Introduction and main results
In their seminal works [4 , 5 ] , Bourgain, Brezis, and Mironescu established that suitably normalized fractional Gagliardo energies converge to their first-order Sobolev counterparts as the fractional order approaches one from below. Since then, BBM-type limiting phenomena have been investigated in various settings, including fractional Orlicz–Sobolev spaces [1 ] , interpolation and extrapolation frameworks [11 ] , whole-space convergence results involving Triebel–Lizorkin spaces [6 ] , generalized fractional Sobolev seminorms on arbitrary domains [21 ] , and ball Banach function spaces [10 ] . Related characterizations of ball Banach Sobolev spaces can be found in [23 ] .
In particular, for n ≥ 1 n \ge1 n ≥ 1 and f ∈ C c ∞ ( R n ) f \in C^\infty_c(\mathbb{R}^n) f ∈ C c ∞ ( R n ) , the BBM formula in the case p = 1 p=1 p = 1 , which is central to the present work, reads
lim α → 1 − ( 1 − α ) ∫ R n ∫ R n ∣ f ( x ) − f ( y ) ∣ ∣ x − y ∣ n + α d x d y = K n ∫ R n ∣ ∇ f ( x ) ∣ d x . \lim_{\alpha\to1^-}(1-\alpha)\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}\frac{|f(x)-f(y)|}{|x-y|^{n+\alpha}}\,dx\,dy=K_n\int_{\mathbb{R}^n}|\nabla f(x)|\,dx. α → 1 − lim ( 1 − α ) ∫ R n ∫ R n ∣ x − y ∣ n + α ∣ f ( x ) − f ( y ) ∣ d x d y = K n ∫ R n ∣∇ f ( x ) ∣ d x .
where
K n = ∫ S n − 1 ∣ ω ⋅ e ∣ d σ ( ω ) = Γ ( n / 2 ) π Γ ( ( n + 1 ) / 2 ) ∣ S n − 1 ∣ , e ∈ S n − 1 . K_n=\int_{\mathbb{S}^{n-1}}|\omega\cdot e|\,d\sigma(\omega)=\frac{\Gamma(n/2)}{\sqrt{\pi}\Gamma((n+1)/2)}|\mathbb{S}^{n-1}|,\qquad e\in\mathbb{S}^{n-1}. K n = ∫ S n − 1 ∣ ω ⋅ e ∣ d σ ( ω ) = π Γ (( n + 1 ) /2 ) Γ ( n /2 ) ∣ S n − 1 ∣ , e ∈ S n − 1 .
d σ d\sigma d σ denotes the standard surface measure on S n − 1 \mathbb{S}^{n-1} S n − 1 , and Γ \Gamma Γ is the Euler Gamma function. The factor 1 − α 1-\alpha 1 − α compensates for the divergence of the fractional energy as α → 1 − \alpha\to1^- α → 1 − .
Quantitative and weighted refinements of BBM-type inequalities have also been investigated. Weighted fractional Poincaré–Sobolev and Hardy-type inequalities, with explicit control of the constants as the fractional parameter approaches one, were established in [19 ] . More recently, Hu, Li, Yang, and Yuan [18 ] established localized weighted inequalities relating Gagliardo and Sobolev seminorms and derived BBM-type formulas in the framework of ball Banach function spaces.
Convergence rates and higher-order asymptotics in BBM-type formulas have also been investigated. Chambolle, Novaga, and Pagliari [8 ] studied convergence rates for nonlocal energies and established the Γ \Gamma Γ -convergence of suitably rescaled functionals to a second-order limit functional. Second-order asymptotic expansions were obtained for fractional perimeters in [7 ] and, more recently, for quadratic fractional Gagliardo energies in [20 ] . Fernández Bonder and Salort [12 ] investigated anisotropic fractional energies, showing how the directional dependence of nonlocal interactions is reflected in the limiting local energy.
Riesz potentials provide a natural framework for Sobolev-type subrepresentation formulas. For 0 < β < n 0 < \beta< n 0 < β < n and a suitable function g g g , let
I β g ( x ) = γ n , β ∫ R n g ( y ) ∣ x − y ∣ n − β d y , I_\beta g(x) = \gamma_{n,\beta} \int_{\mathbb{R}^n} \frac{g(y)}{|x-y|^{n-\beta}}\,dy, I β g ( x ) = γ n , β ∫ R n ∣ x − y ∣ n − β g ( y ) d y ,
where
γ n , β = Γ ( ( n − β ) / 2 ) 2 β π n / 2 Γ ( β / 2 ) . (1) \gamma_{n,\beta} = \frac{\Gamma((n-\beta)/2)}{2^\beta\pi^{n/2}\Gamma(\beta/2)}. \tag*{(1)} γ n , β = 2 β π n /2 Γ ( β /2 ) Γ (( n − β ) /2 ) . (1)
For n ≥ 2 n \ge2 n ≥ 2 , the classical subrepresentation estimate
∣ f ( x ) ∣ ≲ n I 1 ( ∣ ∇ f ∣ ) ( x ) , f ∈ C c ∞ ( R n ) , |f(x)| \lesssim_n I_1(|\nabla f|)(x), \qquad f \in C^\infty_c(\mathbb{R}^n), ∣ f ( x ) ∣ ≲ n I 1 ( ∣∇ f ∣ ) ( x ) , f ∈ C c ∞ ( R n ) ,
provides pointwise control of a function by a Riesz potential of its gradient. Related representation formulas and weighted Poincaré inequalities have been studied in more general settings; see, for instance, [13 ] . Hajłasz and Liu [14 ] investigated maximal potentials and their Sobolev regularity. More recently, Hoang, Moen, and Pérez established pointwise estimates for rough operators in terms of Riesz potentials [15 ] , as well as related estimates involving sparse potential operators [16 ] .
A fractional analogue of this subrepresentation framework involves the nonlinear fractional differential operator introduced by Spector [22 ] ,
D α f ( x ) = ∫ R n ∣ f ( x ) − f ( y ) ∣ ∣ x − y ∣ n + α d y , 0 < α < 1. D^\alpha f(x) = \int_{\mathbb{R}^n} \frac{|f(x)-f(y)|}{|x-y|^{n+\alpha}}\,dy, \qquad0 < \alpha< 1. D α f ( x ) = ∫ R n ∣ x − y ∣ n + α ∣ f ( x ) − f ( y ) ∣ d y , 0 < α < 1.
In connection with the BBM formula, it is natural to consider the normalized Riesz potential
( 1 − α ) I α ( D α f ) (1-\alpha)I_\alpha(D^\alpha f) ( 1 − α ) I α ( D α f )
and investigate its relationship with I 1 ( ∣ ∇ f ∣ ) I_1(|\nabla f|) I 1 ( ∣∇ f ∣ ) as α → 1 − \alpha\to1^{-} α → 1 − . Hoang, Moen, and Pérez [17 ] established that, for n ≥ 2 n \ge2 n ≥ 2 , 0 < α < 1 0 < \alpha< 1 0 < α < 1 , and f ∈ C c ∞ ( R n ) f \in C^\infty_c(\mathbb{R}^n) f ∈ C c ∞ ( R n ) ,
( 1 − α ) I α ( D α f ) ( x ) ≤ C n , α H M P I 1 ( ∣ ∇ f ∣ ) ( x ) , x ∈ R n , (2) (1-\alpha)I_\alpha(D^\alpha f)(x) \le C^{\mathrm{HMP}}_{n,\alpha} I_1(|\nabla f|)(x), \qquad x \in\mathbb{R}^n, \tag*{(2)} ( 1 − α ) I α ( D α f ) ( x ) ≤ C n , α HMP I 1 ( ∣∇ f ∣ ) ( x ) , x ∈ R n , (2)
where, with the normalization (1) ,
C n , α H M P = 2 2 − α π n / 2 Γ ( ( 3 − α ) / 2 ) α Γ ( ( n + α − 1 ) / 2 ) . (3) C^{\mathrm{HMP}}_{n,\alpha} = \frac{2^{2-\alpha}\pi^{n/2}\Gamma((3-\alpha)/2)}{\alpha\Gamma((n+\alpha-1)/2)}. \tag*{(3)} C n , α HMP = α Γ (( n + α − 1 ) /2 ) 2 2 − α π n /2 Γ (( 3 − α ) /2 ) . (3)
In particular,
lim α → 1 − C n , α H M P = ∣ S n − 1 ∣ . \lim_{\alpha\to1^-} C^{\mathrm{HMP}}_{n,\alpha} = |S^{n-1}|. α → 1 − lim C n , α HMP = ∣ S n − 1 ∣.
The constant C n , α H M P C_{n,\alpha}^{\mathrm{HMP}} C n , α HMP is explicit, but its optimality was not addressed in [17 ] . Claros and Pérez [9 ] established the pointwise BBM limit of the normalized Riesz potential. More precisely, for n ≥ 2 n \ge2 n ≥ 2 , f ∈ C c ∞ ( R n ) f \in C_c^\infty(\mathbb{R}^n) f ∈ C c ∞ ( R n ) , and every x ∈ R n x \in\mathbb{R}^n x ∈ R n ,
lim α → 1 − ( 1 − α ) I α ( D α f ) ( x ) = K n I 1 ( ∣ ∇ f ∣ ) ( x ) . (4) \lim_{\alpha\to1^-}(1-\alpha)I_\alpha(D^\alpha f)(x)=K_nI_1(|\nabla f|)(x). \tag*{(4)} α → 1 − lim ( 1 − α ) I α ( D α f ) ( x ) = K n I 1 ( ∣∇ f ∣ ) ( x ) . (4)
They further proved, by a density argument, that for f ∈ W 1 , 1 ( R n ) f \in W^{1,1}(\mathbb{R}^n) f ∈ W 1 , 1 ( R n ) , the same convergence holds in measure on every compact subset of R n \mathbb{R}^n R n . Consequently, every sequence α k → 1 \alpha_k \to1 α k → 1 admits a subsequence along which the convergence holds almost everywhere in R n \mathbb{R}^n R n .
For n ≥ 2 n \ge2 n ≥ 2 , 0 < α < 1 0 < \alpha< 1 0 < α < 1 , f ∈ C c ∞ ( R n ) ∖ { 0 } f \in C_c^\infty(\mathbb{R}^n)\setminus\{0\} f ∈ C c ∞ ( R n ) ∖ { 0 } , and x ∈ R n x \in\mathbb{R}^n x ∈ R n , we consider the quotient
R α ( f , x ) = ( 1 − α ) I α ( D α f ) ( x ) I 1 ( ∣ ∇ f ∣ ) ( x ) . (5) R_\alpha(f,x)=\frac{(1-\alpha)I_\alpha(D^\alpha f)(x)}{I_1(|\nabla f|)(x)}. \tag*{(5)} R α ( f , x ) = I 1 ( ∣∇ f ∣ ) ( x ) ( 1 − α ) I α ( D α f ) ( x ) . (5)
We define the optimal constant associated with (5) by
C n , α ∗ = sup f ∈ C c ∞ ( R n ) ∖ { 0 } x ∈ R n R α ( f , x ) . (6) C_{n,\alpha}^*=\sup_{\substack{f\in C_c^\infty(\mathbb{R}^n)\setminus\{0\}\\x\in\mathbb{R}^n}}R_\alpha(f,x). \tag*{(6)} C n , α ∗ = f ∈ C c ∞ ( R n ) ∖ { 0 } x ∈ R n sup R α ( f , x ) . (6)
Equivalently, C n , α ∗ C_{n,\alpha}^* C n , α ∗ is the smallest constant such that
( 1 − α ) I α ( D α f ) ( x ) ≤ C n , α ∗ I 1 ( ∣ ∇ f ∣ ) ( x ) (1-\alpha)I_\alpha(D^\alpha f)(x)\le C_{n,\alpha}^*I_1(|\nabla f|)(x) ( 1 − α ) I α ( D α f ) ( x ) ≤ C n , α ∗ I 1 ( ∣∇ f ∣ ) ( x )
for every f ∈ C c ∞ ( R n ) f \in C_c^\infty(\mathbb{R}^n) f ∈ C c ∞ ( R n ) and every x ∈ R n x \in\mathbb{R}^n x ∈ R n . By (2) ,
0 < C n , α ∗ ≤ C n , α H M P < ∞ . 0<C_{n,\alpha}^*\le C_{n,\alpha}^{\mathrm{HMP}}<\infty. 0 < C n , α ∗ ≤ C n , α HMP < ∞.
For each fixed admissible pair ( f , x ) (f,x) ( f , x ) , the pointwise limit (4) yields
lim α → 1 − R α ( f , x ) = K n . \lim_{\alpha\to1^-}R_\alpha(f,x)=K_n. α → 1 − lim R α ( f , x ) = K n .
However, this convergence does not justify interchanging the limit with the supremum in (6) , since maximizing pairs may depend on α \alpha α .
The first objective of the present paper is to study the optimal constant C n , α ∗ C_{n,\alpha}^* C n , α ∗ . In dimension two, we establish its leading-order asymptotic behavior as α → 1 − \alpha\to1^- α → 1 − , determine the exact constant in the radial class for every 0 < α < 1 0<\alpha<1 0 < α < 1 , and identify a class of radial extremizers. In dimensions n ≥ 3 n\ge3 n ≥ 3 , we determine the exact value of C n , α ∗ C_{n,\alpha}^* C n , α ∗ for every 0 < α < 1 0<\alpha<1 0 < α < 1 and identify a class of extremizers. The second objective is to refine the leading-order BBM limit (4) by deriving a first-order asymptotic expansion of the normalized potential ( 1 − α ) I α ( D α f ) (1-\alpha)I_\alpha(D^\alpha f) ( 1 − α ) I α ( D α f ) , with a remainder estimate uniform in x ∈ R n x\in\mathbb{R}^n x ∈ R n . Finally, we strengthen the subsequential almost-everywhere convergence result of Claros and Pérez [9 ] by proving convergence almost everywhere as α → 1 − \alpha\to1^- α → 1 − , without passing to a subsequence, for f ∈ W 1 , 1 ( R n ) f\in W^{1,1}(\mathbb{R}^n) f ∈ W 1 , 1 ( R n ) .
We begin with the planar case. Our first result describes the leading-order asymptotic behavior of the unrestricted optimal constant C 2 , α ∗ C_{2,\alpha}^* C 2 , α ∗ .
Theorem 1.1. As α → 1 − \alpha\to1^- α → 1 − ,
C 2 , α ∗ = 4 + O ( 1 − α ) . (7) C_{2,\alpha}^*=4+O(1-\alpha). \tag*{(7)} C 2 , α ∗ = 4 + O ( 1 − α ) . (7) In particular, (7) yields
lim α → 1 − C 2 , α ∗ = 4 = K 2 , \lim_{\alpha\to1^-}C_{2,\alpha}^*=4=K_2, α → 1 − lim C 2 , α ∗ = 4 = K 2 , in agreement with the pointwise BBM limit (4) .
We next restrict the optimization problem to radially symmetric functions centered at the evaluation point. For 0 < α < 1 0 < \alpha< 1 0 < α < 1 , define
C 2 , α r a d = sup x 0 ∈ R 2 f ( y ) = F ( ∣ y − x 0 ∣ ) ∈ C c ∞ ( R 2 ) ∖ { 0 } R α ( f , x 0 ) . C^{\mathrm{rad}}_{2,\alpha}=\sup_{\substack{x_0\in\mathbb{R}^2\\ f(y)=F(|y-x_0|)\in C^\infty_c(\mathbb{R}^2)\setminus\{0\}}}R_\alpha(f,x_0). C 2 , α rad = x 0 ∈ R 2 f ( y ) = F ( ∣ y − x 0 ∣ ) ∈ C c ∞ ( R 2 ) ∖ { 0 } sup R α ( f , x 0 ) .
Our next result provides an explicit formula for this constant and identifies a class of extremizers.
Theorem 1.2. For every 0 < α < 1 0 < \alpha< 1 0 < α < 1 ,
C 2 , α r a d = 2 2 − α π ( 1 − α ) α 2 Γ ( 1 − α / 2 ) Γ ( α / 2 ) ( Γ ( 1 − α ) 2 Γ ( 1 − α / 2 ) 2 − 1 ) . (8) C^{\mathrm{rad}}_{2,\alpha}=\frac{2^{2-\alpha}\pi(1-\alpha)}{\alpha^2}\frac{\Gamma(1-\alpha/2)}{\Gamma(\alpha/2)}\left(\frac{\Gamma(1-\alpha)}{2\Gamma(1-\alpha/2)^2}-1\right). \tag*{(8)} C 2 , α rad = α 2 2 2 − α π ( 1 − α ) Γ ( α /2 ) Γ ( 1 − α /2 ) ( 2Γ ( 1 − α /2 ) 2 Γ ( 1 − α ) − 1 ) . (8) Moreover, for every x 0 ∈ R 2 x_0\in\mathbb{R}^2 x 0 ∈ R 2 and every nontrivial function
f ( y ) = F ( ∣ y − x 0 ∣ ) ∈ C c ∞ ( R 2 ) , f(y)=F(|y-x_0|)\in C^\infty_c(\mathbb{R}^2), f ( y ) = F ( ∣ y − x 0 ∣ ) ∈ C c ∞ ( R 2 ) , with F F F monotone, we have
R α ( f , x 0 ) = C 2 , α r a d . R_\alpha(f,x_0)=C^{\mathrm{rad}}_{2,\alpha}. R α ( f , x 0 ) = C 2 , α rad .
Theorem Theorem 1.2 provides an explicit formula for the sharp constant in the radial class and identifies a family of extremizers for every 0 < α < 1 0 < \alpha< 1 0 < α < 1 . The determination of the unrestricted optimal constant C 2 , α ∗ C^*_{2,\alpha} C 2 , α ∗ for arbitrary 0 < α < 1 0 < \alpha< 1 0 < α < 1 remains unresolved by the present analysis.
We now turn to the case n ≥ 3 n \ge3 n ≥ 3 . For 0 < α < 1 0 < \alpha< 1 0 < α < 1 , define
C n , α r a d = ∣ S n − 2 ∣ ∫ 0 ∞ ∫ 0 π 2 cosh ( n − α 2 τ ) ( sin θ ) n − 2 [ 2 ( cosh τ − cos θ ) ] ( n + α ) / 2 d θ d τ . (9) C^{\mathrm{rad}}_{n,\alpha}=|S^{n-2}|\int_0^\infty\int_0^\pi\frac{2\cosh\left(\frac{n-\alpha}{2}\tau\right)(\sin\theta)^{n-2}}{[2(\cosh\tau-\cos\theta)]^{(n+\alpha)/2}}\,d\theta\,d\tau. \tag*{(9)} C n , α rad = ∣ S n − 2 ∣ ∫ 0 ∞ ∫ 0 π [ 2 ( cosh τ − cos θ ) ] ( n + α ) /2 2 cosh ( 2 n − α τ ) ( sin θ ) n − 2 d θ d τ . (9)
The integral in (9) is finite for every 0 < α < 1 0 < \alpha< 1 0 < α < 1 .
Our next main result is the following.
Theorem 1.3. Let n ≥ 3 n\ge3 n ≥ 3 . For every 0 < α < 1 0 < \alpha< 1 0 < α < 1 ,
C n , α ∗ = ( 1 − α ) γ n , α γ n , 1 C n , α r a d . (10) C^*_{n,\alpha}=(1-\alpha)\frac{\gamma_{n,\alpha}}{\gamma_{n,1}}C^{\mathrm{rad}}_{n,\alpha}. \tag*{(10)} C n , α ∗ = ( 1 − α ) γ n , 1 γ n , α C n , α rad . (10) Moreover, for every x 0 ∈ R n x_0\in\mathbb{R}^n x 0 ∈ R n and every nontrivial function
f ( y ) = F ( ∣ y − x 0 ∣ ) ∈ C c ∞ ( R n ) , f(y)=F(|y-x_0|)\in C^\infty_c(\mathbb{R}^n), f ( y ) = F ( ∣ y − x 0 ∣ ) ∈ C c ∞ ( R n ) , with F F F monotone, we have
R α ( f , x 0 ) = C n , α ∗ . R_\alpha(f,x_0)=C^*_{n,\alpha}. R α ( f , x 0 ) = C n , α ∗ .
Theorem Theorem 1.3 shows that, for n ≥ 3 n\ge3 n ≥ 3 , the unrestricted optimization problem admits an exact reduction to the class of radially monotone functions. In fact, any fixed nontrivial radially monotone function in C c ∞ ( R n ) C^\infty_c(\mathbb{R}^n) C c ∞ ( R n ) , centered at the origin, attains the optimal constant for every 0 < α < 1 0 < \alpha< 1 0 < α < 1 . Consequently, by (4) ,
lim α → 1 − C n , α ∗ = K n . \lim_{\alpha\to1^-}C^*_{n,\alpha}=K_n. α → 1 − lim C n , α ∗ = K n .
The following corollary shows that the sharp constant provides a strict improvement over the constant in the Hoang–Moen–Pérez inequality (2) for every 0 < α < 1 0 < \alpha< 1 0 < α < 1 , and quantifies this improvement as α → 1 − \alpha\to1^- α → 1 − .
Corollary 1.4. Let n ≥ 3 n \ge3 n ≥ 3 . Then, for every 0 < α < 1 0 < \alpha< 1 0 < α < 1 ,
C n , α ∗ < C n , α H M P . C_{n,\alpha}^{*} < C_{n,\alpha}^{\mathrm{HMP}}. C n , α ∗ < C n , α HMP . Moreover,
lim α → 1 − C n , α ∗ C n , α H M P = K n ∣ S n − 1 ∣ < 1. \lim_{\alpha\to1^{-}} \frac{C_{n,\alpha}^{*}}{C_{n,\alpha}^{\mathrm{HMP}}} = \frac{K_n}{|\mathbb{S}^{n-1}|} < 1. α → 1 − lim C n , α HMP C n , α ∗ = ∣ S n − 1 ∣ K n < 1.
Our next main result refines the pointwise BBM limit (4) by identifying the first-order asymptotic correction for each fixed f ∈ C c 2 ( R n ) f \in C_c^2(\mathbb{R}^n) f ∈ C c 2 ( R n ) and establishing a remainder estimate uniform in x ∈ R n x \in\mathbb{R}^n x ∈ R n .
For n ≥ 2 n \ge2 n ≥ 2 , f ∈ C c 2 ( R n ) f \in C_c^2(\mathbb{R}^n) f ∈ C c 2 ( R n ) , and y ∈ R n y \in\mathbb{R}^n y ∈ R n , define
F f ( y ) = ∫ ∣ h ∣ < 1 ∣ f ( y + h ) − f ( y ) ∣ − ∣ ∇ f ( y ) ⋅ h ∣ ∣ h ∣ n + 1 d h + ∫ ∣ h ∣ ≥ 1 ∣ f ( y + h ) − f ( y ) ∣ ∣ h ∣ n + 1 d h . (11) \mathcal{F}_f(y) = \int_{|h|<1} \frac{\lvert f(y+h)-f(y)\rvert-\lvert\nabla f(y)\cdot h\rvert}{|h|^{n+1}}\,dh + \int_{|h|\ge1} \frac{\lvert f(y+h)-f(y)\rvert}{|h|^{n+1}}\,dh. \tag*{(11)} F f ( y ) = ∫ ∣ h ∣ < 1 ∣ h ∣ n + 1 ∣ f ( y + h ) − f ( y )∣ − ∣ ∇ f ( y ) ⋅ h ∣ d h + ∫ ∣ h ∣ ≥ 1 ∣ h ∣ n + 1 ∣ f ( y + h ) − f ( y )∣ d h . (11)
Both integrals in (11) are absolutely convergent for every y ∈ R n y \in\mathbb{R}^n y ∈ R n . Moreover,
F f ∈ L 1 ( R n ) ∩ L ∞ ( R n ) . \mathcal{F}_f \in L^1(\mathbb{R}^n)\cap L^\infty(\mathbb{R}^n). F f ∈ L 1 ( R n ) ∩ L ∞ ( R n ) .
Let
λ n = − 1 2 ψ ( n − 1 2 ) − log 2 − 1 2 ψ ( 1 2 ) , (12) \lambda_n=-\frac{1}{2}\psi\left(\frac{n-1}{2}\right)-\log2-\frac{1}{2}\psi\left(\frac{1}{2}\right), \tag*{(12)} λ n = − 2 1 ψ ( 2 n − 1 ) − log 2 − 2 1 ψ ( 2 1 ) , (12)
where ψ = Γ ′ / Γ \psi=\Gamma'/\Gamma ψ = Γ ′ /Γ denotes the digamma function.
For x ∈ R n x \in\mathbb{R}^n x ∈ R n , we then define
A f ( x ) = I 1 ( F f ) ( x ) − K n γ n , 1 ∫ R n log ∣ x − y ∣ + λ n ∣ x − y ∣ n − 1 ∣ ∇ f ( y ) ∣ d y . (13) A_f(x)=I_1(\mathcal{F}_f)(x)-K_n\gamma_{n,1}\int_{\mathbb{R}^n}\frac{\log|x-y|+\lambda_n}{|x-y|^{n-1}}|\nabla f(y)|\,dy. \tag*{(13)} A f ( x ) = I 1 ( F f ) ( x ) − K n γ n , 1 ∫ R n ∣ x − y ∣ n − 1 log ∣ x − y ∣ + λ n ∣∇ f ( y ) ∣ d y . (13)
Both terms on the right-hand side of (13) are absolutely convergent for every x ∈ R n x \in\mathbb{R}^n x ∈ R n . Moreover,
A f ∈ L ∞ ( R n ) . A_f\in L^\infty(\mathbb{R}^n). A f ∈ L ∞ ( R n ) .
Theorem 1.5. Let n ≥ 2 n \ge2 n ≥ 2 and f ∈ C c 2 ( R n ) f \in C_c^2(\mathbb{R}^n) f ∈ C c 2 ( R n ) . Then, as α → 1 − \alpha\to1^{-} α → 1 − ,
∥ ( 1 − α ) I α ( D α f ) − K n I 1 ( ∣ ∇ f ∣ ) − ( 1 − α ) A f ∥ L ∞ ( R n ) = O ( ( 1 − α ) 2 ) . (14) \left\|(1-\alpha)I_\alpha(D^\alpha f)-K_nI_1(|\nabla f|)-(1-\alpha)A_f\right\|_{L^\infty(\mathbb{R}^n)}=O\left((1-\alpha)^2\right). \tag*{(14)} ∥ ( 1 − α ) I α ( D α f ) − K n I 1 ( ∣∇ f ∣ ) − ( 1 − α ) A f ∥ L ∞ ( R n ) = O ( ( 1 − α ) 2 ) . (14)
The correction term A f A_f A f consists of two contributions: I 1 ( F f ) I_1(\mathcal{F}_f) I 1 ( F f ) arises from the finite remainder in the expansion of the nonlinear fractional differential operator D α f D^\alpha f D α f , while the logarithmic term in (13) reflects the first-order variation of the Riesz kernel and its normalization as α → 1 − \alpha\to1^{-} α → 1 − .
Our final main result strengthens the W 1 , 1 W^{1,1} W 1 , 1 -extension of Claros and Pérez [9 ] by establishing almost everywhere convergence as α → 1 − \alpha\to1^{-} α → 1 − , without passing to a subsequence.
Theorem 1.6. Let n ≥ 2 n \ge2 n ≥ 2 and f ∈ W 1 , 1 ( R n ) f \in W^{1,1}(\mathbb{R}^n) f ∈ W 1 , 1 ( R n ) . Then, for almost every x ∈ R n x \in\mathbb{R}^n x ∈ R n ,
lim α → 1 − ( 1 − α ) I α ( D α f ) ( x ) = K n I 1 ( ∣ ∇ f ∣ ) ( x ) . \lim_{\alpha\to1^{-}}(1-\alpha)I_\alpha(D^\alpha f)(x)=K_nI_1(|\nabla f|)(x). α → 1 − lim ( 1 − α ) I α ( D α f ) ( x ) = K n I 1 ( ∣∇ f ∣ ) ( x ) .
We now briefly outline the main ideas underlying the proofs of our results. Our approach to the sharp-constant problem is based on logarithmic-polar coordinates, which allow us to reformulate the problem as an integral inequality on the cylinder R × S n − 1 \mathbb{R}\times\mathbb{S}^{n-1} R × S n − 1 .
In dimension two, the main challenge is to control the directional dependence of the variational upper bound as α → 1 − \alpha\to1^{-} α → 1 − . We overcome this difficulty by establishing a uniform asymptotic estimate for the associated directional functional, which yields Theorem 1.1 . Furthermore, we solve the optimization problem in the radial class for every 0 < α < 1 0 < \alpha< 1 0 < α < 1 , obtaining an explicit formula for the sharp radial constant and identifying a family of extremizers, as stated in Theorem 1.2 .
In dimensions n ≥ 3 n \ge3 n ≥ 3 , the main challenge is to establish the global optimality of the radial direction for every 0 < α < 1 0 < \alpha< 1 0 < α < 1 . Exploiting the invariance of the natural measure under the geodesic flow on the unit tangent bundle of S n − 1 \mathbb{S}^{n-1} S n − 1 , we derive an upper bound for the unrestricted problem in terms of a directional maximization problem. The key step is to prove that the associated directional functional attains its maximum in the radial direction. This is achieved by establishing a monotonicity property of a suitably transformed kernel and applying a circular rearrangement inequality. The resulting sharp bound is attained by radially monotone functions, yielding Theorem 1.3 .
Turning to the first-order asymptotics, the main difficulty is to identify the correction arising from the simultaneous variation of the nonlinear fractional operator D α D^\alpha D α and the Riesz kernel as α → 1 − \alpha\to1^{-} α → 1 − . Our approach consists in isolating the leading singular contribution of D α f D^\alpha f D α f and controlling its remainder in both L 1 L^1 L 1 and L ∞ L^\infty L ∞ . We then combine this analysis with a first-order expansion of the normalized Riesz kernel, including its logarithmic correction. This allows us to identify the explicit correction term A f A_f A f and establish the uniform quadratic remainder estimate in Theorem 1.5 .
Finally, the proof of Theorem 1.6 relies on a pointwise stability estimate, uniform with respect to α \alpha α near 1 1 1 , derived from the Hoang–Moen–Pérez inequality (2) . Combined with a suitable approximation by smooth functions in W 1 , 1 ( R n ) W^{1,1}(\mathbb{R}^n) W 1 , 1 ( R n ) , this estimate allows us to construct a common set of full measure, independent of α \alpha α , and thereby establish almost-everywhere convergence without passing to a subsequence.
The rest of the paper is organized as follows. Section 2 2 2 is devoted to the planar case, where we prove Theorems 1.1 and 1.2 . Section 3 3 3 contains the proof of Theorem 1.3 and the derivation of Corollary 1.4. Theorem 1.5 is proved in Section 4 4 4 , while the proof of Theorem 1.6 is given in Section 5 5 5 . Finally, Section 6 6 6 discusses the question of determining the unrestricted sharp constant in dimension two and establishes an explicit series representation for the sharp constant in higher dimensions.
The constants C 2 , α ∗ C^*_{2,\alpha} C 2 , α ∗ and C 2 , α r a d C^{\mathrm{rad}}_{2,\alpha} C 2 , α rad
This section is devoted to the proofs of Theorems 1.1 and 1.2 . We begin with several auxiliary lemmas.
For 0 < α < 1 0 < \alpha< 1 0 < α < 1 and
( τ , ϑ ) ∈ ( R × [ − π , π ] ) ∖ { ( 0 , 0 ) } , (\tau,\vartheta) \in\left(\mathbb{R} \times[-\pi,\pi]\right) \setminus\{(0,0)\}, ( τ , ϑ ) ∈ ( R × [ − π , π ] ) ∖ {( 0 , 0 )} ,
define
K α ( τ , ϑ ) = 2 cosh ( 2 − α 2 τ ) [ 2 ( cosh τ − cos ϑ ) ] 1 + α / 2 . (15) \mathcal{K}_{\alpha}(\tau,\vartheta)=\frac{2\cosh\left(\frac{2-\alpha}{2}\tau\right)}{\left[2(\cosh\tau-\cos\vartheta)\right]^{1+\alpha/2}}. \tag*{(15)} K α ( τ , ϑ ) = [ 2 ( cosh τ − cos ϑ ) ] 1 + α /2 2 cosh ( 2 2 − α τ ) . (15)
For β ∈ [ 0 , π / 2 ] \beta\in[0,\pi/2] β ∈ [ 0 , π /2 ] , set
A α ( β ) = 2 ∫ 0 ∞ ∫ 0 π max { τ cos β , ϑ sin β } K α ( τ , ϑ ) d ϑ d τ . (16) A_{\alpha}(\beta)=2\int_{0}^{\infty}\int_{0}^{\pi}\max\{\tau\cos\beta,\vartheta\sin\beta\}\mathcal{K}_{\alpha}(\tau,\vartheta)\,\mathrm{d}\vartheta\,\mathrm{d}\tau. \tag*{(16)} A α ( β ) = 2 ∫ 0 ∞ ∫ 0 π max { τ cos β , ϑ sin β } K α ( τ , ϑ ) d ϑ d τ . (16)
The integral defining A α ( β ) A_{\alpha}(\beta) A α ( β ) is finite for every 0 < α < 1 0 < \alpha< 1 0 < α < 1 and β ∈ [ 0 , π / 2 ] \beta\in[0,\pi/2] β ∈ [ 0 , π /2 ] .
Lemma 2.1. For every 0 < α < 1 0 < \alpha< 1 0 < α < 1 ,
C 2 , α ∗ ≤ ( 1 − α ) γ 2 , α γ 2 , 1 sup 0 ≤ β ≤ π / 2 A α ( β ) . (17) C_{2,\alpha}^{*} \le(1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}\sup_{0\le\beta\le\pi/2} A_{\alpha}(\beta). \tag*{(17)} C 2 , α ∗ ≤ ( 1 − α ) γ 2 , 1 γ 2 , α 0 ≤ β ≤ π /2 sup A α ( β ) . (17) Proof. Fix 0 < α < 1 0 < \alpha< 1 0 < α < 1 . By translation invariance, (6) gives
C 2 , α ∗ = sup f ∈ C c ∞ ( R 2 ) ∖ { 0 } R α ( f , 0 ) . C_{2,\alpha}^{*}=\sup_{f\in C_{c}^{\infty}(\mathbb{R}^{2})\setminus\{0\}}R_{\alpha}(f,0). C 2 , α ∗ = f ∈ C c ∞ ( R 2 ) ∖ { 0 } sup R α ( f , 0 ) . Let f ∈ C c ∞ ( R 2 ) ∖ { 0 } f\in C_{c}^{\infty}(\mathbb{R}^{2})\setminus\{0\} f ∈ C c ∞ ( R 2 ) ∖ { 0 } . We introduce logarithmic-polar coordinates
y = e t ω ( θ ) , ω ( θ ) = ( cos θ , sin θ ) , y=e^{t}\omega(\theta),\qquad\omega(\theta)=(\cos\theta,\sin\theta), y = e t ω ( θ ) , ω ( θ ) = ( cos θ , sin θ ) , and set
u ( t , θ ) = f ( e t ω ( θ ) ) , ( t , θ ) ∈ R × S 1 . u(t,\theta)=f(e^{t}\omega(\theta)),\qquad(t,\theta)\in\mathbb{R}\times\mathbb{S}^{1}. u ( t , θ ) = f ( e t ω ( θ )) , ( t , θ ) ∈ R × S 1 . Then
∣ ∇ f ( e t ω ( θ ) ) ∣ = e − t ∣ ∂ t u ( t , θ ) ∣ 2 + ∣ ∂ θ u ( t , θ ) ∣ 2 , |\nabla f(e^{t}\omega(\theta))|=e^{-t}\sqrt{|\partial_tu(t,\theta)|^{2}+|\partial_\theta u(t,\theta)|^{2}}, ∣∇ f ( e t ω ( θ )) ∣ = e − t ∣ ∂ t u ( t , θ ) ∣ 2 + ∣ ∂ θ u ( t , θ ) ∣ 2 , and therefore
1 γ 2 , 1 I 1 ( ∣ ∇ f ∣ ) ( 0 ) = ∫ R ∫ 0 2 π ∣ ∇ c u ( t , θ ) ∣ d θ d t . (18) \frac{1}{\gamma_{2,1}}I_{1}(|\nabla f|)(0)=\int_{\mathbb{R}}\int_{0}^{2\pi}|\nabla_{\mathrm{c}}u(t,\theta)|\,d\theta\,dt. \tag*{(18)} γ 2 , 1 1 I 1 ( ∣∇ f ∣ ) ( 0 ) = ∫ R ∫ 0 2 π ∣ ∇ c u ( t , θ ) ∣ d θ d t . (18) where
∇ c u = ( ∂ t u , ∂ θ u ) \nabla_{\mathrm{c}}u=(\partial_tu,\partial_\theta u) ∇ c u = ( ∂ t u , ∂ θ u ) denotes the gradient on the cylinder R × S 1 \mathbb{R}\times\mathbb{S}^{1} R × S 1 .
Writing
y = e t ω ( θ ) , z = e s ω ( φ ) , y=e^{t}\omega(\theta),\qquad z=e^{s}\omega(\varphi), y = e t ω ( θ ) , z = e s ω ( φ ) , we have
∣ y − z ∣ 2 = 2 e t + s ( cosh ( t − s ) − cos ( θ − φ ) ) . |y-z|^{2}=2e^{t+s}\bigl(\cosh(t-s)-\cos(\theta-\varphi)\bigr). ∣ y − z ∣ 2 = 2 e t + s ( cosh ( t − s ) − cos ( θ − φ ) ) . Since
d y = e 2 t d t d θ , d z = e 2 s d s d φ , dy=e^{2t}\,dt\,d\theta,\qquad dz=e^{2s}\,ds\,d\varphi, d y = e 2 t d t d θ , d z = e 2 s d s d φ , it follows from the definitions of I α I_{\alpha} I α and D α \mathcal{D}^{\alpha} D α that
1 γ 2 , α I α ( D α f ) ( 0 ) = ∫ R ∫ 0 2 π ∫ R ∫ 0 2 π e α t + 2 s ∣ u ( t , θ ) − u ( s , φ ) ∣ [ 2 e t + s ( cosh ( t − s ) − cos ( θ − φ ) ) ] 1 + α / 2 d φ d s d θ d t . \frac{1}{\gamma_{2,\alpha}}I_{\alpha}(\mathcal{D}^{\alpha}f)(0)=\int_{\mathbb{R}}\int_{0}^{2\pi}\int_{\mathbb{R}}\int_{0}^{2\pi}\frac{e^{\alpha t+2s}|u(t,\theta)-u(s,\varphi)|}{\left[2e^{t+s}\bigl(\cosh(t-s)-\cos(\theta-\varphi)\bigr)\right]^{1+\alpha/2}}\,d\varphi\,ds\,d\theta\,dt. γ 2 , α 1 I α ( D α f ) ( 0 ) = ∫ R ∫ 0 2 π ∫ R ∫ 0 2 π [ 2 e t + s ( cosh ( t − s ) − cos ( θ − φ ) ) ] 1 + α /2 e α t + 2 s ∣ u ( t , θ ) − u ( s , φ ) ∣ d φ d s d θ d t . With the change of variables
τ = s − t , ϑ = φ − θ , \tau=s-t,\qquad\vartheta=\varphi-\theta, τ = s − t , ϑ = φ − θ , and using the 2 π 2\pi 2 π -periodicity in the angular variable, we obtain
1 γ 2 , α I α ( D α f ) ( 0 ) = ∫ R ∫ − π π e ( 1 − α / 2 ) τ [ 2 ( cosh τ − cos ϑ ) ] 1 + α / 2 [ ∫ R ∫ 0 2 π ∣ u ( t + τ , θ + ϑ ) − u ( t , θ ) ∣ d θ d t ] d ϑ d τ . \frac{1}{\gamma_{2,\alpha}}I_{\alpha}(\mathcal{D}^{\alpha}f)(0)
=\int_{\mathbb{R}}\int_{-\pi}^{\pi}\frac{e^{(1-\alpha/2)\tau}}{\left[2(\cosh\tau-\cos\vartheta)\right]^{1+\alpha/2}}
\left[\int_{\mathbb{R}}\int_{0}^{2\pi}|u(t+\tau,\theta+\vartheta)-u(t,\theta)|\,d\theta\,dt\right]d\vartheta\,d\tau. γ 2 , α 1 I α ( D α f ) ( 0 ) = ∫ R ∫ − π π [ 2 ( cosh τ − cos ϑ ) ] 1 + α /2 e ( 1 − α /2 ) τ [ ∫ R ∫ 0 2 π ∣ u ( t + τ , θ + ϑ ) − u ( t , θ ) ∣ d θ d t ] d ϑ d τ .
On the other hand, interchanging the pairs ( t , θ ) (t,\theta) ( t , θ ) and ( s , φ ) (s,\varphi) ( s , φ ) in the original integral and performing the same change of variables gives
1 γ 2 , α I α ( D α f ) ( 0 ) = ∫ R ∫ − π π e − ( 1 − α / 2 ) τ [ 2 ( cosh τ − cos ϑ ) ] 1 + α / 2 [ ∫ R ∫ 0 2 π ∣ u ( t + τ , θ + ϑ ) − u ( t , θ ) ∣ d θ d t ] d ϑ d τ . \frac{1}{\gamma_{2,\alpha}}I_\alpha(D^\alpha f)(0)
= \int_{\mathbb{R}}\int_{-\pi}^{\pi}\frac{e^{-(1-\alpha/2)\tau}}{[2(\cosh\tau-\cos\vartheta)]^{1+\alpha/2}}\left[\int_{\mathbb{R}}\int_0^{2\pi}|u(t+\tau,\theta+\vartheta)-u(t,\theta)|\,d\theta\,dt\right]d\vartheta\,d\tau. γ 2 , α 1 I α ( D α f ) ( 0 ) = ∫ R ∫ − π π [ 2 ( cosh τ − cos ϑ ) ] 1 + α /2 e − ( 1 − α /2 ) τ [ ∫ R ∫ 0 2 π ∣ u ( t + τ , θ + ϑ ) − u ( t , θ ) ∣ d θ d t ] d ϑ d τ .
Averaging the two identities and using (15) , we obtain
1 γ 2 , α I α ( D α f ) ( 0 ) = 1 2 ∫ R ∫ − π π K α ( τ , ϑ ) [ ∫ R ∫ 0 2 π ∣ u ( t + τ , θ + ϑ ) − u ( t , θ ) ∣ d θ d t ] d ϑ d τ . (19) \frac{1}{\gamma_{2,\alpha}}I_\alpha(D^\alpha f)(0)=\frac{1}{2}\int_{\mathbb{R}}\int_{-\pi}^{\pi}\mathcal{K}_\alpha(\tau,\vartheta)\left[\int_{\mathbb{R}}\int_0^{2\pi}|u(t+\tau,\theta+\vartheta)-u(t,\theta)|\,d\theta\,dt\right]d\vartheta\,d\tau. \tag*{(19)} γ 2 , α 1 I α ( D α f ) ( 0 ) = 2 1 ∫ R ∫ − π π K α ( τ , ϑ ) [ ∫ R ∫ 0 2 π ∣ u ( t + τ , θ + ϑ ) − u ( t , θ ) ∣ d θ d t ] d ϑ d τ . (19)
For ( τ , ϑ ) ∈ R × [ − π , π ] (\tau,\vartheta)\in\mathbb{R}\times[-\pi,\pi] ( τ , ϑ ) ∈ R × [ − π , π ] , the fundamental theorem of calculus yields
u ( t + τ , θ + ϑ ) − u ( t , θ ) = ∫ 0 1 ∇ c u ( t + r τ , θ + r ϑ ) ⋅ ( τ , ϑ ) d r . u(t+\tau,\theta+\vartheta)-u(t,\theta)=\int_0^1\nabla_cu(t+r\tau,\theta+r\vartheta)\cdot(\tau,\vartheta)\,dr. u ( t + τ , θ + ϑ ) − u ( t , θ ) = ∫ 0 1 ∇ c u ( t + r τ , θ + r ϑ ) ⋅ ( τ , ϑ ) d r .
Therefore,
∫ R ∫ 0 2 π ∣ u ( t + τ , θ + ϑ ) − u ( t , θ ) ∣ d θ d t ≤ ∫ 0 1 ∫ R ∫ 0 2 π ∣ ∇ c u ( t + r τ , θ + r ϑ ) ⋅ ( τ , ϑ ) ∣ d θ d t d r . \int_{\mathbb{R}}\int_0^{2\pi}|u(t+\tau,\theta+\vartheta)-u(t,\theta)|\,d\theta\,dt\le\int_0^1\int_{\mathbb{R}}\int_0^{2\pi}|\nabla_cu(t+r\tau,\theta+r\vartheta)\cdot(\tau,\vartheta)|\,d\theta\,dt\,dr. ∫ R ∫ 0 2 π ∣ u ( t + τ , θ + ϑ ) − u ( t , θ ) ∣ d θ d t ≤ ∫ 0 1 ∫ R ∫ 0 2 π ∣ ∇ c u ( t + r τ , θ + r ϑ ) ⋅ ( τ , ϑ ) ∣ d θ d t d r .
For each r ∈ [ 0 , 1 ] r\in[0,1] r ∈ [ 0 , 1 ] , translation invariance on R × S 1 \mathbb{R}\times S^1 R × S 1 gives
∫ R ∫ 0 2 π ∣ ∇ c u ( t + r τ , θ + r ϑ ) ⋅ ( τ , ϑ ) ∣ d θ d t = ∫ R ∫ 0 2 π ∣ ∇ c u ( t , θ ) ⋅ ( τ , ϑ ) ∣ d θ d t . \int_{\mathbb{R}}\int_0^{2\pi}|\nabla_cu(t+r\tau,\theta+r\vartheta)\cdot(\tau,\vartheta)|\,d\theta\,dt=\int_{\mathbb{R}}\int_0^{2\pi}|\nabla_cu(t,\theta)\cdot(\tau,\vartheta)|\,d\theta\,dt. ∫ R ∫ 0 2 π ∣ ∇ c u ( t + r τ , θ + r ϑ ) ⋅ ( τ , ϑ ) ∣ d θ d t = ∫ R ∫ 0 2 π ∣ ∇ c u ( t , θ ) ⋅ ( τ , ϑ ) ∣ d θ d t .
Hence
∫ R ∫ 0 2 π ∣ u ( t + τ , θ + ϑ ) − u ( t , θ ) ∣ d θ d t ≤ ∫ R ∫ 0 2 π ∣ ∇ c u ( t , θ ) ⋅ ( τ , ϑ ) ∣ d θ d t . \int_{\mathbb{R}}\int_0^{2\pi}|u(t+\tau,\theta+\vartheta)-u(t,\theta)|\,d\theta\,dt\le\int_{\mathbb{R}}\int_0^{2\pi}|\nabla_cu(t,\theta)\cdot(\tau,\vartheta)|\,d\theta\,dt. ∫ R ∫ 0 2 π ∣ u ( t + τ , θ + ϑ ) − u ( t , θ ) ∣ d θ d t ≤ ∫ R ∫ 0 2 π ∣ ∇ c u ( t , θ ) ⋅ ( τ , ϑ ) ∣ d θ d t .
Together with (19) , this gives
1 γ 2 , α I α ( D α f ) ( 0 ) ≤ ∫ R ∫ 0 2 π [ 1 2 ∫ R ∫ − π π ∣ ∇ c u ( t , θ ) ⋅ ( τ , ϑ ) ∣ K α ( τ , ϑ ) d ϑ d τ ] d θ d t . (20) \frac{1}{\gamma_{2,\alpha}}I_\alpha(D^\alpha f)(0)\le\int_{\mathbb{R}}\int_0^{2\pi}\left[\frac{1}{2}\int_{\mathbb{R}}\int_{-\pi}^{\pi}|\nabla_cu(t,\theta)\cdot(\tau,\vartheta)|\mathcal{K}_\alpha(\tau,\vartheta)\,d\vartheta\,d\tau\right]d\theta\,dt. \tag*{(20)} γ 2 , α 1 I α ( D α f ) ( 0 ) ≤ ∫ R ∫ 0 2 π [ 2 1 ∫ R ∫ − π π ∣ ∇ c u ( t , θ ) ⋅ ( τ , ϑ ) ∣ K α ( τ , ϑ ) d ϑ d τ ] d θ d t . (20)
Setting
E = { ( t , θ ) ∈ R × S 1 : ∣ ∇ c u ( t , θ ) ∣ > 0 } , E=\{(t,\theta)\in\mathbb{R}\times S^1:|\nabla_cu(t,\theta)|>0\}, E = {( t , θ ) ∈ R × S 1 : ∣ ∇ c u ( t , θ ) ∣ > 0 } ,
and observing that the integrand in (20) vanishes on E c E^c E c , we may rewrite (20) as
1 γ 2 , α I α ( D α f ) ( 0 ) ≤ ∫ E [ 1 2 ∫ R ∫ − π π ∣ ∇ c u ( t , θ ) ⋅ ( τ , ϑ ) ∣ K α ( τ , ϑ ) d ϑ d τ ] d θ d t . (21) \frac{1}{\gamma_{2,\alpha}}I_\alpha(D^\alpha f)(0)\le\int_E\left[\frac{1}{2}\int_{\mathbb{R}}\int_{-\pi}^{\pi}|\nabla_cu(t,\theta)\cdot(\tau,\vartheta)|\mathcal{K}_\alpha(\tau,\vartheta)\,d\vartheta\,d\tau\right]d\theta\,dt. \tag*{(21)} γ 2 , α 1 I α ( D α f ) ( 0 ) ≤ ∫ E [ 2 1 ∫ R ∫ − π π ∣ ∇ c u ( t , θ ) ⋅ ( τ , ϑ ) ∣ K α ( τ , ϑ ) d ϑ d τ ] d θ d t . (21)
For ξ = ( ξ 1 , ξ 2 ) ∈ S 1 \xi=(\xi_1,\xi_2)\in S^1 ξ = ( ξ 1 , ξ 2 ) ∈ S 1 , define
Φ α ( ξ ) = 1 2 ∫ R ∫ − π π ∣ τ ξ 1 + ϑ ξ 2 ∣ K α ( τ , ϑ ) d ϑ d τ . (22) \Phi_\alpha(\xi)=\frac{1}{2}\int_{\mathbb{R}}\int_{-\pi}^{\pi}|\tau\xi_1+\vartheta\xi_2|\mathcal{K}_\alpha(\tau,\vartheta)\,d\vartheta\,d\tau. \tag*{(22)} Φ α ( ξ ) = 2 1 ∫ R ∫ − π π ∣ τ ξ 1 + ϑ ξ 2 ∣ K α ( τ , ϑ ) d ϑ d τ . (22)
By (15) , the kernel K α \mathcal{K}_\alpha K α is even in each variable. Hence
Φ α ( ξ 1 , ξ 2 ) = Φ α ( ∣ ξ 1 ∣ , ∣ ξ 2 ∣ ) . \Phi_\alpha(\xi_1,\xi_2)=\Phi_\alpha(|\xi_1|,|\xi_2|). Φ α ( ξ 1 , ξ 2 ) = Φ α ( ∣ ξ 1 ∣ , ∣ ξ 2 ∣ ) .
For every ξ ∈ S 1 \xi\in S^1 ξ ∈ S 1 , there therefore exists a unique β ∈ [ 0 , π / 2 ] \beta\in[0,\pi/2] β ∈ [ 0 , π /2 ] such that
( ∣ ξ 1 ∣ , ∣ ξ 2 ∣ ) = ( cos β , sin β ) . (|\xi_1|,|\xi_2|)=(\cos\beta,\sin\beta). ( ∣ ξ 1 ∣ , ∣ ξ 2 ∣ ) = ( cos β , sin β ) .
Using again the evenness of K α \mathcal{K}_{\alpha} K α in each variable, we obtain
Φ α ( cos β , sin β ) = ∫ 0 ∞ ∫ 0 π ( ∣ τ cos β + ϑ sin β ∣ + ∣ τ cos β − ϑ sin β ∣ ) K α ( τ , ϑ ) d ϑ d τ = 2 ∫ 0 ∞ ∫ 0 π max { τ cos β , ϑ sin β } K α ( τ , ϑ ) d ϑ d τ = A α ( β ) , \begin{aligned}
\Phi_{\alpha}(\cos\beta,\sin\beta)=\int_{0}^{\infty}\int_{0}^{\pi}\left(|\tau\cos\beta+\vartheta\sin\beta|+|\tau\cos\beta-\vartheta\sin\beta|\right)\mathcal{K}_{\alpha}(\tau,\vartheta)\,\mathrm{d}\vartheta\,\mathrm{d}\tau\\
&=2\int_{0}^{\infty}\int_{0}^{\pi}\max\{\tau\cos\beta,\vartheta\sin\beta\}\mathcal{K}_{\alpha}(\tau,\vartheta)\,\mathrm{d}\vartheta\,\mathrm{d}\tau\\
&=A_{\alpha}(\beta),
\end{aligned} Φ α ( cos β , sin β ) = ∫ 0 ∞ ∫ 0 π ( ∣ τ cos β + ϑ sin β ∣ + ∣ τ cos β − ϑ sin β ∣ ) K α ( τ , ϑ ) d ϑ d τ = 2 ∫ 0 ∞ ∫ 0 π max { τ cos β , ϑ sin β } K α ( τ , ϑ ) d ϑ d τ = A α ( β ) ,
where we used the elementary identity
∣ a + b ∣ + ∣ a − b ∣ = 2 max { ∣ a ∣ , ∣ b ∣ } , a , b ∈ R . |a+b|+|a-b|=2\max\{|a|,|b|\},\qquad a,b\in\mathbb{R}. ∣ a + b ∣ + ∣ a − b ∣ = 2 max { ∣ a ∣ , ∣ b ∣ } , a , b ∈ R .
Consequently,
sup ξ ∈ S 1 Φ α ( ξ ) = sup 0 ≤ β ≤ π / 2 A α ( β ) . (23) \sup_{\xi\in S^1}\Phi_{\alpha}(\xi)=\sup_{0\leq\beta\leq\pi/2}A_{\alpha}(\beta). \tag*{(23)} ξ ∈ S 1 sup Φ α ( ξ ) = 0 ≤ β ≤ π /2 sup A α ( β ) . (23)
For every ( t , θ ) ∈ E (t,\theta)\in E ( t , θ ) ∈ E , set
ξ ( t , θ ) = ∇ c u ( t , θ ) ∣ ∇ c u ( t , θ ) ∣ ∈ S 1 . \xi(t,\theta)=\frac{\nabla_c u(t,\theta)}{|\nabla_c u(t,\theta)|}\in S^1. ξ ( t , θ ) = ∣ ∇ c u ( t , θ ) ∣ ∇ c u ( t , θ ) ∈ S 1 .
Then, by (22) ,
1 2 ∫ R ∫ − π π ∣ ∇ c u ( t , θ ) ⋅ ( τ , ϑ ) ∣ K α ( τ , ϑ ) d ϑ d τ = ∣ ∇ c u ( t , θ ) ∣ Φ α ( ξ ( t , θ ) ) . \frac{1}{2}\int_{\mathbb{R}}\int_{-\pi}^{\pi}|\nabla_cu(t,\theta)\cdot(\tau,\vartheta)|\mathcal{K}_{\alpha}(\tau,\vartheta)\,\mathrm{d}\vartheta\,\mathrm{d}\tau=|\nabla_cu(t,\theta)|\Phi_{\alpha}(\xi(t,\theta)). 2 1 ∫ R ∫ − π π ∣ ∇ c u ( t , θ ) ⋅ ( τ , ϑ ) ∣ K α ( τ , ϑ ) d ϑ d τ = ∣ ∇ c u ( t , θ ) ∣ Φ α ( ξ ( t , θ )) .
Hence, by (21) and (23) ,
1 γ 2 , α I α ( D α f ) ( 0 ) ≤ ( sup 0 ≤ β ≤ π / 2 A α ( β ) ) ∫ E ∣ ∇ c u ( t , θ ) ∣ d θ d t = ( sup 0 ≤ β ≤ π / 2 A α ( β ) ) ∫ R ∫ 0 2 π ∣ ∇ c u ( t , θ ) ∣ d θ d t . \begin{aligned}
\frac{1}{\gamma_{2,\alpha}}I_{\alpha}(\mathcal{D}^{\alpha}f)(0)\leq\left(\sup_{0\leq\beta\leq\pi/2}A_{\alpha}(\beta)\right)\int_E|\nabla_cu(t,\theta)|\,\mathrm{d}\theta\,\mathrm{d}t \\
&=\left(\sup_{0\leq\beta\leq\pi/2}A_{\alpha}(\beta)\right)\int_{\mathbb{R}}\int_{0}^{2\pi}|\nabla_cu(t,\theta)|\,\mathrm{d}\theta\,\mathrm{d}t.
\end{aligned} γ 2 , α 1 I α ( D α f ) ( 0 ) ≤ ( 0 ≤ β ≤ π /2 sup A α ( β ) ) ∫ E ∣ ∇ c u ( t , θ ) ∣ d θ d t = ( 0 ≤ β ≤ π /2 sup A α ( β ) ) ∫ R ∫ 0 2 π ∣ ∇ c u ( t , θ ) ∣ d θ d t .
Together with (18) , this gives
R α ( f , 0 ) ≤ ( 1 − α ) γ 2 , α γ 2 , 1 sup 0 ≤ β ≤ π / 2 A α ( β ) . R_{\alpha}(f,0)\leq(1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}\sup_{0\leq\beta\leq\pi/2}A_{\alpha}(\beta). R α ( f , 0 ) ≤ ( 1 − α ) γ 2 , 1 γ 2 , α 0 ≤ β ≤ π /2 sup A α ( β ) .
Taking the supremum over f ∈ C c ∞ ( R 2 ) ∖ { 0 } f\in C^\infty_c(\mathbb{R}^2)\setminus\{0\} f ∈ C c ∞ ( R 2 ) ∖ { 0 } yields (17) . □ \square □
Lemma 2.2. As α → 1 − \alpha\to1^{-} α → 1 − ,
A α ( β ) = 4 1 − α + O ( 1 ) , (24) A_{\alpha}(\beta)=\frac{4}{1-\alpha}+O(1), \tag*{(24)} A α ( β ) = 1 − α 4 + O ( 1 ) , (24) uniformly for β ∈ [ 0 , π / 2 ] \beta\in[0,\pi/2] β ∈ [ 0 , π /2 ] .
Proof. Throughout the proof, C > 0 C>0 C > 0 denotes a constant independent of α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) , β ∈ [ 0 , π / 2 ] \beta\in[0,\pi/2] β ∈ [ 0 , π /2 ] , and the integration variables, whose value may change from line to line.
For τ > 0 \tau>0 τ > 0 and 0 < ϑ < π 0<\vartheta<\pi 0 < ϑ < π , set
ρ = ( τ 2 + ϑ 2 ) 1 / 2 . \rho=(\tau^2+\vartheta^2)^{1/2}. ρ = ( τ 2 + ϑ 2 ) 1/2 .
By (16) , for every α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) and β ∈ [ 0 , π / 2 ] \beta\in[0,\pi/2] β ∈ [ 0 , π /2 ] , we may write
A α ( β ) = 4 ∫ τ , ϑ > 0 τ 2 + ϑ 2 < 1 max { τ cos β , ϑ sin β } ( τ 2 + ϑ 2 ) 1 + α / 2 d ϑ d τ + R α ( β ) , (25) A_\alpha(\beta) = 4 \int_{\substack{\tau,\vartheta>0\\ \tau^2+\vartheta^2<1}} \frac{\max\{\tau\cos\beta,\vartheta\sin\beta\}}{(\tau^2+\vartheta^2)^{1+\alpha/2}}\,d\vartheta\,d\tau+ R_\alpha(\beta), \tag*{(25)} A α ( β ) = 4 ∫ τ , ϑ > 0 τ 2 + ϑ 2 < 1 ( τ 2 + ϑ 2 ) 1 + α /2 max { τ cos β , ϑ sin β } d ϑ d τ + R α ( β ) , (25)
where
R α ( β ) = 2 ∫ τ , ϑ > 0 ρ > 1 max { τ cos β , ϑ sin β } ( K α ( τ , ϑ ) − 2 ρ 2 + α ) d ϑ d τ + 2 ∫ τ > 0 , 0 < ϑ < π ρ ≥ 1 max { τ cos β , ϑ sin β } K α ( τ , ϑ ) d ϑ d τ . (26) \begin{aligned}
R_\alpha(\beta) = 2 \int_{\substack{\tau,\vartheta>0\\ \rho>1}} \max\{\tau\cos\beta,\vartheta\sin\beta\}\left(\mathcal{K}_\alpha(\tau,\vartheta)-\frac{2}{\rho^{2+\alpha}}\right)\,d\vartheta\,d\tau\\
&+ 2\int_{\substack{\tau>0,\ 0<\vartheta<\pi\\ \rho\geq1}}\max\{\tau\cos\beta,\vartheta\sin\beta\}\mathcal{K}_\alpha(\tau,\vartheta)\,d\vartheta\,d\tau. \tag*{(26)}
\end{aligned} R α ( β ) = 2 ∫ τ , ϑ > 0 ρ > 1 max { τ cos β , ϑ sin β } ( K α ( τ , ϑ ) − ρ 2 + α 2 ) d ϑ d τ + 2 ∫ τ > 0 , 0 < ϑ < π ρ ≥ 1 max { τ cos β , ϑ sin β } K α ( τ , ϑ ) d ϑ d τ . (26)
We claim that
sup 1 / 2 ≤ α < 1 0 ≤ β ≤ π / 2 ∣ R α ( β ) ∣ < ∞ . (27) \sup_{\substack{1/2\leq\alpha<1\\0\leq\beta\leq\pi/2}} \left|R_\alpha(\beta)\right|<\infty. \tag*{(27)} 1/2 ≤ α < 1 0 ≤ β ≤ π /2 sup ∣ R α ( β ) ∣ < ∞. (27)
Indeed, fix α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) and β ∈ [ 0 , π / 2 ] \beta\in[0,\pi/2] β ∈ [ 0 , π /2 ] . By (15) and elementary Taylor estimates, we obtain
∣ K α ( τ , ϑ ) − 2 ρ 2 + α ∣ ≤ C ρ − α , 0 < ρ < 1. (28) \left|\mathcal{K}_\alpha(\tau,\vartheta)-\frac{2}{\rho^{2+\alpha}}\right|\leq C\rho^{-\alpha},\qquad0<\rho<1. \tag*{(28)} K α ( τ , ϑ ) − ρ 2 + α 2 ≤ C ρ − α , 0 < ρ < 1. (28)
Moreover, for every τ > 0 \tau>0 τ > 0 and 0 < ϑ < π 0<\vartheta<\pi 0 < ϑ < π , we have
max { τ cos β , ϑ sin β } ≤ ρ . \max\{\tau\cos\beta,\vartheta\sin\beta\}\leq\rho. max { τ cos β , ϑ sin β } ≤ ρ .
Thus, by (28) and using polar coordinates in the ( τ , ϑ ) (\tau,\vartheta) ( τ , ϑ ) -plane, we obtain
∣ ∫ τ , ϑ > 0 ρ < 1 max { τ cos β , ϑ sin β } ( K α ( τ , ϑ ) − 2 ρ 2 + α ) d ϑ d τ ∣ ≤ C ∫ 0 1 ρ 2 − α d ρ ≤ C . (29) \left|\int_{\substack{\tau,\vartheta>0\\\rho<1}}\max\{\tau\cos\beta,\vartheta\sin\beta\}\left(\mathcal{K}_\alpha(\tau,\vartheta)-\frac{2}{\rho^{2+\alpha}}\right)\,d\vartheta\,d\tau\right|
\leq C\int_0^1\rho^{2-\alpha}\,d\rho
\leq C. \tag*{(29)} ∫ τ , ϑ > 0 ρ < 1 max { τ cos β , ϑ sin β } ( K α ( τ , ϑ ) − ρ 2 + α 2 ) d ϑ d τ ≤ C ∫ 0 1 ρ 2 − α d ρ ≤ C . (29)
On the complementary region ρ ≥ 1 \rho\geq1 ρ ≥ 1 , the kernel K α \mathcal{K}_\alpha K α is uniformly bounded for 0 ≤ τ ≤ 2 0\leq\tau\leq2 0 ≤ τ ≤ 2 , since its denominator is bounded away from zero. Moreover, for τ ≥ 2 \tau\geq2 τ ≥ 2 ,
K α ( τ , ϑ ) ≤ C e − τ / 2 , 0 < ϑ < π . \mathcal{K}_\alpha(\tau,\vartheta)\leq Ce^{-\tau/2},\qquad0<\vartheta<\pi. K α ( τ , ϑ ) ≤ C e − τ /2 , 0 < ϑ < π .
Since
max { τ cos β , ϑ sin β } ≤ τ + π , τ > 0 , 0 < ϑ < π , \max\{\tau\cos\beta,\vartheta\sin\beta\}\leq\tau+\pi,\qquad\tau>0,\quad0<\vartheta<\pi, max { τ cos β , ϑ sin β } ≤ τ + π , τ > 0 , 0 < ϑ < π ,
and
∫ 2 ∞ ( τ + π ) e − τ / 2 d τ < ∞ , \int_2^\infty(\tau+\pi)e^{-\tau/2}\,d\tau<\infty, ∫ 2 ∞ ( τ + π ) e − τ /2 d τ < ∞ ,
we obtain
∣ ∫ τ > 0 , 0 < ϑ < π ρ ≥ 1 max { τ cos β , ϑ sin β } K α ( τ , ϑ ) d ϑ d τ ∣ ≤ C . (30) \left|\int_{\substack{\tau>0,\ 0<\vartheta<\pi\\\rho\geq1}}\max\{\tau\cos\beta,\vartheta\sin\beta\}\mathcal{K}_\alpha(\tau,\vartheta)\,d\vartheta\,d\tau\right|\leq C. \tag*{(30)} ∫ τ > 0 , 0 < ϑ < π ρ ≥ 1 max { τ cos β , ϑ sin β } K α ( τ , ϑ ) d ϑ d τ ≤ C . (30)
Combining (26) , (29) , and (30) , we obtain (27) .
We now evaluate the principal term in (25) . Using the polar change of variables
τ = ρ cos φ , ϑ = ρ sin φ , 0 < ρ < 1 , 0 < φ < π 2 , \tau=\rho\cos\varphi,\qquad\vartheta=\rho\sin\varphi,\qquad0<\rho<1,\qquad0<\varphi<\frac{\pi}{2}, τ = ρ cos φ , ϑ = ρ sin φ , 0 < ρ < 1 , 0 < φ < 2 π ,
we obtain for α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) ,
4 ∫ τ , ϑ > 0 τ 2 + ϑ 2 < 1 max { τ cos β , ϑ sin β } ( τ 2 + ϑ 2 ) 1 + α / 2 d ϑ d τ = 4 ∫ 0 1 ρ − α d ρ ∫ 0 π / 2 max { cos φ cos β , sin φ sin β } d φ . 4\int_{\substack{\tau,\vartheta>0\\ \tau^2+\vartheta^2<1}}\frac{\max\{\tau\cos\beta,\vartheta\sin\beta\}}{(\tau^2+\vartheta^2)^{1+\alpha/2}}\,d\vartheta\,d\tau
=4\int_0^1\rho^{-\alpha}\,d\rho\int_0^{\pi/2}\max\{\cos\varphi\cos\beta,\sin\varphi\sin\beta\}\,d\varphi. 4 ∫ τ , ϑ > 0 τ 2 + ϑ 2 < 1 ( τ 2 + ϑ 2 ) 1 + α /2 max { τ cos β , ϑ sin β } d ϑ d τ = 4 ∫ 0 1 ρ − α d ρ ∫ 0 π /2 max { cos φ cos β , sin φ sin β } d φ .
Splitting the integral at φ = π / 2 − β \varphi=\pi/2-\beta φ = π /2 − β yields
∫ 0 π / 2 max { cos φ cos β , sin φ sin β } d φ = cos β sin ( π 2 − β ) + sin β cos ( π 2 − β ) = 1. \int_0^{\pi/2}\max\{\cos\varphi\cos\beta,\sin\varphi\sin\beta\}\,d\varphi
=\cos\beta\sin\left(\frac{\pi}{2}-\beta\right)+\sin\beta\cos\left(\frac{\pi}{2}-\beta\right)
=1. ∫ 0 π /2 max { cos φ cos β , sin φ sin β } d φ = cos β sin ( 2 π − β ) + sin β cos ( 2 π − β ) = 1.
Consequently,
4 ∫ τ , ϑ > 0 τ 2 + ϑ 2 < 1 max { τ cos β , ϑ sin β } ( τ 2 + ϑ 2 ) 1 + α / 2 d ϑ d τ = 4 ∫ 0 1 ρ − α d ρ = 4 1 − α . 4\int_{\substack{\tau,\vartheta>0\\ \tau^2+\vartheta^2<1}}\frac{\max\{\tau\cos\beta,\vartheta\sin\beta\}}{(\tau^2+\vartheta^2)^{1+\alpha/2}}\,d\vartheta\,d\tau
=4\int_0^1\rho^{-\alpha}\,d\rho=\frac{4}{1-\alpha}. 4 ∫ τ , ϑ > 0 τ 2 + ϑ 2 < 1 ( τ 2 + ϑ 2 ) 1 + α /2 max { τ cos β , ϑ sin β } d ϑ d τ = 4 ∫ 0 1 ρ − α d ρ = 1 − α 4 .
Combining this identity with (25) and (27) , we obtain (24) . □
Lemma 2.3. For every 0 < α < 1 0<\alpha<1 0 < α < 1 ,
C 2 , α r a d = ( 1 − α ) γ 2 , α γ 2 , 1 A α ( 0 ) . (31) C^{\mathrm{rad}}_{2,\alpha}=(1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}A_\alpha(0). \tag*{(31)} C 2 , α rad = ( 1 − α ) γ 2 , 1 γ 2 , α A α ( 0 ) . (31) Moreover, the supremum defining C 2 , α r a d C^{\mathrm{rad}}_{2,\alpha} C 2 , α rad is attained by every nontrivial function f ∈ C c ∞ ( R 2 ) f\in C^\infty_c(\mathbb{R}^2) f ∈ C c ∞ ( R 2 ) that is radially symmetric about the evaluation point and has a monotone radial profile.
Proof. Fix 0 < α < 1 0<\alpha<1 0 < α < 1 . By (16) ,
A α ( 0 ) = 2 ∫ 0 ∞ ∫ 0 π K α ( τ , ϑ ) d ϑ d τ . (32) A_\alpha(0)=2\int_0^\infty\int_0^\pi\mathcal{K}_\alpha(\tau,\vartheta)\,d\vartheta\,d\tau. \tag*{(32)} A α ( 0 ) = 2 ∫ 0 ∞ ∫ 0 π K α ( τ , ϑ ) d ϑ d τ . (32)
By translation invariance, it suffices to consider radial functions centered at the origin. Let
f ( x ) = F ( ∣ x ∣ ) ∈ C c ∞ ( R 2 ) ∖ { 0 } , f(x)=F(|x|)\in C^\infty_c(\mathbb{R}^2)\setminus\{0\}, f ( x ) = F ( ∣ x ∣ ) ∈ C c ∞ ( R 2 ) ∖ { 0 } ,
and set
U ( t ) = F ( e t ) , t ∈ R . U(t)=F(e^t),\qquad t\in\mathbb{R}. U ( t ) = F ( e t ) , t ∈ R .
By (18) ,
I 1 ( ∣ ∇ f ∣ ) ( 0 ) γ 2 , 1 = 2 π ∫ R ∣ U ′ ( t ) ∣ d t . (33) \frac{I_1(|\nabla f|)(0)}{\gamma_{2,1}}=2\pi\int_{\mathbb{R}}|U'(t)|\,dt. \tag*{(33)} γ 2 , 1 I 1 ( ∣∇ f ∣ ) ( 0 ) = 2 π ∫ R ∣ U ′ ( t ) ∣ d t . (33)
Using the evenness of the function
τ ↦ ∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t \tau\mapsto\int_{\mathbb{R}}|U(t+\tau)-U(t)|\,dt τ ↦ ∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t
and the evenness of K α \mathcal{K}_\alpha K α in each variable, together with (19) , we obtain
I α ( D α f ) ( 0 ) γ 2 , α = 4 π ∫ 0 ∞ ∫ 0 π K α ( τ , ϑ ) ( ∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t ) d ϑ d τ . (34) \frac{I_\alpha(D^\alpha f)(0)}{\gamma_{2,\alpha}}
=4\pi\int_0^\infty\int_0^\pi\mathcal{K}_\alpha(\tau,\vartheta)\left(\int_{\mathbb{R}}|U(t+\tau)-U(t)|\,dt\right)\,d\vartheta\,d\tau. \tag*{(34)} γ 2 , α I α ( D α f ) ( 0 ) = 4 π ∫ 0 ∞ ∫ 0 π K α ( τ , ϑ ) ( ∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t ) d ϑ d τ . (34)
Moreover, by Tonelli’s theorem, for every τ > 0 \tau>0 τ > 0 ,
∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t ≤ ∫ R ∫ t t + τ ∣ U ′ ( s ) ∣ d s d t = τ ∫ R ∣ U ′ ( s ) ∣ d s . (35) \int_{\mathbb{R}}|U(t+\tau)-U(t)|\,dt
\leq\int_{\mathbb{R}}\int_t^{t+\tau}|U'(s)|\,ds\,dt
=\tau\int_{\mathbb{R}}|U'(s)|\,ds. \tag*{(35)} ∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t ≤ ∫ R ∫ t t + τ ∣ U ′ ( s ) ∣ d s d t = τ ∫ R ∣ U ′ ( s ) ∣ d s . (35)
Combining this estimate with (32) , (33) , and (34) , we obtain
R α ( f , 0 ) ≤ ( 1 − α ) γ 2 , α γ 2 , 1 2 ∫ 0 ∞ ∫ 0 π τ K α ( τ , ϑ ) d ϑ d τ = ( 1 − α ) γ 2 , α γ 2 , 1 A α ( 0 ) . \begin{aligned}
R_{\alpha}(f,0) \le(1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}2\int_{0}^{\infty}\int_{0}^{\pi}\tau\mathcal{K}_{\alpha}(\tau,\vartheta)\,d\vartheta\,d\tau\\
&= (1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}A_{\alpha}(0).
\end{aligned} R α ( f , 0 ) ≤ ( 1 − α ) γ 2 , 1 γ 2 , α 2 ∫ 0 ∞ ∫ 0 π τ K α ( τ , ϑ ) d ϑ d τ = ( 1 − α ) γ 2 , 1 γ 2 , α A α ( 0 ) .
Taking the supremum over the radial class yields
C 2 , α r a d ≤ ( 1 − α ) γ 2 , α γ 2 , 1 A α ( 0 ) . (36) C^{\mathrm{rad}}_{2,\alpha}\le(1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}A_{\alpha}(0). \tag*{(36)} C 2 , α rad ≤ ( 1 − α ) γ 2 , 1 γ 2 , α A α ( 0 ) . (36)
Assume now that F F F is monotone. Then U U U is monotone and, for every τ > 0 \tau>0 τ > 0 and t ∈ R t\in\mathbb{R} t ∈ R ,
∣ U ( t + τ ) − U ( t ) ∣ = ∫ t t + τ ∣ U ′ ( s ) ∣ d s . \lvert U(t+\tau)-U(t)\rvert=\int_{t}^{t+\tau}\lvert U'(s)\rvert\,ds. ∣ U ( t + τ ) − U ( t )∣ = ∫ t t + τ ∣ U ′ ( s )∣ d s .
Consequently, equality holds in (35) , and hence
R α ( f , 0 ) = ( 1 − α ) γ 2 , α γ 2 , 1 A α ( 0 ) . R_{\alpha}(f,0)=(1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}A_{\alpha}(0). R α ( f , 0 ) = ( 1 − α ) γ 2 , 1 γ 2 , α A α ( 0 ) .
Thus, the upper bound in (36) is attained by every such function. This proves (31) and shows that the supremum is a maximum. □
We next compute A α ( 0 ) A_{\alpha}(0) A α ( 0 ) . Recall that, for a ∈ R a\in\mathbb{R} a ∈ R and k ∈ N 0 k\in\mathbb{N}_{0} k ∈ N 0 , the Pochhammer symbol is defined by
( a ) k = { 1 , k = 0 , a ( a + 1 ) ⋯ ( a + k − 1 ) , k ≥ 1. (a)_k=
\begin{cases}
1, & k=0,\\
a(a+1)\cdots(a+k-1), & k\ge1.
\end{cases} ( a ) k = { 1 , a ( a + 1 ) ⋯ ( a + k − 1 ) , k = 0 , k ≥ 1.
For a ∉ { 0 , − 1 , − 2 , … } a\notin\{0,-1,-2,\ldots\} a ∈ / { 0 , − 1 , − 2 , … } , we also have
( a ) k = Γ ( a + k ) Γ ( a ) . (a)_k=\frac{\Gamma(a+k)}{\Gamma(a)}. ( a ) k = Γ ( a ) Γ ( a + k ) .
Lemma 2.4. For every 0 < α < 1 0<\alpha<1 0 < α < 1 , we have
A α ( 0 ) = 2 π α 2 ( 2 Γ ( 1 − α ) Γ ( 1 − α / 2 ) 2 − 1 ) . (37) A_{\alpha}(0)=\frac{2\pi}{\alpha^{2}}\left(\frac{2\Gamma(1-\alpha)}{\Gamma(1-\alpha/2)^{2}}-1\right). \tag*{(37)} A α ( 0 ) = α 2 2 π ( Γ ( 1 − α /2 ) 2 2Γ ( 1 − α ) − 1 ) . (37)
Proof. Fix 0 < α < 1 0<\alpha<1 0 < α < 1 . By (32) ,
A α ( 0 ) = 2 ∫ 0 ∞ ∫ 0 π τ K α ( τ , ϑ ) d ϑ d τ . A_{\alpha}(0)=2\int_{0}^{\infty}\int_{0}^{\pi}\tau\mathcal{K}_{\alpha}(\tau,\vartheta)\,d\vartheta\,d\tau. A α ( 0 ) = 2 ∫ 0 ∞ ∫ 0 π τ K α ( τ , ϑ ) d ϑ d τ .
Using (15) and
2 ( cosh τ − cos ϑ ) = e τ ( 1 − 2 e − τ cos ϑ + e − 2 τ ) , 2(\cosh\tau-\cos\vartheta)=e^{\tau}(1-2e^{-\tau}\cos\vartheta+e^{-2\tau}), 2 ( cosh τ − cos ϑ ) = e τ ( 1 − 2 e − τ cos ϑ + e − 2 τ ) ,
we obtain
K α ( τ , ϑ ) = e − α τ + e − 2 τ ( 1 − 2 e − τ cos ϑ + e − 2 τ ) 1 + α / 2 . \mathcal{K}_{\alpha}(\tau,\vartheta)=\frac{e^{-\alpha\tau}+e^{-2\tau}}{(1-2e^{-\tau}\cos\vartheta+e^{-2\tau})^{1+\alpha/2}}. K α ( τ , ϑ ) = ( 1 − 2 e − τ cos ϑ + e − 2 τ ) 1 + α /2 e − α τ + e − 2 τ .
Hence
A α ( 0 ) = 2 ∫ 0 ∞ τ ( e − α τ + e − 2 τ ) [ ∫ 0 π d ϑ ( 1 − 2 e − τ cos ϑ + e − 2 τ ) 1 + α / 2 ] d τ . (38) A_{\alpha}(0)=2\int_{0}^{\infty}\tau(e^{-\alpha\tau}+e^{-2\tau})\left[\int_{0}^{\pi}\frac{d\vartheta}{(1-2e^{-\tau}\cos\vartheta+e^{-2\tau})^{1+\alpha/2}}\right]d\tau. \tag*{(38)} A α ( 0 ) = 2 ∫ 0 ∞ τ ( e − α τ + e − 2 τ ) [ ∫ 0 π ( 1 − 2 e − τ cos ϑ + e − 2 τ ) 1 + α /2 d ϑ ] d τ . (38)
Set
a = α 2 ∈ ( 0 , 1 2 ) , r = e − τ ∈ ( 0 , 1 ) . a=\frac{\alpha}{2}\in\left(0,\frac{1}{2}\right),\qquad r=e^{-\tau}\in(0,1). a = 2 α ∈ ( 0 , 2 1 ) , r = e − τ ∈ ( 0 , 1 ) .
Then
∫ 0 π d ϑ ( 1 − 2 e − τ cos ϑ + e − 2 τ ) 1 + α / 2 = ∫ 0 π d ϑ ( 1 − 2 r cos ϑ + r 2 ) 1 + a . \int_{0}^{\pi}\frac{d\vartheta}{\left(1-2e^{-\tau}\cos\vartheta+e^{-2\tau}\right)^{1+\alpha/2}}
=
\int_{0}^{\pi}\frac{d\vartheta}{\left(1-2r\cos\vartheta+r^2\right)^{1+a}}. ∫ 0 π ( 1 − 2 e − τ cos ϑ + e − 2 τ ) 1 + α /2 d ϑ = ∫ 0 π ( 1 − 2 r cos ϑ + r 2 ) 1 + a d ϑ .
Using the standard hypergeometric identity (see, e.g., [2, Section 2.1})
∫ 0 π d ϑ ( 1 − 2 r cos ϑ + r 2 ) 1 + a = π ∑ k = 0 ∞ ( ( 1 + a ) k k ! ) 2 r 2 k , 0 < r < 1 , \int_{0}^{\pi}\frac{d\vartheta}{\left(1-2r\cos\vartheta+r^2\right)^{1+a}}
=
\pi\sum_{k=0}^{\infty}\left(\frac{(1+a)_k}{k!}\right)^2r^{2k},
\qquad0<r<1, ∫ 0 π ( 1 − 2 r cos ϑ + r 2 ) 1 + a d ϑ = π k = 0 ∑ ∞ ( k ! ( 1 + a ) k ) 2 r 2 k , 0 < r < 1 ,
we obtain
∫ 0 π d ϑ ( 1 − 2 e − τ cos ϑ + e − 2 τ ) 1 + α / 2 = π ∑ k = 0 ∞ ( ( 1 + a ) k k ! ) 2 e − 2 k τ . \int_{0}^{\pi}\frac{d\vartheta}{\left(1-2e^{-\tau}\cos\vartheta+e^{-2\tau}\right)^{1+\alpha/2}}
=
\pi\sum_{k=0}^{\infty}\left(\frac{(1+a)_k}{k!}\right)^2e^{-2k\tau}. ∫ 0 π ( 1 − 2 e − τ cos ϑ + e − 2 τ ) 1 + α /2 d ϑ = π k = 0 ∑ ∞ ( k ! ( 1 + a ) k ) 2 e − 2 k τ .
Combining this identity with (38) , applying Tonelli’s theorem, and using
∫ 0 ∞ τ e − λ τ d τ = 1 λ 2 , λ > 0 , \int_{0}^{\infty}\tau e^{-\lambda\tau}\,d\tau=\frac{1}{\lambda^2},\qquad\lambda>0, ∫ 0 ∞ τ e − λ τ d τ = λ 2 1 , λ > 0 ,
we obtain
A α ( 0 ) = π 2 ∑ k = 0 ∞ ( ( 1 + a ) k k ! ) 2 [ 1 ( k + a ) 2 + 1 ( k + 1 ) 2 ] . A_\alpha(0)=\frac{\pi}{2}\sum_{k=0}^{\infty}\left(\frac{(1+a)_k}{k!}\right)^2\left[\frac{1}{(k+a)^2}+\frac{1}{(k+1)^2}\right]. A α ( 0 ) = 2 π k = 0 ∑ ∞ ( k ! ( 1 + a ) k ) 2 [ ( k + a ) 2 1 + ( k + 1 ) 2 1 ] .
Using
( 1 + a ) k k + a = ( a ) k a , ( 1 + a ) k k + 1 = ( a ) k + 1 a ( k + 1 ) , \frac{(1+a)_k}{k+a}=\frac{(a)_k}{a},\qquad
\frac{(1+a)_k}{k+1}=\frac{(a)_{k+1}}{a(k+1)}, k + a ( 1 + a ) k = a ( a ) k , k + 1 ( 1 + a ) k = a ( k + 1 ) ( a ) k + 1 ,
we obtain
A α ( 0 ) = π 2 a 2 [ 2 ∑ k = 0 ∞ ( ( a ) k k ! ) 2 − 1 ] . A_\alpha(0)=\frac{\pi}{2a^2}\left[2\sum_{k=0}^{\infty}\left(\frac{(a)_k}{k!}\right)^2-1\right]. A α ( 0 ) = 2 a 2 π [ 2 k = 0 ∑ ∞ ( k ! ( a ) k ) 2 − 1 ] .
Since 0 < a < 1 / 2 0<a<1/2 0 < a < 1/2 , Gauss’ summation formula (see, e.g., [2, Theorem 2.2.2}) gives
∑ k = 0 ∞ ( ( a ) k k ! ) 2 = 2 F 1 ( a , a ; 1 ; 1 ) = Γ ( 1 − 2 a ) Γ ( 1 − a ) 2 . \sum_{k=0}^{\infty}\left(\frac{(a)_k}{k!}\right)^2
={}_2F_1(a,a;1;1)
=\frac{\Gamma(1-2a)}{\Gamma(1-a)^2}. k = 0 ∑ ∞ ( k ! ( a ) k ) 2 = 2 F 1 ( a , a ; 1 ; 1 ) = Γ ( 1 − a ) 2 Γ ( 1 − 2 a ) .
Substituting this identity and recalling that a = α / 2 a=\alpha/2 a = α /2 , we obtain (37) .
We now prove Theorems 1.1 and 1.2.
Proof of Theorem 1.1
Proof. By Lemmas 2.1 and 2.3, we obtain, for every 0 < α < 1 0<\alpha<1 0 < α < 1 ,
( 1 − α ) γ 2 , α γ 2 , 1 A α ( 0 ) = C 2 , α r a d ≤ C 2 , α ∗ ≤ ( 1 − α ) γ 2 , α γ 2 , 1 sup β ∈ [ 0 , π / 2 ] A α ( β ) . (1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}A_\alpha(0)
=C_{2,\alpha}^{\mathrm{rad}}
\le C_{2,\alpha}^{*}
\le(1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}
\sup_{\beta\in[0,\pi/2]}A_\alpha(\beta). ( 1 − α ) γ 2 , 1 γ 2 , α A α ( 0 ) = C 2 , α rad ≤ C 2 , α ∗ ≤ ( 1 − α ) γ 2 , 1 γ 2 , α β ∈ [ 0 , π /2 ] sup A α ( β ) .
By Lemma 2.2,
A α ( β ) = 4 1 − α + O ( 1 ) , α → 1 − , A_\alpha(\beta)=\frac{4}{1-\alpha}+O(1),\qquad\alpha\to1^-, A α ( β ) = 1 − α 4 + O ( 1 ) , α → 1 − ,
uniformly for β ∈ [ 0 , π / 2 ] \beta\in[0,\pi/2] β ∈ [ 0 , π /2 ] . On the other hand, the normalization (1) yields
γ 2 , α γ 2 , 1 = 2 1 − α Γ ( 1 − α / 2 ) Γ ( α / 2 ) , 0 < α < 1. \frac{\gamma_{2,\alpha}}{\gamma_{2,1}}=2^{1-\alpha}\frac{\Gamma(1-\alpha/2)}{\Gamma(\alpha/2)},\qquad0<\alpha<1. γ 2 , 1 γ 2 , α = 2 1 − α Γ ( α /2 ) Γ ( 1 − α /2 ) , 0 < α < 1.
By the smoothness of the Gamma function,
γ 2 , α γ 2 , 1 = 1 + O ( 1 − α ) , α → 1 − . \frac{\gamma_{2,\alpha}}{\gamma_{2,1}}=1+O(1-\alpha),\qquad\alpha\to1^{-}. γ 2 , 1 γ 2 , α = 1 + O ( 1 − α ) , α → 1 − .
Combining these estimates with the preceding two-sided inequality, we obtain (7) . □
Proof of Theorem 1.2
Proof. By Lemmas 2.3 and 2.4, and recalling (1) , we obtain, for every 0 < α < 1 0<\alpha<1 0 < α < 1 ,
C 2 , α r a d = ( 1 − α ) γ 2 , α γ 2 , 1 A α ( 0 ) = 2 2 − α π ( 1 − α ) α 2 Γ ( 1 − α / 2 ) Γ ( α / 2 ) ( 2 Γ ( 1 − α ) Γ ( 1 − α / 2 ) 2 − 1 ) . \begin{aligned}
C^{\mathrm{rad}}_{2,\alpha}=(1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}A_\alpha(0) \\
&=\frac{2^{2-\alpha}\pi(1-\alpha)}{\alpha^2}\frac{\Gamma(1-\alpha/2)}{\Gamma(\alpha/2)}
\left(2\frac{\Gamma(1-\alpha)}{\Gamma(1-\alpha/2)^2}-1\right).
\end{aligned} C 2 , α rad = ( 1 − α ) γ 2 , 1 γ 2 , α A α ( 0 ) = α 2 2 2 − α π ( 1 − α ) Γ ( α /2 ) Γ ( 1 − α /2 ) ( 2 Γ ( 1 − α /2 ) 2 Γ ( 1 − α ) − 1 ) .
This establishes (8) . The assertion concerning the attainment of the supremum follows directly from Lemma 2.3. □
The sharp constant C n , α ∗ C^*_{n,\alpha} C n , α ∗ for n ≥ 3 n\ge3 n ≥ 3
This section is devoted to the proof of Theorem 1.3 and the derivation of Corollary 1.4. Throughout this section, we assume that n ≥ 3 n\ge3 n ≥ 3 . We begin with several auxiliary lemmas.
We first derive logarithmic-polar representations of the Riesz potentials involved in our problem. For 0 < α < 1 0<\alpha<1 0 < α < 1 , τ ∈ R \tau\in\mathbb{R} τ ∈ R , and 0 ≤ ϑ ≤ π 0\le\vartheta\le\pi 0 ≤ ϑ ≤ π , with ( τ , ϑ ) ≠ ( 0 , 0 ) (\tau,\vartheta)\ne(0,0) ( τ , ϑ ) = ( 0 , 0 ) , define
K n , α ( τ , ϑ ) = 2 cosh ( n − α 2 τ ) [ 2 ( cosh τ − cos ϑ ) ] ( n + α ) / 2 . (39) \mathcal{K}_{n,\alpha}(\tau,\vartheta)=\frac{2\cosh\left(\frac{n-\alpha}{2}\tau\right)}{\left[2(\cosh\tau-\cos\vartheta)\right]^{(n+\alpha)/2}}. \tag*{(39)} K n , α ( τ , ϑ ) = [ 2 ( cosh τ − cos ϑ ) ] ( n + α ) /2 2 cosh ( 2 n − α τ ) . (39)
Lemma 3.1. Let 0 < α < 1 0<\alpha<1 0 < α < 1 and f ∈ C c ∞ ( R n ) f\in C^\infty_c(\mathbb{R}^n) f ∈ C c ∞ ( R n ) . Set
u ( t , ω ) = f ( e t ω ) , ( t , ω ) ∈ R × S n − 1 . u(t,\omega)=f(e^t\omega),\qquad(t,\omega)\in\mathbb{R}\times\mathbb{S}^{n-1}. u ( t , ω ) = f ( e t ω ) , ( t , ω ) ∈ R × S n − 1 .
Then
I 1 ( ∣ ∇ f ∣ ) ( 0 ) γ n , 1 = ∫ R ∫ S n − 1 ∣ ∇ c u ( t , ω ) ∣ d σ ( ω ) d t . (40) \frac{I_1(|\nabla f|)(0)}{\gamma_{n,1}}=\int_{\mathbb{R}}\int_{\mathbb{S}^{n-1}}|\nabla_cu(t,\omega)|\,d\sigma(\omega)\,dt. \tag*{(40)} γ n , 1 I 1 ( ∣∇ f ∣ ) ( 0 ) = ∫ R ∫ S n − 1 ∣ ∇ c u ( t , ω ) ∣ d σ ( ω ) d t . (40)
where ∇ c \nabla_c ∇ c denotes the gradient on the cylinder R × S n − 1 \mathbb{R}\times\mathbb{S}^{n-1} R × S n − 1 , defined by
∇ c u ( t , ω ) = ( ∂ t u ( t , ω ) , ∇ S n − 1 u ( t , ω ) ) , ( t , ω ) ∈ R × S n − 1 , \nabla_cu(t,\omega)=\left(\partial_tu(t,\omega),\nabla_{\mathbb{S}^{n-1}}u(t,\omega)\right),\qquad(t,\omega)\in\mathbb{R}\times\mathbb{S}^{n-1}, ∇ c u ( t , ω ) = ( ∂ t u ( t , ω ) , ∇ S n − 1 u ( t , ω ) ) , ( t , ω ) ∈ R × S n − 1 ,
so that
∣ ∇ c u ( t , ω ) ∣ 2 = ∣ ∂ t u ( t , ω ) ∣ 2 + ∣ ∇ S n − 1 u ( t , ω ) ∣ 2 . |\nabla_cu(t,\omega)|^2=|\partial_tu(t,\omega)|^2+|\nabla_{\mathbb{S}^{n-1}}u(t,\omega)|^2. ∣ ∇ c u ( t , ω ) ∣ 2 = ∣ ∂ t u ( t , ω ) ∣ 2 + ∣ ∇ S n − 1 u ( t , ω ) ∣ 2 .
Moreover,
I α ( D α f ) ( 0 ) γ n , α = 1 2 ∫ R ∫ R ∫ S n − 1 ∫ S n − 1 K n , α ( τ , arccos ( ω ⋅ ν ) ) ∣ u ( t + τ , ν ) − u ( t , ω ) ∣ d σ ( ν ) d σ ( ω ) d τ d t . (41) \frac{I_\alpha(D^\alpha f)(0)}{\gamma_{n,\alpha}}
=\frac{1}{2}\int_{\mathbb{R}}\int_{\mathbb{R}}\int_{\mathbb{S}^{n-1}}\int_{\mathbb{S}^{n-1}}
\mathcal{K}_{n,\alpha}\left(\tau,\arccos(\omega\cdot\nu)\right)
|u(t+\tau,\nu)-u(t,\omega)|\,d\sigma(\nu)\,d\sigma(\omega)\,d\tau\,dt. \tag*{(41)} γ n , α I α ( D α f ) ( 0 ) = 2 1 ∫ R ∫ R ∫ S n − 1 ∫ S n − 1 K n , α ( τ , arccos ( ω ⋅ ν ) ) ∣ u ( t + τ , ν ) − u ( t , ω ) ∣ d σ ( ν ) d σ ( ω ) d τ d t . (41)
Proof. We first establish (40) . For every t ∈ R t \in\mathbb{R} t ∈ R and ω ∈ S n − 1 \omega\in S^{n-1} ω ∈ S n − 1 , the definition of u u u gives
∂ t u ( t , ω ) = e t ∇ f ( e t ω ) ⋅ ω \partial_t u(t,\omega)=e^t\nabla f(e^t\omega)\cdot\omega ∂ t u ( t , ω ) = e t ∇ f ( e t ω ) ⋅ ω
and
∇ S n − 1 u ( t , ω ) = e t ( ∇ f ( e t ω ) − ( ∇ f ( e t ω ) ⋅ ω ) ω ) . \nabla_{S^{n-1}}u(t,\omega)=e^t\left(\nabla f(e^t\omega)-(\nabla f(e^t\omega)\cdot\omega)\omega\right). ∇ S n − 1 u ( t , ω ) = e t ( ∇ f ( e t ω ) − ( ∇ f ( e t ω ) ⋅ ω ) ω ) .
Consequently,
∣ ∇ c u ( t , ω ) ∣ 2 = ∣ ∂ t u ( t , ω ) ∣ 2 + ∣ ∇ S n − 1 u ( t , ω ) ∣ 2 = e 2 t ∣ ∇ f ( e t ω ) ∣ 2 , \begin{aligned}
|\nabla_c u(t,\omega)|^2&=|\partial_tu(t,\omega)|^2+|\nabla_{S^{n-1}}u(t,\omega)|^2\\
&=e^{2t}|\nabla f(e^t\omega)|^2,
\end{aligned} ∣ ∇ c u ( t , ω ) ∣ 2 = ∣ ∂ t u ( t , ω ) ∣ 2 + ∣ ∇ S n − 1 u ( t , ω ) ∣ 2 = e 2 t ∣∇ f ( e t ω ) ∣ 2 ,
and hence
∣ ∇ c u ( t , ω ) ∣ = e t ∣ ∇ f ( e t ω ) ∣ , ( t , ω ) ∈ R × S n − 1 . (42) |\nabla_cu(t,\omega)|=e^t|\nabla f(e^t\omega)|,\qquad(t,\omega)\in\mathbb{R}\times S^{n-1}. \tag*{(42)} ∣ ∇ c u ( t , ω ) ∣ = e t ∣∇ f ( e t ω ) ∣ , ( t , ω ) ∈ R × S n − 1 . (42)
Using the polar change of variables
y = e t ω , ( t , ω ) ∈ R × S n − 1 , y=e^t\omega,\qquad(t,\omega)\in\mathbb{R}\times S^{n-1}, y = e t ω , ( t , ω ) ∈ R × S n − 1 ,
we obtain
I 1 ( ∣ ∇ f ∣ ) ( 0 ) γ n , 1 = ∫ R n ∣ ∇ f ( y ) ∣ ∣ y ∣ n − 1 d y = ∫ R ∫ S n − 1 e t ∣ ∇ f ( e t ω ) ∣ d σ ( ω ) d t = ∫ R ∫ S n − 1 ∣ ∇ c u ( t , ω ) ∣ d σ ( ω ) d t , \begin{aligned}
\frac{I_1(|\nabla f|)(0)}{\gamma_{n,1}}
&=\int_{\mathbb{R}^n}\frac{|\nabla f(y)|}{|y|^{n-1}}\,dy\\
&=\int_{\mathbb{R}}\int_{S^{n-1}}e^t|\nabla f(e^t\omega)|\,d\sigma(\omega)\,dt\\
&=\int_{\mathbb{R}}\int_{S^{n-1}}|\nabla_cu(t,\omega)|\,d\sigma(\omega)\,dt,
\end{aligned} γ n , 1 I 1 ( ∣∇ f ∣ ) ( 0 ) = ∫ R n ∣ y ∣ n − 1 ∣∇ f ( y ) ∣ d y = ∫ R ∫ S n − 1 e t ∣∇ f ( e t ω ) ∣ d σ ( ω ) d t = ∫ R ∫ S n − 1 ∣ ∇ c u ( t , ω ) ∣ d σ ( ω ) d t ,
where the last equality follows from (42) . This proves (40) .
We now prove (41) . By the definitions of I α I_\alpha I α and D α D^\alpha D α ,
I α ( D α f ) ( 0 ) γ n , α = ∫ R n ∫ R n ∣ f ( y ) − f ( z ) ∣ ∣ y ∣ n − α ∣ y − z ∣ n + α d z d y . \frac{I_\alpha(D^\alpha f)(0)}{\gamma_{n,\alpha}}=\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}\frac{|f(y)-f(z)|}{|y|^{n-\alpha}|y-z|^{n+\alpha}}\,dz\,dy. γ n , α I α ( D α f ) ( 0 ) = ∫ R n ∫ R n ∣ y ∣ n − α ∣ y − z ∣ n + α ∣ f ( y ) − f ( z ) ∣ d z d y .
Using the changes of variables
y = e t ω , z = e s ν , y=e^t\omega,\qquad z=e^s\nu, y = e t ω , z = e s ν ,
where t , s ∈ R t,s\in\mathbb{R} t , s ∈ R and ω , ν ∈ S n − 1 \omega,\nu\in S^{n-1} ω , ν ∈ S n − 1 , we have
d y = e n t d t d σ ( ω ) , d z = e n s d s d σ ( ν ) , dy=e^{nt}\,dt\,d\sigma(\omega),\qquad dz=e^{ns}\,ds\,d\sigma(\nu), d y = e n t d t d σ ( ω ) , d z = e n s d s d σ ( ν ) ,
and
∣ y − z ∣ 2 = 2 e t + s ( cosh ( s − t ) − ω ⋅ ν ) . |y-z|^2=2e^{t+s}\left(\cosh(s-t)-\omega\cdot\nu\right). ∣ y − z ∣ 2 = 2 e t + s ( cosh ( s − t ) − ω ⋅ ν ) .
Consequently,
I α ( D α f ) ( 0 ) γ n , α = ∫ R ∫ R ∫ S n − 1 ∫ S n − 1 e ( n − α ) ( s − t ) / 2 ∣ u ( s , ν ) − u ( t , ω ) ∣ [ 2 ( cosh ( s − t ) − ω ⋅ ν ) ] ( n + α ) / 2 d σ ( ν ) d σ ( ω ) d s d t . \frac{I_\alpha(D^\alpha f)(0)}{\gamma_{n,\alpha}}=\int_{\mathbb{R}}\int_{\mathbb{R}}\int_{S^{n-1}}\int_{S^{n-1}}\frac{e^{(n-\alpha)(s-t)/2}|u(s,\nu)-u(t,\omega)|}{[2(\cosh(s-t)-\omega\cdot\nu)]^{(n+\alpha)/2}}\,d\sigma(\nu)\,d\sigma(\omega)\,ds\,dt. γ n , α I α ( D α f ) ( 0 ) = ∫ R ∫ R ∫ S n − 1 ∫ S n − 1 [ 2 ( cosh ( s − t ) − ω ⋅ ν ) ] ( n + α ) /2 e ( n − α ) ( s − t ) /2 ∣ u ( s , ν ) − u ( t , ω ) ∣ d σ ( ν ) d σ ( ω ) d s d t .
Making the change of variables
( t , s ) ↦ ( t , τ ) , τ = s − t , (t,s)\mapsto(t,\tau),\qquad\tau=s-t, ( t , s ) ↦ ( t , τ ) , τ = s − t ,
we obtain
I α ( D α f ) ( 0 ) γ n , α = ∫ R ∫ R ∫ S n − 1 ∫ S n − 1 e ( n − α ) τ / 2 ∣ u ( t + τ , ν ) − u ( t , ω ) ∣ [ 2 ( cosh τ − ω ⋅ ν ) ] ( n + α ) / 2 d σ ( ν ) d σ ( ω ) d τ d t . (43) \frac{I_\alpha(D^\alpha f)(0)}{\gamma_{n,\alpha}}=\int_{\mathbb{R}}\int_{\mathbb{R}}\int_{S^{n-1}}\int_{S^{n-1}}\frac{e^{(n-\alpha)\tau/2}|u(t+\tau,\nu)-u(t,\omega)|}{[2(\cosh\tau-\omega\cdot\nu)]^{(n+\alpha)/2}}\,d\sigma(\nu)\,d\sigma(\omega)\,d\tau\,dt. \tag*{(43)} γ n , α I α ( D α f ) ( 0 ) = ∫ R ∫ R ∫ S n − 1 ∫ S n − 1 [ 2 ( cosh τ − ω ⋅ ν ) ] ( n + α ) /2 e ( n − α ) τ /2 ∣ u ( t + τ , ν ) − u ( t , ω ) ∣ d σ ( ν ) d σ ( ω ) d τ d t . (43)
Making the change of variables
( t , τ , ω , ν ) ↦ ( t + τ , − τ , ν , ω ) (t,\tau,\omega,\nu)\mapsto(t+\tau,-\tau,\nu,\omega) ( t , τ , ω , ν ) ↦ ( t + τ , − τ , ν , ω )
in (43) , we also obtain
I α ( D α f ) ( 0 ) γ n , α = ∫ R ∫ R ∫ S n − 1 ∫ S n − 1 e − ( n − α ) τ / 2 ∣ u ( t + τ , ν ) − u ( t , ω ) ∣ [ 2 ( cosh τ − ω ⋅ ν ) ] ( n + α ) / 2 d σ ( ν ) d σ ( ω ) d τ d t . \frac{I_{\alpha}(D^{\alpha}f)(0)}{\gamma_{n,\alpha}}=\int_{\mathbb{R}}\int_{\mathbb{R}}\int_{\mathbb{S}^{n-1}}\int_{\mathbb{S}^{n-1}}\frac{e^{-(n-\alpha)\tau/2}|u(t+\tau,\nu)-u(t,\omega)|}{[2(\cosh\tau-\omega\cdot\nu)]^{(n+\alpha)/2}}\,d\sigma(\nu)\,d\sigma(\omega)\,d\tau\,dt. γ n , α I α ( D α f ) ( 0 ) = ∫ R ∫ R ∫ S n − 1 ∫ S n − 1 [ 2 ( cosh τ − ω ⋅ ν ) ] ( n + α ) /2 e − ( n − α ) τ /2 ∣ u ( t + τ , ν ) − u ( t , ω ) ∣ d σ ( ν ) d σ ( ω ) d τ d t .
Averaging these two representations yields
I α ( D α f ) ( 0 ) γ n , α = ∫ R ∫ R ∫ S n − 1 ∫ S n − 1 cosh ( n − α 2 τ ) [ 2 ( cosh τ − ω ⋅ ν ) ] ( n + α ) / 2 ∣ u ( t + τ , ν ) − u ( t , ω ) ∣ d σ ( ν ) d σ ( ω ) d τ d t . \frac{I_{\alpha}(D^{\alpha}f)(0)}{\gamma_{n,\alpha}}
=\int_{\mathbb{R}}\int_{\mathbb{R}}\int_{\mathbb{S}^{n-1}}\int_{\mathbb{S}^{n-1}}\frac{\cosh\left(\frac{n-\alpha}{2}\tau\right)}{[2(\cosh\tau-\omega\cdot\nu)]^{(n+\alpha)/2}}|u(t+\tau,\nu)-u(t,\omega)|\,d\sigma(\nu)\,d\sigma(\omega)\,d\tau\,dt. γ n , α I α ( D α f ) ( 0 ) = ∫ R ∫ R ∫ S n − 1 ∫ S n − 1 [ 2 ( cosh τ − ω ⋅ ν ) ] ( n + α ) /2 cosh ( 2 n − α τ ) ∣ u ( t + τ , ν ) − u ( t , ω ) ∣ d σ ( ν ) d σ ( ω ) d τ d t .
Finally, by (39) , we obtain (41) . This completes the proof. □
Lemma 3.2. Let 0 < α < 1 0<\alpha<1 0 < α < 1 and
f ( x ) = F ( ∣ x ∣ ) ∈ C c ∞ ( R n ) ∖ { 0 } . f(x)=F(|x|)\in C^\infty_c(\mathbb{R}^n)\setminus\{0\}. f ( x ) = F ( ∣ x ∣ ) ∈ C c ∞ ( R n ) ∖ { 0 } . Then
R α ( f , 0 ) ≤ ( 1 − α ) γ n , α γ n , 1 e n , α g r a d . (44) R_\alpha(f,0)\leq(1-\alpha)\frac{\gamma_{n,\alpha}}{\gamma_{n,1}}\mathfrak{e}^{\mathrm{grad}}_{n,\alpha}. \tag*{(44)} R α ( f , 0 ) ≤ ( 1 − α ) γ n , 1 γ n , α e n , α grad . (44) Moreover, equality holds whenever F F F is monotone.
Proof. We first establish (44) . Set
U ( t ) = F ( e t ) , t ∈ R . U(t)=F(e^t),\qquad t\in\mathbb{R}. U ( t ) = F ( e t ) , t ∈ R .
Since u ( t , ω ) = U ( t ) u(t,\omega)=U(t) u ( t , ω ) = U ( t ) for every ( t , ω ) ∈ R × S n − 1 (t,\omega)\in\mathbb{R}\times\mathbb{S}^{n-1} ( t , ω ) ∈ R × S n − 1 , we have
∂ t u ( t , ω ) = U ′ ( t ) , ∇ S n − 1 u ( t , ω ) = 0. \partial_tu(t,\omega)=U'(t),\qquad\nabla_{\mathbb{S}^{n-1}}u(t,\omega)=0. ∂ t u ( t , ω ) = U ′ ( t ) , ∇ S n − 1 u ( t , ω ) = 0.
Consequently, (40) gives
I 1 ( ∣ ∇ f ∣ ) ( 0 ) γ n , 1 = ∣ S n − 1 ∣ ∫ R ∣ U ′ ( t ) ∣ d t . (45) \frac{I_1(|\nabla f|)(0)}{\gamma_{n,1}}=|\mathbb{S}^{n-1}|\int_{\mathbb{R}}|U'(t)|\,dt. \tag*{(45)} γ n , 1 I 1 ( ∣∇ f ∣ ) ( 0 ) = ∣ S n − 1 ∣ ∫ R ∣ U ′ ( t ) ∣ d t . (45)
On the other hand, for every ω ∈ S n − 1 \omega\in\mathbb{S}^{n-1} ω ∈ S n − 1 , using spherical coordinates with θ = arccos ( ω ⋅ ν ) \theta=\arccos(\omega\cdot\nu) θ = arccos ( ω ⋅ ν ) , we obtain
∫ S n − 1 K n , α ( τ , arccos ( ω ⋅ ν ) ) d σ ( ν ) = ∣ S n − 2 ∣ ∫ 0 π K n , α ( τ , θ ) ( sin θ ) n − 2 d θ , τ ∈ R ∖ { 0 } . \int_{\mathbb{S}^{n-1}}\mathcal{K}_{n,\alpha}(\tau,\arccos(\omega\cdot\nu))\,d\sigma(\nu)=|\mathbb{S}^{n-2}|\int_0^\pi\mathcal{K}_{n,\alpha}(\tau,\theta)(\sin\theta)^{n-2}\,d\theta,\qquad\tau\in\mathbb{R}\setminus\{0\}. ∫ S n − 1 K n , α ( τ , arccos ( ω ⋅ ν )) d σ ( ν ) = ∣ S n − 2 ∣ ∫ 0 π K n , α ( τ , θ ) ( sin θ ) n − 2 d θ , τ ∈ R ∖ { 0 } .
Using this identity, the evenness of K n , α \mathcal{K}_{n,\alpha} K n , α in τ \tau τ , and the evenness of the function
τ ⟼ ∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t , \tau\longmapsto\int_{\mathbb{R}}|U(t+\tau)-U(t)|\,dt, τ ⟼ ∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t ,
we deduce from (41) that
I α ( D α f ) ( 0 ) γ n , α = ∣ S n − 1 ∣ ∣ S n − 2 ∣ ∫ 0 ∞ ∫ 0 π K n , α ( τ , θ ) ( sin θ ) n − 2 ( ∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t ) d θ d τ . (46) \frac{I_\alpha(D^\alpha f)(0)}{\gamma_{n,\alpha}}
=|\mathbb{S}^{n-1}||\mathbb{S}^{n-2}|\int_0^\infty\int_0^\pi\mathcal{K}_{n,\alpha}(\tau,\theta)(\sin\theta)^{n-2}\left(\int_{\mathbb{R}}|U(t+\tau)-U(t)|\,dt\right)\,d\theta\,d\tau. \tag*{(46)} γ n , α I α ( D α f ) ( 0 ) = ∣ S n − 1 ∣∣ S n − 2 ∣ ∫ 0 ∞ ∫ 0 π K n , α ( τ , θ ) ( sin θ ) n − 2 ( ∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t ) d θ d τ . (46)
Moreover, by the fundamental theorem of calculus and Tonelli’s theorem, for every τ > 0 \tau>0 τ > 0 ,
∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t ≤ τ ∫ R ∣ U ′ ( s ) ∣ d s . \int_{\mathbb{R}}|U(t+\tau)-U(t)|\,dt\leq\tau\int_{\mathbb{R}}|U'(s)|\,ds. ∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t ≤ τ ∫ R ∣ U ′ ( s ) ∣ d s .
Consequently, (9) and (46) yield
I α ( D α f ) ( 0 ) γ n , α ≤ ∣ S n − 1 ∣ C n , α r a d ∫ R ∣ U ′ ( t ) ∣ d t . (47) \frac{I_{\alpha}(D^{\alpha}f)(0)}{\gamma_{n,\alpha}} \leq|\mathbb{S}^{n-1}|\mathcal{C}_{n,\alpha}^{\mathrm{rad}}\int_{\mathbb{R}} |U'(t)|\,dt. \tag*{(47)} γ n , α I α ( D α f ) ( 0 ) ≤ ∣ S n − 1 ∣ C n , α rad ∫ R ∣ U ′ ( t ) ∣ d t . (47)
Combining (45) and (47) , we obtain (44) .
Assume now that F F F is monotone. Then U U U is monotone and, for every τ > 0 \tau> 0 τ > 0 and t ∈ R t \in\mathbb{R} t ∈ R ,
∣ U ( t + τ ) − U ( t ) ∣ = ∫ t t + τ ∣ U ′ ( s ) ∣ d s . |U(t+\tau)-U(t)|=\int_t^{t+\tau}|U'(s)|\,ds. ∣ U ( t + τ ) − U ( t ) ∣ = ∫ t t + τ ∣ U ′ ( s ) ∣ d s .
Consequently,
∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t = τ ∫ R ∣ U ′ ( s ) ∣ d s . \int_{\mathbb{R}}|U(t+\tau)-U(t)|\,dt=\tau\int_{\mathbb{R}}|U'(s)|\,ds. ∫ R ∣ U ( t + τ ) − U ( t ) ∣ d t = τ ∫ R ∣ U ′ ( s ) ∣ d s .
Thus, equality holds in (47) , and hence
R α ( f , 0 ) = ( 1 − α ) γ n , α γ n , 1 C n , α r a d . R_{\alpha}(f,0)=(1-\alpha)\frac{\gamma_{n,\alpha}}{\gamma_{n,1}}\mathcal{C}_{n,\alpha}^{\mathrm{rad}}. R α ( f , 0 ) = ( 1 − α ) γ n , 1 γ n , α C n , α rad .
This completes the proof. □
For 0 < α < 1 0 < \alpha< 1 0 < α < 1 and β ∈ [ 0 , π / 2 ] \beta\in[0,\pi/2] β ∈ [ 0 , π /2 ] , define
A n , α ( β ) = ∫ 0 ∞ ∫ 0 π K n , α ( τ , θ ) ( sin θ ) n − 2 ∫ S n − 2 max { τ cos β , θ sin β ∣ η ⋅ e ∣ } d σ ( η ) d θ d τ , (48) A_{n,\alpha}(\beta)=\int_0^\infty\int_0^\pi\mathcal{K}_{n,\alpha}(\tau,\theta)(\sin\theta)^{n-2}\int_{\mathbb{S}^{n-2}}\max\{\tau\cos\beta,\theta\sin\beta|\eta\cdot e|\}\,d\sigma(\eta)\,d\theta\,d\tau, \tag*{(48)} A n , α ( β ) = ∫ 0 ∞ ∫ 0 π K n , α ( τ , θ ) ( sin θ ) n − 2 ∫ S n − 2 max { τ cos β , θ sin β ∣ η ⋅ e ∣ } d σ ( η ) d θ d τ , (48)
where e ∈ S n − 2 e\in\mathbb{S}^{n-2} e ∈ S n − 2 is any fixed unit vector. By rotational invariance, A n , α ( β ) A_{n,\alpha}(\beta) A n , α ( β ) does not depend on the choice of e e e .
Lemma 3.3. For every 0 < α < 1 0 < \alpha< 1 0 < α < 1 and every f ∈ C c ∞ ( R n ) ∖ { 0 } f \in C^\infty_c(\mathbb{R}^n)\setminus\{0\} f ∈ C c ∞ ( R n ) ∖ { 0 } ,
R α ( f , 0 ) ≤ ( 1 − α ) γ n , α γ n , 1 sup β ∈ [ 0 , π / 2 ] A n , α ( β ) . (49) R_{\alpha}(f,0)\leq(1-\alpha)\frac{\gamma_{n,\alpha}}{\gamma_{n,1}}\sup_{\beta\in[0,\pi/2]}A_{n,\alpha}(\beta). \tag*{(49)} R α ( f , 0 ) ≤ ( 1 − α ) γ n , 1 γ n , α β ∈ [ 0 , π /2 ] sup A n , α ( β ) . (49) Moreover,
A n , α ( 0 ) = C n , α r a d . (50) A_{n,\alpha}(0)=\mathcal{C}_{n,\alpha}^{\mathrm{rad}}. \tag*{(50)} A n , α ( 0 ) = C n , α rad . (50) Proof. Fix 0 < α < 1 0 < \alpha< 1 0 < α < 1 and f ∈ C c ∞ ( R n ) ∖ { 0 } f \in C^\infty_c(\mathbb{R}^n)\setminus\{0\} f ∈ C c ∞ ( R n ) ∖ { 0 } . Set
u ( t , ω ) = f ( e t ω ) , ( t , ω ) ∈ R × S n − 1 . u(t,\omega)=f(e^t\omega),\qquad(t,\omega)\in\mathbb{R}\times\mathbb{S}^{n-1}. u ( t , ω ) = f ( e t ω ) , ( t , ω ) ∈ R × S n − 1 . For each ω ∈ S n − 1 \omega\in\mathbb{S}^{n-1} ω ∈ S n − 1 , let
S ω n − 2 = { η ∈ S n − 1 : η ⋅ ω = 0 } , \mathbb{S}^{n-2}_{\omega}=\{\eta\in\mathbb{S}^{n-1}:\eta\cdot\omega=0\}, S ω n − 2 = { η ∈ S n − 1 : η ⋅ ω = 0 } , and denote by d σ ω d\sigma_{\omega} d σ ω the induced surface measure on S ω n − 2 \mathbb{S}^{n-2}_{\omega} S ω n − 2 . Using spherical coordinates centered at ω \omega ω , we write
ν = cos θ ω + sin θ η , 0 < θ < π , η ∈ S ω n − 2 . \nu=\cos\theta\,\omega+\sin\theta\,\eta,\qquad0<\theta<\pi,\quad\eta\in\mathbb{S}^{n-2}_{\omega}. ν = cos θ ω + sin θ η , 0 < θ < π , η ∈ S ω n − 2 . Then
d σ ( ν ) = ( sin θ ) n − 2 d σ ω ( η ) d θ . d\sigma(\nu)=(\sin\theta)^{n-2}\,d\sigma_{\omega}(\eta)\,d\theta. d σ ( ν ) = ( sin θ ) n − 2 d σ ω ( η ) d θ . By (41) , the evenness of K n , α \mathcal{K}_{n,\alpha} K n , α in its first variable, and the symmetry obtained by interchanging the two spherical variables and translating the t t t -variable, we have
I α ( D α f ) ( 0 ) γ n , α = ∫ 0 ∞ ∫ 0 π K n , α ( τ , θ ) ( sin θ ) n − 2 J u ( τ , θ ) d θ d τ , (51) \frac{I_{\alpha}(D^{\alpha}f)(0)}{\gamma_{n,\alpha}}=\int_0^\infty\int_0^\pi\mathcal{K}_{n,\alpha}(\tau,\theta)(\sin\theta)^{n-2}J_u(\tau,\theta)\,d\theta\,d\tau, \tag*{(51)} γ n , α I α ( D α f ) ( 0 ) = ∫ 0 ∞ ∫ 0 π K n , α ( τ , θ ) ( sin θ ) n − 2 J u ( τ , θ ) d θ d τ , (51)
where
J u ( τ , θ ) = ∫ R ∫ S n − 1 ∫ S ω n − 2 ∣ u ( t + τ , cos θ ω + sin θ η ) − u ( t , ω ) ∣ d σ ω ( η ) d σ ( ω ) d t . \mathcal{J}_u(\tau,\theta)=\int_{\mathbb{R}}\int_{\mathbb{S}^{n-1}}\int_{\mathbb{S}^{n-2}_\omega}|u(t+\tau,\cos\theta\,\omega+\sin\theta\,\eta)-u(t,\omega)|\,d\sigma_\omega(\eta)\,d\sigma(\omega)\,dt. J u ( τ , θ ) = ∫ R ∫ S n − 1 ∫ S ω n − 2 ∣ u ( t + τ , cos θ ω + sin θ η ) − u ( t , ω ) ∣ d σ ω ( η ) d σ ( ω ) d t .
Fix τ > 0 \tau>0 τ > 0 , 0 < θ < π 0<\theta<\pi 0 < θ < π , ω ∈ S n − 1 \omega\in\mathbb{S}^{n-1} ω ∈ S n − 1 , and η ∈ S ω n − 2 \eta\in\mathbb{S}^{n-2}_\omega η ∈ S ω n − 2 . For s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] , define
ω s = cos ( s θ ) ω + sin ( s θ ) η \omega_s=\cos(s\theta)\omega+\sin(s\theta)\eta ω s = cos ( s θ ) ω + sin ( s θ ) η
and
η s = − sin ( s θ ) ω + cos ( s θ ) η . \eta_s=-\sin(s\theta)\omega+\cos(s\theta)\eta. η s = − sin ( s θ ) ω + cos ( s θ ) η .
Then
η s ∈ S ω s n − 2 , d ω s d s = θ η s . \eta_s\in\mathbb{S}^{n-2}_{\omega_s},\qquad\frac{d\omega_s}{ds}=\theta\eta_s. η s ∈ S ω s n − 2 , d s d ω s = θ η s .
By the fundamental theorem of calculus,
u ( t + τ , ω 1 ) − u ( t , ω ) = ∫ 0 1 [ τ ∂ t u ( t + s τ , ω s ) + θ ∇ S n − 1 u ( t + s τ , ω s ) ⋅ η s ] d s . u(t+\tau,\omega_1)-u(t,\omega)=\int_0^1\left[\tau\,\partial_tu(t+s\tau,\omega_s)+\theta\,\nabla_{\mathbb{S}^{n-1}}u(t+s\tau,\omega_s)\cdot\eta_s\right]\,ds. u ( t + τ , ω 1 ) − u ( t , ω ) = ∫ 0 1 [ τ ∂ t u ( t + s τ , ω s ) + θ ∇ S n − 1 u ( t + s τ , ω s ) ⋅ η s ] d s .
Consequently,
∣ u ( t + τ , ω 1 ) − u ( t , ω ) ∣ ≤ ∫ 0 1 ∣ τ ∂ t u ( t + s τ , ω s ) + θ ∇ S n − 1 u ( t + s τ , ω s ) ⋅ η s ∣ d s . (52) |u(t+\tau,\omega_1)-u(t,\omega)|\leq\int_0^1\left|\tau\,\partial_tu(t+s\tau,\omega_s)+\theta\,\nabla_{\mathbb{S}^{n-1}}u(t+s\tau,\omega_s)\cdot\eta_s\right|\,ds. \tag*{(52)} ∣ u ( t + τ , ω 1 ) − u ( t , ω ) ∣ ≤ ∫ 0 1 ∣ τ ∂ t u ( t + s τ , ω s ) + θ ∇ S n − 1 u ( t + s τ , ω s ) ⋅ η s ∣ d s . (52)
For each s ∈ [ 0 , 1 ] s\in[0,1] s ∈ [ 0 , 1 ] , the transformation
( ω , η ) ⟼ ( ω s , η s ) (\omega,\eta)\longmapsto(\omega_s,\eta_s) ( ω , η ) ⟼ ( ω s , η s )
is the geodesic flow at time s θ s\theta s θ on the unit tangent bundle of S n − 1 \mathbb{S}^{n-1} S n − 1 . In particular, it preserves the natural measure
d σ ( ω ) d σ ω ( η ) . d\sigma(\omega)\,d\sigma_\omega(\eta). d σ ( ω ) d σ ω ( η ) .
Combining this measure invariance with translation invariance of Lebesgue measure in the t t t -variable and using (52) , we obtain
J u ( τ , θ ) ≤ ∫ R ∫ S n − 1 ∫ S ω n − 2 ∣ τ ∂ t u ( t , ω ) + θ ∇ S n − 1 u ( t , ω ) ⋅ η ∣ d σ ω ( η ) d σ ( ω ) d t . (53) \mathcal{J}_u(\tau,\theta)\leq\int_{\mathbb{R}}\int_{\mathbb{S}^{n-1}}\int_{\mathbb{S}^{n-2}_\omega}\left|\tau\,\partial_tu(t,\omega)+\theta\,\nabla_{\mathbb{S}^{n-1}}u(t,\omega)\cdot\eta\right|\,d\sigma_\omega(\eta)\,d\sigma(\omega)\,dt. \tag*{(53)} J u ( τ , θ ) ≤ ∫ R ∫ S n − 1 ∫ S ω n − 2 ∣ τ ∂ t u ( t , ω ) + θ ∇ S n − 1 u ( t , ω ) ⋅ η ∣ d σ ω ( η ) d σ ( ω ) d t . (53)
Next, using the symmetry η ↦ − η \eta\mapsto-\eta η ↦ − η and the elementary identity
∣ a + b ∣ + ∣ a − b ∣ 2 = max { ∣ a ∣ , ∣ b ∣ } , a , b ∈ R , \frac{|a+b|+|a-b|}{2}=\max\{|a|,|b|\},\qquad a,b\in\mathbb{R}, 2 ∣ a + b ∣ + ∣ a − b ∣ = max { ∣ a ∣ , ∣ b ∣ } , a , b ∈ R ,
we deduce that
∫ S ω n − 2 ∣ τ ∂ t u ( t , ω ) + θ ∇ S n − 1 u ( t , ω ) ⋅ η ∣ d σ ω ( η ) = ∫ S ω n − 2 max { τ ∣ ∂ t u ( t , ω ) ∣ , θ ∣ ∇ S n − 1 u ( t , ω ) ⋅ η ∣ } d σ ω ( η ) . (54) \begin{aligned}
\int_{\mathbb{S}^{n-2}_\omega}\left|\tau\,\partial_tu(t,\omega)+\theta\,\nabla_{\mathbb{S}^{n-1}}u(t,\omega)\cdot\eta\right|\,d\sigma_\omega(\eta)
&=\int_{\mathbb{S}^{n-2}_\omega}\max\{\tau|\partial_tu(t,\omega)|,\theta|\nabla_{\mathbb{S}^{n-1}}u(t,\omega)\cdot\eta|\}\,d\sigma_\omega(\eta). \tag*{(54)}
\end{aligned} ∫ S ω n − 2 ∣ τ ∂ t u ( t , ω ) + θ ∇ S n − 1 u ( t , ω ) ⋅ η ∣ d σ ω ( η ) = ∫ S ω n − 2 max { τ ∣ ∂ t u ( t , ω ) ∣ , θ ∣ ∇ S n − 1 u ( t , ω ) ⋅ η ∣ } d σ ω ( η ) . (54)
Set
G ( t , ω ) = ∣ ∇ c u ( t , ω ) ∣ . G(t,\omega)=|\nabla_cu(t,\omega)|. G ( t , ω ) = ∣ ∇ c u ( t , ω ) ∣.
At every point where G ( t , ω ) > 0 G(t,\omega)>0 G ( t , ω ) > 0 , choose β ( t , ω ) ∈ [ 0 , π / 2 ] \beta(t,\omega)\in[0,\pi/2] β ( t , ω ) ∈ [ 0 , π /2 ] such that
cos β ( t , ω ) = ∣ ∂ t u ( t , ω ) ∣ G ( t , ω ) \cos\beta(t,\omega)=\frac{|\partial_tu(t,\omega)|}{G(t,\omega)} cos β ( t , ω ) = G ( t , ω ) ∣ ∂ t u ( t , ω ) ∣
and
sin β ( t , ω ) = ∣ ∇ S n − 1 u ( t , ω ) ∣ G ( t , ω ) . \sin\beta(t,\omega)=\frac{\lvert\nabla_{\mathbb{S}^{n-1}}u(t,\omega)\rvert}{G(t,\omega)}. sin β ( t , ω ) = G ( t , ω ) ∣ ∇ S n − 1 u ( t , ω )∣ .
At points where G ( t , ω ) = 0 G(t,\omega)=0 G ( t , ω ) = 0 , set β ( t , ω ) = 0 \beta(t,\omega)=0 β ( t , ω ) = 0 . By rotational invariance of the surface measure on S ω n − 2 \mathbb{S}^{n-2}_{\omega} S ω n − 2 , (54) becomes
∫ S ω n − 2 ∣ τ ∂ t u ( t , ω ) + θ ∇ S n − 1 u ( t , ω ) ⋅ η ∣ d σ ω ( η ) = G ( t , ω ) ∫ S n − 2 max { τ cos β ( t , ω ) , θ sin β ( t , ω ) ∣ η ⋅ e ∣ } d σ ( η ) , (55) \begin{aligned}
\int_{\mathbb{S}^{n-2}_{\omega}}\lvert\tau\,\partial_tu(t,\omega)+\theta\nabla_{\mathbb{S}^{n-1}}u(t,\omega)\cdot\eta\rvert\,d\sigma_{\omega}(\eta)
&=G(t,\omega)\int_{\mathbb{S}^{n-2}}\max\{\tau\cos\beta(t,\omega),\theta\sin\beta(t,\omega)\lvert\eta\cdot e\rvert\}\,d\sigma(\eta), \tag*{(55)}
\end{aligned} ∫ S ω n − 2 ∣ τ ∂ t u ( t , ω ) + θ ∇ S n − 1 u ( t , ω ) ⋅ η ∣ d σ ω ( η ) = G ( t , ω ) ∫ S n − 2 max { τ cos β ( t , ω ) , θ sin β ( t , ω ) ∣ η ⋅ e ∣} d σ ( η ) , (55)
where e ∈ S n − 2 e\in\mathbb{S}^{n-2} e ∈ S n − 2 is any fixed unit vector.
Combining (51) , (53) , and (55) , and applying Tonelli’s theorem, we obtain
I α ( D α f ) ( 0 ) γ n , α ≤ ∫ R ∫ S n − 1 G ( t , ω ) A n , α ( β ( t , ω ) ) d σ ( ω ) d t ≤ ( sup β ∈ [ 0 , π / 2 ] A n , α ( β ) ) ∫ R ∫ S n − 1 ∣ ∇ c u ( t , ω ) ∣ d σ ( ω ) d t . \begin{aligned}
\frac{I_{\alpha}(D^{\alpha}f)(0)}{\gamma_{n,\alpha}}
\leq\int_{\mathbb{R}}\int_{\mathbb{S}^{n-1}}G(t,\omega)A_{n,\alpha}(\beta(t,\omega))\,d\sigma(\omega)\,dt \\
&\leq\left(\sup_{\beta\in[0,\pi/2]}A_{n,\alpha}(\beta)\right)\int_{\mathbb{R}}\int_{\mathbb{S}^{n-1}}\lvert\nabla_cu(t,\omega)\rvert\,d\sigma(\omega)\,dt.
\end{aligned} γ n , α I α ( D α f ) ( 0 ) ≤ ∫ R ∫ S n − 1 G ( t , ω ) A n , α ( β ( t , ω )) d σ ( ω ) d t ≤ ( β ∈ [ 0 , π /2 ] sup A n , α ( β ) ) ∫ R ∫ S n − 1 ∣ ∇ c u ( t , ω )∣ d σ ( ω ) d t .
By (40) ,
∫ R ∫ S n − 1 ∣ ∇ c u ( t , ω ) ∣ d σ ( ω ) d t = I 1 ( ∣ ∇ f ∣ ) ( 0 ) γ n , 1 . \int_{\mathbb{R}}\int_{\mathbb{S}^{n-1}}\lvert\nabla_cu(t,\omega)\rvert\,d\sigma(\omega)\,dt=\frac{I_1(\lvert\nabla f\rvert)(0)}{\gamma_{n,1}}. ∫ R ∫ S n − 1 ∣ ∇ c u ( t , ω )∣ d σ ( ω ) d t = γ n , 1 I 1 (∣ ∇ f ∣) ( 0 ) .
Hence,
I α ( D α f ) ( 0 ) ≤ γ n , α γ n , 1 ( sup β ∈ [ 0 , π / 2 ] A n , α ( β ) ) I 1 ( ∣ ∇ f ∣ ) ( 0 ) . I_{\alpha}(D^{\alpha}f)(0)\leq\frac{\gamma_{n,\alpha}}{\gamma_{n,1}}\left(\sup_{\beta\in[0,\pi/2]}A_{n,\alpha}(\beta)\right)I_1(\lvert\nabla f\rvert)(0). I α ( D α f ) ( 0 ) ≤ γ n , 1 γ n , α ( β ∈ [ 0 , π /2 ] sup A n , α ( β ) ) I 1 (∣ ∇ f ∣) ( 0 ) .
Since f ≢ 0 f\not\equiv0 f ≡ 0 , we have I 1 ( ∣ ∇ f ∣ ) ( 0 ) > 0 I_1(\lvert\nabla f\rvert)(0)>0 I 1 (∣ ∇ f ∣) ( 0 ) > 0 . Multiplying by 1 − α 1-\alpha 1 − α and dividing by I 1 ( ∣ ∇ f ∣ ) ( 0 ) I_1(\lvert\nabla f\rvert)(0) I 1 (∣ ∇ f ∣) ( 0 ) , we obtain (49) .
Finally, taking β = 0 \beta=0 β = 0 in (48) , we find
A n , α ( 0 ) = ∫ 0 ∞ ∫ 0 π K n , α ( τ , θ ) ( sin θ ) n − 2 ∫ S n − 2 τ d σ ( η ) d θ d τ = ∣ S n − 2 ∣ ∫ 0 ∞ ∫ 0 π τ K n , α ( τ , θ ) ( sin θ ) n − 2 d θ d τ = C n , α r a d , \begin{aligned}
A_{n,\alpha}(0)
&=\int_0^\infty\int_0^\pi\mathcal{K}_{n,\alpha}(\tau,\theta)(\sin\theta)^{n-2}\int_{\mathbb{S}^{n-2}}\tau\,d\sigma(\eta)\,d\theta\,d\tau\\
&=\lvert\mathbb{S}^{n-2}\rvert\int_0^\infty\int_0^\pi\tau\mathcal{K}_{n,\alpha}(\tau,\theta)(\sin\theta)^{n-2}\,d\theta\,d\tau\\
&=\mathcal{C}^{\mathrm{rad}}_{n,\alpha},
\end{aligned} A n , α ( 0 ) = ∫ 0 ∞ ∫ 0 π K n , α ( τ , θ ) ( sin θ ) n − 2 ∫ S n − 2 τ d σ ( η ) d θ d τ = ∣ S n − 2 ∣ ∫ 0 ∞ ∫ 0 π τ K n , α ( τ , θ ) ( sin θ ) n − 2 d θ d τ = C n , α rad ,
where the last equality follows from (9) . This proves (50) and completes the proof. □
Lemma 3.4. For every 0 < α < 1 0<\alpha<1 0 < α < 1 and every β ∈ [ 0 , π / 2 ] \beta\in[0,\pi/2] β ∈ [ 0 , π /2 ] ,
A n , α ( β ) ≤ A n , α ( 0 ) . (56) A_{n,\alpha}(\beta)\leq A_{n,\alpha}(0). \tag*{(56)} A n , α ( β ) ≤ A n , α ( 0 ) . (56) Consequently,
sup β ∈ [ 0 , π / 2 ] A n , α ( β ) = C n , α r a d . \sup_{\beta\in[0,\pi/2]}A_{n,\alpha}(\beta)=\mathcal{C}^{\mathrm{rad}}_{n,\alpha}. β ∈ [ 0 , π /2 ] sup A n , α ( β ) = C n , α rad .
Proof. Fix 0 < α < 1 0<\alpha<1 0 < α < 1 and 0 ≤ β ≤ π / 2 0\leq\beta\leq\pi/2 0 ≤ β ≤ π /2 . Define
ρ ( τ , θ ) = K n , α ( τ , θ ) ( sin θ θ ) n − 2 , τ ∈ R , 0 < θ < π . (57) \rho(\tau,\theta)=\mathcal{K}_{n,\alpha}(\tau,\theta)\left(\frac{\sin\theta}{\theta}\right)^{n-2},\qquad\tau\in\mathbb{R},\quad0<\theta<\pi. \tag*{(57)} ρ ( τ , θ ) = K n , α ( τ , θ ) ( θ sin θ ) n − 2 , τ ∈ R , 0 < θ < π . (57)
We extend ρ \rho ρ by zero for θ ≥ π \theta\ge\pi θ ≥ π .
By rotational invariance, we may assume that
e = ( 1 , 0 , … , 0 ) ∈ R n − 1 . e=(1,0,\ldots,0)\in\mathbb{R}^{n-1}. e = ( 1 , 0 , … , 0 ) ∈ R n − 1 .
For 0 < θ < π 0<\theta<\pi 0 < θ < π and η ∈ S n − 2 \eta\in S^{n-2} η ∈ S n − 2 , set
y = θ η ∈ R n − 1 . y=\theta\eta\in\mathbb{R}^{n-1}. y = θ η ∈ R n − 1 .
Then
∣ y ∣ = θ , y 1 = θ ( η ⋅ e ) , d y = θ n − 2 d θ d σ ( η ) . |y|=\theta,\qquad y_1=\theta(\eta\cdot e),\qquad dy=\theta^{n-2}\,d\theta\,d\sigma(\eta). ∣ y ∣ = θ , y 1 = θ ( η ⋅ e ) , d y = θ n − 2 d θ d σ ( η ) .
Consequently, by (48) and the definition of ρ \rho ρ , we obtain
A n , α ( β ) = ∫ 0 ∞ ∫ ∣ y ∣ < π max { τ cos β , ∣ y 1 ∣ sin β } ρ ( τ , ∣ y ∣ ) d y d τ . A_{n,\alpha}(\beta)=\int_0^\infty\int_{|y|<\pi}\max\{\tau\cos\beta,|y_1|\sin\beta\}\rho(\tau,|y|)\,dy\,d\tau. A n , α ( β ) = ∫ 0 ∞ ∫ ∣ y ∣ < π max { τ cos β , ∣ y 1 ∣ sin β } ρ ( τ , ∣ y ∣ ) d y d τ .
Since ρ ( τ , θ ) \rho(\tau,\theta) ρ ( τ , θ ) has been extended by zero for θ ≥ π \theta\ge\pi θ ≥ π , we may equivalently write
A n , α ( β ) = ∫ 0 ∞ ∫ R n − 1 max { τ cos β , ∣ y 1 ∣ sin β } ρ ( τ , ∣ y ∣ ) d y d τ . A_{n,\alpha}(\beta)=\int_0^\infty\int_{\mathbb{R}^{n-1}}\max\{\tau\cos\beta,|y_1|\sin\beta\}\rho(\tau,|y|)\,dy\,d\tau. A n , α ( β ) = ∫ 0 ∞ ∫ R n − 1 max { τ cos β , ∣ y 1 ∣ sin β } ρ ( τ , ∣ y ∣ ) d y d τ .
Using the elementary identity
max { a , ∣ b ∣ } = ∣ a + b ∣ + ∣ a − b ∣ 2 , a ≥ 0 , b ∈ R , \max\{a,|b|\}=\frac{|a+b|+|a-b|}{2},\qquad a\ge0,\quad b\in\mathbb{R}, max { a , ∣ b ∣ } = 2 ∣ a + b ∣ + ∣ a − b ∣ , a ≥ 0 , b ∈ R ,
and the evenness of ρ \rho ρ in τ \tau τ , we obtain
A n , α ( β ) = 1 2 ∫ 0 ∞ ∫ R n − 1 ( ∣ τ cos β + y 1 sin β ∣ + ∣ τ cos β − y 1 sin β ∣ ) ρ ( τ , ∣ y ∣ ) d y d τ = 1 2 ∫ R ∫ R n − 1 ∣ τ cos β + y 1 sin β ∣ ρ ( τ , ∣ y ∣ ) d y d τ . (58) \begin{aligned}
A_{n,\alpha}(\beta)
&=\frac{1}{2}\int_0^\infty\int_{\mathbb{R}^{n-1}}\bigl(|\tau\cos\beta+y_1\sin\beta|+|\tau\cos\beta-y_1\sin\beta|\bigr)\rho(\tau,|y|)\,dy\,d\tau\\
&=\frac{1}{2}\int_{\mathbb{R}}\int_{\mathbb{R}^{n-1}}|\tau\cos\beta+y_1\sin\beta|\rho(\tau,|y|)\,dy\,d\tau. \tag*{(58)}
\end{aligned} A n , α ( β ) = 2 1 ∫ 0 ∞ ∫ R n − 1 ( ∣ τ cos β + y 1 sin β ∣ + ∣ τ cos β − y 1 sin β ∣ ) ρ ( τ , ∣ y ∣ ) d y d τ = 2 1 ∫ R ∫ R n − 1 ∣ τ cos β + y 1 sin β ∣ ρ ( τ , ∣ y ∣ ) d y d τ . (58)
Set
z = ( τ , y ) ∈ R × R n − 1 , v β = ( cos β , sin β , 0 , … , 0 ) ∈ S n − 1 . z=(\tau,y)\in\mathbb{R}\times\mathbb{R}^{n-1},\qquad v_\beta=(\cos\beta,\sin\beta,0,\ldots,0)\in S^{n-1}. z = ( τ , y ) ∈ R × R n − 1 , v β = ( cos β , sin β , 0 , … , 0 ) ∈ S n − 1 .
For every r > 0 r>0 r > 0 , define
F r ( s ) = ρ ( r s , r 1 − s 2 ) , − 1 < s < 1 , F_r(s)=\rho\left(rs,r\sqrt{1-s^2}\right),\qquad-1<s<1, F r ( s ) = ρ ( r s , r 1 − s 2 ) , − 1 < s < 1 ,
with endpoint values understood by continuity. Since ρ \rho ρ is even in its first variable, the function F r F_r F r is even.
Using polar coordinates z = r ξ z=r\xi z = r ξ , where r > 0 r>0 r > 0 and ξ ∈ S n − 1 \xi\in S^{n-1} ξ ∈ S n − 1 , we have
τ = r ξ 1 , ∣ y ∣ = r 1 − ξ 1 2 , d z = r n − 1 d r d σ ( ξ ) . \tau=r\xi_1,\qquad|y|=r\sqrt{1-\xi_1^2},\qquad dz=r^{n-1}\,dr\,d\sigma(\xi). τ = r ξ 1 , ∣ y ∣ = r 1 − ξ 1 2 , d z = r n − 1 d r d σ ( ξ ) .
Consequently, by (3.19), we obtain
A n , α ( β ) = 1 2 ∫ 0 ∞ r n ∫ S n − 1 ∣ ξ ⋅ v β ∣ F r ( ∣ ξ 1 ∣ ) d σ ( ξ ) d r . (59) A_{n,\alpha}(\beta)=\frac{1}{2}\int_0^\infty r^n\int_{S^{n-1}}|\xi\cdot v_\beta|F_r(|\xi_1|)\,d\sigma(\xi)\,dr. \tag*{(59)} A n , α ( β ) = 2 1 ∫ 0 ∞ r n ∫ S n − 1 ∣ ξ ⋅ v β ∣ F r ( ∣ ξ 1 ∣ ) d σ ( ξ ) d r . (59)
We now show that, for every r > 0 r>0 r > 0 , the function F r F_r F r is nondecreasing on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Fix r > 0 r>0 r > 0 . For 0 < s < 1 0<s<1 0 < s < 1 , set
θ = r 1 − s 2 . \theta=r\sqrt{1-s^2}. θ = r 1 − s 2 .
Then
r s = r 2 − θ 2 . (60) rs=\sqrt{r^2-\theta^2}. \tag*{(60)} r s = r 2 − θ 2 . (60)
Since
d θ d s = − r s 1 − s 2 < 0 \frac{d\theta}{ds}=-\frac{rs}{\sqrt{1-s^2}}<0 d s d θ = − 1 − s 2 r s < 0
and ρ \rho ρ has a continuous zero extension for θ ≥ π \theta\ge\pi θ ≥ π , it suffices to prove that the function
θ ⟼ ρ ( r 2 − θ 2 , θ ) \theta\longmapsto\rho\left(\sqrt{r^2-\theta^2},\theta\right) θ ⟼ ρ ( r 2 − θ 2 , θ )
is decreasing on 0 < θ < min { r , π } 0<\theta<\min\{r,\pi\} 0 < θ < min { r , π } .
Set
a = n − α 2 , p = a + α . a=\frac{n-\alpha}{2},\qquad p=a+\alpha. a = 2 n − α , p = a + α .
For 0 < θ < min { r , π } 0<\theta<\min\{r,\pi\} 0 < θ < min { r , π } , let
τ = r 2 − θ 2 . \tau=\sqrt{r^2-\theta^2}. τ = r 2 − θ 2 .
Recalling the definitions of ρ \rho ρ and K n , α \mathcal{K}_{n,\alpha} K n , α , and observing that
d τ d θ = − θ τ , \frac{d\tau}{d\theta}=-\frac{\theta}{\tau}, d θ d τ = − τ θ ,
direct differentiation gives
d d θ log ρ ( r 2 − θ 2 , θ ) = − a θ τ tanh ( a τ ) + p θ sinh τ / τ − sin θ cosh τ − cos θ + ( n − 2 ) ( cot θ − 1 θ ) . \frac{d}{d\theta}\log\rho\left(\sqrt{r^2-\theta^2},\theta\right)
=-\frac{a\theta}{\tau}\tanh(a\tau)+p\frac{\theta\sinh\tau/\tau-\sin\theta}{\cosh\tau-\cos\theta}+(n-2)\left(\cot\theta-\frac{1}{\theta}\right). d θ d log ρ ( r 2 − θ 2 , θ ) = − τ a θ tanh ( a τ ) + p cosh τ − cos θ θ sinh τ / τ − sin θ + ( n − 2 ) ( cot θ − θ 1 ) .
Using the elementary inequalities
θ sinh τ / τ − sin θ cosh τ − cos θ ≤ θ τ tanh τ 2 , 1 θ − cot θ ≥ θ 3 , \frac{\theta\sinh\tau/\tau-\sin\theta}{\cosh\tau-\cos\theta}\le\frac{\theta}{\tau}\tanh\frac{\tau}{2},\qquad
\frac{1}{\theta}-\cot\theta\ge\frac{\theta}{3}, cosh τ − cos θ θ sinh τ / τ − sin θ ≤ τ θ tanh 2 τ , θ 1 − cot θ ≥ 3 θ ,
we obtain
d d θ log ρ ( r 2 − θ 2 , θ ) ≤ θ τ D n , α ( τ ) (61) \frac{d}{d\theta}\log\rho\left(\sqrt{r^2-\theta^2},\theta\right)\le\frac{\theta}{\tau}D_{n,\alpha}(\tau) \tag*{(61)} d θ d log ρ ( r 2 − θ 2 , θ ) ≤ τ θ D n , α ( τ ) (61)
where
D n , α ( τ ) = p tanh τ 2 − a tanh ( a τ ) − n − 2 3 τ . D_{n,\alpha}(\tau)=p\tanh\frac{\tau}{2}-a\tanh(a\tau)-\frac{n-2}{3}\tau. D n , α ( τ ) = p tanh 2 τ − a tanh ( a τ ) − 3 n − 2 τ .
Using the elementary inequalities
2 tanh τ 2 − tanh τ < τ 3 , tanh τ 2 − tanh τ < 0 , τ > 0 , 2\tanh\frac{\tau}{2}-\tanh\tau<\frac{\tau}{3},\qquad
\tanh\frac{\tau}{2}-\tanh\tau<0,\qquad
\tau>0, 2 tanh 2 τ − tanh τ < 3 τ , tanh 2 τ − tanh τ < 0 , τ > 0 ,
the monotonicity of tanh \tanh tanh , a > 1 a>1 a > 1 , and p = a + α p=a+\alpha p = a + α , we obtain for every τ > 0 \tau>0 τ > 0 ,
D n , α ( τ ) = a [ tanh τ 2 − tanh ( a τ ) ] + α tanh τ 2 − n − 2 3 τ ≤ a [ tanh τ 2 − tanh τ ] + tanh τ 2 − τ 3 ≤ 2 tanh τ 2 − tanh τ − τ 3 < 0. \begin{aligned}
D_{n,\alpha}(\tau)
&=a\left[\tanh\frac{\tau}{2}-\tanh(a\tau)\right]+\alpha\tanh\frac{\tau}{2}-\frac{n-2}{3}\tau\\
&\le a\left[\tanh\frac{\tau}{2}-\tanh\tau\right]+\tanh\frac{\tau}{2}-\frac{\tau}{3}\\
&\le2\tanh\frac{\tau}{2}-\tanh\tau-\frac{\tau}{3}\\
&<0.
\end{aligned} D n , α ( τ ) = a [ tanh 2 τ − tanh ( a τ ) ] + α tanh 2 τ − 3 n − 2 τ ≤ a [ tanh 2 τ − tanh τ ] + tanh 2 τ − 3 τ ≤ 2 tanh 2 τ − tanh τ − 3 τ < 0.
Since ρ ( τ , θ ) > 0 \rho(\tau,\theta)>0 ρ ( τ , θ ) > 0 for τ > 0 \tau>0 τ > 0 and 0 < θ < π 0<\theta<\pi 0 < θ < π , it follows from (61) that
d d θ ρ ( r 2 − θ 2 , θ ) < 0 , 0 < θ < min { r , π } . \frac{d}{d\theta}\rho\left(\sqrt{r^2-\theta^2},\theta\right)<0,\qquad0<\theta<\min\{r,\pi\}. d θ d ρ ( r 2 − θ 2 , θ ) < 0 , 0 < θ < min { r , π } .
Consequently, F r F_r F r is nondecreasing on [ 0 , 1 ] [0,1] [ 0 , 1 ] .
We next prove that, for every r > 0 r>0 r > 0 ,
∫ S n − 1 ∣ ξ ⋅ v β ∣ F r ( ∣ ξ 1 ∣ ) d σ ( ξ ) ≤ ∫ S n − 1 ∣ ξ 1 ∣ F r ( ∣ ξ 1 ∣ ) d σ ( ξ ) . (62) \int_{\mathbb{S}^{n-1}} \lvert\xi\cdot v_{\beta}\rvert F_r(\lvert\xi_1\rvert)\,\mathrm{d}\sigma(\xi)\leq\int_{\mathbb{S}^{n-1}}\lvert\xi_1\rvert F_r(\lvert\xi_1\rvert)\,\mathrm{d}\sigma(\xi). \tag*{(62)} ∫ S n − 1 ∣ ξ ⋅ v β ∣ F r (∣ ξ 1 ∣) d σ ( ξ ) ≤ ∫ S n − 1 ∣ ξ 1 ∣ F r (∣ ξ 1 ∣) d σ ( ξ ) . (62)
To this end, for almost every ξ ∈ S n − 1 \xi\in\mathbb{S}^{n-1} ξ ∈ S n − 1 , write
ξ = ( q cos φ , q sin φ , ζ ) , \xi=(q\cos\varphi,q\sin\varphi,\zeta), ξ = ( q cos φ , q sin φ , ζ ) ,
where
ζ ∈ B n − 2 , q = 1 − ∣ ζ ∣ 2 , 0 ≤ φ < 2 π , \zeta\in B^{n-2},\qquad q=\sqrt{1-\lvert\zeta\rvert^2},\qquad0\leq\varphi<2\pi, ζ ∈ B n − 2 , q = 1 − ∣ ζ ∣ 2 , 0 ≤ φ < 2 π ,
and B n − 2 B^{n-2} B n − 2 denotes the unit ball in R n − 2 \mathbb{R}^{n-2} R n − 2 . In these coordinates, the surface measure satisfies
d σ ( ξ ) = d φ d ζ . \mathrm{d}\sigma(\xi)=\mathrm{d}\varphi\,\mathrm{d}\zeta. d σ ( ξ ) = d φ d ζ .
Moreover,
∣ ξ ⋅ v β ∣ = q ∣ cos ( φ − β ) ∣ , ∣ ξ 1 ∣ = q ∣ cos φ ∣ . \lvert\xi\cdot v_{\beta}\rvert=q\lvert\cos(\varphi-\beta)\rvert,\qquad\lvert\xi_1\rvert=q\lvert\cos\varphi\rvert. ∣ ξ ⋅ v β ∣ = q ∣ cos ( φ − β )∣ , ∣ ξ 1 ∣ = q ∣ cos φ ∣ .
Consequently,
∫ S n − 1 ∣ ξ ⋅ v β ∣ F r ( ∣ ξ 1 ∣ ) d σ ( ξ ) = ∫ B n − 2 ∫ 0 2 π q ∣ cos ( φ − β ) ∣ F r ( q ∣ cos φ ∣ ) d φ d ζ . (63) \int_{\mathbb{S}^{n-1}}\lvert\xi\cdot v_{\beta}\rvert F_r(\lvert\xi_1\rvert)\,\mathrm{d}\sigma(\xi)
=\int_{B^{n-2}}\int_0^{2\pi}q\lvert\cos(\varphi-\beta)\rvert F_r(q\lvert\cos\varphi\rvert)\,\mathrm{d}\varphi\,\mathrm{d}\zeta. \tag*{(63)} ∫ S n − 1 ∣ ξ ⋅ v β ∣ F r (∣ ξ 1 ∣) d σ ( ξ ) = ∫ B n − 2 ∫ 0 2 π q ∣ cos ( φ − β )∣ F r ( q ∣ cos φ ∣) d φ d ζ . (63)
For each fixed q ∈ [ 0 , 1 ] q\in[0,1] q ∈ [ 0 , 1 ] , the functions
φ ⟼ F r ( q ∣ cos φ ∣ ) \varphi\longmapsto F_r(q\lvert\cos\varphi\rvert) φ ⟼ F r ( q ∣ cos φ ∣)
and
φ ⟼ ∣ cos φ ∣ \varphi\longmapsto\lvert\cos\varphi\rvert φ ⟼ ∣ cos φ ∣
are nonnegative, even, and π \pi π -periodic on R \mathbb{R} R . Moreover, since F r F_r F r is nondecreasing on [ 0 , 1 ] [0,1] [ 0 , 1 ] , both functions are nonincreasing on [ 0 , π / 2 ] [0,\pi/2] [ 0 , π /2 ] . Hence, by the circular rearrangement inequality (see [3 ] ), we obtain
∫ 0 2 π F r ( q ∣ cos φ ∣ ) ∣ cos ( φ − β ) ∣ d φ ≤ ∫ 0 2 π F r ( q ∣ cos φ ∣ ) ∣ cos φ ∣ d φ . \int_0^{2\pi}F_r(q\lvert\cos\varphi\rvert)\lvert\cos(\varphi-\beta)\rvert\,\mathrm{d}\varphi
\leq
\int_0^{2\pi}F_r(q\lvert\cos\varphi\rvert)\lvert\cos\varphi\rvert\,\mathrm{d}\varphi. ∫ 0 2 π F r ( q ∣ cos φ ∣) ∣ cos ( φ − β )∣ d φ ≤ ∫ 0 2 π F r ( q ∣ cos φ ∣) ∣ cos φ ∣ d φ .
Together with (63) , this gives
∫ S n − 1 ∣ ξ ⋅ v β ∣ F r ( ∣ ξ 1 ∣ ) d σ ( ξ ) ≤ ∫ B n − 2 ∫ 0 2 π q ∣ cos φ ∣ F r ( q ∣ cos φ ∣ ) d φ d ζ = ∫ S n − 1 ∣ ξ 1 ∣ F r ( ∣ ξ 1 ∣ ) d σ ( ξ ) , \begin{aligned}
\int_{\mathbb{S}^{n-1}}\lvert\xi\cdot v_{\beta}\rvert F_r(\lvert\xi_1\rvert)\,\mathrm{d}\sigma(\xi)
\leq
\int_{B^{n-2}}\int_0^{2\pi}q\lvert\cos\varphi\rvert F_r(q\lvert\cos\varphi\rvert)\,\mathrm{d}\varphi\,\mathrm{d}\zeta\\
&=\int_{\mathbb{S}^{n-1}}\lvert\xi_1\rvert F_r(\lvert\xi_1\rvert)\,\mathrm{d}\sigma(\xi),
\end{aligned} ∫ S n − 1 ∣ ξ ⋅ v β ∣ F r (∣ ξ 1 ∣) d σ ( ξ ) ≤ ∫ B n − 2 ∫ 0 2 π q ∣ cos φ ∣ F r ( q ∣ cos φ ∣) d φ d ζ = ∫ S n − 1 ∣ ξ 1 ∣ F r (∣ ξ 1 ∣) d σ ( ξ ) ,
which proves (62) .
Finally, combining (59) and (62) , we obtain
A n , α ( β ) ≤ 1 2 ∫ 0 ∞ r n ∫ S n − 1 ∣ ξ 1 ∣ F r ( ∣ ξ 1 ∣ ) d σ ( ξ ) d r = A n , α ( 0 ) . \begin{aligned}
A_{n,\alpha}(\beta)\leq\frac{1}{2}\int_0^\infty r^n\int_{\mathbb{S}^{n-1}}\lvert\xi_1\rvert F_r(\lvert\xi_1\rvert)\,\mathrm{d}\sigma(\xi)\,\mathrm{d}r \\
&=A_{n,\alpha}(0).
\end{aligned} A n , α ( β ) ≤ 2 1 ∫ 0 ∞ r n ∫ S n − 1 ∣ ξ 1 ∣ F r (∣ ξ 1 ∣) d σ ( ξ ) d r = A n , α ( 0 ) .
This proves (56) . The final assertion follows from (50) . □
We are now ready to prove Theorem 1.3 .
Proof of Theorem 1.3
Proof. By Lemmas 3.3 and 3.4 , for every 0 < α < 1 0 < \alpha< 1 0 < α < 1 and every f ∈ C c ∞ ( R n ) ∖ { 0 } f \in C^\infty_c(\mathbb{R}^n) \setminus\{0\} f ∈ C c ∞ ( R n ) ∖ { 0 } , we have
R α ( f , 0 ) ≤ ( 1 − α ) γ n , α γ n , 1 sup β ∈ [ 0 , π / 2 ] A n , α ( β ) = ( 1 − α ) γ n , α γ n , 1 C n , α r a d . R_\alpha(f,0) \le(1-\alpha)\frac{\gamma_{n,\alpha}}{\gamma_{n,1}}\sup_{\beta\in[0,\pi/2]} A_{n,\alpha}(\beta)
= (1-\alpha)\frac{\gamma_{n,\alpha}}{\gamma_{n,1}}\mathcal{C}^{\mathrm{rad}}_{n,\alpha}. R α ( f , 0 ) ≤ ( 1 − α ) γ n , 1 γ n , α β ∈ [ 0 , π /2 ] sup A n , α ( β ) = ( 1 − α ) γ n , 1 γ n , α C n , α rad .
By translation invariance, taking the supremum over all admissible functions and evaluation points yields
C n , α ∗ ≤ ( 1 − α ) γ n , α γ n , 1 C n , α r a d . C^*_{n,\alpha} \le(1-\alpha)\frac{\gamma_{n,\alpha}}{\gamma_{n,1}}\mathcal{C}^{\mathrm{rad}}_{n,\alpha}. C n , α ∗ ≤ ( 1 − α ) γ n , 1 γ n , α C n , α rad .
On the other hand, Lemma 3.2 shows that equality holds for every nontrivial smooth radially monotone function centered at the evaluation point. Consequently,
C n , α ∗ = ( 1 − α ) γ n , α γ n , 1 C n , α r a d . C^*_{n,\alpha} = (1-\alpha)\frac{\gamma_{n,\alpha}}{\gamma_{n,1}}\mathcal{C}^{\mathrm{rad}}_{n,\alpha}. C n , α ∗ = ( 1 − α ) γ n , 1 γ n , α C n , α rad .
This proves (10) and the assertion concerning the attainment of the optimal constant. □
Proof of Corollary 1.4
Proof. Fix 0 < α < 1 0 < \alpha< 1 0 < α < 1 . Let
f ( y ) = F ( ∣ y ∣ ) ∈ C c ∞ ( R n ) ∖ { 0 } , f(y) = F(|y|) \in C^\infty_c(\mathbb{R}^n) \setminus\{0\}, f ( y ) = F ( ∣ y ∣ ) ∈ C c ∞ ( R n ) ∖ { 0 } ,
where F F F is monotone. By Theorem 1.3 ,
R α ( f , 0 ) = C n , α ∗ . (64) R_\alpha(f,0) = C^*_{n,\alpha}. \tag*{(64)} R α ( f , 0 ) = C n , α ∗ . (64)
Choose R > 0 R > 0 R > 0 such that supp f ⊂ B ( 0 , R ) \operatorname{supp} f \subset B(0,R) supp f ⊂ B ( 0 , R ) , where B ( 0 , R ) B(0,R) B ( 0 , R ) denotes the open ball of radius R R R centered at the origin in R n \mathbb{R}^n R n . Since F F F is monotone and nontrivial, we have f ( 0 ) ≠ 0 f(0) \ne0 f ( 0 ) = 0 . For every y ∈ R n y \in\mathbb{R}^n y ∈ R n with ∣ y ∣ > R |y| > R ∣ y ∣ > R , taking h = − 2 y h = -2y h = − 2 y gives
f ( y + h ) = f ( y ) = 0 , f(y+h) = f(y) = 0, f ( y + h ) = f ( y ) = 0 ,
whereas
∣ h ∣ ∫ 0 1 ∣ ∇ f ( y + t h ) ∣ d t > 0 , |h|\int_0^1 |\nabla f(y+th)|\,dt > 0, ∣ h ∣ ∫ 0 1 ∣∇ f ( y + t h ) ∣ d t > 0 ,
since the segment joining y y y to − y -y − y passes through the origin, where f ( 0 ) ≠ 0 f(0) \ne0 f ( 0 ) = 0 . Hence, by continuity, there exists a neighborhood V − 2 y V_{-2y} V − 2 y of − 2 y -2y − 2 y such that
∣ f ( y + h ) − f ( y ) ∣ < ∣ h ∣ ∫ 0 1 ∣ ∇ f ( y + t h ) ∣ d t , h ∈ V − 2 y . |f(y+h)-f(y)| < |h|\int_0^1 |\nabla f(y+th)|\,dt,\qquad h \in V_{-2y}. ∣ f ( y + h ) − f ( y ) ∣ < ∣ h ∣ ∫ 0 1 ∣∇ f ( y + t h ) ∣ d t , h ∈ V − 2 y .
On the other hand, the fundamental theorem of calculus gives
∣ f ( y + h ) − f ( y ) ∣ ≤ ∣ h ∣ ∫ 0 1 ∣ ∇ f ( y + t h ) ∣ d t |f(y+h)-f(y)| \le|h|\int_0^1 |\nabla f(y+th)|\,dt ∣ f ( y + h ) − f ( y ) ∣ ≤ ∣ h ∣ ∫ 0 1 ∣∇ f ( y + t h ) ∣ d t
for every h ∈ R n h \in\mathbb{R}^n h ∈ R n . Consequently, for ∣ y ∣ > R |y| > R ∣ y ∣ > R , splitting the integral defining D α f ( y ) D^\alpha f(y) D α f ( y ) over V − 2 y V_{-2y} V − 2 y and its complement, we obtain
D α f ( y ) = ∫ V − 2 y ∣ f ( y + h ) − f ( y ) ∣ ∣ h ∣ n + α d h + ∫ R n ∖ V − 2 y ∣ f ( y + h ) − f ( y ) ∣ ∣ h ∣ n + α d h < ∫ R n 1 ∣ h ∣ n + α − 1 ∫ 0 1 ∣ ∇ f ( y + t h ) ∣ d t d h = ∫ 0 1 t α − 1 d t ∫ R n ∣ ∇ f ( z ) ∣ ∣ y − z ∣ n + α − 1 d z = 1 α γ n , 1 − α I 1 − α ( ∣ ∇ f ∣ ) ( y ) , \begin{aligned}
D^\alpha f(y) &= \int_{V_{-2y}} \frac{|f(y+h)-f(y)|}{|h|^{n+\alpha}}\,dh+\int_{\mathbb{R}^n\setminus V_{-2y}} \frac{|f(y+h)-f(y)|}{|h|^{n+\alpha}}\,dh\\
&< \int_{\mathbb{R}^n}\frac{1}{|h|^{n+\alpha-1}}\int_0^1|\nabla f(y+th)|\,dt\,dh\\
&=\int_0^1 t^{\alpha-1}\,dt\int_{\mathbb{R}^n}\frac{|\nabla f(z)|}{|y-z|^{n+\alpha-1}}\,dz\\
&=\frac{1}{\alpha\gamma_{n,1-\alpha}}I_{1-\alpha}(|\nabla f|)(y),
\end{aligned} D α f ( y ) = ∫ V − 2 y ∣ h ∣ n + α ∣ f ( y + h ) − f ( y ) ∣ d h + ∫ R n ∖ V − 2 y ∣ h ∣ n + α ∣ f ( y + h ) − f ( y ) ∣ d h < ∫ R n ∣ h ∣ n + α − 1 1 ∫ 0 1 ∣∇ f ( y + t h ) ∣ d t d h = ∫ 0 1 t α − 1 d t ∫ R n ∣ y − z ∣ n + α − 1 ∣∇ f ( z ) ∣ d z = α γ n , 1 − α 1 I 1 − α ( ∣∇ f ∣ ) ( y ) ,
that is,
D α f ( y ) < 1 α γ n , 1 − α I 1 − α ( ∣ ∇ f ∣ ) ( y ) , ∣ y ∣ > R . (65) D^\alpha f(y)<\frac{1}{\alpha\gamma_{n,1-\alpha}}I_{1-\alpha}(|\nabla f|)(y),\qquad|y|>R. \tag*{(65)} D α f ( y ) < α γ n , 1 − α 1 I 1 − α ( ∣∇ f ∣ ) ( y ) , ∣ y ∣ > R . (65)
Moreover, the same argument yields
D α f ( y ) ≤ 1 α γ n , 1 − α I 1 − α ( ∣ ∇ f ∣ ) ( y ) , y ∈ R n . (66) D^\alpha f(y)\le\frac{1}{\alpha\gamma_{n,1-\alpha}}I_{1-\alpha}(|\nabla f|)(y),\qquad y\in\mathbb{R}^n. \tag*{(66)} D α f ( y ) ≤ α γ n , 1 − α 1 I 1 − α ( ∣∇ f ∣ ) ( y ) , y ∈ R n . (66)
Since (65) holds on the set { ∣ y ∣ > R } \{|y|>R\} { ∣ y ∣ > R } of positive measure, while (66) holds on R n \mathbb{R}^n R n , the strict positivity of the Riesz kernel yields
I α ( D α f ) ( 0 ) < 1 α γ n , 1 − α I α I 1 − α ( ∣ ∇ f ∣ ) ( 0 ) . I_\alpha(D^\alpha f)(0)<\frac{1}{\alpha\gamma_{n,1-\alpha}}I_\alpha I_{1-\alpha}(|\nabla f|)(0). I α ( D α f ) ( 0 ) < α γ n , 1 − α 1 I α I 1 − α ( ∣∇ f ∣ ) ( 0 ) .
By the semigroup property I α I 1 − α = I 1 I_\alpha I_{1-\alpha}=I_1 I α I 1 − α = I 1 , we obtain
( 1 − α ) I α ( D α f ) ( 0 ) < 1 − α α γ n , 1 − α I 1 ( ∣ ∇ f ∣ ) ( 0 ) = C n , α H M P I 1 ( ∣ ∇ f ∣ ) ( 0 ) , \begin{aligned}
(1-\alpha)I_\alpha(D^\alpha f)(0)&<\frac{1-\alpha}{\alpha\gamma_{n,1-\alpha}}I_1(|\nabla f|)(0)\\
&=C_{n,\alpha}^{\mathrm{HMP}}I_1(|\nabla f|)(0),
\end{aligned} ( 1 − α ) I α ( D α f ) ( 0 ) < α γ n , 1 − α 1 − α I 1 ( ∣∇ f ∣ ) ( 0 ) = C n , α HMP I 1 ( ∣∇ f ∣ ) ( 0 ) ,
where the last equality follows from (1) and (3) . Since I 1 ( ∣ ∇ f ∣ ) ( 0 ) > 0 I_1(|\nabla f|)(0)>0 I 1 ( ∣∇ f ∣ ) ( 0 ) > 0 , (64) yields
C n , α ∗ < C n , α H M P . C_{n,\alpha}^*<C_{n,\alpha}^{\mathrm{HMP}}. C n , α ∗ < C n , α HMP .
Finally, recalling that
lim α → 1 − C n , α ∗ = K n , lim α → 1 − C n , α H M P = ∣ S n − 1 ∣ , \lim_{\alpha\to1^-}C_{n,\alpha}^*=K_n,\qquad\lim_{\alpha\to1^-}C_{n,\alpha}^{\mathrm{HMP}}=|\mathbb{S}^{n-1}|, α → 1 − lim C n , α ∗ = K n , α → 1 − lim C n , α HMP = ∣ S n − 1 ∣ ,
we conclude that
lim α → 1 − C n , α ∗ C n , α H M P = K n ∣ S n − 1 ∣ < 1. \lim_{\alpha\to1^-}\frac{C_{n,\alpha}^*}{C_{n,\alpha}^{\mathrm{HMP}}}=\frac{K_n}{|\mathbb{S}^{n-1}|}<1. α → 1 − lim C n , α HMP C n , α ∗ = ∣ S n − 1 ∣ K n < 1.
This completes the proof.
First-order asymptotics of the normalized Riesz potential
This section is devoted to the proof of Theorem 1.5 . Throughout this section, we assume that n ≥ 2 n \ge2 n ≥ 2 . We begin with two auxiliary lemmas.
Lemma 4.1. Let f ∈ C c 2 ( R n ) f \in C_c^2(\mathbb{R}^n) f ∈ C c 2 ( R n ) and p ∈ { 1 , ∞ } p \in\{1,\infty\} p ∈ { 1 , ∞ } . Then, as α → 1 − \alpha\to1^- α → 1 − ,
∥ ( 1 − α ) D α f − K n ∣ ∇ f ∣ − ( 1 − α ) F f ∥ L p ( R n ) = O ( ( 1 − α ) 2 ) . (67) \left\|(1-\alpha)D^\alpha f-K_n|\nabla f|-(1-\alpha)\mathcal{F}_f\right\|_{L^p(\mathbb{R}^n)}=O\left((1-\alpha)^2\right). \tag*{(67)} ∥ ( 1 − α ) D α f − K n ∣∇ f ∣ − ( 1 − α ) F f ∥ L p ( R n ) = O ( ( 1 − α ) 2 ) . (67) where F f \mathcal{F}_f F f is defined by (11) .
Proof. Fix 0 < α < 1 0<\alpha<1 0 < α < 1 . For x , h ∈ R n x,h\in\mathbb{R}^n x , h ∈ R n , set
Q f ( x , h ) = ∣ f ( x + h ) − f ( x ) ∣ − ∣ ∇ f ( x ) ⋅ h ∣ . Q_f(x,h)=|f(x+h)-f(x)|-|\nabla f(x)\cdot h|. Q f ( x , h ) = ∣ f ( x + h ) − f ( x ) ∣ − ∣∇ f ( x ) ⋅ h ∣.
By the definition of D α f D^\alpha f D α f , for every x ∈ R n x\in\mathbb{R}^n x ∈ R n ,
D α f ( x ) = ∫ ∣ h ∣ < 1 ∣ ∇ f ( x ) ⋅ h ∣ ∣ h ∣ n + α d h + ∫ ∣ h ∣ < 1 Q f ( x , h ) ∣ h ∣ n + α d h + ∫ ∣ h ∣ ≥ 1 ∣ f ( x + h ) − f ( x ) ∣ ∣ h ∣ n + α d h . D^\alpha f(x)=\int_{|h|<1}\frac{|\nabla f(x)\cdot h|}{|h|^{n+\alpha}}\,dh+\int_{|h|<1}\frac{Q_f(x,h)}{|h|^{n+\alpha}}\,dh+\int_{|h|\ge1}\frac{|f(x+h)-f(x)|}{|h|^{n+\alpha}}\,dh. D α f ( x ) = ∫ ∣ h ∣ < 1 ∣ h ∣ n + α ∣∇ f ( x ) ⋅ h ∣ d h + ∫ ∣ h ∣ < 1 ∣ h ∣ n + α Q f ( x , h ) d h + ∫ ∣ h ∣ ≥ 1 ∣ h ∣ n + α ∣ f ( x + h ) − f ( x ) ∣ d h .
Using polar coordinates and the definition of K n K_n K n , we obtain
∫ ∣ h ∣ < 1 ∣ ∇ f ( x ) ⋅ h ∣ ∣ h ∣ n + α d h = K n ∣ ∇ f ( x ) ∣ ∫ 0 1 r − α d r = K n 1 − α ∣ ∇ f ( x ) ∣ . \begin{aligned}
\int_{|h|<1}\frac{|\nabla f(x)\cdot h|}{|h|^{n+\alpha}}\,dh=K_n|\nabla f(x)|\int_0^1r^{-\alpha}\,dr \\
&=\frac{K_n}{1-\alpha}|\nabla f(x)|.
\end{aligned} ∫ ∣ h ∣ < 1 ∣ h ∣ n + α ∣∇ f ( x ) ⋅ h ∣ d h = K n ∣∇ f ( x ) ∣ ∫ 0 1 r − α d r = 1 − α K n ∣∇ f ( x ) ∣.
Consequently, by (11) ,
( 1 − α ) D α f ( x ) − K n ∣ ∇ f ( x ) ∣ − ( 1 − α ) F f ( x ) = ( 1 − α ) ∫ ∣ h ∣ < 1 Q f ( x , h ) ( 1 ∣ h ∣ n + α − 1 ∣ h ∣ n + 1 ) d h + ( 1 − α ) ∫ ∣ h ∣ ≥ 1 ∣ f ( x + h ) − f ( x ) ∣ ( 1 ∣ h ∣ n + α − 1 ∣ h ∣ n + 1 ) d h . (68) \begin{aligned}
(1-\alpha)D^\alpha f(x)-K_n|\nabla f(x)|-(1-\alpha)\mathcal{F}_f(x)
={}&(1-\alpha)\int_{|h|<1}Q_f(x,h)\left(\frac{1}{|h|^{n+\alpha}}-\frac{1}{|h|^{n+1}}\right)\,dh\\
&+(1-\alpha)\int_{|h|\ge1}|f(x+h)-f(x)|\left(\frac{1}{|h|^{n+\alpha}}-\frac{1}{|h|^{n+1}}\right)\,dh. \tag*{(68)}
\end{aligned} ( 1 − α ) D α f ( x ) − K n ∣∇ f ( x ) ∣ − ( 1 − α ) F f ( x ) = ( 1 − α ) ∫ ∣ h ∣ < 1 Q f ( x , h ) ( ∣ h ∣ n + α 1 − ∣ h ∣ n + 1 1 ) d h + ( 1 − α ) ∫ ∣ h ∣ ≥ 1 ∣ f ( x + h ) − f ( x ) ∣ ( ∣ h ∣ n + α 1 − ∣ h ∣ n + 1 1 ) d h . (68)
We now estimate the two integrals on the right-hand side. By Taylor’s theorem, for every x , h ∈ R n x,h\in\mathbb{R}^n x , h ∈ R n ,
∣ Q f ( x , h ) ∣ ≤ ∣ h ∣ 2 ∫ 0 1 ( 1 − s ) ∣ D 2 f ( x + s h ) ∣ d s , |Q_f(x,h)|\le|h|^2\int_0^1(1-s)|D^2f(x+sh)|\,ds, ∣ Q f ( x , h ) ∣ ≤ ∣ h ∣ 2 ∫ 0 1 ( 1 − s ) ∣ D 2 f ( x + s h ) ∣ d s ,
where ∣ D 2 f ∣ |D^2f| ∣ D 2 f ∣ denotes the operator norm of the Hessian matrix of f f f . Hence, taking the supremum for p = ∞ p=\infty p = ∞ , and applying Tonelli’s theorem and translation invariance for p = 1 p=1 p = 1 , we obtain, for every h ∈ R n h\in\mathbb{R}^n h ∈ R n ,
∥ Q f ( ⋅ , h ) ∥ L p ( R n ) ≤ ∣ h ∣ 2 2 ∥ D 2 f ∥ L p ( R n ) . (69) \|Q_f(\cdot,h)\|_{L^p(\mathbb{R}^n)}\le\frac{|h|^2}{2}\|D^2f\|_{L^p(\mathbb{R}^n)}. \tag*{(69)} ∥ Q f ( ⋅ , h ) ∥ L p ( R n ) ≤ 2 ∣ h ∣ 2 ∥ D 2 f ∥ L p ( R n ) . (69)
Moreover,
∥ f ( ⋅ + h ) − f ( ⋅ ) ∥ L p ( R n ) ≤ 2 ∥ f ∥ L p ( R n ) . (70) \|f(\cdot+h)-f(\cdot)\|_{L^p(\mathbb{R}^n)}\le2\|f\|_{L^p(\mathbb{R}^n)}. \tag*{(70)} ∥ f ( ⋅ + h ) − f ( ⋅ ) ∥ L p ( R n ) ≤ 2∥ f ∥ L p ( R n ) . (70)
On the other hand, for 0 < r < 1 0<r<1 0 < r < 1 ,
∣ 1 r n + α − 1 r n + 1 ∣ = 1 − r 1 − α r n + 1 ≤ ( 1 − α ) ∣ log r ∣ r n + 1 . \left|\frac{1}{r^{n+\alpha}}-\frac{1}{r^{n+1}}\right|=\frac{1-r^{1-\alpha}}{r^{n+1}}
\leq(1-\alpha)\frac{|\log r|}{r^{n+1}}. r n + α 1 − r n + 1 1 = r n + 1 1 − r 1 − α ≤ ( 1 − α ) r n + 1 ∣ log r ∣ .
Similarly, for r ≥ 1 r\geq1 r ≥ 1 ,
∣ 1 r n + α − 1 r n + 1 ∣ = 1 − r − ( 1 − α ) r n + α ≤ ( 1 − α ) log r r n + α . \left|\frac{1}{r^{n+\alpha}}-\frac{1}{r^{n+1}}\right|=\frac{1-r^{-(1-\alpha)}}{r^{n+\alpha}}
\leq(1-\alpha)\frac{\log r}{r^{n+\alpha}}. r n + α 1 − r n + 1 1 = r n + α 1 − r − ( 1 − α ) ≤ ( 1 − α ) r n + α log r .
Combining these estimates with (68) , (69) , and (70) , taking the L p L^p L p -norm, and using polar coordinates, we obtain
∥ ( 1 − α ) D α f − K n ∣ ∇ f ∣ − ( 1 − α ) F f ∥ L p ( R n ) ≤ ( 1 − α ) 2 ∣ S n − 1 ∣ ( 1 2 ∥ D 2 f ∥ L p ( R n ) + 2 α 2 ∥ f ∥ L p ( R n ) ) . \left\|(1-\alpha)\mathcal{D}^{\alpha}f-K_n|\nabla f|-(1-\alpha)\mathcal{F}_f\right\|_{L^p(\mathbb{R}^n)}
\leq(1-\alpha)^2|\mathbb{S}^{n-1}|\left(\frac{1}{2}\|D^2f\|_{L^p(\mathbb{R}^n)}+\frac{2}{\alpha^2}\|f\|_{L^p(\mathbb{R}^n)}\right). ∥ ( 1 − α ) D α f − K n ∣∇ f ∣ − ( 1 − α ) F f ∥ L p ( R n ) ≤ ( 1 − α ) 2 ∣ S n − 1 ∣ ( 2 1 ∥ D 2 f ∥ L p ( R n ) + α 2 2 ∥ f ∥ L p ( R n ) ) .
Since f ∈ C c 2 ( R n ) f\in C_c^2(\mathbb{R}^n) f ∈ C c 2 ( R n ) and α − 2 \alpha^{-2} α − 2 remains bounded as α → 1 − \alpha\to1^- α → 1 − , we obtain (67) . This completes the proof. □ \square □
Lemma 4.2. Let g ∈ L 1 ( R n ) ∩ L ∞ ( R n ) g\in L^1(\mathbb{R}^n)\cap L^\infty(\mathbb{R}^n) g ∈ L 1 ( R n ) ∩ L ∞ ( R n ) . Then, as α → 1 − \alpha\to1^- α → 1 − ,
I α g ( x ) = I 1 g ( x ) − ( 1 − α ) γ n , 1 ∫ R n log ∣ x − y ∣ + λ n ∣ x − y ∣ n − 1 g ( y ) d y + O ( ( 1 − α ) 2 ) , (71) I_\alpha g(x)=I_1g(x)-(1-\alpha)\gamma_{n,1}\int_{\mathbb{R}^n}\frac{\log|x-y|+\lambda_n}{|x-y|^{n-1}}g(y)\,dy+O((1-\alpha)^2), \tag*{(71)} I α g ( x ) = I 1 g ( x ) − ( 1 − α ) γ n , 1 ∫ R n ∣ x − y ∣ n − 1 log ∣ x − y ∣ + λ n g ( y ) d y + O (( 1 − α ) 2 ) , (71) uniformly for x ∈ R n x\in\mathbb{R}^n x ∈ R n , where λ n \lambda_n λ n is defined by (12) . Moreover, there exists a constant C n > 0 C_n>0 C n > 0 , depending only on n n n , such that
∥ I α g ∥ L ∞ ( R n ) ≤ C n ( ∥ g ∥ L 1 ( R n ) + ∥ g ∥ L ∞ ( R n ) ) (72) \|I_\alpha g\|_{L^\infty(\mathbb{R}^n)}\leq C_n\left(\|g\|_{L^1(\mathbb{R}^n)}+\|g\|_{L^\infty(\mathbb{R}^n)}\right) \tag*{(72)} ∥ I α g ∥ L ∞ ( R n ) ≤ C n ( ∥ g ∥ L 1 ( R n ) + ∥ g ∥ L ∞ ( R n ) ) (72) for every α ∈ [ 1 / 2 , 1 ] \alpha\in[1/2,1] α ∈ [ 1/2 , 1 ] .
Proof. Throughout the proof, C n > 0 C_n>0 C n > 0 denotes a constant depending only on n n n , whose value may change from line to line.
For 1 / 2 ≤ s ≤ 1 1/2\leq s\leq1 1/2 ≤ s ≤ 1 and z ∈ R n ∖ { 0 } z\in\mathbb{R}^n\setminus\{0\} z ∈ R n ∖ { 0 } , set
k s ( z ) = γ n , s ∣ z ∣ s − n . k_s(z)=\gamma_{n,s}|z|^{s-n}. k s ( z ) = γ n , s ∣ z ∣ s − n .
By (1) , we have
d d s log γ n , s = − 1 2 ψ ( n − s 2 ) − log 2 − 1 2 ψ ( s 2 ) . \frac{d}{ds}\log\gamma_{n,s}=-\frac{1}{2}\psi\left(\frac{n-s}{2}\right)-\log2-\frac{1}{2}\psi\left(\frac{s}{2}\right). d s d log γ n , s = − 2 1 ψ ( 2 n − s ) − log 2 − 2 1 ψ ( 2 s ) .
In particular, by (12) ,
d d s log γ n , s ∣ s = 1 = λ n . \left.\frac{d}{ds}\log\gamma_{n,s}\right|_{s=1}=\lambda_n. d s d log γ n , s s = 1 = λ n .
Consequently,
∂ k s ( z ) ∂ s ∣ s = 1 = γ n , 1 log ∣ z ∣ + λ n ∣ z ∣ n − 1 . (73) \left.\frac{\partial k_s(z)}{\partial s}\right|_{s=1}
=\gamma_{n,1}\frac{\log|z|+\lambda_n}{|z|^{n-1}}. \tag*{(73)} ∂ s ∂ k s ( z ) s = 1 = γ n , 1 ∣ z ∣ n − 1 log ∣ z ∣ + λ n . (73)
We now estimate the second derivative of k s k_s k s with respect to s s s . Set
b n ( s ) = d d s log γ n , s . b_n(s)=\frac{d}{ds}\log\gamma_{n,s}. b n ( s ) = d s d log γ n , s .
Direct differentiation gives
∂ 2 k s ( z ) ∂ s 2 = k s ( z ) [ ( log ∣ z ∣ + b n ( s ) ) 2 + b n ′ ( s ) ] . \frac{\partial^2 k_s(z)}{\partial s^2}=k_s(z)\left[(\log|z|+b_n(s))^2+b'_n(s)\right]. ∂ s 2 ∂ 2 k s ( z ) = k s ( z ) [ ( log ∣ z ∣ + b n ( s ) ) 2 + b n ′ ( s ) ] .
Since s ↦ γ n , s s\mapsto\gamma_{n,s} s ↦ γ n , s is positive and smooth on [ 1 / 2 , 1 ] [1/2,1] [ 1/2 , 1 ] , the functions γ n , s \gamma_{n,s} γ n , s , b n ( s ) b_n(s) b n ( s ) , and b n ′ ( s ) b'_n(s) b n ′ ( s ) are bounded on this interval. Consequently, there exists C n > 0 C_n>0 C n > 0 such that
∣ ∂ 2 k s ( z ) ∂ s 2 ∣ ≤ C n { ∣ z ∣ 1 / 2 − n ( 1 + log 2 ∣ z ∣ ) , 0 < ∣ z ∣ < 1 , ∣ z ∣ 1 − n ( 1 + log 2 ∣ z ∣ ) , ∣ z ∣ ≥ 1 , (74) \left|\frac{\partial^2 k_s(z)}{\partial s^2}\right|\le C_n
\begin{cases}
|z|^{1/2-n}(1+\log^2|z|), & 0<|z|<1,\\
|z|^{1-n}(1+\log^2|z|), & |z|\ge1,
\end{cases} \tag*{(74)} ∂ s 2 ∂ 2 k s ( z ) ≤ C n { ∣ z ∣ 1/2 − n ( 1 + log 2 ∣ z ∣ ) , ∣ z ∣ 1 − n ( 1 + log 2 ∣ z ∣ ) , 0 < ∣ z ∣ < 1 , ∣ z ∣ ≥ 1 , (74)
uniformly for s ∈ [ 1 / 2 , 1 ] s\in[1/2,1] s ∈ [ 1/2 , 1 ] .
For 1 / 2 ≤ α < 1 1/2\le\alpha<1 1/2 ≤ α < 1 and z ∈ R n ∖ { 0 } z\in\mathbb{R}^n\setminus\{0\} z ∈ R n ∖ { 0 } , set
R n , α ( z ) = k α ( z ) − k 1 ( z ) + ( 1 − α ) ∂ k s ( z ) ∂ s ∣ s = 1 . \mathcal{R}_{n,\alpha}(z)=k_\alpha(z)-k_1(z)+(1-\alpha)\left.\frac{\partial k_s(z)}{\partial s}\right|_{s=1}. R n , α ( z ) = k α ( z ) − k 1 ( z ) + ( 1 − α ) ∂ s ∂ k s ( z ) s = 1 .
Taylor’s theorem gives
∣ R n , α ( z ) ∣ ≤ ( 1 − α ) 2 2 sup α ≤ s ≤ 1 ∣ ∂ 2 k s ( z ) ∂ s 2 ∣ . |\mathcal{R}_{n,\alpha}(z)|\le\frac{(1-\alpha)^2}{2}\sup_{\alpha\le s\le1}\left|\frac{\partial^2 k_s(z)}{\partial s^2}\right|. ∣ R n , α ( z ) ∣ ≤ 2 ( 1 − α ) 2 α ≤ s ≤ 1 sup ∂ s 2 ∂ 2 k s ( z ) .
By (74) , it follows that
∣ R n , α ( z ) ∣ ≤ C n ( 1 − α ) 2 { ∣ z ∣ 1 / 2 − n ( 1 + log 2 ∣ z ∣ ) , 0 < ∣ z ∣ < 1 , ∣ z ∣ 1 − n ( 1 + log 2 ∣ z ∣ ) , ∣ z ∣ ≥ 1. (75) |\mathcal{R}_{n,\alpha}(z)|\le C_n(1-\alpha)^2
\begin{cases}
|z|^{1/2-n}(1+\log^2|z|), & 0<|z|<1,\\
|z|^{1-n}(1+\log^2|z|), & |z|\ge1.
\end{cases} \tag*{(75)} ∣ R n , α ( z ) ∣ ≤ C n ( 1 − α ) 2 { ∣ z ∣ 1/2 − n ( 1 + log 2 ∣ z ∣ ) , ∣ z ∣ 1 − n ( 1 + log 2 ∣ z ∣ ) , 0 < ∣ z ∣ < 1 , ∣ z ∣ ≥ 1. (75)
Using the definitions of I α I_\alpha I α and I 1 I_1 I 1 , together with (73) , we have, for 1 / 2 ≤ α < 1 1/2\le\alpha<1 1/2 ≤ α < 1 and x ∈ R n x\in\mathbb{R}^n x ∈ R n ,
I α g ( x ) − I 1 g ( x ) = ∫ R n ( k α ( x − y ) − k 1 ( x − y ) ) g ( y ) d y = ∫ R n ( k α ( z ) − k 1 ( z ) ) g ( x − z ) d z = ∫ R n R n , α ( z ) g ( x − z ) d z − ( 1 − α ) ∫ R n ∂ k s ( z ) ∂ s ∣ s = 1 g ( x − z ) d z = ∫ R n R n , α ( z ) g ( x − z ) d z − ( 1 − α ) γ n , 1 ∫ R n log ∣ x − y ∣ + λ n ∣ x − y ∣ n − 1 g ( y ) d y . \begin{aligned}
I_\alpha g(x)-I_1g(x)&=\int_{\mathbb{R}^n}(k_\alpha(x-y)-k_1(x-y))g(y)\,dy\\
&=\int_{\mathbb{R}^n}(k_\alpha(z)-k_1(z))g(x-z)\,dz\\
&=\int_{\mathbb{R}^n}\mathcal{R}_{n,\alpha}(z)g(x-z)\,dz-(1-\alpha)\int_{\mathbb{R}^n}\left.\frac{\partial k_s(z)}{\partial s}\right|_{s=1}g(x-z)\,dz\\
&=\int_{\mathbb{R}^n}\mathcal{R}_{n,\alpha}(z)g(x-z)\,dz-(1-\alpha)\gamma_{n,1}\int_{\mathbb{R}^n}\frac{\log|x-y|+\lambda_n}{|x-y|^{n-1}}g(y)\,dy.
\end{aligned} I α g ( x ) − I 1 g ( x ) = ∫ R n ( k α ( x − y ) − k 1 ( x − y )) g ( y ) d y = ∫ R n ( k α ( z ) − k 1 ( z )) g ( x − z ) d z = ∫ R n R n , α ( z ) g ( x − z ) d z − ( 1 − α ) ∫ R n ∂ s ∂ k s ( z ) s = 1 g ( x − z ) d z = ∫ R n R n , α ( z ) g ( x − z ) d z − ( 1 − α ) γ n , 1 ∫ R n ∣ x − y ∣ n − 1 log ∣ x − y ∣ + λ n g ( y ) d y .
Hence,
I α g ( x ) − I 1 g ( x ) + ( 1 − α ) γ n , 1 ∫ R n log ∣ x − y ∣ + λ n ∣ x − y ∣ n − 1 g ( y ) d y = ∫ R n R n , α ( z ) g ( x − z ) d z . I_\alpha g(x)-I_1g(x)+(1-\alpha)\gamma_{n,1}\int_{\mathbb{R}^n}\frac{\log|x-y|+\lambda_n}{|x-y|^{n-1}}g(y)\,dy=\int_{\mathbb{R}^n}\mathcal{R}_{n,\alpha}(z)g(x-z)\,dz. I α g ( x ) − I 1 g ( x ) + ( 1 − α ) γ n , 1 ∫ R n ∣ x − y ∣ n − 1 log ∣ x − y ∣ + λ n g ( y ) d y = ∫ R n R n , α ( z ) g ( x − z ) d z .
Together with (75) , this yields
∣ I α g ( x ) − I 1 g ( x ) + ( 1 − α ) γ n , 1 ∫ R n log ∣ x − y ∣ + λ n ∣ x − y ∣ n − 1 g ( y ) d y ∣ ≤ C n ( 1 − α ) 2 [ ∥ g ∥ L ∞ ( R n ) ∫ ∣ z ∣ < 1 1 + log 2 ∣ z ∣ ∣ z ∣ n − 1 / 2 d z + ∥ g ∥ L 1 ( R n ) sup ∣ z ∣ ≥ 1 1 + log 2 ∣ z ∣ ∣ z ∣ n − 1 ] . \begin{aligned}
\left|I_\alpha g(x)-I_1g(x)+(1-\alpha)\gamma_{n,1}\int_{\mathbb{R}^n}\frac{\log|x-y|+\lambda_n}{|x-y|^{n-1}}g(y)\,dy\right|\\
\le C_n(1-\alpha)^2\left[\|g\|_{L^\infty(\mathbb{R}^n)}\int_{|z|<1}\frac{1+\log^2|z|}{|z|^{n-1/2}}\,dz+\|g\|_{L^1(\mathbb{R}^n)}\sup_{|z|\ge1}\frac{1+\log^2|z|}{|z|^{n-1}}\right].
\end{aligned} I α g ( x ) − I 1 g ( x ) + ( 1 − α ) γ n , 1 ∫ R n ∣ x − y ∣ n − 1 log ∣ x − y ∣ + λ n g ( y ) d y ≤ C n ( 1 − α ) 2 [ ∥ g ∥ L ∞ ( R n ) ∫ ∣ z ∣ < 1 ∣ z ∣ n − 1/2 1 + log 2 ∣ z ∣ d z + ∥ g ∥ L 1 ( R n ) ∣ z ∣ ≥ 1 sup ∣ z ∣ n − 1 1 + log 2 ∣ z ∣ ] .
Since n ≥ 2 n \ge2 n ≥ 2 , we have
∫ ∣ z ∣ < 1 1 + log 2 ∣ z ∣ ∣ z ∣ n − 1 / 2 d z = ∣ S n − 1 ∣ ∫ 0 1 r − 1 / 2 ( 1 + log 2 r ) d r < ∞ \int_{|z|<1} \frac{1+\log^2|z|}{|z|^{n-1/2}}\,dz=|\mathbb{S}^{n-1}|\int_0^1 r^{-1/2}(1+\log^2r)\,dr<\infty ∫ ∣ z ∣ < 1 ∣ z ∣ n − 1/2 1 + log 2 ∣ z ∣ d z = ∣ S n − 1 ∣ ∫ 0 1 r − 1/2 ( 1 + log 2 r ) d r < ∞
and
sup ∣ z ∣ ≥ 1 1 + log 2 ∣ z ∣ ∣ z ∣ n − 1 < ∞ . \sup_{|z|\ge1}\frac{1+\log^2|z|}{|z|^{n-1}}<\infty. ∣ z ∣ ≥ 1 sup ∣ z ∣ n − 1 1 + log 2 ∣ z ∣ < ∞.
Therefore,
∥ I α g − I 1 g + ( 1 − α ) γ n , 1 ∫ R n log ∣ ⋅ − y ∣ + λ n ∣ ⋅ − y ∣ n − 1 g ( y ) d y ∥ L ∞ ( R n ) ≤ C n ( 1 − α ) 2 ( ∥ g ∥ L 1 ( R n ) + ∥ g ∥ L ∞ ( R n ) ) . \left\|I_\alpha g-I_1g+(1-\alpha)\gamma_{n,1}\int_{\mathbb{R}^n}\frac{\log|\cdot-y|+\lambda_n}{|\cdot-y|^{n-1}}g(y)\,dy\right\|_{L^\infty(\mathbb{R}^n)}
\le C_n(1-\alpha)^2\left(\|g\|_{L^1(\mathbb{R}^n)}+\|g\|_{L^\infty(\mathbb{R}^n)}\right). I α g − I 1 g + ( 1 − α ) γ n , 1 ∫ R n ∣ ⋅ − y ∣ n − 1 log ∣ ⋅ − y ∣ + λ n g ( y ) d y L ∞ ( R n ) ≤ C n ( 1 − α ) 2 ( ∥ g ∥ L 1 ( R n ) + ∥ g ∥ L ∞ ( R n ) ) .
This proves (71) .
We now prove the second assertion. For 1 / 2 ≤ α ≤ 1 1/2\le\alpha\le1 1/2 ≤ α ≤ 1 and x ∈ R n x\in\mathbb{R}^n x ∈ R n , we have
∣ I α g ( x ) ∣ ≤ γ n , α ∫ R n ∣ g ( y ) ∣ ∣ x − y ∣ n − α d y ≤ γ n , α ( ∥ g ∥ L ∞ ( R n ) ∫ ∣ z ∣ < 1 d z ∣ z ∣ n − α + ∥ g ∥ L 1 ( R n ) ) = γ n , α ( ∣ S n − 1 ∣ α ∥ g ∥ L ∞ ( R n ) + ∥ g ∥ L 1 ( R n ) ) . \begin{aligned}
|I_\alpha g(x)|\le\gamma_{n,\alpha}\int_{\mathbb{R}^n}\frac{|g(y)|}{|x-y|^{n-\alpha}}\,dy \\
&\le\gamma_{n,\alpha}\left(\|g\|_{L^\infty(\mathbb{R}^n)}\int_{|z|<1}\frac{dz}{|z|^{n-\alpha}}+\|g\|_{L^1(\mathbb{R}^n)}\right) \\
&=\gamma_{n,\alpha}\left(\frac{|\mathbb{S}^{n-1}|}{\alpha}\|g\|_{L^\infty(\mathbb{R}^n)}+\|g\|_{L^1(\mathbb{R}^n)}\right).
\end{aligned} ∣ I α g ( x ) ∣ ≤ γ n , α ∫ R n ∣ x − y ∣ n − α ∣ g ( y ) ∣ d y ≤ γ n , α ( ∥ g ∥ L ∞ ( R n ) ∫ ∣ z ∣ < 1 ∣ z ∣ n − α d z + ∥ g ∥ L 1 ( R n ) ) = γ n , α ( α ∣ S n − 1 ∣ ∥ g ∥ L ∞ ( R n ) + ∥ g ∥ L 1 ( R n ) ) .
Since γ n , α \gamma_{n,\alpha} γ n , α is bounded on [ 1 / 2 , 1 ] [1/2,1] [ 1/2 , 1 ] , we obtain (72) . This completes the proof. □
Proof of Theorem 1.5
Proof. Let f ∈ C c 2 ( R n ) f\in C_c^2(\mathbb{R}^n) f ∈ C c 2 ( R n ) and set
ε = 1 − α . \varepsilon=1-\alpha. ε = 1 − α .
We first note that
F f ∈ L 1 ( R n ) ∩ L ∞ ( R n ) . \mathcal{F}f\in L^1(\mathbb{R}^n)\cap L^\infty(\mathbb{R}^n). F f ∈ L 1 ( R n ) ∩ L ∞ ( R n ) .
Indeed, by (11) , (69) , and (70) , for p ∈ { 1 , ∞ } p\in\{1,\infty\} p ∈ { 1 , ∞ } , we have
∥ F f ∥ L p ( R n ) ≤ 1 2 ∥ D 2 f ∥ L p ( R n ) ∫ ∣ h ∣ < 1 d h ∣ h ∣ n − 1 + 2 ∥ f ∥ L p ( R n ) ∫ ∣ h ∣ ≥ 1 d h ∣ h ∣ n + 1 = ∣ S n − 1 ∣ ( 1 2 ∥ D 2 f ∥ L p ( R n ) + 2 ∥ f ∥ L p ( R n ) ) < ∞ . \begin{aligned}
\|\mathcal{F}f\|_{L^p(\mathbb{R}^n)}
\le\frac{1}{2}\|D^2f\|_{L^p(\mathbb{R}^n)}\int_{|h|<1}\frac{dh}{|h|^{n-1}}
+2\|f\|_{L^p(\mathbb{R}^n)}\int_{|h|\ge1}\frac{dh}{|h|^{n+1}} \\
&=|\mathbb{S}^{n-1}|\left(\frac{1}{2}\|D^2f\|_{L^p(\mathbb{R}^n)}+2\|f\|_{L^p(\mathbb{R}^n)}\right)
<\infty.
\end{aligned} ∥ F f ∥ L p ( R n ) ≤ 2 1 ∥ D 2 f ∥ L p ( R n ) ∫ ∣ h ∣ < 1 ∣ h ∣ n − 1 d h + 2∥ f ∥ L p ( R n ) ∫ ∣ h ∣ ≥ 1 ∣ h ∣ n + 1 d h = ∣ S n − 1 ∣ ( 2 1 ∥ D 2 f ∥ L p ( R n ) + 2∥ f ∥ L p ( R n ) ) < ∞.
For g ∈ L 1 ( R n ) ∩ L ∞ ( R n ) g\in L^1(\mathbb{R}^n)\cap L^\infty(\mathbb{R}^n) g ∈ L 1 ( R n ) ∩ L ∞ ( R n ) , define
L n g ( x ) = γ n , 1 ∫ R n log ∣ x − y ∣ + λ n ∣ x − y ∣ n − 1 g ( y ) d y . \mathcal{L}_n g(x)=\gamma_{n,1}\int_{\mathbb{R}^n}\frac{\log|x-y|+\lambda_n}{|x-y|^{n-1}}g(y)\,dy. L n g ( x ) = γ n , 1 ∫ R n ∣ x − y ∣ n − 1 log ∣ x − y ∣ + λ n g ( y ) d y .
Since n ≥ 2 n\ge2 n ≥ 2 ,
∫ ∣ z ∣ < 1 1 + ∣ log ∣ z ∣ ∣ ∣ z ∣ n − 1 d z < ∞ , sup ∣ z ∣ ≥ 1 1 + log ∣ z ∣ ∣ z ∣ n − 1 < ∞ . \int_{|z|<1}\frac{1+|\log|z||}{|z|^{n-1}}\,dz<\infty,\qquad
\sup_{|z|\ge1}\frac{1+\log|z|}{|z|^{n-1}}<\infty. ∫ ∣ z ∣ < 1 ∣ z ∣ n − 1 1 + ∣ log ∣ z ∣∣ d z < ∞ , ∣ z ∣ ≥ 1 sup ∣ z ∣ n − 1 1 + log ∣ z ∣ < ∞.
Consequently, there exists a constant C n > 0 C_n>0 C n > 0 , depending only on n n n , such that
∥ L n g ∥ L ∞ ( R n ) ≤ C n ( ∥ g ∥ L 1 ( R n ) + ∥ g ∥ L ∞ ( R n ) ) . (76) \|\mathcal{L}_n g\|_{L^\infty(\mathbb{R}^n)}
\le C_n\left(\|g\|_{L^1(\mathbb{R}^n)}+\|g\|_{L^\infty(\mathbb{R}^n)}\right). \tag*{(76)} ∥ L n g ∥ L ∞ ( R n ) ≤ C n ( ∥ g ∥ L 1 ( R n ) + ∥ g ∥ L ∞ ( R n ) ) . (76)
By Lemma 4.1, we can write
ε D α f = K n ∣ ∇ f ∣ + ε F f + E α , \varepsilon D^\alpha f = K_n|\nabla f| + \varepsilon\mathcal{F}_f + E_\alpha, ε D α f = K n ∣∇ f ∣ + ε F f + E α ,
where
∥ E α ∥ L 1 ( R n ) + ∥ E α ∥ L ∞ ( R n ) = O ( ε 2 ) , ε → 0 + . (77) \|E_\alpha\|_{L^1(\mathbb{R}^n)}+\|E_\alpha\|_{L^\infty(\mathbb{R}^n)}=O(\varepsilon^2),\qquad\varepsilon\to0^+. \tag*{(77)} ∥ E α ∥ L 1 ( R n ) + ∥ E α ∥ L ∞ ( R n ) = O ( ε 2 ) , ε → 0 + . (77)
Applying I α I_\alpha I α and using its linearity, we obtain
ε I α ( D α f ) = K n I α ( ∣ ∇ f ∣ ) + ε I α ( F f ) + I α E α . (78) \varepsilon I_\alpha(D^\alpha f)=K_nI_\alpha(|\nabla f|)+\varepsilon I_\alpha(\mathcal{F}_f)+I_\alpha E_\alpha. \tag*{(78)} ε I α ( D α f ) = K n I α ( ∣∇ f ∣ ) + ε I α ( F f ) + I α E α . (78)
We now estimate the three terms on the right-hand side. By the first assertion of Lemma 4.2,
I α ( ∣ ∇ f ∣ ) = I 1 ( ∣ ∇ f ∣ ) − ε L n ( ∣ ∇ f ∣ ) + O L ∞ ( ε 2 ) . (79) I_\alpha(|\nabla f|)=I_1(|\nabla f|)-\varepsilon\mathcal{L}_n(|\nabla f|)+O_{L^\infty}(\varepsilon^2). \tag*{(79)} I α ( ∣∇ f ∣ ) = I 1 ( ∣∇ f ∣ ) − ε L n ( ∣∇ f ∣ ) + O L ∞ ( ε 2 ) . (79)
Applying the same lemma to g = F f g=\mathcal{F}_f g = F f and using (76) , gives
I α ( F f ) = I 1 ( F f ) + O L ∞ ( ε ) . (80) I_\alpha(\mathcal{F}_f)=I_1(\mathcal{F}_f)+O_{L^\infty}(\varepsilon). \tag*{(80)} I α ( F f ) = I 1 ( F f ) + O L ∞ ( ε ) . (80)
Finally, the second assertion of Lemma 4.2, together with (77) , yields
∥ I α E α ∥ L ∞ ( R n ) ≤ C n ( ∥ E α ∥ L 1 ( R n ) + ∥ E α ∥ L ∞ ( R n ) ) = O ( ε 2 ) . (81) \|I_\alpha E_\alpha\|_{L^\infty(\mathbb{R}^n)}\leq C_n\left(\|E_\alpha\|_{L^1(\mathbb{R}^n)}+\|E_\alpha\|_{L^\infty(\mathbb{R}^n)}\right)=O(\varepsilon^2). \tag*{(81)} ∥ I α E α ∥ L ∞ ( R n ) ≤ C n ( ∥ E α ∥ L 1 ( R n ) + ∥ E α ∥ L ∞ ( R n ) ) = O ( ε 2 ) . (81)
Substituting (79) , (80) , and (81) into (78) , we obtain
ε I α ( D α f ) = K n I 1 ( ∣ ∇ f ∣ ) + ε ( I 1 ( F f ) − K n L n ( ∣ ∇ f ∣ ) ) + O L ∞ ( ε 2 ) . \varepsilon I_\alpha(D^\alpha f)=K_nI_1(|\nabla f|)+\varepsilon\left(I_1(\mathcal{F}_f)-K_n\mathcal{L}_n(|\nabla f|)\right)+O_{L^\infty}(\varepsilon^2). ε I α ( D α f ) = K n I 1 ( ∣∇ f ∣ ) + ε ( I 1 ( F f ) − K n L n ( ∣∇ f ∣ ) ) + O L ∞ ( ε 2 ) .
By (13) ,
I 1 ( F f ) − K n L n ( ∣ ∇ f ∣ ) = A f . I_1(\mathcal{F}_f)-K_n\mathcal{L}_n(|\nabla f|)=A_f. I 1 ( F f ) − K n L n ( ∣∇ f ∣ ) = A f .
Therefore, (14) follows. □
Almost-everywhere convergence for W1,1 functions
In this section, we prove the almost-everywhere convergence stated in Theorem 1.6 for functions in W 1 , 1 ( R n ) W^{1,1}(\mathbb{R}^n) W 1 , 1 ( R n ) , without passing to a subsequence. Throughout this section, we assume that n ≥ 2 n\geq2 n ≥ 2 .
We first establish the following lemma.
Lemma 5.1. There exists a constant C n > 0 C_n>0 C n > 0 , depending only on n n n , with the following property: for every u , v ∈ W 1 , 1 ( R n ) u,v\in W^{1,1}(\mathbb{R}^n) u , v ∈ W 1 , 1 ( R n ) , there exists a set E u , v ⊂ R n E_{u,v}\subset\mathbb{R}^n E u , v ⊂ R n of full measure such that
∣ ( 1 − α ) I α ( D α u ) ( x ) − ( 1 − α ) I α ( D α v ) ( x ) ∣ ≤ C n I 1 ( ∣ ∇ ( u − v ) ∣ ) ( x ) < ∞ (82) \left|(1-\alpha)I_\alpha(D^\alpha u)(x)-(1-\alpha)I_\alpha(D^\alpha v)(x)\right|\leq C_nI_1(|\nabla(u-v)|)(x)<\infty \tag*{(82)} ∣ ( 1 − α ) I α ( D α u ) ( x ) − ( 1 − α ) I α ( D α v ) ( x ) ∣ ≤ C n I 1 ( ∣∇ ( u − v ) ∣ ) ( x ) < ∞ (82) for every x ∈ E u , v x\in E_{u,v} x ∈ E u , v and every α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) .
Proof. By (2) and the fact that C n , α H M P C_{n,\alpha}^{\mathrm{HMP}} C n , α HMP is uniformly bounded for α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) , there exists a constant C n > 0 C_n>0 C n > 0 , depending only on n n n , such that
( 1 − α ) I α ( D α φ ) ( x ) ≤ C n I 1 ( ∣ ∇ φ ∣ ) ( x ) (83) (1-\alpha)I_\alpha(D^\alpha\varphi)(x)\leq C_nI_1(|\nabla\varphi|)(x) \tag*{(83)} ( 1 − α ) I α ( D α φ ) ( x ) ≤ C n I 1 ( ∣∇ φ ∣ ) ( x ) (83)
for every φ ∈ C c ∞ ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) , every x ∈ R n x\in\mathbb{R}^n x ∈ R n , and every α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) .
We claim that for every w ∈ W 1 , 1 ( R n ) w\in W^{1,1}(\mathbb{R}^n) w ∈ W 1 , 1 ( R n ) , there exists a set E w ⊂ R n E_w\subset\mathbb{R}^n E w ⊂ R n of full measure such that
( 1 − α ) I α ( D α w ) ( x ) ≤ C n I 1 ( ∣ ∇ w ∣ ) ( x ) < ∞ (84) (1-\alpha)I_\alpha(D^\alpha w)(x)\leq C_nI_1(|\nabla w|)(x)<\infty \tag*{(84)} ( 1 − α ) I α ( D α w ) ( x ) ≤ C n I 1 ( ∣∇ w ∣ ) ( x ) < ∞ (84)
for every x ∈ E w x\in E_w x ∈ E w and every α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) . Indeed, fix w ∈ W 1 , 1 ( R n ) w \in W^{1,1}(\mathbb{R}^{n}) w ∈ W 1 , 1 ( R n ) . By the density of C c ∞ ( R n ) C_{c}^{\infty}(\mathbb{R}^{n}) C c ∞ ( R n ) in W 1 , 1 ( R n ) W^{1,1}(\mathbb{R}^{n}) W 1 , 1 ( R n ) , we can choose ( w j ) ⊂ C c ∞ ( R n ) (w_j) \subset C_{c}^{\infty}(\mathbb{R}^{n}) ( w j ) ⊂ C c ∞ ( R n ) such that
∥ w j − w ∥ W 1 , 1 ( R n ) ≤ 2 − j , j ≥ 1. (85) \lVert w_j-w\rVert_{W^{1,1}(\mathbb{R}^{n})} \le2^{-j}, \qquad j \ge1. \tag*{(85)} ∥ w j − w ∥ W 1 , 1 ( R n ) ≤ 2 − j , j ≥ 1. (85)
Consequently,
G = ∑ j = 1 ∞ ∣ ∇ ( w j − w ) ∣ ∈ L 1 ( R n ) . G=\sum_{j=1}^{\infty}|\nabla(w_j-w)|\in L^1(\mathbb{R}^{n}). G = j = 1 ∑ ∞ ∣∇ ( w j − w ) ∣ ∈ L 1 ( R n ) .
Since the Riesz potential I 1 h I_1h I 1 h is finite almost everywhere whenever h ∈ L 1 ( R n ) h\in L^1(\mathbb{R}^{n}) h ∈ L 1 ( R n ) , the set
E w = { x ∈ R n : I 1 ( G + ∣ ∇ w ∣ ) ( x ) < ∞ } E_w=\{x\in\mathbb{R}^{n}: I_1(G+|\nabla w|)(x)<\infty\} E w = { x ∈ R n : I 1 ( G + ∣∇ w ∣ ) ( x ) < ∞ }
has full measure. Moreover, for every x ∈ E w x\in E_w x ∈ E w , Tonelli’s theorem gives
∑ j = 1 ∞ I 1 ( ∣ ∇ ( w j − w ) ∣ ) ( x ) = I 1 G ( x ) < ∞ . \sum_{j=1}^{\infty}I_1(|\nabla(w_j-w)|)(x)=I_1G(x)<\infty. j = 1 ∑ ∞ I 1 ( ∣∇ ( w j − w ) ∣ ) ( x ) = I 1 G ( x ) < ∞.
Hence, for every x ∈ E w x\in E_w x ∈ E w ,
∣ I 1 ( ∣ ∇ w j ∣ ) ( x ) − I 1 ( ∣ ∇ w ∣ ) ( x ) ∣ ≤ I 1 ( ∣ ∇ ( w j − w ) ∣ ) ( x ) ⟶ 0 (86) |I_1(|\nabla w_j|)(x)-I_1(|\nabla w|)(x)|\le I_1(|\nabla(w_j-w)|)(x)\longrightarrow0 \tag*{(86)} ∣ I 1 ( ∣∇ w j ∣ ) ( x ) − I 1 ( ∣∇ w ∣ ) ( x ) ∣ ≤ I 1 ( ∣∇ ( w j − w ) ∣ ) ( x ) ⟶ 0 (86)
as j → ∞ j\to\infty j → ∞ .
Now fix x ∈ E w x\in E_w x ∈ E w and α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) . By (85) , w j → w w_j\to w w j → w almost everywhere in R n \mathbb{R}^{n} R n as j → ∞ j\to\infty j → ∞ . Hence, applying Fatou’s lemma to the double integral defining I α ( D α w ) ( x ) I_\alpha(D^\alpha w)(x) I α ( D α w ) ( x ) , and using (83) and (86) , we obtain
( 1 − α ) I α ( D α w ) ( x ) ≤ lim inf j → ∞ ( 1 − α ) I α ( D α w j ) ( x ) ≤ C n lim j → ∞ I 1 ( ∣ ∇ w j ∣ ) ( x ) = C n I 1 ( ∣ ∇ w ∣ ) ( x ) < ∞ . \begin{aligned}
(1-\alpha)I_\alpha(D^\alpha w)(x)
&\le\liminf_{j\to\infty}(1-\alpha)I_\alpha(D^\alpha w_j)(x)\\
&\le C_n\lim_{j\to\infty}I_1(|\nabla w_j|)(x)\\
&=C_nI_1(|\nabla w|)(x)\\
&<\infty.
\end{aligned} ( 1 − α ) I α ( D α w ) ( x ) ≤ j → ∞ lim inf ( 1 − α ) I α ( D α w j ) ( x ) ≤ C n j → ∞ lim I 1 ( ∣∇ w j ∣ ) ( x ) = C n I 1 ( ∣∇ w ∣ ) ( x ) < ∞.
Since E w E_w E w and the sequence ( w j ) (w_j) ( w j ) are independent of α \alpha α , this inequality holds for every x ∈ E w x\in E_w x ∈ E w and every α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) . This proves (84) .
Finally, let u , v ∈ W 1 , 1 ( R n ) u,v\in W^{1,1}(\mathbb{R}^{n}) u , v ∈ W 1 , 1 ( R n ) and set
E u , v = E u ∩ E v ∩ E u − v . E_{u,v}=E_u\cap E_v\cap E_{u-v}. E u , v = E u ∩ E v ∩ E u − v .
Since each of these sets has full measure, so does E u , v E_{u,v} E u , v . Moreover, the reverse triangle inequality yields, for every 0 < α < 1 0<\alpha<1 0 < α < 1 ,
∣ D α u ( y ) − D α v ( y ) ∣ ≤ D α ( u − v ) ( y ) |D^\alpha u(y)-D^\alpha v(y)|\le D^\alpha(u-v)(y) ∣ D α u ( y ) − D α v ( y ) ∣ ≤ D α ( u − v ) ( y )
for almost every y ∈ R n y\in\mathbb{R}^{n} y ∈ R n . Consequently, by (84) , for every x ∈ E u , v x\in E_{u,v} x ∈ E u , v and every α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) ,
∣ ( 1 − α ) I α ( D α u ) ( x ) − ( 1 − α ) I α ( D α v ) ( x ) ∣ ≤ ( 1 − α ) I α ( D α ( u − v ) ) ( x ) ≤ C n I 1 ( ∣ ∇ ( u − v ) ∣ ) ( x ) < ∞ . \begin{aligned}
\left|(1-\alpha)I_\alpha(D^\alpha u)(x)-(1-\alpha)I_\alpha(D^\alpha v)(x)\right|
&\le(1-\alpha)I_\alpha(D^\alpha(u-v))(x)\\
&\le C_nI_1(|\nabla(u-v)|)(x)\\
&<\infty.
\end{aligned} ∣ ( 1 − α ) I α ( D α u ) ( x ) − ( 1 − α ) I α ( D α v ) ( x ) ∣ ≤ ( 1 − α ) I α ( D α ( u − v )) ( x ) ≤ C n I 1 ( ∣∇ ( u − v ) ∣ ) ( x ) < ∞.
This proves (82) .
Proof of Theorem 1.6
Proof. Let f ∈ W 1 , 1 ( R n ) f \in W^{1,1}(\mathbb{R}^n) f ∈ W 1 , 1 ( R n ) . By the density of C ∞ ( R n ) C^\infty(\mathbb{R}^n) C ∞ ( R n ) in W 1 , 1 ( R n ) W^{1,1}(\mathbb{R}^n) W 1 , 1 ( R n ) , we can choose ( f j ) ⊂ C c ∞ ( R n ) (f_j) \subset C^\infty_c(\mathbb{R}^n) ( f j ) ⊂ C c ∞ ( R n ) such that
∥ f j − f ∥ W 1 , 1 ( R n ) ≤ 2 − j , j ≥ 1. \lVert f_j-f\rVert_{W^{1,1}(\mathbb{R}^n)} \le2^{-j}, \qquad j \ge1. ∥ f j − f ∥ W 1 , 1 ( R n ) ≤ 2 − j , j ≥ 1.
Consequently,
G = ∑ j = 1 ∞ ∣ ∇ ( f j − f ) ∣ ∈ L 1 ( R n ) . G=\sum_{j=1}^{\infty} \lvert\nabla(f_j-f)\rvert\in L^1(\mathbb{R}^n). G = j = 1 ∑ ∞ ∣ ∇ ( f j − f )∣ ∈ L 1 ( R n ) .
As in the proof of Lemma 5.1 , the set
E 0 = { x ∈ R n : I 1 ( G + ∣ ∇ f ∣ ) ( x ) < ∞ } E_0=\{x\in\mathbb{R}^n:I_1(G+\lvert\nabla f\rvert)(x)<\infty\} E 0 = { x ∈ R n : I 1 ( G + ∣ ∇ f ∣) ( x ) < ∞ }
has full measure, and
∑ j = 1 ∞ I 1 ( ∣ ∇ ( f j − f ) ∣ ) ( x ) = I 1 G ( x ) < ∞ \sum_{j=1}^{\infty}I_1(\lvert\nabla(f_j-f)\rvert)(x)=I_1G(x)<\infty j = 1 ∑ ∞ I 1 (∣ ∇ ( f j − f )∣) ( x ) = I 1 G ( x ) < ∞
for every x ∈ E 0 x\in E_0 x ∈ E 0 . Consequently,
I 1 ( ∣ ∇ ( f j − f ) ∣ ) ( x ) ⟶ 0 as j → ∞ (87) I_1(\lvert\nabla(f_j-f)\rvert)(x)\longrightarrow0 \qquad\text{as }j\to\infty \tag*{(87)} I 1 (∣ ∇ ( f j − f )∣) ( x ) ⟶ 0 as j → ∞ (87)
for every x ∈ E 0 x\in E_0 x ∈ E 0 .
For each j ≥ 1 j\ge1 j ≥ 1 , let E j E_j E j be the set of full measure provided by Lemma 5.1 for the pair ( f , f j ) (f,f_j) ( f , f j ) . Set
E = E 0 ∩ ⋂ j = 1 ∞ E j . E=E_0\cap\bigcap_{j=1}^{\infty}E_j. E = E 0 ∩ j = 1 ⋂ ∞ E j .
Then E E E has full measure. Fix x ∈ E x\in E x ∈ E . By Lemma 5.1 , for every j ≥ 1 j\ge1 j ≥ 1 and every α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) ,
∣ ( 1 − α ) I α ( D α f ) ( x ) − ( 1 − α ) I α ( D α f j ) ( x ) ∣ ≤ C n I 1 ( ∣ ∇ ( f − f j ) ∣ ) ( x ) . \left|(1-\alpha)I_\alpha(D^\alpha f)(x)-(1-\alpha)I_\alpha(D^\alpha f_j)(x)\right|\le C_nI_1(\lvert\nabla(f-f_j)\rvert)(x). ∣ ( 1 − α ) I α ( D α f ) ( x ) − ( 1 − α ) I α ( D α f j ) ( x ) ∣ ≤ C n I 1 (∣ ∇ ( f − f j )∣) ( x ) .
Moreover,
∣ I 1 ( ∣ ∇ f j ∣ ) ( x ) − I 1 ( ∣ ∇ f ∣ ) ( x ) ∣ ≤ I 1 ( ∣ ∇ ( f j − f ) ∣ ) ( x ) . \left|I_1(\lvert\nabla f_j\rvert)(x)-I_1(\lvert\nabla f\rvert)(x)\right|\le I_1(\lvert\nabla(f_j-f)\rvert)(x). ∣ I 1 (∣ ∇ f j ∣) ( x ) − I 1 (∣ ∇ f ∣) ( x ) ∣ ≤ I 1 (∣ ∇ ( f j − f )∣) ( x ) .
Combining these inequalities, we obtain
∣ ( 1 − α ) I α ( D α f ) ( x ) − K n I 1 ( ∣ ∇ f ∣ ) ( x ) ∣ ≤ ∣ ( 1 − α ) I α ( D α f j ) ( x ) − K n I 1 ( ∣ ∇ f j ∣ ) ( x ) ∣ + ( C n + K n ) I 1 ( ∣ ∇ ( f j − f ) ∣ ) ( x ) . (88) \begin{aligned}
\left|(1-\alpha)I_\alpha(D^\alpha f)(x)-K_nI_1(\lvert\nabla f\rvert)(x)\right|
&\le\left|(1-\alpha)I_\alpha(D^\alpha f_j)(x)-K_nI_1(\lvert\nabla f_j\rvert)(x)\right|\\
&\quad+(C_n+K_n)I_1(\lvert\nabla(f_j-f)\rvert)(x). \tag*{(88)}
\end{aligned} ∣ ( 1 − α ) I α ( D α f ) ( x ) − K n I 1 (∣ ∇ f ∣) ( x ) ∣ ≤ ∣ ( 1 − α ) I α ( D α f j ) ( x ) − K n I 1 (∣ ∇ f j ∣) ( x ) ∣ + ( C n + K n ) I 1 (∣ ∇ ( f j − f )∣) ( x ) . (88)
For each fixed j ≥ 1 j\ge1 j ≥ 1 , since f j ∈ C c ∞ ( R n ) f_j\in C^\infty_c(\mathbb{R}^n) f j ∈ C c ∞ ( R n ) , Theorem 1.5 yields
lim α → 1 − ( 1 − α ) I α ( D α f j ) ( x ) = K n I 1 ( ∣ ∇ f j ∣ ) ( x ) . \lim_{\alpha\to1^-}(1-\alpha)I_\alpha(D^\alpha f_j)(x)=K_nI_1(\lvert\nabla f_j\rvert)(x). α → 1 − lim ( 1 − α ) I α ( D α f j ) ( x ) = K n I 1 (∣ ∇ f j ∣) ( x ) .
Hence, taking the limit superior as α → 1 − \alpha\to1^- α → 1 − in (88) , we obtain
lim sup α → 1 − ∣ ( 1 − α ) I α ( D α f ) ( x ) − K n I 1 ( ∣ ∇ f ∣ ) ( x ) ∣ ≤ ( C n + K n ) I 1 ( ∣ ∇ ( f j − f ) ∣ ) ( x ) . \limsup_{\alpha\to1^-}\left|(1-\alpha)I_\alpha(D^\alpha f)(x)-K_nI_1(\lvert\nabla f\rvert)(x)\right|\le(C_n+K_n)I_1(\lvert\nabla(f_j-f)\rvert)(x). α → 1 − lim sup ∣ ( 1 − α ) I α ( D α f ) ( x ) − K n I 1 (∣ ∇ f ∣) ( x ) ∣ ≤ ( C n + K n ) I 1 (∣ ∇ ( f j − f )∣) ( x ) .
Finally, letting j → ∞ j\to\infty j → ∞ and using (87) , we conclude that
lim α → 1 − ( 1 − α ) I α ( D α f ) ( x ) = K n I 1 ( ∣ ∇ f ∣ ) ( x ) . \lim_{\alpha\to1^-}(1-\alpha)I_\alpha(D^\alpha f)(x)=K_nI_1(\lvert\nabla f\rvert)(x). α → 1 − lim ( 1 − α ) I α ( D α f ) ( x ) = K n I 1 (∣ ∇ f ∣) ( x ) .
Since E E E has full measure, the desired almost-everywhere convergence follows. ∎
Further remarks on the sharp constants
In this section, we discuss two further aspects of the sharp-constant problem. We first examine the difficulty of determining the unrestricted optimal constant in dimension two for arbitrary 0 < α < 1 0 < \alpha< 1 0 < α < 1 . We then derive an explicit convergent series representation for the sharp constant C n , α ∗ C_{n,\alpha}^{*} C n , α ∗ in dimensions n ≥ 3 n \ge3 n ≥ 3 .
The unrestricted sharp constant in dimension two
Although Theorems 1.1 and 1.2 provide a precise description of the sharp-constant problem in dimension two, the exact determination of C 2 , α ∗ C_{2,\alpha}^{*} C 2 , α ∗ for arbitrary 0 < α < 1 0 < \alpha< 1 0 < α < 1 remains unresolved by our approach. Indeed, Lemmas 2.1 and 2.3 yield
C 2 , α r a d = ( 1 − α ) γ 2 , α γ 2 , 1 A α ( 0 ) ≤ C 2 , α ∗ ≤ ( 1 − α ) γ 2 , α γ 2 , 1 sup β ∈ [ 0 , π / 2 ] A α ( β ) . C_{2,\alpha}^{\mathrm{rad}}=(1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}A_{\alpha}(0)\le C_{2,\alpha}^{*}\le(1-\alpha)\frac{\gamma_{2,\alpha}}{\gamma_{2,1}}\sup_{\beta\in[0,\pi/2]}A_{\alpha}(\beta). C 2 , α rad = ( 1 − α ) γ 2 , 1 γ 2 , α A α ( 0 ) ≤ C 2 , α ∗ ≤ ( 1 − α ) γ 2 , 1 γ 2 , α β ∈ [ 0 , π /2 ] sup A α ( β ) .
By Lemma 2.2 ,
A α ( β ) = 4 1 − α + O ( 1 ) A_{\alpha}(\beta)=\frac{4}{1-\alpha}+O(1) A α ( β ) = 1 − α 4 + O ( 1 )
uniformly for β ∈ [ 0 , π / 2 ] \beta\in[0,\pi/2] β ∈ [ 0 , π /2 ] as α → 1 − \alpha\to1^{-} α → 1 − . Since
γ 2 , α γ 2 , 1 = 1 + O ( 1 − α ) as α → 1 − , \frac{\gamma_{2,\alpha}}{\gamma_{2,1}}=1+O(1-\alpha)\qquad\text{as }\alpha\to1^{-}, γ 2 , 1 γ 2 , α = 1 + O ( 1 − α ) as α → 1 − ,
we obtain
0 ≤ C 2 , α ∗ − C 2 , α r a d = O ( 1 − α ) as α → 1 − . 0\le C_{2,\alpha}^{*}-C_{2,\alpha}^{\mathrm{rad}}=O(1-\alpha)\qquad\text{as }\alpha\to1^{-}. 0 ≤ C 2 , α ∗ − C 2 , α rad = O ( 1 − α ) as α → 1 − .
Thus, the radial class is asymptotically optimal at leading order, even though its global optimality for each fixed α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) has not been established.
A natural question is whether
C 2 , α ∗ = C 2 , α r a d , 0 < α < 1. C_{2,\alpha}^{*}=C_{2,\alpha}^{\mathrm{rad}},\qquad0<\alpha<1. C 2 , α ∗ = C 2 , α rad , 0 < α < 1.
A sufficient condition would be
A α ( β ) ≤ A α ( 0 ) , β ∈ [ 0 , π / 2 ] , A_{\alpha}(\beta)\le A_{\alpha}(0),\qquad\beta\in[0,\pi/2], A α ( β ) ≤ A α ( 0 ) , β ∈ [ 0 , π /2 ] ,
which would identify the radial direction as a maximizer of the associated directional functional. However, the monotonicity argument used in the proof of Lemma 3.4 for n ≥ 3 n\ge3 n ≥ 3 does not extend directly to dimension two. Indeed, that argument relies on
a = n − α 2 > 1 a=\frac{n-\alpha}{2}>1 a = 2 n − α > 1
and on the additional term
− n − 2 3 τ -\frac{n-2}{3}\tau − 3 n − 2 τ
in the estimate for the logarithmic derivative of the transformed kernel. In dimension two, we instead have
a = 2 − α 2 < 1 , n − 2 = 0 , a=\frac{2-\alpha}{2}<1,\qquad n-2=0, a = 2 2 − α < 1 , n − 2 = 0 ,
so the same reasoning no longer establishes the required monotonicity.
It remains an interesting question whether
C 2 , α ∗ = C 2 , α r a d C_{2,\alpha}^{*}=C_{2,\alpha}^{\mathrm{rad}} C 2 , α ∗ = C 2 , α rad
for every 0 < α < 1 0<\alpha<1 0 < α < 1 , or at least for α \alpha α sufficiently close to 1.
An explicit series representation in higher dimensions
We now complement the integral representation (9) with an absolutely convergent series for C n , α r a d \mathcal{C}_{n,\alpha}^{\mathrm{rad}} C n , α rad . Combined with (10) , this yields an explicit series representation for the sharp constant C n , α ∗ \mathcal{C}_{n,\alpha}^{*} C n , α ∗ for every n ≥ 3 n \ge3 n ≥ 3 and 0 < α < 1 0 < \alpha< 1 0 < α < 1 .
Proposition 6.1. Let n ≥ 3 n \ge3 n ≥ 3 and 0 < α < 1 0 < \alpha< 1 0 < α < 1 . Then
C n , α r a d = ∣ S n − 1 ∣ ∑ k = 0 ∞ ( n + α 2 ) k ( 1 + α 2 ) k ( n 2 ) k k ! ( 1 ( α + 2 k ) 2 + 1 ( n + 2 k ) 2 ) . (89) \mathcal{C}_{n,\alpha}^{\mathrm{rad}} = |\mathbb{S}^{n-1}| \sum_{k=0}^{\infty} \frac{\left(\frac{n+\alpha}{2}\right)_k \left(1+\frac{\alpha}{2}\right)_k}{\left(\frac{n}{2}\right)_k k!} \left(\frac{1}{(\alpha+2k)^2}+\frac{1}{(n+2k)^2}\right). \tag*{(89)} C n , α rad = ∣ S n − 1 ∣ k = 0 ∑ ∞ ( 2 n ) k k ! ( 2 n + α ) k ( 1 + 2 α ) k ( ( α + 2 k ) 2 1 + ( n + 2 k ) 2 1 ) . (89) The series converges absolutely for every 0 < α < 1 0 < \alpha< 1 0 < α < 1 .
Proof. Fix 0 < α < 1 0 < \alpha< 1 0 < α < 1 and set
p = n + α 2 . p = \frac{n+\alpha}{2}. p = 2 n + α . By (39) , for τ > 0 \tau> 0 τ > 0 and 0 ≤ θ ≤ π 0 \le\theta\le\pi 0 ≤ θ ≤ π , we have
K n , α ( τ , θ ) = e − α τ + e − n τ ( 1 − 2 e − τ cos θ + e − 2 τ ) p . \mathcal{K}_{n,\alpha}(\tau,\theta) = \frac{e^{-\alpha\tau}+e^{-n\tau}}{\left(1-2e^{-\tau}\cos\theta+e^{-2\tau}\right)^p}. K n , α ( τ , θ ) = ( 1 − 2 e − τ cos θ + e − 2 τ ) p e − α τ + e − n τ . By (9) ,
C n , α r a d = ∣ S n − 2 ∣ ∫ 0 ∞ τ ( e − α τ + e − n τ ) ∫ 0 π ( sin θ ) n − 2 ( 1 − 2 e − τ cos θ + e − 2 τ ) p d θ d τ . \mathcal{C}_{n,\alpha}^{\mathrm{rad}} = |\mathbb{S}^{n-2}| \int_{0}^{\infty} \tau\left(e^{-\alpha\tau}+e^{-n\tau}\right)\int_{0}^{\pi}\frac{(\sin\theta)^{n-2}}{\left(1-2e^{-\tau}\cos\theta+e^{-2\tau}\right)^p}\,d\theta\,d\tau. C n , α rad = ∣ S n − 2 ∣ ∫ 0 ∞ τ ( e − α τ + e − n τ ) ∫ 0 π ( 1 − 2 e − τ cos θ + e − 2 τ ) p ( sin θ ) n − 2 d θ d τ . Making the change of variables t = e − τ t=e^{-\tau} t = e − τ , we obtain
C n , α r a d = ∣ S n − 2 ∣ ∫ 0 1 ( − log t ) ( t α − 1 + t n − 1 ) ∫ 0 π ( sin θ ) n − 2 ( 1 − 2 t cos θ + t 2 ) p d θ d t . \mathcal{C}_{n,\alpha}^{\mathrm{rad}} = |\mathbb{S}^{n-2}| \int_{0}^{1}(-\log t)\left(t^{\alpha-1}+t^{n-1}\right)\int_{0}^{\pi}\frac{(\sin\theta)^{n-2}}{\left(1-2t\cos\theta+t^2\right)^p}\,d\theta\,dt. C n , α rad = ∣ S n − 2 ∣ ∫ 0 1 ( − log t ) ( t α − 1 + t n − 1 ) ∫ 0 π ( 1 − 2 t cos θ + t 2 ) p ( sin θ ) n − 2 d θ d t . Now, using the classical hypergeometric identity (see, e.g., [2 Sections 2.2 and 3.1] )
∣ S n − 2 ∣ ∫ 0 π ( sin θ ) n − 2 ( 1 − 2 t cos θ + t 2 ) p d θ = ∣ S n − 1 ∣ 2 F 1 ( p , 1 + α 2 ; n 2 ; t 2 ) , 0 < t < 1 , |\mathbb{S}^{n-2}|\int_{0}^{\pi}\frac{(\sin\theta)^{n-2}}{\left(1-2t\cos\theta+t^2\right)^p}\,d\theta= |\mathbb{S}^{n-1}|\,{}_2F_1\left(p,1+\frac{\alpha}{2};\frac{n}{2};t^2\right), \qquad0<t<1, ∣ S n − 2 ∣ ∫ 0 π ( 1 − 2 t cos θ + t 2 ) p ( sin θ ) n − 2 d θ = ∣ S n − 1 ∣ 2 F 1 ( p , 1 + 2 α ; 2 n ; t 2 ) , 0 < t < 1 , we deduce that
C n , α r a d = ∣ S n − 1 ∣ ∫ 0 1 ( − log t ) ( t α − 1 + t n − 1 ) 2 F 1 ( p , 1 + α 2 ; n 2 ; t 2 ) d t . \mathcal{C}_{n,\alpha}^{\mathrm{rad}} = |\mathbb{S}^{n-1}|\int_{0}^{1}(-\log t)\left(t^{\alpha-1}+t^{n-1}\right){}_2F_1\left(p,1+\frac{\alpha}{2};\frac{n}{2};t^2\right)\,dt. C n , α rad = ∣ S n − 1 ∣ ∫ 0 1 ( − log t ) ( t α − 1 + t n − 1 ) 2 F 1 ( p , 1 + 2 α ; 2 n ; t 2 ) d t . Expanding the hypergeometric function into its defining series (see, e.g., [2 Section 2.1] ) and using
∫ 0 1 ( − log t ) t λ − 1 d t = 1 λ 2 , λ > 0 , \int_{0}^{1}(-\log t)t^{\lambda-1}\,dt=\frac{1}{\lambda^2}, \qquad\lambda>0, ∫ 0 1 ( − log t ) t λ − 1 d t = λ 2 1 , λ > 0 , we obtain (89) , where the interchange of summation and integration is justified by the nonnegativity of all terms.
Finally, by the standard asymptotic properties of the Gamma function, we have
( n + α 2 ) k ( 1 + α 2 ) k ( n 2 ) k k ! = O ( k α ) as k → ∞ . \frac{\left(\frac{n+\alpha}{2}\right)_k\left(1+\frac{\alpha}{2}\right)_k}{\left(\frac{n}{2}\right)_k k!}=O(k^\alpha) \qquad\text{as } k\to\infty. ( 2 n ) k k ! ( 2 n + α ) k ( 1 + 2 α ) k = O ( k α ) as k → ∞. Thus, the summands in (89) are O ( k α − 2 ) O(k^{\alpha-2}) O ( k α − 2 ) as k → ∞ k\to\infty k → ∞ . Since 0 < α < 1 0<\alpha<1 0 < α < 1 , the series converges absolutely.
By Legendre’s duplication formula,
Γ ( n − 1 2 ) Γ ( n 2 ) = 2 2 − n π Γ ( n − 1 ) , \Gamma\left(\frac{n-1}{2}\right)\Gamma\left(\frac{n}{2}\right)=2^{2-n}\sqrt{\pi}\Gamma(n-1), Γ ( 2 n − 1 ) Γ ( 2 n ) = 2 2 − n π Γ ( n − 1 ) ,
and (1) , we obtain
γ n , α γ n , 1 ∣ S n − 1 ∣ = 2 n − α π n / 2 Γ ( n − α 2 ) Γ ( α / 2 ) Γ ( n − 1 ) . \frac{\gamma_{n,\alpha}}{\gamma_{n,1}}|S^{n-1}|=\frac{2^{n-\alpha}\pi^{n/2}\Gamma\left(\frac{n-\alpha}{2}\right)}{\Gamma(\alpha/2)\Gamma(n-1)}. γ n , 1 γ n , α ∣ S n − 1 ∣ = Γ ( α /2 ) Γ ( n − 1 ) 2 n − α π n /2 Γ ( 2 n − α ) .
Combining this identity with (10) and Proposition 6.1, we obtain the following explicit series representation of the sharp constant C n , α ∗ C_{n,\alpha}^{*} C n , α ∗ .
Corollary 6.2. Let n ≥ 3 n \ge3 n ≥ 3 and 0 < α < 1 0 < \alpha< 1 0 < α < 1 . Then
C n , α ∗ = 2 n − α π n / 2 ( 1 − α ) Γ ( n − α 2 ) Γ ( α / 2 ) Γ ( n − 1 ) ∑ k = 0 ∞ ( n + α 2 ) k ( n 2 ) k k ! ( 1 ( α + 2 k ) 2 + 1 ( n + 2 k ) 2 ) . C_{n,\alpha}^{*}=\frac{2^{n-\alpha}\pi^{n/2}(1-\alpha)\Gamma\left(\frac{n-\alpha}{2}\right)}{\Gamma(\alpha/2)\Gamma(n-1)}
\sum_{k=0}^{\infty}\frac{\left(\frac{n+\alpha}{2}\right)_{k}}{\left(\frac{n}{2}\right)_{k}k!}\left(\frac{1}{(\alpha+2k)^{2}}+\frac{1}{(n+2k)^{2}}\right). C n , α ∗ = Γ ( α /2 ) Γ ( n − 1 ) 2 n − α π n /2 ( 1 − α ) Γ ( 2 n − α ) k = 0 ∑ ∞ ( 2 n ) k k ! ( 2 n + α ) k ( ( α + 2 k ) 2 1 + ( n + 2 k ) 2 1 ) . The series converges absolutely.
Data availability. No data was used for the research described in the article.
Conflict of interest. The authors declare that they have no conflict of interest.
References
(M. Jleli) Department of Mathematics, College of Science, King Saud University, Riyadh 11451, Saudi Arabia
Email address: jleli@ksu.edu.sa
(B. Samet) Department of Mathematics, College of Science, King Saud University, Riyadh 11451, Saudi Arabia
Email address: bessem.samet@gmail.com