Classicality for the alternating map
We use Castillo’s classification and dual reconstruction for the alternating map [1 Theorem 5.1(ii) and Section 8] . The dual pullback transfers orthogonality, concentration forms and operator symmetries from the input functional. For related polynomial mappings developed by Petronilho and his students, see [5 Theorems 2.1 and 4.2] . The measure formula there assumes compact input support.
Let h ∈ C ∖ { 0 } h \in\mathbb{C} \setminus\{0\} h ∈ C ∖ { 0 } and U ⊆ C U \subseteq\mathbb{C} U ⊆ C satisfy
U ± h 2 ⊆ U , X ( s ) = e π i s / h ( s ∈ U ) , # X ( U ) = ∞ . U \pm\frac{h}{2} \subseteq U,\qquad X(s)=e^{\pi i s/h}\ (s \in U),\qquad\#X(U)=\infty. U ± 2 h ⊆ U , X ( s ) = e π i s / h ( s ∈ U ) , # X ( U ) = ∞.
For example, U = h Q U=h\mathbb{Q} U = h Q has infinite image. Let P \mathcal{P} P be the space of polynomials with complex coefficients and P n = { p ∈ P : deg p ≤ n } \mathcal{P}_n=\{p \in\mathcal{P}:\deg p \le n\} P n = { p ∈ P : deg p ≤ n } , with P − 1 = { 0 } \mathcal{P}_{-1}=\{0\} P − 1 = { 0 } . Write P ′ \mathcal{P}' P ′ for its linear dual.
Since X ( s ± h / 2 ) = ± i X ( s ) X(s \pm h/2)=\pm iX(s) X ( s ± h /2 ) = ± i X ( s ) , the divided-difference and averaging operators are
( D p ) ( x ) = p ( i x ) − p ( − i x ) 2 i x , ( S p ) ( x ) = p ( i x ) + p ( − i x ) 2 . (Dp)(x)=\frac{p(ix)-p(-ix)}{2ix},\qquad(Sp)(x)=\frac{p(ix)+p(-ix)}{2}. ( D p ) ( x ) = 2 i x p ( i x ) − p ( − i x ) , ( S p ) ( x ) = 2 p ( i x ) + p ( − i x ) .
The first quotient extends polynomially through x = 0 x=0 x = 0 . In Maroni’s dual notation [10 Section 1.1] , the transpose of an operator A A A and multiplication of a functional are defined by
⟨ A ′ u , p ⟩ = ⟨ u , A p ⟩ , ⟨ q u , p ⟩ = ⟨ u , q p ⟩ . \langle A'u,p\rangle=\langle u,Ap\rangle,\qquad\langle qu,p\rangle=\langle u,qp\rangle. ⟨ A ′ u , p ⟩ = ⟨ u , A p ⟩ , ⟨ q u , p ⟩ = ⟨ u , q p ⟩ .
For a ∈ P ′ a \in\mathcal{P}' a ∈ P ′ , define J a : P → P ′ J_a:\mathcal{P}\to\mathcal{P}' J a : P → P ′ by
J a f = f a , ⟨ J a f , g ⟩ = ⟨ a , f g ⟩ . J_af=fa,\qquad\langle J_af,g\rangle=\langle a,fg\rangle. J a f = f a , ⟨ J a f , g ⟩ = ⟨ a , f g ⟩ .
An operator A A A is symmetric for a a a when J a A = A ′ J a J_aA=A'J_a J a A = A ′ J a . All transposes refer to this bilinear pairing. A functional is regular when the form is nondegenerate on each P n \mathcal{P}_n P n ; equivalently, it has a monic orthogonal sequence with nonzero squared norms in every degree. A regular functional is classical for this alternating map when there exist polynomials ϕ , ψ \phi,\psi ϕ , ψ , not both zero, satisfying [2 Definition 1.2]
− D ′ ( ϕ u ) = S ′ ( ψ u ) , deg ϕ ≤ 2 , deg ψ ≤ 1. -D'(\phi u)=S'(\psi u),\qquad\deg\phi\le2,\qquad\deg\psi\le1. − D ′ ( ϕ u ) = S ′ ( ψ u ) , deg ϕ ≤ 2 , deg ψ ≤ 1.
2020 Mathematics Subject Classification. 42C05, 33C45.
Key words and phrases. Alternating map, half-step-invariant sets, admissible maps, polynomial mappings, time-and-band limiting.
For real τ \tau τ , we use ϕ = 0 \phi=0 ϕ = 0 , ψ = x − τ \psi=x-\tau ψ = x − τ . The equation becomes S ′ ( ( x − τ ) u ) = 0 S'((x-\tau)u)=0 S ′ (( x − τ ) u ) = 0 .
Let ν \nu ν be a finite positive Borel measure on ( 0 , b ) (0,b) ( 0 , b ) , 0 < b ≤ ∞ 0<b\le\infty 0 < b ≤ ∞ , with infinite support and all moments finite, and let ( R n ) n ≥ 0 (R_n)_{n\geq0} ( R n ) n ≥ 0 be the monic real orthogonal polynomials of
⟨ v , p ⟩ = ∫ ( 0 , b ) p ( t ) d ν ( t ) . \langle\mathbf v,p\rangle=\int_{(0,b)}p(t)\,d\nu(t). ⟨ v , p ⟩ = ∫ ( 0 , b ) p ( t ) d ν ( t ) .
Set, for n ≥ 0 n\geq0 n ≥ 0 ,
S n ( t ) = R n + 1 ( t ) − ρ n R n ( t ) t , ρ n = R n + 1 ( 0 ) R n ( 0 ) , S_n(t)=\frac{R_{n+1}(t)-\rho_nR_n(t)}{t},\qquad\rho_n=\frac{R_{n+1}(0)}{R_n(0)}, S n ( t ) = t R n + 1 ( t ) − ρ n R n ( t ) , ρ n = R n ( 0 ) R n + 1 ( 0 ) ,
h n = ⟨ v , R n 2 ⟩ (1) h_n=\langle\mathbf v,R_n^2\rangle \tag*{(1)} h n = ⟨ v , R n 2 ⟩ (1)
R ^ n = R n h n , S ^ n = S n − ρ n h n . \widehat R_n=\frac{R_n}{\sqrt{h_n}},\qquad\widehat S_n=\frac{S_n}{\sqrt{-\rho_nh_n}}. R n = h n R n , S n = − ρ n h n S n .
Here R n ( 0 ) ≠ 0 R_n(0)\neq0 R n ( 0 ) = 0 : otherwise R n ( t ) = t Q ( t ) R_n(t)=tQ(t) R n ( t ) = tQ ( t ) and 0 = ⟨ v , R n Q ⟩ = ∫ ( 0 , b ) t Q ( t ) 2 d ν ( t ) > 0 0=\langle\mathbf v,R_nQ\rangle=\int_{(0,b)}tQ(t)^2\,d\nu(t)>0 0 = ⟨ v , R n Q ⟩ = ∫ ( 0 , b ) tQ ( t ) 2 d ν ( t ) > 0 . The sequence ( S n ) (S_n) ( S n ) is orthogonal for x v x\mathbf v x v , with squared norms − ρ n h n > 0 -\rho_nh_n>0 − ρ n h n > 0 .
For τ ∈ R \tau\in\mathbb R τ ∈ R , put ⟨ v τ , p ⟩ = ⟨ v , p ( x + τ 2 ) ⟩ \langle\mathbf v_\tau,p\rangle=\langle\mathbf v,p(x+\tau^2)\rangle ⟨ v τ , p ⟩ = ⟨ v , p ( x + τ 2 )⟩ . Every polynomial F ( x ) = E ( x 2 ) + x O ( x 2 ) F(x)=E(x^2)+xO(x^2) F ( x ) = E ( x 2 ) + x O ( x 2 ) has the unique form F ( x ) = A ( x 2 ) + ( x − τ ) B ( x 2 ) F(x)=A(x^2)+(x-\tau)B(x^2) F ( x ) = A ( x 2 ) + ( x − τ ) B ( x 2 ) , with B = O B=O B = O and A = E + τ O A=E+\tau O A = E + τ O . Following [1 Theorem 8.6] and [2 Proposition 1.3] we define u τ \mathbf u_\tau u τ by
⟨ u τ , p ( x 2 ) ⟩ = ⟨ v τ , p ⟩ , S ′ ( ( x − τ ) u τ ) = 0. (2) \langle\mathbf u_\tau,p(x^2)\rangle=\langle\mathbf v_\tau,p\rangle,\qquad S'((x-\tau)\mathbf u_\tau)=0. \tag*{(2)} ⟨ u τ , p ( x 2 )⟩ = ⟨ v τ , p ⟩ , S ′ (( x − τ ) u τ ) = 0. (2)
Write y = x 2 − τ 2 y=x^2-\tau^2 y = x 2 − τ 2 and E τ = { x ∈ R : 0 < x 2 − τ 2 < b } E_\tau=\{x\in\mathbb R:0<x^2-\tau^2<b\} E τ = { x ∈ R : 0 < x 2 − τ 2 < b } . Since S ( P ) S(\mathcal P) S ( P ) is the space of even polynomials, the alternating condition in (2) means ⟨ u τ , ( x − τ ) p ( x 2 ) ⟩ = 0 \langle\mathbf u_\tau,(x-\tau)p(x^2)\rangle=0 ⟨ u τ , ( x − τ ) p ( x 2 )⟩ = 0 for every p ∈ P p\in\mathcal P p ∈ P . In particular, ⟨ u τ , x 2 k + 1 ⟩ = τ ⟨ u τ , x 2 k ⟩ \langle\mathbf u_\tau,x^{2k+1}\rangle=\tau\langle\mathbf u_\tau,x^{2k}\rangle ⟨ u τ , x 2 k + 1 ⟩ = τ ⟨ u τ , x 2 k ⟩ for every k ≥ 0 k\geq0 k ≥ 0 , so the even moments determine the odd moments. Let E τ F = f E_\tau F=f E τ F = f and O τ F = g O_\tau F=g O τ F = g in the unique decomposition F ( x ) = f ( y ) + ( x − τ ) g ( y ) F(x)=f(y)+(x-\tau)g(y) F ( x ) = f ( y ) + ( x − τ ) g ( y ) . Then
u τ = E τ ′ v . \mathbf u_\tau=E'_\tau\mathbf v. u τ = E τ ′ v .
This proves uniqueness directly in P ′ \mathcal P' P ′ .
Theorem 1.1. The functional u τ \mathbf u_\tau u τ is positive definite and classical for the alternating map, with structural pair ϕ = 0 \phi=0 ϕ = 0 , ψ = x − τ \psi=x-\tau ψ = x − τ . With r = t + τ 2 r=\sqrt{t+\tau^2} r = t + τ 2 , a representing measure is determined, for bounded Borel F F F , by
∫ E τ F ( x ) d μ τ ( x ) = ∫ ( 0 , b ) ( r + τ 2 r F ( r ) + r − τ 2 r F ( − r ) ) d ν ( t ) . (3) \int_{E_\tau}F(x)\,d\mu_\tau(x)=\int_{(0,b)}\left(\frac{r+\tau}{2r}F(r)+\frac{r-\tau}{2r}F(-r)\right)\,d\nu(t). \tag*{(3)} ∫ E τ F ( x ) d μ τ ( x ) = ∫ ( 0 , b ) ( 2 r r + τ F ( r ) + 2 r r − τ F ( − r ) ) d ν ( t ) . (3) Its orthonormal polynomials are
P ^ 2 n ( τ ) ( x ) = R ^ n ( y ) , (4) \widehat P_{2n}^{(\tau)}(x)=\widehat R_n(y), \tag*{(4)} P 2 n ( τ ) ( x ) = R n ( y ) , (4) P ^ 2 n + 1 ( τ ) ( x ) = ( x − τ ) S ^ n ( y ) . \widehat P_{2n+1}^{(\tau)}(x)=(x-\tau)\widehat S_n(y). P 2 n + 1 ( τ ) ( x ) = ( x − τ ) S n ( y ) . The operator U τ U_\tau U τ defined by
( U τ ( f , g ) ) ( x ) = f ( y ) + ( x − τ ) g ( y ) (5) (U_\tau(f,g))(x)=f(y)+(x-\tau)g(y) \tag*{(5)} ( U τ ( f , g )) ( x ) = f ( y ) + ( x − τ ) g ( y ) (5) is unitary from L 2 ( ( 0 , b ) , d ν ( t ) ) ⊕ L 2 ( ( 0 , b ) , t d ν ( t ) ) L^2((0,b),d\nu(t))\oplus L^2((0,b),t\,d\nu(t)) L 2 (( 0 , b ) , d ν ( t )) ⊕ L 2 (( 0 , b ) , t d ν ( t )) onto L 2 ( E τ , d μ τ ( x ) ) L^2(E_\tau,d\mu_\tau(x)) L 2 ( E τ , d μ τ ( x )) and intertwines multiplication by a bounded measurable function χ \chi χ on ( 0 , b ) (0,b) ( 0 , b ) with multiplication by χ ( y ) \chi(y) χ ( y ) .
For an integer N ≥ 1 N \ge1 N ≥ 1 and measurable χ : ( 0 , b ) → R \chi:(0,b)\to\mathbb{R} χ : ( 0 , b ) → R with 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 , define
⟨ v χ , p ⟩ = ∫ ( 0 , b ) p ( t ) χ ( t ) d ν ( t ) , \langle v_{\chi},p\rangle=\int_{(0,b)}p(t)\chi(t)\,\mathrm{d}\nu(t), ⟨ v χ , p ⟩ = ∫ ( 0 , b ) p ( t ) χ ( t ) d ν ( t ) ,
⟨ u τ , χ , p ⟩ = ∫ E τ p ( x ) χ ( x 2 − τ 2 ) d μ τ ( x ) , \langle u_{\tau,\chi},p\rangle=\int_{E_{\tau}}p(x)\chi(x^{2}-\tau^{2})\,\mathrm{d}\mu_{\tau}(x), ⟨ u τ , χ , p ⟩ = ∫ E τ p ( x ) χ ( x 2 − τ 2 ) d μ τ ( x ) ,
( M τ ( χ ) ) i j = ⟨ u τ , χ , P ^ i ( τ ) P ^ j ( τ ) ⟩ , 0 ≤ i , j < N . (M_{\tau}(\chi))_{ij}=\langle u_{\tau,\chi},\widehat{P}_{i}^{(\tau)}\widehat{P}_{j}^{(\tau)}\rangle,\qquad0\le i,j<N. ( M τ ( χ ) ) ij = ⟨ u τ , χ , P i ( τ ) P j ( τ ) ⟩ , 0 ≤ i , j < N .
Then
M τ ( χ ) = M 0 ( χ ) , Q M τ ( χ ) Q T = G R ( χ ) ⊕ G S ( χ ) , (6) M_{\tau}(\chi)=M_{0}(\chi),\qquad QM_{\tau}(\chi)Q^{\mathsf{T}}=G_{R}(\chi)\oplus G_{S}(\chi), \tag*{(6)} M τ ( χ ) = M 0 ( χ ) , Q M τ ( χ ) Q T = G R ( χ ) ⊕ G S ( χ ) , (6)
where Q Q Q orders the even indices increasingly, followed by the odd indices increasingly. The block sizes are N R = ⌈ N / 2 ⌉ N_{R}=\lceil N/2\rceil N R = ⌈ N /2 ⌉ and N S = ⌊ N / 2 ⌋ N_{S}=\lfloor N/2\rfloor N S = ⌊ N /2 ⌋ , with entries
( G R ) m n = ⟨ v χ , R ^ m R ^ n ⟩ , 0 ≤ m , n < N R , (G_{R})_{mn}=\langle v_{\chi},\widehat{R}_{m}\widehat{R}_{n}\rangle,\qquad0\le m,n<N_{R}, ( G R ) mn = ⟨ v χ , R m R n ⟩ , 0 ≤ m , n < N R ,
( G S ) m n = ⟨ x v χ , S ^ m S ^ n ⟩ , 0 ≤ m , n < N S . (G_{S})_{mn}=\langle xv_{\chi},\widehat{S}_{m}\widehat{S}_{n}\rangle,\qquad0\le m,n<N_{S}. ( G S ) mn = ⟨ x v χ , S m S n ⟩ , 0 ≤ m , n < N S .
We suppress the argument χ \chi χ in G R , G S G_{R},G_{S} G R , G S when the mask is fixed.
Proof. For f ℓ , g ℓ , q ∈ P f_{\ell},g_{\ell},q\in\mathcal{P} f ℓ , g ℓ , q ∈ P , put F ℓ ( x ) = f ℓ ( y ) + ( x − τ ) g ℓ ( y ) F_{\ell}(x)=f_{\ell}(y)+(x-\tau)g_{\ell}(y) F ℓ ( x ) = f ℓ ( y ) + ( x − τ ) g ℓ ( y ) , ℓ = 1 , 2 \ell=1,2 ℓ = 1 , 2 . The alternating condition annihilates ( x − τ ) h ( y ) (x-\tau)h(y) ( x − τ ) h ( y ) for every h ∈ P h\in\mathcal{P} h ∈ P . Using ( x − τ ) 2 = y − 2 τ ( x − τ ) (x-\tau)^{2}=y-2\tau(x-\tau) ( x − τ ) 2 = y − 2 τ ( x − τ ) gives
⟨ u τ , q ( y ) F 1 F 2 ⟩ = ⟨ v , q f 1 f 2 + x q g 1 g 2 ⟩ , \langle u_{\tau},q(y)F_{1}F_{2}\rangle=\langle v,qf_{1}f_{2}+xqg_{1}g_{2}\rangle, ⟨ u τ , q ( y ) F 1 F 2 ⟩ = ⟨ v , q f 1 f 2 + x q g 1 g 2 ⟩ ,
U τ ′ J q ( y ) u τ U τ = J q v ⊕ J x q v . (7) U_{\tau}'J_{q(y)}u_{\tau}U_{\tau}=J_{qv}\oplus J_{xqv}. \tag*{(7)} U τ ′ J q ( y ) u τ U τ = J q v ⊕ J x q v . (7)
Here U τ ′ U_{\tau}' U τ ′ acts from P ′ \mathcal{P}' P ′ to ( P ⊕ P ) ′ = P ′ ⊕ P ′ (\mathcal{P}\oplus\mathcal{P})'=\mathcal{P}'\oplus\mathcal{P}' ( P ⊕ P ) ′ = P ′ ⊕ P ′ . Taking q = 1 q=1 q = 1 proves the polynomial orthogonality in (4). The branch weights in (3) are positive because r > ∣ τ ∣ r>|\tau| r > ∣ τ ∣ , sum to one and have weighted mean τ \tau τ and variance t t t . Hence (3) satisfies (2). All moments are finite since
∫ E τ ∣ x ∣ k d μ τ ( x ) = ∫ ( 0 , b ) ( t + τ 2 ) k / 2 d ν ( t ) . \int_{E_{\tau}}|x|^{k}\,\mathrm{d}\mu_{\tau}(x)=\int_{(0,b)}(t+\tau^{2})^{k/2}\,\mathrm{d}\nu(t). ∫ E τ ∣ x ∣ k d μ τ ( x ) = ∫ ( 0 , b ) ( t + τ 2 ) k /2 d ν ( t ) .
The same branch calculation gives, for measurable complex-valued f , g f,g f , g ,
∫ E τ ∣ f ( y ) + ( x − τ ) g ( y ) ∣ 2 d μ τ ( x ) = ∫ ( 0 , b ) ( ∣ f ( t ) ∣ 2 + t ∣ g ( t ) ∣ 2 ) d ν ( t ) (8) \int_{E_{\tau}}|f(y)+(x-\tau)g(y)|^{2}\,\mathrm{d}\mu_{\tau}(x)=\int_{(0,b)}\left(|f(t)|^{2}+t|g(t)|^{2}\right)\,\mathrm{d}\nu(t) \tag*{(8)} ∫ E τ ∣ f ( y ) + ( x − τ ) g ( y ) ∣ 2 d μ τ ( x ) = ∫ ( 0 , b ) ( ∣ f ( t ) ∣ 2 + t ∣ g ( t ) ∣ 2 ) d ν ( t ) (8)
with both integrals allowed to be infinite. For a nonzero polynomial F F F , at least one of its components f , g f,g f , g is nonzero. Infinite support makes its squared norm strictly positive. Thus u τ u_{\tau} u τ is regular and (2) gives its alternating classicality. (8) proves the isometry. For F ∈ L 2 ( E τ , d μ τ ( x ) ) F\in L^{2}(E_{\tau},\mathrm{d}\mu_{\tau}(x)) F ∈ L 2 ( E τ , d μ τ ( x )) its inverse is given ν \nu ν -almost everywhere by
g ( t ) = F ( r ) − F ( − r ) 2 r , f ( t ) = F ( r ) + F ( − r ) 2 + τ g ( t ) . g(t)=\frac{F(r)-F(-r)}{2r},\qquad f(t)=\frac{F(r)+F(-r)}{2}+\tau g(t). g ( t ) = 2 r F ( r ) − F ( − r ) , f ( t ) = 2 F ( r ) + F ( − r ) + τ g ( t ) .
The positive branch weights make F ( ± r ) F(\pm r) F ( ± r ) well defined ν \nu ν -almost everywhere. Substitution gives U τ ( f , g ) = F U_{\tau}(f,g)=F U τ ( f , g ) = F , and (8) puts both components in the required spaces. For the mask χ \chi χ , the branch measure gives the dual identities
u τ , χ = E τ ′ v χ , u_{\tau,\chi}=E_{\tau}'v_{\chi}, u τ , χ = E τ ′ v χ ,
U τ ′ J u τ , χ U τ = J v χ ⊕ J x v χ . (9) U_{\tau}'J_{u_{\tau,\chi}}U_{\tau}=J_{v_{\chi}}\oplus J_{xv_{\chi}}. \tag*{(9)} U τ ′ J u τ , χ U τ = J v χ ⊕ J x v χ . (9)
These identities hold for the constructed measures, without a moment determinacy assumption. Multiplication by χ ( y ) \chi(y) χ ( y ) acts componentwise. Taking the matrices of the two sides of (9) in the reconstructed orthonormal bases proves (6) .
Put r R = N R − 1 r_R=N_R-1 r R = N R − 1 , r S = N S − 1 r_S=N_S-1 r S = N S − 1 , with empty matrix blocks omitted. The map (5) carries P r R ⊕ P r S \mathcal{P}_{r_R}\oplus\mathcal{P}_{r_S} P r R ⊕ P r S onto P N − 1 \mathcal{P}_{N-1} P N − 1 . Let Π R \Pi_R Π R , Π S \Pi_S Π S , Π τ \Pi_\tau Π τ be the orthogonal projections onto these three spaces in their respective L 2 L^2 L 2 spaces, and let Mult a \operatorname{Mult}_a Mult a denote multiplication by a bounded measurable a a a . Unitarity and the multiplier identity give
U τ − 1 Π τ U τ = Π R ⊕ Π S , U_\tau^{-1}\Pi_\tau U_\tau=\Pi_R\oplus\Pi_S, U τ − 1 Π τ U τ = Π R ⊕ Π S ,
U τ − 1 Π τ Mult χ ( y ) Π τ U τ = ( Π R Mult χ Π R ) ⊕ ( Π S Mult χ Π S ) . U_\tau^{-1}\Pi_\tau\operatorname{Mult}_{\chi(y)}\Pi_\tau U_\tau=(\Pi_R\operatorname{Mult}_\chi\Pi_R)\oplus(\Pi_S\operatorname{Mult}_\chi\Pi_S). U τ − 1 Π τ Mult χ ( y ) Π τ U τ = ( Π R Mult χ Π R ) ⊕ ( Π S Mult χ Π S ) .
For N = 1 N=1 N = 1 , Π S = 0 \Pi_S=0 Π S = 0 . These identities use the measure μ τ \mu_\tau μ τ constructed in (3) ; no determinacy or polynomial density assumption is needed.
By (8) , the concentration problem is
λ max ( M τ ( χ ) ) = max 0 ≠ F ∈ P N − 1 ∫ E τ ∣ F ( x ) ∣ 2 χ ( x 2 − τ 2 ) d μ τ ( x ) ∫ E τ ∣ F ( x ) ∣ 2 d μ τ ( x ) = max 0 ≠ ( f , g ) ∈ P r R ⊕ P r S ∫ ( 0 , b ) χ ( t ) ( ∣ f ( t ) ∣ 2 + t ∣ g ( t ) ∣ 2 ) d ν ( t ) ∫ ( 0 , b ) ( ∣ f ( t ) ∣ 2 + t ∣ g ( t ) ∣ 2 ) d ν ( t ) . (10) \begin{aligned}
\lambda_{\max}(M_\tau(\chi))=\max_{0\ne F\in\mathcal{P}_{N-1}}\frac{\int_{E_\tau}|F(x)|^2\chi(x^2-\tau^2)\,d\mu_\tau(x)}{\int_{E_\tau}|F(x)|^2\,d\mu_\tau(x)} \\
&=\max_{0\ne(f,g)\in\mathcal{P}_{r_R}\oplus\mathcal{P}_{r_S}}\frac{\int_{(0,b)}\chi(t)(|f(t)|^2+t|g(t)|^2)\,d\nu(t)}{\int_{(0,b)}(|f(t)|^2+t|g(t)|^2)\,d\nu(t)}. \tag*{(10)}
\end{aligned} λ m a x ( M τ ( χ )) = 0 = F ∈ P N − 1 max ∫ E τ ∣ F ( x ) ∣ 2 d μ τ ( x ) ∫ E τ ∣ F ( x ) ∣ 2 χ ( x 2 − τ 2 ) d μ τ ( x ) = 0 = ( f , g ) ∈ P r R ⊕ P r S max ∫ ( 0 , b ) ( ∣ f ( t ) ∣ 2 + t ∣ g ( t ) ∣ 2 ) d ν ( t ) ∫ ( 0 , b ) χ ( t ) ( ∣ f ( t ) ∣ 2 + t ∣ g ( t ) ∣ 2 ) d ν ( t ) . (10)
The optimal value is max { λ max ( G R ) , λ max ( G S ) } \max\{\lambda_{\max}(G_R),\lambda_{\max}(G_S)\} max { λ m a x ( G R ) , λ m a x ( G S )} , with empty blocks omitted. Every maximizing pair reconstructs an extremal polynomial for each τ \tau τ , of degree max { 2 deg f , 2 deg g + 1 } \max\{2\deg f,2\deg g+1\} max { 2 deg f , 2 deg g + 1 } , with deg 0 = − ∞ \deg0=-\infty deg 0 = − ∞ .
Theorem 1.2. Let 0 < s < b 0<s<b 0 < s < b and χ = 1 ( 0 , s ) \chi=\mathbf{1}_{(0,s)} χ = 1 ( 0 , s ) . Let m − m_- m − and m + m_+ m + be the numbers of distinct support points of the restrictions of ν \nu ν to ( 0 , s ) (0,s) ( 0 , s ) and [ s , b ) [s,b) [ s , b ) , respectively, allowing the value ∞ \infty ∞ . Every eigenvalue of M τ ( χ ) M_\tau(\chi) M τ ( χ ) strictly between zero and one is simple. The multiplicities at the endpoints are
mult ( 0 ) = N − min ( N R , m − ) − min ( N S , m − ) , \operatorname{mult}(0)=N-\min(N_R,m_-)-\min(N_S,m_-), mult ( 0 ) = N − min ( N R , m − ) − min ( N S , m − ) , mult ( 1 ) = N − min ( N R , m + ) − min ( N S , m + ) , (11) \operatorname{mult}(1)=N-\min(N_R,m_+)-\min(N_S,m_+), \tag*{(11)} mult ( 1 ) = N − min ( N R , m + ) − min ( N S , m + ) , (11) where min ( n , ∞ ) = n \min(n,\infty)=n min ( n , ∞ ) = n . Thus M τ ( χ ) M_\tau(\chi) M τ ( χ ) has simple spectrum precisely when both m − , m + ≥ ⌊ N / 2 ⌋ m_-,m_+\ge\lfloor N/2\rfloor m − , m + ≥ ⌊ N /2 ⌋ . In particular, its spectrum is simple for every N N N if both restricted measures have infinite support.
Proof. Fix 0 < λ < 1 0<\lambda<1 0 < λ < 1 and define
a λ = v χ − λ v , w λ = ( s − x ) a λ . (12) a_\lambda=v_\chi-\lambda v,\qquad w_\lambda=(s-x)a_\lambda. \tag*{(12)} a λ = v χ − λ v , w λ = ( s − x ) a λ . (12)
The measure representing w λ w_\lambda w λ is ( s − t ) ( χ ( t ) − λ ) d ν ( t ) (s-t)(\chi(t)-\lambda)\,d\nu(t) ( s − t ) ( χ ( t ) − λ ) d ν ( t ) . It is positive away from s s s ; removing a possible atom at s s s leaves infinite support. Thus w λ w_\lambda w λ is positive definite. Let F k F_k F k be its monic orthogonal polynomials and put c k = ⟨ a λ , F k ⟩ c_k=\langle a_\lambda,F_k\rangle c k = ⟨ a λ , F k ⟩ .
An eigenvector of G R G_R G R with eigenvalue λ \lambda λ represents a nonzero real polynomial f ∈ P N R − 1 f\in\mathcal{P}_{N_R-1} f ∈ P N R − 1 satisfying ⟨ a λ , f q ⟩ = 0 \langle a_\lambda,fq\rangle=0 ⟨ a λ , f q ⟩ = 0 for q ∈ P N R − 1 q\in\mathcal{P}_{N_R-1} q ∈ P N R − 1 . Testing with q = ( s − x ) h q=(s-x)h q = ( s − x ) h gives ⟨ w λ , f h ⟩ = 0 \langle w_\lambda,fh\rangle=0 ⟨ w λ , f h ⟩ = 0 for h ∈ P N R − 2 h\in\mathcal{P}_{N_R-2} h ∈ P N R − 2 . Positivity forces deg f = N R − 1 \deg f=N_R-1 deg f = N R − 1 , so f f f is a multiple of F N R − 1 F_{N_R-1} F N R − 1 . The same argument for x a λ xa_\lambda x a λ treats G S G_S G S . Each block eigenspace therefore has dimension one.
Suppose N ≥ 2 N\ge2 N ≥ 2 and λ \lambda λ belongs to both blocks. Write n = N S n=N_S n = N S , so N R = n N_R=n N R = n or n + 1 n+1 n + 1 , and let G n − 1 G_{n-1} G n − 1 be monic orthogonal for x w λ xw_\lambda x w λ . Applying (1) to w λ w_\lambda w λ gives
x G n − 1 = F n − ρ F n − 1 , ρ = F n ( 0 ) F n − 1 ( 0 ) ≠ 0. xG_{n-1}=F_n-\rho F_{n-1},\qquad\rho=\frac{F_n(0)}{F_{n-1}(0)}\ne0. x G n − 1 = F n − ρ F n − 1 , ρ = F n − 1 ( 0 ) F n ( 0 ) = 0.
The two eigenconditions, tested with q = 1 q=1 q = 1 , imply c N R − 1 = 0 c_{N_R-1}=0 c N R − 1 = 0 and c n − ρ c n − 1 = 0 c_n-\rho c_{n-1}=0 c n − ρ c n − 1 = 0 . Hence c n − 1 = c n = 0 c_{n-1}=c_n=0 c n − 1 = c n = 0 . Set F − 1 = 0 F_{-1}=0 F − 1 = 0 , C 0 = 0 C_0=0 C 0 = 0 , and write the recurrence for F k F_k F k as
x F k = F k + 1 + B k F k + C k F k − 1 , C k > 0 ( k ≥ 1 ) . xF_k=F_{k+1}+B_kF_k+C_kF_{k-1},\qquad C_k>0\quad(k\geq1). x F k = F k + 1 + B k F k + C k F k − 1 , C k > 0 ( k ≥ 1 ) .
Pairing with a λ a_\lambda a λ gives
c k + 1 = ( s − B k ) c k − C k c k − 1 ( k ≥ 1 ) , c 1 = ( s − B 0 ) c 0 − ⟨ w λ , 1 ⟩ . \begin{aligned}
c_{k+1}&=(s-B_k)c_k-C_kc_{k-1}\quad(k\geq1),\\
c_1&=(s-B_0)c_0-\langle w_\lambda,1\rangle.
\end{aligned} c k + 1 c 1 = ( s − B k ) c k − C k c k − 1 ( k ≥ 1 ) , = ( s − B 0 ) c 0 − ⟨ w λ , 1 ⟩ .
Consecutive zeros propagate backwards to c 0 = c 1 = 0 c_0=c_1=0 c 0 = c 1 = 0 , contradicting ⟨ w λ , 1 ⟩ > 0 \langle w_\lambda,1\rangle>0 ⟨ w λ , 1 ⟩ > 0 . The blocks therefore have no common eigenvalue in ( 0 , 1 ) (0,1) ( 0 , 1 ) . For N = 1 N=1 N = 1 there is only a scalar block.
For a restriction supported at m m m distinct points, its polynomial Gram matrix of order n n n has rank min ( n , m ) \min(n,m) min ( n , m ) : evaluation at the support points has this rank. Infinite support gives rank n n n . Multiplication of the measure by t > 0 t>0 t > 0 does not change its support. Apply these facts to the blocks of M τ M_\tau M τ and I − M τ I-M_\tau I − M τ to obtain (11). Since all remaining eigenvalues are simple, the displayed criterion follows.
Proposition 1.3. For a fixed N ≥ 1 N\geq1 N ≥ 1 , the finite reconstruction and Theorem 1.2 require only a nonzero finite positive measure ν \nu ν on ( 0 , b ) (0,b) ( 0 , b ) with at least N R = ⌈ N / 2 ⌉ N_R=\lceil N/2\rceil N R = ⌈ N /2 ⌉ distinct support points and
∫ ( 0 , b ) t N − 1 d ν ( t ) < ∞ . \int_{(0,b)}t^{N-1}\,d\nu(t)<\infty. ∫ ( 0 , b ) t N − 1 d ν ( t ) < ∞. Define the real orthonormal polynomials R ^ 0 , … , R ^ N R − 1 \widehat{R}_0,\ldots,\widehat{R}_{N_R-1} R 0 , … , R N R − 1 for ν \nu ν and S ^ 0 , … , S ^ N S − 1 \widehat{S}_0,\ldots,\widehat{S}_{N_S-1} S 0 , … , S N S − 1 for t d ν t\,d\nu t d ν , with positive leading coefficients. Then (3), (4), (8) and (6) hold for these finite systems. Both assumptions are necessary for a nondegenerate concentration problem on all of P N − 1 \mathcal{P}_{N-1} P N − 1 .
Proof. The required scalar moments have maximal order max ( 2 N R − 2 , 2 N S − 1 ) = N − 1 \max(2N_R-2,2N_S-1)=N-1 max ( 2 N R − 2 , 2 N S − 1 ) = N − 1 ; for N = 1 N=1 N = 1 omit the second term. The support condition makes both finite polynomial inner products nondegenerate. The branch calculation proves the finite reconstruction. Only the multiplicity argument needs a finite-moment proof. We choose all eigenpolynomials below real.
For a positive measure with at least k k k support points, an interior eigenpolynomial of its size-k k k concentration compression has degree k − 1 k-1 k − 1 . Otherwise, testing with ( s − x ) f (s-x)f ( s − x ) f gives ∫ ( s − t ) ( χ ( t ) − λ ) f ( t ) 2 d ν ( t ) = 0 \int(s-t)(\chi(t)-\lambda)f(t)^2\,d\nu(t)=0 ∫ ( s − t ) ( χ ( t ) − λ ) f ( t ) 2 d ν ( t ) = 0 . It forces f f f to vanish on at least k − 1 k-1 k − 1 points away from s s s , although deg f ≤ k − 2 \deg f\leq k-2 deg f ≤ k − 2 . This also proves simplicity within each scalar block. Two consecutive compressions have no common interior eigenvalue: an eigenvector of the larger one has nonzero last coordinate; the residual of an eigenvector of the smaller one, extended by zero, is supported only in that coordinate. Pairing the two vectors forces the residual to vanish, contradicting the degree assertion.
Put a = v χ − λ v a=v_\chi-\lambda v a = v χ − λ v , understood only on the degrees for which its moments exist. If N R = N S = n N_R=N_S=n N R = N S = n and f , g f,g f , g are common eigenpolynomials, choose c c c so h = f − c g h=f-cg h = f − c g has degree at most n − 2 n-2 n − 2 . For q ∈ P n − 2 q\in\mathcal{P}_{n-2} q ∈ P n − 2 , both eigenconditions give ⟨ x a , h q ⟩ = 0 \langle xa,hq\rangle=0 ⟨ x a , h q ⟩ = 0 . The consecutive t d ν t\,d\nu t d ν compressions force h = 0 h=0 h = 0 . For n = 1 n=1 n = 1 this is immediate. Consequently ⟨ a , f 2 ⟩ = ⟨ x a , f 2 ⟩ = 0 \langle a,f^2\rangle=\langle xa,f^2\rangle=0 ⟨ a , f 2 ⟩ = ⟨ x a , f 2 ⟩ = 0 . Their difference with factor s s s forces f f f to vanish off s s s ; the first equality forces it to vanish at any cutoff atom as well. This contradicts its positive norm.
If N R = n + 1 N_R=n+1 N R = n + 1 , N S = n N_S=n N S = n , choose c c c so h = x g − c f ∈ P n − 1 h=xg-cf\in\mathcal{P}_{n-1} h = xg − c f ∈ P n − 1 . Now ⟨ a , h q ⟩ = 0 \langle a,hq\rangle=0 ⟨ a , h q ⟩ = 0 for q ∈ P n − 1 q\in\mathcal{P}_{n-1} q ∈ P n − 1 . The consecutive ν \nu ν compressions force h = 0 h=0 h = 0 . After rescaling, f = x g f=xg f = xg ; testing its eigencondition with f f f and g g g gives ⟨ x 2 a , g 2 ⟩ = ⟨ x a , g 2 ⟩ = 0 \langle x^2a,g^2\rangle=\langle xa,g^2\rangle=0 ⟨ x 2 a , g 2 ⟩ = ⟨ x a , g 2 ⟩ = 0 . Since t ↦ s χ ( t ) ( t − λ ) < 0 t\mapsto s\chi(t)(t-\lambda)<0 t ↦ sχ ( t ) ( t − λ ) < 0 off s s s , these equalities again contradict the positive norm, including at a cutoff atom. The case N = 1 N=1 N = 1 is scalar. Every product used has degree at most N − 1 N-1 N − 1 ; the endpoint ranks are unchanged.
Finally, M M M input support points give 2 M 2M 2 M reconstructed points, and
∫ E τ x 2 N − 2 d μ τ ( x ) = ∫ ( 0 , b ) ( t + τ 2 ) N − 1 d ν ( t ) . \int_{E_\tau} x^{2N-2}\,d\mu_\tau(x)=\int_{(0,b)}(t+\tau^2)^{N-1}\,d\nu(t). ∫ E τ x 2 N − 2 d μ τ ( x ) = ∫ ( 0 , b ) ( t + τ 2 ) N − 1 d ν ( t ) .
These prove necessity of the support and moment conditions. This finite statement asserts neither an infinite orthogonal sequence nor a functional on all polynomials. □
Remark 1.4. The simplicity proof uses the sign change rather than the sharp indicator itself. It applies whenever each candidate λ \lambda λ admits c λ c_\lambda c λ such that ( c λ − t ) ( χ ( t ) − λ ) d ν ( t ) (c_\lambda-t)(\chi(t)-\lambda)\,d\nu(t) ( c λ − t ) ( χ ( t ) − λ ) d ν ( t ) , possibly with reversed sign, is positive with infinite support. For input measures of infinite support, every bounded strictly monotone χ : ( 0 , b ) → ( 0 , 1 ) \chi:(0,b)\to(0,1) χ : ( 0 , b ) → ( 0 , 1 ) gives simple concentration spectrum for every N N N : choose its sign-change separator c λ c_\lambda c λ and repeat (12) and the consecutive-zero argument. Arbitrary indicators do not have this property. For ν = d t \nu=dt ν = d t on ( 0 , 2 ) (0,2) ( 0 , 2 ) , N = 3 N=3 N = 3 and χ = 1 ( 1 / 2 , 3 / 2 ) \chi=\mathbf{1}_{(1/2,3/2)} χ = 1 ( 1/2 , 3/2 ) , the bases R ^ 0 = 1 / 2 \widehat{R}_0=1/\sqrt{2} R 0 = 1/ 2 , R ^ 1 = 3 / 2 ( t − 1 ) \widehat{R}_1=\sqrt{3/2}(t-1) R 1 = 3/2 ( t − 1 ) and S ^ 0 = 1 / 2 \widehat{S}_0=1/\sqrt{2} S 0 = 1/ 2 give M τ = diag ( 1 / 2 , 1 / 2 , 1 / 8 ) M_\tau=\operatorname{diag}(1/2,1/2,1/8) M τ = diag ( 1/2 , 1/2 , 1/8 ) . For fixed N N N , all moments are unnecessary: d ν ( t ) = t − ( N + 1 ) 1 ( 1 , ∞ ) ( t ) d t d\nu(t)=t^{-(N+1)}\mathbf{1}_{(1,\infty)}(t)\,dt d ν ( t ) = t − ( N + 1 ) 1 ( 1 , ∞ ) ( t ) d t has a finite moment of order N − 1 N-1 N − 1 and an infinite moment of order N N N .
Corollary 1.5. For N ≥ 2 N\ge2 N ≥ 2 , let K R , K S K_R,K_S K R , K S be real symmetric matrices of orders N R , N S N_R,N_S N R , N S , commuting with G R ( χ ) , G S ( χ ) G_R(\chi),G_S(\chi) G R ( χ ) , G S ( χ ) , respectively. Then Q T ( K R ⊕ K S ) Q Q^{\mathsf T}(K_R\oplus K_S)Q Q T ( K R ⊕ K S ) Q commutes with every M τ ( χ ) M_\tau(\chi) M τ ( χ ) . Half-bandwidths m R , m S m_R,m_S m R , m S become at most 2 max ( m R , m S ) 2\max(m_R,m_S) 2 max ( m R , m S ) . Every real symmetric commutant whose nonzero diagonals have even offsets arises this way. If K R , K S K_R,K_S K R , K S have simple spectra, replacing K S K_S K S by K S + c I K_S+cI K S + c I , where
c = 1 + ∥ K R ∥ ∞ + ∥ K S ∥ ∞ , (13) c=1+\lVert K_R\rVert_\infty+\lVert K_S\rVert_\infty, \tag*{(13)} c = 1 + ∥ K R ∥ ∞ + ∥ K S ∥ ∞ , (13) makes the full spectrum simple, with distance at least one between the two resulting block spectra. Here ∥ ⋅ ∥ ∞ \lVert\cdot\rVert_\infty ∥ ⋅ ∥ ∞ is the maximum absolute row sum.
Proof. Use (6). Index differences double, and c − ∥ K S ∥ ∞ > ∥ K R ∥ ∞ c-\lVert K_S\rVert_\infty>\lVert K_R\rVert_\infty c − ∥ K S ∥ ∞ > ∥ K R ∥ ∞ separates the two spectra. Conversely, the stated zero entries make Q L Q T QLQ^{\mathsf T} Q L Q T block diagonal; its commutator with G R ⊕ G S G_R\oplus G_S G R ⊕ G S vanishes precisely when both scalar commutators do. □
Proposition 1.6. Under Corollary 1.5, suppose both scalar matrices have simple spectra. For j = R , S j=R,S j = R , S , put
ζ j ≥ max { 1 , ∥ K j ∥ ∞ } , δ j = gap ( K j ) , \zeta_j\ge\max\{1,\lVert K_j\rVert_\infty\},\qquad\delta_j=\operatorname{gap}(K_j), ζ j ≥ max { 1 , ∥ K j ∥ ∞ } , δ j = gap ( K j ) , where gap is the minimum spacing between consecutive ordered eigenvalues, with value ∞ \infty ∞ for a scalar matrix. Define
K ^ R = ζ R − 1 K R − 2 I , K ^ S = ζ S − 1 K S + 2 I , (14) \widehat{K}_R=\zeta_R^{-1}K_R-2I,\qquad\widehat{K}_S=\zeta_S^{-1}K_S+2I, \tag*{(14)} K R = ζ R − 1 K R − 2 I , K S = ζ S − 1 K S + 2 I , (14) L ^ = Q T ( K ^ R ⊕ K ^ S ) Q . \widehat{L}=Q^{\mathsf T}(\widehat{K}_R\oplus\widehat{K}_S)Q. L = Q T ( K R ⊕ K S ) Q . Then L ^ \widehat{L} L commutes with every M τ ( χ ) M_\tau(\chi) M τ ( χ ) , has simple spectrum and satisfies the bandwidth bound of Corollary 1.5. The distance between its two block spectra is at least two, and
∥ L ^ ∥ 2 ≤ 3 , (15) \lVert\widehat{L}\rVert_2\le3, \tag*{(15)} ∥ L ∥ 2 ≤ 3 , (15) gap ( L ^ ) ≥ min { δ R / ζ R , δ S / ζ S , 2 } . \operatorname{gap}(\widehat{L})\ge\min\{\delta_R/\zeta_R,\delta_S/\zeta_S,2\}. gap ( L ) ≥ min { δ R / ζ R , δ S / ζ S , 2 } . Here ∥ A ∥ 2 = max ∥ z ∥ 2 = 1 ∥ A z ∥ 2 \lVert A\rVert_2=\max_{\lVert z\rVert_2=1}\lVert Az\rVert_2 ∥ A ∥ 2 = max ∥ z ∥ 2 = 1 ∥ A z ∥ 2 . For tridiagonal scalar matrices, L ^ \widehat{L} L has nonzero entries only on the main and second diagonals.
Proof. The absolute row-sum bound gives ∣ λ ∣ ≤ ∥ K j ∥ ∞ |\lambda|\le\lVert K_j\rVert_\infty ∣ λ ∣ ≤ ∥ K j ∥ ∞ for each scalar eigenvalue. Thus the eigenvalues of K ^ R \widehat{K}_R K R lie between − 3 -3 − 3 and − 1 -1 − 1 , and those of K ^ S \widehat{K}_S K S between 1 1 1 and 3 3 3 . The
affine changes preserve commutation and divide the internal spacings by ζ j \zeta_j ζ j . Orthogonal reconstruction preserves spectra and the Euclidean operator norm. Corollary 1.5 gives the commutation and bandwidth assertions. □
This separation does not give a uniform lower bound for the internal spacings as N N N grows. Indeed, gap ( L ^ ) ≤ 6 / ( N − 1 ) \operatorname{gap}(\widehat{L}) \le6/(N-1) gap ( L ) ≤ 6/ ( N − 1 ) for N ≥ 2 N \ge2 N ≥ 2 , since its N N N eigenvalues lie between − 3 -3 − 3 and 3 3 3 . Certified upper bounds may be used for ζ j \zeta_j ζ j ; powers of two permit exact binary rescaling.
Proposition 1.7. Under Proposition 1.6 , suppose the scalar matrices are tridiagonal with nonzero off-diagonal entries. In the block K j K_j K j of order n j ≥ 2 n_j \ge2 n j ≥ 2 , write d j , i d_{j,i} d j , i for the diagonal entries (1 ≤ i ≤ n j 1 \le i \le n_j 1 ≤ i ≤ n j ) and b j , i b_{j,i} b j , i for the links (1 ≤ i < n j 1 \le i < n_j 1 ≤ i < n j ), with b j , 0 = b j , n j = 0 b_{j,0}=b_{j,n_j}=0 b j , 0 = b j , n j = 0 . Put
r j , i = ∣ b j , i − 1 ∣ + ∣ b j , i ∣ , W j = max i ( d j , i + r j , i ) − min i ( d j , i − r j , i ) , g j = 2 ∏ i = 1 n j − 1 ∣ b j , i ∣ ζ j W j n j − 2 . (16) \begin{aligned}
r_{j,i} &= |b_{j,i-1}|+|b_{j,i}|,\\
W_j &= \max_i(d_{j,i}+r_{j,i})-\min_i(d_{j,i}-r_{j,i}),\\
g_j &= \frac{2\prod_{i=1}^{n_j-1}|b_{j,i}|}{\zeta_j W_j^{n_j-2}}. \tag*{(16)}
\end{aligned} r j , i W j g j = ∣ b j , i − 1 ∣ + ∣ b j , i ∣ , = i max ( d j , i + r j , i ) − i min ( d j , i − r j , i ) , = ζ j W j n j − 2 2 ∏ i = 1 n j − 1 ∣ b j , i ∣ . (16) For a scalar block set g j = ∞ g_j=\infty g j = ∞ . Then
gap ( L ^ ) ≥ min { g R , g S , 2 } > 0. (17) \operatorname{gap}(\widehat{L}) \ge\min\{g_R,g_S,2\}>0. \tag*{(17)} gap ( L ) ≥ min { g R , g S , 2 } > 0. (17) If symmetric stored blocks B j B_j B j satisfy ∥ B j − K ^ j ∥ 2 ≤ e j \lVert B_j-\widehat{K}_j\rVert_2\le e_j ∥ B j − K j ∥ 2 ≤ e j , the reconstructed matrix Q T ( B R ⊕ B S ) Q Q^{\mathsf T}(B_R\oplus B_S)Q Q T ( B R ⊕ B S ) Q has gap at least
min { g R − 2 e R , g S − 2 e S , 2 − e R − e S } , (18) \min\{g_R-2e_R,g_S-2e_S,2-e_R-e_S\}, \tag*{(18)} min { g R − 2 e R , g S − 2 e S , 2 − e R − e S } , (18) whenever this minimum is positive.
Proof. Let K K K be one scalar block of order n ≥ 2 n\ge2 n ≥ 2 , let p ( λ ) = det ( λ I − K ) p(\lambda)=\det(\lambda I-K) p ( λ ) = det ( λ I − K ) and let v i v_i v i be a unit eigenvector at λ i \lambda_i λ i . The corner cofactor and the spectral expansion of the adjugate give
∏ k = 1 n − 1 ∣ b k ∣ = ∣ p ′ ( λ i ) v i ( 1 ) v i ( n ) ∣ ≤ 1 2 ∣ p ′ ( λ i ) ∣ . \prod_{k=1}^{n-1}|b_k|=|p'(\lambda_i)v_i(1)v_i(n)|\le\tfrac{1}{2}|p'(\lambda_i)|. k = 1 ∏ n − 1 ∣ b k ∣ = ∣ p ′ ( λ i ) v i ( 1 ) v i ( n ) ∣ ≤ 2 1 ∣ p ′ ( λ i ) ∣.
If λ i \lambda_i λ i is adjacent to a smallest gap δ \delta δ , then ∣ p ′ ( λ i ) ∣ ≤ δ W n − 2 |p'(\lambda_i)|\le\delta W^{n-2} ∣ p ′ ( λ i ) ∣ ≤ δ W n − 2 for any upper bound W W W on the spectral diameter. Testing an eigenvector at a coordinate of maximal modulus places every eigenvalue between min i ( d i − r i ) \min_i(d_i-r_i) min i ( d i − r i ) and max i ( d i + r i ) \max_i(d_i+r_i) max i ( d i + r i ) ; hence the stated W j W_j W j is such a bound. Divide the resulting scalar gap by ζ j \zeta_j ζ j and use the block separation in Proposition 1.6 . A symmetric perturbation of norm e j e_j e j moves each ordered eigenvalue by at most e j e_j e j , which proves (18). □
These bounds use only matrix entries. They give a sufficient rounding threshold before eigenvalue isolation; they need not be sharp for a given input and may be very small as N N N grows.
Reconstructing commuting operators
Take Laguerre or Jacobi input in Theorem 1.1 , with d ν ( t ) = w ( t ) d t d\nu(t)=w(t)\,dt d ν ( t ) = w ( t ) d t on ( 0 , b ) (0,b) ( 0 , b ) , where
w ( t ) = { t β e − t , b = ∞ , β > − 1 , Laguerre , t β ( 1 − t ) α , b = 1 , α , β > − 1 , Jacobi . (19) w(t)=
\begin{cases}
t^\beta e^{-t}, & b=\infty,\quad\beta>-1,\quad\text{Laguerre},\\
t^\beta(1-t)^\alpha, & b=1,\quad\alpha,\beta>-1,\quad\text{Jacobi}.
\end{cases} \tag*{(19)} w ( t ) = { t β e − t , t β ( 1 − t ) α , b = ∞ , β > − 1 , Laguerre , b = 1 , α , β > − 1 , Jacobi . (19)
Then d μ τ ( x ) = W τ ( x ) d x d\mu_{\tau}(x)=W_{\tau}(x)\,dx d μ τ ( x ) = W τ ( x ) d x on E τ E_{\tau} E τ , with
W τ ( x ) = ∣ x + τ ∣ w ( x 2 − τ 2 ) . (20) W_{\tau}(x)=|x+\tau|w(x^{2}-\tau^{2}). \tag*{(20)} W τ ( x ) = ∣ x + τ ∣ w ( x 2 − τ 2 ) . (20)
For 0 < s < b 0<s<b 0 < s < b , let χ = 1 ( 0 , s ) \chi=\mathbf{1}_{(0,s)} χ = 1 ( 0 , s ) be the indicator of ( 0 , s ) (0,s) ( 0 , s ) and abbreviate v s = v χ \mathbf{v}_{s}=\mathbf{v}_{\chi} v s = v χ and u τ , s = u τ , χ \mathbf{u}_{\tau,s}=\mathbf{u}_{\tau,\chi} u τ , s = u τ , χ . The coefficients in ( σ w ) ′ = η w (\sigma w)'=\eta w ( σ w ) ′ = η w and the cutoff constants are
Table 1.
Proposition 2.1. The operators
T R p = ( t − s ) σ p ′ ′ + ( ( t − s ) η + σ ) p ′ + κ R t p , T S p = ( t − s ) σ p ′ ′ + ( ( t − s ) ( η + σ / t ) + σ ) p ′ + κ S t p (21) \begin{aligned}
T_{R}p&=(t-s)\sigma p''+((t-s)\eta+\sigma)p'+\kappa_{R}tp,\\
T_{S}p&=(t-s)\sigma p''+((t-s)(\eta+\sigma/t)+\sigma)p'+\kappa_{S}tp \tag*{(21)}
\end{aligned} T R p T S p = ( t − s ) σ p ′′ + (( t − s ) η + σ ) p ′ + κ R tp , = ( t − s ) σ p ′′ + (( t − s ) ( η + σ / t ) + σ ) p ′ + κ S tp (21) have symmetric tridiagonal matrices K R , K S K_{R},K_{S} K R , K S on P r R , P r S \mathcal{P}_{r_{R}},\mathcal{P}_{r_{S}} P r R , P r S in the R ^ , S ^ \widehat{R},\widehat{S} R , S bases, with positive off-diagonal entries and hence simple spectra. Moreover, K R G R = G R K R K_{R}G_{R}=G_{R}K_{R} K R G R = G R K R and K S G S = G S K S K_{S}G_{S}=G_{S}K_{S} K S G S = G S K S .
Proof. Let ∂ p = p ′ \partial p=p' ∂ p = p ′ . For a = v , v s \mathbf{a}=\mathbf{v},\mathbf{v}_{s} a = v , v s , integration by parts gives
− ∂ ′ ( ( x − s ) σ a ) = ( ( x − s ) η + σ ) a . (22) -\partial'((x-s)\sigma\mathbf{a})=((x-s)\eta+\sigma)\mathbf{a}. \tag*{(22)} − ∂ ′ (( x − s ) σ a ) = (( x − s ) η + σ ) a . (22)
At the full endpoints, ( t − s ) σ ( t ) w ( t ) (t-s)\sigma(t)w(t) ( t − s ) σ ( t ) w ( t ) times any polynomial tends to zero; at s s s , the cutoff factor vanishes. For f , g ∈ P f,g\in\mathcal{P} f , g ∈ P , put q = f ′ g − f g ′ q=f'g-fg' q = f ′ g − f g ′ . Pairing (22) with q q q gives
⟨ a , ( T R f ) g − f ( T R g ) ⟩ = ⟨ a , ( x − s ) σ q ′ + ( ( x − s ) η + σ ) q ⟩ = 0 , \langle\mathbf{a},(T_{R}f)g-f(T_{R}g)\rangle
=\langle\mathbf{a},(x-s)\sigma q'+((x-s)\eta+\sigma)q\rangle=0, ⟨ a , ( T R f ) g − f ( T R g )⟩ = ⟨ a , ( x − s ) σ q ′ + (( x − s ) η + σ ) q ⟩ = 0 ,
J a T R = T R ′ J a . J_{\mathbf{a}}T_{R}=T_{R}'J_{\mathbf{a}}. J a T R = T R ′ J a .
For a = x v , x v s \mathbf{a}=x\mathbf{v},x\mathbf{v}_{s} a = x v , x v s , replace η \eta η by η + σ / x \eta+\sigma/x η + σ / x ; this proves both symmetries of T S T_{S} T S . The coefficient of t j + 1 t^{j+1} t j + 1 in T R ( t j ) T_{R}(t^{j}) T R ( t j ) is r R − j r_{R}-j r R − j for Laguerre and ( r R − j ) ( r R + j + α + β + 2 ) (r_{R}-j)(r_{R}+j+\alpha+\beta+2) ( r R − j ) ( r R + j + α + β + 2 ) for Jacobi. For T S T_{S} T S , replace r R r_{R} r R by r S r_{S} r S and β \beta β by β + 1 \beta+1 β + 1 . It is positive below the cutoff degree and zero at the cutoff. Symmetry for the full functionals and this degree bound give symmetric tridiagonal matrices with positive off-diagonal entries. In their orthonormal bases, the truncated symmetries give
K R G R = G R K R , K S G S = G S K S . K_{R}G_{R}=G_{R}K_{R},\qquad K_{S}G_{S}=G_{S}K_{S}. K R G R = G R K R , K S G S = G S K S .
An eigenvector of a tridiagonal matrix with nonzero off-diagonal entries is determined by its first entry, so each symmetric block has simple spectrum. □
For A p = σ p ′ ′ + η p ′ Ap=\sigma p''+\eta p' A p = σ p ′′ + η p ′ , X p = t p Xp=tp X p = tp and η = η 0 + η 1 t \eta=\eta_{0}+\eta_{1}t η = η 0 + η 1 t , the operators in [7 , 8 ] can be written as
T R = 1 2 ( A X + X A ) − s A + ( κ R − 1 2 η 1 ) X − 1 2 η 0 I . (23) T_{R}=\frac{1}{2}(AX+XA)-sA+\left(\kappa_{R}-\frac{1}{2}\eta_{1}\right)X-\frac{1}{2}\eta_{0}I. \tag*{(23)} T R = 2 1 ( A X + X A ) − s A + ( κ R − 2 1 η 1 ) X − 2 1 η 0 I . (23)
with the corresponding substitutions for T S T_{S} T S .
Corollary 2.2. For the weights (19), every τ ∈ R \tau\in\mathbb{R} τ ∈ R and N ≥ 2 N\geq2 N ≥ 2 , the matrix M τ ( 1 ( 0 , s ) ) M_{\tau}(\mathbf{1}_{(0,s)}) M τ ( 1 ( 0 , s ) ) has a real symmetric commuting matrix with only the main and second diagonals and simple
spectrum. It represents a second-order differential–reflection operator preserving P N − 1 P_{N-1} P N − 1 and commuting on polynomials with
( C τ , N , s F ) ( x ) = ∑ j = 0 N − 1 P ^ j ( τ ) ( x ) ⟨ u τ , s , P ^ j ( τ ) F ⟩ . (24) (\mathcal{C}_{\tau,N,s}F)(x)=\sum_{j=0}^{N-1}\widehat{P}^{(\tau)}_j(x)\langle u_{\tau,s},\widehat{P}^{(\tau)}_jF\rangle. \tag*{(24)} ( C τ , N , s F ) ( x ) = j = 0 ∑ N − 1 P j ( τ ) ( x ) ⟨ u τ , s , P j ( τ ) F ⟩ . (24)
For N = 1 N=1 N = 1 the matrix is scalar. On P N − 1 P_{N-1} P N − 1 , (24) agrees with the compressed multiplier above and has matrix M τ ( χ ) M_{\tau}(\chi) M τ ( χ ) . The normalized choice in Proposition 1.6 also represents an operator of order at most two with reflection.
Proof. For F ( x ) = f ( y ) + ( x − τ ) g ( y ) F(x)=f(y)+(x-\tau)g(y) F ( x ) = f ( y ) + ( x − τ ) g ( y ) , define
( T τ , N F ) ( x ) = ( T R f ) ( y ) + ( x − τ ) ( ( T S + c I ) g ) ( y ) . (25) (T_{\tau,N}F)(x)=(T_Rf)(y)+(x-\tau)((T_S+cI)g)(y). \tag*{(25)} ( T τ , N F ) ( x ) = ( T R f ) ( y ) + ( x − τ ) (( T S + c I ) g ) ( y ) . (25)
Its restriction to P N − 1 P_{N-1} P N − 1 has matrix L = Q T ( K R ⊕ ( K S + c I ) ) Q L=Q^{\mathsf T}(K_R\oplus(K_S+cI))Q L = Q T ( K R ⊕ ( K S + c I )) Q , with c c c from (13); Corollary 1.5 gives simplicity. The identities
g ( y ) = F ( x ) − F ( − x ) 2 x , g(y)=\frac{F(x)-F(-x)}{2x}, g ( y ) = 2 x F ( x ) − F ( − x ) ,
f ( y ) = F ( x ) + F ( − x ) 2 + τ g ( y ) f(y)=\frac{F(x)+F(-x)}{2}+\tau g(y) f ( y ) = 2 F ( x ) + F ( − x ) + τ g ( y )
show that it is a second-order differential–reflection operator with ordinary leading coefficient ( y − s ) σ ( y ) / ( 4 x 2 ) (y-s)\sigma(y)/(4x^2) ( y − s ) σ ( y ) / ( 4 x 2 ) . The quotients for f , g f,g f , g extend polynomially through x = 0 x=0 x = 0 ; the action there is defined by (25) . Let B = T R ⊕ ( T S + c I ) B=T_R\oplus(T_S+cI) B = T R ⊕ ( T S + c I ) . For a = u τ , u τ , s a=u_{\tau},u_{\tau,s} a = u τ , u τ , s , the corresponding dual pullback (7) or (9) gives
U τ T J a T τ , N U τ = ( U τ T J a U τ ) B = B ′ ( U τ T J a U τ ) = U τ T T τ , N ′ J a U τ . \begin{aligned}
U_{\tau}^{\mathsf T}J_aT_{\tau,N}U_{\tau}=(U_{\tau}^{\mathsf T}J_aU_{\tau})B \\
&=B^{\prime}(U_{\tau}^{\mathsf T}J_aU_{\tau})=U_{\tau}^{\mathsf T}T_{\tau,N}^{\prime}J_aU_{\tau}.
\end{aligned} U τ T J a T τ , N U τ = ( U τ T J a U τ ) B = B ′ ( U τ T J a U τ ) = U τ T T τ , N ′ J a U τ .
Since U τ U_{\tau} U τ and its transpose are invertible on the polynomial spaces and their duals, T τ , N T_{\tau,N} T τ , N is symmetric for both functionals. In particular, for j < N j<N j < N and every F ∈ P F\in P F ∈ P ,
⟨ u τ , s , P ^ j ( τ ) T τ , N F ⟩ = ⟨ u τ , s , ( T τ , N P ^ j ( τ ) ) F ⟩ . \langle u_{\tau,s},\widehat{P}^{(\tau)}_jT_{\tau,N}F\rangle=\langle u_{\tau,s},(T_{\tau,N}\widehat{P}^{(\tau)}_j)F\rangle. ⟨ u τ , s , P j ( τ ) T τ , N F ⟩ = ⟨ u τ , s , ( T τ , N P j ( τ ) ) F ⟩ .
Substitution in (24) , using the cutoff preservation and the symmetric matrix L L L , proves commutation on all polynomials. The normalized realization is
T ^ τ , N = U τ ( ( ζ R − 1 T R − 2 I ) ⊕ ( ζ S − 1 T S + 2 I ) ) U τ − 1 . \widehat{T}_{\tau,N}=U_{\tau}\left((\zeta_R^{-1}T_R-2I)\oplus(\zeta_S^{-1}T_S+2I)\right)U_{\tau}^{-1}. T τ , N = U τ ( ( ζ R − 1 T R − 2 I ) ⊕ ( ζ S − 1 T S + 2 I ) ) U τ − 1 .
The formulas for U τ − 1 U_{\tau}^{-1} U τ − 1 above show that its order is at most two. Both dual symmetries and the cutoff preservation persist under these affine changes, giving the same commutation argument. □
Theorem 2.3. Let N ≥ 2 N\ge2 N ≥ 2 and let χ \chi χ be a mask for which M τ ( χ ) M_{\tau}(\chi) M τ ( χ ) has simple spectrum. Suppose each G j G_j G j , j = R , S j=R,S j = R , S , admits a real symmetric tridiagonal commutant K j K_j K j with nonzero internal off-diagonal entries. Then every real symmetric pentadiagonal matrix commuting with M τ ( χ ) M_{\tau}(\chi) M τ ( χ ) has the form
Q T ( ( a R K R + b R I ) ⊕ ( a S K S + b S I ) ) Q , a R , b R , a S , b S ∈ R , (26) Q^{\mathsf T}\left((a_RK_R+b_RI)\oplus(a_SK_S+b_SI)\right)Q,\qquad a_R,b_R,a_S,b_S\in\mathbb{R}, \tag*{(26)} Q T ( ( a R K R + b R I ) ⊕ ( a S K S + b S I ) ) Q , a R , b R , a S , b S ∈ R , (26) Conversely every matrix in (26) commutes with M τ M_{\tau} M τ . This space has dimension min ( N , 4 ) \min(N,4) min ( N , 4 ) . For N ≥ 3 N\ge3 N ≥ 3 , half-bandwidth two is minimal among real symmetric commuting matrices with simple spectrum in the degree order P ^ 0 ( τ ) , … , P ^ N − 1 ( τ ) \widehat{P}^{(\tau)}_0,\ldots,\widehat{P}^{(\tau)}_{N-1} P 0 ( τ ) , … , P N − 1 ( τ ) .
Proof. The simple full concentration spectrum gives simple spectra for the two blocks and no eigenvalue in common. A commuting matrix therefore preserves each block: in orthonormal eigenbases, an off-block entry would satisfy ( λ R , i − λ S , k ) z i k = 0 (\lambda_{R,i}-\lambda_{S,k})z_{ik}=0 ( λ R , i − λ S , k ) z ik = 0 . A pentadiagonal matrix consequently has scalar tridiagonal blocks A R , A S A_R,A_S A R , A S , with no entries on its first diagonals.
Each G j G_j G j has simple spectrum, so A j A_j A j and K j K_j K j are diagonal in the same eigenbasis and commute with each other. Write b i ≠ 0 b_i \ne0 b i = 0 for the links of K j K_j K j and c i c_i c i for those of A j A_j A j . Comparing second diagonals in their commutator gives b i c i + 1 = c i b i + 1 b_i c_{i+1}=c_i b_{i+1} b i c i + 1 = c i b i + 1 . Thus c i = a j b i c_i=a_j b_i c i = a j b i for a constant a j a_j a j . The diagonal matrix A j − a j K j A_j-a_jK_j A j − a j K j still commutes with K j K_j K j ; its adjacent diagonal entries must be equal. Hence A j = a j K j + b j I A_j=a_jK_j+b_jI A j = a j K j + b j I . This also holds for a scalar block, where the parameters are redundant. The converse follows from Corollary 1.5 . Each block contributes two dimensions if its order is at least two, and one otherwise, giving min ( N , 4 ) \min(N,4) min ( N , 4 ) .
A commuting matrix of half-bandwidth at most one must be diagonal, by the same block preservation. Formula (26) then makes it constant on each parity block. For N ≥ 3 N \ge3 N ≥ 3 at least one block has order at least two, so its spectrum is not simple. Corollary 1.5 attains half-bandwidth two with simple spectrum. □
For Laguerre and Jacobi input, both restricted supports are infinite and Theorem 1.2 gives simple spectrum. Proposition 2.1 supplies the scalar matrices. Thus Theorem 2.3 applies for every interior cutoff.
Applications
We compute in the scalar coordinates of (9) , using
t p k ( t ) = a k + 1 p k + 1 ( t ) + b k p k ( t ) + a k p k − 1 ( t ) , p − 1 = 0 , a 0 = 0 , t p_k(t)=a_{k+1}p_{k+1}(t)+b_kp_k(t)+a_kp_{k-1}(t),\qquad p_{-1}=0,\qquad a_0=0, t p k ( t ) = a k + 1 p k + 1 ( t ) + b k p k ( t ) + a k p k − 1 ( t ) , p − 1 = 0 , a 0 = 0 ,
with p = R ^ p=\widehat{R} p = R or S ^ \widehat{S} S . These coefficients are evaluated from the scalar input at the working precision; no monomial moment matrix is factored. The scalar commutants are formed from (23) or (29) .
For an n × n n \times n n × n real symmetric tridiagonal matrix with diagonal d i d_i d i and off-diagonal b i b_i b i , define
D 0 ( ξ ) = 1 , D 1 ( ξ ) = d 1 − ξ , D_0(\xi)=1,\qquad D_1(\xi)=d_1-\xi, D 0 ( ξ ) = 1 , D 1 ( ξ ) = d 1 − ξ ,
D k ( ξ ) = ( d k − ξ ) D k − 1 ( ξ ) − b k − 1 2 D k − 2 ( ξ ) , 2 ≤ k ≤ n . (27) D_k(\xi)=(d_k-\xi)D_{k-1}(\xi)-b_{k-1}^{2}D_{k-2}(\xi),\qquad2\le k\le n. \tag*{(27)} D k ( ξ ) = ( d k − ξ ) D k − 1 ( ξ ) − b k − 1 2 D k − 2 ( ξ ) , 2 ≤ k ≤ n . (27)
Only finitely many endpoints are zeros of these polynomials. At an endpoint where all D k ( ξ ) ≠ 0 D_k(\xi)\ne0 D k ( ξ ) = 0 , eliminating successive coordinates gives a diagonal quadratic form with coefficients D k ( ξ ) / D k − 1 ( ξ ) D_k(\xi)/D_{k-1}(\xi) D k ( ξ ) / D k − 1 ( ξ ) . The number of negative coefficients is the number of eigenvalues below ξ \xi ξ . We call this the sign count. Interval arithmetic verifies the signs; an undecidable endpoint is replaced or evaluated at higher precision.
Algorithm 1 Alternating concentration from scalar recurrences
Require: N ≥ 1 N\ge1 N ≥ 1 , τ ∈ R \tau\in\mathbb{R} τ ∈ R , a cutoff 0 < s < b 0<s<b 0 < s < b , scalar recurrence formulas with initial constants p 0 p_0 p 0 , symmetric tridiagonal commutants with nonzero off-diagonal entries (scalar blocks allowed), a relative value tolerance ϵ > 0 \epsilon>0 ϵ > 0 and, when eigenpolynomials are requested, a tolerance ϵ v > 0 \epsilon_v>0 ϵ v > 0 for the sine of the angle.
Ensure: Enclosures of concentration values or deficits and, if requested, unit eigen-polynomial approximations in the exact recurrence bases.
Set N R = ⌊ N / 2 ⌋ N_R=\lfloor N/2\rfloor N R = ⌊ N /2 ⌋ , N S = ⌈ N / 2 ⌉ N_S=\lceil N/2\rceil N S = ⌈ N /2 ⌉ and γ = x 2 − τ 2 \gamma=x^2-\tau^2 γ = x 2 − τ 2 . Omit empty blocks.
for each nonempty scalar block j = R , S j=R,S j = R , S do
Form the symmetric tridiagonal K j K_j K j from the scalar recurrence.
Solve for approximate eigenpairs ( η i , z i ) (\eta_i,z_i) ( η i , z i ) separately in this block.
Algorithm 1 (continued)
5: For a scalar block, set g i = ∞ g_i = \infty g i = ∞ and its angle bound to zero.
6: for each requested positive target d i d_i d i do
7: repeat
8: For a block of order greater than one, validate the non-target distance g i g_i g i for the current η i \eta_i η i by sign counts.
9: Evaluate p z i ( t ) = ∑ k ( z i ) k p k ( t ) p_{z_i}(t) = \sum_k(z_i)_k p_k(t) p z i ( t ) = ∑ k ( z i ) k p k ( t ) by recurrence. Obtain d ^ i \widehat{d}_i d i from the requested positive integral
q i n ( z i ) = ∫ ( 0 , s ] ∣ p z i ( t ) ∣ 2 d ν j ( t ) ∥ z i ∥ 2 2 , q o u t ( z i ) = ∫ [ s , b ) ∣ p z i ( t ) ∣ 2 d ν j ( t ) ∥ z i ∥ 2 2 . q_{\mathrm{in}}(z_i) = \frac{\int_{(0,s]} |p_{z_i}(t)|^2\,d\nu_j(t)}{\lVert z_i\rVert_2^2}, \qquad q_{\mathrm{out}}(z_i) = \frac{\int_{[s,b)} |p_{z_i}(t)|^2\,d\nu_j(t)}{\lVert z_i\rVert_2^2}. q in ( z i ) = ∥ z i ∥ 2 2 ∫ ( 0 , s ] ∣ p z i ( t ) ∣ 2 d ν j ( t ) , q out ( z i ) = ∥ z i ∥ 2 2 ∫ [ s , b ) ∣ p z i ( t ) ∣ 2 d ν j ( t ) . 10: Bound the true normalized residual by ρ i \rho_i ρ i and the integral error by e Q , i e_{Q,i} e Q , i , including construction and rounding errors.
11: Set B i = e Q , i + Ω j ( ρ i / g i ) 2 B_i = e_{Q,i} + \Omega_j(\rho_i/g_i)^2 B i = e Q , i + Ω j ( ρ i / g i ) 2 as in (37) . In a one-dimensional block, omit the residual term.
12: If the tolerance fails, increase working precision, rebuild the recurrence data and refine the eigenpair or integration enclosure.
13: until d ^ i > B i \widehat{d}_i > B_i d i > B i , B i ≤ ϵ d ^ i / ( 1 + ϵ ) B_i \le\epsilon\widehat{d}_i/(1+\epsilon) B i ≤ ϵ d i / ( 1 + ϵ ) , and ρ i / g i ≤ ϵ v \rho_i/g_i \le\epsilon_v ρ i / g i ≤ ϵ v if eigenpolynomials are requested
14: end for
15: end for
16: Enclose an extremal value by taking the maximum or minimum of the respective interval endpoints over all candidates.
17: Reconstruct the unit approximations as p z i ( y ) / ∥ z i ∥ 2 p_{z_i}(y)/\lVert z_i\rVert_2 p z i ( y ) / ∥ z i ∥ 2 in the R R R block and ( x − τ ) p z i ( y ) / ∥ z i ∥ 2 (x-\tau)p_{z_i}(y)/\lVert z_i\rVert_2 ( x − τ ) p z i ( y ) / ∥ z i ∥ 2 in the S S S block; interlace the normalized coefficient vectors with Q T Q^{\mathsf{T}} Q T .
18: return the scalar enclosures and reconstructed coefficient vectors.
Here d ν R = d ν d\nu_R=d\nu d ν R = d ν and d ν S = t d ν d\nu_S=t\,d\nu d ν S = t d ν . The complement [ s , b ) [s,b) [ s , b ) includes any mass at the cutoff. Known zero values use absolute enclosures rather than a relative test. In scalar blocks the angle test is automatic. The scalar calculation is independent of τ \tau τ ; (8) transports its L 2 L^2 L 2 errors unchanged. For branch evaluation retain the scalar variable t t t and compute the smaller difference by r − ∣ τ ∣ = t / ( r + ∣ τ ∣ ) r-|\tau|=t/(r+|\tau|) r − ∣ τ ∣ = t / ( r + ∣ τ ∣ ) , where r = t + τ 2 r=\sqrt{t+\tau^2} r = t + τ 2 .
The Laguerre case gives the ( − 1 ) (-1) ( − 1 ) -Meixner–Pollaczek family [2 ] , also described as deformed Hermite polynomials in [6 ] ; the Jacobi case gives the Chihara family [6 ] . The Chihara weight, branch cancellation and parameter-independent norms are given in [6 ] . For fixed s s s , all τ \tau τ give the same concentration matrix on ∣ τ ∣ < ∣ x ∣ < τ 2 + s |\tau|<|x|<\sqrt{\tau^2+s} ∣ τ ∣ < ∣ x ∣ < τ 2 + s . If instead the outer endpoint T = τ 2 + s T=\sqrt{\tau^2+s} T = τ 2 + s is fixed, the input cutoff is s = T 2 − τ 2 s=T^2-\tau^2 s = T 2 − τ 2 .
Scalar Meixner concentration and its tridiagonal commutants were treated in [12 ] . We use the operator construction in [8 ] .
Example 3.1. Let β > 0 \beta>0 β > 0 , 0 < q < 1 0<q<1 0 < q < 1 and let m ≥ 0 m\ge0 m ≥ 0 be an integer. On the grid { β + k : k = 0 , 1 , … } \{\beta+k:k=0,1,\ldots\} { β + k : k = 0 , 1 , … } , set
ν j = ∑ k = 0 ∞ ( β + j ) k q k k ! δ β + k , j = 0 , 1. (28) \nu_j=\sum_{k=0}^{\infty}\frac{(\beta+j)_k q^k}{k!}\,\delta_{\beta+k},\qquad j=0,1. \tag*{(28)} ν j = k = 0 ∑ ∞ k ! ( β + j ) k q k δ β + k , j = 0 , 1. (28)
Here δ a \delta_a δ a is the unit point mass at a a a , ( a ) k = a ( a + 1 ) ⋯ ( a + k − 1 ) (a)_k = a(a+1)\cdots(a+k-1) ( a ) k = a ( a + 1 ) ⋯ ( a + k − 1 ) and ( a ) 0 = 1 (a)_0 = 1 ( a ) 0 = 1 . Both measures have all moments finite, and t d ν 0 ( t ) = β d ν 1 ( t ) t\,d\nu_0(t) = \beta\,d\nu_1(t) t d ν 0 ( t ) = β d ν 1 ( t ) . The ordinary Meixner operators [11 Section 18.22(ii)] are
A j p ( t ) = q ( t + j ) ( p ( t + 1 ) − p ( t ) ) + ( t − β ) ( p ( t − 1 ) − p ( t ) ) . A_jp(t)=q(t+j)\bigl(p(t+1)-p(t)\bigr)+(t-\beta)\bigl(p(t-1)-p(t)\bigr). A j p ( t ) = q ( t + j ) ( p ( t + 1 ) − p ( t ) ) + ( t − β ) ( p ( t − 1 ) − p ( t ) ) .
Take b = ∞ b=\infty b = ∞ , ν = ν 0 \nu=\nu_0 ν = ν 0 , χ = 1 ( 0 , s ) \chi=\mathbf{1}_{(0,s)} χ = 1 ( 0 , s ) , s = β + m + 1 2 s=\beta+m+\frac{1}{2} s = β + m + 2 1 , N ≥ 2 N\ge2 N ≥ 2 and r 0 = r R r_0=r_R r 0 = r R , r 1 = r S r_1=r_S r 1 = r S . With X p = t p Xp=tp X p = tp , put
H j = 1 2 ( A j X + X A j ) − s A j + ( 1 − q ) ( r j + 1 2 ) X . (29) H_j=\frac{1}{2}(A_jX+XA_j)-sA_j+(1-q)\left(r_j+\frac{1}{2}\right)X. \tag*{(29)} H j = 2 1 ( A j X + X A j ) − s A j + ( 1 − q ) ( r j + 2 1 ) X . (29)
Its coefficient raising degree n n n is ( 1 − q ) ( r j − n ) (1-q)(r_j-n) ( 1 − q ) ( r j − n ) . Writing w j ( k ) = ( β + j ) k q k / k ! w_j(k)=(\beta+j)_kq^k/k! w j ( k ) = ( β + j ) k q k / k ! , the balance identity w j ( k ) q ( k + β + j ) = ( k + 1 ) w j ( k + 1 ) w_j(k)q(k+\beta+j)=(k+1)w_j(k+1) w j ( k ) q ( k + β + j ) = ( k + 1 ) w j ( k + 1 ) cancels adjacent summands. The coefficient q ( t + j ) ( t + 1 2 − s ) q(t+j)(t+\frac{1}{2}-s) q ( t + j ) ( t + 2 1 − s ) of p ( t + 1 ) p(t+1) p ( t + 1 ) vanishes at t = β + m t=\beta+m t = β + m , and the lower link vanishes at t = β t=\beta t = β . For nonzero polynomials f , g f,g f , g , the boundary term at infinity is O ( q k k β + j + 1 + deg f + deg g ) O(q^k k^{\beta+j+1+\deg f+\deg g}) O ( q k k β + j + 1 + d e g f + d e g g ) and tends to zero. Consequently the full and truncated sums give
J a H j = H j ′ J a , a = x j v , x j v s , j = 0 , 1. J_aH_j=H'_jJ_a,\qquad a=x^jv,x^jv_s,\qquad j=0,1. J a H j = H j ′ J a , a = x j v , x j v s , j = 0 , 1.
Let K R , K S K_R,K_S K R , K S be the matrices of H 0 , H 1 H_0,H_1 H 0 , H 1 in the R ^ , S ^ \widehat{R},\widehat{S} R , S bases. They are symmetric tridiagonal with positive off-diagonal entries, have simple spectra, and commute with G R , G S G_R,G_S G R , G S , respectively. The reconstructed measure is supported at
x = ± β + k + τ 2 , k = 0 , 1 , … . x=\pm\sqrt{\beta+k+\tau^2},\qquad k=0,1,\ldots. x = ± β + k + τ 2 , k = 0 , 1 , … .
The mask retains k = 0 , … , m k=0,\ldots,m k = 0 , … , m . The blocks G R , G S G_R,G_S G R , G S have ranks min ( N R , m + 1 ) , min ( N S , m + 1 ) \min(N_R,m+1),\min(N_S,m+1) min ( N R , m + 1 ) , min ( N S , m + 1 ) , respectively. The full concentration spectrum is simple precisely when m + 1 ≥ ⌊ N / 2 ⌋ m+1\ge\lfloor N/2\rfloor m + 1 ≥ ⌊ N /2 ⌋ . Under this condition, Theorem 2.3 classifies all its real symmetric pentadiagonal commutants. The measure and orthonormal polynomials are given by (3) and (4). For τ = 0 \tau=0 τ = 0 , this is the Meixner system of [9 Section 3, Case 3] , up to normalization and a scaling of the variable. Corollary 1.5 gives a commuting matrix with only the main and second diagonals and simple spectrum after (13).
Write M = M τ ( χ ) M=M_\tau(\chi) M = M τ ( χ ) , which is independent of τ \tau τ . For β = 3 / 2 \beta=3/2 β = 3/2 , m = 5 m=5 m = 5 and N = 12 N=12 N = 12 , the extremal concentration values are
Table 1 (PDF p. 12).
The values were computed at 100 digits and checked at 130.
The relative commutator norms ∥ K j G j − G j K j ∥ F / ( ∥ K j ∥ F ∥ G j ∥ F ) \lVert K_jG_j-G_jK_j\rVert_F/(\lVert K_j\rVert_F\lVert G_j\rVert_F) ∥ K j G j − G j K j ∥ F / (∥ K j ∥ F ∥ G j ∥ F ) were below 3 × 10 − 101 3\times10^{-101} 3 × 1 0 − 101 at 100 digits, for j = R , S j=R,S j = R , S . Here ∥ A ∥ F 2 = ∑ i , k ∣ A i k ∣ 2 \lVert A\rVert_F^2=\sum_{i,k}|A_{ik}|^2 ∥ A ∥ F 2 = ∑ i , k ∣ A ik ∣ 2 .
Remark 3.2. The simplicity condition in Theorem 2.3 cannot be omitted in general. In Example 3.1, take β = 1 \beta=1 β = 1 , q = 1 / 2 q=1/2 q = 1/2 , m = 0 m=0 m = 0 and N = 5 N=5 N = 5 . Then
G R = r r T , r = ( 1 / 2 , − 1 / 2 , 1 / ( 2 2 ) ) T , A = ( 1 / 2 1 0 1 2 0 0 0 0 ) . G_R=rr^{\mathsf T},\qquad r=(1/\sqrt{2},-1/2,1/(2\sqrt{2}))^{\mathsf T},\qquad A=
\begin{pmatrix}
1/\sqrt{2} & 1 & 0\\
1 & \sqrt{2} & 0\\
0 & 0 & 0
\end{pmatrix}. G R = r r T , r = ( 1/ 2 , − 1/2 , 1/ ( 2 2 ) ) T , A = 1/ 2 1 0 1 2 0 0 0 0 .
satisfies A r = 0 Ar=0 A r = 0 . Thus A A A commutes with G R G_R G R , but it is not affine in K R K_R K R : its second link is zero and its first is nonzero, whereas both links of K R K_R K R are nonzero. The matrix K R + A K_R+A K R + A remains irreducible tridiagonal. A sufficiently large shift of K S K_S K S therefore gives a simple-spectrum pentadiagonal commutant outside (26) .
The Hermite blocks in (30) also determine every cutoff at which the unmodified full matrix has repeated eigenvalues. For N ≥ 2 N\ge2 N ≥ 2 , put a = μ + 1 2 > 0 a=\mu+\frac{1}{2}>0 a = μ + 2 1 > 0 , s = T 2 s=T^2 s = T 2 and m 0 = N R m_0=N_R m 0 = N R , m 1 = N S m_1=N_S m 1 = N S . Write L 0 = L R L_0=L_R L 0 = L R and L 1 = L S L_1=L_S L 1 = L S . In parity ϵ = 0 , 1 \epsilon=0,1 ϵ = 0 , 1 , their diagonal entries d ϵ , j d_{\epsilon,j} d ϵ , j and positive off-diagonal entries b ϵ , j b_{\epsilon,j} b ϵ , j satisfy
d ϵ , j = 2 ( m ϵ − 1 − j ) ( 2 j + ϵ + a ) + ( ϵ + 2 j ) ( s + 1 ) , d_{\epsilon,j}=2(m_\epsilon-1-j)(2j+\epsilon+a)+(\epsilon+2j)(s+1), d ϵ , j = 2 ( m ϵ − 1 − j ) ( 2 j + ϵ + a ) + ( ϵ + 2 j ) ( s + 1 ) ,
b ϵ , j 2 = 4 ( m ϵ − 1 − j ) 2 ( j + 1 ) ( j + a + ϵ ) . (31) b_{\epsilon,j}^{2}=4(m_\epsilon-1-j)^2(j+1)(j+a+\epsilon). \tag*{(31)} b ϵ , j 2 = 4 ( m ϵ − 1 − j ) 2 ( j + 1 ) ( j + a + ϵ ) . (31)
Here 0 ≤ j < m ϵ 0\le j<m_\epsilon 0 ≤ j < m ϵ in the first row and 0 ≤ j < m ϵ − 1 0\le j<m_\epsilon-1 0 ≤ j < m ϵ − 1 in the second. Define the monic polynomials
p R ( λ ; a , s ) = det ( λ I − L R ) , p S ( λ ; a , s ) = det ( λ I − L S ) . p_R(\lambda;a,s)=\det(\lambda I-L_R),\qquad p_S(\lambda;a,s)=\det(\lambda I-L_S). p R ( λ ; a , s ) = det ( λ I − L R ) , p S ( λ ; a , s ) = det ( λ I − L S ) .
Theorem 3.4. For N ≥ 2 N\ge2 N ≥ 2 , the resultant
Δ N ( a , s ) = Res λ ( p R , p S ) (32) \Delta_N(a,s)=\operatorname{Res}_{\lambda}(p_R,p_S) \tag*{(32)} Δ N ( a , s ) = Res λ ( p R , p S ) (32) is a polynomial in a , s a,s a , s with integer coefficients. Its degree in s s s is m 0 m 1 = ⌊ N 2 / 4 ⌋ m_0m_1=\lfloor N^2/4\rfloor m 0 m 1 = ⌊ N 2 /4 ⌋ , with leading coefficient
c N = ∏ i = 0 m 0 − 1 ∏ j = 0 m 1 − 1 ( 2 i − 2 j − 1 ) ≠ 0. c_N=\prod_{i=0}^{m_0-1}\prod_{j=0}^{m_1-1}(2i-2j-1)\ne0. c N = i = 0 ∏ m 0 − 1 j = 0 ∏ m 1 − 1 ( 2 i − 2 j − 1 ) = 0.
The full Hermite matrix L L L has simple spectrum precisely when Δ N ( a , s ) ≠ 0 \Delta_N(a,s) \ne0 Δ N ( a , s ) = 0 . Thus, for each fixed μ > − 1 / 2 \mu> -1/2 μ > − 1/2 , at most ⌊ N 2 / 4 ⌋ \lfloor N^2/4 \rfloor ⌊ N 2 /4 ⌋ positive cutoffs s = T 2 s = T^2 s = T 2 have repeated eigenvalues. In particular,
s > 2 ∥ L ( a , 0 ) ∥ ∞ (33) s > 2\lVert L(a,0)\rVert_\infty \tag*{(33)} s > 2 ∥ L ( a , 0 ) ∥ ∞ (33)
guarantees simple spectrum.
Proof. The tridiagonal determinant recurrence is
P ε , 0 = 1 , P ε , 1 = λ − d ε , 0 , P_{\varepsilon,0}=1,\qquad P_{\varepsilon,1}=\lambda-d_{\varepsilon,0}, P ε , 0 = 1 , P ε , 1 = λ − d ε , 0 ,
P ε , j + 1 = ( λ − d ε , j ) P ε , j − b ε , j − 1 2 P ε , j − 1 ( 1 ≤ j < m ε ) . P_{\varepsilon,j+1}=(\lambda-d_{\varepsilon,j})P_{\varepsilon,j}-b_{\varepsilon,j-1}^{2}P_{\varepsilon,j-1}\qquad(1\le j<m_\varepsilon). P ε , j + 1 = ( λ − d ε , j ) P ε , j − b ε , j − 1 2 P ε , j − 1 ( 1 ≤ j < m ε ) .
It gives p R = P 0 , m 0 p_R=P_{0,m_0} p R = P 0 , m 0 and p S = P 1 , m 1 p_S=P_{1,m_1} p S = P 1 , m 1 with integer coefficients in a , s , λ a,s,\lambda a , s , λ . The resultant is a determinant in these coefficients, so its coefficients are integers too. Equivalently, if λ ε , j \lambda_{\varepsilon,j} λ ε , j are the increasingly ordered eigenvalues of L ε L_\varepsilon L ε ,
Δ N ( a , s ) = ∏ i = 0 m 0 − 1 ∏ j = 0 m 1 − 1 ( λ 0 , i − λ 1 , j ) . \Delta_N(a,s)=\prod_{i=0}^{m_0-1}\prod_{j=0}^{m_1-1}(\lambda_{0,i}-\lambda_{1,j}). Δ N ( a , s ) = i = 0 ∏ m 0 − 1 j = 0 ∏ m 1 − 1 ( λ 0 , i − λ 1 , j ) .
Write L ε = B ε + s D ε L_\varepsilon=B_\varepsilon+sD_\varepsilon L ε = B ε + s D ε , where B ε = L ε ( a , 0 ) B_\varepsilon=L_\varepsilon(a,0) B ε = L ε ( a , 0 ) and
D ε = diag ( ε , ε + 2 , … , ε + 2 m ε − 2 ) . D_\varepsilon=\operatorname{diag}(\varepsilon,\varepsilon+2,\ldots,\varepsilon+2m_\varepsilon-2). D ε = diag ( ε , ε + 2 , … , ε + 2 m ε − 2 ) .
For each fixed a a a , ∣ λ ε , j − ( ε + 2 j ) s ∣ ≤ ∥ B ε ∥ 2 \lvert\lambda_{\varepsilon,j}-(\varepsilon+2j)s\rvert\le\lVert B_\varepsilon\rVert_2 ∣ λ ε , j − ( ε + 2 j ) s ∣ ≤ ∥ B ε ∥ 2 , so λ ε , j = ( ε + 2 j ) s + O ( 1 ) \lambda_{\varepsilon,j}=(\varepsilon+2j)s+O(1) λ ε , j = ( ε + 2 j ) s + O ( 1 ) as s → ∞ s\to\infty s → ∞ . Consequently,
Δ N ( a , s ) = c N s m 0 m 1 + O ( s m 0 m 1 − 1 ) . \Delta_N(a,s)=c_Ns^{m_0m_1}+O(s^{m_0m_1-1}). Δ N ( a , s ) = c N s m 0 m 1 + O ( s m 0 m 1 − 1 ) .
Since this holds for every a > 0 a>0 a > 0 , it proves the polynomial degree and the constant leading coefficient. The positive off-diagonal entries give simple spectra within the blocks. Thus a repeated eigenvalue occurs precisely when p R , p S p_R,p_S p R , p S have a common root. The degree bound gives the finite exceptional set. Finally, L ( a , s ) = L ( a , 0 ) + s diag ( 0 , 1 , … , N − 1 ) L(a,s)=L(a,0)+s\operatorname{diag}(0,1,\ldots,N-1) L ( a , s ) = L ( a , 0 ) + s diag ( 0 , 1 , … , N − 1 ) . Its ordered eigenvalues differ from 0 , s , … , ( N − 1 ) s 0,s,\ldots,(N-1)s 0 , s , … , ( N − 1 ) s by at most ∥ L ( a , 0 ) ∥ 2 ≤ ∥ L ( a , 0 ) ∥ ∞ \lVert L(a,0)\rVert_2\le\lVert L(a,0)\rVert_\infty ∥ L ( a , 0 ) ∥ 2 ≤ ∥ L ( a , 0 ) ∥ ∞ , proving (33). □
For N = 3 N=3 N = 3 , the recurrence gives
Δ 3 ( a , s ) = 2 a ( s − 1 ) − ( s + 1 ) 2 . \Delta_3(a,s)=2a(s-1)-(s+1)^2. Δ 3 ( a , s ) = 2 a ( s − 1 ) − ( s + 1 ) 2 .
For a > 4 a>4 a > 4 its roots s = a − 1 ± a ( a − 4 ) s=a-1\pm\sqrt{a(a-4)} s = a − 1 ± a ( a − 4 ) are positive and distinct, so the exceptional-cutoff bound is attained. For a = 4 a=4 a = 4 the sole exceptional cutoff is s = 3 s=3 s = 3 ; for 0 < a < 4 0<a<4 0 < a < 4 there are none.
Proposition 3.5. For every odd N ≥ 3 N\ge3 N ≥ 3 and every S > 0 S>0 S > 0 , there are parameters a > 0 a>0 a > 0 and s > S s>S s > S for which the full Hermite matrix has an eigenvalue of multiplicity two.
Proof. Write N = 2 m + 1 N=2m+1 N = 2 m + 1 and let ℓ R ( a , s ) \ell_R(a,s) ℓ R ( a , s ) and ℓ S ( a , s ) \ell_S(a,s) ℓ S ( a , s ) be the smallest eigenvalues of the two blocks. Fix s > 1 s>1 s > 1 and isolate the last coordinate of either block as
H a = ( A a b a b a T d ) . H_a=\begin{pmatrix} A_a & b_a \\ b_a^{\mathsf{T}} & d \end{pmatrix}. H a = ( A a b a T b a d ) .
For m = 1 m=1 m = 1 the odd block is the scalar s + 1 s+1 s + 1 and needs no decomposition. Formula (31) gives
A R , a 2 a ⟶ diag ( m , … , 1 ) , b R , a 2 a ⟶ 2 m e m , \frac{A_{R,a}}{2a} \longrightarrow\operatorname{diag}(m,\ldots,1), \qquad\frac{b_{R,a}}{\sqrt{2a}} \longrightarrow\sqrt{2m}e_m, 2 a A R , a ⟶ diag ( m , … , 1 ) , 2 a b R , a ⟶ 2 m e m ,
A S , a 2 a ⟶ diag ( m − 1 , … , 1 ) ( m ≥ 2 ) , \frac{A_{S,a}}{2a} \longrightarrow\operatorname{diag}(m-1,\ldots,1) \quad(m\ge2), 2 a A S , a ⟶ diag ( m − 1 , … , 1 ) ( m ≥ 2 ) ,
b S , a 2 a ⟶ 2 ( m − 1 ) e m − 1 ( m ≥ 2 ) . \frac{b_{S,a}}{\sqrt{2a}} \longrightarrow\sqrt{2(m-1)}e_{m-1} \quad(m\ge2). 2 a b S , a ⟶ 2 ( m − 1 ) e m − 1 ( m ≥ 2 ) .
Here a → ∞ a\to\infty a → ∞ , e j e_j e j is the last coordinate vector in R j \mathbb{R}^j R j , d R = 2 m ( s + 1 ) d_R=2m(s+1) d R = 2 m ( s + 1 ) and d S = ( 2 m − 1 ) ( s + 1 ) d_S=(2m-1)(s+1) d S = ( 2 m − 1 ) ( s + 1 ) . The smallest eigenvalue of each nonempty A a A_a A a tends to infinity. For sufficiently large a a a , completing the square gives
min { 0 , d − b a T A a − 1 b a } ≤ λ min ( H a ) ≤ d . \min\{0,d-b_a^{\mathsf T}A_a^{-1}b_a\}\le\lambda_{\min}(H_a)\le d. min { 0 , d − b a T A a − 1 b a } ≤ λ m i n ( H a ) ≤ d .
The lower bounds stay finite, so these eigenvalues remain bounded. Eliminating the first block gives
λ = d − b a T ( A a − λ I ) − 1 b a . \lambda=d-b_a^{\mathsf T}(A_a-\lambda I)^{-1}b_a. λ = d − b a T ( A a − λ I ) − 1 b a .
The limits above yield
ℓ R ( a , s ) ⟶ 2 m s , ℓ S ( a , s ) ⟶ ( 2 m − 1 ) s + 1. (34) \ell_R(a,s)\longrightarrow2ms,\qquad\ell_S(a,s)\longrightarrow(2m-1)s+1. \tag*{(34)} ℓ R ( a , s ) ⟶ 2 m s , ℓ S ( a , s ) ⟶ ( 2 m − 1 ) s + 1. (34)
Choose s 0 > max { 1 , S } s_0>\max\{1,S\} s 0 > max { 1 , S } . For sufficiently large finite a a a , ℓ R ( a , s 0 ) − ℓ S ( a , s 0 ) > 0 \ell_R(a,s_0)-\ell_S(a,s_0)>0 ℓ R ( a , s 0 ) − ℓ S ( a , s 0 ) > 0 . Hold this a a a fixed. Testing the first coordinate in the even block and using D 1 ≥ I D_1\ge I D 1 ≥ I in the odd block gives
ℓ R ( a , s ) ≤ 2 m a , ℓ S ( a , s ) ≥ s + ℓ S ( a , 0 ) . \ell_R(a,s)\le2ma,\qquad\ell_S(a,s)\ge s+\ell_S(a,0). ℓ R ( a , s ) ≤ 2 ma , ℓ S ( a , s ) ≥ s + ℓ S ( a , 0 ) .
Thus the difference is negative for sufficiently large s s s . The difference therefore vanishes at some s > s 0 s>s_0 s > s 0 . Each block has simple spectrum, so their common smallest eigenvalue has multiplicity two in the full matrix. □
The parameters in this proposition depend on N N N and S S S . Proposition 1.6 separates the Hermite blocks for every admissible N , μ , T N,\mu,T N , μ , T .
For the Jacobi reconstruction, the exceptional cutoffs also form a finite set, including on the parameter line where two diagonal slopes coincide. Fix α > − 1 \alpha>-1 α > − 1 , μ > − 1 / 2 \mu>-1/2 μ > − 1/2 , put β = μ − 1 2 \beta=\mu-\frac{1}{2} β = μ − 2 1 , s = T 2 s=T^2 s = T 2 and
ϑ = 2 α + 2 μ + 1 > − 2 , m 0 = N R , m 1 = N S , K = m 0 m 1 = ⌊ N 2 / 4 ⌋ , \vartheta=2\alpha+2\mu+1>-2,\qquad m_0=N_R,\qquad m_1=N_S,\qquad K=m_0m_1=\lfloor N^2/4\rfloor, ϑ = 2 α + 2 μ + 1 > − 2 , m 0 = N R , m 1 = N S , K = m 0 m 1 = ⌊ N 2 /4 ⌋ ,
and use the Jacobi blocks of (30). Write u n > 0 u_n>0 u n > 0 (n ≥ 1 n\ge1 n ≥ 1 ) for the full recurrence coefficients, so that
x P ^ n ( 0 ) = u n + 1 P ^ n + 1 ( 0 ) + u n P ^ n − 1 ( 0 ) , u 0 = 0 , P ^ − 1 ( 0 ) = 0. x\widehat{P}^{(0)}_n=u_{n+1}\widehat{P}^{(0)}_{n+1}+u_n\widehat{P}^{(0)}_{n-1},\qquad u_0=0,\qquad\widehat{P}^{(0)}_{-1}=0. x P n ( 0 ) = u n + 1 P n + 1 ( 0 ) + u n P n − 1 ( 0 ) , u 0 = 0 , P − 1 ( 0 ) = 0.
Define
p R ( λ ; s ) = det ( λ I − L R ( s ) ) , p S ( λ ; s ) = det ( λ I − L S ( s ) ) , Ξ N ( s ) = Res λ ( p R , p S ) . p_R(\lambda;s)=\det(\lambda I-L_R(s)),\qquad p_S(\lambda;s)=\det(\lambda I-L_S(s)),\qquad\Xi_N(s)=\operatorname{Res}_{\lambda}(p_R,p_S). p R ( λ ; s ) = det ( λ I − L R ( s )) , p S ( λ ; s ) = det ( λ I − L S ( s )) , Ξ N ( s ) = Res λ ( p R , p S ) .
Theorem 3.6. For every N ≥ 2 N\ge2 N ≥ 2 and every fixed admissible pair α , μ \alpha,\mu α , μ , Ξ N \Xi_N Ξ N is a nonzero real polynomial. The full Jacobi matrix of [3 (14)] has simple spectrum precisely when Ξ N ( T 2 ) ≠ 0 \Xi_N(T^2)\ne0 Ξ N ( T 2 ) = 0 . Its degree is K K K if ϑ ≠ − 1 \vartheta\ne-1 ϑ = − 1 . If ϑ = − 1 \vartheta=-1 ϑ = − 1 , its degree is K − 1 K-1 K − 1 , except that for
N = 2 m + 1 ≥ 3 , μ = − 6 m 8 m 2 + 8 m − 1 N=2m+1\ge3,\qquad\mu=-\frac{6m}{8m^2+8m-1} N = 2 m + 1 ≥ 3 , μ = − 8 m 2 + 8 m − 1 6 m
the degree is K − 2 K-2 K − 2 . Thus the number of exceptional cutoffs 0 < T < 1 0<T<1 0 < T < 1 is at most the corresponding degree. Every repeated eigenvalue has multiplicity two.
Proof. Write L ( s ) = A + s D L(s)=A+sD L ( s ) = A + sD , where
A = L ( 0 ) , D = diag ( q 0 , … , q N − 1 ) , q n = n ( n + ϑ ) . A=L(0),\qquad D=\operatorname{diag}(q_0,\ldots,q_{N-1}),\qquad q_n=n(n+\vartheta). A = L ( 0 ) , D = diag ( q 0 , … , q N − 1 ) , q n = n ( n + ϑ ) .
The entries on the second diagonals are independent of s s s and strictly positive. Indeed, in [3 (14)] they are Λ n u n + 1 u n + 2 \Lambda_n u_{n+1}u_{n+2} Λ n u n + 1 u n + 2 , with
k n = ⌊ N − 1 − n 2 ⌋ , Λ n = 4 k n ( n + k n + ϑ + 2 2 ) ( 0 ≤ n < N ) . k_n=\left\lfloor\frac{N-1-n}{2}\right\rfloor,\qquad\Lambda_n=4k_n\left(n+k_n+\frac{\vartheta+2}{2}\right)\qquad(0\le n<N). k n = ⌊ 2 N − 1 − n ⌋ , Λ n = 4 k n ( n + k n + 2 ϑ + 2 ) ( 0 ≤ n < N ) .
For the internal links 0 ≤ n ≤ N − 3 0\le n\le N-3 0 ≤ n ≤ N − 3 , one has Λ n > 0 \Lambda_n>0 Λ n > 0 . Here the recurrence coefficients are taken in the positive-leading orthonormal normalization, with removable quotients evaluated by continuity. Hence each parity block is an irreducible symmetric tridiagonal matrix and has simple spectrum. Its determinant recurrence gives a polynomial in s , λ s,\lambda s , λ . The resultant therefore is a polynomial in s s s , and vanishes exactly when the blocks have a common eigenvalue. Such an eigenvalue has multiplicity two in the full matrix.
The slopes are strictly increasing within each parity. The eigenvalue in parity ε \varepsilon ε corresponding to the slope q 2 j + ε q_{2j+\varepsilon} q 2 j + ε is q 2 j + ε s + O ( 1 ) q_{2j+\varepsilon}s+O(1) q 2 j + ε s + O ( 1 ) as s → ∞ s\to\infty s → ∞ . Consequently, if ϑ ≠ − 1 \vartheta\ne-1 ϑ = − 1 , the leading coefficient of the resultant is
∏ i = 0 m 0 − 1 ∏ j = 0 m 1 − 1 ( q 2 i − q 2 j + 1 ) = ∏ i = 0 m 0 − 1 ∏ j = 0 m 1 − 1 ( 2 i − 2 j − 1 ) ( 2 i + 2 j + 1 + ϑ ) ≠ 0. \prod_{i=0}^{m_0-1}\prod_{j=0}^{m_1-1}(q_{2i}-q_{2j+1})
=
\prod_{i=0}^{m_0-1}\prod_{j=0}^{m_1-1}(2i-2j-1)(2i+2j+1+\vartheta)\ne0. i = 0 ∏ m 0 − 1 j = 0 ∏ m 1 − 1 ( q 2 i − q 2 j + 1 ) = i = 0 ∏ m 0 − 1 j = 0 ∏ m 1 − 1 ( 2 i − 2 j − 1 ) ( 2 i + 2 j + 1 + ϑ ) = 0.
This proves that its degree is K K K .
Suppose now that ϑ = − 1 \vartheta=-1 ϑ = − 1 . Then − 1 2 < μ < 0 -\frac12<\mu<0 − 2 1 < μ < 0 , and the sole coincidence between slopes of different parities is q 0 = q 1 = 0 q_0=q_1=0 q 0 = q 1 = 0 . Put A n = A n , n A_n=A_{n,n} A n = A n , n . The recurrence coefficients needed below are
u 1 2 = 1 + 2 μ , u 2 2 = − 4 μ 3 , u 3 2 = 3 + 2 μ 15 . u_1^2=1+2\mu,\qquad u_2^2=-\frac{4\mu}{3},\qquad u_3^2=\frac{3+2\mu}{15}. u 1 2 = 1 + 2 μ , u 2 2 = − 3 4 μ , u 3 2 = 15 3 + 2 μ .
For N = 2 m N=2m N = 2 m , direct substitution in [3 (14)] gives
A 1 − A 0 = 4 m − 2 + 2 μ − 16 μ 3 ( m − 1 ) 2 > 0. A_1-A_0=4m-2+2\mu-\frac{16\mu}{3}(m-1)^2>0. A 1 − A 0 = 4 m − 2 + 2 μ − 3 16 μ ( m − 1 ) 2 > 0.
All other cross-parity eigenvalue differences grow linearly with s s s . The exceptional pair tends to A 0 − A 1 ≠ 0 A_0-A_1\ne0 A 0 − A 1 = 0 , so the resultant has degree K − 1 K-1 K − 1 .
For N = 2 m + 1 N=2m+1 N = 2 m + 1 , substitution instead gives
A 1 − A 0 = − 4 m − 2 μ 3 ( 8 m 2 + 8 m − 1 ) . A_1-A_0=-4m-\frac{2\mu}{3}(8m^2+8m-1). A 1 − A 0 = − 4 m − 3 2 μ ( 8 m 2 + 8 m − 1 ) .
It is nonzero except at the parameter stated in the theorem; this again gives degree K − 1 K-1 K − 1 . At that parameter, denote by η 0 ( s ) \eta_0(s) η 0 ( s ) and η 1 ( s ) \eta_1(s) η 1 ( s ) the eigenvalues belonging to the zero slopes. Isolating the first coordinate and eliminating the remaining coordinates gives
η 0 ( s ) = A 0 − C 0 s + O ( s − 2 ) , η 1 ( s ) = A 1 − C 1 s + O ( s − 2 ) , \eta_0(s)=A_0-\frac{C_0}{s}+O(s^{-2}),\qquad\eta_1(s)=A_1-\frac{C_1}{s}+O(s^{-2}), η 0 ( s ) = A 0 − s C 0 + O ( s − 2 ) , η 1 ( s ) = A 1 − s C 1 + O ( s − 2 ) ,
where
C 0 = A 0 2 u 1 2 u 2 2 q 2 = − 8 μ 3 m 2 ( 2 m + 1 ) 2 ( 1 + 2 μ ) , C_0=\frac{A_0^2u_1^2u_2^2}{q_2}
=-\frac{8\mu}{3}m^2(2m+1)^2(1+2\mu), C 0 = q 2 A 0 2 u 1 2 u 2 2 = − 3 8 μ m 2 ( 2 m + 1 ) 2 ( 1 + 2 μ ) ,
C 1 = A 1 2 u 2 2 u 3 2 q 3 = − 8 μ 135 ( m − 1 ) 2 ( 2 m + 1 ) 2 ( 3 + 2 μ ) . C_1=\frac{A_1^2u_2^2u_3^2}{q_3}
=-\frac{8\mu}{135}(m-1)^2(2m+1)^2(3+2\mu). C 1 = q 3 A 1 2 u 2 2 u 3 2 = − 135 8 μ ( m − 1 ) 2 ( 2 m + 1 ) 2 ( 3 + 2 μ ) .
For m = 1 m = 1 m = 1 the odd block is scalar and C 1 = 0 C_1 = 0 C 1 = 0 . The same formula therefore covers all m ≥ 1 m \ge1 m ≥ 1 . At the specified parameter one has − 2 / 5 ≤ μ < 0 -2/5 \le\mu< 0 − 2/5 ≤ μ < 0 , since
8 m 2 + 8 m − 1 − 15 m = ( m − 1 ) ( 8 m + 1 ) ≥ 0. 8m^2 + 8m - 1 - 15m = (m - 1)(8m + 1) \ge0. 8 m 2 + 8 m − 1 − 15 m = ( m − 1 ) ( 8 m + 1 ) ≥ 0.
It follows that
45 m 2 ( 1 + 2 μ ) − ( m − 1 ) 2 ( 3 + 2 μ ) ≥ 9 m 2 − 3 ( m − 1 ) 2 > 0 , 45m^2(1 + 2\mu) - (m - 1)^2(3 + 2\mu) \ge9m^2 - 3(m - 1)^2 > 0, 45 m 2 ( 1 + 2 μ ) − ( m − 1 ) 2 ( 3 + 2 μ ) ≥ 9 m 2 − 3 ( m − 1 ) 2 > 0 ,
so C 0 > C 1 C_0 > C_1 C 0 > C 1 . Thus the exceptional eigenvalue difference is a nonzero multiple of s − 1 s^{-1} s − 1 , whereas all the other K − 1 K - 1 K − 1 cross-parity differences are nonzero multiples of s s s to leading order. The resultant has degree exactly K − 2 K - 2 K − 2 . This also proves that it never vanishes identically. □ \square □
For example, at N = 3 N = 3 N = 3 , α = − 3 / 5 \alpha= -3/5 α = − 3/5 and μ = − 2 / 5 \mu= -2/5 μ = − 2/5 , the resultant is the nonzero constant Ξ 3 ( s ) = − 96 / 25 \Xi_3(s) = -96/25 Ξ 3 ( s ) = − 96/25 . The reduction of its degree therefore creates no persistent degeneracy.
The scalar residual bounds are as follows [4 ] .
Proposition 3.7. Under Corollary 1.5, suppose each K j K_j K j , j = R , S j = R, S j = R , S , has simple spectrum, and let V j V_j V j be an orthogonal matrix of its eigenvectors. Then D j = V j T G j V j D_j = V_j^{\mathsf T}G_jV_j D j = V j T G j V j is diagonal and, for every τ \tau τ ,
Y = Q T ( V R ⊕ V S ) , M τ ( χ ) Y = Y ( D R ⊕ D S ) . (35) Y = Q^{\mathsf T}(V_R \oplus V_S), \qquad M_\tau(\chi)Y = Y(D_R \oplus D_S). \tag*{(35)} Y = Q T ( V R ⊕ V S ) , M τ ( χ ) Y = Y ( D R ⊕ D S ) . (35) The spectra of K R K_R K R and K S K_S K S may overlap. In a block of size at least two, let ( κ i , v i ) (\kappa_i, v_i) ( κ i , v i ) be an eigenpair of K j K_j K j , with ∥ v i ∥ 2 = 1 \lVert v_i\rVert_2 = 1 ∥ v i ∥ 2 = 1 , and set λ i = v i T G j v i \lambda_i = v_i^{\mathsf T}G_jv_i λ i = v i T G j v i . For a real unit vector z z z and η ∈ R \eta\in\mathbb{R} η ∈ R , put r = K j z − η z r = K_jz - \eta z r = K j z − η z and δ = min k ≠ i ∣ κ k − η ∣ \delta= \min_{k \ne i} \lvert\kappa_k - \eta\rvert δ = min k = i ∣ κ k − η ∣ . Write ω j = λ max ( G j ) − λ min ( G j ) \omega_j = \lambda_{\max}(G_j) - \lambda_{\min}(G_j) ω j = λ m a x ( G j ) − λ m i n ( G j ) ; then ω j ≤ min { 1 , tr G j , tr ( I − G j ) } \omega_j \le\min\{1,\operatorname{tr}G_j,\operatorname{tr}(I - G_j)\} ω j ≤ min { 1 , tr G j , tr ( I − G j )} . If δ > 0 \delta> 0 δ > 0 , then
sin ∠ ( z , v i ) ≤ ∥ r ∥ 2 δ , ∣ z T G j z − λ i ∣ ≤ ω j ∥ r ∥ 2 2 δ 2 . (36) \sin\angle(z,v_i) \le\frac{\lVert r\rVert_2}{\delta}, \qquad\lvert z^{\mathsf T}G_jz - \lambda_i\rvert\le\omega_j\frac{\lVert r\rVert_2^2}{\delta^2}. \tag*{(36)} sin ∠ ( z , v i ) ≤ δ ∥ r ∥ 2 , ∣ z T G j z − λ i ∣ ≤ ω j δ 2 ∥ r ∥ 2 2 . (36) Both bounds are preserved by reconstruction with U τ U_\tau U τ . Here ∠ \angle ∠ denotes the acute angle between the spans of the vectors.
For a nonzero stored vector z z z , let d i d_i d i be either λ i \lambda_i λ i or 1 − λ i 1 - \lambda_i 1 − λ i . Suppose ρ \rho ρ bounds the true residual ∥ K j z − η z ∥ 2 / ∥ z ∥ 2 \lVert K_jz - \eta z\rVert_2/\lVert z\rVert_2 ∥ K j z − η z ∥ 2 / ∥ z ∥ 2 , 0 < g ≤ δ 0 < g \le\delta 0 < g ≤ δ , Ω j ≥ ω j \Omega_j \ge\omega_j Ω j ≥ ω j , and d ^ \widehat d d approximates the corresponding normalized positive form with absolute error at most e Q e_Q e Q . Then
∣ d ^ − d i ∣ ≤ B = e Q + Ω j ( ρ / g ) 2 , \lvert\widehat d - d_i\rvert\le B = e_Q + \Omega_j(\rho/g)^2, ∣ d − d i ∣ ≤ B = e Q + Ω j ( ρ / g ) 2 , ∣ d ^ d i − 1 ∣ ≤ B d ^ − B ( d ^ > B ) . (37) \left\lvert\frac{\widehat d}{d_i} - 1\right\rvert\le\frac{B}{\widehat d - B}\quad(\widehat d > B). \tag*{(37)} d i d − 1 ≤ d − B B ( d > B ) . (37) For a positive target, B ≤ ϵ d / d ^ ( 1 + ϵ ) B \le\epsilon d/\widehat d(1 + \epsilon) B ≤ ϵ d / d ( 1 + ϵ ) with d ^ > B \widehat d > B d > B guarantees relative error at most ϵ \epsilon ϵ . Construction and rounding errors must be included in ρ \rho ρ and e Q e_Q e Q ; Ω j = 1 \Omega_j = 1 Ω j = 1 is always admissible.
Proof. Commutation and simplicity make every eigenvector of K j K_j K j an eigenvector of G j G_j G j . This proves (35). Expand z = ∑ k a k v k z = \sum_k a_kv_k z = ∑ k a k v k in their common orthonormal eigenbasis. Since 0 ≤ G j ≤ I 0 \le G_j \le I 0 ≤ G j ≤ I ,
∥ r ∥ 2 2 = ∑ k ∣ κ k − η ∣ 2 ∣ a k ∣ 2 ≥ δ 2 ∑ k ≠ i ∣ a k ∣ 2 , \lVert r\rVert_2^2 = \sum_k \lvert\kappa_k - \eta\rvert^2\lvert a_k\rvert^2 \ge\delta^2\sum_{k \ne i}\lvert a_k\rvert^2, ∥ r ∥ 2 2 = k ∑ ∣ κ k − η ∣ 2 ∣ a k ∣ 2 ≥ δ 2 k = i ∑ ∣ a k ∣ 2 , ∣ z T G j z − λ i ∣ = ∣ ∑ k ≠ i ( λ k − λ i ) ∣ a k ∣ 2 ∣ ≤ ω j ∑ k ≠ i ∣ a k ∣ 2 . \lvert z^{\mathsf T}G_jz - \lambda_i\rvert= \left\lvert\sum_{k \ne i}(\lambda_k - \lambda_i)\lvert a_k\rvert^2\right\rvert\le\omega_j\sum_{k \ne i}\lvert a_k\rvert^2. ∣ z T G j z − λ i ∣ = k = i ∑ ( λ k − λ i ) ∣ a k ∣ 2 ≤ ω j k = i ∑ ∣ a k ∣ 2 .
Use (8) for the lifted vectors. In dimension one, both errors vanish. Apply the same argument to z / ∥ z ∥ 2 z/\lVert z\rVert_2 z / ∥ z ∥ 2 and to I − G ~ j I-\widetilde{G}_j I − G j for a deficit, then add e Q e_Q e Q . Finally d i ≥ d ^ − B > 0 d_i \ge\widehat{d}-B>0 d i ≥ d − B > 0 gives the relative bound. If ∥ K j − K ^ j ∥ 2 ≤ e K \lVert K_j-\widehat{K}_j\rVert_2\le e_K ∥ K j − K j ∥ 2 ≤ e K , one may use ρ ≥ ∥ K ^ j z − η z ∥ 2 / ∥ z ∥ 2 + e K \rho\ge\lVert\widehat{K}_jz-\eta z\rVert_2/\lVert z\rVert_2+e_K ρ ≥ ∥ K j z − η z ∥ 2 / ∥ z ∥ 2 + e K .
The affine changes in Proposition 1.6 preserve the scalar eigenvectors. Applied also to the approximate eigenvalue, they divide both the residual and the within-block spectral distance by ζ j \zeta_j ζ j , leaving their ratio unchanged.
Proposition 3.8. Under Proposition 1.6 , let B j B_j B j be stored real symmetric approximations to K ^ j \widehat{K}_j K j , with the same band structure, satisfying
∥ B j − K ^ j ∥ 2 ≤ e j , e j ≥ 0 , j = R , S . \lVert B_j-\widehat{K}_j\rVert_2\le e_j,\qquad e_j\ge0,\qquad j=R,S. ∥ B j − K j ∥ 2 ≤ e j , e j ≥ 0 , j = R , S . Suppose a complete ordered list of closed intervals I j , k I_{j,k} I j , k , with endpoints a j , k , b j , k a_{j,k},b_{j,k} a j , k , b j , k , each containing exactly one eigenvalue of B j B_j B j , has been certified. For tridiagonal B j B_j B j , use (27) . Put
γ j = min 1 ≤ k < N j ( a j , k + 1 − b j , k ) , \gamma_j=\min_{1\le k<N_j}(a_{j,k+1}-b_{j,k}), γ j = 1 ≤ k < N j min ( a j , k + 1 − b j , k ) , with γ j = ∞ \gamma_j=\infty γ j = ∞ for a scalar block. If
γ j > 2 e j ( j = R , S ) , e R + e S < 2 , \gamma_j>2e_j\quad(j=R,S),\qquad e_R+e_S<2, γ j > 2 e j ( j = R , S ) , e R + e S < 2 , then L ~ = Q T ( B R ⊕ B S ) Q \widetilde{L}=Q^{\mathsf T}(B_R\oplus B_S)Q L = Q T ( B R ⊕ B S ) Q has simple spectrum, and
gap ( L ~ ) ≥ min { γ R , γ S , 2 − e R − e S } , gap ( L ^ ) ≥ min { γ R − 2 e R , γ S − 2 e S , 2 } . (38) \begin{aligned}
\operatorname{gap}(\widetilde{L})&\ge\min\{\gamma_R,\gamma_S,2-e_R-e_S\},\\
\operatorname{gap}(\widehat{L})&\ge\min\{\gamma_R-2e_R,\gamma_S-2e_S,2\}. \tag*{(38)}
\end{aligned} gap ( L ) gap ( L ) ≥ min { γ R , γ S , 2 − e R − e S } , ≥ min { γ R − 2 e R , γ S − 2 e S , 2 } . (38) In a block of order at least two, for a computed pair ( ξ , z ) (\xi,z) ( ξ , z ) with ξ ∈ I j , i \xi\in I_{j,i} ξ ∈ I j , i and z ≠ 0 z\ne0 z = 0 , take certified bounds
ρ ≥ ∥ B j z − ξ z ∥ 2 ∥ z ∥ 2 + e j , 0 < g ≤ min k ≠ i dist ( ξ , I j , k ) − e j . (39) \rho\ge\frac{\lVert B_jz-\xi z\rVert_2}{\lVert z\rVert_2}+e_j,
\qquad
0<g\le\min_{k\ne i}\operatorname{dist}(\xi,I_{j,k})-e_j. \tag*{(39)} ρ ≥ ∥ z ∥ 2 ∥ B j z − ξ z ∥ 2 + e j , 0 < g ≤ k = i min dist ( ξ , I j , k ) − e j . (39) The angle and positive-form bounds of Proposition 3.7 then apply with these ρ , g \rho,g ρ , g . For a scalar block the angle is zero and the positive-form bound uses B = e Q B=e_Q B = e Q . Construction, normalization and rounding errors are included in e j e_j e j ; residual-evaluation errors are included in ρ \rho ρ . The stored matrix need not commute exactly with M τ ( χ ) M_\tau(\chi) M τ ( χ ) , but, with ω j \omega_j ω j as in Proposition 3.7 ,
∥ L ~ M τ − M τ L ~ ∥ 2 ≤ max j = R , S ( ω j e j ) ≤ max { e R , e S } . (40) \lVert\widetilde{L}M_\tau-M_\tau\widetilde{L}\rVert_2\le\max_{j=R,S}(\omega_je_j)\le\max\{e_R,e_S\}. \tag*{(40)} ∥ L M τ − M τ L ∥ 2 ≤ j = R , S max ( ω j e j ) ≤ max { e R , e S } . (40)
Proof. The ordered eigenvalues of K ^ j \widehat{K}_j K j differ from those of B j B_j B j by at most e j e_j e j . Their intervals expanded by e j e_j e j therefore remain disjoint, proving the exact-block gap bounds. The stored R R R spectrum lies between − 3 − e R -3-e_R − 3 − e R and − 1 + e R -1+e_R − 1 + e R , and the stored S S S spectrum between 1 − e S 1-e_S 1 − e S and 3 + e S 3+e_S 3 + e S . This proves (38) . The residual for the exact block is bounded by ρ \rho ρ , and its distance to every non-target eigenvalue is bounded below by g g g , proving (39) . For the commutator, use the alternating decomposition (6) . In each scalar block, center G j G_j G j at the midpoint of its spectral extremes. Its centered norm is ω j / 2 \omega_j/2 ω j /2 , so the commutator norm is at most ω j e j \omega_je_j ω j e j .
For tridiagonal data, a certified upper bound for e j e_j e j is the largest row sum of the three entry-error bounds. All quantities in this certificate can therefore be obtained from recurrence enclosures and sign counts. If an exact eigenvector of the stored block is used, the sine of its angle from the corresponding exact eigenvector is at most e j / ( γ j − e j ) e_j/(\gamma_j-e_j) e j / ( γ j − e j ) .
Corollary 3.9. Under the input assumptions of Algorithm 1, fix a finite N N N . Suppose the initial constants, recurrence coefficients, scalar commutant entries and positive-form evaluations at computable coefficient vectors admit computable enclosures of arbitrarily small width. With scalar eigenpairs refined by sign counts and the tridiagonal recurrence, Algorithm 1 terminates for every strictly positive target, every ϵ > 0 \epsilon> 0 ϵ > 0 and every requested ϵ v > 0 \epsilon_v > 0 ϵ v > 0 . After a sign choice, the reconstructed unit eigenpolynomial, represented in the exact orthonormal bases, has L 2 L^2 L 2 distance at most 2 ϵ v \sqrt{2}\epsilon_v 2 ϵ v from its exact counterpart, for every τ \tau τ .
Proof. Each scalar eigenvalue is simple and has positive distance from its neighbors. Sign counts isolate each eigenvalue, and the recurrence refines its eigenvector. The certified residual tends to zero while the non-target distance remains positive. Matrix and integration errors also tend to zero. Thus ρ / g → 0 \rho/g \to0 ρ / g → 0 and B → 0 B \to0 B → 0 , whereas d ^ → d > 0 \widehat{d} \to d > 0 d → d > 0 . Both stopping tests eventually hold. For unit vectors with acute angle θ \theta θ , the norm distance after sign choice is 2 sin ( θ / 2 ) ≤ 2 sin θ 2\sin(\theta/2) \le\sqrt{2}\sin\theta 2 sin ( θ /2 ) ≤ 2 sin θ . The alternating map preserves this distance by (8). □
This is an input-dependent finite-dimensional assertion. It gives no uniform working precision as N N N grows. Tied concentration values do not affect the scalar eigenpair certificates. Taking maxima or minima of the respective enclosure endpoints also encloses the extremal value without identifying a unique extremal index.
For (19), write d ν R ( t ) = d ν ( t ) d\nu_R(t)=d\nu(t) d ν R ( t ) = d ν ( t ) , d ν S ( t ) = t d ν ( t ) d\nu_S(t)=t\,d\nu(t) d ν S ( t ) = t d ν ( t ) and let p i p_i p i be the normalized scalar eigenpolynomial with coefficient vector v i v_i v i in the R ^ \widehat{R} R or S ^ \widehat{S} S basis. Then
λ i = ∫ 0 s ∣ p i ( t ) ∣ 2 d ν j ( t ) , 1 − λ i = ∫ s b ∣ p i ( t ) ∣ 2 d ν j ( t ) , j = R , S . (41) \lambda_i=\int_0^s \lvert p_i(t)\rvert^2\,d\nu_j(t),\qquad1-\lambda_i=\int_s^b \lvert p_i(t)\rvert^2\,d\nu_j(t),\qquad j=R,S. \tag*{(41)} λ i = ∫ 0 s ∣ p i ( t ) ∣ 2 d ν j ( t ) , 1 − λ i = ∫ s b ∣ p i ( t ) ∣ 2 d ν j ( t ) , j = R , S . (41)
Each integral has a positive N j N_j N j -node quadrature rule exact through degree 2 N j − 1 2N_j-1 2 N j − 1 [11 Section 3.5(v)] , whereas deg p i 2 ≤ 2 N j − 2 \deg p_i^2\le2N_j-2 deg p i 2 ≤ 2 N j − 2 . Summing (41) over an orthonormal eigenbasis gives tr G j \operatorname{tr}G_j tr G j and tr ( I − G j ) \operatorname{tr}(I-G_j) tr ( I − G j ) . To compute these integrals without restricted-weight moment matrices, put β R = β \beta_R=\beta β R = β , β S = β + 1 \beta_S=\beta+1 β S = β + 1 . For Jacobi input and a scalar polynomial p p p , use
∫ 0 s p ( t ) 2 t β j ( 1 − t ) α d t = s β j + 1 ∫ 0 1 u β j ( 1 − s u ) α p ( s u ) 2 d u . (42) \int_0^s p(t)^2t^{\beta_j}(1-t)^\alpha\,dt=s^{\beta_j+1}\int_0^1u^{\beta_j}(1-su)^\alpha p(su)^2\,du. \tag*{(42)} ∫ 0 s p ( t ) 2 t β j ( 1 − t ) α d t = s β j + 1 ∫ 0 1 u β j ( 1 − s u ) α p ( s u ) 2 d u . (42)
For the weight u β j u^{\beta_j} u β j , the nodes are the eigenvalues of its tridiagonal recurrence matrix. If α = d \alpha=d α = d is a nonnegative integer, m m m nodes are exact when 2 m − 1 ≥ 2 deg p + d 2m-1\ge2\deg p+d 2 m − 1 ≥ 2 deg p + d . For Laguerre deficits, use
∫ s ∞ p ( t ) 2 t β j e − t d t = e − s ∫ 0 ∞ e − u ( s + u ) β j p ( s + u ) 2 d u (43) \int_s^\infty p(t)^2t^{\beta_j}e^{-t}\,dt=e^{-s}\int_0^\infty e^{-u}(s+u)^{\beta_j}p(s+u)^2\,du \tag*{(43)} ∫ s ∞ p ( t ) 2 t β j e − t d t = e − s ∫ 0 ∞ e − u ( s + u ) β j p ( s + u ) 2 d u (43)
and quadrature for the weight e − u e^{-u} e − u . The weights are w ℓ − 1 = ∑ k = 0 m − 1 π k ( u ℓ ) 2 w_\ell^{-1}=\sum_{k=0}^{m-1}\pi_k(u_\ell)^2 w ℓ − 1 = ∑ k = 0 m − 1 π k ( u ℓ ) 2 , where ( π k ) (\pi_k) ( π k ) is orthonormal for the base quadrature weight [11 Sections 3.5(v)–(vi)] . The error e Q e_Q e Q includes the quadrature remainder when the rule is not exact, as well as node, weight and recurrence-evaluation errors. For Meixner, the interior integral is the finite positive sum in Example 3.1. Deficits are integrated directly, without subtracting a rounded value from one.
Proposition 3.10. Let a , c > − 1 a,c>-1 a , c > − 1 , 0 < h < 1 0<h<1 0 < h < 1 , and let p p p be a nonzero real polynomial of degree at most d d d . Set
J ( p ) = ∫ 0 1 u c ( 1 − h u ) a p ( u ) 2 d u , J 0 ( p ) = ∫ 0 1 u c p ( u ) 2 d u . J(p)=\int_0^1u^c(1-hu)^a p(u)^2\,du,\qquad J_0(p)=\int_0^1u^c p(u)^2\,du. J ( p ) = ∫ 0 1 u c ( 1 − h u ) a p ( u ) 2 d u , J 0 ( p ) = ∫ 0 1 u c p ( u ) 2 d u .
Let ( u ℓ , w ℓ ) ℓ = 1 m (u_\ell,w_\ell)_{\ell=1}^{m} ( u ℓ , w ℓ ) ℓ = 1 m , with 0 < u ℓ < 1 0<u_\ell<1 0 < u ℓ < 1 , be the positive quadrature rule for u c d u u^{c}\,du u c d u exact through degree 2 m − 1 2m-1 2 m − 1 , and put
Q m ( p ) = ∑ ℓ = 1 m w ℓ ( 1 − h u ℓ ) a p ( u ℓ ) 2 . Q_m(p)=\sum_{\ell=1}^{m}w_\ell(1-hu_\ell)^a p(u_\ell)^2. Q m ( p ) = ℓ = 1 ∑ m w ℓ ( 1 − h u ℓ ) a p ( u ℓ ) 2 .
Choose an integer K ≥ 0 K\geq0 K ≥ 0 with K + 1 > a K+1>a K + 1 > a and 2 m − 1 ≥ K + 2 d 2m-1\geq K+2d 2 m − 1 ≥ K + 2 d , and define
E K = ∣ ( − a ) K + 1 ∣ h K + 1 ( K + 1 ) ! ( 1 − h ) , ℓ a = min { 1 , ( 1 − h ) a } . E_K=\frac{|(-a)_{K+1}|h^{K+1}}{(K+1)!(1-h)},\qquad\ell_a=\min\{1,(1-h)^a\}. E K = ( K + 1 )! ( 1 − h ) ∣ ( − a ) K + 1 ∣ h K + 1 , ℓ a = min { 1 , ( 1 − h ) a } .
Then J 0 ( p ) = ∑ ℓ w ℓ p ( u ℓ ) 2 J_0(p)=\sum_\ell w_\ell p(u_\ell)^2 J 0 ( p ) = ∑ ℓ w ℓ p ( u ℓ ) 2 and
∣ J ( p ) − Q m ( p ) ∣ ≤ 2 E K J 0 ( p ) , ∣ Q m ( p ) J ( p ) − 1 ∣ ≤ 2 E K ℓ a . (44) |J(p)-Q_m(p)|\leq2E_KJ_0(p),\qquad\left|\frac{Q_m(p)}{J(p)}-1\right|\leq\frac{2E_K}{\ell_a}. \tag*{(44)} ∣ J ( p ) − Q m ( p ) ∣ ≤ 2 E K J 0 ( p ) , J ( p ) Q m ( p ) − 1 ≤ ℓ a 2 E K . (44)
If a a a is a nonnegative integer and K ≥ a K\geq a K ≥ a , the rule is exact.
Proof. In the binomial series ( 1 − h u ) a = ∑ k ≥ 0 ( − a ) k ( h u ) k / k ! (1-hu)^a=\sum_{k\geq0}(-a)_k(hu)^k/k! ( 1 − h u ) a = ∑ k ≥ 0 ( − a ) k ( h u ) k / k ! , the absolute ratios of successive tail terms are at most h h h once k ≥ K + 1 > a k\geq K+1>a k ≥ K + 1 > a . The remainder after degree K K K is therefore bounded uniformly on 0 ≤ u ≤ 1 0\leq u\leq1 0 ≤ u ≤ 1 by E K E_K E K . The rule integrates the truncated factor times p 2 p^2 p 2 exactly. Bounding the remainder in both the integral and the quadrature gives 2 E K J 0 ( p ) 2E_KJ_0(p) 2 E K J 0 ( p ) . Since ( 1 − h u ) a ≥ ℓ a (1-hu)^a\geq\ell_a ( 1 − h u ) a ≥ ℓ a , the relative bound follows. □
For Jacobi input, the proposition applies to (42) . Deficits have the direct form
∫ s 1 p ( t ) 2 t 2 β j ( 1 − t ) α d t = ( 1 − s ) α + 1 ∫ 0 1 u α ( 1 − ( 1 − s ) u ) β j p ( 1 − ( 1 − s ) u ) 2 d u . (45) \int_s^1 p(t)^2t^{2\beta_j}(1-t)^\alpha\,dt
=(1-s)^{\alpha+1}\int_0^1 u^\alpha\bigl(1-(1-s)u\bigr)^{\beta_j}p\bigl(1-(1-s)u\bigr)^2\,du. \tag*{(45)} ∫ s 1 p ( t ) 2 t 2 β j ( 1 − t ) α d t = ( 1 − s ) α + 1 ∫ 0 1 u α ( 1 − ( 1 − s ) u ) β j p ( 1 − ( 1 − s ) u ) 2 d u . (45)
The same bound applies with h = 1 − s h=1-s h = 1 − s , a = β j a=\beta_j a = β j and c = α c=\alpha c = α . It is independent of the polynomial coefficients and the magnitude of the positive form. Node, weight and evaluation errors must still be added to obtain e Q e_Q e Q in (37) .
Example 3.11. Take N = 3 N=3 N = 3 , μ = 4 \mu=4 μ = 4 and s = 2 + ε s=2+\varepsilon s = 2 + ε , 0 < ε < 3 / 2 0<\varepsilon<3/2 0 < ε < 3/2 . The unmodified Hermite matrix and its relevant gaps are
L = ( 9 0 3 2 0 3 + ε 0 3 2 0 6 + 2 ε ) , D = 81 − 12 ε + 4 ε 2 , L=\begin{pmatrix}
9&0&3\sqrt{2}\\
0&3+\varepsilon&0\\
3\sqrt{2}&0&6+2\varepsilon
\end{pmatrix},
\qquad D=\sqrt{81-12\varepsilon+4\varepsilon^2}, L = 9 0 3 2 0 3 + ε 0 3 2 0 6 + 2 ε , D = 81 − 12 ε + 4 ε 2 , δ = 2 ε ( 3 − ε ) 9 + D ∼ ε 3 . \delta=\frac{2\varepsilon(3-\varepsilon)}{9+D}\sim\frac{\varepsilon}{3}. δ = 9 + D 2 ε ( 3 − ε ) ∼ 3 ε . Here D D D is the even-block gap and δ \delta δ separates its lower eigenvalue from the odd eigenvalue. Let v v v be the corresponding unit even vector and let e e e be the odd coordinate vector. For η > 0 \eta>0 η > 0 , the symmetric perturbation η ( v e T + e v T ) \eta(ve^{\mathsf T}+ev^{\mathsf T}) η ( v e T + e v T ) has norm η \eta η and rotates the nearby pair through an angle θ \theta θ satisfying tan ( 2 θ ) = 2 η / δ \tan(2\theta)=2\eta/\delta tan ( 2 θ ) = 2 η / δ , 0 < θ < π / 4 0<\theta<\pi/4 0 < θ < π /4 . At ε = 10 − 12 \varepsilon=10^{-12} ε = 1 0 − 12 and η = 10 − 14 \eta=10^{-14} η = 1 0 − 14 , sin θ > 0.02995 \sin\theta>0.02995 sin θ > 0.02995 . Every parity-preserving perturbation of norm at most η \eta η , however, has even-vector sine error at most η / ( D − η ) < 1.112 × 10 − 15 \eta/(D-\eta)<1.112\times10^{-15} η / ( D − η ) < 1.112 × 1 0 − 15 ; the odd vector is unchanged.
Choose ζ R = ( 9 + 3 2 ) / 2 \zeta_R=(9+3\sqrt{2})/2 ζ R = ( 9 + 3 2 ) /2 , ζ S = 1 \zeta_S=1 ζ S = 1 in (14) . The normalized full gap is D / ( 9 + 3 2 ) > 0.6796 D/(9+3\sqrt{2})>0.6796 D / ( 9 + 3 2 ) > 0.6796 at these parameters. Even an arbitrary symmetric perturbation of norm at most η \eta η then gives sine error below 1.472 × 10 − 14 1.472\times10^{-14} 1.472 × 1 0 − 14 . These compare equal absolute perturbation norms. A full-matrix solver that preserves parity also has the block protection; the example quantifies sensitivity rather than a universal difference between solvers.
Example 3.12. Take the N = 12 N = 12 N = 12 examples of [3 Section 4] : Laguerre input ( β , s ) = ( 7 / 2 , 49 ) (\beta, s) = (7/2, 49) ( β , s ) = ( 7/2 , 49 ) and Jacobi input ( α , β , s ) = ( 1 , 1 / 2 , 1 / 256 ) (\alpha, \beta, s) = (1, 1/2, 1/256) ( α , β , s ) = ( 1 , 1/2 , 1/256 ) . We compare full diagonalization of L = Q T ( L R ⊕ L S ) Q L = Q^{\mathsf{T}}(L_R \oplus L_S)Q L = Q T ( L R ⊕ L S ) Q with its two blocks (30) , and write M = M 0 ( χ ) M = M_0(\chi) M = M 0 ( χ ) . The second block uses β + 1 \beta+ 1 β + 1 .
After rounding 100-digit L to binary64 and extracting its blocks, both solves used LAPACK dsyev (Apple Accelerate, lower triangle). Reference eigenvectors of the unrounded L, normalization, polynomial bases and integration use 100 digits. Vectors are compared in R N \mathbb{R}^{N} R N , with block vectors embedded using Q T Q^{\mathsf{T}} Q T . Adaptive quadrature in (41) was checked against incomplete-gamma or incomplete-beta moments; all values were checked at 130 digits.
Let v i , v ^ i v_i,\hat{v}_i v i , v ^ i be the unit reference and computed eigenvectors, matched by their eigenvalues of L. Set E v = max i sin ∠ ( v ^ i , v i ) E_v = \max_i \sin\angle(\hat{v}_i,v_i) E v = max i sin ∠ ( v ^ i , v i ) and E r e l = ∣ d ^ / d − 1 ∣ E_{\mathrm{rel}} = |\hat{d}/d - 1| E rel = ∣ d ^ / d − 1∣ . Here d = 1 − λ max ( M ) d = 1 - \lambda_{\max}(M) d = 1 − λ m a x ( M ) for Hermite and d = λ min ( M ) d = \lambda_{\min}(M) d = λ m i n ( M ) for Gegenbauer, with reference values
d H = 1.09428890155 … × 10 − 29 , d G = 2.93714201786 … × 10 − 35 . d_{\mathrm{H}} = 1.09428890155\ldots\times10^{-29}, \qquad d_{\mathrm{G}} = 2.93714201786\ldots\times10^{-35}. d H = 1.09428890155 … × 1 0 − 29 , d G = 2.93714201786 … × 1 0 − 35 . We evaluate d ^ \hat{d} d ^ by the appropriate positive integral in (41) , at the computed vector matched to the reference extremal eigenpair. For full-L vectors, both scalar components are included. The errors are
Table 1 (PDF p. 21).
The separate-block solves give smaller E v E_v E v in both tests. For the smallest Gegenbauer eigenvalue, they reduce E r e l E_{\mathrm{rel}} E rel from 116% to 1.01%; the full solve gives the smaller relative error in the Hermite deficit.
Example 3.13. For Jacobi input ( α , β ) = ( 1 / 2 , 1 / 4 ) (\alpha,\beta) = (1/2, 1/4) ( α , β ) = ( 1/2 , 1/4 ) and scalar order 14, direct binary64 Cholesky factorization of the monomial moment matrix fails. The basis defined by binary64 recurrence data has Frobenius orthogonality error 3.52 × 10 − 15 3.52 \times10^{-15} 3.52 × 1 0 − 15 , checked with 100-digit inner products.
For the N = 12 N = 12 N = 12 Gegenbauer example above, Algorithm 1 computed the scalar eigenvectors at 30 decimal digits and verified the residuals and positive forms by outward-rounded interval arithmetic at 90 digits. The nodes were enclosed by sign counts; weights were evaluated by the reciprocal sums above. Since α = 1 \alpha= 1 α = 1 and deg p ≤ 5 \deg p \le5 deg p ≤ 5 , the integrand in (42) , apart from u β u^{\beta} u β , has degree at most 11. Six nodes per block therefore suffice exactly. All twelve concentration values were enclosed, isolating the minimum. For its computed value
d ^ = 2.93714201785750961572874550537485 … × 10 − 35 , \hat{d} = 2.93714201785750961572874550537485\ldots\times10^{-35}, d ^ = 2.93714201785750961572874550537485 … × 1 0 − 35 , the verified bounds, with Ω S = 1 \Omega_S = 1 Ω S = 1 , were
ρ < 1.080 × 10 − 29 , g > 11.476 , e Q < 2.20 × 10 − 86 , \rho< 1.080 \times10^{-29}, \qquad g > 11.476, \qquad e_Q < 2.20 \times10^{-86}, ρ < 1.080 × 1 0 − 29 , g > 11.476 , e Q < 2.20 × 1 0 − 86 , B < 8.86 × 10 − 61 , ∣ d ^ λ min ( M ) − 1 ∣ < 3.02 × 10 − 26 . B < 8.86 \times10^{-61}, \qquad\left|\frac{\hat{d}}{\lambda_{\min}(M)} - 1\right| < 3.02 \times10^{-26}. B < 8.86 × 1 0 − 61 , λ m i n ( M ) d ^ − 1 < 3.02 × 1 0 − 26 . This computation uses higher precision than the preceding binary64 eigensolves. The tolerance is verified by (37) ; the alternating reconstruction transports the enclosure to every τ \tau τ . It does not change the intrinsic conditioning of the scalar problem.
Example 3.14. For N = 24 N = 24 N = 24 , we tested noninteger Jacobi parameters near − 1 -1 − 1 and cutoffs near the endpoints. Scalar eigenvectors were computed at 50 decimal digits; outward interval arithmetic at 110 digits certified their residuals, all within-block eigenvalue separations and the positive forms. Proposition 3.10 bounds the quadrature remainder. Every concentration value, or deficit, was enclosed before isolating the minimum.
The first two rows give d = λ min ( M ) d = \lambda_{\min}(\mathrm{M}) d = λ m i n ( M ) and the last gives d = 1 − λ max ( M ) d = 1 - \lambda_{\max}(\mathrm{M}) d = 1 − λ m a x ( M ) . Leading digits of d ^ \widehat{d} d are shown; the relative bounds apply to the complete computed values.
Table 2.
Here m R , m S m_R,m_S m R , m S are quadrature-node counts. The truncation degrees K K K were ( 108 , 108 ) (108,108) ( 108 , 108 ) , ( 26 , 26 ) (26,26) ( 26 , 26 ) and ( 27 , 25 ) (27,25) ( 27 , 25 ) , respectively. An independent calculation at 165 digits using incomplete-beta moments lay inside each certified enclosure. The reconstruction transports these bounds to every τ \tau τ .
Example 3.15. The normalized construction was also tested at larger dimensions for the original Hermite parameters μ = 4 \mu= 4 μ = 4 , s = 49 s = 49 s = 49 , and Jacobi parameters α = 1 \alpha= 1 α = 1 , β = 1 / 2 \beta= 1/2 β = 1/2 , s = 1 / 256 s = 1/256 s = 1/256 . Here both routes diagonalize the same L ^ \widehat{\mathrm{L}} L from (14) : the full route uses the symmetric band driver DSBEV with half-bandwidth two, and the scalar route uses two calls to the tridiagonal driver DSTEV, all in binary64. Outward interval arithmetic at 80 decimal digits, applied to the exact recurrence coefficients, encloses the matrix construction error, the residual of every returned vector, and all scalar eigenvalues using sign counts. Let v i v_i v i be the exact unit eigenvectors of L ^ \widehat{\mathrm{L}} L and v ^ i \widehat{v}_i v i the normalized computed vectors with the corresponding ordered labels. The table gives upper bounds for the matrix norm and eigenvector sine error, and lower bounds for the smallest gap.
Table 3.
The two drivers returned identical binary64 eigenpairs, up to signs, on these six inputs, so these tests show no accuracy advantage for the scalar route over an efficient banded driver. The scalar representation uses 2 N − 2 2N - 2 2 N − 2 diagonal and off-diagonal entries, compared with 3 N 3N 3 N allocated entries in the band format; retaining the two eigenvector matrices in scalar form uses N R 2 + N S 2 N_R^2 + N_S^2 N R 2 + N S 2 entries, compared with N 2 N^2 N 2 for the full eigenvector matrix. The certificates concern eigenvectors and gaps; no relative accuracy for the extreme concentration values is asserted at these larger dimensions. By (8) , the angle bounds also hold for the reconstructed unit eigenpolynomials for every τ \tau τ .
At N = 256 N = 256 N = 256 , median times over nine runs on Apple Accelerate were 4.36 versus 2.21 milliseconds for Hermite and 4.06 versus 2.18 for Jacobi, for the full and block routes respectively. Both wrappers include allocation and copies; the block route also interlaces its vectors. At N = 64 N = 64 N = 64 the timings were essentially equal. These timings exclude recurrence construction and interval verification, and are specific to the inputs, drivers and platform.
Using the binary64 protocol of Example 3.12, Figure 1 compares four inputs at N = 6 , 9 , 12 , 17 , 24 N = 6, 9, 12, 17, 24 N = 6 , 9 , 12 , 17 , 24 . The blocks give smaller E v E_v E v in all twenty tests, by factors between 1.20 and 8.53; their largest error is 5.44 × 10 − 15 5.44 \times10^{-15} 5.44 × 1 0 − 15 .
Figure 1. Eigenvector errors computed here for full (horizontal) and separate-block (vertical) diagonalizations of the commuting matrices in [3 ] . Both use identical binary64 data and the same QR driver. Points below the diagonal have smaller E v E_v E v for the block solves.
The ordering of E v E_v E v depends on the input. With the same protocol, N = 5 N = 5 N = 5 and Jacobi input ( α , β , s ) = ( 1 / 2 , 1 / 4 , 1 / 4 ) (\alpha, \beta, s) = (1/2, 1/4, 1/4) ( α , β , s ) = ( 1/2 , 1/4 , 1/4 ) , the full solve gives E v = 2.85 × 10 − 16 E_v = 2.85 \times10^{-16} E v = 2.85 × 1 0 − 16 , compared with 3.00 × 10 − 16 3.00 \times10^{-16} 3.00 × 1 0 − 16 for the block solves. These values were checked at 100 and 130 digits.
For the Jacobi input ( α , β , s ) = ( 1 / 2 , 1 / 4 , 1 / 4 ) (\alpha, \beta, s) = (1/2, 1/4, 1/4) ( α , β , s ) = ( 1/2 , 1/4 , 1/4 ) , N = 9 N = 9 N = 9 and τ = 1 \tau= 1 τ = 1 , we also integrated the extremals of the two scalar blocks, reconstructed in x x x . The density is positive on ( − 2 , − 1 ) ∪ ( 1 , 2 ) (-\sqrt{2}, -1) \cup(1, \sqrt{2}) ( − 2 , − 1 ) ∪ ( 1 , 2 ) , and the window is ( − 5 / 2 , − 1 ) ∪ ( 1 , 5 / 2 ) (-\sqrt{5}/2, -1) \cup(1, \sqrt{5}/2) ( − 5 /2 , − 1 ) ∪ ( 1 , 5 /2 ) . The largest block eigenvalues are
λ R = λ max ( G R ) = 0.997377870710549 … \lambda_R = \lambda_{\max}(G_R) = 0.997377870710549\ldots λ R = λ m a x ( G R ) = 0.997377870710549 …
λ S = λ max ( G S ) = 0.917749195280867 … \lambda_S = \lambda_{\max}(G_S) = 0.917749195280867\ldots λ S = λ m a x ( G S ) = 0.917749195280867 …
At 100 digits, the reconstructed norms and retained energies agree with 1 and these values within 10 − 95 10^{-95} 1 0 − 95 .
Acknowledgements
The author is deeply grateful to one of his thesis supervisors, Kenier Castillo, for teaching him to seek simplicity and generality through duality. Castillo’s handwritten notes on this problem already contained the central ideas developed here. He declined coauthorship. The author was supported by FAPESP, grants 2024/01546-1 and 2024/23674-1.
Declaration on the Use of Artificial Intelligence
ChatGPT assisted with writing. Responsibility for the final manuscript rests with the author. Further extensions may be explored with AI when a specific problem calls for them. Our aim is to use the alternating map to identify the common structure behind a broad range of problems.
Department of Mathematics and Computer Science, São Paulo State University (UNESP), Presidente Prudente, SP, Brazil
Email address: lucas.tertuliano@unesp.br