Keywords. Buckling eigenvalue; Pólya–Szegő conjecture; Isoperimetric inequality; Bessel functions; Sobolev capacity
2020 Mathematics Subject Classification. Primary 35P15; Secondary 35J40, 49Q10, 33C10, 46E35
Introduction
In 1951, Pólya and Szegő conjectured that, among clamped plates of prescribed area subjected to uniform compression in the plane of the plate, the disk has the least critical buckling load [13 ] . Its higher-dimensional form asks whether the ball minimizes the first clamped buckling eigenvalue at fixed volume. This is a fourth-order counterpart of the Faber–Krahn theorem for the Dirichlet Laplacian. The question tests how much geometric information can be retained when both the bending energy and the gradient energy enter a variational quotient.
In the planar model, a buckling eigenfunction describes the transverse deflection of a homogeneous thin plate, and the critical compressive load is proportional to the first eigenvalue [5 Section 1] . A trial deflection gives an upper bound immediately; a lower bound must control every admissible deflection. A sharp bound depending only on area therefore gives a shape-independent threshold below which this idealized plate remains stable. In higher dimensions the same problem belongs to spectral geometry and shape optimization: it connects a fourth-order variational problem with sharp second-order spectral inequalities and the geometry of Sobolev spaces.
Let Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n be open, with 0 < ∣ Ω ∣ < ∞ 0 < |\Omega| < \infty 0 < ∣Ω∣ < ∞ , where ∣ Ω ∣ |\Omega| ∣Ω∣ denotes Lebesgue measure. For j = 1 , 2 j = 1,2 j = 1 , 2 , the Sobolev space H j ( Ω ) H^{j}(\Omega) H j ( Ω ) consists of functions whose weak derivatives through order j j j are square integrable, and H 0 j ( Ω ) H_{0}^{j}(\Omega) H 0 j ( Ω ) is the closure of the smooth, compactly supported functions C c ∞ ( Ω ) C_{c}^{\infty}(\Omega) C c ∞ ( Ω ) in this space. Define
Λ 1 ( Ω ) = inf 0 ≠ U ∈ H 0 2 ( Ω ) ∫ Ω ( Δ U ) 2 d x ∫ Ω ∣ ∇ U ∣ 2 d x , λ 1 D ( Ω ) = inf 0 ≠ v ∈ H 0 1 ( Ω ) ∫ Ω ∣ ∇ v ∣ 2 d x ∫ Ω v 2 d x . (1) \begin{aligned}
\Lambda_{1}(\Omega) &= \inf_{0\ne U\in H_{0}^{2}(\Omega)} \frac{\int_{\Omega}(\Delta U)^{2}\,dx}{\int_{\Omega}|\nabla U|^{2}\,dx},\\
\lambda_{1}^{\mathrm{D}}(\Omega) &= \inf_{0\ne v\in H_{0}^{1}(\Omega)} \frac{\int_{\Omega}|\nabla v|^{2}\,dx}{\int_{\Omega}v^{2}\,dx}. \tag*{(1)}
\end{aligned} Λ 1 ( Ω ) λ 1 D ( Ω ) = 0 = U ∈ H 0 2 ( Ω ) inf ∫ Ω ∣∇ U ∣ 2 d x ∫ Ω ( Δ U ) 2 d x , = 0 = v ∈ H 0 1 ( Ω ) inf ∫ Ω v 2 d x ∫ Ω ∣∇ v ∣ 2 d x . (1)
Here ∇ \nabla ∇ and Δ \Delta Δ are the Euclidean gradient and Laplacian. On a smooth bounded domain, the first quotient gives the eigenvalue problem
Δ 2 U + Λ Δ U = 0 in Ω , U = ∂ n U = 0 on ∂ Ω . \Delta^{2}U+\Lambda\Delta U=0 \quad\text{in }\Omega,\qquad U=\partial_{n}U=0 \quad\text{on }\partial\Omega. Δ 2 U + ΛΔ U = 0 in Ω , U = ∂ n U = 0 on ∂ Ω.
where n \mathbf{n} n is the outward unit normal. These are the clamped boundary conditions. The closure definition in (1) continues to make sense for arbitrary open sets.
Write B R ( x 0 ) B_{R}(x_{0}) B R ( x 0 ) for the open ball of radius R R R centered at x 0 x_{0} x 0 , and B R = B R ( 0 ) B_{R}=B_{R}(0) B R = B R ( 0 ) . Let J η J_{\eta} J η denote the Bessel function of the first kind of order η \eta η , and j η , l j_{\eta,l} j η , l its l l l th positive zero. The classical ball formulas are
λ 1 D ( B R ) = j n / 2 − 1 , 1 2 R 2 , Λ 1 ( B R ) = j n / 2 , 1 2 R 2 . \lambda_{1}^{\mathrm{D}}(B_{R})=\frac{j_{n/2-1,1}^{2}}{R^{2}},\qquad\Lambda_{1}(B_{R})=\frac{j_{n/2,1}^{2}}{R^{2}}. λ 1 D ( B R ) = R 2 j n /2 − 1 , 1 2 , Λ 1 ( B R ) = R 2 j n /2 , 1 2 .
see [5 Sections 1 and 4] and [8 Theorem 3.2 and Section 6] . Writing ω n = ∣ B 1 ∣ \omega_{n}=|B_{1}| ω n = ∣ B 1 ∣ , the volume comparison takes the explicit form
∣ Ω ∣ 2 / n Λ 1 ( Ω ) ≥ ω n 2 / n j n / 2 , 1 2 , ∣ B R ∣ = ∣ Ω ∣ . |\Omega|^{2/n}\Lambda_{1}(\Omega)\ge\omega_{n}^{2/n}j_{n/2,1}^{2},\qquad|B_{R}|=|\Omega|. ∣Ω ∣ 2/ n Λ 1 ( Ω ) ≥ ω n 2/ n j n /2 , 1 2 , ∣ B R ∣ = ∣Ω∣.
For smooth bounded domains the conjectured equality case is a ball. For arbitrary open sets the precise equality statement involves Sobolev capacity, as explained below. The exponent 2 / n 2/n 2/ n is forced by the dilation law Λ 1 ( t Ω ) = t − 2 Λ 1 ( Ω ) \Lambda_{1}(t\Omega)=t^{-2}\Lambda_{1}(\Omega) Λ 1 ( t Ω ) = t − 2 Λ 1 ( Ω ) for t > 0 t>0 t > 0 ; the Dirichlet eigenvalue obeys the same law.
The related Rayleigh conjecture concerns a different quotient,
Γ 1 ( Ω ) = inf 0 ≠ U ∈ H 0 2 ( Ω ) ∫ Ω ( Δ U ) 2 d x ∫ Ω U 2 d x \Gamma_1(\Omega)=\inf_{0\ne U\in H_0^2(\Omega)}\frac{\int_\Omega(\Delta U)^2\,dx}{\int_\Omega U^2\,dx} Γ 1 ( Ω ) = 0 = U ∈ H 0 2 ( Ω ) inf ∫ Ω U 2 d x ∫ Ω ( Δ U ) 2 d x
which describes the vibration of a clamped plate. Talenti’s comparison method [18 ] was a decisive advance: it treated the positive and negative regions separately and reduced the fourth-order problem to radial comparison problems on two balls. Nadirashvili [11 ] recovered the sharp planar bound by retaining additional coupling between the rearranged problems. Ashbaugh and Benguria [2 ] gave a detailed Bessel-function analysis that also proved the three-dimensional vibration inequality. These results established that sign changes need not prevent a sharp fourth-order isoperimetric theorem. They do not, however, establish the corresponding buckling inequality: replacing ∫ U 2 d x \int U^2\,dx ∫ U 2 d x by ∫ ∣ ∇ U ∣ 2 d x \int|\nabla U|^2\,dx ∫ ∣∇ U ∣ 2 d x changes the comparison problem.
For buckling, Szegő’s pioneering symmetrization argument [16 , 17 ] proves the planar result under an additional sign assumption on a first eigenfunction. That assumption is unavailable on general domains; the examples and the role of admissibility are discussed in [5 Section 8] and [3 ] . In particular, taking an absolute value does not generally preserve H 2 H^2 H 2 , and ordinary decreasing rearrangement does not supply the required control of ∫ ( Δ U ) 2 d x \int(\Delta U)^2\,dx ∫ ( Δ U ) 2 d x . The obstacle is therefore not simply a missing maximum-principle argument: both the clamped function space and the two energies in the quotient must be retained.
A second approach compares buckling and membrane spectra. Extending the observation of Bramble and Payne [6 ] , Ashbaugh and Laugesen [5 ] obtained, on bounded domains,
Λ 1 ( Ω ) ≥ c n Λ 1 ( B R ) , c n = 2 2 / n j n / 2 − 1 , 1 2 j n / 2 , 1 2 < 1 , c n ⟶ 1 ( n ⟶ ∞ ) . \Lambda_1(\Omega)\ge c_n\Lambda_1(B_R),\qquad c_n=2^{2/n}\frac{j_{n/2-1,1}^2}{j_{n/2,1}^2}<1,\qquad c_n\longrightarrow1\quad(n\longrightarrow\infty). Λ 1 ( Ω ) ≥ c n Λ 1 ( B R ) , c n = 2 2/ n j n /2 , 1 2 j n /2 − 1 , 1 2 < 1 , c n ⟶ 1 ( n ⟶ ∞ ) .
Their result provides an explicit bound in every dimension and is asymptotically sharp. It also identifies a concrete limitation of that route: the passage through the second Dirichlet eigenvalue and its two-ball volume bound leaves a loss at each fixed dimension. Ashbaugh, Benguria and Laugesen [3 ] explain why the analogous two-ball reduction is sharp for vibration in dimensions two and three but gives only a partial result for buckling. This is a limitation of those reductions, rather than an obstruction to the conjecture itself.
Shape optimization offers a complementary approach. Ashbaugh and Bucur [4 ] developed planar existence and relaxation methods and examined the Willms–Weinberger characterization of smooth optimizers. Stollenwerk [14 ] proved existence of an optimal open set in every dimension, allowing unbounded competitors. Stollenwerk and Wagner [15 ] proved that a smooth minimizing domain with connected boundary is a ball, using the second domain variation. More recently, Lamboley and Nahon [10 ] obtained planar boundary regularity near points where the limiting local area fraction is less than one, with a description of the possible angular singularities. These results isolate an important difficulty: existence does not by itself provide all the regularity and geometric hypotheses needed to characterize an optimizer.
Tang and Zhang [19 ] have announced a sharp planar buckling–Dirichlet comparison in a public working preprint. Their construction uses radial eigenfunctions to define a local transform of a test function and its gradient, together with two divergence corrections. We use this calibration and weak variational framework. Our contribution is its higher-dimensional realization and the analytic estimates that control all Hessian components through dimension forty-seven. We also give a coefficient proof for the planar case.
Our aim is to preserve the ball’s sharp spectral ratio while avoiding assumptions on a minimizing eigenfunction or on an optimal domain. The result is stronger than the volume comparison, because it retains the first Dirichlet eigenvalue of the original set.
We first compare the two eigenvalues on the same open set.
Theorem 1.1. Let 2 ≤ n ≤ 47 2 \le n \le47 2 ≤ n ≤ 47 . For every open set Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n of finite positive volume,
Λ 1 ( Ω ) ≥ j n / 2 , 1 2 j n / 2 − 1 , 1 2 λ 1 D ( Ω ) . (2) \Lambda_1(\Omega) \ge\frac{j_{n/2,1}^{2}}{j_{n/2-1,1}^{2}}\lambda_1^{D}(\Omega). \tag*{(2)} Λ 1 ( Ω ) ≥ j n /2 − 1 , 1 2 j n /2 , 1 2 λ 1 D ( Ω ) . (2) The constant is optimal and is attained on balls.
The Faber–Krahn inequality compares λ 1 D ( Ω ) \lambda_1^{D}(\Omega) λ 1 D ( Ω ) with that of an equal-volume ball; see [7 , 13 ] . It converts Theorem 1.1 into the desired volume bound. For arbitrary open sets, equality in the resulting buckling bound requires an H 2 H^2 H 2 -capacity condition. We write C 2 , 2 C_{2,2} C 2 , 2 for the Bessel capacity associated with H 2 ( R n ) H^2(\mathbb{R}^{n}) H 2 ( R n ) . A set of zero such capacity is negligible for the value trace of H 2 H^2 H 2 functions. A precise definition and the relevant trace criterion are given in § 2.4 .
The spectral comparison gives the following sharp volume inequality and equality class.
Corollary 1.2. Under the hypotheses of Theorem 1.1 , let ∣ B R ∣ = ∣ Ω ∣ |B_R|=|\Omega| ∣ B R ∣ = ∣Ω∣ . Then
Λ 1 ( Ω ) ≥ j n / 2 , 1 2 R 2 = Λ 1 ( B R ) . \Lambda_1(\Omega) \ge\frac{j_{n/2,1}^{2}}{R^{2}}=\Lambda_1(B_R). Λ 1 ( Ω ) ≥ R 2 j n /2 , 1 2 = Λ 1 ( B R ) . Equality holds if and only if
Ω = B R ( x 0 ) ∖ E , C 2 , 2 ( E ) = 0 , \Omega=B_R(x_0)\setminus E,\qquad C_{2,2}(E)=0, Ω = B R ( x 0 ) ∖ E , C 2 , 2 ( E ) = 0 , where E E E is relatively closed in B R ( x 0 ) B_R(x_0) B R ( x 0 ) . For n = 2 , 3 n=2,3 n = 2 , 3 , this means that Ω \Omega Ω is a translated ball. For n ≥ 4 n\ge4 n ≥ 4 , sets of zero capacity may be removed.
The proof starts with the first buckling mode and the first Dirichlet mode of a ball. Their quotient determines a scalar profile f f f and a local transform F = ∣ ∇ U ∣ f ( U / ∣ ∇ U ∣ ) F=|\nabla U|f(U/|\nabla U|) F = ∣∇ U ∣ f ( U /∣∇ U ∣ ) , extended continuously where the gradient vanishes. Two divergence terms turn the difference between the fourth-order and second-order energies into a quadratic expression in the Hessian D 2 U D^{2}U D 2 U , the matrix of second derivatives. We call this identity a calibration : the ball profile is a null direction, and a nonnegative remainder yields the sharp spectral comparison after integration.
The new difficulty in dimension n n n lies in the n − 1 n-1 n − 1 tangential Hessian directions. Their trace couples to the radial entry, while their trace-free part, obtained by subtracting the scalar matrix with the same trace, has its own coefficient. Controlling only the planar block would leave this term untreated. We prove the required coefficient signs by a residue expansion in which the first remaining pole dominates a possibly signed infinite tail. Positive products for the radial quotient then give a rational energy estimate. This already controls the matrix through dimension ten. In the higher dimensions, a positive Laplace representation, together with two exact endpoint moments, gives the stronger estimate needed for the determinant. Together, these estimates make the Hessian remainder nonnegative for 2 ≤ n ≤ 47 2 \le n \le47 2 ≤ n ≤ 47 . They follow directly from the radial equations, independently of the spectral comparison, and control entire parameter intervals and infinite coefficient tails.
The global step is equally necessary. The transform is made globally Lipschitz in the independent pair ( U , ∇ U ) (U,\nabla U) ( U , ∇ U ) , with the appropriate parities across U = 0 U=0 U = 0 . Locality of weak derivatives handles { ∇ U = 0 } \{\nabla U=0\} { ∇ U = 0 } even when it has positive measure. Hence no flux is lost across nodal or critical sets. Dilation and Sobolev approximation then prove (2) without assuming that a first buckling eigenfunction is positive, or even using existence of one. Applying Faber–Krahn then gives the volume inequality. The equality argument combines Faber–Krahn rigidity with the clamped Sobolev trace criterion, which explains why zero capacity, rather than zero volume, is the appropriate qualification.
The dimension restriction is intrinsic to this particular calibration. With q = j n / 2 − 1 , 1 q=j_{n/2-1,1} q = j n /2 − 1 , 1 and k = j n / 2 , 1 k=j_{n/2,1} k = j n /2 , 1 , the leading endpoint sign of its coupled determinant is the sign of 3 ( k 2 − q 2 ) − 4 ( n − 1 ) 3\left(k^{2}-q^{2}\right)-4(n-1) 3 ( k 2 − q 2 ) − 4 ( n − 1 ) . We prove that it is positive for 3 ≤ n ≤ 47 3 \le n \le47 3 ≤ n ≤ 47 and negative for every n ≥ 48 n \ge48 n ≥ 48 . Thus the method has an exact dimensional endpoint. This does not disprove the buckling conjecture in higher dimensions, nor does the present result settle the separate Rayleigh vibration problem.
Section § 2 contains the calibration construction, the sharp spectral comparison, and the equality argument. § 3 proves the coefficient signs in dimensions three through forty-seven. § 4 supplies the radial bounds and determinant estimates, and § 5 identifies the exact dimensional endpoint of the calibration. Appendix § 6 contains the planar coefficient estimates.
The calibration and the sharp comparison
The proof begins with the ball profiles and a correction of the Hessian energy. We establish positivity of the resulting matrix, pass to a global functional inequality, and then identify the equality cases. The coefficient and multiplier estimates used here are proved in § 3 and § 4 , with the planar coefficient argument in Appendix § 6 .
Radial construction
All functions are real. We identify H 0 j ( Ω ) H_0^j(\Omega) H 0 j ( Ω ) with the closure of the zero extensions of C c ∞ ( Ω ) C_c^\infty(\Omega) C c ∞ ( Ω ) in H j ( R n ) H^j(\mathbb{R}^n) H j ( R n ) ; the norms agree on these test functions. Integrals without a domain are over the ambient Euclidean space, or over the interval explicitly specified by their limits. The norm of a matrix is its Frobenius norm, ∣ H ∣ 2 = ∑ i , j H i j 2 |H|^2 = \sum_{i,j} H_{ij}^2 ∣ H ∣ 2 = ∑ i , j H ij 2 . A symmetric matrix M M M is positive semidefinite if v T M v ≥ 0 v^{\mathsf T}Mv \ge0 v T M v ≥ 0 for every real vector v v v . A prime or the subscript r r r denotes a radial derivative; derivatives with respect to other variables are indicated explicitly. The notation O ( r a ) O(r^a) O ( r a ) has its usual local meaning, with the dimension fixed. The relation f ≍ g f \asymp g f ≍ g means that the ratio of the two positive quantities stays between positive constants. We use the Bessel series, recurrence relations, zero interlacing, and canonical products in [12 Sections 10.2, 10.6 and 10.21] . For Sturm comparison, which compares the spacing of zeros of second-order equations by ordering their potentials, we use [20 Chapter 5] .
For the radial construction put ν = n / 2 − 1 \nu= n/2 - 1 ν = n /2 − 1 , q = j ν , 1 q = j_{\nu,1} q = j ν , 1 , k = j ν + 1 , 1 k = j_{\nu+1,1} k = j ν + 1 , 1 and λ = q 2 / k 2 \lambda= q^2/k^2 λ = q 2 / k 2 . Put m = n − 1 m = n - 1 m = n − 1 . Define ψ n ( r ) = 2 ν Γ ( ν + 1 ) r − ν J ν ( r ) \psi_n(r) = 2^\nu\Gamma(\nu+ 1)r^{-\nu}J_\nu(r) ψ n ( r ) = 2 ν Γ ( ν + 1 ) r − ν J ν ( r ) , continued by ψ n ( 0 ) = 1 \psi_n(0) = 1 ψ n ( 0 ) = 1 , where
Γ ( z ) = ∫ 0 ∞ t z − 1 e − t d t \Gamma(z) = \int_0^\infty t^{z-1}e^{-t}\,\mathrm{d}t Γ ( z ) = ∫ 0 ∞ t z − 1 e − t d t
for z > 0 z > 0 z > 0 is the Euler gamma function [12 Section 5.2(i)] . The same formulas define the radial profiles for real n ≥ 2 n \ge2 n ≥ 2 , as used in some parameter estimates below. Put
c = − ψ n ( k ) > 0 , σ = 1 + c , α = q / k , λ = α 2 , β = − α ψ n ′ ( q ) c > 0 , K = β 2 . c = -\psi_n(k) > 0,\qquad\sigma= 1 + c,\qquad\alpha= q/k,\qquad\lambda= \alpha^2,\qquad\beta= -\frac{\alpha\psi_n'(q)}{c} > 0,\qquad K = \beta^2. c = − ψ n ( k ) > 0 , σ = 1 + c , α = q / k , λ = α 2 , β = − c α ψ n ′ ( q ) > 0 , K = β 2 .
Let X = ψ n X = \psi_n X = ψ n , with X ( 0 ) = 1 X(0) = 1 X ( 0 ) = 1 , and put
p = − X r = ( r / n ) ψ n + 2 ( r ) , u = X + c , ϕ = X ( α r ) , 0 < r < k . p = -X_r = (r/n)\psi_{n+2}(r),\qquad u = X + c,\qquad\phi= X(\alpha r),\qquad0 < r < k. p = − X r = ( r / n ) ψ n + 2 ( r ) , u = X + c , ϕ = X ( α r ) , 0 < r < k .
The recurrence and interlacing give q < k < j ν , 2 q < k < j_{\nu,2} q < k < j ν , 2 , p > 0 p > 0 p > 0 , u > 0 u > 0 u > 0 , and X ( k ) = − c X(k) = -c X ( k ) = − c with c > 0 c > 0 c > 0 . The equations are
X r r + m X r / r + X = 0 , p r r + m p r / r + ( 1 − m / r 2 ) p = 0 , ϕ r r + m ϕ r / r + λ ϕ = 0. X_{rr} + mX_r/r + X = 0,\qquad p_{rr} + mp_r/r + (1 - m/r^2)p = 0,\qquad\phi_{rr} + m\phi_r/r + \lambda\phi= 0. X r r + m X r / r + X = 0 , p r r + m p r / r + ( 1 − m / r 2 ) p = 0 , ϕ r r + m ϕ r / r + λ ϕ = 0.
Also Δ u = − ( u − c ) \Delta u = -(u - c) Δ u = − ( u − c ) and Δ 2 u + Δ u = 0 \Delta^2u + \Delta u = 0 Δ 2 u + Δ u = 0 on B k B_k B k . The functions u u u and ϕ \phi ϕ are the ball buckling and Dirichlet modes at the common scale B k B_k B k . Their distinct orders of vanishing are shown in Figure 1 . The quotient u / p u/p u / p turns these radial data into a profile defined on the whole positive half-line.
Figure 1. The radial buckling and Dirichlet profiles in dimension three, normalized to have value one at the center. At the boundary r = k r=k r = k , the buckling profile has a double zero, u ( k ) = u r ( k ) = 0 u(k)=u_r(k)=0 u ( k ) = u r ( k ) = 0 , whereas ϕ ( k ) = 0 \phi(k)=0 ϕ ( k ) = 0 and ϕ r ( k ) < 0 \phi_r(k)<0 ϕ r ( k ) < 0 . This difference explains why the calibration uses the quotient u / ( − u r ) u/(-u_r) u / ( − u r ) to compare the two profiles. The curves illustrate the explicit definitions and endpoint expansions in § 2.1 ; no sign assertion is inferred from the plot.
The next lemma makes the radial quotient a global coordinate for the calibration profiles.
Lemma 2.1. The function t = u / p t=u/p t = u / p is a smooth strictly decreasing bijection from ( 0 , k ) (0,k) ( 0 , k ) onto ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) .
Proof. Set S = p 2 + u p r S=p^2+up_r S = p 2 + u p r . The canonical product gives
p ( r ) = r n ∏ l ≥ 1 ( 1 − r 2 κ l 2 ) , κ l = j ν + 1 , l , κ 1 = k , p(r)=\frac{r}{n}\prod_{l\geq1}\left(1-\frac{r^2}{\kappa_l^2}\right),\qquad\kappa_l=j_{\nu+1,l},\qquad\kappa_1=k, p ( r ) = n r l ≥ 1 ∏ ( 1 − κ l 2 r 2 ) , κ l = j ν + 1 , l , κ 1 = k ,
( log p ) r r = − r − 2 − ∑ l ≥ 1 2 ( κ l 2 + r 2 ) ( κ l 2 − r 2 ) 2 < 0. (\log p)_{rr}=-r^{-2}-\sum_{l\geq1}\frac{2(\kappa_l^2+r^2)}{(\kappa_l^2-r^2)^2}<0. ( log p ) r r = − r − 2 − l ≥ 1 ∑ ( κ l 2 − r 2 ) 2 2 ( κ l 2 + r 2 ) < 0.
If p r ≥ 0 p_r\geq0 p r ≥ 0 , then S > 0 S>0 S > 0 . Otherwise A = − p r / p > 0 A=-p_r/p>0 A = − p r / p > 0 and strict logarithmic concavity gives p ( v ) < p ( r ) e − A ( r ) ( v − r ) p(v)<p(r)e^{-A(r)(v-r)} p ( v ) < p ( r ) e − A ( r ) ( v − r ) for v > r v>r v > r . Hence u = ∫ r k p ( v ) d v < p / A u=\int_r^k p(v)\,\mathrm{d}v<p/A u = ∫ r k p ( v ) d v < p / A and S = p 2 − A u p > 0 S=p^2-Aup>0 S = p 2 − A u p > 0 . Thus t r = − S / p 2 < 0 t_r=-S/p^2<0 t r = − S / p 2 < 0 . The endpoint limits t ∼ n σ / r t\sim n\sigma/r t ∼ nσ / r and t ∼ ( k − r ) / 2 t\sim(k-r)/2 t ∼ ( k − r ) /2 prove the assertion.
We use s = t ( r ) s=t(r) s = t ( r ) as independent calibration variable. Subscripts r r r denote radial derivatives, and subscripts s s s derivatives after this change.
The corrections below are chosen so that the radial ball Hessian lies in the kernel of the energy difference. This determines the relations proved in Proposition 2.3. Define, as functions of r r r and equivalently of s = t ( r ) s=t(r) s = t ( r ) ,
A = − p r p , C = − 1 r , f = ϕ p , z = f s , g = f − s z , w = ϕ r p . A=-\frac{p_r}{p},\qquad C=-\frac{1}{r},\qquad f=\frac{\phi}{p},\qquad z=f_s,\qquad g=f-sz,\qquad w=\frac{\phi_r}{p}. A = − p p r , C = − r 1 , f = p ϕ , z = f s , g = f − sz , w = p ϕ r .
For the derivation it is useful to retain m = n − 1 m=n-1 m = n − 1 in the radial equations; 1 ≤ m ≤ 46 1\le m\le46 1 ≤ m ≤ 46 will be used in the main argument. The corresponding radial equations give
t r = t A − 1 , C r = C 2 , A r = A 2 + m A C + 1 − m C 2 (3) t_r=tA-1,\qquad C_r=C^2,\qquad A_r=A^2+mAC+1-mC^2 \tag*{(3)} t r = t A − 1 , C r = C 2 , A r = A 2 + m A C + 1 − m C 2 (3)
f r = w + A f = t r z , w r = ( A + m C ) w − λ f , w = − z − A g . (4) f_r=w+Af=t_rz,\qquad w_r=(A+mC)w-\lambda f,\qquad w=-z-Ag. \tag*{(4)} f r = w + A f = t r z , w r = ( A + m C ) w − λ f , w = − z − A g . (4)
The last identity follows by writing f = g + s z f=g+sz f = g + sz and t r = s A − 1 t_r=sA-1 t r = s A − 1 .
Introduce the scalar V 0 V_0 V 0 and the two correction profiles
V 0 = w 2 − m C f w + λ f 2 , G = V 0 − K A r m C 2 , b = − K A − m G C − f w . (5) V_0=w^2-mCfw+\lambda f^2,\qquad G=\frac{V_0-KA_r}{mC^2},\qquad\mathfrak{b}=-KA-mGC-fw. \tag*{(5)} V 0 = w 2 − m C f w + λ f 2 , G = m C 2 V 0 − K A r , b = − K A − m GC − f w . (5)
Here V 0 V_0 V 0 is a scalar, distinct from the vector-valued perspective map used later. Since C ≠ 0 C\ne0 C = 0 , these formulas define smooth profiles in the interior, without an additional differential-equation existence argument.
The differential identities below provide the cancellation needed in the divergence correction.
Lemma 2.2. The profiles satisfy
G r = b + 2 { A + ( m − 1 ) C } G + 2 K C , (6) G_r=\mathfrak{b}+2\{A+(m-1)C\}G+2KC, \tag*{(6)} G r = b + 2 { A + ( m − 1 ) C } G + 2 K C , (6) b r + m C G r = − 2 f r w . (7) \mathfrak{b}_r+mCG_r=-2f_rw. \tag*{(7)} b r + m C G r = − 2 f r w . (7) Proof. Differentiate using (3) –(4) . The terms group as
( V 0 ) r = ( 2 A + m C ) V 0 − m C 2 f w , (V_0)_r=(2A+mC)V_0-mC^2fw, ( V 0 ) r = ( 2 A + m C ) V 0 − m C 2 f w , A r r = ( 2 A + m C ) A r + m A C 2 − 2 m C 3 . A_{rr}=(2A+mC)A_r+mAC^2-2mC^3. A r r = ( 2 A + m C ) A r + m A C 2 − 2 m C 3 . Differentiate V 0 = K A r + m G C 2 V_0=KA_r+mGC^2 V 0 = K A r + m G C 2 and cancel the common K ( 2 A + m C ) A r K(2A+mC)A_r K ( 2 A + m C ) A r term. Division by m C 2 mC^2 m C 2 yields
G r = ( 2 A + ( m − 2 ) C ) G − K A + 2 K C − f w , G_r=(2A+(m-2)C)G-KA+2KC-fw, G r = ( 2 A + ( m − 2 ) C ) G − K A + 2 K C − f w , which is (6) after substituting b \mathfrak{b} b . Also f r w − f w r = V 0 f_rw-fw_r=V_0 f r w − f w r = V 0 . Differentiating the definition of b \mathfrak{b} b therefore gives
b r + m C G r = − K A r − m G C 2 − f r w − f w r = − V 0 − f r w − f w r = − 2 f r w . \mathfrak{b}_r+mCG_r=-KA_r-mGC^2-f_rw-fw_r=-V_0-f_rw-fw_r=-2f_rw. b r + m C G r = − K A r − m G C 2 − f r w − f w r = − V 0 − f r w − f w r = − 2 f r w .
Set
a = K − g 2 , d = 2 G − s G s , E = K + ( m − 1 ) G m , (8) a=K-g^2,\qquad d=2G-sG_s,\qquad E=\frac{K+(m-1)G}{m}, \tag*{(8)} a = K − g 2 , d = 2 G − s G s , E = m K + ( m − 1 ) G , (8)
L A = 2 b − s b s − 2 z g , L C = b + G s , h = − K − z 2 + λ f 2 + b s . L_A=2\mathfrak{b}-s\mathfrak{b}_s-2zg,\qquad L_C=\mathfrak{b}+G_s,\qquad h=-K-z^2+\lambda f^2+\mathfrak{b}_s. L A = 2 b − s b s − 2 z g , L C = b + G s , h = − K − z 2 + λ f 2 + b s .
The symmetric matrix
M m = ( a d / 2 L A / 2 d / 2 E L C / 2 L A / 2 L C / 2 h ) M_m=
\begin{pmatrix}
a & d/2 & L_A/2\\
d/2 & E & L_C/2\\
L_A/2 & L_C/2 & h
\end{pmatrix} M m = a d /2 L A /2 d /2 E L C /2 L A /2 L C /2 h
will act on a radial Hessian entry, a tangential trace, and a gradient magnitude.
The next identities determine its kernel exactly.
Proposition 2.3. For 0 < r < k 0<r<k 0 < r < k ,
L A = − 2 a A − d m C , L C = − d A − 2 E m C , h = a A 2 + d m A C + E m 2 C 2 . (9) L_A=-2aA-dmC,\qquad L_C=-dA-2EmC,\qquad h=aA^2+dmAC+Em^2C^2. \tag*{(9)} L A = − 2 a A − d m C , L C = − d A − 2 E m C , h = a A 2 + d m A C + E m 2 C 2 . (9) Thus M m ( A , m C , 1 ) T = 0 M_m(A,mC,1)^{\mathsf{T}}=0 M m ( A , m C , 1 ) T = 0 , and
( X , τ , P ) M m ( X , τ , P ) T = a ( X − A P ) 2 + d ( X − A P ) ( τ − m C P ) + E ( τ − m C P ) 2 . (X,\tau,P)M_m(X,\tau,P)^{\mathsf{T}}=a(X-AP)^2+d(X-AP)(\tau-mCP)+E(\tau-mCP)^2. ( X , τ , P ) M m ( X , τ , P ) T = a ( X − A P ) 2 + d ( X − A P ) ( τ − m C P ) + E ( τ − m C P ) 2 .
Proof. For the second equation in (9) , use ( 1 − s A ) G s = − G r (1-sA)G_s=-G_r ( 1 − s A ) G s = − G r :
L C + d A + 2 E m C = b + ( 1 − s A ) G s + 2 A G + 2 { K + ( m − 1 ) G } C = b − G r + 2 A G + 2 { K + ( m − 1 ) G } C = 0. \begin{aligned}
L_C+dA+2EmC&=b+(1-sA)G_s+2AG+2\{K+(m-1)G\}C\\
&=b-G_r+2AG+2\{K+(m-1)G\}C=0.
\end{aligned} L C + d A + 2 E m C = b + ( 1 − s A ) G s + 2 A G + 2 { K + ( m − 1 ) G } C = b − G r + 2 A G + 2 { K + ( m − 1 ) G } C = 0.
Dividing (7) by t r t_r t r gives b s + m C G s = − 2 z w b_s+mCG_s=-2zw b s + m C G s = − 2 z w . Consequently
L A + 2 a A + d m C = 2 ( b + K A + m G C ) − s ( b s + m C G s ) − 2 g ( z + A g ) = − 2 f w + 2 s z w + 2 g w = 0 , \begin{aligned}
L_A+2aA+dmC&=2(b+KA+mGC)-s(b_s+mCG_s)-2g(z+Ag)\\
&=-2fw+2szw+2gw=0,
\end{aligned} L A + 2 a A + d m C = 2 ( b + K A + m GC ) − s ( b s + m C G s ) − 2 g ( z + A g ) = − 2 f w + 2 sz w + 2 g w = 0 ,
using f = g + s z f=g+sz f = g + sz and w = − z − A g w=-z-Ag w = − z − A g .
For the last equation, the identities b s = − 2 z w − m C G s b_s=-2zw-mCG_s b s = − 2 z w − m C G s and − z 2 − 2 z w = w 2 − A 2 g 2 -z^2-2zw=w^2-A^2g^2 − z 2 − 2 z w = w 2 − A 2 g 2 reduce the difference to
h − a A 2 − d m A C − E m 2 C 2 = w 2 + λ f 2 − K ( 1 + A 2 + m C 2 ) − 2 m A C G − m ( m − 1 ) G C 2 + m C G r . \begin{aligned}
h-aA^2-dmAC-Em^2C^2
&=w^2+\lambda f^2-K(1+A^2+mC^2)\\
&\quad-2mACG-m(m-1)GC^2+mCG_r.
\end{aligned} h − a A 2 − d m A C − E m 2 C 2 = w 2 + λ f 2 − K ( 1 + A 2 + m C 2 ) − 2 m A C G − m ( m − 1 ) G C 2 + m C G r .
Substitute (6) , followed by b = − K A − m G C − f w b=-KA-mGC-fw b = − K A − m GC − f w . The result is
V 0 − K ( A 2 + m A C + 1 − m C 2 ) − m G C 2 = V 0 − K A r − m G C 2 = 0. V_0-K(A^2+mAC+1-mC^2)-mGC^2=V_0-KA_r-mGC^2=0. V 0 − K ( A 2 + m A C + 1 − m C 2 ) − m G C 2 = V 0 − K A r − m G C 2 = 0.
This proves all three equations and the completed-square expression.
We record both endpoint expansions, since they will justify the global Lipschitz extension. These are convergent power series after removal of the displayed leading factors. Their differentiated versions are therefore valid. At the radial center,
p = r n − r 3 2 n ( n + 2 ) + O ( r 5 ) , u = σ − r 2 2 n + O ( r 4 ) , t = n σ r + ( n σ 2 ( n + 2 ) − 1 2 ) r + O ( r 3 ) , f = n r + ( n 2 ( n + 2 ) − λ 2 ) r + O ( r 3 ) , w = − λ + O ( r 2 ) , A = − r − 1 + O ( r ) , G = G 0 + O ( r 2 ) , G 0 = n λ − K m , b = 2 m λ r + O ( r ) . \begin{aligned}
p&=\frac{r}{n}-\frac{r^3}{2n(n+2)}+O(r^5),&
u&=\sigma-\frac{r^2}{2n}+O(r^4),\\
t&=\frac{n\sigma}{r}+\left(\frac{n\sigma}{2(n+2)}-\frac12\right)r+O(r^3),&
f&=\frac{n}{r}+\left(\frac{n}{2(n+2)}-\frac{\lambda}{2}\right)r+O(r^3),\\
w&=-\lambda+O(r^2),&
A&=-r^{-1}+O(r),\\
G&=G_0+O(r^2),&
G_0&=\frac{n\lambda-K}{m},\qquad b=\frac{2m\lambda}{r}+O(r).
\end{aligned} p t w G = n r − 2 n ( n + 2 ) r 3 + O ( r 5 ) , = r nσ + ( 2 ( n + 2 ) nσ − 2 1 ) r + O ( r 3 ) , = − λ + O ( r 2 ) , = G 0 + O ( r 2 ) , u f A G 0 = σ − 2 n r 2 + O ( r 4 ) , = r n + ( 2 ( n + 2 ) n − 2 λ ) r + O ( r 3 ) , = − r − 1 + O ( r ) , = m nλ − K , b = r 2 mλ + O ( r ) .
In particular 1 / t 1/t 1/ t is analytic with a nonzero derivative at zero. Its analytic inverse gives, as s → ∞ s \to\infty s → ∞ ,
f = s / σ + O ( s − 1 ) , f s = σ − 1 + O ( s − 2 ) , g = O ( s − 1 ) , G = G 0 + O ( s − 2 ) , G s = O ( s − 3 ) , b = ( 2 λ / σ ) s + O ( s − 1 ) , b s = 2 λ / σ + O ( s − 2 ) , b − s b s = O ( s − 1 ) . (10) \begin{aligned}
f &= s/\sigma+ O\left(s^{-1}\right), & f_s &= \sigma^{-1} + O\left(s^{-2}\right), & g &= O\left(s^{-1}\right),\\
G &= G_0 + O\left(s^{-2}\right), & G_s &= O\left(s^{-3}\right),\\
\mathbf{b} &= (2\lambda/\sigma)s + O\left(s^{-1}\right), & \mathbf{b}_s &= 2\lambda/\sigma+ O\left(s^{-2}\right), & \mathbf{b} - s\mathbf{b}_s &= O\left(s^{-1}\right). \tag*{(10)}
\end{aligned} f G b = s / σ + O ( s − 1 ) , = G 0 + O ( s − 2 ) , = ( 2 λ / σ ) s + O ( s − 1 ) , f s G s b s = σ − 1 + O ( s − 2 ) , = O ( s − 3 ) , = 2 λ / σ + O ( s − 2 ) , g b − s b s = O ( s − 1 ) , = O ( s − 1 ) . (10)
For the outer endpoint let δ = k − r \delta= k-r δ = k − r and temporarily keep m m m in the formulas (with m = n − 1 m=n-1 m = n − 1 ). The radial equations give
p = c δ ( 1 + a 0 δ + b 0 δ 2 + O ( δ 3 ) ) , ϕ = β c δ ( 1 + a 0 δ + d 0 δ 2 + O ( δ 3 ) ) , (11) \begin{aligned}
p &= c\delta\left(1+a_0\delta+b_0\delta^2+O\left(\delta^3\right)\right),\\
\phi&= \beta c\delta\left(1+a_0\delta+d_0\delta^2+O\left(\delta^3\right)\right), \tag*{(11)}
\end{aligned} p ϕ = cδ ( 1 + a 0 δ + b 0 δ 2 + O ( δ 3 ) ) , = β cδ ( 1 + a 0 δ + d 0 δ 2 + O ( δ 3 ) ) , (11)
a 0 = m 2 k , b 0 = m ( m + 2 ) / k 2 − 1 6 , d 0 = m ( m + 1 ) / k 2 − λ 6 . a_0=\frac{m}{2k},\qquad b_0=\frac{m(m+2)/k^2-1}{6},\qquad d_0=\frac{m(m+1)/k^2-\lambda}{6}. a 0 = 2 k m , b 0 = 6 m ( m + 2 ) / k 2 − 1 , d 0 = 6 m ( m + 1 ) / k 2 − λ .
Also u = c δ 2 / 2 + c a 0 δ 3 / 3 + O ( δ 4 ) u=c\delta^2/2+ca_0\delta^3/3+O\left(\delta^4\right) u = c δ 2 /2 + c a 0 δ 3 /3 + O ( δ 4 ) and t = δ / 2 + O ( δ 2 ) t=\delta/2+O\left(\delta^2\right) t = δ /2 + O ( δ 2 ) . With h b = 1 − λ − m / k 2 h_b=1-\lambda-m/k^2 h b = 1 − λ − m / k 2 ,
f = β + β h b 6 δ 2 + O ( δ 3 ) , z = 2 β h b 3 δ + O ( δ 2 ) , f=\beta+\frac{\beta h_b}{6}\delta^2+O\left(\delta^3\right),\qquad z=\frac{2\beta h_b}{3}\delta+O\left(\delta^2\right), f = β + 6 β h b δ 2 + O ( δ 3 ) , z = 3 2 β h b δ + O ( δ 2 ) ,
g = β − β h b 6 δ 2 + O ( δ 3 ) . g=\beta-\frac{\beta h_b}{6}\delta^2+O\left(\delta^3\right). g = β − 6 β h b δ 2 + O ( δ 3 ) .
To check the removable singularity of G G G , use
A = δ − 1 + a 0 + ( 2 b 0 − a 0 2 ) δ + O ( δ 2 ) , w = − β { δ − 1 + a 0 + ( 3 d 0 − a 0 2 − b 0 ) δ + O ( δ 2 ) } . A=\delta^{-1}+a_0+\left(2b_0-a_0^2\right)\delta+O\left(\delta^2\right),\qquad
w=-\beta\left\{\delta^{-1}+a_0+\left(3d_0-a_0^2-b_0\right)\delta+O\left(\delta^2\right)\right\}. A = δ − 1 + a 0 + ( 2 b 0 − a 0 2 ) δ + O ( δ 2 ) , w = − β { δ − 1 + a 0 + ( 3 d 0 − a 0 2 − b 0 ) δ + O ( δ 2 ) } .
In V 0 − K A r V_0-KA_r V 0 − K A r the two pole coefficients vanish; its constant is K { 6 d 0 − m ( m + 1 ) / k 2 + λ } = 0 K\{6d_0-m(m+1)/k^2+\lambda\}=0 K { 6 d 0 − m ( m + 1 ) / k 2 + λ } = 0 . Thus G = O ( δ ) G=O(\delta) G = O ( δ ) analytically. The pole and constant of − K A − f w -KA-fw − K A − f w cancel too, so b = O ( δ ) \mathbf{b}=O(\delta) b = O ( δ ) . Substitution into the correction ODE then gives
G = 2 K 3 k δ + O ( δ 2 ) , d = 2 K 3 k δ + O ( δ 2 ) , b = 2 K ( 1 − λ ) 3 δ + O ( δ 2 ) . G=\frac{2K}{3k}\delta+O\left(\delta^2\right),\qquad
d=\frac{2K}{3k}\delta+O\left(\delta^2\right),\qquad
\mathbf{b}=\frac{2K(1-\lambda)}{3}\delta+O\left(\delta^2\right). G = 3 k 2 K δ + O ( δ 2 ) , d = 3 k 2 K δ + O ( δ 2 ) , b = 3 2 K ( 1 − λ ) δ + O ( δ 2 ) .
The equality of the linear factors in (11) ensures f s ( 0 ) = 0 f_s(0)=0 f s ( 0 ) = 0 . Extend f , G f,G f , G evenly and b \mathbf{b} b oddly to the whole s s s axis. These extensions are locally Lipschitz; the even extension of G G G may have a corner at zero. With derivatives understood almost everywhere, the endpoint expansions and interior smoothness give the global bounds needed in the weak argument:
f s , f − s f s , b s , b − s b s , G , G s , s G s f_s,\quad f-sf_s,\quad\mathbf{b}_s,\quad\mathbf{b}-s\mathbf{b}_s,\quad G,\quad G_s,\quad sG_s f s , f − s f s , b s , b − s b s , G , G s , s G s
are essentially bounded on R \mathbb{R} R . (2.10)
Positivity of the calibration matrix
We collect the analytic inputs used by the matrix proof. All quantities come from the radial profiles already defined, and S = p 2 + u p r > 0 S=p^2+up_r>0 S = p 2 + u p r > 0 . Write x = r 2 x=r^2 x = r 2 , b = k 2 b=k^2 b = k 2 , y = x / b y=x/b y = x / b and
F ( x ) = r ϕ ( r ) p ( r ) , T ( x ) = r u ( r ) p ( r ) , C = u p S , L = 1 − y 1 + y . F(x)=\frac{r\phi(r)}{p(r)},\qquad
T(x)=\frac{ru(r)}{p(r)},\qquad
C=\frac{up}{S},\qquad
L=\frac{1-y}{1+y}. F ( x ) = p ( r ) r ϕ ( r ) , T ( x ) = p ( r ) r u ( r ) , C = S u p , L = 1 + y 1 − y .
The use of T T T here is unrelated to the tangential Hessian block in the later pointwise identity.
The following scalar properties supply the inputs for the matrix argument.
Proposition 2.4. For every 2 ≤ n ≤ 47 2 \le n \le47 2 ≤ n ≤ 47 , the following hold on 0 < r < k 0 < r < k 0 < r < k .
The correction has the expansion with positive weights
G = ∑ j ≥ 1 B j ( 1 − y j ) , B j > 0 , ∑ j ≥ 1 j B j = K / 3. G = \sum_{j \ge1} B_j \left(1-y^j\right), \qquad B_j > 0, \qquad\sum_{j \ge1} jB_j = K/3. G = j ≥ 1 ∑ B j ( 1 − y j ) , B j > 0 , j ≥ 1 ∑ j B j = K /3. In particular G r < 0 G_r < 0 G r < 0 and 0 < G < K 0 < G < K 0 < G < K .
F F F has positive nonconstant Taylor coefficients. The function T T T is positive, decreasing and strictly concave as a function of x x x , has negative nonconstant Taylor coefficients, and satisfies T ( b ) = 0 T(b) = 0 T ( b ) = 0 .
The radial multiplier satisfies C / r > L C/r > L C / r > L , and
0 < d = 2 G + C G r < 2 K 3 L . 0 < d = 2G + CG_r < \frac{2K}{3}L. 0 < d = 2 G + C G r < 3 2 K L .
Proof. In dimension two, use (A.6) and the endpoint slope in Appendix § 6 . In dimensions n ≥ 3 n \ge3 n ≥ 3 , combine Lemma 3.1 with Theorems 3.12 and 3.15 in § 3 , use G ( k ) = 0 G(k) = 0 G ( k ) = 0 , and put B j = a j k 2 j B_j = a_j k^{2j} B j = a j k 2 j . The endpoint slope gives the sum. The product formula in § 4.1 and Proposition 4.1 prove the second assertion. Lemma 4.2 , Corollary 4.3 and Corollary 4.8 there prove the third assertion. These scalar proofs do not use matrix positivity.
We now use the positive Taylor coefficients of F = r ϕ / p F = r\phi/p F = r ϕ / p , the positive weights in the expansion of G G G , and the negative nonconstant Taylor coefficients of T = r t T = rt T = r t . Put m = n − 1 m = n - 1 m = n − 1 , b = k 2 b = k^2 b = k 2 , y = r 2 / b y = r^2/b y = r 2 / b , h = 1 − y h = 1-y h = 1 − y , and
L = 1 − y 1 + y , C = u p / S , a = K − g 2 , E = K + ( m − 1 ) G m , d = 2 G + C G r . L = \frac{1-y}{1+y}, \qquad C = up/S, \qquad a = K-g^2, \qquad E = \frac{K+(m-1)G}{m}, \qquad d = 2G + CG_r. L = 1 + y 1 − y , C = u p / S , a = K − g 2 , E = m K + ( m − 1 ) G , d = 2 G + C G r .
For 2 ≤ n ≤ 47 2 \le n \le47 2 ≤ n ≤ 47 , set A ( y ) = F ( b y ) / ( K k ) \mathcal{A}(y) = F(by)/(\sqrt{K}\,k) A ( y ) = F ( b y ) / ( K k ) . We use this normalized profile throughout the subsection.
We first estimate the radial energy by a rational function of the normalized radius.
Lemma 2.5. For every 0 < y < 1 0 < y < 1 0 < y < 1 ,
g K < 2 y 1 + y , a K > L 2 . \frac{g}{\sqrt{K}} < \frac{2\sqrt{y}}{1+y}, \qquad\frac{a}{K} > L^2. K g < 1 + y 2 y , K a > L 2 .
Proof. Put T ( y ) = T ( b y ) T(y) = T(by) T ( y ) = T ( b y ) and D = − T ′ / T D = -T'/T D = − T ′ / T . Endpoint regularity gives A ( 1 ) = 1 \mathcal{A}(1) = 1 A ( 1 ) = 1 and A ′ ( 1 ) = 1 / 2 \mathcal{A}'(1) = 1/2 A ′ ( 1 ) = 1/2 . The positivity of all derivatives of F F F implies convexity of A \mathcal{A} A . Consequently
A ′ ( y ) ≤ 1 − A ( y ) 1 − y , A ( y ) ≥ 1 + y 2 . \mathcal{A}'(y) \le\frac{1-\mathcal{A}(y)}{1-y}, \qquad\mathcal{A}(y) \ge\frac{1+y}{2}. A ′ ( y ) ≤ 1 − y 1 − A ( y ) , A ( y ) ≥ 2 1 + y .
The strictly negative nonconstant coefficients of T \mathcal{T} T imply that T \mathcal{T} T is decreasing and strictly concave in y y y . Since T ( 1 ) = T ( b ) = 0 \mathcal{T}(1)=\mathcal{T}(b)=0 T ( 1 ) = T ( b ) = 0 ,
0 < D < 1 1 − y 0<D<\frac{1}{1-y} 0 < D < 1 − y 1
The exact formula g = 2 x ( T F x − F T x ) / ( T − 2 x T x ) g=2\sqrt{x}\,(\mathcal{T}F_x-F\mathcal{T}_x)/(\mathcal{T}-2x\mathcal{T}_x) g = 2 x ( T F x − F T x ) / ( T − 2 x T x ) gives
g K y = 2 ( A ′ + D A ) 1 + 2 y D ≤ 1 + ( 1 + y ) D 1 + 2 y D < 2 1 + y . \frac{g}{\sqrt{K}\sqrt{y}}=\frac{2(\mathcal{A}'+D\mathcal{A})}{1+2yD}\leq\frac{1+(1+y)D}{1+2yD}<\frac{2}{1+y}. K y g = 1 + 2 y D 2 ( A ′ + D A ) ≤ 1 + 2 y D 1 + ( 1 + y ) D < 1 + y 2 .
The same formula is strictly positive because A , A ′ , D > 0 \mathcal{A},\mathcal{A}',D>0 A , A ′ , D > 0 , so it may be squared. For the first inequality, substitute the secant bound for A ′ \mathcal{A}' A ′ and then the lower tangent bound for A \mathcal{A} A ; its coefficient D − 1 / ( 1 − y ) D-1/(1-y) D − 1/ ( 1 − y ) is negative. The last strict inequality is equivalent to ( 1 − y ) D < 1 (1-y)D<1 ( 1 − y ) D < 1 . Squaring proves the assertion for a a a .
This rational estimate already controls the coupled determinant through dimension ten.
Theorem 2.6. For every 2 ≤ n ≤ 10 2\leq n\leq10 2 ≤ n ≤ 10 , throughout 0 < r < k 0<r<k 0 < r < k ,
( K − g 2 ) E − d 2 / 4 > 0. (K-g^2)E-d^2/4>0. ( K − g 2 ) E − d 2 /4 > 0.
Proof. Proposition 2.4 gives 0 < d < ( 2 K / 3 ) L 0<d<(2K/3)L 0 < d < ( 2 K /3 ) L . Also E ≥ K / m E\geq K/m E ≥ K / m , with equality only when m = 1 m=1 m = 1 . Lemma 2.5 therefore yields
d 2 4 a E < m 9 ≤ 1. \frac{d^2}{4aE}<\frac{m}{9}\leq1. 4 a E d 2 < 9 m ≤ 1.
The strict bounds for a a a and d d d ensure strictness also at n = 2 n=2 n = 2 and n = 10 n=10 n = 10 .
To reach higher dimensions, we estimate the correction and its derivative together.
Lemma 2.7. Suppose G = ∑ j ≥ 1 B j ( 1 − y j ) G=\sum_{j\geq1}B_j(1-y^j) G = ∑ j ≥ 1 B j ( 1 − y j ) , B j > 0 B_j>0 B j > 0 , and G r ( k ) = − 2 K / ( 3 k ) G_r(k)=-2K/(3k) G r ( k ) = − 2 K / ( 3 k ) . If d ≥ 0 d\geq0 d ≥ 0 , then
d 2 4 a E ≤ m L 2 3 ( a / K ) { 3 + ( m − 1 ) L } . \frac{d^2}{4aE}\leq\frac{mL^2}{3(a/K)\{3+(m-1)L\}}. 4 a E d 2 ≤ 3 ( a / K ) { 3 + ( m − 1 ) L } m L 2 . Proof. The endpoint derivative gives ∑ j j B j = K / 3 \sum_j jB_j=K/3 ∑ j j B j = K /3 . Thus w j = 3 j B j / K w_j=3jB_j/K w j = 3 j B j / K are probability weights. Set
A j = 1 − y j j , c = C r > L , B j ∗ = A j − c y j . A_j=\frac{1-y^j}{j},\qquad c=\frac{C}{r}>L,\qquad B_j^*=A_j-cy^j. A j = j 1 − y j , c = r C > L , B j ∗ = A j − c y j . Then
G K = 1 3 ∑ w j A j , d K = 2 3 ∑ w j B j ∗ , E K = ∑ w j { 1 m + m − 1 3 m A j } . \frac{G}{K}=\frac{1}{3}\sum w_jA_j,\qquad\frac{d}{K}=\frac{2}{3}\sum w_jB_j^*,\qquad\frac{E}{K}=\sum w_j\left\{\frac{1}{m}+\frac{m-1}{3m}A_j\right\}. K G = 3 1 ∑ w j A j , K d = 3 2 ∑ w j B j ∗ , K E = ∑ w j { m 1 + 3 m m − 1 A j } . The geometric-sum inequality ( 1 + y ) ( 1 − y j ) ≤ j ( 1 − y ) ( 1 + y j ) (1+y)(1-y^j)\leq j(1-y)(1+y^j) ( 1 + y ) ( 1 − y j ) ≤ j ( 1 − y ) ( 1 + y j ) gives
( B j ∗ ) + ≤ min { A j , L } . \left(B_j^*\right)_+\leq\min\{A_j,L\}. ( B j ∗ ) + ≤ min { A j , L } .
Here ( t ) + = max { t , 0 } (t)_{+}=\max\{t,0\} ( t ) + = max { t , 0 } denotes the positive part. Writing e ( A ) = 1 / m + ( m − 1 ) A / ( 3 m ) e(A)=1/m+(m-1)A/(3m) e ( A ) = 1/ m + ( m − 1 ) A / ( 3 m ) , the function A 2 / e ( A ) A^2/e(A) A 2 / e ( A ) is increasing, whereas L 2 / e ( A ) L^2/e(A) L 2 / e ( A ) is decreasing. Hence ( B j ∗ ) + 2 / e ( A j ) ≤ L 2 / e ( L ) (B_j^*)_{+}^{2}/e(A_j)\leq L^2/e(L) ( B j ∗ ) + 2 / e ( A j ) ≤ L 2 / e ( L ) . Cauchy–Schwarz yields
( ∑ w j B j ∗ ) 2 ≤ ( ∑ w j ( B j ∗ ) + ) 2 ≤ ( ∑ w j e ( A j ) ) L 2 e ( L ) . \left(\sum w_jB_j^*\right)^2\leq\left(\sum w_j(B_j^*)_{+}\right)^2\leq\left(\sum w_je(A_j)\right)\frac{L^2}{e(L)}. ( ∑ w j B j ∗ ) 2 ≤ ( ∑ w j ( B j ∗ ) + ) 2 ≤ ( ∑ w j e ( A j ) ) e ( L ) L 2 .
The first inequality uses ∑ w j B j ∗ ≥ 0 \sum w_jB_j^*\geq0 ∑ w j B j ∗ ≥ 0 . Substitution proves the result. Convergence follows from ∑ j B j = K / 3 \sum jB_j=K/3 ∑ j B j = K /3 .
Positive exponential integrals give a stronger curvature estimate, with constants fixed by two exact endpoint moments.
Lemma 2.8. The function A ′ ′ \mathcal{A}'' A ′′ is strictly positive and logarithmically convex (that is, its logarithm is convex) on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Its endpoint moments are
M 2 = A ′ ′ ( 1 ) = k 2 − q 2 − m − 3 12 , M 3 = A ′ ′ ′ ( 1 ) = m + 3 8 . M_2=\mathcal{A}''(1)=\frac{k^2-q^2-m-3}{12},\qquad M_3=\mathcal{A}'''(1)=\frac{m+3}{8}. M 2 = A ′′ ( 1 ) = 12 k 2 − q 2 − m − 3 , M 3 = A ′′′ ( 1 ) = 8 m + 3 . In particular,
A ′ ′ ( y ) ≥ M 2 exp ( − M 3 M 2 ( 1 − y ) ) . \mathcal{A}''(y)\geq M_2\exp\left(-\frac{M_3}{M_2}(1-y)\right). A ′′ ( y ) ≥ M 2 exp ( − M 2 M 3 ( 1 − y ) ) . Proof. The zero-ratio product for F F F consists, after cancellation of the first zero, of factors with a > v > 1 a>v>1 a > v > 1 of the form
1 − y / a 1 − y / v = v a + ( 1 − v a ) v v − y . \frac{1-y/a}{1-y/v}=\frac{v}{a}+\left(1-\frac{v}{a}\right)\frac{v}{v-y}. 1 − y / v 1 − y / a = a v + ( 1 − a v ) v − y v . Indeed,
v v − y = ∫ 0 ∞ e y t v e − v t d t , y < v . \frac{v}{v-y}=\int_0^\infty e^{yt}ve^{-vt}\,dt,\qquad y<v. v − y v = ∫ 0 ∞ e y t v e − v t d t , y < v . Thus each factor is the exponential integral of a positive measure: a mixture of a unit point mass at zero and an exponential probability distribution of rate v v v . A finite product has the same form, with the product represented by convolution of the measures. If P P P is such a product, Cauchy–Schwarz applied to its second derivative gives
( P ( 3 ) ) 2 ≤ P ′ ′ P ( 4 ) . (P^{(3)})^2\leq P''P^{(4)}. ( P ( 3 ) ) 2 ≤ P ′′ P ( 4 ) . This is precisely the differential criterion for logarithmic convexity of P ′ ′ P'' P ′′ . The canonical products and their derivatives converge locally uniformly before the second pole. Passing to the limit therefore gives the same inequality for A ′ ′ \mathcal{A}'' A ′′ ; its strict positivity follows from the positive Taylor coefficients. No limiting measure is needed for this conclusion.
For an explicit check of the moments, set h ( x ) = p ( x ) / x h(x)=p(\sqrt{x})/\sqrt{x} h ( x ) = p ( x ) / x , and normalize H ( y ) = h ( b y ) H(y)=h(by) H ( y ) = h ( b y ) and Φ ( y ) = ϕ ( k 2 y ) \Phi(y)=\phi(k_2\sqrt{y}) Φ ( y ) = ϕ ( k 2 y ) by separate constant factors so that H ( 1 ) = Φ ( 1 ) = 0 H(1)=\Phi(1)=0 H ( 1 ) = Φ ( 1 ) = 0 and H ′ ( 1 ) = Φ ′ ( 1 ) = 1 H'(1)=\Phi'(1)=1 H ′ ( 1 ) = Φ ′ ( 1 ) = 1 . The harmless scaling makes Φ / H = A \Phi/H=\mathcal{A} Φ/ H = A . Their equations are
4 y H ′ ′ + 2 ( n + 2 ) H ′ + b H = 0 , 4 y Φ ′ ′ + 2 n Φ ′ + q 2 Φ = 0. 4yH''+2(n+2)H'+bH=0,\qquad4y\Phi''+2n\Phi'+q^2\Phi=0. 4 y H ′′ + 2 ( n + 2 ) H ′ + b H = 0 , 4 y Φ ′′ + 2 n Φ ′ + q 2 Φ = 0.
Writing their Taylor coefficients at y = 1 y=1 y = 1 as H j , Φ j H_j,\Phi_j H j , Φ j gives
H j + 2 = − ( j + 1 ) ( 4 j + 2 n + 4 ) H j + 1 + b H j 4 ( j + 1 ) ( j + 2 ) , H_{j+2}=-\frac{(j+1)(4j+2n+4)H_{j+1}+bH_j}{4(j+1)(j+2)}, H j + 2 = − 4 ( j + 1 ) ( j + 2 ) ( j + 1 ) ( 4 j + 2 n + 4 ) H j + 1 + b H j ,
and the same recurrence with 2 n + 4 , b 2n+4,b 2 n + 4 , b replaced by 2 n , q 2 2n,q^2 2 n , q 2 . Division through order three gives
A ( 1 + s ) = 1 + s 2 + k 2 − q 2 − m − 3 24 s 2 + m + 3 48 s 3 + O ( s 4 ) . A(1+s)=1+\frac{s}{2}+\frac{k^2-q^2-m-3}{24}s^2+\frac{m+3}{48}s^3+O\left(s^4\right). A ( 1 + s ) = 1 + 2 s + 24 k 2 − q 2 − m − 3 s 2 + 48 m + 3 s 3 + O ( s 4 ) .
This proves the exact moments. The tangent inequality for log A ′ ′ \log A'' log A ′′ at 1 proves the exponential lower bound.
A lower bound for this curvature strengthens the preceding rational energy estimate.
Lemma 2.9. For any 0 < C ≤ A ′ ′ ( y ) 0<C\le A''(y) 0 < C ≤ A ′′ ( y ) ,
g K y < 2 − C ( 1 − y ) 2 1 + y , a / K L 2 > 1 + 4 y C − y C 2 ( 1 − y ) 2 . \frac{g}{\sqrt{K}\sqrt{y}}<\frac{2-C(1-y)^2}{1+y},\qquad\frac{a/K}{L^2}>1+4yC-yC^2(1-y)^2. K y g < 1 + y 2 − C ( 1 − y ) 2 , L 2 a / K > 1 + 4 y C − y C 2 ( 1 − y ) 2 . The right side of the first inequality is positive.
Proof. The nonnegative Taylor coefficients make A ′ ′ A'' A ′′ increasing, so A ′ ′ ≥ C A''\ge C A ′′ ≥ C on [ y , 1 ] [y,1] [ y , 1 ] . Strong convexity gives
A ( y ) ≥ 1 − h 2 + C h 2 2 , A ′ ( y ) ≤ 1 − A ( y ) h − C h 2 . A(y)\ge1-\frac{h}{2}+\frac{Ch^2}{2},\qquad A'(y)\le\frac{1-A(y)}{h}-\frac{Ch}{2}. A ( y ) ≥ 1 − 2 h + 2 C h 2 , A ′ ( y ) ≤ h 1 − A ( y ) − 2 C h .
Substitute in the exact formula for g g g as in Lemma 2.5 . The result is
g K y ≤ 1 − 2 C h + ( 1 + y + C h 2 ) D 1 + 2 y D . \frac{g}{\sqrt{K}\sqrt{y}}\le\frac{1-2Ch+(1+y+Ch^2)D}{1+2yD}. K y g ≤ 1 + 2 y D 1 − 2 C h + ( 1 + y + C h 2 ) D .
This fractional-linear expression has positive derivative in D D D : its numerator after differentiation is h + C h ( h + 4 y ) > 0 h+Ch(h+4y)>0 h + C h ( h + 4 y ) > 0 . The strict bound D < 1 / h D<1/h D < 1/ h now proves the first assertion. Its right side is positive since g > 0 g>0 g > 0 . Squaring it and subtracting from one gives the second assertion.
In dimensions above ten, combine Lemma 2.9 with Lemma 2.7 and the scalar exponential comparisons proved in § 4.2 . The resulting determinant theorem is given in Theorem 2.10 .
The rational estimate in low dimensions and the Laplace estimate in the remaining range now give the determinant sign throughout the required dimension range. We then check the mixed and trace-free Hessian coefficients.
Theorem 2.10. For every integer 2 ≤ n ≤ 47 2\le n\le47 2 ≤ n ≤ 47 , on the full interval 0 < r < k 0<r<k 0 < r < k ,
d 2 < 4 E ( K − g 2 ) . d^2<4E\left(K-g^2\right). d 2 < 4 E ( K − g 2 ) .
Proof. For 2 ≤ n ≤ 10 2 \le n \le10 2 ≤ n ≤ 10 use Theorem 2.6 . For 11 ≤ n ≤ 47 11 \le n \le47 11 ≤ n ≤ 47 , put m = n − 1 m=n-1 m = n − 1 and h = 1 − r 2 / k 2 h=1-r^2/k^2 h = 1 − r 2 / k 2 . Corollary 4.8 gives d > 0 d>0 d > 0 . The rational energy bound gives a = K − g 2 > 0 a=K-g^2>0 a = K − g 2 > 0 . Choose M M M from the corresponding row of the following table:
Table 3.
The gap estimates in § 4.2 prove M 2 > M > 0 M_2>M>0 M 2 > M > 0 . Since M ↦ M exp ( − M 3 h / M ) M\mapsto M\exp(-M_3h/M) M ↦ M exp ( − M 3 h / M ) is increasing, Lemma 2.8 yields
A ′ ′ ( 1 − h ) > C : = M exp ( − m + 3 8 M h ) . \mathcal{A}''(1-h)>C:=M\exp\left(-\frac{m+3}{8M}h\right). A ′′ ( 1 − h ) > C := M exp ( − 8 M m + 3 h ) .
The strong energy estimate therefore gives
a / K L 2 > 1 + 4 ( 1 − h ) C − ( 1 − h ) C 2 h 2 , L = h 2 − h . \frac{a/K}{L^2}>1+4(1-h)C-(1-h)C^2h^2,\qquad L=\frac{h}{2-h}. L 2 a / K > 1 + 4 ( 1 − h ) C − ( 1 − h ) C 2 h 2 , L = 2 − h h .
The elementary exponential comparisons in § 4.2 , namely Lemmas 4.9 , 4.10 and 4.11, prove for these four ranges
1 + 4 ( 1 − h ) C − ( 1 − h ) C 2 h 2 > m 3 { 3 + ( m − 1 ) L } . 1+4(1-h)C-(1-h)C^2h^2>\frac{m}{3\{3+(m-1)L\}}. 1 + 4 ( 1 − h ) C − ( 1 − h ) C 2 h 2 > 3 { 3 + ( m − 1 ) L } m .
Substitute this strict inequality in Lemma 2.7 . It gives d 2 / ( 4 a E ) < 1 d^2/(4aE)<1 d 2 / ( 4 a E ) < 1 , as required. □
The determinant bound also yields the signs needed for the mixed and trace-free Hessian terms.
Corollary 2.11. For every 2 ≤ n ≤ 47 2\le n\le47 2 ≤ n ≤ 47 and 0 < r < k 0<r<k 0 < r < k ,
a = K − g 2 > 0 , E > 0 , a E − d 2 / 4 > 0 , 2 K − g 2 − d > 0 , K − G > 0. a=K-g^2>0,\qquad E>0,\qquad aE-d^2/4>0,\qquad2K-g^2-d>0,\qquad K-G>0. a = K − g 2 > 0 , E > 0 , a E − d 2 /4 > 0 , 2 K − g 2 − d > 0 , K − G > 0. The null direction is given by (9) .
Proof. The determinant estimate has been proved. For the remaining signs, use the positive weights B j B_j B j in the expansion of G G G . The endpoint derivative gives ∑ j B j = K / 3 \sum jB_j=K/3 ∑ j B j = K /3 , so
0 < G < G 0 = ∑ B j < ∑ j B j = K / 3. 0<G<G_0=\sum B_j<\sum jB_j=K/3. 0 < G < G 0 = ∑ B j < ∑ j B j = K /3.
The strict inequality uses B 2 > 0 B_2>0 B 2 > 0 . Thus E > 0 E>0 E > 0 , K − G > 0 K-G>0 K − G > 0 , and d = 2 G + C G r < 2 G < 2 K / 3 d=2G+C G_r<2G<2K/3 d = 2 G + C G r < 2 G < 2 K /3 . The energy bound g 2 < K g^2<K g 2 < K now gives 2 K − g 2 − d > K / 3 > 0 2K-g^2-d>K/3>0 2 K − g 2 − d > K /3 > 0 . The kernel identities depend only on the defining radial equations. □
The estimates now control every Hessian component. We next turn this pointwise information into an inequality for arbitrary test functions.
The functional comparison
The profiles are now fixed in a dimension 2 ≤ n ≤ 47 2 \le n \le47 2 ≤ n ≤ 47 . We first compute the local energy identity, then justify it across the exceptional sets and pass to the variational infimum.
For the moment assume U > 0 U > 0 U > 0 and P = ∣ ∇ U ∣ > 0 P = |\nabla U| > 0 P = ∣∇ U ∣ > 0 at a point, and evaluate all profiles at s = U / P s = U/P s = U / P . In an orthonormal frame with first vector e = ∇ U / P e = \nabla U/P e = ∇ U / P , write
H = D 2 U = ( X ζ T ζ T ) , τ = tr T , T 0 = T − τ m I m . H = D^2U = \begin{pmatrix} X & \zeta^{\mathsf{T}} \\ \zeta& T \end{pmatrix}, \qquad\tau= \operatorname{tr} T, \qquad T_0 = T - \frac{\tau}{m}I_m. H = D 2 U = ( X ζ ζ T T ) , τ = tr T , T 0 = T − m τ I m .
Here I m I_m I m is the identity matrix, ζ ∈ R m \zeta\in\mathbb{R}^m ζ ∈ R m , and T 0 T_0 T 0 is the trace-free part of the tangential block T T T . For m = 1 m = 1 m = 1 , T 0 = 0 T_0 = 0 T 0 = 0 , so the same calculation includes dimension two. The symbol A : B = tr ( A T B ) A:B = \operatorname{tr}(A^{\mathsf{T}}B) A : B = tr ( A T B ) denotes the matrix inner product. Define T 1 ( H ) = ( tr H ) I n − H T_1(H) = (\operatorname{tr} H)I_n - H T 1 ( H ) = ( tr H ) I n − H , the first Newton tensor. For a smooth Hessian its divergence vanishes, since ∂ i ( Δ U δ i j − U i j ) = 0 \partial_i(\Delta U\delta_{ij} - U_{ij}) = 0 ∂ i ( Δ U δ ij − U ij ) = 0 (summed over i i i ). Consider
F = P f ( s ) , J 1 = P b ( s ) ∇ U , J 2 = G ( s ) T 1 ( H ) ∇ U . F = P f(s), \qquad J_1 = P b(s)\nabla U, \qquad J_2 = G(s)T_1(H)\nabla U. F = P f ( s ) , J 1 = P b ( s ) ∇ U , J 2 = G ( s ) T 1 ( H ) ∇ U .
The identities ∇ P = H e \nabla P = He ∇ P = H e and ∇ s = e − ( s / P ) H e \nabla s = e - (s/P)He ∇ s = e − ( s / P ) H e imply
∇ F = z P e + g H e , ∣ ∇ F ∣ 2 = ( z P + g X ) 2 + g 2 ∣ ζ ∣ 2 . (12) \nabla F = zP e + gHe, \qquad|\nabla F|^2 = (zP + gX)^2 + g^2|\zeta|^2. \tag*{(12)} ∇ F = z P e + g H e , ∣∇ F ∣ 2 = ( z P + g X ) 2 + g 2 ∣ ζ ∣ 2 . (12)
Direct differentiation of the first vector field gives
div J 1 = b s P 2 + ( 2 b − s b s ) P X + b P τ . \operatorname{div} J_1 = b_sP^2 + (2b - sb_s)PX + bP\tau. div J 1 = b s P 2 + ( 2 b − s b s ) P X + b P τ .
For the second, use
e T T 1 ( H ) e = τ , e T H T 1 ( H ) e = X τ − ∣ ζ ∣ 2 , e^{\mathsf{T}}T_1(H)e = \tau, \qquad e^{\mathsf{T}}HT_1(H)e = X\tau- |\zeta|^2, e T T 1 ( H ) e = τ , e T H T 1 ( H ) e = X τ − ∣ ζ ∣ 2 ,
and T 1 ( H ) : H = 2 X τ + τ 2 − ∣ T ∣ 2 − 2 ∣ ζ ∣ 2 T_1(H):H = 2X\tau+ \tau^2 - |T|^2 - 2|\zeta|^2 T 1 ( H ) : H = 2 X τ + τ 2 − ∣ T ∣ 2 − 2∣ ζ ∣ 2 . The result is
div J 2 = d X τ + G ( τ 2 − ∣ T ∣ 2 ) − d ∣ ζ ∣ 2 + G s P τ . (13) \operatorname{div} J_2 = dX\tau+ G(\tau^2 - |T|^2) - d|\zeta|^2 + G_sP\tau. \tag*{(13)} div J 2 = d X τ + G ( τ 2 − ∣ T ∣ 2 ) − d ∣ ζ ∣ 2 + G s P τ . (13)
Using ∣ T ∣ 2 = ∣ T 0 ∣ 2 + τ 2 / m |T|^2 = |T_0|^2 + \tau^2/m ∣ T ∣ 2 = ∣ T 0 ∣ 2 + τ 2 / m , combine (12) –(13) to obtain
K ( ∣ H ∣ 2 − P 2 ) − ( ∣ ∇ F ∣ 2 − λ F 2 ) + div J 1 + div J 2 = a X 2 + d X τ + E τ 2 + L A P X + L C P τ + h P 2 + ( 2 K − g 2 − d ) ∣ ζ ∣ 2 + ( K − G ) ∣ T 0 ∣ 2 . (14) \begin{aligned}
K\left(|H|^2-P^2\right)-\left(|\nabla F|^2-\lambda F^2\right)+\operatorname{div}J_1+\operatorname{div}J_2
&= aX^2+dX\tau+E\tau^2+L_APX+L_CP\tau+hP^2 \\
&\quad+(2K-g^2-d)|\zeta|^2+(K-G)|T_0|^2. \tag*{(14)}
\end{aligned} K ( ∣ H ∣ 2 − P 2 ) − ( ∣∇ F ∣ 2 − λ F 2 ) + div J 1 + div J 2 = a X 2 + d X τ + E τ 2 + L A P X + L C P τ + h P 2 + ( 2 K − g 2 − d ) ∣ ζ ∣ 2 + ( K − G ) ∣ T 0 ∣ 2 . (14)
All coefficients are those of (8) , with m = n − 1 m = n - 1 m = n − 1 . By Proposition 2.3, the right side equals
D = a ( X − A P ) 2 + d ( X − A P ) ( τ − m C P ) + E ( τ − m C P ) 2 + ( 2 K − g 2 − d ) ∣ ζ ∣ 2 + ( K − G ) ∣ T 0 ∣ 2 . (15) \begin{aligned}
\mathcal{D}
&= a(X-AP)^2+d(X-AP)(\tau-mCP)+E(\tau-mCP)^2 \\
&\quad+(2K-g^2-d)|\zeta|^2+(K-G)|T_0|^2. \tag*{(15)}
\end{aligned} D = a ( X − A P ) 2 + d ( X − A P ) ( τ − m C P ) + E ( τ − m C P ) 2 + ( 2 K − g 2 − d ) ∣ ζ ∣ 2 + ( K − G ) ∣ T 0 ∣ 2 . (15)
Corollary 2.11 shows D ≥ 0 \mathcal{D} \ge0 D ≥ 0 . The trace-free term in (15) is part of the n n n -dimensional identity and has been retained throughout the scalar reduction.
It remains to justify the fields globally, including P = 0 P = 0 P = 0 . Simply integrating the pointwise identity off that set would be insufficient: a smooth function can have a critical set of positive measure.
We begin by extending the perspective maps across v = 0 v = 0 v = 0 . For v ∈ R n v \in\mathbb{R}^{n} v ∈ R n , P = ∣ v ∣ > 0 P = \lvert v\rvert> 0 P = ∣ v ∣ > 0 , and y ∈ R y \in\mathbb{R} y ∈ R , define
Φ ( y , v ) = P f ( y / P ) , B ( y , v ) = P b ( y / P ) , V ( y , v ) = G ( y / P ) v . \Phi(y,v) = P f(y/P), \qquad B(y,v) = P b(y/P), \qquad\mathcal{V}(y,v) = G(y/P)v. Φ ( y , v ) = P f ( y / P ) , B ( y , v ) = P b ( y / P ) , V ( y , v ) = G ( y / P ) v .
Using the even/odd extensions from § 2.1 , set
Φ ( y , 0 ) = ∣ y ∣ / σ , B ( y , 0 ) = ( 2 λ / σ ) y , V ( y , 0 ) = 0. (16) \Phi(y,0) = \lvert y\rvert/\sigma, \qquad B(y,0) = (2\lambda/\sigma)y, \qquad\mathcal{V}(y,0) = 0. \tag*{(16)} Φ ( y , 0 ) = ∣ y ∣ / σ , B ( y , 0 ) = ( 2 λ / σ ) y , V ( y , 0 ) = 0. (16)
These are continuous extensions by (10) . Continuity at ( 0 , 0 ) (0,0) ( 0 , 0 ) follows from the linear growth bounds for f f f , b b b and boundedness of G G G .
Off v = 0 v = 0 v = 0 and the possible corner at y = 0 y = 0 y = 0 , differentiation gives
Φ y = f s , ∇ v Φ = ( f − s f s ) v P , B y = b s , ∇ v B = ( b − s b s ) v P , V y = G s v P , D v V = G I − s G s v v T P 2 . \begin{aligned}
\Phi_y &= f_s, & \nabla_v\Phi&= (f-sf_s)\frac{v}{P}, \\
B_y &= b_s, & \nabla_v B &= (b-sb_s)\frac{v}{P}, \\
\mathcal{V}_y &= G_s\frac{v}{P}, & D_v\mathcal{V} &= GI-sG_s\frac{vv^{\mathsf T}}{P^2}.
\end{aligned} Φ y B y V y = f s , = b s , = G s P v , ∇ v Φ ∇ v B D v V = ( f − s f s ) P v , = ( b − s b s ) P v , = G I − s G s P 2 v v T .
Equation (2.10) bounds these derivatives uniformly. To conclude global Lipschitz continuity, integrate along line segments: the equations v = 0 v = 0 v = 0 and y = 0 y = 0 y = 0 each cut such a segment in at most one point unless they hold on the entire segment. In the latter case the explicit restrictions (16) , or the values at y = 0 y = 0 y = 0 , are Lipschitz. Subdivision at the remaining exceptional points and continuity give the same uniform bound across the whole segment.
Now fix U ∈ C c ∞ ( Ω ) U \in C^\infty_c(\Omega) U ∈ C c ∞ ( Ω ) and define globally
F = Φ ( U , ∇ U ) , J 1 = B ( U , ∇ U ) ∇ U , J 2 = T 1 ( D 2 U ) V ( U , ∇ U ) . (17) F = \Phi(U,\nabla U), \qquad J_1 = B(U,\nabla U)\nabla U, \qquad J_2 = T_1(D^2U)\mathcal{V}(U,\nabla U). \tag*{(17)} F = Φ ( U , ∇ U ) , J 1 = B ( U , ∇ U ) ∇ U , J 2 = T 1 ( D 2 U ) V ( U , ∇ U ) . (17)
The pair ( U , ∇ U ) (U,\nabla U) ( U , ∇ U ) is globally Lipschitz. The smooth factor T 1 ( D 2 U ) T_1(D^2U) T 1 ( D 2 U ) and its derivatives are bounded on the support. Hence F F F , J 1 J_1 J 1 , J 2 J_2 J 2 are Lipschitz and compactly supported in Ω \Omega Ω . In particular F ∈ H 0 1 ( Ω ) F \in H^1_0(\Omega) F ∈ H 0 1 ( Ω ) : mollification with support smaller than the distance of supp U \operatorname{supp} U supp U to Ω c \Omega^c Ω c gives the required approximation. Also J 1 , J 2 ∈ W 1 , ∞ J_1,J_2 \in W^{1,\infty} J 1 , J 2 ∈ W 1 , ∞ , the space of bounded functions with bounded weak first derivatives, with compact support, so their weak divergences have zero integral.
The same fields also accommodate sign changes of U U U . Replacing U U U by − U -U − U changes both ∇ U \nabla U ∇ U and D 2 U D^2U D 2 U to their negatives. Evenness of f f f gives the same F F F . Oddness of b b b gives B ( − U , − ∇ U ) = − B ( U , ∇ U ) B(-U,-\nabla U)=-B(U,\nabla U) B ( − U , − ∇ U ) = − B ( U , ∇ U ) , so J 1 J_1 J 1 is unchanged. Evenness of G G G changes V \mathcal{V} V to its negative, and T 1 ( − H ) = − T 1 ( H ) T_1(-H)=-T_1(H) T 1 ( − H ) = − T 1 ( H ) , so J 2 J_2 J 2 is unchanged as well. Therefore the left side of (14) is invariant, and its nonnegativity on { U < 0 , P > 0 } \{U<0,P>0\} { U < 0 , P > 0 } follows by applying the positive calculation to − U -U − U . The set { U = 0 , ∇ U ≠ 0 } \{U=0,\nabla U\ne0\} { U = 0 , ∇ U = 0 } is covered by countably many smooth local hypersurfaces, by the implicit function theorem, and has measure zero. The fields themselves are globally Lipschitz across it, so no surface distribution or unaccounted nodal flux appears.
It remains to verify the inequality on the critical set. Let Z = { ∇ U = 0 } Z=\{\nabla U=0\} Z = { ∇ U = 0 } . We use Sobolev locality: two Sobolev functions that agree on a measurable set have equal weak gradients almost everywhere there. The chain rule, locality, and the approximation facts used here are standard; see [21 Chapter 2] .
Apply this fact to each component of ∇ U \nabla U ∇ U . It gives D 2 U = 0 D^2U=0 D 2 U = 0 almost everywhere on Z Z Z . Moreover J 1 = J 2 = 0 J_1=J_2=0 J 1 = J 2 = 0 pointwise on Z Z Z , by (17) and V ( U , 0 ) = 0 V(U,0)=0 V ( U , 0 ) = 0 . Since these fields are Lipschitz, every weak first derivative of their components vanishes almost everywhere on that common zero set. Thus div J 1 = div J 2 = 0 \operatorname{div}J_1=\operatorname{div}J_2=0 div J 1 = div J 2 = 0 almost everywhere on Z Z Z .
The Lipschitz function F − ∣ U ∣ / σ F-|U|/\sigma F − ∣ U ∣/ σ also vanishes on Z Z Z , so its weak gradient is zero there. Since ∇ ∣ U ∣ = 0 \nabla|U|=0 ∇∣ U ∣ = 0 almost everywhere on Z Z Z , it follows that ∇ F = 0 \nabla F=0 ∇ F = 0 there. The left side of (14) consequently reduces to
λ F 2 = λ U 2 / σ 2 ≥ 0 \lambda F^2=\lambda U^2/\sigma^2\ge0 λ F 2 = λ U 2 / σ 2 ≥ 0
almost everywhere on Z Z Z .
This proof does not differentiate a perspective map at its corner and does not assume the critical set has measure zero.
The preceding pointwise and regularity arguments establish, almost everywhere in Ω \Omega Ω ,
K ( ∣ D 2 U ∣ 2 − ∣ ∇ U ∣ 2 ) − ( ∣ ∇ F ∣ 2 − λ F 2 ) + div J 1 + div J 2 ≥ 0. K\left(|D^2U|^2-|\nabla U|^2\right)-\left(|\nabla F|^2-\lambda F^2\right)+\operatorname{div}J_1+\operatorname{div}J_2\ge0. K ( ∣ D 2 U ∣ 2 − ∣∇ U ∣ 2 ) − ( ∣∇ F ∣ 2 − λ F 2 ) + div J 1 + div J 2 ≥ 0.
Integrate and use the compact support of both vector fields. For compactly supported smooth U U U , integration by parts twice gives
∫ U i i U j j d x = ∫ U i j 2 d x , ∫ ( Δ U ) 2 d x = ∫ ∣ D 2 U ∣ 2 d x . \int U_{ii}U_{jj}\,\mathrm{d}x=\int U_{ij}^2\,\mathrm{d}x,\qquad\int(\Delta U)^2\,\mathrm{d}x=\int|D^2U|^2\,\mathrm{d}x. ∫ U ii U j j d x = ∫ U ij 2 d x , ∫ ( Δ U ) 2 d x = ∫ ∣ D 2 U ∣ 2 d x .
We obtain the functional inequality
∫ ∣ ∇ F ∣ 2 d x − λ ∫ F 2 d x ≤ K { ∫ ( Δ U ) 2 d x − ∫ ∣ ∇ U ∣ 2 d x } . (18) \int|\nabla F|^2\,\mathrm{d}x-\lambda\int F^2\,\mathrm{d}x\le K\left\{\int(\Delta U)^2\,\mathrm{d}x-\int|\nabla U|^2\,\mathrm{d}x\right\}. \tag*{(18)} ∫ ∣∇ F ∣ 2 d x − λ ∫ F 2 d x ≤ K { ∫ ( Δ U ) 2 d x − ∫ ∣∇ U ∣ 2 d x } . (18)
The function F F F is not zero when U U U is not zero. At a point with U ≠ 0 U\ne0 U = 0 , either P > 0 P>0 P > 0 , in which case F = P f ( ∣ U ∣ / P ) > 0 F=Pf(|U|/P)>0 F = P f ( ∣ U ∣/ P ) > 0 , or P = 0 P=0 P = 0 , in which case F = ∣ U ∣ / σ > 0 F=|U|/\sigma>0 F = ∣ U ∣/ σ > 0 . By continuity it is positive on a set of positive measure. It is therefore an admissible nonzero Dirichlet trial function.
For 0 ≠ U ∈ C c ∞ ( Ω ) 0\ne U\in C^\infty_c(\Omega) 0 = U ∈ C c ∞ ( Ω ) put
Q = ∫ ( Δ U ) 2 d x ∫ ∣ ∇ U ∣ 2 d x > 0 , ρ = Q , W ( x ) = U ( x / ρ ) , x ∈ ρ Ω . Q=\frac{\int(\Delta U)^2\,\mathrm{d}x}{\int|\nabla U|^2\,\mathrm{d}x}>0,\qquad\rho=\sqrt{Q},\qquad W(x)=U(x/\rho),\quad x\in\rho\Omega. Q = ∫ ∣∇ U ∣ 2 d x ∫ ( Δ U ) 2 d x > 0 , ρ = Q , W ( x ) = U ( x / ρ ) , x ∈ ρ Ω.
In dimension n n n the numerator of the buckling quotient scales by ρ n − 4 \rho^{n-4} ρ n − 4 and its denominator by ρ n − 2 \rho^{n-2} ρ n − 2 , so W W W has quotient one. Apply (18) to W W W on ρ Ω \rho\Omega ρ Ω . Its right side is zero, and the Dirichlet variational principle yields
λ 1 D ( ρ Ω ) ≤ ∫ ∣ ∇ F W ∣ 2 d x ∫ F W 2 d x ≤ λ . \lambda_1^{\mathrm{D}}(\rho\Omega) \le\frac{\int|\nabla F_W|^2\,\mathrm{d}x}{\int F_W^2\,\mathrm{d}x} \le\lambda. λ 1 D ( ρ Ω ) ≤ ∫ F W 2 d x ∫ ∣∇ F W ∣ 2 d x ≤ λ .
Since λ 1 D ( ρ Ω ) = λ 1 D ( Ω ) / Q \lambda_1^{\mathrm{D}}(\rho\Omega)=\lambda_1^{\mathrm{D}}(\Omega)/Q λ 1 D ( ρ Ω ) = λ 1 D ( Ω ) / Q , we obtain Q ≥ λ 1 D ( Ω ) / λ Q\ge\lambda_1^{\mathrm{D}}(\Omega)/\lambda Q ≥ λ 1 D ( Ω ) / λ .
For any nonzero U ∈ H 0 2 ( Ω ) U\in H_0^2(\Omega) U ∈ H 0 2 ( Ω ) choose smooth compactly supported U j → U U_j\to U U j → U in H 2 H^2 H 2 . Both quadratic integrals in its quotient converge, and the limiting gradient denominator is positive. The quotients therefore converge as well. Taking the infimum proves Theorem 1.1 for 2 ≤ n ≤ 47 2\le n\le47 2 ≤ n ≤ 47 .
The n n n -dimensional Faber–Krahn inequality gives λ 1 D ( Ω ) ≥ q 2 / R 2 \lambda_1^{\mathrm{D}}(\Omega)\ge q^2/R^2 λ 1 D ( Ω ) ≥ q 2 / R 2 for ∣ B R ∣ = ∣ Ω ∣ |B_R|=|\Omega| ∣ B R ∣ = ∣Ω∣ . This is the standard form-domain version of Faber–Krahn, valid for arbitrary finite-volume open sets; see [7 ] . The ball formula Λ 1 ( B R ) = k 2 / R 2 \Lambda_1(B_R)=k^2/R^2 Λ 1 ( B R ) = k 2 / R 2 quoted in the introduction therefore proves the inequality in Corollary 1.2. It also shows that the constant in Theorem 1.1 is optimal. The remaining issue is the exact equality class.
Equality and Sobolev capacity
For s = 1 , 2 s=1,2 s = 1 , 2 , let G s G_s G s be the Bessel kernel, whose Fourier transform is ( 1 + ∣ ξ ∣ 2 ) − s / 2 (1+|\xi|^2)^{-s/2} ( 1 + ∣ ξ ∣ 2 ) − s /2 . The Bessel capacity is
C s , 2 ( E ) = inf { ∥ h ∥ L 2 ( R n ) 2 : h ≥ 0 , G s ∗ h ≥ 1 on E } . C_{s,2}(E)=\inf\left\{\|h\|_{L^2(\mathbb{R}^n)}^2:h\ge0,\ G_s*h\ge1\text{ on }E\right\}. C s , 2 ( E ) = inf { ∥ h ∥ L 2 ( R n ) 2 : h ≥ 0 , G s ∗ h ≥ 1 on E } .
The convolution is defined for nonnegative h h h , with the value + ∞ +\infty + ∞ allowed. We use only capacity-zero sets, so the normalization of the Fourier transform is immaterial. A statement holds quasi-everywhere , abbreviated q.e., if its exceptional set has zero specified capacity. A quasicontinuous representative is one that becomes continuous after removal of an open set of arbitrarily small capacity. These representatives exist for Sobolev functions and are unique up to capacity-zero sets; see [1 ] . In particular, C 2 , 2 ( E ) = 0 C_{2,2}(E)=0 C 2 , 2 ( E ) = 0 implies ∣ E ∣ = 0 |E|=0 ∣ E ∣ = 0 and C 1 , 2 ( E ) = 0 C_{1,2}(E)=0 C 1 , 2 ( E ) = 0 .
The equality argument depends on the spectral comparison only through the following hypothesis; the converse assertion holds in every dimension. Put q n = j n / 2 − 1 , 1 q_n=j_{n/2-1,1} q n = j n /2 − 1 , 1 and k n = j n / 2 , 1 k_n=j_{n/2,1} k n = j n /2 , 1 .
The following statement separates the capacity argument from the construction of the calibration.
Theorem 2.12. Fix a dimension n ≥ 2 n\ge2 n ≥ 2 in which the strong comparison
Λ 1 ( D ) ≥ k n 2 q n 2 λ 1 D ( D ) for every open D ⊂ R n , 0 < ∣ D ∣ < ∞ (19) \Lambda_1(D)\ge\frac{k_n^2}{q_n^2}\lambda_1^{\mathrm{D}}(D)\qquad\text{for every open }D\subset\mathbb{R}^n,\quad0<|D|<\infty \tag*{(19)} Λ 1 ( D ) ≥ q n 2 k n 2 λ 1 D ( D ) for every open D ⊂ R n , 0 < ∣ D ∣ < ∞ (19)
has been proved. Let Ω \Omega Ω be such an open set and let ∣ B R ∣ = ∣ Ω ∣ |B_R|=|\Omega| ∣ B R ∣ = ∣Ω∣ . Then
Λ 1 ( Ω ) = Λ 1 ( B R ) ⟺ Ω = B R ( x 0 ) ∖ E for some x 0 ∈ R n , C 2 , 2 ( E ) = 0. (20) \Lambda_1(\Omega)=\Lambda_1(B_R)\quad\Longleftrightarrow\quad\Omega=B_R(x_0)\setminus E\text{ for some }x_0\in\mathbb{R}^n,\quad C_{2,2}(E)=0. \tag*{(20)} Λ 1 ( Ω ) = Λ 1 ( B R ) ⟺ Ω = B R ( x 0 ) ∖ E for some x 0 ∈ R n , C 2 , 2 ( E ) = 0. (20)
Here E E E is relatively closed in B R ( x 0 ) B_R(x_0) B R ( x 0 ) , as is forced by openness of Ω \Omega Ω . Consequently this applies in every dimension 2 ≤ n ≤ 47 2\le n\le47 2 ≤ n ≤ 47 by Theorem 1.1 . The sufficiency direction in (20) is valid in every dimension, independently of (19) .
We use the clamped Sobolev trace criterion
v ∈ H 0 2 ( D ) ⟺ { v ~ = 0 C 2 , 2 -q.e. on D c , ∂ i v ~ = 0 C 1 , 2 -q.e. on D c , 1 ≤ i ≤ n , v ∈ H 2 ( R n ) . (21) v\in H^2_0(D)\quad\Longleftrightarrow\quad
\begin{cases}
\widetilde v=0 & C_{2,2}\text{-q.e. on }D^c,\\
\widetilde{\partial_i v}=0 & C_{1,2}\text{-q.e. on }D^c,\quad1\le i\le n,
\end{cases}
\qquad v\in H^2(\mathbb{R}^n). \tag*{(21)} v ∈ H 0 2 ( D ) ⟺ { v = 0 ∂ i v = 0 C 2 , 2 -q.e. on D c , C 1 , 2 -q.e. on D c , 1 ≤ i ≤ n , v ∈ H 2 ( R n ) . (21)
The representatives in the two lines are the appropriate, possibly different, quasicontinuous representatives. This is the m = p = 2 m=p=2 m = p = 2 closed-set synthesis theorem of Hedberg–Wolff [9 Theorem 5] ; in particular it is valid without boundary regularity and without an embedding into C 1 C^1 C 1 .
Proof of Theorem 2.12. Suppose first that Λ 1 ( Ω ) = k n 2 / R 2 \Lambda_1(\Omega)=k_n^2/R^2 Λ 1 ( Ω ) = k n 2 / R 2 . The strong comparison and Faber–Krahn give
k n 2 R 2 = Λ 1 ( Ω ) ≥ k n 2 q n 2 λ 1 D ( Ω ) ≥ k n 2 q n 2 q n 2 R 2 . \frac{k_n^2}{R^2}=\Lambda_1(\Omega)\ge\frac{k_n^2}{q_n^2}\lambda_1^{D}(\Omega)\ge\frac{k_n^2}{q_n^2}\frac{q_n^2}{R^2}. R 2 k n 2 = Λ 1 ( Ω ) ≥ q n 2 k n 2 λ 1 D ( Ω ) ≥ q n 2 k n 2 R 2 q n 2 .
Thus λ 1 D ( Ω ) = λ 1 D ( B R ) \lambda_1^D(\Omega)=\lambda_1^D(B_R) λ 1 D ( Ω ) = λ 1 D ( B R ) . The Fraenkel asymmetry of Ω \Omega Ω is
A ( Ω ) = inf x 0 ∈ R n ∣ Ω △ B R ( x 0 ) ∣ ∣ Ω ∣ , \mathcal{A}(\Omega)=\inf_{x_0\in\mathbb{R}^n}\frac{|\Omega\mathbin{\triangle} B_R(x_0)|}{|\Omega|}, A ( Ω ) = x 0 ∈ R n inf ∣Ω∣ ∣Ω △ B R ( x 0 ) ∣ ,
where △ \triangle △ denotes symmetric difference and ∣ B R ∣ = ∣ Ω ∣ |B_R|=|\Omega| ∣ B R ∣ = ∣Ω∣ . The quantitative Faber–Krahn theorem [7 Main Theorem, $q=2$] implies zero Fraenkel asymmetry. Hence Ω \Omega Ω agrees almost everywhere with a translate of B R B_R B R . For completeness, the centers of balls approaching zero symmetric difference cannot escape to infinity: intersection of such a ball with the fixed finite-measure set Ω \Omega Ω would then tend to zero. A bounded subsequence of centers converges; continuity of ball translations in L 1 L^1 L 1 gives one ball B = B R ( x 0 ) B=B_R(x_0) B = B R ( x 0 ) with ∣ Ω △ B ∣ = 0 |\Omega\mathbin{\triangle} B|=0 ∣Ω △ B ∣ = 0 . Since Ω \Omega Ω is open, it follows that Ω ⊂ B \Omega\subset B Ω ⊂ B pointwise. Indeed, a point of Ω ∖ B \Omega\setminus B Ω ∖ B , including a point on ∂ B \partial B ∂ B , has an open neighborhood whose intersection with B c B^c B c has positive measure. This would contradict the almost-everywhere equality.
We next justify attainment on Ω \Omega Ω , rather than assume it from an unbounded-domain statement. Take a minimizing sequence in H 0 2 ( Ω ) H^2_0(\Omega) H 0 2 ( Ω ) with ∫ ∣ ∇ v j ∣ 2 d x = 1 \int|\nabla v_j|^2\,dx=1 ∫ ∣∇ v j ∣ 2 d x = 1 . Its zero extensions belong to H 0 2 ( B ) H^2_0(B) H 0 2 ( B ) . The identity ∫ ∣ D 2 v j ∣ 2 d x = ∫ ∣ Δ v j ∣ 2 d x \int|D^2v_j|^2\,dx=\int|\Delta v_j|^2\,dx ∫ ∣ D 2 v j ∣ 2 d x = ∫ ∣Δ v j ∣ 2 d x , followed by Poincaré on the bounded ball, gives boundedness in H 2 ( B ) H^2(B) H 2 ( B ) . Weak compactness, closedness of the zero-extended subspace H 0 2 ( Ω ) H^2_0(\Omega) H 0 2 ( Ω ) , and the Rellich compact embedding H 2 ( B ) ↪ H 1 ( B ) H^2(B)\hookrightarrow H^1(B) H 2 ( B ) ↪ H 1 ( B ) [21 ] give a nonzero minimizer v ∈ H 0 2 ( Ω ) v\in H^2_0(\Omega) v ∈ H 0 2 ( Ω ) with ∫ ∣ ∇ v ∣ 2 d x = 1 \int|\nabla v|^2\,dx=1 ∫ ∣∇ v ∣ 2 d x = 1 . It is also a first buckling eigenfunction of B B B .
The first buckling eigenspace of a ball is one-dimensional and is spanned by a smooth radial function U B U_B U B that is strictly positive in the interior; see [8 Theorem 3.2 and Section 6, $\kappa = 0$] . Thus v = c U B v = cU_B v = c U B in H 2 ( R n ) H^2(\mathbb{R}^n) H 2 ( R n ) for a nonzero constant c c c . Let E = B ∖ Ω E = B \setminus\Omega E = B ∖ Ω . By the first line of (21) , v ~ = 0 \tilde{v} = 0 v ~ = 0 C 2 , 2 C_{2,2} C 2 , 2 -q.e. on E E E . But the continuous representative c U B cU_B c U B is nonzero at every point of E E E , and it agrees with v ~ \tilde{v} v ~ outside a C 2 , 2 C_{2,2} C 2 , 2 -null set. Consequently C 2 , 2 ( E ) = 0 C_{2,2}(E) = 0 C 2 , 2 ( E ) = 0 . Conversely, suppose Ω = B ∖ E \Omega= B \setminus E Ω = B ∖ E and C 2 , 2 ( E ) = 0 C_{2,2}(E) = 0 C 2 , 2 ( E ) = 0 . For any v ∈ H 0 2 ( B ) v \in H^2_0(B) v ∈ H 0 2 ( B ) , its value trace vanishes C 2 , 2 C_{2,2} C 2 , 2 -q.e. on B c B^c B c and its derivative traces vanish C 1 , 2 C_{1,2} C 1 , 2 -q.e. there. The same statements hold on B c ∪ E = Ω c B^c \cup E = \Omega^c B c ∪ E = Ω c since C 1 , 2 ( E ) = 0 C_{1,2}(E) = 0 C 1 , 2 ( E ) = 0 as well. Formula (21) yields v ∈ H 0 2 ( Ω ) v \in H^2_0(\Omega) v ∈ H 0 2 ( Ω ) . The reverse inclusion is immediate, so that
H 0 2 ( Ω ) = H 0 2 ( B ) as zero-extended subspaces of H 2 ( R n ) . H^2_0(\Omega) = H^2_0(B) \quad\text{as zero-extended subspaces of } H^2(\mathbb{R}^n). H 0 2 ( Ω ) = H 0 2 ( B ) as zero-extended subspaces of H 2 ( R n ) .
Their buckling infima are equal, and ∣ Ω ∣ = ∣ B ∣ |\Omega| = |B| ∣Ω∣ = ∣ B ∣ because C 2 , 2 C_{2,2} C 2 , 2 -null sets have zero Lebesgue measure.
Equality in the volume inequality must be distinguished from equality in the spectral ratio.
Remark 2.13. Theorem 2.12 classifies equality in the volume bound. The ratio comparison itself also attains equality on a finite disjoint union of equal-radius balls: both first variational values are the minimum of the corresponding component values. Its volume bound is strict when there is more than one component. The theorem makes no claim to classify equality in the ratio comparison alone.
A singleton has positive C s , 2 C_{s,2} C s , 2 capacity exactly when 2 s > n 2s > n 2 s > n [1 Chapter 2] . Applying this fact with s = 2 s = 2 s = 2 to Theorem 2.12 shows that equality forces a translated ball in dimensions two and three. In dimensions n ≥ 4 n \ge4 n ≥ 4 , any interior point x 1 x_1 x 1 can be removed without changing H 0 2 ( B ) H^2_0(B) H 0 2 ( B ) or the first buckling eigenvalue. This sufficiency assertion does not require the spectral comparison.
The qualification by capacity cannot be replaced by equality almost everywhere. For 2 ≤ n ≤ 47 2 \le n \le47 2 ≤ n ≤ 47 , the set D = B R ∖ { x n = 0 } D = B_R \setminus\{x_n = 0\} D = B R ∖ { x n = 0 } has the same volume as B R B_R B R but has strictly larger first buckling eigenvalue. Indeed, the deleted hyperplane disk has positive C 2 , 2 C_{2,2} C 2 , 2 capacity, by the continuous local hyperplane trace map for H 2 H^2 H 2 [21 ] , and Theorem 2.12 applies. Figure 2 illustrates this distinction. The equality proof uses the C 2 , 2 C_{2,2} C 2 , 2 value trace as well as the C 1 , 2 C_{1,2} C 1 , 2 derivative traces; Dirichlet rigidity alone does not determine the clamped equality class.
Figure 2. Schematic sections of equal-volume sets. Removing a set of zero volume need not preserve the clamped function space. A point has zero H 2 H^2 H 2 capacity for n ≥ 4 n \ge4 n ≥ 4 , whereas a hyperplane disk has positive capacity. The circles represent sections, not a planar puncture example.
Coefficient signs in dimensions three through forty-seven
We now establish the coefficient estimates used in Proposition 2.4 for 3 ≤ n ≤ 47 3 \le n \le47 3 ≤ n ≤ 47 . The three-dimensional profiles admit an explicit trigonometric form. For higher dimensions, a residue expansion reduces the proof to comparison of the first retained pole with the signed tail; the same structure then yields the estimates in dimensions four and five. Appendix § 6 supplies the separate planar argument.
We write [ r j ] H [r^j]H [ r j ] H for the coefficient of r j r^j r j in the Taylor series of H H H .
We shall use 157 / 50 < π < 22 / 7 157/50 < \pi< 22/7 157/50 < π < 22/7 . The tangent addition formula gives Machin’s identity
π = 16 arctan ( 1 / 5 ) − 4 arctan ( 1 / 239 ) , \pi= 16\arctan(1/5) - 4\arctan(1/239), π = 16 arctan ( 1/5 ) − 4 arctan ( 1/239 ) ,
and its alternating series give
π > 16 ( 1 5 − 1 3 ⋅ 5 3 ) − 4 239 > 157 50 , \pi> 16\left(\frac{1}{5}-\frac{1}{3\cdot5^3}\right)-\frac{4}{239}>\frac{157}{50}, π > 16 ( 5 1 − 3 ⋅ 5 3 1 ) − 239 4 > 50 157 ,
π < 16 ( 1 5 − 1 3 ⋅ 5 3 + 1 5 ⋅ 5 5 ) − 4 ( 1 239 − 1 3 ⋅ 239 3 ) < 22 7 . \pi< 16\left(\frac{1}{5}-\frac{1}{3\cdot5^3}+\frac{1}{5\cdot5^5}\right)-4\left(\frac{1}{239}-\frac{1}{3\cdot239^3}\right)<\frac{22}{7}. π < 16 ( 5 1 − 3 ⋅ 5 3 1 + 5 ⋅ 5 5 1 ) − 4 ( 239 1 − 3 ⋅ 23 9 3 1 ) < 7 22 .
The three-dimensional correction coefficients
For n = 3 n=3 n = 3 the common profiles become X ( r ) = sin r / r X(r)=\sin r/r X ( r ) = sin r / r and p ( r ) = ( sin r − r cos r ) / r 2 p(r)=(\sin r-r\cos r)/r^2 p ( r ) = ( sin r − r cos r ) / r 2 , with q = π q=\pi q = π . For this subsection put
D ( r ) = sin r − r cos r , L ( r ) = ( K − 1 ) r 2 − K sin 2 r + α − 2 sin 2 ( α r ) . D(r)=\sin r-r\cos r,\qquad\mathcal{L}(r)=(K-1)r^2-K\sin^2r+\alpha^{-2}\sin^2(\alpha r). D ( r ) = sin r − r cos r , L ( r ) = ( K − 1 ) r 2 − K sin 2 r + α − 2 sin 2 ( α r ) .
Direct substitution into (5) gives
G = K − r 2 L 2 D 2 . (22) G=K-\frac{r^2\mathcal{L}}{2D^2}. \tag*{(22)} G = K − 2 D 2 r 2 L . (22)
Write κ l \kappa_l κ l for the positive zeros of D D D , with κ 1 = k \kappa_1=k κ 1 = k . Since tan k = k \tan k=k tan k = k and cos k < 0 \cos k<0 cos k < 0 , one has
c = ( 1 + k 2 ) − 1 / 2 , K = 1 + k − 2 > 1. c=(1+k^2)^{-1/2},\qquad K=1+k^{-2}>1. c = ( 1 + k 2 ) − 1/2 , K = 1 + k − 2 > 1.
The following coarse bounds suffice:
4.49 < k < 4.50 , 12 25 < λ < 49 100 < 1 2 , 27 125 < c < 11 50 , K < 21 20 . (23) 4.49<k<4.50,\qquad\frac{12}{25}<\lambda<\frac{49}{100}<\frac{1}{2},\qquad\frac{27}{125}<c<\frac{11}{50},\qquad K<\frac{21}{20}. \tag*{(23)} 4.49 < k < 4.50 , 25 12 < λ < 100 49 < 2 1 , 125 27 < c < 50 11 , K < 20 21 . (23)
For completeness the root enclosure can be obtained directly from the alternating sine and cosine series: if S N ( x ) = ∑ j = 0 N ( − 1 ) j x 2 j + 1 / ( 2 j + 1 ) ! S_N(x)=\sum_{j=0}^{N}(-1)^j x^{2j+1}/(2j+1)! S N ( x ) = ∑ j = 0 N ( − 1 ) j x 2 j + 1 / ( 2 j + 1 )! and C N ( x ) = ∑ j = 0 N ( − 1 ) j x 2 j / ( 2 j ) ! C_N(x)=\sum_{j=0}^{N}(-1)^j x^{2j}/(2j)! C N ( x ) = ∑ j = 0 N ( − 1 ) j x 2 j / ( 2 j )! , the two rational inequalities
S 11 ( 449 / 100 ) − ( 449 / 100 ) C 12 ( 449 / 100 ) > 0 , S 12 ( 9 / 2 ) − ( 9 / 2 ) C 11 ( 9 / 2 ) < 0 S_{11}(449/100)-(449/100)C_{12}(449/100)>0,\qquad S_{12}(9/2)-(9/2)C_{11}(9/2)<0 S 11 ( 449/100 ) − ( 449/100 ) C 12 ( 449/100 ) > 0 , S 12 ( 9/2 ) − ( 9/2 ) C 11 ( 9/2 ) < 0
give the assertion by the alternating remainder bound and D ′ = r sin r < 0 D'=r\sin r<0 D ′ = r sin r < 0 on ( π , 3 π / 2 ) (\pi,3\pi/2) ( π , 3 π /2 ) . The remaining bounds in (23) follow from 157 / 50 < π < 22 / 7 157/50<\pi<22/7 157/50 < π < 22/7 and the explicit expressions for c c c , K K K , λ \lambda λ . The bounds for π \pi π were established at the start of this section.
These explicit profiles yield the coefficient signs required for the three-dimensional calibration.
Lemma 3.1. There are a j > 0 a_j>0 a j > 0 such that, with convergence of all derivatives on [ 0 , k ] [0,k] [ 0 , k ] ,
G ( r ) = G ( 0 ) − ∑ j ≥ 1 a j r 2 j = ∑ j ≥ 1 a j ( k 2 j − r 2 j ) . G(r)=G(0)-\sum_{j\geq1}a_jr^{2j}=\sum_{j\geq1}a_j\left(k^{2j}-r^{2j}\right). G ( r ) = G ( 0 ) − j ≥ 1 ∑ a j r 2 j = j ≥ 1 ∑ a j ( k 2 j − r 2 j ) . In particular G > 0 G>0 G > 0 , G r < 0 G_r<0 G r < 0 and G r r < 0 G_{rr}<0 G r r < 0 on ( 0 , k ) (0,k) ( 0 , k ) .
Proof. Use the entire function L \mathcal{L} L and (22) . The singularities at 0 , ± k 0,\pm k 0 , ± k are removable by § 2.1 . All other nonzero zeros of D D D are the nonzero zeros of J 3 / 2 J_{3/2} J 3/2 , and are real and simple [12 Section 10.21(i)] .
At κ = κ l \kappa=\kappa_l κ = κ l , l ≥ 2 l\geq2 l ≥ 2 , one has D ′ ′ ( κ ) / D ′ ( κ ) = 2 / κ D''(\kappa)/D'(\kappa)=2/\kappa D ′′ ( κ ) / D ′ ( κ ) = 2/ κ . Consequently r 2 / D 2 r^2/D^2 r 2 / D 2 has no simple pole term, and the principal part of G G G at κ \kappa κ is
− ( 1 + κ 2 ) L ( κ ) 2 κ 2 ( r − κ ) 2 − ( 1 + κ 2 ) L ′ ( κ ) 2 κ 2 ( r − κ ) . -\frac{(1+\kappa^2)\mathcal{L}(\kappa)}{2\kappa^2(r-\kappa)^2}-\frac{(1+\kappa^2)\mathcal{L}'(\kappa)}{2\kappa^2(r-\kappa)}. − 2 κ 2 ( r − κ ) 2 ( 1 + κ 2 ) L ( κ ) − 2 κ 2 ( r − κ ) ( 1 + κ 2 ) L ′ ( κ ) .
Evenness determines the principal part at − κ -\kappa − κ . For j ≥ 2 j\geq2 j ≥ 2 , paired residues therefore give
a j = ∑ l ≥ 2 1 + κ l 2 κ l 2 j + 4 ( ( 2 j + 1 ) L ( κ l ) − κ l L ′ ( κ l ) ) . (24) a_j=\sum_{l\geq2}\frac{1+\kappa_l^2}{\kappa_l^{2j+4}}\left((2j+1)\mathcal{L}(\kappa_l)-\kappa_l\mathcal{L}'(\kappa_l)\right). \tag*{(24)} a j = l ≥ 2 ∑ κ l 2 j + 4 1 + κ l 2 ( ( 2 j + 1 ) L ( κ l ) − κ l L ′ ( κ l ) ) . (24)
Here is a justification of the contour limit. Take the square with vertical sides ℜ r = ± N π \Re r=\pm N\pi ℜ r = ± N π and horizontal sides ℑ r = ± N π \Im r=\pm N\pi ℑ r = ± N π . On a vertical side, with y = ℑ r y=\Im r y = ℑ r ,
∣ D ∣ 2 = ( N π ) 2 cosh 2 y + ( sinh y − y cosh y ) 2 . |D|^2=(N\pi)^2\cosh^2 y+(\sinh y-y\cosh y)^2. ∣ D ∣ 2 = ( N π ) 2 cosh 2 y + ( sinh y − y cosh y ) 2 .
On a horizontal side the inequality ∣ D ∣ ≥ ∣ r cos r ∣ − ∣ sin r ∣ |D|\geq|r\cos r|-|\sin r| ∣ D ∣ ≥ ∣ r cos r ∣ − ∣ sin r ∣ gives ∣ D ∣ ≥ c 1 ∣ r ∣ e N π |D|\geq c_1|r|e^{N\pi} ∣ D ∣ ≥ c 1 ∣ r ∣ e N π for large N N N . Since α < 1 \alpha<1 α < 1 , these bounds give G = O ( N 2 ) G=O(N^2) G = O ( N 2 ) uniformly on the contour. The integral of G ( r ) / r 2 j + 1 G(r)/r^{2j+1} G ( r ) / r 2 j + 1 is thus O ( N 2 − 2 j ) → 0 O(N^{2-2j})\to0 O ( N 2 − 2 j ) → 0 for j ≥ 2 j\geq2 j ≥ 2 . The residue series converges absolutely, its terms being O ( κ l − 2 j ) O(\kappa_l^{-2j}) O ( κ l − 2 j ) . No contour claim for j = 1 j=1 j = 1 is needed.
At a pole put x = α κ x=\alpha\kappa x = α κ and
P = κ 2 ( κ 2 − k 2 ) k 2 ( 1 + κ 2 ) > 0. P=\frac{\kappa^2(\kappa^2-k^2)}{k^2(1+\kappa^2)}>0. P = k 2 ( 1 + κ 2 ) κ 2 ( κ 2 − k 2 ) > 0.
Using sin 2 κ = κ 2 / ( 1 + κ 2 ) \sin^2 \kappa= \kappa^2/(1+\kappa^2) sin 2 κ = κ 2 / ( 1 + κ 2 ) gives
L ( κ ) = P + λ − 1 sin 2 x > 0 , \mathcal{L}(\kappa)=P+\lambda^{-1}\sin^2 x>0, L ( κ ) = P + λ − 1 sin 2 x > 0 ,
5 L ( κ ) − κ L ′ ( κ ) = 3 P + λ − 1 ( 5 sin 2 x − x sin ( 2 x ) ) . 5\mathcal{L}(\kappa)-\kappa\mathcal{L}'(\kappa)=3P+\lambda^{-1}(5\sin^2 x-x\sin(2x)). 5 L ( κ ) − κ L ′ ( κ ) = 3 P + λ − 1 ( 5 sin 2 x − x sin ( 2 x )) .
This last expression is positive for every l ≥ 2 l\ge2 l ≥ 2 , as follows. For l = 2 l=2 l = 2 , the bounds 9 π / 4 < κ 2 < 5 π / 2 9\pi/4<\kappa_2<5\pi/2 9 π /4 < κ 2 < 5 π /2 give 3 π / 2 < x < 2 π 3\pi/2<x<2\pi 3 π /2 < x < 2 π , hence sin ( 2 x ) < 0 \sin(2x)<0 sin ( 2 x ) < 0 . For l = 3 l=3 l = 3 , 13 π / 4 < κ 3 < 7 π / 2 13\pi/4<\kappa_3<7\pi/2 13 π /4 < κ 3 < 7 π /2 gives P > 4 P>4 P > 4 , 9 π / 4 < x < 11 π / 4 9\pi/4<x<11\pi/4 9 π /4 < x < 11 π /4 , and x < 7.8 x<7.8 x < 7.8 ; therefore the expression exceeds
12 − 7.8 − 2.5 12 / 25 = 23 24 > 0. 12-\frac{7.8-2.5}{12/25}=\frac{23}{24}>0. 12 − 12/25 7.8 − 2.5 = 24 23 > 0.
The bounds on κ 2 , κ 3 \kappa_2,\kappa_3 κ 2 , κ 3 use tan r = 1 < r \tan r=1<r tan r = 1 < r at the respective lower endpoints. For l ≥ 4 l\ge4 l ≥ 4 , κ l > 14 \kappa_l>14 κ l > 14 : indeed ϵ = 9 π / 2 − 14 > 0.13 \epsilon=9\pi/2-14>0.13 ϵ = 9 π /2 − 14 > 0.13 gives tan 14 = cot ϵ < 1 / ϵ < 14 \tan14=\cot\epsilon<1/\epsilon<14 tan 14 = cot ϵ < 1/ ϵ < 14 . Also
P > 4 κ 2 81 − 21 20 , α − 1 < 3 2 , P>\frac{4\kappa^2}{81}-\frac{21}{20},\qquad\alpha^{-1}<\frac{3}{2}, P > 81 4 κ 2 − 20 21 , α − 1 < 2 3 ,
so the expression is greater than 4 κ 2 / 27 − 63 / 20 − 3 κ / 2 > 0 4\kappa^2/27-63/20-3\kappa/2>0 4 κ 2 /27 − 63/20 − 3 κ /2 > 0 for κ ≥ 14 \kappa\ge14 κ ≥ 14 . Thus every term in (24) is positive for j ≥ 2 j\ge2 j ≥ 2 . Direct expansion at zero gives
a 1 = K − 3 λ + 2 λ 2 10 > 0 , a_1=\frac{K-3\lambda+2\lambda^2}{10}>0, a 1 = 10 K − 3 λ + 2 λ 2 > 0 ,
since 3 λ − 2 λ 2 < 1 < K 3\lambda-2\lambda^2<1<K 3 λ − 2 λ 2 < 1 < K . The nearest possible remaining pole lies at κ 2 > k \kappa_2>k κ 2 > k , which justifies convergence and differentiation on [ 0 , k ] [0,k] [ 0 , k ] . Finally G ( k ) = 0 G(k)=0 G ( k ) = 0 gives the second series representation.
The first pole and the tail in higher dimensions
This subsection proves the calibration coefficient signs in every integer dimension 6 ≤ n ≤ 47 6\le n\le47 6 ≤ n ≤ 47 . Uniform pole estimates, quantitative quarter-phase comparisons, and rational polynomial inequalities in a continuous dimension parameter give the proof. Put m = n − 1 m=n-1 m = n − 1 , ν = ( n − 2 ) / 2 \nu=(n-2)/2 ν = ( n − 2 ) /2 , X = ψ n X=\psi_n X = ψ n , p = − X r = r ψ n + 2 / n p=-X_r=r\psi_{n+2}/n p = − X r = r ψ n + 2 / n , q l = j ν , l q_l=j_{\nu,l} q l = j ν , l , κ l = j ν + 1 , l \kappa_l=j_{\nu+1,l} κ l = j ν + 1 , l , q = q 1 q=q_1 q = q 1 , k = κ 1 k=\kappa_1 k = κ 1 , α = q / k \alpha=q/k α = q / k , λ = α 2 \lambda=\alpha^2 λ = α 2 , c = − X ( k ) c=-X(k) c = − X ( k ) , and K = λ p ( q ) 2 / c 2 K=\lambda p(q)^2/c^2 K = λ p ( q ) 2 / c 2 . Write
B = X 2 + p 2 − m X p / r , Q ( r ) = K B ( r ) − λ B ( α r ) , W ( r ) = r m B ( r ) , G = K − r 2 Q m p 2 . B=X^2+p^2-mXp/r,\qquad Q(r)=KB(r)-\lambda B(\alpha r),\qquad W(r)=r^mB(r),\qquad G=K-\frac{r^2Q}{mp^2}. B = X 2 + p 2 − m X p / r , Q ( r ) = K B ( r ) − λ B ( α r ) , W ( r ) = r m B ( r ) , G = K − m p 2 r 2 Q .
The first step is to determine the sign of the residue factor at every pole.
Lemma 3.2. For every real n ≥ 3 n\ge3 n ≥ 3 and every l ≥ 2 l\ge2 l ≥ 2 , Q ( κ l ) > 0 Q(\kappa_l)>0 Q ( κ l ) > 0 . Moreover q l / κ l q_l/\kappa_l q l / κ l strictly increases with l l l .
Proof. The identities
( X 2 + p 2 ) r = − 2 m p 2 / r , W r = m r m − 2 X p , \left(X^2+p^2\right)_r=-2mp^2/r,\qquad W_r=mr^{m-2}Xp, ( X 2 + p 2 ) r = − 2 m p 2 / r , W r = m r m − 2 X p ,
follow from X r = − p X_r=-p X r = − p and p r = X − m p / r p_r=X-mp/r p r = X − m p / r . One also has
B ( r ) = r − m − 1 ∫ 0 r s m { X ( s ) 2 + p ( s ) 2 } d s > 0. B(r)=r^{-m-1}\int_0^r s^m\{X(s)^2+p(s)^2\}\,\mathrm{d}s>0. B ( r ) = r − m − 1 ∫ 0 r s m { X ( s ) 2 + p ( s ) 2 } d s > 0.
Indeed the derivative of r m + 1 B r^{m+1}B r m + 1 B is r m ( X 2 + p 2 ) r^m(X^2+p^2) r m ( X 2 + p 2 ) , and its value at zero is zero.
Choose a continuous polar phase X = ρ cos θ X=\rho\cos\theta X = ρ cos θ , p = ρ sin θ p=\rho\sin\theta p = ρ sin θ . On the quarter-cycle from q l q_l q l to κ l \kappa_l κ l set φ = θ − ( l − 1 ) π − π / 2 \varphi=\theta-(l-1)\pi-\pi/2 φ = θ − ( l − 1 ) π − π /2 . Then 0 ≤ φ ≤ π / 2 0\leq\varphi\leq\pi/2 0 ≤ φ ≤ π /2 and
d r l d φ = 2 r l 2 r l + m sin 2 φ . \frac{\mathrm{d}r_l}{\mathrm{d}\varphi}=\frac{2r_l}{2r_l+m\sin2\varphi}. d φ d r l = 2 r l + m sin 2 φ 2 r l .
The solutions r l ( φ ) r_l(\varphi) r l ( φ ) strictly increase with l l l , by uniqueness of this scalar differential equation and the initial values r l ( 0 ) = q l r_l(0)=q_l r l ( 0 ) = q l . Consequently
log κ l q l = ∫ 0 π / 2 2 d φ 2 r l + m sin 2 φ \log\frac{\kappa_l}{q_l}=\int_0^{\pi/2}\frac{2\,\mathrm{d}\varphi}{2r_l+m\sin2\varphi} log q l κ l = ∫ 0 π /2 2 r l + m sin 2 φ 2 d φ
strictly decreases with l l l , proving the zero-ratio assertion. Also
log W ( q l ) W ( κ l ) = ∫ 0 π / 2 2 m sin 2 φ ( 2 r l + m sin 2 φ ) 2 d φ \log\frac{W(q_l)}{W(\kappa_l)}=\int_0^{\pi/2}\frac{2m\sin2\varphi}{(2r_l+m\sin2\varphi)^2}\,\mathrm{d}\varphi log W ( κ l ) W ( q l ) = ∫ 0 π /2 ( 2 r l + m sin 2 φ ) 2 2 m sin 2 φ d φ
strictly decreases with l l l . Its first value is
R : = W ( q ) W ( κ ) = K α m − 2 > 1. R:=\frac{W(q)}{W(\kappa)}=K\alpha^{m-2}>1. R := W ( κ ) W ( q ) = K α m − 2 > 1.
The successive local maxima W ( q l ) W(q_l) W ( q l ) are nondecreasing. To see this, put Y ν = r J ν Y_\nu=\sqrt{r}J_\nu Y ν = r J ν , v ν = 1 − ( ν 2 − 1 / 4 ) / r 2 v_\nu=1-(\nu^2-1/4)/r^2 v ν = 1 − ( ν 2 − 1/4 ) / r 2 and E ν = Y ν , r 2 + v ν Y ν 2 \mathcal{E}_\nu=Y_{\nu,r}^2+v_\nu Y_\nu^2 E ν = Y ν , r 2 + v ν Y ν 2 . Then E ν , r = 2 ( ν 2 − 1 / 4 ) Y ν 2 / r 3 ≥ 0 \mathcal{E}_{\nu,r}=2(\nu^2-1/4)Y_\nu^2/r^3\geq0 E ν , r = 2 ( ν 2 − 1/4 ) Y ν 2 / r 3 ≥ 0 , and W ( q l ) = N 2 E ν ( q l ) W(q_l)=N^2\mathcal{E}_\nu(q_l) W ( q l ) = N 2 E ν ( q l ) with N = 2 ν Γ ( ν + 1 ) N=2^\nu\Gamma(\nu+1) N = 2 ν Γ ( ν + 1 ) . The other positive critical points of W W W are the local minima κ l \kappa_l κ l , while W ( 0 ) = 0 W(0)=0 W ( 0 ) = 0 . Hence W ( x ) ≤ W ( q l ) W(x)\leq W(q_l) W ( x ) ≤ W ( q l ) whenever 0 ≤ x ≤ q l 0\leq x\leq q_l 0 ≤ x ≤ q l . Since α κ l < q l \alpha\kappa_l<q_l α κ l < q l for l ≥ 2 l\geq2 l ≥ 2 , we obtain
W ( α κ l ) W ( κ l ) ≤ W ( q l ) W ( κ l ) < R . \frac{W(\alpha\kappa_l)}{W(\kappa_l)}\leq\frac{W(q_l)}{W(\kappa_l)}<R. W ( κ l ) W ( α κ l ) ≤ W ( κ l ) W ( q l ) < R .
Multiplying by λ α − m W ( κ l ) \lambda\alpha^{-m}W(\kappa_l) λ α − m W ( κ l ) gives κ l m Q ( κ l ) > 0 \kappa_l^mQ(\kappa_l)>0 κ l m Q ( κ l ) > 0 , as required. The proof includes n = 3 n=3 n = 3 , for which the local maxima are constant and the strictly decreasing quarter-cycle ratio still supplies the strict conclusion.
Before estimating the higher-order residues, we determine the sign of the first coefficient at the center.
Lemma 3.3. For every real n ≥ 6 n\geq6 n ≥ 6 , with the canonical ball parameters and m = n − 1 m=n-1 m = n − 1 , one has
K − n λ + m λ 2 > 0. K-n\lambda+m\lambda^2>0. K − nλ + m λ 2 > 0. Consequently
a 1 = K − n λ + m λ 2 m ( n + 2 ) > 0. a_1=\frac{K-n\lambda+m\lambda^2}{m(n+2)}>0. a 1 = m ( n + 2 ) K − nλ + m λ 2 > 0.
Proof. Two analytic facts already proved independently of the calibration coefficient signs will be used. First, the strictly decreasing quarter-cycle energy gives
R = W ( q ) W ( k ) = K λ m / 2 − 1 > 1. R = \frac{W(q)}{W(k)} = K\lambda^{m/2-1} > 1. R = W ( k ) W ( q ) = K λ m /2 − 1 > 1.
Second, the zero ratios q l / κ l q_l/\kappa_l q l / κ l strictly increase with l l l . For d ≥ 2 d \ge2 d ≥ 2 , define the inverse fourth-power zero sum
σ 2 ( d ) : = ∑ l ≥ 1 j d / 2 − 1 , l − 4 = 1 2 d 2 ( d + 2 ) . \sigma_{2}^{(d)} := \sum_{l \ge1} j_{d/2-1,l}^{-4} = \frac{1}{2d^{2}(d+2)}. σ 2 ( d ) := l ≥ 1 ∑ j d /2 − 1 , l − 4 = 2 d 2 ( d + 2 ) 1 .
The value follows by comparing the coefficient of r 4 r^4 r 4 in the logarithm of the Bessel series with that of the canonical product [12 Sections 10.2(i) and 10.21(iii)] . Since q 1 / κ 1 = α q_1/\kappa_1 = \alpha q 1 / κ 1 = α and q l / κ l > α q_l/\kappa_l > \alpha q l / κ l > α for l ≥ 2 l \ge2 l ≥ 2 , the first terms in the following comparison agree and every remaining term is strictly smaller:
λ 2 σ 2 ( n ) = ∑ l ≥ 1 α 4 q l 4 < ∑ l ≥ 1 1 κ l 4 = σ 2 ( n + 2 ) . \lambda^2 \sigma_{2}^{(n)} = \sum_{l \ge1} \frac{\alpha^4}{q_l^4} < \sum_{l \ge1} \frac{1}{\kappa_l^4} = \sigma_{2}^{(n+2)}. λ 2 σ 2 ( n ) = l ≥ 1 ∑ q l 4 α 4 < l ≥ 1 ∑ κ l 4 1 = σ 2 ( n + 2 ) .
It follows that
0 < λ < n ( n + 2 ) ( n + 4 ) < λ H : = n ( n + 3 ) ( n + 2 ) ( n + 4 ) < 1. 0 < \lambda< \frac{n}{\sqrt{(n+2)(n+4)}} < \lambda_H := \frac{n(n+3)}{(n+2)(n+4)} < 1. 0 < λ < ( n + 2 ) ( n + 4 ) n < λ H := ( n + 2 ) ( n + 4 ) n ( n + 3 ) < 1.
The second strict comparison uses ( n + 3 ) 2 = ( n + 2 ) ( n + 4 ) + 1 (n+3)^2 = (n+2)(n+4) + 1 ( n + 3 ) 2 = ( n + 2 ) ( n + 4 ) + 1 . Put a = m / 2 a = m/2 a = m /2 and x = 1 − λ x = 1-\lambda x = 1 − λ . We claim
( 1 − x ) − a > 1 + 2 a x . (1-x)^{-a} > 1 + 2ax. ( 1 − x ) − a > 1 + 2 a x .
All coefficients of its binomial series are positive. Define the increasing polynomial in x > 0 x > 0 x > 0
Q a ( x ) = − 1 + ∑ j = 2 5 ( a + 1 ) ( a + 2 ) ⋯ ( a + j − 1 ) j ! x j − 1 . Q_a(x) = -1 + \sum_{j=2}^{5} \frac{(a+1)(a+2)\cdots(a+j-1)}{j!}x^{j-1}. Q a ( x ) = − 1 + j = 2 ∑ 5 j ! ( a + 1 ) ( a + 2 ) ⋯ ( a + j − 1 ) x j − 1 .
Since x > x H = ( 3 n + 8 ) / ( ( n + 2 ) ( n + 4 ) ) x > x_H = (3n+8)/((n+2)(n+4)) x > x H = ( 3 n + 8 ) / (( n + 2 ) ( n + 4 )) , it suffices to show Q a ( x H ) > 0 Q_a(x_H) > 0 Q a ( x H ) > 0 . The exact polynomial identity is
Q a ( x H ) = P 8 ( n ) 1920 ( n + 2 ) 4 ( n + 4 ) 4 , Q_a(x_H) = \frac{P_8(n)}{1920(n+2)^4(n+4)^4}, Q a ( x H ) = 1920 ( n + 2 ) 4 ( n + 4 ) 4 P 8 ( n ) ,
where
P 8 ( n ) = 591 n 8 + 8850 n 7 + 38876 n 6 − 64850 n 5 − 1218075 n 4 − 4848304 n 3 − 9493760 n 2 − 9483264 n − 3870720. P_8(n) = 591n^8 + 8850n^7 + 38876n^6 - 64850n^5
- 1218075n^4 - 4848304n^3 - 9493760n^2
- 9483264n - 3870720. P 8 ( n ) = 591 n 8 + 8850 n 7 + 38876 n 6 − 64850 n 5 − 1218075 n 4 − 4848304 n 3 − 9493760 n 2 − 9483264 n − 3870720.
Its nonzero coefficients have exactly one sign change in descending order. By Descartes’ rule it has at most one positive root. Since
P 8 ( 0 ) < 0 , P 8 ( 6 ) = 1751207184 > 0 , P_8(0) < 0,\qquad P_8(6) = 1751207184 > 0, P 8 ( 0 ) < 0 , P 8 ( 6 ) = 1751207184 > 0 ,
that root lies below 6, and P 8 ( n ) > 0 P_8(n)>0 P 8 ( n ) > 0 for all real n ≥ 6 n\ge6 n ≥ 6 . Thus Q a ( x ) > 0 Q_a(x)>0 Q a ( x ) > 0 . Keeping the binomial series through degree five gives
( 1 − x ) − a > 1 + a x + a x { 1 + Q a ( x ) } > 1 + 2 a x , (1-x)^{-a}>1+ax+ax\{1+Q_a(x)\}>1+2ax, ( 1 − x ) − a > 1 + a x + a x { 1 + Q a ( x )} > 1 + 2 a x ,
proving the claim.
Finally R > 1 R>1 R > 1 implies K > λ 1 − a K>\lambda^{1-a} K > λ 1 − a . Therefore
K − n λ + m λ 2 > λ { λ − a − n + m λ } = λ { ( 1 − x ) − a − 1 − 2 a x } > 0. \begin{aligned}
K-n\lambda+m\lambda^2>\lambda\{\lambda^{-a}-n+m\lambda\} \\
&=\lambda\{(1-x)^{-a}-1-2ax\}>0.
\end{aligned} K − nλ + m λ 2 > λ { λ − a − n + mλ } = λ {( 1 − x ) − a − 1 − 2 a x } > 0.
This proves the lemma.
The next bounds locate the first retained pole and keep the denominators in the tail estimate positive.
Lemma 3.4. For every real n ≥ 6 n\ge6 n ≥ 6 , let q = j n / 2 − 1 , 1 q=j_{n/2-1,1} q = j n /2 − 1 , 1 , k = j n / 2 , 1 k=j_{n/2,1} k = j n /2 , 1 , κ 2 = j n / 2 , 2 \kappa_2=j_{n/2,2} κ 2 = j n /2 , 2 , and set
α = q / k , A = ( n 2 − 1 ) / 4 , L = π 1 − α , D = α 2 − A / L 2 . \alpha=q/k,\qquad A=(n^2-1)/4,\qquad L=\frac{\pi}{1-\alpha},\qquad D=\alpha^2-A/L^2. α = q / k , A = ( n 2 − 1 ) /4 , L = 1 − α π , D = α 2 − A / L 2 . Then
κ 2 < 2 k < L , D > 1 5 . \kappa_2<2k<L,\qquad D>\frac{1}{5}. κ 2 < 2 k < L , D > 5 1 . Proof. For η ≥ 1 / 2 \eta\ge1/2 η ≥ 1/2 , the first Dirichlet eigenvalue of − ∂ s 2 + ( η 2 − 1 / 4 ) / s 2 -\partial_s^2+(\eta^2-1/4)/s^2 − ∂ s 2 + ( η 2 − 1/4 ) / s 2 on ( 0 , 1 ) (0,1) ( 0 , 1 ) is j η , 1 2 j_{\eta,1}^2 j η , 1 2 . The one-dimensional Wirtinger inequality gives, when η > 1 / 2 \eta>1/2 η > 1/2 ,
j η , 1 2 > η 2 − 1 4 + π 2 . j_{\eta,1}^2>\eta^2-\frac{1}{4}+\pi^2. j η , 1 2 > η 2 − 4 1 + π 2 . In particular k 2 > A + π 2 k^2>A+\pi^2 k 2 > A + π 2 . The function Y ( r ) = r J n / 2 ( r ) Y(r)=\sqrt{r}J_{n/2}(r) Y ( r ) = r J n /2 ( r ) satisfies
Y ′ ′ + ( 1 − A r 2 ) Y = 0. Y''+\left(1-\frac{A}{r^2}\right)Y=0. Y ′′ + ( 1 − r 2 A ) Y = 0. On r ≥ k r\ge k r ≥ k , the coefficient is strictly greater than or equal to ω 2 = 1 − A / k 2 > 0 \omega^2=1-A/k^2>0 ω 2 = 1 − A / k 2 > 0 , with strict inequality inside the interval. Sturm comparison from the zero k k k therefore gives
κ 2 − k < π ω < k . \kappa_2-k<\frac{\pi}{\omega}<k. κ 2 − k < ω π < k . This proves κ 2 < 2 k \kappa_2<2k κ 2 < 2 k .
For the canonical radial pair X , p X,p X , p on the quarter-cycle from q q q to k k k , write X = ρ cos θ X=\rho\cos\theta X = ρ cos θ and p = ρ sin θ p=\rho\sin\theta p = ρ sin θ . Here π / 2 < θ < π \pi/2<\theta<\pi π /2 < θ < π , and
θ r = 1 − n − 1 2 r sin 2 θ > 1. \theta_r=1-\frac{n-1}{2r}\sin2\theta>1. θ r = 1 − 2 r n − 1 sin 2 θ > 1. Consequently k − q < π / 2 k-q<\pi/2 k − q < π /2 and L = π k / ( k − q ) > 2 k L=\pi k/(k-q)>2k L = π k / ( k − q ) > 2 k .
Apply the same eigenvalue lower bound to the order n / 2 − 1 n/2-1 n /2 − 1 :
q 2 > n 2 − 4 n + 3 4 + π 2 . q^2 > \frac{n^2-4n+3}{4}+\pi^2. q 2 > 4 n 2 − 4 n + 3 + π 2 .
Since L > 2 k L>2k L > 2 k and π 2 > 9 \pi^2>9 π 2 > 9 , it follows that
D > q 2 − A / 4 k 2 > 3 n 2 − 16 n + 157 16 k 2 . D>\frac{q^2-A/4}{k^2}>\frac{3n^2-16n+157}{16k^2}. D > k 2 q 2 − A /4 > 16 k 2 3 n 2 − 16 n + 157 .
The numerator is positive for n ≥ 6 n\geq6 n ≥ 6 . For the upper bound, use the weighted radial Dirichlet problem with weight s d − 1 s^{d-1} s d − 1 , where d = n + 2 d=n+2 d = n + 2 . The trial v ( s ) = ( 1 − s 2 ) 2 v(s)=(1-s^2)^2 v ( s ) = ( 1 − s 2 ) 2 has quotient
16 ∫ 0 1 s d + 1 ( 1 − s 2 ) 2 d s ∫ 0 1 s d − 1 ( 1 − s 2 ) 4 d s = d ( d + 8 ) 3 . \frac{16\int_0^1s^{d+1}(1-s^2)^2\,\mathrm{d}s}{\int_0^1s^{d-1}(1-s^2)^4\,\mathrm{d}s}=\frac{d(d+8)}{3}. ∫ 0 1 s d − 1 ( 1 − s 2 ) 4 d s 16 ∫ 0 1 s d + 1 ( 1 − s 2 ) 2 d s = 3 d ( d + 8 ) .
The equality follows by the substitution z = s 2 z=s^2 z = s 2 and integration of the resulting polynomials. The variational bound is strict since this polynomial does not satisfy the first eigenfunction equation. Thus k 2 < ( n + 2 ) ( n + 10 ) / 3 k^2<(n+2)(n+10)/3 k 2 < ( n + 2 ) ( n + 10 ) /3 , and
D > 3 ( 3 n 2 − 16 n + 157 ) 16 ( n + 2 ) ( n + 10 ) > 1 5 . D>\frac{3(3n^2-16n+157)}{16(n+2)(n+10)}>\frac{1}{5}. D > 16 ( n + 2 ) ( n + 10 ) 3 ( 3 n 2 − 16 n + 157 ) > 5 1 .
The last comparison is the positive polynomial identity
29 n 2 − 432 n + 2035 = 29 ( n − 216 29 ) 2 + 12359 29 > 0. 29n^2-432n+2035=29\left(n-\frac{216}{29}\right)^2+\frac{12359}{29}>0. 29 n 2 − 432 n + 2035 = 29 ( n − 29 216 ) 2 + 29 12359 > 0.
This proves the lemma.
These geometric bounds give a uniform upper estimate for the possible negative tail.
Lemma 3.5. In every integer dimension 6 ≤ n ≤ 47 6\leq n\leq47 6 ≤ n ≤ 47 , the constant
V = m 2 D ( κ 2 L ) 4 ( 1 L + 1 4 π ) , L = π 1 − α , D = α 2 − A L 2 , A = n 2 − 1 4 , V=\frac{m}{2\sqrt{D}}\left(\frac{\kappa_2}{L}\right)^4\left(\frac{1}{L}+\frac{1}{4\pi}\right),\qquad L=\frac{\pi}{1-\alpha},\qquad D=\alpha^2-\frac{A}{L^2},\qquad A=\frac{n^2-1}{4}, V = 2 D m ( L κ 2 ) 4 ( L 1 + 4 π 1 ) , L = 1 − α π , D = α 2 − L 2 A , A = 4 n 2 − 1 , satisfies V < 1 / 6 V<1/6 V < 1/6 . Here m = n − 1 m=n-1 m = n − 1 , α = q / k \alpha=q/k α = q / k and κ 2 \kappa_2 κ 2 is the second positive zero of J n / 2 J_{n/2} J n /2 .
Proof. Put b = k 2 b=k^2 b = k 2 and δ = k − q \delta=k-q δ = k − q . The following elementary geometric bounds will be used:
b < ( n + 2 ) ( n + 10 ) 3 , λ = q 2 k 2 > ( n − 2 n ) 2 , κ 2 L < δ ( 1 π + 1 b − A ) . (25) b<\frac{(n+2)(n+10)}{3},\qquad\lambda=\frac{q^2}{k^2}>\left(\frac{n-2}{n}\right)^2,\qquad\frac{\kappa_2}{L}<\delta\left(\frac{1}{\pi}+\frac{1}{\sqrt{b-A}}\right). \tag*{(25)} b < 3 ( n + 2 ) ( n + 10 ) , λ = k 2 q 2 > ( n n − 2 ) 2 , L κ 2 < δ ( π 1 + b − A 1 ) . (25) The first follows from the weighted radial Dirichlet trial in Lemma 3.4 . For the second, the one-dimensional Rayleigh formula and Hardy’s inequality show that j ν , 1 2 / ν 2 j_{\nu,1}^2/\nu^2 j ν , 1 2 / ν 2 strictly decreases with ν > 0 \nu>0 ν > 0 : after division by ν 2 \nu^2 ν 2 , its kinetic-minus-Hardy part has a positive coefficient which strictly decreases. Apply this at ν = ( n − 2 ) / 2 \nu=(n-2)/2 ν = ( n − 2 ) /2 and ν + 1 \nu+1 ν + 1 . For the last inequality, the transformed J n / 2 J_{n/2} J n /2 equation has potential 1 − A / r 2 1-A/r^2 1 − A / r 2 . Sturm comparison to
the constant potential 1 − A / b > 0 1-A/b>0 1 − A / b > 0 on [ k , κ 2 ] [k,\kappa_2] [ k , κ 2 ] gives κ 2 − k < π k / b − A \kappa_2-k<\pi k/\sqrt{b-A} κ 2 − k < π k / b − A , and L = π k / δ L=\pi k/\delta L = π k / δ gives the result. The bounds on b − A b-A b − A below ensure the stated positivity.
The quarter-cycle polar equation gives
δ < J ( m / k ) , J ( a ) = ∫ 0 π / 2 d θ 1 + a sin θ cos θ . \delta<J(m/k),\qquad J(a)=\int_{0}^{\pi/2}\frac{d\theta}{1+a\sin\theta\cos\theta}. δ < J ( m / k ) , J ( a ) = ∫ 0 π /2 1 + a sin θ cos θ d θ .
This is decreasing in a a a . For 0 < a < 2 0<a<2 0 < a < 2 , set z 2 = ( 2 − a ) / ( 2 + a ) z^2=(2-a)/(2+a) z 2 = ( 2 − a ) / ( 2 + a ) . Elementary integration and the alternating arctangent estimate give
J ( a ) = 4 arctan z 4 − a 2 < 4 2 + a ( 1 − z 2 3 + z 4 5 ) . J(a)=\frac{4\arctan z}{\sqrt{4-a^2}}<\frac{4}{2+a}\left(1-\frac{z^2}{3}+\frac{z^4}{5}\right). J ( a ) = 4 − a 2 4 arctan z < 2 + a 4 ( 1 − 3 z 2 + 5 z 4 ) .
From the first bound in (25) , one has respectively m / k > 3 / 4 m/k>3/4 m / k > 3/4 for n ≥ 6 n\ge6 n ≥ 6 , m / k > 1 m/k>1 m / k > 1 for n ≥ 11 n\ge11 n ≥ 11 , and m / k > 7 / 6 m/k>7/6 m / k > 7/6 for n ≥ 15 n\ge15 n ≥ 15 . The cleared polynomial inequalities are
39 n 2 − 204 n − 132 > 0 , 2 n 2 − 18 n − 17 > 0 , 59 n 2 − 804 n − 872 > 0 39n^2-204n-132>0,\qquad2n^2-18n-17>0,\qquad59n^2-804n-872>0 39 n 2 − 204 n − 132 > 0 , 2 n 2 − 18 n − 17 > 0 , 59 n 2 − 804 n − 872 > 0
on their stated ranges. Each is increasing there and positive at its left endpoint. Thus
δ < 13 10 ( n ≥ 6 ) , δ < 11 9 ( n ≥ 11 ) , δ < 47 40 ( n ≥ 15 ) , \delta<\frac{13}{10}\ (n\ge6),\qquad\delta<\frac{11}{9}\ (n\ge11),\qquad\delta<\frac{47}{40}\ (n\ge15), δ < 10 13 ( n ≥ 6 ) , δ < 9 11 ( n ≥ 11 ) , δ < 40 47 ( n ≥ 15 ) ,
because the three arctangent upper bounds are respectively 5168 / 3993 < 13 / 10 5168/3993<13/10 5168/3993 < 13/10 , 164 / 135 < 11 / 9 164/135<11/9 164/135 < 11/9 , and 8024 / 6859 < 47 / 40 8024/6859<47/40 8024/6859 < 47/40 .
We also need two direct Barta estimates. The positive comparison function w = r s ( 1 − r 2 ) w=r^s(1-r^2) w = r s ( 1 − r 2 ) in the transformed order-n / 2 n/2 n /2 equation has
− w ′ ′ w + A r 2 = A − s ( s − 1 ) r 2 + 4 s + 2 1 − r 2 . -\frac{w''}{w}+\frac{A}{r^2}=\frac{A-s(s-1)}{r^2}+\frac{4s+2}{1-r^2}. − w w ′′ + r 2 A = r 2 A − s ( s − 1 ) + 1 − r 2 4 s + 2 .
Its infimum is the square of the sum of the square roots of the two positive coefficients. For s = 2 s=2 s = 2 , n ≥ 6 n\ge6 n ≥ 6 , this gives
b > n 2 + 31 4 + 10 ( n 2 − 9 ) > n 2 + 10 n + 35 4 , b − A > 5 n + 18 2 . b>\frac{n^2+31}{4}+\sqrt{10(n^2-9)}>\frac{n^2+10n+35}{4},\qquad b-A>\frac{5n+18}{2}. b > 4 n 2 + 31 + 10 ( n 2 − 9 ) > 4 n 2 + 10 n + 35 , b − A > 2 5 n + 18 .
The last strict inequality follows from 10 ( n 2 − 9 ) > 5 n / 2 + 1 \sqrt{10(n^2-9)}>5n/2+1 10 ( n 2 − 9 ) > 5 n /2 + 1 , whose squared difference is 15 n 2 / 4 − 5 n − 91 > 0 15n^2/4-5n-91>0 15 n 2 /4 − 5 n − 91 > 0 at and beyond n = 6 n=6 n = 6 . For s = 4 s=4 s = 4 , n ≥ 15 n\ge15 n ≥ 15 , use n 2 − 49 > n − 2 \sqrt{n^2-49}>n-2 n 2 − 49 > n − 2 and 2 > 7 / 5 \sqrt{2}>7/5 2 > 7/5 to obtain
b > 5 n 2 + 84 n − 53 20 , b − A > 21 n − 12 5 . b>\frac{5n^2+84n-53}{20},\qquad b-A>\frac{21n-12}{5}. b > 20 5 n 2 + 84 n − 53 , b − A > 5 21 n − 12 .
These Barta estimates follow directly from the ground-state factorization on compactly supported tests and form-domain density; the comparison function need not have the behavior r ( n + 1 ) / 2 r^{(n+1)/2} r ( n + 1 ) /2 of the regular solution at zero.
Write p 0 = 50 / 157 > 1 / π p_0=50/157>1/\pi p 0 = 50/157 > 1/ π . Since A / b < 1 A/b<1 A / b < 1 , the preceding bounds give
D > ( n − 2 n ) 2 − δ 2 p 0 2 . D>\left(\frac{n-2}{n}\right)^2-\delta^2p_0^2. D > ( n n − 2 ) 2 − δ 2 p 0 2 .
For three of the four required ranges, the following constant bounds are immediate by substitution at the left endpoint:
Table 4.
To make the last column fully explicit, its three rational upper bounds before rounding are
65 7 { 11 9 ( p 0 + 1 / 6 ) } 4 p 0 ( 11 72 + 1 4 ) < 3 20 , \frac{65}{7}\left\{\frac{11}{9}(p_0+1/6)\right\}^{4}p_0\left(\frac{11}{72}+\frac{1}{4}\right)<\frac{3}{20}, 7 65 { 9 11 ( p 0 + 1/6 ) } 4 p 0 ( 72 11 + 4 1 ) < 20 3 ,
700 39 { 47 40 ( p 0 + 4 / 31 ) } 4 p 0 ( 47 430 + 1 4 ) < 4 25 , \frac{700}{39}\left\{\frac{47}{40}(p_0+4/31)\right\}^{4}p_0\left(\frac{47}{430}+\frac{1}{4}\right)<\frac{4}{25}, 39 700 { 40 47 ( p 0 + 4/31 ) } 4 p 0 ( 430 47 + 4 1 ) < 25 4 ,
460 17 { 47 40 ( p 0 + 1 / 11 ) } 4 p 0 ( 47 740 + 1 4 ) < 3 20 . \frac{460}{17}\left\{\frac{47}{40}(p_0+1/11)\right\}^{4}p_0\left(\frac{47}{740}+\frac{1}{4}\right)<\frac{3}{20}. 17 460 { 40 47 ( p 0 + 1/11 ) } 4 p 0 ( 740 47 + 4 1 ) < 20 3 .
Each is a direct rational inequality; clearing the displayed positive denominators verifies it. They are all less than 1 / 6 1/6 1/6 .
It remains to consider the real interval 6 ≤ n ≤ 10 6 \leq n \leq10 6 ≤ n ≤ 10 . The Barta bound gives k > ( n + 5 ) / 2 k>(n+5)/2 k > ( n + 5 ) /2 and b − A > ( n + 13 ) / 4 \sqrt{b-A}>(n+13)/4 b − A > ( n + 13 ) /4 : for the latter square the inequality n 2 − 14 n + 25 < 0 n^2-14n+25<0 n 2 − 14 n + 25 < 0 , which holds throughout [ 6 , 10 ] [6,10] [ 6 , 10 ] . Moreover
D > n + 7 25 . \sqrt{D}>\frac{n+7}{25}. D > 25 n + 7 .
Indeed f ( n ) = ( 1 − 2 / n ) 2 − ( 65 / 157 ) 2 f(n)=\sqrt{(1-2/n)^2-(65/157)^2} f ( n ) = ( 1 − 2/ n ) 2 − ( 65/157 ) 2 is strictly concave for n ≥ 6 n\geq6 n ≥ 6 , as follows by differentiating f = w 2 − c 2 f=\sqrt{w^2-c^2} f = w 2 − c 2 with w = 1 − 2 / n w=1-2/n w = 1 − 2/ n , w ′ > 0 w'>0 w ′ > 0 , w ′ ′ < 0 w''<0 w ′′ < 0 . The direct rational comparisons f ( 6 ) > 13 / 25 f(6)>13/25 f ( 6 ) > 13/25 and f ( 10 ) > 17 / 25 f(10)>17/25 f ( 10 ) > 17/25 give the stated lower chord. Consequently V < W ( n ) V<W(n) V < W ( n ) , where
W ( n ) = 25 ( n − 1 ) 2 ( n + 7 ) { 13 10 ( p 0 + 4 n + 13 ) } 4 p 0 { 13 5 ( n + 5 ) + 1 4 } . W(n)=\frac{25(n-1)}{2(n+7)}\left\{\frac{13}{10}\left(p_0+\frac{4}{n+13}\right)\right\}^{4}p_0\left\{\frac{13}{5(n+5)}+\frac{1}{4}\right\}. W ( n ) = 2 ( n + 7 ) 25 ( n − 1 ) { 10 13 ( p 0 + n + 13 4 ) } 4 p 0 { 5 ( n + 5 ) 13 + 4 1 } .
This rational function strictly decreases for every n ≥ 6 n\geq6 n ≥ 6 . For a short check, put
D = ( n − 1 ) ( n + 7 ) ( n + 13 ) { p 0 ( n + 13 ) + 4 } ( n + 5 ) ( n + 77 / 5 ) > 0. \mathcal{D}=(n-1)(n+7)(n+13)\{p_0(n+13)+4\}(n+5)(n+77/5)>0. D = ( n − 1 ) ( n + 7 ) ( n + 13 ) { p 0 ( n + 13 ) + 4 } ( n + 5 ) ( n + 77/5 ) > 0.
With v = n − 6 v=n-6 v = n − 6 , direct logarithmic differentiation gives
− D W ′ W = 11209208 + 53618064 v + 10245920 v 2 + 645360 v 3 + 13160 v 4 785 > 0. -\mathcal{D}\frac{W'}{W}=\frac{11209208+53618064v+10245920v^2+645360v^3+13160v^4}{785}>0. − D W W ′ = 785 11209208 + 53618064 v + 10245920 v 2 + 645360 v 3 + 13160 v 4 > 0.
Finally direct substitution gives
1 6 − W ( 6 ) = 183971814833017 1640916934670745204 > 1 9000 > 0. \frac{1}{6}-W(6)=\frac{183971814833017}{1640916934670745204}>\frac{1}{9000}>0. 6 1 − W ( 6 ) = 1640916934670745204 183971814833017 > 9000 1 > 0.
Thus V < W ( 6 ) < 1 / 6 V<W(6)<1/6 V < W ( 6 ) < 1/6 also in this last range, completing the proof.
Figure 3 shows the separation of the first retained pole from the tail.
Figure 3. The pole decomposition, schematically and not to scale. The inequalities κ 2 < 2 k < L \kappa_2 < 2k < L κ 2 < 2 k < L separate the first retained pole from the region where the residue contribution may change sign. The argument compares a lower bound for the first contribution with an upper bound for the entire adverse tail.
The next criterion reduces positivity of all higher coefficients to domination of one infinite tail.
Lemma 3.6. Let n ≥ 4 n \ge4 n ≥ 4 , and define
L = π 1 − α , A = m ( m + 2 ) 4 , D = α 2 − A L 2 , V = m 2 D ( κ 2 L ) 4 ( 1 L + 1 4 π ) . L = \frac{\pi}{1-\alpha}, \qquad A = \frac{m(m+2)}{4}, \qquad D = \alpha^2 - \frac{A}{L^2}, \qquad V = \frac{m}{2\sqrt{D}}\left(\frac{\kappa_2}{L}\right)^4\left(\frac{1}{L}+\frac{1}{4\pi}\right). L = 1 − α π , A = 4 m ( m + 2 ) , D = α 2 − L 2 A , V = 2 D m ( L κ 2 ) 4 ( L 1 + 4 π 1 ) . T l = 3 Q ( κ l ) + m α κ l X ( α κ l ) p ( α κ l ) , U l = T l K X ( κ l ) 2 . T_l = 3Q(\kappa_l) + \frac{m\alpha}{\kappa_l}X(\alpha\kappa_l)p(\alpha\kappa_l), \qquad U_l = \frac{T_l}{KX(\kappa_l)^2}. T l = 3 Q ( κ l ) + κ l m α X ( α κ l ) p ( α κ l ) , U l = K X ( κ l ) 2 T l . Suppose D > 0 D > 0 D > 0 , κ 2 < L \kappa_2 < L κ 2 < L , and U 2 > V U_2 > V U 2 > V . Then for every j ≥ 2 j \ge2 j ≥ 2 ,
∑ l ≥ 2 T l X ( κ l ) 2 κ l 2 j > 0. \sum_{l\ge2}\frac{T_l}{X(\kappa_l)^2\kappa_l^{2j}} > 0. l ≥ 2 ∑ X ( κ l ) 2 κ l 2 j T l > 0. Proof. If κ l < L \kappa_l < L κ l < L , separation of consecutive positive zeros by more than π \pi π gives κ l − 1 < κ l − π < α κ l < q l \kappa_{l-1} < \kappa_l-\pi< \alpha\kappa_l < q_l κ l − 1 < κ l − π < α κ l < q l . On ( κ l − 1 , q l ) (\kappa_{l-1},q_l) ( κ l − 1 , q l ) the factors X , p X,p X , p have the same sign. Lemma 3.2 therefore gives T l > 0 T_l > 0 T l > 0 in this range.
For the remaining poles put x = α κ l x=\alpha\kappa_l x = α κ l , Y = x J ν + 1 ( x ) Y=\sqrt{x}J_{\nu+1}(x) Y = x J ν + 1 ( x ) , v = 1 − A / x 2 v=1-A/x^2 v = 1 − A / x 2 , and E = Y r 2 + v Y 2 \mathcal{E}=Y_r^2+vY^2 E = Y r 2 + v Y 2 . Since x ≥ α L > A x\ge\alpha L>\sqrt{A} x ≥ α L > A , this is a positive quadratic form. The identities
X ( x ) = N x − m / 2 { Y r + m Y / ( 2 x ) } , p ( x ) = N x − m / 2 Y X(x)=Nx^{-m/2}\{Y_r+mY/(2x)\}, \qquad p(x)=Nx^{-m/2}Y X ( x ) = N x − m /2 { Y r + mY / ( 2 x )} , p ( x ) = N x − m /2 Y give
κ l m T l = 3 K κ l m X ( κ l ) 2 − 3 N 2 λ α − m { Y r 2 + ( 1 − 5 m 2 / ( 12 x 2 ) ) Y 2 − m Y Y r / ( 3 x ) } . \kappa_l^mT_l=3K\kappa_l^mX(\kappa_l)^2-3N^2\lambda\alpha^{-m}\{Y_r^2+(1-5m^2/(12x^2))Y^2-mYY_r/(3x)\}. κ l m T l = 3 K κ l m X ( κ l ) 2 − 3 N 2 λ α − m { Y r 2 + ( 1 − 5 m 2 / ( 12 x 2 )) Y 2 − mY Y r / ( 3 x )} . Relative to E \mathcal{E} E , the Y 2 Y^2 Y 2 diagonal correction is − m ( m − 3 ) / ( 6 x 2 ) ≤ 0 -m(m-3)/(6x^2)\le0 − m ( m − 3 ) / ( 6 x 2 ) ≤ 0 . Dropping that nonpositive correction shows that the largest relative eigenvalue is at most
1 + m 6 x 2 − A . 1+\frac{m}{6\sqrt{x^2-A}}. 1 + 6 x 2 − A m . Moreover E ( x ) < E ( κ l ) \mathcal{E}(x)<\mathcal{E}(\kappa_l) E ( x ) < E ( κ l ) by the Sonin derivative, and N 2 E ( κ l ) = κ l m X ( κ l ) 2 N^2\mathcal{E}(\kappa_l)=\kappa_l^mX(\kappa_l)^2 N 2 E ( κ l ) = κ l m X ( κ l ) 2 , because J ν + 1 ( κ l ) = 0 J_{\nu+1}(\kappa_l)=0 J ν + 1 ( κ l ) = 0 . Using R > 1 R>1 R > 1 from Lemma 3.2, this yields
U l > − m 2 R α 2 κ l 2 − A > − m 2 D κ l ( κ l ≥ L ) . U_l>-\frac{m}{2R\sqrt{\alpha^2\kappa_l^2-A}}>-\frac{m}{2\sqrt{D}\kappa_l}\qquad(\kappa_l\ge L). U l > − 2 R α 2 κ l 2 − A m > − 2 D κ l m ( κ l ≥ L ) .
The spacing bound supplies
∑ κ l ≥ L κ l − 5 ≤ L − 5 + L − 4 4 π . \sum_{\kappa_l\ge L}\kappa_l^{-5}\le L^{-5}+\frac{L^{-4}}{4\pi}. κ l ≥ L ∑ κ l − 5 ≤ L − 5 + 4 π L − 4 .
Indeed the first term is at most L − 5 L^{-5} L − 5 , and comparison of each subsequent term with the integral of the decreasing function r − 5 r^{-5} r − 5 over the preceding gap gives the second term. For every j ≥ 2 j\ge2 j ≥ 2 and κ l ≥ L > κ 2 \kappa_l\ge L>\kappa_2 κ l ≥ L > κ 2 ,
κ 2 2 j κ l 2 j + 1 ≤ κ 2 4 κ l 5 . \frac{\kappa_2^{2j}}{\kappa_l^{2j+1}}\le\frac{\kappa_2^4}{\kappa_l^5}. κ l 2 j + 1 κ 2 2 j ≤ κ l 5 κ 2 4 .
Keep the positive l = 2 l=2 l = 2 term, discard all the other positive terms below L L L , and use the tail bound. The resulting normalized sum is strictly larger than U 2 − V > 0 U_2-V>0 U 2 − V > 0 , proving the claim.
For the following phase comparison, let n ≥ 4 n\ge4 n ≥ 4 and assume κ 2 < L = π / ( 1 − α ) \kappa_2<L=\pi/(1-\alpha) κ 2 < L = π / ( 1 − α ) . Write m = n − 1 m=n-1 m = n − 1 , A = ( n 2 − 1 ) / 4 A=(n^2-1)/4 A = ( n 2 − 1 ) /4 , A ν = ( ( n − 2 ) 2 − 1 ) / 4 A_\nu=((n-2)^2-1)/4 A ν = (( n − 2 ) 2 − 1 ) /4 , and retain the notation q = q 1 q=q_1 q = q 1 , k = κ 1 k=\kappa_1 k = κ 1 , α = q / k \alpha=q/k α = q / k , κ = κ 2 \kappa=\kappa_2 κ = κ 2 , x = α κ x=\alpha\kappa x = α κ and δ = k − q \delta=k-q δ = k − q . In particular k < x < q 2 < κ k<x<q_2<\kappa k < x < q 2 < κ . The first inequality follows from κ < L \kappa<L κ < L , the zero spacing κ − k > π \kappa-k>\pi κ − k > π , and α κ > κ − π \alpha\kappa>\kappa-\pi α κ > κ − π .
A quantitative phase comparison supplies the lower bound for the first retained pole.
Lemma 3.7. Let n ≥ 4 n\ge4 n ≥ 4 and κ < L \kappa<L κ < L . Suppose that K 0 > k K_0>k K 0 > k , C 0 > κ C_0>\kappa C 0 > κ , d 0 > δ d_0>\delta d 0 > δ and B 0 < b − A B_0<b-A B 0 < b − A , where b = k 2 b=k^2 b = k 2 . Put p − = 157 / 50 p_-=157/50 p − = 157/50 and p + = 22 / 7 p_+=22/7 p + = 22/7 , and define
s = A ν C 0 2 , g = p − ( 1 + s 2 ) , I = 2 p − g ( 2 K 0 + 2 m / p − ) ( 2 C 0 + 2 m / p − ) . s=\frac{A_\nu}{C_0^2},\qquad g=p_-\left(1+\frac{s}{2}\right),\qquad I=\frac{2p_-g}{(2K_0+2m/p_-)(2C_0+2m/p_-)}. s = C 0 2 A ν , g = p − ( 1 + 2 s ) , I = ( 2 K 0 + 2 m / p − ) ( 2 C 0 + 2 m / p − ) 2 p − g . D k = 2 K 0 + m p + 4 , D c = 2 C 0 + m p + 4 , E = 4 m g ( D k + D c ) D k 2 D c 2 , c 0 = 1 − 2 d 0 2 p + 2 3 B 0 . D_k=2K_0+\frac{mp_+}{4},\qquad D_c=2C_0+\frac{mp_+}{4},\qquad E=\frac{4mg(D_k+D_c)}{D_k^2D_c^2},\qquad c_0=1-\frac{2d_0^2p_+^2}{3B_0}. D k = 2 K 0 + 4 m p + , D c = 2 C 0 + 4 m p + , E = D k 2 D c 2 4 m g ( D k + D c ) , c 0 = 1 − 3 B 0 2 d 0 2 p + 2 . S = m I c 0 1 + m I , u 0 = S 1 − S , U = u 0 + 3 E 1 + E . S=\frac{mIc_0}{1+mI},\qquad u_0=\frac{S}{1-S},\qquad U=\frac{u_0+3E}{1+E}. S = 1 + m I m I c 0 , u 0 = 1 − S S , U = 1 + E u 0 + 3 E . If 0 < s < 1 0<s<1 0 < s < 1 , 0 < c 0 < 1 0<c_0<1 0 < c 0 < 1 and m I < 1 mI<1 m I < 1 , then U 2 > U U_2>U U 2 > U .
Proof. Let r l ( ϕ ) r_l(\phi) r l ( ϕ ) be the quarter-phase trajectories from q l q_l q l to κ l \kappa_l κ l , as in Lemma 3.2. They satisfy
r l ′ = 2 r l 2 r l + m sin 2 ϕ , 0 ≤ ϕ ≤ π 2 . r_l'=\frac{2r_l}{2r_l+m\sin2\phi},\qquad0\le\phi\le\frac{\pi}{2}. r l ′ = 2 r l + m sin 2 ϕ 2 r l , 0 ≤ ϕ ≤ 2 π .
The right side is increasing in r l r_l r l . Thus r 2 − r 1 ≥ q 2 − q 1 r_2-r_1\ge q_2-q_1 r 2 − r 1 ≥ q 2 − q 1 . The transformed J ν J_\nu J ν equation on [ q 1 , q 2 ] [q_1,q_2] [ q 1 , q 2 ] has potential at most 1 − A ν / C 0 2 1-A_\nu/C_0^2 1 − A ν / C 0 2 . Sturm comparison and ( 1 − s ) − 1 / 2 > 1 + s / 2 (1-s)^{-1/2}>1+s/2 ( 1 − s ) − 1/2 > 1 + s /2 give
q 2 − q 1 > π 1 − A ν / C 0 2 > g . q_2-q_1>\frac{\pi}{\sqrt{1-A_\nu/C_0^2}}>g. q 2 − q 1 > 1 − A ν / C 0 2 π > g .
Let R l = W ( q l ) / W ( κ l ) R_l=W(q_l)/W(\kappa_l) R l = W ( q l ) / W ( κ l ) . The exact quarter-phase identities give
I ∗ : = log q 2 / κ q / k = ∫ 0 π / 2 4 ( r 2 − r 1 ) ( 2 r 1 + m sin 2 ϕ ) ( 2 r 2 + m sin 2 ϕ ) d ϕ , I_*:=\log\frac{q_2/\kappa}{q/k}=\int_0^{\pi/2}\frac{4(r_2-r_1)}{(2r_1+m\sin2\phi)(2r_2+m\sin2\phi)}\,\mathrm{d}\phi, I ∗ := log q / k q 2 / κ = ∫ 0 π /2 ( 2 r 1 + m sin 2 ϕ ) ( 2 r 2 + m sin 2 ϕ ) 4 ( r 2 − r 1 ) d ϕ ,
E ∗ : = log R 1 R 2 = ∫ 0 π / 2 2 m sin 2 ϕ { 1 ( 2 r 1 + m sin 2 ϕ ) 2 − 1 ( 2 r 2 + m sin 2 ϕ ) 2 } d ϕ . E_*:=\log\frac{R_1}{R_2}=\int_0^{\pi/2}2m\sin2\phi\left\{\frac{1}{(2r_1+m\sin2\phi)^2}-\frac{1}{(2r_2+m\sin2\phi)^2}\right\}\,\mathrm{d}\phi. E ∗ := log R 2 R 1 = ∫ 0 π /2 2 m sin 2 ϕ { ( 2 r 1 + m sin 2 ϕ ) 2 1 − ( 2 r 2 + m sin 2 ϕ ) 2 1 } d ϕ .
For the first integral, bound r 1 < K 0 r_1<K_0 r 1 < K 0 , r 2 < C 0 r_2<C_0 r 2 < C 0 , and r 2 − r 1 > g r_2-r_1>g r 2 − r 1 > g . The function
t ⟼ 1 ( 2 K 0 + m t ) ( 2 C 0 + m t ) t\longmapsto\frac{1}{(2K_0+mt)(2C_0+mt)} t ⟼ ( 2 K 0 + m t ) ( 2 C 0 + m t ) 1
is convex for t ≥ 0 t\ge0 t ≥ 0 . The average of sin 2 φ \sin2\varphi sin 2 φ on this interval is 2 / π 2/\pi 2/ π . Jensen’s inequality, followed by p − < π p_-<\pi p − < π , yields I ∗ > I I_* > I I ∗ > I .
For the second integral use
1 a 2 − 1 b 2 = ( b − a ) ( a + b ) a 2 b 2 , b − a = 2 ( r 2 − r 1 ) > 2 g . \frac{1}{a^2}-\frac{1}{b^2}=\frac{(b-a)(a+b)}{a^2b^2},\qquad b-a=2(r_2-r_1)>2g. a 2 1 − b 2 1 = a 2 b 2 ( b − a ) ( a + b ) , b − a = 2 ( r 2 − r 1 ) > 2 g .
The function ( a + b ) / ( a 2 b 2 ) = 1 / ( a b 2 ) + 1 / ( a 2 b ) (a+b)/(a^2b^2)=1/(ab^2)+1/(a^2b) ( a + b ) / ( a 2 b 2 ) = 1/ ( a b 2 ) + 1/ ( a 2 b ) is decreasing in each positive variable. After their upper substitution, it is convex in t = sin 2 φ t=\sin2\varphi t = sin 2 φ : each summand is a product of positive inverse powers of positive affine functions, with nonnegative second derivative. The measure sin 2 φ d φ \sin2\varphi\,d\varphi sin 2 φ d φ has total mass one and mean value π / 4 \pi/4 π /4 for t t t . Jensen’s inequality and π < p + \pi<p_+ π < p + therefore give E ∗ > E E_*>E E ∗ > E .
Let θ \theta θ be the polar phase of ( X , p ) (X,p) ( X , p ) , and set ϑ = 3 π / 2 − θ ( x ) ∈ ( 0 , π / 2 ) \vartheta=3\pi/2-\theta(x)\in(0,\pi/2) ϑ = 3 π /2 − θ ( x ) ∈ ( 0 , π /2 ) and d = q 2 − x d=q_2-x d = q 2 − x . Since q 2 / x = e I ∗ q_2/x=e^{I_*} q 2 / x = e I ∗ , one has d > x I d>xI d > x I . Run the phase equation backwards from q 2 q_2 q 2 to x x x . At distance t t t from q 2 q_2 q 2 , its phase deficit satisfies
ϑ t = 1 − m sin 2 ϑ 2 ( q 2 − t ) ≥ 1 − m x ϑ . \vartheta_t=1-\frac{m\sin2\vartheta}{2(q_2-t)}\ge1-\frac{m}{x}\vartheta. ϑ t = 1 − 2 ( q 2 − t ) m sin 2 ϑ ≥ 1 − x m ϑ .
Integration and e − z ≤ ( 1 + z ) − 1 e^{-z}\le(1+z)^{-1} e − z ≤ ( 1 + z ) − 1 show
ϑ ≥ x m ( 1 − e − m d / x ) ≥ d 1 + m d / x > x I 1 + m I . \vartheta\ge\frac{x}{m}(1-e^{-md/x})\ge\frac{d}{1+md/x}>\frac{xI}{1+mI}. ϑ ≥ m x ( 1 − e − m d / x ) ≥ 1 + m d / x d > 1 + m I x I .
On the other hand ϑ ≤ d \vartheta\le d ϑ ≤ d . The quarter-cycle lengths strictly increase with l l l , so κ − q 2 > k − q = δ \kappa-q_2>k-q=\delta κ − q 2 > k − q = δ . Hence
ϑ ≤ d < δ κ − k k < d 0 π b − A < d 0 p + B 0 . \vartheta\le d<\delta\frac{\kappa-k}{k}<\frac{d_0\pi}{\sqrt{b-A}}<\frac{d_0p_+}{\sqrt{B_0}}. ϑ ≤ d < δ k κ − k < b − A d 0 π < B 0 d 0 p + .
Here the second strict bound is the Sturm spacing estimate for the order-n / 2 n/2 n /2 equation. Using sin t ≥ t − t 3 / 6 \sin t\ge t-t^3/6 sin t ≥ t − t 3 /6 gives
m sin 2 ϑ 2 x ≥ m ϑ x ( 1 − 2 ϑ 2 3 ) > m I 1 + m I c 0 = S . \frac{m\sin2\vartheta}{2x}\ge\frac{m\vartheta}{x}\left(1-\frac{2\vartheta^2}{3}\right)>\frac{mI}{1+mI}c_0=S. 2 x m sin 2 ϑ ≥ x m ϑ ( 1 − 3 2 ϑ 2 ) > 1 + m I m I c 0 = S .
The positive quantity
u = x W ′ ( x ) W ( x ) = m sin 2 ϑ / ( 2 x ) 1 − m sin 2 ϑ / ( 2 x ) u=\frac{xW'(x)}{W(x)}=\frac{m\sin2\vartheta/(2x)}{1-m\sin2\vartheta/(2x)} u = W ( x ) x W ′ ( x ) = 1 − m sin 2 ϑ / ( 2 x ) m sin 2 ϑ / ( 2 x )
therefore exceeds u 0 u_0 u 0 . Its denominator is positive because W ( x ) > 0 W(x)>0 W ( x ) > 0 .
Finally put σ = W ( x ) / ( R 1 W ( κ ) ) \sigma=W(x)/(R_1W(\kappa)) σ = W ( x ) / ( R 1 W ( κ )) . Since W ( x ) < W ( q 2 ) = R 2 W ( κ ) W(x)<W(q_2)=R_2W(\kappa) W ( x ) < W ( q 2 ) = R 2 W ( κ ) , one has 0 < σ < e − E ∗ < ( 1 + E ) − 1 0<\sigma<e^{-E_*}<(1+E)^{-1} 0 < σ < e − E ∗ < ( 1 + E ) − 1 . The exact expression for the first pole is
U 2 = 3 − σ ( 3 − u ) . U_2=3-\sigma(3-u). U 2 = 3 − σ ( 3 − u ) .
The hypotheses imply 0 < S < 1 / 2 0<S<1/2 0 < S < 1/2 and u 0 < 1 u_0<1 u 0 < 1 , hence U < 3 U<3 U < 3 . If u ≥ 3 u\ge3 u ≥ 3 the conclusion is immediate; if u < 3 u<3 u < 3 , the preceding upper bound for σ \sigma σ gives U 2 > ( u 0 + 3 E ) / ( 1 + E ) U_2>(u_0+3E)/(1+E) U 2 > ( u 0 + 3 E ) / ( 1 + E ) , as claimed.
We next replace the phase parameters by elementary bounds that hold on whole dimension intervals.
Lemma 3.8. The following choices satisfy every hypothesis of Lemma 3.7 throughout the indicated real intervals:
In addition V < ν V<\nu V < ν , where
ν = m 2 D 0 { d 0 ( 50 157 + 1 H 0 ) } 4 50 157 ( d 0 k 0 + 1 4 ) \nu=\frac{m}{2D_0}\left\{d_0\left(\frac{50}{157}+\frac{1}{H_0}\right)\right\}^{4}\frac{50}{157}\left(\frac{d_0}{k_0}+\frac{1}{4}\right) ν = 2 D 0 m { d 0 ( 157 50 + H 0 1 ) } 4 157 50 ( k 0 d 0 + 4 1 ) with the following choices, except that on [ 11 , 14 ] [11,14] [ 11 , 14 ] one can take the simpler bound ν = 3 / 20 \nu=3/20 ν = 3/20 :
Proof. Lemma 3.4 gives κ < L \kappa<L κ < L for every real n ≥ 6 n\geq6 n ≥ 6 . The Barta and quarter-phase bounds in the proof of Lemma 3.5 give b − A > B 0 b-A>B_0 b − A > B 0 and δ < d 0 \delta<d_0 δ < d 0 on these real intervals. Its Dirichlet trial gives b < b + = ( n + 2 ) ( n + 10 ) / 3 b<b_+=(n+2)(n+10)/3 b < b + = ( n + 2 ) ( n + 10 ) /3 . The two upper bounds for k k k follow from
{ 3 ( n + 5 ) 5 } 2 − b + = 2 n 2 − 30 n + 175 75 > 0 , \left\{\frac{3(n+5)}{5}\right\}^{2}-b_+=\frac{2n^2-30n+175}{75}>0, { 5 3 ( n + 5 ) } 2 − b + = 75 2 n 2 − 30 n + 175 > 0 ,
{ 7 n + 38 12 } 2 − b + = ( n − 22 ) 2 144 ≥ 0. \left\{\frac{7n+38}{12}\right\}^{2}-b_+=\frac{(n-22)^2}{144}\geq0. { 12 7 n + 38 } 2 − b + = 144 ( n − 22 ) 2 ≥ 0.
To bound the second zero, let f ( t ) = t + π t / t 2 − A f(t)=t+\pi t/\sqrt{t^2-A} f ( t ) = t + π t / t 2 − A . It is increasing for t ≥ k t\geq k t ≥ k . Indeed it suffices that π 2 A 2 < ( b − A ) 3 \pi^2A^2<(b-A)^3 π 2 A 2 < ( b − A ) 3 ; for 6 ≤ n ≤ 14 6\leq n\leq14 6 ≤ n ≤ 14 this follows from
π 2 A 2 < 5 n 4 8 ≤ 70 n 3 8 < ( 5 n 2 ) 3 < ( b − A ) 3 , \pi^2A^2<\frac{5n^4}{8}\leq\frac{70n^3}{8}<\left(\frac{5n}{2}\right)^3<(b-A)^3, π 2 A 2 < 8 5 n 4 ≤ 8 70 n 3 < ( 2 5 n ) 3 < ( b − A ) 3 ,
and for 15 ≤ n ≤ 47 15\leq n\leq47 15 ≤ n ≤ 47 from 5 n 4 / 8 ≤ 235 n 3 / 8 < ( 4 n ) 3 < ( b − A ) 3 5n^4/8\leq235n^3/8<(4n)^3<(b-A)^3 5 n 4 /8 ≤ 235 n 3 /8 < ( 4 n ) 3 < ( b − A ) 3 . Sturm comparison gives κ < f ( k ) < f ( b + ) \kappa<f(k)<f(\sqrt{b_+}) κ < f ( k ) < f ( b + ) . The function
R ( n ) = n 2 + 12 n + 20 n 2 + 48 n + 83 R(n)=\frac{n^2+12n+20}{n^2+48n+83} R ( n ) = n 2 + 48 n + 83 n 2 + 12 n + 20
is increasing, since the numerator of its derivative is 36 n 2 + 126 n + 36 36n^2+126n+36 36 n 2 + 126 n + 36 . Thus
f ( b + ) < K 0 + 2 p + R ( n ) < C 0 . f(\sqrt{b_+})<K_0+2p+\sqrt{R(n)}<C_0. f ( b + ) < K 0 + 2 p + R ( n ) < C 0 .
For the four intervals, the exact positive differences at their right endpoints between ( C 0 − K 0 ) 2 (C_0-K_0)^2 ( C 0 − K 0 ) 2 and 4 p + 2 R ( n ) 4p_+^2R(n) 4 p + 2 R ( n ) are respectively
37269 270725 , 720 15533 , 126453 236425 , 279 379 . \frac{37269}{270725},\quad\frac{720}{15533},\quad\frac{126453}{236425},\quad\frac{279}{379}. 270725 37269 , 15533 720 , 236425 126453 , 379 279 .
Clearly 0 < s < 1 0<s<1 0 < s < 1 . Also 0 < c 0 < 1 0<c_0<1 0 < c 0 < 1 : its smallest coarse comparison is the low-dimensional endpoint n = 6 n=6 n = 6 , where 2 ( 13 / 10 ) 2 p + 2 / ( 3 ⋅ 24 ) < 1 2(13/10)^2p_+^2/(3\cdot24)<1 2 ( 13/10 ) 2 p + 2 / ( 3 ⋅ 24 ) < 1 . For the remaining hypothesis, g < 3 π / 2 g<3\pi/2 g < 3 π /2 and K 0 ≥ ( n + 5 ) / 2 K_0\ge(n+5)/2 K 0 ≥ ( n + 5 ) /2 , C 0 > ( n + 11 ) / 2 C_0>(n+11)/2 C 0 > ( n + 11 ) /2 imply
m I < 30 ( n − 1 ) ( n + 5 ) ( n + 11 ) < 1. m_I<\frac{30(n-1)}{(n+5)(n+11)}<1. m I < ( n + 5 ) ( n + 11 ) 30 ( n − 1 ) < 1.
because ( n + 5 ) ( n + 11 ) − 30 ( n − 1 ) = ( n − 7 ) 2 + 36 > 0 (n+5)(n+11)-30(n-1)=(n-7)^2+36>0 ( n + 5 ) ( n + 11 ) − 30 ( n − 1 ) = ( n − 7 ) 2 + 36 > 0 . For the upper bound on V \mathcal V V , it remains to check D > D 0 \sqrt D>D_0 D > D 0 , k > k 0 k>k_0 k > k 0 , and b − A > H 0 \sqrt{b-A}>H_0 b − A > H 0 . The Hardy estimate and the quarter-phase bound give
D > ( 1 − 2 n ) 2 − ( d 0 ⋅ 50 157 ) 2 . \sqrt D>\sqrt{\left(1-\frac{2}{n}\right)^2-\left(d_0\cdot\frac{50}{157}\right)^2}. D > ( 1 − n 2 ) 2 − ( d 0 ⋅ 157 50 ) 2 .
For fixed d 0 d_0 d 0 the right side is concave in n n n where it is positive, so its endpoint lower chords apply. The squared positive differences at n = 6 , 10 n=6,10 n = 6 , 10 from ( 13 / 25 ) 2 (13/25)^2 ( 13/25 ) 2 and ( 17 / 25 ) 2 (17/25)^2 ( 17/25 ) 2 are, respectively,
365746 138650625 , 95414 15405625 . \frac{365746}{138650625},\quad\frac{95414}{15405625}. 138650625 365746 , 15405625 95414 .
For the high intervals, use endpoints 15 , 30 , 47 15,30,47 15 , 30 , 47 and chord values $39/50,17/20,22/25. The squared positive differences are
5951131 2218410000 , 76159 8873640 , 1239214471 544496410000 . \frac{5951131}{2218410000},\quad\frac{76159}{8873640},\quad\frac{1239214471}{544496410000}. 2218410000 5951131 , 8873640 76159 , 544496410000 1239214471 .
The printed D 0 D_0 D 0 are exactly these chords, using [ 15 , 30 ] [15,30] [ 15 , 30 ] for the third row. The Barta lower bound for b b b gives k > k 0 k>k_0 k > k 0 ; in the low interval the squared difference is ( 2093 − 200 n ) / 1250 > 0 (2093-200n)/1250>0 ( 2093 − 200 n ) /1250 > 0 , and in the high intervals it is ( 24 n − 233 ) / 20 > 0 (24n-233)/20>0 ( 24 n − 233 ) /20 > 0 . Finally the differences B 0 − H 0 2 B_0-H_0^2 B 0 − H 0 2 are concave quadratics. At the endpoints of [ 6 , 10 ] [6,10] [ 6 , 10 ] they equal 71 / 225 , 9 / 25 71/225,9/25 71/225 , 9/25 ; at those of [ 15 , 30 ] [15,30] [ 15 , 30 ] they equal 43 / 405 , 58 / 405 43/405,58/405 43/405 , 58/405 ; and at those of [ 30 , 47 ] [30,47] [ 30 , 47 ] they equal 58 / 405 , 179 / 324 58/405,179/324 58/405 , 179/324 . Hence all are positive. Substituting these bounds into the definition of V \mathcal V V proves V < V \mathcal V<\mathcal V V < V . The bound V < 3 / 20 \mathcal V<3/20 V < 3/20 on [ 11 , 14 ] [11,14] [ 11 , 14 ] was proved directly in Lemma 3.5 . □
We first compare the lower and upper bounds on the two lower-dimensional parameter intervals.
Lemma 3.9. For the parameters and scalar functions U , V \mathcal U,\mathcal V U , V in Lemma 3.8 , one has
U > V throughout [ 6 , 10 ] and throughout [ 11 , 14 ] . \mathcal U>\mathcal V\quad\text{throughout }[6,10]\text{ and throughout }[11,14]. U > V throughout [ 6 , 10 ] and throughout [ 11 , 14 ] .
Proof. The following polynomial recipe specifies both exact expansions used in the proof. Put
p = 157 50 , P = 22 7 , r = 50 157 , m = n − 1 , K 0 = 3 ( n + 5 ) 5 . p = \frac{157}{50},\quad P = \frac{22}{7},\quad r = \frac{50}{157},\quad m = n-1,\quad K_0 = \frac{3(n+5)}{5}. p = 50 157 , P = 7 22 , r = 157 50 , m = n − 1 , K 0 = 5 3 ( n + 5 ) .
On the first interval take C 0 = K 0 + 19 / 5 C_0=K_0+19/5 C 0 = K 0 + 19/5 and d 0 = 13 / 10 d_0=13/10 d 0 = 13/10 ; on the second take C 0 = K 0 + 4 C_0=K_0+4 C 0 = K 0 + 4 and d 0 = 11 / 9 d_0=11/9 d 0 = 11/9 . Define the polynomials
H = 2 C 0 2 + ( n − 2 ) 2 − 1 4 , I N = p 4 H , I D = 4 C 0 2 ( p K 0 + m ) ( p C 0 + m ) , D k = 2 K 0 + m P 4 , D c = 2 C 0 + m P 4 , E N = 2 m p H ( D k + D c ) , E D = C 0 2 D k 2 D c 2 , C D = 3 ( 5 n + 18 ) , C N = C D − 4 d 0 2 P 2 , u N = m I N C N , u D = I D C D + m I N ( C D − C N ) , U N = u N E D + 3 E N u D , U D = u D ( E D + E N ) . \begin{aligned}
H &= 2C_0^2+\frac{(n-2)^2-1}{4}, &
I_N &= p^4H, &
I_D &= 4C_0^2(pK_0+m)(pC_0+m), \\
D_k &= 2K_0+\frac{mP}{4}, &
D_c &= 2C_0+\frac{mP}{4}, \\
E_N &= 2mpH(D_k+D_c), &
E_D &= C_0^2D_k^2D_c^2, \\
C_D &= 3(5n+18), &
C_N &= C_D-4d_0^2P^2, \\
u_N &= mI_NC_N, &
u_D &= I_DC_D+mI_N(C_D-C_N), \\
U_N &= u_NE_D+3E_Nu_D, &
U_D &= u_D(E_D+E_N).
\end{aligned} H D k E N C D u N U N = 2 C 0 2 + 4 ( n − 2 ) 2 − 1 , = 2 K 0 + 4 m P , = 2 m p H ( D k + D c ) , = 3 ( 5 n + 18 ) , = m I N C N , = u N E D + 3 E N u D , I N D c E D C N u D U D = p 4 H , = 2 C 0 + 4 m P , = C 0 2 D k 2 D c 2 , = C D − 4 d 0 2 P 2 , = I D C D + m I N ( C D − C N ) , = u D ( E D + E N ) . I D = 4 C 0 2 ( p K 0 + m ) ( p C 0 + m ) ,
The formulas in Lemma 3.7 give exactly I = I N / I D I=I_N/I_D I = I N / I D , E = E N / E D E=E_N/E_D E = E N / E D , u 0 = u N / u D u_0=u_N/u_D u 0 = u N / u D , and U = U N / U D \mathcal{U}=U_N/U_D U = U N / U D . All displayed denominators are positive.
On [ 6 , 10 ] [6,10] [ 6 , 10 ] , let
V N = 25 m d 0 4 8 { r ( 7 n + 104 ) + 30 } 4 ( 25 n + 133 + 200 d 0 ) , V_N=\frac{25md_0^4}{8}\{r(7n+104)+30\}^4(25n+133+200d_0), V N = 8 25 m d 0 4 { r ( 7 n + 104 ) + 30 } 4 ( 25 n + 133 + 200 d 0 ) ,
V D = ( n + 7 ) ( 7 n + 104 ) 4 ( 25 n + 133 ) . V_D=(n+7)(7n+104)^4(25n+133). V D = ( n + 7 ) ( 7 n + 104 ) 4 ( 25 n + 133 ) .
Then V = V N / V D \mathcal{V}=V_N/V_D V = V N / V D exactly. On [ 11 , 14 ] [11,14] [ 11 , 14 ] , take simply V N = 3 V_N=3 V N = 3 , V D = 20 V_D=20 V D = 20 . In either case put
Q ( n ) = U N V D − V N U D . Q(n)=U_NV_D-V_NU_D. Q ( n ) = U N V D − V N U D .
Thus it suffices to prove that this explicitly defined polynomial is positive on the corresponding interval.
For the first interval write Q ( 6 + v ) = 10 25 ∑ j = 0 17 q j v j Q(6+v)=10^{25}\sum_{j=0}^{17}q_jv^j Q ( 6 + v ) = 1 0 25 ∑ j = 0 17 q j v j . Multiplication of the displayed polynomial factors gives the following strict rational lower bounds. Powers of ten in this table are exact.
Table 1.
Every term through degree fourteen is nonnegative for v ≥ 0 v\geq0 v ≥ 0 , with a positive constant. On 0 ≤ v ≤ 4 0\leq v\leq4 0 ≤ v ≤ 4 , the remaining terms obey
q 15 v 15 + q 16 v 16 + q 17 v 17 ≥ v 15 { 2 ⋅ 10 − 15 − 28 ⋅ 10 − 18 − 32 ⋅ 10 − 19 } ≥ 0. q_{15}v^{15}+q_{16}v^{16}+q_{17}v^{17}\geq v^{15}\{2\cdot10^{-15}-28\cdot10^{-18}-32\cdot10^{-19}\}\geq0. q 15 v 15 + q 16 v 16 + q 17 v 17 ≥ v 15 { 2 ⋅ 1 0 − 15 − 28 ⋅ 1 0 − 18 − 32 ⋅ 1 0 − 19 } ≥ 0.
Consequently Q > 0 Q > 0 Q > 0 on the entire first interval.
For the second interval write Q ( 11 + v ) = 10 16 ∑ j = 0 11 q j v j Q(11+v)=10^{16}\sum_{j=0}^{11}q_jv^j Q ( 11 + v ) = 1 0 16 ∑ j = 0 11 q j v j . The same factor recipe gives these strict lower bounds:
For 0 ≤ v ≤ 3 0\le v\le3 0 ≤ v ≤ 3 , group all terms from degree five onwards. Their coefficient after factoring v 5 v^5 v 5 is greater than
3 ⋅ 10 − 4 − 3 ( 9 ⋅ 10 − 6 ) − 9 ( 18 ⋅ 10 − 7 ) − 27 ⋅ 10 − 7 − 81 ( 3 ⋅ 10 − 9 ) − 243 ( 5 ⋅ 10 − 11 ) − 729 ( 4 ⋅ 10 − 13 ) > 1 4000 > 0. 3\cdot10^{-4}-3(9\cdot10^{-6})-9(18\cdot10^{-7})-27\cdot10^{-7}
-81(3\cdot10^{-9})-243(5\cdot10^{-11})-729(4\cdot10^{-13})>\frac{1}{4000}>0. 3 ⋅ 1 0 − 4 − 3 ( 9 ⋅ 1 0 − 6 ) − 9 ( 18 ⋅ 1 0 − 7 ) − 27 ⋅ 1 0 − 7 − 81 ( 3 ⋅ 1 0 − 9 ) − 243 ( 5 ⋅ 1 0 − 11 ) − 729 ( 4 ⋅ 1 0 − 13 ) > 4000 1 > 0.
The lower-degree terms are nonnegative with a positive constant. Thus Q > 0 Q>0 Q > 0 on the second interval as well. This proves the lemma.
The same comparison on the two remaining intervals completes the parameter estimates through dimension forty-seven.
Lemma 3.10. Use the quantities U , V U,V U , V in Lemmas 3.7 –3.8 . Then
U > V ( 15 ≤ n ≤ 29 or 30 ≤ n ≤ 47 ) . U>V\qquad(15\le n\le29\text{ or }30\le n\le47). U > V ( 15 ≤ n ≤ 29 or 30 ≤ n ≤ 47 ) .
Proof. We give factored polynomials and small coefficient bounds for the whole intervals. Put
p = 157 50 , z = 1 p , P = 22 7 , δ = 47 40 , m = n − 1 , p=\frac{157}{50},\qquad z=\frac{1}{p},\qquad P=\frac{22}{7},\qquad\delta=\frac{47}{40},\qquad m=n-1, p = 50 157 , z = p 1 , P = 7 22 , δ = 40 47 , m = n − 1 ,
K 0 = 7 n + 38 12 , B = 21 n − 12 5 , b ∗ = 2 δ 2 P 2 3 . K_0=\frac{7n+38}{12},\qquad B=\frac{21n-12}{5},\qquad b_*=\frac{2\delta^2P^2}{3}. K 0 = 12 7 n + 38 , B = 5 21 n − 12 , b ∗ = 3 2 δ 2 P 2 .
Choose the remaining affine polynomials by the following table:
Define
A ν = ( n − 2 ) 2 − 1 4 , R = 2 C 0 2 + A ν , B k = K 0 + m p , B c = C 0 + m p , A_\nu=\frac{(n-2)^2-1}{4},\qquad R=2C_0^2+A_\nu,\qquad B_k=K_0+\frac{m}{p},\qquad B_c=C_0+\frac{m}{p}, A ν = 4 ( n − 2 ) 2 − 1 , R = 2 C 0 2 + A ν , B k = K 0 + p m , B c = C 0 + p m ,
D k = 2 K 0 + m P 4 , D c = 2 C 0 + m P 4 , D_k=2K_0+\frac{mP}{4},\qquad D_c=2C_0+\frac{mP}{4}, D k = 2 K 0 + 4 m P , D c = 2 C 0 + 4 m P ,
I N = p 2 R , I D = 4 C 0 2 B k B c , E N = 2 m p R ( D k + D c ) , E D = C 0 2 D k 2 D c 2 , I_N=p^2R,\qquad I_D=4C_0^2B_kB_c,\qquad E_N=2mpR(D_k+D_c),\qquad E_D=C_0^2D_k^2D_c^2, I N = p 2 R , I D = 4 C 0 2 B k B c , E N = 2 m pR ( D k + D c ) , E D = C 0 2 D k 2 D c 2 ,
J = m I N ( B − b ∗ ) , Q = I D B + m I N b ∗ . J=mI_N(B-b_*),\qquad Q=I_DB+mI_Nb_*. J = m I N ( B − b ∗ ) , Q = I D B + m I N b ∗ .
Every denominator and every displayed factor is positive on the stated intervals. Direct substitution in the phase formulas gives
I = I N I D , E = E N E D , w 0 = J Q , U = J E D + 3 E N Q Q ( E D + E N ) . I = \frac{I_N}{I_D}, \qquad E = \frac{E_N}{E_D}, \qquad w_0 = \frac{J}{Q}, \qquad\mathcal{U} = \frac{J E_D + 3E_N Q}{Q(E_D + E_N)}. I = I D I N , E = E D E N , w 0 = Q J , U = Q ( E D + E N ) J E D + 3 E N Q .
The following additional polynomials remove a common positive factor m m m :
Z = Q ( E D + E N ) , Z = Q(E_D + E_N), Z = Q ( E D + E N ) ,
H = p R { p ( B − b ∗ ) E D + 6 ( D k + D c ) Q } , H = pR\{p(B-b_*)E_D + 6(D_k+D_c)Q\}, H = pR { p ( B − b ∗ ) E D + 6 ( D k + D c ) Q } ,
N = δ z 4 ( z H 0 + 1 ) 4 ( 4 δ + k 0 ) , T = 8 D 0 H 0 4 k 0 , Π = H T − N Z . N = \delta_z^4(zH_0+1)^4(4\delta+k_0), \qquad T = 8D_0H_0^4k_0, \qquad\Pi= HT-NZ. N = δ z 4 ( z H 0 + 1 ) 4 ( 4 δ + k 0 ) , T = 8 D 0 H 0 4 k 0 , Π = H T − N Z .
Indeed J E D + 3 E N Q = m H J E_D+3E_NQ=mH J E D + 3 E N Q = m H , so that the exact identity is
U − V = m Π ( n ) Z T . (26) \mathcal{U}-\mathcal{V}=\frac{m\Pi(n)}{ZT}. \tag*{(26)} U − V = Z T m Π ( n ) . (26)
In particular Π \Pi Π is a polynomial of degree 16 with completely specified rational coefficients. Here are the two endpoint checks used only to normalize this polynomial; all quantities in them are the rational functions just displayed:
Table 5.
They follow by multiplication of positive rational denominators. Thus Π ( 15 ) > 0 \Pi(15)>0 Π ( 15 ) > 0 in the first block, and Π ( 30 ) > 0 \Pi(30)>0 Π ( 30 ) > 0 in the second.
For a block [ ℓ , r ] [\ell,r] [ ℓ , r ] , set t = ( n − ℓ ) / ( r − ℓ ) t=(n-\ell)/(r-\ell) t = ( n − ℓ ) / ( r − ℓ ) and write
Π ( ℓ + ( r − ℓ ) t ) Π ( ℓ ) = ∑ j = 0 16 q j t j , q j = ( r − ℓ ) j Π ( j ) ( ℓ ) j ! Π ( ℓ ) . \frac{\Pi(\ell+(r-\ell)t)}{\Pi(\ell)}=\sum_{j=0}^{16}q_jt^j,\qquad q_j=\frac{(r-\ell)^j\Pi^{(j)}(\ell)}{j!\Pi(\ell)}. Π ( ℓ ) Π ( ℓ + ( r − ℓ ) t ) = j = 0 ∑ 16 q j t j , q j = j ! Π ( ℓ ) ( r − ℓ ) j Π ( j ) ( ℓ ) .
The following are strict lower bounds for q j q_j q j ; in both rows q 0 = 1 q_0=1 q 0 = 1 . They are obtained simply by expanding the factored polynomial Π \Pi Π above and using the displayed coefficient formula.
Table 6.
Table 7.
For the first interval and 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , discard the nine positive nonconstant terms and use t j ≤ 1 t^j\le1 t j ≤ 1 on the remaining terms. This gives
Π ( 15 + 14 t ) Π ( 15 ) ≥ 1 − 284 + 474 + 178 + 39 + 6 + 1 + 1 1000 = 17 1000 > 0. \frac{\Pi(15+14t)}{\Pi(15)}\ge1-\frac{284+474+178+39+6+1+1}{1000}=\frac{17}{1000}>0. Π ( 15 ) Π ( 15 + 14 t ) ≥ 1 − 1000 284 + 474 + 178 + 39 + 6 + 1 + 1 = 1000 17 > 0.
For the second interval, t j ≥ t 5 t^j \ge t^5 t j ≥ t 5 when 1 ≤ j ≤ 4 1 \le j \le4 1 ≤ j ≤ 4 , whereas t j ≤ t 5 t^j \le t^5 t j ≤ t 5 when j ≥ 5 j \ge5 j ≥ 5 . The positive lower bounds sum to 4 + 10 + 12 + 7 = 33 4 + 10 + 12 + 7 = 33 4 + 10 + 12 + 7 = 33 ; the absolute values of all negative lower bounds sum to 1 + 5 + 5 + 3 + 2 + 7 = 23 1 + 5 + 5 + 3 + 2 + 7 = 23 1 + 5 + 5 + 3 + 2 + 7 = 23 . Therefore
∏ ( 30 + 17 t ) ∏ ( 30 ) ≥ 1 + ( 33 − 23 ) t 5 ≥ 1 > 0. \frac{\prod(30+17t)}{\prod(30)} \ge1 + (33-23)t^5 \ge1 > 0. ∏ ( 30 ) ∏ ( 30 + 17 t ) ≥ 1 + ( 33 − 23 ) t 5 ≥ 1 > 0.
Equation (26) now proves both continuous blocks.
The preceding estimates now compare the first pole with the entire possible negative tail.
Lemma 3.11. For every integer 6 ≤ n ≤ 47 6 \le n \le47 6 ≤ n ≤ 47 , one has U 2 > V U_2 > V U 2 > V .
Proof. Choose the row of Lemma 3.8 containing n n n . Lemma 3.7 gives U 2 > U U_2 > \mathcal{U} U 2 > U and Lemma 3.8 gives V < V V < \mathcal{V} V < V . Lemmas 3.9 and 3.10 give U > V \mathcal{U} > \mathcal{V} U > V in all four rows. Hence U 2 > V U_2 > V U 2 > V .
The residue formula and this comparison establish all coefficient signs in the higher-dimensional range.
Theorem 3.12. For every integer 6 ≤ n ≤ 47 6 \le n \le47 6 ≤ n ≤ 47 the calibration has a convergent expansion on 0 ≤ r < κ 2 0 \le r < \kappa_2 0 ≤ r < κ 2 ,
G ( r ) = n λ − K m − ∑ j ≥ 1 a j r 2 j , a j > 0. G(r) = \frac{n\lambda-K}{m} - \sum_{j\ge1} a_j r^{2j}, \qquad a_j > 0. G ( r ) = m nλ − K − j ≥ 1 ∑ a j r 2 j , a j > 0. In particular on 0 < r < k 0 < r < k 0 < r < k ,
0 < G < K , G r < 0 , G r r < 0 , G r ( k ) = − 2 K 3 k . 0 < G < K, \qquad G_r < 0, \qquad G_{rr} < 0, \qquad G_r(k) = -\frac{2K}{3k}. 0 < G < K , G r < 0 , G r r < 0 , G r ( k ) = − 3 k 2 K .
Proof. The points 0 , ± k 0,\pm k 0 , ± k are removable. At k k k one has Q = Q r = 0 Q=Q_r=0 Q = Q r = 0 , Q r r = 2 m K c 2 / k 2 Q_{rr}=2mKc^2/k^2 Q r r = 2 m K c 2 / k 2 , and hence G ( k ) = 0 G(k)=0 G ( k ) = 0 . For j ≥ 2 j\ge2 j ≥ 2 the paired residue identity is
a j = ∑ l ≥ 2 2 m X ( κ l ) κ l 2 j { ( 2 j − 1 ) Q ( κ l ) + m α κ l X ( α κ l ) p ( α κ l ) } . a_j=\sum_{l\ge2}\frac{2}{mX(\kappa_l)\kappa_l^{2j}}\left\{(2j-1)Q(\kappa_l)+\frac{m\alpha}{\kappa_l}X(\alpha\kappa_l)p(\alpha\kappa_l)\right\}. a j = l ≥ 2 ∑ m X ( κ l ) κ l 2 j 2 { ( 2 j − 1 ) Q ( κ l ) + κ l m α X ( α κ l ) p ( α κ l ) } .
To check the residue formula, the double- and simple-pole coefficients of K − G K-G K − G at r = κ r=\kappa r = κ are
A κ = κ 2 Q ( κ ) m X ( κ ) 2 , D κ = κ 2 Q r ( κ ) + ( m + 2 ) κ Q ( κ ) m X ( κ ) 2 . A_\kappa=\frac{\kappa^2Q(\kappa)}{mX(\kappa)^2}, \qquad D_\kappa=\frac{\kappa^2Q_r(\kappa)+(m+2)\kappa Q(\kappa)}{mX(\kappa)^2}. A κ = m X ( κ ) 2 κ 2 Q ( κ ) , D κ = m X ( κ ) 2 κ 2 Q r ( κ ) + ( m + 2 ) κ Q ( κ ) .
The paired residue at ± κ \pm\kappa ± κ is 2 { ( 2 j + 1 ) A κ / κ 2 j + 2 − D κ / κ 2 j + 1 } 2\{(2j+1)A_\kappa/\kappa^{2j+2}-D_\kappa/\kappa^{2j+1}\} 2 {( 2 j + 1 ) A κ / κ 2 j + 2 − D κ / κ 2 j + 1 } . Substitution gives the stated formula, using p r ( κ ) = X ( κ ) p_r(\kappa)=X(\kappa) p r ( κ ) = X ( κ ) , p r r ( κ ) / p r ( κ ) = − m / κ p_{rr}(\kappa)/p_r(\kappa)=-m/\kappa p r r ( κ ) / p r ( κ ) = − m / κ and
Q r ( κ ) = − m Q ( κ ) / κ − m α X ( α κ ) p ( α κ ) / κ 2 . Q_r(\kappa)=-mQ(\kappa)/\kappa-m\alpha X(\alpha\kappa)p(\alpha\kappa)/\kappa^2. Q r ( κ ) = − m Q ( κ ) / κ − m α X ( α κ ) p ( α κ ) / κ 2 .
To justify the identity, integrate G ( r ) / r 2 j + 1 G(r)/r^{2j+1} G ( r ) / r 2 j + 1 around squares whose vertical sides are
ℜ r = ± ( N + n + 1 4 ) π \Re r=\pm\left(N+\frac{n+1}{4}\right)\pi ℜ r = ± ( N + 4 n + 1 ) π
and whose horizontal sides have the same absolute height. The Bessel asymptotics in [12 Section 10.17] give ∣ p ( r ) ∣ ≥ c 0 ∣ r ∣ − m / 2 e ∣ ℑ r ∣ |p(r)| \ge c_0 |r|^{-m/2}e^{|\Im r|} ∣ p ( r ) ∣ ≥ c 0 ∣ r ∣ − m /2 e ∣ℑ r ∣ and G = O ( ∣ r ∣ 2 ) G = O(|r|^2) G = O ( ∣ r ∣ 2 ) there. The contour integral tends to zero for j ≥ 2 j \ge2 j ≥ 2 . The residue series converges absolutely because Q ( κ ) = O ( κ − m ) Q(\kappa) = O(\kappa^{-m}) Q ( κ ) = O ( κ − m ) , X ( κ ) 2 ≍ κ − m X(\kappa)^2 \asymp\kappa^{-m} X ( κ ) 2 ≍ κ − m and X ( α κ ) p ( α κ ) = O ( κ − m ) X(\alpha\kappa)p(\alpha\kappa) = O(\kappa^{-m}) X ( α κ ) p ( α κ ) = O ( κ − m ) . Every brace is T l + 2 ( j − 2 ) Q ( κ l ) T_{l+2}(j-2)Q(\kappa_l) T l + 2 ( j − 2 ) Q ( κ l ) . Lemma 3.11 gives U 2 > V U_2 > V U 2 > V . Together with Lemmas 3.2, 3.4 , and 3.6 , this proves a j > 0 a_j > 0 a j > 0 . The center expansion supplies
a 1 = K − n λ + m λ 2 m ( n + 2 ) > 0 a_1 = \frac{K-n\lambda+m\lambda^2}{m(n+2)} > 0 a 1 = m ( n + 2 ) K − nλ + m λ 2 > 0
by Lemma 3.3, independently of Lemma 3.11 . The positivity of the a j a_j a j and G ( k ) = 0 G(k)=0 G ( k ) = 0 give G > 0 G>0 G > 0 , G r < 0 G_r<0 G r < 0 , G r r < 0 G_{rr}<0 G r r < 0 on ( 0 , k ) (0,k) ( 0 , k ) .
The endpoint expansion in § 2.1 , derived directly from the radial equations, gives G r ( k ) = − 2 K / ( 3 k ) G_r(k)=-2K/(3k) G r ( k ) = − 2 K / ( 3 k ) . Since k < κ 2 k<\kappa_2 k < κ 2 , the coefficient series may be differentiated at k k k , and therefore
∑ j ≥ 1 j a j k 2 j = − k G r ( k ) 2 = K 3 . \sum_{j\ge1} ja_j k^{2j}=-\frac{kG_r(k)}{2}=\frac{K}{3}. j ≥ 1 ∑ j a j k 2 j = − 2 k G r ( k ) = 3 K .
Together with G ( k ) = 0 G(k)=0 G ( k ) = 0 , this gives, for 0 < r < k 0<r<k 0 < r < k ,
0 < G ( r ) < ∑ j ≥ 1 a j k 2 j < K 3 < K . 0<G(r)<\sum_{j\ge1}a_j k^{2j}<\frac{K}{3}<K. 0 < G ( r ) < j ≥ 1 ∑ a j k 2 j < 3 K < K .
The penultimate inequality is strict because a 2 > 0 a_2>0 a 2 > 0 .
The coefficients in dimensions four and five
We now establish the coefficient signs in dimensions four and five using the same residue formula and tail criterion. The phase lemma applies once the condition κ 2 < L \kappa_2<L κ 2 < L has been checked.
For dimensions four and five, the same argument uses the following explicit root bounds.
Lemma 3.13. Write q = j n / 2 − 1 , 1 q=j_{n/2-1,1} q = j n /2 − 1 , 1 , k = j n / 2 , 1 k=j_{n/2,1} k = j n /2 , 1 and A = ( n 2 − 1 ) / 4 A=(n^2-1)/4 A = ( n 2 − 1 ) /4 . The data
Table 8.
satisfy q > q − q>q_- q > q − , k − < k < k + k_-<k<k_+ k − < k < k + , q < 3.84 q<3.84 q < 3.84 for n = 4 n=4 n = 4 , and q < 4.50 q<4.50 q < 4.50 for n = 5 n=5 n = 5 . With L = π / ( 1 − q / k ) L=\pi/(1-q/k) L = π / ( 1 − q / k ) and D = ( q / k ) 2 − A / L 2 D=(q/k)^2-A/L^2 D = ( q / k ) 2 − A / L 2 ,
κ 2 < C 0 < L , D > d D 2 > 0. \kappa_2<C_0<L,\qquad D>d_D^2>0. κ 2 < C 0 < L , D > d D 2 > 0.
Proof. For d ≥ 3 d \ge3 d ≥ 3 put
P d , N ( r ) = ∑ j = 0 N ( − 1 ) j r 2 j 4 j j ! ( d / 2 ) j . P_{d,N}(r)=\sum_{j=0}^{N}\frac{(-1)^j r^{2j}}{4^j j!(d/2)_j}. P d , N ( r ) = j = 0 ∑ N 4 j j ! ( d /2 ) j ( − 1 ) j r 2 j .
Here ( a ) j = a ( a + 1 ) ⋯ ( a + j − 1 ) (a)_j=a(a+1)\cdots(a+j-1) ( a ) j = a ( a + 1 ) ⋯ ( a + j − 1 ) for j ≥ 1 j\ge1 j ≥ 1 and ( a ) 0 = 1 (a)_0=1 ( a ) 0 = 1 is the rising factorial. The decreasing alternating tail gives P d , 9 < ψ d < P d , 8 P_{d,9}<\psi_d<P_{d,8} P d , 9 < ψ d < P d , 8 at all arguments in the following exact rational table:
Table 9.
Each comparison is specified by substituting into the displayed finite sum and clearing its positive denominator. The initial sign is positive, so these positive-to-negative crossings have odd index. Consecutive positive zeros are separated by more than π \pi π ; a third zero is greater than 2 π 2\pi 2 π , whereas all the table endpoints are less than 6 < 2 π 6<2\pi 6 < 2 π . Thus the crossings identify first zeros. Use p − = 157 / 50 < π < p + = 22 / 7 p_-=157/50<\pi<p_+=22/7 p − = 157/50 < π < p + = 22/7 , as established at the start of § 3 .
Put B 0 = k − 2 − A B_0=k_-^2-A B 0 = k − 2 − A and L 0 = p − / ( 1 − q − / k + ) L_0=p_-/(1-q_-/k_+) L 0 = p − / ( 1 − q − / k + ) . Sturm comparison gives
κ 2 < k + + π k − k − 2 − A < k + + p + k − k − 2 − A < C 0 . \kappa_2<k_++\frac{\pi k_-}{\sqrt{k_-^2-A}}<k_++\frac{p_+k_-}{\sqrt{k_-^2-A}}<C_0. κ 2 < k + + k − 2 − A π k − < k + + k − 2 − A p + k − < C 0 .
The last comparison follows from the following positive values of ( C 0 − k + ) 2 B 0 − p + 2 k − 2 (C_0-k_+)^2B_0-p_+^2k_-^2 ( C 0 − k + ) 2 B 0 − p + 2 k − 2 , for n = 4 , 5 n=4,5 n = 4 , 5 respectively:
93456794469 1225000000 , 24281435853 153125000 . \frac{93456794469}{1225000000},\qquad\frac{24281435853}{153125000}. 1225000000 93456794469 , 153125000 24281435853 .
Also L > L 0 L>L_0 L > L 0 , whose respective values are 40349 / 3275 40349/3275 40349/3275 and 90589 / 6400 90589/6400 90589/6400 ; each exceeds C 0 C_0 C 0 . Finally D > ( q − / k + ) 2 − A / L 0 2 > d D 2 D>(q_-/k_+)^2-A/L_0^2>d_D^2 D > ( q − / k + ) 2 − A / L 0 2 > d D 2 ; the two positive margins in the last comparison are
24670408577 2035052251250 , 1718746495 131301870736 . \frac{24670408577}{2035052251250},\qquad\frac{1718746495}{131301870736}. 2035052251250 24670408577 , 131301870736 1718746495 .
This proves all assertions.
These root bounds retain the strict first-pole dominance in dimensions four and five.
Lemma 3.14. For both dimensions, U 2 > V U_2>V U 2 > V in Lemma 3.6 .
Proof. Apply Lemma 3.7 with K 0 = k + K_0=k_+ K 0 = k + , C 0 C_0 C 0 from Lemma 3.13, d 0 = k + − q − d_0=k_+-q_- d 0 = k + − q − and B 0 = k − 2 − A B_0=k_-^2-A B 0 = k − 2 − A . Substitution in its displayed rational formulas gives
m I > 6 / 25 , c 0 > 49 / 100 , E > 1 / 61. m_I>6/25,\qquad c_0>49/100,\qquad E>1/61. m I > 6/25 , c 0 > 49/100 , E > 1/61.
For these data, 0 < s < 1 0 < s < 1 0 < s < 1 , K 0 ≥ ( n + 5 ) / 2 K_0 \ge(n + 5)/2 K 0 ≥ ( n + 5 ) /2 and C 0 > ( n + 11 ) / 2 C_0 > (n + 11)/2 C 0 > ( n + 11 ) /2 . Since g < 3 π / 2 g < 3\pi/2 g < 3 π /2 , the formula for I I I gives m I < 30 ( n − 1 ) / [ ( n + 5 ) ( n + 11 ) ] < 1 mI < 30(n - 1)/[(n + 5)(n + 11)] < 1 m I < 30 ( n − 1 ) / [( n + 5 ) ( n + 11 )] < 1 . Also c 0 < 1 c_0 < 1 c 0 < 1 by its definition, so every hypothesis of the phase lemma holds. In particular 0 < S < 1 / 2 0 < S < 1/2 0 < S < 1/2 , and the phase estimate gives
U 2 > U > 108 / 713 > 3 / 25. U_2 > U > 108/713 > 3/25. U 2 > U > 108/713 > 3/25.
Here S = m I c 0 / ( 1 + m I ) S = mIc_0/(1 + mI) S = m I c 0 / ( 1 + m I ) , u 0 = S / ( 1 − S ) u_0 = S/(1 - S) u 0 = S / ( 1 − S ) and U = ( u 0 + 3 E ) / ( 1 + E ) U = (u_0 + 3E)/(1 + E) U = ( u 0 + 3 E ) / ( 1 + E ) increase in the three lower inputs; for the displayed data u 0 < 1 < 3 u_0 < 1 < 3 u 0 < 1 < 3 .
The tail is smaller than
W n = m 2 d D ( C 0 L 0 ) 4 ( 1 L 0 + 1 4 p − ) . \mathcal{W}_n = \frac{m}{2d_D}\left(\frac{C_0}{L_0}\right)^4\left(\frac{1}{L_0} + \frac{1}{4p_-}\right). W n = 2 d D m ( L 0 C 0 ) 4 ( L 0 1 + 4 p − 1 ) .
The exact positive differences 11 / 100 − W n 11/100 - \mathcal{W}_n 11/100 − W n for n = 4 , 5 n = 4,5 n = 4 , 5 are
312567793029496310535353 21389167149811660795349800 , 6206044125316845926731439 610066710957425872061194900 . \frac{312567793029496310535353}{21389167149811660795349800}, \qquad
\frac{6206044125316845926731439}{610066710957425872061194900}. 21389167149811660795349800 312567793029496310535353 , 610066710957425872061194900 6206044125316845926731439 .
Thus U 2 > 3 / 25 > 11 / 100 > V U_2 > 3/25 > 11/100 > V U 2 > 3/25 > 11/100 > V , as required.
The coefficient signs in these two dimensions follow from the same residue structure.
Theorem 3.15. The conclusions of Theorem 3.12 , including every coefficient sign, 0 < G < K 0 < G < K 0 < G < K and G r ( k ) = − 2 K / ( 3 k ) G_r(k) = -2K/(3k) G r ( k ) = − 2 K / ( 3 k ) , also hold for n = 4 , 5 n = 4,5 n = 4 , 5 .
Proof. The residue identity, removable singularities, contour limit and absolute convergence in that theorem apply unchanged. Lemmas 3.2, 3.6 and 3.14 give a j > 0 a_j > 0 a j > 0 for all j ≥ 2 j \ge2 j ≥ 2 . For a 1 a_1 a 1 , the independent quarter-cycle comparison gives K > λ 1 − m / 2 K > \lambda^{1-m/2} K > λ 1 − m /2 . The root bounds give λ < 9 / 16 \lambda< 9/16 λ < 9/16 for n = 4 n = 4 n = 4 and λ < 16 / 25 \lambda< 16/25 λ < 16/25 for n = 5 n = 5 n = 5 . Hence
K > λ − 1 / 2 > 4 / 3 ≥ 4 λ − 3 λ 2 , n = 4 , K > \lambda^{-1/2} > 4/3 \ge4\lambda- 3\lambda^2,\qquad n = 4, K > λ − 1/2 > 4/3 ≥ 4 λ − 3 λ 2 , n = 4 ,
K > λ − 1 > 25 / 16 ≥ 5 λ − 4 λ 2 , n = 5. K > \lambda^{-1} > 25/16 \ge5\lambda- 4\lambda^2,\qquad n = 5. K > λ − 1 > 25/16 ≥ 5 λ − 4 λ 2 , n = 5.
Thus a 1 = ( K − n λ + m λ 2 ) / [ m ( n + 2 ) ] > 0 a_1 = (K - n\lambda+ m\lambda^2)/[m(n + 2)] > 0 a 1 = ( K − nλ + m λ 2 ) / [ m ( n + 2 )] > 0 in both cases. The endpoint expansion and the positive-coefficient argument at the end of that proof give all remaining conclusions, including G < K G < K G < K .
Radial bounds and determinant estimates
The coefficient signs from § 3 and Appendix § 6 reduce the matrix problem to bounds for the radial multiplier and the normalized Dirichlet profile. We first prove the product and multiplier estimates, then establish the scalar exponential inequalities used in the determinant argument.
Product expansions and radial bounds
For n ≥ 2 n \ge2 n ≥ 2 let x = r 2 x=r^2 x = r 2 , b = k 2 b=k^2 b = k 2 , and F ( x ) = r ϕ ( r ) / p ( r ) F(x)=r\phi(r)/p(r) F ( x ) = r ϕ ( r ) / p ( r ) . The zero-ratio comparison proved in Appendix § 6 for n = 2 n=2 n = 2 , and in Lemma 3.2 for n ≥ 3 n \ge3 n ≥ 3 , yields
F ( x ) = n ∏ l ≥ 2 1 − x / ( q l / α ) 2 1 − x / κ l 2 , ( q l / α ) 2 > κ l 2 . F(x)=n\prod_{l\ge2}\frac{1-x/(q_l/\alpha)^2}{1-x/\kappa_l^2},\qquad(q_l/\alpha)^2>\kappa_l^2. F ( x ) = n l ≥ 2 ∏ 1 − x / κ l 2 1 − x / ( q l / α ) 2 , ( q l / α ) 2 > κ l 2 .
Each factor has strictly positive Taylor coefficients after its constant term. The products converge uniformly on compact subsets of ∣ x ∣ < κ 2 2 |x|<\kappa_2^2 ∣ x ∣ < κ 2 2 . Thus F F F has strictly positive Taylor coefficients there. In particular it is absolutely monotone: every derivative is nonnegative on [ 0 , b ] [0,b] [ 0 , b ] .
We next examine the coefficients of the radial parameter. Put x = r 2 x=r^2 x = r 2 , b = k 2 b=k^2 b = k 2 , p = r h ( x ) p=rh(x) p = r h ( x ) and T ( x ) = r t = u / h T(x)=rt=u/h T ( x ) = r t = u / h . Let κ l \kappa_l κ l be the positive zeros of p / r p/r p / r , with κ 1 = k \kappa_1=k κ 1 = k .
The sign of this series will control the radial multiplier through the logarithmic derivative of T T T .
Proposition 4.1. In every dimension n ≥ 2 n\ge2 n ≥ 2 ,
T ( x ) = n ( 1 + c ) − ∑ j ≥ 1 s j x j , s j > 0 , ∣ x ∣ < κ 2 2 . T(x)=n(1+c)-\sum_{j\ge1}s_jx^j,\qquad s_j>0,\qquad|x|<\kappa_2^2. T ( x ) = n ( 1 + c ) − j ≥ 1 ∑ s j x j , s j > 0 , ∣ x ∣ < κ 2 2 . Consequently T ′ < 0 T'<0 T ′ < 0 and T ′ ′ < 0 T''<0 T ′′ < 0 on 0 < x < b 0<x<b 0 < x < b . In fact every positive-order derivative of T T T is strictly negative on this interval.
Proof. Write
h ( x ) = 1 n ( 1 − x / b ) P ( x ) , P ( x ) = ∏ l ≥ 2 ( 1 − x / κ l 2 ) . h(x)=\frac{1}{n}(1-x/b)P(x),\qquad P(x)=\prod_{l\ge2}(1-x/\kappa_l^2). h ( x ) = n 1 ( 1 − x / b ) P ( x ) , P ( x ) = l ≥ 2 ∏ ( 1 − x / κ l 2 ) .
The radial equation gives X = n h + 2 x h ′ X=nh+2xh' X = nh + 2 x h ′ . Evaluating at b b b proves
n c = 2 P ( b ) . nc=2P(b). n c = 2 P ( b ) .
Hence
T = n + 2 x h ′ h + c h . T=n+2x\frac{h'}{h}+\frac{c}{h}. T = n + 2 x h h ′ + h c .
For the coefficient computation normalize a l = b / κ l 2 a_l=b/\kappa_l^2 a l = b / κ l 2 , so a 1 = 1 a_1=1 a 1 = 1 and 0 < a l < 1 0<a_l<1 0 < a l < 1 for l ≥ 2 l\ge2 l ≥ 2 , and put P 0 = ∏ l ≥ 2 ( 1 − a l ) > 0 P_0=\prod_{l\ge2}(1-a_l)>0 P 0 = ∏ l ≥ 2 ( 1 − a l ) > 0 . Let p j = ∑ l ≥ 1 a l j p_j=\sum_{l\ge1}a_l^j p j = ∑ l ≥ 1 a l j . Define the complete homogeneous symmetric functions h j h_j h j by
∏ l ≥ 1 ( 1 − a l z ) − 1 = ∑ j ≥ 0 h j z j . \prod_{l\ge1}(1-a_lz)^{-1}=\sum_{j\ge0}h_jz^j. l ≥ 1 ∏ ( 1 − a l z ) − 1 = j ≥ 0 ∑ h j z j .
Thus h j h_j h j is the sum of all degree-j j j monomials with nondecreasing indices, with h 0 = 1 h_0=1 h 0 = 1 . The absolutely convergent product expansions give, for every j ≥ 1 j\ge1 j ≥ 1 ,
b j [ x j ] T = 2 ( P 0 h j − p j ) . b^j[x^j]T=2(P_0h_j-p_j). b j [ x j ] T = 2 ( P 0 h j − p j ) .
Since a 1 = 1 a_1=1 a 1 = 1 ,
h j ( a 1 , a 2 , … ) = ∑ i = 0 j h i ( a 2 , a 3 , … ) < ∑ i ≥ 0 h i ( a 2 , a 3 , … ) = 1 P 0 . h_j(a_1,a_2,\ldots)=\sum_{i=0}^{j}h_i(a_2,a_3,\ldots)<\sum_{i\geq0}h_i(a_2,a_3,\ldots)=\frac{1}{P_0}. h j ( a 1 , a 2 , … ) = i = 0 ∑ j h i ( a 2 , a 3 , … ) < i ≥ 0 ∑ h i ( a 2 , a 3 , … ) = P 0 1 .
Also p j > 1 p_j>1 p j > 1 , so every positive-order coefficient of T T T is strictly negative. The singularity at x = b x=b x = b is removable because u u u has a double zero while h h h has a simple zero there; the next possible poles are at κ 2 2 \kappa_2^2 κ 2 2 . Thus the Taylor expansion has the stated disk of convergence. The derivative signs follow termwise.
Let x = r 2 x=r^2 x = r 2 , b = k 2 b=k^2 b = k 2 , p = r h ( x ) p=rh(x) p = r h ( x ) , and write the Bessel product as
h ( x ) = 1 n ( 1 − x / b ) P ( x ) , P ( x ) = ∏ l ≥ 2 ( 1 − x / κ l 2 ) . h(x)=\frac{1}{n}(1-x/b)P(x),\qquad P(x)=\prod_{l\geq2}(1-x/\kappa_l^2). h ( x ) = n 1 ( 1 − x / b ) P ( x ) , P ( x ) = l ≥ 2 ∏ ( 1 − x / κ l 2 ) .
Set T ( x ) = r t = u / h T(x)=rt=u/h T ( x ) = r t = u / h and C = u p / S C=up/S C = u p / S . The product representation yields the lower multiplier bound used in the coupling estimate.
Lemma 4.2. On 0 < r < k 0<r<k 0 < r < k , in every dimension n ≥ 2 n\geq2 n ≥ 2 ,
C r > k 2 − r 2 k 2 + r 2 . \frac{C}{r}>\frac{k^2-r^2}{k^2+r^2}. r C > k 2 + r 2 k 2 − r 2 . Moreover T ( x ) / ( b − x ) T(x)/(b-x) T ( x ) / ( b − x ) has nonnegative Taylor coefficients, and its derivative is strictly positive on [ 0 , b ] [0,b] [ 0 , b ] .
Proof. Since u ( r ) = ∫ r k p ( s ) d s u(r)=\int_r^k p(s)\,\mathrm{d}s u ( r ) = ∫ r k p ( s ) d s , substitution of the product gives
R ( x ) : = T ( x ) b − x = 1 2 ∫ 0 1 ( 1 − v ) P ( x + ( b − x ) v ) P ( x ) d v . R(x):=\frac{T(x)}{b-x}=\frac{1}{2}\int_0^1(1-v)\frac{P(x+(b-x)v)}{P(x)}\,\mathrm{d}v. R ( x ) := b − x T ( x ) = 2 1 ∫ 0 1 ( 1 − v ) P ( x ) P ( x + ( b − x ) v ) d v .
Each factor of the quotient is
1 − v ( b − x ) κ l 2 − x = ( 1 − v ) + v ( κ l 2 − b ) κ l 2 − x . 1-\frac{v(b-x)}{\kappa_l^2-x}=(1-v)+\frac{v(\kappa_l^2-b)}{\kappa_l^2-x}. 1 − κ l 2 − x v ( b − x ) = ( 1 − v ) + κ l 2 − x v ( κ l 2 − b ) .
It has nonnegative Taylor coefficients and strictly positive x x x derivative when 0 < v < 1 0<v<1 0 < v < 1 . Uniform convergence on compact subintervals of x < κ 2 2 x<\kappa_2^2 x < κ 2 2 justifies multiplication and differentiation of the products, and then integration. Thus R ′ > 0 R'>0 R ′ > 0 .
Direct radial differentiation gives
C r = 1 1 − 2 x T ′ / T , − 2 x T ′ T = 2 x b − x − 2 x R ′ R < 2 x b − x . \frac{C}{r}=\frac{1}{1-2xT'/T},\qquad-\frac{2xT'}{T}=\frac{2x}{b-x}-\frac{2xR'}{R}<\frac{2x}{b-x}. r C = 1 − 2 x T ′ / T 1 , − T 2 x T ′ = b − x 2 x − R 2 x R ′ < b − x 2 x .
The first denominator is positive because S > 0 S>0 S > 0 . Taking reciprocals proves the strict lower bound.
The lower multiplier bound gives the upper estimate for the coupling coefficient.
Corollary 4.3. Suppose G = ∑ j ≥ 1 B j ( 1 − ( r / k ) 2 j ) G = \sum_{j \ge1} B_j\left(1-(r/k)^{2j}\right) G = ∑ j ≥ 1 B j ( 1 − ( r / k ) 2 j ) with B j > 0 B_j > 0 B j > 0 and G r ( k ) = − 2 K / ( 3 k ) G_r(k) = -2K/(3k) G r ( k ) = − 2 K / ( 3 k ) . Then, in every such dimension,
d = 2 G + C G r < 2 K 3 1 − ( r / k ) 2 1 + ( r / k ) 2 d = 2G + CG_r < \frac{2K}{3}\frac{1-(r/k)^2}{1+(r/k)^2} d = 2 G + C G r < 3 2 K 1 + ( r / k ) 2 1 − ( r / k ) 2
Proof. Put y = r 2 / k 2 y = r^2/k^2 y = r 2 / k 2 . The endpoint derivative gives ∑ j ≥ 1 j B j = K / 3 \sum_{j \ge1} jB_j = K/3 ∑ j ≥ 1 j B j = K /3 , while the preceding lemma gives
d 2 < ∑ j ≥ 1 j B j { 1 − y j j − 1 − y 1 + y y j } . \frac{d}{2} < \sum_{j \ge1} jB_j\left\{\frac{1-y^j}{j}-\frac{1-y}{1+y}y^j\right\}. 2 d < j ≥ 1 ∑ j B j { j 1 − y j − 1 + y 1 − y y j } .
For every integer j ≥ 1 j \ge1 j ≥ 1 ,
( 1 + y ) ( 1 − y j ) ≤ j ( 1 − y ) ( 1 + y j ) . (1+y)(1-y^j) \le j(1-y)(1+y^j). ( 1 + y ) ( 1 − y j ) ≤ j ( 1 − y ) ( 1 + y j ) .
Indeed summing y i + y j − i ≤ 1 + y j y^i+y^{j-i}\le1+y^j y i + y j − i ≤ 1 + y j over 1 ≤ i < j 1\le i<j 1 ≤ i < j gives the equivalent geometric-sum inequality. Hence each brace is at most ( 1 − y ) / ( 1 + y ) (1-y)/(1+y) ( 1 − y ) / ( 1 + y ) . Substitution of the endpoint sum proves the claim. □
Let x = r 2 x=r^2 x = r 2 , b = k 2 b=k^2 b = k 2 , y = x / b y=x/b y = x / b , and use R ( x ) = T ( x ) / ( b − x ) R(x)=T(x)/(b-x) R ( x ) = T ( x ) / ( b − x ) from the lower multiplier bound.
Logarithmic convexity of R R R gives a complementary upper multiplier bound through dimension twenty-two.
Proposition 4.4. In every dimension 2 ≤ n ≤ 22 2\le n\le22 2 ≤ n ≤ 22 , throughout 0 < r < k 0<r<k 0 < r < k ,
0 < C = w p S < k u 2 − r 2 r . 0<C=\frac{wp}{S}<\frac{k_u^2-r^2}{r}. 0 < C = S w p < r k u 2 − r 2 . More generally, for arbitrary n ≥ 2 n\ge2 n ≥ 2 , whenever the displayed denominator is positive,
C r < 1 − y 1 + y − n + 2 6 y ( 1 − y ) . \frac{C}{r}<\frac{1-y}{1+y-\frac{n+2}{6}y(1-y)}. r C < 1 + y − 6 n + 2 y ( 1 − y ) 1 − y .
Proof. We justify the logarithmic convexity used here. In the product-integral formula for R R R , each factor, for a = κ l 2 > b a=\kappa_l^2>b a = κ l 2 > b and 0 < v < 1 0<v<1 0 < v < 1 , is
h v ( x ) = ( 1 − v ) + v ( a − b ) a − x . h_v(x)=(1-v)+\frac{v(a-b)}{a-x}. h v ( x ) = ( 1 − v ) + a − x v ( a − b ) .
It is strictly logarithmically convex: writing A = 1 − v > 0 A=1-v>0 A = 1 − v > 0 , B = v ( a − b ) > 0 B=v(a-b)>0 B = v ( a − b ) > 0 and t = a − x > 0 t=a-x>0 t = a − x > 0 gives
h v h v ′ ′ − ( h v ′ ) 2 = 2 A B t 3 + B 2 t 4 > 0. h_vh_v''-(h_v')^2=\frac{2AB}{t^3}+\frac{B^2}{t^4}>0. h v h v ′′ − ( h v ′ ) 2 = t 3 2 A B + t 4 B 2 > 0.
Products preserve logarithmic convexity, and so does integration against the positive measure ( 1 − v ) d v / 2 (1-v)\,dv/2 ( 1 − v ) d v /2 , by Hölder’s inequality. Uniform convergence of the product and its first two derivatives on [ 0 , b ] [0,b] [ 0 , b ] gives logarithmic convexity for the infinite product-integral. For the strict endpoint comparison below, one may keep one factor: its strict logarithmic convexity for 0 < v < 1 0<v<1 0 < v < 1 makes Hölder’s bound strict when the two x x x arguments differ. Thus ( log R ) ′ (\log R)' ( log R ) ′ strictly increases. For the endpoint evaluation, put
Σ ( x ) = ∑ l ≥ 2 ( κ l 2 − x ) − 1 . \Sigma(x)=\sum_{l\ge2}(\kappa_l^2-x)^{-1}. Σ ( x ) = l ≥ 2 ∑ ( κ l 2 − x ) − 1 .
The function h ( x ) = p ( x ) / x h(x)=p(\sqrt{x})/\sqrt{x} h ( x ) = p ( x ) / x solves 4 x h ′ ′ + 2 ( n + 2 ) h ′ + h = 0 4xh''+2(n+2)h'+h=0 4 x h ′′ + 2 ( n + 2 ) h ′ + h = 0 . Since h ( b ) = 0 h(b)=0 h ( b ) = 0 , its product gives
− 2 Σ ( b ) = h ′ ′ ( b ) h ′ ( b ) = − n + 2 2 b , Σ ( b ) = n + 2 4 b . -2\Sigma(b)=\frac{h''(b)}{h'(b)}=-\frac{n+2}{2b},\qquad\Sigma(b)=\frac{n+2}{4b}. − 2Σ ( b ) = h ′ ( b ) h ′′ ( b ) = − 2 b n + 2 , Σ ( b ) = 4 b n + 2 .
At the endpoint the product factors have value 1 1 1 and their x x x derivatives are v / ( κ l 2 − b ) v/(\kappa_l^2-b) v / ( κ l 2 − b ) . Hence
R ( b ) = 1 4 , R ′ ( b ) = 1 2 Σ ( b ) ∫ 0 1 v ( 1 − v ) d v = n + 2 48 b . R(b)=\frac{1}{4},\qquad R'(b)=\frac{1}{2}\Sigma(b)\int_0^1v(1-v)\,dv=\frac{n+2}{48b}. R ( b ) = 4 1 , R ′ ( b ) = 2 1 Σ ( b ) ∫ 0 1 v ( 1 − v ) d v = 48 b n + 2 .
Consequently, for every x < b x<b x < b ,
R ′ ( x ) R ( x ) < R ′ ( b ) R ( b ) = n + 2 12 b . \frac{R'(x)}{R(x)}<\frac{R'(b)}{R(b)}=\frac{n+2}{12b}. R ( x ) R ′ ( x ) < R ( b ) R ′ ( b ) = 12 b n + 2 .
In the exact multiplier identity this gives
r C = 1 − 2 x T ′ T = 1 + 2 x b − x − 2 x R ′ R > 1 + y − n + 2 6 y ( 1 − y ) 1 − y . \frac{r}{C}=1-\frac{2xT'}{T}=1+\frac{2x}{b-x}-\frac{2xR'}{R}>\frac{1+y-\frac{n+2}{6}y(1-y)}{1-y}. C r = 1 − T 2 x T ′ = 1 + b − x 2 x − R 2 x R ′ > 1 − y 1 + y − 6 n + 2 y ( 1 − y ) .
Taking reciprocals proves the second assertion where its denominator is positive. If n ≤ 22 n\le22 n ≤ 22 , then
1 − n + 2 6 y ( 1 − y ) ≥ 1 − 4 y ( 1 − y ) ≥ 0. 1-\frac{n+2}{6}y(1-y)\ge1-4y(1-y)\ge0. 1 − 6 n + 2 y ( 1 − y ) ≥ 1 − 4 y ( 1 − y ) ≥ 0.
Thus the strict preceding inequality gives r / C > y / ( 1 − y ) > 0 r/C>y/(1-y)>0 r / C > y / ( 1 − y ) > 0 , including when n = 22 n=22 n = 22 , y = 1 / 2 y=1/2 y = 1/2 . It follows that C / r < ( 1 − y ) / y C/r<(1-y)/y C / r < ( 1 − y ) / y , as asserted. □
For 23 ≤ n ≤ 47 23\le n\le47 23 ≤ n ≤ 47 , we control the radial multiplier by the first three coefficients of a positive power series. Normalizing their sum to one turns the required estimate into an elementary moment inequality.
Put x = r 2 x=r^2 x = r 2 , b = k 2 b=k^2 b = k 2 , y = x / b y=x/b y = x / b , ε = 1 − y \varepsilon=1-y ε = 1 − y , T ( x ) = r t ( r ) T(x)=rt(r) T ( x ) = r t ( r ) , ρ = − t / t r = C \rho=-t/t_r=C ρ = − t / t r = C , and
R ( x ) = T ( x ) / ( b − x ) , Q ( y ) = 4 R ( b y ) = ∑ j ≥ 0 ν j y j . R(x)=T(x)/(b-x),\qquad Q(y)=4R(by)=\sum_{j\ge0}\nu_jy^j. R ( x ) = T ( x ) / ( b − x ) , Q ( y ) = 4 R ( b y ) = j ≥ 0 ∑ ν j y j .
The product-integral formula for R R R proves ν j ≥ 0 \nu_j\ge0 ν j ≥ 0 , and the endpoint identity R ( b ) = 1 / 4 R(b)=1/4 R ( b ) = 1/4 gives ∑ ν j = 1 \sum\nu_j=1 ∑ ν j = 1 . The exact multiplier formula is
ρ r = ( 1 + 2 y ε − 2 y Q ′ ( y ) Q ( y ) ) − 1 . \frac{\rho}{r}=\left(1+\frac{2y}{\varepsilon}-2y\frac{Q'(y)}{Q(y)}\right)^{-1}. r ρ = ( 1 + ε 2 y − 2 y Q ( y ) Q ′ ( y ) ) − 1 .
Consequently ρ / r < ε / y \rho/r<\varepsilon/y ρ / r < ε / y is equivalent to
∑ j ≥ 0 ν j f j ( y ) < 0 , f j ( y ) = { 2 j ( 1 − y ) − 1 } y j . (27) \sum_{j\ge0}\nu_jf_j(y)<0,\qquad f_j(y)=\{2j(1-y)-1\}y^j. \tag*{(27)} j ≥ 0 ∑ ν j f j ( y ) < 0 , f j ( y ) = { 2 j ( 1 − y ) − 1 } y j . (27)
Three weighted coefficient sums suffice for the upper multiplier bound.
Lemma 4.5. For every real parameter 23 ≤ n ≤ 47 23 \le n \le47 23 ≤ n ≤ 47 in the radial formulas,
ν 0 > 1 5 , ν 1 > 3 25 , ν 2 > 89 1000 . \nu_0 > \frac{1}{5}, \qquad\nu_1 > \frac{3}{25}, \qquad\nu_2 > \frac{89}{1000}. ν 0 > 5 1 , ν 1 > 25 3 , ν 2 > 1000 89 . Proof. Write d = n + 2 d=n+2 d = n + 2 , h = n + 4 h=n+4 h = n + 4 and c = − X ( k ) > 0 c=-X(k)>0 c = − X ( k ) > 0 , and expand T ( x ) = T 0 + T 1 x + T 2 x 2 + O ( x 3 ) T(x)=T_0+T_1x+T_2x^2+O(x^3) T ( x ) = T 0 + T 1 x + T 2 x 2 + O ( x 3 ) at the center. Then
ν 0 = 4 T 0 b , ν 1 = 4 ( T 0 b + T 1 ) , ν 2 = 4 ( T 0 b + T 1 + b T 2 ) , \nu_0 = \frac{4T_0}{b}, \qquad\nu_1 = 4\left(\frac{T_0}{b}+T_1\right), \qquad\nu_2 = 4\left(\frac{T_0}{b}+T_1+bT_2\right), ν 0 = b 4 T 0 , ν 1 = 4 ( b T 0 + T 1 ) , ν 2 = 4 ( b T 0 + T 1 + b T 2 ) , T 0 = n ( 1 + c ) , T 1 = n c − 2 2 d , T 2 = n c ( n + 6 ) − 8 8 d 2 h . T_0=n(1+c), \qquad T_1=\frac{nc-2}{2d}, \qquad T_2=\frac{nc(n+6)-8}{8d^2h}. T 0 = n ( 1 + c ) , T 1 = 2 d n c − 2 , T 2 = 8 d 2 h n c ( n + 6 ) − 8 . All three expressions increase with c c c . For real n n n , the radial problem is the one-dimensional problem with weight r n + 1 r^{n+1} r n + 1 . The trial function v ( r ) = ( 1 − r 2 ) 2 v(r)=(1-r^2)^2 v ( r ) = ( 1 − r 2 ) 2 gives
b < 16 ∫ 0 1 r n + 3 ( 1 − r 2 ) 2 d r ∫ 0 1 r n + 1 ( 1 − r 2 ) 4 d r = ( n + 2 ) ( n + 10 ) 3 . b < \frac{16\int_0^1 r^{n+3}(1-r^2)^2\,\mathrm{d}r}{\int_0^1 r^{n+1}(1-r^2)^4\,\mathrm{d}r}=\frac{(n+2)(n+10)}{3}. b < ∫ 0 1 r n + 1 ( 1 − r 2 ) 4 d r 16 ∫ 0 1 r n + 3 ( 1 − r 2 ) 2 d r = 3 ( n + 2 ) ( n + 10 ) . The inequality is strict because the trial function has a double zero at 1 1 1 , whereas the first radial eigenfunction has a simple zero. After setting c = 0 c=0 c = 0 , each lower expression for ν 0 , ν 1 , ν 2 \nu_0,\nu_1,\nu_2 ν 0 , ν 1 , ν 2 decreases with b b b . Substituting this upper bound therefore gives
ν 0 > 12 n d ( n + 10 ) , ν 1 > 8 ( n − 5 ) d ( n + 10 ) , ν 2 > 8 ( n − 5 ) d ( n + 10 ) − 4 ( n + 10 ) 3 d h . \nu_0 > \frac{12n}{d(n+10)}, \qquad\nu_1 > \frac{8(n-5)}{d(n+10)}, \qquad\nu_2 > \frac{8(n-5)}{d(n+10)}-\frac{4(n+10)}{3dh}. ν 0 > d ( n + 10 ) 12 n , ν 1 > d ( n + 10 ) 8 ( n − 5 ) , ν 2 > d ( n + 10 ) 8 ( n − 5 ) − 3 d h 4 ( n + 10 ) . For the first two desired comparisons, clear their positive denominators. The resulting inequalities are
n 2 − 48 n + 20 < 0 , − 3 n 2 + 164 n − 1060 > 0. n^2-48n+20<0, \qquad-3n^2+164n-1060>0. n 2 − 48 n + 20 < 0 , − 3 n 2 + 164 n − 1060 > 0. The first convex polynomial is negative at both 23 , 47 23,47 23 , 47 , and the second concave polynomial is positive there. The third comparison reduces to
P 3 ( n ) = − 267 n 3 + 15728 n 2 − 122156 n − 901360 > 0. P_3(n)=-267n^3+15728n^2-122156n-901360>0. P 3 ( n ) = − 267 n 3 + 15728 n 2 − 122156 n − 901360 > 0. It is concave on [ 23 , 47 ] [23,47] [ 23 , 47 ] , since P 3 ′ ′ = 31456 − 1602 n < 0 P_3''=31456-1602n<0 P 3 ′′ = 31456 − 1602 n < 0 , and P 3 ( 23 ) = 1360575 P_3(23)=1360575 P 3 ( 23 ) = 1360575 , P 3 ( 47 ) = 379719 P_3(47)=379719 P 3 ( 47 ) = 379719 .
We bound the remaining mass by elementary estimates for the coefficient tail.
Lemma 4.6. For the intervals in the next table, f j ( y ) < β i f_j(y)<\beta_i f j ( y ) < β i holds for all integers j ≥ 0 j\ge0 j ≥ 0 and all y y y up to the indicated right endpoint. The last row holds on the whole interval 0 < y < 1 0<y<1 0 < y < 1 .
Table 10.
Proof. Set ℓ = − log y \ell=-\log y ℓ = − log y and r ( y ) = ℓ / [ 2 ( 1 − y ) ] r(y)=\ell/[2(1-y)] r ( y ) = ℓ / [ 2 ( 1 − y )] . Maximizing in the continuous variable u = j ℓ ≥ 0 u=j\ell\ge0 u = j ℓ ≥ 0 gives
f j ( y ) = { u / r ( y ) − 1 } e − u ≤ e − 1 − r ( y ) r ( y ) . f_j(y)=\{u/r(y)-1\}e^{-u}\le\frac{e^{-1-r(y)}}{r(y)}. f j ( y ) = { u / r ( y ) − 1 } e − u ≤ r ( y ) e − 1 − r ( y ) .
The last expression decreases as r ( y ) r(y) r ( y ) increases, and r ( y ) r(y) r ( y ) strictly decreases with y y y . At y = b i < 1 y=b_i<1 y = b i < 1 , put ξ = ( 1 − b i ) / ( 1 + b i ) \xi=(1-b_i)/(1+b_i) ξ = ( 1 − b i ) / ( 1 + b i ) . The elementary logarithm series gives
r ( b i ) ≥ 1 + ξ 2 / 3 1 + b i > r i . r(b_i)\ge\frac{1+\xi^2/3}{1+b_i}>r_i. r ( b i ) ≥ 1 + b i 1 + ξ 2 /3 > r i .
Every last rational comparison is immediate from the displayed small fractions. For b i = 1 b_i=1 b i = 1 , use r ( y ) > 1 / 2 r(y)>1/2 r ( y ) > 1/2 . To check the exponential bounds without any decimal constant, let E 6 ( v ) = ∑ l = 0 6 v l / l ! E_6(v)=\sum_{l=0}^{6}v^l/l! E 6 ( v ) = ∑ l = 0 6 v l / l ! . In each of the seven rows, direct rational multiplication gives
β i r i E 6 ( 1 + r i ) > 1 + 3 1000 > 1. \beta_i r_i E_6(1+r_i)>1+\frac{3}{1000}>1. β i r i E 6 ( 1 + r i ) > 1 + 1000 3 > 1.
Since e 1 + r i > E 6 ( 1 + r i ) e^{1+r_i}>E_6(1+r_i) e 1 + r i > E 6 ( 1 + r i ) , this proves all the bounds.
Combining the three masses with the tail bounds gives the required upper multiplier estimate.
Theorem 4.7. For every integer 2 ≤ n ≤ 47 2\le n\le47 2 ≤ n ≤ 47 and every 0 < r < k 0<r<k 0 < r < k ,
0 < ρ < k 2 − r 2 r . 0<\rho<\frac{k^2-r^2}{r}. 0 < ρ < r k 2 − r 2 . Proof. Proposition 4.4 proves the result for n ≤ 22 n\le22 n ≤ 22 . For 23 ≤ n ≤ 47 23\le n\le47 23 ≤ n ≤ 47 use Lemmas 4.5 and 4.6. If 0 < y ≤ 1 / 2 0<y\le1/2 0 < y ≤ 1/2 , then f 1 ≤ 1 / 8 f_1\le1/8 f 1 ≤ 1/8 and f j ≤ 1 / 4 f_j\le1/4 f j ≤ 1/4 for all j ≥ 2 j\ge2 j ≥ 2 . For the latter assertion the maximum is at y = 1 / 2 y=1/2 y = 1/2 , where ( j − 1 ) 2 − j ≤ 1 / 4 (j-1)2^{-j}\le1/4 ( j − 1 ) 2 − j ≤ 1/4 ; the case j = 2 j=2 j = 2 also has its maximum there. Thus
∑ ν j f j ≤ 1 4 − 5 4 ν 0 − 1 8 ν 1 < − 3 200 < 0. \sum\nu_jf_j\le\frac{1}{4}-\frac{5}{4}\nu_0-\frac{1}{8}\nu_1<-\frac{3}{200}<0. ∑ ν j f j ≤ 4 1 − 4 5 ν 0 − 8 1 ν 1 < − 200 3 < 0. On each remaining interval [ a i , b i ] [a_i,b_i] [ a i , b i ] , the first two polynomials f 1 , f 2 f_1,f_2 f 1 , f 2 decrease for y ≥ 1 / 2 y\ge1/2 y ≥ 1/2 . Hence
∑ ν j f j ≤ β i − ( 1 + β i ) ν 0 − ( β i − f 1 ( a i ) ) ν 1 − ( β i − f 2 ( a i ) ) ν 2 . \sum\nu_jf_j\le\beta_i-(1+\beta_i)\nu_0-(\beta_i-f_1(a_i))\nu_1-(\beta_i-f_2(a_i))\nu_2. ∑ ν j f j ≤ β i − ( 1 + β i ) ν 0 − ( β i − f 1 ( a i )) ν 1 − ( β i − f 2 ( a i )) ν 2 . All the subtracted coefficients are positive. Substituting the three mass lower bounds gives the following strictly negative upper bounds:
Table 11.
This proves (27) on all of ( 0 , 1 ) (0,1) ( 0 , 1 ) .
Theorem 4.7 supplies the strict positivity of the coupling coefficient.
Corollary 4.8. In every 2 ≤ n ≤ 47 2 \le n \le47 2 ≤ n ≤ 47 , d > 0 d > 0 d > 0 throughout 0 < r < k 0 < r < k 0 < r < k , using only the positive weights in the established expansion of G G G .
Proof. Write G = ∑ j ≥ 1 a j ( b j − r 2 j ) G = \sum_{j \ge1} a_j \left(b^j-r^{2j}\right) G = ∑ j ≥ 1 a j ( b j − r 2 j ) with a j > 0 a_j > 0 a j > 0 . Then
d = 2 ∑ j ≥ 1 a j { b j − r 2 j − j ρ r 2 j − 1 } > 0 , d = 2\sum_{j \ge1} a_j\left\{b^j-r^{2j}-j\rho r^{2j-1}\right\} > 0, d = 2 j ≥ 1 ∑ a j { b j − r 2 j − j ρ r 2 j − 1 } > 0 , since b j − r 2 j ≥ j ( b − r 2 ) r 2 j − 2 > j ρ r 2 j − 1 b^j-r^{2j} \ge j\left(b-r^2\right)r^{2j-2} > j\rho r^{2j-1} b j − r 2 j ≥ j ( b − r 2 ) r 2 j − 2 > j ρ r 2 j − 1 .
Bessel gaps and exponential estimates
For m = n − 1 m=n-1 m = n − 1 , put Δ = k 2 − q 2 \Delta=k^2-q^2 Δ = k 2 − q 2 . The following lower bounds are sufficient for the scalar matrix proof:
Table 1 (PDF p. 49).
Only the four right endpoints of the dimensional ranges need to be estimated. Indeed the variational secant argument in § 5 for L ( t ) = j t , 1 2 L(t)=j_{\sqrt{t},1}^2 L ( t ) = j t , 1 2 gives that
j ν + 1 , 1 2 − j ν , 1 2 2 ν + 1 \frac{j_{\nu+1,1}^2-j_{\nu,1}^2}{2\nu+1} 2 ν + 1 j ν + 1 , 1 2 − j ν , 1 2
is nonincreasing for ν ≥ 1 / 2 \nu\ge1/2 ν ≥ 1/2 . Here ν = n / 2 − 1 \nu=n/2-1 ν = n /2 − 1 .
Table 2.
For P d , N ( r ) = ∑ j = 0 N t j P_{d,N}(r)=\sum_{j=0}^{N}t_j P d , N ( r ) = ∑ j = 0 N t j , with
t 0 = 1 , t j = − r 2 2 j ( d + 2 j − 2 ) t j − 1 , t_0=1,\qquad t_j=-\frac{r^2}{2j(d+2j-2)}t_{j-1}, t 0 = 1 , t j = − 2 j ( d + 2 j − 2 ) r 2 t j − 1 ,
the alternating remainder gives P d , 61 < ψ d < P d , 60 P_{d,61}<\psi_d<P_{d,60} P d , 61 < ψ d < P d , 60 at these arguments: r + 2 < 2 ⋅ 61 ( d + 120 ) r_+^2<2\cdot61(d+120) r + 2 < 2 ⋅ 61 ( d + 120 ) ensures decreasing tail magnitudes from the required index onward; monotonicity of the early terms is not needed. The following are explicit finite rational comparisons; terminating decimals denote exact rationals:
Table 12 (PDF p. 50).
All entries mean precisely the displayed 61- or 62-term rational sum. For an integer-only check, write r = u / v r=u/v r = u / v , put D N = ∏ j = 1 N 2 j ( d + 2 j − 2 ) D_N=\prod_{j=1}^{N}2j(d+2j-2) D N = ∏ j = 1 N 2 j ( d + 2 j − 2 ) , and use
D N v 2 N P d , N ( u / v ) = ∑ j = 0 N ( − 1 ) j u 2 j v 2 ( N − j ) ∏ i = j + 1 N 2 i ( d + 2 i − 2 ) . D_Nv^{2N}P_{d,N}(u/v)=\sum_{j=0}^{N}(-1)^j u^{2j}v^{2(N-j)}\prod_{i=j+1}^{N}2i(d+2i-2). D N v 2 N P d , N ( u / v ) = j = 0 ∑ N ( − 1 ) j u 2 j v 2 ( N − j ) i = j + 1 ∏ N 2 i ( d + 2 i − 2 ) .
Here is an elementary identification of the roots as the first ones. Let ν = d / 2 − 1 \nu=d/2-1 ν = d /2 − 1 , a ν = ν 2 − 1 / 4 a_\nu=\sqrt{\nu^2-1/4} a ν = ν 2 − 1/4 , and Y = r ν + 1 / 2 ψ d ( r ) Y=r^{\nu+1/2}\psi_d(r) Y = r ν + 1/2 ψ d ( r ) . Then
Y ′ ′ + V ν Y = 0 , V ν = 1 − a ν 2 / r 2 . Y''+V_\nu Y=0,\qquad V_\nu=1-a_\nu^2/r^2. Y ′′ + V ν Y = 0 , V ν = 1 − a ν 2 / r 2 .
The regular solution and its derivative are positive up to a ν a_\nu a ν . On [ a ν , r + ] [a_\nu,r_+] [ a ν , r + ] one has V ν < ω 2 V_\nu<\omega^2 V ν < ω 2 , with
Table 13 (PDF p. 50).
These assertions are the rational comparisons ( 1 − ω 2 ) r + 2 < a ν 2 (1-\omega^2)r_+^2<a_\nu^2 ( 1 − ω 2 ) r + 2 < a ν 2 . Duhamel comparison with frequency ω \omega ω shows that the first zero is after a ν + π / ( 2 ω ) a_\nu+\pi/(2\omega) a ν + π / ( 2 ω ) . Since V ν < 1 V_\nu<1 V ν < 1 , successive zeros have distance greater than π \pi π . A second zero before r + r_+ r + would therefore exceed
a ν + π ( 1 + 1 2 ω ) > ν − 1 40 + 314 100 ( 1 + 1 2 ω ) > r + . a_\nu+\pi\left(1+\frac{1}{2\omega}\right)>\nu-\frac{1}{40}+\frac{314}{100}\left(1+\frac{1}{2\omega}\right)>r_+. a ν + π ( 1 + 2 ω 1 ) > ν − 40 1 + 100 314 ( 1 + 2 ω 1 ) > r + .
The inequalities hold for every row above; a ν > ν − 1 / 40 a_\nu>\nu-1/40 a ν > ν − 1/40 follows by squaring. This contradiction, together with the Taylor sign change, proves each stated first-root bracket. Finally the margins obtained from these brackets are
Table 14.
They are strictly positive. The secant monotonicity proves all four dimensional ranges in the first table.
The preceding estimates complete the Bessel-function part of the determinant argument. The remaining scalar exponential inequalities are proved by rational bounds and polynomial positivity. Each coefficient comparison below applies on a whole interval.
For the first dimension range above ten, the following exponential comparison closes the determinant estimate.
Lemma 4.9. Let 10 ≤ m ≤ 26 10 \le m \le26 10 ≤ m ≤ 26 , put M = ( m − 6 ) / 24 M = (m - 6)/24 M = ( m − 6 ) /24 , and for 0 ≤ h ≤ 1 0 \le h \le1 0 ≤ h ≤ 1 set
c = M exp ( − m + 3 8 M h ) . c = M \exp\left(-\frac{m+3}{8M}h\right). c = M exp ( − 8 M m + 3 h ) . Then
1 + 4 ( 1 − h ) c − ( 1 − h ) c 2 h 2 > m 3 { 3 + ( m − 1 ) h / ( 2 − h ) } . 1 + 4(1-h)c - (1-h)c^2h^2 > \frac{m}{3\{3+(m-1)h/(2-h)\}}. 1 + 4 ( 1 − h ) c − ( 1 − h ) c 2 h 2 > 3 { 3 + ( m − 1 ) h / ( 2 − h )} m .
Proof. Write t = m − 6 t=m-6 t = m − 6 , so 4 ≤ t ≤ 20 4 \le t \le20 4 ≤ t ≤ 20 , and abbreviate
R = ( t + 6 ) ( 2 − h ) 3 { 6 + ( t + 2 ) h } , H = t − 3 2 ( t + 3 ) , z = 3 ( t + 9 ) t h . R = \frac{(t+6)(2-h)}{3\{6+(t+2)h\}}, \qquad H = \frac{t-3}{2(t+3)}, \qquad z = \frac{3(t+9)}{t}h. R = 3 { 6 + ( t + 2 ) h } ( t + 6 ) ( 2 − h ) , H = 2 ( t + 3 ) t − 3 , z = t 3 ( t + 9 ) h .
Thus c = ( t / 24 ) e − z c=(t/24)e^{-z} c = ( t /24 ) e − z . Since 0 < c ≤ 5 / 6 0<c\le5/6 0 < c ≤ 5/6 ,
1 + 4 ( 1 − h ) c − ( 1 − h ) c 2 h 2 = 1 + ( 1 − h ) c ( 4 − c h 2 ) ≥ 1. 1+4(1-h)c-(1-h)c^2h^2=1+(1-h)c(4-ch^2)\ge1. 1 + 4 ( 1 − h ) c − ( 1 − h ) c 2 h 2 = 1 + ( 1 − h ) c ( 4 − c h 2 ) ≥ 1.
If h ≥ H h\ge H h ≥ H , then R ≤ 1 R\le1 R ≤ 1 ; the desired inequality is strict for h < 1 h<1 h < 1 by the positive extra term, and for h = 1 h=1 h = 1 because R = m / { 3 ( m + 2 ) } < 1 R=m/\{3(m+2)\}<1 R = m / { 3 ( m + 2 )} < 1 .
It remains to consider 0 ≤ h ≤ H < 1 / 2 0\le h\le H<1/2 0 ≤ h ≤ H < 1/2 . In this range c h 2 < 5 / 24 < 1 / 4 ch^2<5/24<1/4 c h 2 < 5/24 < 1/4 , so
1 + ( 1 − h ) c ( 4 − c h 2 ) > 1 + 15 4 ( 1 − h ) c . 1+(1-h)c(4-ch^2)>1+\frac{15}{4}(1-h)c. 1 + ( 1 − h ) c ( 4 − c h 2 ) > 1 + 4 15 ( 1 − h ) c .
Moreover
z ≤ 3 ( t + 9 ) ( t − 3 ) 2 t ( t + 3 ) < 2 , z\le\frac{3(t+9)(t-3)}{2t(t+3)}<2, z ≤ 2 t ( t + 3 ) 3 ( t + 9 ) ( t − 3 ) < 2 ,
since the positive numerator of the last difference is t 2 − 6 t + 81 = ( t − 3 ) 2 + 72 t^2-6t+81=(t-3)^2+72 t 2 − 6 t + 81 = ( t − 3 ) 2 + 72 . For 0 ≤ z < 2 0\le z<2 0 ≤ z < 2 one has the elementary lower bound
e − z ≥ 2 − z 2 + z . e^{-z}\ge\frac{2-z}{2+z}. e − z ≥ 2 + z 2 − z .
Indeed log ( ( 2 + z ) / ( 2 − z ) ) − z \log((2+z)/(2-z))-z log (( 2 + z ) / ( 2 − z )) − z vanishes at zero and has derivative z 2 / ( 4 − z 2 ) ≥ 0 z^2/(4-z^2)\ge0 z 2 / ( 4 − z 2 ) ≥ 0 . Consequently it suffices to show
1 + 5 t 32 ( 1 − h ) 2 t − 3 ( t + 9 ) h 2 t + 3 ( t + 9 ) h > R . 1+\frac{5t}{32}(1-h)\frac{2t-3(t+9)h}{2t+3(t+9)h}>R. 1 + 32 5 t ( 1 − h ) 2 t + 3 ( t + 9 ) h 2 t − 3 ( t + 9 ) h > R .
All denominators are positive. Upon multiplication this becomes Q ( t , h ) > 0 Q(t,h)>0 Q ( t , h ) > 0 , where
Q ( t , h ) = 3 { 6 + ( t + 2 ) h } { 2 t + 3 ( t + 9 ) h } + 15 t 32 ( 1 − h ) { 6 + ( t + 2 ) h } { 2 t − 3 ( t + 9 ) h } − ( t + 6 ) ( 2 − h ) { 2 t + 3 ( t + 9 ) h } . \begin{aligned}
Q(t,h)={}&3\{6+(t+2)h\}\{2t+3(t+9)h\}\\
&+\frac{15t}{32}(1-h)\{6+(t+2)h\}\{2t-3(t+9)h\}\\
&-(t+6)(2-h)\{2t+3(t+9)h\}.
\end{aligned} Q ( t , h ) = 3 { 6 + ( t + 2 ) h } { 2 t + 3 ( t + 9 ) h } + 32 15 t ( 1 − h ) { 6 + ( t + 2 ) h } { 2 t − 3 ( t + 9 ) h } − ( t + 6 ) ( 2 − h ) { 2 t + 3 ( t + 9 ) h } .
Set h = H u h=Hu h = H u , 0 ≤ u ≤ 1 0\leq u\leq1 0 ≤ u ≤ 1 . The following cubic identity is exact:
( t + 3 ) 3 Q ( t , H u ) = ∑ j = 0 3 B j ( t ) ( 3 j ) u j ( 1 − u ) 3 − j . (t+3)^3Q(t,Hu)=\sum_{j=0}^{3}B_j(t)\binom{3}{j}u^j(1-u)^{3-j}. ( t + 3 ) 3 Q ( t , H u ) = j = 0 ∑ 3 B j ( t ) ( j 3 ) u j ( 1 − u ) 3 − j .
where
B 0 ( t ) = t ( t + 3 ) 3 ( 13 t + 96 ) 8 , B_0(t)=\frac{t(t+3)^3(13t+96)}{8}, B 0 ( t ) = 8 t ( t + 3 ) 3 ( 13 t + 96 ) ,
B 1 ( t ) = ( t + 3 ) 2 96 { t 3 ( 15 t − 52 ) + 702 t 2 + 10269 t − 7776 } , B_1(t)=\frac{(t+3)^2}{96}\{t^3(15t-52)+702t^2+10269t-7776\}, B 1 ( t ) = 96 ( t + 3 ) 2 { t 3 ( 15 t − 52 ) + 702 t 2 + 10269 t − 7776 } ,
B 2 ( t ) = t + 3 384 { t 4 ( 45 t − 131 ) + 975 t 3 + 39627 t 2 + 136620 t − 93312 } , B_2(t)=\frac{t+3}{384}\{t^4(45t-131)+975t^3+39627t^2+136620t-93312\}, B 2 ( t ) = 384 t + 3 { t 4 ( 45 t − 131 ) + 975 t 3 + 39627 t 2 + 136620 t − 93312 } ,
B 3 ( t ) = 15 t ( t + 5 ) ( t + 6 ) ( t + 9 ) 256 { ( t − 3 ) 2 + 72 } . B_3(t)=\frac{15t(t+5)(t+6)(t+9)}{256}\{(t-3)^2+72\}. B 3 ( t ) = 256 15 t ( t + 5 ) ( t + 6 ) ( t + 9 ) {( t − 3 ) 2 + 72 } .
All four are positive for t ≥ 4 t\geq4 t ≥ 4 . For the middle two, use
15 t − 52 ≥ 8 , 10269 t − 7776 ≥ 33300 , 15t-52\geq8,\qquad10269t-7776\geq33300, 15 t − 52 ≥ 8 , 10269 t − 7776 ≥ 33300 ,
45 t − 131 ≥ 49 , 136620 t − 93312 ≥ 453168. 45t-131\geq49,\qquad136620t-93312\geq453168. 45 t − 131 ≥ 49 , 136620 t − 93312 ≥ 453168.
The Bernstein basis functions ( d j ) u j ( 1 − u ) d − j \binom{d}{j}u^j(1-u)^{d-j} ( j d ) u j ( 1 − u ) d − j on [ 0 , 1 ] [0,1] [ 0 , 1 ] are nonnegative and sum to one. Thus Q ( t , H u ) > 0 Q(t,Hu)>0 Q ( t , H u ) > 0 , completing the proof.
The next comparison treats the two middle-dimensional ranges with the same functional inequality.
Lemma 4.10. Let either
27 ≤ m ≤ 36 , M = m − 8 32 , or 37 ≤ m ≤ 41 , M = 13 m − 108 432 . 27\leq m\leq36,\qquad M=\frac{m-8}{32},\qquad\text{or}\qquad37\leq m\leq41,\qquad M=\frac{13m-108}{432}. 27 ≤ m ≤ 36 , M = 32 m − 8 , or 37 ≤ m ≤ 41 , M = 432 13 m − 108 . For 0 ≤ h ≤ 1 0\leq h\leq1 0 ≤ h ≤ 1 put
c = M exp ( − m + 3 8 M h ) . c=M\exp\left(-\frac{m+3}{8M}h\right). c = M exp ( − 8 M m + 3 h ) . Then
1 + 4 ( 1 − h ) c − ( 1 − h ) c 2 h 2 > m ( 2 − h ) 3 { 6 + ( m − 4 ) h } = m 3 { 3 + ( m − 1 ) h / ( 2 − h ) } . (28) 1+4(1-h)c-(1-h)c^2h^2>
\frac{m(2-h)}{3\{6+(m-4)h\}}
=\frac{m}{3\{3+(m-1)h/(2-h)\}}. \tag*{(28)} 1 + 4 ( 1 − h ) c − ( 1 − h ) c 2 h 2 > 3 { 6 + ( m − 4 ) h } m ( 2 − h ) = 3 { 3 + ( m − 1 ) h / ( 2 − h )} m . (28) The two ranges are continuous real intervals.
We record the short polynomial checks explicitly. If P ( h ) = ∑ i = 0 d p i h i P(h)=\sum_{i=0}^{d}p_i h^i P ( h ) = ∑ i = 0 d p i h i , its Bernstein coefficients on [ 0 , H ] [0,H] [ 0 , H ] are
b j ( P ; H ) = ∑ i = 0 j p i H i ( j i ) ( d i ) , 0 ≤ j ≤ d . (29) b_j(P;H)=\sum_{i=0}^{j}p_iH^i\frac{\binom{j}{i}}{\binom{d}{i}},\qquad0\leq j\leq d. \tag*{(29)} b j ( P ; H ) = i = 0 ∑ j p i H i ( i d ) ( i j ) , 0 ≤ j ≤ d . (29)
The binomial theorem gives P ( H u ) = ∑ j = 0 d b j ( P ; H ) ( d j ) u j ( 1 − u ) d − j P(Hu)=\sum_{j=0}^{d}b_j(P;H)\binom{d}{j}u^j(1-u)^{d-j} P ( H u ) = ∑ j = 0 d b j ( P ; H ) ( j d ) u j ( 1 − u ) d − j . Consequently, positive b j b_j b j imply P > 0 P>0 P > 0 on [ 0 , H ] [0,H] [ 0 , H ] .
Define the following factored polynomials:
D 1 ( h ) = 239 ( 1 − h ) ( 19 + 44 h ) ( 6 + 23 h ) 2 ( 19 − 15 h ) 4 − 12160 ( 2 − h ) ( 6 − 4 h ) ( 19 + 15 h ) 4 , D_1(h)=239(1-h)(19+44h)(6+23h)^2(19-15h)^4
-12160(2-h)(6-4h)(19+15h)^4, D 1 ( h ) = 239 ( 1 − h ) ( 19 + 44 h ) ( 6 + 23 h ) 2 ( 19 − 15 h ) 4 − 12160 ( 2 − h ) ( 6 − 4 h ) ( 19 + 15 h ) 4 ,
P 1 ( h ) = ( − 36 + 96 h ) ( 19 + 15 h ) 4 + 4541 640 ( 6 + 23 h ) ( 1 − h ) ( 19 − 15 h ) 4 , P_1(h)=(-36+96h)(19+15h)^4+\frac{4541}{640}(6+23h)(1-h)(19-15h)^4, P 1 ( h ) = ( − 36 + 96 h ) ( 19 + 15 h ) 4 + 640 4541 ( 6 + 23 h ) ( 1 − h ) ( 19 − 15 h ) 4 ,
D 2 ( h ) = 2587 7200 ( 1 − h ) ( 1 + 7938 4849 h ) ( 6 + 33 h ) 2 ( 373 − 270 h ) 4 − ( 2 − h ) ( 6 − 4 h ) ( 373 + 270 h ) 4 , D_2(h)=\frac{2587}{7200}(1-h)\left(1+\frac{7938}{4849}h\right)(6+33h)^2(373-270h)^4
-(2-h)(6-4h)(373+270h)^4, D 2 ( h ) = 7200 2587 ( 1 − h ) ( 1 + 4849 7938 h ) ( 6 + 33 h ) 2 ( 373 − 270 h ) 4 − ( 2 − h ) ( 6 − 4 h ) ( 373 + 270 h ) 4 ,
P 2 ( h ) = ( − 56 + 136 h ) ( 373 + 270 h ) 4 + 74227 7200 ( 6 + 33 h ) ( 1 − h ) ( 373 − 270 h ) 4 . P_2(h)=(-56+136h)(373+270h)^4+\frac{74227}{7200}(6+33h)(1-h)(373-270h)^4. P 2 ( h ) = ( − 56 + 136 h ) ( 373 + 270 h ) 4 + 7200 74227 ( 6 + 33 h ) ( 1 − h ) ( 373 − 270 h ) 4 .
Here are strictly positive integer lower bounds obtained from (29) ; the entries run from j = 0 j=0 j = 0 to j = d j=d j = d . Thus each row is a direct rational polynomial identity and inequality.
Table 1 (PDF p. 53).
In particular, D 1 > 0 D_1>0 D 1 > 0 on [ 0 , 9 / 22 ] [0,9/22] [ 0 , 9/22 ] , D 2 > 0 D_2>0 D 2 > 0 on [ 0 , 8 / 19 ] [0,8/19] [ 0 , 8/19 ] , and P 2 > 0 P_2>0 P 2 > 0 on [ 0 , 7 / 17 ] [0,7/17] [ 0 , 7/17 ] .
For completeness, the closer polynomial P 1 P_1 P 1 has a convenient calculus proof of positivity on [ 0 , 3 / 8 ] [0,3/8] [ 0 , 3/8 ] . Its expanded coefficients in ascending order are
274065063 320 , − 2627387963 640 , − 18045957883 640 , 9764885517 64 , − 1368225945 16 , 6761625525 128 , − 1057485375 128 . \frac{274065063}{320},\quad-\frac{2627387963}{640},\quad-\frac{18045957883}{640},\quad\frac{9764885517}{64},\quad-\frac{1368225945}{16},\quad\frac{6761625525}{128},\quad-\frac{1057485375}{128}. 320 274065063 , − 640 2627387963 , − 640 18045957883 , 64 9764885517 , − 16 1368225945 , 128 6761625525 , − 128 1057485375 .
They give, throughout 0 ≤ h ≤ 3 / 8 0\leq h\leq3/8 0 ≤ h ≤ 3/8 ,
P 1 ′ ′ ′ ( h ) > 10 6 { 915 − 2053 h − 992 h 3 } ≥ 10 6 { 915 − 2053 3 8 − 992 ( 3 8 ) 3 } > 0. P_1^{\prime\prime\prime}(h)>10^6\{915-2053h-992h^3\}
\geq10^6\left\{915-2053\frac{3}{8}-992\left(\frac{3}{8}\right)^3\right\}>0. P 1 ′′′ ( h ) > 1 0 6 { 915 − 2053 h − 992 h 3 } ≥ 1 0 6 { 915 − 2053 8 3 − 992 ( 8 3 ) 3 } > 0.
Direct substitution of the rational points gives
P 1 ′ ( 0 ) < 0 , P 1 ′ ( 1 / 8 ) < 0 , P 1 ( 1 / 8 ) > 181000 , P 1 ′ ′ ( 1 / 8 ) > 44000000 , P_1'(0)<0,\qquad P_1'(1/8)<0,\qquad P_1(1/8)>181000,\qquad P_1''(1/8)>44000000, P 1 ′ ( 0 ) < 0 , P 1 ′ ( 1/8 ) < 0 , P 1 ( 1/8 ) > 181000 , P 1 ′′ ( 1/8 ) > 44000000 ,
P 1 ( 3 / 16 ) > 7300 , ∣ P 1 ′ ( 3 / 16 ) ∣ < 530000. P_1(3/16)>7300,\qquad\left|P_1'(3/16)\right|<530000. P 1 ( 3/16 ) > 7300 , ∣ P 1 ′ ( 3/16 ) ∣ < 530000.
Since P 1 ′ P_1' P 1 ′ is convex, it is negative on [ 0 , 1 / 8 ] [0,1/8] [ 0 , 1/8 ] , where P 1 ≥ P 1 ( 1 / 8 ) > 0 P_1 \ge P_1(1/8)>0 P 1 ≥ P 1 ( 1/8 ) > 0 . On [ 1 / 8 , 3 / 8 ] [1/8,3/8] [ 1/8 , 3/8 ] the increasing function P 1 ′ ′ P_1'' P 1 ′′ is greater than 44000000 44000000 44000000 . Taylor’s integral formula, in either direction from 3 / 16 3/16 3/16 , therefore gives
P 1 ( h ) > 7300 − 530000 ∣ h − 3 16 ∣ + 22000000 ( h − 3 16 ) 2 ≥ 7300 − 530000 2 88000000 > 4100. P_1(h)>7300-530000\left|h-\frac{3}{16}\right|+22000000\left(h-\frac{3}{16}\right)^2\ge7300-\frac{530000^2}{88000000}>4100. P 1 ( h ) > 7300 − 530000 h − 16 3 + 22000000 ( h − 16 3 ) 2 ≥ 7300 − 88000000 53000 0 2 > 4100.
This proves P 1 > 0 P_1>0 P 1 > 0 on the required interval.
Proof of Lemma 4.10. Write
R m ( h ) = m ( 2 − h ) 3 { 6 + ( m − 4 ) h } , a = m + 3 8 M , c = M e − a h . R_m(h)=\frac{m(2-h)}{3\{6+(m-4)h\}},\qquad a=\frac{m+3}{8M},\qquad c=Me^{-ah}. R m ( h ) = 3 { 6 + ( m − 4 ) h } m ( 2 − h ) , a = 8 M m + 3 , c = M e − ah .
First suppose 27 ≤ m ≤ 36 27\le m\le36 27 ≤ m ≤ 36 . The elementary maximum of h 2 e − a h h^2e^{-ah} h 2 e − ah on h ≥ 0 h\ge0 h ≥ 0 , together with e 2 > 7 e^2>7 e 2 > 7 , gives
c h 2 ≤ 4 M a 2 e 2 < 4 ( 7 / 8 ) ( 39 / 7 ) 2 7 = 49 3042 < 1 60 . ch^2\le\frac{4M}{a^2e^2}<\frac{4(7/8)}{(39/7)^2\,7}=\frac{49}{3042}<\frac{1}{60}. c h 2 ≤ a 2 e 2 4 M < ( 39/7 ) 2 7 4 ( 7/8 ) = 3042 49 < 60 1 .
Consequently the left side of (28) is greater than 1 + κ ( 1 − h ) c 1+\kappa(1-h)c 1 + κ ( 1 − h ) c for h < 1 h<1 h < 1 , where κ = 239 / 60 \kappa=239/60 κ = 239/60 . At h = 1 h=1 h = 1 the desired inequality holds directly. Also R m ( h ) ≤ 1 R_m(h)\le1 R m ( h ) ≤ 1 when h ≥ ( m − 9 ) / ( 2 m − 6 ) h\ge(m-9)/(2m-6) h ≥ ( m − 9 ) / ( 2 m − 6 ) ; thus h ≥ 9 / 22 h\ge9/22 h ≥ 9/22 is already settled. For 0 ≤ h ≤ 9 / 22 0\le h\le9/22 0 ≤ h ≤ 9/22 consider
Φ m ( h ) = 1 + κ ( 1 − h ) c − R m ( h ) . \Phi_m(h)=1+\kappa(1-h)c-R_m(h). Φ m ( h ) = 1 + κ ( 1 − h ) c − R m ( h ) .
For fixed h h h , differentiation with respect to the real parameter m m m gives
c m = 1 32 e − 4 h ( 1 + z ) e − z , z = 44 h m − 8 . c_m=\frac{1}{32}e^{-4h}(1+z)e^{-z},\qquad z=\frac{44h}{m-8}. c m = 32 1 e − 4 h ( 1 + z ) e − z , z = m − 8 44 h .
The function ( 1 + z ) e − z (1+z)e^{-z} ( 1 + z ) e − z is decreasing on z ≥ 0 z\ge0 z ≥ 0 . Since z z z decreases with m m m , c m c_m c m increases with m m m . Moreover
∂ m R m ( h ) = ( 2 − h ) ( 6 − 4 h ) 3 { 6 + ( m − 4 ) h } 2 \partial_mR_m(h)=\frac{(2-h)(6-4h)}{3\{6+(m-4)h\}^2} ∂ m R m ( h ) = 3 { 6 + ( m − 4 ) h } 2 ( 2 − h ) ( 6 − 4 h )
decreases with m m m on this interval. It follows that
∂ m Φ m ≥ 239 1920 ( 1 − h ) ( 1 + 44 h 19 ) e − 120 h / 19 − ( 2 − h ) ( 6 − 4 h ) 3 ( 6 + 23 h ) 2 . \partial_m\Phi_m\ge\frac{239}{1920}(1-h)\left(1+\frac{44h}{19}\right)e^{-120h/19}-\frac{(2-h)(6-4h)}{3(6+23h)^2}. ∂ m Φ m ≥ 1920 239 ( 1 − h ) ( 1 + 19 44 h ) e − 120 h /19 − 3 ( 6 + 23 h ) 2 ( 2 − h ) ( 6 − 4 h ) .
For 0 ≤ x < 8 0\le x<8 0 ≤ x < 8 we shall use the elementary inequality
e − x ≥ ( 1 − x / 8 1 + x / 8 ) 4 . (30) e^{-x}\ge\left(\frac{1-x/8}{1+x/8}\right)^4. \tag*{(30)} e − x ≥ ( 1 + x /8 1 − x /8 ) 4 . (30)
Indeed, if u = x / 8 u=x/8 u = x /8 , then log ( ( 1 − u ) / ( 1 + u ) ) ≤ − 2 u \log((1-u)/(1+u))\le-2u log (( 1 − u ) / ( 1 + u )) ≤ − 2 u , as follows by differentiation. Using (30) in the preceding estimate and multiplying by the positive denominator 36480 ( 6 + 23 h ) 2 ( 19 + 15 h ) 4 36480(6+23h)^2(19+15h)^4 36480 ( 6 + 23 h ) 2 ( 19 + 15 h ) 4 proves ∂ m Φ m > 0 \partial_m\Phi_m>0 ∂ m Φ m > 0 , because its numerator is D 1 ( h ) > 0 D_1(h)>0 D 1 ( h ) > 0 . Hence Φ m ≥ Φ 27 \Phi_m\ge\Phi_{27} Φ m ≥ Φ 27 . For h ≥ 3 / 8 h\ge3/8 h ≥ 3/8 , R 27 ( h ) ≤ 1 R_{27}(h)\le1 R 27 ( h ) ≤ 1 and Φ 27 > 0 \Phi_{27}>0 Φ 27 > 0 . For 0 ≤ h ≤ 3 / 8 0\le h\le3/8 0 ≤ h ≤ 3/8 , applying (30) once more shows
3 ( 6 + 23 h ) ( 19 + 15 h ) 4 Φ 27 ( h ) ≥ P 1 ( h ) > 0. 3(6+23h)(19+15h)^4\Phi_{27}(h)\ge P_1(h)>0. 3 ( 6 + 23 h ) ( 19 + 15 h ) 4 Φ 27 ( h ) ≥ P 1 ( h ) > 0.
This completes the first range.
Now suppose 37 ≤ m ≤ 41 37 \le m \le41 37 ≤ m ≤ 41 . We have M < 1 M < 1 M < 1 , a ≥ 2376 / 425 > 11 / 2 a \ge2376/425 > 11/2 a ≥ 2376/425 > 11/2 , and therefore
c h 2 ≤ 4 M a 2 e 2 < 16 847 < 1 50 . ch^2 \le\frac{4M}{a^2e^2} < \frac{16}{847} < \frac{1}{50}. c h 2 ≤ a 2 e 2 4 M < 847 16 < 50 1 .
Use κ = 199 / 50 \kappa= 199/50 κ = 199/50 and the same definition of Φ m \Phi_m Φ m . Here R m ( h ) ≤ 1 R_m(h) \le1 R m ( h ) ≤ 1 for h ≥ 8 / 19 h \ge8/19 h ≥ 8/19 , and for smaller h h h
c m = 13 432 e − 54 h / 13 ( 1 + z ) e − z , z = 7938 h 13 ( 13 m − 108 ) . c_m = \frac{13}{432}e^{-54h/13}(1+z)e^{-z}, \qquad z = \frac{7938h}{13(13m-108)}. c m = 432 13 e − 54 h /13 ( 1 + z ) e − z , z = 13 ( 13 m − 108 ) 7938 h .
Exactly the same monotonicity argument yields
∂ m Φ m ≥ 199 50 13 432 ( 1 − h ) ( 1 + 7938 h 4849 ) e − 2160 h / 373 − ( 2 − h ) ( 6 − 4 h ) 3 ( 6 + 33 h ) 2 . \partial_m\Phi_m \ge\frac{199}{50}\frac{13}{432}(1-h)\left(1+\frac{7938h}{4849}\right)e^{-2160h/373}-\frac{(2-h)(6-4h)}{3(6+33h)^2}. ∂ m Φ m ≥ 50 199 432 13 ( 1 − h ) ( 1 + 4849 7938 h ) e − 2160 h /373 − 3 ( 6 + 33 h ) 2 ( 2 − h ) ( 6 − 4 h ) .
Use (30) , now with x = 2160 h / 373 x = 2160h/373 x = 2160 h /373 . Multiplication by 3 ( 6 + 33 h ) 2 ( 373 + 270 h ) 4 3(6+33h)^2(373+270h)^4 3 ( 6 + 33 h ) 2 ( 373 + 270 h ) 4 gives the lower bound D 2 ( h ) > 0 D_2(h) > 0 D 2 ( h ) > 0 . Thus Φ m ≥ Φ 37 \Phi_m \ge\Phi_{37} Φ m ≥ Φ 37 . The latter is positive if h ≥ 7 / 17 h \ge7/17 h ≥ 7/17 , because R 37 ( h ) ≤ 1 R_{37}(h) \le1 R 37 ( h ) ≤ 1 . For 0 ≤ h ≤ 7 / 17 0 \le h \le7/17 0 ≤ h ≤ 7/17 , the same exponential bound gives
3 ( 6 + 33 h ) ( 373 + 270 h ) 4 Φ 37 ( h ) ≥ P 2 ( h ) > 0. 3(6+33h)(373+270h)^4\Phi_{37}(h) \ge P_2(h) > 0. 3 ( 6 + 33 h ) ( 373 + 270 h ) 4 Φ 37 ( h ) ≥ P 2 ( h ) > 0.
This proves the second range and the lemma.
For the highest five dimensions, the determinant estimate reduces to the following exponential comparison. Its parameter interval covers 43 ≤ n ≤ 47 43 \le n \le47 43 ≤ n ≤ 47 after setting m = n − 1 m = n-1 m = n − 1 .
Lemma 4.11. Let 42 ≤ m ≤ 46 42 \le m \le46 42 ≤ m ≤ 46 , u = m − 9 u = m-9 u = m − 9 , and put
μ = u 36 , α = 9 2 + 54 u , c ( h ) = μ e − α h , Θ m ( h ) = m ( 2 − h ) 3 { 6 + ( m − 4 ) h } . \mu= \frac{u}{36}, \qquad\alpha= \frac{9}{2}+\frac{54}{u}, \qquad c(h)=\mu e^{-\alpha h}, \qquad\Theta_m(h)=\frac{m(2-h)}{3\{6+(m-4)h\}}. μ = 36 u , α = 2 9 + u 54 , c ( h ) = μ e − α h , Θ m ( h ) = 3 { 6 + ( m − 4 ) h } m ( 2 − h ) . For every 0 < h < 1 0<h<1 0 < h < 1 ,
1 + 4 ( 1 − h ) c ( h ) − ( 1 − h ) c ( h ) 2 h 2 > Θ m ( h ) . (31) 1+4(1-h)c(h)-(1-h)c(h)^2h^2>\Theta_m(h). \tag*{(31)} 1 + 4 ( 1 − h ) c ( h ) − ( 1 − h ) c ( h ) 2 h 2 > Θ m ( h ) . (31) At h = 0 h=0 h = 0 the two sides agree.
Proof. Write the difference in (31) as F ( u , h ) \mathcal{F}(u,h) F ( u , h ) . Since Θ m ( h ) ≤ 1 \Theta_m(h) \le1 Θ m ( h ) ≤ 1 exactly when
h ≥ h 0 ( u ) : = u 2 ( u + 6 ) , h \ge h_0(u) := \frac{u}{2(u+6)}, h ≥ h 0 ( u ) := 2 ( u + 6 ) u ,
and c ( h ) h 2 ≤ 37 / 36 < 4 c(h)h^2 \le37/36 < 4 c ( h ) h 2 ≤ 37/36 < 4 , this part of the interval is immediate. It remains to consider
0 < h ≤ h 0 ( u ) ≤ 37 86 < 13 30 < 1 2 . 0<h\le h_0(u)\le\frac{37}{86}<\frac{13}{30}<\frac{1}{2}. 0 < h ≤ h 0 ( u ) ≤ 86 37 < 30 13 < 2 1 .
We first prove that F \mathcal{F} F increases with u u u on the larger rectangle 33 ≤ u ≤ 37 33 \le u \le37 33 ≤ u ≤ 37 , 0 < h ≤ 37 / 86 0<h\le37/86 0 < h ≤ 37/86 . Use the elementary Padé inequality
e − x > S ( x ) N ( x ) , N ( x ) = 1 + x 2 + x 2 10 + x 3 120 , S ( x ) = 1 − x 2 + x 2 10 − x 3 120 ( 0 < x ≤ 3 ) . (32) e^{-x} > \frac{S(x)}{N(x)}, \qquad N(x)=1+\frac{x}{2}+\frac{x^2}{10}+\frac{x^3}{120}, \qquad S(x)=1-\frac{x}{2}+\frac{x^2}{10}-\frac{x^3}{120} \qquad(0<x\le3). \tag*{(32)} e − x > N ( x ) S ( x ) , N ( x ) = 1 + 2 x + 10 x 2 + 120 x 3 , S ( x ) = 1 − 2 x + 10 x 2 − 120 x 3 ( 0 < x ≤ 3 ) . (32)
Indeed S > 0 S>0 S > 0 on this interval, and, for the unnormalized polynomials N ^ = 120 N \widehat{N}=120N N = 120 N , S ^ = 120 S \widehat{S}=120S S = 120 S ,
d d x log N ^ S ^ − 1 = x 6 N ^ S ^ > 0. \frac{\mathrm{d}}{\mathrm{d}x}\log\frac{\widehat{N}}{\widehat{S}}-1=\frac{x^6}{\widehat{N}\widehat{S}}>0. d x d log S N − 1 = N S x 6 > 0.
Their quotient is one at zero, proving (32) .
Direct differentiation gives
F u = ( 1 − h ) ( 4 − 2 c h 2 ) e − α h 36 ( 1 + 54 h u ) − ( 2 − h ) ( 6 − 4 h ) 3 { 6 + ( u + 5 ) h } 2 . \mathcal{F}_u=\frac{(1-h)(4-2ch^2)e^{-\alpha h}}{36}\left(1+\frac{54h}{u}\right)-\frac{(2-h)(6-4h)}{3\{6+(u+5)h\}^2}. F u = 36 ( 1 − h ) ( 4 − 2 c h 2 ) e − α h ( 1 + u 54 h ) − 3 { 6 + ( u + 5 ) h } 2 ( 2 − h ) ( 6 − 4 h ) .
Using c ≤ 37 / 36 c\le37/36 c ≤ 37/36 , α ≤ 135 / 22 \alpha\le135/22 α ≤ 135/22 , and u ≥ 33 u\ge33 u ≥ 33 shows that
F u ≥ ( 1 − h ) ( 4 − 37 h 2 / 18 ) ( 1 + 54 h / 37 ) e − 135 h / 22 36 − ( 2 − h ) ( 6 − 4 h ) 3 ( 6 + 38 h ) 2 . \mathcal{F}_u\ge\frac{(1-h)(4-37h^2/18)(1+54h/37)e^{-135h/22}}{36}-\frac{(2-h)(6-4h)}{3(6+38h)^2}. F u ≥ 36 ( 1 − h ) ( 4 − 37 h 2 /18 ) ( 1 + 54 h /37 ) e − 135 h /22 − 3 ( 6 + 38 h ) 2 ( 2 − h ) ( 6 − 4 h ) .
All factors used in this comparison are positive. Apply (32) with x = 135 h / 22 < 3 x=135h/22<3 x = 135 h /22 < 3 and multiply by 36 N ( 135 h / 22 ) ( 6 + 38 h ) 2 36N(135h/22)(6+38h)^2 36 N ( 135 h /22 ) ( 6 + 38 h ) 2 . The resulting numerator is
Q ( h ) = ( 1 − h ) ( 4 − 37 h 2 / 18 ) ( 1 + 54 h / 37 ) ( 6 + 38 h ) 2 S ( 135 h / 22 ) − 12 ( 2 − h ) ( 6 − 4 h ) N ( 135 h / 22 ) . Q(h)=(1-h)(4-37h^2/18)(1+54h/37)(6+38h)^2S(135h/22)-12(2-h)(6-4h)N(135h/22). Q ( h ) = ( 1 − h ) ( 4 − 37 h 2 /18 ) ( 1 + 54 h /37 ) ( 6 + 38 h ) 2 S ( 135 h /22 ) − 12 ( 2 − h ) ( 6 − 4 h ) N ( 135 h /22 ) .
It has zero constant coefficient. The nine coefficients of Q ( h ) / h Q(h)/h Q ( h ) / h have the following strict lower bounds, in increasing order of degree:
( 1174 , 997 , − 13353 , 11623 , 19996 , − 37930 , 9293 , 16303 , − 8342 ) . (1174,\ 997,\ -13353,\ 11623,\ 19996,\ -37930,\ 9293,\ 16303,\ -8342). ( 1174 , 997 , − 13353 , 11623 , 19996 , − 37930 , 9293 , 16303 , − 8342 ) .
These bounds follow by multiplication of the displayed cubic polynomials; all entries are integers. Consequently
Q ( h ) h > R ( h ) + h 4 ( 19996 − 37930 h ) + h 6 ( 9293 + 16303 h − 8342 h 2 ) , \frac{Q(h)}{h}>R(h)+h^4(19996-37930h)+h^6(9293+16303h-8342h^2), h Q ( h ) > R ( h ) + h 4 ( 19996 − 37930 h ) + h 6 ( 9293 + 16303 h − 8342 h 2 ) ,
R ( h ) = 1174 + 997 h − 13353 h 2 + 11623 h 3 . R(h)=1174+997h-13353h^2+11623h^3. R ( h ) = 1174 + 997 h − 13353 h 2 + 11623 h 3 .
The last two terms are nonnegative for 0 ≤ h ≤ 13 / 30 0\le h\le13/30 0 ≤ h ≤ 13/30 and strictly positive when h > 0 h>0 h > 0 . On this same interval the cubic R R R has Bernstein coefficients
1174 , 118621 90 , 187867 300 , 1198921 27000 , 1174,\qquad\frac{118621}{90},\qquad\frac{187867}{300},\qquad\frac{1198921}{27000}, 1174 , 90 118621 , 300 187867 , 27000 1198921 ,
which are positive. Thus F u > 0 \mathcal{F}_u>0 F u > 0 throughout the stated rectangle.
It remains to prove F ( 33 , h ) > 0 \mathcal{F}(33,h)>0 F ( 33 , h ) > 0 for 0 < h ≤ 11 / 26 0<h\le11/26 0 < h ≤ 11/26 . Here μ = 11 / 12 \mu=11/12 μ = 11/12 , α = 135 / 22 \alpha=135/22 α = 135/22 , and, with B = 26 / 11 B=26/11 B = 26/11 , D = 19 / 3 D=19/3 D = 19/3 ,
Θ 42 ( h ) − 1 = 4 μ 1 − B h 1 + D h . \Theta_{42}(h)-1=4\mu\frac{1-Bh}{1+Dh}. Θ 42 ( h ) − 1 = 4 μ 1 + D h 1 − B h .
The map z ↦ 4 z − h 2 z 2 z \mapsto4z-h^{2}z^{2} z ↦ 4 z − h 2 z 2 is increasing for the relevant positive z z z . Hence (32) reduces the claim to P ( h ) > 0 \mathcal{P}(h)>0 P ( h ) > 0 , where
P ( h ) = ( 1 − h ) ( 1 + D h ) { S ( α h ) N ( α h ) − μ 4 h 2 S ( α h ) 2 } − ( 1 − B h ) N ( α h ) 2 . \mathcal{P}(h)=(1-h)(1+Dh)\left\{S(\alpha h)N(\alpha h)-\frac{\mu}{4}h^{2}S(\alpha h)^{2}\right\}-(1-Bh)N(\alpha h)^{2}. P ( h ) = ( 1 − h ) ( 1 + D h ) { S ( α h ) N ( α h ) − 4 μ h 2 S ( α h ) 2 } − ( 1 − B h ) N ( α h ) 2 .
Again the constant coefficient vanishes. The ten coefficients of P ( h ) / h \mathcal{P}(h)/h P ( h ) / h , in increasing order, have strict lower bounds
( 1560 1000 , − 10886 1000 , 3236 1000 , 57078 1000 , 36105 1000 , 63 , − 79 , 78 , − 26 , 0 ) . \left(\frac{1560}{1000},-\frac{10886}{1000},\frac{3236}{1000},\frac{57078}{1000},\frac{36105}{1000},63,-79,78,-26,0\right). ( 1000 1560 , − 1000 10886 , 1000 3236 , 1000 57078 , 1000 36105 , 63 , − 79 , 78 , − 26 , 0 ) .
The pairings h 5 ( 63 − 79 h ) > 0 h^{5}(63-79h)>0 h 5 ( 63 − 79 h ) > 0 and h 7 ( 78 − 26 h ) > 0 h^{7}(78-26h)>0 h 7 ( 78 − 26 h ) > 0 for h < 1 / 2 h<1/2 h < 1/2 give
P ( h ) h > q ( h ) : = 1560 − 10886 h + 3236 h 2 + 57078 h 3 + 36105 h 4 1000 . \frac{\mathcal{P}(h)}{h}>q(h):=\frac{1560-10886h+3236h^{2}+57078h^{3}+36105h^{4}}{1000}. h P ( h ) > q ( h ) := 1000 1560 − 10886 h + 3236 h 2 + 57078 h 3 + 36105 h 4 .
This quartic is positive for every h ≥ 0 h\geq0 h ≥ 0 . To see this without root localization, note that q ′ ′ ( h ) > 6 q''(h)>6 q ′′ ( h ) > 6 there, whereas at a = 13 / 60 a=13/60 a = 13/60 ,
q ( a ) = 11580751 864000000 > 1 100 , q ′ ( a ) = 17057 720000 < 1 30 . q(a)=\frac{11580751}{864000000}>\frac{1}{100},\qquad q'(a)=\frac{17057}{720000}<\frac{1}{30}. q ( a ) = 864000000 11580751 > 100 1 , q ′ ( a ) = 720000 17057 < 30 1 .
Strong convexity therefore gives, with z = h − a z=h-a z = h − a ,
q ( h ) > 1 100 + q ′ ( a ) z + 3 z 2 ≥ 1 100 − 1 10800 > 0. q(h)>\frac{1}{100}+q'(a)z+3z^{2}\geq\frac{1}{100}-\frac{1}{10800}>0. q ( h ) > 100 1 + q ′ ( a ) z + 3 z 2 ≥ 100 1 − 10800 1 > 0.
This proves the base case. Increasing u u u proves the desired inequality when h ≤ h 0 ( 33 ) h\leq h_{0}(33) h ≤ h 0 ( 33 ) . If h > h 0 ( 33 ) h>h_{0}(33) h > h 0 ( 33 ) , the already established trivial part gives F ( 33 , h ) > 0 F(33,h)>0 F ( 33 , h ) > 0 as well; the same monotonicity applies up to 37 / 86 37/86 37/86 . Thus the argument covers every required pair ( u , h ) (u,h) ( u , h ) .
The exact endpoint dimension threshold
This section concerns the calibration used in this paper, rather than the truth of the buckling inequality in arbitrary dimension. It proves an obstruction in every dimension n ≥ 48 n\geq48 n ≥ 48 . The endpoint test is positive in dimension 47; full-interval positivity is supplied by Theorem 2.10 . The conclusion for all higher dimensions follows from a variational concavity argument at the end of this section.
Figure 4 illustrates the endpoint sign. The proof below supplies its exact value and propagates the sign to all larger dimensions.
Figure 4. The endpoint coefficient changes sign between dimensions 47 and 48. The plotted values illustrate the obstruction; the sign assertions are proved below by rational root bounds and variational concavity. They are not inferred from the plot.
Retain the profiles of § 2.1 , now in any fixed dimension n ≥ 2 n\geq2 n ≥ 2 , and put m = n − 1 m=n-1 m = n − 1 . The defining ODEs and endpoint expansions do not require positivity of the calibration. The leading term of the determinant gives a necessary condition for the calibration to remain positive semidefinite.
Proposition 5.1. For every fixed dimension, as δ = k − r ↓ 0 \delta= k-r \downarrow0 δ = k − r ↓ 0 ,
a E − d 2 4 = K 2 δ 2 9 m k 2 { 3 ( k 2 − q 2 ) − 4 m } + O ( δ 3 ) . (33) aE-\frac{d^2}{4}=\frac{K^2\delta^2}{9mk^2}\left\{3\left(k^2-q^2\right)-4m\right\}+O\left(\delta^3\right). \tag*{(33)} a E − 4 d 2 = 9 m k 2 K 2 δ 2 { 3 ( k 2 − q 2 ) − 4 m } + O ( δ 3 ) . (33) Consequently 3 ( k 2 − q 2 ) ≥ 4 m 3\left(k^2-q^2\right)\ge4m 3 ( k 2 − q 2 ) ≥ 4 m is necessary for the positive semidefiniteness of this calibration matrix throughout 0 < r < k 0<r<k 0 < r < k .
Proof. The endpoint expansions in § 2.1 follow solely from the radial equations and hold in every fixed dimension. With h b = 1 − λ − m / k 2 h_b=1-\lambda-m/k^2 h b = 1 − λ − m / k 2 , they give
g = β − β h b 6 δ 2 + O ( δ 3 ) , a = K h b 3 δ 2 + O ( δ 3 ) . g=\beta-\frac{\beta h_b}{6}\delta^2+O\left(\delta^3\right),\qquad a=\frac{K h_b}{3}\delta^2+O\left(\delta^3\right). g = β − 6 β h b δ 2 + O ( δ 3 ) , a = 3 K h b δ 2 + O ( δ 3 ) .
To recover the slope of the correction, substitute G = γ δ + O ( δ 2 ) G=\gamma\delta+O\left(\delta^2\right) G = γ δ + O ( δ 2 ) into (6) . The constant terms give − γ = 2 γ − 2 K / k -\gamma=2\gamma-2K/k − γ = 2 γ − 2 K / k , and hence
G = 2 K 3 k δ + O ( δ 2 ) , d = 2 K 3 k δ + O ( δ 2 ) , E = K m + O ( δ ) . G=\frac{2K}{3k}\delta+O\left(\delta^2\right),\qquad d=\frac{2K}{3k}\delta+O\left(\delta^2\right),\qquad E=\frac{K}{m}+O(\delta). G = 3 k 2 K δ + O ( δ 2 ) , d = 3 k 2 K δ + O ( δ 2 ) , E = m K + O ( δ ) .
Consequently
a E − d 2 4 = K 2 { h b 3 m − 1 9 k 2 } δ 2 + O ( δ 3 ) . aE-\frac{d^2}{4}=K^2\left\{\frac{h_b}{3m}-\frac{1}{9k^2}\right\}\delta^2+O\left(\delta^3\right). a E − 4 d 2 = K 2 { 3 m h b − 9 k 2 1 } δ 2 + O ( δ 3 ) .
Substituting h b = 1 − q 2 / k 2 − m / k 2 h_b=1-q^2/k^2-m/k^2 h b = 1 − q 2 / k 2 − m / k 2 proves (33) . Its leading prefactor is positive, so a negative bracket precludes positive semidefiniteness near k k k . □
Figure 5 resolves the endpoint behavior of the determinant itself. Its limiting values are given exactly by (33) ; the next step proves their signs without using the plotted approximations.
Figure 5. The calibration determinant near the outer endpoint. Here ε = 1 − r / k \varepsilon=1-r/k ε = 1 − r / k and D n ( ε ) = ( a E − d 2 / 4 ) / ( K 2 ε 2 ) D_n(\varepsilon)=(aE-d^2/4)/(K^2\varepsilon^2) D n ( ε ) = ( a E − d 2 /4 ) / ( K 2 ε 2 ) , with all profiles evaluated at r = k ( 1 − ε ) r=k(1-\varepsilon) r = k ( 1 − ε ) . The left panel shows 0 ≤ ε ≤ 0.1 0\leq\varepsilon\leq0.1 0 ≤ ε ≤ 0.1 ; the right panel magnifies 0 ≤ ε ≤ 5 × 10 − 4 0\leq\varepsilon\leq5\times10^{-4} 0 ≤ ε ≤ 5 × 1 0 − 4 and rescales both axes as indicated. The values at zero are the continuous limits supplied by (33) . The negative boundary interval in dimension 48 is visible only at the finer scale. These curves are illustrations; the strict signs are established by the determinant estimates and the exact endpoint analysis.
We give finite Taylor comparisons and an elementary zero-counting argument, so the root identification does not rely on a numerical root finder. Define the rational polynomial
P n , N ( r ) = ∑ j = 0 N t j , t 0 = 1 , t j = − r 2 2 j ( n + 2 j − 2 ) t j − 1 . P_{n,N}(r)=\sum_{j=0}^{N}t_j,\qquad t_0=1,\qquad t_j=-\frac{r^2}{2j(n+2j-2)}t_{j-1}. P n , N ( r ) = j = 0 ∑ N t j , t 0 = 1 , t j = − 2 j ( n + 2 j − 2 ) r 2 t j − 1 .
For all arguments below, r < 30 r<30 r < 30 and n ≥ 47 n\geq47 n ≥ 47 . The magnitudes of the terms decrease strictly from index 61 onward, since their successive ratio is at most 900 / [ 124 ( 47 + 122 ) ] < 1 900/[124(47+122)]<1 900/ [ 124 ( 47 + 122 )] < 1 . The alternating tail therefore gives
P n , 61 ( r ) < ψ n ( r ) < P n , 60 ( r ) . (34) P_{n,61}(r)<\psi_n(r)<P_{n,60}(r). \tag*{(34)} P n , 61 ( r ) < ψ n ( r ) < P n , 60 ( r ) . (34)
The following are finite rational inequalities. All terminating decimals in the table denote exact rational numbers; the recurrence defining P n , N P_{n,N} P n , N specifies the arithmetic completely.
Table 15.
For an integer-arithmetic verification of the table, let r = v / w r=v/w r = v / w in positive integers and N = 60 N=60 N = 60 or 61 61 61 , multiply P n , N ( v / w ) − η P_{n,N}(v/w)-\eta P n , N ( v / w ) − η by the positive integer
M = w 2 N ∏ i = 1 N 2 i ( n + 2 i − 2 ) ⋅ den ( η ) . M=w^{2N}\prod_{i=1}^{N}2i(n+2i-2)\cdot\operatorname{den}(\eta). M = w 2 N i = 1 ∏ N 2 i ( n + 2 i − 2 ) ⋅ den ( η ) .
The required sign is the sign of the integer
den ( η ) ∑ j = 0 N ( − 1 ) j v 2 j w 2 N − 2 j ∏ i = j + 1 N 2 i ( n + 2 i − 2 ) − num ( η ) w 2 N ∏ i = 1 N 2 i ( n + 2 i − 2 ) . \operatorname{den}(\eta)\sum_{j=0}^{N}(-1)^jv^{2j}w^{2N-2j}\prod_{i=j+1}^{N}2i(n+2i-2)-\operatorname{num}(\eta)w^{2N}\prod_{i=1}^{N}2i(n+2i-2). den ( η ) j = 0 ∑ N ( − 1 ) j v 2 j w 2 N − 2 j i = j + 1 ∏ N 2 i ( n + 2 i − 2 ) − num ( η ) w 2 N i = 1 ∏ N 2 i ( n + 2 i − 2 ) .
To identify the roots in these sign-changing intervals as first roots, put ν = n / 2 − 1 \nu=n/2-1 ν = n /2 − 1 , a ν = ν 2 − 1 / 4 a_\nu=\sqrt{\nu^2-1/4} a ν = ν 2 − 1/4 , and
Y ( r ) = r ν + 1 / 2 ψ n ( r ) , Y ′ ′ + V ν ( r ) Y = 0 , V ν ( r ) = 1 − a ν 2 r 2 . Y(r)=r^{\nu+1/2}\psi_n(r),\qquad Y''+V_\nu(r)Y=0,\qquad V_\nu(r)=1-\frac{a_\nu^2}{r^2}. Y ( r ) = r ν + 1/2 ψ n ( r ) , Y ′′ + V ν ( r ) Y = 0 , V ν ( r ) = 1 − r 2 a ν 2 .
The regular expansion at zero gives Y , Y ′ > 0 Y,Y'>0 Y , Y ′ > 0 initially. As long as 0 < r ≤ a ν 0<r\leq a_\nu 0 < r ≤ a ν and Y > 0 Y>0 Y > 0 , one has Y ′ ′ ≥ 0 Y''\geq0 Y ′′ ≥ 0 . Hence Y , Y ′ > 0 Y,Y'>0 Y , Y ′ > 0 throughout this interval, including at a ν a_\nu a ν . Let B B B be the upper argument in the corresponding pair of rows. For each of the four orders ν = 22.5 , 23 , 23.5 , 24 \nu=22.5,23,23.5,24 ν = 22.5 , 23 , 23.5 , 24 , direct rational comparison gives
16 B 2 < 25 ( ν 2 − 1 / 4 ) , V ν ( r ) < 9 / 25 ( a ν ≤ r ≤ B ) . 16B^2<25(\nu^2-1/4),\qquad V_\nu(r)<9/25\quad(a_\nu\leq r\leq B). 16 B 2 < 25 ( ν 2 − 1/4 ) , V ν ( r ) < 9/25 ( a ν ≤ r ≤ B ) .
Comparison with frequency 3 / 5 3/5 3/5 shows that the first zero, if it lies before B B B , is strictly greater than a ν + 5 π / 6 a_\nu+5\pi/6 a ν + 5 π /6 . One can see this without invoking a zero theorem: before its first zero, the Duhamel formula about a ν a_\nu a ν is the positive homogeneous solution
Y ( a ν ) cos ( 3 t / 5 ) + 5 3 Y ′ ( a ν ) sin ( 3 t / 5 ) Y(a_\nu)\cos(3t/5)+\frac{5}{3}Y'(a_\nu)\sin(3t/5) Y ( a ν ) cos ( 3 t /5 ) + 3 5 Y ′ ( a ν ) sin ( 3 t /5 )
plus the integral with kernel 5 3 sin ( 3 ( t − v ) / 5 ) \frac{5}{3}\sin(3(t-v)/5) 3 5 sin ( 3 ( t − v ) /5 ) and forcing ( 9 / 25 − V ν ) Y (9/25-V_\nu)Y ( 9/25 − V ν ) Y . Every term is nonnegative, and the homogeneous solution is strictly positive, for 0 ≤ t ≤ 5 π / 6 0\leq t\leq5\pi/6 0 ≤ t ≤ 5 π /6 . Similarly, successive zeros have distance strictly greater than π \pi π : after a zero, choose the sign to make the derivative positive and use Duhamel comparison with frequency 1 1 1 ; the forcing is ( 1 − V ν ) Y > 0 (1-V_\nu)Y>0 ( 1 − V ν ) Y > 0 before the next zero. A next zero at distance at most π \pi π contradicts the positive sine formula and integral. Consequently a second zero before B B B would exceed
a ν + 11 π 6 > ν − 1 40 + 11 ( 3.14 ) 6 > ν + 5.7316 > B . a_\nu+\frac{11\pi}{6}>\nu-\frac{1}{40}+\frac{11(3.14)}{6}>\nu+5.7316>B. a ν + 6 11 π > ν − 40 1 + 6 11 ( 3.14 ) > ν + 5.7316 > B .
The table and (34) exhibit a zero before B B B , whereas this argument permits at most one zero there. Thus the first-root brackets are
n = 48 : 28.640 < q < 28.641 , 29.710 < k < 29.711 , n = 47 : 28.1043 < q < 28.1044 , 29.1755 < k < 29.1756. (35) \begin{aligned}
n=48:\quad&28.640<q<28.641,\qquad29.710<k<29.711,\\
n=47:\quad&28.1043<q<28.1044,\qquad29.1755<k<29.1756. \tag*{(35)}
\end{aligned} n = 48 : n = 47 : 28.640 < q < 28.641 , 29.710 < k < 29.711 , 28.1043 < q < 28.1044 , 29.1755 < k < 29.1756. (35)
The brackets in dimension forty-eight force a negative leading determinant coefficient.
Corollary 5.2. In dimension 48 the present calibration matrix is not positive semidefinite on the whole radial interval. In fact its radial–trace principal determinant is negative for all r < k r<k r < k sufficiently close to $k.
Proof. Here m = 47 m=47 m = 47 . The brackets (35) give the strict rational inequality
3 ( k 2 − q 2 ) − 188 < 3 { ( 29.711 ) 2 − ( 28.640 ) 2 } − 188 = − 518237 10 6 < 0. 3\left(k^2-q^2\right)-188<3\left\{(29.711)^2-(28.640)^2\right\}-188=-\frac{518237}{10^6}<0. 3 ( k 2 − q 2 ) − 188 < 3 { ( 29.711 ) 2 − ( 28.640 ) 2 } − 188 = − 1 0 6 518237 < 0.
Apply Proposition 5.1; its leading factor K 2 / ( 9 m k 2 ) K^2/(9mk^2) K 2 / ( 9 m k 2 ) is strictly positive.
For comparison, in dimension 47 the equally explicit lower bound is
3 ( k 2 − q 2 ) − 184 > 3 { ( 29.1755 ) 2 − ( 28.1044 ) 2 } − 184 = 5750267 10 8 > 0. 3\left(k^2-q^2\right)-184>3\left\{(29.1755)^2-(28.1044)^2\right\}-184=\frac{5750267}{10^8}>0. 3 ( k 2 − q 2 ) − 184 > 3 { ( 29.1755 ) 2 − ( 28.1044 ) 2 } − 184 = 1 0 8 5750267 > 0.
Thus the boundary principal determinant is positive sufficiently close to k k k in that dimension. This last statement is local: it neither proves full-interval positivity in dimension 47 nor establishes that 48 is the first dimension in which an interior obstruction occurs.
The conclusion for all higher dimensions follows from a variational concavity argument, rather than from a numerical trend in Bessel zeros. For real ν ≥ 1 / 2 \nu\ge1/2 ν ≥ 1/2 , let j ν j_\nu j ν denote the first positive zero of the regular solution of order ν \nu ν . Put
L ( t ) = j t 2 , t ≥ 1 / 4. L(t)=j_{\sqrt{t}}^2,\qquad t\ge1/4. L ( t ) = j t 2 , t ≥ 1/4.
Then
L ( t ) = inf 0 ≠ v ∈ H 0 1 ( 0 , 1 ) ∫ 0 1 { ∣ v ′ ∣ 2 + t − 1 / 4 r 2 ∣ v ∣ 2 } d r ∫ 0 1 ∣ v ∣ 2 d r . (36) L(t)=\inf_{0\ne v\in H_0^1(0,1)}
\frac{\int_0^1\left\{|v'|^2+\frac{t-1/4}{r^2}|v|^2\right\}\,dr}
{\int_0^1|v|^2\,dr}. \tag*{(36)} L ( t ) = 0 = v ∈ H 0 1 ( 0 , 1 ) inf ∫ 0 1 ∣ v ∣ 2 d r ∫ 0 1 { ∣ v ′ ∣ 2 + r 2 t − 1/4 ∣ v ∣ 2 } d r . (36)
All forms have the same domain: the elementary Hardy inequality ∫ 0 1 v 2 / r 2 d r ≤ 4 ∫ 0 1 ∣ v ′ ∣ 2 d r \int_0^1 v^2/r^2\,dr\le4\int_0^1|v'|^2\,dr ∫ 0 1 v 2 / r 2 d r ≤ 4 ∫ 0 1 ∣ v ′ ∣ 2 d r holds on H 0 1 ( 0 , 1 ) H_0^1(0,1) H 0 1 ( 0 , 1 ) . For a direct verification of (36) , the positive function
w ( r ) = r ν + 1 / 2 ψ ν + 2 ( j ν r ) w(r)=r^{\nu+1/2}\psi_{\nu+2}(j_\nu r) w ( r ) = r ν + 1/2 ψ ν + 2 ( j ν r )
satisfies the eigenvalue equation, vanishes at 1, and belongs to H 0 1 ( 0 , 1 ) H_0^1(0,1) H 0 1 ( 0 , 1 ) . For v ∈ C c ∞ ( 0 , 1 ) v\in C_c^\infty(0,1) v ∈ C c ∞ ( 0 , 1 ) , integration by parts gives
∫ 0 1 { ∣ v ′ ∣ 2 + ( ν 2 − 1 / 4 r 2 − j ν 2 ) v 2 } d r = ∫ 0 1 w 2 ∣ ( v w ) ′ ∣ 2 d r ≥ 0. \int_0^1\left\{|v'|^2+\left(\frac{\nu^2-1/4}{r^2}-j_\nu^2\right)v^2\right\}\,dr
=\int_0^1w^2\left|\left(\frac{v}{w}\right)'\right|^2\,dr\ge0. ∫ 0 1 { ∣ v ′ ∣ 2 + ( r 2 ν 2 − 1/4 − j ν 2 ) v 2 } d r = ∫ 0 1 w 2 ( w v ) ′ 2 d r ≥ 0.
Approximation, the Hardy bound, and evaluation on w w w prove (36) . For Hardy’s inequality and the density statement, see [21 ] .
Because the right-hand side of (36) is an infimum of affine functions of t t t , L L L is concave. Thus its secant slope
R ( ν ) = L ( ( ν + 1 ) 2 ) − L ( ν 2 ) ( ν + 1 ) 2 − ν 2 = j ν + 1 2 − j ν 2 2 ν + 1 R(\nu)=\frac{L\left((\nu+1)^2\right)-L\left(\nu^2\right)}{(\nu+1)^2-\nu^2}=\frac{j_{\nu+1}^2-j_\nu^2}{2\nu+1} R ( ν ) = ( ν + 1 ) 2 − ν 2 L ( ( ν + 1 ) 2 ) − L ( ν 2 ) = 2 ν + 1 j ν + 1 2 − j ν 2
is nonincreasing in ν \nu ν . Here one uses the elementary concave-function fact that a secant slope cannot increase when either endpoint of the secant interval is moved to the right (while preserving their order). This also handles overlapping intervals by first moving the right endpoint, then the left endpoint; for disjoint intervals the usual three-point secant inequalities apply.
For integer dimension n ≥ 3 n\geq3 n ≥ 3 , ν = n / 2 − 1 \nu=n/2-1 ν = n /2 − 1 and R ( ν ) = ( k 2 − q 2 ) / ( n − 1 ) R(\nu)=(k^2-q^2)/(n-1) R ( ν ) = ( k 2 − q 2 ) / ( n − 1 ) . The rational comparisons already proved show
R ( 22.5 ) > 4 3 , R ( 23 ) < 4 3 . R(22.5)>\frac{4}{3},\qquad R(23)<\frac{4}{3}. R ( 22.5 ) > 3 4 , R ( 23 ) < 3 4 .
It follows that
{ 3 ( k 2 − q 2 ) − 4 ( n − 1 ) > 0 , 3 ≤ n ≤ 47 , 3 ( k 2 − q 2 ) − 4 ( n − 1 ) < 0 , n ≥ 48. \begin{cases}
3\left(k^2-q^2\right)-4(n-1)>0, & 3\leq n\leq47,\\
3\left(k^2-q^2\right)-4(n-1)<0, & n\geq48.
\end{cases} { 3 ( k 2 − q 2 ) − 4 ( n − 1 ) > 0 , 3 ( k 2 − q 2 ) − 4 ( n − 1 ) < 0 , 3 ≤ n ≤ 47 , n ≥ 48.
In particular the present pointwise calibration fails in every dimension n ≥ 48 n\geq48 n ≥ 48 , already in a left neighborhood of its outer endpoint. This endpoint argument alone gives only a local statement in dimensions 3 through 47. For these dimensions, full-interval positivity is supplied by Theorem 2.10 .
The planar coefficient estimates
The dimension-independent algebra in § 2.1 and § 2.3 also applies when n = 2 n=2 n = 2 . We therefore need only the planar coefficient signs. The perspective map and the correction profile are those of Tang and Zhang [19 ] ; the argument below estimates the residue coefficients by comparing the first pole with the remaining signed tail.
Zero geometry and the first pole
Put
q = j 0 , 1 , k = j 1 , 1 , α = q / k , λ = α 2 , c = − J 0 ( k ) , σ = 1 + c , β = α J 1 ( q ) / c , K = β 2 . q=j_{0,1},\qquad k=j_{1,1},\qquad\alpha=q/k,\qquad\lambda=\alpha^2,\qquad c=-J_0(k),\qquad\sigma=1+c,\qquad\beta=\alpha J_1(q)/c,\qquad K=\beta^2. q = j 0 , 1 , k = j 1 , 1 , α = q / k , λ = α 2 , c = − J 0 ( k ) , σ = 1 + c , β = α J 1 ( q ) / c , K = β 2 .
For 0 < r < k 0<r<k 0 < r < k write
X = J 0 ( r ) , p = J 1 ( r ) , u = X + c , ϕ = J 0 ( α r ) , S = p 2 + u p r , s = u / p , X=J_0(r),\qquad p=J_1(r),\qquad u=X+c,\qquad\phi=J_0(\alpha r),\qquad S=p^2+upr,\qquad s=u/p, X = J 0 ( r ) , p = J 1 ( r ) , u = X + c , ϕ = J 0 ( α r ) , S = p 2 + u p r , s = u / p ,
f ( s ( r ) ) = ϕ / p , z = f s , g = f − s z , ρ = u p / S . f(s(r))=\phi/p,\qquad z=f_s,\qquad g=f-sz,\qquad\rho=up/S. f ( s ( r )) = ϕ / p , z = f s , g = f − sz , ρ = u p / S .
Interlacing gives p , u , ϕ > 0 p,u,\phi>0 p , u , ϕ > 0 . The equations are X r = − p X_r=-p X r = − p , p r = X − p / r p_r=X-p/r p r = X − p / r , and ϕ r r + ϕ r / r + λ ϕ = 0 \phi_{rr}+\phi_r/r+\lambda\phi=0 ϕ r r + ϕ r / r + λ ϕ = 0 . Also u ( r ) = ∫ r k p ( v ) d v u(r)=\int_r^k p(v)\,\mathrm{d}v u ( r ) = ∫ r k p ( v ) d v .
Write q l = j 0 , l q_l=j_{0,l} q l = j 0 , l and κ l = j 1 , l \kappa_l=j_{1,l} κ l = j 1 , l . The ratios q l / κ l q_l/\kappa_l q l / κ l strictly increase. Here is a proof that remains valid in dimension two and does not use a higher-dimensional Sonin sign. On the quarter-cycle from q l q_l q l to κ l \kappa_l κ l , polar coordinates X = R cos θ X=R\cos\theta X = R cos θ , p = R sin θ p=R\sin\theta p = R sin θ give
d r l d φ = 2 r l 2 r l + sin 2 φ , 0 ≤ φ ≤ π / 2. \frac{\mathrm{d}r_l}{\mathrm{d}\varphi}=\frac{2r_l}{2r_l+\sin2\varphi},\qquad0\leq\varphi\leq\pi/2. d φ d r l = 2 r l + sin 2 φ 2 r l , 0 ≤ φ ≤ π /2.
The trajectories strictly increase with l l l , while
log κ l q l = ∫ 0 π / 2 2 d φ 2 r l + sin 2 φ \log\frac{\kappa_l}{q_l}=\int_0^{\pi/2}\frac{2\,\mathrm{d}\varphi}{2r_l+\sin2\varphi} log q l κ l = ∫ 0 π /2 2 r l + sin 2 φ 2 d φ
strictly decreases. This proves the ratio assertion.
Consecutive positive zeros of J 1 J_1 J 1 have distance greater than π \pi π : apply Wirtinger’s inequality between zeros to Y = r J 1 ( r ) Y=\sqrt{r}J_1(r) Y = r J 1 ( r ) , which solves Y ′ ′ + ( 1 − 3 / ( 4 r 2 ) ) Y = 0 Y''+(1-3/(4r^2))Y=0 Y ′′ + ( 1 − 3/ ( 4 r 2 )) Y = 0 . The same variational equation on ( 0 , 1 ) (0,1) ( 0 , 1 ) gives k 2 > 3 / 4 + π 2 k^2>3/4+\pi^2 k 2 > 3/4 + π 2 . Sturm comparison on [ k , κ 2 ] [k,\kappa_2] [ k , κ 2 ] then gives
κ 2 − k < π k k 2 − 3 / 4 < k . \kappa_2-k<\frac{\pi k}{\sqrt{k^2-3/4}}<k. κ 2 − k < k 2 − 3/4 π k < k .
The quarter-phase speed is strictly greater than one, so k − q < π / 2 k-q<\pi/2 k − q < π /2 . Thus
κ 2 < 2 k < π 1 − α , k < α κ 2 < q 2 . \kappa_2<2k<\frac{\pi}{1-\alpha},\qquad k<\alpha\kappa_2<q_2. κ 2 < 2 k < 1 − α π , k < α κ 2 < q 2 .
For the last lower bound use α κ 2 > κ 2 − π > k \alpha\kappa_2>\kappa_2-\pi>k α κ 2 > κ 2 − π > k ; the upper bound is the zero-ratio assertion just proved.
The following rational root bounds give the constants needed for the pole estimates. For ν = 0 , 1 \nu=0,1 ν = 0 , 1 set
S ν , N ( v ) = ∑ j = 0 N ( − 1 ) j ( v / 2 ) 2 j + ν j ! ( j + ν ) ! . S_{\nu,N}(v)=\sum_{j=0}^{N}\frac{(-1)^j(v/2)^{2j+\nu}}{j!(j+\nu)!}. S ν , N ( v ) = j = 0 ∑ N j ! ( j + ν )! ( − 1 ) j ( v /2 ) 2 j + ν .
At the arguments below the alternating tail decreases from index two onwards. Hence S ν , 7 < J ν < S ν , 6 S_{\nu,7}<J_\nu<S_{\nu,6} S ν , 7 < J ν < S ν , 6 , and direct rational multiplication gives just four comparisons:
Table 16.
The series also gives J 1 ( r ) > r / 2 − r 3 / 16 ≥ 0 J_1(r)>r/2-r^3/16\geq0 J 1 ( r ) > r /2 − r 3 /16 ≥ 0 for 0 < r ≤ 8 0<r\leq\sqrt{8} 0 < r ≤ 8 . Zero spacing greater than π \pi π therefore makes the indicated J 1 J_1 J 1 zero the first one; interlacing identifies the J 0 J_0 J 0 zero. Thus
12 5 < q < 49 20 , 19 5 < k < 77 20 , 48 77 < α < 49 76 . \frac{12}{5}<q<\frac{49}{20},\qquad\frac{19}{5}<k<\frac{77}{20},\qquad\frac{48}{77}<\alpha<\frac{49}{76}. 5 12 < q < 20 49 , 5 19 < k < 20 77 , 77 48 < α < 76 49 .
The bounds 157 / 50 < π < 22 / 7 157/50<\pi<22/7 157/50 < π < 22/7 used below were established at the start of Section 3.
For the pole estimates, define the auxiliary quantities
B ( r ) = r ( X 2 + p 2 ) − X p , L ( r ) = K B ( r ) − α B ( α r ) , \mathcal{B}(r) = r\left(X^2+p^2\right)-Xp,\qquad\mathcal{L}(r)=K\mathcal{B}(r)-\alpha\mathcal{B}(\alpha r), B ( r ) = r ( X 2 + p 2 ) − X p , L ( r ) = K B ( r ) − α B ( α r ) ,
Q ( r ) = L ( r ) / r , G = K − r 2 Q / p 2 . Q(r)=\mathcal{L}(r)/r,\qquad G=K-r^2Q/p^2. Q ( r ) = L ( r ) / r , G = K − r 2 Q / p 2 .
Differentiation and integration give
B ′ = X p / r , 2 B ( r ) − p ( r ) 2 / r = 3 ∫ 0 r p ( v ) 2 / v 2 d v , \mathcal{B}'=Xp/r,\qquad2\mathcal{B}(r)-p(r)^2/r=3\int_0^r p(v)^2/v^2\,\mathrm{d}v, B ′ = X p / r , 2 B ( r ) − p ( r ) 2 / r = 3 ∫ 0 r p ( v ) 2 / v 2 d v ,
B ( q l + 1 ) − B ( q l ) = − 1 2 ∫ q l q l + 1 X ( v ) 2 / v 2 d v . \mathcal{B}(q_{l+1})-\mathcal{B}(q_l)=-\frac{1}{2}\int_{q_l}^{q_{l+1}}X(v)^2/v^2\,\mathrm{d}v. B ( q l + 1 ) − B ( q l ) = − 2 1 ∫ q l q l + 1 X ( v ) 2 / v 2 d v .
The successive minima B ( κ l ) \mathcal{B}(\kappa_l) B ( κ l ) increase strictly and the successive maxima B ( q l ) \mathcal{B}(q_l) B ( q l ) decrease strictly. In particular
0 < B ( κ l ) ≤ B ( v ) ≤ B ( q l ) ( v ≥ q l ) , K B ( k ) = α B ( q ) . 0<\mathcal{B}(\kappa_l)\leq\mathcal{B}(v)\leq\mathcal{B}(q_l)\quad(v\geq q_l),\qquad K\mathcal{B}(k)=\alpha\mathcal{B}(q). 0 < B ( κ l ) ≤ B ( v ) ≤ B ( q l ) ( v ≥ q l ) , K B ( k ) = α B ( q ) .
For every l ≥ 2 l\geq2 l ≥ 2 , these identities imply Q ( κ l ) > 0 Q(\kappa_l)>0 Q ( κ l ) > 0 . Also
R 1 : = B ( q ) B ( k ) = K / α > 1. R_1:=\frac{\mathcal{B}(q)}{\mathcal{B}(k)}=K/\alpha>1. R 1 := B ( k ) B ( q ) = K / α > 1.
The energy ratio also determines the sign of the first coefficient at the center. The quarter-phase formula for the energy ratio gives
log R 1 = ∫ 0 π / 2 2 sin 2 φ ( 2 r 1 + sin 2 φ ) 2 d φ . \log R_1=\int_0^{\pi/2}\frac{2\sin2\varphi}{(2r_1+\sin2\varphi)^2}\,\mathrm{d}\varphi. log R 1 = ∫ 0 π /2 ( 2 r 1 + sin 2 φ ) 2 2 sin 2 φ d φ .
The measure sin 2 φ d φ \sin2\varphi\,\mathrm{d}\varphi sin 2 φ d φ has mass one and mean π / 4 \pi/4 π /4 for sin 2 φ \sin2\varphi sin 2 φ . Jensen’s inequality and r 1 < 77 / 20 r_1<77/20 r 1 < 77/20 give
log R 1 > 2 ( 77 / 10 + π / 4 ) 2 > 2 ( 297 / 35 ) 2 > 1 37 . \log R_1>\frac{2}{(77/10+\pi/4)^2}>\frac{2}{(297/35)^2}>\frac{1}{37}. log R 1 > ( 77/10 + π /4 ) 2 2 > ( 297/35 ) 2 2 > 37 1 .
Consequently K > ( 38 / 37 ) α K>(38/37)\alpha K > ( 38/37 ) α . The function 38 / 37 − 2 t + t 3 38/37-2t+t^3 38/37 − 2 t + t 3 decreases on 0 < t ≤ 49 / 76 0<t\leq49/76 0 < t ≤ 49/76 , and
38 37 − 2 49 76 + ( 49 76 ) 3 = 90325 16242112 > 0. \frac{38}{37}-2\frac{49}{76}+\left(\frac{49}{76}\right)^3=\frac{90325}{16242112}>0. 37 38 − 2 76 49 + ( 76 49 ) 3 = 16242112 90325 > 0.
The center coefficient therefore has the sign
a 1 : = K − 2 λ + λ 2 4 > 0. (37) a_1:=\frac{K-2\lambda+\lambda^2}{4}>0. \tag*{(37)} a 1 := 4 K − 2 λ + λ 2 > 0. (37)
The higher coefficients require a quantitative lower bound for the first retained pole. Put κ = κ 2 \kappa=\kappa_2 κ = κ 2 , x = α κ x=\alpha\kappa x = α κ , δ = k − q \delta=k-q δ = k − q , A = 3 / 4 A=3/4 A = 3/4 , and L = π / ( 1 − α ) L=\pi/(1-\alpha) L = π / ( 1 − α ) . The root bounds and Sturm comparison give
δ < 29 / 20 , k 2 − A > ( 37 / 10 ) 2 , \delta<29/20,\qquad k^2-A>(37/10)^2, δ < 29/20 , k 2 − A > ( 37/10 ) 2 ,
κ < k + π k k 2 − A < 77 20 + 7 7 22 37 < 71 10 , L > 12089 1450 > 25 3 . \kappa<k+\frac{\pi k}{\sqrt{k^2-A}}<\frac{77}{20}+\frac{7}{7}\frac{22}{37}<\frac{71}{10},\qquad L>\frac{12089}{1450}>\frac{25}{3}. κ < k + k 2 − A π k < 20 77 + 7 7 37 22 < 10 71 , L > 1450 12089 > 3 25 .
Here t / t 2 − A t/\sqrt{t^2-A} t / t 2 − A decreases, which explains the substitutions in the second line. It follows that
D : = α 2 − A / L 2 > ( 48 77 ) 2 − 27 2500 = 3 8 + 82959 29645000 > 3 8 . (38) D := \alpha^2-A/L^2 > \left(\frac{48}{77}\right)^2-\frac{27}{2500}=\frac{3}{8}+\frac{82959}{29645000}>\frac{3}{8}. \tag*{(38)} D := α 2 − A / L 2 > ( 77 48 ) 2 − 2500 27 = 8 3 + 29645000 82959 > 8 3 . (38)
We next obtain two phase gaps without a q 2 q_2 q 2 or κ 2 \kappa_2 κ 2 root table. The transformed J 0 J_0 J 0 potential on [ q , q 2 ] [q,q_2] [ q , q 2 ] is 1 + 1 / ( 4 r 2 ) < 601 / 576 1+1/(4r^2)<601/576 1 + 1/ ( 4 r 2 ) < 601/576 . Sturm comparison gives
q 2 − q > 24 π 601 > 3. q_2-q>\frac{24\pi}{\sqrt{601}}>3. q 2 − q > 601 24 π > 3.
Let r 1 , r 2 r_1,r_2 r 1 , r 2 be the quarter-phase trajectories from q , k q,k q , k and q 2 , κ q_2,\kappa q 2 , κ , respectively. Their differential equation has an increasing right side, so r 2 − r 1 ≥ q 2 − q > 3 r_2-r_1\ge q_2-q>3 r 2 − r 1 ≥ q 2 − q > 3 . Set
I ∗ : = log q 2 / κ q / k , E ∗ : = log R 1 R 2 , R 2 = B ( q 2 ) / B ( κ ) . I_*:=\log\frac{q_2/\kappa}{q/k},\qquad E_*:=\log\frac{R_1}{R_2},\qquad R_2=\mathcal{B}(q_2)/\mathcal{B}(\kappa). I ∗ := log q / k q 2 / κ , E ∗ := log R 2 R 1 , R 2 = B ( q 2 ) / B ( κ ) .
Differentiating these ratios along the quarter-phase gives
I ∗ = ∫ 0 π / 2 4 ( r 2 − r 1 ) ( 2 r 1 + sin 2 φ ) ( 2 r 2 + sin 2 φ ) d φ , I_*=\int_0^{\pi/2}\frac{4(r_2-r_1)}{(2r_1+\sin2\varphi)(2r_2+\sin2\varphi)}\,\mathrm{d}\varphi, I ∗ = ∫ 0 π /2 ( 2 r 1 + sin 2 φ ) ( 2 r 2 + sin 2 φ ) 4 ( r 2 − r 1 ) d φ ,
E ∗ = ∫ 0 π / 2 2 sin 2 φ { 1 ( 2 r 1 + sin 2 φ ) 2 − 1 ( 2 r 2 + sin 2 φ ) 2 } d φ . E_*=\int_0^{\pi/2}2\sin2\varphi\left\{\frac{1}{(2r_1+\sin2\varphi)^2}-\frac{1}{(2r_2+\sin2\varphi)^2}\right\}\,\mathrm{d}\varphi. E ∗ = ∫ 0 π /2 2 sin 2 φ { ( 2 r 1 + sin 2 φ ) 2 1 − ( 2 r 2 + sin 2 φ ) 2 1 } d φ .
Use r 1 < 77 / 20 r_1<77/20 r 1 < 77/20 , r 2 < 71 / 10 r_2<71/10 r 2 < 71/10 , and the gap > 3 >3 > 3 . Jensen’s inequality in the first integral, with mean 2 / π 2/\pi 2/ π , gives
I ∗ > 18 ( 77 / 10 + 2 / 3 ) ( 71 / 5 + 2 / 3 ) = 8100 55973 > 1 7 . I_*>\frac{18}{(77/10+2/3)(71/5+2/3)}=\frac{8100}{55973}>\frac{1}{7}. I ∗ > ( 77/10 + 2/3 ) ( 71/5 + 2/3 ) 18 = 55973 8100 > 7 1 .
In the second integral, use the probability measure sin 2 φ d φ \sin2\varphi\,\mathrm{d}\varphi sin 2 φ d φ with mean π / 4 \pi/4 π /4 and the decreasing convex kernel ( a + b ) / ( a 2 b 2 ) (a+b)/(a^2b^2) ( a + b ) / ( a 2 b 2 ) . With D k = 297 / 35 D_k=297/35 D k = 297/35 , D c = 1049 / 70 D_c=1049/70 D c = 1049/70 one obtains
E ∗ > 12 ( D k + D c ) D k 2 D c 2 > 12 ( 117 / 5 ) ( 17 / 2 ) 2 15 2 > 1 58 . E_*>\frac{12(D_k+D_c)}{D_k^2D_c^2}>\frac{12(117/5)}{(17/2)^2 15^2}>\frac{1}{58}. E ∗ > D k 2 D c 2 12 ( D k + D c ) > ( 17/2 ) 2 1 5 2 12 ( 117/5 ) > 58 1 .
These kernels are convex because they are sums of products of positive inverse powers of positive affine functions.
Let θ \theta θ be the polar phase and ϑ = 3 π / 2 − θ ( x ) ∈ ( 0 , π / 2 ) \vartheta=3\pi/2-\theta(x)\in(0,\pi/2) ϑ = 3 π /2 − θ ( x ) ∈ ( 0 , π /2 ) . Writing d = q 2 − x d=q_2-x d = q 2 − x , the backwards phase equation gives
ϑ ≥ x ( 1 − e − d / x ) ≥ d 1 + d / x > x 8 > 1 2 . \vartheta\ge x\left(1-e^{-d/x}\right)\ge\frac{d}{1+d/x}>\frac{x}{8}>\frac{1}{2}. ϑ ≥ x ( 1 − e − d / x ) ≥ 1 + d / x d > 8 x > 2 1 .
Indeed d = x ( e I ∗ − 1 ) > x / 7 d=x(e^{I_*}-1)>x/7 d = x ( e I ∗ − 1 ) > x /7 , and x > α ( k + π ) > ( 48 / 77 ) ( 34 / 5 ) > 4 x>\alpha(k+\pi)>(48/77)(34/5)>4 x > α ( k + π ) > ( 48/77 ) ( 34/5 ) > 4 . The increasing quarter-cycle lengths and the Sturm spacing bound give
ϑ ≤ d < δ κ − k k < δ π k 2 − A < 319 259 . \vartheta\le d<\delta\frac{\kappa-k}{k}<\frac{\delta\pi}{\sqrt{k^2-A}}<\frac{319}{259}. ϑ ≤ d < δ k κ − k < k 2 − A δ π < 259 319 .
Thus 1 < 2 ϑ < 638 / 259 < π 1 < 2\vartheta< 638/259 < \pi 1 < 2 ϑ < 638/259 < π . Concavity of sine shows
sin 2 ϑ > 49 80 . (39) \sin2\vartheta> \frac{49}{80}. \tag*{(39)} sin 2 ϑ > 80 49 . (39)
For a fully elementary check of the upper endpoint, the complementary angle t = π − 638 / 259 t=\pi-638/259 t = π − 638/259 lies between 2 / 3 2/3 2/3 and 1 1 1 , by the displayed rational bounds on π \pi π . Hence sin t > t − t 3 / 6 > 50 / 81 > 49 / 80 \sin t>t-t^3/6>50/81>49/80 sin t > t − t 3 /6 > 50/81 > 49/80 ; at the lower endpoint use sin 1 > 5 / 6 > 49 / 80 \sin1>5/6>49/80 sin 1 > 5/6 > 49/80 .
The exact logarithmic derivative at x x x is
u ∗ : = x B ′ ( x ) B ( x ) = sin 2 ϑ / ( 2 x ) 1 − sin 2 ϑ / ( 2 x ) . u_* := \frac{xB'(x)}{B(x)}=\frac{\sin2\vartheta/(2x)}{1-\sin2\vartheta/(2x)}. u ∗ := B ( x ) x B ′ ( x ) = 1 − sin 2 ϑ / ( 2 x ) sin 2 ϑ / ( 2 x ) .
Since x < ( 49 / 76 ) ( 71 / 10 ) x<(49/76)(71/10) x < ( 49/76 ) ( 71/10 ) , (A.3) gives
sin 2 ϑ 2 x > 19 284 , u ∗ > 19 265 . \frac{\sin2\vartheta}{2x}>\frac{19}{284}, \qquad u_*>\frac{19}{265}. 2 x sin 2 ϑ > 284 19 , u ∗ > 265 19 .
Put σ ∗ : = B ( x ) / ( R 1 B ( κ ) ) \sigma_*:=B(x)/(R_1B(\kappa)) σ ∗ := B ( x ) / ( R 1 B ( κ )) . Then 0 < σ ∗ < e − E ∗ < 58 / 59 0<\sigma_*<e^{-E_*}<58/59 0 < σ ∗ < e − E ∗ < 58/59 , and the normalized first pole is exactly
U 2 : = 3 Q ( κ ) + α X ( x ) p ( x ) / κ K X ( κ ) 2 = 3 − σ ∗ ( 3 − u ∗ ) . U_2:=\frac{3Q(\kappa)+\alpha X(x)p(x)/\kappa}{KX(\kappa)^2}=3-\sigma_*(3-u_*). U 2 := K X ( κ ) 2 3 Q ( κ ) + α X ( x ) p ( x ) / κ = 3 − σ ∗ ( 3 − u ∗ ) .
If u ∗ ≥ 3 u_*\ge3 u ∗ ≥ 3 this is at least three. Otherwise the previous bounds give
U 2 > 19 / 265 + 3 / 58 1 + 1 / 58 = 1897 15635 > 3 25 . (40) U_2>\frac{19/265+3/58}{1+1/58}=\frac{1897}{15635}>\frac{3}{25}. \tag*{(40)} U 2 > 1 + 1/58 19/265 + 3/58 = 15635 1897 > 25 3 . (40)
Tail control and the coefficient expansion
We now compare this positive contribution with the possible negative tail. For each ℓ ≥ 2 \ell\ge2 ℓ ≥ 2 put
U ℓ = 3 Q ( κ ℓ ) + α X ( α κ ℓ ) p ( α κ ℓ ) / κ ℓ K X ( κ ℓ ) 2 . U_\ell=\frac{3Q(\kappa_\ell)+\alpha X(\alpha\kappa_\ell)p(\alpha\kappa_\ell)/\kappa_\ell}{KX(\kappa_\ell)^2}. U ℓ = K X ( κ ℓ ) 2 3 Q ( κ ℓ ) + α X ( α κ ℓ ) p ( α κ ℓ ) / κ ℓ .
If κ ℓ < L \kappa_\ell<L κ ℓ < L , zero spacing and the zero-ratio theorem put κ ℓ − 1 < α κ ℓ < q ℓ \kappa_{\ell-1}<\alpha\kappa_\ell<q_\ell κ ℓ − 1 < α κ ℓ < q ℓ . Then X , p X,p X , p have the same sign, so U ℓ > 0 U_\ell>0 U ℓ > 0 . For κ ℓ ≥ L \kappa_\ell\ge L κ ℓ ≥ L , set v = α κ ℓ v=\alpha\kappa_\ell v = α κ ℓ , Y ( v ) = v J 1 ( v ) Y(v)=\sqrt{v}J_1(v) Y ( v ) = v J 1 ( v ) and E ( v ) = Y ′ 2 + ( 1 − A / v 2 ) Y 2 \mathcal{E}(v)=Y'^2+(1-A/v^2)Y^2 E ( v ) = Y ′2 + ( 1 − A / v 2 ) Y 2 . The Sonin identity E ′ = 2 A Y 2 / v 3 ≥ 0 \mathcal{E}'=2AY^2/v^3\ge0 E ′ = 2 A Y 2 / v 3 ≥ 0 , integrated over a nonempty interval on which Y Y Y is not identically zero, gives E ( v ) < E ( κ ℓ ) = κ ℓ X ( κ ℓ ) 2 \mathcal{E}(v)<\mathcal{E}(\kappa_\ell)=\kappa_\ell X(\kappa_\ell)^2 E ( v ) < E ( κ ℓ ) = κ ℓ X ( κ ℓ ) 2 . A direct substitution gives the exact numerator
κ ℓ K X ( κ ℓ ) 2 U ℓ = 3 K κ ℓ X ( κ ℓ ) 2 − 3 α { E ( v ) + Y 2 3 v 2 − Y Y ′ 3 v } . \kappa_\ell KX(\kappa_\ell)^2U_\ell=3K\kappa_\ell X(\kappa_\ell)^2-3\alpha\left\{\mathcal{E}(v)+\frac{Y^2}{3v^2}-\frac{YY'}{3v}\right\}. κ ℓ K X ( κ ℓ ) 2 U ℓ = 3 K κ ℓ X ( κ ℓ ) 2 − 3 α { E ( v ) + 3 v 2 Y 2 − 3 v Y Y ′ } .
Relative to E \mathcal{E} E , the two correction eigenvalues are bounded above by the sum 1 / [ 3 ( v 2 − A ) ] + 1 / [ 6 v 2 − A ] 1/[3(v^2-A)]+1/[6\sqrt{v^2-A}] 1/ [ 3 ( v 2 − A )] + 1/ [ 6 v 2 − A ] . Using K / α = R 1 > 1 K/\alpha=R_1>1 K / α = R 1 > 1 and (A.2) therefore gives
U ℓ > − 1 2 D κ ℓ − 1 D κ ℓ 2 . U_\ell>-\frac{1}{2\sqrt{D}\,\kappa_\ell}-\frac{1}{D\kappa_\ell^2}. U ℓ > − 2 D κ ℓ 1 − D κ ℓ 2 1 .
The zero spacing greater than π \pi π implies, for s > 1 s>1 s > 1 ,
∑ κ l ≥ L κ l − s ≤ L − s + L 1 − s ( s − 1 ) π . \sum_{\kappa_l\ge L}\kappa_l^{-s}\le L^{-s}+\frac{L^{1-s}}{(s-1)\pi}. κ l ≥ L ∑ κ l − s ≤ L − s + ( s − 1 ) π L 1 − s .
Consequently, for every j ≥ 2 j\ge2 j ≥ 2 ,
κ 2 2 j ∑ l ≥ 2 U l κ l 2 j > U 2 − V ~ , \kappa_2^{2j}\sum_{l\ge2}\frac{U_l}{\kappa_l^{2j}} > U_2-\widetilde V, κ 2 2 j l ≥ 2 ∑ κ l 2 j U l > U 2 − V ,
where
V ~ = 1 2 D ( κ 2 L ) 4 ( 1 L + 1 4 π ) + 1 D ( κ 2 L ) 4 ( 1 L 2 + 1 5 π L ) . \widetilde V=\frac{1}{2\sqrt D}\left(\frac{\kappa_2}{L}\right)^4\left(\frac{1}{L}+\frac{1}{4\pi}\right)+\frac{1}{D}\left(\frac{\kappa_2}{L}\right)^4\left(\frac{1}{L^2}+\frac{1}{5\pi L}\right). V = 2 D 1 ( L κ 2 ) 4 ( L 1 + 4 π 1 ) + D 1 ( L κ 2 ) 4 ( L 2 1 + 5 π L 1 ) .
Use D > 3 / 8 D>3/8 D > 3/8 , D > 3 / 5 \sqrt D>3/5 D > 3/5 , L > 25 / 3 L>25/3 L > 25/3 , κ 2 < 71 / 10 \kappa_2<71/10 κ 2 < 71/10 , and π > 25 / 8 \pi>25/8 π > 25/8 . These yield the single rational bound
V ~ < ( 1 6 + 184 3125 ) ( 213 250 ) 4 < 3 25 . \widetilde V<\left(\frac{1}{6}+\frac{184}{3125}\right)\left(\frac{213}{250}\right)^4<\frac{3}{25}. V < ( 6 1 + 3125 184 ) ( 250 213 ) 4 < 25 3 .
The last positive difference is 28105528377 / 24414062500000 28105528377/24414062500000 28105528377/24414062500000 . Together with (A.4), this proves that the weighted sum of the U l U_l U l is positive at every order j ≥ 2 j\ge2 j ≥ 2 , without requiring an individual sign at all far poles.
We obtain the coefficient signs by expressing the Taylor coefficients as sums of paired residues. The points 0 , ± k 0,\pm k 0 , ± k are removable and G G G is even meromorphic. The ODE Taylor expansions, for δ = k − r \delta=k-r δ = k − r , begin
p c = δ + δ 2 2 k + 3 / k 2 − 1 6 δ 3 + ( 1 2 k 3 − 1 12 k ) δ 4 + O ( δ 5 ) . \frac{p}{c}=\delta+\frac{\delta^2}{2k}+\frac{3/k^2-1}{6}\delta^3+\left(\frac{1}{2k^3}-\frac{1}{12k}\right)\delta^4+O(\delta^5). c p = δ + 2 k δ 2 + 6 3/ k 2 − 1 δ 3 + ( 2 k 3 1 − 12 k 1 ) δ 4 + O ( δ 5 ) .
ϕ β c = δ + δ 2 2 k + 2 / k 2 − λ 6 δ 3 + ( 1 4 k 3 − λ 12 k ) δ 4 + O ( δ 5 ) . \frac{\phi}{\beta c}=\delta+\frac{\delta^2}{2k}+\frac{2/k^2-\lambda}{6}\delta^3+\left(\frac{1}{4k^3}-\frac{\lambda}{12k}\right)\delta^4+O(\delta^5). β c ϕ = δ + 2 k δ 2 + 6 2/ k 2 − λ δ 3 + ( 4 k 3 1 − 12 k λ ) δ 4 + O ( δ 5 ) .
Substitution gives G ( k ) = 0 G(k)=0 G ( k ) = 0 and G r ( k ) = − 2 K / ( 3 k ) G_r(k)=-2K/(3k) G r ( k ) = − 2 K / ( 3 k ) . At each further zero κ \kappa κ , one has p r ( κ ) = X ( κ ) p_r(\kappa)=X(\kappa) p r ( κ ) = X ( κ ) , p r r ( κ ) = − X ( κ ) / κ p_{rr}(\kappa)=-X(\kappa)/\kappa p r r ( κ ) = − X ( κ ) / κ , and κ L ′ ( κ ) = ϕ ( κ ) ϕ r ( κ ) \kappa L'(\kappa)=\phi(\kappa)\phi_r(\kappa) κ L ′ ( κ ) = ϕ ( κ ) ϕ r ( κ ) . Thus the paired-pole calculation gives, for j ≥ 2 j\ge2 j ≥ 2 ,
a j : = − [ r 2 j ] G = 2 ∑ l ≥ 2 K X ( κ l ) 2 U l + 2 ( j − 2 ) Q ( κ l ) X ( κ l ) 2 κ l 2 j > 0. (41) a_j:=-[r^{2j}]G=2\sum_{l\ge2}\frac{KX(\kappa_l)^2U_l+2(j-2)Q(\kappa_l)}{X(\kappa_l)^2\kappa_l^{2j}}>0. \tag*{(41)} a j := − [ r 2 j ] G = 2 l ≥ 2 ∑ X ( κ l ) 2 κ l 2 j K X ( κ l ) 2 U l + 2 ( j − 2 ) Q ( κ l ) > 0. (41)
To justify the contour passage, use the classical fixed-order asymptotic formula
J ν ( ζ ) = ( 2 π ζ ) 1 / 2 { cos ( ζ − ν π / 2 − π / 4 ) + O ( ∣ ζ ∣ − 1 e ∣ Im ζ ∣ ) } , ν = 0 , 1 , ∣ arg ζ ∣ ≤ π / 2. J_\nu(\zeta)=\left(\frac{2}{\pi\zeta}\right)^{1/2}\left\{\cos\left(\zeta-\nu\pi/2-\pi/4\right)+O\left(|\zeta|^{-1}e^{|\operatorname{Im}\zeta|}\right)\right\},\qquad\nu=0,1,\quad|\arg\zeta|\le\pi/2. J ν ( ζ ) = ( π ζ 2 ) 1/2 { cos ( ζ − ν π /2 − π /4 ) + O ( ∣ ζ ∣ − 1 e ∣ Im ζ ∣ ) } , ν = 0 , 1 , ∣ arg ζ ∣ ≤ π /2.
This fixed-order Hankel expansion and its complex remainder bound are given in [12 10.17.3, 10.17.5–10.17.6 and §10.17(iv)] . On ∣ ζ ∣ = ( N + 3 / 4 ) π |\zeta|=(N+3/4)\pi ∣ ζ ∣ = ( N + 3/4 ) π , its leading J 1 J_1 J 1 cosine has modulus bounded below by a fixed multiple of e ∣ Im ζ ∣ e^{|\operatorname{Im}\zeta|} e ∣ Im ζ ∣ . For ∣ Im ζ ∣ ≥ 1 |\operatorname{Im}\zeta|\ge1 ∣ Im ζ ∣ ≥ 1 use sinh ∣ Im ζ ∣ \sinh|\operatorname{Im}\zeta| sinh ∣ Im ζ ∣ ; otherwise the real part differs from the radius by O ( N − 1 ) O(N^{-1}) O ( N − 1 ) . Parity handles the left half-plane. Therefore ∣ J 1 ( ζ ) ∣ ≥ C N − 1 / 2 e ∣ Im ζ ∣ |J_1(\zeta)|\ge CN^{-1/2}e^{|\operatorname{Im}\zeta|} ∣ J 1 ( ζ ) ∣ ≥ C N − 1/2 e ∣ Im ζ ∣ , B ( ζ ) , B ( α ζ ) = O ( e 2 ∣ Im ζ ∣ ) \mathcal B(\zeta),\mathcal B(\alpha\zeta)=O(e^{2|\operatorname{Im}\zeta|}) B ( ζ ) , B ( α ζ ) = O ( e 2∣ Im ζ ∣ ) , and G ( ζ ) = O ( N 2 ) G(\zeta)=O(N^2) G ( ζ ) = O ( N 2 ) there. The contour integral of G ( ζ ) / ζ 2 j + 1 G(\zeta)/\zeta^{2j+1} G ( ζ ) / ζ 2 j + 1 consequently tends to zero for j ≥ 2 j\ge2 j ≥ 2 .
The nonzero zeros of J 1 J_1 J 1 are real and simple [12 Section 10.21(i)] . On the real zeros the same asymptotic gives X ( κ l ) 2 ≍ κ l − 1 X(\kappa_l)^2 \asymp\kappa_l^{-1} X ( κ l ) 2 ≍ κ l − 1 , L ( κ l ) = O ( 1 ) \mathcal{L}(\kappa_l)=O(1) L ( κ l ) = O ( 1 ) and ϕ ( κ l ) ϕ r ( κ l ) = O ( κ l − 1 ) \phi(\kappa_l)\phi_r(\kappa_l)=O(\kappa_l^{-1}) ϕ ( κ l ) ϕ r ( κ l ) = O ( κ l − 1 ) .
The paired series is absolutely convergent by zero spacing. At κ \kappa κ the principal coefficients of K − G K-G K − G are
A κ = κ L ( κ ) X ( κ ) 2 , D κ = 2 L ( κ ) + κ L ′ ( κ ) X ( κ ) 2 . A_\kappa=\frac{\kappa\mathcal{L}(\kappa)}{X(\kappa)^2}, \qquad D_\kappa=\frac{2\mathcal{L}(\kappa)+\kappa\mathcal{L}'(\kappa)}{X(\kappa)^2}. A κ = X ( κ ) 2 κ L ( κ ) , D κ = X ( κ ) 2 2 L ( κ ) + κ L ′ ( κ ) .
The paired residue of G ( ζ ) / ζ 2 j + 1 G(\zeta)/\zeta^{2j+1} G ( ζ ) / ζ 2 j + 1 is 2 { ( 2 j − 1 ) L − κ L ′ } / { κ 2 j + 1 X ( κ ) 2 } 2\{(2j-1)\mathcal{L}-\kappa\mathcal{L}'\}/\{\kappa^{2j+1}X(\kappa)^2\} 2 {( 2 j − 1 ) L − κ L ′ } / { κ 2 j + 1 X ( κ ) 2 } , which is exactly the summand in (A.5). This proves the coefficient identity.
Finally (A.1), (A.5) and analyticity before κ 2 \kappa_2 κ 2 give
G ( r ) = 2 λ − K − ∑ j ≥ 1 a j r 2 j , a j > 0 , ∣ r ∣ < κ 2 . (42) G(r)=2\lambda-K-\sum_{j\geq1}a_jr^{2j}, \qquad a_j>0,\qquad|r|<\kappa_2. \tag*{(42)} G ( r ) = 2 λ − K − j ≥ 1 ∑ a j r 2 j , a j > 0 , ∣ r ∣ < κ 2 . (42)
Since G ( κ ) = 0 G(\kappa)=0 G ( κ ) = 0 , this also proves G > 0 G>0 G > 0 on ( 0 , κ ) (0,\kappa) ( 0 , κ ) and every positive-order derivative of G G G with respect to r 2 r^2 r 2 is negative.
Data availability. All quantities used in the proofs are specified in this article.
Conflict of interest. The authors declare that they have no conflicts of interest.