. Proposition 2.3, with exponent ss, gives

∥vk∥Ls(gk)s=ℓkβs∥uk∥ss≤Cs+Bsℓkβs.\|v_k\|_{L^s(g_k)}^{s}=\ell_k^{\beta_s}\|u_k\|_s^s\le C_s+B_s\ell_k^{\beta_s}.

If the supremum in (38) tended to zero, the vanishing estimate and interpolation with L2L^2, or with an exponent s>max⁡{q,r0}s>\max\{q,r_0\}, would give ∥vk∥Lq(gk)→0\|v_k\|_{L^q(g_k)}\to0, contradicting (36). Hence there exist zkz_k and c>0c>0 such that

∫Bg(zk,ρℓk)uk2 dVg≥c.\int_{B_g(z_k,\rho\ell_k)}u_k^2\,dV_g\ge c.

Pass to a subsequence for which dg(zk,∂M)/ℓkd_g(z_k,\partial M)/\ell_k either tends to infinity or remains bounded. In the first case interior normal charts have rescaled radii tending to infinity. In the second, Fermi charts at aka_k contain the indicated balls inside a fixed rescaled half-ball. The pulled-back metrics converge to the Euclidean metric on every fixed ball or half-ball.

The bounds (36) give weak H1H^1 convergence and, by the subcritical Rellich theorem, strong local L2L^2 and LqL^q convergence. The displayed lower bound makes the limit nonzero. Lower semicontinuity on expanding balls gives (37). In the boundary case local trace continuity gives zero trace, and the global H1H^1 bound puts ww in H01(R+n)H^1_0(\mathbb{R}^n_+). A nonzero L2L^2 function on either model cannot have zero gradient. □

Lemma 3.3 (Norm decomposition). Let w,m,ew,m,e be as in Lemma 3.2. There are cutoffs of its pushforwards wk∈H01(M)w_k\in H^1_0(M), supported in coordinate neighborhoods whose radii in the original metric tend to zero, such that, for rk=uk−wkr_k=u_k-w_k,

∥rk∥22→1−m,(39)\|r_k\|_2^2\to1-m, \tag*{(39)}
ℓk2∥∇rk∥22→1−e,(40)\ell_k^2\|\nabla r_k\|_2^2\to1-e, \tag*{(40)}
ℓkβ(∥uk∥qq−∥rk∥qq)→∥w∥qq.(41)\ell_k^\beta\left(\lVert u_k\rVert_q^q-\lVert r_k\rVert_q^q\right)\to\lVert w\rVert_q^q. \tag*{(41)}

Proof. At a fixed rescaled radius RR, let wk,R=Pzk,ℓk(χRw)w_{k,R}=P_{z_k,\ell_k}(\chi_R w) in the interior case and wk,R=Pak,ℓk(χRw)w_{k,R}=P_{a_k,\ell_k}(\chi_R w) in the boundary case, using the charts from Lemma 3.2; here χR(y)=χ(y/R)\chi_R(y)=\chi(y/R). Weak local convergence and the local metric comparison give

⟨uk,wk,R⟩2→⟨w,χRw⟩2,∥wk,R∥22→∥χRw∥22,\langle u_k,w_{k,R}\rangle_2\to\langle w,\chi_Rw\rangle_2,\qquad\lVert w_{k,R}\rVert_2^2\to\lVert\chi_Rw\rVert_2^2,
ℓk2⟨∇uk,∇wk,R⟩2→⟨∇w,∇(χRw)⟩2,ℓk2∥∇wk,R∥22→∥∇(χRw)∥22.\ell_k^2\langle\nabla u_k,\nabla w_{k,R}\rangle_2\to\langle\nabla w,\nabla(\chi_Rw)\rangle_2,\qquad\ell_k^2\lVert\nabla w_{k,R}\rVert_2^2\to\lVert\nabla(\chi_Rw)\rVert_2^2.

Strong local LqL^q convergence also gives

ℓkβ(∥uk∥qq−∥uk−wk,R∥qq)⟶∫(∣w∣q−∣w−χRw∣q).\ell_k^\beta\left(\lVert u_k\rVert_q^q-\lVert u_k-w_{k,R}\rVert_q^q\right)\longrightarrow\int\left(|w|^q-|w-\chi_Rw|^q\right).

This is the localized form of the nonlinear splitting of Brézis and Lieb [6]. Since w∈H1w\in H^1, one has χRw→w\chi_Rw\to w in H1H^1 and LqL^q as R→∞R\to\infty. In particular, the cutoff-gradient term is bounded by CR−1∥w∥L2(R<∣y∣<2R)CR^{-1}\lVert w\rVert_{L^2(R<|y|<2R)} and tends to zero.

Choose R=Rk→∞R=R_k\to\infty sufficiently slowly that the preceding limits hold along the diagonal, ℓkRk→0\ell_kR_k\to0, and the supports remain in the coordinate domains. In the interior case require in addition 2ℓkRk<dg(zk,∂M)2\ell_kR_k<d_g(z_k,\partial M). Expanding the two squared norms proves (39) and (40); the nonlinear limit gives (41). □

Lemma 3.4 (An elementary inequality). Let a,b>0a,b>0 with s=a+b>1s=a+b>1. If 0≤m,e≤10\le m,e\le1, then

maeb+(1−m)a(1−e)b≤1.(42)m^ae^b+(1-m)^a(1-e)^b\le1. \tag*{(42)}

If m,e>0m,e>0, equality holds only when m=e=1m=e=1.

Proof. Weighted Young’s inequality bounds the left-hand side by

as(ms+(1−m)s)+bs(es+(1−e)s)≤1.\frac{a}{s}\left(m^s+(1-m)^s\right)+\frac{b}{s}\left(e^s+(1-e)^s\right)\le1.

Since ts+(1−t)s=1t^s+(1-t)^s=1 only at t=0,1t=0,1, positive m,em,e force m=e=1m=e=1 in the equality case. □

Proof of the concentration part of Theorem B. First assume (35). Apply the almost-sharp inequality to rkr_k in Lemma 3.3, multiply by ℓkβ\ell_k^\beta, and let k→∞k\to\infty and then η↓0\eta\downarrow0. Since ∥rk∥2\lVert r_k\rVert_2 is bounded, the lower-order term tends to zero. Thus

lim sup⁡kℓkβ∥rk∥qq≤C∗(1−m)α/2(1−e)β/2.\limsup_k\ell_k^\beta\lVert r_k\rVert_q^q\le C_*(1-m)^{\alpha/2}(1-e)^{\beta/2}.

The flat inequality, with zero extension if necessary, gives ∥w∥qq≤C∗mα/2eβ/2\lVert w\rVert_q^q\le C_*m^{\alpha/2}e^{\beta/2}. The nonlinear splitting now yields

C∗≤C∗(mα/2eβ/2+(1−m)α/2(1−e)β/2)≤C∗.C_*\le C_*\left(m^{\alpha/2}e^{\beta/2}+(1-m)^{\alpha/2}(1-e)^{\beta/2}\right)\le C_*.

Since (α+β)/2=q/2>1(\alpha+\beta)/2=q/2>1, Lemma 3.4 gives m=e=1m=e=1. Moreover ww attains the flat sharp constant. The half-space case is impossible by Lemma 2.1. Hence dg(zk,∂M)/ℓk→∞d_g(z_k,\partial M)/\ell_k\to\infty.

The whole-space equality classification and m=e=1m=e=1 give w=sQ0(⋅−y0)w=sQ_0(\mathord\cdot-y_0) for some s∈{−1,1}s\in\{-1,1\} and y0∈Rny_0\in\mathbb{R}^n. Include the fixed translation in the choice of centers. Smooth dependence of the coordinate maps, applied to the fixed profile and followed by a diagonal choice of cutoff radii, changes the approximation by o(1)o(1) in L2L^2 and by o(ℓk−1)o(\ell_k^{-1}) in the gradient norm. Equations (39) and (40) give (14), and the translated centers still satisfy (15).

For a maximizing sequence at CM=C∗C_M=C_*, a subsequence with bounded gradient norms converges strongly by the direct argument used in the strict case. Otherwise select a subsequence with gradients tending to infinity and apply the preceding proof. The bounded Euclidean-domain statement is Lemma 2.1. □

Scalar curvature and attainment

Moment identities and cutoffs

For x∈K⊂⊂M∘x \in K \subset\subset M^\circ and sufficiently small ε\varepsilon, define

Ux,ε(z)=ε−n/2χ(Ψx−1(z)εR(ε))Q(Ψx−1(z)ε),R(ε)=L∣log⁡ε∣,(43)U_{x,\varepsilon}(z)=\varepsilon^{-n/2}\chi\left(\frac{\Psi_x^{-1}(z)}{\varepsilon R(\varepsilon)}\right)Q\left(\frac{\Psi_x^{-1}(z)}{\varepsilon}\right),\qquad R(\varepsilon)=L|\log\varepsilon|, \tag*{(43)}

extended by zero outside the normal chart, where LL is a sufficiently large fixed constant. For j=0,2j=0,2, write

Mj=∫Rn∣y∣jQ2,Ej=∫Rn∣y∣j∣∇Q∣2,Nj=∫Rn∣y∣jQq.M_j=\int_{\mathbb{R}^n}|y|^jQ^2,\qquad E_j=\int_{\mathbb{R}^n}|y|^j|\nabla Q|^2,\qquad N_j=\int_{\mathbb{R}^n}|y|^jQ^q.

Exponential decay gives, for every fixed integer j≥0j\geq0,

∫∣y∣>R(1+∣y∣)j(Q2+∣∇Q∣2+Qq) dy≤Cje−cR,R≥1.(44)\int_{|y|>R}(1+|y|)^j\left(Q^2+|\nabla Q|^2+Q^q\right)\,dy\leq C_je^{-cR},\qquad R\geq1. \tag*{(44)}

The cutoff argument of [21 Lemma A.1], using (44), bounds all required moment errors by O(ε−c′R)O(\varepsilon^{-c'R}). Choose LL so these are O(ε4)O(\varepsilon^4). The cutoff profiles have uniformly bounded third weighted moments.

Lemma 4.1 (Identities for the Euclidean optimizer). One has

M0=αqN0,E0=βqN0,(45)M_0=\frac{\alpha}{q}N_0,\qquad E_0=\frac{\beta}{q}N_0, \tag*{(45)}

and

E2+M2=N2+nM0.(46)E_2+M_2=N_2+nM_0. \tag*{(46)}

Proof. Testing (16) by QQ gives E0+M0=N0E_0+M_0=N_0. For Qt(y)=tn/2Q(ty)Q_t(y)=t^{n/2}Q(ty), differentiating 12∥∇Qt∥22+12∥Qt∥22−1q∥Qt∥qq\frac{1}{2}\|\nabla Q_t\|_2^2+\frac{1}{2}\|Q_t\|_2^2-\frac{1}{q}\|Q_t\|_q^q at t=1t=1 gives E0=βN0/qE_0=\beta N_0/q. Testing the equation by ∣y∣2Q|y|^2Q and using 2∫Q2y⋅∇Q=−nM02\int Q^2y\cdot\nabla Q=-nM_0 gives (46). Exponential decay justifies these operations. □

The interior expansion

Proof of Theorem C. Let fε=χR(ε)Qf_\varepsilon=\chi_{R(\varepsilon)}Q. In normal coordinates at x∈Kx\in K, uniformly on the shrinking support,

∣g(εy)∣=1−ε26Ric⁡g(x)ijyiyj+OK(ε3∣y∣3).(47)\sqrt{|g(\varepsilon y)|}=1-\frac{\varepsilon^2}{6}\operatorname{Ric}_g(x)_{ij}y_iy_j+O_K(\varepsilon^3|y|^3). \tag*{(47)}

For radial fεf_\varepsilon, the Gauss lemma gives gij(εy)∂ifε∂jfε=∣∂rfε∣2g^{ij}(\varepsilon y)\partial_i f_\varepsilon\partial_j f_\varepsilon=|\partial_r f_\varepsilon|^2. Radial integration contracts the quadratic volume term to Scal⁡g(x)/n\operatorname{Scal}_g(x)/n. Thus

∥Ux,ε∥22=M0−Scal⁡g(x)6nε2M2+OK(ε3),(48)\|U_{x,\varepsilon}\|_2^2=M_0-\frac{\operatorname{Scal}_g(x)}{6n}\varepsilon^2M_2+O_K(\varepsilon^3), \tag*{(48)}
ε2∥∇Ux,ε∥22=E0−Scal⁡g(x)6nε2E2+OK(ε3),(49)\varepsilon^2\|\nabla U_{x,\varepsilon}\|_2^2=E_0-\frac{\operatorname{Scal}_g(x)}{6n}\varepsilon^2E_2+O_K(\varepsilon^3), \tag*{(49)}
εβ∥Ux,ε∥qq=N0−Scal⁡g(x)6nε2N2+OK(ε3).(50)\varepsilon^\beta\|U_{x,\varepsilon}\|_q^q=N_0-\frac{\operatorname{Scal}_g(x)}{6n}\varepsilon^2N_2+O_K(\varepsilon^3). \tag*{(50)}

The remainders are uniform because the third weighted moments are bounded independently of the cutoff radius. Since N0/(M0α/2E0β/2)=C∗N_0/(M_0^{\alpha/2}E_0^{\beta/2})=C_*, expansion of the quotient yields

WM(Ux,ε)C∗=1−Scal⁡g(x)6nε2(N2N0−α2M2M0−β2E2E0)+OK(ε3).(51)\frac{W_M(U_{x,\varepsilon})}{C_*}=1-\frac{\operatorname{Scal}_g(x)}{6n}\varepsilon^2\left(\frac{N_2}{N_0}-\frac{\alpha}{2}\frac{M_2}{M_0}-\frac{\beta}{2}\frac{E_2}{E_0}\right)+O_K(\varepsilon^3). \tag*{(51)}

By Lemma 4.1,

N2N0−α2M2M0−β2E2E0=N2N0−q(M2+E2)2N0=−q−22N2N0−nα2.\begin{aligned} \frac{N_2}{N_0}-\frac{\alpha}{2}\frac{M_2}{M_0}-\frac{\beta}{2}\frac{E_2}{E_0} =\frac{N_2}{N_0}-\frac{q(M_2+E_2)}{2N_0} \\ &=-\frac{q-2}{2}\frac{N_2}{N_0}-\frac{n\alpha}{2}. \end{aligned}

Substituting into (51) gives (17) and (18). If scalar curvature is positive at an interior point, Ux,εU_{x,\varepsilon} has quotient larger than C∗C_{*} for sufficiently small ε\varepsilon. Theorem B gives attainment. The last assertion is the contrapositive. □

Corollary 4.2 (Small balls). For a fixed interior point xx, CBg(x,r)→C∗C_{B_g(x,r)} \to C_{*} as r↓0r \downarrow0. If Scal⁡g(x)>0\operatorname{Scal}_g(x)>0, then every sufficiently small such ball has CBg(x,r)>C∗C_{B_g(x,r)}>C_{*} and admits an extremal.

Proof. Zero extension in a normal chart and the metric comparison give CBg(x,r)≤C∗(1+Cr2)C_{B_g(x,r)} \le C_{*}(1+Cr^{2}), while local insertion gives the reverse lower bound C∗C_{*}. If scalar curvature is positive at xx, choose ε\varepsilon so that Ux,εU_{x,\varepsilon} is supported in Bg(x,r)B_g(x,r) and use Theorem C. □

Extremals on manifolds with small holes

Throughout this section (N,g)(N,g) is fixed, smooth, closed, and connected, and V=Vol⁡(N)V=\operatorname{Vol}(N). Put Λ=λ+(N)>0\Lambda=\lambda_{+}(N)>0, the first positive Laplace eigenvalue. Then

∥f−fˉ∥22≤Λ−1∥∇f∥22,fˉ=V−1∫Nf.(52)\|f-\bar f\|_{2}^{2}\leq\Lambda^{-1}\|\nabla f\|_{2}^{2},\qquad\bar f=V^{-1}\int_N f. \tag*{(52)}

We identify Dirichlet functions with their zero extensions to NN and suppress the tildes in the proofs. The H1H^{1} norm is ∥f∥H1(N)2=∥f∥22+∥∇f∥22\|f\|_{H^{1}(N)}^{2}=\|f\|_{2}^{2}+\|\nabla f\|_{2}^{2}. All constants in this section depend only on NN and the fixed exponent qq.

Small energy and localization near the first eigenspace

Lemma 5.1 (Nonlinear norm at small energy). There are E∗>0E_{*}>0 and K<∞K<\infty such that, if f∈H1(N)f\in H^{1}(N), f≥0f\geq0, ∥f∥2=1\|f\|_{2}=1, and E=∥∇f∥22≤E∗E=\|\nabla f\|_{2}^{2}\leq E_{*}, then

V1−q/2≤∥f∥qq≤V1−q/2+KE,∥f−V−1/2∥H1(N)≤KE1/2.(53)V^{1-q/2}\leq\|f\|_{q}^{q}\leq V^{1-q/2}+KE,\qquad\|f-V^{-1/2}\|_{H^{1}(N)}\leq KE^{1/2}. \tag*{(53)}

Proof. Write f=a+hf=a+h, with a=fˉ≥0a=\bar f\geq0 and ∫Nh=0\int_N h=0. The Poincaré inequality and the Sobolev embedding give ∥h∥22≤E/Λ\|h\|_{2}^{2}\leq E/\Lambda and ∥h∥q≤KE1/2\|h\|_{q}\leq KE^{1/2}. Since Va2=1−∥h∥22Va^{2}=1-\|h\|_{2}^{2}, one has a=V−1/2+O(E)a=V^{-1/2}+O(E). The inequality ∣(a+t)q−aq−qaq−1t∣≤K(t2+∣t∣q)\lvert(a+t)^{q}-a^{q}-qa^{q-1}t\rvert\leq K(t^{2}+\lvert t\rvert^{q}), valid for these aa and a+t≥0a+t\geq0, yields

∫Nfq=Vaq+O(∥h∥22+∥h∥qq)=V1−q/2+O(E).\int_N f^{q}=Va^{q}+O(\|h\|_{2}^{2}+\|h\|_{q}^{q})=V^{1-q/2}+O(E).

Hölder’s inequality gives the lower bound; the same decomposition gives the H1H^{1} estimate. □

Fix a smooth hole DD with connected complement. Write λ=λD\lambda=\lambda_D, φ=φD\varphi=\varphi_D, and S=V1−q/2S=V^{1-q/2}. Corollary 2.4 gives

CD≥Sλ−β/2.(54)C_D\geq S\lambda^{-\beta/2}. \tag*{(54)}

For small λ\lambda this exceeds C∗C_{*}, so Theorem B supplies a nonnegative normalized maximizer.

Lemma 5.2 (Energy of near-maximizers). For sufficiently small λ\lambda, every nonnegative f∈H01(MD)f\in H^{1}_{0}(M_D) with ∥f∥2=1\|f\|_{2}=1 and δ=δD(f)≤1/2\delta=\delta_D(f)\leq1/2 satisfies

λ≤E:=∥∇f∥22≤λ(1+K(δ+λ)).(55)\lambda\leq E:=\|\nabla f\|_{2}^{2}\leq\lambda(1+K(\delta+\lambda)). \tag*{(55)}

Moreover,

S≤CDλβ/2≤S+Kλ.(56)S\leq C_D\lambda^{\beta/2}\leq S+K\lambda. \tag*{(56)}

Proof. The inequality on NN (30) gives

12Sλ−β/2≤WMD(f)≤K(1+E−β/2),\frac{1}{2}S\lambda^{-\beta/2}\leq W_{M_D}(f)\leq K(1+E^{-\beta/2}),

so E≤KλE\leq K\lambda for small λ\lambda. The Rayleigh principle gives E≥λE\geq\lambda. By Lemma 5.1, ∥f∥qq≤S+Kλ\|f\|_{q}^{q}\leq S+K\lambda, whence

Eλ≤(1+Kλ1−δ)2/β≤1+K′(δ+λ).\frac{E}{\lambda}\leq\left(\frac{1+K\lambda}{1-\delta}\right)^{2/\beta}\leq1+K'(\delta+\lambda).

This proves (55). For a maximizer uu, it also gives ∥u∥qq≤S+Kλ\|u\|_{q}^{q}\leq S+K\lambda. Together with ∥∇u∥22≥λ\|\nabla u\|_{2}^{2}\geq\lambda and (54), this proves (56). □

The second Dirichlet eigenvalue satisfies

λ2(MD)≥Λ.(57)\lambda_2(M_D) \ge\Lambda. \tag*{(57)}

Indeed, every two-dimensional subspace of zero-extended Dirichlet functions contains a nonzero function of mean zero on NN, and the min–max principle applies. Set

XD={z∈H01(MD):⟨z,φ⟩2=0},BD(z,h)=∫N∇z⋅∇h−λ∫Nzh.X_D = \{z \in H^1_0(M_D) : \langle z,\varphi\rangle_2 = 0\}, \qquad B_D(z,h) = \int_N \nabla z \cdot\nabla h - \lambda\int_N zh.

For λ≤Λ/2\lambda\le\Lambda/2, (57) gives

cN∥z∥H1(N)2≤BD(z,z)≤∥z∥H1(N)2(z∈XD),(58)c_N\|z\|_{H^1(N)}^2 \le B_D(z,z) \le\|z\|_{H^1(N)}^2 \qquad(z \in X_D), \tag*{(58)}

where cN>0c_N > 0 is independent of DD. If f≥0f \ge0 and ∥f∥2=1\|f\|_2 = 1, write

f=FD(z):=a(z)φ+z,a(z)=1−∥z∥22,z∈XD.(59)f = F_D(z) := a(z)\varphi+ z, \qquad a(z) = \sqrt{1-\|z\|_2^2}, \qquad z \in X_D. \tag*{(59)}

This representation uses a(z)=⟨f,φ⟩2>0a(z) = \langle f,\varphi\rangle_2 > 0. Orthogonality and the eigenfunction equation give the exact identity

∥∇FD(z)∥22=λ+BD(z,z).(60)\|\nabla F_D(z)\|_2^2 = \lambda+ B_D(z,z). \tag*{(60)}

Strict concavity and quantitative stability

For sufficiently small λ\lambda, define the convex neighborhood

UD={z∈XD:BD(z,z)<λ/2}.U_D = \{z \in X_D : B_D(z,z) < \lambda/2\}.

By (58), a(z)a(z) is uniformly bounded away from zero on this set. We use (59) for all real z∈UDz \in U_D, and put

JD(z)=∫N∣FD(z)∣q,LD(z)=log⁡JD(z)−β2log⁡(λ+BD(z,z)).(61)J_D(z) = \int_N |F_D(z)|^q, \qquad\mathcal{L}_D(z) = \log J_D(z) - \frac{\beta}{2}\log(\lambda+ B_D(z,z)). \tag*{(61)}

Thus LD(z)=log⁡WMD(FD(z))\mathcal{L}_D(z) = \log W_{M_D}(F_D(z)).

Lemma 5.3 (Uniform strict concavity). For sufficiently small λ\lambda,

D2LD(z)[h,h]≤−cλ∥h∥H1(N)2(z∈UD, h∈XD),(62)D^2\mathcal{L}_D(z)[h,h] \le-\frac{c}{\lambda}\|h\|_{H^1(N)}^2 \qquad(z \in U_D,\ h \in X_D), \tag*{(62)}

where c>0c > 0 depends only on NN and qq. On UDU_D, the map FDF_D is uniformly Lipschitz from H1H^1 to H1H^1.

Proof. The derivatives of aa are

Da(z)[h]=−⟨z,h⟩2a(z),D2a(z)[h,h]=−∥h∥22a(z)−⟨z,h⟩22a(z)3.Da(z)[h] = -\frac{\langle z,h\rangle_2}{a(z)}, \qquad D^2a(z)[h,h] = -\frac{\|h\|_2^2}{a(z)} - \frac{\langle z,h\rangle_2^2}{a(z)^3}.

Since aa is bounded below and ∥φ∥H12=1+λ\|\varphi\|_{H^1}^2 = 1+\lambda, these give

∥DFD(z)[h]∥H1≤K∥h∥H1,∥D2FD(z)[h,h]∥H1≤K∥h∥H12.(63)\|DF_D(z)[h]\|_{H^1} \le K\|h\|_{H^1}, \qquad\|D^2F_D(z)[h,h]\|_{H^1} \le K\|h\|_{H^1}^2. \tag*{(63)}

The first bound gives Lipschitz continuity. Since 2<q<2∗2 < q < 2^*, the nonlinear integral is C2C^2 on H1(N)H^1(N). Writing F=FD(z)F = F_D(z),

DJD(z)[h]=q∫N∣F∣q−2F DFD(z)[h],DJ_D(z)[h] = q\int_N |F|^{q-2}F\,DF_D(z)[h],
D2JD(z)[h,h]=q(q−1)∫N∣F∣q−2(DFD(z)[h])2+q∫N∣F∣q−2F D2FD(z)[h,h].D^2J_D(z)[h,h] = q(q-1)\int_N |F|^{q-2}(DF_D(z)[h])^2 + q\int_N |F|^{q-2}F\,D^2F_D(z)[h,h].

Since ∥F∥2=1\|F\|_2 = 1, ∥F∥H12≤1+3λ/2\|F\|_{H^1}^2 \le1+3\lambda/2, and JD(z)≥SJ_D(z) \ge S by Hölder’s inequality, the Sobolev embedding on NN gives

∣D2log⁡JD(z)[h,h]∣≤K∥h∥H12.(64)|D^2\log J_D(z)[h,h]| \le K\|h\|_{H^1}^2. \tag*{(64)}

This holds also for sign-changing FF.

Put t=BD(z,z)t = B_D(z,z). The Cauchy–Schwarz inequality for the positive form BDB_D gives

D2log⁡(λ+BD(z,z))[h,h]=2BD(h,h)λ+t−4BD(z,h)2(λ+t)2.≥2(λ−t)(λ+t)2BD(h,h)≥49λBD(h,h),\begin{aligned} D^2\log(\lambda+B_D(z,z))[h,h] = \frac{2B_D(h,h)}{\lambda+t} - \frac{4B_D(z,h)^2}{(\lambda+t)^2}. \\ &\ge\frac{2(\lambda-t)}{(\lambda+t)^2}B_D(h,h)\ge\frac{4}{9\lambda}B_D(h,h), \end{aligned}

since 0≤t<λ/20\le t<\lambda/2. Together with (64) and (58), this gives (62) for sufficiently small λ\lambda. □

Uniqueness and stability in Theorem A. By (55), choose δ0\delta_0 and then λ0\lambda_0 such that every admissible ff with δD(f)≤δ0\delta_D(f)\le\delta_0 is represented by f=FD(z)f=F_D(z) with BD(z,z)<λ/4B_D(z,z)<\lambda/4. In particular, this applies to every nonnegative normalized maximizer. Taking absolute values preserves the quotient on real functions, so each such maximizer also maximizes the quotient over all real functions. Its coordinate is therefore an interior critical point of LD\mathcal L_D. Existence follows from (54) and Theorem B. Strict concavity gives a unique such point zDz_D; set uD=FD(zD)u_D=F_D(z_D).

For f=FD(z)f=F_D(z) with δ=δD(f)≤δ0\delta=\delta_D(f)\le\delta_0, integrate (62) on the segment from zDz_D to zz, using DLD(zD)=0D\mathcal L_D(z_D)=0. This gives

c2λ∥z−zD∥H12≤LD(zD)−LD(z)=−log⁡(1−δ)≤2δ.\frac{c}{2\lambda}\lVert z-z_D\rVert_{H^1}^2\le\mathcal L_D(z_D)-\mathcal L_D(z)=-\log(1-\delta)\le2\delta.

The Lipschitz estimate for FDF_D now proves (11). □

Approximation by the first eigenfunction

Lemma 5.4 (First variation at the eigenfunction). For small λ\lambda,

∣DLD(0)[h]∣≤Kλ1/2∥h∥H1(N)(h∈XD).(65)\lvert D\mathcal L_D(0)[h]\rvert\le K\lambda^{1/2}\lVert h\rVert_{H^1(N)}\qquad(h\in X_D). \tag*{(65)}

Proof. The first derivative of λ+BD(z,z)\lambda+B_D(z,z) vanishes at zero and DFD(0)[h]=hDF_D(0)[h]=h. Let c0=V−1/2c_0=V^{-1/2}. Orthogonality to φ\varphi gives

DLD(0)[h]=q∫Nφq∫N(φq−1−c0q−2φ)h.D\mathcal L_D(0)[h]=\frac{q}{\int_N\varphi^q}\int_N(\varphi^{q-1}-c_0^{q-2}\varphi)h.

The pointwise bound ∣tq−1−c0q−2t∣≤K(1+tq−2)∣t−c0∣\lvert t^{q-1}-c_0^{q-2}t\rvert\le K(1+t^{q-2})\lvert t-c_0\rvert for t≥0t\ge0 and Hölder’s inequality, with q′=q/(q−1)q'=q/(q-1), give

∥φq−1−c0q−2φ∥q′≤K(1+∥φ∥qq−2)∥φ−c0∥q.\lVert\varphi^{q-1}-c_0^{q-2}\varphi\rVert_{q'}\le K(1+\lVert\varphi\rVert_q^{q-2})\lVert\varphi-c_0\rVert_q.

Lemma 5.1, applied to φ\varphi, gives ∥φ−c0∥H1≤Kλ1/2\lVert\varphi-c_0\rVert_{H^1}\le K\lambda^{1/2}. The Sobolev embedding and ∫Nφq≥S\int_N\varphi^q\ge S prove (65). □

Spectral estimates in Theorem A. Equation (56) proves (6). By (55) with zero deficit and Lemma 5.1, ∥uD−V−1/2∥H1=O(λ1/2)\lVert u_D-V^{-1/2}\rVert_{H^1}=O(\lambda^{1/2}), proving (10).

Using DLD(zD)=0D\mathcal L_D(z_D)=0, integrate the Hessian on the segment from 00 to zDz_D. Lemmas 5.3 and 5.4 give

cλ∥zD∥H12≤DLD(0)[zD]≤Kλ1/2∥zD∥H1.\frac{c}{\lambda}\lVert z_D\rVert_{H^1}^2\le D\mathcal L_D(0)[z_D]\le K\lambda^{1/2}\lVert z_D\rVert_{H^1}.

Thus ∥zD∥H1=O(λ3/2)\lVert z_D\rVert_{H^1}=O(\lambda^{3/2}). Since a(zD)=1+O(∥zD∥22)=1+O(λ3)a(z_D)=1+O(\lVert z_D\rVert_2^2)=1+O(\lambda^3) and uD=a(zD)φ+zDu_D=a(z_D)\varphi+z_D, this proves (7). The exact identity (60) gives

0≤∥∇uD∥22−λ=BD(zD,zD)≤∥zD∥H12=O(λ3),0\le\lVert\nabla u_D\rVert_2^2-\lambda=B_D(z_D,z_D)\le\lVert z_D\rVert_{H^1}^2=O(\lambda^3),

which is (9). Finally, concavity and maximality imply

0≤log⁡CDWMD(φD)=LD(zD)−LD(0)≤DLD(0)[zD]=O(λ2).0\le\log\frac{C_D}{W_{M_D}(\varphi_D)}=\mathcal L_D(z_D)-\mathcal L_D(0)\le D\mathcal L_D(0)[z_D]=O(\lambda^2).

Exponentiation proves (8). □

Remark 5.5. Since φ~D=0\widetilde{\varphi}_D=0 on DD, Lemma 5.1 gives V−1Vol⁡(D)≤∥φ~D−V−1/2∥22=O(λD)V^{-1}\operatorname{Vol}(D)\le\lVert\widetilde{\varphi}_D-V^{-1/2}\rVert_2^2=O(\lambda_D). Thus vanishing hole volume follows from the spectral assumption.

Capacity and the first eigenvalue

The admissible class in (4) is nonempty and weakly closed. Poincaré gives a unique minimizer of positive energy. Smoothness of ∂D\partial D ensures that 1−v∈H01(MD)1-v\in H^1_0(M_D) for every admissible vv.

Courtois [15 Theorem 1.3] proved (67) for this capacity. We use the following two-sided estimate.

Proposition 5.6 (Spectral-capacity comparison). Let cD=cap⁡N(D)c_D=\operatorname{cap}_N(D). Then

cDV+cD/Λ≤λD≤cDV.(66)\frac{c_D}{V+c_D/\Lambda}\leq\lambda_D\leq\frac{c_D}{V}. \tag*{(66)}

Consequently cD→0c_D\to0 if and only if λD→0\lambda_D\to0, and in this limit

λD=cDV(1+O(cD)).(67)\lambda_D=\frac{c_D}{V}(1+O(c_D)). \tag*{(67)}

Proof. For the capacitary minimizer vDv_D, the Dirichlet test function 1−vD1-v_D has squared norm V+∥vD∥22V+\lVert v_D\rVert_2^2 and energy cDc_D, giving λD≤cD/V\lambda_D\leq c_D/V.

For the other direction, let a=V−1∫NφD>0a=V^{-1}\int_N\varphi_D>0. The function v=1−φD/av=1-\varphi_D/a is admissible for the capacity and cDa2≤λDc_Da^2\leq\lambda_D. By (52),

Va2=1−∥φD−a∥22≥1−λD/Λ.Va^2=1-\lVert\varphi_D-a\rVert_2^2\geq1-\lambda_D/\Lambda.

Combining the last two inequalities and rearranging gives the lower bound in (66). The equivalence and (67) follow. □

Substituting (67) into (6), with 1−q/2+β/2=1−α/21-q/2+\beta/2=1-\alpha/2, proves (12) and completes Theorem A.

Remark 5.7 (Torsional representation of the capacity). If −ΔgtD=1-\Delta_g t_D=1 in MDM_D and tD∈H01(MD)t_D\in H^1_0(M_D), then cap⁡N(D)=V2/∫MDtD\operatorname{cap}_N(D)=V^2/\int_{M_D}t_D. Indeed, w=1−vw=1-v has ∫w=V\int w=V and V2=∫∇w⋅∇tDV^2=\int\nabla w\cdot\nabla t_D. Cauchy–Schwarz gives the lower bound, with equality at w=VtD/∫tDw=Vt_D/\int t_D.

Geodesic balls

Write ωn−1=∣Sn−1∣\omega_{n-1}=|\mathbb{S}^{n-1}| and κn=(n−2)ωn−1\kappa_n=(n-2)\omega_{n-1} for n≥3n\geq3.

Lemma 5.8 (Capacity of a small ball). For a fixed x∈Nx\in N,

cap⁡N(Bg(x,r))∼{κnrn−2,n≥3,2π/∣log⁡r∣,n=2,r↓0.(68)\operatorname{cap}_N(B_g(x,r))\sim \begin{cases} \kappa_n r^{n-2}, & n\geq3,\\ 2\pi/|\log r|, & n=2, \end{cases} \qquad r\downarrow0. \tag*{(68)}

Proof. Fix 0<δ0<\delta below the normal-coordinate radius at xx. Let frf_r equal one on BrB_r, zero outside BδB_\delta, and on the intervening annulus set

fr(s)={s2−n−δ2−nr2−n−δ2−n,n≥3,log⁡(δ/s)log⁡(δ/r),n=2.f_r(s)= \begin{cases} \dfrac{s^{2-n}-\delta^{2-n}}{r^{2-n}-\delta^{2-n}}, & n\geq3,\\ \dfrac{\log(\delta/s)}{\log(\delta/r)}, & n=2. \end{cases}

The radial energy in the Riemannian metric differs by a factor 1+O(δ2)1+O(\delta^2) from its Euclidean value

ωn−1∫rδs1−n ds.\frac{\omega_{n-1}}{\int_r^\delta s^{1-n}\,\mathrm{d}s}.

Since 0≤fr≤10\leq f_r\leq1 and fr→0f_r\to0 almost everywhere, ar=V−1∫Nfr→0a_r=V^{-1}\int_Nf_r\to0. The function (fr−ar)/(1−ar)(f_r-a_r)/(1-a_r) is admissible for the mean-zero capacity. This gives the required upper bound, up to a factor tending to 1+O(δ2)1+O(\delta^2).

Let vrv_r be the mean-zero capacitary minimizer. The upper bound and (52) imply vr→0v_r\to0 in H1(N)H^1(N). Hence its trace on the fixed sphere of radius δ\delta tends to zero in L2L^2. In Euclidean polar coordinates inside the normal chart, let br(s)b_r(s) be the spherical average of vrv_r. Its trace at s=rs=r is one, while br(δ)→0b_r(\delta)\to0. The Gauss lemma, the volume-density comparison, and Cauchy–Schwarz give

cap⁡N(Bg(x,r))≥(1−Cδ2)ωn−1∫rδsn−1∣br′(s)∣2 ds≥(1−Cδ2)ωn−1∣1−br(δ)∣2∫rδs1−n ds.\begin{aligned} \operatorname{cap}_N(B_g(x,r))\geq(1-C\delta^2)\omega_{n-1}\int_r^\delta s^{n-1}|b_r'(s)|^2\,\mathrm{d}s \\ &\ge(1-C\delta^2)\frac{\omega_{n-1}|1-b_r(\delta)|^2}{\int_r^\delta s^{1-n}\,ds}. \end{aligned}

First let r↓0r \downarrow0 and then δ↓0\delta\downarrow0. Evaluating the denominator proves (68) in both dimension ranges.

Corollary 5.9 (Complements of small geodesic balls). As r↓0r \downarrow0,

CN∖Bg(x,r)∼{V1−α/2κn−β/2r−(n−2)β/2,n≥3,(2π)−β/2∣log⁡r∣β/2,n=2.(69)C_{N\setminus B_g(x,r)} \sim \begin{cases} V^{1-\alpha/2}\kappa_n^{-\beta/2}r^{-(n-2)\beta/2}, & n\ge3,\\ (2\pi)^{-\beta/2}|\log r|^{\beta/2}, & n=2. \end{cases} \tag*{(69)}

Proof. Combine Theorem A and Lemma 5.8. In dimension two, α=2\alpha=2, so the volume factor is one.

Remark 5.10 (Examples). If NN is a flat torus and Mr=N∖Bg(x,r)‾M_r=N\setminus\overline{B_g(x,r)}, then CMr→∞C_{M_r}\to\infty as r↓0r\downarrow0, and the nonnegative normalized extremal is unique for sufficiently small rr. In particular, attainment does not imply positive scalar curvature. Bounded Euclidean domains, on the other hand, have CM=C∗C_M=C_* and no extremal.

The L2L^2-critical nonlinear Schrödinger energy

Assume in this section that q=2+4/nq=2+4/n. Then β=2\beta=2, α=4/n\alpha=4/n, and the energy and masses are defined by (19) and (20). The same infimum is obtained for real or complex functions: passing from uu to ∣u∣|u| preserves mass and the nonlinear integral and cannot increase the gradient energy.

Proof of Theorem D. For a function of mass mm, the definition of WMW_M gives the exact identity

EM(u)=(12−m2/nqWM(u))∥∇u∥22.(70)\mathcal{E}_M(u)=\left(\frac{1}{2}-\frac{m^{2/n}}{q}W_M(u)\right)\|\nabla u\|_2^2. \tag*{(70)}

The energy is therefore nonnegative for m≤m0(M)m\le m_0(M). For m>m0(M)m>m_0(M), choose vv with WM(v)>q/(2m2/n)W_M(v)>q/(2m^{2/n}) and multiply it by a constant to obtain mass mm. The resulting function has negative energy.

Suppose 0<m<m∗0<m<m_*. Choose η>0\eta>0 such that (C∗+η)m2/n<q/2(C_*+\eta)m^{2/n}<q/2. The almost-sharp inequality gives, under the constraint ∥u∥22=m\|u\|_2^2=m,

EM(u)≥(12−C∗+ηqm2/n)∥∇u∥22−Bηqmq/2.(71)\mathcal{E}_M(u)\ge\left(\frac{1}{2}-\frac{C_*+\eta}{q}m^{2/n}\right)\|\nabla u\|_2^2-\frac{B_\eta}{q}m^{q/2}. \tag*{(71)}

A minimizing sequence is therefore bounded in H01(M)H_0^1(M). Compactness into L2L^2 and LqL^q preserves the mass and the nonlinear integral, while weak lower semicontinuity of the gradient term gives a minimizer.

For m>m∗m>m_*, a compactly supported approximation to a suitable constant multiple of a Euclidean optimizer gives a function ff of mass mm with negative Euclidean energy. Apply Px,εP_{x,\varepsilon} at an interior point xx and multiply by 1+O(ε2)1+O(\varepsilon^2) so that the resulting function has mass mm. Local comparison gives

EM(Px,εf)=ε−2(12∥∇f∥22−1q∥f∥qq)+O(1),\mathcal{E}_M(P_{x,\varepsilon}f)=\varepsilon^{-2}\left(\frac{1}{2}\|\nabla f\|_2^2-\frac{1}{q}\|f\|_q^q\right)+O(1),

with the same leading term after normalization. Hence IM(m)=−∞I_M(m)=-\infty.

For CM>C∗C_M>C_*, the interval (m0(M),m∗)(m_0(M),m_*) has negative-energy minimizers by the preceding arguments. At β=2\beta=2, (6) reads CD=V−2/nλD−1(1+O(λD))C_D=V^{-2/n}\lambda_D^{-1}(1+O(\lambda_D)); substitution in (20) gives (21).

Corollary 6.1 (Zero and negative energy standing waves). If CM>C∗C_M>C_*, then IM(m0(M))=0I_M(m_0(M))=0 is attained by a suitably normalized GN extremal. For m0(M)≤m<m∗m_0(M)\le m<m_*, every nonnegative constrained minimizer is positive in M∘M^\circ and solves

−Δgu+ωu=uq−1,u∣∂M=0.(72)-\Delta_g u+\omega u=u^{q-1},\qquad u|_{\partial M}=0. \tag*{(72)}

for some ω>0\omega> 0. It gives the standing wave ψ(t,z)=eiωtu(z)\psi(t,z)=e^{i\omega t}u(z) for the focusing Dirichlet equation i∂tψ+Δgψ+∣ψ∣q−2ψ=0i\partial_t\psi+\Delta_g\psi+|\psi|^{q-2}\psi=0.

Proof. At mass m0(M)m_0(M), the energy is nonnegative by (70), and a suitably normalized GN extremal makes it zero. For a constrained minimizer, the multiplier in (72) satisfies ωm=∥u∥qq−∥∇u∥22\omega m=\|u\|_q^q-\|\nabla u\|_2^2. Since ∥u∥qq≥(q/2)∥∇u∥22\|u\|_q^q\ge(q/2)\|\nabla u\|_2^2 at nonpositive energy, ω>0\omega>0. Elliptic regularity and the maximum principle give positivity; substitution verifies the nonlinear Schrödinger equation. □

Corollary 6.2 (Capacity and the threshold for negative energy). For a family MD=N∖D‾M_D=N\setminus\overline{D} with connected complements and λD→0\lambda_D\to0,

m0(MD)∼(q2)n/2V1−n/2cap⁡N(D)n/2.(73)m_0(M_D)\sim\left(\frac{q}{2}\right)^{n/2}V^{1-n/2}\operatorname{cap}_N(D)^{n/2}. \tag*{(73)}

Consequently, for every fixed m∈(0,m∗)m\in(0,m_*), all sufficiently small holes admit a negative-energy constrained minimizer of mass mm.

Proof. Use (21) and (67). The last conclusion follows from m0(MD)→0m_0(M_D)\to0 and Theorem D. □

Further applications

The sharp constants and stability estimates above also provide input for nonlinear evolution equations. We indicate two applications and the steps needed to develop them. We do not undertake the corresponding analysis of the evolution problems here.

Fast diffusion and extinction estimates

Consider the fast diffusion equation ∂tρ=Δg(ρm)\partial_t\rho=\Delta_g(\rho^m) with ρ≥0\rho\ge0 and ρm∈H01(M)\rho^m\in H_0^1(M), where 0<m<10<m<1. Choose q=1+1/mq=1+1/m and assume q<2∗q<2^*; thus m>(n−2)/(n+2)m>(n-2)/(n+2) when n≥3n\ge3. In this subsection, CMC_M, α\alpha, and β\beta refer to this exponent qq. The natural starting point is the dissipation identity

Y(t):=∫Mρm+1 dVg,Y′(t)=−(m+1)∫M∣∇(ρm)∣2 dVg.Y(t):=\int_M\rho^{m+1}\,dV_g,\qquad Y'(t)=-(m+1)\int_M|\nabla(\rho^m)|^2\,dV_g.

Applying the GN inequality to f=ρmf=\rho^m and then using ∥f∥2≤Vol⁡(M)1/2−1/q∥f∥q\|f\|_2\le\operatorname{Vol}(M)^{1/2-1/q}\|f\|_q gives

Y′(t)≤−(m+1)KMY(t)2m/(m+1),KM=CM−2/βVol⁡(M)−2α/(nq).(74)Y'(t)\le-(m+1)K_MY(t)^{2m/(m+1)},\qquad K_M=C_M^{-2/\beta}\operatorname{Vol}(M)^{-2\alpha/(nq)}. \tag*{(74)}

For solutions satisfying the dissipation identity, integration yields the extinction-time upper bound Text≤Y(0)(1−m)/(m+1)/((1−m)KM)T_{\mathrm{ext}}\le Y(0)^{(1-m)/(m+1)}/((1-m)K_M). A complementary lower bound follows by testing the equation against the positive first Dirichlet eigenfunction Φ1\Phi_1, normalized by ∫MΦ1 dVg=1\int_M\Phi_1\,dV_g=1. Concavity of s↦sms\mapsto s^m gives

ddt∫MρΦ1 dVg≥−λ1(M)(∫MρΦ1 dVg)m,Text≥(∫Mρ(0)Φ1 dVg)1−m(1−m)λ1(M).\frac{d}{dt}\int_M\rho\Phi_1\,dV_g\ge-\lambda_1(M)\left(\int_M\rho\Phi_1\,dV_g\right)^m,\qquad T_{\mathrm{ext}}\ge\frac{\left(\int_M\rho(0)\Phi_1\,dV_g\right)^{1-m}}{(1-m)\lambda_1(M)}.

For MD=N∖D‾M_D=N\setminus\overline{D} as in Theorem A, the sharp-constant asymptotic and Vol⁡(MD)→V\operatorname{Vol}(M_D)\to V give

KMD∼λDV(1−m)/(m+1),λD−1∼Vcap⁡N(D).(75)K_{M_D}\sim\lambda_D V^{(1-m)/(m+1)},\qquad\lambda_D^{-1}\sim\frac{V}{\operatorname{cap}_N(D)}. \tag*{(75)}

For initial data obtained by restricting a fixed bounded, nonnegative, nonzero function on NN, the normalized eigenfunction converges to V−1V^{-1}, so the lower bound has a positive limiting numerator. The two estimates therefore give the same order λD−1\lambda_D^{-1} for the extinction time. This supplies a direct route from the GN asymptotic to quantitative dependence of extinction on hole capacity. A treatment of weak solutions, rescaled extinction profiles, and convergence rates would require further analysis. The use of interpolation and stability inequalities in entropy methods is developed in [11, 4].

Stability of standing waves

Return to q=2+4/nq = 2 + 4/n. § 6 establishes the variational mass thresholds and existence of standing waves. These results also provide a starting point for studying dynamical stability. At the zero-energy mass on MDM_D, put UD=m0(MD)uDU_D = \sqrt{m_0(M_D)}u_D. For U=m0(MD)fU = \sqrt{m_0(M_D)}f with f≥0f \ge0 and ∥f∥2=1\lVert f\rVert_2 = 1, (70) gives

EMD(U)=m0(MD)2∥∇f∥22δD(f).\mathcal{E}_{M_D}(U) = \frac{m_0(M_D)}{2}\lVert\nabla f\rVert_2^2\delta_D(f).

Together with ∥∇f∥22≥λD\lVert\nabla f\rVert_2^2 \ge\lambda_D, the deficit estimate in Theorem A bounds ∥U−UD∥H12\lVert U - U_D\rVert_{H^1}^2 by a constant times EMD(U)\mathcal{E}_{M_D}(U) when the relative deficit is small. This gives quantitative control of nonnegative near-minimizers. To study orbital stability, one would combine compactness of constrained minimizing sequences with conservation of mass and energy for a well-posed Dirichlet flow, and account for the complex phase. The capacity law for m0(MD)m_0(M_D) and the approximation of uDu_D by φD\varphi_D make the small-hole limit a natural setting for this analysis. We leave these dynamical questions for future work.

Acknowledgements

Acknowledgements. The authors thank IIT Bombay for providing ideal working conditions.

References

Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai 400076, India Email address: [email protected], [email protected]

Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai 400076, India Email address: [email protected], [email protected]

References

  1. [1]Laura Abatangelo, Veronica Felli, Luc Hillairet, and Corentin Léna. Spectral stability under removal of small capacity sets and applications to Aharonov–Bohm operators. J. Spectr. Theory, 9(2):379–427, 2019.arxiv.org/abs/1611.06750
  2. [2]Riccardo Adami, Enrico Serra, and Paolo Tilli. Negative energy ground states for the L²-critical NLSE on metric graphs. Comm. Math. Phys., 352(1):387–406, 2017.
  3. [3]V. Banica. The nonlinear Schrödinger equation on hyperbolic space. Comm. Partial Differential Equations, 32(10–12):1643–1677, 2007.arxiv.org/abs/math/0406058
  4. [4]Matteo Bonforte, Jean Dolbeault, Bruno Nazaret, and Nikita Simonov. Stability in Gagliardo–Nirenberg–Sobolev inequalities: flows, regularity, and the entropy method. Mem. Amer. Math. Soc., 308(1554):viii+166, 2025.
  5. [5]Gabriele Bianchi and Henrik Egnell. A note on the Sobolev inequality. J. Funct. Anal., 100(1):18–24, 1991.DOI
  6. [6]Haïm Brézis and Elliott Lieb. A relation between pointwise convergence of functions and convergence of functionals. Proc. Amer. Math. Soc., 88(3):486–490, 1983.
  7. [7]Christophe Brouttelande. The best–constant problem for a family of Gagliardo–Nirenberg inequalities on a compact Riemannian manifold. Proc. Edinb. Math. Soc. (2), 46(1):117–146, 2003.DOI
  8. [8]Jurandir Ceccon and Carlos E. Durán. Sharp constants in Riemannian Lᵖ-Gagliardo–Nirenberg inequalities. J. Math. Anal. Appl., 433(1):260–281, 2016.DOI
  9. [9]Jurandir Ceccon, Carlos E. Durán, and Juliano D. B. de Godoi. First and second sharp constants in Riemannian Gagliardo–Nirenberg inequalities. Math. Nachr., 299(3):617–636, 2026.DOI
  10. [11]Eric A. Carlen and Alessio Figalli. Stability for a GNS inequality and the log–HLS inequality, with application to the critical mass Keller–Segel equation. Duke Math. J., 162(3):579–625, 2013.DOI
  11. [12]Jeff Cheeger. A lower bound for the smallest eigenvalue of the Laplacian. In Problems in analysis (Sympos. in honor of Salomon Bochner, Princeton Univ., Princeton, N.J., 1969), pages 195–199. Princeton Univ. Press, Princeton, NJ, 1970.
  12. [13]Liang Cheng. On the rigidity of manifolds with respect to the Gagliardo–Nirenberg inequalities. J. Math. Pures Appl. (9), 215:Paper No. 103960, 43, 2026.arxiv.org/abs/2507.05908
  13. [14]Hans Christianson, Jeremy Marzuola, Jason Metcalfe, and Michael Taylor. Nonlinear bound states on weakly homogeneous spaces. Comm. Partial Differential Equations, 39(1):34–97, 2014.arxiv.org/abs/1203.3612
  14. [15]Gilles Courtois. Spectrum of manifolds with holes. J. Funct. Anal., 134(1):194–221, 1995.DOI
  15. [16]Manuel Del Pino and Jean Dolbeault. Best constants for Gagliardo–Nirenberg inequalities and applications to nonlinear diffusions. J. Math. Pures Appl. (9), 81(9):847–875, 2002.DOI
  16. [17]Emilio Gagliardo. Proprietà di alcune classi di funzioni in più variabili. Ricerche Mat., 7:102–137, 1958.
  17. [18]Man Kam Kwong. Uniqueness of positive solutions of Δu − u + uᵖ = 0 in Rⁿ. Arch. Rational Mech. Anal., 105(3):243–266, 1989.DOI
  18. [19]P.-L. Lions. The concentration–compactness principle in the calculus of variations. The locally compact case. I. Ann. Inst. H. Poincaré Anal. Non Linéaire, 1(2):109–145, 1984.
  19. [20]P.-L. Lions. The concentration–compactness principle in the calculus of variations. The locally compact case. II. Ann. Inst. H. Poincaré Anal. Non Linéaire, 1(4):223–283, 1984.
  20. [21]Mayukh Mukherjee and Utsab Sarkar. Quantitative stability and strict inequalities for the Escobar functional. arXiv e-prints, October 2026. arXiv:2601.22665.
  21. [22]Mayukh Mukherjee. Extremal values of the (fractional) Weinstein functional on the hyperbolic space. Forum Math., 29(4):959–970, 2017.arxiv.org/abs/1406.4931
  22. [23]L. Nirenberg. On elliptic partial differential equations. Ann. Scuola Norm. Sup. Pisa Cl. Sci. (3), 13:115–162, 1959.
  23. [24]Walter A. Strauss. Existence of solitary waves in higher dimensions. Comm. Math. Phys., 55(2):149–162, 1977.DOI

Paper details

vlnc:W_n9erwrya
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