Introduction
Background and motivation
Assume that n ∈ N + n \in\mathbb{N}^{+} n ∈ N + , 1 ≤ p < n 1 \le p < n 1 ≤ p < n , and set p ∗ = n p n − p p^{*} = \frac{np}{n-p} p ∗ = n − p n p . The classical Sobolev inequality states that for any function u ∈ W 1 , p ( R n ) u \in W^{1,p}(\mathbb{R}^{n}) u ∈ W 1 , p ( R n ) ,
∥ u ∥ L p ∗ ( R n ) ≤ S n , p ∥ ∇ u ∥ L p ( R n ) , (1) \lVert u \rVert_{L^{p^{*}}(\mathbb{R}^{n})} \le S_{n,p}\lVert\nabla u \rVert_{L^{p}(\mathbb{R}^{n})}, \tag*{(1)} ∥ u ∥ L p ∗ ( R n ) ≤ S n , p ∥ ∇ u ∥ L p ( R n ) , (1)
where S n , p S_{n,p} S n , p denotes the optimal Sobolev constant. It follows from [4 , 52 ] that
S n , 1 = 1 n ( n ω n − 1 ) 1 n , S_{n,1} = \frac{1}{n}\left(\frac{n}{\omega_{n-1}}\right)^{\frac{1}{n}}, S n , 1 = n 1 ( ω n − 1 n ) n 1 ,
and for p > 1 p > 1 p > 1 ,
S n , p = 1 n ( n ( p − 1 ) n − p ) 1 − 1 p ( n ! Γ ( n p ) Γ ( n + 1 − n p ) ω n − 1 ) 1 n , S_{n,p} = \frac{1}{n}\left(\frac{n(p-1)}{n-p}\right)^{1-\frac{1}{p}}\left(\frac{n!}{\Gamma\left(\frac{n}{p}\right)\Gamma\left(n+1-\frac{n}{p}\right)\omega_{n-1}}\right)^{\frac{1}{n}}, S n , p = n 1 ( n − p n ( p − 1 ) ) 1 − p 1 Γ ( p n ) Γ ( n + 1 − p n ) ω n − 1 n ! n 1 ,
where ω n − 1 \omega_{n-1} ω n − 1 denotes the volume of S n − 1 \mathbb{S}^{n-1} S n − 1 and Γ ( s ) : = ∫ 0 ∞ t s − 1 e − t d t \Gamma(s) := \int_{0}^{\infty}t^{s-1}e^{-t}\,\mathrm{d}t Γ ( s ) := ∫ 0 ∞ t s − 1 e − t d t for s > 0 s > 0 s > 0 . The Moser–Trudinger inequality, which can be viewed as the endpoint case of (1) , states that, for any bounded open set Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n with n ≥ 2 n \ge2 n ≥ 2 and any function u ∈ W 0 1 , n ( Ω ) ∖ { 0 } u \in W^{1,n}_{0}(\Omega) \setminus\{0\} u ∈ W 0 1 , n ( Ω ) ∖ { 0 } ,
∫ Ω exp ( a n ∣ u ∣ n n − 1 ∥ ∇ u ∥ L n ( Ω ) n n − 1 ) d x ≤ C ( n ) ∣ Ω ∣ . (2) \int_{\Omega} \exp\left(a_n\frac{|u|^{\frac{n}{n-1}}}{\|\nabla u\|_{L^n(\Omega)}^{\frac{n}{n-1}}}\right)\,dx \le C(n)|\Omega|. \tag*{(2)} ∫ Ω exp a n ∥∇ u ∥ L n ( Ω ) n − 1 n ∣ u ∣ n − 1 n d x ≤ C ( n ) ∣Ω∣. (2)
Here a n = n ω n − 1 1 n − 1 a_n = n\omega_{n-1}^{\frac{1}{n-1}} a n = n ω n − 1 n − 1 1 is the optimal constant; see [43 , 54 ] . In [2 ] , Adams generalized (2) to higher-order Sobolev spaces. Assume that s , n ∈ N + s,n \in\mathbb{N}_{+} s , n ∈ N + and 1 ≤ s < n 1 \le s < n 1 ≤ s < n . For any bounded open set Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n and any function u ∈ W 0 s , n s ( Ω ) ∖ { 0 } u \in W^{s,\frac{n}{s}}_{0}(\Omega) \setminus\{0\} u ∈ W 0 s , s n ( Ω ) ∖ { 0 } ,
∫ Ω exp ( a s , n ∣ u ∣ n n − s ∥ Δ s 2 u ∥ L n s ( Ω ) n n − s ) d x ≤ C ( n , s ) ∣ Ω ∣ . (3) \int_{\Omega} \exp\left(a_{s,n}\frac{|u|^{\frac{n}{n-s}}}{\|\Delta^{\frac{s}{2}}u\|_{L^{\frac{n}{s}}(\Omega)}^{\frac{n}{n-s}}}\right)\,dx \le C(n,s)|\Omega|. \tag*{(3)} ∫ Ω exp a s , n ∥ Δ 2 s u ∥ L s n ( Ω ) n − s n ∣ u ∣ n − s n d x ≤ C ( n , s ) ∣Ω∣. (3)
The optimal constant a s , n a_{s,n} a s , n is given by
a s , n = n ω n − 1 ( π n 2 2 s Γ ( s + 1 2 ) Γ ( n 2 − ⌊ s 2 ⌋ ) ) n n − s . a_{s,n}=\frac{n}{\omega_{n-1}}\left(\frac{\pi^{\frac{n}{2}}2^{s}\Gamma\left(\frac{s+1}{2}\right)}{\Gamma\left(\frac{n}{2}-\left\lfloor\frac{s}{2}\right\rfloor\right)}\right)^{\frac{n}{n-s}}. a s , n = ω n − 1 n ( Γ ( 2 n − ⌊ 2 s ⌋ ) π 2 n 2 s Γ ( 2 s + 1 ) ) n − s n .
When s s s is odd, Δ s 2 u \Delta^{\frac{s}{2}}u Δ 2 s u should be understood as ∇ Δ s − 1 2 u \nabla\Delta^{\frac{s-1}{2}}u ∇ Δ 2 s − 1 u .
In this paper, we are concerned with generalizations of (1) , (2) , and (3) to general smooth, closed, connected manifolds. Let ( M , g ) (M,g) ( M , g ) be a smooth closed manifold of dimension n ≥ 2 n \ge2 n ≥ 2 . For 1 ≤ p < n 1 \le p < n 1 ≤ p < n , the Sobolev inequality for functions u ∈ W 1 , p ( M ) u \in W^{1,p}(M) u ∈ W 1 , p ( M ) takes the form
∥ u ∥ L p ∗ ( M ) ≤ A ∥ ∇ u ∥ L p ( M ) + B ∥ u ∥ L p ( M ) . (4) \|u\|_{L^{p^*}(M)} \le A\|\nabla u\|_{L^p(M)}+B\|u\|_{L^p(M)}. \tag*{(4)} ∥ u ∥ L p ∗ ( M ) ≤ A ∥∇ u ∥ L p ( M ) + B ∥ u ∥ L p ( M ) . (4)
Unlike the Euclidean case (1) , it is easy to see that the term ∥ u ∥ L p ( M ) \|u\|_{L^p(M)} ∥ u ∥ L p ( M ) on the right-hand side of (4) cannot be removed. The problem of optimizing the coefficients A A A and B B B in (4) is known as the A B AB A B -program, and a detailed introduction can be found in the monographs [6 , 31 , 32 ] .
In the pioneering work [4 ] , Aubin showed that for any ε > 0 \varepsilon> 0 ε > 0 there exists a positive constant B ( M , p , ε ) > 0 B(M,p,\varepsilon)>0 B ( M , p , ε ) > 0 such that
∥ u ∥ L p ∗ ( M ) p ≤ ( S n , p p + ε ) ∥ ∇ u ∥ L p ( M ) p + B ( M , p , ε ) ∥ u ∥ L p ( M ) p . (5) \|u\|_{L^{p^*}(M)}^p \le(S_{n,p}^p+\varepsilon)\|\nabla u\|_{L^p(M)}^p+B(M,p,\varepsilon)\|u\|_{L^p(M)}^p. \tag*{(5)} ∥ u ∥ L p ∗ ( M ) p ≤ ( S n , p p + ε ) ∥∇ u ∥ L p ( M ) p + B ( M , p , ε ) ∥ u ∥ L p ( M ) p . (5)
Moreover, this inequality is almost sharp in the sense that the coefficient S n , p p + ε S_{n,p}^p+\varepsilon S n , p p + ε cannot be replaced by any number smaller than S n , p p S_{n,p}^p S n , p p . It is a delicate issue whether one can take ε = 0 \varepsilon=0 ε = 0 in (5) and obtain the following sharp inequality :
∥ u ∥ L p ∗ ( M ) p ≤ S n , p p ∥ ∇ u ∥ L p ( M ) p + B ( M , p ) ∥ u ∥ L p ( M ) p . (6) \|u\|_{L^{p^*}(M)}^p \le S_{n,p}^p\|\nabla u\|_{L^p(M)}^p+B(M,p)\|u\|_{L^p(M)}^p. \tag*{(6)} ∥ u ∥ L p ∗ ( M ) p ≤ S n , p p ∥∇ u ∥ L p ( M ) p + B ( M , p ) ∥ u ∥ L p ( M ) p . (6)
On S n \mathbb{S}^{n} S n , it was shown in [4 ] (see also [33 ] ) that (6) holds for every 1 ≤ p ≤ 2 1 \le p \le2 1 ≤ p ≤ 2 . However, it follows from [19 ] that (6) cannot hold for 2 < p < n 2 < p < \sqrt{n} 2 < p < n . We refer interested readers to [32 Section 4] for further related results.
For the higher-order Moser–Trudinger inequality, it follows from [9 , 21 ] that the following sharp inequality
∫ M exp ( a s , n ∣ u ∣ n n − s ∥ Δ s 2 u ∥ L s n s ( M ) n n − s ) d μ ≤ c ( M , s ) (7) \int_M \exp\left(a_{s,n}\frac{|u|^{\frac{n}{n-s}}}{\left\lVert\Delta^{\frac{s}{2}}u\right\rVert_{L_s^{\frac{n}{s}}(M)}^{\frac{n}{n-s}}}\right)\,\mathrm{d}\mu\leq c(M,s) \tag*{(7)} ∫ M exp a s , n Δ 2 s u L s s n ( M ) n − s n ∣ u ∣ n − s n d μ ≤ c ( M , s ) (7)
holds for any u ∈ W s , n s ( M ) ∖ { 0 } u \in W^{s,\frac{n}{s}}(M)\setminus\{0\} u ∈ W s , s n ( M ) ∖ { 0 } satisfying ∫ M u d μ = 0 \int_M u\,\mathrm{d}\mu=0 ∫ M u d μ = 0 . Here μ \mu μ is the measure associated with the metric g g g . Combining (7) with Young’s inequality, we obtain the higher-order Moser–Trudinger–Onofri inequality (see [29 , 43 , 45 ] )
log ∫ M e n u d μ ≤ α s , n ∥ Δ s 2 u ∥ L s n s ( M ) n s + C ( M , s ) . (8) \log\int_M e^{nu}\,\mathrm{d}\mu\leq\alpha_{s,n}\left\lVert\Delta^{\frac{s}{2}}u\right\rVert_{L_s^{\frac{n}{s}}(M)}^{\frac{n}{s}}+C(M,s). \tag*{(8)} log ∫ M e n u d μ ≤ α s , n Δ 2 s u L s s n ( M ) s n + C ( M , s ) . (8)
Here
α s , n = s ( ( n − s ) ω n − 1 n ) n − s s ( Γ ( n 2 − ⌊ s 2 ⌋ ) π n 2 2 s Γ ( s + 1 2 ) ) n s \alpha_{s,n}=s\left(\frac{(n-s)\omega_{n-1}}{n}\right)^{\frac{n-s}{s}}\left(\frac{\Gamma\left(\frac{n}{2}-\left\lfloor\frac{s}{2}\right\rfloor\right)}{\pi^{\frac{n}{2}}2^s\Gamma\left(\frac{s+1}{2}\right)}\right)^{\frac{n}{s}} α s , n = s ( n ( n − s ) ω n − 1 ) s n − s ( π 2 n 2 s Γ ( 2 s + 1 ) Γ ( 2 n − ⌊ 2 s ⌋ ) ) s n
is the optimal constant. A careful analysis of the constants appearing in (7) and (8) can be found in [58 ] .
In the seminal work [5 ] , Aubin showed that by imposing vanishing first-order moment constraints, one can improve the constants in both the Sobolev inequality and the first-order Moser–Trudinger–Onofri inequality on S n S^n S n . More precisely, for any ε > 0 \varepsilon>0 ε > 0 and any function u ∈ W 1 , p ( S n ) u\in W^{1,p}(S^n) u ∈ W 1 , p ( S n ) satisfying
∫ S n x i ∣ u ∣ p ∗ d μ = 0 , 1 ≤ i ≤ n + 1 , \int_{S^n}x_i|u|^{p^*}\,\mathrm{d}\mu=0,\qquad1\leq i\leq n+1, ∫ S n x i ∣ u ∣ p ∗ d μ = 0 , 1 ≤ i ≤ n + 1 ,
there exists a positive constant B ( n , p , ε ) B(n,p,\varepsilon) B ( n , p , ε ) such that
∥ u ∥ L p ∗ ( S n ) p ≤ ( 2 − p n S n , p p + ε ) ∥ ∇ u ∥ L p ( S n ) p + B ( n , p , ε ) ∥ u ∥ L p ( S n ) p . (9) \left\lVert u\right\rVert_{L^{p^*}(S^n)}^p\leq\left(2^{-\frac{p}{n}}S_{n,p}^p+\varepsilon\right)\left\lVert\nabla u\right\rVert_{L^p(S^n)}^p+B(n,p,\varepsilon)\left\lVert u\right\rVert_{L^p(S^n)}^p. \tag*{(9)} ∥ u ∥ L p ∗ ( S n ) p ≤ ( 2 − n p S n , p p + ε ) ∥ ∇ u ∥ L p ( S n ) p + B ( n , p , ε ) ∥ u ∥ L p ( S n ) p . (9)
Similarly, for any ε > 0 \varepsilon>0 ε > 0 and any function u ∈ W 1 , n ( S n ) u\in W^{1,n}(S^n) u ∈ W 1 , n ( S n ) satisfying
∫ S n x i e n u d μ = 0 , 1 ≤ i ≤ n + 1 , \int_{S^n}x_i e^{nu}\,\mathrm{d}\mu=0,\qquad1\leq i\leq n+1, ∫ S n x i e n u d μ = 0 , 1 ≤ i ≤ n + 1 ,
there exists a positive constant C ( n , ε ) C(n,\varepsilon) C ( n , ε ) such that
log ∫ S n e n u d μ ≤ ( α 1 , n 2 + ε ) ∥ ∇ u ∥ L n ( S n ) n + C ( n , ε ) . (10) \log\int_{S^n}e^{nu}\,\mathrm{d}\mu\leq\left(\frac{\alpha_{1,n}}{2}+\varepsilon\right)\left\lVert\nabla u\right\rVert_{L^n(S^n)}^n+C(n,\varepsilon). \tag*{(10)} log ∫ S n e n u d μ ≤ ( 2 α 1 , n + ε ) ∥ ∇ u ∥ L n ( S n ) n + C ( n , ε ) . (10)
Moreover, the inequalities (9) and (10) are almost sharp in the same sense as (5) . These inequalities play an important role in the analysis of prescribing scalar curvature problems on S n S^n S n ; see [5 , 6 , 11 , 12 , 32 ] and the references therein for further discussions.
Recently, generalizations of (9) and (10) incorporating higher-order moment constraints and higher-order Moser–Trudinger–Onofri inequalities have been established in a series of works [10 , 29 , 30 ] . To state these results, we first introduce some notation.
For m , n ∈ N + m,n \in\mathbb{N}_{+} m , n ∈ N + , we set
P m , n = { all polynomials on R n + 1 of degree at most m } (11) \mathcal{P}_{m,n}=\{ \text{all polynomials on } \mathbb{R}^{n+1} \text{ of degree at most }m\} \tag*{(11)} P m , n = { all polynomials on R n + 1 of degree at most m } (11)
and
P ˚ m , n = { f ∈ P m , n : ∫ S n f d μ = 0 } . (12) \mathring{\mathcal{P}}_{m,n}=\left\{f\in\mathcal{P}_{m,n}:\int_{\mathbb{S}^{n}}f\,\mathrm{d}\mu=0\right\}. \tag*{(12)} P ˚ m , n = { f ∈ P m , n : ∫ S n f d μ = 0 } . (12)
Let Δ ( S n ) \Delta(\mathbb{S}^{n}) Δ ( S n ) denote the space of probability measures on S n \mathbb{S}^{n} S n supported on countably many points. For ν = ∑ i ν i δ ξ i ∈ Δ ( S n ) \nu=\sum_i\nu_i\delta_{\xi_i}\in\Delta(\mathbb{S}^{n}) ν = ∑ i ν i δ ξ i ∈ Δ ( S n ) and 0 ≤ θ < 1 0\le\theta<1 0 ≤ θ < 1 , we define ∥ ν ∥ θ : = ∑ i ν i θ \|\nu\|_{\theta}:=\sum_i\nu_i^{\theta} ∥ ν ∥ θ := ∑ i ν i θ .
The following two notions, which originate from numerical analysis and scientific computing (see, for instance, [16 , 17 , 46 ] ), are useful for formulating improved inequalities under vanishing higher-order moment constraints:
M m , n c = { ν ∈ Δ ( S n ) : ∫ S n f d ν = 0 for all f ∈ P ˚ m , n } , (13) \mathcal{M}^{c}_{m,n}=\left\{\nu\in\Delta(\mathbb{S}^{n}):\int_{\mathbb{S}^{n}}f\,\mathrm{d}\nu=0\text{ for all }f\in\mathring{\mathcal{P}}_{m,n}\right\}, \tag*{(13)} M m , n c = { ν ∈ Δ ( S n ) : ∫ S n f d ν = 0 for all f ∈ P ˚ m , n } , (13)
Θ ( m , θ , n ) = inf { ∥ ν ∥ θ : ν ∈ M m , n c } . (14) \Theta(m,\theta,n)=\inf\left\{\|\nu\|_{\theta}:\nu\in\mathcal{M}^{c}_{m,n}\right\}. \tag*{(14)} Θ ( m , θ , n ) = inf { ∥ ν ∥ θ : ν ∈ M m , n c } . (14)
The elements of M m , n c \mathcal{M}^{c}_{m,n} M m , n c are usually referred to as cubature formulas of strength m m m . Estimates for Θ ( m , θ , n ) \Theta(m,\theta,n) Θ ( m , θ , n ) , particularly in the case θ = 0 \theta=0 θ = 0 , have been widely studied. It is known that Θ ( 1 , θ , n ) = 2 1 − θ \Theta(1,\theta,n)=2^{1-\theta} Θ ( 1 , θ , n ) = 2 1 − θ , Θ ( 2 , θ , n ) = ( n + 2 ) 1 − θ \Theta(2,\theta,n)=(n+2)^{1-\theta} Θ ( 2 , θ , n ) = ( n + 2 ) 1 − θ , and Θ ( 3 , θ , n ) = ( 2 n + 2 ) 1 − θ \Theta(3,\theta,n)=(2n+2)^{1-\theta} Θ ( 3 , θ , n ) = ( 2 n + 2 ) 1 − θ . We refer interested readers to [16 , 17 , 30 , 46 ] and the references therein for a more detailed discussion of cubature formulas, spherical designs, and estimates for Θ \Theta Θ .
We now state the almost sharp improved Sobolev and Moser–Trudinger–Onofri inequalities under vanishing higher-order moment constraints. In [30 Theorem 1.1] , Hang and Wang showed that for m , n ∈ N + m,n\in\mathbb{N}_{+} m , n ∈ N + , 1 < p < n 1<p<n 1 < p < n , ε > 0 \varepsilon>0 ε > 0 , and any function u ∈ W 1 , p ( S n ) u\in W^{1,p}(\mathbb{S}^{n}) u ∈ W 1 , p ( S n ) satisfying
∫ S n f ∣ u ∣ p ∗ d μ = 0 for all f ∈ P ˚ m , n , (15) \int_{\mathbb{S}^{n}}f|u|^{p^{*}}\,\mathrm{d}\mu=0\quad\text{for all }f\in\mathring{\mathcal{P}}_{m,n}, \tag*{(15)} ∫ S n f ∣ u ∣ p ∗ d μ = 0 for all f ∈ P ˚ m , n , (15)
the following almost sharp improved Sobolev inequality holds:
∥ u ∥ L p ∗ ( S n ) p ≤ ( S n , p p Θ ( m , n n − p , n ) + ε ) ∥ ∇ u ∥ L p ( S n ) p + B ( m , n , p , ε ) ∥ u ∥ L p ( S n ) p , (16) \|u\|_{L^{p^{*}}(\mathbb{S}^{n})}^{p}\leq\left(\frac{S_{n,p}^{p}}{\Theta\left(m,\frac{n}{n-p},n\right)}+\varepsilon\right)\|\nabla u\|_{L^{p}(\mathbb{S}^{n})}^{p}+B(m,n,p,\varepsilon)\|u\|_{L^{p}(\mathbb{S}^{n})}^{p}, \tag*{(16)} ∥ u ∥ L p ∗ ( S n ) p ≤ Θ ( m , n − p n , n ) S n , p p + ε ∥∇ u ∥ L p ( S n ) p + B ( m , n , p , ε ) ∥ u ∥ L p ( S n ) p , (16)
for some positive constant B ( m , n , p , ε ) B(m,n,p,\varepsilon) B ( m , n , p , ε ) .
For the higher-order Moser–Trudinger–Onofri inequality, it follows from Chang–Hang [10 Theorem 1.2] and Hang [29 Theorems 3.2, 4.5, 4.11] that for s , m , n ∈ N + s,m,n\in\mathbb{N}_{+} s , m , n ∈ N + , 1 ≤ s < n 1\leq s<n 1 ≤ s < n , ε > 0 \varepsilon>0 ε > 0 , and any function u ∈ W s , n s ( S n ) u\in W^{s,\frac{n}{s}}(\mathbb{S}^{n}) u ∈ W s , s n ( S n ) satisfying ∫ S n u d μ = 0 \int_{\mathbb{S}^{n}}u\,\mathrm{d}\mu=0 ∫ S n u d μ = 0 and
∫ S n f e n u d μ = 0 for all f ∈ P ˚ m , n , (17) \int_{\mathbb{S}^{n}}f e^{nu}\,\mathrm{d}\mu=0\quad\text{for all }f\in\mathring{\mathcal{P}}_{m,n}, \tag*{(17)} ∫ S n f e n u d μ = 0 for all f ∈ P ˚ m , n , (17)
the following almost sharp improved inequality holds:
log ∫ S n e n u d μ ≤ ( α s , n Θ ( m , 0 , n ) + ε ) ∥ Δ s 2 u ∥ L n s ( S n ) n s + C ( m , n , s , ε ) , (18) \log\int_{\mathbb{S}^{n}} e^{nu}\,\mathrm{d}\mu\le\left(\frac{\alpha_{s,n}}{\Theta(m,0,n)}+\varepsilon\right)\left\|\Delta^{\frac{s}{2}}u\right\|_{L^{\frac{n}{s}}(\mathbb{S}^{n})}^{\frac{n}{s}}+C(m,n,s,\varepsilon), \tag*{(18)} log ∫ S n e n u d μ ≤ ( Θ ( m , 0 , n ) α s , n + ε ) Δ 2 s u L s n ( S n ) s n + C ( m , n , s , ε ) , (18)
for some positive constant C ( m , n , s , ε ) C(m,n,s,\varepsilon) C ( m , n , s , ε ) . Since Θ ( 1 , θ , n ) = 2 1 − θ \Theta(1,\theta,n)=2^{1-\theta} Θ ( 1 , θ , n ) = 2 1 − θ for any 0 ≤ θ < 1 0\le\theta<1 0 ≤ θ < 1 , one can readily recover (9) and (10) by taking m = 1 m=1 m = 1 in (16) and (18) , respectively.
Motivated by the A B AB A B -program for the inequality (4) , it is natural to ask whether one can take ε = 0 \varepsilon=0 ε = 0 in (16) and (18) , and what can be said about the remainders B m B_m B m and C m C_m C m . The study of these questions is not only important from the perspective of functional inequalities, but also closely related to various prescribing curvature problems on S n \mathbb{S}^{n} S n . So far, however, only limited progress has been made. The only completely solved case is due to Gui and Moradifam [28 ] . In [28 ] , they applied the sphere covering inequality to give a complete analysis of (18) in the case m = s = 1 m=s=1 m = s = 1 and n = 2 n=2 n = 2 , thereby confirming a conjecture proposed in [11 ] . More recently, Gui, Li, Wei, and Ye [26 ] generalized the results in [28 ] to Beckner’s inequality involving the GJMS operator (see also [23 , 24 , 25 , 27 , 39 ] for partial results). Since the GJMS energy is strictly larger than the fractional Sobolev energy when n ≥ 3 n\ge3 n ≥ 3 , their results do not imply that one can take ε = 0 \varepsilon=0 ε = 0 in (18) when s = n 2 s=\frac{n}{2} s = 2 n .
It is also worth mentioning that the recent proofs of (16) and (18) in [10 , 29 , 30 ] are based on careful use of concentration–compactness principles (see, for instance, [29 , 41 , 42 ] ). These compactness-based arguments not only allow one to handle vanishing higher-order moment constraints and yield concise and transparent proofs, but are also highly flexible and can be adapted to a wide range of inequalities on general compact manifolds (see, e.g., [1 , 14 , 15 , 35 , 38 , 50 , 56 , 57 ] ). However, the reliance on concentration–compactness principles precludes any quantitative information on the remaining constants B m B_m B m and C m C_m C m appearing in (16) and (18) , let alone the possibility of determining whether the parameter ε \varepsilon ε can be taken to be zero.
In this paper, we make further progress on these problems. By developing a new approach to (16) and (18) , we establish sharpness results in several cases. We also provide non-sharp examples through a general failure principle. Moreover, we obtain quantitative estimates for the remainders in general cases, yielding sharp inequalities at different scales. Our approach also applies to improved inequalities under symmetry constraints.
Statements of main results
Let us summarize our main results. Our first result implies that one can take ε = 0 \varepsilon=0 ε = 0 in (16) when m ∈ { 1 , 2 , 3 } m\in\{1,2,3\} m ∈ { 1 , 2 , 3 } , p = 2 p=2 p = 2 , and n ≥ 3 n\ge3 n ≥ 3 .
Theorem 1.1. Assume m ∈ { 1 , 2 , 3 } m\in\{1,2,3\} m ∈ { 1 , 2 , 3 } , p = 2 p=2 p = 2 , and n ≥ 3 n\ge3 n ≥ 3 . Then there is a constant B = B ( m , n ) < ∞ B=B(m,n)<\infty B = B ( m , n ) < ∞ such that every u ∈ H 1 ( S n ) u\in H^{1}(\mathbb{S}^{n}) u ∈ H 1 ( S n ) satisfying (15) also satisfies
∥ u ∥ L 2 ∗ ( S n ) 2 ≤ S n , 2 2 Θ ( m , n − 2 n , n ) ∥ ∇ u ∥ L 2 ( S n ) 2 + B ∥ u ∥ L 2 ( S n ) 2 . (19) \|u\|_{L^{2^{*}}(\mathbb{S}^{n})}^{2}\le\frac{S_{n,2}^{2}}{\Theta\left(m,\frac{n-2}{n},n\right)}\|\nabla u\|_{L^{2}(\mathbb{S}^{n})}^{2}+B\|u\|_{L^{2}(\mathbb{S}^{n})}^{2}. \tag*{(19)} ∥ u ∥ L 2 ∗ ( S n ) 2 ≤ Θ ( m , n n − 2 , n ) S n , 2 2 ∥∇ u ∥ L 2 ( S n ) 2 + B ∥ u ∥ L 2 ( S n ) 2 . (19)
To our knowledge, (19) gives the first sharp improved Sobolev inequality under moment constraints. The restriction m ∈ { 1 , 2 , 3 } m \in\{1,2,3\} m ∈ { 1 , 2 , 3 } arises from a technical issue: our proof relies on properties of extremal measures attaining the value Θ \Theta Θ when m ∈ { 1 , 2 , 3 } m \in\{1,2,3\} m ∈ { 1 , 2 , 3 } (see Remark 2.5 for details).
Our second result provides a general failure principle: if the sharp Sobolev inequality (6) fails, then the corresponding sharp improved inequality also fails.
Theorem 1.2. Assume 1 ≤ p < n 1 \le p < n 1 ≤ p < n and the sharp Sobolev inequality (6) fails on S n \mathbb{S}^n S n . Then for any m ∈ N + m \in\mathbb{N}_+ m ∈ N + , there does not exist a finite constant B = B ( m , n , p ) B = B(m,n,p) B = B ( m , n , p ) such that
∥ u ∥ L p ∗ ( S n ) p ≤ S n , p p Θ ( m , n − p n , n ) ∥ ∇ u ∥ L p ( S n ) p + B ∥ u ∥ L p ( S n ) p (20) \lVert u\rVert_{L^{p^*}(\mathbb{S}^n)}^p \le\frac{S_{n,p}^p}{\Theta\left(m,\frac{n-p}{n},n\right)}\lVert\nabla u\rVert_{L^p(\mathbb{S}^n)}^p + B\lVert u\rVert_{L^p(\mathbb{S}^n)}^p \tag*{(20)} ∥ u ∥ L p ∗ ( S n ) p ≤ Θ ( m , n n − p , n ) S n , p p ∥ ∇ u ∥ L p ( S n ) p + B ∥ u ∥ L p ( S n ) p (20) holds for every u ∈ W 1 , p ( S n ) u \in W^{1,p}(\mathbb{S}^n) u ∈ W 1 , p ( S n ) satisfying (15) .
By [19 Theorem 1] , we have the following direct corollary.
Corollary 1.3. Assume 2 < p < n 2 < p < \sqrt{n} 2 < p < n . Then for any m ∈ N + m \in\mathbb{N}_+ m ∈ N + , there does not exist a finite constant B = B ( m , n , p ) B = B(m,n,p) B = B ( m , n , p ) such that (20) holds for every u ∈ W 1 , p ( S n ) u \in W^{1,p}(\mathbb{S}^n) u ∈ W 1 , p ( S n ) satisfying (15) .
To our knowledge, these two results provide the first non-sharp cases of improved Sobolev inequalities under moment constraints.
Our third result implies that one can take ε = 0 \varepsilon= 0 ε = 0 in (18) for s = n 2 s = \frac{n}{2} s = 2 n , n ≥ 2 n \ge2 n ≥ 2 , and every m ∈ N + m \in\mathbb{N}_+ m ∈ N + . In particular, we allow s ∉ N + s \notin\mathbb{N}_+ s ∈ / N + when n n n is odd.
Theorem 1.4. Assume m , n ∈ N + m,n \in\mathbb{N}_+ m , n ∈ N + , n ≥ 2 n \ge2 n ≥ 2 , and s = n 2 s = \frac{n}{2} s = 2 n . Then there is a constant C = C ( m , n ) < ∞ C = C(m,n) < \infty C = C ( m , n ) < ∞ such that every u ∈ H s ( S n ) u \in H^s(\mathbb{S}^n) u ∈ H s ( S n ) satisfying ∫ S n u d μ = 0 \int_{\mathbb{S}^n} u \,\mathrm{d}\mu= 0 ∫ S n u d μ = 0 and (17) also satisfies
log ( ω n − 1 ∫ S n e n u d μ ) ≤ n 2 ( n − 1 ) ! ω n Θ ( m , 0 , n ) ∥ ( − Δ ) n 4 u ∥ L 2 ( S n ) 2 + C , (21) \log\left(\omega_n^{-1}\int_{\mathbb{S}^n} e^{nu}\,\mathrm{d}\mu\right) \le\frac{n}{2(n-1)!\omega_n\Theta(m,0,n)}\left\lVert(-\Delta)^{\frac{n}{4}}u\right\rVert_{L^2(\mathbb{S}^n)}^2 + C, \tag*{(21)} log ( ω n − 1 ∫ S n e n u d μ ) ≤ 2 ( n − 1 )! ω n Θ ( m , 0 , n ) n ( − Δ ) 4 n u L 2 ( S n ) 2 + C , (21)
To our knowledge, (21) gives the first sharp improved Moser–Trudinger–Onofri inequalities since the work of Gui and Moradifam [28 ] , which treats the special case s = 1 s = 1 s = 1 , n = 2 n = 2 n = 2 , m = 1 m = 1 m = 1 .
Our next result concerns the best remainder in (21) , which we denote by C b e s t ( m , n ) C_{\mathrm{best}}(m,n) C best ( m , n ) . It follows from [28 ] that C b e s t ( 1 , 2 ) = 0 C_{\mathrm{best}}(1,2) = 0 C best ( 1 , 2 ) = 0 and that all extremal functions are constant. In recent work [26 ] , Gui, Li, Wei, and Ye investigated a higher-order Beckner’s inequality. They showed that, for the sharp improved Beckner’s inequality, the remainder is 0 and all extremal functions are constant. Based on these results, one might expect that C b e s t ( m , n ) = 0 C_{\mathrm{best}}(m,n) = 0 C best ( m , n ) = 0 for general m , n m,n m , n . However, we show that this is false, at least for large n n n .
Theorem 1.5. C b e s t ( m , n ) > 0 C_{\mathrm{best}}(m,n) > 0 C best ( m , n ) > 0 in the following four cases: m = 1 m = 1 m = 1 , n ≥ 5 n \ge5 n ≥ 5 ; m = 2 m = 2 m = 2 , n ≥ 12 n \ge12 n ≥ 12 ; m = 3 m = 3 m = 3 , n ≥ 18 n \ge18 n ≥ 18 ; m ≥ 4 m \ge4 m ≥ 4 , n ≥ 18 ( m + 1 ) n \ge18(m+1) n ≥ 18 ( m + 1 ) .
We remark that for m ≥ 2 m \ge2 m ≥ 2 , our lower bound on n n n is rough. For m = 1 m = 1 m = 1 , we devote particular attention to obtaining the lower bound n ≥ 5 n \ge5 n ≥ 5 , since C b e s t ( 1 , 2 ) = 0 C_{\mathrm{best}}(1,2) = 0 C best ( 1 , 2 ) = 0 is already known. We leave open the interesting question of whether C b e s t ( 1 , 3 ) C_{\mathrm{best}}(1,3) C best ( 1 , 3 ) and C b e s t ( 1 , 4 ) C_{\mathrm{best}}(1,4) C best ( 1 , 4 ) vanish.
The following two theorems provide quantitative remainders for (16) and (18) , respectively, over the full range of parameters.
Theorem 1.6. Assume m , n ∈ N + m,n \in\mathbb{N}_{+} m , n ∈ N + , 1 ≤ p < n 1 \le p < n 1 ≤ p < n , and 0 < ε < 1 0 < \varepsilon< 1 0 < ε < 1 . Assume further that the sharp Sobolev inequality (6) holds on S n \mathbb{S}^{n} S n . Then there exists a positive constant B = B ( m , n , p ) B=B(m,n,p) B = B ( m , n , p ) such that, for every u ∈ W 1 , p ( S n ) u \in W^{1,p}(\mathbb{S}^{n}) u ∈ W 1 , p ( S n ) satisfying (15) ,
∥ u ∥ L p ∗ ( S n ) p ≤ ( S n , p p Θ ( m , n − p n , n ) + ε ) ∥ ∇ u ∥ L p ( S n ) p + B ε − γ ( p ) ∥ u ∥ L p ( S n ) p . \lVert u\rVert_{L^{p^{*}}(\mathbb{S}^{n})}^{p} \le\left(\frac{S_{n,p}^{p}}{\Theta\left(m,\frac{n-p}{n},n\right)}+\varepsilon\right)\lVert\nabla u\rVert_{L^{p}(\mathbb{S}^{n})}^{p}+B\varepsilon^{-\gamma(p)}\lVert u\rVert_{L^{p}(\mathbb{S}^{n})}^{p}. ∥ u ∥ L p ∗ ( S n ) p ≤ ( Θ ( m , n n − p , n ) S n , p p + ε ) ∥ ∇ u ∥ L p ( S n ) p + B ε − γ ( p ) ∥ u ∥ L p ( S n ) p . where
γ ( p ) = { p 2 2 ( n − p ) , 1 ≤ p ≤ 2 , n p 2 ( n − p ) − 1 , 2 < p < n . (22) \gamma(p)=
\begin{cases}
\frac{p^{2}}{2(n-p)}, & 1\le p\le2,\\
\frac{np}{2(n-p)}-1, & 2<p<n.
\end{cases} \tag*{(22)} γ ( p ) = { 2 ( n − p ) p 2 , 2 ( n − p ) n p − 1 , 1 ≤ p ≤ 2 , 2 < p < n . (22)
Theorem 1.7. Assume s , m , n ∈ N + s,m,n \in\mathbb{N}_{+} s , m , n ∈ N + , 1 ≤ s < n 1 \le s < n 1 ≤ s < n , and 0 < ε < 1 0 < \varepsilon< 1 0 < ε < 1 . Then there exists a positive constant C = C ( m , n , s ) C=C(m,n,s) C = C ( m , n , s ) such that every function u ∈ W s , n s ( S n ) u \in W^{s,\frac{n}{s}}(\mathbb{S}^{n}) u ∈ W s , s n ( S n ) satisfying ∫ S n u d μ = 0 \int_{\mathbb{S}^{n}}u\,\mathrm{d}\mu=0 ∫ S n u d μ = 0 and (17) also satisfies
log ∫ S n e n u d μ ≤ ( α s , n Θ ( m , 0 , n ) + ε ) ∥ Δ s 2 u ∥ L n s ( S n ) s n + C ε − s n − s . \log\int_{\mathbb{S}^{n}}e^{nu}\,\mathrm{d}\mu\le\left(\frac{\alpha_{s,n}}{\Theta(m,0,n)}+\varepsilon\right)\lVert\Delta^{\frac{s}{2}}u\rVert_{L^{\frac{n}{s}}(\mathbb{S}^{n})}^{\frac{s}{n}}+C\varepsilon^{-\frac{s}{n-s}}. log ∫ S n e n u d μ ≤ ( Θ ( m , 0 , n ) α s , n + ε ) ∥ Δ 2 s u ∥ L s n ( S n ) n s + C ε − n − s s .
We remark that the assumption that (6) holds in Theorem 1.6 is made only for simplicity: when it does not hold, the remainder involves the term B ( m , n , p , ε ) B(m,n,p,\varepsilon) B ( m , n , p , ε ) in (16) . Moreover, we prove stronger results than the two theorems above (see Theorems 5.1 and 6.2 ), in which the orthogonality conditions are replaced by almost-orthogonality conditions.
As a direct consequence of the two theorems above, we obtain sharp improved inequalities at different scales.
Corollary 1.8. Under the same assumptions as in Theorem 1.6 , there exists a positive constant B = B ( m , n , p ) B=B(m,n,p) B = B ( m , n , p ) such that
∥ u ∥ L p ∗ ( S n ) γ ‾ ( p ) ≤ ( S n , p Θ ( m , n − p n , n ) 1 p ) γ ‾ ( p ) ∥ ∇ u ∥ L p ( S n ) γ ‾ ( p ) + B ∥ u ∥ L p ( S n ) γ ‾ ( p ) . (23) \lVert u\rVert_{L^{p^{*}}(\mathbb{S}^{n})}^{\overline{\gamma}(p)} \le\left(\frac{S_{n,p}}{\Theta\left(m,\frac{n-p}{n},n\right)^{\frac{1}{p}}}\right)^{\overline{\gamma}(p)}\lVert\nabla u\rVert_{L^{p}(\mathbb{S}^{n})}^{\overline{\gamma}(p)}+B\lVert u\rVert_{L^{p}(\mathbb{S}^{n})}^{\overline{\gamma}(p)}. \tag*{(23)} ∥ u ∥ L p ∗ ( S n ) γ ( p ) ≤ Θ ( m , n n − p , n ) p 1 S n , p γ ( p ) ∥ ∇ u ∥ L p ( S n ) γ ( p ) + B ∥ u ∥ L p ( S n ) γ ( p ) . (23) where
γ ‾ ( p ) = p 1 + γ ( p ) = { 2 p ( n − p ) 2 ( n − p ) + p 2 , 1 ≤ p ≤ 2 , 2 ( n − p ) n , 2 < p < n . (24) \overline{\gamma}(p)=\frac{p}{1+\gamma(p)}=
\begin{cases}
\frac{2p(n-p)}{2(n-p)+p^{2}}, & 1\le p\le2,\\
\frac{2(n-p)}{n}, & 2<p<n.
\end{cases} \tag*{(24)} γ ( p ) = 1 + γ ( p ) p = { 2 ( n − p ) + p 2 2 p ( n − p ) , n 2 ( n − p ) , 1 ≤ p ≤ 2 , 2 < p < n . (24)
Corollary 1.9. Under the same assumptions as in Theorem 1.7 , there exists a positive constant C = C ( m , n , s ) C=C(m,n,s) C = C ( m , n , s ) such that
log 1 − s n ( 1 ω n ∫ S n e n u d μ ) ≤ ( α s , n Θ ( m , 0 , n ) ) 1 − s n ∥ Δ s 2 u ∥ L n s ( S n ) n s − 1 + C . (25) \log^{1-\frac{s}{n}}\left(\frac{1}{\omega_{n}}\int_{\mathbb{S}^{n}}e^{nu}\,\mathrm{d}\mu\right)\le\left(\frac{\alpha_{s,n}}{\Theta(m,0,n)}\right)^{1-\frac{s}{n}}\lVert\Delta^{\frac{s}{2}}u\rVert_{L^{\frac{n}{s}}(\mathbb{S}^{n})}^{\frac{n}{s}-1}+C. \tag*{(25)} log 1 − n s ( ω n 1 ∫ S n e n u d μ ) ≤ ( Θ ( m , 0 , n ) α s , n ) 1 − n s ∥ Δ 2 s u ∥ L s n ( S n ) s n − 1 + C . (25)
The inequalities (23) and (25) are connected to the AB-program for Sobolev-type inequalities at different scales. We refer interested readers to [32 Section 4.4] for a detailed discussion on this topic.
Finally, we mention two other types of improved inequalities to which our approach can be extended.
The first type comes from the recent work [15 ] . In [15 Theorem 1.1] , Chen and Zhang established almost sharp improved first-order Moser–Trudinger–Onofri inequalities under the constraint that the second-order moments deviate suitably from zero on S 2 \mathbb{S}^{2} S 2 . This constraint can be viewed as an interesting variant of (17) . Here we generalize their results to higher-order inequalities on S n \mathbb{S}^{n} S n for all n ≥ 2 n \ge2 n ≥ 2 and obtain quantitative remainders (see Subsection 6.2 for details).
The second type arises from symmetry constraints, which form another classical topic. A brief introduction is given in § 7 . We focus on improved Sobolev and higher-order Moser–Trudinger–Onofri inequalities on general smooth, closed manifolds under symmetry constraints and obtain various new results on sharpness, non-sharpness, and quantitative remainder estimates (see Theorems 7.1 , 7.2 , and 7.8 for details).
Outline of strategies
We briefly describe the strategies used in this paper.
The starting point for the proof of Theorem 1.1 is the argument of Hang and Wang in [30 ] : we argue by contradiction and apply the concentration–compactness principle to obtain a sequence of functions u j u_j u j with suitable convergence properties. The key step is to decompose u j u_j u j into a sum of several Talenti bubbles and a remainder that is orthogonal to the tangent spaces of the Talenti bubbles and satisfies a spectral inequality (see Proposition 2.3 ). Through careful expansions, we derive delicate error estimates for this decomposition (see Lemmas 2.4 and 2.6 ). Finally, we expand the Sobolev inequality using this decomposition and apply the error estimates to obtain a contradiction.
To prove Theorem 1.2 , we use Druet’s localization theorem [32 Theorem 4.9] , which states that the failure of the sharp Sobolev inequality (6) is equivalent to its local failure near some point. Placing this local failure phenomenon around the support of an extremal measure yields the desired test function.
The proof of Theorem 1.4 relies on three ingredients. The first is a penalized higher-order Moser–Trudinger–Onofri inequality (see Lemma 4.2 ), which strengthens the classical inequality (8) by an L 2 L^{2} L 2 -term. The second is a mass separation lemma (see Lemma 4.3 ), which provides a uniform description of the mass distribution of the measure e n u ∫ S n e n u d μ d μ \frac{e^{nu}}{\int_{\mathbb{S}^{n}}e^{nu}\,d\mu}\,d\mu ∫ S n e n u d μ e n u d μ . The third (see Lemma 4.5) provides estimates for an orthogonal decomposition of H s ( S n ) H^{s}(\mathbb{S}^{n}) H s ( S n ) adapted to the mass distributions supplied by the second ingredient. To deduce Theorem 1.4 , we first apply the mass separation lemma to reduce the left-hand side of (21) to log ∫ U i e n u d μ \log\int_{U_i}e^{nu}\,d\mu log ∫ U i e n u d μ for some subset U i ⊂ S n U_i \subset\mathbb{S}^{n} U i ⊂ S n . The estimates for the orthogonal decomposition then reduce this expression further to log ∫ U i e n u i d μ \log\int_{U_i}e^{nu_i}\,d\mu log ∫ U i e n u i d μ , where u i u_i u i is an orthogonal component of u u u . We next apply the penalized higher-order Moser–Trudinger–Onofri inequality to each u i u_i u i . The additional L 2 L^{2} L 2 -term allows us to control the errors and ultimately deduce (21) .
For the proof of Theorem 1.5 , we construct test functions for general m m m by suitably perturbing constant functions. This construction is effective for large n n n . For m = 1 m=1 m = 1 , we also construct explicit test functions to cover the small-n n n case.
To prove Theorems 1.6 and 5.1 , we construct a suitable l p l^p l p -partition of unity η j \eta_j η j , apply (6) to each η j u \eta_j u η j u , and sum the resulting estimates. The key is to establish an effective lower bound for the term in (82) (see Proposition 5.7). This lower bound follows from a careful choice of partition functions (see Lemma 5.4 ) and a quantitative analysis of almost-orthogonality (see Lemmas 5.2 and 5.6).
For Theorems 1.7 and 6.2 , we combine the mass separation lemma (Lemma 4.3 ) with a quantitative “distribution of mass” principle for the higher-order Moser–Trudinger–Onofri inequality (8) (see Lemma 6.1 ). This type of principle originates in the work of Chen and Li [13 ] and has been widely used in the study of improved inequalities. However, standard approaches to this principle, such as truncation arguments and eigenspace decompositions, do not apply in our setting. We therefore develop a different method: we first reduce the estimate to a Riesz-type inequality using Green’s representation and then apply a truncation argument to the resulting inequality.
Finally, to prove Corollaries 1.8 and 1.9, respectively, we combine Theorems 1.6 and 1.7 with the elementary inequality ( x + y ) q ≥ x q + q x q − 1 y (x+y)^q \ge x^q+qx^{q-1}y ( x + y ) q ≥ x q + q x q − 1 y for x , y ≥ 0 x,y \ge0 x , y ≥ 0 and q ≥ 1 q \ge1 q ≥ 1 .
Structure of the paper
In Section 2, we prove Theorem 2.1 , which is equivalent to Theorem 1.1 . Theorem 1.2 is proved in Section 3. Section 4 is devoted to the proofs of Theorems 4.1 and 4.6 , which imply Theorems 1.4 and 1.5 , respectively. In Section 5 and Subsection 6.1, we establish Theorems 5.1 and 6.2 , which imply Theorems 1.6 and 1.7 , respectively. The proofs of Corollaries 1.8 and 1.9 are also given there. In Subsection 6.2, we generalize Theorem 1.1 of [15 ] to higher-order inequalities on S n \mathbb{S}^n S n for all n ≥ 2 n \ge2 n ≥ 2 . Finally, in Section 7, we study improved inequalities under symmetry constraints and obtain various new results on sharpness, non-sharpness, and quantitative remainder estimates.
Sharp improved Sobolev inequalities under low-order moment constraints
Throughout this section, we assume n ≥ 3 n \ge3 n ≥ 3 and let d w = d μ ω n dw = \frac{d\mu}{\omega_n} d w = ω n d μ be the normalized measure on S n \mathbb{S}^n S n . All norms are taken with respect to d w dw d w . We set
q = 2 n n − 2 , α = 2 q = n − 2 n , κ = n ( n − 2 ) 4 . q = \frac{2n}{n-2}, \qquad\alpha= \frac{2}{q} = \frac{n-2}{n}, \qquad\kappa= \frac{n(n-2)}{4}. q = n − 2 2 n , α = q 2 = n n − 2 , κ = 4 n ( n − 2 ) .
and equip H 1 ( S n ) H^1(\mathbb{S}^n) H 1 ( S n ) with the norm
∥ v ∥ H 2 = E ( v ) : = ∫ S n ( ∣ ∇ v ∣ 2 + κ v 2 ) d w . \lVert v\rVert_H^2 = E(v) := \int_{\mathbb{S}^n} \left(|\nabla v|^2 + \kappa v^2\right)\,dw. ∥ v ∥ H 2 = E ( v ) := ∫ S n ( ∣∇ v ∣ 2 + κ v 2 ) d w .
The sharp Sobolev inequality on S n \mathbb{S}^n S n reads (see, for example, [7 ] )
E ( v ) ≥ κ ∥ v ∥ L q ( S n ) 2 . (26) E(v) \ge\kappa\lVert v\rVert_{L^q(\mathbb{S}^n)}^2. \tag*{(26)} E ( v ) ≥ κ ∥ v ∥ L q ( S n ) 2 . (26)
For m ∈ { 1 , 2 , 3 } m \in\{1,2,3\} m ∈ { 1 , 2 , 3 } , let
N m = { 2 , m = 1 , n + 2 , m = 2 , 2 n + 2 , m = 3 , c m = N m − 1 , Θ m = Θ ( m , α , n ) = N m 1 − α . N_m =
\begin{cases}
2, & m=1,\\
n+2, & m=2,\\
2n+2, & m=3,
\end{cases}
\qquad c_m=N_m^{-1}, \qquad\Theta_m=\Theta(m,\alpha,n)=N_m^{1-\alpha}. N m = ⎩ ⎨ ⎧ 2 , n + 2 , 2 n + 2 , m = 1 , m = 2 , m = 3 , c m = N m − 1 , Θ m = Θ ( m , α , n ) = N m 1 − α .
For z , ξ ∈ S n z,\xi\in\mathbb{S}^n z , ξ ∈ S n , put s = z ⋅ ξ s=z\cdot\xi s = z ⋅ ξ and define three polynomials
H m , z ( ξ ) = { 1 + s 2 , m = 1 , [ 1 + ( n + 1 ) s ] 2 ( n + 2 ) 2 , m = 2 , s 2 ( 1 + s ) 2 , m = 3. (27) H_{m,z}(\xi)=
\begin{cases}
\dfrac{1+s}{2}, & m=1,\\
\dfrac{[1+(n+1)s]^2}{(n+2)^2}, & m=2,\\
\dfrac{s^2(1+s)}{2}, & m=3.
\end{cases} \tag*{(27)} H m , z ( ξ ) = ⎩ ⎨ ⎧ 2 1 + s , ( n + 2 ) 2 [ 1 + ( n + 1 ) s ] 2 , 2 s 2 ( 1 + s ) , m = 1 , m = 2 , m = 3. (27)
They satisfy
H m , z ≥ 0 , H m , z ( z ) = 1 , ∫ S n H m , z d w = c m , H m , z − c m ∈ P ˚ m , n . (28) H_{m,z}\ge0,\qquad H_{m,z}(z)=1,\qquad\int_{\mathbb{S}^n}H_{m,z}\,\mathrm{d}w=c_m,\qquad H_{m,z}-c_m\in\mathring{\mathcal{P}}_{m,n}. \tag*{(28)} H m , z ≥ 0 , H m , z ( z ) = 1 , ∫ S n H m , z d w = c m , H m , z − c m ∈ P ˚ m , n . (28)
Our main result is the following, which is equivalent to Theorem 1.1 .
Theorem 2.1. For every m ∈ { 1 , 2 , 3 } m\in\{1,2,3\} m ∈ { 1 , 2 , 3 } , there is a constant B = B ( m , n ) < ∞ B=B(m,n)<\infty B = B ( m , n ) < ∞ such that every u ∈ H 1 ( S n ) u\in H^1(\mathbb{S}^n) u ∈ H 1 ( S n ) satisfying (15) obeys
∥ u ∥ L q ( S n ) 2 ≤ 1 κ Θ m ∥ ∇ u ∥ L 2 ( S n ) 2 + B ∥ u ∥ L 2 ( S n ) 2 . (29) \lVert u\rVert_{L^q(\mathbb{S}^n)}^2\le\frac{1}{\kappa\Theta_m}\lVert\nabla u\rVert_{L^2(\mathbb{S}^n)}^2+B\lVert u\rVert_{L^2(\mathbb{S}^n)}^2. \tag*{(29)} ∥ u ∥ L q ( S n ) 2 ≤ κ Θ m 1 ∥ ∇ u ∥ L 2 ( S n ) 2 + B ∥ u ∥ L 2 ( S n ) 2 . (29)
The proof of Theorem 2.1 consists of a series of lemmas. The first lemma follows from [46 Theorem 4.1] , which provides useful estimates for cubature formulas. We include a simple proof for completeness.
Lemma 2.2. Suppose that ν = ∑ i w i δ z i ∈ M m , n c \nu=\sum_i w_i\delta_{z_i}\in\mathcal{M}^c_{m,n} ν = ∑ i w i δ z i ∈ M m , n c , where w i > 0 w_i>0 w i > 0 and ∑ i w i = 1 \sum_i w_i=1 ∑ i w i = 1 . Then
w i ≤ c m for every i , ∑ i w i α ≥ Θ m . (30) w_i\le c_m\quad\text{for every }i,\qquad\sum_i w_i^\alpha\ge\Theta_m. \tag*{(30)} w i ≤ c m for every i , i ∑ w i α ≥ Θ m . (30) Moreover, equality in the second inequality holds if and only if ν \nu ν has exactly N m N_m N m atoms, all of weight c m = N m − 1 c_m=N_m^{-1} c m = N m − 1 , and the configurations of the points z i z_i z i are, respectively, an antipodal pair when m = 1 m=1 m = 1 , the vertices of a regular simplex when m = 2 m=2 m = 2 , and the vertices of an orthogonal cross-polytope when m = 3 m=3 m = 3 .
Proof. From (28) we have
w i ≤ ∑ k w k H m , z i ( z k ) = ∫ S n H m , z i d w = c m . w_i\le\sum_k w_kH_{m,z_i}(z_k)=\int_{\mathbb{S}^n}H_{m,z_i}\,\mathrm{d}w=c_m. w i ≤ k ∑ w k H m , z i ( z k ) = ∫ S n H m , z i d w = c m .
Since α − 1 < 0 \alpha-1<0 α − 1 < 0 ,
∑ i w i α = ∑ i w i w i α − 1 ≥ c m α − 1 ∑ i w i = N m 1 − α = Θ m . \sum_i w_i^\alpha=\sum_i w_iw_i^{\alpha-1}\ge c_m^{\alpha-1}\sum_iw_i=N_m^{1-\alpha}=\Theta_m. i ∑ w i α = i ∑ w i w i α − 1 ≥ c m α − 1 i ∑ w i = N m 1 − α = Θ m .
Equality forces every positive weight to equal c m c_m c m , and the converse is immediate. Notice also that equality forces H m , z i ( z k ) = 0 H_{m,z_i}(z_k)=0 H m , z i ( z k ) = 0 for i ≠ k i\ne k i = k . Thus the equality configurations are, respectively, an antipodal pair, the vertices of a regular simplex, and the vertices of an orthogonal cross-polytope.
Our next proposition provides a crucial refinement of the concentration–compactness argument used by Hang and Wang in [30 ] . Before stating it, we recall some basic facts about Sobolev extremizers on S n \mathbb{S}^n S n . For 0 < ε ≤ 1 0<\varepsilon\le1 0 < ε ≤ 1 and z ∈ S n z\in\mathbb{S}^n z ∈ S n , let
U ε , z ( ξ ) = ( 2 ε 1 + ε 2 − ( 1 − ε 2 ) ξ ⋅ z ) n − 2 2 . (31) U_{\varepsilon,z}(\xi)=\left(\frac{2\varepsilon}{1+\varepsilon^2-(1-\varepsilon^2)\xi\cdot z}\right)^{\frac{n-2}{2}}. \tag*{(31)} U ε , z ( ξ ) = ( 1 + ε 2 − ( 1 − ε 2 ) ξ ⋅ z 2 ε ) 2 n − 2 . (31)
These are the normalized conformal Talenti bubbles. If we set A = − Δ S n + κ A=-\Delta_{\mathbb{S}^n}+\kappa A = − Δ S n + κ , then
∥ U ε , z ∥ L q ( S n ) = 1 , A U ε , z = κ U ε , z q − 1 , E ( U ε , z ) = κ . (32) \lVert U_{\varepsilon,z}\rVert_{L^q(\mathbb{S}^n)}=1,\qquad AU_{\varepsilon,z}=\kappa U_{\varepsilon,z}^{q-1},\qquad E(U_{\varepsilon,z})=\kappa. \tag*{(32)} ∥ U ε , z ∥ L q ( S n ) = 1 , A U ε , z = κ U ε , z q − 1 , E ( U ε , z ) = κ . (32)
Let T ε , z T_{\varepsilon,z} T ε , z be a conformal isometry of both H 1 ( S n ) H^1(\mathbb{S}^n) H 1 ( S n ) and L q ( S n ) L^q(\mathbb{S}^n) L q ( S n ) such that T ε , z 1 = U ε , z T_{\varepsilon,z}1=U_{\varepsilon,z} T ε , z 1 = U ε , z . If H k \mathcal{H}_k H k denotes the space of spherical harmonics of degree k k k , the tangent space to the bubble manifold at U = U ε , z U=U_{\varepsilon,z} U = U ε , z is
T ( U ) = T ε , z ( H 0 ⊕ H 1 ) . (33) \mathcal{T}(U)=T_{\varepsilon,z}(\mathcal{H}_0\oplus\mathcal{H}_1). \tag*{(33)} T ( U ) = T ε , z ( H 0 ⊕ H 1 ) . (33)
Proposition 2.3. Let m ∈ { 1 , 2 , 3 } m\in\{1,2,3\} m ∈ { 1 , 2 , 3 } , and let u j ∈ H 1 ( S n ) u_j\in H^1(\mathbb{S}^n) u j ∈ H 1 ( S n ) with u j ≥ 0 u_j\ge0 u j ≥ 0 satisfy (15) ,
∥ u j ∥ L q ( S n ) = 1 , ∥ u j ∥ L 2 ( S n ) ⟶ 0 , lim sup j → ∞ E ( u j ) ≤ κ Θ m . (34) \lVert u_j\rVert_{L^q(\mathbb{S}^n)}=1,\qquad\lVert u_j\rVert_{L^2(\mathbb{S}^n)}\longrightarrow0,\qquad\limsup_{j\to\infty}E(u_j)\le\kappa\Theta_m. \tag*{(34)} ∥ u j ∥ L q ( S n ) = 1 , ∥ u j ∥ L 2 ( S n ) ⟶ 0 , j → ∞ lim sup E ( u j ) ≤ κ Θ m . (34) After passing to a subsequence, there are distinct points z 1 ∗ , … , z N m ∗ ∈ S n z_1^*,\ldots,z_{N_m}^*\in\mathbb{S}^n z 1 ∗ , … , z N m ∗ ∈ S n such that
∣ u j ∣ q d ω ⇀ ∗ 1 N m ∑ i = 1 N m δ z i ∗ , E ( u j ) ⟶ κ Θ m . (35) |u_j|^q\,d\omega\mathop{\rightharpoonup^*} \frac{1}{N_m}\sum_{i=1}^{N_m}\delta_{z_i^*},\qquad E(u_j)\longrightarrow\kappa\Theta_m. \tag*{(35)} ∣ u j ∣ q d ω ⇀ ∗ N m 1 i = 1 ∑ N m δ z i ∗ , E ( u j ) ⟶ κ Θ m . (35) Moreover, one can write
u j = ∑ i = 1 N m a i , j U i , j + r j , U i , j = U ε i , j , z i , j , (36) u_j=\sum_{i=1}^{N_m}a_{i,j}U_{i,j}+r_j,\qquad U_{i,j}=U_{\varepsilon_{i,j},z_{i,j}}, \tag*{(36)} u j = i = 1 ∑ N m a i , j U i , j + r j , U i , j = U ε i , j , z i , j , (36) where
a i , j ⟶ N m − 1 q , ε i , j ⟶ 0 , z i , j ⟶ z i ∗ , ∥ r j ∥ H ⟶ 0 , (37) a_{i,j}\longrightarrow N_m^{-\frac{1}{q}},\qquad\varepsilon_{i,j}\longrightarrow0,\qquad z_{i,j}\longrightarrow z_i^*,\qquad\lVert r_j\rVert_H\longrightarrow0, \tag*{(37)} a i , j ⟶ N m − q 1 , ε i , j ⟶ 0 , z i , j ⟶ z i ∗ , ∥ r j ∥ H ⟶ 0 , (37) and
r j ⊥ H T ( U i , j ) ( 1 ≤ i ≤ N m ) . (38) r_j\perp_H\mathcal{T}(U_{i,j})\qquad(1\le i\le N_m). \tag*{(38)} r j ⊥ H T ( U i , j ) ( 1 ≤ i ≤ N m ) . (38) Furthermore, for every 0 < γ < 4 n + 4 0<\gamma<\frac{4}{n+4} 0 < γ < n + 4 4 and all sufficiently large j j j ,
∥ r j ∥ H 2 − ( q − 1 ) κ ∑ i = 1 N m ∫ S n U i , j q − 2 r j 2 d ω ≥ γ ∥ r j ∥ H 2 . (39) \lVert r_j\rVert_H^2-(q-1)\kappa\sum_{i=1}^{N_m}\int_{\mathbb{S}^n}U_{i,j}^{q-2}r_j^2\,d\omega\ge\gamma\lVert r_j\rVert_H^2. \tag*{(39)} ∥ r j ∥ H 2 − ( q − 1 ) κ i = 1 ∑ N m ∫ S n U i , j q − 2 r j 2 d ω ≥ γ ∥ r j ∥ H 2 . (39)
Proof. Since u j ⇀ 0 u_j\rightharpoonup0 u j ⇀ 0 in H 1 ( S n ) H^1(\mathbb{S}^n) H 1 ( S n ) , the concentration–compactness principle (see for example [30, Theorem 2.1) yields that, up to a subsequence,
∣ u j ∣ q d w ⇀ ∗ ν = ∑ ℓ ν ℓ δ x ℓ , ∣ ∇ u j ∣ 2 d w ⇀ ∗ μ , μ ≥ κ ∑ ℓ ν ℓ α δ x ℓ . |u_j|^q\,dw \mathrel{\overset{*}{\rightharpoonup}} \nu=\sum_{\ell}\nu_\ell\delta_{x_\ell},\qquad|\nabla u_j|^2\,dw \mathrel{\overset{*}{\rightharpoonup}} \mu,\qquad\mu\geq\kappa\sum_{\ell}\nu_\ell^\alpha\delta_{x_\ell}. ∣ u j ∣ q d w ⇀ ∗ ν = ℓ ∑ ν ℓ δ x ℓ , ∣∇ u j ∣ 2 d w ⇀ ∗ μ , μ ≥ κ ℓ ∑ ν ℓ α δ x ℓ .
The moment constraints pass to the limit, so ν ∈ M m , n c \nu\in\mathcal{M}_{m,n}^c ν ∈ M m , n c . Lemma 2.2 and (34) therefore give
κ Θ m ≥ μ ( S n ) ≥ κ ∑ ℓ ν ℓ α ≥ κ Θ m , \kappa\Theta_m\geq\mu(S^n)\geq\kappa\sum_{\ell}\nu_\ell^\alpha\geq\kappa\Theta_m, κ Θ m ≥ μ ( S n ) ≥ κ ℓ ∑ ν ℓ α ≥ κ Θ m ,
Thus every inequality is an equality. Lemma 2.2 then yields (35) .
Choose fixed disjoint neighborhoods of the points z i ∗ z_i^* z i ∗ and cut off u j u_j u j in these neighborhoods. Write u j = ∑ i = 1 N m χ i u j + u ~ j u_j=\sum_{i=1}^{N_m}\chi_i u_j+\widetilde{u}_j u j = ∑ i = 1 N m χ i u j + u j , where χ i \chi_i χ i is a cutoff function near z i ∗ z_i^* z i ∗ . The convergence properties imply that ∥ u ~ j ∥ H → 0 \|\widetilde{u}_j\|_H\to0 ∥ u j ∥ H → 0 , and each χ i u j \chi_i u_j χ i u j has L q L^q L q -mass tending to N m − 1 N_m^{-1} N m − 1 and energy tending to κ N m − α \kappa N_m^{-\alpha} κ N m − α . Thus the Sobolev deficit of χ i u j \chi_i u_j χ i u j tends to zero. By the spherical form of the Bianchi–Egnell stability theorem [8 ] , we can replace χ i u j \chi_i u_j χ i u j with a i , j 0 U i , j 0 a_{i,j}^0U_{i,j}^0 a i , j 0 U i , j 0 , where a i , j 0 → N m − 1 q a_{i,j}^0\to N_m^{-\frac{1}{q}} a i , j 0 → N m − q 1 , U i , j 0 = U ε i , j 0 , z i , j 0 U_{i,j}^0=U_{\varepsilon_{i,j}^0,z_{i,j}^0} U i , j 0 = U ε i , j 0 , z i , j 0 is a Talenti bubble, ε i , j 0 → 0 \varepsilon_{i,j}^0\to0 ε i , j 0 → 0 , and z i , j 0 → z i ∗ z_{i,j}^0\to z_i^* z i , j 0 → z i ∗ . Moreover, we have r j 0 : = u j − ∑ i = 1 N m a i , j 0 U i , j 0 → 0 r_j^0:=u_j-\sum_{i=1}^{N_m}a_{i,j}^0U_{i,j}^0\to0 r j 0 := u j − ∑ i = 1 N m a i , j 0 U i , j 0 → 0 in H 1 ( S n ) H^1(S^n) H 1 ( S n ) .
We next modify this decomposition to obtain the orthogonality condition (38) . Write T i , j 0 = T ε i , j 0 , z i , j 0 T_{i,j}^0=T_{\varepsilon_{i,j}^0,z_{i,j}^0} T i , j 0 = T ε i , j 0 , z i , j 0 . We use normalized coordinates on the bubble manifold. For ϑ = ( s , b ) ∈ R × R n + 1 \vartheta=(s,b)\in\mathbb{R}\times\mathbb{R}^{n+1} ϑ = ( s , b ) ∈ R × R n + 1 near the origin, set
W ( s , b ) ( ξ ) = ( 1 + s ) ( 1 − ∣ b ∣ 2 1 + ∣ b ∣ 2 − 2 b ⋅ ξ ) n − 2 2 . (40) W(s,b)(\xi)=(1+s)\left(\frac{1-|b|^2}{1+|b|^2-2b\cdot\xi}\right)^{\frac{n-2}{2}}. \tag*{(40)} W ( s , b ) ( ξ ) = ( 1 + s ) ( 1 + ∣ b ∣ 2 − 2 b ⋅ ξ 1 − ∣ b ∣ 2 ) 2 n − 2 . (40)
Thus W ( 0 , 0 ) = 1 W(0,0)=1 W ( 0 , 0 ) = 1 , and the derivatives of W W W at the origin span H 0 ⊕ H 1 \mathcal{H}_0\oplus\mathcal{H}_1 H 0 ⊕ H 1 . Consequently, for ϑ = ( ϑ 1 , … , ϑ N m ) ∈ R d \vartheta=(\vartheta_1,\ldots,\vartheta_{N_m})\in\mathbb{R}^d ϑ = ( ϑ 1 , … , ϑ N m ) ∈ R d , d = N m ( n + 2 ) d=N_m(n+2) d = N m ( n + 2 ) , the map
Ψ j ( ϑ ) = ∑ i = 1 N m a i , j 0 T i , j 0 W ( ϑ i ) (41) \Psi_j(\vartheta)=\sum_{i=1}^{N_m}a_{i,j}^0T_{i,j}^0W(\vartheta_i) \tag*{(41)} Ψ j ( ϑ ) = i = 1 ∑ N m a i , j 0 T i , j 0 W ( ϑ i ) (41)
parametrizes a neighborhood of the initial multi-bubble ∑ i = 1 N m a i , j 0 U i , j 0 \sum_{i=1}^{N_m}a_{i,j}^0U_{i,j}^0 ∑ i = 1 N m a i , j 0 U i , j 0 .
Index the coordinates of ϑ \vartheta ϑ by β = ( i , k ) \beta=(i,k) β = ( i , k ) , 1 ≤ i ≤ N m 1\leq i\leq N_m 1 ≤ i ≤ N m , 0 ≤ k ≤ n + 1 0\leq k\leq n+1 0 ≤ k ≤ n + 1 , and put
ϕ j , β ( ϑ ) = ∂ ϑ β Ψ j ( ϑ ) . \phi_{j,\beta}(\vartheta)=\partial_{\vartheta_\beta}\Psi_j(\vartheta). ϕ j , β ( ϑ ) = ∂ ϑ β Ψ j ( ϑ ) .
For a fixed i i i , the vectors { ϕ j , ( i , k ) ( ϑ ) : 0 ≤ k ≤ n + 1 } \{\phi_{j,(i,k)}(\vartheta):0\leq k\leq n+1\} { ϕ j , ( i , k ) ( ϑ ) : 0 ≤ k ≤ n + 1 } span the tangent space of the i i i th bubble factor. Define the finite-dimensional map
F j ( ϑ ) = ( ⟨ u j − Ψ j ( ϑ ) , ϕ j , β ( ϑ ) ⟩ H ) β ∈ R d . (42) F_j(\vartheta)=\bigl(\langle u_j-\Psi_j(\vartheta),\phi_{j,\beta}(\vartheta)\rangle_H\bigr)_\beta\in\mathbb{R}^d. \tag*{(42)} F j ( ϑ ) = ( ⟨ u j − Ψ j ( ϑ ) , ϕ j , β ( ϑ ) ⟩ H ) β ∈ R d . (42)
Hence F j ( ϑ ) = 0 F_j(\vartheta)=0 F j ( ϑ ) = 0 is equivalent to the orthogonality conditions in (38) . By conformal invariance and our choice of coordinates, the C 3 C^3 C 3 -norm of Ψ j \Psi_j Ψ j is uniformly bounded in a fixed neighborhood of 0. We therefore compute the derivative at the initial parameter ϑ = 0 \vartheta=0 ϑ = 0 :
∂ ϑ α F j , β ( 0 ) = − ⟨ ϕ j , α ( 0 ) , ϕ j , β ( 0 ) ⟩ H + ⟨ r j 0 , ∂ ϑ α ϕ j , β ( 0 ) ⟩ H = − ( G j ) α β + o ( 1 ) , (43) \partial_{\vartheta_\alpha}F_{j,\beta}(0)=-\langle\phi_{j,\alpha}(0),\phi_{j,\beta}(0)\rangle_H+\langle r_j^0,\partial_{\vartheta_\alpha}\phi_{j,\beta}(0)\rangle_H=-(G_j)_{\alpha\beta}+o(1), \tag*{(43)} ∂ ϑ α F j , β ( 0 ) = − ⟨ ϕ j , α ( 0 ) , ϕ j , β ( 0 ) ⟩ H + ⟨ r j 0 , ∂ ϑ α ϕ j , β ( 0 ) ⟩ H = − ( G j ) α β + o ( 1 ) , (43)
where ( G j ) α β = ⟨ ϕ j , α ( 0 ) , ϕ j , β ( 0 ) ⟩ H (G_j)_{\alpha\beta}=\langle\phi_{j,\alpha}(0),\phi_{j,\beta}(0)\rangle_H ( G j ) α β = ⟨ ϕ j , α ( 0 ) , ϕ j , β ( 0 ) ⟩ H . The diagonal blocks of G j G_j G j converge to positive multiples of the fixed single-bubble Gram matrix
G ∗ = ( ⟨ ∂ ϑ k W ( 0 ) , ∂ ϑ ℓ W ( 0 ) ⟩ H ) 0 ≤ k , ℓ ≤ n + 1 , G_*=\left(\langle\partial_{\vartheta_k}W(0),\partial_{\vartheta_\ell}W(0)\rangle_H\right)_{0\leq k,\ell\leq n+1}, G ∗ = ( ⟨ ∂ ϑ k W ( 0 ) , ∂ ϑ ℓ W ( 0 ) ⟩ H ) 0 ≤ k , ℓ ≤ n + 1 ,
which is positive definite. Since ε i , j 0 → 0 \varepsilon_{i,j}^0\to0 ε i , j 0 → 0 , z i , j 0 → z i ∗ z_{i,j}^0\to z_i^* z i , j 0 → z i ∗ , and z i ∗ ≠ z j ∗ z_i^*\ne z_j^* z i ∗ = z j ∗ for i ≠ j i\ne j i = j , all off-diagonal blocks tend to zero. Therefore,
G j ⟶ N m − 2 n G ∗ ⊕ ⋯ ⊕ N m − 2 n G ∗ , (44) G_j\longrightarrow N_m^{-\frac{2}{n}}G_*\oplus\cdots\oplus N_m^{-\frac{2}{n}}G_*, \tag*{(44)} G j ⟶ N m − n 2 G ∗ ⊕ ⋯ ⊕ N m − n 2 G ∗ , (44)
so D F j ( 0 ) DF_j(0) D F j ( 0 ) is invertible for large j j j and ∥ D F j ( 0 ) − 1 ∥ ≤ C \|DF_j(0)^{-1}\|\leq C ∥ D F j ( 0 ) − 1 ∥ ≤ C uniformly in j j j .
Since the C 3 C^3 C 3 -norm of Ψ j \Psi_j Ψ j is uniformly bounded in a fixed neighborhood of 0 and ∣ F j ( 0 ) ∣ → 0 |F_j(0)|\to0 ∣ F j ( 0 ) ∣ → 0 , we may use the local inverse function theorem in the following form: if F : R d → R d F:\mathbb{R}^d\to\mathbb{R}^d F : R d → R d is C 1 C^1 C 1 , D F ( 0 ) − 1 DF(0)^{-1} D F ( 0 ) − 1 is bounded by K K K , D F DF D F is sufficiently close to D F ( 0 ) DF(0) D F ( 0 ) in a fixed ball, and ∣ F ( 0 ) ∣ |F(0)| ∣ F ( 0 ) ∣ is small, then there is a unique zero ϑ \vartheta ϑ of F F F with ∣ ϑ ∣ ≤ 2 K ∣ F ( 0 ) ∣ |\vartheta|\leq2K|F(0)| ∣ ϑ ∣ ≤ 2 K ∣ F ( 0 ) ∣ . Applying this theorem to F j F_j F j for large j j j , we find a unique ϑ j \vartheta_j ϑ j such that F j ( ϑ j ) = 0 F_j(\vartheta_j)=0 F j ( ϑ j ) = 0 and ∣ ϑ j ∣ → 0 |\vartheta_j|\to0 ∣ ϑ j ∣ → 0 . Set a i , j U i , j = a i , j 0 T i , j 0 W ( ϑ i , j ) a_{i,j}U_{i,j}=a_{i,j}^0T_{i,j}^0W(\vartheta_{i,j}) a i , j U i , j = a i , j 0 T i , j 0 W ( ϑ i , j ) and
r j = u j − ∑ i = 1 N m a i , j U i , j . r_j=u_j-\sum_{i=1}^{N_m}a_{i,j}U_{i,j}. r j = u j − i = 1 ∑ N m a i , j U i , j .
Then (37) and (38) follow.
It remains to prove (39) . If not, we could normalize ∥ r j ∥ H = 1 \|r_j\|_H=1 ∥ r j ∥ H = 1 and, after pulling back by each bubble, assume
T ε i , j , z i , j − 1 r j ⇀ R i in H 1 ( S n ) , T ε i , j , z i , j − 1 r j ⟶ R i in L 2 ( S n ) . T_{\varepsilon_{i,j},z_{i,j}}^{-1}r_j\rightharpoonup R_i\quad\text{in }H^1(S^n),\qquad T_{\varepsilon_{i,j},z_{i,j}}^{-1}r_j\longrightarrow R_i\quad\text{in }L^2(S^n). T ε i , j , z i , j − 1 r j ⇀ R i in H 1 ( S n ) , T ε i , j , z i , j − 1 r j ⟶ R i in L 2 ( S n ) .
By (38) , R i ⊥ H ( H 0 ⊕ H 1 ) R_i\perp_H(\mathcal{H}_0\oplus\mathcal{H}_1) R i ⊥ H ( H 0 ⊕ H 1 ) . Expanding
∥ r j − ∑ i = 1 N m T ε i , j , z i , j R i ∥ H 2 ≥ 0 , \left\|r_j-\sum_{i=1}^{N_m}T_{\varepsilon_{i,j},z_{i,j}}R_i\right\|_H^2\geq0, r j − i = 1 ∑ N m T ε i , j , z i , j R i H 2 ≥ 0 ,
we obtain
∑ i = 1 N m ∥ R i ∥ H 2 ≤ 1. \sum_{i=1}^{N_m}\|R_i\|_H^2\leq1. i = 1 ∑ N m ∥ R i ∥ H 2 ≤ 1.
Since the first spherical eigenvalue outside H 0 ⊕ H 1 \mathcal{H}_0\oplus\mathcal{H}_1 H 0 ⊕ H 1 is 2 ( n + 1 ) 2(n+1) 2 ( n + 1 ) ,
( q − 1 ) κ ∑ i = 1 N m ∥ R i ∥ L 2 ( S n ) 2 ≤ ( q − 1 ) κ κ + 2 ( n + 1 ) ∑ i = 1 N m ∥ R i ∥ H 2 ≤ n n + 4 . (q-1)\kappa\sum_{i=1}^{N_m}\|R_i\|_{L^2(S^n)}^2\leq\frac{(q-1)\kappa}{\kappa+2(n+1)}\sum_{i=1}^{N_m}\|R_i\|_H^2\leq\frac{n}{n+4}. ( q − 1 ) κ i = 1 ∑ N m ∥ R i ∥ L 2 ( S n ) 2 ≤ κ + 2 ( n + 1 ) ( q − 1 ) κ i = 1 ∑ N m ∥ R i ∥ H 2 ≤ n + 4 n .
Finally, conformal invariance gives
∫ S n U i , j q − 2 r j 2 d w = ∥ T ε i , j , z i , j − 1 r j ∥ L 2 ( S n ) 2 ⟶ ∥ R i ∥ L 2 ( S n ) 2 . \int_{S^n}U_{i,j}^{q-2}r_j^2\,\mathrm{d}w=\left\|T_{\varepsilon_{i,j},z_{i,j}}^{-1}r_j\right\|_{L^2(S^n)}^2\longrightarrow\|R_i\|_{L^2(S^n)}^2. ∫ S n U i , j q − 2 r j 2 d w = T ε i , j , z i , j − 1 r j L 2 ( S n ) 2 ⟶ ∥ R i ∥ L 2 ( S n ) 2 .
Thus the lower limit of the left-hand side of (39) is at least 4 n + 4 \frac{4}{n+4} n + 4 4 , which yields a contradiction. This completes the proof. □ \square □
In what follows, we work along a sequence with the decomposition (36) –(38) . We henceforth assume that j j j is sufficiently large and suppress the index j j j . Set
V = ∑ i = 1 N m a i U i , t = ∥ r ∥ H , M = ∑ i = 1 N m ℓ n ( ε i ) , V=\sum_{i=1}^{N_m}a_iU_i,\qquad t=\|r\|_H,\qquad M=\sum_{i=1}^{N_m}\ell_n(\varepsilon_i), V = i = 1 ∑ N m a i U i , t = ∥ r ∥ H , M = i = 1 ∑ N m ℓ n ( ε i ) ,
where
ℓ n ( ε ) = { ε , n = 3 , ε 2 log ( 1 / ε ) , n = 4 , ε 2 , n ≥ 5. (45) \ell_n(\varepsilon)=
\begin{cases}
\varepsilon, & n=3,\\
\varepsilon^2\log(1/\varepsilon), & n=4,\\
\varepsilon^2, & n\ge5.
\end{cases} \tag*{(45)} ℓ n ( ε ) = ⎩ ⎨ ⎧ ε , ε 2 log ( 1/ ε ) , ε 2 , n = 3 , n = 4 , n ≥ 5. (45)
Our next two lemmas establish further estimates for this decomposition.
Lemma 2.4. Assume in addition that ∥ u ∥ L q ( S n ) = 1 \lVert u\rVert_{L^q(S^n)}=1 ∥ u ∥ L q ( S n ) = 1 and that u = V + r u=V+r u = V + r satisfies (15) . Then there exist positive constants b b b and C C C such that
∣ ∑ i = 1 N m a i q − 1 ∣ + max 1 ≤ i ≤ N m ∣ a i q − c m ∣ ≤ C ( M + t 2 ) , (46) \left|\sum_{i=1}^{N_m}a_i^q-1\right|+\max_{1\le i\le N_m}\left|a_i^q-c_m\right|\le C(M+t^2), \tag*{(46)} i = 1 ∑ N m a i q − 1 + 1 ≤ i ≤ N m max ∣ a i q − c m ∣ ≤ C ( M + t 2 ) , (46) and
E ( u ) − κ Θ m ∥ u ∥ L q ( S n ) 2 ≥ b ∥ r ∥ H 2 − C M . (47) E(u)-\kappa\Theta_m\lVert u\rVert_{L^q(S^n)}^2\ge b\lVert r\rVert_H^2-CM. \tag*{(47)} E ( u ) − κ Θ m ∥ u ∥ L q ( S n ) 2 ≥ b ∥ r ∥ H 2 − C M . (47)
Proof. With q ′ = q / ( q − 1 ) q'=q/(q-1) q ′ = q / ( q − 1 ) , set
I i k = ∫ S n U i q − 1 U k d w , δ = ∥ V q − 1 − ∑ i = 1 N m a i q − 1 U i q − 1 ∥ L q ′ ( S n ) . I_{ik}=\int_{S^n}U_i^{q-1}U_k\,dw,\qquad
\delta=\left\lVert V^{q-1}-\sum_{i=1}^{N_m}a_i^{q-1}U_i^{q-1}\right\rVert_{L^{q'}(S^n)}. I ik = ∫ S n U i q − 1 U k d w , δ = V q − 1 − i = 1 ∑ N m a i q − 1 U i q − 1 L q ′ ( S n ) .
The explicit decay in (31) gives
∑ i < k I i k ≤ C M , δ 2 ≤ C M , ∥ V q − 2 − ∑ i = 1 N m a i q − 2 U i q − 2 ∥ L n 2 ( S n ) = o ( 1 ) . (48) \sum_{i<k}I_{ik}\le CM,\qquad\delta^2\le CM,\qquad
\left\lVert V^{q-2}-\sum_{i=1}^{N_m}a_i^{q-2}U_i^{q-2}\right\rVert_{L^{\frac n2}(S^n)}=o(1). \tag*{(48)} i < k ∑ I ik ≤ C M , δ 2 ≤ C M , V q − 2 − i = 1 ∑ N m a i q − 2 U i q − 2 L 2 n ( S n ) = o ( 1 ) . (48)
Moreover, uniformly for f f f in a bounded subset of C 2 ( S n ) C^2(S^n) C 2 ( S n ) , we have
∫ S n f U i q d w = f ( z i ) + O f ( ε i 2 ) , ∥ ( f − f ( z i ) ) U i q − 1 ∥ L q ′ ( S n ) ≤ C f ε i . (49) \int_{S^n}fU_i^q\,dw=f(z_i)+O_f(\varepsilon_i^2),\qquad
\lVert(f-f(z_i))U_i^{q-1}\rVert_{L^{q'}(S^n)}\le C_f\varepsilon_i. \tag*{(49)} ∫ S n f U i q d w = f ( z i ) + O f ( ε i 2 ) , ∥( f − f ( z i )) U i q − 1 ∥ L q ′ ( S n ) ≤ C f ε i . (49)
The orthogonality condition (38) yields
∫ S n U i q − 1 r d w = 0. (50) \int_{S^n}U_i^{q-1}r\,dw=0. \tag*{(50)} ∫ S n U i q − 1 r d w = 0. (50)
For f f f in a bounded subset of C 2 ( S n ) C^2(S^n) C 2 ( S n ) , the first-order expansion of ∣ V + r ∣ q |V+r|^q ∣ V + r ∣ q gives
∫ S n f ∣ V + r ∣ q d w = ∫ S n f V q d w + q ∫ S n f V q − 1 r d w + O f ( t 2 ) . \int_{S^n}f|V+r|^q\,dw
=
\int_{S^n}fV^q\,dw
+
q\int_{S^n}fV^{q-1}r\,dw
+
O_f(t^2). ∫ S n f ∣ V + r ∣ q d w = ∫ S n f V q d w + q ∫ S n f V q − 1 r d w + O f ( t 2 ) .
Combining (48) , (49) , and (50) with a standard expansion argument, we obtain
∣ ∫ S n f V q − 1 r d w ∣ ≤ C f ( δ + ∑ i = 1 N m ε i ) t , ∫ S n f V q d w = ∑ i = 1 N m a i q f ( z i ) + O f ( M ) . \left|\int_{S^n}fV^{q-1}r\,dw\right|
\le C_f\left(\delta+\sum_{i=1}^{N_m}\varepsilon_i\right)t,\qquad
\int_{S^n}fV^q\,dw=\sum_{i=1}^{N_m}a_i^qf(z_i)+O_f(M). ∫ S n f V q − 1 r d w ≤ C f ( δ + i = 1 ∑ N m ε i ) t , ∫ S n f V q d w = i = 1 ∑ N m a i q f ( z i ) + O f ( M ) .
Since δ 2 + ∑ i = 1 N m ϵ i 2 ≤ C M \delta^2+\sum_{i=1}^{N_m}\epsilon_i^2\le CM δ 2 + ∑ i = 1 N m ϵ i 2 ≤ C M , Young’s inequality yields
∫ S n f ∣ u ∣ q d w = ∑ i = 1 N m a i q f ( z i ) + O f ( M + t 2 ) . (51) \int_{\mathbb{S}^n} f|u|^q\,\mathrm{d}w=\sum_{i=1}^{N_m}a_i^qf(z_i)+O_f(M+t^2). \tag*{(51)} ∫ S n f ∣ u ∣ q d w = i = 1 ∑ N m a i q f ( z i ) + O f ( M + t 2 ) . (51)
Taking f = 1 f=1 f = 1 gives
T : = ∑ i = 1 N m a i q = 1 + O ( M + t 2 ) . (52) T:=\sum_{i=1}^{N_m}a_i^q=1+O(M+t^2). \tag*{(52)} T := i = 1 ∑ N m a i q = 1 + O ( M + t 2 ) . (52)
Next, fix i i i and take f = H m , z i f=H_{m,z_i} f = H m , z i . Equations (28) and (15) give
c m = ∫ S n H m , z i ∣ u ∣ q d w = ∑ k = 1 N m a k q H m , z i ( z k ) + O ( M + t 2 ) . c_m=\int_{\mathbb{S}^n}H_{m,z_i}|u|^q\,\mathrm{d}w=\sum_{k=1}^{N_m}a_k^qH_{m,z_i}(z_k)+O(M+t^2). c m = ∫ S n H m , z i ∣ u ∣ q d w = k = 1 ∑ N m a k q H m , z i ( z k ) + O ( M + t 2 ) .
The nonnegativity of H m , z i H_{m,z_i} H m , z i and the identity H m , z i ( z i ) = 1 H_{m,z_i}(z_i)=1 H m , z i ( z i ) = 1 imply a i q − c m ≤ C ( M + t 2 ) a_i^q-c_m\le C(M+t^2) a i q − c m ≤ C ( M + t 2 ) . From (52) and N m c m = 1 N_mc_m=1 N m c m = 1 , we also obtain a i q − c m ≥ − C ( M + t 2 ) a_i^q-c_m\ge-C(M+t^2) a i q − c m ≥ − C ( M + t 2 ) . This proves (46) .
We turn to the estimate (47) . From (32) we have
E ( u ) = κ ∑ i = 1 N m a i 2 + 2 κ ∑ i < k a i a k I i k + t 2 . (53) E(u)=\kappa\sum_{i=1}^{N_m}a_i^2+2\kappa\sum_{i<k}a_ia_kI_{ik}+t^2. \tag*{(53)} E ( u ) = κ i = 1 ∑ N m a i 2 + 2 κ i < k ∑ a i a k I ik + t 2 . (53)
The second-order expansion yields
∫ S n ∣ V + r ∣ q d w = ∫ S n V q d w + q ∫ S n V q − 1 r d w + q ( q − 1 ) 2 ∫ S n V q − 2 r 2 d w + o ( t 2 ) . (54) \int_{\mathbb{S}^n}|V+r|^q\,\mathrm{d}w=\int_{\mathbb{S}^n}V^q\,\mathrm{d}w+q\int_{\mathbb{S}^n}V^{q-1}r\,\mathrm{d}w+\frac{q(q-1)}{2}\int_{\mathbb{S}^n}V^{q-2}r^2\,\mathrm{d}w+o(t^2). \tag*{(54)} ∫ S n ∣ V + r ∣ q d w = ∫ S n V q d w + q ∫ S n V q − 1 r d w + 2 q ( q − 1 ) ∫ S n V q − 2 r 2 d w + o ( t 2 ) . (54)
Using the concavity of s ↦ s α s\mapsto s^\alpha s ↦ s α together with (48) and (53) , we obtain
E ( u ) − κ Θ m ∥ u ∥ L q ( S n ) 2 ≥ κ ( ∑ i = 1 N m a i 2 − Θ m T α ) + t 2 − ( q − 1 ) κ Θ m T α − 1 ∫ S n V q − 2 r 2 d w − C ( M + δ t ) + o ( t 2 ) . (55) \begin{aligned}
E(u)-\kappa\Theta_m\|u\|_{L^q(\mathbb{S}^n)}^2
&\ge\kappa\left(\sum_{i=1}^{N_m}a_i^2-\Theta_mT^\alpha\right)+t^2\\
&\quad-(q-1)\kappa\Theta_mT^{\alpha-1}\int_{\mathbb{S}^n}V^{q-2}r^2\,\mathrm{d}w-C(M+\delta t)+o(t^2). \tag*{(55)}
\end{aligned} E ( u ) − κ Θ m ∥ u ∥ L q ( S n ) 2 ≥ κ ( i = 1 ∑ N m a i 2 − Θ m T α ) + t 2 − ( q − 1 ) κ Θ m T α − 1 ∫ S n V q − 2 r 2 d w − C ( M + δ t ) + o ( t 2 ) . (55)
To estimate the first term on the right, set x i = a i q x_i=a_i^q x i = a i q and x ˉ = T N m \bar{x}=\frac{T}{N_m} x ˉ = N m T . Combining (46) and Taylor’s formula with ∑ i = 1 N m ( x i − x ˉ ) = 0 \sum_{i=1}^{N_m}(x_i-\bar{x})=0 ∑ i = 1 N m ( x i − x ˉ ) = 0 gives
∑ i = 1 N m a i 2 − Θ m T α = ∑ i = 1 N m x i α − N m x ˉ α ≥ − C ∑ i = 1 N m ( x i − x ˉ ) 2 ≥ − C M + o ( t 2 ) . (56) \sum_{i=1}^{N_m}a_i^2-\Theta_mT^\alpha=\sum_{i=1}^{N_m}x_i^\alpha-N_m\bar{x}^\alpha\ge-C\sum_{i=1}^{N_m}(x_i-\bar{x})^2\ge-CM+o(t^2). \tag*{(56)} i = 1 ∑ N m a i 2 − Θ m T α = i = 1 ∑ N m x i α − N m x ˉ α ≥ − C i = 1 ∑ N m ( x i − x ˉ ) 2 ≥ − C M + o ( t 2 ) . (56)
On the other hand, (46) gives
Θ m T α − 1 a i q − 2 = 1 + O ( M + t 2 ) . (57) \Theta_mT^{\alpha-1}a_i^{q-2}=1+O(M+t^2). \tag*{(57)} Θ m T α − 1 a i q − 2 = 1 + O ( M + t 2 ) . (57)
It follows from (56) , (48) and (57) that
E ( u ) − κ Θ m ∥ u ∥ L q ( S n ) 2 ≥ t 2 − ( q − 1 ) κ ∑ i = 1 N m ∫ S n U i q − 2 r 2 d w − C ( M + δ t ) + o ( t 2 ) . E(u)-\kappa\Theta_m\lVert u\rVert_{L^q(\mathbb{S}^n)}^2 \ge t^2-(q-1)\kappa\sum_{i=1}^{N_m}\int_{\mathbb{S}^n}U_i^{q-2}r^2\,\mathrm{d}w-C(M+\delta t)+o(t^2). E ( u ) − κ Θ m ∥ u ∥ L q ( S n ) 2 ≥ t 2 − ( q − 1 ) κ i = 1 ∑ N m ∫ S n U i q − 2 r 2 d w − C ( M + δ t ) + o ( t 2 ) .
By applying (39) , (48) , and Young’s inequality δ t ≤ η t 2 + C η δ 2 \delta t\leq\eta t^2+C_\eta\delta^2 δ t ≤ η t 2 + C η δ 2 for η \eta η small, we obtain (47) .
Remark 2.5. We now explain why the assumption m ∈ { 1 , 2 , 3 } m\in\{1,2,3\} m ∈ { 1 , 2 , 3 } is crucial in our proof. By Lemma 2.2 , every extremal measure ν \nu ν has exactly N m N_m N m atoms, all of weight c m = N m − 1 c_m=N_m^{-1} c m = N m − 1 . This property makes the first term in the expansion of E ( u ) − κ Θ m ∥ u ∥ L q ( S n ) 2 E(u)-\kappa\Theta_m\lVert u\rVert_{L^q(\mathbb{S}^n)}^2 E ( u ) − κ Θ m ∥ u ∥ L q ( S n ) 2 small (see (56) ) and makes the quadratic term exactly the left-hand side of (39) , yielding c t 2 ct^2 c t 2 as a lower bound. When m ≥ 4 m\geq4 m ≥ 4 , the configurations of extremal measures are generally difficult to determine (see, for example, [16 , 17 , 46 ] ); if the weights of ν \nu ν are not equal, one may not be able to apply (39) .
Lemma 2.6. For every fixed sufficiently small constant ρ > 0 \rho>0 ρ > 0 , we have
M ≤ C ρ ∥ u ∥ L 2 ( S n ) + C 0 η n ( ρ ) ∥ r ∥ H 2 , η n ( ρ ) = { ρ , n = 3 , ρ 2 , n ≥ 4 , (58) M\leq C_\rho\lVert u\rVert_{L^2(\mathbb{S}^n)}+C_0\eta_n(\rho)\lVert r\rVert_H^2,\qquad
\eta_n(\rho)=
\begin{cases}
\rho, & n=3,\\
\rho^2, & n\geq4,
\end{cases} \tag*{(58)} M ≤ C ρ ∥ u ∥ L 2 ( S n ) + C 0 η n ( ρ ) ∥ r ∥ H 2 , η n ( ρ ) = { ρ , ρ 2 , n = 3 , n ≥ 4 , (58) where C 0 C_0 C 0 is independent of ρ \rho ρ .
Proof. Since j j j is sufficiently large, the balls B i = B ( z i , ρ ) B_i=B(z_i,\rho) B i = B ( z i , ρ ) are disjoint, and a i ≥ N m − 1 / q 2 a_i\geq\frac{N_m^{-1/q}}{2} a i ≥ 2 N m − 1/ q . We then have
∑ i = 1 N m ∫ B i a i 2 U i 2 d w ≤ ∑ i = 1 N m ∫ B i V 2 d w ≤ 2 ∥ u ∥ L 2 ( S n ) 2 + 2 ∑ i = 1 N m ∫ B i r 2 d w . (59) \sum_{i=1}^{N_m}\int_{B_i}a_i^2U_i^2\,\mathrm{d}w\leq\sum_{i=1}^{N_m}\int_{B_i}V^2\,\mathrm{d}w\leq2\lVert u\rVert_{L^2(\mathbb{S}^n)}^2+2\sum_{i=1}^{N_m}\int_{B_i}r^2\,\mathrm{d}w. \tag*{(59)} i = 1 ∑ N m ∫ B i a i 2 U i 2 d w ≤ i = 1 ∑ N m ∫ B i V 2 d w ≤ 2 ∥ u ∥ L 2 ( S n ) 2 + 2 i = 1 ∑ N m ∫ B i r 2 d w . (59)
Hölder’s inequality and the sharp Sobolev inequality (26) yield
∑ i = 1 N m ∫ B i r 2 d w ≤ C ρ 2 ∥ r ∥ H 2 . \sum_{i=1}^{N_m}\int_{B_i}r^2\,\mathrm{d}w\leq C\rho^2\lVert r\rVert_H^2. i = 1 ∑ N m ∫ B i r 2 d w ≤ C ρ 2 ∥ r ∥ H 2 .
Direct radial integration in (31) gives
∫ B i U i 2 d w ≥ { c ρ ε i , n = 3 , − c ε i 2 log ( ε i ) , n = 4 , c ε i 2 , n ≥ 5. \int_{B_i}U_i^2\,\mathrm{d}w\geq
\begin{cases}
c\rho\varepsilon_i, & n=3,\\
-c\varepsilon_i^2\log(\varepsilon_i), & n=4,\\
c\varepsilon_i^2, & n\geq5.
\end{cases} ∫ B i U i 2 d w ≥ ⎩ ⎨ ⎧ c ρ ε i , − c ε i 2 log ( ε i ) , c ε i 2 , n = 3 , n = 4 , n ≥ 5.
Inserting these two estimates into (59) yields (58) .
Now we are ready to give the proof of Theorem 2.1 .
Proof of Theorem 2.1. Fix m ∈ { 1 , 2 , 3 } m\in\{1,2,3\} m ∈ { 1 , 2 , 3 } and define
D m ( u ) = E ( u ) − κ Θ m ∥ u ∥ L q ( S n ) 2 . D_m(u)=E(u)-\kappa\Theta_m\lVert u\rVert_{L^q(\mathbb{S}^n)}^2. D m ( u ) = E ( u ) − κ Θ m ∥ u ∥ L q ( S n ) 2 .
It suffices to prove that D m ( u ) ≥ − B m , n ∥ u ∥ L 2 2 D_m(u) \ge-B_{m,n}\|u\|_{L^2}^2 D m ( u ) ≥ − B m , n ∥ u ∥ L 2 2 for some finite constant B m , n B_{m,n} B m , n . Otherwise, we can find nonnegative functions u j ∈ H 1 ( S n ) u_j \in H^1(\mathbb{S}^n) u j ∈ H 1 ( S n ) satisfying (15) such that
∥ u j ∥ L q ( S n ) = 1 , D m ( u j ) < − j ∥ u j ∥ L 2 ( S n ) 2 . (60) \|u_j\|_{L^q(\mathbb{S}^n)}=1,\qquad D_m(u_j)<-j\|u_j\|_{L^2(\mathbb{S}^n)}^2. \tag*{(60)} ∥ u j ∥ L q ( S n ) = 1 , D m ( u j ) < − j ∥ u j ∥ L 2 ( S n ) 2 . (60)
Consequently,
E ( u j ) + j ∥ u j ∥ L 2 ( S n ) 2 < κ Θ m , E(u_j)+j\|u_j\|_{L^2(\mathbb{S}^n)}^2<\kappa\Theta_m, E ( u j ) + j ∥ u j ∥ L 2 ( S n ) 2 < κ Θ m ,
so (34) holds. Proposition 2.3 and Lemma 2.4 give
D m ( u j ) ≥ b ∥ r j ∥ H 2 − C M j . D_m(u_j)\ge b\|r_j\|_H^2-CM_j. D m ( u j ) ≥ b ∥ r j ∥ H 2 − C M j .
Choose ρ > 0 \rho>0 ρ > 0 so small that C C 0 η n ( ρ ) < b / 2 CC_0\eta_n(\rho)<b/2 C C 0 η n ( ρ ) < b /2 . Lemma 2.6 then gives, for all sufficiently large j j j ,
D m ( u j ) ≥ b 2 ∥ r j ∥ H 2 − C C ρ ∥ u j ∥ L 2 ( S n ) 2 ≥ − C C ρ ∥ u j ∥ L 2 ( S n ) 2 , D_m(u_j)\ge\frac{b}{2}\|r_j\|_H^2-CC_\rho\|u_j\|_{L^2(\mathbb{S}^n)}^2\ge-CC_\rho\|u_j\|_{L^2(\mathbb{S}^n)}^2, D m ( u j ) ≥ 2 b ∥ r j ∥ H 2 − C C ρ ∥ u j ∥ L 2 ( S n ) 2 ≥ − C C ρ ∥ u j ∥ L 2 ( S n ) 2 ,
contradicting (60) . This completes the proof.
A failure principle for sharp improved Sobolev inequalities
In this section, we prove Theorem 1.2. We first recall Druet’s localization theorem [32 Theorem 4.9] , whose main idea originates in Druet’s work [19 ] .
Theorem 3.1. Let ( M , g ) (M,g) ( M , g ) be a smooth closed manifold of dimension n ≥ 2 n\ge2 n ≥ 2 . Suppose that, for some 1 ≤ p < n 1\le p<n 1 ≤ p < n , the sharp Sobolev inequality (6) is locally valid in the sense that each x ∈ M x\in M x ∈ M possesses an open neighborhood Ω x \Omega_x Ω x such that, for every u ∈ C c ∞ ( Ω x ) u\in C^\infty_c(\Omega_x) u ∈ C c ∞ ( Ω x ) ,
∥ u ∥ L p ∗ ( M ) p ≤ S n , p p ∥ ∇ u ∥ L p ( M ) p + B x ∥ u ∥ L p ( M ) p . \|u\|_{L^{p^*}(M)}^p\le S_{n,p}^p\|\nabla u\|_{L^p(M)}^p+B_x\|u\|_{L^p(M)}^p. ∥ u ∥ L p ∗ ( M ) p ≤ S n , p p ∥∇ u ∥ L p ( M ) p + B x ∥ u ∥ L p ( M ) p . for some constant B x B_x B x independent of u u u . Then the sharp Sobolev inequality (6) holds globally on M M M for some constant B ( M , p ) B(M,p) B ( M , p ) .
The following corollary is crucial to our construction.
Corollary 3.2. Assume 1 ≤ p < n 1\le p<n 1 ≤ p < n and that the sharp Sobolev inequality (6) fails on S n \mathbb{S}^n S n . Then, for any z ∈ S n z\in\mathbb{S}^n z ∈ S n , there exist a sequence r k → 0 + r_k\to0^+ r k → 0 + and nonnegative functions w k ∈ W 1 , p ( S n ) w_k\in W^{1,p}(\mathbb{S}^n) w k ∈ W 1 , p ( S n ) such that supp w k ⊂ B ( z , r k ) \operatorname{supp}w_k\subset B(z,r_k) supp w k ⊂ B ( z , r k ) , each w k w_k w k is radially symmetric with respect to z z z , and
∥ w k ∥ L p ∗ ( S n ) p − S n , p p ∥ ∇ w k ∥ L p ( S n ) p ∥ w k ∥ L p ( S n ) p → + ∞ . \frac{\|w_k\|_{L^{p^*}(\mathbb{S}^n)}^p-S_{n,p}^p\|\nabla w_k\|_{L^p(\mathbb{S}^n)}^p}{\|w_k\|_{L^p(\mathbb{S}^n)}^p}\to+\infty. ∥ w k ∥ L p ( S n ) p ∥ w k ∥ L p ∗ ( S n ) p − S n , p p ∥∇ w k ∥ L p ( S n ) p → + ∞. Proof. By Theorem 3.1 and the rotation invariance of S n \mathbb{S}^n S n , for any z ∈ S n z\in\mathbb{S}^n z ∈ S n , we can find a sequence r k → 0 + r_k\to0^+ r k → 0 + and functions v k ∈ C c ∞ ( S n ) v_k\in C^\infty_c(\mathbb{S}^n) v k ∈ C c ∞ ( S n ) such that supp v k ⊂ B ( z , r k ) \operatorname{supp}v_k\subset B(z,r_k) supp v k ⊂ B ( z , r k ) and
∥ v k ∥ L p ∗ ( S n ) p − S n , p p ∥ ∇ v k ∥ L p ( S n ) p ∥ v k ∥ L p ( S n ) p → + ∞ . \frac{\|v_k\|_{L^{p^*}(\mathbb{S}^n)}^p-S_{n,p}^p\|\nabla v_k\|_{L^p(\mathbb{S}^n)}^p}{\|v_k\|_{L^p(\mathbb{S}^n)}^p}\to+\infty. ∥ v k ∥ L p ( S n ) p ∥ v k ∥ L p ∗ ( S n ) p − S n , p p ∥∇ v k ∥ L p ( S n ) p → + ∞. By the spherical Pólya–Szegő inequality, it suffices to take w k = ∣ v k ∣ ∗ w_k=|v_k|^* w k = ∣ v k ∣ ∗ , where w k w_k w k is the spherical decreasing rearrangement of ∣ v k ∣ |v_k| ∣ v k ∣ with respect to z z z .
We now prove Theorem 1.2 .
Proof of Theorem 1.2. It suffices to find a sequence of test functions u k ∈ W 1 , p ( S n ) u_k \in W^{1,p}(\mathbb{S}^{n}) u k ∈ W 1 , p ( S n ) satisfying (15) such that
∥ u k ∥ L p ∗ ( S n ) p − S p , n p Θ ( m , n − p n , n ) ∥ ∇ u k ∥ L p ( S n ) p ∥ u k ∥ L p ( S n ) p → + ∞ . \frac{\lVert u_k\rVert_{L^{p^*}(\mathbb{S}^{n})}^{p}-\dfrac{S_{p,n}^{p}}{\Theta\left(m,\frac{n-p}{n},n\right)}\lVert\nabla u_k\rVert_{L^{p}(\mathbb{S}^{n})}^{p}}{\lVert u_k\rVert_{L^{p}(\mathbb{S}^{n})}^{p}} \to+\infty. ∥ u k ∥ L p ( S n ) p ∥ u k ∥ L p ∗ ( S n ) p − Θ ( m , n n − p , n ) S p , n p ∥ ∇ u k ∥ L p ( S n ) p → + ∞.
By [46 Corollary 3.2] , there exists a finitely supported measure ν ∈ M m , n c \nu\in\mathcal{M}^{c}_{m,n} ν ∈ M m , n c such that
∥ ν ∥ n − p n = Θ ( m , n − p n , n ) . \lVert\nu\rVert_{\frac{n-p}{n}}=\Theta\left(m,\frac{n-p}{n},n\right). ∥ ν ∥ n n − p = Θ ( m , n n − p , n ) .
Write ν = ∑ i = 1 N ν i δ ξ i \nu=\sum_{i=1}^{N}\nu_i\delta_{\xi_i} ν = ∑ i = 1 N ν i δ ξ i .
Fix a point z ∈ S n z \in\mathbb{S}^{n} z ∈ S n and let w k ( y ) = ϕ k ( d ( y , z ) ) w_k(y)=\phi_k(d(y,z)) w k ( y ) = ϕ k ( d ( y , z )) be the sequence of functions given by Corollary 3.2. Set w i , k ( y ) = ϕ k ( d ( y , ξ i ) ) w_{i,k}(y)=\phi_k(d(y,\xi_i)) w i , k ( y ) = ϕ k ( d ( y , ξ i )) for 1 ≤ i ≤ N 1\leq i\leq N 1 ≤ i ≤ N . For sufficiently large k k k , the supports of w i , k w_{i,k} w i , k are mutually disjoint. For such k k k , define
u k = ∑ i = 1 N ν i 1 / p ∗ w i , k . u_k=\sum_{i=1}^{N}\nu_i^{1/p^*}w_{i,k}. u k = i = 1 ∑ N ν i 1/ p ∗ w i , k .
We claim that these u k u_k u k are the required test functions.
We first verify the moment constraints. Let H k \mathcal{H}_k H k denote the space of spherical harmonics of degree k k k on S n \mathbb{S}^{n} S n . We have the orthogonal decomposition P ˚ m , n = ⨁ k = 1 m H k \mathring{\mathcal{P}}_{m,n}=\bigoplus_{k=1}^{m}\mathcal{H}_k P ˚ m , n = ⨁ k = 1 m H k (see [49 ] for details). Since w k w_k w k is radial, the Funk–Hecke formula gives, for every Y ∈ H ℓ Y\in\mathcal{H}_{\ell} Y ∈ H ℓ , 1 ≤ ℓ ≤ m 1\leq\ell\leq m 1 ≤ ℓ ≤ m ,
∫ S n Y ( x ) w i , k p ∗ ( x ) d μ ( x ) = b ℓ , k Y ( ξ i ) \int_{\mathbb{S}^{n}}Y(x)w_{i,k}^{p^*}(x)\,\mathrm{d}\mu(x)=b_{\ell,k}Y(\xi_i) ∫ S n Y ( x ) w i , k p ∗ ( x ) d μ ( x ) = b ℓ , k Y ( ξ i )
for some constant b ℓ , k b_{\ell,k} b ℓ , k independent of i i i . Therefore,
∫ S n Y ∣ u k ∣ p ∗ d μ = b ℓ , k ∑ i = 1 N ν i Y ( ξ i ) = b ℓ , k ∫ S n Y d ν = 0. \int_{\mathbb{S}^{n}}Y|u_k|^{p^*}\,\mathrm{d}\mu=b_{\ell,k}\sum_{i=1}^{N}\nu_iY(\xi_i)=b_{\ell,k}\int_{\mathbb{S}^{n}}Y\,\mathrm{d}\nu=0. ∫ S n Y ∣ u k ∣ p ∗ d μ = b ℓ , k i = 1 ∑ N ν i Y ( ξ i ) = b ℓ , k ∫ S n Y d ν = 0.
Hence (15) holds for u k u_k u k .
Since the supports of w i , k w_{i,k} w i , k are disjoint, our choice of ν \nu ν gives
∥ u k ∥ L p ∗ ( S n ) p ∗ = ∑ i = 1 N ν i ∥ w k ∥ L p ∗ ( S n ) p ∗ = ∥ w k ∥ L p ∗ ( S n ) p ∗ . \lVert u_k\rVert_{L^{p^*}(\mathbb{S}^{n})}^{p^*}=\sum_{i=1}^{N}\nu_i\lVert w_k\rVert_{L^{p^*}(\mathbb{S}^{n})}^{p^*}=\lVert w_k\rVert_{L^{p^*}(\mathbb{S}^{n})}^{p^*}. ∥ u k ∥ L p ∗ ( S n ) p ∗ = i = 1 ∑ N ν i ∥ w k ∥ L p ∗ ( S n ) p ∗ = ∥ w k ∥ L p ∗ ( S n ) p ∗ .
Similarly, we have
∥ ∇ u k ∥ L p ( S n ) p = ∑ i = 1 N ν i p / p ∗ ∥ ∇ w k ∥ L p ( S n ) p = Θ ( m , n − p n , n ) ∥ ∇ w k ∥ L p ( S n ) p , \lVert\nabla u_k\rVert_{L^{p}(\mathbb{S}^{n})}^{p}=\sum_{i=1}^{N}\nu_i^{p/p^*}\lVert\nabla w_k\rVert_{L^{p}(\mathbb{S}^{n})}^{p}=\Theta\left(m,\frac{n-p}{n},n\right)\lVert\nabla w_k\rVert_{L^{p}(\mathbb{S}^{n})}^{p}, ∥ ∇ u k ∥ L p ( S n ) p = i = 1 ∑ N ν i p / p ∗ ∥ ∇ w k ∥ L p ( S n ) p = Θ ( m , n n − p , n ) ∥ ∇ w k ∥ L p ( S n ) p ,
and
∥ u k ∥ L p ( S n ) p = ∑ i = 1 N ν i p / p ∗ ∥ w k ∥ L p ( S n ) p = Θ ( m , n − p n , n ) ∥ w k ∥ L p ( S n ) p . \lVert u_k\rVert_{L^{p}(\mathbb{S}^{n})}^{p}=\sum_{i=1}^{N}\nu_i^{p/p^*}\lVert w_k\rVert_{L^{p}(\mathbb{S}^{n})}^{p}=\Theta\left(m,\frac{n-p}{n},n\right)\lVert w_k\rVert_{L^{p}(\mathbb{S}^{n})}^{p}. ∥ u k ∥ L p ( S n ) p = i = 1 ∑ N ν i p / p ∗ ∥ w k ∥ L p ( S n ) p = Θ ( m , n n − p , n ) ∥ w k ∥ L p ( S n ) p .
Therefore,
∥ u k ∥ L p ∗ ( S n ) p − S n , p p Θ ( m , n − p n p , n ) ∥ ∇ u k ∥ L p ( S n ) p ∥ u k ∥ L p ( S n ) p = 1 Θ ( m , n − p n p , n ) ∥ w k ∥ p ∗ p − S n , p p ∥ ∇ w k ∥ p p ∥ w k ∥ p p . \frac{\lVert u_k\rVert_{L^{p^*}(\mathbb{S}^n)}^p-\dfrac{S_{n,p}^p}{\Theta\left(m,\frac{n-p}{n}p,n\right)}\lVert\nabla u_k\rVert_{L^p(\mathbb{S}^n)}^p}{\lVert u_k\rVert_{L^p(\mathbb{S}^n)}^p}
=
\frac{1}{\Theta\left(m,\frac{n-p}{n}p,n\right)}
\frac{\lVert w_k\rVert_{p^*}^p-S_{n,p}^p\lVert\nabla w_k\rVert_p^p}{\lVert w_k\rVert_p^p}. ∥ u k ∥ L p ( S n ) p ∥ u k ∥ L p ∗ ( S n ) p − Θ ( m , n n − p p , n ) S n , p p ∥ ∇ u k ∥ L p ( S n ) p = Θ ( m , n n − p p , n ) 1 ∥ w k ∥ p p ∥ w k ∥ p ∗ p − S n , p p ∥ ∇ w k ∥ p p .
Since the right-hand side tends to + ∞ +\infty + ∞ by Corollary 3.2, this completes the proof. □
Sharp improved Moser–Trudinger–Onofri inequalities under general moment constraints
In this section, we prove Theorems 1.4 and 1.5 . The first subsection deals with Theorem 1.4 , and the second is devoted to Theorem 1.5 .
Throughout this section, we write d w = d μ ω n \mathrm{d}w=\frac{\mathrm{d}\mu}{\omega_n} d w = ω n d μ for the normalized surface measure on the sphere S n \mathbb{S}^n S n , and set L = − Δ S n L=-\Delta_{\mathbb{S}^n} L = − Δ S n for notational simplicity. All norms are taken with respect to d w \mathrm{d}w d w . Let H k \mathcal{H}_k H k be the space of spherical harmonics of degree k k k . Then
L ∣ H k = λ k Id , λ k = k ( k + n − 1 ) . L\big|_{\mathcal{H}_k}=\lambda_k\operatorname{Id},\qquad\lambda_k=k(k+n-1). L H k = λ k Id , λ k = k ( k + n − 1 ) .
For u ∈ H n 2 ( S n ) u\in H^{\frac{n}{2}}(\mathbb{S}^n) u ∈ H 2 n ( S n ) , define
E n ( u ) : = ⟨ u , L n 2 u ⟩ L 2 ( d w ) = ∑ k = 1 ∞ λ k n 2 ∥ u k ∥ L 2 ( S n ) 2 . E_n(u):=\left\langle u,L^{\frac{n}{2}}u\right\rangle_{L^2(\mathrm{d}w)}
=\sum_{k=1}^{\infty}\lambda_k^{\frac{n}{2}}\lVert u_k\rVert_{L^2(\mathbb{S}^n)}^2. E n ( u ) := ⟨ u , L 2 n u ⟩ L 2 ( d w ) = k = 1 ∑ ∞ λ k 2 n ∥ u k ∥ L 2 ( S n ) 2 .
where the operator L n 2 L^{\frac{n}{2}} L 2 n is defined by
L n 2 u = ∑ k = 1 ∞ λ k n 2 u k . L^{\frac{n}{2}}u=\sum_{k=1}^{\infty}\lambda_k^{\frac{n}{2}}u_k. L 2 n u = k = 1 ∑ ∞ λ k 2 n u k .
Here u = ∑ k = 0 ∞ u k u=\sum_{k=0}^{\infty}u_k u = ∑ k = 0 ∞ u k is the spherical harmonic decomposition. For u , v ∈ H n 2 ( S n ) u,v\in H^{\frac{n}{2}}(\mathbb{S}^n) u , v ∈ H 2 n ( S n ) , we also define
B n ( u , v ) : = 1 4 ( E n ( u + v ) − E n ( u − v ) ) = ⟨ L n 4 u , L n 4 v ⟩ L 2 ( d w ) . B_n(u,v):=\frac{1}{4}\left(E_n(u+v)-E_n(u-v)\right)
=\left\langle L^{\frac{n}{4}}u,L^{\frac{n}{4}}v\right\rangle_{L^2(\mathrm{d}w)}. B n ( u , v ) := 4 1 ( E n ( u + v ) − E n ( u − v ) ) = ⟨ L 4 n u , L 4 n v ⟩ L 2 ( d w ) .
If ∫ S n u d w = 0 \int_{\mathbb{S}^n}u\,\mathrm{d}w=0 ∫ S n u d w = 0 , that is, if u 0 = 0 u_0=0 u 0 = 0 , then the Poincaré inequality gives E n ( u ) ≥ n n 2 ∥ u ∥ L 2 ( S n ) 2 E_n(u)\geq n^{\frac{n}{2}}\lVert u\rVert_{L^2(\mathbb{S}^n)}^2 E n ( u ) ≥ n 2 n ∥ u ∥ L 2 ( S n ) 2 .
Throughout what follows, we fix m ≥ 1 m\geq1 m ≥ 1 and n ≥ 2 n\geq2 n ≥ 2 , and set
N ∗ = Θ ( m , 0 , n ) , A m , n = n 2 ( n − 1 ) ! Θ ( m , 0 , n ) . (61) N_*=\Theta(m,0,n),\qquad A_{m,n}=\frac{n}{2(n-1)!\Theta(m,0,n)}. \tag*{(61)} N ∗ = Θ ( m , 0 , n ) , A m , n = 2 ( n − 1 )! Θ ( m , 0 , n ) n . (61)
The sharpness result
The following result is equivalent to Theorem 1.4 .
Theorem 4.1. There exists a positive constant C = C ( m , n ) C=C(m,n) C = C ( m , n ) such that for any u ∈ H n 2 ( S n ) u\in H^{\frac{n}{2}}(\mathbb{S}^n) u ∈ H 2 n ( S n )
satisfying ∫ S n u d w = 0 \int_{\mathbb{S}^{n}} u\,\mathrm{d}w = 0 ∫ S n u d w = 0 and (17) ,
log ∫ S n e n u d w ≤ A m , n E n ( u ) + C . (62) \log\int_{\mathbb{S}^{n}} e^{nu}\,\mathrm{d}w \le A_{m,n}E_n(u) + C. \tag*{(62)} log ∫ S n e n u d w ≤ A m , n E n ( u ) + C . (62)
The proof of Theorem 4.1 has three ingredients. The first is the following penalized Adams inequality and its consequence, a penalized higher-order Moser–Trudinger–Onofri inequality. These strengthen the classical inequalities (7) and (8) , respectively, in the case s = n 2 s = \frac{n}{2} s = 2 n .
Lemma 4.2. Assume 0 ≤ κ < n n / 2 0 \le\kappa< n^{n/2} 0 ≤ κ < n n /2 and define
Q κ ( v ) : = E n ( v ) − κ ∥ v ∥ L 2 ( S n ) 2 Q_{\kappa}(v) := E_n(v) - \kappa\|v\|_{L^2(\mathbb{S}^{n})}^{2} Q κ ( v ) := E n ( v ) − κ ∥ v ∥ L 2 ( S n ) 2 for every nonzero v ∈ H n 2 ( S n ) v \in H^{\frac{n}{2}}(\mathbb{S}^{n}) v ∈ H 2 n ( S n ) satisfying ∫ S n v d w = 0 \int_{\mathbb{S}^{n}} v\,\mathrm{d}w = 0 ∫ S n v d w = 0 . Then there exists a positive constant C ( n , κ ) C(n,\kappa) C ( n , κ ) such that
∫ S n exp ( n ! 2 v 2 Q κ ( v ) ) d w ≤ C ( n , κ ) . (63) \int_{\mathbb{S}^{n}} \exp\left(\frac{n!}{2}\frac{v^2}{Q_{\kappa}(v)}\right)\,\mathrm{d}w \le C(n,\kappa). \tag*{(63)} ∫ S n exp ( 2 n ! Q κ ( v ) v 2 ) d w ≤ C ( n , κ ) . (63) Consequently,
log ∫ S n e n v d w ≤ n 2 ( n − 1 ) ! Q κ ( v ) + log C ( n , κ ) . (64) \log\int_{\mathbb{S}^{n}} e^{nv}\,\mathrm{d}w \le\frac{n}{2(n-1)!}Q_{\kappa}(v) + \log C(n,\kappa). \tag*{(64)} log ∫ S n e n v d w ≤ 2 ( n − 1 )! n Q κ ( v ) + log C ( n , κ ) . (64)
Similar penalized Moser–Trudinger-type inequalities have been widely studied in the Euclidean setting; see, for example, [3 , 53 , 55 ] . To the best of our knowledge, the inequalities (63) and (64) have not appeared in this form. The key idea is to reduce (63) to a special case of the abstract Adams inequality of Fontana and Morpurgo [22 Theorem 1] . This reduction uses some basic facts from the theory of pseudodifferential operators. We refer the reader to [37 Sections 18.1–18.2] and [48 Chapter 1] for background.
Proof. Let Π 0 \Pi_0 Π 0 be the L 2 ( d w ) L^2(\mathrm{d}w) L 2 ( d w ) -orthogonal projection onto the constants. We define an operator
B κ : = L ~ n / 2 − κ ( I − Π 0 ) = L n / 2 − κ ( I − Π 0 ) + Π 0 , B_{\kappa} := \widetilde{L}^{n/2} - \kappa(I-\Pi_0) = L^{n/2} - \kappa(I-\Pi_0) + \Pi_0, B κ := L n /2 − κ ( I − Π 0 ) = L n /2 − κ ( I − Π 0 ) + Π 0 ,
where L ~ : = L + Π 0 \widetilde{L} := L + \Pi_0 L := L + Π 0 . Note that B κ B_{\kappa} B κ acts as the identity on the constant functions, while on H k \mathcal{H}_k H k , k ≥ 1 k \ge1 k ≥ 1 , it acts by λ k n / 2 − κ \lambda_k^{n/2} - \kappa λ k n /2 − κ . Since λ 1 = n \lambda_1 = n λ 1 = n and κ < n n / 2 \kappa< n^{n/2} κ < n n /2 , all eigenvalues are strictly positive. Hence B κ B_{\kappa} B κ is positive definite, self-adjoint, and invertible.
Since L L L is a second-order elliptic differential operator and Π 0 \Pi_0 Π 0 has finite rank, L ~ = L + Π 0 \widetilde{L} = L + \Pi_0 L = L + Π 0 is a positive classical elliptic pseudodifferential operator of order 2 2 2 , with principal symbol ∣ ζ ∣ g 2 |\zeta|_g^2 ∣ ζ ∣ g 2 . Seeley’s theorem on complex powers of elliptic operators (see, for example, [47 ] and [48 Chapter 2] ) implies that L ~ n / 2 \widetilde{L}^{n/2} L n /2 is a classical elliptic pseudodifferential operator of order n n n , with principal symbol ∣ ζ ∣ g n |\zeta|_g^n ∣ ζ ∣ g n . It follows that B κ B_{\kappa} B κ is also a classical elliptic pseudodifferential operator of order n n n , with principal symbol ∣ ζ ∣ g n |\zeta|_g^n ∣ ζ ∣ g n .
Applying Seeley’s theorem on complex powers of elliptic operators once more (see, for example, [47 ] and [48 Chapter 2] ) to B κ B_{\kappa} B κ , we find that T κ : = B κ − 1 / 2 T_{\kappa} := B_{\kappa}^{-1/2} T κ := B κ − 1/2 is a classical self-adjoint elliptic pseudodifferential operator of order − n 2 -\frac{n}{2} − 2 n , with principal symbol ∣ ζ ∣ g − n / 2 |\zeta|_g^{-n/2} ∣ ζ ∣ g − n /2 . By [36 Theorem 5.2.1] , we may take K κ ( x , y ) K_{\kappa}(x,y) K κ ( x , y ) to be the Schwartz kernel of T κ T_{\kappa} T κ relative to d w \mathrm{d}w d w . Since T κ T_{\kappa} T κ is a classical pseudodifferential operator, K κ ( x , y ) K_{\kappa}(x,y) K κ ( x , y ) is smooth away from the diagonal. Moreover, since the principal symbol is ∣ ζ ∣ g − n 2 |\zeta|_{\mathrm{g}^{-\frac{n}{2}}} ∣ ζ ∣ g − 2 n , a local kernel expansion near the diagonal shows that, for some δ > 0 \delta>0 δ > 0 ,
K κ ( x , y ) = ω n ( 2 π ) − n 2 d g ( x , y ) − n 2 + O ( d g ( x , y ) − n 2 + δ ) , y → x , K_{\kappa}(x,y)=\omega_n(2\pi)^{-\frac{n}{2}}d_{\mathrm{g}}(x,y)^{-\frac{n}{2}}+O\left(d_{\mathrm{g}}(x,y)^{-\frac{n}{2}+\delta}\right),\qquad y\to x, K κ ( x , y ) = ω n ( 2 π ) − 2 n d g ( x , y ) − 2 n + O ( d g ( x , y ) − 2 n + δ ) , y → x ,
uniformly with respect to x ∈ S n x\in\mathbb{S}^n x ∈ S n (the factor ω n \omega_n ω n appears because the kernel is taken with respect to the normalized measure d w dw d w ).
To apply the abstract Adams inequality [22 Theorem 1] to T κ T_\kappa T κ , it suffices to check that conditions (13) and (14) therein hold for the kernel K κ K_\kappa K κ .
For fixed x ∈ S n x\in\mathbb{S}^n x ∈ S n , let K κ , x ( y ) : = K κ ( x , y ) K_{\kappa,x}(y):=K_\kappa(x,y) K κ , x ( y ) := K κ ( x , y ) and denote by
m ( K κ , x , s ) = w ( { y ∈ S n : ∣ K κ , x ( y ) ∣ > s } ) , m(K_{\kappa,x},s)=w\left(\left\{y\in\mathbb{S}^n:|K_{\kappa,x}(y)|>s\right\}\right), m ( K κ , x , s ) = w ( { y ∈ S n : ∣ K κ , x ( y ) ∣ > s } ) ,
the distribution function of K κ , x K_{\kappa,x} K κ , x . Let K κ , x ∗ K_{\kappa,x}^{*} K κ , x ∗ denote the decreasing rearrangement of K κ , x K_{\kappa,x} K κ , x :
K κ , x ∗ ( t ) = inf { s > 0 : m ( K κ , x , s ) ≤ t } . K_{\kappa,x}^{*}(t)=\inf\left\{s>0:m(K_{\kappa,x},s)\leq t\right\}. K κ , x ∗ ( t ) = inf { s > 0 : m ( K κ , x , s ) ≤ t } .
Combining the expansion above with the small-geodesic-ball volume estimate
w ( B g ( x , r ) ) = ω n − 1 n ω n r n ( 1 + O ( r 2 ) ) w(B_{\mathrm{g}}(x,r))=\frac{\omega_{n-1}}{n\omega_n}r^n(1+O(r^2)) w ( B g ( x , r )) = n ω n ω n − 1 r n ( 1 + O ( r 2 ))
gives
sup x ∈ S n m ( K κ , x , s ) ≤ ω n − 1 ω n ( 2 π ) n s − 2 ( 1 + O ( s − δ ′ ) ) , s → + ∞ , \sup_{x\in\mathbb{S}^n}m(K_{\kappa,x},s)\leq\frac{\omega_{n-1}\omega_n}{(2\pi)^n}s^{-2}(1+O(s^{-\delta'})),\qquad s\to+\infty, x ∈ S n sup m ( K κ , x , s ) ≤ ( 2 π ) n ω n − 1 ω n s − 2 ( 1 + O ( s − δ ′ )) , s → + ∞ ,
for some δ ′ > 0 \delta'>0 δ ′ > 0 . Set
Φ ( s ) : = ω n − 1 ω n ( 2 π ) n s − 2 ( 1 + O ( s − δ ′ ) ) . \Phi(s):=\frac{\omega_{n-1}\omega_n}{(2\pi)^n}s^{-2}(1+O(s^{-\delta'})). Φ ( s ) := ( 2 π ) n ω n − 1 ω n s − 2 ( 1 + O ( s − δ ′ )) .
For any t > Φ ( s ) t>\Phi(s) t > Φ ( s ) and x ∈ S n x\in\mathbb{S}^n x ∈ S n , from
m ( K κ , x , s ) ≤ Φ ( s ) < t m(K_{\kappa,x},s)\leq\Phi(s)<t m ( K κ , x , s ) ≤ Φ ( s ) < t
we obtain K κ , x ∗ ( t ) ≤ s K_{\kappa,x}^{*}(t)\leq s K κ , x ∗ ( t ) ≤ s . Thus,
k 1 ∗ ( t ) : = sup x ∈ S n K κ , x ∗ ( t ) ≤ s k_1^{*}(t):=\sup_{x\in\mathbb{S}^n}K_{\kappa,x}^{*}(t)\leq s k 1 ∗ ( t ) := x ∈ S n sup K κ , x ∗ ( t ) ≤ s
whenever t > Φ ( s ) t>\Phi(s) t > Φ ( s ) , and hence
m ( k 1 ∗ , s ) ≤ Φ ( s ) . m(k_1^{*},s)\leq\Phi(s). m ( k 1 ∗ , s ) ≤ Φ ( s ) .
This is precisely the first distribution estimate required in [22 Theorem 1] . Since T κ T_\kappa T κ is self-adjoint, the second required distribution estimate follows in the same way:
m ( k 2 ∗ , s ) ≤ Φ ( s ) . m(k_2^{*},s)\leq\Phi(s). m ( k 2 ∗ , s ) ≤ Φ ( s ) .
Consequently, applying [22 Theorem 1] to T κ T_\kappa T κ gives
∫ S n exp ( n ! 2 ∣ T κ f ∣ 2 ∥ f ∥ L 2 ( S n ) 2 ) d w ≤ C ( n , κ ) \int_{\mathbb{S}^n}\exp\left(\frac{n!}{2}\frac{|T_\kappa f|^2}{\|f\|_{L^2(\mathbb{S}^n)}^2}\right)dw\leq C(n,\kappa) ∫ S n exp ( 2 n ! ∥ f ∥ L 2 ( S n ) 2 ∣ T κ f ∣ 2 ) d w ≤ C ( n , κ )
for every nonzero function f ∈ L 2 ( S n ) f \in L^2(S^n) f ∈ L 2 ( S n ) (here we also use the identity ω n − 1 ω n = 2 n + 1 π n ( n − 1 ) ! \omega_{n-1}\omega_n = \frac{2^{n+1}\pi^n}{(n-1)!} ω n − 1 ω n = ( n − 1 )! 2 n + 1 π n ). For every v ∈ H n 2 ( S n ) v \in H^{\frac{n}{2}}(S^n) v ∈ H 2 n ( S n ) satisfying ∫ S n v d w = 0 \int_{S^n} v\,\mathrm{d}w = 0 ∫ S n v d w = 0 , taking f = B κ 1 / 2 v f = B_\kappa^{1/2}v f = B κ 1/2 v yields (63) .
Finally, Young’s inequality gives
n v ≤ n ! 2 v 2 Q κ ( v ) + n 2 2 n ! Q κ ( v ) . nv \le\frac{n!}{2}\frac{v^2}{Q_\kappa(v)}+\frac{n^2}{2n!}Q_\kappa(v). n v ≤ 2 n ! Q κ ( v ) v 2 + 2 n ! n 2 Q κ ( v ) .
Therefore,
log ∫ S n e n v d w ≤ n 2 2 n ! Q κ ( v ) + log C ( n , κ ) = n 2 ( n − 1 ) ! Q κ ( v ) + log C ( n , κ ) , \log\int_{S^n} e^{nv}\,\mathrm{d}w \le\frac{n^2}{2n!}Q_\kappa(v)+\log C(n,\kappa)=\frac{n}{2(n-1)!}Q_\kappa(v)+\log C(n,\kappa), log ∫ S n e n v d w ≤ 2 n ! n 2 Q κ ( v ) + log C ( n , κ ) = 2 ( n − 1 )! n Q κ ( v ) + log C ( n , κ ) ,
which is (64) . This completes the proof.
The second ingredient is the following mass separation lemma.
Lemma 4.3. Assume σ \sigma σ is a Borel probability measure on S n S^n S n satisfying
∫ S n f d σ = 0 for every f ∈ P ˚ m , n . \int_{S^n} f\,\mathrm{d}\sigma=0 \quad\text{for every } f\in\mathring{\mathcal{P}}_{m,n}. ∫ S n f d σ = 0 for every f ∈ P ˚ m , n . Then there exists a positive constant R = R ( m , n ) > 0 R=R(m,n)>0 R = R ( m , n ) > 0 such that, for any 0 < r < R 100 0<r<\frac{R}{100} 0 < r < 100 R , one can find N ∗ N_* N ∗ points x 1 , … , x N ∗ ∈ S n x_1,\ldots,x_{N_*}\in S^n x 1 , … , x N ∗ ∈ S n such that d ( x i , x j ) ≥ R 2 d(x_i,x_j)\ge\frac{R}{2} d ( x i , x j ) ≥ 2 R for i ≠ j i\ne j i = j and σ ( B ( x i , r ) ) ≥ δ r \sigma(B(x_i,r))\ge\delta_r σ ( B ( x i , r )) ≥ δ r for every i i i , where δ r = δ r ( m , n ) > 0 \delta_r=\delta_r(m,n)>0 δ r = δ r ( m , n ) > 0 .
Proof. Let f 1 , … , f t f_1,\ldots,f_t f 1 , … , f t be an orthonormal basis of P ˚ m , n \mathring{\mathcal{P}}_{m,n} P ˚ m , n in L 2 ( d w ) L^2(\mathrm{d}w) L 2 ( d w ) , and define F = ( f 1 , … , f t ) : S n → R t F=(f_1,\ldots,f_t):S^n\to\mathbb{R}^t F = ( f 1 , … , f t ) : S n → R t . Set C 1 = ∥ F ∥ L ∞ ( S n ) C_1=\lVert F\rVert_{L^\infty(S^n)} C 1 = ∥ F ∥ L ∞ ( S n ) and C 2 = ∥ ∇ F ∥ L ∞ ( S n ) C_2=\lVert\nabla F\rVert_{L^\infty(S^n)} C 2 = ∥ ∇ F ∥ L ∞ ( S n ) . By the definition of N ∗ N_* N ∗ in (61) and a standard compactness argument,
C 3 : = min { ∣ ∑ i = 1 N ∗ − 1 w i F ( x i ) ∣ : w i ≥ 0 , ∑ i = 1 N ∗ − 1 w i = 1 , x i ∈ S n , 1 ≤ i ≤ N ∗ − 1 } C_3:=\min\left\{\left|\sum_{i=1}^{N_*-1}w_iF(x_i)\right|:w_i\ge0,\ \sum_{i=1}^{N_*-1}w_i=1,\ x_i\in S^n,\ 1\le i\le N_*-1\right\} C 3 := min { i = 1 ∑ N ∗ − 1 w i F ( x i ) : w i ≥ 0 , i = 1 ∑ N ∗ − 1 w i = 1 , x i ∈ S n , 1 ≤ i ≤ N ∗ − 1 }
is a positive number. Set R = min { π 16 , C 3 16 C 2 } R=\min\left\{\frac{\pi}{16},\frac{C_3}{16C_2}\right\} R = min { 16 π , 16 C 2 C 3 } and η = C 3 16 C 1 \eta=\frac{C_3}{16C_1} η = 16 C 1 C 3 . We claim that for any N ∗ − 1 N_*-1 N ∗ − 1 points x 1 , … , x N ∗ − 1 ∈ S n x_1,\ldots,x_{N_*-1}\in S^n x 1 , … , x N ∗ − 1 ∈ S n ,
σ ( S n ∖ ⋃ i = 1 N ∗ − 1 B ( x i , R ) ) ≥ η . \sigma\left(S^n\setminus\bigcup_{i=1}^{N_*-1}B(x_i,R)\right)\ge\eta. σ ( S n ∖ i = 1 ⋃ N ∗ − 1 B ( x i , R ) ) ≥ η .
Indeed, if not, we take mutually disjoint Borel sets A i ⊂ B ( x i , R ) A_i\subset B(x_i,R) A i ⊂ B ( x i , R ) such that
⋃ i = 1 N ∗ − 1 A i = ⋃ i = 1 N ∗ − 1 B ( x i , R ) \bigcup_{i=1}^{N_*-1}A_i=\bigcup_{i=1}^{N_*-1}B(x_i,R) i = 1 ⋃ N ∗ − 1 A i = i = 1 ⋃ N ∗ − 1 B ( x i , R )
and define a discrete probability measure
σ 0 : = ∑ i = 1 N ∗ − 1 σ ( A i ) δ x i + ( 1 − ∑ i = 1 N ∗ − 1 σ ( A i ) ) δ x 1 . \sigma_0:=\sum_{i=1}^{N_*-1}\sigma(A_i)\delta_{x_i}+\left(1-\sum_{i=1}^{N_*-1}\sigma(A_i)\right)\delta_{x_1}. σ 0 := i = 1 ∑ N ∗ − 1 σ ( A i ) δ x i + ( 1 − i = 1 ∑ N ∗ − 1 σ ( A i ) ) δ x 1 .
Since ∫ S n F d σ = 0 \int_{\mathbb{S}^{n}} F\,d\sigma=0 ∫ S n F d σ = 0 , we have
∣ ∫ S n F d σ 0 ∣ ≤ ∑ i = 1 N ∗ − 1 ∣ ∫ A i F d σ − σ ( A i ) F ( x i ) ∣ + ( 1 − ∑ i = 1 N ∗ − 1 σ ( A i ) ) ∣ F ( x 1 ) ∣ + ∣ ∫ S n ∖ ⋃ A i F d σ ∣ ≤ ∑ i = 1 N ∗ − 1 C 2 R σ ( A i ) + 2 C 1 η ≤ C 3 2 , \begin{aligned}
\left|\int_{\mathbb{S}^{n}}F\,d\sigma_{0}\right|
\leq\sum_{i=1}^{N_{*}-1}\left|\int_{A_i}F\,d\sigma-\sigma(A_i)F(x_i)\right|+\left(1-\sum_{i=1}^{N_{*}-1}\sigma(A_i)\right)|F(x_1)|+\left|\int_{\mathbb{S}^{n}\setminus\bigcup A_i}F\,d\sigma\right| \\
&\leq\sum_{i=1}^{N_{*}-1}C_2R\sigma(A_i)+2C_1\eta\leq\frac{C_3}{2},
\end{aligned} ∫ S n F d σ 0 ≤ i = 1 ∑ N ∗ − 1 ∫ A i F d σ − σ ( A i ) F ( x i ) + ( 1 − i = 1 ∑ N ∗ − 1 σ ( A i ) ) ∣ F ( x 1 ) ∣ + ∫ S n ∖ ⋃ A i F d σ ≤ i = 1 ∑ N ∗ − 1 C 2 R σ ( A i ) + 2 C 1 η ≤ 2 C 3 ,
which contradicts the definition of C 3 C_3 C 3 above.
To finish the proof, choose finitely many points y 1 , ⋯ , y s y_1,\cdots,y_s y 1 , ⋯ , y s such that S n ⊂ ⋃ i B ( y i , r ) \mathbb{S}^{n}\subset\bigcup_i B(y_i,r) S n ⊂ ⋃ i B ( y i , r ) . We may assume that s ≤ C r − n s\leq Cr^{-n} s ≤ C r − n for some positive constant C = C ( n ) C=C(n) C = C ( n ) . First choose x 1 ∈ { y i : 1 ≤ i ≤ s } x_1\in\{y_i:1\leq i\leq s\} x 1 ∈ { y i : 1 ≤ i ≤ s } such that σ ( B ( x 1 , r ) ) ≥ s − 1 \sigma(B(x_1,r))\geq s^{-1} σ ( B ( x 1 , r )) ≥ s − 1 . By the claim above, we can then find x 2 ∈ { y i : 1 ≤ i ≤ s } x_2\in\{y_i:1\leq i\leq s\} x 2 ∈ { y i : 1 ≤ i ≤ s } such that d ( x 1 , x 2 ) ≥ R − r d(x_1,x_2)\geq R-r d ( x 1 , x 2 ) ≥ R − r and σ ( B ( x 2 , r ) ) ≥ η s − 1 \sigma(B(x_2,r))\geq\eta s^{-1} σ ( B ( x 2 , r )) ≥ η s − 1 . Iterating this procedure, we obtain N ∗ N_* N ∗ points x 1 , ⋯ , x N ∗ ∈ S n x_1,\cdots,x_{N_*}\in\mathbb{S}^{n} x 1 , ⋯ , x N ∗ ∈ S n such that d ( x i , x j ) ≥ R − r ≥ R 2 d(x_i,x_j)\geq R-r\geq\frac{R}{2} d ( x i , x j ) ≥ R − r ≥ 2 R for i ≠ j i\neq j i = j and σ ( B ( x i , r ) ) ≥ η s − 1 \sigma(B(x_i,r))\geq\eta s^{-1} σ ( B ( x i , r )) ≥ η s − 1 . This completes the proof.
Remark 4.4. The preceding proof also shows that the same conclusion holds under the relaxed condition
∣ ∫ S n f d σ ∣ ≤ ε 0 ∥ f ∥ L ∞ for every f ∈ P ˚ m , n , \left|\int_{\mathbb{S}^{n}}f\,d\sigma\right|\leq\varepsilon_0\|f\|_{L^\infty}\qquad\text{for every }f\in\mathring{P}_{m,n}, ∫ S n f d σ ≤ ε 0 ∥ f ∥ L ∞ for every f ∈ P ˚ m , n ,
where ε 0 = ε 0 ( m , n ) \varepsilon_0=\varepsilon_0(m,n) ε 0 = ε 0 ( m , n ) is a small positive number. This observation is useful in the proof of Theorem 6.2 .
The third ingredient provides crucial estimates for suitable orthogonal decompositions of H n 2 ( S n ) H^{\frac{n}{2}}(\mathbb{S}^{n}) H 2 n ( S n ) . Specifically, let x 1 , ⋯ , x N ∗ ∈ S n x_1,\cdots,x_{N_*}\in\mathbb{S}^{n} x 1 , ⋯ , x N ∗ ∈ S n be the points obtained in Lemma 4.3 , and assume that 0 < r < R 100 0<r<\frac{R}{100} 0 < r < 100 R . Set U i = B ( x i , 2 r ) U_i=B(x_i,2r) U i = B ( x i , 2 r ) and define
X i : = { φ ∈ C c ∞ ( U i ) : ∫ S n φ d w = 0 } ⊂ H n 2 ( S n ) , X : = ⨁ i = 1 N ∗ X i . X_i:=\left\{\varphi\in C^\infty_c(U_i):\int_{\mathbb{S}^{n}}\varphi\,dw=0\right\}\subset H^{\frac{n}{2}}(\mathbb{S}^{n}),\qquad X:=\bigoplus_{i=1}^{N_*}X_i. X i := { φ ∈ C c ∞ ( U i ) : ∫ S n φ d w = 0 } ⊂ H 2 n ( S n ) , X := i = 1 ⨁ N ∗ X i .
For any v ∈ H n 2 ( S n ) v\in H^{\frac{n}{2}}(\mathbb{S}^{n}) v ∈ H 2 n ( S n ) with ∫ S n v d w = 0 \int_{\mathbb{S}^{n}}v\,dw=0 ∫ S n v d w = 0 , there is a unique decomposition v = ∑ i = 1 N ∗ v i + h v=\sum_{i=1}^{N_*}v_i+h v = ∑ i = 1 N ∗ v i + h such that v i ∈ X i v_i\in X_i v i ∈ X i for every i i i and h h h is orthogonal to X X X , that is, B n ( h , f ) = 0 B_n(h,f)=0 B n ( h , f ) = 0 for every f ∈ X f\in X f ∈ X . We then have the following estimates.
Lemma 4.5. There exist two positive constants C 1 = C 1 ( n , R ) C_1=C_1(n,R) C 1 = C 1 ( n , R ) and C 2 = C 2 ( n , r ) C_2=C_2(n,r) C 2 = C 2 ( n , r ) such that
∣ B n ( v i , v j ) ∣ ≤ C 1 r n ∥ v i ∥ L 2 ( S n ) ∥ v j ∥ L 2 ( S n ) (65) |B_n(v_i,v_j)|\leq C_1r^n\|v_i\|_{L^2(\mathbb{S}^{n})}\|v_j\|_{L^2(\mathbb{S}^{n})} \tag*{(65)} ∣ B n ( v i , v j ) ∣ ≤ C 1 r n ∥ v i ∥ L 2 ( S n ) ∥ v j ∥ L 2 ( S n ) (65)
for every i ≠ j i\neq j i = j , and
∥ h ∥ L ∞ ( B ( x i , r ) ) ≤ C 2 E n ( h ) 1 2 (66) \|h\|_{L^\infty(B(x_i,r))}\leq C_2E_n(h)^{\frac{1}{2}} \tag*{(66)} ∥ h ∥ L ∞ ( B ( x i , r )) ≤ C 2 E n ( h ) 2 1 (66)
for every i i i .
Proof. We first prove (65) . If n n n is even, L n 2 L^{\frac{n}{2}} L 2 n is a local differential operator and preserves supports. Since the supports of v i v_i v i and v j v_j v j are separated, we have B n ( v i , v j ) = ⟨ v i , L n 2 v j ⟩ = 0 B_n(v_i,v_j)=\langle v_i,L^{\frac{n}{2}}v_j\rangle=0 B n ( v i , v j ) = ⟨ v i , L 2 n v j ⟩ = 0 for every i ≠ j i \ne j i = j . Suppose now that n n n is odd. Then L n 2 L^{\frac{n}{2}} L 2 n is a classical pseudodifferential operator of order n n n . Its Schwartz kernel is smooth away from the diagonal. Hence, on B ( x i , 2 r ) × B ( x j , 2 r ) B(x_i,2r) \times B(x_j,2r) B ( x i , 2 r ) × B ( x j , 2 r ) , it is uniformly bounded by a constant C = C ( n , R ) C=C(n,R) C = C ( n , R ) , which further implies that
∣ B n ( v i , v j ) ∣ ≤ C ( n , R ) ∥ v i ∥ L 1 ( S n ) ∥ v j ∥ L 1 ( S n ) . \lvert B_n(v_i,v_j)\rvert\le C(n,R)\lVert v_i\rVert_{L^1(S^n)}\lVert v_j\rVert_{L^1(S^n)}. ∣ B n ( v i , v j )∣ ≤ C ( n , R ) ∥ v i ∥ L 1 ( S n ) ∥ v j ∥ L 1 ( S n ) .
Hölder’s inequality now gives (65) .
We next prove (66) . Choose ψ i ∈ C c ∞ ( B ( x i , 2 r ) ) \psi_i \in C^\infty_c(B(x_i,2r)) ψ i ∈ C c ∞ ( B ( x i , 2 r )) with ∫ S n ψ i d w = 1 \int_{S^n}\psi_i\,dw=1 ∫ S n ψ i d w = 1 and put c i : = B n ( h , ψ i ) c_i := B_n(h,\psi_i) c i := B n ( h , ψ i ) . For any φ ∈ C c ∞ ( B ( x i , 2 r ) ) \varphi\in C^\infty_c(B(x_i,2r)) φ ∈ C c ∞ ( B ( x i , 2 r )) , define
φ 0 : = φ − ( ∫ S n φ d w ) ψ i . \varphi_0 := \varphi- \left(\int_{S^n}\varphi\,dw\right)\psi_i. φ 0 := φ − ( ∫ S n φ d w ) ψ i .
Then φ 0 ∈ X i \varphi_0 \in X_i φ 0 ∈ X i and we have
0 = B n ( h , φ 0 ) = B n ( h , φ ) − c i ∫ S n φ d w . 0 = B_n(h,\varphi_0) = B_n(h,\varphi) - c_i\int_{S^n}\varphi\,dw. 0 = B n ( h , φ 0 ) = B n ( h , φ ) − c i ∫ S n φ d w .
Thus, L n 2 h = c i L^{\frac{n}{2}}h=c_i L 2 n h = c i weakly in B ( x i , 2 r ) B(x_i,2r) B ( x i , 2 r ) . A standard interior estimate for elliptic pseudodifferential operators (see, for instance, [37 Chapter XVIII] or [48 Theorem 7.2] ) yields
∥ h ∥ H n ( B ( x i , r ) ) ≤ C ( n , r ) ( ∥ h ∥ L 2 ( S n ) + ∣ c i ∣ ) . \lVert h\rVert_{H^n(B(x_i,r))} \le C(n,r)\left(\lVert h\rVert_{L^2(S^n)}+\lvert c_i\rvert\right). ∥ h ∥ H n ( B ( x i , r )) ≤ C ( n , r ) ( ∥ h ∥ L 2 ( S n ) + ∣ c i ∣ ) .
Recall that the Poincaré inequality gives ∥ h ∥ L 2 ( S n ) 2 ≤ C ( n ) E n ( h ) \lVert h\rVert_{L^2(S^n)}^2 \le C(n)E_n(h) ∥ h ∥ L 2 ( S n ) 2 ≤ C ( n ) E n ( h ) . Moreover, the Cauchy–Schwarz inequality gives
∣ c i ∣ = ∣ B n ( h , ψ i ) ∣ ≤ E n ( h ) 1 2 E n ( ψ i ) 1 2 . \lvert c_i\rvert= \lvert B_n(h,\psi_i)\rvert\le E_n(h)^{\frac{1}{2}}E_n(\psi_i)^{\frac{1}{2}}. ∣ c i ∣ = ∣ B n ( h , ψ i )∣ ≤ E n ( h ) 2 1 E n ( ψ i ) 2 1 .
Choosing ψ i \psi_i ψ i to depend only on n n n and r r r , and then applying the embedding H n ↪ L ∞ H^n \hookrightarrow L^\infty H n ↪ L ∞ , we obtain (66) .
We are now ready to prove Theorem 4.1.
Proof of Theorem 4.1. Throughout the proof, C C C denotes a positive constant depending only on m m m and n n n , whose value may vary from line to line. Define the measure
d σ = e n u ∫ S n e n u d w d w . d\sigma= \frac{e^{nu}}{\int_{S^n}e^{nu}\,dw}\,dw. d σ = ∫ S n e n u d w e n u d w .
on S n S^n S n . The assumptions on u u u imply that σ \sigma σ satisfies the hypotheses of Lemma 4.3. We may therefore apply Lemmas 4.3 and 4.5, retaining their notation. Choose 0 < r = r ( m , n ) < R 100 0<r=r(m,n)<\frac{R}{100} 0 < r = r ( m , n ) < 100 R , to be fixed sufficiently small below, and write u = ∑ i = 1 N ∗ v i + h u=\sum_{i=1}^{N_*}v_i+h u = ∑ i = 1 N ∗ v i + h as in the above decomposition. Lemma 4.3 gives
log ∫ S n e n u d w ≤ log ∫ B ( x i , r ) e n u d w − log δ r . \log\int_{S^n}e^{nu}\,dw \le\log\int_{B(x_i,r)}e^{nu}\,dw-\log\delta_r. log ∫ S n e n u d w ≤ log ∫ B ( x i , r ) e n u d w − log δ r .
As u = v i + h u=v_i+h u = v i + h on B ( x i , r ) B(x_i,r) B ( x i , r ) , applying (64) with κ = 1 \kappa=1 κ = 1 and (66) yields
log ∫ S n e n u d w ≤ log ∫ S n e n v i d w + n ∥ h ∥ L ∞ ( B ( x i , r ) ) − log δ r ≤ n 2 ( n − 1 ) ! ( E n ( v i ) − ∥ v i ∥ L 2 ( S n ) 2 ) + C E n ( h ) 1 2 + C . \begin{aligned}
\log\int_{\mathbb{S}^n}e^{nu}\,dw &\leq\log\int_{\mathbb{S}^n}e^{nv_i}\,dw+n\|h\|_{L^\infty(B(x_i,r))}-\log\delta_r\\
&\leq\frac{n}{2(n-1)!}\left(E_n(v_i)-\|v_i\|_{L^2(\mathbb{S}^n)}^2\right)+CE_n(h)^{\frac{1}{2}}+C.
\end{aligned} log ∫ S n e n u d w ≤ log ∫ S n e n v i d w + n ∥ h ∥ L ∞ ( B ( x i , r )) − log δ r ≤ 2 ( n − 1 )! n ( E n ( v i ) − ∥ v i ∥ L 2 ( S n ) 2 ) + C E n ( h ) 2 1 + C .
Averaging this estimate over 1 ≤ i ≤ N ∗ 1\leq i\leq N_* 1 ≤ i ≤ N ∗ gives
log ∫ S n e n u d w ≤ A m , n ∑ i = 1 N ∗ ( E n ( v i ) − ∥ v i ∥ L 2 ( S n ) 2 ) + C E n ( h ) 1 2 + C = A m , n ( E n ( u ) − ∑ i = 1 N ∗ ∥ v i ∥ L 2 ( S n ) 2 − ∑ i ≠ j B n ( v i , v j ) − E n ( h ) ) + C E n ( h ) 1 2 + C . \begin{aligned}
\log\int_{\mathbb{S}^n}e^{nu}\,dw &\leq A_{m,n}\sum_{i=1}^{N_*}\left(E_n(v_i)-\|v_i\|_{L^2(\mathbb{S}^n)}^2\right)+CE_n(h)^{\frac{1}{2}}+C\\
&=A_{m,n}\left(E_n(u)-\sum_{i=1}^{N_*}\|v_i\|_{L^2(\mathbb{S}^n)}^2-\sum_{i\ne j}B_n(v_i,v_j)-E_n(h)\right)+CE_n(h)^{\frac{1}{2}}+C.
\end{aligned} log ∫ S n e n u d w ≤ A m , n i = 1 ∑ N ∗ ( E n ( v i ) − ∥ v i ∥ L 2 ( S n ) 2 ) + C E n ( h ) 2 1 + C = A m , n E n ( u ) − i = 1 ∑ N ∗ ∥ v i ∥ L 2 ( S n ) 2 − i = j ∑ B n ( v i , v j ) − E n ( h ) + C E n ( h ) 2 1 + C .
By choosing r = r ( m , n ) r=r(m,n) r = r ( m , n ) sufficiently small and combining (65) with Hölder’s inequality, we obtain
∑ i ≠ j ∣ B n ( v i , v j ) ∣ ≤ ∑ i = 1 N ∗ ∥ v i ∥ L 2 ( S n ) 2 . \sum_{i\ne j}|B_n(v_i,v_j)|\leq\sum_{i=1}^{N_*}\|v_i\|_{L^2(\mathbb{S}^n)}^2. i = j ∑ ∣ B n ( v i , v j ) ∣ ≤ i = 1 ∑ N ∗ ∥ v i ∥ L 2 ( S n ) 2 .
Therefore,
log ∫ S n e n u d w ≤ A m , n ( E n ( u ) − E n ( h ) ) + C E n ( h ) 1 2 + C = A m , n E n ( u ) − A m , n E n ( h ) + C E n ( h ) 1 2 + C ≤ A m , n E n ( u ) + C . \begin{aligned}
\log\int_{\mathbb{S}^n}e^{nu}\,dw&\leq A_{m,n}\left(E_n(u)-E_n(h)\right)+CE_n(h)^{\frac{1}{2}}+C\\
&=A_{m,n}E_n(u)-A_{m,n}E_n(h)+CE_n(h)^{\frac{1}{2}}+C\\
&\leq A_{m,n}E_n(u)+C.
\end{aligned} log ∫ S n e n u d w ≤ A m , n ( E n ( u ) − E n ( h ) ) + C E n ( h ) 2 1 + C = A m , n E n ( u ) − A m , n E n ( h ) + C E n ( h ) 2 1 + C ≤ A m , n E n ( u ) + C .
The proof is complete.
The best remainder
In this subsection, we set
C b e s t ( m , n ) : = sup u { log ∫ S n e n u d w − A m , n E n ( u ) } . C_{\mathrm{best}}(m,n):=\sup_u\left\{\log\int_{\mathbb{S}^n}e^{nu}\,dw-A_{m,n}E_n(u)\right\}. C best ( m , n ) := u sup { log ∫ S n e n u d w − A m , n E n ( u ) } .
where the supremum is taken over all u ∈ H n 2 ( S n ) u\in H^{\frac n2}(\mathbb{S}^n) u ∈ H 2 n ( S n ) satisfying ∫ S n u d w = 0 \int_{\mathbb{S}^n}u\,dw=0 ∫ S n u d w = 0 and (17) . This definition is equivalent to the one used in Theorem 1.5.
Our main result is the following, which is slightly stronger than Theorem 1.5.
Theorem 4.6. Assume m ≥ 1 m\geq1 m ≥ 1 and n ≥ 2 n\geq2 n ≥ 2 . If
[ ( m + 1 ) ( m + n ) ] n 2 < N ∗ n ! , (67) [(m+1)(m+n)]^{\frac n2}<N_*n!, \tag*{(67)} [( m + 1 ) ( m + n ) ] 2 n < N ∗ n ! , (67) then C b e s t ( m , n ) > 0 C_{\mathrm{best}}(m,n)>0 C best ( m , n ) > 0 . In particular, (67) holds in the following four cases: m = 1 , n ≥ 9 m=1,n\geq9 m = 1 , n ≥ 9 ; m = 2 , n ≥ 12 m=2,n\geq12 m = 2 , n ≥ 12 ; m = 3 , n ≥ 18 m=3,n\geq18 m = 3 , n ≥ 18 ; m ≥ 4 , n ≥ 18 ( m + 1 ) m\geq4,n\geq18(m+1) m ≥ 4 , n ≥ 18 ( m + 1 ) . Moreover, C b e s t ( m , n ) > 0 C_{\mathrm{best}}(m,n)>0 C best ( m , n ) > 0 when m = 1 m=1 m = 1 and 5 ≤ n ≤ 8 5\leq n\leq8 5 ≤ n ≤ 8 .
The proof of Theorem 4.6 has two parts. In the first, we derive (67) by constructing test functions that are suitable perturbations of constant functions. In the second, we treat the case m = 1 m=1 m = 1 and 5 ≤ n ≤ 8 5 \le n \le8 5 ≤ n ≤ 8 by constructing explicit test functions.
Proof of the first part. Let Y Y Y be a nonzero function in H m + 1 \mathcal{H}_{m+1} H m + 1 . For ∣ t ∣ |t| ∣ t ∣ sufficiently small, define the smooth function u t = 1 n log ( 1 + t Y ) − 1 n ∫ S n log ( 1 + t Y ) d w u_t=\frac{1}{n}\log(1+tY)-\frac{1}{n}\int_{\mathbb{S}^n}\log(1+tY)\,dw u t = n 1 log ( 1 + t Y ) − n 1 ∫ S n log ( 1 + t Y ) d w . It is straightforward to verify that ∫ S n u t d w = 0 \int_{\mathbb{S}^n}u_t\,dw=0 ∫ S n u t d w = 0 and that u t u_t u t satisfies (17) . Thus, u t u_t u t is an admissible test function. A Taylor expansion gives
log ∫ S n e n u t d w = t 2 2 ∥ Y ∥ L 2 ( S n ) 2 + O ( t 3 ) . \log\int_{\mathbb{S}^n}e^{nu_t}\,dw=\frac{t^2}{2}\|Y\|_{L^2(\mathbb{S}^n)}^2+O(t^3). log ∫ S n e n u t d w = 2 t 2 ∥ Y ∥ L 2 ( S n ) 2 + O ( t 3 ) .
Moreover, since u t = t n Y + O ( t 2 ) u_t=\frac{t}{n}Y+O(t^2) u t = n t Y + O ( t 2 ) and Y ∈ H m + 1 Y\in\mathcal{H}_{m+1} Y ∈ H m + 1 , we have
E n ( u t ) = t 2 n 2 E n ( Y ) + O ( t 3 ) = t 2 n 2 λ m + 1 n 2 ∥ Y ∥ L 2 ( S n ) 2 + O ( t 3 ) . E_n(u_t)=\frac{t^2}{n^2}E_n(Y)+O(t^3)=\frac{t^2}{n^2}\lambda_{m+1}^{\frac{n}{2}}\|Y\|_{L^2(\mathbb{S}^n)}^2+O(t^3). E n ( u t ) = n 2 t 2 E n ( Y ) + O ( t 3 ) = n 2 t 2 λ m + 1 2 n ∥ Y ∥ L 2 ( S n ) 2 + O ( t 3 ) .
Combining these two estimates and recalling that λ m + 1 = ( m + 1 ) ( m + n ) \lambda_{m+1}=(m+1)(m+n) λ m + 1 = ( m + 1 ) ( m + n ) , we obtain
log ∫ S n e n u t d w − A m , n E n ( u t ) = ( 1 2 − A m , n n 2 [ ( m + 1 ) ( m + n ) ] n 2 ) t 2 ∥ Y ∥ L 2 ( S n ) 2 + O ( t 3 ) = ( N ∗ n ! − [ ( m + 1 ) ( m + n ) ] n 2 A m , n n 2 ) t 2 ∥ Y ∥ L 2 ( S n ) 2 + O ( t 3 ) . \begin{aligned}
\log\int_{\mathbb{S}^n}e^{nu_t}\,dw-A_{m,n}E_n(u_t)=\left(\frac{1}{2}-\frac{A_{m,n}}{n^2}[(m+1)(m+n)]^{\frac{n}{2}}\right)t^2\|Y\|_{L^2(\mathbb{S}^n)}^2+O(t^3) \\
&=\left(N_*n!-[(m+1)(m+n)]^{\frac{n}{2}}\frac{A_{m,n}}{n^2}\right)t^2\|Y\|_{L^2(\mathbb{S}^n)}^2+O(t^3).
\end{aligned} log ∫ S n e n u t d w − A m , n E n ( u t ) = ( 2 1 − n 2 A m , n [( m + 1 ) ( m + n ) ] 2 n ) t 2 ∥ Y ∥ L 2 ( S n ) 2 + O ( t 3 ) = ( N ∗ n ! − [( m + 1 ) ( m + n ) ] 2 n n 2 A m , n ) t 2 ∥ Y ∥ L 2 ( S n ) 2 + O ( t 3 ) .
Therefore, (67) implies that C b e s t ( m , n ) > 0 C_{\mathrm{best}}(m,n)>0 C best ( m , n ) > 0 .
Finally, since Θ ( 1 , 0 , n ) = 2 \Theta(1,0,n)=2 Θ ( 1 , 0 , n ) = 2 , Θ ( 2 , 0 , n ) = n + 2 \Theta(2,0,n)=n+2 Θ ( 2 , 0 , n ) = n + 2 , and Θ ( 3 , 0 , n ) = 2 n + 2 \Theta(3,0,n)=2n+2 Θ ( 3 , 0 , n ) = 2 n + 2 , substituting these explicit values into (67) and performing elementary calculations shows that (67) holds in the following three cases: m = 1 m=1 m = 1 , n ≥ 9 n\ge9 n ≥ 9 ; m = 2 m=2 m = 2 , n ≥ 12 n\ge12 n ≥ 12 ; and m = 3 m=3 m = 3 , n ≥ 18 n\ge18 n ≥ 18 . For m ≥ 4 m\ge4 m ≥ 4 and n ≥ 18 ( m + 1 ) n\ge18(m+1) n ≥ 18 ( m + 1 ) , we have
n ! > n n e n > n n 3 n ≥ [ 2 ( m + 1 ) n ] n 2 ≥ [ ( m + 1 ) ( n + m ) ] n 2 , n!>\frac{n^n}{e^n}>\frac{n^n}{3^n}\ge[2(m+1)n]^{\frac{n}{2}}\ge[(m+1)(n+m)]^{\frac{n}{2}}, n ! > e n n n > 3 n n n ≥ [ 2 ( m + 1 ) n ] 2 n ≥ [( m + 1 ) ( n + m ) ] 2 n ,
which also implies that (67) holds. □
Proof of the second part. Put z = x n + 1 z=x_{n+1} z = x n + 1 and α = n − 1 2 \alpha=\frac{n-1}{2} α = 2 n − 1 , and let C k α C_k^\alpha C k α be the standard Gegenbauer polynomial, normalized by
C 0 α = 1 , C 1 α = 2 α z , C_0^\alpha=1,\qquad C_1^\alpha=2\alpha z, C 0 α = 1 , C 1 α = 2 α z ,
and
( k + 1 ) C k + 1 α = 2 ( k + α ) z C k α − ( k + 2 α − 1 ) C k − 1 α . (k+1)C_{k+1}^\alpha=2(k+\alpha)zC_k^\alpha-(k+2\alpha-1)C_{k-1}^\alpha. ( k + 1 ) C k + 1 α = 2 ( k + α ) z C k α − ( k + 2 α − 1 ) C k − 1 α .
We define
u 5 = 1789 10000 C 2 2 + 41 2000 C 4 2 + 3 625 C 6 2 + 1 625 C 8 2 + 3 5000 C 10 2 + 1 5000 C 12 2 = 6656 625 z 12 − 12672 625 z 10 + 9504 625 z 8 − 2912 625 z 6 + 169 125 z 4 + 3681 2500 z 2 − 3097 10000 , \begin{aligned}
u_5=\frac{1789}{10000}C_2^2+\frac{41}{2000}C_4^2+\frac{3}{625}C_6^2+\frac{1}{625}C_8^2+\frac{3}{5000}C_{10}^2+\frac{1}{5000}C_{12}^2 \\
&=\frac{6656}{625}z^{12}-\frac{12672}{625}z^{10}+\frac{9504}{625}z^8-\frac{2912}{625}z^6+\frac{169}{125}z^4+\frac{3681}{2500}z^2-\frac{3097}{10000},
\end{aligned} u 5 = 10000 1789 C 2 2 + 2000 41 C 4 2 + 625 3 C 6 2 + 625 1 C 8 2 + 5000 3 C 10 2 + 5000 1 C 12 2 = 625 6656 z 12 − 625 12672 z 10 + 625 9504 z 8 − 625 2912 z 6 + 125 169 z 4 + 2500 3681 z 2 − 10000 3097 ,
u 6 = 6 5 z 4 + 8 5 z 2 − 2 7 = 248 1925 C 2 5 / 2 + 16 1925 C 4 5 / 2 , u_6=\frac{6}{5}z^4+\frac{8}{5}z^2-\frac{2}{7}=\frac{248}{1925}C_2^{5/2}+\frac{16}{1925}C_4^{5/2}, u 6 = 5 6 z 4 + 5 8 z 2 − 7 2 = 1925 248 C 2 5/2 + 1925 16 C 4 5/2 ,
u 7 = 3 ( z 2 − 1 8 ) , u 8 = 4 ( z 2 − 1 9 ) . u_7=3\left(z^2-\frac{1}{8}\right),\qquad u_8=4\left(z^2-\frac{1}{9}\right). u 7 = 3 ( z 2 − 8 1 ) , u 8 = 4 ( z 2 − 9 1 ) .
We refer to [51 ] for background on Gegenbauer polynomials. Since C k α ( z ) ∈ H k C_k^\alpha(z)\in\mathcal{H}_k C k α ( z ) ∈ H k , it is straightforward to verify that u n u_n u n satisfies ∫ S n u n d w = 0 \int_{\mathbb{S}^n}u_n\,\mathrm{d}w=0 ∫ S n u n d w = 0 and (17) for every 5 ≤ n ≤ 8 5\le n\le8 5 ≤ n ≤ 8 . Moreover, direct computations give
A 1 , 5 E 5 ( u 5 ) < 225 32 , A 1 , 6 E 6 ( u 6 ) = 17504 3003 , A 1 , 7 E 7 ( u 7 ) = 196 25 , A 1 , 8 E 8 ( u 8 ) = 4608 385 . A_{1,5}E_5(u_5)<\frac{225}{32},\qquad A_{1,6}E_6(u_6)=\frac{17504}{3003},\qquad A_{1,7}E_7(u_7)=\frac{196}{25},\qquad A_{1,8}E_8(u_8)=\frac{4608}{385}. A 1 , 5 E 5 ( u 5 ) < 32 225 , A 1 , 6 E 6 ( u 6 ) = 3003 17504 , A 1 , 7 E 7 ( u 7 ) = 25 196 , A 1 , 8 E 8 ( u 8 ) = 385 4608 .
To obtain lower bounds for ∫ S n e n u n d w \int_{\mathbb{S}^n}e^{nu_n}\,\mathrm{d}w ∫ S n e n u n d w , we use the elementary estimate
T N ( x ) : = ∑ j = 0 N x j j ! ≤ e x , ∀ x ∈ R , ∀ N odd . T_N(x):=\sum_{j=0}^{N}\frac{x^j}{j!}\le e^x,\qquad\forall x\in\mathbb{R},\quad\forall N\ \text{odd}. T N ( x ) := j = 0 ∑ N j ! x j ≤ e x , ∀ x ∈ R , ∀ N odd .
Set I n , N ( u ) = ∫ S n T N ( n u ) d w I_{n,N}(u)=\int_{\mathbb{S}^n}T_N(nu)\,\mathrm{d}w I n , N ( u ) = ∫ S n T N ( n u ) d w . Direct computations yield
I 5 , 19 ( u 5 ) ≥ 1200 , I 6 , 13 ( u 6 ) ≥ 400 , I 7 , 13 ( u 7 ) ≥ 3000 , I 8 , 13 ( u 8 ) ≥ 200000. I_{5,19}(u_5)\ge1200,\qquad I_{6,13}(u_6)\ge400,\qquad I_{7,13}(u_7)\ge3000,\qquad I_{8,13}(u_8)\ge200000. I 5 , 19 ( u 5 ) ≥ 1200 , I 6 , 13 ( u 6 ) ≥ 400 , I 7 , 13 ( u 7 ) ≥ 3000 , I 8 , 13 ( u 8 ) ≥ 200000.
Since
e 225 32 < 1200 , e 17504 3003 < 400 , e 196 25 < 3000 , e 4608 385 < 200000 , e^{\frac{225}{32}}<1200,\qquad e^{\frac{17504}{3003}}<400,\qquad e^{\frac{196}{25}}<3000,\qquad e^{\frac{4608}{385}}<200000, e 32 225 < 1200 , e 3003 17504 < 400 , e 25 196 < 3000 , e 385 4608 < 200000 ,
we conclude that C b e s t ( 1 , n ) > 0 C_{\mathrm{best}}(1,n)>0 C best ( 1 , n ) > 0 for 5 ≤ n ≤ 8 5\le n\le8 5 ≤ n ≤ 8 .
Quantitative Sobolev remainder estimates
Throughout this section, we fix m , n ∈ N + m,n\in\mathbb{N}_+ m , n ∈ N + and 1 ≤ p < n 1\le p<n 1 ≤ p < n , and set p ∗ = n p n − p p^*=\frac{np}{n-p} p ∗ = n − p n p and θ = p p ∗ = n − p n \theta=\frac{p}{p^*}=\frac{n-p}{n} θ = p ∗ p = n n − p . For notational simplicity, we write
Θ = Θ ( m , θ , n ) . \Theta=\Theta(m,\theta,n). Θ = Θ ( m , θ , n ) .
Our main result is the following, which is stronger than Theorem 1.6 .
Theorem 5.1. Assume that the sharp Sobolev inequality (6) holds on S n \mathbb{S}^n S n . Then there exists a positive constant B = B ( m , n , p ) B=B(m,n,p) B = B ( m , n , p ) such that, for every $0<\varepsilon<1$, every linear functional T : P ˚ m , n → R T:\mathring{\mathcal{P}}_{m,n}\to\mathbb{R} T : P ˚ m , n → R , and every $u\in W^{1,p}(\mathbb{S}^n) satisfying
∣ ∫ S n f ∣ u ∣ p ∗ d μ ∣ ≤ ∣ T ( f ) ∣ ∥ u ∥ L p ∗ ( S n ) p ∗ for all f ∈ P ˚ m , n , (68) \left|\int_{\mathbb{S}^n}f|u|^{p^*}\,\mathrm{d}\mu\right|\le|T(f)|\|u\|_{L^{p^*}(\mathbb{S}^n)}^{p^*}\qquad\text{for all }f\in\mathring{\mathcal{P}}_{m,n}, \tag*{(68)} ∫ S n f ∣ u ∣ p ∗ d μ ≤ ∣ T ( f ) ∣∥ u ∥ L p ∗ ( S n ) p ∗ for all f ∈ P ˚ m , n , (68) one has
∥ u ∥ L p ∗ ( S n ) p ≤ ( S n , p p Θ ( m , n − p n , n ) + ε ) ∥ ∇ u ∥ L p ( S n ) p + B ( ε − γ ( p ) + ∥ T ∥ ∞ θ ) ∥ u ∥ L p ( S n ) p , (69) \|u\|_{L^{p^*}(\mathbb{S}^n)}^p\le\left(\frac{S_{n,p}^p}{\Theta\left(m,\frac{n-p}{n},n\right)}+\varepsilon\right)\|\nabla u\|_{L^p(\mathbb{S}^n)}^p+B\left(\varepsilon^{-\gamma(p)}+\|T\|_\infty^\theta\right)\|u\|_{L^p(\mathbb{S}^n)}^p, \tag*{(69)} ∥ u ∥ L p ∗ ( S n ) p ≤ ( Θ ( m , n n − p , n ) S n , p p + ε ) ∥∇ u ∥ L p ( S n ) p + B ( ε − γ ( p ) + ∥ T ∥ ∞ θ ) ∥ u ∥ L p ( S n ) p , (69)
where γ ( p ) \gamma(p) γ ( p ) is defined in (22) , and
∥ T ∥ ∞ = sup f ∈ P ˚ m , n ∥ f ∥ L ∞ ( S n ) ≤ 1 ∣ T ( f ) ∣ . (70) \lVert T\rVert_{\infty}=\sup_{\substack{f\in\mathring{\mathcal{P}}_{m,n}\\ \lVert f\rVert_{L^{\infty}(\mathbb{S}^{n})}\leq1}}|T(f)|. \tag*{(70)} ∥ T ∥ ∞ = f ∈ P ˚ m , n ∥ f ∥ L ∞ ( S n ) ≤ 1 sup ∣ T ( f ) ∣. (70)
The proof of Theorem 5.1 requires several quantitative estimates. Let d w = d μ ω n \mathrm{d}w=\frac{\mathrm{d}\mu}{\omega_n} d w = ω n d μ denote the normalized surface measure on S n \mathbb{S}^{n} S n . We set
D = dim ( P m , n ∣ S n ) = ( n + m n ) + ( n + m − 1 n ) , d = D − 1. (71) D=\dim(\mathcal{P}_{m,n}|_{\mathbb{S}^{n}})=\binom{n+m}{n}+\binom{n+m-1}{n},\qquad d=D-1. \tag*{(71)} D = dim ( P m , n ∣ S n ) = ( n n + m ) + ( n n + m − 1 ) , d = D − 1. (71)
Choose an orthonormal basis f 1 , … , f d f_1,\ldots,f_d f 1 , … , f d of P ˚ m , n \mathring{\mathcal{P}}_{m,n} P ˚ m , n in L 2 ( d w ) L^2(\mathrm{d}w) L 2 ( d w ) and define F = ( f 1 , … , f d ) F=(f_1,\ldots,f_d) F = ( f 1 , … , f d ) . The rotational invariance of P ˚ m , n \mathring{\mathcal{P}}_{m,n} P ˚ m , n shows that ∣ F ( x ) ∣ 2 |F(x)|^2 ∣ F ( x ) ∣ 2 is constant on S n \mathbb{S}^{n} S n . Consequently,
∣ F ( x ) ∣ 2 = d , ∫ S n F d w = 0 , ∫ S n F F T d w = Id . (72) |F(x)|^2=d,\qquad\int_{\mathbb{S}^{n}}F\,\mathrm{d}w=0,\qquad\int_{\mathbb{S}^{n}}FF^{\mathsf{T}}\,\mathrm{d}w=\operatorname{Id}. \tag*{(72)} ∣ F ( x ) ∣ 2 = d , ∫ S n F d w = 0 , ∫ S n F F T d w = Id . (72)
For a probability measure σ \sigma σ on S n \mathbb{S}^{n} S n , define
τ ( σ ) = sup f ∈ P ˚ m , n ∥ f ∥ L ∞ ( S n ) ≤ 1 ∣ ∫ S n f d σ ∣ . (73) \tau(\sigma)=\sup_{\substack{f\in\mathring{\mathcal{P}}_{m,n}\\ \lVert f\rVert_{L^{\infty}(\mathbb{S}^{n})}\leq1}}\left|\int_{\mathbb{S}^{n}}f\,\mathrm{d}\sigma\right|. \tag*{(73)} τ ( σ ) = f ∈ P ˚ m , n ∥ f ∥ L ∞ ( S n ) ≤ 1 sup ∫ S n f d σ . (73)
If e = ∫ S n F d σ e=\int_{\mathbb{S}^{n}}F\,\mathrm{d}\sigma e = ∫ S n F d σ , then we have
∣ e ∣ ≤ d τ ( σ ) . (74) |e|\leq\sqrt{d}\,\tau(\sigma). \tag*{(74)} ∣ e ∣ ≤ d τ ( σ ) . (74)
The first lemma gives a lower bound for ∥ ⋅ ∥ θ \lVert\cdot\rVert_{\theta} ∥ ⋅ ∥ θ on general discrete measures.
Lemma 5.2. For every finitely supported probability measure ν = ∑ i ν i δ x i \nu=\sum_i\nu_i\delta_{x_i} ν = ∑ i ν i δ x i on S n \mathbb{S}^{n} S n ,
∑ i ν i θ ≥ Θ − D 1 − θ d θ τ ( ν ) θ . (75) \sum_i\nu_i^{\theta}\geq\Theta-D^{1-\theta}d^{\theta}\tau(\nu)^{\theta}. \tag*{(75)} i ∑ ν i θ ≥ Θ − D 1 − θ d θ τ ( ν ) θ . (75) Proof. Let e = ∫ S n F d ν e=\int_{\mathbb{S}^{n}}F\,\mathrm{d}\nu e = ∫ S n F d ν . If e = 0 e=0 e = 0 , then (75) follows directly from the definition of Θ \Theta Θ in (14) . Suppose that e ≠ 0 e\neq0 e = 0 . We correct ν \nu ν by adding a suitable measure. Consider the probability measure
d ρ ( x ) = ( 1 − e ⋅ F ( x ) d ∣ e ∣ ) d w ( x ) . \mathrm{d}\rho(x)=\left(1-\frac{e\cdot F(x)}{\sqrt{d}|e|}\right)\mathrm{d}w(x). d ρ ( x ) = ( 1 − d ∣ e ∣ e ⋅ F ( x ) ) d w ( x ) . The density is nonnegative by (72) , and the same identities give ∫ S n F d ρ = − e d ∣ e ∣ \int_{\mathbb{S}^{n}}F\,\mathrm{d}\rho=-\frac{e}{\sqrt{d}|e|} ∫ S n F d ρ = − d ∣ e ∣ e . By Carathéodory’s theorem, there is a probability measure ρ a \rho_a ρ a supported on at most d + 1 = D d+1=D d + 1 = D points and satisfying ∫ S n F d ρ a = ∫ S n F d ρ \int_{\mathbb{S}^{n}}F\,\mathrm{d}\rho_a=\int_{\mathbb{S}^{n}}F\,\mathrm{d}\rho ∫ S n F d ρ a = ∫ S n F d ρ .
Set t = d ∣ e ∣ t=\sqrt{d}|e| t = d ∣ e ∣ and ν ^ = ( ν + t ρ a ) / ( 1 + t ) \widehat{\nu}=(\nu+t\rho_a)/(1+t) ν = ( ν + t ρ a ) / ( 1 + t ) . Then ν ^ ∈ M m , n c \widehat{\nu}\in\mathcal{M}^{c}_{m,n} ν ∈ M m , n c . Since ( a + b ) θ ≤ a θ + b θ (a+b)^{\theta}\leq a^{\theta}+b^{\theta} ( a + b ) θ ≤ a θ + b θ for a , b ≥ 0 a,b\geq0 a , b ≥ 0 , we have
Θ ≤ ∥ ν ^ ∥ θ ≤ ∥ ν ∥ θ + t θ ∥ ρ a ∥ θ ( 1 + t ) θ ≤ ∥ ν ∥ θ + D 1 − θ t θ . \Theta\leq\lVert\widehat{\nu}\rVert_{\theta}\leq\frac{\lVert\nu\rVert_{\theta}+t^{\theta}\lVert\rho_a\rVert_{\theta}}{(1+t)^{\theta}}\leq\lVert\nu\rVert_{\theta}+D^{1-\theta}t^{\theta}. Θ ≤ ∥ ν ∥ θ ≤ ( 1 + t ) θ ∥ ν ∥ θ + t θ ∥ ρ a ∥ θ ≤ ∥ ν ∥ θ + D 1 − θ t θ . By (74) , t ≤ d τ ( ν ) t\leq d\tau(\nu) t ≤ d τ ( ν ) , which proves the lemma.
Let q = θ − 1 = p ∗ p q=\theta^{-1}=\frac{p^*}{p} q = θ − 1 = p p ∗ . Given finitely many continuous functions a j ≥ 0 a_j\geq0 a j ≥ 0 with ∑ j a j = 1 \sum_j a_j=1 ∑ j a j = 1 , define (the specific choice of the a j a_j a j will be made in the proof of Theorem 5.1 )
Φ ( σ ) = ∑ j ( ∫ S n a j q d σ ) θ . \Phi(\sigma)=\sum_j\left(\int_{\mathbb{S}^n}a_j^q\,\mathrm{d}\sigma\right)^\theta. Φ ( σ ) = j ∑ ( ∫ S n a j q d σ ) θ .
The following lemma provides a finite-atom reduction principle.
Lemma 5.3. For every probability measure σ \sigma σ on S n \mathbb{S}^n S n , there is a probability measure ν = ∑ i = 1 K w i δ x i \nu=\sum_{i=1}^{K}w_i\delta_{x_i} ν = ∑ i = 1 K w i δ x i such that
K ≤ D , w i > 0 , ∫ S n F d ν = ∫ S n F d σ , Φ ( ν ) ≤ Φ ( σ ) . (76) K\leq D,\qquad w_i>0,\qquad\int_{\mathbb{S}^n}F\,\mathrm{d}\nu=\int_{\mathbb{S}^n}F\,\mathrm{d}\sigma,\qquad\Phi(\nu)\leq\Phi(\sigma). \tag*{(76)} K ≤ D , w i > 0 , ∫ S n F d ν = ∫ S n F d σ , Φ ( ν ) ≤ Φ ( σ ) . (76)
Proof. Apply Carathéodory’s theorem to the continuous map ( F , a 1 q , … , a J q ) (F,a_1^q,\ldots,a_J^q) ( F , a 1 q , … , a J q ) to replace σ \sigma σ by a finitely supported probability measure ν 0 \nu_0 ν 0 preserving the integrals of all its components. Suppose that the resulting measure has K > D = d + 1 K>D=d+1 K > D = d + 1 positive weights. There is then a nonzero vector ( h i ) i = 1 K (h_i)_{i=1}^{K} ( h i ) i = 1 K such that
∑ i h i = 0 , ∑ i h i F ( x i ) = 0. \sum_i h_i=0,\qquad\sum_i h_iF(x_i)=0. i ∑ h i = 0 , i ∑ h i F ( x i ) = 0.
The set of parameters s s s for which w i + s h i ≥ 0 w_i+sh_i\geq0 w i + s h i ≥ 0 for every i i i is a closed interval [ s − , s + ] [s_-,s_+] [ s − , s + ] with s − < 0 < s + s_-<0<s_+ s − < 0 < s + , and at each endpoint at least one positive weight vanishes. Moreover, the function
s ⟼ ∑ j ( ∑ i ( w i + s h i ) a j ( x i ) q ) θ s\longmapsto\sum_j\left(\sum_i(w_i+sh_i)a_j(x_i)^q\right)^\theta s ⟼ j ∑ ( i ∑ ( w i + s h i ) a j ( x i ) q ) θ
is concave. Hence its value at one of the two endpoints does not exceed its value at s = 0 s=0 s = 0 . Choosing that endpoint preserves the mass and the F F F -moment, does not increase Φ \Phi Φ , and decreases the number of positive weights. Iteration gives (76) .
Our third lemma constructs a suitable ℓ p \ell^p ℓ p -partition of unity.
Lemma 5.4. There exist positive constants c , C , r 1 c,C,r_1 c , C , r 1 , depending only on n n n and p p p , such that for every 0 < r < r 1 0<r<r_1 0 < r < r 1 there are finitely many nonnegative C 1 C^1 C 1 functions η j \eta_j η j supported in geodesic balls of radius r r r for which, with a j = η j p a_j=\eta_j^p a j = η j p ,
∑ j η j p = 1 , ∑ j ∣ ∇ η j ∣ p ≤ C r − p , (77) \sum_j\eta_j^p=1,\qquad\sum_j\lvert\nabla\eta_j\rvert^p\leq Cr^{-p}, \tag*{(77)} j ∑ η j p = 1 , j ∑ ∣ ∇ η j ∣ p ≤ C r − p , (77) and
∑ j ∣ a j ( x ) − a j ( y ) ∣ 2 ≥ c min { ∣ x − y ∣ 2 r 2 , 1 } for all x , y ∈ S n . (78) \sum_j\lvert a_j(x)-a_j(y)\rvert^2\geq c\min\left\{\frac{\lvert x-y\rvert^2}{r^2},1\right\}\qquad\text{for all }x,y\in\mathbb{S}^n. \tag*{(78)} j ∑ ∣ a j ( x ) − a j ( y ) ∣ 2 ≥ c min { r 2 ∣ x − y ∣ 2 , 1 } for all x , y ∈ S n . (78)
Proof. Take a maximal r 8 \frac{r}{8} 8 r -separated set { z ℓ } \{z_\ell\} { z ℓ } and choose nonnegative smooth cutoff functions ψ ℓ \psi_\ell ψ ℓ which are equal to one on B ( z ℓ , r 2 ) B(z_\ell,\frac{r}{2}) B ( z ℓ , 2 r ) , are supported in B ( z ℓ , r ) B(z_\ell,r) B ( z ℓ , r ) , and satisfy ∣ ∇ ψ ℓ ∣ ≤ C r \lvert\nabla\psi_\ell\rvert\leq\frac{C}{r} ∣ ∇ ψ ℓ ∣ ≤ r C . These balls have overlap bounded by a constant M = M ( n ) M = M(n) M = M ( n ) . Set
η ℓ 0 = ψ ℓ ( ∑ k ψ k p ) 1 / p , λ ℓ = ( η ℓ 0 ) p . \eta_\ell^0 = \frac{\psi_\ell}{\left(\sum_k \psi_k^p\right)^{1/p}}, \qquad\lambda_\ell= \left(\eta_\ell^0\right)^p. η ℓ 0 = ( ∑ k ψ k p ) 1/ p ψ ℓ , λ ℓ = ( η ℓ 0 ) p .
Then ∑ ℓ λ ℓ = 1 \sum_\ell\lambda_\ell= 1 ∑ ℓ λ ℓ = 1 , ∑ ℓ ∣ ∇ η ℓ 0 ∣ p ≤ C r − p \sum_\ell|\nabla\eta_\ell^0|^p \leq Cr^{-p} ∑ ℓ ∣∇ η ℓ 0 ∣ p ≤ C r − p , and for every x ∈ S n x \in S^n x ∈ S n there is an index ℓ \ell ℓ such that λ ℓ ( x ) ≥ M − 1 \lambda_\ell(x) \geq M^{-1} λ ℓ ( x ) ≥ M − 1 . To obtain (78) , we modify the functions η ℓ 0 \eta_\ell^0 η ℓ 0 as follows. Set N = n + 1 N = n + 1 N = n + 1 and choose a smooth function χ \chi χ such that χ ( t ) = t \chi(t) = t χ ( t ) = t for ∣ t ∣ ≤ 1 / 2 |t| \leq1/2 ∣ t ∣ ≤ 1/2 , ∣ χ ∣ ≤ 3 / 4 |\chi| \leq3/4 ∣ χ ∣ ≤ 3/4 , and ∣ χ ′ ∣ ≤ C |\chi'| \leq C ∣ χ ′ ∣ ≤ C . For 1 ≤ k ≤ N 1 \leq k \leq N 1 ≤ k ≤ N and each sign, let
h ℓ , k , ± ( x ) = 1 ± χ ( ( x k − z ℓ , k ) / ( 2 r ) ) 2 N , η ℓ , k , ± = η ℓ 0 h ℓ , k , ± 1 / p . h_{\ell,k,\pm}(x) = \frac{1 \pm\chi((x_k-z_{\ell,k})/(2r))}{2N}, \qquad\eta_{\ell,k,\pm} = \eta_\ell^0 h_{\ell,k,\pm}^{1/p}. h ℓ , k , ± ( x ) = 2 N 1 ± χ (( x k − z ℓ , k ) / ( 2 r )) , η ℓ , k , ± = η ℓ 0 h ℓ , k , ± 1/ p .
Since ∑ k , ± h ℓ , k , ± = 1 \sum_{k,\pm} h_{\ell,k,\pm} = 1 ∑ k , ± h ℓ , k , ± = 1 , h ℓ , k , ± ≥ ( 8 N ) − 1 h_{\ell,k,\pm} \geq(8N)^{-1} h ℓ , k , ± ≥ ( 8 N ) − 1 , and ∣ ∇ h ℓ , k , ± ∣ ≤ C r |\nabla h_{\ell,k,\pm}| \leq\frac{C}{r} ∣∇ h ℓ , k , ± ∣ ≤ r C , the functions η ℓ , k , ± \eta_{\ell,k,\pm} η ℓ , k , ± satisfy (77) . Denote their p p p -th powers by a ℓ , k , ± = λ ℓ h ℓ , k , ± a_{\ell,k,\pm} = \lambda_\ell h_{\ell,k,\pm} a ℓ , k , ± = λ ℓ h ℓ , k , ± .
It remains to establish (78) . Fix x , y ∈ S n x,y \in S^n x , y ∈ S n and choose ℓ \ell ℓ so that λ ℓ ( x ) ≥ M − 1 \lambda_\ell(x) \geq M^{-1} λ ℓ ( x ) ≥ M − 1 . Let Δ a \Delta a Δ a be the vector consisting of the 2 N 2N 2 N differences a ℓ , k , ± ( x ) − a ℓ , k , ± ( y ) a_{\ell,k,\pm}(x)-a_{\ell,k,\pm}(y) a ℓ , k , ± ( x ) − a ℓ , k , ± ( y ) . If λ ℓ ( y ) = 0 \lambda_\ell(y)=0 λ ℓ ( y ) = 0 , then the sum of these differences is λ ℓ ( x ) \lambda_\ell(x) λ ℓ ( x ) , and hence ∥ Δ a ∥ ℓ 2 ≥ ( M 2 N ) − 1 \|\Delta a\|_{\ell^2} \geq(M\sqrt{2N})^{-1} ∥Δ a ∥ ℓ 2 ≥ ( M 2 N ) − 1 . If λ ℓ ( y ) > 0 \lambda_\ell(y)>0 λ ℓ ( y ) > 0 , then x , y ∈ B ( z ℓ , r ) x,y \in B(z_\ell,r) x , y ∈ B ( z ℓ , r ) , so χ \chi χ is linear at all the arguments involved. With A = λ ℓ ( x ) − λ ℓ ( y ) A=\lambda_\ell(x)-\lambda_\ell(y) A = λ ℓ ( x ) − λ ℓ ( y ) , one has
A = ∑ k , ± Δ a k , ± , 2 N r ( Δ a k , + − Δ a k , − ) = λ ℓ ( x ) ( x k − y k ) + A ( y k − z ℓ , k ) . A = \sum_{k,\pm} \Delta a_{k,\pm}, \qquad2Nr(\Delta a_{k,+}-\Delta a_{k,-}) = \lambda_\ell(x)(x_k-y_k) + A(y_k-z_{\ell,k}). A = k , ± ∑ Δ a k , ± , 2 N r ( Δ a k , + − Δ a k , − ) = λ ℓ ( x ) ( x k − y k ) + A ( y k − z ℓ , k ) .
Since ∣ y − z ℓ ∣ ≤ r |y-z_\ell| \leq r ∣ y − z ℓ ∣ ≤ r , Hölder’s inequality gives λ ℓ ( x ) ∣ x − y ∣ ≤ C ( n ) r ∥ Δ a ∥ ℓ 2 \lambda_\ell(x)|x-y| \leq C(n)r\|\Delta a\|_{\ell^2} λ ℓ ( x ) ∣ x − y ∣ ≤ C ( n ) r ∥Δ a ∥ ℓ 2 . Combining the two cases and adding the nonnegative contributions corresponding to the other indices proves (78) . Relabeling the triples ( ℓ , k , ± ) (\ell,k,\pm) ( ℓ , k , ± ) as a single index concludes the proof. □
The fourth lemma provides an elementary but useful inequality.
Lemma 5.5. Let 0 ≤ X ≤ 1 0 \leq X \leq1 0 ≤ X ≤ 1 be a random variable on a probability space and let q > 1 q>1 q > 1 . Then
∥ X ∥ L q − E X ≥ c q Var ( X ) , c q = 1 2 min { q − 1 , 1 } . (79) \|X\|_{L^q} - \mathbb{E}X \geq c_q \operatorname{Var}(X), \qquad c_q = \frac{1}{2}\min\{q-1,1\}. \tag*{(79)} ∥ X ∥ L q − E X ≥ c q Var ( X ) , c q = 2 1 min { q − 1 , 1 } . (79) Proof. If q ≥ 2 q \geq2 q ≥ 2 , then ∥ X ∥ L q ≥ ∥ X ∥ L 2 \|X\|_{L^q} \geq\|X\|_{L^2} ∥ X ∥ L q ≥ ∥ X ∥ L 2 and
∥ X ∥ L 2 − E X = Var ( X ) ∥ X ∥ L 2 + E X ≥ 1 2 Var ( X ) . \|X\|_{L^2} - \mathbb{E}X = \frac{\operatorname{Var}(X)}{\|X\|_{L^2}+\mathbb{E}X} \geq\frac{1}{2}\operatorname{Var}(X). ∥ X ∥ L 2 − E X = ∥ X ∥ L 2 + E X Var ( X ) ≥ 2 1 Var ( X ) . If 1 < q < 2 1<q<2 1 < q < 2 , the second derivative of t ↦ t q t \mapsto t^q t ↦ t q on ( 0 , 1 ] (0,1] ( 0 , 1 ] is at least q ( q − 1 ) q(q-1) q ( q − 1 ) , so Taylor’s theorem gives
E X q ≥ ( E X ) q + q ( q − 1 ) 2 Var ( X ) . \mathbb{E}X^q \geq(\mathbb{E}X)^q + \frac{q(q-1)}{2}\operatorname{Var}(X). E X q ≥ ( E X ) q + 2 q ( q − 1 ) Var ( X ) . The elementary inequality ( a + b ) 1 / q ≥ a 1 / q + b q (a+b)^{1/q} \geq a^{1/q} + \frac{b}{q} ( a + b ) 1/ q ≥ a 1/ q + q b for a , b ≥ 0 a,b \geq0 a , b ≥ 0 and a + b ≤ 1 a+b \leq1 a + b ≤ 1 now yields (79) . □
The next lemma gives a useful error estimate comparing a measure with small support to a suitably chosen point mass.
Lemma 5.6. Let π \pi π be a probability measure supported in a spherical region of small diameter d 0 = d 0 ( m , n ) d_0=d_0(m,n) d 0 = d 0 ( m , n ) . Define
b = ∫ S n x d π ( x ) , z = b ∣ b ∣ , v = 1 − ∣ b ∣ 2 = 1 2 ∬ S n × S n ∣ x − y ∣ 2 d π ( x ) d π ( y ) . (80) b=\int_{\mathbb{S}^n}x\,\mathrm{d}\pi(x),\qquad z=\frac{b}{\lvert b\rvert},\qquad v=1-\lvert b\rvert^2=\frac{1}{2}\iint_{\mathbb{S}^n\times\mathbb{S}^n}\lvert x-y\rvert^2\,\mathrm{d}\pi(x)\,\mathrm{d}\pi(y). \tag*{(80)} b = ∫ S n x d π ( x ) , z = ∣ b ∣ b , v = 1 − ∣ b ∣ 2 = 2 1 ∬ S n × S n ∣ x − y ∣ 2 d π ( x ) d π ( y ) . (80)
Then, for every f ∈ P m , n f\in\mathcal{P}_{m,n} f ∈ P m , n ,
∣ ∫ S n f d π − f ( z ) ∣ ≤ C ( m , n ) ∥ f ∥ L ∞ ( S n ) v . (81) \left\lvert\int_{\mathbb{S}^n}f\,\mathrm{d}\pi-f(z)\right\rvert\leq C(m,n)\lVert f\rVert_{L^\infty(\mathbb{S}^n)}v. \tag*{(81)} ∫ S n f d π − f ( z ) ≤ C ( m , n ) ∥ f ∥ L ∞ ( S n ) v . (81)
Proof. Let C = C ( m , n ) C=C(m,n) C = C ( m , n ) be a positive constant that may vary from line to line. Since P m , n \mathcal{P}_{m,n} P m , n is finite-dimensional, we have
∥ f ∥ C 2 ( S n ) ≤ C ∥ f ∥ L ∞ ( S n ) . \lVert f\rVert_{C^2(\mathbb{S}^n)}\leq C\lVert f\rVert_{L^\infty(\mathbb{S}^n)}. ∥ f ∥ C 2 ( S n ) ≤ C ∥ f ∥ L ∞ ( S n ) .
Taking d 0 d_0 d 0 sufficiently small and using the Taylor expansion of f f f at z z z together with the definition of z z z , we obtain
∣ ∫ S n f d π − f ( z ) ∣ ≤ C ∥ f ∥ L ∞ ( S n ) ( ∣ ∫ S n ( x − z ) d π ( x ) ∣ + ∫ S n ∣ x − z ∣ 2 d π ( x ) ) ≤ C ∥ f ∥ L ∞ ( S n ) ( 1 − ∣ b ∣ ) ≤ C ∥ f ∥ L ∞ ( S n ) v . \begin{aligned}
\left\lvert\int_{\mathbb{S}^n}f\,\mathrm{d}\pi-f(z)\right\rvert\leq C\lVert f\rVert_{L^\infty(\mathbb{S}^n)}\left(\left\lvert\int_{\mathbb{S}^n}(x-z)\,\mathrm{d}\pi(x)\right\rvert+\int_{\mathbb{S}^n}\lvert x-z\rvert^2\,\mathrm{d}\pi(x)\right) \\
&\leq C\lVert f\rVert_{L^\infty(\mathbb{S}^n)}(1-\lvert b\rvert)\leq C\lVert f\rVert_{L^\infty(\mathbb{S}^n)}v.
\end{aligned} ∫ S n f d π − f ( z ) ≤ C ∥ f ∥ L ∞ ( S n ) ( ∫ S n ( x − z ) d π ( x ) + ∫ S n ∣ x − z ∣ 2 d π ( x ) ) ≤ C ∥ f ∥ L ∞ ( S n ) ( 1 − ∣ b ∣) ≤ C ∥ f ∥ L ∞ ( S n ) v .
The proof is complete.
For the partition in Lemma 5.4 , define
Φ r ( σ ) = ∑ j ( ∫ S n η j p ∗ d σ ) θ = ∑ j ( ∫ S n a j q d σ ) 1 / q . (82) \Phi_r(\sigma)=\sum_j\left(\int_{\mathbb{S}^n}\eta_j^{p^*}\,\mathrm{d}\sigma\right)^\theta=\sum_j\left(\int_{\mathbb{S}^n}a_j^q\,\mathrm{d}\sigma\right)^{1/q}. \tag*{(82)} Φ r ( σ ) = j ∑ ( ∫ S n η j p ∗ d σ ) θ = j ∑ ( ∫ S n a j q d σ ) 1/ q . (82)
Combining the five lemmas above, we obtain the following lower bound for Φ r ( σ ) \Phi_r(\sigma) Φ r ( σ ) , which is crucial for proving (69) .
Proposition 5.7. There exist r 0 = r 0 ( m , n , p ) > 0 r_0=r_0(m,n,p)>0 r 0 = r 0 ( m , n , p ) > 0 and C = C ( m , n , p ) > 0 C=C(m,n,p)>0 C = C ( m , n , p ) > 0 such that, for every 0 < r < r 0 0<r<r_0 0 < r < r 0 and every probability measure σ \sigma σ on S n \mathbb{S}^n S n ,
Φ r ( σ ) ≥ Θ − C r 2 θ 1 − θ − C τ ( σ ) θ . (83) \Phi_r(\sigma)\geq\Theta-Cr^{\frac{2\theta}{1-\theta}}-C\tau(\sigma)^\theta. \tag*{(83)} Φ r ( σ ) ≥ Θ − C r 1 − θ 2 θ − C τ ( σ ) θ . (83)
Proof. By Lemma 5.3 , there is a probability measure ν = ∑ i = 1 K w i δ x i \nu=\sum_{i=1}^{K}w_i\delta_{x_i} ν = ∑ i = 1 K w i δ x i with K ≤ D K\leq D K ≤ D , τ ( ν ) = τ ( σ ) \tau(\nu)=\tau(\sigma) τ ( ν ) = τ ( σ ) , and Φ r ( ν ) ≤ Φ r ( σ ) \Phi_r(\nu)\leq\Phi_r(\sigma) Φ r ( ν ) ≤ Φ r ( σ ) . Thus it suffices to establish a lower bound for Φ r ( ν ) \Phi_r(\nu) Φ r ( ν ) . Join two atoms x i , x j x_i,x_j x i , x j if d ( x i , x j ) ≤ 2 r d(x_i,x_j)\leq2r d ( x i , x j ) ≤ 2 r , and denote the connected components by I 1 , … , I L I_1,\ldots,I_L I 1 , … , I L . Each component has diameter at most 2 ( D − 1 ) r 2(D-1)r 2 ( D − 1 ) r . We choose r 0 < min { r 1 , 1 2 ( D − 1 ) } r_0<\min\{r_1,\frac{1}{2(D-1)}\} r 0 < min { r 1 , 2 ( D − 1 ) 1 } so that every component lies in the range of Lemma 5.6. Since each a j a_j a j is supported in a ball of radius r r r , it can be nonzero at atoms belonging to at most one component. For each component, set
A ℓ = ∑ i ∈ I ℓ w i , π ℓ = A ℓ − 1 ∑ i ∈ I ℓ w i δ x i , b ℓ = ∫ S n x d π ℓ ( x ) , A_\ell=\sum_{i\in I_\ell}w_i,\qquad\pi_\ell=A_\ell^{-1}\sum_{i\in I_\ell}w_i\delta_{x_i},\qquad b_\ell=\int_{\mathbb{S}^n}x\,\mathrm{d}\pi_\ell(x), A ℓ = i ∈ I ℓ ∑ w i , π ℓ = A ℓ − 1 i ∈ I ℓ ∑ w i δ x i , b ℓ = ∫ S n x d π ℓ ( x ) ,
and let z ℓ = b ℓ ∣ b ℓ ∣ z_\ell= \frac{b_\ell}{|b_\ell|} z ℓ = ∣ b ℓ ∣ b ℓ and v ℓ = 1 − ∣ b ℓ ∣ 2 v_\ell= 1 - |b_\ell|^2 v ℓ = 1 − ∣ b ℓ ∣ 2 . Define
ν ˉ = ∑ ℓ = 1 L A ℓ δ z ℓ , H = ∑ ℓ = 1 L A ℓ θ , V = ∑ ℓ = 1 L A ℓ v ℓ . (84) \bar{\nu} = \sum_{\ell=1}^{L} A_\ell\delta_{z_\ell}, \qquad H = \sum_{\ell=1}^{L} A_\ell^\theta, \qquad V = \sum_{\ell=1}^{L} A_\ell v_\ell. \tag*{(84)} ν ˉ = ℓ = 1 ∑ L A ℓ δ z ℓ , H = ℓ = 1 ∑ L A ℓ θ , V = ℓ = 1 ∑ L A ℓ v ℓ . (84)
Applying Lemma 5.5 to each a j a_j a j on each component and using ∑ j ∫ S n a j d π ℓ = 1 \sum_j \int_{\mathbb{S}^n} a_j\,d\pi_\ell=1 ∑ j ∫ S n a j d π ℓ = 1 , we obtain
∑ j : a j ∣ I ℓ ≢ 0 ( ∫ S n a j q d ν ) θ = A ℓ θ ∑ j ∥ a j ∥ L q ( π ℓ ) ≥ A ℓ θ ( 1 + c q ∑ j Var π ℓ ( a j ) ) . (85) \sum_{j:a_j|_{I_\ell}\not\equiv0} \left(\int_{\mathbb{S}^n} a_j^q\,d\nu\right)^\theta= A_\ell^\theta\sum_j \|a_j\|_{L^q(\pi_\ell)} \ge A_\ell^\theta\left(1+c_q\sum_j \operatorname{Var}_{\pi_\ell}(a_j)\right). \tag*{(85)} j : a j ∣ I ℓ ≡ 0 ∑ ( ∫ S n a j q d ν ) θ = A ℓ θ j ∑ ∥ a j ∥ L q ( π ℓ ) ≥ A ℓ θ ( 1 + c q j ∑ Var π ℓ ( a j ) ) . (85)
By (78) and the diameter bound on I ℓ I_\ell I ℓ ,
∑ j Var π ℓ ( a j ) = 1 2 ∬ ∑ j ∣ a j ( x ) − a j ( y ) ∣ 2 d π ℓ ( x ) d π ℓ ( y ) ≥ c D 2 r 2 1 2 ∬ ∣ x − y ∣ 2 d π ℓ ( x ) d π ℓ ( y ) = c D 2 r 2 v ℓ . \begin{aligned}
\sum_j \operatorname{Var}_{\pi_\ell}(a_j) = \frac{1}{2}\iint\sum_j |a_j(x)-a_j(y)|^2\,d\pi_\ell(x)\,d\pi_\ell(y) \\
&\ge\frac{c}{D^2r^2}\frac{1}{2}\iint|x-y|^2\,d\pi_\ell(x)\,d\pi_\ell(y) = \frac{c}{D^2r^2}v_\ell.
\end{aligned} j ∑ Var π ℓ ( a j ) = 2 1 ∬ j ∑ ∣ a j ( x ) − a j ( y ) ∣ 2 d π ℓ ( x ) d π ℓ ( y ) ≥ D 2 r 2 c 2 1 ∬ ∣ x − y ∣ 2 d π ℓ ( x ) d π ℓ ( y ) = D 2 r 2 c v ℓ .
Summing (85) over ℓ \ell ℓ and using A ℓ θ ≥ A ℓ A_\ell^\theta\ge A_\ell A ℓ θ ≥ A ℓ gives
Φ r ( σ ) ≥ Φ r ( ν ) ≥ H + c r − 2 ∑ ℓ A ℓ θ v ℓ ≥ H + c r − 2 V . (86) \Phi_r(\sigma) \ge\Phi_r(\nu) \ge H+cr^{-2}\sum_\ell A_\ell^\theta v_\ell\ge H+cr^{-2}V. \tag*{(86)} Φ r ( σ ) ≥ Φ r ( ν ) ≥ H + c r − 2 ℓ ∑ A ℓ θ v ℓ ≥ H + c r − 2 V . (86)
By Lemma 5.6, for every f ∈ P ˚ m , n f\in\mathring{\mathcal{P}}_{m,n} f ∈ P ˚ m , n with ∥ f ∥ L ∞ ( S n ) ≤ 1 \|f\|_{L^\infty(\mathbb{S}^n)}\le1 ∥ f ∥ L ∞ ( S n ) ≤ 1 ,
∣ ∫ S n f d ν ˉ − ∫ S n f d ν ∣ ≤ C ∑ ℓ A ℓ v ℓ = C V . \left|\int_{\mathbb{S}^n} f\,d\bar{\nu}-\int_{\mathbb{S}^n} f\,d\nu\right|\le C\sum_\ell A_\ell v_\ell=CV. ∫ S n f d ν ˉ − ∫ S n f d ν ≤ C ℓ ∑ A ℓ v ℓ = C V .
Thus τ ( ν ˉ ) ≤ τ ( σ ) + C V \tau(\bar{\nu})\le\tau(\sigma)+CV τ ( ν ˉ ) ≤ τ ( σ ) + C V . Applying Lemma 5.2 to ν ˉ \bar{\nu} ν ˉ , we find
H ≥ ∥ ν ˉ ∥ θ ≥ Θ − C ( τ ( σ ) + V ) θ ≥ Θ − C τ ( σ ) θ − C V θ . (87) H\ge\|\bar{\nu}\|_\theta\ge\Theta-C(\tau(\sigma)+V)^\theta\ge\Theta-C\tau(\sigma)^\theta-CV^\theta. \tag*{(87)} H ≥ ∥ ν ˉ ∥ θ ≥ Θ − C ( τ ( σ ) + V ) θ ≥ Θ − C τ ( σ ) θ − C V θ . (87)
Combining (86) and (87) yields
Φ r ( σ ) ≥ Θ − C τ ( σ ) θ − C V θ + c r − 2 V . (88) \Phi_r(\sigma)\ge\Theta-C\tau(\sigma)^\theta-CV^\theta+cr^{-2}V. \tag*{(88)} Φ r ( σ ) ≥ Θ − C τ ( σ ) θ − C V θ + c r − 2 V . (88)
For positive constants C 1 , c 1 C_1,c_1 C 1 , c 1 , we have
sup V ≥ 0 { C 1 V θ − c 1 r − 2 V } = ( 1 − θ ) C 1 ( θ C 1 c 1 ) θ 1 − θ r − 2 θ 1 − θ . \sup_{V\ge0}\{C_1V^\theta-c_1r^{-2}V\}=(1-\theta)C_1\left(\frac{\theta C_1}{c_1}\right)^{\frac{\theta}{1-\theta}}r^{-\frac{2\theta}{1-\theta}}. V ≥ 0 sup { C 1 V θ − c 1 r − 2 V } = ( 1 − θ ) C 1 ( c 1 θ C 1 ) 1 − θ θ r − 1 − θ 2 θ .
Optimizing the last two terms in (88) proves (83) .
We now prove Theorem 5.1 .
Proof of Theorem 5.1. Let B = B ( n , p ) B=B(n,p) B = B ( n , p ) be a fixed constant for which (6) holds on S n \mathbb{S}^n S n , and take the partition from Lemma 5.4 . When 1 ≤ p ≤ 2 1 \le p \le2 1 ≤ p ≤ 2 , we apply the elementary estimate
∣ a + b ∣ p ≤ ∣ a ∣ p + p ∣ a ∣ p − 2 a ⋅ b + C p ∣ b ∣ p |a+b|^p \le|a|^p+p|a|^{p-2}a\cdot b+C_p|b|^p ∣ a + b ∣ p ≤ ∣ a ∣ p + p ∣ a ∣ p − 2 a ⋅ b + C p ∣ b ∣ p
with a = η j ∇ u a=\eta_j\nabla u a = η j ∇ u and b = u ∇ η j b=u\nabla\eta_j b = u ∇ η j . The first-order terms cancel after summation because
p ∣ ∇ u ∣ p − 2 u ∇ u ⋅ ∑ j η j p − 1 ∇ η j = 1 p ∣ ∇ u ∣ p − 2 u ∇ u ⋅ ∇ ∑ j η j p = 0. (89) p|\nabla u|^{p-2}u\nabla u\cdot\sum_j\eta_j^{p-1}\nabla\eta_j=\frac{1}{p}|\nabla u|^{p-2}u\nabla u\cdot\nabla\sum_j\eta_j^p=0. \tag*{(89)} p ∣∇ u ∣ p − 2 u ∇ u ⋅ j ∑ η j p − 1 ∇ η j = p 1 ∣∇ u ∣ p − 2 u ∇ u ⋅ ∇ j ∑ η j p = 0. (89)
Consequently, we have
∑ j ∥ ∇ ( η j u ) ∥ L p ( S n ) p ≤ ∥ ∇ u ∥ L p ( S n ) p + C r − p ∥ u ∥ L p ( S n ) p . (90) \sum_j\|\nabla(\eta_j u)\|_{L^p(S^n)}^p\le\|\nabla u\|_{L^p(S^n)}^p+Cr^{-p}\|u\|_{L^p(S^n)}^p. \tag*{(90)} j ∑ ∥∇ ( η j u ) ∥ L p ( S n ) p ≤ ∥∇ u ∥ L p ( S n ) p + C r − p ∥ u ∥ L p ( S n ) p . (90)
For p > 2 p>2 p > 2 , we combine the vector inequality
∣ a + b ∣ p ≤ ∣ a ∣ p + p ∣ a ∣ p − 2 a ⋅ b + C p ( ∣ a ∣ p − 2 ∣ b ∣ 2 + ∣ b ∣ p ) |a+b|^p\le|a|^p+p|a|^{p-2}a\cdot b+C_p\left(|a|^{p-2}|b|^2+|b|^p\right) ∣ a + b ∣ p ≤ ∣ a ∣ p + p ∣ a ∣ p − 2 a ⋅ b + C p ( ∣ a ∣ p − 2 ∣ b ∣ 2 + ∣ b ∣ p )
with Young’s inequality, valid for 0 < δ ≤ 1 0<\delta\le1 0 < δ ≤ 1 ,
C p ∣ a ∣ p − 2 ∣ b ∣ 2 ≤ δ ∣ a ∣ p + C p δ − p − 2 2 ∣ b ∣ p . C_p|a|^{p-2}|b|^2\le\delta|a|^p+C_p\delta^{-\frac{p-2}{2}}|b|^p. C p ∣ a ∣ p − 2 ∣ b ∣ 2 ≤ δ ∣ a ∣ p + C p δ − 2 p − 2 ∣ b ∣ p .
Using the cancellation in (89) , we similarly obtain
∑ j ∥ ∇ ( η j u ) ∥ L p ( S n ) p ≤ ( 1 + δ ) ∥ ∇ u ∥ L p ( S n ) p + C δ 1 − p 2 r − p ∥ u ∥ L p ( S n ) p . (91) \sum_j\|\nabla(\eta_j u)\|_{L^p(S^n)}^p\le(1+\delta)\|\nabla u\|_{L^p(S^n)}^p+C\delta^{1-\frac{p}{2}}r^{-p}\|u\|_{L^p(S^n)}^p. \tag*{(91)} j ∑ ∥∇ ( η j u ) ∥ L p ( S n ) p ≤ ( 1 + δ ) ∥∇ u ∥ L p ( S n ) p + C δ 1 − 2 p r − p ∥ u ∥ L p ( S n ) p . (91)
The assertion is immediate for u = 0 u=0 u = 0 , so assume that u ≠ 0 u\ne0 u = 0 and set
U = ∥ u ∥ L p ∗ ( S n ) p , E = ∥ ∇ u ∥ L p ( S n ) p , L = ∥ u ∥ L p ( S n ) p , d σ = ∣ u ∣ p ∗ d μ ∥ u ∥ L p ∗ ( S n ) p ∗ . U=\|u\|_{L^{p^*}(S^n)}^p,\qquad E=\|\nabla u\|_{L^p(S^n)}^p,\qquad L=\|u\|_{L^p(S^n)}^p,\qquad d\sigma=\frac{|u|^{p^*}\,d\mu}{\|u\|_{L^{p^*}(S^n)}^{p^*}}. U = ∥ u ∥ L p ∗ ( S n ) p , E = ∥∇ u ∥ L p ( S n ) p , L = ∥ u ∥ L p ( S n ) p , d σ = ∥ u ∥ L p ∗ ( S n ) p ∗ ∣ u ∣ p ∗ d μ .
It follows from (68) that
τ ( σ ) θ U ≤ ∥ T ∥ ∞ θ L . (92) \tau(\sigma)^\theta U\le\|T\|_\infty^\theta L. \tag*{(92)} τ ( σ ) θ U ≤ ∥ T ∥ ∞ θ L . (92)
Applying (6) to each η j u \eta_j u η j u and summing, we obtain
Φ r ( σ ) U ≤ S n , p p ∑ j ∥ ∇ ( η j u ) ∥ L p ( S n ) p + B L . (93) \Phi_r(\sigma)U\le S_{n,p}^p\sum_j\|\nabla(\eta_j u)\|_{L^p(S^n)}^p+BL. \tag*{(93)} Φ r ( σ ) U ≤ S n , p p j ∑ ∥∇ ( η j u ) ∥ L p ( S n ) p + B L . (93)
Combining Proposition 5.7 with (92) and (93) , and setting
α = 2 θ 1 − θ = 2 ( n − p ) p , (94) \alpha=\frac{2\theta}{1-\theta}=\frac{2(n-p)}{p}, \tag*{(94)} α = 1 − θ 2 θ = p 2 ( n − p ) , (94)
we obtain
( Θ − C r α ) U ≤ S n , p p ∑ j ∥ ∇ ( η j u ) ∥ L p ( S n ) p + ( B + C ∥ T ∥ ∞ θ ) L . (95) (\Theta-Cr^\alpha)U\le S_{n,p}^p\sum_j\|\nabla(\eta_j u)\|_{L^p(S^n)}^p+(B+C\|T\|_\infty^\theta)L. \tag*{(95)} ( Θ − C r α ) U ≤ S n , p p j ∑ ∥∇ ( η j u ) ∥ L p ( S n ) p + ( B + C ∥ T ∥ ∞ θ ) L . (95)
When 1 ≤ p ≤ 2 1 \le p \le2 1 ≤ p ≤ 2 , take r = c ε 1 α r = c\varepsilon^{\frac{1}{\alpha}} r = c ε α 1 for a sufficiently small positive constant c = c ( m , n , p ) c = c(m,n,p) c = c ( m , n , p ) and apply (90) to obtain (69) . When p > 2 p > 2 p > 2 , take r = c ε 1 α r = c\varepsilon^{\frac{1}{\alpha}} r = c ε α 1 as above and apply (91) with δ = c ε \delta= c\varepsilon δ = c ε to obtain the same conclusion.
We end this section with the proof of Corollary 1.8.
Proof of Corollary 1.8. For simplicity, let C C C denote a constant depending only on n , p , m n,p,m n , p , m ; its value may change from line to line.
For any 0 < ε < 1 0 < \varepsilon< 1 0 < ε < 1 , Theorem 1.6 gives
∥ u ∥ L p ∗ ( S n ) p ≤ ( S n , p p Θ ( m , n − p n , n ) + ε ) ∥ ∇ u ∥ L p ( S n ) p + C ε − γ ( p ) ∥ u ∥ L p ( S n ) p . (96) \lVert u\rVert_{L^{p^*}(\mathbb{S}^n)}^p \leq\left(\frac{S_{n,p}^p}{\Theta\left(m,\frac{n-p}{n},n\right)}+\varepsilon\right)\lVert\nabla u\rVert_{L^p(\mathbb{S}^n)}^p+C\varepsilon^{-\gamma(p)}\lVert u\rVert_{L^p(\mathbb{S}^n)}^p. \tag*{(96)} ∥ u ∥ L p ∗ ( S n ) p ≤ ( Θ ( m , n n − p , n ) S n , p p + ε ) ∥ ∇ u ∥ L p ( S n ) p + C ε − γ ( p ) ∥ u ∥ L p ( S n ) p . (96)
We consider the case ∥ ∇ u ∥ L p ( S n ) > ∥ u ∥ L p ( S n ) \lVert\nabla u\rVert_{L^p(\mathbb{S}^n)}>\lVert u\rVert_{L^p(\mathbb{S}^n)} ∥ ∇ u ∥ L p ( S n ) > ∥ u ∥ L p ( S n ) . Choose
ε = ( ∥ u ∥ L p ( S n ) ∥ ∇ u ∥ L p ( S n ) ) p 1 + γ ( p ) = ( ∥ u ∥ L p ( S n ) ∥ ∇ u ∥ L p ( S n ) ) γ ˉ ( p ) . \varepsilon=\left(\frac{\lVert u\rVert_{L^p(\mathbb{S}^n)}}{\lVert\nabla u\rVert_{L^p(\mathbb{S}^n)}}\right)^{\frac{p}{1+\gamma(p)}}=\left(\frac{\lVert u\rVert_{L^p(\mathbb{S}^n)}}{\lVert\nabla u\rVert_{L^p(\mathbb{S}^n)}}\right)^{\bar{\gamma}(p)}. ε = ( ∥ ∇ u ∥ L p ( S n ) ∥ u ∥ L p ( S n ) ) 1 + γ ( p ) p = ( ∥ ∇ u ∥ L p ( S n ) ∥ u ∥ L p ( S n ) ) γ ˉ ( p ) .
It follows from (96) that
∥ u ∥ L p ∗ ( S n ) p ≤ S n , p p Θ ( m , n − p n , n ) ∥ ∇ u ∥ L p ( S n ) p + C ∥ ∇ u ∥ L p ( S n ) p − γ ˉ ( p ) ∥ u ∥ L p ( S n ) γ ˉ ( p ) . (97) \lVert u\rVert_{L^{p^*}(\mathbb{S}^n)}^p\leq\frac{S_{n,p}^p}{\Theta\left(m,\frac{n-p}{n},n\right)}\lVert\nabla u\rVert_{L^p(\mathbb{S}^n)}^p+C\lVert\nabla u\rVert_{L^p(\mathbb{S}^n)}^{p-\bar{\gamma}(p)}\lVert u\rVert_{L^p(\mathbb{S}^n)}^{\bar{\gamma}(p)}. \tag*{(97)} ∥ u ∥ L p ∗ ( S n ) p ≤ Θ ( m , n n − p , n ) S n , p p ∥ ∇ u ∥ L p ( S n ) p + C ∥ ∇ u ∥ L p ( S n ) p − γ ˉ ( p ) ∥ u ∥ L p ( S n ) γ ˉ ( p ) . (97)
We use the elementary inequality
( x + y ) q ≥ x q + q x q − 1 y , x , y ≥ 0 , q ≥ 1 , (98) (x+y)^q\geq x^q+qx^{q-1}y,\qquad x,y\geq0,\quad q\geq1, \tag*{(98)} ( x + y ) q ≥ x q + q x q − 1 y , x , y ≥ 0 , q ≥ 1 , (98)
with q = 1 + γ ( p ) q=1+\gamma(p) q = 1 + γ ( p ) ,
x = ( S n , p Θ ( m , n − p n , n ) 1 p ) γ ˉ ( p ) ∥ ∇ u ∥ L p ( S n ) γ ˉ ( p ) x=\left(\frac{S_{n,p}}{\Theta\left(m,\frac{n-p}{n},n\right)^{\frac{1}{p}}}\right)^{\bar{\gamma}(p)}\lVert\nabla u\rVert_{L^p(\mathbb{S}^n)}^{\bar{\gamma}(p)} x = Θ ( m , n n − p , n ) p 1 S n , p γ ˉ ( p ) ∥ ∇ u ∥ L p ( S n ) γ ˉ ( p )
and
y = C ∥ u ∥ L p ( S n ) γ ˉ ( p ) , y=C\lVert u\rVert_{L^p(\mathbb{S}^n)}^{\bar{\gamma}(p)}, y = C ∥ u ∥ L p ( S n ) γ ˉ ( p ) ,
we see that the right-hand side of (97) is bounded by ( x + y ) q (x+y)^q ( x + y ) q . Hence,
∥ u ∥ L p ∗ ( S n ) p q ≤ x + y , \lVert u\rVert_{L^{p^*}(\mathbb{S}^n)}^{\frac{p}{q}}\leq x+y, ∥ u ∥ L p ∗ ( S n ) q p ≤ x + y ,
which yields (23) .
Quantitative Moser–Trudinger–Onofri remainder estimates
In the first subsection, we prove Theorems 1.7 and 6.2 , as well as Corollary 1.9. In the second subsection, we generalize [15 Theorem 1.1] .
General moment constraints
The proofs of Theorems 1.7 and 6.2 rely on the following quantitative “distribution of mass” principle for the higher-order Moser–Trudinger–Onofri inequality (8) .
Lemma 6.1. Let ( M , g ) (M,g) ( M , g ) be a smooth closed manifold of dimension n ≥ 2 n \ge2 n ≥ 2 . Assume s ∈ N + s \in\mathbb{N}_{+} s ∈ N + , 1 ≤ s < n 1 \le s < n 1 ≤ s < n , and u ∈ W s , n s ( M ) u \in W^{s,\frac{n}{s}}(M) u ∈ W s , s n ( M ) satisfies ∫ M u d μ = 0 \int_M u\,\mathrm{d}\mu= 0 ∫ M u d μ = 0 . Suppose there exist open subsets Ω 1 , … , Ω k ⊂ M \Omega_1,\ldots,\Omega_k \subset M Ω 1 , … , Ω k ⊂ M and positive constants r < 1 r < 1 r < 1 and δ \delta δ such that dist ( Ω i , Ω j ) ≥ 2 r \operatorname{dist}(\Omega_i,\Omega_j) \ge2r dist ( Ω i , Ω j ) ≥ 2 r whenever i ≠ j i \ne j i = j , and
∫ Ω i e n u d μ ∫ M e n u d μ ≥ δ (99) \frac{\int_{\Omega_i} e^{nu}\,\mathrm{d}\mu}{\int_M e^{nu}\,\mathrm{d}\mu} \ge\delta \tag*{(99)} ∫ M e n u d μ ∫ Ω i e n u d μ ≥ δ (99) for all 1 ≤ i ≤ k 1 \le i \le k 1 ≤ i ≤ k . Then for any 0 < ε < 1 0 < \varepsilon< 1 0 < ε < 1 , we have
log ∫ M e n u d μ ≤ ( α s , n k + ε ) ∥ Δ s 2 u ∥ L n s ( M ) n s + c ( n , s , M , δ , r ) ε − s n − s . (100) \log\int_M e^{nu}\,\mathrm{d}\mu\le\left(\frac{\alpha_{s,n}}{k}+\varepsilon\right)\left\|\Delta^{\frac{s}{2}}u\right\|_{L^{\frac{n}{s}}(M)}^{\frac{n}{s}}+c(n,s,M,\delta,r)\varepsilon^{-\frac{s}{n-s}}. \tag*{(100)} log ∫ M e n u d μ ≤ ( k α s , n + ε ) Δ 2 s u L s n ( M ) s n + c ( n , s , M , δ , r ) ε − n − s s . (100)
The “distribution of mass” method dates back to the seminal work of Chen and Li [13 ] and has been widely used in the study of improved inequalities and prescribing curvature problems (see, for instance, [18 , 44 , 50 ] and the references therein). To the best of our knowledge, previous work has mainly focused on the first-order case (i.e. s = 1 s=1 s = 1 ) and the Hilbert case (i.e. n = 2 s n=2s n = 2 s ). For the first-order case (see, for example, [13 , 50 ] ), the proofs typically rely on a truncation argument originating in [5 ] . For the Hilbert case (see, for example, [18 , 44 ] ), the proofs are usually based on suitable eigenspace decompositions. However, these methods do not extend to the general setting.
To establish Lemma 6.1 , we draw on Fontana’s proof of (7) in [21 ] , whose ideas trace back to [2 ] . We use the Green representation of u ∈ W s , n s ( M ) u \in W^{s,\frac{n}{s}}(M) u ∈ W s , s n ( M ) satisfying ∫ M u d μ = 0 \int_M u\,\mathrm{d}\mu= 0 ∫ M u d μ = 0 :
u ( x ) = ∫ M G s ( x , y ) ⋅ Δ s 2 u ( y ) d μ ( y ) = : J s ( Δ s 2 u ) ( x ) . (101) u(x)=\int_M G_s(x,y)\cdot\Delta^{\frac{s}{2}}u(y)\,\mathrm{d}\mu(y)=:J_s\left(\Delta^{\frac{s}{2}}u\right)(x). \tag*{(101)} u ( x ) = ∫ M G s ( x , y ) ⋅ Δ 2 s u ( y ) d μ ( y ) =: J s ( Δ 2 s u ) ( x ) . (101)
Here G s G_s G s is a scalar function when s s s is even, and a vector field when s s s is odd. For any v ∈ L n s ( M ) v \in L^{\frac{n}{s}}(M) v ∈ L s n ( M ) , we define
I s ( v ) ( x ) = ∫ M ∣ G s ( x , y ) ∣ ∣ v ( y ) ∣ d μ ( y ) . (102) I_s(v)(x)=\int_M |G_s(x,y)||v(y)|\,\mathrm{d}\mu(y). \tag*{(102)} I s ( v ) ( x ) = ∫ M ∣ G s ( x , y ) ∣∣ v ( y ) ∣ d μ ( y ) . (102)
From [21 ] , Theorems 1.8 and 1.9 and Young’s inequality, we obtain the following Riesz-type Moser–Trudinger–Onofri inequality:
log ∫ M e n I s ( v ) d μ ≤ α s , n ∥ v ∥ L n s ( M ) n s + C ( M , s ) . (103) \log\int_M e^{nI_s(v)}\,\mathrm{d}\mu\le\alpha_{s,n}\|v\|_{L^{\frac{n}{s}}(M)}^{\frac{n}{s}}+C(M,s). \tag*{(103)} log ∫ M e n I s ( v ) d μ ≤ α s , n ∥ v ∥ L s n ( M ) s n + C ( M , s ) . (103)
By (101) and (102) , inequality (103) implies (8) . Our main idea is to reduce the estimates for (8) to those for (103) , which can then be handled using classical cut-off techniques similar to those in [50 Theorem 3.3] .
Proof of Lemma 6.1. For simplicity, let c 0 c_0 c 0 denote a positive constant depending only on n n n , s s s , M M M , δ \delta δ , and r r r and allowed to vary from line to line. Choose η i ∈ C ∞ ( M ) \eta_i \in C^\infty(M) η i ∈ C ∞ ( M ) (1 ≤ i ≤ k 1 \le i \le k 1 ≤ i ≤ k ) such that 0 ≤ η i ≤ 1 0 \le\eta_i \le1 0 ≤ η i ≤ 1 , η i ∣ Ω ~ i = 1 \eta_i|_{\widetilde{\Omega}_i}=1 η i ∣ Ω i = 1 , where Ω ~ i : = { x ∈ M : dist ( x , Ω i ) ≤ r 3 } \widetilde{\Omega}_i:=\{x \in M:\operatorname{dist}(x,\Omega_i)\le\frac{r}{3}\} Ω i := { x ∈ M : dist ( x , Ω i ) ≤ 3 r } , ∣ ∇ η i ∣ ≤ c 0 r − 1 |\nabla\eta_i|\le c_0r^{-1} ∣∇ η i ∣ ≤ c 0 r − 1 , and supp ( η i ) ∩ supp ( η j ) = ∅ \operatorname{supp}(\eta_i)\cap\operatorname{supp}(\eta_j)=\varnothing supp ( η i ) ∩ supp ( η j ) = ∅ whenever i ≠ j i\ne j i = j . For each 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k , equations (101) and (99) give
∫ M e n u d μ ≤ δ − 1 ∫ Ω i e n u d μ = δ − 1 ∫ Ω i e n J s ( Δ s 2 u ) d μ = δ − 1 ∫ Ω i e n J s ( η i Δ s 2 u ) + n J s ( ( 1 − η i ) Δ s 2 u ) d μ . (104) \begin{aligned}
\int_M e^{nu}\,d\mu&\le\delta^{-1}\int_{\Omega_i}e^{nu}\,d\mu\\
&=\delta^{-1}\int_{\Omega_i}e^{nJ_s(\Delta^{\frac{s}{2}}u)}\,d\mu\\
&=\delta^{-1}\int_{\Omega_i}e^{nJ_s(\eta_i\Delta^{\frac{s}{2}}u)+nJ_s((1-\eta_i)\Delta^{\frac{s}{2}}u)}\,d\mu. \tag*{(104)}
\end{aligned} ∫ M e n u d μ ≤ δ − 1 ∫ Ω i e n u d μ = δ − 1 ∫ Ω i e n J s ( Δ 2 s u ) d μ = δ − 1 ∫ Ω i e n J s ( η i Δ 2 s u ) + n J s (( 1 − η i ) Δ 2 s u ) d μ . (104)
By [21 Theorem 1.8] and the properties of η i \eta_i η i , for any x ∈ Ω i x\in\Omega_i x ∈ Ω i ,
J s ( ( 1 − η i ) Δ s 2 u ) ( x ) ≤ c 0 ∥ Δ s 2 u ∥ L 1 ( M ) . (105) J_s((1-\eta_i)\Delta^{\frac{s}{2}}u)(x)\le c_0\|\Delta^{\frac{s}{2}}u\|_{L^1(M)}. \tag*{(105)} J s (( 1 − η i ) Δ 2 s u ) ( x ) ≤ c 0 ∥ Δ 2 s u ∥ L 1 ( M ) . (105)
Using (102) , (103) , (104) , and (105) , we obtain
∫ M e n u d μ ≤ e C ( M , s ) δ − 1 exp ( α s , n ∥ η i Δ s 2 u ∥ L s n s ( M ) n s + c 0 ∥ Δ s 2 u ∥ L s n s ( M ) n s ) . (106) \int_M e^{nu}\,d\mu\le e^{C(M,s)}\delta^{-1}\exp\left(\alpha_{s,n}\|\eta_i\Delta^{\frac{s}{2}}u\|_{L_s^{\frac ns}(M)}^{\frac ns}+c_0\|\Delta^{\frac{s}{2}}u\|_{L_s^{\frac ns}(M)}^{\frac ns}\right). \tag*{(106)} ∫ M e n u d μ ≤ e C ( M , s ) δ − 1 exp ( α s , n ∥ η i Δ 2 s u ∥ L s s n ( M ) s n + c 0 ∥ Δ 2 s u ∥ L s s n ( M ) s n ) . (106)
By Young’s inequality,
c 0 ∥ Δ s 2 u ∥ L s n s ( M ) n s ≤ ε ∥ Δ s 2 u ∥ L s n s ( M ) n s + c 0 n s ε − s n − s . (107) c_0\|\Delta^{\frac{s}{2}}u\|_{L_s^{\frac ns}(M)}^{\frac ns}\le\varepsilon\|\Delta^{\frac{s}{2}}u\|_{L_s^{\frac ns}(M)}^{\frac ns}+c_0^{\frac ns}\varepsilon^{-\frac{s}{n-s}}. \tag*{(107)} c 0 ∥ Δ 2 s u ∥ L s s n ( M ) s n ≤ ε ∥ Δ 2 s u ∥ L s s n ( M ) s n + c 0 s n ε − n − s s . (107)
Moreover, there exists an index 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k such that
∥ η i Δ s 2 u ∥ L s n s ( M ) n s ≤ 1 k ∥ Δ s 2 u ∥ L s n s ( M ) n s . (108) \|\eta_i\Delta^{\frac{s}{2}}u\|_{L_s^{\frac ns}(M)}^{\frac ns}\le\frac{1}{k}\|\Delta^{\frac{s}{2}}u\|_{L_s^{\frac ns}(M)}^{\frac ns}. \tag*{(108)} ∥ η i Δ 2 s u ∥ L s s n ( M ) s n ≤ k 1 ∥ Δ 2 s u ∥ L s s n ( M ) s n . (108)
Combining (106) , (107) , and (108) yields (100) .
Combining Remark 4.4 with Lemma 6.1 gives the following result.
Theorem 6.2. Assume that s , m , n ∈ N + s,m,n\in\mathbb{N}_+ s , m , n ∈ N + , 1 ≤ s < n 1\le s<n 1 ≤ s < n , and 0 < ε < 1 0<\varepsilon<1 0 < ε < 1 . Then, for any function u ∈ W s n s ( S n ) u\in W_s^{\frac ns}(\mathbb{S}^n) u ∈ W s s n ( S n ) satisfying ∫ S n u d μ = 0 \int_{\mathbb{S}^n}u\,d\mu=0 ∫ S n u d μ = 0 and
∣ ∫ S n f e n u d μ ∣ ≤ ∣ T ( f ) ∣ for all f ∈ P ˚ m , n , (109) \left|\int_{\mathbb{S}^n}fe^{nu}\,d\mu\right|\le|T(f)|\quad\text{for all }f\in\mathring{\mathcal{P}}_{m,n}, \tag*{(109)} ∫ S n f e n u d μ ≤ ∣ T ( f ) ∣ for all f ∈ P ˚ m , n , (109) where T T T is a linear functional on P ˚ m , n \mathring{\mathcal{P}}_{m,n} P ˚ m , n , the following inequality holds:
log ∫ S n e n u d μ ≤ ( α s , n Θ ( m , 0 , n ) + ε ) ∥ Δ s 2 u ∥ L s n s ( S n ) s + c ε , (110) \log\int_{\mathbb{S}^n}e^{nu}\,d\mu\le\left(\frac{\alpha_{s,n}}{\Theta(m,0,n)}+\varepsilon\right)\|\Delta^{\frac{s}{2}}u\|_{L_s^{\frac ns}(\mathbb{S}^n)}^s+c_\varepsilon, \tag*{(110)} log ∫ S n e n u d μ ≤ ( Θ ( m , 0 , n ) α s , n + ε ) ∥ Δ 2 s u ∥ L s s n ( S n ) s + c ε , (110) where the constant c ε c_\varepsilon c ε can be chosen as
c ε = log ( ∥ T ∥ ∞ + 1 ) + C ( m , n , s ) ε − s n − s , (111) c_\varepsilon=\log(\|T\|_\infty+1)+C(m,n,s)\varepsilon^{-\frac{s}{n-s}}, \tag*{(111)} c ε = log ( ∥ T ∥ ∞ + 1 ) + C ( m , n , s ) ε − n − s s , (111) and ∥ T ∥ ∞ \|T\|_\infty ∥ T ∥ ∞ is defined in (70) .
Proof. Take ε 0 = ε 0 ( m , n ) \varepsilon_0=\varepsilon_0(m,n) ε 0 = ε 0 ( m , n ) to be the positive constant in Remark 4.4. If
∫ S n e n u d μ ≤ ∥ T ∥ ∞ ε 0 − 1 , (112) \int_{\mathbb{S}^n} e^{nu}\,d\mu\le\|\mathcal{T}\|_{\infty}\varepsilon_0^{-1}, \tag*{(112)} ∫ S n e n u d μ ≤ ∥ T ∥ ∞ ε 0 − 1 , (112)
then (110) holds provided that
c ε ≥ log ( ∥ T ∥ ∞ + 1 ) − log ε 0 . c_{\varepsilon}\ge\log\left(\|\mathcal{T}\|_{\infty}+1\right)-\log\varepsilon_0. c ε ≥ log ( ∥ T ∥ ∞ + 1 ) − log ε 0 .
Assume now that (112) does not hold. Define
d σ = e n u d μ ∫ S n e n u d μ . d\sigma=\frac{e^{nu}\,d\mu}{\int_{\mathbb{S}^n}e^{nu}\,d\mu}. d σ = ∫ S n e n u d μ e n u d μ .
Then σ \sigma σ is a Borel probability measure on S n \mathbb{S}^n S n . Moreover, by (109) , it satisfies the relaxed moment condition in Remark 4.4. Therefore, we can find N ∗ = Θ ( m , 0 , n ) N_*=\Theta(m,0,n) N ∗ = Θ ( m , 0 , n ) points x 1 , … , x N ∗ ∈ S n x_1,\ldots,x_{N_*}\in\mathbb{S}^n x 1 , … , x N ∗ ∈ S n such that d ( x i , x j ) ≥ R = R ( m , n ) > 0 d(x_i,x_j)\ge R=R(m,n)>0 d ( x i , x j ) ≥ R = R ( m , n ) > 0 whenever i ≠ j i\ne j i = j , and σ ( B ( x i , R 200 ) ) ≥ δ ( m , n ) > 0 \sigma\left(B\left(x_i,\frac{R}{200}\right)\right)\ge\delta(m,n)>0 σ ( B ( x i , 200 R ) ) ≥ δ ( m , n ) > 0 for every i i i .
Applying Lemma 6.1 to the domains Ω i : = B ( x i , R 200 ) \Omega_i:=B\left(x_i,\frac{R}{200}\right) Ω i := B ( x i , 200 R ) (1 ≤ i ≤ N ∗ 1\le i\le N_* 1 ≤ i ≤ N ∗ ), we obtain (110) from (100) . The estimate (111) follows from the preceding argument.
We conclude this subsection by proving Corollary 1.9.
Proof of Corollary 1.9. For simplicity, let C C C denote a constant depending only on n n n , s s s , m m m and allowed to vary from line to line.
From Theorem 1.7 , for any 0 < ε < 1 0<\varepsilon<1 0 < ε < 1 ,
log ∫ S n e n u d μ ≤ ( α s , n Θ ( m , 0 , n ) + ε ) ∥ Δ s 2 u ∥ L n s ( S n ) s + C ε − s n − s . (113) \log\int_{\mathbb{S}^n}e^{nu}\,d\mu\le\left(\frac{\alpha_{s,n}}{\Theta(m,0,n)}+\varepsilon\right)\left\|\Delta^{\frac{s}{2}}u\right\|_{L^{\frac{n}{s}}(\mathbb{S}^n)}^s+C\varepsilon^{-\frac{s}{n-s}}. \tag*{(113)} log ∫ S n e n u d μ ≤ ( Θ ( m , 0 , n ) α s , n + ε ) Δ 2 s u L s n ( S n ) s + C ε − n − s s . (113)
It suffices to consider the case ∥ Δ s 2 u ∥ L n s ( S n ) > 1 \left\|\Delta^{\frac{s}{2}}u\right\|_{L^{\frac{n}{s}}(\mathbb{S}^n)}>1 Δ 2 s u L s n ( S n ) > 1 . Setting
ε = ∥ Δ s 2 u ∥ L n s ( S n ) 1 − n s , \varepsilon=\left\|\Delta^{\frac{s}{2}}u\right\|_{L^{\frac{n}{s}}(\mathbb{S}^n)}^{1-\frac{n}{s}}, ε = Δ 2 s u L s n ( S n ) 1 − s n ,
we derive from (113) that
log ∫ S n e n u d μ ≤ α s , n Θ ( m , 0 , n ) ∥ Δ s 2 u ∥ L n s ( S n ) s + C ∥ Δ s 2 u ∥ L n s ( S n ) . (114) \log\int_{\mathbb{S}^n}e^{nu}\,d\mu\le\frac{\alpha_{s,n}}{\Theta(m,0,n)}\left\|\Delta^{\frac{s}{2}}u\right\|_{L^{\frac{n}{s}}(\mathbb{S}^n)}^s+C\left\|\Delta^{\frac{s}{2}}u\right\|_{L^{\frac{n}{s}}(\mathbb{S}^n)}. \tag*{(114)} log ∫ S n e n u d μ ≤ Θ ( m , 0 , n ) α s , n Δ 2 s u L s n ( S n ) s + C Δ 2 s u L s n ( S n ) . (114)
Choosing q = n n − s q=\frac{n}{n-s} q = n − s n ,
x = ( α s , n Θ ( m , 0 , n ) ) 1 − s n ∥ Δ s 2 u ∥ L n s ( S n ) x=\left(\frac{\alpha_{s,n}}{\Theta(m,0,n)}\right)^{1-\frac{s}{n}}\left\|\Delta^{\frac{s}{2}}u\right\|_{L^{\frac{n}{s}}(\mathbb{S}^n)} x = ( Θ ( m , 0 , n ) α s , n ) 1 − n s Δ 2 s u L s n ( S n )
and
we see from (98) that the right-hand side of (114) is bounded by ( x + y ) q (x+y)^q ( x + y ) q . Hence,
log 1 q ( 1 ω n ∫ S n e n u d μ ) ≤ x + y , \log^{\frac{1}{q}}\left(\frac{1}{\omega_n}\int_{S^n}e^{nu}\,d\mu\right)\le x+y, log q 1 ( ω n 1 ∫ S n e n u d μ ) ≤ x + y ,
which gives (25) .
Second-moment refinements
In this subsection, we revisit the recent work [15 ] . For u ∈ W 1 , 2 ( S 2 ) u\in W^{1,2}(S^2) u ∈ W 1 , 2 ( S 2 ) , set Λ ( u ) = ( Λ i j ( u ) ) i , j \Lambda(u)=(\Lambda_{ij}(u))_{i,j} Λ ( u ) = ( Λ ij ( u ) ) i , j , where, for 1 ≤ i , j ≤ 3 1\le i,j\le3 1 ≤ i , j ≤ 3 ,
Λ i j ( u ) = ∫ S 2 e 2 u ( x ) ( x i x j − 1 3 δ i j ) d μ ∫ S 2 e 2 u d μ . \Lambda_{ij}(u)=\frac{\int_{S^2}e^{2u(x)}\left(x_i x_j-\frac{1}{3}\delta_{ij}\right)\,d\mu}{\int_{S^2}e^{2u}\,d\mu}. Λ ij ( u ) = ∫ S 2 e 2 u d μ ∫ S 2 e 2 u ( x ) ( x i x j − 3 1 δ ij ) d μ .
Define ∥ Λ ( u ) ∥ 2 = ∑ i , j Λ i j ( u ) 2 \|\Lambda(u)\|^2=\sum_{i,j}\Lambda_{ij}(u)^2 ∥Λ ( u ) ∥ 2 = ∑ i , j Λ ij ( u ) 2 , which measures the deviation of the second-order moments from zero. In [15 Theorem 1.1] , Chen and Zhang proved that if u u u satisfies ∫ S 2 u d μ = 0 \int_{S^2}u\,d\mu=0 ∫ S 2 u d μ = 0 and (17) with m = 1 m=1 m = 1 and n = 2 n=2 n = 2 , and if ∥ Λ ( u ) ∥ 2 ≤ 2 3 − δ \|\Lambda(u)\|^2\le\frac{2}{3}-\delta ∥Λ ( u ) ∥ 2 ≤ 3 2 − δ for some positive constant δ \delta δ , then, for any 0 < ε < 1 0<\varepsilon<1 0 < ε < 1 ,
log ∫ S 2 e 2 u d μ ≤ ( α 1 , 2 3 + ε ) ∥ ∇ u ∥ L 2 ( S 2 ) 2 + c ε , δ (115) \log\int_{S^2}e^{2u}\,d\mu\le\left(\frac{\alpha_{1,2}}{3}+\varepsilon\right)\|\nabla u\|_{L^2(S^2)}^2+c_{\varepsilon,\delta} \tag*{(115)} log ∫ S 2 e 2 u d μ ≤ ( 3 α 1 , 2 + ε ) ∥∇ u ∥ L 2 ( S 2 ) 2 + c ε , δ (115)
for some positive constant c ε , δ c_{\varepsilon,\delta} c ε , δ . If, in addition, ∥ Λ ( u ) ∥ 2 ≤ 1 6 − δ \|\Lambda(u)\|^2\le\frac{1}{6}-\delta ∥Λ ( u ) ∥ 2 ≤ 6 1 − δ , then
log ∫ S 2 e 2 u d μ ≤ ( α 1 , 2 4 + ε ) ∥ ∇ u ∥ L 2 ( S 2 ) 2 + c ε , δ . (116) \log\int_{S^2}e^{2u}\,d\mu\le\left(\frac{\alpha_{1,2}}{4}+\varepsilon\right)\|\nabla u\|_{L^2(S^2)}^2+c_{\varepsilon,\delta}. \tag*{(116)} log ∫ S 2 e 2 u d μ ≤ ( 4 α 1 , 2 + ε ) ∥∇ u ∥ L 2 ( S 2 ) 2 + c ε , δ . (116)
Note that Θ ( 1 , 0 , 2 ) = 2 \Theta(1,0,2)=2 Θ ( 1 , 0 , 2 ) = 2 and Θ ( 2 , 0 , 2 ) = 4 \Theta(2,0,2)=4 Θ ( 2 , 0 , 2 ) = 4 (see § 2 ). Hence (115) represents an intermediate case between (18) with m = 1 , n = 2 m=1,n=2 m = 1 , n = 2 and (18) with m = 2 , n = 2 m=2,n=2 m = 2 , n = 2 , while (116) improves (18) by relaxing the condition (17) in the case m = 2 , n = 2 m=2,n=2 m = 2 , n = 2 . Moreover, both (115) and (116) are almost sharp, and the upper bounds 2 3 − δ \frac{2}{3}-\delta 3 2 − δ and 1 6 − δ \frac{1}{6}-\delta 6 1 − δ cannot be replaced by any constants larger than 2 3 \frac{2}{3} 3 2 and 1 6 \frac{1}{6} 6 1 , respectively.
The proofs of (115) and (116) in [15 ] rely on a careful application of the concentration–compactness principles developed in [10 , 29 ] . Here, we give a new quantitative proof of these results and further generalize them to the higher-order inequality (8) on S n S^n S n for n ≥ 2 n\ge2 n ≥ 2 .
We first introduce some notation. Let σ \sigma σ be a Radon probability measure on S n S^n S n . Define Λ ( n ) ( σ ) = ( Λ i j ( n ) ( σ ) ) i , j \Lambda^{(n)}(\sigma)=(\Lambda_{ij}^{(n)}(\sigma))_{i,j} Λ ( n ) ( σ ) = ( Λ ij ( n ) ( σ ) ) i , j , where for 1 ≤ i , j ≤ n + 1 1\le i,j\le n+1 1 ≤ i , j ≤ n + 1 ,
Λ i j ( n ) ( σ ) = ∫ S n ( x i x j − 1 n + 1 δ i j ) d σ . \Lambda_{ij}^{(n)}(\sigma)=\int_{S^n}\left(x_i x_j-\frac{1}{n+1}\delta_{ij}\right)\,d\sigma. Λ ij ( n ) ( σ ) = ∫ S n ( x i x j − n + 1 1 δ ij ) d σ .
Define ∥ Λ ( n ) ( σ ) ∥ 2 = ∑ i , j Λ i j ( n ) ( σ ) 2 \|\Lambda^{(n)}(\sigma)\|^2=\sum_{i,j}\Lambda_{ij}^{(n)}(\sigma)^2 ∥ Λ ( n ) ( σ ) ∥ 2 = ∑ i , j Λ ij ( n ) ( σ ) 2 . For any function u ∈ W s s , n ( S n ) u\in W_s^{s,n}(S^n) u ∈ W s s , n ( S n ) with ∫ S n u d μ = 0 \int_{S^n}u\,d\mu=0 ∫ S n u d μ = 0 , set Λ ( n ) ( u ) = Λ ( n ) ( σ u ) \Lambda^{(n)}(u)=\Lambda^{(n)}(\sigma_u) Λ ( n ) ( u ) = Λ ( n ) ( σ u ) , where
d σ u = e n u d μ ∫ S n e n u d μ . d\sigma_u=\frac{e^{nu}\,d\mu}{\int_{S^n}e^{nu}\,d\mu}. d σ u = ∫ S n e n u d μ e n u d μ .
From § 2 , we know that Θ ( 1 , 0 , n ) = 2 \Theta(1,0,n)=2 Θ ( 1 , 0 , n ) = 2 and Θ ( 2 , 0 , n ) = n + 2 \Theta(2,0,n)=n+2 Θ ( 2 , 0 , n ) = n + 2 . For any integer $2\le k\le n + 2 n+2 n + 2 , define
g n ( k ) = inf { ∥ Λ ( n ) ( ν ) ∥ 2 : ν ∈ M 1 , n c , ∥ ν ∥ 0 ≤ k } . g_n(k)=\inf\left\{\left\|\Lambda^{(n)}(\nu)\right\|^2:\nu\in\mathcal{M}_{1,n}^c,\ \|\nu\|_0\leq k\right\}. g n ( k ) = inf { Λ ( n ) ( ν ) 2 : ν ∈ M 1 , n c , ∥ ν ∥ 0 ≤ k } .
Recall that ∥ ν ∥ 0 \|\nu\|_0 ∥ ν ∥ 0 denotes the cardinality of the support of ν \nu ν . It was shown in [15 ] that g 2 ( 2 ) = 2 3 g_2(2)=\frac{2}{3} g 2 ( 2 ) = 3 2 , g 2 ( 3 ) = 1 6 g_2(3)=\frac{1}{6} g 2 ( 3 ) = 6 1 , and g 2 ( 4 ) = 0 g_2(4)=0 g 2 ( 4 ) = 0 . The following lemma determines g n ( k ) g_n(k) g n ( k ) in all dimensions.
Lemma 6.3. For any integer 2 ≤ k ≤ n + 2 2\leq k\leq n+2 2 ≤ k ≤ n + 2 , we have
g n ( k ) = 1 k − 1 − 1 n + 1 = n − k + 2 ( k − 1 ) ( n + 1 ) . g_n(k)=\frac{1}{k-1}-\frac{1}{n+1}=\frac{n-k+2}{(k-1)(n+1)}. g n ( k ) = k − 1 1 − n + 1 1 = ( k − 1 ) ( n + 1 ) n − k + 2 . Moreover, equality is attained if and only if ν \nu ν is the uniform measure supported on the vertices of a regular ( k − 1 ) (k-1) ( k − 1 ) -simplex contained in a ( k − 1 ) (k-1) ( k − 1 ) -dimensional linear subspace of R n + 1 \mathbb{R}^{n+1} R n + 1 . In particular, the minimizing measure is unique up to orthogonal transformations of R n + 1 \mathbb{R}^{n+1} R n + 1 .
Proof. Let
ν = ∑ a = 1 s w a δ x a , 1 ≤ s ≤ k , \nu=\sum_{a=1}^{s}w_a\delta_{x_a},\qquad1\leq s\leq k, ν = a = 1 ∑ s w a δ x a , 1 ≤ s ≤ k ,
be an admissible measure, where x a ∈ S n x_a\in\mathbb{S}^n x a ∈ S n , w a > 0 w_a>0 w a > 0 , and ∑ a = 1 s w a = 1 \sum_{a=1}^{s}w_a=1 ∑ a = 1 s w a = 1 . Since ν ∈ M 1 , n c \nu\in\mathcal{M}_{1,n}^c ν ∈ M 1 , n c , we have
∑ a = 1 s w a x a = 0. \sum_{a=1}^{s}w_ax_a=0. a = 1 ∑ s w a x a = 0.
Define the second moment matrix
M = ∫ S n x ⊗ x d ν = ∑ a = 1 s w a x a x a T . M=\int_{\mathbb{S}^n}x\otimes x\,d\nu=\sum_{a=1}^{s}w_ax_ax_a^{T}. M = ∫ S n x ⊗ x d ν = a = 1 ∑ s w a x a x a T .
Then M M M is symmetric and positive semidefinite. Moreover, tr M = ∑ a = 1 s w a ∣ x a ∣ 2 = 1 \operatorname{tr}M=\sum_{a=1}^{s}w_a|x_a|^2=1 tr M = ∑ a = 1 s w a ∣ x a ∣ 2 = 1 . Since
Λ ( n ) ( ν ) = M − 1 n + 1 I n + 1 , \Lambda^{(n)}(\nu)=M-\frac{1}{n+1}I_{n+1}, Λ ( n ) ( ν ) = M − n + 1 1 I n + 1 ,
we obtain
∥ Λ ( n ) ( ν ) ∥ 2 = tr ( M 2 ) − 1 n + 1 . \left\|\Lambda^{(n)}(\nu)\right\|^2=\operatorname{tr}(M^2)-\frac{1}{n+1}. Λ ( n ) ( ν ) 2 = tr ( M 2 ) − n + 1 1 .
It remains to estimate tr ( M 2 ) \operatorname{tr}(M^2) tr ( M 2 ) . Let r = rank M r=\operatorname{rank}M r = rank M . Since
Ran ( M ) ⊂ span { x 1 , … , x s } , \operatorname{Ran}(M)\subset\operatorname{span}\{x_1,\ldots,x_s\}, Ran ( M ) ⊂ span { x 1 , … , x s } ,
we have
r ≤ dim span { x 1 , … , x s } . r\leq\dim\operatorname{span}\{x_1,\ldots,x_s\}. r ≤ dim span { x 1 , … , x s } .
On the other hand, the relation ∑ a = 1 s w a x a = 0 \sum_{a=1}^{s}w_ax_a=0 ∑ a = 1 s w a x a = 0 is a nontrivial linear dependence relation among the vectors x 1 , … , x s x_1,\ldots,x_s x 1 , … , x s . Therefore,
r ≤ dim span { x 1 , … , x s } ≤ s − 1 ≤ k − 1. r\leq\dim\operatorname{span}\{x_1,\ldots,x_s\}\leq s-1\leq k-1. r ≤ dim span { x 1 , … , x s } ≤ s − 1 ≤ k − 1.
Let λ 1 , … , λ r \lambda_1,\ldots,\lambda_r λ 1 , … , λ r be the nonzero eigenvalues of M M M . Since ∑ i = 1 r λ i = tr M = 1 \sum_{i=1}^{r}\lambda_i=\operatorname{tr}M=1 ∑ i = 1 r λ i = tr M = 1 , the Cauchy–Schwarz inequality gives
tr ( M 2 ) = ∑ i = 1 r λ i 2 ≥ 1 r ( ∑ i = 1 r λ i ) 2 = 1 r . \operatorname{tr}(M^2)=\sum_{i=1}^{r}\lambda_i^2\geq\frac{1}{r}\left(\sum_{i=1}^{r}\lambda_i\right)^2=\frac{1}{r}. tr ( M 2 ) = i = 1 ∑ r λ i 2 ≥ r 1 ( i = 1 ∑ r λ i ) 2 = r 1 .
Combining this with r ≤ k − 1 r\leq k-1 r ≤ k − 1 , we obtain
tr ( M 2 ) ≥ 1 k − 1 . \operatorname{tr}(M^2)\geq\frac{1}{k-1}. tr ( M 2 ) ≥ k − 1 1 .
Consequently,
∥ Λ ( n ) ( ν ) ∥ 2 ≥ 1 k − 1 − 1 n + 1 , \|\Lambda^{(n)}(\nu)\|^2\geq\frac{1}{k-1}-\frac{1}{n+1}, ∥ Λ ( n ) ( ν ) ∥ 2 ≥ k − 1 1 − n + 1 1 ,
which proves
g n ( k ) ≥ 1 k − 1 − 1 n + 1 . g_n(k)\geq\frac{1}{k-1}-\frac{1}{n+1}. g n ( k ) ≥ k − 1 1 − n + 1 1 .
We next show that this lower bound is attained. Since k ≤ n + 2 k\leq n+2 k ≤ n + 2 , there exists a ( k − 1 ) (k-1) ( k − 1 ) -dimensional linear subspace V ⊂ R n + 1 V\subset\mathbb{R}^{n+1} V ⊂ R n + 1 and a regular ( k − 1 ) (k-1) ( k − 1 ) -simplex v 1 , … , v k ∈ V ∩ S n v_1,\ldots,v_k\in V\cap S^n v 1 , … , v k ∈ V ∩ S n satisfying
∑ a = 1 k v a = 0 , ⟨ v a , v b ⟩ = − 1 k − 1 ( a ≠ b ) . \sum_{a=1}^{k}v_a=0,\qquad\langle v_a,v_b\rangle=-\frac{1}{k-1}\quad(a\ne b). a = 1 ∑ k v a = 0 , ⟨ v a , v b ⟩ = − k − 1 1 ( a = b ) .
Let ν ∗ = 1 k ∑ a = 1 k δ v a \nu_*=\frac{1}{k}\sum_{a=1}^{k}\delta_{v_a} ν ∗ = k 1 ∑ a = 1 k δ v a . Then ν ∗ ∈ M 1 , n c \nu_*\in\mathcal{M}^c_{1,n} ν ∗ ∈ M 1 , n c . Moreover, the vertices of a regular simplex satisfy the identity
1 k ∑ a = 1 k v a v a T = 1 k − 1 P V , \frac{1}{k}\sum_{a=1}^{k}v_av_a^T=\frac{1}{k-1}P_V, k 1 a = 1 ∑ k v a v a T = k − 1 1 P V ,
where P V P_V P V is the orthogonal projection onto V V V . Hence
M ∗ = 1 k − 1 P V M_*=\frac{1}{k-1}P_V M ∗ = k − 1 1 P V
and therefore
tr ( M ∗ 2 ) = 1 ( k − 1 ) 2 tr ( P V ) = 1 k − 1 . \operatorname{tr}(M_*^2)=\frac{1}{(k-1)^2}\operatorname{tr}(P_V)=\frac{1}{k-1}. tr ( M ∗ 2 ) = ( k − 1 ) 2 1 tr ( P V ) = k − 1 1 .
Thus
∥ Λ ( n ) ( ν ∗ ) ∥ 2 = 1 k − 1 − 1 n + 1 , \|\Lambda^{(n)}(\nu_*)\|^2=\frac{1}{k-1}-\frac{1}{n+1}, ∥ Λ ( n ) ( ν ∗ ) ∥ 2 = k − 1 1 − n + 1 1 ,
which proves the desired formula for g n ( k ) g_n(k) g n ( k ) .
It remains to characterize the equality case. Suppose that an admissible measure ν \nu ν attains the above lower bound. Then equality must hold in the Cauchy–Schwarz inequality, and hence all nonzero eigenvalues of M M M are equal. Since
tr ( M 2 ) = 1 k − 1 , \operatorname{tr}(M^2)=\frac{1}{k-1}, tr ( M 2 ) = k − 1 1 ,
we must have rank M = k − 1 \operatorname{rank}M=k-1 rank M = k − 1 . In particular, the support of ν \nu ν contains exactly k k k points.
Let
A = ( w 1 x 1 , … , w k x k ) A = \left(\sqrt{w_1}x_1,\ldots,\sqrt{w_k}x_k\right) A = ( w 1 x 1 , … , w k x k )
and set
q = ( w 1 , … , w k ) T . q = \left(\sqrt{w_1},\ldots,\sqrt{w_k}\right)^{T}. q = ( w 1 , … , w k ) T .
The barycenter condition implies A q = 0 Aq=0 A q = 0 . Since rank A = k − 1 \operatorname{rank} A=k-1 rank A = k − 1 , we have ker A = span { q } \ker A=\operatorname{span}\{q\} ker A = span { q } . Moreover, A A T = M AA^{T}=M A A T = M has exactly k − 1 k-1 k − 1 nonzero eigenvalues, all equal to 1 / ( k − 1 ) 1/(k-1) 1/ ( k − 1 ) . Hence A T A A^{T}A A T A has the same nonzero eigenvalues and kernel span { q } \operatorname{span}\{q\} span { q } , which yields
A T A = 1 k − 1 ( I k − q q T ) . A^{T}A=\frac{1}{k-1}\left(I_k-qq^{T}\right). A T A = k − 1 1 ( I k − q q T ) .
Comparing the diagonal entries, we obtain
w i = 1 k − 1 ( 1 − w i ) , w_i=\frac{1}{k-1}(1-w_i), w i = k − 1 1 ( 1 − w i ) ,
and hence w i = 1 k w_i=\frac{1}{k} w i = k 1 for every i i i . Comparing the off-diagonal entries gives
w i w j ⟨ x i , x j ⟩ = − 1 k − 1 w i w j , \sqrt{w_iw_j}\langle x_i,x_j\rangle=-\frac{1}{k-1}\sqrt{w_iw_j}, w i w j ⟨ x i , x j ⟩ = − k − 1 1 w i w j ,
and therefore
⟨ x i , x j ⟩ = − 1 k − 1 , i ≠ j . \langle x_i,x_j\rangle=-\frac{1}{k-1},\qquad i\ne j. ⟨ x i , x j ⟩ = − k − 1 1 , i = j .
Thus the support points form a regular ( k − 1 ) (k-1) ( k − 1 ) -simplex, and the measure is the uniform measure on its vertices. Conversely, we have already shown that every such measure attains the lower bound. The proof is complete.
We are now ready to state our main result.
Theorem 6.4. Assume that s , n , k ∈ N + s,n,k\in\mathbb{N}_+ s , n , k ∈ N + , 1 ≤ s < n 1\le s<n 1 ≤ s < n , 3 ≤ k ≤ n + 2 3\le k\le n+2 3 ≤ k ≤ n + 2 , 0 < δ < 1 ( k − 1 ) ( k − 2 ) 0<\delta<\frac{1}{(k-1)(k-2)} 0 < δ < ( k − 1 ) ( k − 2 ) 1 , and 0 < ε < 1 0<\varepsilon<1 0 < ε < 1 . Then, for any function u ∈ W s s , n ( S n ) u\in W_s^{s,n}(\mathbb{S}^n) u ∈ W s s , n ( S n ) satisfying ∫ S n u d μ = 0 \int_{\mathbb{S}^n}u\,d\mu=0 ∫ S n u d μ = 0 , ∥ Λ ( n ) ( u ) ∥ 2 ≤ g n ( k − 1 ) − δ \|\Lambda^{(n)}(u)\|^2\le g_n(k-1)-\delta ∥ Λ ( n ) ( u ) ∥ 2 ≤ g n ( k − 1 ) − δ , and (17) with m = 1 m=1 m = 1 , we have
log ∫ S n e n u d μ ≤ ( α s , n k + ε ) ∥ Δ s 2 u ∥ L s n s ( S n ) n s + C ( n , s , δ , k ) ε − s n − s . (117) \log\int_{\mathbb{S}^n}e^{nu}\,d\mu\le\left(\frac{\alpha_{s,n}}{k}+\varepsilon\right)\|\Delta^{\frac{s}{2}}u\|_{L_s^{\frac{n}{s}}(\mathbb{S}^n)}^{\frac{n}{s}}+C(n,s,\delta,k)\varepsilon^{-\frac{s}{n-s}}. \tag*{(117)} log ∫ S n e n u d μ ≤ ( k α s , n + ε ) ∥ Δ 2 s u ∥ L s s n ( S n ) s n + C ( n , s , δ , k ) ε − n − s s . (117)
Standard constructions as in [10 , 15 , 29 , 50 ] show that (6.19) is almost sharp and that the upper bound g n ( k − 1 ) − δ g_n(k-1)-\delta g n ( k − 1 ) − δ cannot be replaced by any constant larger than g n ( k − 1 ) g_n(k-1) g n ( k − 1 ) . More precisely, choose probability measures ν k \nu_k ν k and ν k ′ \nu'_{k} ν k ′ attaining g n ( k ) g_n(k) g n ( k ) and g n ( k − 1 ) g_n(k-1) g n ( k − 1 ) , respectively. Following [15 Propositions 2.3 and 2.4] and [29 Theorem 3.3] , one constructs test functions concentrating near the supports of ν k \nu_k ν k and ν k ′ \nu'_k ν k ′ , respectively, and performs direct expansions. Since these computations are standard, we omit the details.
As in the proof of Theorem 6.2 , it suffices to combine Lemma 6.1 with the following result to deduce (6.19).
Lemma 6.5. Assume that n , k ∈ N + n,k\in\mathbb{N}_+ n , k ∈ N + , 3 ≤ k ≤ n + 2 3\le k\le n+2 3 ≤ k ≤ n + 2 , and 0 < δ < 1 ( k − 1 ) ( k − 2 ) 0<\delta<\frac{1}{(k-1)(k-2)} 0 < δ < ( k − 1 ) ( k − 2 ) 1 . If σ \sigma σ is a Radon probability measure on S n \mathbb{S}^n S n whose first moments vanish and such that ∥ Λ ( n ) ( σ ) ∥ 2 ≤ g n ( k − 1 ) − δ \|\Lambda^{(n)}(\sigma)\|^2\le g_n(k-1)-\delta ∥ Λ ( n ) ( σ ) ∥ 2 ≤ g n ( k − 1 ) − δ ,
then there exist positive constants γ 1 = γ 1 ( n ) \gamma_{1}=\gamma_{1}(n) γ 1 = γ 1 ( n ) and γ 2 = γ 2 ( n ) \gamma_{2}=\gamma_{2}(n) γ 2 = γ 2 ( n ) and k k k open subsets Ω 1 , … , Ω k ⊂ S n \Omega_{1},\ldots,\Omega_{k}\subset S^{n} Ω 1 , … , Ω k ⊂ S n satisfying dist ( Ω i , Ω j ) ≥ γ 1 δ \operatorname{dist}(\Omega_{i},\Omega_{j})\geq\gamma_{1}\delta dist ( Ω i , Ω j ) ≥ γ 1 δ whenever i ≠ j i\neq j i = j and σ ( Ω i ) ≥ γ 2 δ n + 1 \sigma(\Omega_{i})\geq\gamma_{2}\delta^{n+1} σ ( Ω i ) ≥ γ 2 δ n + 1 for every i i i .
Proof. For simplicity, let C n C_{n} C n denote a positive constant depending only on n n n and allowed to vary from line to line. By Lemma 6.3 , we have g n ( k − 1 ) = 1 k − 2 − 1 n + 1 g_{n}(k-1)=\frac{1}{k-2}-\frac{1}{n+1} g n ( k − 1 ) = k − 2 1 − n + 1 1 . Set M ( σ ) = ∫ S n x ⊗ x d σ M(\sigma)=\int_{S^{n}}x\otimes x\,\mathrm{d}\sigma M ( σ ) = ∫ S n x ⊗ x d σ . The inequality ∥ Λ ( n ) ( σ ) ∥ 2 ≤ g n ( k − 1 ) − δ \lVert\Lambda^{(n)}(\sigma)\rVert^{2}\leq g_{n}(k-1)-\delta ∥ Λ ( n ) ( σ ) ∥ 2 ≤ g n ( k − 1 ) − δ gives
tr ( M ( σ ) 2 ) ≤ 1 k − 2 − δ . (118) \operatorname{tr}(M(\sigma)^{2})\leq\frac{1}{k-2}-\delta. \tag*{(118)} tr ( M ( σ ) 2 ) ≤ k − 2 1 − δ . (118)
We call a spherical ball B ( x , γ 1 δ ) B(x,\gamma_{1}\delta) B ( x , γ 1 δ ) heavy if σ ( B ( x , γ 1 δ ) ) ≥ γ 2 δ n + 1 \sigma(B(x,\gamma_{1}\delta))\geq\gamma_{2}\delta^{n+1} σ ( B ( x , γ 1 δ )) ≥ γ 2 δ n + 1 . Here γ 1 = γ 1 ( n ) \gamma_{1}=\gamma_{1}(n) γ 1 = γ 1 ( n ) and γ 2 = γ 2 ( n ) \gamma_{2}=\gamma_{2}(n) γ 2 = γ 2 ( n ) are positive constants to be determined. Choose a maximal family of heavy balls B ( x 1 , γ 1 δ ) , … , B ( x ℓ , γ 1 δ ) B(x_{1},\gamma_{1}\delta),\ldots,B(x_{\ell},\gamma_{1}\delta) B ( x 1 , γ 1 δ ) , … , B ( x ℓ , γ 1 δ ) whose centers satisfy ∣ x i − x j ∣ ≥ 4 γ 1 δ \lvert x_{i}-x_{j}\rvert\geq4\gamma_{1}\delta ∣ x i − x j ∣ ≥ 4 γ 1 δ whenever i ≠ j i\neq j i = j . If ℓ ≥ k \ell\geq k ℓ ≥ k , then any k k k of these balls give the desired conclusion. Suppose, for contradiction, that ℓ ≤ k − 1 \ell\leq k-1 ℓ ≤ k − 1 . Define
U = ⋃ i = 1 ℓ B ( x i , 6 γ 1 δ ) . U=\bigcup_{i=1}^{\ell}B(x_{i},6\gamma_{1}\delta). U = i = 1 ⋃ ℓ B ( x i , 6 γ 1 δ ) .
Choose a maximal γ 1 δ \gamma_{1}\delta γ 1 δ -separated set { y 1 , … , y N } \{y_{1},\ldots,y_{N}\} { y 1 , … , y N } in S n ∖ U S^{n}\setminus U S n ∖ U . A standard covering estimate gives N ≤ C n ( γ 1 δ ) − n N\leq C_{n}(\gamma_{1}\delta)^{-n} N ≤ C n ( γ 1 δ ) − n . Moreover, the balls B ( y j , γ 1 δ ) B(y_{j},\gamma_{1}\delta) B ( y j , γ 1 δ ) are not heavy, since otherwise they could be added to the maximal family. Hence σ ( B ( y j , γ 1 δ ) ) < γ 2 δ n + 1 \sigma(B(y_{j},\gamma_{1}\delta))<\gamma_{2}\delta^{n+1} σ ( B ( y j , γ 1 δ )) < γ 2 δ n + 1 . Because these balls cover S n ∖ U S^{n}\setminus U S n ∖ U , we obtain
σ ( S n ∖ U ) ≤ C n ( γ 1 δ ) − n γ 2 δ n + 1 = C n γ 1 − n γ 2 δ . (119) \sigma(S^{n}\setminus U)\leq C_{n}(\gamma_{1}\delta)^{-n}\gamma_{2}\delta^{n+1}=C_{n}\gamma_{1}^{-n}\gamma_{2}\delta. \tag*{(119)} σ ( S n ∖ U ) ≤ C n ( γ 1 δ ) − n γ 2 δ n + 1 = C n γ 1 − n γ 2 δ . (119)
Let E 1 , … , E ℓ E_{1},\ldots,E_{\ell} E 1 , … , E ℓ be a disjoint Borel partition of U U U satisfying E i ⊂ B ( x i , 6 γ 1 δ ) E_{i}\subset B(x_{i},6\gamma_{1}\delta) E i ⊂ B ( x i , 6 γ 1 δ ) and set
p i = σ ( E i ) , P = ∑ i = 1 ℓ p i . p_{i}=\sigma(E_{i}),\qquad P=\sum_{i=1}^{\ell}p_{i}. p i = σ ( E i ) , P = i = 1 ∑ ℓ p i .
Then, by (119) , we have P ≥ 1 − C n γ 1 − n γ 2 δ P\geq1-C_{n}\gamma_{1}^{-n}\gamma_{2}\delta P ≥ 1 − C n γ 1 − n γ 2 δ . Define
a = ∑ i = 1 ℓ p i x i , C = ∑ i = 1 ℓ p i x i x i T . a=\sum_{i=1}^{\ell}p_{i}x_{i},\qquad C=\sum_{i=1}^{\ell}p_{i}x_{i}x_{i}^{T}. a = i = 1 ∑ ℓ p i x i , C = i = 1 ∑ ℓ p i x i x i T .
Since the first moments of σ \sigma σ vanish, we can write
a = ∑ i ∫ E i ( x i − x ) d σ − ∫ S n ∖ U x d σ . a=\sum_{i}\int_{E_{i}}(x_{i}-x)\,\mathrm{d}\sigma-\int_{S^{n}\setminus U}x\,\mathrm{d}\sigma. a = i ∑ ∫ E i ( x i − x ) d σ − ∫ S n ∖ U x d σ .
Therefore, ∣ a ∣ ≤ 6 γ 1 δ + C n γ 1 − n γ 2 δ \lvert a\rvert\leq6\gamma_{1}\delta+C_{n}\gamma_{1}^{-n}\gamma_{2}\delta ∣ a ∣ ≤ 6 γ 1 δ + C n γ 1 − n γ 2 δ . Similarly, we have
M ( σ ) − C = ∑ i ∫ E i ( x x T − x i x i T ) d σ + ∫ S n ∖ U x x T d σ . M(\sigma)-C=\sum_{i}\int_{E_{i}}(xx^{T}-x_{i}x_{i}^{T})\,\mathrm{d}\sigma+\int_{S^{n}\setminus U}xx^{T}\,\mathrm{d}\sigma. M ( σ ) − C = i ∑ ∫ E i ( x x T − x i x i T ) d σ + ∫ S n ∖ U x x T d σ .
By Hölder’s inequality we have ∥ M ( σ ) − C ∥ ≤ C n γ 1 δ + C n γ 1 − n γ 2 δ \lVert M(\sigma)-C\rVert\leq C_{n}\gamma_{1}\delta+C_{n}\gamma_{1}^{-n}\gamma_{2}\delta ∥ M ( σ ) − C ∥ ≤ C n γ 1 δ + C n γ 1 − n γ 2 δ .
Now let A A A be the matrix whose columns are p i x i \sqrt{p_i}x_i p i x i , and let
v = ( p 1 , … , p ℓ ) T . v = (\sqrt{p_1},\ldots,\sqrt{p_\ell})^T. v = ( p 1 , … , p ℓ ) T .
Then
A A T = C , A v = a , ∣ v ∣ 2 = P . AA^T=C,\qquad Av=a,\qquad|v|^2=P. A A T = C , A v = a , ∣ v ∣ 2 = P .
Set A T A = G A^TA=G A T A = G and let λ 1 \lambda_1 λ 1 be the smallest eigenvalue. By the Rayleigh quotient and the estimates of ∣ a ∣ |a| ∣ a ∣ and P P P above, we obtain
λ 1 ≤ v T G v ∣ v ∣ 2 = ∣ a ∣ 2 P ≤ e : = ( 6 γ 1 δ + C n γ 1 − n γ 2 δ ) 2 1 − C n γ 1 − n γ 2 δ . (120) \lambda_1\leq\frac{v^TGv}{|v|^2}=\frac{|a|^2}{P}\leq e:=\frac{(6\gamma_1\delta+C_n\gamma_1^{-n}\gamma_2\delta)^2}{1-C_n\gamma_1^{-n}\gamma_2\delta}. \tag*{(120)} λ 1 ≤ ∣ v ∣ 2 v T G v = P ∣ a ∣ 2 ≤ e := 1 − C n γ 1 − n γ 2 δ ( 6 γ 1 δ + C n γ 1 − n γ 2 δ ) 2 . (120)
Since ℓ ≤ k − 1 \ell\leq k-1 ℓ ≤ k − 1 , after removing the smallest eigenvalue, G G G has at most k − 2 k-2 k − 2 nonzero eigenvalues. By (120) , their sum is at least P − e P-e P − e . Therefore, by the Cauchy–Schwarz inequality, we have
tr ( C 2 ) = tr ( G 2 ) ≥ ( P − e ) 2 k − 2 . \operatorname{tr}(C^2)=\operatorname{tr}(G^2)\geq\frac{(P-e)^2}{k-2}. tr ( C 2 ) = tr ( G 2 ) ≥ k − 2 ( P − e ) 2 .
Since both M ( σ ) M(\sigma) M ( σ ) and C C C are positive semidefinite and have trace bounded by one,
∣ tr ( M ( σ ) 2 ) − tr ( C 2 ) ∣ ≤ 2 ∥ M ( σ ) − C ∥ ≤ 2 C n γ 1 δ + 2 C n γ 1 − n γ 2 δ . \left|\operatorname{tr}(M(\sigma)^2)-\operatorname{tr}(C^2)\right|\leq2\lVert M(\sigma)-C\rVert\leq2C_n\gamma_1\delta+2C_n\gamma_1^{-n}\gamma_2\delta. tr ( M ( σ ) 2 ) − tr ( C 2 ) ≤ 2 ∥ M ( σ ) − C ∥ ≤ 2 C n γ 1 δ + 2 C n γ 1 − n γ 2 δ .
Combining the preceding two estimates yields
tr ( M ( σ ) 2 ) ≥ ( P − e ) 2 k − 2 − ( 2 C n γ 1 δ + 2 C n γ 1 − n γ 2 δ ) . \operatorname{tr}(M(\sigma)^2)\geq\frac{(P-e)^2}{k-2}-(2C_n\gamma_1\delta+2C_n\gamma_1^{-n}\gamma_2\delta). tr ( M ( σ ) 2 ) ≥ k − 2 ( P − e ) 2 − ( 2 C n γ 1 δ + 2 C n γ 1 − n γ 2 δ ) .
Using the estimate P ≥ 1 − C n γ 1 − n γ 2 δ P\geq1-C_n\gamma_1^{-n}\gamma_2\delta P ≥ 1 − C n γ 1 − n γ 2 δ and the expression for e e e in (120) , we deduce, provided that both γ 1 \gamma_1 γ 1 and γ 1 − n γ 2 \gamma_1^{-n}\gamma_2 γ 1 − n γ 2 are sufficiently small, that
tr ( M ( σ ) 2 ) ≥ 1 k − 2 − δ 2 , \operatorname{tr}(M(\sigma)^2)\geq\frac{1}{k-2}-\frac{\delta}{2}, tr ( M ( σ ) 2 ) ≥ k − 2 1 − 2 δ ,
which contradicts the estimate (118) . The proof is thus complete.
An argument similar to the proof of Corollary 1.9 also gives the following sharp improved inequality at a different scale.
Corollary 6.6. Assume that s , n , k ∈ N + s,n,k\in\mathbb{N}_+ s , n , k ∈ N + , 1 ≤ s < n 1\leq s<n 1 ≤ s < n , 3 ≤ k ≤ n + 2 3\leq k\leq n+2 3 ≤ k ≤ n + 2 , and 0 < δ < 1 ( k − 1 ) ( k − 2 ) 0<\delta<\frac{1}{(k-1)(k-2)} 0 < δ < ( k − 1 ) ( k − 2 ) 1 . Then, for any function u ∈ W s , n s ( S n ) u\in W^{s,\frac{n}{s}}(S^n) u ∈ W s , s n ( S n ) satisfying ∫ S n u d μ = 0 \int_{S^n}u\,d\mu=0 ∫ S n u d μ = 0 , ∥ Λ ( n ) ( u ) ∥ 2 ≤ g n ( k − 1 ) − δ \lVert\Lambda^{(n)}(u)\rVert^2\leq g_n(k-1)-\delta ∥ Λ ( n ) ( u ) ∥ 2 ≤ g n ( k − 1 ) − δ , and (17) with m = 1 m=1 m = 1 , we have
log 1 − s n ( 1 ω n ∫ S n e n u d μ ) ≤ ( α s , n k ) 1 − s n ∥ Λ s 2 u ∥ L n s ( S n ) n s + C ( n , s , δ , k ) . \log^{1-\frac{s}{n}}\left(\frac{1}{\omega_n}\int_{S^n}e^{nu}\,d\mu\right)\leq\left(\frac{\alpha_{s,n}}{k}\right)^{1-\frac{s}{n}}\lVert\Lambda^{\frac{s}{2}}u\rVert_{L^{\frac{n}{s}}(S^n)}^{\frac{n}{s}}+C(n,s,\delta,k). log 1 − n s ( ω n 1 ∫ S n e n u d μ ) ≤ ( k α s , n ) 1 − n s ∥ Λ 2 s u ∥ L s n ( S n ) s n + C ( n , s , δ , k ) .
Improved Sobolev-type inequalities under symmetry constraints
In this section, we study how symmetry can improve Sobolev-type embeddings. The first systematic investigation of this subject on R n \mathbb{R}^{n} R n was carried out by Lions in the celebrated work [40 ] . Here we focus on smooth closed connected manifolds. In general, two types of improvement are possible for functions invariant under the action of a group G G G .
First, if every orbit of G G G is infinite, one may obtain Sobolev-type embeddings into higher L p L^{p} L p spaces and, in particular, compact embeddings. Second, if G G G has finite orbits, embeddings into higher L p L^{p} L p spaces are no longer expected; one may instead seek improved embedding constants. We focus on the second type. We refer to [31 Chapter 5] , [32 Chapter 9] , and [20 , 34 ] for further discussion, including detailed treatments of noncompact manifolds.
We begin with some notation and preliminaries; see also [32 Section 9.1] . Let ( M , g ) (M,g) ( M , g ) be a smooth closed manifold of dimension n ≥ 2 n \ge2 n ≥ 2 , and let Iso ( M ) \operatorname{Iso}(M) Iso ( M ) denote its isometry group. Suppose that G G G is a compact subgroup of Iso ( M ) \operatorname{Iso}(M) Iso ( M ) . For x ∈ M x \in M x ∈ M , define its orbit by
O G x = { g x : g ∈ G } . O_{G}^{x} = \{gx : g \in G\}. O G x = { g x : g ∈ G } .
We set
k G = min x ∈ M ∣ O G x ∣ . k_{G} = \min_{x \in M} \lvert O_{G}^{x} \rvert. k G = x ∈ M min ∣ O G x ∣ .
where ∣ O G x ∣ \lvert O_{G}^{x} \rvert ∣ O G x ∣ denotes the cardinality of O G x O_{G}^{x} O G x . Throughout this section, we assume k G < + ∞ k_{G} < +\infty k G < + ∞ . For s ∈ N + s \in\mathbb{N}_{+} s ∈ N + and 1 ≤ p ≤ + ∞ 1 \le p \le+\infty 1 ≤ p ≤ + ∞ , define
W G s , p ( M ) = { u ∈ W s , p ( M ) : u ∘ g = u for all g ∈ G } . W_{G}^{s,p}(M) = \{u \in W^{s,p}(M) : u \circ g = u \text{ for all } g \in G\}. W G s , p ( M ) = { u ∈ W s , p ( M ) : u ∘ g = u for all g ∈ G } .
In the pioneering work [31 Theorem 5.8] , Hebey and Vaugon established the following improved almost sharp Sobolev inequality. If 1 ≤ p < n 1 \le p < n 1 ≤ p < n and u ∈ W G 1 , p ( M ) u \in W_{G}^{1,p}(M) u ∈ W G 1 , p ( M ) , then, for every 0 < ε < 1 0 < \varepsilon< 1 0 < ε < 1 , there exists a positive constant b ε b_{\varepsilon} b ε such that
∥ u ∥ L p ∗ ( M ) p ≤ ( S n , p p k G p / n + ε ) ∥ ∇ u ∥ L p ( M ) p + b ε ∥ u ∥ L p ( M ) p . (121) \lVert u\rVert_{L^{p^{*}}(M)}^{p} \le\left(\frac{S_{n,p}^{p}}{k_{G}^{p/n}}+\varepsilon\right)\lVert\nabla u\rVert_{L^{p}(M)}^{p}+b_{\varepsilon}\lVert u\rVert_{L^{p}(M)}^{p}. \tag*{(121)} ∥ u ∥ L p ∗ ( M ) p ≤ ( k G p / n S n , p p + ε ) ∥ ∇ u ∥ L p ( M ) p + b ε ∥ u ∥ L p ( M ) p . (121)
In the special case p = 2 p=2 p = 2 , they also proved the sharp inequality (see [31 Theorem 5.10] )
∥ u ∥ L p ∗ ( M ) 2 ≤ S n , 2 2 k G 2 / n ∥ ∇ u ∥ L 2 ( M ) 2 + b ∥ u ∥ L 2 ( M ) 2 (122) \lVert u\rVert_{L^{p^{*}}(M)}^{2} \le\frac{S_{n,2}^{2}}{k_{G}^{2/n}}\lVert\nabla u\rVert_{L^{2}(M)}^{2}+b\lVert u\rVert_{L^{2}(M)}^{2} \tag*{(122)} ∥ u ∥ L p ∗ ( M ) 2 ≤ k G 2/ n S n , 2 2 ∥ ∇ u ∥ L 2 ( M ) 2 + b ∥ u ∥ L 2 ( M ) 2 (122)
for some positive constant b b b . Their proof relies on a careful construction of a partition of unity and a detailed local analysis along the G G G -orbits.
In [20 Theorem 2] , Faget established a first-order Moser–Trudinger–Onofri analogue of (121) . More precisely, if u ∈ W G 1 , n ( M ) u \in W_{G}^{1,n}(M) u ∈ W G 1 , n ( M ) and ∫ M u d μ = 0 \int_{M} u\,\mathrm{d}\mu= 0 ∫ M u d μ = 0 , then, for every 0 < ε < 1 0 < \varepsilon< 1 0 < ε < 1 , there exists a positive constant c ε c_{\varepsilon} c ε such that
log ∫ M e n u d μ ≤ ( α 1 , n k G + ε ) ∥ ∇ u ∥ L n ( M ) n + c ε . (123) \log\int_{M} e^{nu}\,\mathrm{d}\mu\le\left(\frac{\alpha_{1,n}}{k_{G}}+\varepsilon\right)\lVert\nabla u\rVert_{L^{n}(M)}^{n}+c_{\varepsilon}. \tag*{(123)} log ∫ M e n u d μ ≤ ( k G α 1 , n + ε ) ∥ ∇ u ∥ L n ( M ) n + c ε . (123)
This inequality is almost sharp . Its proof in [20 ] combines the techniques developed in [31 ] with a truncation method introduced in [5 ] .
The strategies used to prove Theorems 5.1 and 6.2 yield the following generalizations of (122) and (123) .
Theorem 7.1. Assume 1 ≤ p < n 1 \le p < n 1 ≤ p < n and u ∈ W G 1 , p ( M ) u \in W^{1,p}_{G}(M) u ∈ W G 1 , p ( M ) . Suppose that the sharp Sobolev inequality (6) holds. If 1 ≤ p ≤ 2 1 \le p \le2 1 ≤ p ≤ 2 , then
∥ u ∥ L p ∗ ( M ) p ≤ S n , p p k G p / n ∥ ∇ u ∥ L p ( M ) p + b G ( M , p ) ∥ u ∥ L p ( M ) p . \lVert u\rVert_{L^{p^*}(M)}^p \le\frac{S_{n,p}^p}{k_G^{p/n}}\lVert\nabla u\rVert_{L^p(M)}^p+b_G(M,p)\lVert u\rVert_{L^p(M)}^p. ∥ u ∥ L p ∗ ( M ) p ≤ k G p / n S n , p p ∥ ∇ u ∥ L p ( M ) p + b G ( M , p ) ∥ u ∥ L p ( M ) p . whereas if 2 < p < n 2 < p < n 2 < p < n , then, for every 0 < ε < 1 0 < \varepsilon< 1 0 < ε < 1 ,
∥ u ∥ L p ∗ ( M ) p ≤ ( S n , p p k G p / n + ε ) ∥ ∇ u ∥ L p ( M ) p + b G ( M , p ) ε 1 − p 2 ∥ u ∥ L p ( M ) p . \lVert u\rVert_{L^{p^*}(M)}^p \le\left(\frac{S_{n,p}^p}{k_G^{p/n}}+\varepsilon\right)\lVert\nabla u\rVert_{L^p(M)}^p+b_G(M,p)\varepsilon^{1-\frac{p}{2}}\lVert u\rVert_{L^p(M)}^p. ∥ u ∥ L p ∗ ( M ) p ≤ ( k G p / n S n , p p + ε ) ∥ ∇ u ∥ L p ( M ) p + b G ( M , p ) ε 1 − 2 p ∥ u ∥ L p ( M ) p .
Theorem 7.2. Assume s ∈ N + s \in\mathbb{N}_+ s ∈ N + , 1 ≤ s < n 1 \le s < n 1 ≤ s < n , and u ∈ W G s , n s ( M ) u \in W^{s,\frac{n}{s}}_{G}(M) u ∈ W G s , s n ( M ) with ∫ M u d μ = 0 \int_M u\,d\mu=0 ∫ M u d μ = 0 . Then, for every 0 < ε < 1 0 < \varepsilon< 1 0 < ε < 1 ,
log ∫ M e n u d μ ≤ ( α s , n k G + ε ) ∥ Δ s 2 u ∥ L n s ( M ) n s + c G ( M , s ) ε − s n − s . \log\int_M e^{nu}\,d\mu\le\left(\frac{\alpha_{s,n}}{k_G}+\varepsilon\right)\lVert\Delta^{\frac{s}{2}}u\rVert_{L^{\frac{n}{s}}(M)}^{\frac{n}{s}}+c_G(M,s)\varepsilon^{-\frac{s}{n-s}}. log ∫ M e n u d μ ≤ ( k G α s , n + ε ) ∥ Δ 2 s u ∥ L s n ( M ) s n + c G ( M , s ) ε − n − s s .
Theorem 7.1 gives a general criterion for the validity of improved sharp Sobolev inequalities under symmetry constraints. In particular, (122) is a special case of Theorem 7.1, because Hebey–Vaugon [33 ] proved that, when p = 2 p=2 p = 2 , the sharp Sobolev inequality (6) holds on every smooth closed manifold. To our knowledge, Theorem 7.1 is the first general sharpness result since the work of Hebey–Vaugon [33 ] . Theorem 7.2, in turn, extends (123) to the higher-order setting.
The key ingredients in the proofs are the following three lemmas.
Lemma 7.3. Let σ \sigma σ be a Borel probability measure on M M M . If σ \sigma σ is invariant under the action of G G G , then there exists r G > 0 r_G > 0 r G > 0 such that
σ ( B ( x , r G ) ) ≤ 1 k G \sigma(B(x,r_G)) \le\frac{1}{k_G} σ ( B ( x , r G )) ≤ k G 1 for any x ∈ M x \in M x ∈ M .
Proof. Fix x ∈ M x \in M x ∈ M . Since ∣ O G x ∣ ≥ k G \lvert O_G^x\rvert\ge k_G ∣ O G x ∣ ≥ k G , choose distinct points y 1 = x , y 2 , … , y k G y_1=x,y_2,\ldots,y_{k_G} y 1 = x , y 2 , … , y k G in O G x O_G^x O G x . There exists r x > 0 r_x>0 r x > 0 such that
B ( y i , r x ) ∩ B ( y j , r x ) = ∅ for i ≠ j . B(y_i,r_x)\cap B(y_j,r_x)=\varnothing\qquad\text{for } i\ne j. B ( y i , r x ) ∩ B ( y j , r x ) = ∅ for i = j . Because σ \sigma σ is invariant under the action of G G G , we have
σ ( B ( y i , r x ) ) = σ ( B ( y j , r x ) ) \sigma(B(y_i,r_x))=\sigma(B(y_j,r_x)) σ ( B ( y i , r x )) = σ ( B ( y j , r x )) for all i ≠ j i\ne j i = j . Hence
σ ( B ( x , r x ) ) ≤ 1 k G . \sigma(B(x,r_x))\le\frac{1}{k_G}. σ ( B ( x , r x )) ≤ k G 1 .
Covering M M M by finitely many balls B ( x i , r x i 2 ) B(x_i,\frac{r^{x_i}}{2}) B ( x i , 2 r x i ) , 1 ≤ i ≤ I 1 \le i \le I 1 ≤ i ≤ I , and setting r G : = min i { r x i 2 } r_G := \min_i\{\frac{r^{x_i}}{2}\} r G := min i { 2 r x i } gives the desired conclusion.
Lemma 7.4
Lemma 7.4. Assume N > 0 N > 0 N > 0 , 1 ≤ p < n 1 \le p < n 1 ≤ p < n , r > 0 r > 0 r > 0 , and σ \sigma σ is a Borel probability measure on M M M satisfying
σ ( B ( x , r ) ) ≤ 1 N for all x ∈ M . (124) \sigma(B(x,r)) \le\frac{1}{N} \quad\text{for all } x \in M. \tag*{(124)} σ ( B ( x , r )) ≤ N 1 for all x ∈ M . (124)
Then there exist finitely many points x j ∈ M x_j \in M x j ∈ M , 1 ≤ j ≤ J 1 \le j \le J 1 ≤ j ≤ J , and an ℓ p \ell^p ℓ p -partition of unity { η j } j = 1 J \{\eta_j\}_{j=1}^J { η j } j = 1 J such that supp η j ⊂ B ( x j , r ) \operatorname{supp}\eta_j \subset B(x_j,r) supp η j ⊂ B ( x j , r ) ,
∑ j = 1 J η j p = 1 , ∑ j = 1 J ∣ ∇ η j ∣ p ≤ C r − p , ∑ j = 1 J β j n − p n ≥ N p n , (125) \sum_{j=1}^{J}\eta_j^p = 1,\qquad\sum_{j=1}^{J}|\nabla\eta_j|^p \le Cr^{-p},\qquad\sum_{j=1}^{J}\beta_j^{\frac{n-p}{n}} \ge N^{\frac{p}{n}}, \tag*{(125)} j = 1 ∑ J η j p = 1 , j = 1 ∑ J ∣∇ η j ∣ p ≤ C r − p , j = 1 ∑ J β j n n − p ≥ N n p , (125)
where β j : = ∫ M η j p ∗ d σ \beta_j := \int_M \eta_j^{p^*}\,d\sigma β j := ∫ M η j p ∗ d σ and C = C ( M , p ) C = C(M,p) C = C ( M , p ) .
Proof. Throughout the proof, C C C denotes a positive constant depending only on M M M and p p p , whose value may vary from line to line. Take a maximal r 8 \frac{r}{8} 8 r -separated set { x j } j = 1 J \{x_j\}_{j=1}^J { x j } j = 1 J of M M M . We can find a family of smooth functions { φ j } j = 1 J \{\varphi_j\}_{j=1}^J { φ j } j = 1 J such that supp φ j ⊂ B ( x j , r ) \operatorname{supp}\varphi_j \subset B(x_j,r) supp φ j ⊂ B ( x j , r ) , 0 ≤ φ j ≤ 1 0 \le\varphi_j \le1 0 ≤ φ j ≤ 1 , φ j = 1 \varphi_j = 1 φ j = 1 on B ( x j , r 2 ) B(x_j,\frac{r}{2}) B ( x j , 2 r ) , and ∣ ∇ φ j ∣ ≤ C r − 1 |\nabla\varphi_j| \le Cr^{-1} ∣∇ φ j ∣ ≤ C r − 1 . Define
η j = φ j ( ∑ k = 1 J φ k p ) 1 / p . \eta_j = \frac{\varphi_j}{\left(\sum_{k=1}^{J}\varphi_k^p\right)^{1/p}}. η j = ( ∑ k = 1 J φ k p ) 1/ p φ j .
Then ∑ j = 1 J η j p = 1 \sum_{j=1}^{J}\eta_j^p = 1 ∑ j = 1 J η j p = 1 . By the choice of { x j } j = 1 J \{x_j\}_{j=1}^J { x j } j = 1 J , every point y ∈ M y \in M y ∈ M belongs to at most C C C of the balls B ( x i , r ) B(x_i,r) B ( x i , r ) . Therefore,
∑ j = 1 J ∣ ∇ η j ∣ p ≤ C r − p . \sum_{j=1}^{J}|\nabla\eta_j|^p \le Cr^{-p}. j = 1 ∑ J ∣∇ η j ∣ p ≤ C r − p .
It remains to prove the estimate involving β j \beta_j β j . By Hölder’s inequality and (124) ,
∫ M η j p d σ ≤ ( ∫ M η j p ∗ d σ ) p / p ∗ σ ( supp η j ) 1 − p / p ∗ ≤ β j p / p ∗ N − p / n . \int_M \eta_j^p\,d\sigma\le\left(\int_M \eta_j^{p^*}\,d\sigma\right)^{p/p^*}\sigma(\operatorname{supp}\eta_j)^{1-p/p^*} \le\beta_j^{p/p^*}N^{-p/n}. ∫ M η j p d σ ≤ ( ∫ M η j p ∗ d σ ) p / p ∗ σ ( supp η j ) 1 − p / p ∗ ≤ β j p / p ∗ N − p / n .
Summing over j j j and using ∑ j = 1 J η j p = 1 \sum_{j=1}^{J}\eta_j^p = 1 ∑ j = 1 J η j p = 1 , we obtain
∑ j = 1 J β j n − p n ≥ N p / n , \sum_{j=1}^{J}\beta_j^{\frac{n-p}{n}} \ge N^{p/n}, j = 1 ∑ J β j n n − p ≥ N p / n ,
which completes the proof.
Lemma 7.5
Lemma 7.5. Let σ \sigma σ be a Borel probability measure on M M M . If σ \sigma σ is invariant under the action of G G G , then there exist open subsets Ω 1 , … , Ω k G ⊂ M \Omega_1,\ldots,\Omega_{k_G} \subset M Ω 1 , … , Ω k G ⊂ M and positive constants R = R ( M , G ) R = R(M,G) R = R ( M , G ) and δ = δ ( M , G ) \delta= \delta(M,G) δ = δ ( M , G ) such that dist ( Ω i , Ω j ) ≥ R \operatorname{dist}(\Omega_i,\Omega_j) \ge R dist ( Ω i , Ω j ) ≥ R whenever i ≠ j i \ne j i = j , and σ ( Ω i ) ≥ δ \sigma(\Omega_i) \ge\delta σ ( Ω i ) ≥ δ for every 1 ≤ i ≤ k G 1 \le i \le k_G 1 ≤ i ≤ k G . Proof. By Lemma 7.3 , the argument is similar to that used in the proof of Lemma 4.3 ; we therefore omit the details.
Replacing Proposition 5.7 and Lemma 4.3 by Lemmas 7.4 and 7.5, respectively, in the proofs of Theorems 5.1 and 6.2 yields Theorems 7.1 and 7.2 . We omit the details.
Arguments similar to those used in the proofs of Corollaries 1.8 and 1.9 also give the following sharp improved inequalities at different scales.
Corollary 7.6. Assume that 2 < p < n 2 < p < n 2 < p < n and u ∈ W G 1 , p ( M ) u \in W_{G}^{1,p}(M) u ∈ W G 1 , p ( M ) . If {(1.6)} holds, then
∥ u ∥ L p ∗ ( M ) 2 ≤ ( S n , p k G 1 / n ) 2 ∥ ∇ u ∥ L p ( M ) 2 + D ( M , p , G ) ∥ u ∥ L p ( M ) 2 . \lVert u\rVert_{L^{p^{*}}(M)}^{2} \leq\left(\frac{S_{n,p}}{k_{G}^{1/n}}\right)^{2}\lVert\nabla u\rVert_{L^{p}(M)}^{2} + D(M,p,G)\lVert u\rVert_{L^{p}(M)}^{2}. ∥ u ∥ L p ∗ ( M ) 2 ≤ ( k G 1/ n S n , p ) 2 ∥ ∇ u ∥ L p ( M ) 2 + D ( M , p , G ) ∥ u ∥ L p ( M ) 2 .
Corollary 7.7. Let s ∈ N + s \in\mathbb{N}_{+} s ∈ N + with 1 ≤ s < n 1 \leq s < n 1 ≤ s < n , and let u ∈ W G s , n s ( M ) u \in W_{G}^{s,\frac{n}{s}}(M) u ∈ W G s , s n ( M ) satisfy ∫ M u d μ = 0 \int_{M}u\,\mathrm{d}\mu= 0 ∫ M u d μ = 0 . Then
log 1 − s n ( 1 V ( M ) ∫ M e n u d μ ) ≤ ( α s , n k G ) 1 − s n ∥ Δ s 2 u ∥ L n s ( M ) n s − 1 + D ( M , s , G ) . (126) \log^{1-\frac{s}{n}}\left(\frac{1}{V(M)}\int_{M}e^{nu}\,\mathrm{d}\mu\right) \leq\left(\frac{\alpha_{s,n}}{k_{G}}\right)^{1-\frac{s}{n}}\lVert\Delta^{\frac{s}{2}}u\rVert_{L^{\frac{n}{s}}(M)}^{\frac{n}{s}-1} + D(M,s,G). \tag*{(126)} log 1 − n s ( V ( M ) 1 ∫ M e n u d μ ) ≤ ( k G α s , n ) 1 − n s ∥ Δ 2 s u ∥ L s n ( M ) s n − 1 + D ( M , s , G ) . (126)
where V ( M ) V(M) V ( M ) denotes the volume of ( M , g ) (M,g) ( M , g ) .
We conclude this section with a general criterion for the failure of sharp improved Sobolev inequalities under symmetry constraints. To our knowledge, it gives the first non-sharp case of an improved Sobolev inequality under symmetry constraints.
Theorem 7.8. Assume 2 < p < n 2 < p < \sqrt{n} 2 < p < n and that there exists a point x ∈ M x \in M x ∈ M such that k G = ∣ O G x ∣ k_{G} = \lvert O_{G}^{x}\rvert k G = ∣ O G x ∣ and the scalar curvature at x x x is positive. Then there is no finite constant b b b such that
∥ u ∥ L p ∗ ( M ) p ≤ S n , p p k G p / n ∥ ∇ u ∥ L p ( M ) p + b ∥ u ∥ L p ( M ) p \lVert u\rVert_{L^{p^{*}}(M)}^{p} \leq\frac{S_{n,p}^{p}}{k_{G}^{p/n}}\lVert\nabla u\rVert_{L^{p}(M)}^{p} + b\lVert u\rVert_{L^{p}(M)}^{p} ∥ u ∥ L p ∗ ( M ) p ≤ k G p / n S n , p p ∥ ∇ u ∥ L p ( M ) p + b ∥ u ∥ L p ( M ) p holds for every u ∈ W G 1 , p ( M ) u \in W_{G}^{1,p}(M) u ∈ W G 1 , p ( M ) .
Proof. The proof is analogous to that of Theorem 1.2 . The proof of Theorem 1 gives a sequence of nonnegative functions u k ∈ C c ∞ ( M ) u_{k} \in C_{c}^{\infty}(M) u k ∈ C c ∞ ( M ) such that supp u k ⊂ B ( x , k − 1 ) \operatorname{supp}u_{k} \subset B(x,k^{-1}) supp u k ⊂ B ( x , k − 1 ) , each u k ( y ) u_{k}(y) u k ( y ) depends only on d g ( x , y ) d_{g}(x,y) d g ( x , y ) , and
∥ u k ∥ L p ∗ ( M ) p − S n , p p ∥ ∇ u k ∥ L p ( M ) p ∥ u k ∥ L p ( M ) p → + ∞ . \frac{\lVert u_{k}\rVert_{L^{p^{*}}(M)}^{p}-S_{n,p}^{p}\lVert\nabla u_{k}\rVert_{L^{p}(M)}^{p}}{\lVert u_{k}\rVert_{L^{p}(M)}^{p}} \to+\infty. ∥ u k ∥ L p ( M ) p ∥ u k ∥ L p ∗ ( M ) p − S n , p p ∥ ∇ u k ∥ L p ( M ) p → + ∞.
Write O G x = { g 1 x , … , g k G x } O_{G}^{x}=\{g_{1}x,\ldots,g_{k_{G}}x\} O G x = { g 1 x , … , g k G x } , where g i ∈ G g_{i}\in G g i ∈ G , and set v k , i = u k ∘ ( g i ) − 1 v_{k,i}=u_{k}\circ(g_{i})^{-1} v k , i = u k ∘ ( g i ) − 1 . Then supp v k , i ⊂ B ( g i x , k − 1 ) \operatorname{supp}v_{k,i}\subset B(g_{i}x,k^{-1}) supp v k , i ⊂ B ( g i x , k − 1 ) , and, for all sufficiently large k k k , supp v k , i ∩ supp v k , j = ∅ \operatorname{supp}v_{k,i}\cap\operatorname{supp}v_{k,j}=\varnothing supp v k , i ∩ supp v k , j = ∅ whenever i ≠ j i\ne j i = j . For such k k k , define w k = ∑ i = 1 k G v k , i ∈ W G 1 , p ( M ) w_{k}=\sum_{i=1}^{k_{G}}v_{k,i}\in W_{G}^{1,p}(M) w k = ∑ i = 1 k G v k , i ∈ W G 1 , p ( M ) . Then
∥ w k ∥ L p ∗ ( M ) p − S n , p p k G − p n ∥ ∇ w k ∥ L p ( M ) p ∥ w k ∥ L p ( M ) p = ∥ u k ∥ L p ∗ ( M ) p − S n , p p ∥ ∇ u k ∥ L p ( M ) p k G p n ∥ u k ∥ L p ( M ) p → + ∞ , \frac{\lVert w_{k}\rVert_{L^{p^{*}}(M)}^{p}-S_{n,p}^{p}k_{G}^{-\frac{p}{n}}\lVert\nabla w_{k}\rVert_{L^{p}(M)}^{p}}{\lVert w_{k}\rVert_{L^{p}(M)}^{p}}
=
\frac{\lVert u_{k}\rVert_{L^{p^{*}}(M)}^{p}-S_{n,p}^{p}\lVert\nabla u_{k}\rVert_{L^{p}(M)}^{p}}{k_{G}^{\frac{p}{n}}\lVert u_{k}\rVert_{L^{p}(M)}^{p}}
\to+\infty, ∥ w k ∥ L p ( M ) p ∥ w k ∥ L p ∗ ( M ) p − S n , p p k G − n p ∥ ∇ w k ∥ L p ( M ) p = k G n p ∥ u k ∥ L p ( M ) p ∥ u k ∥ L p ∗ ( M ) p − S n , p p ∥ ∇ u k ∥ L p ( M ) p → + ∞ ,
which completes the proof.
Declarations
Declaration of generative AI and AI-assisted technologies. During the preparation of this work, the authors used ChatGPT (OpenAI) to assist with language editing and locating relevant references, and to assist in searching for explicit examples for the case m = 1 m = 1 m = 1 , 5 ≤ n ≤ 8 5 \le n \le8 5 ≤ n ≤ 8 , in Theorem 4.6 . The suggested references and examples, together with the associated mathematical calculations, were independently checked and verified by the authors. The authors reviewed and edited the content as needed and take full responsibility for the contents of the article.
Conflict of interest. On behalf of all authors, the corresponding author states that there is no conflict of interest.
Data availability statement. No data were used for the research described in this article.
References