*Corresponding author(s). E-mail(s): hhxie@lsec.cc.ac.cn;
Contributing authors: wangyifan@lsec.cc.ac.cn;
∑ j = 1 k λ j ≥ c k ( 2 ) ( Ω ) n n + 2 ( 2 π ) 2 ω n − 2 / n ∣ Ω ∣ − 2 / n k 1 + 2 / n , c k ( 2 ) ( Ω ) > 1 , \sum_{j=1}^{k}\lambda_j \ge c_k^{(2)}(\Omega)\frac{n}{n+2}(2\pi)^2\omega_n^{-2/n}|\Omega|^{-2/n}k^{1+2/n}, \qquad c_k^{(2)}(\Omega)>1, j = 1 ∑ k λ j ≥ c k ( 2 ) ( Ω ) n + 2 n ( 2 π ) 2 ω n − 2/ n ∣Ω ∣ − 2/ n k 1 + 2/ n , c k ( 2 ) ( Ω ) > 1 ,
where c k ( 2 ) ( Ω ) c_k^{(2)}(\Omega) c k ( 2 ) ( Ω ) is strictly decreasing in k k k and converges to one, and
c k ( 2 ) ( Ω ) − 1 ≥ n + 2 2 ( n + 1 ) 2 ( ω n − 1 ω n ) 2 ( 3 4 ) n − 1 k − 2 / n . c_k^{(2)}(\Omega)-1 \ge\frac{n+2}{2(n+1)^2}\left(\frac{\omega_{n-1}}{\omega_n}\right)^2\left(\frac{3}{4}\right)^{n-1}k^{-2/n}. c k ( 2 ) ( Ω ) − 1 ≥ 2 ( n + 1 ) 2 n + 2 ( ω n ω n − 1 ) 2 ( 4 3 ) n − 1 k − 2/ n .
Thus the sharp leading term in the Berezin–Li–Yau sum is supplemented by an explicit positive, volume-only correction of order k k k . Apart from boundedness in the concentration argument, no boundary regularity, convexity, or connectedness is assumed. These improvements do not prove Pólya’s conjecture: the individual coefficient is below one, and the sum multiplier tends to one.
keywords: Dirichlet eigenvalues; Berezin–Li–Yau inequality; Pólya conjecture; spectral deficit; commutator identities; coupled Fourier projections; frequency-concentration operators; 35P15; 35P20; 42B37; 47B10; 47B34
Introduction
Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n , n ≥ 2 n \ge2 n ≥ 2 , be an open set of finite positive measure V = ∣ Ω ∣ V=|\Omega| V = ∣Ω∣ , and let
0 < λ 1 ≤ λ 2 ≤ ⋯ 0 < \lambda_1 \le\lambda_2 \le\cdots 0 < λ 1 ≤ λ 2 ≤ ⋯
be the Dirichlet eigenvalues, repeated according to multiplicity. With ω n = ∣ B 1 ( 0 ) ∣ \omega_n=|B_1(0)| ω n = ∣ B 1 ( 0 ) ∣ , set
C n = ( 2 π ) 2 ω n − 2 / n . (1) C_n=(2\pi)^2\omega_n^{-2/n}. \tag*{(1)} C n = ( 2 π ) 2 ω n − 2/ n . (1)
Weyl’s law states that λ k ∼ C n V − 2 / n k 2 / n \lambda_k \sim C_nV^{-2/n}k^{2/n} λ k ∼ C n V − 2/ n k 2/ n as k → ∞ k \to\infty k → ∞ ; see, for example, [19 ] . Pólya’s conjecture [18 ] asks whether the corresponding termwise inequality
λ k ≥ C n V − 2 / n k 2 / n (2) \lambda_k \ge C_nV^{-2/n}k^{2/n} \tag*{(2)} λ k ≥ C n V − 2/ n k 2/ n (2)
holds for every k k k . This remains open for general domains. The classical Berezin–Li–Yau inequality [1 , 15 ] gives
∑ j = 1 k λ j ≥ n n + 2 C n V − 2 / n k 1 + 2 / n . (3) \sum_{j=1}^{k}\lambda_j \ge\frac{n}{n+2}C_nV^{-2/n}k^{1+2/n}. \tag*{(3)} j = 1 ∑ k λ j ≥ n + 2 n C n V − 2/ n k 1 + 2/ n . (3)
Since ∑ j = 1 k λ j ≤ k λ k \sum_{j=1}^{k}\lambda_j \le k\lambda_k ∑ j = 1 k λ j ≤ k λ k , it implies
λ k ≥ n n + 2 C n V − 2 / n k 2 / n . (4) \lambda_k \ge\frac{n}{n+2}C_nV^{-2/n}k^{2/n}. \tag*{(4)} λ k ≥ n + 2 n C n V − 2/ n k 2/ n . (4)
The leading constant in the sum inequality is asymptotically sharp, but the deduction of (4) discards the spectral deficit k λ k − ∑ j = 1 k λ j k\lambda_k-\sum_{j=1}^{k}\lambda_j k λ k − ∑ j = 1 k λ j . This distinction motivates the two results of the present paper.
The first result is a universal individual-eigenvalue estimate. Exact coordinate-commutator identities relate the spectral deficit to a projected-gradient density. At the same time, the scalar and vector Fourier projections of a plane wave satisfy a nonlinear pointwise compatibility condition. After normalization, these relations yield two densities f f f and g g g satisfying
t f ≤ f g + ( 1 − f ) ( f − g ) , ∫ R n ( t f + 2 n g ) d η ≤ ∫ R n f d η . \sqrt{t}f \le\sqrt{fg}+\sqrt{(1-f)(f-g)}, \qquad\int_{\mathbb{R}^n}\left(tf+\frac{2}{n}g\right)\,d\eta\le\int_{\mathbb{R}^n}f\,d\eta. t f ≤ f g + ( 1 − f ) ( f − g ) , ∫ R n ( t f + n 2 g ) d η ≤ ∫ R n f d η .
Eliminating g g g pointwise reduces the problem to a one-dimensional variational integral and gives
λ k ≥ C n n e w C n V − 2 / n k 2 / n , n n + 2 < c n e n v < C n n e w < 1. (5) \lambda_k \ge C_n^{\mathrm{new}}C_nV^{-2/n}k^{2/n}, \qquad\frac{n}{n+2}<c_n^{\mathrm{env}}<C_n^{\mathrm{new}}<1. \tag*{(5)} λ k ≥ C n new C n V − 2/ n k 2/ n , n + 2 n < c n env < C n new < 1. (5)
Here n / ( n + 2 ) n/(n+2) n / ( n + 2 ) is the direct Li–Yau coefficient, while c n e n v c_n^{\mathrm{env}} c n env is the coefficient generated by the uncoupled envelope F k ( ξ ) ≤ M Ω min { 1 , λ k / ∣ ξ ∣ 2 } F_k(\xi)\le M_\Omega\min\{1,\lambda_k/|\xi|^2\} F k ( ξ ) ≤ M Ω min { 1 , λ k /∣ ξ ∣ 2 } , M Ω = V / ( 2 π ) n M_\Omega=V/(2\pi)^n M Ω = V / ( 2 π ) n , studied in [26 ] . We recall its variational definition in (61) so that all constants are self-contained.
The two strict inequalities in (5) separate the gain from the scalar energy cap from the additional gain due to commutator coupling; see Proposition 4.2 and Table 1 for the proof and numerical comparison.
Table 1. Comparison of the classical Li–Yau individual coefficient, the uncoupled scalar-envelope coefficient, and numerical trial values of the commutator-coupled quotient.
The second result improves the eigenvalue sum at each finite index. For a bounded open set we introduce a positive trace-class frequency-concentration operator T r T_r T r on L 2 ( B r ) L^2(B_r) L 2 ( B r ) . At the Li–Yau filling radius R k R_k R k , one has Tr ( T R k ) = k \operatorname{Tr}(T_{R_k})=k Tr ( T R k ) = k , whereas analytic uniqueness implies Tr ( T R k 2 ) < k \operatorname{Tr}(T_{R_k}^2)<k Tr ( T R k 2 ) < k . The Ky Fan principle then forces some of the Fourier mass of the first k k k eigenfunctions to lie outside B R k B_{R_k} B R k . A two-region bathtub problem converts this leakage into
∑ j = 1 k λ j ≥ C k ( 2 ) ( Ω ) n n + 2 C n V − 2 / n k 1 + 2 / n , C k ( 2 ) ( Ω ) > 1. (6) \sum_{j=1}^{k}\lambda_j \ge C_k^{(2)}(\Omega)\frac{n}{n+2}C_nV^{-2/n}k^{1+2/n}, \qquad C_k^{(2)}(\Omega)>1. \tag*{(6)} j = 1 ∑ k λ j ≥ C k ( 2 ) ( Ω ) n + 2 n C n V − 2/ n k 1 + 2/ n , C k ( 2 ) ( Ω ) > 1. (6)
The multiplier has an explicit squared-trace representation, decreases strictly to one, and satisfies
C k ( 2 ) ( Ω ) − 1 ≥ κ n k − 2 / n , κ n = n + 2 2 ( n + 1 ) 2 ( ω n − 1 ω n ) 2 ( 3 4 ) n − 1 . (7) C_k^{(2)}(\Omega)-1\ge\kappa_n k^{-2/n}, \qquad\kappa_n=\frac{n+2}{2(n+1)^2}\left(\frac{\omega_{n-1}}{\omega_n}\right)^2\left(\frac{3}{4}\right)^{n-1}. \tag*{(7)} C k ( 2 ) ( Ω ) − 1 ≥ κ n k − 2/ n , κ n = 2 ( n + 1 ) 2 n + 2 ( ω n ω n − 1 ) 2 ( 4 3 ) n − 1 . (7)
Equivalently, (6) contains the additional volume-only term
n n + 2 κ n C n V − 2 / n k . \frac{n}{n+2}\kappa_n C_nV^{-2/n}k. n + 2 n κ n C n V − 2/ n k .
The decay of the relative correction is consistent with the sharp semiclassical leading coefficient.
These two statements recover different pieces of information. The commutator argument strengthens a Weyl-scale lower bound for a single eigenvalue, while the concentration argument strengthens the sharp sum bound without changing its leading coefficient. Neither result proves Pólya’s conjecture: C n n e w < 1 C_n^{\mathrm{new}} < 1 C n new < 1 , and C k ( 2 ) ( Ω ) → 1 C_k^{(2)}(\Omega) \to1 C k ( 2 ) ( Ω ) → 1 . The role of the present estimates is therefore to quantify two losses in the classical Fourier argument and to identify where additional information would have to enter.
We briefly compare these conclusions with the closest available results. For individual eigenvalues, the strict chain in (5) improves both the direct Li–Yau coefficient and the scalar-envelope coefficient from our previous work [26 ] . Its advantage is uniformity: it holds for every k k k on every finite-measure open set. By contrast, Jiang and Lin proved that for every ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) and every bounded Lipschitz domain,
λ k ≥ ( 1 + ε ) − 2 / n C n V − 2 / n k 2 / n whenever λ k ≥ Λ ε ( Ω ) , (8) \lambda_k \ge(1+\varepsilon)^{-2/n} C_n V^{-2/n} k^{2/n} \qquad\text{whenever } \lambda_k \ge\Lambda_\varepsilon(\Omega), \tag*{(8)} λ k ≥ ( 1 + ε ) − 2/ n C n V − 2/ n k 2/ n whenever λ k ≥ Λ ε ( Ω ) , (8)
where the explicit threshold depends on ε \varepsilon ε and the geometry of Ω \Omega Ω [12 ] . Thus (8) approaches the Pólya coefficient at high energy, whereas our fixed coefficient is valid without a threshold or boundary regularity. The full coefficient one is known for special geometries, including Euclidean balls and sufficiently thin product domains [2 , 9 ] ; other universal individual bounds of a different form appear in [24 ] .
For eigenvalue sums, write
L k L Y : = n n + 2 C n V − 2 / n k 1 + 2 / n . L_k^{\mathrm{LY}} := \frac{n}{n+2} C_n V^{-2/n} k^{1+2/n}. L k LY := n + 2 n C n V − 2/ n k 1 + 2/ n .
Melas proved the inertia-dependent correction
∑ j = 1 k λ j ≥ L k L Y + c n M V I ( Ω ) k , I ( Ω ) : = inf a ∈ R n ∫ Ω ∣ x − a ∣ 2 d x , (9) \sum_{j=1}^{k} \lambda_j \ge L_k^{\mathrm{LY}} + c_n^{\mathrm{M}} \frac{V}{I(\Omega)} k,\qquad I(\Omega) := \inf_{a\in\mathbb{R}^n} \int_{\Omega} |x-a|^2\,dx, \tag*{(9)} j = 1 ∑ k λ j ≥ L k LY + c n M I ( Ω ) V k , I ( Ω ) := a ∈ R n inf ∫ Ω ∣ x − a ∣ 2 d x , (9)
with subsequent refinements in [6 , 10 , 11 , 17 ] . Our quantitative estimate instead gives
∑ j = 1 k λ j ≥ L k L Y + n n + 2 κ n C n V − 2 / n k (10) \sum_{j=1}^{k} \lambda_j \ge L_k^{\mathrm{LY}} + \frac{n}{n+2}\kappa_n C_n V^{-2/n} k \tag*{(10)} j = 1 ∑ k λ j ≥ L k LY + n + 2 n κ n C n V − 2/ n k (10)
for every k k k on every bounded open set. Both remainders are linear in k k k , but (10) depends only on the volume and not on I ( Ω ) I(\Omega) I ( Ω ) .
Several stronger-looking corrections belong to different regimes. Volume-only strict improvements for order-one Riesz means under minimal assumptions were obtained in [5 ] , with an exponentially small relative gain at high energy. Boundary-sensitive estimates and two-term Riesz-mean asymptotics can reach the larger second Weyl scale under additional geometric, convexity, or Lipschitz hypotheses [3 , 4 , 7 , 13 , 27 ] . These results exploit information unavailable on an arbitrary bounded open set. The point of (6) and (7) is instead an explicit finite-k k k , volume-only improvement requiring no boundary smoothness, convexity, connectedness, or moment-of-inertia term. The paper is organized as follows. § 2 records the Fourier identities and the radial-envelope bathtub principle. § 3 establishes the commutator identity, the coupled projection inequality, and the normalized slack formula. § 4 defines C n e w C^{\mathrm{new}} C new , proves the individual bound, and compares the three variational coefficients. § 5 develops the concentration operator and its trace formulas, and § 6 proves the strict sum estimate and its explicit excess. The final section summarizes the scope and limitations of the two methods in this paper.
Preliminaries
We begin by recording the spectral and Fourier identities on which the argument rests. Let { ϕ j } j ≥ 1 \{\phi_j\}_{j\ge1} { ϕ j } j ≥ 1 be a real-valued orthonormal basis of Dirichlet eigenfunctions in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) . We use complex Hilbert spaces when pairing these functions with plane waves, with inner product linear in its first argument. Thus
− Δ ϕ j = λ j ϕ j , ϕ j ∈ H 0 1 ( Ω ) , ∫ Ω ϕ i ϕ j d x = δ i j . (11) -\Delta\phi_j=\lambda_j\phi_j,\qquad\phi_j\in H_0^1(\Omega),\qquad\int_{\Omega}\phi_i\phi_j\,\mathrm{d}x=\delta_{ij}. \tag*{(11)} − Δ ϕ j = λ j ϕ j , ϕ j ∈ H 0 1 ( Ω ) , ∫ Ω ϕ i ϕ j d x = δ ij . (11)
We use the unitary Fourier transform
u ^ ( ξ ) = ( 2 π ) − n 2 ∫ R n u ( x ) e − i x ⋅ ξ d x . (12) \widehat{u}(\xi)=(2\pi)^{-\frac{n}{2}}\int_{\mathbb{R}^n}u(x)e^{-\mathrm{i}x\cdot\xi}\,\mathrm{d}x. \tag*{(12)} u ( ξ ) = ( 2 π ) − 2 n ∫ R n u ( x ) e − i x ⋅ ξ d x . (12)
where every function in H 0 1 ( Ω ) H_0^1(\Omega) H 0 1 ( Ω ) is extended by zero to R n \mathbb{R}^n R n . The Fourier density associated with the first k k k eigenfunctions is
F k ( ξ ) = ∑ j = 1 k ∣ ϕ j ^ ( ξ ) ∣ 2 . (13) F_k(\xi)=\sum_{j=1}^{k}\lvert\widehat{\phi_j}(\xi)\rvert^2. \tag*{(13)} F k ( ξ ) = j = 1 ∑ k ∣ ϕ j ( ξ ) ∣ 2 . (13)
Plancherel’s theorem and the eigenvalue equation then give the two basic identities
∫ R n F k ( ξ ) d ξ = k , ∫ R n ∣ ξ ∣ 2 F k ( ξ ) d ξ = ∑ j = 1 k λ j . (14) \int_{\mathbb{R}^n}F_k(\xi)\,\mathrm{d}\xi=k,\qquad\int_{\mathbb{R}^n}\lvert\xi\rvert^2F_k(\xi)\,\mathrm{d}\xi=\sum_{j=1}^{k}\lambda_j. \tag*{(14)} ∫ R n F k ( ξ ) d ξ = k , ∫ R n ∣ ξ ∣ 2 F k ( ξ ) d ξ = j = 1 ∑ k λ j . (14)
Thus the spectral problem has been converted into a constrained moment problem: the total mass of F k F_k F k is known exactly, its second moment is the eigenvalue sum, and the remaining task is to obtain the sharpest available pointwise majorant.
Before deriving that majorant, we formulate the rearrangement principle that will turn it into a lower bound for the second moment. The following statement is a radial-obstacle version of the standard bathtub principle; see [26 ] and Lieb and Loss [16 ] .
Lemma 2.1 (Bathtub principle under a radial envelope) Let h : [ 0 , ∞ ) → [ 0 , ∞ ) h : [0,\infty) \to[0,\infty) h : [ 0 , ∞ ) → [ 0 , ∞ ) be measurable and nonincreasing, and suppose that h ( ∣ ⋅ ∣ ) ∈ L l o c 1 ( R n ) h(|\cdot|) \in L^1_{\mathrm{loc}}(\mathbb{R}^n) h ( ∣ ⋅ ∣ ) ∈ L loc 1 ( R n ) . Let 0 ≤ M < ∞ 0 \le M < \infty 0 ≤ M < ∞ and assume that
∫ R n h ( ∣ ξ ∣ ) d ξ ≥ M . \int_{\mathbb{R}^n} h(|\xi|)\,\mathrm{d}\xi\ge M. ∫ R n h ( ∣ ξ ∣ ) d ξ ≥ M . If a measurable function f : R n → [ 0 , ∞ ) f : \mathbb{R}^n \to[0,\infty) f : R n → [ 0 , ∞ ) satisfies
0 ≤ f ( ξ ) ≤ h ( ∣ ξ ∣ ) a.e. , ∫ R n f ( ξ ) d ξ = M , 0 \le f(\xi) \le h(|\xi|) \quad\text{a.e.}, \qquad\int_{\mathbb{R}^n} f(\xi)\,\mathrm{d}\xi= M, 0 ≤ f ( ξ ) ≤ h ( ∣ ξ ∣ ) a.e. , ∫ R n f ( ξ ) d ξ = M , then there exists R ∈ [ 0 , ∞ ] R \in[0,\infty] R ∈ [ 0 , ∞ ] such that
∫ B R h ( ∣ ξ ∣ ) d ξ = M . \int_{B_R} h(|\xi|)\,\mathrm{d}\xi= M. ∫ B R h ( ∣ ξ ∣ ) d ξ = M . Consequently, if
g R ( ξ ) : = h ( ∣ ξ ∣ ) 1 B R ( ξ ) , g_R(\xi) := h(|\xi|)\mathbf{1}_{B_R}(\xi), g R ( ξ ) := h ( ∣ ξ ∣ ) 1 B R ( ξ ) , then
∫ R n ∣ ξ ∣ 2 f ( ξ ) d ξ ≥ ∫ B R ∣ ξ ∣ 2 h ( ∣ ξ ∣ ) d ξ . (15) \int_{\mathbb{R}^n} |\xi|^2 f(\xi)\,\mathrm{d}\xi\ge\int_{B_R} |\xi|^2 h(|\xi|)\,\mathrm{d}\xi. \tag*{(15)} ∫ R n ∣ ξ ∣ 2 f ( ξ ) d ξ ≥ ∫ B R ∣ ξ ∣ 2 h ( ∣ ξ ∣ ) d ξ . (15) The integrals in (15) may take the value + ∞ +\infty + ∞ .
Proof. Set
H ( R ) = ∫ B R h ( ∣ ξ ∣ ) d ξ . H(R) = \int_{B_R} h(|\xi|)\,\mathrm{d}\xi. H ( R ) = ∫ B R h ( ∣ ξ ∣ ) d ξ .
Local integrability implies that H H H is continuous and nondecreasing, with H ( 0 ) = 0 H(0) = 0 H ( 0 ) = 0 and H ( R ) ↑ ∫ R n h ( ∣ ξ ∣ ) d ξ H(R) \mathrel{\uparrow} \int_{\mathbb{R}^n} h(|\xi|)\,\mathrm{d}\xi H ( R ) ↑ ∫ R n h ( ∣ ξ ∣ ) d ξ as R → ∞ R \to\infty R → ∞ . Hence a radius R ∈ [ 0 , ∞ ] R \in[0,\infty] R ∈ [ 0 , ∞ ] with H ( R ) = M H(R) = M H ( R ) = M exists. If 0 < R < ∞ 0 < R < \infty 0 < R < ∞ , then g R = h ( ∣ ⋅ ∣ ) 1 B R g_R = h(|\cdot|)\mathbf{1}_{B_R} g R = h ( ∣ ⋅ ∣ ) 1 B R has the same mass as f f f , and
( ∣ ξ ∣ 2 − R 2 ) ( f ( ξ ) − g R ( ξ ) ) ≥ 0 for a.e. ξ ∈ R n . (|\xi|^2 - R^2)(f(\xi) - g_R(\xi)) \ge0 \quad\text{for a.e. } \xi\in\mathbb{R}^n. ( ∣ ξ ∣ 2 − R 2 ) ( f ( ξ ) − g R ( ξ )) ≥ 0 for a.e. ξ ∈ R n .
Indeed, both factors are nonpositive on B R B_R B R , whereas both are nonnegative on its complement. Integration and cancellation of the equal-mass terms multiplied by R 2 R^2 R 2 give (15) . If R = 0 R = 0 R = 0 , then M = 0 M = 0 M = 0 and the assertion is immediate. If R = ∞ R = \infty R = ∞ , then M = ∫ h ( ∣ ξ ∣ ) d ξ M = \int h(|\xi|)\,\mathrm{d}\xi M = ∫ h ( ∣ ξ ∣ ) d ξ ; the inequalities 0 ≤ f ≤ h ( ∣ ⋅ ∣ ) 0 \le f \le h(|\cdot|) 0 ≤ f ≤ h ( ∣ ⋅ ∣ ) and equality of their integrals imply f = h ( ∣ ⋅ ∣ ) f = h(|\cdot|) f = h ( ∣ ⋅ ∣ ) almost everywhere, so equality holds in (15) . □
The lemma isolates the classical variational part of the argument. The usual L 2 L^2 L 2 and energy estimates supply a scalar radial cap for F k F_k F k , whereas the next section retains more information by coupling that density to the low-mode gradient projection and measuring the integrated slack through a commutator identity.
Commutator Identities and Coupled Fourier Projections
The preceding radial bathtub principle becomes substantially stronger when the scalar Fourier density is not treated in isolation. We now retain its interaction with the low-mode part of the gradient. For the commutator calculation we first assume that Ω \Omega Ω is bounded. The final eigenvalue bound is extended to arbitrary finite-measure open sets by exhaustion.
Coordinate commutators and an exact deficit identity
For 1 ≤ α ≤ n 1 \le\alpha\le n 1 ≤ α ≤ n and i , j ≥ 1 i,j \ge1 i , j ≥ 1 , set
a i j ( α ) = ∫ Ω x α ϕ i ϕ j d x , d i j ( α ) = ∫ Ω ϕ i ∂ α ϕ j d x . (16) a_{ij}^{(\alpha)}=\int_{\Omega}x_\alpha\phi_i\phi_j\,\mathrm{d}x,\qquad d_{ij}^{(\alpha)}=\int_{\Omega}\phi_i\partial_\alpha\phi_j\,\mathrm{d}x. \tag*{(16)} a ij ( α ) = ∫ Ω x α ϕ i ϕ j d x , d ij ( α ) = ∫ Ω ϕ i ∂ α ϕ j d x . (16)
The first identity below is the eigenfunction matrix-element form of the coordinate commutator; see [14 Eq. (2.6) and Example 4.1] . The second is the coordinate Thomas–Reiche–Kuhn sum rule [14 Eq. (1.2)] . We give a direct proof using the weak eigenvalue equation.
Lemma 3.1 (Coordinate commutator identities) For every i , j ≥ 1 i,j \ge1 i , j ≥ 1 and 1 ≤ α ≤ n 1 \le\alpha\le n 1 ≤ α ≤ n ,
( λ i − λ j ) a i j ( α ) = − 2 d i j ( α ) , (17) (\lambda_i-\lambda_j)a_{ij}^{(\alpha)}=-2d_{ij}^{(\alpha)}, \tag*{(17)} ( λ i − λ j ) a ij ( α ) = − 2 d ij ( α ) , (17) ∑ i = 1 ∞ ( λ i − λ j ) ∣ a i j ( α ) ∣ 2 = 1. (18) \sum_{i=1}^{\infty}(\lambda_i-\lambda_j)\lvert a_{ij}^{(\alpha)}\rvert^2=1. \tag*{(18)} i = 1 ∑ ∞ ( λ i − λ j ) ∣ a ij ( α ) ∣ 2 = 1. (18) The series in (18) is absolutely convergent.
Proof. Because Ω \Omega Ω is bounded, x α ∈ W 1 , ∞ ( Ω ) x_\alpha\in W^{1,\infty}(\Omega) x α ∈ W 1 , ∞ ( Ω ) , and multiplication by either x α x_\alpha x α or x α 2 x_\alpha^2 x α 2 preserves H 0 1 ( Ω ) H_0^1(\Omega) H 0 1 ( Ω ) . The weak eigenvalue equation for ϕ i \phi_i ϕ i , tested with x α ϕ j x_\alpha\phi_j x α ϕ j , and the equation for ϕ j \phi_j ϕ j , tested with x α ϕ i x_\alpha\phi_i x α ϕ i , give
λ i a i j ( α ) = ∫ Ω x α ∇ ϕ i ⋅ ∇ ϕ j d x + ∫ Ω ϕ j ∂ α ϕ i d x , (19) \lambda_i a_{ij}^{(\alpha)}=\int_{\Omega}x_\alpha\nabla\phi_i\cdot\nabla\phi_j\,\mathrm{d}x+\int_{\Omega}\phi_j\partial_\alpha\phi_i\,\mathrm{d}x, \tag*{(19)} λ i a ij ( α ) = ∫ Ω x α ∇ ϕ i ⋅ ∇ ϕ j d x + ∫ Ω ϕ j ∂ α ϕ i d x , (19)
λ j a i j ( α ) = ∫ Ω x α ∇ ϕ i ⋅ ∇ ϕ j d x + ∫ Ω ϕ i ∂ α ϕ j d x . (20) \lambda_j a_{ij}^{(\alpha)}=\int_{\Omega}x_\alpha\nabla\phi_i\cdot\nabla\phi_j\,\mathrm{d}x+\int_{\Omega}\phi_i\partial_\alpha\phi_j\,\mathrm{d}x. \tag*{(20)} λ j a ij ( α ) = ∫ Ω x α ∇ ϕ i ⋅ ∇ ϕ j d x + ∫ Ω ϕ i ∂ α ϕ j d x . (20)
Here ϕ i ϕ j ∈ W 0 1 , 1 ( Ω ) \phi_i\phi_j \in W_0^{1,1}(\Omega) ϕ i ϕ j ∈ W 0 1 , 1 ( Ω ) , so weak integration by parts implies
∫ Ω ϕ j ∂ α ϕ i d x = − ∫ Ω ϕ i ∂ α ϕ j d x = − d i j ( α ) . (21) \int_{\Omega}\phi_j\partial_\alpha\phi_i\,\mathrm{d}x=-\int_{\Omega}\phi_i\partial_\alpha\phi_j\,\mathrm{d}x=-d_{ij}^{(\alpha)}. \tag*{(21)} ∫ Ω ϕ j ∂ α ϕ i d x = − ∫ Ω ϕ i ∂ α ϕ j d x = − d ij ( α ) . (21)
With (21) and subtracting the preceding two identities (19) and (20) therefore proves (17) . Notice also that a j i ( α ) = a i j ( α ) a_{ji}^{(\alpha)}=a_{ij}^{(\alpha)} a j i ( α ) = a ij ( α ) and d j i ( α ) = − d i j ( α ) d_{ji}^{(\alpha)}=-d_{ij}^{(\alpha)} d j i ( α ) = − d ij ( α ) ; in particular, d i i ( α ) = 0 d_{ii}^{(\alpha)}=0 d ii ( α ) = 0 .
For the sum rule, put u = x α ϕ j ∈ H 0 1 ( Ω ) u=x_\alpha\phi_j\in H_0^1(\Omega) u = x α ϕ j ∈ H 0 1 ( Ω ) . The spectral representation of the Dirichlet form x α ϕ j = ∑ i = 1 ∞ a i j ( α ) ϕ i x_\alpha\phi_j=\sum_{i=1}^{\infty}a_{ij}^{(\alpha)}\phi_i x α ϕ j = ∑ i = 1 ∞ a ij ( α ) ϕ i and then Parseval’s identity yield
∑ i = 1 ∞ λ i ∣ a i j ( α ) ∣ 2 = ∫ Ω ∣ ∇ u ∣ 2 d x , (22) \sum_{i=1}^{\infty}\lambda_i\lvert a_{ij}^{(\alpha)}\rvert^2=\int_{\Omega}\lvert\nabla u\rvert^2\,\mathrm{d}x, \tag*{(22)} i = 1 ∑ ∞ λ i ∣ a ij ( α ) ∣ 2 = ∫ Ω ∣ ∇ u ∣ 2 d x , (22)
∑ i = 1 ∞ ∣ a i j ( α ) ∣ 2 = ∫ Ω x α 2 ϕ j 2 d x . (23) \sum_{i=1}^{\infty}\lvert a_{ij}^{(\alpha)}\rvert^2=\int_{\Omega}x_\alpha^2\phi_j^2\,\mathrm{d}x. \tag*{(23)} i = 1 ∑ ∞ ∣ a ij ( α ) ∣ 2 = ∫ Ω x α 2 ϕ j 2 d x . (23)
Expanding the first integral makes the constant in the sum rule explicit:
∫ Ω ∣ ∇ ( x α ϕ j ) ∣ 2 d x = ∫ Ω x α 2 ∣ ∇ ϕ j ∣ 2 d x + 2 ∫ Ω x α ϕ j ∂ α ϕ j d x + ∫ Ω ϕ j 2 d x . (24) \int_{\Omega}\lvert\nabla(x_\alpha\phi_j)\rvert^2\,\mathrm{d}x=\int_{\Omega}x_\alpha^2\lvert\nabla\phi_j\rvert^2\,\mathrm{d}x+2\int_{\Omega}x_\alpha\phi_j\partial_\alpha\phi_j\,\mathrm{d}x+\int_{\Omega}\phi_j^2\,\mathrm{d}x. \tag*{(24)} ∫ Ω ∣ ∇ ( x α ϕ j ) ∣ 2 d x = ∫ Ω x α 2 ∣ ∇ ϕ j ∣ 2 d x + 2 ∫ Ω x α ϕ j ∂ α ϕ j d x + ∫ Ω ϕ j 2 d x . (24)
On the other hand, testing the j j j -th eigenvalue equation with x α 2 ϕ j x_{\alpha}^{2}\phi_j x α 2 ϕ j gives
λ j ∫ Ω x α 2 ϕ j 2 d x = ∫ Ω x α 2 ∣ ∇ ϕ j ∣ 2 d x + 2 ∫ Ω x α ϕ j ∂ α ϕ j d x . (25) \lambda_j \int_{\Omega} x_{\alpha}^{2}\phi_j^{2}\,dx = \int_{\Omega} x_{\alpha}^{2}|\nabla\phi_j|^{2}\,dx + 2\int_{\Omega}x_{\alpha}\phi_j\partial_{\alpha}\phi_j\,dx. \tag*{(25)} λ j ∫ Ω x α 2 ϕ j 2 d x = ∫ Ω x α 2 ∣∇ ϕ j ∣ 2 d x + 2 ∫ Ω x α ϕ j ∂ α ϕ j d x . (25)
After subtraction, The combination of (22) , (23) , (24) , (25) and the L 2 L^2 L 2 normalization ∫ Ω ϕ j 2 d x = 1 \int_{\Omega}\phi_j^2\,dx=1 ∫ Ω ϕ j 2 d x = 1 proves (18) . Finally,
∑ i = 1 ∞ ∣ λ i − λ j ∣ ∣ a i j ( α ) ∣ 2 ≤ ∑ i = 1 ∞ ( λ i + λ j ) ∣ a i j ( α ) ∣ 2 < ∞ , \sum_{i=1}^{\infty}|\lambda_i-\lambda_j||a_{ij}^{(\alpha)}|^2\leq\sum_{i=1}^{\infty}(\lambda_i+\lambda_j)|a_{ij}^{(\alpha)}|^2<\infty, i = 1 ∑ ∞ ∣ λ i − λ j ∣∣ a ij ( α ) ∣ 2 ≤ i = 1 ∑ ∞ ( λ i + λ j ) ∣ a ij ( α ) ∣ 2 < ∞ ,
by the two spectral identities above. Thus the series in (18) is absolutely convergent. □
Fix k ≥ 1 k\geq1 k ≥ 1 , write
Λ = λ k , S k = ∑ j = 1 k λ j , D k = k Λ − S k , (26) \Lambda=\lambda_k,\qquad S_k=\sum_{j=1}^{k}\lambda_j,\qquad D_k=k\Lambda-S_k, \tag*{(26)} Λ = λ k , S k = j = 1 ∑ k λ j , D k = k Λ − S k , (26)
and introduce
I k = ∑ α = 1 n ∑ i = 1 k ∑ j = 1 k ∣ d i j ( α ) ∣ 2 , (27) I_k=\sum_{\alpha=1}^{n}\sum_{i=1}^{k}\sum_{j=1}^{k}|d_{ij}^{(\alpha)}|^2, \tag*{(27)} I k = α = 1 ∑ n i = 1 ∑ k j = 1 ∑ k ∣ d ij ( α ) ∣ 2 , (27)
B k = ∑ α = 1 n ∑ i = 1 k ∑ j > k ( Λ − λ j ) ( λ i − λ j ) ∣ a i j ( α ) ∣ 2 . (28) B_k=\sum_{\alpha=1}^{n}\sum_{i=1}^{k}\sum_{j>k}(\Lambda-\lambda_j)(\lambda_i-\lambda_j)|a_{ij}^{(\alpha)}|^2. \tag*{(28)} B k = α = 1 ∑ n i = 1 ∑ k j > k ∑ ( Λ − λ j ) ( λ i − λ j ) ∣ a ij ( α ) ∣ 2 . (28)
Every term in B k B_k B k is nonnegative, so B k ≥ 0 B_k\geq0 B k ≥ 0 .
The following identity is the Dirichlet Laplacian specialization of the trace formula of Harrell and Stubbe [8 ] , Theorem 1, Eq. (4). We give a direct proof in our notation, which also covers the case where the cutoff at k k k splits a multiple eigenvalue.
Proposition 3.2 (Exact commutator slack) The quantities in (26) to (28) satisfy
n D k = 2 I k + B k . (29) nD_k=2I_k+B_k. \tag*{(29)} n D k = 2 I k + B k . (29) In particular,
I k ≤ n 2 D k . (30) I_k\leq\frac{n}{2}D_k. \tag*{(30)} I k ≤ 2 n D k . (30)
Proof. Multiply (18) by Λ − λ j \Lambda-\lambda_j Λ − λ j and sum over 1 ≤ j ≤ k 1\leq j\leq k 1 ≤ j ≤ k and 1 ≤ α ≤ n 1\leq\alpha\leq n 1 ≤ α ≤ n . The absolute convergence established in Lemma 3.1 justifies splitting each i i i -sum at k k k , and hence
n D k = ∑ α = 1 n ∑ j = 1 k ( Λ − λ j ) ∑ i = 1 ∞ ( λ i − λ j ) ∣ a i j ( α ) ∣ 2 = ∑ α = 1 n ∑ i = 1 k ∑ j = 1 k ( Λ − λ j ) ( λ i − λ j ) ∣ a i j ( α ) ∣ 2 + B k . (31) \begin{aligned}
nD_k=\sum_{\alpha=1}^{n}\sum_{j=1}^{k}(\Lambda-\lambda_j)\sum_{i=1}^{\infty}(\lambda_i-\lambda_j)|a_{ij}^{(\alpha)}|^2 \\
&=\sum_{\alpha=1}^{n}\sum_{i=1}^{k}\sum_{j=1}^{k}(\Lambda-\lambda_j)(\lambda_i-\lambda_j)|a_{ij}^{(\alpha)}|^2+B_k. \tag*{(31)}
\end{aligned} n D k = α = 1 ∑ n j = 1 ∑ k ( Λ − λ j ) i = 1 ∑ ∞ ( λ i − λ j ) ∣ a ij ( α ) ∣ 2 = α = 1 ∑ n i = 1 ∑ k j = 1 ∑ k ( Λ − λ j ) ( λ i − λ j ) ∣ a ij ( α ) ∣ 2 + B k . (31)
It remains to identify the finite double sum. Since a i j ( α ) = a j i ( α ) a_{ij}^{(\alpha)}=a_{ji}^{(\alpha)} a ij ( α ) = a j i ( α ) , symmetrization in i i i and j j j gives, for each α \alpha α ,
∑ i = 1 k ∑ j = 1 k ( Λ − λ j ) ( λ i − λ j ) ∣ a i j ( α ) ∣ 2 = 1 2 ∑ i = 1 k ∑ j = 1 k [ ( Λ − λ j ) ( λ i − λ j ) + ( Λ − λ i ) ( λ j − λ i ) ] ∣ a i j ( α ) ∣ 2 = 1 2 ∑ i = 1 k ∑ j = 1 k ( λ i − λ j ) 2 ∣ a i j ( α ) ∣ 2 = 2 ∑ i = 1 k ∑ j = 1 k ∣ d i j ( α ) ∣ 2 , \begin{aligned}
\sum_{i=1}^{k}\sum_{j=1}^{k}(\Lambda-\lambda_j)(\lambda_i-\lambda_j)|a_{ij}^{(\alpha)}|^2 \\
&=\frac{1}{2}\sum_{i=1}^{k}\sum_{j=1}^{k}\left[(\Lambda-\lambda_j)(\lambda_i-\lambda_j)+(\Lambda-\lambda_i)(\lambda_j-\lambda_i)\right]|a_{ij}^{(\alpha)}|^2 \\
&=\frac{1}{2}\sum_{i=1}^{k}\sum_{j=1}^{k}(\lambda_i-\lambda_j)^2|a_{ij}^{(\alpha)}|^2=2\sum_{i=1}^{k}\sum_{j=1}^{k}|d_{ij}^{(\alpha)}|^2,
\end{aligned} i = 1 ∑ k j = 1 ∑ k ( Λ − λ j ) ( λ i − λ j ) ∣ a ij ( α ) ∣ 2 = 2 1 i = 1 ∑ k j = 1 ∑ k [ ( Λ − λ j ) ( λ i − λ j ) + ( Λ − λ i ) ( λ j − λ i ) ] ∣ a ij ( α ) ∣ 2 = 2 1 i = 1 ∑ k j = 1 ∑ k ( λ i − λ j ) 2 ∣ a ij ( α ) ∣ 2 = 2 i = 1 ∑ k j = 1 ∑ k ∣ d ij ( α ) ∣ 2 ,
where the last equality follows from (17) . Summing over α \alpha α proves (29) . Since B k ≥ 0 B_k\geq0 B k ≥ 0 , (3.15) follows immediately. In Proposition 3.7`, the same identity will reappear as the exact slack in the normalized integral constraint. □
The coupled scalar–vector projection inequality
Set
M Ω = V ( 2 π ) n (32) M_{\Omega}=\frac{V}{(2\pi)^n} \tag*{(32)} M Ω = ( 2 π ) n V (32)
and let P k P_k P k be the orthogonal projection in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) onto E k = span { ϕ 1 , … , ϕ k } E_k=\operatorname{span}\{\phi_1,\ldots,\phi_k\} E k = span { ϕ 1 , … , ϕ k } . Its componentwise extension to L 2 ( Ω ; C n ) L^2(\Omega;\mathbb{C}^n) L 2 ( Ω ; C n ) is denoted by P k P_k P k . For ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n , define
E ξ ( x ) = ( 2 π ) − n / 2 e − i x ⋅ ξ , q ξ = P k E ξ , E_{\xi}(x)=(2\pi)^{-n/2}e^{-ix\cdot\xi},\qquad q_{\xi}=P_kE_{\xi}, E ξ ( x ) = ( 2 π ) − n /2 e − i x ⋅ ξ , q ξ = P k E ξ ,
and retain the notation F k ( ξ ) = ∥ q ξ ∥ 2 2 F_k(\xi)=\|q_{\xi}\|_2^2 F k ( ξ ) = ∥ q ξ ∥ 2 2 . We also set
T k ( ξ ) = ∥ ∇ q ξ ∥ 2 2 , G k ( ξ ) = ∥ P k ∇ q ξ ∥ 2 2 . (33) T_k(\xi)=\|\nabla q_{\xi}\|_2^2,\qquad G_k(\xi)=\|P_k\nabla q_{\xi}\|_2^2. \tag*{(33)} T k ( ξ ) = ∥∇ q ξ ∥ 2 2 , G k ( ξ ) = ∥ P k ∇ q ξ ∥ 2 2 . (33)
If c ( ξ ) = ( c j ( ξ ) ) 1 ≤ j ≤ k = ( ϕ j ^ ( ξ ) ) 1 ≤ j ≤ k c(\xi)=(c_j(\xi))_{1\leq j\leq k}=(\widehat{\phi_j}(\xi))_{1\leq j\leq k} c ( ξ ) = ( c j ( ξ ) ) 1 ≤ j ≤ k = ( ϕ j ( ξ ) ) 1 ≤ j ≤ k and D ( α ) = ( d i j ( α ) ) 1 ≤ i , j ≤ k D^{(\alpha)}=(d_{ij}^{(\alpha)})_{1\leq i,j\leq k} D ( α ) = ( d ij ( α ) ) 1 ≤ i , j ≤ k , then
G k ( ξ ) = ∑ α = 1 n ∥ D ( α ) c ( ξ ) ∥ c k 2 . (34) G_k(\xi)=\sum_{\alpha=1}^{n}\left\|D^{(\alpha)}c(\xi)\right\|_{c_k}^{2}. \tag*{(34)} G k ( ξ ) = α = 1 ∑ n D ( α ) c ( ξ ) c k 2 . (34)
Consequently, Plancherel’s identity gives the exact integral relation
∫ R n G k ( ξ ) d ξ = I k . (35) \int_{\mathbb{R}^{n}}G_k(\xi)\,d\xi=I_k. \tag*{(35)} ∫ R n G k ( ξ ) d ξ = I k . (35)
Moreover,
0 ≤ G k ( ξ ) ≤ T k ( ξ ) = ∑ j = 1 k λ j ∣ ϕ j ^ ( ξ ) ∣ 2 ≤ Λ F k ( ξ ) . (36) 0\leq G_k(\xi)\leq T_k(\xi)=\sum_{j=1}^{k}\lambda_j|\widehat{\phi_j}(\xi)|^2\leq\Lambda F_k(\xi). \tag*{(36)} 0 ≤ G k ( ξ ) ≤ T k ( ξ ) = j = 1 ∑ k λ j ∣ ϕ j ( ξ ) ∣ 2 ≤ Λ F k ( ξ ) . (36)
Proposition 3.3 (Coupled Fourier projection) For every ξ ≠ 0 \xi\ne0 ξ = 0 ,
∣ ξ ∣ F k ( ξ ) ≤ F k ( ξ ) G k ( ξ ) + ( M Ω − F k ( ξ ) ) ( T k ( ξ ) − G k ( ξ ) ) . (37) |\xi|F_k(\xi)\leq\sqrt{F_k(\xi)}G_k(\xi)+\sqrt{(M_{\Omega}-F_k(\xi))(T_k(\xi)-G_k(\xi))}. \tag*{(37)} ∣ ξ ∣ F k ( ξ ) ≤ F k ( ξ ) G k ( ξ ) + ( M Ω − F k ( ξ )) ( T k ( ξ ) − G k ( ξ )) . (37)
In particular,
∣ ξ ∣ F k ( ξ ) ≤ F k ( ξ ) G k ( ξ ) + ( M Ω − F k ( ξ ) ) ( Λ F k ( ξ ) − G k ( ξ ) ) . (38) |\xi|F_k(\xi) \le\sqrt{F_k(\xi)G_k(\xi)}+\sqrt{\left(M_{\Omega}-F_k(\xi)\right)\left(\Lambda F_k(\xi)-G_k(\xi)\right)}. \tag*{(38)} ∣ ξ ∣ F k ( ξ ) ≤ F k ( ξ ) G k ( ξ ) + ( M Ω − F k ( ξ ) ) ( Λ F k ( ξ ) − G k ( ξ ) ) . (38)
Proof. Fix ξ ≠ 0 \xi\ne0 ξ = 0 and set
Y ξ = − i ξ ∣ ξ ∣ E ξ ∈ L 2 ( Ω ; C n ) . Y_{\xi}=-i\frac{\xi}{|\xi|}E_{\xi}\in L^2(\Omega;\mathbb{C}^n). Y ξ = − i ∣ ξ ∣ ξ E ξ ∈ L 2 ( Ω ; C n ) .
Since ∣ E ξ ∣ = ( 2 π ) − n / 2 |E_{\xi}|=(2\pi)^{-n/2} ∣ E ξ ∣ = ( 2 π ) − n /2 , the definition (3.16) gives
∥ Y ξ ∥ 2 2 = M Ω . (39) \|Y_{\xi}\|_2^2=M_{\Omega}. \tag*{(39)} ∥ Y ξ ∥ 2 2 = M Ω . (39)
The componentwise projection acts only on the scalar factor of Y ξ Y_{\xi} Y ξ , and therefore
P k Y ξ = − i ξ ∣ ξ ∣ P k E ξ = − i ξ ∣ ξ ∣ q ξ , ∥ P k Y ξ ∥ 2 2 = ∥ q ξ ∥ 2 2 = F k ( ξ ) . (40) P_kY_{\xi}=-i\frac{\xi}{|\xi|}P_kE_{\xi}=-i\frac{\xi}{|\xi|}q_{\xi},\qquad\|P_kY_{\xi}\|_2^2=\|q_{\xi}\|_2^2=F_k(\xi). \tag*{(40)} P k Y ξ = − i ∣ ξ ∣ ξ P k E ξ = − i ∣ ξ ∣ ξ q ξ , ∥ P k Y ξ ∥ 2 2 = ∥ q ξ ∥ 2 2 = F k ( ξ ) . (40)
We next compute the relevant pairing. Because q ξ ∈ E k ⊂ H 0 1 ( Ω ) q_{\xi}\in E_k\subset H^1_0(\Omega) q ξ ∈ E k ⊂ H 0 1 ( Ω ) , weak integration by parts is legitimate. Using div Y ξ = − ∣ ξ ∣ E ξ \operatorname{div}Y_{\xi}=-|\xi|E_{\xi} div Y ξ = − ∣ ξ ∣ E ξ , we obtain
⟨ ∇ q ξ , Y ξ ⟩ = − ∫ Ω q ξ div Y ξ d x = ∣ ξ ∣ ⟨ q ξ , E ξ ⟩ = ∣ ξ ∣ ⟨ P k E ξ , E ξ ⟩ = ∣ ξ ∣ ∥ P k E ξ ∥ 2 2 = ∣ ξ ∣ F k ( ξ ) . (41) \begin{aligned}
\langle\nabla q_{\xi},Y_{\xi}\rangle
&=-\int_{\Omega}q_{\xi}\operatorname{div}Y_{\xi}\,dx\\
&=|\xi|\langle q_{\xi},E_{\xi}\rangle=|\xi|\langle P_kE_{\xi},E_{\xi}\rangle\\
&=|\xi|\|P_kE_{\xi}\|_2^2=|\xi|F_k(\xi). \tag*{(41)}
\end{aligned} ⟨ ∇ q ξ , Y ξ ⟩ = − ∫ Ω q ξ div Y ξ d x = ∣ ξ ∣ ⟨ q ξ , E ξ ⟩ = ∣ ξ ∣ ⟨ P k E ξ , E ξ ⟩ = ∣ ξ ∣∥ P k E ξ ∥ 2 2 = ∣ ξ ∣ F k ( ξ ) . (41)
Here the penultimate equality uses that P k P_k P k is an orthogonal projection.
Decompose both vector fields with respect to the orthogonal splitting
L 2 ( Ω ; C n ) = Ran P k ⊕ Ker P k . L^2(\Omega;\mathbb{C}^n)=\operatorname{Ran}P_k\oplus\operatorname{Ker}P_k. L 2 ( Ω ; C n ) = Ran P k ⊕ Ker P k .
The mixed pairings vanish, so
⟨ ∇ q ξ , Y ξ ⟩ = ⟨ P k ∇ q ξ , P k Y ξ ⟩ + ⟨ ( I − P k ) ∇ q ξ , ( I − P k ) Y ξ ⟩ . (42) \langle\nabla q_{\xi},Y_{\xi}\rangle=\langle P_k\nabla q_{\xi},P_kY_{\xi}\rangle+\langle(I-P_k)\nabla q_{\xi},(I-P_k)Y_{\xi}\rangle. \tag*{(42)} ⟨ ∇ q ξ , Y ξ ⟩ = ⟨ P k ∇ q ξ , P k Y ξ ⟩ + ⟨( I − P k ) ∇ q ξ , ( I − P k ) Y ξ ⟩ . (42)
By (33) , (39) and (40) and orthogonality,
∥ P k ∇ q ξ ∥ 2 2 = G k ( ξ ) , ∥ ( I − P k ) ∇ q ξ ∥ 2 2 = T k ( ξ ) − G k ( ξ ) , \|P_k\nabla q_{\xi}\|_2^2=G_k(\xi),\qquad\|(I-P_k)\nabla q_{\xi}\|_2^2=T_k(\xi)-G_k(\xi), ∥ P k ∇ q ξ ∥ 2 2 = G k ( ξ ) , ∥ ( I − P k ) ∇ q ξ ∥ 2 2 = T k ( ξ ) − G k ( ξ ) ,
∥ P k Y ξ ∥ 2 2 = F k ( ξ ) , ∥ ( I − P k ) Y ξ ∥ 2 2 = M Ω − F k ( ξ ) . \|P_kY_{\xi}\|_2^2=F_k(\xi),\qquad\|(I-P_k)Y_{\xi}\|_2^2=M_{\Omega}-F_k(\xi). ∥ P k Y ξ ∥ 2 2 = F k ( ξ ) , ∥ ( I − P k ) Y ξ ∥ 2 2 = M Ω − F k ( ξ ) .
Applying these identities, the triangle inequality and Cauchy–Schwarz to the two orthogonal components in (42) , and then using (41) , proves (37) . Finally, (36) gives T k ( ξ ) − G k ( ξ ) ≤ Λ F k ( ξ ) − G k ( ξ ) T_k(\xi)-G_k(\xi)\le\Lambda F_k(\xi)-G_k(\xi) T k ( ξ ) − G k ( ξ ) ≤ Λ F k ( ξ ) − G k ( ξ ) . Substitution into (37) proves (38) . □
The scalar consequences include both familiar Bessel bounds.
Corollary 3.4 (The uncoupled envelope) For almost every ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n ,
F k ( ξ ) ≤ M Ω , (43) F_k(\xi)\le M_{\Omega}, \tag*{(43)} F k ( ξ ) ≤ M Ω , (43) and, for ξ ≠ 0 \xi\ne0 ξ = 0 ,
F k ( ξ ) ≤ M Ω min { 1 , Λ ∣ ξ ∣ 2 } . (44) F_k(\xi)\le M_{\Omega}\min\left\{1,\frac{\Lambda}{|\xi|^2}\right\}. \tag*{(44)} F k ( ξ ) ≤ M Ω min { 1 , ∣ ξ ∣ 2 Λ } . (44)
Proof. The first estimate is the scalar Bessel inequality. Applying Cauchy–Schwarz directly to (37) , or squaring the projection identity before splitting, gives ∣ ξ ∣ 2 F k ( ξ ) 2 ≤ M Ω T k ( ξ ) ≤ M Ω Λ F k ( ξ ) |\xi|^2F_k(\xi)^2\le M_{\Omega}T_k(\xi)\le M_{\Omega}\Lambda F_k(\xi) ∣ ξ ∣ 2 F k ( ξ ) 2 ≤ M Ω T k ( ξ ) ≤ M Ω Λ F k ( ξ ) . Combining these two estimates proves the assertion. □
The scalar envelope (44) was also used in [26 ] . Its associated variational coefficient is denoted here by c n e n v c_n^{\mathrm{env}} c n env in (5) and is defined independently in (61) .
Normalized constraints and the exact integrated slack
Put p = n / 2 p=n/2 p = n /2 , rescale ξ = Λ η \xi=\sqrt{\Lambda}\eta ξ = Λ η , and write
t = ∣ η ∣ 2 , f ( η ) = F k ( Λ η ) M Ω , g ( η ) = G k ( Λ η ) Λ M Ω . (45) t=|\eta|^2,\qquad f(\eta)=\frac{F_k(\sqrt{\Lambda}\eta)}{M_{\Omega}},\qquad g(\eta)=\frac{G_k(\sqrt{\Lambda}\eta)}{\Lambda M_{\Omega}}. \tag*{(45)} t = ∣ η ∣ 2 , f ( η ) = M Ω F k ( Λ η ) , g ( η ) = Λ M Ω G k ( Λ η ) . (45)
For t ≥ 0 t\geq0 t ≥ 0 define
m ( t ) = { 1 , t ∈ [ 0 , 1 ] , t − 1 , t > 1 , Ψ ( t , u ) = u [ t u − ( 1 − u ) ( 1 − t u ) ] + 2 . (46) m(t)=
\begin{cases}
1, & t\in[0,1],\\
t^{-1}, & t>1,
\end{cases}
\qquad
\Psi(t,u)=u\left[\sqrt{t}u-\sqrt{(1-u)(1-tu)}\right]_{+}^{2}. \tag*{(46)} m ( t ) = { 1 , t − 1 , t ∈ [ 0 , 1 ] , t > 1 , Ψ ( t , u ) = u [ t u − ( 1 − u ) ( 1 − t u ) ] + 2 . (46)
Lemma 3.5 (Basic properties of the penalty) On its natural domain 0 ≤ u ≤ m ( t ) 0\leq u\leq m(t) 0 ≤ u ≤ m ( t ) , the function Ψ \Psi Ψ is continuous, nonnegative, and nondecreasing separately in t t t and u u u . For the monotonicity in t t t , this means that t 1 ≤ t 2 t_1\leq t_2 t 1 ≤ t 2 and u ≤ m ( t 2 ) u\leq m(t_2) u ≤ m ( t 2 ) , so that both ( t 1 , u ) (t_1,u) ( t 1 , u ) and ( t 2 , u ) (t_2,u) ( t 2 , u ) lie in the natural domain. Moreover,
Ψ ( t , u ) = 0 ⟺ u ≤ 1 1 + t , (47) \Psi(t,u)=0\quad\Longleftrightarrow\quad u\leq\frac{1}{1+t}, \tag*{(47)} Ψ ( t , u ) = 0 ⟺ u ≤ 1 + t 1 , (47) and, for t > 0 t>0 t > 0 ,
Ψ ( t , m ( t ) ) = { t , 0 < t ≤ 1 , t − 2 , t ≥ 1. (48) \Psi(t,m(t))=
\begin{cases}
t, & 0<t\leq1,\\
t^{-2}, & t\geq1.
\end{cases} \tag*{(48)} Ψ ( t , m ( t )) = { t , t − 2 , 0 < t ≤ 1 , t ≥ 1. (48)
Proof. Continuity and nonnegativity are immediate. Put
B ( t , u ) = t u − ( 1 − u ) ( 1 − t u ) . B(t,u)=\sqrt{t}u-\sqrt{(1-u)(1-tu)}. B ( t , u ) = t u − ( 1 − u ) ( 1 − t u ) .
For fixed u u u , the first term is nondecreasing in t t t and the expression under the second square root is nonincreasing. Hence B B B and u [ B ] + 2 u[B]_{+}^{2} u [ B ] + 2 are nondecreasing in t t t . For fixed t t t , note that
∂ ∂ u [ ( 1 − u ) ( 1 − t u ) ] = − ( 1 + t ) + 2 t u ≤ 0 \frac{\partial}{\partial u}\left[(1-u)(1-tu)\right]=-(1+t)+2tu\leq0 ∂ u ∂ [ ( 1 − u ) ( 1 − t u ) ] = − ( 1 + t ) + 2 t u ≤ 0
on 0 ≤ u ≤ m ( t ) 0\leq u\leq m(t) 0 ≤ u ≤ m ( t ) : if t ≤ 1 t\leq1 t ≤ 1 , then 2 t u ≤ 2 t ≤ 1 + t 2tu\leq2t\leq1+t 2 t u ≤ 2 t ≤ 1 + t , while if t ≥ 1 t\geq1 t ≥ 1 , then 2 t u ≤ 2 ≤ 1 + t 2tu\leq2\leq1+t 2 t u ≤ 2 ≤ 1 + t . Thus B B B is nondecreasing in u u u ; since both u u u and [ B ] + 2 [B]_{+}^{2} [ B ] + 2 are nonnegative and nondecreasing, so is Ψ ( t , u ) = u [ B ( t , u ) ] + 2 \Psi(t,u)=u[B(t,u)]_{+}^{2} Ψ ( t , u ) = u [ B ( t , u ) ] + 2 . Finally, the bracket is nonpositive exactly when
t u 2 ≤ ( 1 − u ) ( 1 − t u ) , tu^2\leq(1-u)(1-tu), t u 2 ≤ ( 1 − u ) ( 1 − t u ) ,
which reduces to ( 1 + t ) u ≤ 1 (1+t)u\leq1 ( 1 + t ) u ≤ 1 and proves (47) . Substitution of m ( t ) = 1 m(t)=1 m ( t ) = 1 for t ≤ 1 t\leq1 t ≤ 1 and m ( t ) = t − 1 m(t)=t^{-1} m ( t ) = t − 1 for t ≥ 1 t\geq1 t ≥ 1 proves (48) . □
Lemma 3.6 (Pointwise elimination of the gradient variable) The normalized densities satisfy
0 ≤ f ≤ m ( t ) , 0 ≤ g ≤ f , t f ≤ f g + ( 1 − f ) ( f − g ) , (49) 0\leq f\leq m(t),\qquad0\leq g\leq f,\qquad\sqrt{t}f\leq\sqrt{fg}+\sqrt{(1-f)(f-g)}, \tag*{(49)} 0 ≤ f ≤ m ( t ) , 0 ≤ g ≤ f , t f ≤ f g + ( 1 − f ) ( f − g ) , (49) and hence
g ( η ) ≥ Ψ ( t , f ( η ) ) . (50) g(\eta)\geq\Psi(t,f(\eta)). \tag*{(50)} g ( η ) ≥ Ψ ( t , f ( η )) . (50)
Proof. By (44) , evaluated at ξ = Λ η \xi=\sqrt{\Lambda}\eta ξ = Λ η ,
0 ≤ f ( η ) ≤ min { 1 , 1 ∣ η ∣ 2 } = m ( t ) , 0\le f(\eta)\le\min\left\{1,\frac{1}{|\eta|^2}\right\}=m(t), 0 ≤ f ( η ) ≤ min { 1 , ∣ η ∣ 2 1 } = m ( t ) ,
where the value at t = 0 t=0 t = 0 is understood through the first branch of m m m . Likewise, (36) gives 0 ≤ g ≤ f 0\le g\le f 0 ≤ g ≤ f . Finally, substituting F k = M Ω f F_k=M_{\Omega}f F k = M Ω f , G k = Λ M Ω g G_k=\Lambda M_{\Omega}g G k = Λ M Ω g , and ∣ ξ ∣ = Λ t |\xi|=\sqrt{\Lambda}\sqrt{t} ∣ ξ ∣ = Λ t into (38) , and then dividing by Λ M Ω \sqrt{\Lambda}M_{\Omega} Λ M Ω , yields
t f ≤ f g + ( 1 − f ) ( f − g ) . \sqrt{t}f\le\sqrt{fg}+\sqrt{(1-f)(f-g)}. t f ≤ f g + ( 1 − f ) ( f − g ) .
At t = 0 t=0 t = 0 this inequality is automatic. Thus (49) holds.
It remains to eliminate g g g . The bound f ≤ m ( t ) f\le m(t) f ≤ m ( t ) implies both 0 ≤ f ≤ 1 0\le f\le1 0 ≤ f ≤ 1 and 0 ≤ t f ≤ 1 0\le tf\le1 0 ≤ t f ≤ 1 , so every square root below is real. If f > 0 f>0 f > 0 , set
s = g f , a = f , b = 1 − f , r = t f . s=\sqrt{\frac{g}{f}},\qquad a=\sqrt{f},\qquad b=\sqrt{1-f},\qquad r=\sqrt{tf}. s = f g , a = f , b = 1 − f , r = t f .
Then a 2 + b 2 = 1 a^2+b^2=1 a 2 + b 2 = 1 , 0 ≤ r , s ≤ 1 0\le r,s\le1 0 ≤ r , s ≤ 1 , and division of the preceding inequality by f \sqrt{f} f gives the scalar constraint
r ≤ a s + b 1 − s 2 . r\le as+b\sqrt{1-s^2}. r ≤ a s + b 1 − s 2 .
Write a = sin α a=\sin\alpha a = sin α and s = sin θ s=\sin\theta s = sin θ with α , θ ∈ [ 0 , π / 2 ] \alpha,\theta\in[0,\pi/2] α , θ ∈ [ 0 , π /2 ] . Its right-hand side is
sin α sin θ + cos α cos θ = cos ( θ − α ) . \sin\alpha\sin\theta+\cos\alpha\cos\theta=\cos(\theta-\alpha). sin α sin θ + cos α cos θ = cos ( θ − α ) .
If r ≤ b r\le b r ≤ b , then s = 0 s=0 s = 0 is admissible. If r > b r>b r > b , the first admissible angle is θ = α − arccos r \theta=\alpha-\arccos r θ = α − arccos r , and hence
s ≥ a r − b 1 − r 2 . s\ge ar-b\sqrt{1-r^2}. s ≥ a r − b 1 − r 2 .
Both cases can therefore be written as
s ≥ [ a r − b 1 − r 2 ] + = [ t f − ( 1 − f ) ( 1 − t f ) ] + . s\ge\left[ar-b\sqrt{1-r^2}\right]_+
=\left[\sqrt{tf}-\sqrt{(1-f)(1-tf)}\right]_+. s ≥ [ a r − b 1 − r 2 ] + = [ t f − ( 1 − f ) ( 1 − t f ) ] + .
Since g = f s 2 g=fs^2 g = f s 2 , squaring and multiplying by f f f proves g ≥ Ψ ( t , f ) g\ge\Psi(t,f) g ≥ Ψ ( t , f ) . If f = 0 f=0 f = 0 , the already proved inequality 0 ≤ g ≤ f 0\le g\le f 0 ≤ g ≤ f gives g = 0 = Ψ ( t , 0 ) g=0=\Psi(t,0) g = 0 = Ψ ( t , 0 ) . The inactive branch in this argument is exactly the zero set described in (47) . □ \square □
Proposition 3.7 (Moments and exact slack) The zeroth and second moments and the projected-gradient mass are
∫ R n f d η = k M Ω Λ p , (51) \int_{\mathbb{R}^n} f\,d\eta=\frac{k}{M_{\Omega}\Lambda^p}, \tag*{(51)} ∫ R n f d η = M Ω Λ p k , (51) ∫ R n t f d η = S k M Ω Λ p + 1 , (52) \int_{\mathbb{R}^n} tf\,d\eta=\frac{S_k}{M_{\Omega}\Lambda^{p+1}}, \tag*{(52)} ∫ R n t f d η = M Ω Λ p + 1 S k , (52) ∫ R n g d η = I k M Ω Λ p + 1 . (53) \int_{\mathbb{R}^n} g\,d\eta=\frac{I_k}{M_{\Omega}\Lambda^{p+1}}. \tag*{(53)} ∫ R n g d η = M Ω Λ p + 1 I k . (53) Moreover,
∫ R n ( t f + 1 p g ) d η ≤ ∫ R n f d η , (54) \int_{\mathbb{R}^n}\left(tf+\frac{1}{p}g\right)d\eta\le\int_{\mathbb{R}^n}f\,d\eta, \tag*{(54)} ∫ R n ( t f + p 1 g ) d η ≤ ∫ R n f d η , (54) and the slack is exactly
∫ R n f d η − ∫ R n ( t f + 1 p g ) d η = B k n M Ω Λ p + 1 . (55) \int_{\mathbb{R}^n}f\,d\eta-\int_{\mathbb{R}^n}\left(tf+\frac{1}{p}g\right)d\eta=\frac{\mathcal{B}_k}{nM_{\Omega}\Lambda^{p+1}}. \tag*{(55)} ∫ R n f d η − ∫ R n ( t f + p 1 g ) d η = n M Ω Λ p + 1 B k . (55)
Proof. Under the change of variables ξ = Λ η \xi= \sqrt{\Lambda}\eta ξ = Λ η , one has d η = Λ − n / 2 d ξ = Λ − p d ξ \mathrm{d}\eta= \Lambda^{-n/2}\mathrm{d}\xi= \Lambda^{-p}\mathrm{d}\xi d η = Λ − n /2 d ξ = Λ − p d ξ and t = ∣ η ∣ 2 = ∣ ξ ∣ 2 / Λ t = |\eta|^2 = |\xi|^2/\Lambda t = ∣ η ∣ 2 = ∣ ξ ∣ 2 /Λ . Consequently,
∫ R n f ( η ) d η = Λ − p M Ω ∫ R n F k ( ξ ) d ξ = k M Ω Λ p , ∫ R n t f ( η ) d η = Λ − p − 1 M Ω ∫ R n ∣ ξ ∣ 2 F k ( ξ ) d ξ = S k M Ω Λ p + 1 , ∫ R n g ( η ) d η = Λ − p − 1 M Ω ∫ R n G k ( ξ ) d ξ = I k M Ω Λ p + 1 . \begin{aligned}
\int_{\mathbb{R}^n} f(\eta)\,\mathrm{d}\eta&= \frac{\Lambda^{-p}}{M_{\Omega}}\int_{\mathbb{R}^n}F_k(\xi)\,\mathrm{d}\xi= \frac{k}{M_{\Omega}\Lambda^p},\\
\int_{\mathbb{R}^n}t f(\eta)\,\mathrm{d}\eta&= \frac{\Lambda^{-p-1}}{M_{\Omega}}\int_{\mathbb{R}^n}|\xi|^2F_k(\xi)\,\mathrm{d}\xi= \frac{S_k}{M_{\Omega}\Lambda^{p+1}},\\
\int_{\mathbb{R}^n}g(\eta)\,\mathrm{d}\eta&= \frac{\Lambda^{-p-1}}{M_{\Omega}}\int_{\mathbb{R}^n}G_k(\xi)\,\mathrm{d}\xi= \frac{I_k}{M_{\Omega}\Lambda^{p+1}}.
\end{aligned} ∫ R n f ( η ) d η ∫ R n t f ( η ) d η ∫ R n g ( η ) d η = M Ω Λ − p ∫ R n F k ( ξ ) d ξ = M Ω Λ p k , = M Ω Λ − p − 1 ∫ R n ∣ ξ ∣ 2 F k ( ξ ) d ξ = M Ω Λ p + 1 S k , = M Ω Λ − p − 1 ∫ R n G k ( ξ ) d ξ = M Ω Λ p + 1 I k .
Here the first two final equalities use (14) , and the third uses (35) . This proves (51) to (53) and also exposes all scaling factors in the normalization.
Using p = n / 2 p=n/2 p = n /2 and the definition of D k D_k D k , the difference between the two sides of (54) is
∫ R n f d η − ∫ R n ( t f + 1 p g ) d η = k Λ − S k − p − 1 I k M Ω Λ p + 1 = D k − 2 n I k M Ω Λ p + 1 = B k n M Ω Λ p + 1 . \begin{aligned}
\int_{\mathbb{R}^n}f\,\mathrm{d}\eta-\int_{\mathbb{R}^n}\left(tf+\frac{1}{p}g\right)\,\mathrm{d}\eta
&=\frac{k\Lambda-S_k-p^{-1}I_k}{M_{\Omega}\Lambda^{p+1}}\\
&=\frac{D_k-\frac{2}{n}I_k}{M_{\Omega}\Lambda^{p+1}}
=\frac{\mathcal{B}_k}{nM_{\Omega}\Lambda^{p+1}}.
\end{aligned} ∫ R n f d η − ∫ R n ( t f + p 1 g ) d η = M Ω Λ p + 1 k Λ − S k − p − 1 I k = M Ω Λ p + 1 D k − n 2 I k = n M Ω Λ p + 1 B k .
where the last equality is precisely (29) . This proves (55) ; its nonnegativity, already recorded after (28) , gives (54) . □
Remark 3.8 (What is gained over the scalar envelope) The cap (44) remembers only f ≤ m ( t ) f\le m(t) f ≤ m ( t ) . The new system additionally forces g ≥ Ψ ( t , f ) g\ge\Psi(t,f) g ≥ Ψ ( t , f ) and charges the integral of g g g through (54) . It is this simultaneous pointwise and integrated restriction, rather than a new scalar cap alone, that improves the principal coefficient below.
A Universal Commutator-Coupled Eigenvalue Constant
We now solve the dimensionless variational problem created by the coupled constraints. For a > 1 a>1 a > 1 , define
L n , a ( t ) = max 0 ≤ u ≤ m ( t ) { ( a − t ) u − 1 p Ψ ( t , u ) } . (56) L_{n,a}(t)=\max_{0\le u\le m(t)}\left\{(a-t)u-\frac{1}{p}\Psi(t,u)\right\}. \tag*{(56)} L n , a ( t ) = 0 ≤ u ≤ m ( t ) max { ( a − t ) u − p 1 Ψ ( t , u ) } . (56)
J n ( a ) = p ∫ 0 a L n , a ( t ) t p − 1 d t . (57) J_n(a)=p\int_0^a L_{n,a}(t)t^{p-1}\,\mathrm{d}t. \tag*{(57)} J n ( a ) = p ∫ 0 a L n , a ( t ) t p − 1 d t . (57)
C n n e w = sup a > 1 ( a − 1 J n ( a ) ) 1 / p . (58) C_n^{\mathrm{new}}=\sup_{a>1}\left(\frac{a-1}{J_n(a)}\right)^{1/p}. \tag*{(58)} C n new = a > 1 sup ( J n ( a ) a − 1 ) 1/ p . (58)
The maximization in (56) is one-dimensional and takes place on a compact interval. Moreover, L n , a ( t ) = 0 L_{n,a}(t)=0 L n , a ( t ) = 0 for t ≥ a t\ge a t ≥ a , because the objective is then nonpositive and vanishes at u = 0 u=0 u = 0 ; for 0 < t < a 0<t<a 0 < t < a it is positive, since one may choose 0 < u ≤ min { m ( t ) , ( 1 + t ) − 1 } 0<u\le\min\{m(t),(1+t)^{-1}\} 0 < u ≤ min { m ( t ) , ( 1 + t ) − 1 } , for which Ψ ( t , u ) = 0 \Psi(t,u)=0 Ψ ( t , u ) = 0 . Consequently 0 < J n ( a ) < ∞ 0<J_n(a)<\infty 0 < J n ( a ) < ∞ , and J n ( a ) J_n(a) J n ( a ) is directly computable without any domain-dependent quantity.
Theorem 4.1 (Commutator-coupled individual bound) Let Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n , n ≥ 2 n \ge2 n ≥ 2 , be an open set with 0 < V = ∣ Ω ∣ < ∞ 0 < V = |\Omega| < \infty 0 < V = ∣Ω∣ < ∞ . Then, for every k ≥ 1 k \ge1 k ≥ 1 ,
λ k ( Ω ) ≥ C n e n v C n V − 2 / n k 2 / n . (59) \lambda_k(\Omega) \ge C_n^{\mathrm{env}} C_n V^{-2/n} k^{2/n}. \tag*{(59)} λ k ( Ω ) ≥ C n env C n V − 2/ n k 2/ n . (59)
Proof. Assume first that Ω \Omega Ω is bounded and use the functions in (45) . For every a > 1 a>1 a > 1 , (54) gives
( a − 1 ) ∫ R n f d η ≤ ∫ R n ( ( a − t ) f − 1 p g ) d η ≤ ∫ R n L n , a ( ∣ η ∣ 2 ) d η = ω n J n ( a ) , \begin{aligned}
(a-1)\int_{\mathbb{R}^{n}} f\,\mathrm{d}\eta\le\int_{\mathbb{R}^{n}}\left((a-t)f-\frac{1}{p}g\right)\,\mathrm{d}\eta\\
&\le\int_{\mathbb{R}^{n}} L_{n,a}(|\eta|^{2})\,\mathrm{d}\eta=\omega_n J_n(a),
\end{aligned} ( a − 1 ) ∫ R n f d η ≤ ∫ R n ( ( a − t ) f − p 1 g ) d η ≤ ∫ R n L n , a ( ∣ η ∣ 2 ) d η = ω n J n ( a ) ,
where Lemma 3.6 was used in the second step. By (51) ,
k M Ω Λ p ≤ ω n J n ( a ) a − 1 . \frac{k}{M_{\Omega}\Lambda^p}\le\frac{\omega_n J_n(a)}{a-1}. M Ω Λ p k ≤ a − 1 ω n J n ( a ) .
If
W k = 4 π 2 ( k ω n V ) 2 / n , W_k=4\pi^2\left(\frac{k}{\omega_n V}\right)^{2/n}, W k = 4 π 2 ( ω n V k ) 2/ n ,
then W k p = k / ( M Ω ω n ) W_k^p=k/(M_{\Omega}\omega_n) W k p = k / ( M Ω ω n ) , and therefore
Λ W k ≥ ( a − 1 J n ( a ) ) 1 / p . \frac{\Lambda}{W_k}\ge\left(\frac{a-1}{J_n(a)}\right)^{1/p}. W k Λ ≥ ( J n ( a ) a − 1 ) 1/ p .
Taking the supremum over a > 1 a>1 a > 1 proves the result for bounded domains. For a general finite-measure open set, let Ω R = Ω ∩ B R \Omega_R=\Omega\cap B_R Ω R = Ω ∩ B R . The spaces H 0 1 ( Ω R ) H^1_0(\Omega_R) H 0 1 ( Ω R ) increase to a form core of H 0 1 ( Ω ) H^1_0(\Omega) H 0 1 ( Ω ) , so the min–max principle gives λ k ( Ω R ) ↓ λ k ( Ω ) \lambda_k(\Omega_R)\downarrow\lambda_k(\Omega) λ k ( Ω R ) ↓ λ k ( Ω ) , while ∣ Ω R ∣ ↑ V |\Omega_R|\uparrow V ∣ Ω R ∣ ↑ V . Applying the bounded-domain result and passing to the limit completes the proof. □
Strict comparison with the uncoupled two-envelope method
If the nonlinear penalty in (56) is discarded, one obtains
J n , 0 ( a ) = p ∫ 0 a ( a − t ) m ( t ) t p − 1 d t . (60) J_{n,0}(a)=p\int_0^a(a-t)m(t)t^{p-1}\,\mathrm{d}t. \tag*{(60)} J n , 0 ( a ) = p ∫ 0 a ( a − t ) m ( t ) t p − 1 d t . (60)
c n e n v = max a > 1 ( a − 1 J n , 0 ( a ) ) 1 / p . (61) c_n^{\mathrm{env}}=\max_{a>1}\left(\frac{a-1}{J_{n,0}(a)}\right)^{1/p}. \tag*{(61)} c n env = a > 1 max ( J n , 0 ( a ) a − 1 ) 1/ p . (61)
A direct integration gives [26 ]
J n , 0 ( a ) = { a log a + 1 2 , p = 1 , a p − a p − 1 + 1 p + 1 , p > 1. (62) J_{n,0}(a)=
\begin{cases}
a\log a+\dfrac{1}{2}, & p=1,\\
\dfrac{a^p-a}{p-1}+\dfrac{1}{p+1}, & p>1.
\end{cases} \tag*{(62)} J n , 0 ( a ) = ⎩ ⎨ ⎧ a log a + 2 1 , p − 1 a p − a + p + 1 1 , p = 1 , p > 1. (62)
This is the coefficient generated by the uncoupled scalar envelope (44) , which was analyzed in [26 ] .
For comparison, define the constant-density relaxation
J n , L 2 ( a ) = p ∫ 0 a ( a − t ) t p − 1 d t = a p + 1 p + 1 . (63) J_{n,L2}(a)=p\int_{0}^{a}(a-t)t^{p-1}\,\mathrm{d}t=\frac{a^{p+1}}{p+1}. \tag*{(63)} J n , L 2 ( a ) = p ∫ 0 a ( a − t ) t p − 1 d t = p + 1 a p + 1 . (63)
It is the variational integral corresponding to the constant cap used in the Li–Yau argument [15 ] .
Proposition 4.2 (Strict improvement and Pólya obstruction) For every n ≥ 2 n\ge2 n ≥ 2 ,
n n + 2 < c n e n v < c n n e w < 1. (64) \frac{n}{n+2}<c_n^{\mathrm{env}}<c_n^{\mathrm{new}}<1. \tag*{(64)} n + 2 n < c n env < c n new < 1. (64) Proof. Put
Φ a ( t , u ) = ( a − t ) u − 1 p Ψ ( t , u ) , 0 ≤ u ≤ m ( t ) . \Phi_a(t,u)=(a-t)u-\frac{1}{p}\Psi(t,u),\qquad0\le u\le m(t). Φ a ( t , u ) = ( a − t ) u − p 1 Ψ ( t , u ) , 0 ≤ u ≤ m ( t ) . We prove the three comparisons separately.
Step 1: the coupled coefficient is strictly larger than the scalar-envelope coefficient. Fix a > 1 a>1 a > 1 and 0 < t < a 0<t<a 0 < t < a . Since a − t > 0 a-t>0 a − t > 0 and Ψ ≥ 0 \Psi\ge0 Ψ ≥ 0 ,
Φ a ( t , u ) ≤ ( a − t ) u ≤ ( a − t ) m ( t ) . (65) \Phi_a(t,u)\le(a-t)u\le(a-t)m(t). \tag*{(65)} Φ a ( t , u ) ≤ ( a − t ) u ≤ ( a − t ) m ( t ) . (65) The admissible interval is compact and Φ a ( t , ⋅ ) \Phi_a(t,\cdot) Φ a ( t , ⋅ ) is continuous, so the maximum defining L n , a ( t ) L_{n,a}(t) L n , a ( t ) is attained. Equality in both inequalities in (65) would force u = m ( t ) u=m(t) u = m ( t ) and Ψ ( t , m ( t ) ) = 0 \Psi(t,m(t))=0 Ψ ( t , m ( t )) = 0 . This is impossible because (48) gives Ψ ( t , m ( t ) ) > 0 \Psi(t,m(t))>0 Ψ ( t , m ( t )) > 0 for t > 0 t>0 t > 0 . Therefore
L n , a ( t ) < ( a − t ) m ( t ) , 0 < t < a . L_{n,a}(t)<(a-t)m(t),\qquad0<t<a. L n , a ( t ) < ( a − t ) m ( t ) , 0 < t < a . Multiplication by p t p − 1 pt^{p-1} p t p − 1 and integration yields
J n ( a ) < J n , 0 ( a ) , a > 1. (66) J_n(a)<J_{n,0}(a),\qquad a>1. \tag*{(66)} J n ( a ) < J n , 0 ( a ) , a > 1. (66) The explicit formula (62) shows that
r 0 ( a ) : = ( a − 1 J n , 0 ( a ) ) 1 / p r_0(a):=\left(\frac{a-1}{J_{n,0}(a)}\right)^{1/p} r 0 ( a ) := ( J n , 0 ( a ) a − 1 ) 1/ p is continuous on ( 1 , ∞ ) (1,\infty) ( 1 , ∞ ) and tends to zero as a ↓ 1 a\downarrow1 a ↓ 1 and as a → ∞ a\to\infty a → ∞ . Hence it attains its maximum at some a e n v > 1 a_{\mathrm{env}}>1 a env > 1 . By (66) ,
c n n e w ≥ ( a e n v − 1 J n ( a e n v ) ) 1 / p > r 0 ( a e n v ) = c n e n v . c_n^{\mathrm{new}}\ge\left(\frac{a_{\mathrm{env}}-1}{J_n(a_{\mathrm{env}})}\right)^{1/p}>r_0(a_{\mathrm{env}})=c_n^{\mathrm{env}}. c n new ≥ ( J n ( a env ) a env − 1 ) 1/ p > r 0 ( a env ) = c n env . Step 2: the scalar envelope strictly improves the direct Li–Yau coefficient. Because m ( t ) ≤ 1 m(t)\le1 m ( t ) ≤ 1 and m ( t ) < 1 m(t)<1 m ( t ) < 1 for t > 1 t>1 t > 1 , for every a > 1 a>1 a > 1 we have
J n , L 2 ( a ) − J n , 0 ( a ) = p ∫ 1 a ( a − t ) ( 1 − t − 1 ) t p − 1 d t > 0. J_{n,L2}(a)-J_{n,0}(a)=p\int_{1}^{a}(a-t)(1-t^{-1})t^{p-1}\,\mathrm{d}t>0. J n , L 2 ( a ) − J n , 0 ( a ) = p ∫ 1 a ( a − t ) ( 1 − t − 1 ) t p − 1 d t > 0. The function associated with the constant-density relaxation is maximized at a L = 1 + 1 / p a_L=1+1/p a L = 1 + 1/ p , and
max a > 1 ( a − 1 J n , L 2 ( a ) ) 1 / p = ( ( p + 1 ) ( a L − 1 ) a L p + 1 ) 1 / p = p p + 1 = n n + 2 . (67) \max_{a>1}\left(\frac{a-1}{J_{n,L2}(a)}\right)^{1/p}
=\left(\frac{(p+1)(a_L-1)}{a_L^{p+1}}\right)^{1/p}
=\frac{p}{p+1}=\frac{n}{n+2}. \tag*{(67)} a > 1 max ( J n , L 2 ( a ) a − 1 ) 1/ p = ( a L p + 1 ( p + 1 ) ( a L − 1 ) ) 1/ p = p + 1 p = n + 2 n . (67) Evaluating the scalar-envelope quotient at a L a_L a L and using the strict integral inequality gives c n e n v > n / ( n + 2 ) c_n^{\mathrm{env}}>n/(n+2) c n env > n / ( n + 2 ) .
Step 3: the coupled relaxation still has coefficient strictly below one. Define
A p ( T ) = p ∫ 0 T t p − 1 1 + t d t , T p p = 2 A p ( T p ) , (68) A_p(T)=p\int_0^T\frac{t^{p-1}}{1+t}\,dt,\qquad T_p^p=2A_p(T_p), \tag*{(68)} A p ( T ) = p ∫ 0 T 1 + t t p − 1 d t , T p p = 2 A p ( T p ) , (68)
where the second equation has a unique solution T p > 1 T_p>1 T p > 1 . To see this, let B p ( T ) = T p − 2 A p ( T ) B_p(T)=T^p-2A_p(T) B p ( T ) = T p − 2 A p ( T ) , so that
B p ′ ( T ) = p T p − 1 T − 1 T + 1 , (69) B_p'(T)=pT^{p-1}\frac{T-1}{T+1}, \tag*{(69)} B p ′ ( T ) = p T p − 1 T + 1 T − 1 , (69)
for T > 0 T>0 T > 0 . Moreover, A p ( 1 ) > 1 / 2 A_p(1)>1/2 A p ( 1 ) > 1/2 , hence B p ( 1 ) < 0 B_p(1)<0 B p ( 1 ) < 0 . If p = 1 p=1 p = 1 , then A 1 ( T ) = log ( 1 + T ) A_1(T)=\log(1+T) A 1 ( T ) = log ( 1 + T ) ; if p > 1 p>1 p > 1 , then A p ( T ) = O ( T p − 1 ) A_p(T)=O(T^{p-1}) A p ( T ) = O ( T p − 1 ) . In either case B p ( T ) → ∞ B_p(T)\to\infty B p ( T ) → ∞ as T → ∞ T\to\infty T → ∞ . Thus (69) gives exactly one zero in ( 1 , ∞ ) (1,\infty) ( 1 , ∞ ) .
Consider the radial pair
f ∗ ( η ) = 1 { ∣ η ∣ 2 ≤ T p } 1 + ∣ η ∣ 2 , g ∗ ( η ) = 0 , f_*(\eta)=\frac{\mathbf{1}_{\{|\eta|^2\leq T_p\}}}{1+|\eta|^2},\qquad g_*(\eta)=0, f ∗ ( η ) = 1 + ∣ η ∣ 2 1 { ∣ η ∣ 2 ≤ T p } , g ∗ ( η ) = 0 ,
and recall the radial integration formula
∫ R n h ( ∣ η ∣ 2 ) d η = p ω n ∫ 0 ∞ h ( t ) t p − 1 d t . (70) \int_{\mathbb{R}^n}h(|\eta|^2)\,d\eta=p\omega_n\int_0^\infty h(t)t^{p-1}\,dt. \tag*{(70)} ∫ R n h ( ∣ η ∣ 2 ) d η = p ω n ∫ 0 ∞ h ( t ) t p − 1 d t . (70)
For 0 ≤ t ≤ T p 0\leq t\leq T_p 0 ≤ t ≤ T p , the identity f ∗ = 1 / ( 1 + t ) f_*=1/(1+t) f ∗ = 1/ ( 1 + t ) gives
t f ∗ = ( 1 − f ∗ ) f ∗ ; \sqrt{t}\,f_*=\sqrt{(1-f_*)f_*}; t f ∗ = ( 1 − f ∗ ) f ∗ ;
outside this ball both sides vanish. Also 0 ≤ f ∗ ≤ m ( t ) 0\leq f_*\leq m(t) 0 ≤ f ∗ ≤ m ( t ) . Consequently ( f ∗ , g ∗ ) (f_*,g_*) ( f ∗ , g ∗ ) satisfies the complete pointwise system in (49) . From (70) and T p p = 2 A p ( T p ) T_p^p=2A_p(T_p) T p p = 2 A p ( T p ) ,
∫ R n f ∗ d η = ω n A p ( T p ) , \int_{\mathbb{R}^n}f_*\,d\eta=\omega_nA_p(T_p), ∫ R n f ∗ d η = ω n A p ( T p ) ,
∫ R n ∣ η ∣ 2 f ∗ d η = ω n ( T p p − A p ( T p ) ) = ω n A p ( T p ) . \int_{\mathbb{R}^n}|\eta|^2f_*\,d\eta=\omega_n\bigl(T_p^p-A_p(T_p)\bigr)=\omega_nA_p(T_p). ∫ R n ∣ η ∣ 2 f ∗ d η = ω n ( T p p − A p ( T p ) ) = ω n A p ( T p ) .
Thus the integrated constraint also holds, with equality.
Since L n , a ( t ) L_{n,a}(t) L n , a ( t ) is the pointwise maximum of Φ a ( t , u ) \Phi_a(t,u) Φ a ( t , u ) and is zero for t ≥ a t\geq a t ≥ a , (70) gives, for every a > 1 a>1 a > 1 ,
ω n J n ( a ) ≥ ∫ R n ( a − ∣ η ∣ 2 ) f ∗ ( η ) d η = ( a − 1 ) ω n A p ( T p ) . \omega_nJ_n(a)\geq\int_{\mathbb{R}^n}(a-|\eta|^2)f_*(\eta)\,d\eta=(a-1)\omega_nA_p(T_p). ω n J n ( a ) ≥ ∫ R n ( a − ∣ η ∣ 2 ) f ∗ ( η ) d η = ( a − 1 ) ω n A p ( T p ) .
It follows that
C n n e w ≤ C n f l a t : = A p ( T p ) − 1 / p = 2 1 / p T p . (71) C_n^{\mathrm{new}}\leq C_n^{\mathrm{flat}}:=A_p(T_p)^{-1/p}=\frac{2^{1/p}}{T_p}. \tag*{(71)} C n new ≤ C n flat := A p ( T p ) − 1/ p = T p 2 1/ p . (71)
It remains to verify that the last number is strictly smaller than one. Let T 0 = 2 1 / p T_0=2^{1/p} T 0 = 2 1/ p . After setting t = T 0 r t=T_0r t = T 0 r , Jensen’s inequality for the strictly convex function r ↦ ( 1 + T 0 r ) − 1 r\mapsto(1+T_0r)^{-1} r ↦ ( 1 + T 0 r ) − 1 and the probability density p r p − 1 pr^{p-1} p r p − 1 on ( 0 , 1 ) (0,1) ( 0 , 1 ) gives
A p ( T 0 ) = 2 ∫ 0 1 p r p − 1 1 + T 0 r d r > 2 1 + T 0 p / ( p + 1 ) ≥ 1 , A_p(T_0)=2\int_0^1\frac{pr^{p-1}}{1+T_0r}\,dr>\frac{2}{1+T_0p/(p+1)}\geq1, A p ( T 0 ) = 2 ∫ 0 1 1 + T 0 r p r p − 1 d r > 1 + T 0 p / ( p + 1 ) 2 ≥ 1 ,
where the last inequality is equivalent to 2 1 / p ≤ 1 + 1 / p 2^{1/p}\leq1+1/p 2 1/ p ≤ 1 + 1/ p , which follows from Bernoulli’s inequality ( 1 + 1 / p ) p ≥ 2 (1+1/p)^p\geq2 ( 1 + 1/ p ) p ≥ 2 . Hence B p ( T 0 ) = 2 − 2 A p ( T 0 ) < 0 B_p(T_0)=2-2A_p(T_0)<0 B p ( T 0 ) = 2 − 2 A p ( T 0 ) < 0 . Since B p B_p B p is strictly increasing on ( 1 , ∞ ) (1,\infty) ( 1 , ∞ ) and vanishes at T p T_p T p , we have T p > T 0 T_p>T_0 T p > T 0 , and therefore C n n e w ≤ C n f l a t < 1 C_n^{\mathrm{new}}\leq C_n^{\mathrm{flat}}<1 C n new ≤ C n flat < 1 .
Combining the three steps proves (64) . □
Remark 4.3 (Meaning of the upper obstruction) The last inequality is a limitation of the present relaxed constraint , not a counterexample to Pólya’s conjecture. A proof with coefficient one must use information not encoded by (49) and (54) , or must quantify the positive slack B k B_k B k in (55) .
Numerical reference values
The first coefficient column of Table 1 records the classical Li–Yau value n / ( n + 2 ) n/(n+2) n / ( n + 2 ) , and the second records the rigorously defined scalar-envelope coefficient c n e n v c_n^{\mathrm{env}} c n env from [26 ] . The final column is a numerical evaluation of the commutator-coupled quotient at the trial value of a a a shown in the preceding column. The values were obtained by maximizing first in u u u , numerically integrating the resulting function of t t t separately on ( 0 , 1 ) (0,1) ( 0 , 1 ) and ( 1 , a ) (1,a) ( 1 , a ) , and then performing the outer one-dimensional search in a a a . They are rounded to six decimal places and are included only to indicate the size of the gain; they are not interval-certified bounds and are not used in any proof.
The table illustrates the strict hierarchy proved in Proposition 4.2 ; the last column is diagnostic and is not used in the proof. This hierarchy concerns individual eigenvalues. The next two sections instead retain the sharp semiclassical coefficient in the eigenvalue sum and improve its finite-k k k remainder.
A residual vector-field refinement
There is a natural place where additional geometry could enter. The gradient space
G = { ∇ u : u ∈ H 0 1 ( Ω ) } ⊂ L 2 ( Ω ; C n ) \mathcal{G}=\{\nabla u:u\in H_0^1(\Omega)\}\subset L^2(\Omega;\mathbb{C}^n) G = { ∇ u : u ∈ H 0 1 ( Ω )} ⊂ L 2 ( Ω ; C n )
is closed by Poincaré’s inequality. Since P k P_k P k is an orthogonal projection defined in § 3.2 , Ker P k \operatorname{Ker}P_k Ker P k is also closed. Hence
H k = G ⊥ ∩ Ker P k \mathcal{H}_k=\mathcal{G}^{\perp}\cap\operatorname{Ker}P_k H k = G ⊥ ∩ Ker P k
is also a closed subspace of L 2 ( Ω ; C n ) L^2(\Omega;\mathbb{C}^n) L 2 ( Ω ; C n ) . We denote by
Π H k : L 2 ( Ω ; C n ) ⟶ H k \Pi_{\mathcal{H}_k}:L^2(\Omega;\mathbb{C}^n)\longrightarrow\mathcal{H}_k Π H k : L 2 ( Ω ; C n ) ⟶ H k
the orthogonal projection onto H k \mathcal{H}_k H k , i.e., for every Z ∈ L 2 ( Ω ; C n ) Z \in L^2(\Omega;\mathbb{C}^n) Z ∈ L 2 ( Ω ; C n ) ,
Π H k Z ∈ H k , ⟨ Z − Π H k Z , W ⟩ L 2 ( Ω ; C n ) = 0 for all W ∈ H k . \Pi_{\mathcal{H}_k} Z \in\mathcal{H}_k,\qquad\langle Z-\Pi_{\mathcal{H}_k}Z,W\rangle_{L^2(\Omega;\mathbb{C}^n)}=0\quad\text{for all }W\in\mathcal{H}_k. Π H k Z ∈ H k , ⟨ Z − Π H k Z , W ⟩ L 2 ( Ω ; C n ) = 0 for all W ∈ H k .
In particular, Π H k 2 = Π H k = Π H k ∗ \Pi_{\mathcal{H}_k}^2=\Pi_{\mathcal{H}_k}=\Pi_{\mathcal{H}_k}^* Π H k 2 = Π H k = Π H k ∗ , and Ran Π H k = H k \operatorname{Ran}\Pi_{\mathcal{H}_k}=\mathcal{H}_k Ran Π H k = H k . Define
ρ k ( ξ ) = M Ω − 1 ∥ Π H k Y ξ ∥ 2 2 . \rho_k(\xi)=M_\Omega^{-1}\|\Pi_{\mathcal{H}_k}Y_\xi\|_2^2. ρ k ( ξ ) = M Ω − 1 ∥ Π H k Y ξ ∥ 2 2 .
Because H k ⊂ Ker P k \mathcal{H}_k\subset\operatorname{Ker}P_k H k ⊂ Ker P k , the projection vanishes on Ran P k \operatorname{Ran}P_k Ran P k , and therefore
Π H k Y ξ = Π H k ( I − P k ) Y ξ , 0 ≤ ρ k ( ξ ) ≤ M Ω − 1 ∥ ( I − P k ) Y ξ ∥ 2 2 = 1 − F k ( ξ ) M Ω . \Pi_{\mathcal{H}_k}Y_\xi=\Pi_{\mathcal{H}_k}(I-P_k)Y_\xi,\qquad0\leq\rho_k(\xi)\leq M_\Omega^{-1}\|(I-P_k)Y_\xi\|_2^2=1-\frac{F_k(\xi)}{M_\Omega}. Π H k Y ξ = Π H k ( I − P k ) Y ξ , 0 ≤ ρ k ( ξ ) ≤ M Ω − 1 ∥ ( I − P k ) Y ξ ∥ 2 2 = 1 − M Ω F k ( ξ ) .
where the last equality follows from (39) and (40) . Thus the quantity 1 − f − ρ k 1-f-\rho_k 1 − f − ρ k below is nonnegative; in the normalized variables we use the abbreviation ρ k = ρ k ( Λ η ) \rho_k=\rho_k(\sqrt{\Lambda}\eta) ρ k = ρ k ( Λ η ) .
Moreover, Π H k Y ξ \Pi_{\mathcal{H}_k}Y_\xi Π H k Y ξ is orthogonal to ∇ q ξ \nabla q_\xi ∇ q ξ and, being in Ker P k \operatorname{Ker}P_k Ker P k , to the componentwise low-mode space Ran P k \operatorname{Ran}P_k Ran P k . Removing this component before the second Cauchy–Schwarz estimate refines (49) to
t f ≤ f g + ( 1 − f − ρ k ) ( f − g ) . (72) \sqrt{t}\,f\leq\sqrt{fg}+\sqrt{(1-f-\rho_k)(f-g)}. \tag*{(72)} t f ≤ f g + ( 1 − f − ρ k ) ( f − g ) . (72)
In particular, at g = 0 g=0 g = 0 one has
( 1 + t ) f ≤ 1 − ρ k . (73) (1+t)f\leq1-\rho_k. \tag*{(73)} ( 1 + t ) f ≤ 1 − ρ k . (73)
This is a genuine improvement whenever ρ k \rho_k ρ k has a positive lower bound on a relevant frequency region. No such lower bound follows from volume alone; accordingly, (72) is recorded as a domain-sensitive refinement and is omitted from the universal coefficient.
Frequency Concentration and Multifrequency Constraints
The second part replaces the pointwise energy envelope by a collective constraint on the first k k k Fourier transforms. Its central object is a bounded-domain analogue of the classical time–frequency concentration operator [21 ] . We establish its strict spectral structure, derive first- and second-trace overlap formulas, and then convert them into a finite-rank restriction for the Fourier projection onto the first k k k eigenfunctions. Throughout Sections 5 5 5 and 6 6 6 , Ω \Omega Ω is assumed bounded.
The frequency-concentration operator
Define
γ Ω ( z ) = 1 V ∫ Ω e i x ⋅ z d x . (74) \gamma_\Omega(z)=\frac{1}{V}\int_\Omega e^{i x\cdot z}\,dx. \tag*{(74)} γ Ω ( z ) = V 1 ∫ Ω e i x ⋅ z d x . (74)
It follows directly that γ Ω ( 0 ) = 1 \gamma_{\Omega}(0)=1 γ Ω ( 0 ) = 1 , while the triangle inequality gives
∣ γ Ω ( z ) ∣ ≤ 1 V ∫ Ω ∣ e i x ⋅ z ∣ d x = 1 V ∫ Ω 1 d x = 1. |\gamma_{\Omega}(z)| \le\frac{1}{V}\int_{\Omega} |e^{i x\cdot z}|\,\mathrm{d}x = \frac{1}{V}\int_{\Omega} 1\,\mathrm{d}x = 1. ∣ γ Ω ( z ) ∣ ≤ V 1 ∫ Ω ∣ e i x ⋅ z ∣ d x = V 1 ∫ Ω 1 d x = 1.
Let 1 Ω \mathbf{1}_{\Omega} 1 Ω denote the indicator function of Ω \Omega Ω . Since V = ∣ Ω ∣ < ∞ V=|\Omega|<\infty V = ∣Ω∣ < ∞ , we have 1 Ω ∈ L 1 ( R n ) ∩ L 2 ( R n ) \mathbf{1}_{\Omega}\in L^{1}(\mathbb{R}^{n})\cap L^{2}(\mathbb{R}^{n}) 1 Ω ∈ L 1 ( R n ) ∩ L 2 ( R n ) . With the Fourier convention in (12) , the definition in (74) becomes
γ Ω ( z ) = ( 2 π ) n / 2 V 1 Ω ^ ( − z ) . \gamma_{\Omega}(z)=\frac{(2\pi)^{n/2}}{V}\widehat{\mathbf{1}_{\Omega}}(-z). γ Ω ( z ) = V ( 2 π ) n /2 1 Ω ( − z ) .
Using this identity and M Ω = V / ( 2 π ) n M_{\Omega}=V/(2\pi)^{n} M Ω = V / ( 2 π ) n , changing z z z to − z -z − z , and applying Plancherel’s theorem [22 ] , we obtain
M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 d z = V ( 2 π ) n ( 2 π ) n V 2 ∫ R n ∣ 1 Ω ^ ( − z ) ∣ 2 d z = 1 V ∫ R n ∣ 1 Ω ^ ( ξ ) ∣ 2 d ξ = 1 V ∫ R n ∣ 1 Ω ( x ) ∣ 2 d x = ∣ Ω ∣ V = 1 , \begin{aligned}
M_{\Omega}\int_{\mathbb{R}^{n}}|\gamma_{\Omega}(z)|^{2}\,\mathrm{d}z
&=\frac{V}{(2\pi)^{n}}\frac{(2\pi)^{n}}{V^{2}}\int_{\mathbb{R}^{n}}\left|\widehat{\mathbf{1}_{\Omega}}(-z)\right|^{2}\,\mathrm{d}z\\
&=\frac{1}{V}\int_{\mathbb{R}^{n}}\left|\widehat{\mathbf{1}_{\Omega}}(\xi)\right|^{2}\,\mathrm{d}\xi\\
&=\frac{1}{V}\int_{\mathbb{R}^{n}}|\mathbf{1}_{\Omega}(x)|^{2}\,\mathrm{d}x=\frac{|\Omega|}{V}=1,
\end{aligned} M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 d z = ( 2 π ) n V V 2 ( 2 π ) n ∫ R n 1 Ω ( − z ) 2 d z = V 1 ∫ R n 1 Ω ( ξ ) 2 d ξ = V 1 ∫ R n ∣ 1 Ω ( x ) ∣ 2 d x = V ∣Ω∣ = 1 ,
which proves
M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 d z = 1. (75) M_{\Omega}\int_{\mathbb{R}^{n}}|\gamma_{\Omega}(z)|^{2}\,\mathrm{d}z=1. \tag*{(75)} M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 d z = 1. (75)
For r > 0 r>0 r > 0 , let H r = L 2 ( B r ) H_{r}=L^{2}(B_{r}) H r = L 2 ( B r ) and G = L 2 ( Ω ) G=L^{2}(\Omega) G = L 2 ( Ω ) , and define
( U r f ) ( x ) = ( 2 π ) − n / 2 ∫ B r e i x ⋅ ξ f ( ξ ) d ξ . (76) (U_{r}f)(x)=(2\pi)^{-n/2}\int_{B_{r}}e^{i x\cdot\xi}f(\xi)\,\mathrm{d}\xi. \tag*{(76)} ( U r f ) ( x ) = ( 2 π ) − n /2 ∫ B r e i x ⋅ ξ f ( ξ ) d ξ . (76)
Its adjoint is the restricted Fourier transform
( U r ∗ u ) ( ξ ) = ( 2 π ) − n / 2 ∫ Ω e − i x ⋅ ξ u ( x ) d x , ξ ∈ B r . (77) (U_{r}^{*}u)(\xi)=(2\pi)^{-n/2}\int_{\Omega}e^{-i x\cdot\xi}u(x)\,\mathrm{d}x,\qquad\xi\in B_{r}. \tag*{(77)} ( U r ∗ u ) ( ξ ) = ( 2 π ) − n /2 ∫ Ω e − i x ⋅ ξ u ( x ) d x , ξ ∈ B r . (77)
Set
T r = U r ∗ U r : H r → H r , A r = U r U r ∗ : G → G . (78) T_{r}=U_{r}^{*}U_{r}:H_{r}\to H_{r},\qquad A_{r}=U_{r}U_{r}^{*}:G\to G. \tag*{(78)} T r = U r ∗ U r : H r → H r , A r = U r U r ∗ : G → G . (78)
Proposition 5.1 (Factorization and kernel) The operator U r U_{r} U r is Hilbert–Schmidt, with
∥ U r ∥ H S 2 = M Ω ∣ B r ∣ = M Ω ω n r n . (79) \|U_{r}\|_{\mathrm{HS}}^{2}=M_{\Omega}|B_{r}|=M_{\Omega}\omega_{n}r^{n}. \tag*{(79)} ∥ U r ∥ HS 2 = M Ω ∣ B r ∣ = M Ω ω n r n . (79) Moreover, T r T_{r} T r and A r A_{r} A r are positive, self-adjoint, compact operators, and
( T r f ) ( ξ ) = M Ω ∫ B r γ Ω ( η − ξ ) f ( η ) d η . (80) (T_{r}f)(\xi)=M_{\Omega}\int_{B_{r}}\gamma_{\Omega}(\eta-\xi)f(\eta)\,\mathrm{d}\eta. \tag*{(80)} ( T r f ) ( ξ ) = M Ω ∫ B r γ Ω ( η − ξ ) f ( η ) d η . (80) Finally,
0 ≤ T r ≤ I , 0 ≤ A r ≤ I . (81) 0\le T_{r}\le I,\qquad0\le A_{r}\le I. \tag*{(81)} 0 ≤ T r ≤ I , 0 ≤ A r ≤ I . (81)
Proof. Write
K r ( x , ξ ) = ( 2 π ) − n / 2 e i x ⋅ ξ , ( x , ξ ) ∈ Ω × B r , K_r(x,\xi)=(2\pi)^{-n/2}e^{ix\cdot\xi},\qquad(x,\xi)\in\Omega\times B_r, K r ( x , ξ ) = ( 2 π ) − n /2 e i x ⋅ ξ , ( x , ξ ) ∈ Ω × B r ,
for the integral kernel of U r U_r U r . Since Ω \Omega Ω and B r B_r B r have finite measure, K r ∈ L 2 ( Ω × B r ) K_r\in L^2(\Omega\times B_r) K r ∈ L 2 ( Ω × B r ) and hence U r U_r U r is Hilbert–Schmidt. More explicitly,
∥ U r ∥ H S 2 = ∫ Ω ∫ B r ∣ K r ( x , ξ ) ∣ 2 d ξ d x = ( 2 π ) − n ∣ Ω ∣ ∣ B r ∣ = V ( 2 π ) n ω n r n = M Ω ω n r n , \begin{aligned}
\|U_r\|_{\mathrm{HS}}^2=\int_{\Omega}\int_{B_r}|K_r(x,\xi)|^2\,d\xi\,dx \\
&=(2\pi)^{-n}|\Omega|\,|B_r|=\frac{V}{(2\pi)^n}\omega_n r^n=M_\Omega\omega_n r^n,
\end{aligned} ∥ U r ∥ HS 2 = ∫ Ω ∫ B r ∣ K r ( x , ξ ) ∣ 2 d ξ d x = ( 2 π ) − n ∣Ω∣ ∣ B r ∣ = ( 2 π ) n V ω n r n = M Ω ω n r n ,
which proves (79) . We use here the standard kernel characterization of Hilbert–Schmidt operators; see, for example, [20 Chapter 2] .
For completeness, the adjoint formula can be checked directly. If f ∈ H r f\in H_r f ∈ H r and u ∈ G u\in G u ∈ G , Fubini’s theorem (or first the calculation on dense L 1 ∩ L 2 L^1\cap L^2 L 1 ∩ L 2 subspaces and then continuity) gives
⟨ U r f , u ⟩ G = ( 2 π ) − n / 2 ∫ Ω ∫ B r e i x ⋅ η f ( η ) u ( x ) ‾ d η d x = ∫ B r f ( η ) ( 2 π ) − n / 2 ∫ Ω e − i x ⋅ η u ( x ) d x d η , \begin{aligned}
\langle U_rf,u\rangle_G=(2\pi)^{-n/2}\int_{\Omega}\int_{B_r}e^{ix\cdot\eta}f(\eta)\overline{u(x)}\,d\eta\,dx \\
&=\int_{B_r}f(\eta)(2\pi)^{-n/2}\int_{\Omega}e^{-ix\cdot\eta}u(x)\,dx\,d\eta,
\end{aligned} ⟨ U r f , u ⟩ G = ( 2 π ) − n /2 ∫ Ω ∫ B r e i x ⋅ η f ( η ) u ( x ) d η d x = ∫ B r f ( η ) ( 2 π ) − n /2 ∫ Ω e − i x ⋅ η u ( x ) d x d η ,
which is (77) . Substituting this formula into T r = U r ∗ U r T_r=U_r^*U_r T r = U r ∗ U r and interchanging the two finite-measure integrals yields
( T r f ) ( ξ ) = ( 2 π ) − n ∫ B r ∫ Ω e i x ⋅ ( η − ξ ) d x f ( η ) d η = V ( 2 π ) n ∫ B r ( 1 V ∫ Ω e i x ⋅ ( η − ξ ) d x ) f ( η ) d η = M Ω ∫ B r γ Ω ( η − ξ ) f ( η ) d η , \begin{aligned}
(T_rf)(\xi)=(2\pi)^{-n}\int_{B_r}\int_{\Omega}e^{ix\cdot(\eta-\xi)}\,dx\,f(\eta)\,d\eta\\
&=\frac{V}{(2\pi)^n}\int_{B_r}\left(\frac{1}{V}\int_{\Omega}e^{ix\cdot(\eta-\xi)}\,dx\right)f(\eta)\,d\eta\\
&=M_\Omega\int_{B_r}\gamma_\Omega(\eta-\xi)f(\eta)\,d\eta,
\end{aligned} ( T r f ) ( ξ ) = ( 2 π ) − n ∫ B r ∫ Ω e i x ⋅ ( η − ξ ) d x f ( η ) d η = ( 2 π ) n V ∫ B r ( V 1 ∫ Ω e i x ⋅ ( η − ξ ) d x ) f ( η ) d η = M Ω ∫ B r γ Ω ( η − ξ ) f ( η ) d η ,
which proves (80) with all normalization constants displayed.
The identities T r = U r ∗ U r T_r=U_r^*U_r T r = U r ∗ U r and A r = U r U r ∗ A_r=U_rU_r^* A r = U r U r ∗ show immediately that both operators are self-adjoint and positive. Since U r U_r U r and U r ∗ U_r^* U r ∗ are Hilbert–Schmidt, these two products are trace class, and in particular compact.
It remains to verify the operator bounds. Let E r : H r → L 2 ( R n ) E_r:H_r\to L^2(\mathbb{R}^n) E r : H r → L 2 ( R n ) denote extension by zero and R Ω : L 2 ( R n ) → L 2 ( Ω ) R_\Omega:L^2(\mathbb{R}^n)\to L^2(\Omega) R Ω : L 2 ( R n ) → L 2 ( Ω ) restriction. With the unitary Fourier convention used here,
U r = R Ω F − 1 E r . U_r=R_\Omega\mathcal{F}^{-1}E_r. U r = R Ω F − 1 E r .
Thus ∥ U r ∥ ≤ 1 \|U_r\|\le1 ∥ U r ∥ ≤ 1 by Plancherel. Equivalently, for every f ∈ H r f\in H_r f ∈ H r ,
⟨ T r f , f ⟩ H r = ∥ U r f ∥ G 2 ≤ ∥ f ∥ H r 2 . (82) \langle T_rf,f\rangle_{H_r}=\|U_rf\|_G^2\le\|f\|_{H_r}^2. \tag*{(82)} ⟨ T r f , f ⟩ H r = ∥ U r f ∥ G 2 ≤ ∥ f ∥ H r 2 . (82)
Together with positivity this is precisely 0 ≤ T r ≤ I 0\le T_r\le I 0 ≤ T r ≤ I . Likewise, U r ∗ U_r^* U r ∗ is the Fourier transform of the zero extension of u ∈ G u\in G u ∈ G , followed by restriction to B r B_r B r , so ∥ U r ∗ u ∥ H r ≤ ∥ u ∥ G \|U_r^*u\|_{H_r}\le\|u\|_G ∥ U r ∗ u ∥ H r ≤ ∥ u ∥ G . Therefore
0 ≤ ⟨ A r u , u ⟩ G = ∥ U r ∗ u ∥ H r 2 ≤ ∥ u ∥ G 2 , 0\le\langle A_ru,u\rangle_G=\|U_r^*u\|_{H_r}^2\le\|u\|_G^2, 0 ≤ ⟨ A r u , u ⟩ G = ∥ U r ∗ u ∥ H r 2 ≤ ∥ u ∥ G 2 ,
which proves 0 ≤ A r ≤ I 0\le A_r\le I 0 ≤ A r ≤ I and completes the proof. □
The strict spectral structure is the essential point. We use the following elementary uniqueness statement.
Lemma 5.2 (Analytic continuation and uniqueness) Let g ∈ L 1 ( R n ) g\in L^1(\mathbb{R}^n) g ∈ L 1 ( R n ) have bounded support. Then g ^ \widehat{g} g extends to an entire function on C n \mathbb{C}^n C n . Consequently, if g ^ \widehat{g} g vanishes on a nonempty open subset of R n \mathbb{R}^n R n , then g = 0 g=0 g = 0 almost everywhere. The same conclusions hold for the inverse Fourier transform of a compactly supported L 2 L^2 L 2 function.
Proof. Choose R > 0 R>0 R > 0 such that supp g ⊂ B R \operatorname{supp} g\subset B_R supp g ⊂ B R . The function
G ( z ) = ( 2 π ) − n / 2 ∫ B R e − i x ⋅ z g ( x ) d x . G(z)=(2\pi)^{-n/2}\int_{B_R}e^{-i x\cdot z}g(x)\,dx. G ( z ) = ( 2 π ) − n /2 ∫ B R e − i x ⋅ z g ( x ) d x .
is entire, because the integrand and all of its z z z -derivatives are locally uniformly dominated by an integrable multiple of ∣ g ∣ |g| ∣ g ∣ . This is the elementary compact-support case of the Paley–Wiener principle; see, for example, [23 Chapter 4, Section 3] .
If G G G vanishes on a nonempty open subset of R n \mathbb{R}^{n} R n , choose a real open rectangle contained in that set and apply the one-variable identity theorem successively in each coordinate. It follows that G ≡ 0 G\equiv0 G ≡ 0 on C n \mathbb{C}^{n} C n . Injectivity of the Fourier transform on L 1 ( R n ) L^{1}(\mathbb{R}^{n}) L 1 ( R n ) then gives g = 0 g=0 g = 0 almost everywhere; see [22 Chapter 6, Section 2] .
Finally, a compactly supported L 2 L^{2} L 2 function belongs to L 1 L^{1} L 1 by Cauchy–Schwarz. The inverse-transform assertions follow from the same argument with the sign of the exponential reversed. □
Theorem 5.3 (Strict spectral structure) The nonzero eigenvalues of T r T_r T r , repeated according to multiplicity, can be arranged as
1 > μ 1 ( r ) ≥ μ 2 ( r ) ≥ ⋯ > 0 , μ ℓ ( r ) ⟶ 0. (83) 1>\mu_1(r)\geq\mu_2(r)\geq\cdots>0,\qquad\mu_\ell(r)\longrightarrow0. \tag*{(83)} 1 > μ 1 ( r ) ≥ μ 2 ( r ) ≥ ⋯ > 0 , μ ℓ ( r ) ⟶ 0. (83) They form the complete spectrum of T r T_r T r apart from the accumulation point zero. The operator A r A_r A r has the same eigenvalues, including multiplicity, and its corresponding eigenfunctions form an orthonormal basis of G G G .
Proof. Let 0 ≠ f ∈ H r 0\ne f\in H_r 0 = f ∈ H r , extend f f f by zero to f ~ ∈ L 2 ( R n ) \widetilde{f}\in L^{2}(\mathbb{R}^{n}) f ∈ L 2 ( R n ) , and put
F = F − 1 f ~ . F=\mathcal{F}^{-1}\widetilde{f}. F = F − 1 f .
Because f ~ \widetilde{f} f is supported in B r ‾ \overline{B_r} B r , it also belongs to L 1 ( R n ) L^{1}(\mathbb{R}^{n}) L 1 ( R n ) , and F F F is real analytic (indeed entire after complexification) by Lemma 5.2 . Moreover,
U r f = F ∣ Ω , ⟨ T r f , f ⟩ H r = ∥ F ∥ L 2 ( Ω ) 2 , ∥ F ∥ L 2 ( R n ) 2 = ∥ f ∥ H r 2 . U_r f=F|_{\Omega},\qquad\langle T_r f,f\rangle_{H_r}=\|F\|_{L^{2}(\Omega)}^{2},\qquad\|F\|_{L^{2}(\mathbb{R}^{n})}^{2}=\|f\|_{H_r}^{2}. U r f = F ∣ Ω , ⟨ T r f , f ⟩ H r = ∥ F ∥ L 2 ( Ω ) 2 , ∥ F ∥ L 2 ( R n ) 2 = ∥ f ∥ H r 2 .
If the lower endpoint were attained, then F = 0 F=0 F = 0 almost everywhere on the nonempty open set Ω \Omega Ω . Since F F F is continuous, almost-everywhere vanishing on an open set implies pointwise vanishing there. Analytic uniqueness would imply f ~ = 0 \widetilde{f}=0 f = 0 , contrary to f ≠ 0 f\ne0 f = 0 . If the upper endpoint were attained, then
0 = ∥ F ∥ L 2 ( R n ) 2 − ∥ F ∥ L 2 ( Ω ) 2 = ∥ F ∥ L 2 ( R n ∖ Ω ) 2 . 0=\|F\|_{L^{2}(\mathbb{R}^{n})}^{2}-\|F\|_{L^{2}(\Omega)}^{2}=\|F\|_{L^{2}(\mathbb{R}^{n}\setminus\Omega)}^{2}. 0 = ∥ F ∥ L 2 ( R n ) 2 − ∥ F ∥ L 2 ( Ω ) 2 = ∥ F ∥ L 2 ( R n ∖ Ω ) 2 .
Since Ω \Omega Ω is bounded, R n ∖ Ω ‾ \mathbb{R}^{n}\setminus\overline{\Omega} R n ∖ Ω is a nonempty open set; F F F vanishes there almost everywhere and hence everywhere by continuity. Analytic uniqueness again forces f = 0 f=0 f = 0 . We have therefore proved the strict quadratic-form bounds
0 < ⟨ T r f , f ⟩ < ∥ f ∥ 2 ( 0 ≠ f ∈ H r ) . (84) 0<\langle T_r f,f\rangle<\|f\|^{2}\qquad(0\ne f\in H_r). \tag*{(84)} 0 < ⟨ T r f , f ⟩ < ∥ f ∥ 2 ( 0 = f ∈ H r ) . (84)
In particular, T r T_r T r is injective: T r f = 0 T_r f=0 T r f = 0 would imply ⟨ T r f , f ⟩ = 0 \langle T_r f,f\rangle=0 ⟨ T r f , f ⟩ = 0 . It also has no eigenvalue equal to one, because T r f = f T_r f=f T r f = f would imply ⟨ T r f , f ⟩ = ∥ f ∥ 2 \langle T_r f,f\rangle=\|f\|^{2} ⟨ T r f , f ⟩ = ∥ f ∥ 2 . By Proposition 5.1 , T r T_r T r is compact, positive, and self-adjoint. The compact self-adjoint spectral theorem [25 ] therefore shows that every nonzero spectral value is a positive eigenvalue of finite multiplicity, that the only possible accumulation point is zero, and that the corresponding eigenspaces span ( Ker T r ) ⊥ = H r (\operatorname{Ker}T_r)^\perp=H_r ( Ker T r ) ⊥ = H r . Since H r H_r H r is infinite dimensional whereas an injective finite-rank operator on H r H_r H r cannot exist, there are infinitely many such eigenvalues. Listing them with multiplicity gives
μ 1 ( r ) ≥ μ 2 ( r ) ≥ ⋯ > 0 , μ ℓ ( r ) ⟶ 0. \mu_1(r)\geq\mu_2(r)\geq\cdots>0,\qquad\mu_\ell(r)\longrightarrow0. μ 1 ( r ) ≥ μ 2 ( r ) ≥ ⋯ > 0 , μ ℓ ( r ) ⟶ 0.
Compactness also implies that a nonzero operator norm is attained as an eigenvalue; the exclusion of the eigenvalue one therefore yields μ 1 ( r ) = ∥ T r ∥ < 1 \mu_1(r)=\lVert T_r\rVert<1 μ 1 ( r ) = ∥ T r ∥ < 1 , proving (83) .
We next compare T r T_r T r and A r A_r A r . For every μ > 0 \mu>0 μ > 0 , the map
U r : Ker ( T r − μ I ) ⟶ Ker ( A r − μ I ) U_r:\operatorname{Ker}(T_r-\mu I)\longrightarrow\operatorname{Ker}(A_r-\mu I) U r : Ker ( T r − μ I ) ⟶ Ker ( A r − μ I )
is well defined, because
A r ( U r f ) = U r ( U r ∗ U r f ) = U r ( T r f ) = μ U r f . A_r(U_rf)=U_r(U_r^*U_rf)=U_r(T_rf)=\mu U_rf. A r ( U r f ) = U r ( U r ∗ U r f ) = U r ( T r f ) = μ U r f .
It is injective on this eigenspace since ∥ U r f ∥ 2 = ⟨ T r f , f ⟩ = μ ∥ f ∥ 2 \lVert U_rf\rVert^2=\langle T_rf,f\rangle=\mu\lVert f\rVert^2 ∥ U r f ∥ 2 = ⟨ T r f , f ⟩ = μ ∥ f ∥ 2 . Conversely, if A r u = μ u A_ru=\mu u A r u = μu , then
T r ( U r ∗ u ) = U r ∗ ( U r U r ∗ u ) = μ U r ∗ u , T_r(U_r^*u)=U_r^*(U_rU_r^*u)=\mu U_r^*u, T r ( U r ∗ u ) = U r ∗ ( U r U r ∗ u ) = μ U r ∗ u ,
and U r ∗ u ≠ 0 U_r^*u\ne0 U r ∗ u = 0 because ∥ U r ∗ u ∥ 2 = ⟨ A r u , u ⟩ = μ ∥ u ∥ 2 \lVert U_r^*u\rVert^2=\langle A_ru,u\rangle=\mu\lVert u\rVert^2 ∥ U r ∗ u ∥ 2 = ⟨ A r u , u ⟩ = μ ∥ u ∥ 2 . The inverse of the first map is μ − 1 U r ∗ \mu^{-1}U_r^* μ − 1 U r ∗ , so the two μ \mu μ -eigenspaces are isomorphic. Thus T r = U r ∗ U r T_r=U_r^*U_r T r = U r ∗ U r and A r = U r U r ∗ A_r=U_rU_r^* A r = U r U r ∗ have the same nonzero eigenvalues, including multiplicities; this is also the standard singular-value correspondence for a compact operator [20 Chapter 1] .
It remains to prove completeness in G G G . Suppose U r ∗ u = 0 U_r^*u=0 U r ∗ u = 0 . If u ~ \widetilde{u} u denotes the zero extension of u u u from Ω \Omega Ω to R n \mathbb{R}^n R n , then u ~ \widetilde{u} u has bounded support and belongs to L 1 L^1 L 1 by Cauchy–Schwarz, while
u ~ ^ ( ξ ) = 0 for almost every ξ ∈ B r . \widehat{\widetilde{u}}(\xi)=0\qquad\text{for almost every }\xi\in B_r. u ( ξ ) = 0 for almost every ξ ∈ B r .
The transform is continuous (indeed entire), so it vanishes everywhere on the open ball B r B_r B r . By Lemma 5.2, u = 0 u=0 u = 0 . Hence Ker U r ∗ = { 0 } \operatorname{Ker}U_r^*=\{0\} Ker U r ∗ = { 0 } and therefore Ker A r = Ker U r ∗ = { 0 } \operatorname{Ker}A_r=\operatorname{Ker}U_r^*=\{0\} Ker A r = Ker U r ∗ = { 0 } . Applying the compact self-adjoint spectral theorem to A r A_r A r gives
G = Ker A r ⊕ ⨁ μ > 0 Ker ( A r − μ I ) = ⨁ μ > 0 Ker ( A r − μ I ) . G=\operatorname{Ker}A_r\oplus\bigoplus_{\mu>0}\operatorname{Ker}(A_r-\mu I)=\bigoplus_{\mu>0}\operatorname{Ker}(A_r-\mu I). G = Ker A r ⊕ μ > 0 ⨁ Ker ( A r − μ I ) = μ > 0 ⨁ Ker ( A r − μ I ) .
Choosing an orthonormal basis in each finite-dimensional eigenspace therefore produces the asserted orthonormal eigenbasis of G G G . □
Remark 5.4 The pointwise strict inequality in (84) does not imply a uniform gap below one: for fixed Ω \Omega Ω , ∥ T r ∥ → 1 \lVert T_r\rVert\to1 ∥ T r ∥ → 1 as r → ∞ r\to\infty r → ∞ . What is used below is the strict inequality for every fixed finite radius.
Trace identities and overlap asymptotics
Theorem 5.5 (First and second traces) The operator T r T_r T r is positive trace class, T r 2 T_r^2 T r 2 is trace class, and
t ( r ) : = Tr ( T r ) = M Ω ∣ B r ∣ = M Ω ω n r n , (85) t(r):=\operatorname{Tr}(T_r)=M_\Omega|B_r|=M_\Omega\omega_n r^n, \tag*{(85)} t ( r ) := Tr ( T r ) = M Ω ∣ B r ∣ = M Ω ω n r n , (85) Q ( r ) : = Tr ( T r 2 ) = ∑ ℓ = 1 ∞ μ ℓ ( r ) 2 = M Ω 2 ∬ B r × B r ∣ γ Ω ( η − ξ ) ∣ 2 d η d ξ . (86) Q(r):=\operatorname{Tr}(T_r^2)=\sum_{\ell=1}^{\infty}\mu_\ell(r)^2=M_\Omega^2\iint_{B_r\times B_r}|\gamma_\Omega(\eta-\xi)|^2\,d\eta\,d\xi. \tag*{(86)} Q ( r ) := Tr ( T r 2 ) = ℓ = 1 ∑ ∞ μ ℓ ( r ) 2 = M Ω 2 ∬ B r × B r ∣ γ Ω ( η − ξ ) ∣ 2 d η d ξ . (86) Moreover,
0 < Q ( r ) < t ( r ) . (87) 0<Q(r)<t(r). \tag*{(87)} 0 < Q ( r ) < t ( r ) . (87)
Proof. Since U r U_r U r is Hilbert–Schmidt, T r = U r ∗ U r T_r=U_r^*U_r T r = U r ∗ U r is trace class and Tr ( T r ) = ∥ U r ∥ H S 2 \operatorname{Tr}(T_r)=\lVert U_r\rVert_{\mathrm{HS}}^2 Tr ( T r ) = ∥ U r ∥ HS 2 , proving (85) . The Hilbert–Schmidt norm of the kernel in (80) gives (86) . Finally, every μ ℓ ( r ) \mu_\ell(r) μ ℓ ( r ) lies in ( 0 , 1 ) (0,1) ( 0 , 1 ) , so 0 < ∑ ℓ μ ℓ ( r ) 2 < ∑ ℓ μ ℓ ( r ) = t ( r ) 0<\sum_\ell\mu_\ell(r)^2<\sum_\ell\mu_\ell(r)=t(r) 0 < ∑ ℓ μ ℓ ( r ) 2 < ∑ ℓ μ ℓ ( r ) = t ( r ) . □
Theorem 5.6 (Overlap representation) Let
a r ( z ) = ∣ B r ∩ ( B r − z ) ∣ . (88) a_r(z)=\lvert B_r\cap(B_r-z)\rvert. \tag*{(88)} a r ( z ) = ∣ B r ∩ ( B r − z )∣ . (88) Then
Q ( r ) = M Ω 2 ∫ R n ∣ γ Ω ( z ) ∣ 2 a r ( z ) d z , (89) Q(r)=M_\Omega^2\int_{\mathbb{R}^n}\lvert\gamma_\Omega(z)\rvert^2a_r(z)\,\mathrm{d}z, \tag*{(89)} Q ( r ) = M Ω 2 ∫ R n ∣ γ Ω ( z ) ∣ 2 a r ( z ) d z , (89) and
Q ( r ) t ( r ) ⟶ 1 ( r → ∞ ) . (90) \frac{Q(r)}{t(r)}\longrightarrow1\qquad(r\to\infty). \tag*{(90)} t ( r ) Q ( r ) ⟶ 1 ( r → ∞ ) . (90)
Figure 1 illustrates the intersection defining a r ( z ) a_r(z) a r ( z ) .
Figure 1. The overlap of the balls B r B_r B r and B r − z B_r-z B r − z with center separation ∣ z ∣ \lvert z\rvert ∣ z ∣ .
Proof Set
Φ ( z ) = ∣ γ Ω ( z ) ∣ 2 . \Phi(z)=\lvert\gamma_\Omega(z)\rvert^2. Φ ( z ) = ∣ γ Ω ( z ) ∣ 2 .
The integrand in (86) is nonnegative, so Tonelli’s theorem allows all changes in the order of integration below without any prior integrability assumption. Extending the integrations by means of the indicator 1 B r \mathbf{1}_{B_r} 1 B r and setting z = η − ξ z=\eta-\xi z = η − ξ , we obtain
∫ B r ∫ B r Φ ( η − ξ ) d η d ξ = ∫ R n ∫ R n Φ ( η − ξ ) 1 B r ( ξ ) 1 B r ( η ) d η d ξ = ∫ R n ∫ R n Φ ( z ) 1 B r ( ξ ) 1 B r ( ξ + z ) d ξ d z = ∫ R n Φ ( z ) ( ∫ R n 1 B r ( ξ ) 1 B r − z ( ξ ) d ξ ) d z = ∫ R n Φ ( z ) ∣ B r ∩ ( B r − z ) ∣ d z = ∫ R n ∣ γ Ω ( z ) ∣ 2 a r ( z ) d z . \begin{aligned}
\int_{B_r}\int_{B_r}\Phi(\eta-\xi)\,\mathrm{d}\eta\,\mathrm{d}\xi
&=\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}\Phi(\eta-\xi)\mathbf{1}_{B_r}(\xi)\mathbf{1}_{B_r}(\eta)\,\mathrm{d}\eta\,\mathrm{d}\xi\\
&=\int_{\mathbb{R}^n}\int_{\mathbb{R}^n}\Phi(z)\mathbf{1}_{B_r}(\xi)\mathbf{1}_{B_r}(\xi+z)\,\mathrm{d}\xi\,\mathrm{d}z\\
&=\int_{\mathbb{R}^n}\Phi(z)\left(\int_{\mathbb{R}^n}\mathbf{1}_{B_r}(\xi)\mathbf{1}_{B_r-z}(\xi)\,\mathrm{d}\xi\right)\,\mathrm{d}z\\
&=\int_{\mathbb{R}^n}\Phi(z)\lvert B_r\cap(B_r-z)\rvert\,\mathrm{d}z\\
&=\int_{\mathbb{R}^n}\lvert\gamma_\Omega(z)\rvert^2a_r(z)\,\mathrm{d}z.
\end{aligned} ∫ B r ∫ B r Φ ( η − ξ ) d η d ξ = ∫ R n ∫ R n Φ ( η − ξ ) 1 B r ( ξ ) 1 B r ( η ) d η d ξ = ∫ R n ∫ R n Φ ( z ) 1 B r ( ξ ) 1 B r ( ξ + z ) d ξ d z = ∫ R n Φ ( z ) ( ∫ R n 1 B r ( ξ ) 1 B r − z ( ξ ) d ξ ) d z = ∫ R n Φ ( z ) ∣ B r ∩ ( B r − z )∣ d z = ∫ R n ∣ γ Ω ( z ) ∣ 2 a r ( z ) d z .
Multiplication by M Ω 2 M_\Omega^2 M Ω 2 and comparison with (86) prove (89) . This calculation also shows directly that a r a_r a r is measurable, 0 ≤ a r ( z ) ≤ ∣ B r ∣ 0\leq a_r(z)\leq\lvert B_r\rvert 0 ≤ a r ( z ) ≤ ∣ B r ∣ , and a r ( z ) = 0 a_r(z)=0 a r ( z ) = 0 when ∣ z ∣ ≥ 2 r \lvert z\rvert\geq2r ∣ z ∣ ≥ 2 r . We next justify the pointwise limit needed for the asymptotic formula. By scaling,
a r ( z ) ∣ B r ∣ = ∣ B 1 ∩ ( B 1 − z / r ) ∣ ∣ B 1 ∣ . \frac{a_r(z)}{\lvert B_r\rvert}=\frac{\lvert B_1\cap(B_1-z/r)\rvert}{\lvert B_1\rvert}. ∣ B r ∣ a r ( z ) = ∣ B 1 ∣ ∣ B 1 ∩ ( B 1 − z / r )∣ .
For any vector h h h ,
∣ B 1 ∣ − ∣ B 1 ∩ ( B 1 − h ) ∣ = 1 2 ∣ B 1 △ ( B 1 − h ) ∣ , \lvert B_1\rvert-\lvert B_1\cap(B_1-h)\rvert=\frac{1}{2}\lvert B_1\mathbin{\triangle}(B_1-h)\rvert, ∣ B 1 ∣ − ∣ B 1 ∩ ( B 1 − h )∣ = 2 1 ∣ B 1 △ ( B 1 − h )∣ ,
where △ \triangle △ denotes symmetric difference. Translation is continuous in L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) , applied here to 1 B 1 \mathbf{1}_{B_1} 1 B 1 ; hence the last quantity tends to zero as h → 0 h \to0 h → 0 . Taking h = z / r h = z/r h = z / r gives, for every fixed z ∈ R n z \in\mathbb{R}^n z ∈ R n ,
a r ( z ) ∣ B r ∣ ⟶ 1 ( r → ∞ ) . \frac{a_r(z)}{|B_r|} \longrightarrow1 \qquad(r \to\infty). ∣ B r ∣ a r ( z ) ⟶ 1 ( r → ∞ ) .
Finally, divide (89) by t ( r ) = M Ω ∣ B r ∣ t(r) = M_\Omega|B_r| t ( r ) = M Ω ∣ B r ∣ to get
Q ( r ) t ( r ) = M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 a r ( z ) ∣ B r ∣ d z . (91) \frac{Q(r)}{t(r)} = M_\Omega\int_{\mathbb{R}^n} |\gamma_\Omega(z)|^2 \frac{a_r(z)}{|B_r|}\, dz. \tag*{(91)} t ( r ) Q ( r ) = M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 ∣ B r ∣ a r ( z ) d z . (91)
The ratio in the integrand lies between zero and one, while M Ω ∣ γ Ω ∣ 2 M_\Omega|\gamma_\Omega|^2 M Ω ∣ γ Ω ∣ 2 is integrable and has integral one by (75) . Dominated convergence therefore gives
lim r → ∞ Q ( r ) t ( r ) = M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 d z = 1 , \lim_{r\to\infty} \frac{Q(r)}{t(r)} = M_\Omega\int_{\mathbb{R}^n} |\gamma_\Omega(z)|^2\, dz = 1, r → ∞ lim t ( r ) Q ( r ) = M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 d z = 1 ,
which is (90) and the proof is complete. □
Finite-rank projections and the multifrequency constraint
Let P P P be an orthogonal projection of rank k k k in G G G , and define
M r ( P ) : = Tr ( P A r ) . (92) \mathcal{M}_r(P) := \operatorname{Tr}(P A_r). \tag*{(92)} M r ( P ) := Tr ( P A r ) . (92)
If { u j } j = 1 k \{u_j\}_{j=1}^k { u j } j = 1 k is an orthonormal basis of Ran ( P ) \operatorname{Ran}(P) Ran ( P ) , then
M r ( P ) = ∑ j = 1 k ∫ B r ∣ u j ^ ( ξ ) ∣ 2 d ξ . (93) \mathcal{M}_r(P) = \sum_{j=1}^k \int_{B_r} |\widehat{u_j}(\xi)|^2\, d\xi. \tag*{(93)} M r ( P ) = j = 1 ∑ k ∫ B r ∣ u j ( ξ ) ∣ 2 d ξ . (93)
Theorem 5.7 (Ky Fan bound) With
B k ( r ) = ∑ ℓ = 1 k μ ℓ ( r ) , (94) B_k(r) = \sum_{\ell=1}^k \mu_\ell(r), \tag*{(94)} B k ( r ) = ℓ = 1 ∑ k μ ℓ ( r ) , (94) one has
M r ( P ) ≤ B k ( r ) , sup rank P = k M r ( P ) = B k ( r ) . (95) \mathcal{M}_r(P) \le B_k(r), \qquad\sup_{\operatorname{rank} P=k} \mathcal{M}_r(P) = B_k(r). \tag*{(95)} M r ( P ) ≤ B k ( r ) , rank P = k sup M r ( P ) = B k ( r ) . (95) Furthermore, we have
B k ( r ) < min { k , t ( r ) } , B k ( r ) ≤ k Q ( r ) . (96) B_k(r) < \min\{k,t(r)\}, \qquad B_k(r) \le\sqrt{kQ(r)}. \tag*{(96)} B k ( r ) < min { k , t ( r )} , B k ( r ) ≤ k Q ( r ) . (96)
Proof For brevity write μ ℓ = μ ℓ ( r ) \mu_\ell= \mu_\ell(r) μ ℓ = μ ℓ ( r ) . By Theorem 5.3 , A r A_r A r has an orthonormal eigenbasis { v ℓ } ℓ ≥ 1 \{v_\ell\}_{\ell\ge1} { v ℓ } ℓ ≥ 1 with
A r v ℓ = μ ℓ v ℓ , 1 > μ 1 ≥ μ 2 ≥ ⋯ > 0. A_r v_\ell= \mu_\ell v_\ell, \qquad1 > \mu_1 \ge\mu_2 \ge\cdots> 0. A r v ℓ = μ ℓ v ℓ , 1 > μ 1 ≥ μ 2 ≥ ⋯ > 0.
Since A r A_r A r is trace class and P P P is bounded, P A r P A_r P A r is trace class, so its trace may be computed in this basis. Put
a ℓ = ⟨ P v ℓ , v ℓ ⟩ = ∥ P v ℓ ∥ 2 . a_\ell= \langle P v_\ell, v_\ell\rangle= \|P v_\ell\|^2. a ℓ = ⟨ P v ℓ , v ℓ ⟩ = ∥ P v ℓ ∥ 2 .
The last equality uses P = P ∗ = P 2 P = P^* = P^2 P = P ∗ = P 2 . Thus 0 ≤ a ℓ ≤ 1 0 \le a_\ell\le1 0 ≤ a ℓ ≤ 1 , and
M r ( P ) = Tr ( P A r ) = ∑ ℓ = 1 ∞ ⟨ P A r v ℓ , v ℓ ⟩ = ∑ ℓ = 1 ∞ μ ℓ ⟨ P v ℓ , v ℓ ⟩ = ∑ ℓ = 1 ∞ μ ℓ a ℓ . \begin{aligned}
\mathcal{M}_r(P) = \operatorname{Tr}(P A_r) = \sum_{\ell=1}^{\infty} \langle P A_r v_\ell, v_\ell\rangle\\
&= \sum_{\ell=1}^{\infty} \mu_{\ell}\langle P v_{\ell},v_{\ell}\rangle= \sum_{\ell=1}^{\infty} \mu_{\ell}a_{\ell}.
\end{aligned} M r ( P ) = Tr ( P A r ) = ℓ = 1 ∑ ∞ ⟨ P A r v ℓ , v ℓ ⟩ = ℓ = 1 ∑ ∞ μ ℓ ⟨ P v ℓ , v ℓ ⟩ = ℓ = 1 ∑ ∞ μ ℓ a ℓ .
We also record why the coefficients have total mass k k k . If { u j } j = 1 k \{u_j\}_{j=1}^{k} { u j } j = 1 k is an orthonormal basis of Ran ( P ) \operatorname{Ran}(P) Ran ( P ) , then
P v ℓ = ∑ j = 1 k ⟨ v ℓ , u j ⟩ u j . P v_{\ell} = \sum_{j=1}^{k}\langle v_{\ell},u_j\rangle u_j. P v ℓ = j = 1 ∑ k ⟨ v ℓ , u j ⟩ u j .
Parseval’s identity in the eigenbasis { v ℓ } \{v_{\ell}\} { v ℓ } therefore gives
∑ ℓ = 1 ∞ a ℓ = ∑ ℓ = 1 ∞ ∑ j = 1 k ∣ ⟨ v ℓ , u j ⟩ ∣ 2 = ∑ j = 1 k ∥ u j ∥ 2 = k . \sum_{\ell=1}^{\infty}a_{\ell}=\sum_{\ell=1}^{\infty}\sum_{j=1}^{k}\lvert\langle v_{\ell},u_j\rangle\rvert^{2}=\sum_{j=1}^{k}\lVert u_j\rVert^{2}=k. ℓ = 1 ∑ ∞ a ℓ = ℓ = 1 ∑ ∞ j = 1 ∑ k ∣⟨ v ℓ , u j ⟩ ∣ 2 = j = 1 ∑ k ∥ u j ∥ 2 = k .
The maximizing arrangement places this unit-bounded mass on the largest k k k eigenvalues. Here is a direct verification. From ∑ ℓ > k a ℓ = k − ∑ ℓ = 1 k a ℓ = ∑ ℓ = 1 k ( 1 − a ℓ ) \sum_{\ell>k}a_{\ell}=k-\sum_{\ell=1}^{k}a_{\ell}=\sum_{\ell=1}^{k}(1-a_{\ell}) ∑ ℓ > k a ℓ = k − ∑ ℓ = 1 k a ℓ = ∑ ℓ = 1 k ( 1 − a ℓ ) and μ ℓ ≤ μ k \mu_{\ell}\leq\mu_k μ ℓ ≤ μ k for ℓ > k \ell>k ℓ > k , we obtain
M r ( P ) = ∑ ℓ = 1 k μ ℓ a ℓ + ∑ ℓ > k μ ℓ a ℓ ≤ ∑ ℓ = 1 k μ ℓ a ℓ + μ k ∑ ℓ = 1 k ( 1 − a ℓ ) = ∑ ℓ = 1 k [ μ ℓ a ℓ + μ k ( 1 − a ℓ ) ] ≤ ∑ ℓ = 1 k μ ℓ = B k ( r ) . \begin{aligned}
\mathcal{M}_r(P)&=\sum_{\ell=1}^{k}\mu_{\ell}a_{\ell}+\sum_{\ell>k}\mu_{\ell}a_{\ell}\\
&\leq\sum_{\ell=1}^{k}\mu_{\ell}a_{\ell}+\mu_k\sum_{\ell=1}^{k}(1-a_{\ell})\\
&=\sum_{\ell=1}^{k}\left[\mu_{\ell}a_{\ell}+\mu_k(1-a_{\ell})\right]\\
&\leq\sum_{\ell=1}^{k}\mu_{\ell}=B_k(r).
\end{aligned} M r ( P ) = ℓ = 1 ∑ k μ ℓ a ℓ + ℓ > k ∑ μ ℓ a ℓ ≤ ℓ = 1 ∑ k μ ℓ a ℓ + μ k ℓ = 1 ∑ k ( 1 − a ℓ ) = ℓ = 1 ∑ k [ μ ℓ a ℓ + μ k ( 1 − a ℓ ) ] ≤ ℓ = 1 ∑ k μ ℓ = B k ( r ) .
Let P k s p P_k^{\mathrm{sp}} P k sp be the orthogonal projection onto span { v 1 , … , v k } \operatorname{span}\{v_1,\ldots,v_k\} span { v 1 , … , v k } . For this projection, a ℓ = 1 a_{\ell}=1 a ℓ = 1 for ℓ ≤ k \ell\leq k ℓ ≤ k and a ℓ = 0 a_{\ell}=0 a ℓ = 0 for ℓ > k \ell>k ℓ > k , so M r ( P k s p ) = B k ( r ) \mathcal{M}_r(P_k^{\mathrm{sp}})=B_k(r) M r ( P k sp ) = B k ( r ) . This proves both assertions in (95) , including attainment of the supremum.
It remains to prove (96) . Since every μ ℓ < 1 \mu_{\ell}<1 μ ℓ < 1 ,
B k ( r ) = ∑ ℓ = 1 k μ ℓ < k . B_k(r)=\sum_{\ell=1}^{k}\mu_{\ell}<k. B k ( r ) = ℓ = 1 ∑ k μ ℓ < k .
Moreover, Theorem 5.3 gives infinitely many strictly positive eigenvalues, and (85) gives t ( r ) = ∑ ℓ = 1 ∞ μ ℓ t(r)=\sum_{\ell=1}^{\infty}\mu_{\ell} t ( r ) = ∑ ℓ = 1 ∞ μ ℓ . Hence
t ( r ) − B k ( r ) = ∑ ℓ > k μ ℓ > 0 , t(r)-B_k(r)=\sum_{\ell>k}\mu_{\ell}>0, t ( r ) − B k ( r ) = ℓ > k ∑ μ ℓ > 0 ,
which proves B k ( r ) < t ( r ) B_k(r)<t(r) B k ( r ) < t ( r ) . Finally, Cauchy–Schwarz and (86) yield
B k ( r ) 2 ≤ k ∑ ℓ = 1 k μ ℓ 2 ≤ k ∑ ℓ = 1 ∞ μ ℓ 2 = k Q ( r ) . B_k(r)^2\leq k\sum_{\ell=1}^{k}\mu_{\ell}^{2}\leq k\sum_{\ell=1}^{\infty}\mu_{\ell}^{2}=kQ(r). B k ( r ) 2 ≤ k ℓ = 1 ∑ k μ ℓ 2 ≤ k ℓ = 1 ∑ ∞ μ ℓ 2 = k Q ( r ) .
Taking square roots completes the proof. □
Let P k P_k P k be the projection onto span { ϕ 1 , … , ϕ k } \operatorname{span}\{\phi_1,\ldots,\phi_k\} span { ϕ 1 , … , ϕ k } . By (13) , (93) , (95) and (96) , we have
∫ B r F k ( ξ ) d ξ ≤ B k ( r ) ≤ k Q ( r ) . (97) \int_{B_r}F_k(\xi)\,d\xi\leq B_k(r)\leq\sqrt{kQ(r)}. \tag*{(97)} ∫ B r F k ( ξ ) d ξ ≤ B k ( r ) ≤ k Q ( r ) . (97)
This collective constraint is absent from the pointwise envelopes in (43) and (44) . To convert the resulting mass leakage into a moment estimate, we need a two-region version of the bathtub principle.
Lemma 5.8 (Minimal moment at prescribed outside mass) Let M 0 , R > 0 M_0, R > 0 M 0 , R > 0 , set k = M 0 ω n R n k = M_0\omega_n R^n k = M 0 ω n R n , and let f : R n → [ 0 , M 0 ] f : \mathbb{R}^n \to[0,M_0] f : R n → [ 0 , M 0 ] satisfy ∫ R n f = k \int_{\mathbb{R}^n} f = k ∫ R n f = k . If
m = ∫ R n ∖ B R f ( ξ ) d ξ , τ = m k ∈ [ 0 , 1 ] , m = \int_{\mathbb{R}^n\setminus B_R} f(\xi)\,\mathrm{d}\xi,\qquad\tau= \frac{m}{k} \in[0,1], m = ∫ R n ∖ B R f ( ξ ) d ξ , τ = k m ∈ [ 0 , 1 ] , then, with q = 1 + 2 / n q = 1 + 2/n q = 1 + 2/ n , we have
∫ R n ∣ ξ ∣ 2 f ( ξ ) d ξ ≥ n n + 2 k R 2 [ ( 1 − τ ) q + ( 1 + τ ) q − 1 ] . (98) \int_{\mathbb{R}^n} |\xi|^2 f(\xi)\,\mathrm{d}\xi\ge\frac{n}{n+2}kR^2\left[(1-\tau)^q+(1+\tau)^q-1\right]. \tag*{(98)} ∫ R n ∣ ξ ∣ 2 f ( ξ ) d ξ ≥ n + 2 n k R 2 [ ( 1 − τ ) q + ( 1 + τ ) q − 1 ] . (98) The right-hand side is optimal under these constraints.
Proof. If ∫ R n ∣ ξ ∣ 2 f ( ξ ) d ξ = + ∞ \int_{\mathbb{R}^n}|\xi|^2f(\xi)\,\mathrm{d}\xi=+\infty ∫ R n ∣ ξ ∣ 2 f ( ξ ) d ξ = + ∞ , the estimate is immediate, so we may assume that the second moment is finite. Since f ≥ 0 f\ge0 f ≥ 0 and its total mass is k k k , one has 0 ≤ m ≤ k 0\le m\le k 0 ≤ m ≤ k and hence 0 ≤ τ ≤ 1 0\le\tau\le1 0 ≤ τ ≤ 1 , as stated.
The prescribed masses in the two regions are
∫ B R f ( ξ ) d ξ = k − m = k ( 1 − τ ) , (99) \int_{B_R}f(\xi)\,\mathrm{d}\xi=k-m=k(1-\tau), \tag*{(99)} ∫ B R f ( ξ ) d ξ = k − m = k ( 1 − τ ) , (99)
∫ R n ∖ B R f ( ξ ) d ξ = m = k τ . (100) \int_{\mathbb{R}^n\setminus B_R}f(\xi)\,\mathrm{d}\xi=m=k\tau. \tag*{(100)} ∫ R n ∖ B R f ( ξ ) d ξ = m = k τ . (100)
Set
ρ = R ( 1 − τ ) 1 / n , σ = R ( 1 + τ ) 1 / n . \rho=R(1-\tau)^{1/n},\qquad\sigma=R(1+\tau)^{1/n}. ρ = R ( 1 − τ ) 1/ n , σ = R ( 1 + τ ) 1/ n .
Thus 0 ≤ ρ ≤ R ≤ σ ≤ 2 1 / n R 0\le\rho\le R\le\sigma\le2^{1/n}R 0 ≤ ρ ≤ R ≤ σ ≤ 2 1/ n R . The two radii are chosen so that a density saturated at the cap M 0 M_0 M 0 fills exactly the required inner and outer masses. Since k = M 0 ω n R n k=M_0\omega_nR^n k = M 0 ω n R n , these definitions give
M 0 ∣ B ρ ∣ = M 0 ω n R n ( 1 − τ ) = k ( 1 − τ ) , M_0|B_\rho|=M_0\omega_nR^n(1-\tau)=k(1-\tau), M 0 ∣ B ρ ∣ = M 0 ω n R n ( 1 − τ ) = k ( 1 − τ ) ,
M 0 ∣ B σ ∖ B R ∣ = M 0 ω n ( σ n − R n ) = k τ . M_0|B_\sigma\setminus B_R|=M_0\omega_n(\sigma^n-R^n)=k\tau. M 0 ∣ B σ ∖ B R ∣ = M 0 ω n ( σ n − R n ) = k τ .
Thus M 0 1 B ρ M_0\mathbf{1}_{B_\rho} M 0 1 B ρ has exactly the same mass as f f f in B R B_R B R , and M 0 1 B σ ∖ B R M_0\mathbf{1}_{B_\sigma\setminus B_R} M 0 1 B σ ∖ B R has exactly the same mass as f f f in R n ∖ B R \mathbb{R}^n\setminus B_R R n ∖ B R .
The least expensive inner configuration is M 0 1 B ρ M_0\mathbf{1}_{B_\rho} M 0 1 B ρ , while the least expensive outer configuration is M 0 1 B σ ∖ B R M_0\mathbf{1}_{B_\sigma\setminus B_R} M 0 1 B σ ∖ B R . To see the first assertion, observe that on B ρ B_\rho B ρ both factors in
( ∣ ξ ∣ 2 − ρ 2 ) ( f − M 0 1 B ρ ) ≥ 0 , (|\xi|^2-\rho^2)(f-M_0\mathbf{1}_{B_\rho})\ge0, ( ∣ ξ ∣ 2 − ρ 2 ) ( f − M 0 1 B ρ ) ≥ 0 ,
are nonpositive, whereas on B R ∖ B ρ B_R\setminus B_\rho B R ∖ B ρ they are both nonnegative. Hence the displayed inequality holds throughout B R B_R B R . To display the cancellation rather than hide it in the rearrangement argument, write
∫ B R ∣ ξ ∣ 2 ( f − M 0 1 B ρ ) d ξ = ∫ B R ( ∣ ξ ∣ 2 − ρ 2 ) ( f − M 0 1 B ρ ) d ξ + ρ 2 ( ∫ B R f d ξ − M 0 ∣ B ρ ∣ ) . \begin{aligned}
\int_{B_R}|\xi|^2(f-M_0\mathbf{1}_{B_\rho})\,\mathrm{d}\xi
={}&\int_{B_R}(|\xi|^2-\rho^2)(f-M_0\mathbf{1}_{B_\rho})\,\mathrm{d}\xi\\
&+\rho^2\left(\int_{B_R}f\,\mathrm{d}\xi-M_0|B_\rho|\right).
\end{aligned} ∫ B R ∣ ξ ∣ 2 ( f − M 0 1 B ρ ) d ξ = ∫ B R ( ∣ ξ ∣ 2 − ρ 2 ) ( f − M 0 1 B ρ ) d ξ + ρ 2 ( ∫ B R f d ξ − M 0 ∣ B ρ ∣ ) .
The first term is nonnegative by the preceding pointwise sign argument, whereas the term in parentheses vanishes by (99) . It follows that
∫ B R ∣ ξ ∣ 2 f ( ξ ) d ξ ≥ M 0 ∫ B ρ ∣ ξ ∣ 2 d ξ . (101) \int_{B_R}|\xi|^2f(\xi)\,\mathrm{d}\xi\ge M_0\int_{B_\rho}|\xi|^2\,\mathrm{d}\xi. \tag*{(101)} ∫ B R ∣ ξ ∣ 2 f ( ξ ) d ξ ≥ M 0 ∫ B ρ ∣ ξ ∣ 2 d ξ . (101)
Similarly, on B σ ∖ B R B_\sigma\setminus B_R B σ ∖ B R both factors in
( ∣ ξ ∣ 2 − σ 2 ) ( f − M 0 1 B σ ∖ B R ) ≥ 0 , \left(|\xi|^2-\sigma^2\right)\left(f-M_0\mathbf{1}_{B_\sigma\setminus B_R}\right)\ge0, ( ∣ ξ ∣ 2 − σ 2 ) ( f − M 0 1 B σ ∖ B R ) ≥ 0 ,
are nonpositive, while on R n ∖ B σ \mathbb{R}^n\setminus B_\sigma R n ∖ B σ both are nonnegative. The corresponding expanded identity is
∫ R n ∖ B R ∣ ξ ∣ 2 ( f − M 0 1 B σ ∖ B R ) d ξ = ∫ R n ∖ B R ( ∣ ξ ∣ 2 − σ 2 ) ( f − M 0 1 B σ ∖ B R ) d ξ + σ 2 ( ∫ R n ∖ B R f d ξ − M 0 ∣ B σ ∖ B R ∣ ) . \begin{aligned}
\int_{\mathbb{R}^n\setminus B_R}|\xi|^2\left(f-M_0\mathbf{1}_{B_\sigma\setminus B_R}\right)\,d\xi
={}&\int_{\mathbb{R}^n\setminus B_R}\left(|\xi|^2-\sigma^2\right)\left(f-M_0\mathbf{1}_{B_\sigma\setminus B_R}\right)\,d\xi\\
&+\sigma^2\left(\int_{\mathbb{R}^n\setminus B_R}f\,d\xi-M_0|B_\sigma\setminus B_R|\right).
\end{aligned} ∫ R n ∖ B R ∣ ξ ∣ 2 ( f − M 0 1 B σ ∖ B R ) d ξ = ∫ R n ∖ B R ( ∣ ξ ∣ 2 − σ 2 ) ( f − M 0 1 B σ ∖ B R ) d ξ + σ 2 ( ∫ R n ∖ B R f d ξ − M 0 ∣ B σ ∖ B R ∣ ) .
The first term is nonnegative by the sign argument and the second term vanishes by (100) . Therefore
∫ R n ∖ B R ∣ ξ ∣ 2 f ( ξ ) d ξ ≥ M 0 ∫ B σ ∖ B R ∣ ξ ∣ 2 d ξ . (102) \int_{\mathbb{R}^n\setminus B_R}|\xi|^2f(\xi)\,d\xi\ge M_0\int_{B_\sigma\setminus B_R}|\xi|^2\,d\xi. \tag*{(102)} ∫ R n ∖ B R ∣ ξ ∣ 2 f ( ξ ) d ξ ≥ M 0 ∫ B σ ∖ B R ∣ ξ ∣ 2 d ξ . (102)
For completeness, polar coordinates give, for every a ≥ 0 a\ge0 a ≥ 0 ,
∫ B a ∣ ξ ∣ 2 d ξ = ∣ S n − 1 ∣ ∫ 0 a r n + 1 d r = n ω n a n + 2 n + 2 . (103) \int_{B_a}|\xi|^2\,d\xi=|\mathbb{S}^{n-1}|\int_0^a r^{n+1}\,dr=n\omega_n\frac{a^{n+2}}{n+2}. \tag*{(103)} ∫ B a ∣ ξ ∣ 2 d ξ = ∣ S n − 1 ∣ ∫ 0 a r n + 1 d r = n ω n n + 2 a n + 2 . (103)
Adding (101) and (102) and then using (103) , we obtain
∫ R n ∣ ξ ∣ 2 f ( ξ ) d ξ ≥ M 0 n ω n n + 2 ( ρ n + 2 + σ n + 2 − R n + 2 ) = M 0 n ω n R n + 2 n + 2 [ ( 1 − τ ) ( n + 2 ) / n + ( 1 + τ ) ( n + 2 ) / n − 1 ] = n n + 2 k R 2 [ ( 1 − τ ) q + ( 1 + τ ) q − 1 ] , \begin{aligned}
\int_{\mathbb{R}^n}|\xi|^2f(\xi)\,d\xi
&\ge\frac{M_0n\omega_n}{n+2}\left(\rho^{n+2}+\sigma^{n+2}-R^{n+2}\right)\\
&=\frac{M_0n\omega_nR^{n+2}}{n+2}\left[(1-\tau)^{(n+2)/n}+(1+\tau)^{(n+2)/n}-1\right]\\
&=\frac{n}{n+2}kR^2\left[(1-\tau)^q+(1+\tau)^q-1\right],
\end{aligned} ∫ R n ∣ ξ ∣ 2 f ( ξ ) d ξ ≥ n + 2 M 0 n ω n ( ρ n + 2 + σ n + 2 − R n + 2 ) = n + 2 M 0 n ω n R n + 2 [ ( 1 − τ ) ( n + 2 ) / n + ( 1 + τ ) ( n + 2 ) / n − 1 ] = n + 2 n k R 2 [ ( 1 − τ ) q + ( 1 + τ ) q − 1 ] ,
where the last line uses M 0 ω n R n = k M_0\omega_nR^n=k M 0 ω n R n = k and q = ( n + 2 ) / n = 1 + 2 / n q=(n+2)/n=1+2/n q = ( n + 2 ) / n = 1 + 2/ n . This is (98) .
Finally,
f ∗ ( ξ ) = M 0 1 B ρ ( ξ ) + M 0 1 B σ ∖ B R ( ξ ) f_*(\xi)=M_0\mathbf{1}_{B_\rho}(\xi)+M_0\mathbf{1}_{B_\sigma\setminus B_R}(\xi) f ∗ ( ξ ) = M 0 1 B ρ ( ξ ) + M 0 1 B σ ∖ B R ( ξ )
takes values in [ 0 , M 0 ] [0,M_0] [ 0 , M 0 ] , has total mass k k k and outside mass m m m , and makes both regional comparisons equalities. It therefore attains the right-hand side of (98) , proving optimality rather than merely a lower estimate. The construction also covers the endpoints without a limiting argument: if τ = 0 \tau=0 τ = 0 , then ρ = σ = R \rho=\sigma=R ρ = σ = R and the outer annulus is empty; if τ = 1 \tau=1 τ = 1 , then ρ = 0 \rho=0 ρ = 0 and all mass is placed in B 2 1 / n R ∖ B R B_{2^{1/n}R}\setminus B_R B 2 1/ n R ∖ B R . □ \square □
Strict Li–Yau Sum Improvements
We now convert the multifrequency leakage into an eigenvalue-sum estimate. We derive the exact spectral and squared-trace multipliers, prove their quantitative behavior in k k k , and record the resulting explicit linear correction.
Define the Li–Yau filling radius by
R k = ( k M Ω ω n ) 1 / n , t ( R k ) = k . (104) R_k=\left(\frac{k}{M_\Omega\omega_n}\right)^{1/n},\qquad t(R_k)=k. \tag*{(104)} R k = ( M Ω ω n k ) 1/ n , t ( R k ) = k . (104)
and set
β k = B k ( R k ) k , θ k = ( Q ( R k ) k ) 1 / 2 . (105) \beta_k=\frac{B_k(R_k)}{k},\qquad\theta_k=\left(\frac{Q(R_k)}{k}\right)^{1/2}. \tag*{(105)} β k = k B k ( R k ) , θ k = ( k Q ( R k ) ) 1/2 . (105)
By (87) and (96) ,
0 < β k ≤ θ k < 1. (106) 0<\beta_k\leq\theta_k<1. \tag*{(106)} 0 < β k ≤ θ k < 1. (106)
At r = R k r=R_k r = R k , (97) therefore implies
1 k ∫ R n ∖ B R k F k ( ξ ) d ξ ≥ 1 − β k ≥ 1 − θ k > 0. (107) \frac{1}{k}\int_{\mathbb{R}^n\setminus B_{R_k}}F_k(\xi)\,\mathrm{d}\xi\geq1-\beta_k\geq1-\theta_k>0. \tag*{(107)} k 1 ∫ R n ∖ B R k F k ( ξ ) d ξ ≥ 1 − β k ≥ 1 − θ k > 0. (107)
Theorem 6.1 (Exact and computable strict improvements) Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n , n ≥ 2 n\geq2 n ≥ 2 , be a nonempty bounded open set. For every k ≥ 1 k\geq1 k ≥ 1 , put q = 1 + 2 / n q=1+2/n q = 1 + 2/ n and define
C k s p ( Ω ) = β k q + ( 2 − β k ) q − 1 , (108) C_k^{\mathrm{sp}}(\Omega)=\beta_k^q+(2-\beta_k)^q-1, \tag*{(108)} C k sp ( Ω ) = β k q + ( 2 − β k ) q − 1 , (108) C k ( 2 ) ( Ω ) = θ k q + ( 2 − θ k ) q − 1. (109) C_k^{(2)}(\Omega)=\theta_k^q+(2-\theta_k)^q-1. \tag*{(109)} C k ( 2 ) ( Ω ) = θ k q + ( 2 − θ k ) q − 1. (109) Then
∑ j = 1 k λ j ≥ C k s p ( Ω ) L k L Y ≥ C k ( 2 ) ( Ω ) L k L Y > L k L Y , (110) \sum_{j=1}^{k}\lambda_j\geq C_k^{\mathrm{sp}}(\Omega)L_k^{\mathrm{LY}}\geq C_k^{(2)}(\Omega)L_k^{\mathrm{LY}}>L_k^{\mathrm{LY}}, \tag*{(110)} j = 1 ∑ k λ j ≥ C k sp ( Ω ) L k LY ≥ C k ( 2 ) ( Ω ) L k LY > L k LY , (110) where, as in the introduction,
L k L Y = n n + 2 C n V − 2 / n k 1 + 2 / n . (111) L_k^{\mathrm{LY}}=\frac{n}{n+2}C_nV^{-2/n}k^{1+2/n}. \tag*{(111)} L k LY = n + 2 n C n V − 2/ n k 1 + 2/ n . (111) For fixed Ω \Omega Ω ,
C k s p ( Ω ) ⟶ 1 , C k ( 2 ) ( Ω ) ⟶ 1 ( k → ∞ ) . (112) C_k^{\mathrm{sp}}(\Omega)\longrightarrow1,\qquad C_k^{(2)}(\Omega)\longrightarrow1\qquad(k\to\infty). \tag*{(112)} C k sp ( Ω ) ⟶ 1 , C k ( 2 ) ( Ω ) ⟶ 1 ( k → ∞ ) . (112)
Proof. Apply Lemma 5.8 to f = F k f=F_k f = F k , M 0 = M Ω M_0=M_\Omega M 0 = M Ω , R = R k R=R_k R = R k . The function
H n ( τ ) = ( 1 − τ ) q + ( 1 + τ ) q − 1 H_n(\tau)=(1-\tau)^q+(1+\tau)^q-1 H n ( τ ) = ( 1 − τ ) q + ( 1 + τ ) q − 1
is strictly increasing on ( 0 , 1 ) (0,1) ( 0 , 1 ) , because
H n ′ ( τ ) = q [ ( 1 + τ ) q − 1 − ( 1 − τ ) q − 1 ] > 0. H_n'(\tau)=q\left[(1+\tau)^{q-1}-(1-\tau)^{q-1}\right]>0. H n ′ ( τ ) = q [ ( 1 + τ ) q − 1 − ( 1 − τ ) q − 1 ] > 0.
In view of (107) , we have τ ≥ 1 − β k ≥ 1 − θ k \tau\geq1-\beta_k\geq1-\theta_k τ ≥ 1 − β k ≥ 1 − θ k and hence
∑ j = 1 k λ j ≥ H n ( 1 − β k ) L k L Y ≥ H n ( 1 − θ k ) L k L Y , \sum_{j=1}^{k}\lambda_j\geq H_n(1-\beta_k)L_k^{\mathrm{LY}}\geq H_n(1-\theta_k)L_k^{\mathrm{LY}}, j = 1 ∑ k λ j ≥ H n ( 1 − β k ) L k LY ≥ H n ( 1 − θ k ) L k LY ,
which gives the first two inequalities in (110) . Since 0 < θ k < 1 0<\theta_k<1 0 < θ k < 1 and q > 1 q>1 q > 1 , strict convexity yields
θ k q + ( 2 − θ k ) q > 2. \theta_k^q+(2-\theta_k)^q>2. θ k q + ( 2 − θ k ) q > 2.
Thus C k ( 2 ) ( Ω ) > 1 C_k^{(2)}(\Omega)>1 C k ( 2 ) ( Ω ) > 1 , proving that the Li–Yau bound is strictly improved for every finite k k k . It remains to verify the limits. Since R k → ∞ R_k\to\infty R k → ∞ as k → ∞ k\to\infty k → ∞ , (90) gives θ k 2 = Q ( R k ) / t ( R k ) → 1 \theta_k^2=Q(R_k)/t(R_k)\to1 θ k 2 = Q ( R k ) / t ( R k ) → 1 . To treat β k \beta_k β k , we use the fractional version of the weight-filling argument from the proof of Theorem 5.7 . If μ 1 ≥ μ 2 ≥ ⋯ ≥ 0 \mu_1\geq\mu_2\geq\cdots\geq0 μ 1 ≥ μ 2 ≥ ⋯ ≥ 0 and a sequence ( b ℓ ) (b_\ell) ( b ℓ ) satisfies
0 ≤ b ℓ ≤ 1 , ∑ ℓ = 1 ∞ b ℓ = k , (113) 0\leq b_\ell\leq1,\qquad\sum_{\ell=1}^{\infty}b_\ell=k, \tag*{(113)} 0 ≤ b ℓ ≤ 1 , ℓ = 1 ∑ ∞ b ℓ = k , (113)
then the total available weight k k k is maximally paired with the k k k largest values of the decreasing sequence ( μ ℓ ) (\mu_\ell) ( μ ℓ ) . We give the full calculation. First define the amount of weight placed beyond the first k k k indices by
δ : = ∑ ℓ > k b ℓ . (114) \delta:= \sum_{\ell>k} b_\ell. \tag*{(114)} δ := ℓ > k ∑ b ℓ . (114)
The mass constraint (113) on ( b ℓ ) (b_\ell) ( b ℓ ) implies
δ = k − ∑ ℓ = 1 k b ℓ = ∑ ℓ = 1 k ( 1 − b ℓ ) . \delta= k-\sum_{\ell=1}^{k}b_\ell=\sum_{\ell=1}^{k}(1-b_\ell). δ = k − ℓ = 1 ∑ k b ℓ = ℓ = 1 ∑ k ( 1 − b ℓ ) .
Thus δ \delta δ is also exactly the unfilled capacity among the first k k k indices. Since (114) and μ ℓ ≤ μ k \mu_\ell\le\mu_k μ ℓ ≤ μ k for ℓ > k \ell>k ℓ > k , the tail contribution is bounded by
∑ ℓ > k μ ℓ b ℓ ≤ μ k ∑ ℓ > k b ℓ = μ k δ . (115) \sum_{\ell>k}\mu_\ell b_\ell\le\mu_k\sum_{\ell>k}b_\ell=\mu_k\delta. \tag*{(115)} ℓ > k ∑ μ ℓ b ℓ ≤ μ k ℓ > k ∑ b ℓ = μ k δ . (115)
On the other hand, μ ℓ ≥ μ k \mu_\ell\ge\mu_k μ ℓ ≥ μ k for 1 ≤ ℓ ≤ k 1\le\ell\le k 1 ≤ ℓ ≤ k , so the unfilled part of the first k k k positions satisfies
∑ ℓ = 1 k μ ℓ ( 1 − b ℓ ) ≥ μ k ∑ ℓ = 1 k ( 1 − b ℓ ) = μ k δ . (116) \sum_{\ell=1}^{k}\mu_\ell(1-b_\ell)\ge\mu_k\sum_{\ell=1}^{k}(1-b_\ell)=\mu_k\delta. \tag*{(116)} ℓ = 1 ∑ k μ ℓ ( 1 − b ℓ ) ≥ μ k ℓ = 1 ∑ k ( 1 − b ℓ ) = μ k δ . (116)
Combining the last two estimates (115) and (116) gives
∑ ℓ > k μ ℓ b ℓ ≤ ∑ ℓ = 1 k μ ℓ ( 1 − b ℓ ) . (117) \sum_{\ell>k}\mu_\ell b_\ell\le\sum_{\ell=1}^{k}\mu_\ell(1-b_\ell). \tag*{(117)} ℓ > k ∑ μ ℓ b ℓ ≤ ℓ = 1 ∑ k μ ℓ ( 1 − b ℓ ) . (117)
Equivalently, from (115) , (116) and (117) , subtracting the weighted sum from the completely filled first k k k positions yields the explicitly nonnegative deficit
∑ ℓ = 1 k μ ℓ − ∑ ℓ = 1 ∞ μ ℓ b ℓ = ∑ ℓ = 1 k μ ℓ ( 1 − b ℓ ) − ∑ ℓ > k μ ℓ b ℓ ≥ μ k δ − μ k δ = 0. \sum_{\ell=1}^{k}\mu_\ell-\sum_{\ell=1}^{\infty}\mu_\ell b_\ell
=\sum_{\ell=1}^{k}\mu_\ell(1-b_\ell)-\sum_{\ell>k}\mu_\ell b_\ell
\ge\mu_k\delta-\mu_k\delta=0. ℓ = 1 ∑ k μ ℓ − ℓ = 1 ∑ ∞ μ ℓ b ℓ = ℓ = 1 ∑ k μ ℓ ( 1 − b ℓ ) − ℓ > k ∑ μ ℓ b ℓ ≥ μ k δ − μ k δ = 0.
This proves
∑ ℓ = 1 ∞ μ ℓ b ℓ ≤ ∑ ℓ = 1 k μ ℓ . (118) \sum_{\ell=1}^{\infty}\mu_\ell b_\ell\le\sum_{\ell=1}^{k}\mu_\ell. \tag*{(118)} ℓ = 1 ∑ ∞ μ ℓ b ℓ ≤ ℓ = 1 ∑ k μ ℓ . (118)
All sums in this argument are finite or absolutely convergent: indeed, 0 ≤ ∑ ℓ μ ℓ b ℓ ≤ μ 1 ∑ ℓ b ℓ = k μ 1 0\le\sum_\ell\mu_\ell b_\ell\le\mu_1\sum_\ell b_\ell=k\mu_1 0 ≤ ∑ ℓ μ ℓ b ℓ ≤ μ 1 ∑ ℓ b ℓ = k μ 1 . This also justifies splitting the series at the index k k k .
At r = R k r=R_k r = R k , the first-trace identity and (104) give
0 < μ ℓ ( R k ) < 1 , ∑ ℓ = 1 ∞ μ ℓ ( R k ) = t ( R k ) = k . 0<\mu_\ell(R_k)<1,\qquad\sum_{\ell=1}^{\infty}\mu_\ell(R_k)=t(R_k)=k. 0 < μ ℓ ( R k ) < 1 , ℓ = 1 ∑ ∞ μ ℓ ( R k ) = t ( R k ) = k .
Thus b ℓ : = μ ℓ ( R k ) b_\ell:=\mu_\ell(R_k) b ℓ := μ ℓ ( R k ) is an admissible fractional weight sequence in (118) : the pointwise constraint 0 < b ℓ < 1 0<b_\ell<1 0 < b ℓ < 1 follows from Theorem 5.3 , and the preceding display verifies its total mass. This is only a choice of test weights; it does not assert that b ℓ = ∥ P v ℓ ∥ 2 b_\ell=\lVert P v_\ell\rVert^2 b ℓ = ∥ P v ℓ ∥ 2 for some rank-k k k projection. With this choice, the left- and right-hand sides of (118) become, respectively,
∑ ℓ = 1 ∞ μ ℓ ( R k ) b ℓ = ∑ ℓ = 1 ∞ μ ℓ ( R k ) 2 = Q ( R k ) \sum_{\ell=1}^{\infty}\mu_\ell(R_k)b_\ell
=\sum_{\ell=1}^{\infty}\mu_\ell(R_k)^2=Q(R_k) ℓ = 1 ∑ ∞ μ ℓ ( R k ) b ℓ = ℓ = 1 ∑ ∞ μ ℓ ( R k ) 2 = Q ( R k )
and
∑ ℓ = 1 k μ ℓ ( R k ) = B k ( R k ) . \sum_{\ell=1}^{k} \mu_{\ell}(R_k)=B_k(R_k). ℓ = 1 ∑ k μ ℓ ( R k ) = B k ( R k ) .
Therefore (118) gives Q ( R k ) ≤ B k ( R k ) Q(R_k) \le B_k(R_k) Q ( R k ) ≤ B k ( R k ) . Dividing by k = t ( R k ) k=t(R_k) k = t ( R k ) and using (105) yields
θ k 2 = Q ( R k ) k ≤ B k ( R k ) k = β k . \theta_k^2=\frac{Q(R_k)}{k}\le\frac{B_k(R_k)}{k}=\beta_k. θ k 2 = k Q ( R k ) ≤ k B k ( R k ) = β k .
Together with the already proved bound β k ≤ θ k \beta_k\le\theta_k β k ≤ θ k , this gives θ k 2 ≤ β k ≤ θ k \theta_k^2\le\beta_k\le\theta_k θ k 2 ≤ β k ≤ θ k . Since θ k → 1 \theta_k\to1 θ k → 1 , the squeeze theorem now shows β k → 1 \beta_k\to1 β k → 1 as well.
The squared-trace version is explicit in terms of the Fourier transform of the domain. Indeed, (91) and (105) gives
θ k 2 = M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 ∣ B R k ∩ ( B R k − z ) ∣ ∣ B R k ∣ d z . (119) \theta_k^2=M_\Omega\int_{\mathbb{R}^n}|\gamma_\Omega(z)|^2\frac{|B_{R_k}\cap(B_{R_k}-z)|}{|B_{R_k}|}\,dz. \tag*{(119)} θ k 2 = M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 ∣ B R k ∣ ∣ B R k ∩ ( B R k − z ) ∣ d z . (119)
The overlap vanishes for ∣ z ∣ ≥ 2 R k |z|\ge2R_k ∣ z ∣ ≥ 2 R k , so only a bounded frequency region occurs in this formula.
The overlap representation contains additional information about the dependence on k k k . The next proposition makes this dependence explicit.
Proposition 6.2 (Monotonicity and a quantitative excess) Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n , n ≥ 2 n\ge2 n ≥ 2 , be a nonempty bounded open set. Then, we have
0 < θ k < θ k + 1 < 1 , C k + 1 ( 2 ) ( Ω ) < C k ( 2 ) ( Ω ) . (120) 0<\theta_k<\theta_{k+1}<1,\qquad C_{k+1}^{(2)}(\Omega)<C_k^{(2)}(\Omega). \tag*{(120)} 0 < θ k < θ k + 1 < 1 , C k + 1 ( 2 ) ( Ω ) < C k ( 2 ) ( Ω ) . (120) Moreover, for every k ≥ 1 k\ge1 k ≥ 1 ,
C k ( 2 ) ( Ω ) − 1 ≥ κ n k − 2 / n , (121) C_k^{(2)}(\Omega)-1\ge\kappa_n k^{-2/n}, \tag*{(121)} C k ( 2 ) ( Ω ) − 1 ≥ κ n k − 2/ n , (121) where
κ n = n + 2 2 ( n + 1 ) 2 ( ω n − 1 ω n ) 2 ( 3 4 ) n − 1 . (122) \kappa_n=\frac{n+2}{2(n+1)^2}\left(\frac{\omega_{n-1}}{\omega_n}\right)^2\left(\frac{3}{4}\right)^{n-1}. \tag*{(122)} κ n = 2 ( n + 1 ) 2 n + 2 ( ω n ω n − 1 ) 2 ( 4 3 ) n − 1 . (122) In particular, C k ( 2 ) ( Ω ) C_k^{(2)}(\Omega) C k ( 2 ) ( Ω ) has the following asymptotic property
C k ( 2 ) ( Ω ) ↓ 1. (123) C_k^{(2)}(\Omega)\downarrow1. \tag*{(123)} C k ( 2 ) ( Ω ) ↓ 1. (123)
Proof. Rescaling the configuration in Figure 1 by 1 / r 1/r 1/ r gives two unit balls with center separation s = ∣ z ∣ / r s=|z|/r s = ∣ z ∣/ r . By rotational invariance, we may take them to be B 1 B_1 B 1 and B 1 − s e 1 B_1-se_1 B 1 − s e 1 . The normalized overlap volume is then given by
Φ n ( s ) = ∣ B 1 ∩ ( B 1 − s e 1 ) ∣ ω n = 2 ω n − 1 ω n ∫ s / 2 1 ( 1 − u 2 ) ( n − 1 ) / 2 d u , 0 ≤ s ≤ 2 , (124) \Phi_n(s)=\frac{|B_1\cap(B_1-se_1)|}{\omega_n}=\frac{2\omega_{n-1}}{\omega_n}\int_{s/2}^{1}(1-u^2)^{(n-1)/2}\,du,\qquad0\le s\le2, \tag*{(124)} Φ n ( s ) = ω n ∣ B 1 ∩ ( B 1 − s e 1 ) ∣ = ω n 2 ω n − 1 ∫ s /2 1 ( 1 − u 2 ) ( n − 1 ) /2 d u , 0 ≤ s ≤ 2 , (124)
and set Φ n ( s ) = 0 \Phi_n(s)=0 Φ n ( s ) = 0 for s ≥ 2 s\ge2 s ≥ 2 . Thus
∣ B r ∩ ( B r − z ) ∣ ∣ B r ∣ = Φ n ( ∣ z ∣ r ) . \frac{|B_r\cap(B_r-z)|}{|B_r|}=\Phi_n\left(\frac{|z|}{r}\right). ∣ B r ∣ ∣ B r ∩ ( B r − z ) ∣ = Φ n ( r ∣ z ∣ ) .
The function Φ n \Phi_n Φ n is strictly decreasing on ( 0 , 2 ) (0,2) ( 0 , 2 ) . Consequently, the function
q ( r ) : = Q ( r ) t ( r ) = M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 Φ n ( ∣ z ∣ r ) d z (125) q(r):=\frac{Q(r)}{t(r)}=M_\Omega\int_{\mathbb{R}^n}|\gamma_\Omega(z)|^2\Phi_n\left(\frac{|z|}{r}\right)\,dz \tag*{(125)} q ( r ) := t ( r ) Q ( r ) = M Ω ∫ R n ∣ γ Ω ( z ) ∣ 2 Φ n ( r ∣ z ∣ ) d z (125)
is nondecreasing in r r r . The increase is strict: since γ Ω ( 0 ) = 1 \gamma_{\Omega}(0)=1 γ Ω ( 0 ) = 1 and γ Ω \gamma_{\Omega} γ Ω is continuous, its modulus is bounded away from zero in a neighborhood of the origin, where the overlap profile increases strictly when r r r increases. Since R k + 1 > R k R_{k+1}>R_k R k + 1 > R k and θ k 2 = q ( R k ) \theta_k^2=q(R_k) θ k 2 = q ( R k ) , this proves the first inequality in (120) .
Now put
ρ Ω ( z ) = M Ω ∣ γ Ω ( z ) ∣ 2 . (126) \rho_{\Omega}(z)=M_{\Omega}|\gamma_{\Omega}(z)|^2. \tag*{(126)} ρ Ω ( z ) = M Ω ∣ γ Ω ( z ) ∣ 2 . (126)
By (75) and ∣ γ Ω ∣ ≤ 1 |\gamma_{\Omega}|\le1 ∣ γ Ω ∣ ≤ 1 ,
0 ≤ ρ Ω ≤ M Ω , ∫ R n ρ Ω ( z ) d z = 1. (127) 0\le\rho_{\Omega}\le M_{\Omega},\qquad\int_{\mathbb{R}^n}\rho_{\Omega}(z)\,dz=1. \tag*{(127)} 0 ≤ ρ Ω ≤ M Ω , ∫ R n ρ Ω ( z ) d z = 1. (127)
The radial cost
z ⟼ 1 − Φ n ( ∣ z ∣ R k ) z\longmapsto1-\Phi_n\left(\frac{|z|}{R_k}\right) z ⟼ 1 − Φ n ( R k ∣ z ∣ )
is nondecreasing in ∣ z ∣ |z| ∣ z ∣ . The bathtub principle applied to (127) therefore shows that its integral is minimized by M Ω 1 B R 1 M_{\Omega}\mathbf{1}_{B_{R_1}} M Ω 1 B R 1 , where M Ω ω n R 1 n = 1 M_{\Omega}\omega_nR_1^n=1 M Ω ω n R 1 n = 1 . Since R k = k 1 / n R 1 R_k=k^{1/n}R_1 R k = k 1/ n R 1 , we obtain the domain-independent estimate
1 − θ k 2 ≥ d n , k : = n ∫ 0 1 [ 1 − Φ n ( s k 1 / n ) ] s n − 1 d s . (128) 1-\theta_k^2\ge d_{n,k}:=n\int_0^1\left[1-\Phi_n\left(\frac{s}{k^{1/n}}\right)\right]s^{n-1}\,ds. \tag*{(128)} 1 − θ k 2 ≥ d n , k := n ∫ 0 1 [ 1 − Φ n ( k 1/ n s ) ] s n − 1 d s . (128)
For 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , (124) gives
1 − Φ n ( t ) = 2 ω n − 1 ω n ∫ 0 t / 2 ( 1 − u 2 ) ( n − 1 ) / 2 d u ≥ ω n − 1 ω n ( 3 4 ) ( n − 1 ) / 2 t . (129) \begin{aligned}
1-\Phi_n(t)&=\frac{2\omega_{n-1}}{\omega_n}\int_0^{t/2}(1-u^2)^{(n-1)/2}\,du\\
&\ge\frac{\omega_{n-1}}{\omega_n}\left(\frac{3}{4}\right)^{(n-1)/2}t. \tag*{(129)}
\end{aligned} 1 − Φ n ( t ) = ω n 2 ω n − 1 ∫ 0 t /2 ( 1 − u 2 ) ( n − 1 ) /2 d u ≥ ω n ω n − 1 ( 4 3 ) ( n − 1 ) /2 t . (129)
Since s / k 1 / n ≤ 1 s/k^{1/n}\le1 s / k 1/ n ≤ 1 , inserting this estimate into (128) yields
1 − θ k 2 ≥ A n k − 1 / n , A n = n n + 1 ω n − 1 ω n ( 3 4 ) ( n − 1 ) / 2 . (130) 1-\theta_k^2\ge A_nk^{-1/n},\qquad A_n=\frac{n}{n+1}\frac{\omega_{n-1}}{\omega_n}\left(\frac{3}{4}\right)^{(n-1)/2}. \tag*{(130)} 1 − θ k 2 ≥ A n k − 1/ n , A n = n + 1 n ω n ω n − 1 ( 4 3 ) ( n − 1 ) /2 . (130)
It remains to pass from the trace deficit to the multiplier. Let δ = 1 − θ k \delta=1-\theta_k δ = 1 − θ k and recall q = 1 + 2 / n ∈ ( 1 , 2 ] q=1+2/n\in(1,2] q = 1 + 2/ n ∈ ( 1 , 2 ] . Convexity gives
( 1 − δ ) q + ( 1 + δ ) q − 2 ≥ q ( q − 1 ) δ 2 , 0 ≤ δ ≤ 1. (131) (1-\delta)^q+(1+\delta)^q-2\ge q(q-1)\delta^2,\qquad0\le\delta\le1. \tag*{(131)} ( 1 − δ ) q + ( 1 + δ ) q − 2 ≥ q ( q − 1 ) δ 2 , 0 ≤ δ ≤ 1. (131)
Indeed, the second derivative of the left-hand side is
q ( q − 1 ) [ ( 1 − δ ) q − 2 + ( 1 + δ ) q − 2 ] ≥ 2 q ( q − 1 ) . q(q-1)\left[(1-\delta)^{q-2}+(1+\delta)^{q-2}\right]\ge2q(q-1). q ( q − 1 ) [ ( 1 − δ ) q − 2 + ( 1 + δ ) q − 2 ] ≥ 2 q ( q − 1 ) .
For 0 ≤ δ < 1 0\le\delta<1 0 ≤ δ < 1 , the last inequality follows from AM–GM because ( 1 − δ ) q − 2 ≥ 1 (1-\delta)^{q-2}\ge1 ( 1 − δ ) q − 2 ≥ 1 when q ≤ 2 q\le2 q ≤ 2 ; the endpoint follows by continuity (or by a one-sided limit when q < 2 q<2 q < 2 ). Since
1 − θ k = 1 − θ k 2 1 + θ k ≥ 1 2 ( 1 − θ k 2 ) , 1-\theta_k=\frac{1-\theta_k^2}{1+\theta_k}\ge\frac{1}{2}(1-\theta_k^2), 1 − θ k = 1 + θ k 1 − θ k 2 ≥ 2 1 ( 1 − θ k 2 ) ,
(130) and (131) imply
c k ( 2 ) ( Ω ) − 1 ≥ q ( q − 1 ) 4 A n 2 k − 2 / n = κ n k − 2 / n . c_k^{(2)}(\Omega)-1\ge\frac{q(q-1)}{4}A_n^2k^{-2/n}=\kappa_nk^{-2/n}. c k ( 2 ) ( Ω ) − 1 ≥ 4 q ( q − 1 ) A n 2 k − 2/ n = κ n k − 2/ n .
Finally, the function
G ( s ) = s q + ( 2 − s ) q − 1 G(s)=s^q+(2-s)^q-1 G ( s ) = s q + ( 2 − s ) q − 1
satisfies G ′ ( s ) < 0 G'(s)<0 G ′ ( s ) < 0 on ( 0 , 1 ) (0,1) ( 0 , 1 ) . The strict increase of θ k \theta_k θ k therefore proves the second assertion in (120) ; its limit, already established in (112) , proves (123) . □ \square □
Combining the quantitative excess with (110) gives a completely explicit correction to the Berezin–Li–Yau sum.
Corollary 6.3 (A universal explicit sum correction) Under the assumptions of Proposition 6.2 ,
∑ j = 1 k λ j ≥ ( 1 + κ n k − 2 / n ) L k L Y \sum_{j=1}^{k} \lambda_j \ge\left(1+\kappa_n k^{-2/n}\right)L_k^{\mathrm{LY}} j = 1 ∑ k λ j ≥ ( 1 + κ n k − 2/ n ) L k LY = L k L Y + n n + 2 κ n C n V − 2 / n k . (132) = L_k^{\mathrm{LY}}+\frac{n}{n+2}\kappa_n C_n V^{-2/n}k. \tag*{(132)} = L k LY + n + 2 n κ n C n V − 2/ n k . (132)
Remark 6.4 (The exact spectral coefficient) The same conclusion is not currently available for C k s p C_k^{\mathrm{sp}} C k sp . Since G G G is strictly decreasing, monotonic decrease of C k s p C_k^{\mathrm{sp}} C k sp is equivalent to β k + 1 ≥ β k \beta_{k+1}\ge\beta_k β k + 1 ≥ β k , or
B k + 1 ( R k + 1 ) ≥ k + 1 k B k ( R k ) . (133) B_{k+1}(R_{k+1})\ge\frac{k+1}{k}B_k(R_k). \tag*{(133)} B k + 1 ( R k + 1 ) ≥ k k + 1 B k ( R k ) . (133) For fixed k k k , B k ( r ) B_k(r) B k ( r ) increases with r r r , whereas for fixed r r r the averages B k ( r ) / k B_k(r)/k B k ( r ) / k decrease with k k k . These two effects compete along r = R k r=R_k r = R k , and the Ky Fan argument alone does not imply (133) . What is proved without an additional cross-index estimate is
C k s p ( Ω ) > 1 , C k s p ( Ω ) ⟶ 1. C_k^{\mathrm{sp}}(\Omega)>1,\qquad C_k^{\mathrm{sp}}(\Omega)\longrightarrow1. C k sp ( Ω ) > 1 , C k sp ( Ω ) ⟶ 1. Thus no monotonicity claim for the exact spectral coefficient is made.
Corollary 6.5 (The two-dimensional quantitative sum bound) If n = 2 n=2 n = 2 , then
∑ j = 1 k λ j ≥ [ 1 + 2 ( 1 − θ k ) 2 ] 2 π k 2 V ≥ ( 1 + 2 3 π 2 k ) 2 π k 2 V = 2 π k 2 V + 4 k 3 π V . (134) \sum_{j=1}^{k}\lambda_j\ge\left[1+2(1-\theta_k)^2\right]\frac{2\pi k^2}{V}\ge\left(1+\frac{2}{3\pi^2 k}\right)\frac{2\pi k^2}{V}=\frac{2\pi k^2}{V}+\frac{4k}{3\pi V}. \tag*{(134)} j = 1 ∑ k λ j ≥ [ 1 + 2 ( 1 − θ k ) 2 ] V 2 π k 2 ≥ ( 1 + 3 π 2 k 2 ) V 2 π k 2 = V 2 π k 2 + 3 π V 4 k . (134)
Remark 6.6 (Domain assumptions for the quantitative gap) The concentration-operator results are stated for bounded open sets, as in § 5 . Once the leakage inequality has been established, however, the proof of (121) uses only 0 < V < ∞ 0<V<\infty 0 < V < ∞ , Parseval’s identity, ∣ γ Ω ∣ ≤ 1 |\gamma_\Omega|\le1 ∣ γ Ω ∣ ≤ 1 , and the Euclidean ball-overlap profile. Thus no boundary smoothness, Lipschitz regularity, convexity, connectedness, or finite-perimeter assumption enters the quantitative gap estimate itself.
Remark 6.7 (Two distinct improvements) The individual coefficient C n n e w C_n^{\mathrm{new}} C n new is universal and remains strictly above the uncoupled envelope coefficient c n e n v c_n^{\mathrm{env}} c n env . It therefore remains strictly above n / ( n + 2 ) n/(n+2) n / ( n + 2 ) as k → ∞ k\to\infty k → ∞ . By contrast, C k ( 2 ) ( Ω ) C_k^{(2)}(\Omega) C k ( 2 ) ( Ω ) improves the sharp Berezin–Li–Yau sum coefficient for each finite k k k , decreases strictly to one, and has the universal excess quantified in (121) . The two statements therefore improve different losses in the classical argument and should not be numerically identified.
Proposition 6.8 (Rigid-motion and scaling behavior) Under the assumptions of Theorem 6.1 , the coefficients C k s p C_k^{\mathrm{sp}} C k sp and C k ( 2 ) C_k^{(2)} C k ( 2 ) are invariant under translations and uniform scalings of the domain.
Proof. If Ω ′ = a Ω + x 0 \Omega^\prime=a\Omega+x_0 Ω ′ = a Ω + x 0 , a > 0 a>0 a > 0 , then γ Ω ′ ( z ) = e i x 0 ⋅ z γ Ω ( a z ) \gamma_{\Omega^\prime}(z)=e^{i x_0\cdot z}\gamma_\Omega(az) γ Ω ′ ( z ) = e i x 0 ⋅ z γ Ω ( a z ) and R k ( Ω ′ ) = a − 1 R k ( Ω ) R_k(\Omega^\prime)=a^{-1}R_k(\Omega) R k ( Ω ′ ) = a − 1 R k ( Ω ) . A change of variables in (89) proves the assertion for C k ( 2 ) C_k^{(2)} C k ( 2 ) ; the concentration operators are unitarily equivalent after rescaling, which proves it for c k s p c_k^{\mathrm{sp}} c k sp .
Conclusion and Outlook
This paper establishes two complementary refinements of the Berezin–Li–Yau Fourier method. For individual eigenvalues, the exact commutator identity in Proposition 3.2 and the coupled scalar–vector projection inequality in Proposition 3.3 lead to Theorem 4.1 . The comparison in Proposition 4.2 and Table 1 makes explicit the two successive gains: the scalar energy envelope first improves the Li–Yau coefficient, and the coupled constraint then improves that envelope by retaining compatibility between the scalar Fourier mass and its projected gradient.
For eigenvalue sums on bounded open sets, the trace identities and Ky Fan estimate in Theorems 5.5 and 5.7 impose a collective frequency constraint on the first k k k eigenfunctions. Together with the leakage estimate, these yield the strict finite-index multiplier in Theorem 6.1 ; Proposition 6.2 quantifies its universal excess, and Corollary 6.3 converts that excess into a volume-only linear correction. This part requires no boundary smoothness, convexity, connectedness, or moment-of-inertia information.
The limitations are equally clear. The individual coefficient obtained from the relaxed coupled constraints is strictly below one, whereas the sum multiplier decreases to one; neither result therefore proves Pólya’s conjecture. Three structures retained by the analysis but not yet exploited quantitatively are the nonnegative commutator slack B k B_k B k , the residual vector field ρ k \rho_k ρ k in (72) , and the concentration spectrum beyond its first two traces. Lower bounds for the first two quantities, or higher-trace and nonradial estimates for the third, are natural next steps. To alter the Pólya coefficient, such refinements would need signed information at the boundary scale or a mechanism that survives the passage from averaged spectral bounds to individual eigenvalues.
Statements and Declarations
Funding. This work was supported by the National Key Research and Development Program of China (2025YFA1016600, 2025YFA1016601), the National Natural Science Foundation of China (12331015), and the National Center for Mathematics and Interdisciplinary Science, CAS.
Competing interests. The authors declare that there is no conflict of interest.
Data availability. This article is theoretical and does not use external datasets.