Introduction
Perturbations of the standard exponential system naturally produce sinc kernels and discrete operators whose shifts depend on the frequency index. Closely related matrix structures also arise when exponential systems are restricted to disconnected spatial domains. A common shifted-sinc mechanism underlies these problems, but the corresponding operators encode different geometric information. One of the main purposes of the present paper is to keep these structures distinct while developing a common operator-theoretic framework for their analysis.
The first object is the genuine Gram operator associated with an index-dependent perturbed exponential family. If the frequencies are
λ n = n + δ n , \lambda_n = n + \delta_n, λ n = n + δ n ,
then the Gram kernel is
sinc π ( n − m + δ n − δ m ) , \operatorname{sinc}_{\pi}(n-m+\delta_n-\delta_m), sinc π ( n − m + δ n − δ m ) ,
and therefore depends simultaneously on the perturbations attached to both indices. This two-index dependence is essential and distinguishes the genuine Gram problem from the auxiliary shifted-sinc model that will be used throughout the paper.
The second object is the sign-modulated shifted-sinc synthesis operator
( N δ a ) m = ∑ n ∈ Z a n sinc π ( n − m + s n δ ) , s n ∈ { − 1 , + 1 } , (N_{\delta}a)_m=\sum_{n\in\mathbb{Z}}a_n\operatorname{sinc}_{\pi}(n-m+s_n\delta),\qquad s_n\in\{-1,+1\}, ( N δ a ) m = n ∈ Z ∑ a n sinc π ( n − m + s n δ ) , s n ∈ { − 1 , + 1 } ,
whose shift depends only on the input index. Equivalently,
N δ = T δ P + + T − δ P − , N_{\delta}=T_{\delta}P_{+}+T_{-\delta}P_{-}, N δ = T δ P + + T − δ P − ,
where P + P_{+} P + and P − P_{-} P − are the coordinate projections determined by the sign pattern. This auxiliary model is not itself the Gram operator of the perturbed family, but it admits an exact block structure that isolates the role of the sign pattern and makes possible a detailed spectral analysis.
A third structure appears for exponential systems generated by unions of shifted lattices on disconnected domains. After folding modulo one, the synthesis problem is represented by finite-dimensional matrix fibers. For unions of intervals of unequal lengths the fiber dimension may vary with the base point, but only finitely many different fibers occur. Thus the infinite-dimensional problem is again governed by explicit finite-dimensional data, now through a geometric rather than a periodic reduction.
These three settings are linked by the same shifted-sinc and Fourier-fiber mechanisms, but they answer different questions. In the perturbative problem the main issue is how the arrangement of local frequency shifts affects stability. In the disconnected-domain problem the relevant information is encoded by the integer fibers of the spatial set and by the corresponding phase matrices. The results below show that periodicity or finite-cell geometry can collapse either problem to exact finite-dimensional calculations, whereas genuinely nonperiodic configurations retain an essentially infinite-dimensional component.
Main results and organization
The first part of the paper concerns perturbed exponential systems and local frequency shifts. Section 2 introduces the genuine Gram framework and records the distinction between the actual Gram operator and the auxiliary models used later. Section 3 develops the shifted-sinc operators
showing that they form a unitary group on ℓ 2 ( Z ) \ell^2(\mathbb{Z}) ℓ 2 ( Z ) . This group structure provides the common operator language used throughout the perturbative part of the paper.
Section 4 then studies the auxiliary binary operator
N δ = T δ P + + T − δ P − . N_{\delta}=T_{\delta}P_{+}+T_{-\delta}P_{-}. N δ = T δ P + + T − δ P − .
Besides estimates that are uniform over all sign patterns, the fixed-pattern problem admits an exact description. Section 5 proves that, if
ρ δ , s = ∥ P − T 2 δ P + ∥ 2 , \rho_{\delta,s}=\frac{\lVert P_{-}T_{2\delta}P_{+}\rVert}{2}, ρ δ , s = 2 ∥ P − T 2 δ P + ∥ ,
then
inf σ ( N δ ∗ N δ ) = 1 − ρ δ , s , sup σ ( N δ ∗ N δ ) = 1 + ρ δ , s . \inf\sigma(N_{\delta}^{*}N_{\delta})=1-\rho_{\delta,s},\qquad\sup\sigma(N_{\delta}^{*}N_{\delta})=1+\rho_{\delta,s}. inf σ ( N δ ∗ N δ ) = 1 − ρ δ , s , sup σ ( N δ ∗ N δ ) = 1 + ρ δ , s .
Thus a single pattern-dependent compression norm determines both optimal squared synthesis bounds. In particular, ρ δ , s < 1 \rho_{\delta,s}<1 ρ δ , s < 1 is exactly the lower-stability condition for the synthesis operator onto its closed range.
Periodicity produces a further and stronger reduction. For a p p p -periodic binary pattern, § 6 proves a factorization
N p , δ , s ( ξ ) = H p , δ , s E δ , s ( ξ ) V ∗ . N_{p,\delta,s}(\xi)=H_{p,\delta,s}E_{\delta,s}(\xi)V^{*}. N p , δ , s ( ξ ) = H p , δ , s E δ , s ( ξ ) V ∗ .
in which the last two factors are unitary. Consequently, all singular values are independent of the fiber variable ξ \xi ξ , and the optimal bounds are the extreme eigenvalues of the single finite Gram matrix
H p , δ , s ∗ H p , δ , s . H_{p,\delta,s}^{*}H_{p,\delta,s}. H p , δ , s ∗ H p , δ , s .
The basis property is equivalent to a finite Vandermonde noncollision condition. This gives, in particular, the exact alternating bounds
1 ± sin ( π δ ) 1\pm\sin(\pi\delta) 1 ± sin ( π δ )
and an explicit period-three example with optimal squared bounds
1 ± 5 3 1\pm\frac{\sqrt{5}}{3} 1 ± 3 5
at δ = 1 / 4 \delta=1/4 δ = 1/4 .
The relation between periodic and nonperiodic binary patterns is developed further in § 7.1 . If A b i n ( δ ) A_{\mathrm{bin}}(\delta) A bin ( δ ) denotes the worst lower squared synthesis bound over all binary patterns and A p ( δ ) A_p(\delta) A p ( δ ) the corresponding minimum over p p p -periodic patterns, we prove
A b i n ( δ ) = inf p ≥ 1 A p ( δ ) . A_{\mathrm{bin}}(\delta)=\inf_{p\geq1}A_p(\delta). A bin ( δ ) = p ≥ 1 inf A p ( δ ) .
Thus the global extremal problem can be approximated by finite-dimensional periodic problems, even when no fixed period realizes the worst behavior. At the critical value δ = 1 / 4 \delta=1/4 δ = 1/4 , a step pattern leads to a half-shifted Hilbert matrix and has vanishing lower bound, whereas every fixed periodic binary pattern remains invertible. Hence the critical loss of stability is invisible at every fixed period and emerges only asymptotically as the period tends to infinity. This leads naturally to finite-dimensional extremal questions and to a conjectural logarithmic critical degeneration, motivated by the asymptotics of finite Hilbert matrices.
We then return to the genuine Gram problem. § 8.1 and § 8.2 separate the two mechanisms that enter near the critical regime. A four-block sesquilinear model first isolates the interaction of two sign partitions. The genuine Gram expansion then distinguishes same-sign drift from opposite-sign half-shift interaction and yields support-dependent first-order estimates. § 9 gives a complementary support-independent perturbation theorem around any already stable structured exponential basis. The two results address different aspects of stability: the former retains detailed first-order information about the perturbation, whereas the latter provides a global neighborhood of a fixed stable system.
Section 10 extends the local-shift theory from the binary alphabet { + δ , − δ } \{+\delta,-\delta\} { + δ , − δ } to finitely many shifts t 1 , … , t L t_1,\ldots,t_L t 1 , … , t L . If P j P_j P j denotes the coordinate projection associated with the shift t j t_j t j , the Gram blocks are
G j k = P j T t k − t j P k . (1) G_{jk}=P_jT_{t_k-t_j}P_k. \tag*{(1)} G j k = P j T t k − t j P k . (1)
A matrix of block norms yields a general sufficient lower-stability criterion. The binary case is exceptional because its self-adjoint off-diagonal structure reduces the two spectral endpoints to a single scalar norm; this symmetry does not persist for three or more local shifts. Under periodicity, however, the finite-alphabet problem again reduces exactly to a single constant Vandermonde-type matrix.
The second part of the paper concerns disconnected spatial domains. We begin with two full shifted lattices on a union of two unit intervals. Sections 11.1 and 11.2 treat respectively integer and noninteger translations and reduce the synthesis operator to explicit 2 × 2 2 \times2 2 × 2 matrix fibers. This gives optimal bounds and complete basis criteria in both regimes.
These calculations are instances of a more general folding principle. Section 12 first extends the construction to finitely many translated unit intervals and then to arbitrary finite unions of bounded intervals. For
E = ⋃ j = 1 J [ a j , b j ) , Λ = ⋃ k = 1 K ( Z + δ k ) , E = \bigcup_{j=1}^{J}[a_j,b_j), \qquad\Lambda= \bigcup_{k=1}^{K}(\mathbb{Z}+\delta_k), E = j = 1 ⋃ J [ a j , b j ) , Λ = k = 1 ⋃ K ( Z + δ k ) ,
we associate to each x ∈ [ 0 , 1 ) x \in[0,1) x ∈ [ 0 , 1 ) the integer fiber
Ω x = { m ∈ Z : x + m ∈ E } . \Omega_x = \{m \in\mathbb{Z}: x+m \in E\}. Ω x = { m ∈ Z : x + m ∈ E } .
Only finitely many different sets Ω x \Omega_x Ω x occur on positive-measure cells. On a cell with fiber Ω r \Omega_r Ω r , the synthesis operator is represented by the finite phase matrix
M r = ( e 2 π i δ k m ) m ∈ Ω r , 1 ≤ k ≤ K . M_r = \left(e^{2\pi i\delta_km}\right)_{m\in\Omega_r,\ 1\leq k\leq K}. M r = ( e 2 π i δ k m ) m ∈ Ω r , 1 ≤ k ≤ K .
The extreme singular values of these matrices give the exact quantitative bounds, while their column rank, row rank, and square invertibility characterize respectively Riesz sequences, frames, and Riesz bases. This formulation also makes clear why total measure alone is insufficient: the local fiber multiplicity must be compatible with the number of frequency channels.
The finite-cell formulation has two further applications. In Section 13.1, periodicity of spectra for finite unions of intervals converts spectrality into simultaneous finite Hadamard conditions on the integer fibers. Comparing adjacent fibers then produces arithmetic restrictions on possible spectra and, under suitable transition hypotheses, a rigidity mechanism related to the one-dimensional Fuglede problem. In Section 13.2, a finite collection of rationally commensurate sampling lattices is decomposed into cosets of a common lattice. The resulting multirate system is therefore governed by the same finite-cell phase matrices, giving exact existence criteria, optimal condition numbers, and an explicit reconstruction procedure.
The paper concludes by separating the regimes in which exact finite-dimensional reductions are available from the genuinely nonperiodic problems that remain infinite-dimensional. Among the latter are the geometric characterization of
ρ δ , s = ∥ P − T 2 δ P + ∥ \rho_{\delta,s} = \lVert P_{-}T_{2\delta}P_{+}\rVert ρ δ , s = ∥ P − T 2 δ P + ∥
for arbitrary nonperiodic patterns, the critical asymptotics of the periodic extremal problem, a support-independent near-critical theory for the genuine Gram operator, and extensions of the finite-cell framework beyond rationally commensurate lattice systems.
The perturbed exponential Gram framework
The one-dimensional Gram operator
Let ( δ n ) n ∈ Z (\delta_n)_{n\in\mathbb{Z}} ( δ n ) n ∈ Z be a real sequence. We consider the exponential family
B ( δ ) : = { e 2 π i ( n + δ n ) x } n ∈ Z ⊂ L 2 ( − 1 2 , 1 2 ) . (2) \mathcal{B}(\delta) := \left\{e^{2\pi i(n+\delta_n)x}\right\}_{n\in\mathbb{Z}} \subset L^2\left(-\frac{1}{2},\frac{1}{2}\right). \tag*{(2)} B ( δ ) := { e 2 π i ( n + δ n ) x } n ∈ Z ⊂ L 2 ( − 2 1 , 2 1 ) . (2)
No smallness assumption is imposed at this stage. Whenever a uniform stability result requires
sup n ∈ Z ∣ δ n ∣ < 1 4 , \sup_{n\in\mathbb{Z}}|\delta_n| < \frac{1}{4}, n ∈ Z sup ∣ δ n ∣ < 4 1 ,
this hypothesis will be stated explicitly. For finitely supported coefficient sequences, all Gram identities below are valid without any smallness assumption on ( δ n ) (\delta_n) ( δ n ) .
We use the normalized sinc function
sinc π ( u ) : = { sin ( π u ) π u , u ≠ 0 , 1 , u = 0. \operatorname{sinc}_{\pi}(u) := \begin{cases} \dfrac{\sin(\pi u)}{\pi u}, & u\ne0, \\ 1, & u=0. \end{cases} sinc π ( u ) := ⎩ ⎨ ⎧ π u sin ( π u ) , 1 , u = 0 , u = 0.
Recall that B \mathcal{B} B is a Riesz sequence if there exist constants A , B > 0 A,B>0 A , B > 0 such that, for every finitely supported sequence a = ( a n ) n ∈ Z a=(a_n)_{n\in\mathbb{Z}} a = ( a n ) n ∈ Z ,
A ∑ n ∈ Z ∣ a n ∣ 2 ≤ ∥ ∑ n ∈ Z a n e 2 π i ( n + δ n ) x ∥ L 2 ( − 1 / 2 , 1 / 2 ) 2 ≤ B ∑ n ∈ Z ∣ a n ∣ 2 . (3) A\sum_{n\in\mathbb{Z}}|a_n|^2 \le\left\|\sum_{n\in\mathbb{Z}}a_ne^{2\pi i(n+\delta_n)x}\right\|_{L^2(-1/2,1/2)}^2 \le B\sum_{n\in\mathbb{Z}}|a_n|^2. \tag*{(3)} A n ∈ Z ∑ ∣ a n ∣ 2 ≤ n ∈ Z ∑ a n e 2 π i ( n + δ n ) x L 2 ( − 1/2 , 1/2 ) 2 ≤ B n ∈ Z ∑ ∣ a n ∣ 2 . (3)
Expanding the L 2 L^2 L 2 norm gives
∥ ∑ n ∈ Z a n e 2 π i ( n + δ n ) x ∥ L 2 ( − 1 / 2 , 1 / 2 ) 2 = ∑ n , m ∈ Z a n a m ‾ ∫ − 1 / 2 1 / 2 e 2 π i ( n − m + δ n − δ m ) x d x = ∑ n , m ∈ Z a n a m ‾ sinc π ( n − m + δ n − δ m ) . (4) \begin{aligned}
\left\|\sum_{n\in\mathbb{Z}}a_ne^{2\pi i(n+\delta_n)x}\right\|_{L^2(-1/2,1/2)}^2
= \sum_{n,m\in\mathbb{Z}}a_n\overline{a_m}\int_{-1/2}^{1/2}e^{2\pi i(n-m+\delta_n-\delta_m)x}\,\mathrm{d}x \\
&= \sum_{n,m\in\mathbb{Z}}a_n\overline{a_m}\operatorname{sinc}_{\pi}(n-m+\delta_n-\delta_m). \tag*{(4)}
\end{aligned} n ∈ Z ∑ a n e 2 π i ( n + δ n ) x L 2 ( − 1/2 , 1/2 ) 2 = n , m ∈ Z ∑ a n a m ∫ − 1/2 1/2 e 2 π i ( n − m + δ n − δ m ) x d x = n , m ∈ Z ∑ a n a m sinc π ( n − m + δ n − δ m ) . (4)
Accordingly, the Gram operator G : ℓ 2 ( Z ) → ℓ 2 ( Z ) G:\ell^2(\mathbb{Z})\to\ell^2(\mathbb{Z}) G : ℓ 2 ( Z ) → ℓ 2 ( Z ) is formally given by
( G a ) m : = ∑ n ∈ Z sinc π ( n − m + δ n − δ m ) a n , m ∈ Z , (5) (Ga)_m := \sum_{n\in\mathbb{Z}}\operatorname{sinc}_{\pi}(n-m+\delta_n-\delta_m)a_n,\qquad m\in\mathbb{Z}, \tag*{(5)} ( G a ) m := n ∈ Z ∑ sinc π ( n − m + δ n − δ m ) a n , m ∈ Z , (5)
and (3) reads
∥ ∑ n ∈ Z a n e 2 π i ( n + δ n ) x ∥ L 2 ( − 1 / 2 , 1 / 2 ) 2 = ⟨ G a , a ⟩ ℓ 2 . \left\|\sum_{n\in\mathbb{Z}}a_ne^{2\pi i(n+\delta_n)x}\right\|_{L^2(-1/2,1/2)}^2 = \langle Ga,a\rangle_{\ell^2}. n ∈ Z ∑ a n e 2 π i ( n + δ n ) x L 2 ( − 1/2 , 1/2 ) 2 = ⟨ G a , a ⟩ ℓ 2 .
Since
sinc π ( k + u ) = ( − 1 ) k sin ( π u ) π ( k + u ) , k ∈ Z , u ∈ R , \operatorname{sinc}_{\pi}(k+u)=(-1)^k\frac{\sin(\pi u)}{\pi(k+u)},\qquad k\in\mathbb{Z},\quad u\in\mathbb{R}, sinc π ( k + u ) = ( − 1 ) k π ( k + u ) sin ( π u ) , k ∈ Z , u ∈ R ,
it is convenient to remove the oscillatory factor ( − 1 ) n − m (-1)^{n-m} ( − 1 ) n − m by the unitary modulation
( M a ) n : = ( − 1 ) n a n , n ∈ Z . (Ma)_n:=(-1)^na_n,\qquad n\in\mathbb{Z}. ( M a ) n := ( − 1 ) n a n , n ∈ Z .
Define the unitarily equivalent operator G ~ : = M G M \widetilde{G} := MGM G := M GM . Then
⟨ G a , a ⟩ ℓ 2 = ⟨ G ~ ( M a ) , M a ⟩ ℓ 2 , \langle Ga,a\rangle_{\ell^2}=\langle\widetilde{G}(Ma),Ma\rangle_{\ell^2}, ⟨ G a , a ⟩ ℓ 2 = ⟨ G ( M a ) , M a ⟩ ℓ 2 ,
and the kernel of G ~ \widetilde{G} G is
K ~ m , n = { 1 , n = m , sin ( π ( δ n − δ m ) ) π ( n − m + δ n − δ m ) , n ≠ m . (6) \widetilde{K}_{m,n}=
\begin{cases}
1, & n=m,\\
\dfrac{\sin(\pi(\delta_n-\delta_m))}{\pi(n-m+\delta_n-\delta_m)}, & n\ne m.
\end{cases} \tag*{(6)} K m , n = ⎩ ⎨ ⎧ 1 , π ( n − m + δ n − δ m ) sin ( π ( δ n − δ m )) , n = m , n = m . (6)
Thus the original Gram form and the modulated form are equivalent up to the unitary map M M M .
Remark 2.1. If δ n = s n δ \delta_n=s_n\delta δ n = s n δ with s n ∈ { + 1 , − 1 } s_n\in\{+1,-1\} s n ∈ { + 1 , − 1 } , the genuine Gram kernel becomes
sinc π ( n − m + δ ( s n − s m ) ) . (7) \operatorname{sinc}_{\pi}\bigl(n-m+\delta(s_n-s_m)\bigr). \tag*{(7)} sinc π ( n − m + δ ( s n − s m ) ) . (7)
By contrast, the auxiliary sign operator studied in § 4 has kernel
sinc π ( n − m + s n δ ) . (8) \operatorname{sinc}_{\pi}(n-m+s_n\delta). \tag*{(8)} sinc π ( n − m + s n δ ) . (8)
The first expression depends on both s n s_n s n and s m s_m s m , whereas the second depends only on s n s_n s n . The auxiliary model is motivated by the Gram problem but is not the Gram operator itself.
Multidimensional shifted sinc operators
For θ = ( θ 1 , … , θ d ) ∈ R d \boldsymbol{\theta}=(\theta_1,\ldots,\theta_d)\in\mathbb{R}^d θ = ( θ 1 , … , θ d ) ∈ R d and finitely supported a = ( a n ) n ∈ Z d \mathbf{a}=(a_n)_{n\in\mathbb{Z}^d} a = ( a n ) n ∈ Z d , define
( T θ a ) m : = ∑ n ∈ Z d a n ∏ j = 1 d sinc π ( n j − m j + θ j ) . (9) (T_{\boldsymbol{\theta}}\mathbf{a})_m:=\sum_{n\in\mathbb{Z}^d}a_n\prod_{j=1}^{d}\operatorname{sinc}_{\pi}(n_j-m_j+\theta_j). \tag*{(9)} ( T θ a ) m := n ∈ Z d ∑ a n j = 1 ∏ d sinc π ( n j − m j + θ j ) . (9)
Lemma 2.2. Let Q d = [ − 1 2 , 1 2 ] d Q^d=[-\frac{1}{2},\frac{1}{2}]^d Q d = [ − 2 1 , 2 1 ] d . Let a , b ∈ ℓ 0 2 ( Z d ) \mathbf{a},\mathbf{b}\in\ell^2_0(\mathbb{Z}^d) a , b ∈ ℓ 0 2 ( Z d ) and s , t ∈ R d \mathbf{s},\mathbf{t}\in\mathbb{R}^d s , t ∈ R d . Then
⟨ ∑ n ∈ Z d a n e 2 π i ⟨ n + s , x ⟩ , ∑ m ∈ Z d b m e 2 π i ⟨ m + t , x ⟩ ⟩ L 2 ( Q d ) = ∑ n , m ∈ Z d a n b m ‾ ∏ j = 1 d sinc π ( n j − m j + s j − t j ) = ⟨ T s − t a , b ⟩ ℓ 2 ( Z d ) . (10) \left\langle\sum_{n\in\mathbb{Z}^d}a_n e^{2\pi i\langle n+s,x\rangle},\sum_{m\in\mathbb{Z}^d}b_m e^{2\pi i\langle m+t,x\rangle}\right\rangle_{L^2(Q^d)}
=
\sum_{n,m\in\mathbb{Z}^d}a_n\overline{b_m}\prod_{j=1}^{d}\operatorname{sinc}_{\pi}(n_j-m_j+s_j-t_j)
=
\langle T_{\mathbf{s}-\mathbf{t}}\mathbf{a},\mathbf{b}\rangle_{\ell^2(\mathbb{Z}^d)}. \tag*{(10)} ⟨ n ∈ Z d ∑ a n e 2 π i ⟨ n + s , x ⟩ , m ∈ Z d ∑ b m e 2 π i ⟨ m + t , x ⟩ ⟩ L 2 ( Q d ) = n , m ∈ Z d ∑ a n b m j = 1 ∏ d sinc π ( n j − m j + s j − t j ) = ⟨ T s − t a , b ⟩ ℓ 2 ( Z d ) . (10) Proof. The integral factorizes coordinatewise:
∫ Q d e 2 π i ⟨ n − m + s − t , x ⟩ d x = ∏ j = 1 d ∫ − 1 / 2 1 / 2 e 2 π i ( n j − m j + s j − t j ) x j d x j \int_{Q^d}e^{2\pi i\langle n-m+s-t,x\rangle}\,\mathrm{d}x
=
\prod_{j=1}^{d}\int_{-1/2}^{1/2}e^{2\pi i(n_j-m_j+s_j-t_j)x_j}\,\mathrm{d}x_j ∫ Q d e 2 π i ⟨ n − m + s − t , x ⟩ d x = j = 1 ∏ d ∫ − 1/2 1/2 e 2 π i ( n j − m j + s j − t j ) x j d x j = ∏ j = 1 d sinc π ( n j − m j + s j − t j ) . =
\prod_{j=1}^{d}\operatorname{sinc}_{\pi}(n_j-m_j+s_j-t_j). = j = 1 ∏ d sinc π ( n j − m j + s j − t j ) . Substituting this identity into the scalar product yields (9).
The shifted sinc group
For t ∈ R t \in\mathbb{R} t ∈ R and finitely supported a = ( a n ) n ∈ Z a=(a_n)_{n\in\mathbb{Z}} a = ( a n ) n ∈ Z , define
( T t a ) m : = ∑ n ∈ Z a n sinc π ( n − m + t ) , m ∈ Z . (11) (T_ta)_m := \sum_{n\in\mathbb{Z}} a_n \operatorname{sinc}_{\pi}(n-m+t), \qquad m\in\mathbb{Z}. \tag*{(11)} ( T t a ) m := n ∈ Z ∑ a n sinc π ( n − m + t ) , m ∈ Z . (11)
The one-parameter family of shifted-sinc operators considered here is closely related to the discrete Hilbert-transform group studied by De Carli and Samad [1 ] . More precisely, if D t D_t D t denotes the family introduced in that work and M M M is the unitary modulation ( M a ) n = ( − 1 ) n a n (Ma)_n=(-1)^n a_n ( M a ) n = ( − 1 ) n a n , then
D t = M T − t M , t ∈ R . D_t=MT_{-t}M,\qquad t\in\mathbb{R}. D t = M T − t M , t ∈ R .
Thus the group law, strong continuity, and ℓ 2 \ell^2 ℓ 2 -isometry properties are already known in an equivalent normalization.
We recall the Fourier-multiplier representation and the resulting group properties below, since this formulation is particularly convenient for the pattern-dependent compressions and finite matrix reductions developed in the sequel.
Let F : ℓ 2 ( Z ) → L 2 ( − 1 2 , 1 2 ) \mathcal{F}:\ell^2(\mathbb{Z})\to L^2\left(-\frac{1}{2},\frac{1}{2}\right) F : ℓ 2 ( Z ) → L 2 ( − 2 1 , 2 1 ) be the discrete Fourier transform
( F a ) ( ξ ) = ∑ m ∈ Z a m e − 2 π i m ξ , − 1 2 < ξ < 1 2 . (12) (\mathcal{F}a)(\xi)=\sum_{m\in\mathbb{Z}}a_m e^{-2\pi i m\xi},\qquad-\frac{1}{2}<\xi<\frac{1}{2}. \tag*{(12)} ( F a ) ( ξ ) = m ∈ Z ∑ a m e − 2 π im ξ , − 2 1 < ξ < 2 1 . (12)
Theorem 3.1 (Fourier multiplier representation). For every t ∈ R t\in\mathbb{R} t ∈ R , the operator T t T_t T t extends uniquely to a unitary operator on ℓ 2 ( Z ) \ell^2(\mathbb{Z}) ℓ 2 ( Z ) and
F ( T t a ) ( ξ ) = e − 2 π i t ξ F a ( ξ ) (13) \mathcal{F}(T_ta)(\xi)=e^{-2\pi i t\xi}\mathcal{F}a(\xi) \tag*{(13)} F ( T t a ) ( ξ ) = e − 2 π i t ξ F a ( ξ ) (13) for almost every ξ ∈ ( − 1 2 , 1 2 ) \xi\in\left(-\frac{1}{2},\frac{1}{2}\right) ξ ∈ ( − 2 1 , 2 1 ) .
Proof. Let
h t ( ξ ) = e − 2 π i t ξ , − 1 2 < ξ < 1 2 , h_t(\xi)=e^{-2\pi i t\xi},\qquad-\frac{1}{2}<\xi<\frac{1}{2}, h t ( ξ ) = e − 2 π i t ξ , − 2 1 < ξ < 2 1 ,
and let M h t \mathcal{M}_{h_t} M h t denote multiplication by h t h_t h t on L 2 ( − 1 2 , 1 2 ) L^2\left(-\frac{1}{2},\frac{1}{2}\right) L 2 ( − 2 1 , 2 1 ) . Define
T ~ t : = F − 1 M h t F . \widetilde{T}_t:=\mathcal{F}^{-1}\mathcal{M}_{h_t}\mathcal{F}. T t := F − 1 M h t F .
Since ∣ h t ( ξ ) ∣ = 1 |h_t(\xi)|=1 ∣ h t ( ξ ) ∣ = 1 almost everywhere, T ~ t \widetilde{T}_t T t is unitary on ℓ 2 ( Z ) \ell^2(\mathbb{Z}) ℓ 2 ( Z ) .
It remains to verify that T ~ t \widetilde{T}_t T t agrees with (10) on finitely supported sequences. Let a ∈ ℓ 0 2 ( Z ) a\in\ell^2_0(\mathbb{Z}) a ∈ ℓ 0 2 ( Z ) . For every m ∈ Z m\in\mathbb{Z} m ∈ Z ,
( T ~ t a ) m = ∫ − 1 / 2 1 / 2 e − 2 π i t ξ ( F a ) ( ξ ) e 2 π i m ξ d ξ = ∑ n ∈ Z a n ∫ − 1 / 2 1 / 2 e 2 π i ( m − n − t ) ξ d ξ = ∑ n ∈ Z a n sinc π ( m − n − t ) = ∑ n ∈ Z a n sinc π ( n − m + t ) . \begin{aligned}
(\widetilde{T}_ta)_m
&=\int_{-1/2}^{1/2}e^{-2\pi i t\xi}(\mathcal{F}a)(\xi)e^{2\pi i m\xi}\,d\xi\\
&=\sum_{n\in\mathbb{Z}}a_n\int_{-1/2}^{1/2}e^{2\pi i(m-n-t)\xi}\,d\xi\\
&=\sum_{n\in\mathbb{Z}}a_n\operatorname{sinc}_{\pi}(m-n-t)\\
&=\sum_{n\in\mathbb{Z}}a_n\operatorname{sinc}_{\pi}(n-m+t).
\end{aligned} ( T t a ) m = ∫ − 1/2 1/2 e − 2 π i t ξ ( F a ) ( ξ ) e 2 π im ξ d ξ = n ∈ Z ∑ a n ∫ − 1/2 1/2 e 2 π i ( m − n − t ) ξ d ξ = n ∈ Z ∑ a n sinc π ( m − n − t ) = n ∈ Z ∑ a n sinc π ( n − m + t ) .
e n d s m i d = 1 {ends_mid=1} e n d s m i d = 1
The interchange between the finite sum and the integral is immediate. Thus T ~ t = T t \widetilde{T}_t = T_t T t = T t on ℓ 0 2 ( Z ) \ell_0^2(\mathbb{Z}) ℓ 0 2 ( Z ) , and consequently
F ( T t a ) ( ξ ) = e − 2 π i t ξ F a ( ξ ) . \mathcal{F}(T_ta)(\xi) = e^{-2\pi it\xi}\mathcal{F}a(\xi). F ( T t a ) ( ξ ) = e − 2 π i t ξ F a ( ξ ) .
The uniqueness of the extension follows from the density of ℓ 0 2 ( Z ) \ell_0^2(\mathbb{Z}) ℓ 0 2 ( Z ) in ℓ 2 ( Z ) \ell^2(\mathbb{Z}) ℓ 2 ( Z ) .
Remark 3.2 (A shift check at t = 1 t = 1 t = 1 ). For t = 1 t = 1 t = 1 ,
( T 1 a ) m = ∑ n ∈ Z a n sinc π ( n − m + 1 ) . (14) (T_1a)_m = \sum_{n\in\mathbb{Z}} a_n \operatorname{sinc}_{\pi}(n-m+1). \tag*{(14)} ( T 1 a ) m = n ∈ Z ∑ a n sinc π ( n − m + 1 ) . (14)
Since n − m + 1 n-m+1 n − m + 1 is an integer, the sinc term vanishes unless n − m + 1 = 0 n-m+1 = 0 n − m + 1 = 0 , that is, unless n = m − 1 n = m-1 n = m − 1 . Hence
( T 1 a ) m = a m − 1 . (15) (T_1a)_m = a_{m-1}. \tag*{(15)} ( T 1 a ) m = a m − 1 . (15)
The Fourier transform of this shift is
F ( T 1 a ) ( ξ ) = ∑ m a m − 1 e − 2 π i m ξ = ∑ k a k e − 2 π i ( k + 1 ) ξ = e − 2 π i ξ F a ( ξ ) , \mathcal{F}(T_1a)(\xi) = \sum_m a_{m-1}e^{-2\pi im\xi} = \sum_k a_k e^{-2\pi i(k+1)\xi} = e^{-2\pi i\xi}\mathcal{F}a(\xi), F ( T 1 a ) ( ξ ) = m ∑ a m − 1 e − 2 π im ξ = k ∑ a k e − 2 π i ( k + 1 ) ξ = e − 2 π i ξ F a ( ξ ) ,
in agreement with (12).
Remark 3.3 (Convolution convention). With the standard discrete convolution
( a ∗ k ) m = ∑ n ∈ Z a n k ( m − n ) , (a * k)_m = \sum_{n\in\mathbb{Z}} a_n k(m-n), ( a ∗ k ) m = n ∈ Z ∑ a n k ( m − n ) ,
one has
( T t a ) m = ∑ n a n sinc π ( n − m + t ) = ∑ n a n sinc π ( m − n − t ) = ∑ n a n k − t ( m − n ) , (T_ta)_m = \sum_n a_n \operatorname{sinc}_{\pi}(n-m+t) = \sum_n a_n \operatorname{sinc}_{\pi}(m-n-t) = \sum_n a_n k_{-t}(m-n), ( T t a ) m = n ∑ a n sinc π ( n − m + t ) = n ∑ a n sinc π ( m − n − t ) = n ∑ a n k − t ( m − n ) ,
where
T t a = a ∗ k − t , k − t ( j ) = sinc π ( j − t ) . (16) T_ta = a * k_{-t}, \qquad k_{-t}(j) = \operatorname{sinc}_{\pi}(j-t). \tag*{(16)} T t a = a ∗ k − t , k − t ( j ) = sinc π ( j − t ) . (16)
Proposition 3.4 (Group law, adjoint, and unitarity). For all s , t ∈ R s,t \in\mathbb{R} s , t ∈ R ,
T t T s = T t + s , T t ∗ = T − t , ∥ T t a ∥ ℓ 2 = ∥ a ∥ ℓ 2 . (17) T_tT_s = T_{t+s}, \qquad T_t^* = T_{-t}, \qquad\lVert T_ta\rVert_{\ell^2} = \lVert a\rVert_{\ell^2}. \tag*{(17)} T t T s = T t + s , T t ∗ = T − t , ∥ T t a ∥ ℓ 2 = ∥ a ∥ ℓ 2 . (17) Proof. By Parseval and Theorem 3.1,
F ( T t T s a ) ( ξ ) = e − 2 π i t ξ F ( T s a ) ( ξ ) = e − 2 π i ( t + s ) ξ F a ( ξ ) = F ( T t + s a ) ( ξ ) , \mathcal{F}(T_tT_sa)(\xi) = e^{-2\pi it\xi}\mathcal{F}(T_sa)(\xi) = e^{-2\pi i(t+s)\xi}\mathcal{F}a(\xi) = \mathcal{F}(T_{t+s}a)(\xi), F ( T t T s a ) ( ξ ) = e − 2 π i t ξ F ( T s a ) ( ξ ) = e − 2 π i ( t + s ) ξ F a ( ξ ) = F ( T t + s a ) ( ξ ) , which gives T t T s = T t + s T_tT_s = T_{t+s} T t T s = T t + s . Multiplication by e − 2 π i t ξ e^{-2\pi it\xi} e − 2 π i t ξ has adjoint multiplication by its complex conjugate e 2 π i t ξ e^{2\pi it\xi} e 2 π i t ξ , which is the multiplier of T − t T_{-t} T − t , and therefore T t ∗ = T − t T_t^* = T_{-t} T t ∗ = T − t . Finally,
∥ T t a ∥ ℓ 2 2 = ∥ F ( T t a ) ∥ L 2 ( − 1 / 2 , 1 / 2 ) 2 = ∫ − 1 / 2 1 / 2 ∣ e − 2 π i t ξ ∣ 2 ∣ F a ( ξ ) ∣ 2 d ξ = ∫ − 1 / 2 1 / 2 ∣ F a ( ξ ) ∣ 2 d ξ = ∥ a ∥ ℓ 2 2 . \lVert T_ta\rVert_{\ell^2}^2 = \lVert\mathcal{F}(T_ta)\rVert_{L^2(-1/2,1/2)}^2 = \int_{-1/2}^{1/2} \lvert e^{-2\pi it\xi}\rvert^2 \lvert\mathcal{F}a(\xi)\rvert^2\,d\xi= \int_{-1/2}^{1/2} \lvert\mathcal{F}a(\xi)\rvert^2\,d\xi= \lVert a\rVert_{\ell^2}^2. ∥ T t a ∥ ℓ 2 2 = ∥ F ( T t a ) ∥ L 2 ( − 1/2 , 1/2 ) 2 = ∫ − 1/2 1/2 ∣ e − 2 π i t ξ ∣ 2 ∣ F a ( ξ ) ∣ 2 d ξ = ∫ − 1/2 1/2 ∣ F a ( ξ ) ∣ 2 d ξ = ∥ a ∥ ℓ 2 2 .
Lemma 3.5 (Distance from the identity). For 0 ≤ ∣ t ∣ ≤ 1 0 \le|t| \le1 0 ≤ ∣ t ∣ ≤ 1 ,
∥ T t − I ∥ ℓ 2 → ℓ 2 = 2 sin ( π ∣ t ∣ 2 ) . (18) \lVert T_t-I\rVert_{\ell^2\to\ell^2}=2\sin\left(\frac{\pi|t|}{2}\right). \tag*{(18)} ∥ T t − I ∥ ℓ 2 → ℓ 2 = 2 sin ( 2 π ∣ t ∣ ) . (18) In particular, ∥ T 2 δ − I ∥ ℓ 2 → ℓ 2 = 2 sin ( π δ ) \lVert T_{2\delta}-I\rVert_{\ell^2\to\ell^2}=2\sin(\pi\delta) ∥ T 2 δ − I ∥ ℓ 2 → ℓ 2 = 2 sin ( π δ ) , 0 ≤ δ < 1 2 0 \le\delta< \frac{1}{2} 0 ≤ δ < 2 1 .
∥ T 2 δ − I ∥ ℓ 2 → ℓ 2 = 2 sin ( π δ ) , 0 ≤ δ < 1 2 . (19) \lVert T_{2\delta}-I\rVert_{\ell^2\to\ell^2}=2\sin(\pi\delta), \qquad0 \le\delta< \frac{1}{2}. \tag*{(19)} ∥ T 2 δ − I ∥ ℓ 2 → ℓ 2 = 2 sin ( π δ ) , 0 ≤ δ < 2 1 . (19)
Proof. From (13) ,
∥ T t − I ∥ ℓ 2 → ℓ 2 = ess sup ∣ ξ ∣ ≤ 1 / 2 ∣ e − 2 π i t ξ − 1 ∣ . (20) \lVert T_t-I\rVert_{\ell^2\to\ell^2}=\operatorname{ess\,sup}_{|\xi|\le1/2}\left|e^{-2\pi i t\xi}-1\right|. \tag*{(20)} ∥ T t − I ∥ ℓ 2 → ℓ 2 = ess sup ∣ ξ ∣ ≤ 1/2 e − 2 π i t ξ − 1 . (20)
Moreover,
∣ e − 2 π i t ξ − 1 ∣ 2 = 4 sin 2 ( π t ξ ) , \left|e^{-2\pi i t\xi}-1\right|^2=4\sin^2(\pi t\xi), e − 2 π i t ξ − 1 2 = 4 sin 2 ( π t ξ ) ,
so
∣ e − 2 π i t ξ − 1 ∣ = 2 ∣ sin ( π t ξ ) ∣ . (21) \left|e^{-2\pi i t\xi}-1\right|=2|\sin(\pi t\xi)|. \tag*{(21)} e − 2 π i t ξ − 1 = 2∣ sin ( π t ξ ) ∣. (21)
If ∣ t ∣ ≤ 1 |t|\le1 ∣ t ∣ ≤ 1 and ∣ ξ ∣ ≤ 1 / 2 |\xi|\le1/2 ∣ ξ ∣ ≤ 1/2 , then 0 ≤ π ∣ t ξ ∣ ≤ π / 2 0\le\pi|t\xi|\le\pi/2 0 ≤ π ∣ t ξ ∣ ≤ π /2 , and the sine is increasing on [ 0 , π / 2 ] [0,\pi/2] [ 0 , π /2 ] . The essential supremum is obtained as ∣ ξ ∣ → 1 / 2 |\xi|\to1/2 ∣ ξ ∣ → 1/2 , giving (18) . Taking t = 2 δ t=2\delta t = 2 δ gives (19) .
The auxiliary sign-modulated shifted sinc operator
Fix 0 ≤ δ < 1 2 0\le\delta<\frac{1}{2} 0 ≤ δ < 2 1 and a sign pattern s = ( s n ) n ∈ Z s=(s_n)_{n\in\mathbb{Z}} s = ( s n ) n ∈ Z with s n ∈ { + 1 , − 1 } s_n\in\{+1,-1\} s n ∈ { + 1 , − 1 } . For finitely supported a = ( a n ) n ∈ Z a=(a_n)_{n\in\mathbb{Z}} a = ( a n ) n ∈ Z define
( N δ a ) m : = ∑ n ∈ Z a n sinc π ( n − m + s n δ ) , m ∈ Z . (22) (N_\delta a)_m:=\sum_{n\in\mathbb{Z}}a_n\operatorname{sinc}_{\pi}(n-m+s_n\delta),\qquad m\in\mathbb{Z}. \tag*{(22)} ( N δ a ) m := n ∈ Z ∑ a n sinc π ( n − m + s n δ ) , m ∈ Z . (22)
Let S + : = { n ∈ Z : s n = 1 } S_+:=\{n\in\mathbb{Z}:s_n=1\} S + := { n ∈ Z : s n = 1 } , S − : = { n ∈ Z : s n = − 1 } S_-:=\{n\in\mathbb{Z}:s_n=-1\} S − := { n ∈ Z : s n = − 1 } , and let P + P_+ P + and P − P_- P − be the corresponding coordinate projections.
Lemma 4.1 (Block decomposition). For every a ∈ ℓ 2 ( Z ) a\in\ell^2(\mathbb{Z}) a ∈ ℓ 2 ( Z ) ,
N δ = T δ P + + T − δ P − . (23) N_\delta=T_\delta P_+ + T_{-\delta}P_-. \tag*{(23)} N δ = T δ P + + T − δ P − . (23)
Proof. If n ∈ S + n\in S_+ n ∈ S + , then
sinc π ( n − m + s n δ ) = sinc π ( n − m + δ ) , \operatorname{sinc}_{\pi}(n-m+s_n\delta)=\operatorname{sinc}_{\pi}(n-m+\delta), sinc π ( n − m + s n δ ) = sinc π ( n − m + δ ) ,
whereas, if n ∈ S − n\in S_- n ∈ S − , then
sinc π ( n − m + s n δ ) = sinc π ( n − m − δ ) . \operatorname{sinc}_{\pi}(n-m+s_n\delta)=\operatorname{sinc}_{\pi}(n-m-\delta). sinc π ( n − m + s n δ ) = sinc π ( n − m − δ ) .
Splitting the sum in (22) according to S + S_+ S + and S − S_- S − gives (23) .
Proposition 4.2 (Uniform boundedness). For every sign pattern and every 0 ≤ δ < 1 2 0\le\delta<\frac{1}{2} 0 ≤ δ < 2 1 , the operator N δ N_\delta N δ extends boundedly to ℓ 2 ( Z ) \ell^2(\mathbb{Z}) ℓ 2 ( Z ) and
∥ N δ ∥ ℓ 2 → ℓ 2 ≤ 2 . (24) \lVert N_\delta\rVert_{\ell^2\to\ell^2}\le\sqrt{2}. \tag*{(24)} ∥ N δ ∥ ℓ 2 → ℓ 2 ≤ 2 . (24)
Proof. Using Lemma 4.1 and the unitarity of T ± δ T_{\pm\delta} T ± δ ,
∥ N δ a ∥ ≤ ∥ T δ ( P + a ) + T − δ ( P − a ) ∥ ≤ ∥ P + a ∥ + ∥ P − a ∥ ≤ 2 ( ∥ P + a ∥ 2 + ∥ P − a ∥ 2 ) = 2 ∥ a ∥ . \lVert N_\delta a\rVert\le\lVert T_\delta(P_+a)+T_{-\delta}(P_-a)\rVert\le\lVert P_+a\rVert+\lVert P_-a\rVert\le\sqrt{2(\lVert P_+a\rVert^2+\lVert P_-a\rVert^2)}=\sqrt{2}\lVert a\rVert. ∥ N δ a ∥ ≤ ∥ T δ ( P + a ) + T − δ ( P − a )∥ ≤ ∥ P + a ∥ + ∥ P − a ∥ ≤ 2 (∥ P + a ∥ 2 + ∥ P − a ∥ 2 ) = 2 ∥ a ∥ .
Proposition 4.3 (Adjoint and square). For an arbitrary sign pattern,
N δ ∗ = P + T − δ + P − T δ , (25) N_\delta^*=P_+T_{-\delta}+P_-T_\delta, \tag*{(25)} N δ ∗ = P + T − δ + P − T δ , (25) and
N δ ∗ N δ = I + P + T − 2 δ P − + P − T 2 δ P + . (26) N_\delta^*N_\delta=I+P_+T_{-2\delta}P_-+P_-T_{2\delta}P_+. \tag*{(26)} N δ ∗ N δ = I + P + T − 2 δ P − + P − T 2 δ P + . (26) Proof. From (22),
N δ ∗ = ( T δ P + + T − δ P − ) ∗ = P + T − δ + P − T δ , N_\delta^*=(T_\delta P_++T_{-\delta}P_-)^*=P_+T_{-\delta}+P_-T_\delta, N δ ∗ = ( T δ P + + T − δ P − ) ∗ = P + T − δ + P − T δ , which proves (24). Hence
N δ ∗ N δ = ( P + T − δ + P − T δ ) ( T δ P + + T − δ P − ) = P + T − δ T δ P + + P + T − δ T − δ P − + P − T δ T δ P + + P − T δ T − δ P − (27) \begin{aligned}
N_\delta^*N_\delta&=(P_+T_{-\delta}+P_-T_\delta)(T_\delta P_++T_{-\delta}P_-) \\
&=P_+T_{-\delta}T_\delta P_++P_+T_{-\delta}T_{-\delta}P_- \\
&\quad+P_-T_\delta T_\delta P_++P_-T_\delta T_{-\delta}P_- \tag*{(27)}
\end{aligned} N δ ∗ N δ = ( P + T − δ + P − T δ ) ( T δ P + + T − δ P − ) = P + T − δ T δ P + + P + T − δ T − δ P − + P − T δ T δ P + + P − T δ T − δ P − (27) = P + + P − + P + T − 2 δ P − + P − T 2 δ P + (28) \begin{aligned}
&=P_++P_-+P_+T_{-2\delta}P_-+P_-T_{2\delta}P_+ \tag*{(28)}
\end{aligned} = P + + P − + P + T − 2 δ P − + P − T 2 δ P + (28) = I + P + T − 2 δ P − + P − T 2 δ P + , (29) =I+P_+T_{-2\delta}P_-+P_-T_{2\delta}P_+, \tag*{(29)} = I + P + T − 2 δ P − + P − T 2 δ P + , (29) as claimed.
Proposition 4.4 (Preliminary quadratic estimate). For every a ∈ ℓ 2 ( Z ) a\in\ell^2(\mathbb{Z}) a ∈ ℓ 2 ( Z ) ,
( 1 − 2 sin ( π δ ) ) ∥ a ∥ 2 2 ≤ ∥ N δ a ∥ 2 2 ≤ ( 1 + 2 sin ( π δ ) ) ∥ a ∥ 2 2 . (30) (1-2\sin(\pi\delta))\lVert a\rVert_2^2\le\lVert N_\delta a\rVert_2^2\le(1+2\sin(\pi\delta))\lVert a\rVert_2^2. \tag*{(30)} ( 1 − 2 sin ( π δ )) ∥ a ∥ 2 2 ≤ ∥ N δ a ∥ 2 2 ≤ ( 1 + 2 sin ( π δ )) ∥ a ∥ 2 2 . (30) Consequently, this estimate gives a positive lower bound whenever
2 sin ( π δ ) < 1 , equivalently 0 ≤ δ < 1 6 . (31) 2\sin(\pi\delta)<1,\qquad\text{equivalently}\qquad0\le\delta<\frac{1}{6}. \tag*{(31)} 2 sin ( π δ ) < 1 , equivalently 0 ≤ δ < 6 1 . (31) Proof. Put
a + = P + a , a − = P − a . a^+=P_+a,\qquad a^-=P_-a. a + = P + a , a − = P − a . The supports of a + a^+ a + and a − a^- a − are disjoint, so
⟨ a + , a − ⟩ = 0 , ∥ a ∥ 2 2 = ∥ a + ∥ 2 2 + ∥ a − ∥ 2 2 . (32) \langle a^+,a^-\rangle=0,\qquad\lVert a\rVert_2^2=\lVert a^+\rVert_2^2+\lVert a^-\rVert_2^2. \tag*{(32)} ⟨ a + , a − ⟩ = 0 , ∥ a ∥ 2 2 = ∥ a + ∥ 2 2 + ∥ a − ∥ 2 2 . (32) By (22),
N δ a = T δ a + + T − δ a − . (33) N_\delta a=T_\delta a^++T_{-\delta}a^-. \tag*{(33)} N δ a = T δ a + + T − δ a − . (33) Therefore
∥ N δ a ∥ 2 2 = ⟨ T δ a + + T − δ a − , T δ a + + T − δ a − ⟩ (34) \lVert N_\delta a\rVert_2^2=\langle T_\delta a^++T_{-\delta}a^-,T_\delta a^++T_{-\delta}a^-\rangle \tag*{(34)} ∥ N δ a ∥ 2 2 = ⟨ T δ a + + T − δ a − , T δ a + + T − δ a − ⟩ (34) = ∥ T δ a + ∥ 2 2 + ∥ T − δ a − ∥ 2 2 + 2 Re ⟨ T δ a + , T − δ a − ⟩ . (35) =\lVert T_\delta a^+\rVert_2^2+\lVert T_{-\delta}a^-\rVert_2^2+2\operatorname{Re}\langle T_\delta a^+,T_{-\delta}a^-\rangle. \tag*{(35)} = ∥ T δ a + ∥ 2 2 + ∥ T − δ a − ∥ 2 2 + 2 Re ⟨ T δ a + , T − δ a − ⟩ . (35)
Since both shifted operators are unitary,
∥ T δ a + ∥ 2 2 + ∥ T − δ a − ∥ 2 2 = ∥ a + ∥ 2 2 + ∥ a − ∥ 2 2 = ∥ a ∥ 2 2 . (36) \lVert T_{\delta}a^{+}\rVert_{2}^{2}+\lVert T_{-\delta}a^{-}\rVert_{2}^{2}=\lVert a^{+}\rVert_{2}^{2}+\lVert a^{-}\rVert_{2}^{2}=\lVert a\rVert_{2}^{2}. \tag*{(36)} ∥ T δ a + ∥ 2 2 + ∥ T − δ a − ∥ 2 2 = ∥ a + ∥ 2 2 + ∥ a − ∥ 2 2 = ∥ a ∥ 2 2 . (36)
Using T − δ ∗ = T δ T_{-\delta}^{*}=T_{\delta} T − δ ∗ = T δ ,
⟨ T δ a + , T − δ a − ⟩ = ⟨ T − δ ∗ T δ a + , a − ⟩ = ⟨ T δ T δ a + , a − ⟩ = ⟨ T 2 δ a + , a − ⟩ . \langle T_{\delta}a^{+},T_{-\delta}a^{-}\rangle=\langle T_{-\delta}^{*}T_{\delta}a^{+},a^{-}\rangle=\langle T_{\delta}T_{\delta}a^{+},a^{-}\rangle=\langle T_{2\delta}a^{+},a^{-}\rangle. ⟨ T δ a + , T − δ a − ⟩ = ⟨ T − δ ∗ T δ a + , a − ⟩ = ⟨ T δ T δ a + , a − ⟩ = ⟨ T 2 δ a + , a − ⟩ .
Since a + ⊥ a − a^{+}\perp a^{-} a + ⊥ a − ,
⟨ T 2 δ a + , a − ⟩ = ⟨ ( T 2 δ − I ) a + , a − ⟩ + ⟨ a + , a − ⟩ = ⟨ ( T 2 δ − I ) a + , a − ⟩ . \langle T_{2\delta}a^{+},a^{-}\rangle=\langle(T_{2\delta}-I)a^{+},a^{-}\rangle+\langle a^{+},a^{-}\rangle=\langle(T_{2\delta}-I)a^{+},a^{-}\rangle. ⟨ T 2 δ a + , a − ⟩ = ⟨( T 2 δ − I ) a + , a − ⟩ + ⟨ a + , a − ⟩ = ⟨( T 2 δ − I ) a + , a − ⟩ .
Thus, by Cauchy–Schwarz and (18),
∣ 2 Re ⟨ T 2 δ a + , a − ⟩ ∣ ≤ 2 ∣ ⟨ ( T 2 δ − I ) a + , a − ⟩ ∣ ≤ 2 ∥ ( T 2 δ − I ) a + ∥ 2 ∥ a − ∥ 2 (37) \left|2\operatorname{Re}\langle T_{2\delta}a^{+},a^{-}\rangle\right|\leq2\left|\langle(T_{2\delta}-I)a^{+},a^{-}\rangle\right|\leq2\lVert(T_{2\delta}-I)a^{+}\rVert_{2}\lVert a^{-}\rVert_{2} \tag*{(37)} 2 Re ⟨ T 2 δ a + , a − ⟩ ≤ 2 ⟨( T 2 δ − I ) a + , a − ⟩ ≤ 2 ∥( T 2 δ − I ) a + ∥ 2 ∥ a − ∥ 2 (37)
≤ 2 ∥ T 2 δ − I ∥ ∥ a + ∥ 2 ∥ a − ∥ 2 = 4 sin ( π δ ) ∥ a + ∥ 2 ∥ a − ∥ 2 . (38) \leq2\lVert T_{2\delta}-I\rVert\lVert a^{+}\rVert_{2}\lVert a^{-}\rVert_{2}=4\sin(\pi\delta)\lVert a^{+}\rVert_{2}\lVert a^{-}\rVert_{2}. \tag*{(38)} ≤ 2 ∥ T 2 δ − I ∥ ∥ a + ∥ 2 ∥ a − ∥ 2 = 4 sin ( π δ ) ∥ a + ∥ 2 ∥ a − ∥ 2 . (38)
The elementary inequality 2 x y ≤ x 2 + y 2 2xy\leq x^{2}+y^{2} 2 x y ≤ x 2 + y 2 , applied to x = ∥ a + ∥ 2 x=\lVert a^{+}\rVert_{2} x = ∥ a + ∥ 2 and y = ∥ a − ∥ 2 y=\lVert a^{-}\rVert_{2} y = ∥ a − ∥ 2 , gives
2 ∥ a + ∥ 2 ∥ a − ∥ 2 ≤ ∥ a + ∥ 2 2 + ∥ a − ∥ 2 2 = ∥ a ∥ 2 2 . (39) 2\lVert a^{+}\rVert_{2}\lVert a^{-}\rVert_{2}\leq\lVert a^{+}\rVert_{2}^{2}+\lVert a^{-}\rVert_{2}^{2}=\lVert a\rVert_{2}^{2}. \tag*{(39)} 2 ∥ a + ∥ 2 ∥ a − ∥ 2 ≤ ∥ a + ∥ 2 2 + ∥ a − ∥ 2 2 = ∥ a ∥ 2 2 . (39)
Consequently,
∣ 2 Re ⟨ T 2 δ a + , a − ⟩ ∣ ≤ 2 sin ( π δ ) ∥ a ∥ 2 2 . (40) \left|2\operatorname{Re}\langle T_{2\delta}a^{+},a^{-}\rangle\right|\leq2\sin(\pi\delta)\lVert a\rVert_{2}^{2}. \tag*{(40)} 2 Re ⟨ T 2 δ a + , a − ⟩ ≤ 2 sin ( π δ ) ∥ a ∥ 2 2 . (40)
Combining (35), (36), and (40) gives (30).
Theorem 4.5 (Refined quadratic estimate). For every sign pattern, every a ∈ ℓ 2 ( Z ) a\in\ell^{2}(\mathbb{Z}) a ∈ ℓ 2 ( Z ) , and every 0 ≤ δ ≤ 1 4 0\leq\delta\leq\frac{1}{4} 0 ≤ δ ≤ 4 1 ,
( 1 − sin ( 2 π δ ) ) ∥ a ∥ 2 2 ≤ ∥ N δ a ∥ 2 2 ≤ ( 1 + sin ( 2 π δ ) ) ∥ a ∥ 2 2 . (41) (1-\sin(2\pi\delta))\lVert a\rVert_{2}^{2}\leq\lVert N_{\delta}a\rVert_{2}^{2}\leq(1+\sin(2\pi\delta))\lVert a\rVert_{2}^{2}. \tag*{(41)} ( 1 − sin ( 2 π δ )) ∥ a ∥ 2 2 ≤ ∥ N δ a ∥ 2 2 ≤ ( 1 + sin ( 2 π δ )) ∥ a ∥ 2 2 . (41) In particular, N δ N_{\delta} N δ is bounded below for every
0 ≤ δ < 1 4 . (42) 0\leq\delta<\frac{1}{4}. \tag*{(42)} 0 ≤ δ < 4 1 . (42) Proof. Orthogonality permits the subtraction of any scalar multiple of the identity. More precisely, for every c ∈ C c\in\mathbb{C} c ∈ C ,
⟨ T 2 δ a + , a − ⟩ = ⟨ ( T 2 δ − c I ) a + , a − ⟩ + c ⟨ a + , a − ⟩ = ⟨ ( T 2 δ − c I ) a + , a − ⟩ . (43) \langle T_{2\delta}a^{+},a^{-}\rangle=\langle(T_{2\delta}-cI)a^{+},a^{-}\rangle+c\langle a^{+},a^{-}\rangle=\langle(T_{2\delta}-cI)a^{+},a^{-}\rangle. \tag*{(43)} ⟨ T 2 δ a + , a − ⟩ = ⟨( T 2 δ − c I ) a + , a − ⟩ + c ⟨ a + , a − ⟩ = ⟨( T 2 δ − c I ) a + , a − ⟩ . (43) Assume
0 ≤ δ ≤ 1 4 . (44) 0\leq\delta\leq\frac{1}{4}. \tag*{(44)} 0 ≤ δ ≤ 4 1 . (44) and set φ = 2 π δ \varphi=2\pi\delta φ = 2 π δ , c = cos φ c=\cos\varphi c = cos φ . By (12), the multiplier of T 2 δ T_{2\delta} T 2 δ is e − 4 π i δ ξ e^{-4\pi i\delta\xi} e − 4 π i δ ξ . Put θ = 4 π δ ξ \theta=4\pi\delta\xi θ = 4 π δ ξ . Since ∣ ξ ∣ ≤ 1 / 2 |\xi|\leq1/2 ∣ ξ ∣ ≤ 1/2 , ∣ θ ∣ ≤ 2 π δ = φ |\theta|\leq2\pi\delta=\varphi ∣ θ ∣ ≤ 2 π δ = φ . Moreover, (44) implies 0 ≤ φ ≤ π / 2 0\leq\varphi\leq\pi/2 0 ≤ φ ≤ π /2 , so cos φ ≥ 0 \cos\varphi\geq0 cos φ ≥ 0 . We compute
∥ T 2 δ − cos φ I ∥ = sup ∣ θ ∣ ≤ φ ∣ e − i θ − cos φ ∣ . \lVert T_{2\delta}-\cos\varphi I\rVert=\sup_{|\theta|\leq\varphi}\left|e^{-i\theta}-\cos\varphi\right|. ∥ T 2 δ − cos φ I ∥ = ∣ θ ∣ ≤ φ sup e − i θ − cos φ . The square of the quantity inside the supremum is
∣ e − i θ − cos φ ∣ 2 = ( e − i θ − cos φ ) ( e i θ − cos φ ) = 1 + cos 2 φ − cos φ ( e i θ + e − i θ ) (45) \left|e^{-i\theta}-\cos\varphi\right|^{2}=(e^{-i\theta}-\cos\varphi)(e^{i\theta}-\cos\varphi)=1+\cos^{2}\varphi-\cos\varphi(e^{i\theta}+e^{-i\theta}) \tag*{(45)} e − i θ − cos φ 2 = ( e − i θ − cos φ ) ( e i θ − cos φ ) = 1 + cos 2 φ − cos φ ( e i θ + e − i θ ) (45)
= 1 + cos 2 φ − 2 cos φ cos θ . (46) = 1+\cos^2\varphi-2\cos\varphi\cos\theta. \tag*{(46)} = 1 + cos 2 φ − 2 cos φ cos θ . (46)
For ∣ θ ∣ ≤ φ ≤ π / 2 |\theta|\le\varphi\le\pi/2 ∣ θ ∣ ≤ φ ≤ π /2 , the cosine is even and decreasing on [ 0 , π ] [0,\pi] [ 0 , π ] , hence
cos θ ≥ cos φ . (47) \cos\theta\ge\cos\varphi. \tag*{(47)} cos θ ≥ cos φ . (47)
Because cos φ ≥ 0 \cos\varphi\ge0 cos φ ≥ 0 , equations (46) –(47) yield
∣ e − i θ − cos φ ∣ 2 ≤ 1 + cos 2 φ − 2 cos 2 φ = 1 − cos 2 φ = sin 2 φ . \left|e^{-i\theta}-\cos\varphi\right|^2\le1+\cos^2\varphi-2\cos^2\varphi=1-\cos^2\varphi=\sin^2\varphi. e − i θ − cos φ 2 ≤ 1 + cos 2 φ − 2 cos 2 φ = 1 − cos 2 φ = sin 2 φ .
Equality occurs at θ = ± φ \theta=\pm\varphi θ = ± φ , so
∥ T 2 δ − cos ( 2 π δ ) I ∥ = sin ( 2 π δ ) , 0 ≤ δ ≤ 1 4 . (48) \left\|T_{2\delta}-\cos(2\pi\delta)I\right\|=\sin(2\pi\delta),\qquad0\le\delta\le\frac{1}{4}. \tag*{(48)} ∥ T 2 δ − cos ( 2 π δ ) I ∥ = sin ( 2 π δ ) , 0 ≤ δ ≤ 4 1 . (48)
Using (43) with this choice of c c c ,
∣ 2 Re ⟨ T 2 δ a + , a − ⟩ ∣ ≤ 2 ∣ ⟨ ( T 2 δ − cos ( 2 π δ ) I ) a + , a − ⟩ ∣ ≤ 2 sin ( 2 π δ ) ∥ a + ∥ 2 ∥ a − ∥ 2 (49) \left|2\operatorname{Re}\left\langle T_{2\delta}a^+,a^-\right\rangle\right|\le2\left|\left\langle\left(T_{2\delta}-\cos(2\pi\delta)I\right)a^+,a^-\right\rangle\right|\le2\sin(2\pi\delta)\left\|a^+\right\|_2\left\|a^-\right\|_2 \tag*{(49)} 2 Re ⟨ T 2 δ a + , a − ⟩ ≤ 2 ⟨ ( T 2 δ − cos ( 2 π δ ) I ) a + , a − ⟩ ≤ 2 sin ( 2 π δ ) a + 2 a − 2 (49)
≤ sin ( 2 π δ ) ( ∥ a + ∥ 2 2 + ∥ a − ∥ 2 2 ) = sin ( 2 π δ ) ∥ a ∥ 2 2 . (50) \le\sin(2\pi\delta)\left(\left\|a^+\right\|_2^2+\left\|a^-\right\|_2^2\right)=\sin(2\pi\delta)\left\|a\right\|_2^2. \tag*{(50)} ≤ sin ( 2 π δ ) ( a + 2 2 + a − 2 2 ) = sin ( 2 π δ ) ∥ a ∥ 2 2 . (50)
Substitution into (35) gives (41) .
For 0 < δ < 1 / 4 0<\delta<1/4 0 < δ < 1/4 ,
sin ( 2 π δ ) = 2 sin ( π δ ) cos ( π δ ) < 2 sin ( π δ ) , (51) \sin(2\pi\delta)=2\sin(\pi\delta)\cos(\pi\delta)<2\sin(\pi\delta), \tag*{(51)} sin ( 2 π δ ) = 2 sin ( π δ ) cos ( π δ ) < 2 sin ( π δ ) , (51)
so (41) is strictly stronger than (30) in that range. The lower coefficient is positive precisely for (42) .
Remark 4.6. At δ = 1 / 4 \delta=1/4 δ = 1/4 , the lower coefficient in (41) is zero and the upper coefficient is 2, consistently with the uniform estimate ∥ N δ ∥ ≤ 2 \left\|N_\delta\right\|\le\sqrt{2} ∥ N δ ∥ ≤ 2 . The theorem concerns the auxiliary kernel (7); it does not by itself prove the corresponding statement for the genuine Gram kernel (6).
Proposition 4.7 (Vectorwise two-sided estimate). For every a ∈ ℓ 2 ( Z ) a\in\ell^2(\mathbb{Z}) a ∈ ℓ 2 ( Z ) ,
∣ ∥ P + a ∥ − ∥ P − a ∥ ∣ ≤ ∥ N δ a ∥ ≤ ∥ P + a ∥ + ∥ P − a ∥ . (52) \left|\left\|P_+a\right\|-\left\|P_-a\right\|\right|\le\left\|N_\delta a\right\|\le\left\|P_+a\right\|+\left\|P_-a\right\|. \tag*{(52)} ∣ ∥ P + a ∥ − ∥ P − a ∥ ∣ ≤ ∥ N δ a ∥ ≤ ∥ P + a ∥ + ∥ P − a ∥ . (52)
Proof. The upper bound is the triangle inequality applied to (22). The lower bound follows from the reverse triangle inequality and the unitarity of T ± δ T_{\pm\delta} T ± δ .
Exact pattern-dependent bounds for the binary shifted-sinc model
We now pass from pattern-uniform estimates to the exact geometry of a fixed sign pattern. Retain the notation of Section 4:
N δ = T δ P + + T − δ P − , N_\delta=T_\delta P_+ + T_{-\delta}P_-, N δ = T δ P + + T − δ P − ,
where P + P_+ P + and P − P_- P − are the coordinate projections associated with
S + = { n : s n = 1 } , S − = { n : s n = − 1 } . S_+=\{n:s_n=1\},\qquad S_-=\{n:s_n=-1\}. S + = { n : s n = 1 } , S − = { n : s n = − 1 } .
Proposition 4.3 gives
N δ ∗ N δ = I + P + T − 2 δ P − + P − T 2 δ P + . N_\delta^*N_\delta=I+P_+T_{-2\delta}P_-+P_-T_{2\delta}P_+. N δ ∗ N δ = I + P + T − 2 δ P − + P − T 2 δ P + .
The only nontrivial interaction is therefore the compression of the unitary shift T 2 δ T_{2\delta} T 2 δ between the two sign subspaces.
Define
C δ , s = P − T 2 δ P + : P + ℓ 2 ( Z ) ⟶ P − ℓ 2 ( Z ) , ρ δ , s = ∥ C δ , s ∥ . (53) C_{\delta,s}=P_-T_{2\delta}P_+:P_+\ell^2(\mathbb{Z})\longrightarrow P_-\ell^2(\mathbb{Z}),\qquad\rho_{\delta,s}=\left\|C_{\delta,s}\right\|. \tag*{(53)} C δ , s = P − T 2 δ P + : P + ℓ 2 ( Z ) ⟶ P − ℓ 2 ( Z ) , ρ δ , s = ∥ C δ , s ∥ . (53)
Theorem 5.1 (Exact block spectrum). Relative to the orthogonal decomposition
ℓ 2 ( Z ) = P + ℓ 2 ( Z ) ⊕ P − ℓ 2 ( Z ) , \ell^2(\mathbb{Z}) = P_+\ell^2(\mathbb{Z}) \oplus P_-\ell^2(\mathbb{Z}), ℓ 2 ( Z ) = P + ℓ 2 ( Z ) ⊕ P − ℓ 2 ( Z ) , one has
N δ ∗ N δ = ( I C δ , s ∗ C δ , s I ) . (54) N_\delta^*N_\delta=
\begin{pmatrix}
I & C_{\delta,s}^* \\
C_{\delta,s} & I
\end{pmatrix}. \tag*{(54)} N δ ∗ N δ = ( I C δ , s C δ , s ∗ I ) . (54) Moreover,
inf σ ( N δ ∗ N δ ) = 1 − ρ δ , s , sup σ ( N δ ∗ N δ ) = 1 + ρ δ , s . (55) \inf\sigma(N_\delta^*N_\delta) = 1-\rho_{\delta,s}, \qquad\sup\sigma(N_\delta^*N_\delta) = 1+\rho_{\delta,s}. \tag*{(55)} inf σ ( N δ ∗ N δ ) = 1 − ρ δ , s , sup σ ( N δ ∗ N δ ) = 1 + ρ δ , s . (55) Consequently,
( 1 − ρ δ , s ) ∥ a ∥ 2 2 ≤ ∥ N δ a ∥ 2 2 ≤ ( 1 + ρ δ , s ) ∥ a ∥ 2 2 , a ∈ ℓ 2 ( Z ) , (56) (1-\rho_{\delta,s})\lVert a\rVert_2^2 \le\lVert N_\delta a\rVert_2^2 \le(1+\rho_{\delta,s})\lVert a\rVert_2^2,\qquad a\in\ell^2(\mathbb{Z}), \tag*{(56)} ( 1 − ρ δ , s ) ∥ a ∥ 2 2 ≤ ∥ N δ a ∥ 2 2 ≤ ( 1 + ρ δ , s ) ∥ a ∥ 2 2 , a ∈ ℓ 2 ( Z ) , (56) and both constants are optimal.
Proof. The block identity follows directly from Proposition 4.3. Put
K δ , s = ( 0 C δ , s ∗ C δ , s 0 ) , N δ ∗ N δ = I + K δ , s . K_{\delta,s} =
\begin{pmatrix}
0 & C_{\delta,s}^* \\
C_{\delta,s} & 0
\end{pmatrix},
\qquad N_\delta^*N_\delta=I+K_{\delta,s}. K δ , s = ( 0 C δ , s C δ , s ∗ 0 ) , N δ ∗ N δ = I + K δ , s .
The operator K δ , s K_{\delta,s} K δ , s is self-adjoint. Introduce the unitary involution
J = ( I 0 0 − I ) . J =
\begin{pmatrix}
I & 0 \\
0 & -I
\end{pmatrix}. J = ( I 0 0 − I ) .
Then
J K δ , s J = − K δ , s . JK_{\delta,s}J=-K_{\delta,s}. J K δ , s J = − K δ , s .
Unitary conjugation preserves the spectrum, hence
σ ( K δ , s ) = σ ( − K δ , s ) = − σ ( K δ , s ) . \sigma(K_{\delta,s})=\sigma(-K_{\delta,s})=-\sigma(K_{\delta,s}). σ ( K δ , s ) = σ ( − K δ , s ) = − σ ( K δ , s ) .
Thus the spectrum is symmetric about the origin. Furthermore,
K δ , s 2 = ( C δ , s ∗ C δ , s 0 0 C δ , s C δ , s ∗ ) , K_{\delta,s}^2 =
\begin{pmatrix}
C_{\delta,s}^*C_{\delta,s} & 0 \\
0 & C_{\delta,s}C_{\delta,s}^*
\end{pmatrix}, K δ , s 2 = ( C δ , s ∗ C δ , s 0 0 C δ , s C δ , s ∗ ) ,
so
∥ K δ , s ∥ 2 = max { ∥ C δ , s ∗ C δ , s ∥ , ∥ C δ , s C δ , s ∗ ∥ } = ∥ C δ , s ∥ 2 . \lVert K_{\delta,s}\rVert^2=\max\{\lVert C_{\delta,s}^*C_{\delta,s}\rVert,\lVert C_{\delta,s}C_{\delta,s}^*\rVert\}=\lVert C_{\delta,s}\rVert^2. ∥ K δ , s ∥ 2 = max {∥ C δ , s ∗ C δ , s ∥ , ∥ C δ , s C δ , s ∗ ∥} = ∥ C δ , s ∥ 2 .
Hence
∥ K δ , s ∥ = ρ δ , s . \lVert K_{\delta,s}\rVert=\rho_{\delta,s}. ∥ K δ , s ∥ = ρ δ , s .
Because a bounded self-adjoint operator has compact real spectrum and its norm is the maximum of the absolute values of the spectral points, at least one of ± ρ δ , s \pm\rho_{\delta,s} ± ρ δ , s belongs to the spectrum. Spectral symmetry then forces both endpoints to belong to the spectrum. Therefore
inf σ ( K δ , s ) = − ρ δ , s , sup σ ( K δ , s ) = ρ δ , s , \inf\sigma(K_{\delta,s})=-\rho_{\delta,s},\qquad\sup\sigma(K_{\delta,s})=\rho_{\delta,s}, inf σ ( K δ , s ) = − ρ δ , s , sup σ ( K δ , s ) = ρ δ , s ,
and shifting by the identity yields (55). The quadratic inequalities and their optimality follow from the spectral theorem.
Corollary 5.3 (Pattern-dependent Riesz criterion). The exponential family
{ e 2 π i ( n + s n δ ) x } n ∈ Z \left\{e^{2\pi i(n+s_n\delta)x}\right\}_{n\in\mathbb{Z}} { e 2 π i ( n + s n δ ) x } n ∈ Z is a Riesz sequence in L 2 ( − 1 / 2 , 1 / 2 ) L^2(-1/2,1/2) L 2 ( − 1/2 , 1/2 ) if and only if
ρ δ , s < 1. (57) \rho_{\delta,s}<1. \tag*{(57)} ρ δ , s < 1. (57) When (57) holds, the condition number of the synthesis operator onto its closed range is
κ = 1 + ρ δ , s 1 − ρ δ , s . (58) \kappa=\sqrt{\frac{1+\rho_{\delta,s}}{1-\rho_{\delta,s}}}. \tag*{(58)} κ = 1 − ρ δ , s 1 + ρ δ , s . (58) The condition ρ δ , s < 1 \rho_{\delta,s}<1 ρ δ , s < 1 alone does not imply completeness. If the synthesis operator is also onto, then the family is a Riesz basis.
Proof. The Fourier-coordinate representation of synthesis is N δ N_\delta N δ . By Theorem 5.1, a positive lower synthesis bound is equivalent to ρ δ , s < 1 \rho_{\delta,s}<1 ρ δ , s < 1 . A bounded-below operator is injective with closed range, which is precisely the synthesis characterization of a Riesz sequence. Surjectivity is the additional condition needed for a Riesz basis. The condition-number formula follows from the optimal singular values 1 − ρ δ , s \sqrt{1-\rho_{\delta,s}} 1 − ρ δ , s and 1 + ρ δ , s \sqrt{1+\rho_{\delta,s}} 1 + ρ δ , s .
The pattern-uniform bound of Theorem 4.5 is recovered from the exact invariant. Since
P − P + = 0 , P_-P_+=0, P − P + = 0 ,
C δ , s = P − ( T 2 δ − c I ) P + C_{\delta,s}=P_-(T_{2\delta}-cI)P_+ C δ , s = P − ( T 2 δ − c I ) P +
for every scalar c c c . With c = cos ( 2 π δ ) c=\cos(2\pi\delta) c = cos ( 2 π δ ) and the scalar-approximation identity used in the proof of Theorem 4.5,
ρ δ , s ≤ sin ( 2 π δ ) , 0 ≤ δ ≤ 1 4 . (59) \rho_{\delta,s}\le\sin(2\pi\delta), \qquad0\le\delta\le\frac{1}{4}. \tag*{(59)} ρ δ , s ≤ sin ( 2 π δ ) , 0 ≤ δ ≤ 4 1 . (59)
Thus sin ( 2 π δ ) \sin(2\pi\delta) sin ( 2 π δ ) is a pattern-uniform upper estimate, whereas ρ δ , s \rho_{\delta,s} ρ δ , s is the exact invariant for the fixed partition S + ∪ S − S_+\cup S_- S + ∪ S − .
Periodic binary patterns and constant finite matrices
Suppose now that s n + p = s n s_{n+p}=s_n s n + p = s n for a period p ≥ 1 p\ge1 p ≥ 1 . The periodic structure permits a complete finite-dimensional reduction. The key point is stronger than the existence of a matrix-valued symbol: the fiber variable can be factored entirely into a unitary input matrix.
Let
J p = [ − 1 2 , − 1 2 + 1 p ) , v r = ( e − 2 π i j r / p ) j = 0 p − 1 , Q r = 1 p v r v r ∗ , J_p=\left[-\frac{1}{2},-\frac{1}{2}+\frac{1}{p}\right),\qquad v_r=\left(e^{-2\pi i jr/p}\right)_{j=0}^{p-1},\qquad Q_r=\frac{1}{p}v_rv_r^*, J p = [ − 2 1 , − 2 1 + p 1 ) , v r = ( e − 2 π ij r / p ) j = 0 p − 1 , Q r = p 1 v r v r ∗ ,
for 0 ≤ r ≤ p − 1 0\le r\le p-1 0 ≤ r ≤ p − 1 . The vectors p − 1 / 2 v r p^{-1/2}v_r p − 1/2 v r form an orthonormal basis of C p \mathbb{C}^p C p . Put
V = 1 p ( v 0 ⋯ v p − 1 ) , (60) V=\frac{1}{\sqrt{p}}(v_0\ \cdots\ v_{p-1}), \tag*{(60)} V = p 1 ( v 0 ⋯ v p − 1 ) , (60)
so that V V V is unitary. For t ∈ R t\in\mathbb{R} t ∈ R and ξ ∈ J p \xi\in J_p ξ ∈ J p , define
D t ( ξ ) = diag ( e − 2 π i t ( ξ + j / p ) ) j = 0 p − 1 . (61) D_t(\xi)=\operatorname{diag}\left(e^{-2\pi i t(\xi+j/p)}\right)_{j=0}^{p-1}. \tag*{(61)} D t ( ξ ) = diag ( e − 2 π i t ( ξ + j / p ) ) j = 0 p − 1 . (61)
The standard period-p p p fiberization sends the residue projection modulo p p p to Q r Q_r Q r and T t T_t T t to multiplication by D t ( ξ ) D_t(\xi) D t ( ξ ) . Hence the binary synthesis symbol is
N p , δ , s ( ξ ) = ∑ r = 0 p − 1 D s r δ ( ξ ) Q r . (62) \mathcal{N}_{p,\delta,s}(\xi)=\sum_{r=0}^{p-1}D_{s_r\delta}(\xi)Q_r. \tag*{(62)} N p , δ , s ( ξ ) = r = 0 ∑ p − 1 D s r δ ( ξ ) Q r . (62)
Define the constant matrix
H p , δ , s = ( 1 p e − 2 π i j ( r + s r δ ) / p ) j , r = 0 p − 1 (63) H_{p,\delta,s}=\left(\frac{1}{\sqrt{p}}e^{-2\pi i j(r+s_r\delta)/p}\right)_{j,r=0}^{p-1} \tag*{(63)} H p , δ , s = ( p 1 e − 2 π ij ( r + s r δ ) / p ) j , r = 0 p − 1 (63)
and the diagonal unitary matrix
E δ , s ( ξ ) = diag ( e − 2 π i s r δ ξ ) r = 0 p − 1 . (64) E_{\delta,s}(\xi)=\operatorname{diag}\left(e^{-2\pi i s_r\delta\xi}\right)_{r=0}^{p-1}. \tag*{(64)} E δ , s ( ξ ) = diag ( e − 2 π i s r δ ξ ) r = 0 p − 1 . (64)
Theorem 6.1 (Constant-matrix reduction for periodic binary patterns). For every ξ ∈ J p \xi\in J_p ξ ∈ J p ,
N p , δ , s ( ξ ) = H p , δ , s E δ , s ( ξ ) V ∗ . (65) \mathcal{N}_{p,\delta,s}(\xi)=H_{p,\delta,s}E_{\delta,s}(\xi)V^*. \tag*{(65)} N p , δ , s ( ξ ) = H p , δ , s E δ , s ( ξ ) V ∗ . (65) Consequently, all singular values of N p , δ , s ( ξ ) \mathcal{N}_{p,\delta,s}(\xi) N p , δ , s ( ξ ) are independent of ξ \xi ξ . The optimal squared synthesis bounds are
A p , δ , s = λ min ( H p , δ , s ∗ H p , δ , s ) , B p , δ , s = λ max ( H p , δ , s ∗ H p , δ , s ) . (66) A_{p,\delta,s}=\lambda_{\min}(H_{p,\delta,s}^*H_{p,\delta,s}),\qquad B_{p,\delta,s}=\lambda_{\max}(H_{p,\delta,s}^*H_{p,\delta,s}). \tag*{(66)} A p , δ , s = λ m i n ( H p , δ , s ∗ H p , δ , s ) , B p , δ , s = λ m a x ( H p , δ , s ∗ H p , δ , s ) . (66) The associated exponential family is a Riesz basis if and only if H p , δ , s H_{p,\delta,s} H p , δ , s is invertible.
Proof. Because Q q v r = δ q r v r Q_qv_r=\delta_{qr}v_r Q q v r = δ q r v r , the r r r -th column of N p , δ , s ( ξ ) V \mathcal{N}_{p,\delta,s}(\xi)V N p , δ , s ( ξ ) V is
D s r δ ( ξ ) v r p . D_{s_r\delta}(\xi)\frac{v_r}{\sqrt{p}}. D s r δ ( ξ ) p v r .
Its j j j -th entry is
1 p e − 2 π i s r δ ( ξ + j / p ) e − 2 π i j r / p = 1 p e − 2 π i j ( r + s r δ ) / p e − 2 π i s r δ ξ , \frac{1}{\sqrt{p}}e^{-2\pi i s_r\delta(\xi+j/p)}e^{-2\pi i jr/p}
=\frac{1}{\sqrt{p}}e^{-2\pi i j(r+s_r\delta)/p}e^{-2\pi i s_r\delta\xi}, p 1 e − 2 π i s r δ ( ξ + j / p ) e − 2 π ij r / p = p 1 e − 2 π ij ( r + s r δ ) / p e − 2 π i s r δ ξ ,
which is exactly the j j j -th entry of the r r r -th column of H p , δ , s E δ , s ( ξ ) H_{p,\delta,s}E_{\delta,s}(\xi) H p , δ , s E δ , s ( ξ ) . Therefore
N p , δ , s ( ξ ) V = H p , δ , s E δ , s ( ξ ) , \mathcal{N}_{p,\delta,s}(\xi)V=H_{p,\delta,s}E_{\delta,s}(\xi), N p , δ , s ( ξ ) V = H p , δ , s E δ , s ( ξ ) ,
and (65) follows.
Both E δ , s ( ξ ) E_{\delta,s}(\xi) E δ , s ( ξ ) and V V V are unitary. Hence
N p , δ , s ( ξ ) ∗ N p , δ , s ( ξ ) = V E δ , s ( ξ ) ∗ H p , δ , s ∗ H p , δ , s E δ , s ( ξ ) V ∗ , \mathcal{N}_{p,\delta,s}(\xi)^{*}\mathcal{N}_{p,\delta,s}(\xi)=VE_{\delta,s}(\xi)^{*}H_{p,\delta,s}^{*}H_{p,\delta,s}E_{\delta,s}(\xi)V^{*}, N p , δ , s ( ξ ) ∗ N p , δ , s ( ξ ) = V E δ , s ( ξ ) ∗ H p , δ , s ∗ H p , δ , s E δ , s ( ξ ) V ∗ ,
so every fiber Gram matrix is unitarily conjugate to the same constant Gram matrix. This proves (66).
To see sharpness directly, let u ∈ C p u\in\mathbb{C}^{p} u ∈ C p be a unit eigenvector of H p , δ , s ∗ H p , δ , s H_{p,\delta,s}^{*}H_{p,\delta,s} H p , δ , s ∗ H p , δ , s with eigenvalue μ \mu μ , and set
F u ( ξ ) = ∣ J p ∣ − 1 / 2 V E δ , s ( ξ ) ∗ u . F_u(\xi)=|J_p|^{-1/2}VE_{\delta,s}(\xi)^{*}u. F u ( ξ ) = ∣ J p ∣ − 1/2 V E δ , s ( ξ ) ∗ u .
Then ∫ J p ∥ F u ( ξ ) ∥ 2 d ξ = 1 \int_{J_p}\lVert F_u(\xi)\rVert^{2}\,\mathrm{d}\xi=1 ∫ J p ∥ F u ( ξ ) ∥ 2 d ξ = 1 and
∫ J p ∥ N p , δ , s ( ξ ) F u ( ξ ) ∥ 2 d ξ = ∥ H p , δ , s u ∥ 2 = μ . \int_{J_p}\lVert\mathcal{N}_{p,\delta,s}(\xi)F_u(\xi)\rVert^{2}\,\mathrm{d}\xi=\lVert H_{p,\delta,s}u\rVert^{2}=\mu. ∫ J p ∥ N p , δ , s ( ξ ) F u ( ξ ) ∥ 2 d ξ = ∥ H p , δ , s u ∥ 2 = μ .
The fiberization is onto, so the smallest and largest eigenvalues are attained as optimal global constants.
If H p , δ , s H_{p,\delta,s} H p , δ , s is invertible, then
N p , δ , s ( ξ ) − 1 = V E δ , s ( ξ ) ∗ H p , δ , s − 1 , \mathcal{N}_{p,\delta,s}(\xi)^{-1}=VE_{\delta,s}(\xi)^{*}H_{p,\delta,s}^{-1}, N p , δ , s ( ξ ) − 1 = V E δ , s ( ξ ) ∗ H p , δ , s − 1 ,
with norm independent of ξ \xi ξ , so the synthesis map is boundedly invertible. If H p , δ , s H_{p,\delta,s} H p , δ , s is singular, choose 0 ≠ u ∈ ker H p , δ , s 0\ne u\in\ker H_{p,\delta,s} 0 = u ∈ ker H p , δ , s and use the same fiber transformation to obtain a nonzero vector in the kernel of synthesis. Thus invertibility of H p , δ , s H_{p,\delta,s} H p , δ , s is equivalent to the Riesz-basis property.
Remark 6.2. The factorization (65) is a right-unitary reduction. It preserves singular values, rank, invertibility, kernel dimension, and the norm of the inverse. It does not assert that the individual matrices N p , δ , s ( ξ ) \mathcal{N}_{p,\delta,s}(\xi) N p , δ , s ( ξ ) are unitarily similar to H p , δ , s H_{p,\delta,s} H p , δ , s .
Corollary 6.3 (Vandermonde collision criterion). The periodic binary family is a Riesz basis if and only if
r + s r δ ≠ q + s q δ ( m o d p ) ( 0 ≤ q < r ≤ p − 1 ) . (67) r+sr\delta\ne q+sq\delta\pmod p\qquad(0\le q<r\le p-1). \tag*{(67)} r + sr δ = q + s q δ ( mod p ) ( 0 ≤ q < r ≤ p − 1 ) . (67) In particular, every fixed periodic binary pattern is a Riesz basis for
0 ≤ δ < 1 2 . 0\le\delta<\frac{1}{2}. 0 ≤ δ < 2 1 .
Proof. Set
z r = e − 2 π i ( r + s r δ ) / p . z_r=e^{-2\pi\mathrm{i}(r+sr\delta)/p}. z r = e − 2 π i ( r + sr δ ) / p .
Then H p , δ , s H_{p,\delta,s} H p , δ , s is the normalized Vandermonde matrix
( H p , δ , s ) j r = p − 1 / 2 z r j , (H_{p,\delta,s})_{jr}=p^{-1/2}z_r^{j}, ( H p , δ , s ) j r = p − 1/2 z r j ,
so
det H p , δ , s = p − p / 2 ∏ 0 ≤ q < r ≤ p − 1 ( z r − z q ) . (68) \det H_{p,\delta,s}=p^{-p/2}\prod_{0\le q<r\le p-1}(z_r-z_q). \tag*{(68)} det H p , δ , s = p − p /2 0 ≤ q < r ≤ p − 1 ∏ ( z r − z q ) . (68)
Thus H p , δ , s H_{p,\delta,s} H p , δ , s is invertible exactly when the nodes are pairwise distinct, which is equivalent to (67) . If 0 ≤ δ < 1 / 2 0 \le\delta< 1/2 0 ≤ δ < 1/2 and a collision occurred, then
r − q + ( s r − s q ) δ ∈ p Z . r-q+(s_r-s_q)\delta\in p\mathbb{Z}. r − q + ( s r − s q ) δ ∈ p Z .
Since r − q r-q r − q is an integer, ( s r − s q ) δ ∈ { 0 , ± 2 δ } (s_r-s_q)\delta\in\{0,\pm2\delta\} ( s r − s q ) δ ∈ { 0 , ± 2 δ } would also be an integer. Its absolute value is strictly less than one, hence it must vanish. Then r − q ∈ p Z r-q \in p\mathbb{Z} r − q ∈ p Z , impossible for distinct residues q , r q,r q , r in { 0 , … , p − 1 } \{0,\ldots,p-1\} { 0 , … , p − 1 } .
Remark 6.4. The preceding corollary is a statement about each fixed periodic pattern. The lower bound may depend on both the period and the pattern; no positive constant uniform over all periods is asserted.
The alternating pattern
Take
s n = ( − 1 ) n . (69) s_n=(-1)^n. \tag*{(69)} s n = ( − 1 ) n . (69)
The frequency pairs are
λ 2 k = 2 k + δ , λ 2 k + 1 = 2 k + 1 − δ , \lambda_{2k}=2k+\delta,\qquad\lambda_{2k+1}=2k+1-\delta, λ 2 k = 2 k + δ , λ 2 k + 1 = 2 k + 1 − δ ,
with within-pair separation
λ 2 k + 1 − λ 2 k = 1 − 2 δ . (70) \lambda_{2k+1}-\lambda_{2k}=1-2\delta. \tag*{(70)} λ 2 k + 1 − λ 2 k = 1 − 2 δ . (70)
Thus the geometry already predicts a collision at δ = 1 / 2 \delta=1/2 δ = 1/2 .
Theorem 6.5 (Exact alternating bounds). For the pattern (68) and 0 ≤ δ ≤ 1 / 2 0 \le\delta\le1/2 0 ≤ δ ≤ 1/2 ,
( 1 − sin ( π δ ) ) ∥ a ∥ 2 2 ≤ ∥ N δ a ∥ 2 2 ≤ ( 1 + sin ( π δ ) ) ∥ a ∥ 2 2 . (71) (1-\sin(\pi\delta))\lVert a\rVert_2^2 \le\lVert N_\delta a\rVert_2^2 \le(1+\sin(\pi\delta))\lVert a\rVert_2^2. \tag*{(71)} ( 1 − sin ( π δ )) ∥ a ∥ 2 2 ≤ ∥ N δ a ∥ 2 2 ≤ ( 1 + sin ( π δ )) ∥ a ∥ 2 2 . (71) Both constants are optimal. The associated exponential family is a Riesz basis for 0 ≤ δ < 1 / 2 0 \le\delta< 1/2 0 ≤ δ < 1/2 . At δ = 1 / 2 \delta=1/2 δ = 1/2 two neighboring frequencies coincide and the lower bound vanishes.
Proof. Use the period-two symbol with s 0 = 1 s_0=1 s 0 = 1 and s 1 = − 1 s_1=-1 s 1 = − 1 . For ξ ∈ [ − 1 / 2 , 0 ) \xi\in[-1/2,0) ξ ∈ [ − 1/2 , 0 ) let
α = 2 π δ ξ , β = 2 π δ ( ξ + 1 / 2 ) = α + π δ . \alpha=2\pi\delta\xi,\qquad\beta=2\pi\delta(\xi+1/2)=\alpha+\pi\delta. α = 2 π δ ξ , β = 2 π δ ( ξ + 1/2 ) = α + π δ .
In the fiber coordinates ( f ( ξ ) , f ( ξ + 1 / 2 ) ) T (f(\xi),f(\xi+1/2))^{T} ( f ( ξ ) , f ( ξ + 1/2 ) ) T ,
N 2 , δ ( ξ ) = ( cos α − i sin α − i sin β cos β ) . N_{2,\delta}(\xi)=
\begin{pmatrix}
\cos\alpha& -i\sin\alpha\\
-i\sin\beta& \cos\beta
\end{pmatrix}. N 2 , δ ( ξ ) = ( cos α − i sin β − i sin α cos β ) .
The Gram matrix has trace 2 2 2 , while
det N 2 , δ ( ξ ) = cos α cos β + sin α sin β = cos ( α − β ) = cos ( π δ ) . \det N_{2,\delta}(\xi)=\cos\alpha\cos\beta+\sin\alpha\sin\beta=\cos(\alpha-\beta)=\cos(\pi\delta). det N 2 , δ ( ξ ) = cos α cos β + sin α sin β = cos ( α − β ) = cos ( π δ ) .
Hence the Gram determinant is cos 2 ( π δ ) \cos^2(\pi\delta) cos 2 ( π δ ) and its eigenvalues solve
λ 2 − 2 λ + cos 2 ( π δ ) = 0. \lambda^2-2\lambda+\cos^2(\pi\delta)=0. λ 2 − 2 λ + cos 2 ( π δ ) = 0.
For 0 ≤ δ ≤ 1 / 2 0 \le\delta\le1/2 0 ≤ δ ≤ 1/2 the eigenvalues are
1 − sin ( π δ ) , 1 + sin ( π δ ) , 1-\sin(\pi\delta),\qquad1+\sin(\pi\delta), 1 − sin ( π δ ) , 1 + sin ( π δ ) ,
independently of ξ \xi ξ . Theorem 6.1 gives optimality and the basis property for δ < 1 / 2 \delta<1/2 δ < 1/2 . At δ = 1 / 2 \delta=1/2 δ = 1/2 , equation (70) gives
2 k + 1 2 = 2 k + 1 − 1 2 , 2k+\frac{1}{2}=2k+1-\frac{1}{2}, 2 k + 2 1 = 2 k + 1 − 2 1 ,
so two exponentials become identical.
Corollary 6.6. For the alternating pattern,
ρ δ , s = sin ( π δ ) , 0 ≤ δ ≤ 1 2 . \rho_{\delta,s}=\sin(\pi\delta), \qquad0\leq\delta\leq\frac{1}{2}. ρ δ , s = sin ( π δ ) , 0 ≤ δ ≤ 2 1 . In particular, at δ = 1 / 4 \delta=1/4 δ = 1/4 the exact lower bound is
1 − 1 2 > 0. 1-\frac{1}{\sqrt{2}}>0. 1 − 2 1 > 0.
Proof. Compare Theorem 6.5 with Theorem 5.1. □
A period-three binary configuration
Consider the period-three pattern
s 3 k = s 3 k + 1 = 1 , s 3 k + 2 = − 1 (72) s_{3k}=s_{3k+1}=1,\qquad s_{3k+2}=-1 \tag*{(72)} s 3 k = s 3 k + 1 = 1 , s 3 k + 2 = − 1 (72)
and take δ = 1 / 4 \delta=1/4 δ = 1/4 . The frequencies in one period are
3 k + 1 4 , 3 k + 5 4 , 3 k + 7 4 . 3k+\frac{1}{4},\qquad3k+\frac{5}{4},\qquad3k+\frac{7}{4}. 3 k + 4 1 , 3 k + 4 5 , 3 k + 4 7 .
Proposition 6.7 (Exact period-three bounds). For (71) at δ = 1 / 4 \delta=1/4 δ = 1/4 , the associated exponential family is a Riesz basis with optimal squared bounds
A = 1 − 5 3 , B = 1 + 5 3 . (73) A=1-\frac{\sqrt{5}}{3},\qquad B=1+\frac{\sqrt{5}}{3}. \tag*{(73)} A = 1 − 3 5 , B = 1 + 3 5 . (73) Moreover,
ρ 1 / 4 , s = 5 3 . (74) \rho_{1/4,s}=\frac{\sqrt{5}}{3}. \tag*{(74)} ρ 1/4 , s = 3 5 . (74)
Proof. Subtract the common shift 1 / 4 1/4 1/4 from the three residues. This multiplies the rows of the constant matrix by unimodular factors and therefore preserves singular values. With
ω = e − 2 π i / 3 , \omega=e^{-2\pi i/3}, ω = e − 2 π i /3 ,
the constant matrix is unitarily equivalent on the left to
H ~ = 1 3 ( 1 1 1 1 ω − 1 1 ω 2 1 ) . \widetilde{H}=\frac{1}{\sqrt{3}}
\begin{pmatrix}
1 & 1 & 1\\
1 & \omega& -1\\
1 & \omega^{2} & 1
\end{pmatrix}. H = 3 1 1 1 1 1 ω ω 2 1 − 1 1 .
Using 1 + ω + ω 2 = 0 1+\omega+\omega^{2}=0 1 + ω + ω 2 = 0 and ω ‾ = ω 2 \overline{\omega}=\omega^{2} ω = ω 2 ,
H ~ ∗ H ~ = ( 1 0 1 / 3 0 1 − 2 ω 2 / 3 1 / 3 − 2 ω / 3 1 ) = ( I 2 c c ∗ 1 ) , c = ( 1 / 3 − 2 ω 2 / 3 ) . \widetilde{H}^{*}\widetilde{H}=
\begin{pmatrix}
1 & 0 & 1/3\\
0 & 1 & -2\omega^{2}/3\\
1/3 & -2\omega/3 & 1
\end{pmatrix}
=
\begin{pmatrix}
I_{2} & c\\
c^{*} & 1
\end{pmatrix},
\qquad
c=
\begin{pmatrix}
1/3\\
-2\omega^{2}/3
\end{pmatrix}. H ∗ H = 1 0 1/3 0 1 − 2 ω /3 1/3 − 2 ω 2 /3 1 = ( I 2 c ∗ c 1 ) , c = ( 1/3 − 2 ω 2 /3 ) .
Since
∥ c ∥ 2 = 1 9 + 4 9 = 5 9 , \lVert c\rVert^{2}=\frac{1}{9}+\frac{4}{9}=\frac{5}{9}, ∥ c ∥ 2 = 9 1 + 9 4 = 9 5 ,
the one-dimensional subspace in the first two coordinates orthogonal to c c c has eigenvalue 1. On the orthogonal complement, the Gram matrix reduces to
( 1 ∥ c ∥ ∥ c ∥ 1 ) . \begin{pmatrix} 1 & \lVert c\rVert\\ \lVert c\rVert& 1 \end{pmatrix}. ( 1 ∥ c ∥ ∥ c ∥ 1 ) .
whose eigenvalues are 1 ± 5 / 3 1 \pm\sqrt{5}/3 1 ± 5 /3 . This proves (72). Comparing with the exact binary formula of Theorem 5.1 gives (73).
Remark 6.8 (General period-three formula). For the same pattern and 0 < δ < 1 / 2 0 < \delta< 1/2 0 < δ < 1/2 , the two positive-sign columns are orthogonal and the negative column interacts with their two-dimensional span. If ρ 3 ( δ ) \rho_3(\delta) ρ 3 ( δ ) denotes the norm of this interaction, then
σ ( H 3 , δ , s ∗ H 3 , δ , s ) = { 1 , 1 − ρ 3 ( δ ) , 1 + ρ 3 ( δ ) } , \sigma(H_{3,\delta,s}^{*}H_{3,\delta,s}) = \{1,1-\rho_3(\delta),1+\rho_3(\delta)\}, σ ( H 3 , δ , s ∗ H 3 , δ , s ) = { 1 , 1 − ρ 3 ( δ ) , 1 + ρ 3 ( δ )} ,
with
ρ 3 ( δ ) = 1 − ( sin ( 2 π δ ) 3 sin ( 2 π δ / 3 ) ) 2 . (75) \rho_3(\delta)=\sqrt{1-\left(\frac{\sin(2\pi\delta)}{3\sin(2\pi\delta/3)}\right)^2}. \tag*{(75)} ρ 3 ( δ ) = 1 − ( 3 sin ( 2 π δ /3 ) sin ( 2 π δ ) ) 2 . (75)
At δ = 1 / 4 \delta= 1/4 δ = 1/4 this equals 5 / 3 \sqrt{5}/3 5 /3 , while at δ = 1 / 2 \delta= 1/2 δ = 1/2 it equals 1. The latter degeneration again corresponds to a collision:
3 k + 1 + 1 2 = 3 k + 2 − 1 2 . 3k+1+\frac{1}{2}=3k+2-\frac{1}{2}. 3 k + 1 + 2 1 = 3 k + 2 − 2 1 .
Critical endpoint and periodic extremizers
Periodic approximation of worst-case binary stability
For a binary pattern s = ( s n ) n ∈ Z s=(s_n)_{n\in\mathbb{Z}} s = ( s n ) n ∈ Z and 0 ≤ δ < 1 / 2 0 \leq\delta< 1/2 0 ≤ δ < 1/2 , write
A ( δ , s ) : = inf 0 ≠ a ∈ ℓ 2 ( Z ) ∥ N δ , s a ∥ 2 2 ∥ a ∥ 2 2 = 1 − ρ δ , s , A(\delta,s):=\inf_{0\ne a\in\ell^2(\mathbb{Z})}\frac{\lVert N_{\delta,s}a\rVert_2^2}{\lVert a\rVert_2^2}=1-\rho_{\delta,s}, A ( δ , s ) := 0 = a ∈ ℓ 2 ( Z ) inf ∥ a ∥ 2 2 ∥ N δ , s a ∥ 2 2 = 1 − ρ δ , s ,
where the last identity follows from the exact block spectral formula of Section 5. Define the worst binary lower bound
A b i n ( δ ) : = inf s ∈ { − 1 , + 1 } Z A ( δ , s ) . (76) A_{\mathrm{bin}}(\delta):=\inf_{s\in\{-1,+1\}^{\mathbb{Z}}}A(\delta,s). \tag*{(76)} A bin ( δ ) := s ∈ { − 1 , + 1 } Z inf A ( δ , s ) . (76)
For p ≥ 1 p\geq1 p ≥ 1 , define the periodic extremal quantity
A p ( δ ) : = min s ∈ { − 1 , + 1 } p λ min ( H p , δ , s ∗ H p , δ , s ) , (77) A_p(\delta):=\min_{s\in\{-1,+1\}^p}\lambda_{\min}(H_{p,\delta,s}^{*}H_{p,\delta,s}), \tag*{(77)} A p ( δ ) := s ∈ { − 1 , + 1 } p min λ m i n ( H p , δ , s ∗ H p , δ , s ) , (77)
where H p , δ , s H_{p,\delta,s} H p , δ , s is the constant matrix in Theorem 6.1.
Theorem 7.1 (Periodic approximation of the worst binary lower bound). For every $0\leq\delta<1/2,
A b i n ( δ ) = inf p ≥ 1 A p ( δ ) . (78) A_{\mathrm{bin}}(\delta)=\inf_{p\geq1}A_p(\delta). \tag*{(78)} A bin ( δ ) = p ≥ 1 inf A p ( δ ) . (78) Equivalently,
sup s ∈ { − 1 , + 1 } Z ρ δ , s = sup p ≥ 1 max s ∈ { − 1 , + 1 } p ρ δ , s . \sup_{s\in\{-1,+1\}^{\mathbb{Z}}}\rho_{\delta,s}
=
\sup_{p\geq1}\max_{s\in\{-1,+1\}^p}\rho_{\delta,s}. s ∈ { − 1 , + 1 } Z sup ρ δ , s = p ≥ 1 sup s ∈ { − 1 , + 1 } p max ρ δ , s .
Proof. Every periodic pattern is in particular a binary pattern on Z \mathbb{Z} Z , so
A b i n ( δ ) ≤ inf p ≥ 1 A p ( δ ) . A_{\mathrm{bin}}(\delta) \le\inf_{p\ge1} A_p(\delta). A bin ( δ ) ≤ p ≥ 1 inf A p ( δ ) .
For the reverse inequality fix an arbitrary binary pattern s s s and ε > 0 \varepsilon> 0 ε > 0 . Since finitely supported sequences are dense in ℓ 0 2 ( Z ) \ell^2_0(\mathbb{Z}) ℓ 0 2 ( Z ) and N δ , s N_{\delta,s} N δ , s is bounded, there exists 0 ≠ a ∈ ℓ 0 2 ( Z ) 0 \ne a \in\ell^2_0(\mathbb{Z}) 0 = a ∈ ℓ 0 2 ( Z ) such that
∥ N δ , s a ∥ 2 2 ∥ a ∥ 2 2 < A ( δ , s ) + ε . \frac{\lVert N_{\delta,s}a\rVert_2^2}{\lVert a\rVert_2^2} < A(\delta,s) + \varepsilon. ∥ a ∥ 2 2 ∥ N δ , s a ∥ 2 2 < A ( δ , s ) + ε .
Choose an integer p p p strictly larger than the diameter of supp a \operatorname{supp} a supp a . There is a p p p -periodic pattern s ~ \widetilde{s} s that agrees with s s s on supp a \operatorname{supp} a supp a . Since the sign in the definition of N δ , s N_{\delta,s} N δ , s is attached only to the input index,
N δ , s ~ a = N δ , s a . N_{\delta,\widetilde{s}}a = N_{\delta,s}a. N δ , s a = N δ , s a .
Hence
A p ( δ ) ≤ A ( δ , s ~ ) ≤ ∥ N δ , s ~ a ∥ 2 2 ∥ a ∥ 2 2 < A ( δ , s ) + ε . A_p(\delta) \le A(\delta,\widetilde{s}) \le\frac{\lVert N_{\delta,\widetilde{s}}a\rVert_2^2}{\lVert a\rVert_2^2} < A(\delta,s) + \varepsilon. A p ( δ ) ≤ A ( δ , s ) ≤ ∥ a ∥ 2 2 ∥ N δ , s a ∥ 2 2 < A ( δ , s ) + ε .
Taking the infimum over s s s and then letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 gives the reverse inequality. □ \square □
Theorem 7.1 is useful because Theorem 6.1 converts every term on the right-hand side of (78) into a finite-dimensional eigenvalue problem. Thus the global binary extremal problem is the limit of a sequence of explicit finite Vandermonde problems, even though no single finite period need realize the worst nonperiodic behavior.
Critical degeneration and extremal conjectures
The critical value δ = 1 / 4 \delta= 1/4 δ = 1/4 makes this phenomenon particularly transparent. Let
s n s t e p = { + 1 , n ≤ 0 , − 1 , n ≥ 1. (79) s_n^{\mathrm{step}} =
\begin{cases}
+1, & n \le0,\\
-1, & n \ge1.
\end{cases} \tag*{(79)} s n step = { + 1 , − 1 , n ≤ 0 , n ≥ 1. (79)
Let P + P_+ P + and P − P_- P − denote the corresponding coordinate projections.
Lemma 7.2 (The critical step pattern and a half-shifted Hilbert matrix). For δ = 1 / 4 \delta= 1/4 δ = 1/4 , the cross compression
C s t e p = P − T 1 / 2 P + C_{\mathrm{step}} = P_-T_{1/2}P_+ C step = P − T 1/2 P + is unitarily equivalent to the Hankel matrix
1 π ( 1 j + k + 1 / 2 ) j , k ≥ 0 . (80) \frac{1}{\pi}\left(\frac{1}{j+k+1/2}\right)_{j,k\ge0}. \tag*{(80)} π 1 ( j + k + 1/2 1 ) j , k ≥ 0 . (80) Moreover,
∥ C s t e p ∥ = 1 , ρ 1 / 4 , s s t e p = 1 , A ( 1 / 4 , s s t e p ) = 0. (81) \lVert C_{\mathrm{step}}\rVert= 1,\qquad\rho_{1/4,s^{\mathrm{step}}} = 1,\qquad A(1/4,s^{\mathrm{step}}) = 0. \tag*{(81)} ∥ C step ∥ = 1 , ρ 1/4 , s step = 1 , A ( 1/4 , s step ) = 0. (81)
Proof. For n = − j ≤ 0 n=-j\leq0 n = − j ≤ 0 and m = k + 1 ≥ 1 m=k+1\geq1 m = k + 1 ≥ 1 ,
sinc π ( n − m + 1 / 2 ) = sinc π ( − ( j + k + 1 / 2 ) ) = ( − 1 ) j + k π ( j + k + 1 / 2 ) . \operatorname{sinc}_{\pi}(n-m+1/2)=\operatorname{sinc}_{\pi}(-(j+k+1/2))=\frac{(-1)^{j+k}}{\pi(j+k+1/2)}. sinc π ( n − m + 1/2 ) = sinc π ( − ( j + k + 1/2 )) = π ( j + k + 1/2 ) ( − 1 ) j + k .
The factor ( − 1 ) j + k (-1)^{j+k} ( − 1 ) j + k is removed by diagonal unitary modulations on the input and output half-lines, which gives (80) . Let
H 1 / 2 = ( 1 j + k + 1 / 2 ) j , k ≥ 0 , H 1 = ( 1 j + k + 1 ) j , k ≥ 0 . \mathcal{H}_{1/2}=\left(\frac{1}{j+k+1/2}\right)_{j,k\geq0},\qquad\mathcal{H}_1=\left(\frac{1}{j+k+1}\right)_{j,k\geq0}. H 1/2 = ( j + k + 1/2 1 ) j , k ≥ 0 , H 1 = ( j + k + 1 1 ) j , k ≥ 0 .
Their difference is Hilbert–Schmidt because
∑ j , k ≥ 0 ∣ 1 j + k + 1 / 2 − 1 j + k + 1 ∣ 2 < ∞ . \sum_{j,k\geq0}\left|\frac{1}{j+k+1/2}-\frac{1}{j+k+1}\right|^2<\infty. j , k ≥ 0 ∑ j + k + 1/2 1 − j + k + 1 1 2 < ∞.
The classical Hilbert matrix H 1 \mathcal{H}_1 H 1 has essential norm π \pi π Rosenblum [2 ] ; hence H 1 / 2 \mathcal{H}_{1/2} H 1/2 has essential norm π \pi π as well. Therefore (79) has norm at least one. On the other hand C s t e p C_{\mathrm{step}} C step is a compression of the unitary operator T 1 / 2 T_{1/2} T 1/2 , so ∥ C s t e p ∥ ≤ 1 \|C_{\mathrm{step}}\|\leq1 ∥ C step ∥ ≤ 1 . Thus equality holds. The remaining statements follow from the exact formula A ( δ , s ) = 1 − ρ δ , s A(\delta,s)=1-\rho_{\delta,s} A ( δ , s ) = 1 − ρ δ , s .
Corollary 7.3 (Critical degeneration through increasing periods). At the critical value δ = 1 / 4 \delta=1/4 δ = 1/4 ,
inf p ≥ 1 min s ∈ { − 1 , + 1 } p λ min ( H p , 1 / 4 , s ∗ H p , 1 / 4 , s ) = 0. (82) \inf_{p\geq1}\ \min_{s\in\{-1,+1\}^p}\lambda_{\min}\left(H_{p,1/4,s}^{*}H_{p,1/4,s}\right)=0. \tag*{(82)} p ≥ 1 inf s ∈ { − 1 , + 1 } p min λ m i n ( H p , 1/4 , s ∗ H p , 1/4 , s ) = 0. (82) Nevertheless, for every fixed p p p and every fixed binary p p p -periodic pattern, the matrix H p , 1 / 4 , s H_{p,1/4,s} H p , 1/4 , s is invertible.
Proof. The first statement follows from Theorem 7.1 and Lemma 7.2 . The second follows from the collision criterion in Corollary 6.3 , since 1 / 4 < 1 / 2 1/4<1/2 1/4 < 1/2 .
Thus the loss of stability at the critical value δ = 1 / 4 \delta=1/4 δ = 1/4 is not detected by any fixed periodic pattern: every fixed-period problem remains invertible, whereas the infimum over increasing periods vanishes. The critical obstruction is therefore genuinely asymptotic in the period. This suggests an extremal finite-dimensional problem in its own right. This is consistent with the critical role of the threshold δ = 1 / 4 \delta=1/4 δ = 1/4 in Kadec’s 1/4 theorem; see Alemany and Nitzan [3 ] . For a related but distinct discussion of the thresholds 1 / 4 1/4 1/4 and 1 / 2 1/2 1/2 in perturbed trigonometric interpolation and quadrature, see Austin and Trefethen [4 ] .
For p ≥ 2 p\geq2 p ≥ 2 , let s b l k , p s^{\mathrm{blk},p} s blk , p be a balanced two-block pattern, for instance
s r b l k , p = { + 1 , 0 ≤ r < ⌈ p / 2 ⌉ , − 1 , ⌈ p / 2 ⌉ ≤ r < p . (83) s^{\mathrm{blk},p}_r=
\begin{cases}
+1, & 0\leq r<\lceil p/2\rceil,\\
-1, & \lceil p/2\rceil\leq r<p.
\end{cases} \tag*{(83)} s r blk , p = { + 1 , − 1 , 0 ≤ r < ⌈ p /2 ⌉ , ⌈ p /2 ⌉ ≤ r < p . (83)
Cyclic translations and a global sign change give equivalent candidates for the extremal problem.
Conjecture 7.4 (Balanced two-block extremizer). At δ = 1 / 4 \delta=1/4 δ = 1/4 , for every p ≥ 2 p\geq2 p ≥ 2 ,
A p ( 1 / 4 ) = λ min ( H p , 1 / 4 , s b l k , p ∗ H p , 1 / 4 , s b l k , p ) . (84) A_p(1/4)=\lambda_{\min}\left(H_{p,1/4,s^{\mathrm{blk},p}}^{*}H_{p,1/4,s^{\mathrm{blk},p}}\right). \tag*{(84)} A p ( 1/4 ) = λ m i n ( H p , 1/4 , s blk , p ∗ H p , 1/4 , s blk , p ) . (84) More generally, among binary patterns with a prescribed number of plus and minus signs, the minimum is attained by a pattern in which each sign occupies one contiguous cyclic block.
Exhaustive numerical enumeration in double precision for 2 ≤ p ≤ 18 2 \le p \le18 2 ≤ p ≤ 18 supports Conjecture 7.4. At δ = 1 / 4 \delta= 1/4 δ = 1/4 the corresponding minima for p = 2 , 4 , 8 , 12 , 16 , 18 p = 2, 4, 8, 12, 16, 18 p = 2 , 4 , 8 , 12 , 16 , 18 are approximately
0.292893 , 0.206647 , 0.150762 , 0.127577 , 0.114147 , 0.109242 , 0.292893,\quad0.206647,\quad0.150762,\quad0.127577,\quad0.114147,\quad0.109242, 0.292893 , 0.206647 , 0.150762 , 0.127577 , 0.114147 , 0.109242 ,
respectively. A reproducible computation is provided separately.
The appearance of the half-shifted Hilbert matrix in Lemma 7.2 suggests a logarithmically slow critical degeneration. For the classical N × N N \times N N × N Hilbert matrix H N = ( 1 / ( j + k + 1 ) ) j , k = 0 N − 1 \mathcal{H}_N = (1/(j+k+1))_{j,k=0}^{N-1} H N = ( 1/ ( j + k + 1 ) ) j , k = 0 N − 1 , de Bruijn and Wilf de Bruijn and Wilf [5 ] proved
π − ∥ H N ∥ = π 5 2 ( log N ) 2 + O ( log log N ( log N ) 3 ) . \pi-\lVert\mathcal{H}_N\rVert=\frac{\pi^5}{2(\log N)^2}+O\left(\frac{\log\log N}{(\log N)^3}\right). π − ∥ H N ∥ = 2 ( log N ) 2 π 5 + O ( ( log N ) 3 log log N ) .
The periodic two-block matrices are not identical to ordinary finite sections of the Hilbert matrix; they contain a cyclic, periodized interaction. Therefore the Hilbert asymptotic cannot be transferred without a comparison theorem.
The currently accessible numerical range is too short to distinguish reliably between competing logarithmic decay rates. Accordingly, the following conjecture should be regarded as a heuristic suggested by the Hilbert-matrix asymptotics, rather than as an asymptotic law strongly supported by the present numerical data.
Conjecture 7.5 (Critical logarithmic degeneration). There exist constants 0 < c < C < ∞ 0 < c < C < \infty 0 < c < C < ∞ such that
c ( log p ) 2 ≤ A p ( 1 / 4 ) ≤ C ( log p ) 2 (85) \frac{c}{(\log p)^2} \le A_p(1/4) \le\frac{C}{(\log p)^2} \tag*{(85)} ( log p ) 2 c ≤ A p ( 1/4 ) ≤ ( log p ) 2 C (85) for all sufficiently large p p p .
A possible route to Conjecture 7.5 is to compare the cross block of the balanced periodic pattern with a finite section of the operator in (79), controlling the wrap-around terms uniformly in p p p . Combined with Conjecture 7.4, such a comparison would turn the sharp critical behavior of the full binary problem into an explicit finite-section asymptotic.
From sign-pattern interactions to the genuine Gram form
The preceding sections give a complete description of the auxiliary binary synthesis model. We now return to the genuine Gram problem. We first introduce a two-pattern sesquilinear form that isolates the interaction of shifted-sinc blocks, and then apply the same structure to the near-critical Gram kernel.
A four-block sesquilinear model
Fix 0 ≤ δ , σ < 1 2 0 \le\delta,\sigma< \frac{1}{2} 0 ≤ δ , σ < 2 1 . Let s = ( s n ) n ∈ Z s=(s_n)_{n\in\mathbb{Z}} s = ( s n ) n ∈ Z and t = ( t m ) m ∈ Z t=(t_m)_{m\in\mathbb{Z}} t = ( t m ) m ∈ Z , with s n , t m ∈ { ± 1 } s_n,t_m\in\{\pm1\} s n , t m ∈ { ± 1 } . For finitely supported sequences a , b ∈ ℓ 0 2 ( Z ) a,b\in\ell_0^2(\mathbb{Z}) a , b ∈ ℓ 0 2 ( Z ) , define
T ( a , b ) : = ∑ n , m ∈ Z a n b m ‾ sinc π ( n − m + s n δ − t m σ ) . (86) \mathcal{T}(a,b):=\sum_{n,m\in\mathbb{Z}}a_n\overline{b_m}\operatorname{sinc}\pi(n-m+s_n\delta-t_m\sigma). \tag*{(86)} T ( a , b ) := n , m ∈ Z ∑ a n b m sinc π ( n − m + s n δ − t m σ ) . (86)
Write
S ± : = { n ∈ Z : s n = ± 1 } S_\pm:=\{n\in\mathbb{Z}:s_n=\pm1\} S ± := { n ∈ Z : s n = ± 1 } , \quad U ± : = { m ∈ Z : t m = ± 1 } U_\pm:=\{m\in\mathbb{Z}:t_m=\pm1\} U ± := { m ∈ Z : t m = ± 1 } ,
and let P ± s P_\pm^s P ± s and P ± t P_\pm^t P ± t be the corresponding coordinate projections.
Lemma 8.1 (Four-block decomposition). The form in (86) extends uniquely to a bounded sesquilinear form on ℓ 2 ( Z ) × ℓ 2 ( Z ) \ell^2(\mathbb{Z}) \times\ell^2(\mathbb{Z}) ℓ 2 ( Z ) × ℓ 2 ( Z ) . More precisely,
∣ T ( a , b ) ∣ ≤ 2 ∥ a ∥ ℓ 2 ∥ b ∥ ℓ 2 . (87) |\mathcal{T}(a,b)| \le2\|a\|_{\ell^2}\|b\|_{\ell^2}. \tag*{(87)} ∣ T ( a , b ) ∣ ≤ 2∥ a ∥ ℓ 2 ∥ b ∥ ℓ 2 . (87) For all a , b ∈ ℓ 2 ( Z ) a,b \in\ell^2(\mathbb{Z}) a , b ∈ ℓ 2 ( Z ) ,
T ( a , b ) = ∑ ε , η ∈ { ± 1 } ⟨ T ε δ − η σ ( P s ε a ) , P t η b ⟩ ℓ 2 . (88) \mathcal{T}(a,b)=\sum_{\varepsilon,\eta\in\{\pm1\}}\left\langle T_{\varepsilon\delta-\eta\sigma}(P^\varepsilon_s a),P^\eta_t b\right\rangle_{\ell^2}. \tag*{(88)} T ( a , b ) = ε , η ∈ { ± 1 } ∑ ⟨ T ε δ − η σ ( P s ε a ) , P t η b ⟩ ℓ 2 . (88)
Proof. For finitely supported sequences, split the defining sum according to the four sign combinations
( s n , t m ) = ( ε , η ) . (s_n,t_m)=(\varepsilon,\eta). ( s n , t m ) = ( ε , η ) .
On each block,
sinc π ( n − m + s n δ − t m σ ) = sinc π ( n − m + ε δ − η σ ) , \operatorname{sinc}_{\pi}(n-m+s_n\delta-t_m\sigma)=\operatorname{sinc}_{\pi}(n-m+\varepsilon\delta-\eta\sigma), sinc π ( n − m + s n δ − t m σ ) = sinc π ( n − m + ε δ − η σ ) ,
which gives (88) . By Cauchy–Schwarz and the unitarity of the shifted-sinc operators,
∣ T ( a , b ) ∣ ≤ ∑ ε , η ∈ { ± 1 } ∥ P s ε a ∥ ℓ 2 ∥ P t η b ∥ ℓ 2 = ( ∥ P s + a ∥ ℓ 2 + ∥ P s − a ∥ ℓ 2 ) ( ∥ P t + b ∥ ℓ 2 + ∥ P t − b ∥ ℓ 2 ) \begin{aligned}
|\mathcal{T}(a,b)|\leq\sum_{\varepsilon,\eta\in\{\pm1\}}\|P^\varepsilon_s a\|_{\ell^2}\|P^\eta_t b\|_{\ell^2} \\
&=\left(\|P^+_s a\|_{\ell^2}+\|P^-_s a\|_{\ell^2}\right)\left(\|P^+_t b\|_{\ell^2}+\|P^-_t b\|_{\ell^2}\right)
\end{aligned} ∣ T ( a , b ) ∣ ≤ ε , η ∈ { ± 1 } ∑ ∥ P s ε a ∥ ℓ 2 ∥ P t η b ∥ ℓ 2 = ( ∥ P s + a ∥ ℓ 2 + ∥ P s − a ∥ ℓ 2 ) ( ∥ P t + b ∥ ℓ 2 + ∥ P t − b ∥ ℓ 2 )
≤ 2 ∥ a ∥ ℓ 2 ∥ b ∥ ℓ 2 . \leq2\|a\|_{\ell^2}\|b\|_{\ell^2}. ≤ 2∥ a ∥ ℓ 2 ∥ b ∥ ℓ 2 .
The extension now follows uniquely by density.
Proposition 8.2 (Structured sign patterns). If both sign patterns are constant, then
∣ T ( a , b ) ∣ ≤ 1 |\mathcal{T}(a,b)|\leq1 ∣ T ( a , b ) ∣ ≤ 1 for unit vectors. If exactly one sign pattern is constant, then
∣ T ( a , b ) ∣ ≤ 2 . |\mathcal{T}(a,b)|\leq\sqrt{2}. ∣ T ( a , b ) ∣ ≤ 2 .
Proof. If both sign patterns are constant, write
s n ≡ ε , t m ≡ η , ε , η ∈ { ± 1 } . s_n\equiv\varepsilon,\qquad t_m\equiv\eta,\qquad\varepsilon,\eta\in\{\pm1\}. s n ≡ ε , t m ≡ η , ε , η ∈ { ± 1 } .
Only one block in (88) survives, and
T ( a , b ) = ⟨ T ε δ − η σ a , b ⟩ . \mathcal{T}(a,b)=\left\langle T_{\varepsilon\delta-\eta\sigma}a,b\right\rangle. T ( a , b ) = ⟨ T ε δ − η σ a , b ⟩ .
Since T ε δ − η σ T_{\varepsilon\delta-\eta\sigma} T ε δ − η σ is unitary,
∣ T ( a , b ) ∣ ≤ ∥ a ∥ 2 ∥ b ∥ 2 = 1. |\mathcal{T}(a,b)|\leq\|a\|_2\|b\|_2=1. ∣ T ( a , b ) ∣ ≤ ∥ a ∥ 2 ∥ b ∥ 2 = 1.
In the special case ε = η \varepsilon=\eta ε = η and δ = σ \delta=\sigma δ = σ , this reduces to T 0 = I T_0=I T 0 = I and hence T ( a , b ) = ⟨ a , b ⟩ \mathcal{T}(a,b)=\langle a,b\rangle T ( a , b ) = ⟨ a , b ⟩ .
If exactly one sign pattern is constant, only two blocks survive, and the same argument used in Lemma 8.1 gives
∣ T ( a , b ) ∣ ≤ 2 . |\mathcal{T}(a,b)|\leq\sqrt{2}. ∣ T ( a , b ) ∣ ≤ 2 .
Proposition 8.3 (Absence of a uniform positive lower bound). There is no constant A > 0 A > 0 A > 0 such that
A ≤ ∣ T ( a , b ) ∣ for all ∥ a ∥ = ∥ b ∥ = 1. A \leq|\mathcal{T}(a,b)| \qquad\text{for all }\|a\|=\|b\|=1. A ≤ ∣ T ( a , b ) ∣ for all ∥ a ∥ = ∥ b ∥ = 1. Proof. Define
S : = ∑ ε , η ∈ { ± 1 } P η η T ε δ − η σ P ε ε . S := \sum_{\varepsilon,\eta\in\{\pm1\}} P_{\eta}^{\eta}T_{\varepsilon\delta-\eta\sigma}P_{\varepsilon}^{\varepsilon}. S := ε , η ∈ { ± 1 } ∑ P η η T ε δ − η σ P ε ε . so that
T ( a , b ) = ⟨ S a , b ⟩ . \mathcal{T}(a,b)=\langle Sa,b\rangle. T ( a , b ) = ⟨ S a , b ⟩ . If S a ≠ 0 Sa\ne0 S a = 0 , choose a unit vector b b b orthogonal to S a Sa S a . If S a = 0 Sa=0 S a = 0 , every unit vector b b b gives the same conclusion. In either case, T ( a , b ) = 0 \mathcal{T}(a,b)=0 T ( a , b ) = 0 .
The near-critical regime of the genuine Gram form
Assume
δ n = s n ( 1 4 − ε n ) , s n ∈ { ± 1 } , ε n ≥ 0. (89) \delta_n=s_n\left(\frac{1}{4}-\varepsilon_n\right),\qquad s_n\in\{\pm1\},\qquad\varepsilon_n\geq0. \tag*{(89)} δ n = s n ( 4 1 − ε n ) , s n ∈ { ± 1 } , ε n ≥ 0. (89)
Let M M M be the unitary modulation introduced in Section 2. It is enough to study
Q ~ ( b ) : = ⟨ G ~ b , b ⟩ = ∑ n , m ∈ Z b n b m ‾ K ~ m , n , (90) \widetilde{Q}(b):=\langle\widetilde{G}b,b\rangle=\sum_{n,m\in\mathbb{Z}}b_n\overline{b_m}\widetilde{K}_{m,n}, \tag*{(90)} Q ( b ) := ⟨ G b , b ⟩ = n , m ∈ Z ∑ b n b m K m , n , (90)
where K ~ m , n \widetilde{K}_{m,n} K m , n is given by (5). Since G ~ = M G M \widetilde{G}=MGM G = M GM ,
Q ~ ( b ) ≥ 0. \widetilde{Q}(b)\geq0. Q ( b ) ≥ 0.
Let
S ± : = { n ∈ Z : s n = ± 1 } , b ± : = P ± b . S_{\pm}:=\{n\in\mathbb{Z}:s_n=\pm1\},\qquad b^{\pm}:=P_{\pm}b. S ± := { n ∈ Z : s n = ± 1 } , b ± := P ± b .
For later use, for θ ∈ R ∖ Z \theta\in\mathbb{R}\setminus\mathbb{Z} θ ∈ R ∖ Z define the modulated shifted sinc operator
H θ : = M T θ M , ( H θ a ) m = ∑ n ∈ Z a n sin ( π θ ) π ( n − m + θ ) . (91) \mathcal{H}_{\theta}:=MT_{\theta}M,\qquad(\mathcal{H}_{\theta}a)_m=\sum_{n\in\mathbb{Z}}a_n\frac{\sin(\pi\theta)}{\pi(n-m+\theta)}. \tag*{(91)} H θ := M T θ M , ( H θ a ) m = n ∈ Z ∑ a n π ( n − m + θ ) sin ( π θ ) . (91)
It is unitary because it is unitarily equivalent to T θ T_{\theta} T θ . In the notation of De Carli and Samad [1 ] , this operator coincides with D − θ D_{-\theta} D − θ . In particular, the half-shifted operator H 1 / 2 \mathcal{H}_{1/2} H 1/2 belongs to their unitary one-parameter group. The new feature of the present analysis is the interaction of these operators with the coordinate projections determined by the sign pattern, rather than the unitarity of the underlying discrete Hilbert transform.
The critical form
If ε n = 0 \varepsilon_n=0 ε n = 0 , that is, δ n = s n / 4 \delta_n=s_n/4 δ n = s n /4 , the same-sign terms vanish for n ≠ m n\ne m n = m , whereas the opposite-sign terms produce a half-integer shift. Hence
Q ~ c r i t ( b ) = ∥ b ∥ 2 2 + 2 Re ⟨ H 1 / 2 b + , b − ⟩ . (92) \widetilde{Q}_{\mathrm{crit}}(b)=\|b\|_2^2+2\operatorname{Re}\langle\mathcal{H}_{1/2}b^+,b^-\rangle. \tag*{(92)} Q crit ( b ) = ∥ b ∥ 2 2 + 2 Re ⟨ H 1/2 b + , b − ⟩ . (92)
Therefore
∥ b ∥ 2 2 − 2 ∥ b + ∥ 2 ∥ b − ∥ 2 ≤ Q ~ c r i t ( b ) ≤ ∥ b ∥ 2 2 + 2 ∥ b + ∥ 2 ∥ b − ∥ 2 ≤ 2 ∥ b ∥ 2 2 . (93) \|b\|_2^2-2\|b^+\|_2\|b^-\|_2\leq\widetilde{Q}_{\mathrm{crit}}(b)\leq\|b\|_2^2+2\|b^+\|_2\|b^-\|_2\leq2\|b\|_2^2. \tag*{(93)} ∥ b ∥ 2 2 − 2∥ b + ∥ 2 ∥ b − ∥ 2 ≤ Q crit ( b ) ≤ ∥ b ∥ 2 2 + 2∥ b + ∥ 2 ∥ b − ∥ 2 ≤ 2∥ b ∥ 2 2 . (93)
Kernel expansions
Let b ∈ ℓ 0 2 ( Z ) b \in\ell_0^2(\mathbb{Z}) b ∈ ℓ 0 2 ( Z ) and write
S : = supp ( b ) , N : = ∣ S ∣ , η S : = sup n ∈ S ε n , ω S : = sup n , m ∈ S ∣ ε n − ε m ∣ . (94) S := \operatorname{supp}(b), \qquad N := |S|, \qquad\eta_S := \sup_{n \in S} \varepsilon_n, \qquad\omega_S := \sup_{n,m \in S} |\varepsilon_n-\varepsilon_m|. \tag*{(94)} S := supp ( b ) , N := ∣ S ∣ , η S := n ∈ S sup ε n , ω S := n , m ∈ S sup ∣ ε n − ε m ∣. (94)
Assume, for convenience, that η S ≤ 1 / 8 \eta_S \le1/8 η S ≤ 1/8 .
Lemma 8.4 (Same-sign and opposite-sign expansions). Let n , m ∈ S n,m \in S n , m ∈ S , n ≠ m n \ne m n = m . If s n = s m s_n=s_m s n = s m , then
K ~ m , n = s n ε m − ε n n − m + r m , n s a m e , (95) \widetilde{K}_{m,n}=s_n\frac{\varepsilon_m-\varepsilon_n}{n-m}+r^{\mathrm{same}}_{m,n}, \tag*{(95)} K m , n = s n n − m ε m − ε n + r m , n same , (95) where
∣ r m , n s a m e ∣ ≤ C ∣ ε m − ε n ∣ 2 ∣ n − m ∣ . (96) |r^{\mathrm{same}}_{m,n}| \le C\frac{|\varepsilon_m-\varepsilon_n|^2}{|n-m|}. \tag*{(96)} ∣ r m , n same ∣ ≤ C ∣ n − m ∣ ∣ ε m − ε n ∣ 2 . (96) If s n ≠ s m s_n \ne s_m s n = s m , then
K ~ m , n = sin ( π ( s n − s m ) / 4 ) π ( n − m + ( s n − s m ) / 4 ) + r m , n c r o s s , (97) \widetilde{K}_{m,n}=\frac{\sin(\pi(s_n-s_m)/4)}{\pi\left(n-m+(s_n-s_m)/4\right)}+r^{\mathrm{cross}}_{m,n}, \tag*{(97)} K m , n = π ( n − m + ( s n − s m ) /4 ) sin ( π ( s n − s m ) /4 ) + r m , n cross , (97) where
∣ r m , n c r o s s ∣ ≤ C ( η S ∣ n − m + ( s n − s m ) / 4 ∣ 2 + η S 2 ∣ n − m + ( s n − s m ) / 4 ∣ ) . (98) |r^{\mathrm{cross}}_{m,n}| \le C\left(\frac{\eta_S}{|n-m+(s_n-s_m)/4|^2}+\frac{\eta_S^2}{|n-m+(s_n-s_m)/4|}\right). \tag*{(98)} ∣ r m , n cross ∣ ≤ C ( ∣ n − m + ( s n − s m ) /4 ∣ 2 η S + ∣ n − m + ( s n − s m ) /4∣ η S 2 ) . (98)
Proof. Suppose first that s n = s m = s s_n=s_m=s s n = s m = s . Set k = n − m k=n-m k = n − m , h = ε m − ε n h=\varepsilon_m-\varepsilon_n h = ε m − ε n . Then δ n − δ m = s h \delta_n-\delta_m=sh δ n − δ m = s h , and, by (5),
K ~ m , n = sin ( π s h ) π ( k + s h ) . (99) \widetilde{K}_{m,n}=\frac{\sin(\pi sh)}{\pi(k+sh)}. \tag*{(99)} K m , n = π ( k + s h ) sin ( π s h ) . (99)
Since sin ( π s h ) = s sin ( π h ) \sin(\pi sh)=s\sin(\pi h) sin ( π s h ) = s sin ( π h ) ,
K ~ m , n − s h k = s ( sin ( π h ) π ( k + s h ) − h k ) = s k sin ( π h ) − π h ( k + s h ) π k ( k + s h ) = s k ( sin ( π h ) − π h ) − π s h 2 π k ( k + s h ) . \widetilde{K}_{m,n}-s\frac{h}{k}
=s\left(\frac{\sin(\pi h)}{\pi(k+sh)}-\frac{h}{k}\right)
=s\frac{k\sin(\pi h)-\pi h(k+sh)}{\pi k(k+sh)}
=s\frac{k(\sin(\pi h)-\pi h)-\pi sh^2}{\pi k(k+sh)}. K m , n − s k h = s ( π ( k + s h ) sin ( π h ) − k h ) = s π k ( k + s h ) k sin ( π h ) − π h ( k + s h ) = s π k ( k + s h ) k ( sin ( π h ) − π h ) − π s h 2 .
Because ∣ h ∣ ≤ 1 / 4 |h|\le1/4 ∣ h ∣ ≤ 1/4 and k ∈ Z ∖ { 0 } k\in\mathbb{Z}\setminus\{0\} k ∈ Z ∖ { 0 } , one has ∣ k + s h ∣ ≥ ∣ k ∣ / 2 |k+sh|\ge|k|/2 ∣ k + s h ∣ ≥ ∣ k ∣/2 . Moreover, ∣ sin ( π h ) − π h ∣ ≤ C ∣ h ∣ 3 |\sin(\pi h)-\pi h|\le C|h|^3 ∣ sin ( π h ) − π h ∣ ≤ C ∣ h ∣ 3 . It follows that
∣ K ~ m , n − s h k ∣ ≤ C ( ∣ h ∣ 3 ∣ k ∣ + ∣ h ∣ 2 ∣ k ∣ 2 ) ≤ C ∣ h ∣ 2 ∣ k ∣ , \left|\widetilde{K}_{m,n}-s\frac{h}{k}\right|
\le C\left(\frac{|h|^3}{|k|}+\frac{|h|^2}{|k|^2}\right)
\le C\frac{|h|^2}{|k|}, K m , n − s k h ≤ C ( ∣ k ∣ ∣ h ∣ 3 + ∣ k ∣ 2 ∣ h ∣ 2 ) ≤ C ∣ k ∣ ∣ h ∣ 2 ,
which proves (94)–(95).
Suppose now that s n ≠ s m s_n\ne s_m s n = s m . Then s m = − s n s_m=-s_n s m = − s n . Set s = s n s=s_n s = s n , k = n − m k=n-m k = n − m , and u = ε n + ε m u=\varepsilon_n+\varepsilon_m u = ε n + ε m . We have δ n − δ m = s ( 1 / 2 − u ) \delta_n-\delta_m=s(1/2-u) δ n − δ m = s ( 1/2 − u ) . Let x 0 = k + s / 2 = k + ( s n − s m ) / 4 x_0=k+s/2=k+(s_n-s_m)/4 x 0 = k + s /2 = k + ( s n − s m ) /4 . Then
K ~ m , n = sin ( π ( s / 2 − s u ) ) π ( x 0 − s u ) = s cos ( π u ) π ( x 0 − s u ) , (100) \widetilde{K}_{m,n}=\frac{\sin(\pi(s/2-su))}{\pi(x_0-su)}=\frac{s\cos(\pi u)}{\pi(x_0-su)}, \tag*{(100)} K m , n = π ( x 0 − s u ) sin ( π ( s /2 − s u )) = π ( x 0 − s u ) s cos ( π u ) , (100)
whereas
sin ( π s / 2 ) π x 0 = s π x 0 . \frac{\sin(\pi s/2)}{\pi x_0}=\frac{s}{\pi x_0}. π x 0 sin ( π s /2 ) = π x 0 s .
Comparing (99) with this critical modulated kernel gives
K ~ m , n − s π x 0 = s π ( cos ( π u ) x 0 − s u − 1 x 0 ) = s π ( cos ( π u ) − 1 x 0 − s u + s u x 0 ( x 0 − s u ) ) . \widetilde{K}_{m,n}-\frac{s}{\pi x_0}
=\frac{s}{\pi}\left(\frac{\cos(\pi u)}{x_0-su}-\frac{1}{x_0}\right)
=\frac{s}{\pi}\left(\frac{\cos(\pi u)-1}{x_0-su}+\frac{su}{x_0(x_0-su)}\right). K m , n − π x 0 s = π s ( x 0 − s u cos ( π u ) − x 0 1 ) = π s ( x 0 − s u cos ( π u ) − 1 + x 0 ( x 0 − s u ) s u ) .
Since ∣ x 0 ∣ ≥ 1 / 2 |x_0| \ge1/2 ∣ x 0 ∣ ≥ 1/2 , u ≤ 2 η S ≤ 1 / 4 u \le2\eta_S \le1/4 u ≤ 2 η S ≤ 1/4 , and therefore ∣ x 0 − s u ∣ ≥ ∣ x 0 ∣ / 2 |x_0-su| \ge|x_0|/2 ∣ x 0 − s u ∣ ≥ ∣ x 0 ∣/2 , while ∣ cos ( π u ) − 1 ∣ ≤ C u 2 |\cos(\pi u)-1| \le Cu^2 ∣ cos ( π u ) − 1∣ ≤ C u 2 , we obtain
∣ r m , n c r o s s ∣ ≤ C ( u ∣ x 0 ∣ 2 + u 2 ∣ x 0 ∣ ) ≤ C ( η S ∣ n − m + ( s n − s m ) / 4 ∣ 2 + η S 2 ∣ n − m + ( s n − s m ) / 4 ∣ ) . |r_{m,n}^{\mathrm{cross}}| \le C\left(\frac{u}{|x_0|^2}+\frac{u^2}{|x_0|}\right) \le C\left(\frac{\eta_S}{|n-m+(s_n-s_m)/4|^2}+\frac{\eta_S^2}{|n-m+(s_n-s_m)/4|}\right). ∣ r m , n cross ∣ ≤ C ( ∣ x 0 ∣ 2 u + ∣ x 0 ∣ u 2 ) ≤ C ( ∣ n − m + ( s n − s m ) /4 ∣ 2 η S + ∣ n − m + ( s n − s m ) /4∣ η S 2 ) .
This proves (96)–(97).
Remark 8.5. The critical shift is ( s n − s m ) / 4 = ± 1 / 2 (s_n-s_m)/4=\pm1/2 ( s n − s m ) /4 = ± 1/2 . Because the alternating factor has already been removed in (5), the relevant critical operator is H 1 / 2 = M T 1 / 2 M \mathcal{H}_{1/2}=MT_{1/2}M H 1/2 = M T 1/2 M , rather than T 1 / 2 T_{1/2} T 1/2 itself.
Define the same-sign drift term
D S ( b ) : = ∑ n , m ∈ S n ≠ m , s n = s m b n b m ‾ s n ε m − ε n n − m . (101) D_S(b):=\sum_{\substack{n,m\in S\\ n\ne m,\ s_n=s_m}} b_n\overline{b_m}s_n\frac{\varepsilon_m-\varepsilon_n}{n-m}. \tag*{(101)} D S ( b ) := n , m ∈ S n = m , s n = s m ∑ b n b m s n n − m ε m − ε n . (101)
Then
Q ~ ( b ) = Q ~ c r i t ( b ) + D S ( b ) + R S ( b ) , (102) \widetilde{Q}(b)=\widetilde{Q}_{\mathrm{crit}}(b)+D_S(b)+R_S(b), \tag*{(102)} Q ( b ) = Q crit ( b ) + D S ( b ) + R S ( b ) , (102)
where R S R_S R S collects the remainders in (95) and (97).
Proposition 8.6 (Support-dependent perturbative estimate). There exists a universal constant C > 0 C>0 C > 0 such that, for every finitely supported b ∈ ℓ 0 2 ( Z ) b\in\ell_0^2(\mathbb{Z}) b ∈ ℓ 0 2 ( Z ) ,
∣ D S ( b ) ∣ ≤ C ω S log ( 1 + N ) ∥ b ∥ 2 2 , (103) |D_S(b)|\le C\omega_S\log(1+N)\|b\|_2^2, \tag*{(103)} ∣ D S ( b ) ∣ ≤ C ω S log ( 1 + N ) ∥ b ∥ 2 2 , (103) ∣ R S ( b ) ∣ ≤ C ( η S + η S 2 log ( 1 + N ) ) ∥ b ∥ 2 2 . (104) |R_S(b)|\le C(\eta_S+\eta_S^2\log(1+N))\|b\|_2^2. \tag*{(104)} ∣ R S ( b ) ∣ ≤ C ( η S + η S 2 log ( 1 + N )) ∥ b ∥ 2 2 . (104) Consequently,
Q ~ ( b ) = ∥ b ∥ 2 2 + 2 Re ⟨ H 1 / 2 b + , b − ⟩ + E S ( b ) , (105) \widetilde{Q}(b)=\|b\|_2^2+2\operatorname{Re}\langle\mathcal{H}_{1/2}b^+,b^-\rangle+E_S(b), \tag*{(105)} Q ( b ) = ∥ b ∥ 2 2 + 2 Re ⟨ H 1/2 b + , b − ⟩ + E S ( b ) , (105) with
∣ E S ( b ) ∣ ≤ C ( ω S log ( 1 + N ) + η S + η S 2 log ( 1 + N ) ) ∥ b ∥ 2 2 . (106) |E_S(b)|\le C(\omega_S\log(1+N)+\eta_S+\eta_S^2\log(1+N))\|b\|_2^2. \tag*{(106)} ∣ E S ( b ) ∣ ≤ C ( ω S log ( 1 + N ) + η S + η S 2 log ( 1 + N )) ∥ b ∥ 2 2 . (106) In particular,
Q ~ ( b ) ≤ [ 2 + C ( ω S log ( 1 + N ) + η S + η S 2 log ( 1 + N ) ) ] ∥ b ∥ 2 2 . (107) \widetilde{Q}(b)\le[2+C(\omega_S\log(1+N)+\eta_S+\eta_S^2\log(1+N))]\|b\|_2^2. \tag*{(107)} Q ( b ) ≤ [ 2 + C ( ω S log ( 1 + N ) + η S + η S 2 log ( 1 + N ))] ∥ b ∥ 2 2 . (107)
Proof. Since S S S is finite, the kernels
K 1 ( m , n ) : = 1 { m ≠ n } ∣ n − m ∣ , K 1 / 2 ( m , n ) : = 1 ∣ n − m + 1 / 2 ∣ , K_1(m,n):=\frac{\mathbf{1}_{\{m\ne n\}}}{|n-m|},\qquad K_{1/2}(m,n):=\frac{1}{|n-m+1/2|}, K 1 ( m , n ) := ∣ n − m ∣ 1 { m = n } , K 1/2 ( m , n ) := ∣ n − m + 1/2∣ 1 ,
restricted to S × S S\times S S × S , satisfy
sup m ∈ S ∑ n ∈ S n ≠ m 1 ∣ n − m ∣ ≤ 2 ∑ r = 1 N − 1 1 r ≤ C log ( 1 + N ) , (108) \sup_{m\in S}\sum_{\substack{n\in S\\n\ne m}}\frac{1}{|n-m|}\le2\sum_{r=1}^{N-1}\frac{1}{r}\le C\log(1+N), \tag*{(108)} m ∈ S sup n ∈ S n = m ∑ ∣ n − m ∣ 1 ≤ 2 r = 1 ∑ N − 1 r 1 ≤ C log ( 1 + N ) , (108)
and
sup m ∈ S ∑ n ∈ S 1 ∣ n − m + 1 / 2 ∣ ≤ C log ( 1 + N ) , sup m ∈ S ∑ n ∈ S 1 ∣ n − m + 1 / 2 ∣ 2 ≤ C . (109) \sup_{m\in S}\sum_{n\in S}\frac{1}{|n-m+1/2|}\le C\log(1+N),\qquad\sup_{m\in S}\sum_{n\in S}\frac{1}{|n-m+1/2|^2}\le C. \tag*{(109)} m ∈ S sup n ∈ S ∑ ∣ n − m + 1/2∣ 1 ≤ C log ( 1 + N ) , m ∈ S sup n ∈ S ∑ ∣ n − m + 1/2 ∣ 2 1 ≤ C . (109)
The same bounds hold with 1 / 2 1/2 1/2 replaced by − 1 / 2 -1/2 − 1/2 . Therefore Schur’s test is applicable on the finite set S S S .
Using (100), the bound ∣ ε m − ε n ∣ ≤ ω S |\varepsilon_m-\varepsilon_n|\le\omega_S ∣ ε m − ε n ∣ ≤ ω S , and (107) yields (102); the factor s n s_n s n has modulus one. Likewise, (95), (97), and (107)–(108) give (103). Formula (104) follows from (91) and (101), and (106) follows from (92).
A support-independent neighborhood of structured systems
The finite-support expansion of § 8.2 provides first-order information on the genuine Gram form, but its error estimates contain a logarithmic factor depending on the support size. We now consider a complementary question: the stability of an exponential system under arbitrary uniformly small perturbations of its frequencies, when the unperturbed system is already known to be a Riesz basis.
The general principle that exponential Riesz bases remain stable under sufficiently small perturbations is classical. In particular, De Carli and Pathak [6 ] established explicit quantitative stability conditions for exponential bases and frames on multidimensional domains, including disconnected domains. Their results provide general perturbation thresholds in terms of the original Riesz bounds and the geometry of the underlying domain. The theorem below is not intended as an improvement of these general stability results.
Our emphasis is instead on the interaction between perturbation stability and the exact spectral information obtained in the preceding sections. For periodic binary patterns, the constant-matrix reduction gives the optimal Riesz bounds explicitly as the extreme eigenvalues of a finite Gram matrix. These bounds can then be used as input for a perturbation estimate around the corresponding structured exponential basis.
For this purpose, we record a direct operator-norm estimate obtained from the Taylor expansion of the exponential factors. Its proof uses only the upper Riesz bound of the reference system and yields explicit lower and upper bounds for the perturbed family. In particular, the resulting estimates are independent of the support size and remain valid for arbitrary uniformly small index-dependent perturbations, even when the perturbed frequency pattern is no longer periodic.
The contribution of this section is therefore the combination of exact pattern-dependent spectral bounds with an explicit stability neighborhood of a structured reference system, rather than a new general perturbation principle. The sufficient conditions obtained below are not claimed to be optimal. Moreover, this approach requires a reference system with a strictly positive lower Riesz bound and therefore does not resolve the support-independent near-critical problem for arbitrary nonperiodic sign configurations.
Theorem 9.1 (Uniform perturbation of a structured exponential basis). Let
( S 0 a ) ( x ) = ∑ n ∈ Z a n e 2 π i n x (S_0a)(x)=\sum_{n\in\mathbb{Z}}a_ne^{2\pi i n x} ( S 0 a ) ( x ) = n ∈ Z ∑ a n e 2 π in x be the synthesis operator of a Riesz basis of L 2 ( − 1 / 2 , 1 / 2 ) L^2(-1/2,1/2) L 2 ( − 1/2 , 1/2 ) with optimal bounds A 0 , B 0 > 0 A_0,B_0>0 A 0 , B 0 > 0 . Let ( r n ) (r_n) ( r n ) be real with
sup n ∣ r n ∣ ≤ η , \sup_n |r_n|\leq\eta, n sup ∣ r n ∣ ≤ η , and define
( S r a ) ( x ) = ∑ n ∈ Z a n e 2 π i ( 0 + r n ) x . (S_ra)(x)=\sum_{n\in\mathbb{Z}}a_ne^{2\pi i(0+r_n)x}. ( S r a ) ( x ) = n ∈ Z ∑ a n e 2 π i ( 0 + r n ) x . Then
∥ S r − S 0 ∥ ℓ 2 → L 2 ≤ B 0 ( e π η − 1 ) . (110) \|S_r-S_0\|_{\ell^2\to L^2}\leq\sqrt{B_0}(e^{\pi\eta}-1). \tag*{(110)} ∥ S r − S 0 ∥ ℓ 2 → L 2 ≤ B 0 ( e π η − 1 ) . (110)
If
B 0 ( e π η − 1 ) < A 0 , (111) \sqrt{B_0}(e^{\pi\eta}-1)<\sqrt{A_0}, \tag*{(111)} B 0 ( e π η − 1 ) < A 0 , (111)
then the perturbed family is a Riesz basis, with bounds
A r ≥ [ A 0 − B 0 ( e π η − 1 ) ] 2 , (112) A_r\geq\left[\sqrt{A_0}-\sqrt{B_0}(e^{\pi\eta}-1)\right]^2, \tag*{(112)} A r ≥ [ A 0 − B 0 ( e π η − 1 ) ] 2 , (112)
B r ≤ B 0 e 2 π η . (113) B_r\leq B_0e^{2\pi\eta}. \tag*{(113)} B r ≤ B 0 e 2 π η . (113)
Proof. For a finitely supported sequence,
S r a − S 0 a = ∑ k = 1 ∞ ( 2 π i x ) k k ! ∑ n ∈ Z a n r n k e 2 π i λ n 0 x . S_ra-S_0a=\sum_{k=1}^{\infty}\frac{(2\pi i x)^k}{k!}\sum_{n\in\mathbb{Z}}a_nr_n^ke^{2\pi i\lambda_n^0x}. S r a − S 0 a = k = 1 ∑ ∞ k ! ( 2 π i x ) k n ∈ Z ∑ a n r n k e 2 π i λ n 0 x .
Since ∣ x ∣ ≤ 1 / 2 |x|\leq1/2 ∣ x ∣ ≤ 1/2 and S 0 S_0 S 0 has upper bound B 0 B_0 B 0 ,
∥ S r a − S 0 a ∥ L 2 ≤ ∑ k = 1 ∞ π k k ! ∥ ∑ n a n r n k e 2 π i λ n 0 x ∥ L 2 ≤ B 0 ∑ k = 1 ∞ ( π η ) k k ! ∥ a ∥ 2 = B 0 ( e π η − 1 ) ∥ a ∥ 2 . \begin{aligned}
\|S_ra-S_0a\|_{L^2}
&\leq\sum_{k=1}^{\infty}\frac{\pi^k}{k!}\left\|\sum_n a_nr_n^ke^{2\pi i\lambda_n^0x}\right\|_{L^2}\\
&\leq\sqrt{B_0}\sum_{k=1}^{\infty}\frac{(\pi\eta)^k}{k!}\|a\|_2\\
&=\sqrt{B_0}(e^{\pi\eta}-1)\|a\|_2.
\end{aligned} ∥ S r a − S 0 a ∥ L 2 ≤ k = 1 ∑ ∞ k ! π k n ∑ a n r n k e 2 π i λ n 0 x L 2 ≤ B 0 k = 1 ∑ ∞ k ! ( π η ) k ∥ a ∥ 2 = B 0 ( e π η − 1 ) ∥ a ∥ 2 .
Density gives (109). The direct and reverse triangle inequalities yield (111)–(112). Finally,
∥ ( S r − S 0 ) S 0 − 1 ∥ ≤ B 0 ( e π η − 1 ) A 0 < 1 , \|(S_r-S_0)S_0^{-1}\|\leq\frac{\sqrt{B_0}(e^{\pi\eta}-1)}{\sqrt{A_0}}<1, ∥ ( S r − S 0 ) S 0 − 1 ∥ ≤ A 0 B 0 ( e π η − 1 ) < 1 ,
so S r S 0 − 1 S_rS_0^{-1} S r S 0 − 1 is invertible by a Neumann-series argument. Hence S r S_r S r is onto as well as bounded below.
Corollary 9.2 (A global alternating near-critical neighborhood). Let
λ n = n + ( − 1 ) n ( 1 4 − ε n ) , 0 ≤ ε n ≤ η . \lambda_n=n+(-1)^n\left(\frac{1}{4}-\varepsilon_n\right),\qquad0\leq\varepsilon_n\leq\eta. λ n = n + ( − 1 ) n ( 4 1 − ε n ) , 0 ≤ ε n ≤ η . If
η < log 2 2 π , (114) \eta<\frac{\log2}{2\pi}, \tag*{(114)} η < 2 π log 2 , (114) then { e 2 π i λ n x } n ∈ Z \{e^{2\pi i\lambda_nx}\}_{n\in\mathbb{Z}} { e 2 π i λ n x } n ∈ Z is a Riesz basis. One may take
A = [ 1 − 2 − 1 / 2 − 1 + 2 − 1 / 2 ( e π η − 1 ) ] 2 , (115) A=\left[\sqrt{1-2^{-1/2}}-\sqrt{1+2^{-1/2}}(e^{\pi\eta}-1)\right]^2, \tag*{(115)} A = [ 1 − 2 − 1/2 − 1 + 2 − 1/2 ( e π η − 1 ) ] 2 , (115) B = ( 1 + 2 − 1 / 2 ) e 2 π η . (116) B=(1+2^{-1/2})e^{2\pi\eta}. \tag*{(116)} B = ( 1 + 2 − 1/2 ) e 2 π η . (116) Proof. At ε n = 0 \varepsilon_n=0 ε n = 0 , Theorem 6.5 with δ = 1 / 4 \delta=1/4 δ = 1/4 gives
A 0 = 1 − 2 − 1 / 2 , B 0 = 1 + 2 − 1 / 2 . A_0=1-2^{-1/2},\qquad B_0=1+2^{-1/2}. A 0 = 1 − 2 − 1/2 , B 0 = 1 + 2 − 1/2 . The additional perturbation is r n = − ( − 1 ) n ε n r_n=-(-1)^n\varepsilon_n r n = − ( − 1 ) n ε n . Condition (110) becomes
e π η − 1 < 1 − 2 − 1 / 2 1 + 2 − 1 / 2 = 2 − 1 , e^{\pi\eta}-1<\sqrt{\frac{1-2^{-1/2}}{1+2^{-1/2}}}=\sqrt{2}-1, e π η − 1 < 1 + 2 − 1/2 1 − 2 − 1/2 = 2 − 1 , which is equivalent to (113).
Three or more local frequency shifts
The binary decomposition extends naturally when the local perturbation takes values in a finite set. The resulting Gram operator reveals exactly what is special about the binary case: with two values there is one off-diagonal operator and one scalar invariant, whereas with three or more values one obtains a genuine block-operator matrix.
Let Γ 1 , … , Γ L \Gamma_1,\ldots,\Gamma_L Γ 1 , … , Γ L be a partition of Z \mathbb{Z} Z , let P j P_j P j be the coordinate projection onto
H j = ℓ 2 ( Γ j ) , \mathcal{H}_j=\ell^2(\Gamma_j), H j = ℓ 2 ( Γ j ) ,
and choose real shifts t 1 , … , t L t_1,\ldots,t_L t 1 , … , t L . Define
τ n = t j ( n ∈ Γ j ) , λ n = n + τ n . \tau_n=t_j \quad(n\in\Gamma_j), \qquad\lambda_n=n+\tau_n. τ n = t j ( n ∈ Γ j ) , λ n = n + τ n .
If frequencies coincide, repetitions are retained. The Fourier-coordinate synthesis operator on L 2 ( − 1 / 2 , 1 / 2 ) L^2(-1/2,1/2) L 2 ( − 1/2 , 1/2 ) is
N = ∑ j = 1 L T t j P j . (117) N=\sum_{j=1}^{L}T_{t_j}P_j. \tag*{(117)} N = j = 1 ∑ L T t j P j . (117)
Since each T t j T_{t_j} T t j is unitary,
∥ N a ∥ ≤ ∑ j = 1 L ∥ P j a ∥ ≤ L ∥ a ∥ . \|Na\|\leq\sum_{j=1}^{L}\|P_ja\|\leq\sqrt{L}\|a\|. ∥ N a ∥ ≤ j = 1 ∑ L ∥ P j a ∥ ≤ L ∥ a ∥.
Proposition 10.1 (Block Gram matrix). Relative to
ℓ 2 ( Z ) = ⨁ j = 1 L H j , \ell^{2}(\mathbb{Z}) = \bigoplus_{j=1}^{L} \mathcal{H}_{j}, ℓ 2 ( Z ) = j = 1 ⨁ L H j , the Gram operator G = N ∗ N G = N^{*}N G = N ∗ N has blocks
G j k = P j T t k − t j P k : H k ⟶ H j . (118) G_{jk} = P_{j}T_{t_{k}-t_{j}}P_{k} : \mathcal{H}_{k} \longrightarrow\mathcal{H}_{j}. \tag*{(118)} G j k = P j T t k − t j P k : H k ⟶ H j . (118) In particular,
G j j = I H j , G k j = G j k ∗ . G_{jj} = I_{\mathcal{H}_{j}}, \qquad G_{kj} = G_{jk}^{*}. G j j = I H j , G k j = G j k ∗ . Define the real symmetric matrix R = ( r j k ) j , k = 1 L \mathbf{R} = (r_{jk})_{j,k=1}^{L} R = ( r j k ) j , k = 1 L by
r j j = 0 , r j k = ∥ G j k ∥ ( j ≠ k ) , η = ∥ R ∥ C L → C L . r_{jj}=0,\qquad r_{jk}=\lVert G_{jk}\rVert\quad(j\ne k),\qquad\eta=\lVert\mathbf{R}\rVert_{\mathcal{C}^{L}\to\mathcal{C}^{L}}. r j j = 0 , r j k = ∥ G j k ∥ ( j = k ) , η = ∥ R ∥ C L → C L . Then
( 1 − η ) ∥ a ∥ 2 ≤ ∥ N a ∥ 2 ≤ ( 1 + η ) ∥ a ∥ 2 . (119) (1-\eta)\lVert a\rVert^{2}\le\lVert Na\rVert^{2}\le(1+\eta)\lVert a\rVert^{2}. \tag*{(119)} ( 1 − η ) ∥ a ∥ 2 ≤ ∥ N a ∥ 2 ≤ ( 1 + η ) ∥ a ∥ 2 . (119) Hence η < 1 \eta<1 η < 1 is sufficient for the associated family to be a Riesz sequence.
Proof. From (116),
N ∗ N = ∑ j , k = 1 L P j T − t j T t k P k = ∑ j , k = 1 L P j T t k − t j P k , N^{*}N=\sum_{j,k=1}^{L}P_{j}T_{-t_{j}}T_{t_{k}}P_{k}=\sum_{j,k=1}^{L}P_{j}T_{t_{k}-t_{j}}P_{k}, N ∗ N = j , k = 1 ∑ L P j T − t j T t k P k = j , k = 1 ∑ L P j T t k − t j P k ,
which gives the block identity and the diagonal and adjoint relations.
Write a j = P j a a_{j}=P_{j}a a j = P j a and
v = ( ∥ a 1 ∥ , … , ∥ a L ∥ ) T . v=(\lVert a_{1}\rVert,\ldots,\lVert a_{L}\rVert)^{T}. v = (∥ a 1 ∥ , … , ∥ a L ∥ ) T .
Then
∣ ⟨ ( G − I ) a , a ⟩ ∣ ≤ ∑ j ≠ k ∥ G j k ∥ ∥ a k ∥ ∥ a j ∥ = v T R v ≤ ∥ R ∥ ∥ v ∥ 2 = η ∥ a ∥ 2 . \begin{aligned}
\left|\langle(G-I)a,a\rangle\right|
&\le\sum_{j\ne k}\lVert G_{jk}\rVert\lVert a_{k}\rVert\lVert a_{j}\rVert\\
&=v^{T}\mathbf{R}v\\
&\le\lVert\mathbf{R}\rVert\lVert v\rVert^{2}=\eta\lVert a\rVert^{2}.
\end{aligned} ∣ ⟨( G − I ) a , a ⟩ ∣ ≤ j = k ∑ ∥ G j k ∥ ∥ a k ∥ ∥ a j ∥ = v T R v ≤ ∥ R ∥ ∥ v ∥ 2 = η ∥ a ∥ 2 .
Since ⟨ G a , a ⟩ = ∥ N a ∥ 2 \langle Ga,a\rangle=\lVert Na\rVert^{2} ⟨ G a , a ⟩ = ∥ N a ∥ 2 , inequality (118) follows.
Remark 10.2. The comparison bound (118) is sufficient and need not be optimal. For a general nonperiodic finite partition it yields a lower synthesis estimate when η < 1 \eta<1 η < 1 , but it does not by itself provide completeness. In the binary case L = 2 L=2 L = 2 , the self-adjoint off-diagonal part has the special form
( 0 C ∗ C 0 ) , \begin{pmatrix}
0 & C^{*}\\
C & 0
\end{pmatrix}, ( 0 C C ∗ 0 ) ,
so one norm ∥ C ∥ \lVert C\rVert ∥ C ∥ gives the exact endpoints. For L ≥ 3 L\ge3 L ≥ 3 , no analogous one-scalar reduction is available in general.
Periodic finite-alphabet patterns
Assume now that the local shift sequence is p p p -periodic:
τ n + p = τ n . \tau_{n+p}=\tau_n. τ n + p = τ n .
Write τ r \tau_r τ r for the shift carried by residue class r r r , 0 ≤ r ≤ p − 1 0\le r\le p-1 0 ≤ r ≤ p − 1 . Define
H p , τ = ( 1 p e − 2 π i j ( r + τ r ) / p ) j , r = 0 p − 1 , (120) H_{p,\tau}=\left(\frac{1}{\sqrt{p}}e^{-2\pi i j(r+\tau_r)/p}\right)_{j,r=0}^{p-1}, \tag*{(120)} H p , τ = ( p 1 e − 2 π ij ( r + τ r ) / p ) j , r = 0 p − 1 , (120)
E τ ( ξ ) = diag ( e − 2 π i τ r ξ ) r = 0 p − 1 . (121) E_\tau(\xi)=\operatorname{diag}\left(e^{-2\pi i\tau_r\xi}\right)_{r=0}^{p-1}. \tag*{(121)} E τ ( ξ ) = diag ( e − 2 π i τ r ξ ) r = 0 p − 1 . (121)
Theorem 10.3 (Exact periodic reduction for finitely many local shifts). The period-p p p synthesis symbol satisfies
N p , τ ( ξ ) = ∑ r = 0 p − 1 D τ r ( ξ ) Q r = H p , τ E τ ( ξ ) V ∗ . (122) \mathcal{N}_{p,\tau}(\xi)=\sum_{r=0}^{p-1}D_{\tau_r}(\xi)Q_r=H_{p,\tau}E_\tau(\xi)V^*. \tag*{(122)} N p , τ ( ξ ) = r = 0 ∑ p − 1 D τ r ( ξ ) Q r = H p , τ E τ ( ξ ) V ∗ . (122) The optimal squared synthesis bounds are
A = λ min ( H p , τ ∗ H p , τ ) , B = λ max ( H p , τ ∗ H p , τ ) . (123) A=\lambda_{\min}(H_{p,\tau}^*H_{p,\tau}),\qquad B=\lambda_{\max}(H_{p,\tau}^*H_{p,\tau}). \tag*{(123)} A = λ m i n ( H p , τ ∗ H p , τ ) , B = λ m a x ( H p , τ ∗ H p , τ ) . (123) The exponential family is a Riesz basis if and only if
r + τ r ≠ q + τ q ( m o d p ) ( 0 ≤ q < r ≤ p − 1 ) . (124) r+\tau_r\ne q+\tau_q\pmod p\qquad(0\le q<r\le p-1). \tag*{(124)} r + τ r = q + τ q ( mod p ) ( 0 ≤ q < r ≤ p − 1 ) . (124)
Proof. The computation of Theorem 6.1 does not use the fact that the shifts take only two values. The j j j -th entry of the r r r -th column of N p , τ ( ξ ) V \mathcal{N}_{p,\tau}(\xi)V N p , τ ( ξ ) V is
1 p e − 2 π i τ r ( ξ + j / p ) e − 2 π i j r / p = 1 p e − 2 π i j ( r + τ r ) / p e − 2 π i τ r ξ . \frac{1}{\sqrt{p}}e^{-2\pi i\tau_r(\xi+j/p)}e^{-2\pi i jr/p}
=
\frac{1}{\sqrt{p}}e^{-2\pi i j(r+\tau_r)/p}e^{-2\pi i\tau_r\xi}. p 1 e − 2 π i τ r ( ξ + j / p ) e − 2 π ij r / p = p 1 e − 2 π ij ( r + τ r ) / p e − 2 π i τ r ξ .
This proves (121). Since the two right-hand factors are unitary, the sharp norm argument gives (122).
Set
z r = e − 2 π i ( r + τ r ) / p . z_r=e^{-2\pi i(r+\tau_r)/p}. z r = e − 2 π i ( r + τ r ) / p .
Then
det H p , τ = p − p / 2 ∏ 0 ≤ q < r ≤ p − 1 ( z r − z q ) . \det H_{p,\tau}=p^{-p/2}\prod_{0\le q<r\le p-1}(z_r-z_q). det H p , τ = p − p /2 0 ≤ q < r ≤ p − 1 ∏ ( z r − z q ) .
so H p , τ H_{p,\tau} H p , τ is invertible exactly under (123). Uniform invertibility of the fiber symbol follows from
N p , τ ( ξ ) − 1 = V E τ ( ξ ) ∗ H p , τ − 1 . \mathcal{N}_{p,\tau}(\xi)^{-1}=VE_\tau(\xi)^*H_{p,\tau}^{-1}. N p , τ ( ξ ) − 1 = V E τ ( ξ ) ∗ H p , τ − 1 .
If the determinant vanishes, a nonzero vector in ker H p , τ \ker H_{p,\tau} ker H p , τ produces a nonzero kernel vector for synthesis after the inverse fiber transformation.
Corollary 10.4 (A sufficient oscillation condition). Every fixed periodic local-shift pattern satisfying
max 0 ≤ r < p τ r − min 0 ≤ r < p τ r < 1 (125) \max_{0\le r<p}\tau_r-\min_{0\le r<p}\tau_r<1 \tag*{(125)} 0 ≤ r < p max τ r − 0 ≤ r < p min τ r < 1 (125) generates a Riesz basis of L 2 ( − 1 / 2 , 1 / 2 ) L^2(-1/2,1/2) L 2 ( − 1/2 , 1/2 ) .
Proof. If a collision occurred, then
r − q + τ r − τ q ∈ p Z . r-q+\tau_r-\tau_q \in p\mathbb{Z}. r − q + τ r − τ q ∈ p Z .
Hence τ r − τ q \tau_r-\tau_q τ r − τ q would be an integer. Under (125) its absolute value is strictly less than one, so it must be zero. This would imply r − q ∈ p Z r-q \in p\mathbb{Z} r − q ∈ p Z , impossible for distinct residues. □ \square □
Remark 10.5. The criterion concerns the range of the local shifts. Adding a common constant to every τ r \tau_r τ r multiplies every exponential by the same unimodular function and therefore leaves all Gram eigenvalues unchanged.
A three-shift example without binary spectral symmetry
Consider the period-three pattern
τ 3 k = 1 4 , τ 3 k + 1 = 0 , τ 3 k + 2 = − 1 4 . \tau_{3k}=\frac{1}{4},\qquad\tau_{3k+1}=0,\qquad\tau_{3k+2}=-\frac{1}{4}. τ 3 k = 4 1 , τ 3 k + 1 = 0 , τ 3 k + 2 = − 4 1 .
Its frequencies in one period are
3 k + 1 4 , 3 k + 1 , 3 k + 7 4 . 3k+\frac{1}{4},\qquad3k+1,\qquad3k+\frac{7}{4}. 3 k + 4 1 , 3 k + 1 , 3 k + 4 7 .
Subtracting the common shift 1 / 4 1/4 1/4 preserves the Gram eigenvalues and reduces the constant matrix, up to a left unitary diagonal factor, to
H ~ = 1 3 ( 1 1 1 1 − i − 1 1 − 1 1 ) . \widetilde{H}=\frac{1}{\sqrt{3}}
\begin{pmatrix}
1 & 1 & 1\\
1 & -i & -1\\
1 & -1 & 1
\end{pmatrix}. H = 3 1 1 1 1 1 − i − 1 1 − 1 1 .
A direct multiplication gives
H ~ ∗ H ~ = ( 1 − i / 3 1 / 3 i / 3 1 − i / 3 1 / 3 i / 3 1 ) = 4 3 I 3 − 1 3 w w ∗ , w = ( 1 − i − 1 ) . (126) \widetilde{H}^{*}\widetilde{H}=
\begin{pmatrix}
1 & -i/3 & 1/3\\
i/3 & 1 & -i/3\\
1/3 & i/3 & 1
\end{pmatrix}
=\frac{4}{3}I_3-\frac{1}{3}ww^{*},\qquad
w=
\begin{pmatrix}
1\\
-i\\
-1
\end{pmatrix}. \tag*{(126)} H ∗ H = 1 i /3 1/3 − i /3 1 i /3 1/3 − i /3 1 = 3 4 I 3 − 3 1 w w ∗ , w = 1 − i − 1 . (126)
Since ∥ w ∥ 2 = 3 \lVert w\rVert^2=3 ∥ w ∥ 2 = 3 , the eigenvalue on span { w } \operatorname{span}\{w\} span { w } is 1 / 3 1/3 1/3 , while both eigenvalues on w ⊥ w^{\perp} w ⊥ are 4 / 3 4/3 4/3 . Therefore the family is a Riesz basis with exact squared bounds
A = 1 3 , B = 4 3 . (127) A=\frac{1}{3},\qquad B=\frac{4}{3}. \tag*{(127)} A = 3 1 , B = 3 4 . (127)
In particular,
σ ( H ~ ∗ H ~ − I ) = { − 2 3 , 1 3 } . \sigma(\widetilde{H}^{*}\widetilde{H}-I)=\left\{-\frac{2}{3},\frac{1}{3}\right\}. σ ( H ∗ H − I ) = { − 3 2 , 3 1 } .
which is not symmetric about zero. Thus the exact binary form 1 ± ρ 1\pm\rho 1 ± ρ does not extend to a general three-shift pattern; the correct periodic invariant is the full finite Gram matrix H p , τ ∗ H p , τ H^{*}_{p,\tau}H_{p,\tau} H p , τ ∗ H p , τ .
Two shifted lattices on two intervals
Integer translation
We now turn from index-dependent local perturbations to the disconnected-domain branch of the paper. The simplest nontrivial configuration consists of two complete frequency lattices acting on a union of two unit intervals. In this setting the synthesis operator reduces to an explicit 2 × 2 2 \times2 2 × 2 matrix field, providing a concrete model for the general finite-cell fiberization developed in the next section. Here the frequency set is the union of two complete lattices, Z \mathbb{Z} Z and Z + δ \mathbb{Z}+\delta Z + δ , while the spatial domain is the union of two intervals. The resulting synthesis map has two channels and is naturally represented by a 2 × 2 2 \times2 2 × 2 matrix.
Let α ∈ Z \alpha\in\mathbb{Z} α ∈ Z with ∣ α ∣ ≥ 1 |\alpha| \ge1 ∣ α ∣ ≥ 1 , let 0 < δ < 1 0 < \delta< 1 0 < δ < 1 , and define E = [ 0 , 1 ] ∪ [ α , α + 1 ] E = [0,1] \cup[\alpha,\alpha+1] E = [ 0 , 1 ] ∪ [ α , α + 1 ] . For finitely supported sequences a = ( a n ) a=(a_n) a = ( a n ) and b = ( b n ) b=(b_n) b = ( b n ) , set
A ( x ) = ∑ n ∈ Z a n e 2 π i n x , B δ ( x ) = ∑ n ∈ Z b n e 2 π i ( n + δ ) x , F ( x ) = A ( x ) + B δ ( x ) . A(x)=\sum_{n\in\mathbb{Z}}a_n e^{2\pi i n x},\qquad B_\delta(x)=\sum_{n\in\mathbb{Z}}b_n e^{2\pi i(n+\delta)x},\qquad F(x)=A(x)+B_\delta(x). A ( x ) = n ∈ Z ∑ a n e 2 π in x , B δ ( x ) = n ∈ Z ∑ b n e 2 π i ( n + δ ) x , F ( x ) = A ( x ) + B δ ( x ) .
Theorem 11.1 (Exact bounds and the integer-translation criterion). For every finitely supported a , b ∈ ℓ 2 ( Z ) a,b \in\ell^2(\mathbb{Z}) a , b ∈ ℓ 2 ( Z ) ,
2 ( 1 − ∣ cos ( π α δ ) ∣ ) ( ∥ a ∥ 2 2 + ∥ b ∥ 2 2 ) ≤ ∥ F ∥ L 2 ( E ) 2 ≤ 2 ( 1 + ∣ cos ( π α δ ) ∣ ) ( ∥ a ∥ 2 2 + ∥ b ∥ 2 2 ) . (128) \begin{aligned}
2(1-|\cos(\pi\alpha\delta)|)(\lVert a\rVert_2^2+\lVert b\rVert_2^2) &\le\lVert F\rVert_{L^2(E)}^2 \\
&\le2(1+|\cos(\pi\alpha\delta)|)(\lVert a\rVert_2^2+\lVert b\rVert_2^2). \tag*{(128)}
\end{aligned} 2 ( 1 − ∣ cos ( π α δ ) ∣ ) (∥ a ∥ 2 2 + ∥ b ∥ 2 2 ) ≤ ∥ F ∥ L 2 ( E ) 2 ≤ 2 ( 1 + ∣ cos ( π α δ ) ∣ ) (∥ a ∥ 2 2 + ∥ b ∥ 2 2 ) . (128) The constants are optimal. Moreover,
{ e 2 π i n x } n ∈ Z ∪ { e 2 π i ( n + δ ) x } n ∈ Z is a Riesz basis of L 2 ( E ) ⟺ α δ ∉ Z . (129) \{e^{2\pi i n x}\}_{n\in\mathbb{Z}} \cup\{e^{2\pi i(n+\delta)x}\}_{n\in\mathbb{Z}}\text{ is a Riesz basis of }L^2(E)
\Longleftrightarrow\alpha\delta\notin\mathbb{Z}. \tag*{(129)} { e 2 π in x } n ∈ Z ∪ { e 2 π i ( n + δ ) x } n ∈ Z is a Riesz basis of L 2 ( E ) ⟺ α δ ∈ / Z . (129)
Proof. Since α ∈ Z \alpha\in\mathbb{Z} α ∈ Z ,
A ( x + α ) = ∑ n a n e 2 π i n ( x + α ) = ∑ n a n e 2 π i n x e 2 π i n α = ∑ n a n e 2 π i n x = A ( x ) , A(x+\alpha)=\sum_n a_n e^{2\pi i n(x+\alpha)}=\sum_n a_n e^{2\pi i n x}e^{2\pi i n\alpha}=\sum_n a_n e^{2\pi i n x}=A(x), A ( x + α ) = n ∑ a n e 2 π in ( x + α ) = n ∑ a n e 2 π in x e 2 π in α = n ∑ a n e 2 π in x = A ( x ) ,
because n α ∈ Z n\alpha\in\mathbb{Z} n α ∈ Z . Similarly,
B δ ( x + α ) = ∑ n b n e 2 π i ( n + δ ) ( x + α ) = ∑ n b n e 2 π i ( n + δ ) x e 2 π i n α e 2 π i α δ = e 2 π i α δ B δ ( x ) . B_\delta(x+\alpha)=\sum_n b_n e^{2\pi i(n+\delta)(x+\alpha)}=\sum_n b_n e^{2\pi i(n+\delta)x}e^{2\pi i n\alpha}e^{2\pi i\alpha\delta}=e^{2\pi i\alpha\delta}B_\delta(x). B δ ( x + α ) = n ∑ b n e 2 π i ( n + δ ) ( x + α ) = n ∑ b n e 2 π i ( n + δ ) x e 2 π in α e 2 π i α δ = e 2 π i α δ B δ ( x ) .
Let ϑ = 2 π α δ \vartheta=2\pi\alpha\delta ϑ = 2 π α δ . For each x ∈ [ 0 , 1 ] x\in[0,1] x ∈ [ 0 , 1 ] , the two values of F F F on the two copies of the unit interval are
F ( x ) = A ( x ) + B δ ( x ) , F(x)=A(x)+B_\delta(x), F ( x ) = A ( x ) + B δ ( x ) ,
F ( x + α ) = A ( x ) + e i ϑ B δ ( x ) . F(x+\alpha)=A(x)+e^{i\vartheta}B_\delta(x). F ( x + α ) = A ( x ) + e i ϑ B δ ( x ) .
Writing these equations as a vector identity gives
\begin{pmatrix}
F(x)\\
F(x+\alpha)
\end{pmatrix}
=
\underbrace{\begin{pmatrix}
1 & 1\\
1 & e^{i\vartheta}
\end{pmatrix}}_{M_\vartheta}
\begin{pmatrix}
A(x)\\
B_\delta(x)
\end{pmatrix}. \tag*{(130)}\end{pmatrix}
The first column ( 1 , 1 ) T (1,1)^T ( 1 , 1 ) T records how A A A appears on the two intervals, while the second column ( 1 , e i ϑ ) T (1,e^{i\vartheta})^T ( 1 , e i ϑ ) T records how B δ B_\delta B δ appears.
Define
z ( x ) = ( A ( x ) B δ ( x ) ) . z(x)=\begin{pmatrix} A(x) \\ B_{\delta}(x) \end{pmatrix}. z ( x ) = ( A ( x ) B δ ( x ) ) .
Since the intervals are disjoint up to endpoints,
∥ F ∥ L 2 ( E ) 2 = ∫ 0 1 ∣ F ( x ) ∣ 2 d x + ∫ α α + 1 ∣ F ( y ) ∣ 2 d y = ∫ 0 1 ∣ F ( x ) ∣ 2 d x + ∫ 0 1 ∣ F ( x + α ) ∣ 2 d x (131) \|F\|_{L^2(E)}^2=\int_0^1 |F(x)|^2\,dx+\int_{\alpha}^{\alpha+1}|F(y)|^2\,dy=\int_0^1 |F(x)|^2\,dx+\int_0^1 |F(x+\alpha)|^2\,dx \tag*{(131)} ∥ F ∥ L 2 ( E ) 2 = ∫ 0 1 ∣ F ( x ) ∣ 2 d x + ∫ α α + 1 ∣ F ( y ) ∣ 2 d y = ∫ 0 1 ∣ F ( x ) ∣ 2 d x + ∫ 0 1 ∣ F ( x + α ) ∣ 2 d x (131)
= ∫ 0 1 ( ∣ F ( x ) ∣ 2 + ∣ F ( x + α ) ∣ 2 ) d x = ∫ 0 1 ∥ M ϑ z ( x ) ∥ C 2 2 d x . (132) =\int_0^1\left(|F(x)|^2+|F(x+\alpha)|^2\right)\,dx=\int_0^1\|M_{\vartheta}z(x)\|_{\mathbb{C}^2}^2\,dx. \tag*{(132)} = ∫ 0 1 ( ∣ F ( x ) ∣ 2 + ∣ F ( x + α ) ∣ 2 ) d x = ∫ 0 1 ∥ M ϑ z ( x ) ∥ C 2 2 d x . (132)
For any complex matrix M M M and vector z z z ,
∥ M z ∥ C 2 2 = ( M z ) ∗ ( M z ) = z ∗ M ∗ M z . (133) \|Mz\|_{\mathbb{C}^2}^2=(Mz)^*(Mz)=z^*M^*Mz. \tag*{(133)} ∥ M z ∥ C 2 2 = ( M z ) ∗ ( M z ) = z ∗ M ∗ M z . (133)
Thus the fiber Gram matrix is H ϑ = M ϑ ∗ M ϑ H_{\vartheta}=M_{\vartheta}^*M_{\vartheta} H ϑ = M ϑ ∗ M ϑ . Computing it entry by entry,
M ϑ ∗ = ( 1 1 1 e − i ϑ ) , H ϑ = ( 1 1 1 e − i ϑ ) ( 1 1 1 e i ϑ ) = ( 2 1 + e i ϑ 1 + e − i ϑ 2 ) . M_{\vartheta}^*=\begin{pmatrix}1&1\\1&e^{-i\vartheta}\end{pmatrix},\qquad H_{\vartheta}=\begin{pmatrix}1&1\\1&e^{-i\vartheta}\end{pmatrix}\begin{pmatrix}1&1\\1&e^{i\vartheta}\end{pmatrix}=\begin{pmatrix}2&1+e^{i\vartheta}\\1+e^{-i\vartheta}&2\end{pmatrix}. M ϑ ∗ = ( 1 1 1 e − i ϑ ) , H ϑ = ( 1 1 1 e − i ϑ ) ( 1 1 1 e i ϑ ) = ( 2 1 + e − i ϑ 1 + e i ϑ 2 ) .
Consequently,
∥ F ∥ L 2 ( E ) 2 = ∫ 0 1 z ( x ) ∗ H ϑ z ( x ) d x . (134) \|F\|_{L^2(E)}^2=\int_0^1 z(x)^*H_{\vartheta}z(x)\,dx. \tag*{(134)} ∥ F ∥ L 2 ( E ) 2 = ∫ 0 1 z ( x ) ∗ H ϑ z ( x ) d x . (134)
To find the best pointwise constants, compute the eigenvalues of H ϑ H_{\vartheta} H ϑ . They are the roots of
0 = det ( H ϑ − λ I ) = det ( 2 − λ 1 + e i ϑ 1 + e − i ϑ 2 − λ ) = ( 2 − λ ) 2 − ( 1 + e i ϑ ) ( 1 + e − i ϑ ) . (135) 0=\det(H_{\vartheta}-\lambda I)=\det\begin{pmatrix}2-\lambda&1+e^{i\vartheta}\\1+e^{-i\vartheta}&2-\lambda\end{pmatrix}=(2-\lambda)^2-(1+e^{i\vartheta})(1+e^{-i\vartheta}). \tag*{(135)} 0 = det ( H ϑ − λ I ) = det ( 2 − λ 1 + e − i ϑ 1 + e i ϑ 2 − λ ) = ( 2 − λ ) 2 − ( 1 + e i ϑ ) ( 1 + e − i ϑ ) . (135)
The product in the second term is
( 1 + e i ϑ ) ( 1 + e − i ϑ ) = 1 + e i ϑ + e − i ϑ + 1 = 2 + 2 cos ϑ = 4 cos 2 ( ϑ 2 ) . (1+e^{i\vartheta})(1+e^{-i\vartheta})=1+e^{i\vartheta}+e^{-i\vartheta}+1=2+2\cos\vartheta=4\cos^2\left(\frac{\vartheta}{2}\right). ( 1 + e i ϑ ) ( 1 + e − i ϑ ) = 1 + e i ϑ + e − i ϑ + 1 = 2 + 2 cos ϑ = 4 cos 2 ( 2 ϑ ) .
Therefore
( 2 − λ ) 2 = 4 cos 2 ( ϑ 2 ) , (2-\lambda)^2=4\cos^2\left(\frac{\vartheta}{2}\right), ( 2 − λ ) 2 = 4 cos 2 ( 2 ϑ ) ,
and the two eigenvalues are
λ − = 2 − 2 ∣ cos ( ϑ 2 ) ∣ , λ + = 2 + 2 ∣ cos ( ϑ 2 ) ∣ . (136) \lambda_-=2-2\left|\cos\left(\frac{\vartheta}{2}\right)\right|,\qquad\lambda_+=2+2\left|\cos\left(\frac{\vartheta}{2}\right)\right|. \tag*{(136)} λ − = 2 − 2 cos ( 2 ϑ ) , λ + = 2 + 2 cos ( 2 ϑ ) . (136)
Since ϑ / 2 = π α δ \vartheta/2=\pi\alpha\delta ϑ /2 = π α δ ,
λ ± = 2 ± 2 ∣ cos ( π α δ ) ∣ . (137) \lambda_{\pm}=2\pm2|\cos(\pi\alpha\delta)|. \tag*{(137)} λ ± = 2 ± 2∣ cos ( π α δ ) ∣. (137)
For every z ∈ C 2 z\in\mathbb{C}^2 z ∈ C 2 , the spectral theorem gives
λ − ∥ z ∥ C 2 2 ≤ z ∗ H ϑ z ≤ λ + ∥ z ∥ C 2 2 . (138) \lambda_-\|z\|_{\mathbb{C}^2}^2\le z^*H_{\vartheta}z\le\lambda_+\|z\|_{\mathbb{C}^2}^2. \tag*{(138)} λ − ∥ z ∥ C 2 2 ≤ z ∗ H ϑ z ≤ λ + ∥ z ∥ C 2 2 . (138)
Applying (137) to z = z ( x ) z=z(x) z = z ( x ) and integrating,
λ − ∫ 0 1 ( ∣ A ( x ) ∣ 2 + ∣ B δ ( x ) ∣ 2 ) d x ≤ ∥ F ∥ L 2 ( E ) 2 ≤ λ + ∫ 0 1 ( ∣ A ( x ) ∣ 2 + ∣ B δ ( x ) ∣ 2 ) d x . (139) \lambda_-\int_0^1\left(|A(x)|^2+|B_{\delta}(x)|^2\right)\,dx\le\|F\|_{L^2(E)}^2\le\lambda_+\int_0^1\left(|A(x)|^2+|B_{\delta}(x)|^2\right)\,dx. \tag*{(139)} λ − ∫ 0 1 ( ∣ A ( x ) ∣ 2 + ∣ B δ ( x ) ∣ 2 ) d x ≤ ∥ F ∥ L 2 ( E ) 2 ≤ λ + ∫ 0 1 ( ∣ A ( x ) ∣ 2 + ∣ B δ ( x ) ∣ 2 ) d x . (139)
Parseval gives
∫ 0 1 ∣ A ( x ) ∣ 2 d x = ∑ n ∣ a n ∣ 2 = ∥ a ∥ 2 2 , (140) \int_0^1 |A(x)|^2\,\mathrm{d}x=\sum_n |a_n|^2=\lVert a\rVert_2^2, \tag*{(140)} ∫ 0 1 ∣ A ( x ) ∣ 2 d x = n ∑ ∣ a n ∣ 2 = ∥ a ∥ 2 2 , (140)
and
∫ 0 1 ∣ B δ ( x ) ∣ 2 d x = ∫ 0 1 ∣ e 2 π i δ x ∑ n b n e 2 π i n x ∣ 2 d x = ∫ 0 1 ∣ ∑ n b n e 2 π i n x ∣ 2 d x = ∑ n ∣ b n ∣ 2 = ∥ b ∥ 2 2 . (141) \int_0^1 |B_\delta(x)|^2\,\mathrm{d}x=\int_0^1\left|e^{2\pi i\delta x}\sum_n b_n e^{2\pi i n x}\right|^2\,\mathrm{d}x=\int_0^1\left|\sum_n b_n e^{2\pi i n x}\right|^2\,\mathrm{d}x=\sum_n |b_n|^2=\lVert b\rVert_2^2. \tag*{(141)} ∫ 0 1 ∣ B δ ( x ) ∣ 2 d x = ∫ 0 1 e 2 π i δ x n ∑ b n e 2 π in x 2 d x = ∫ 0 1 n ∑ b n e 2 π in x 2 d x = n ∑ ∣ b n ∣ 2 = ∥ b ∥ 2 2 . (141)
Substituting (136), (139), and (140) into (138) gives (127).
The constants are optimal because A A A and B δ B_\delta B δ can be chosen independently as arbitrary L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) functions: multiplication by e 2 π i δ x e^{2\pi i\delta x} e 2 π i δ x is unitary, and both ordinary exponential systems are orthonormal bases of L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) . Thus z ( x ) z(x) z ( x ) may be chosen in an eigenvector direction of H ϑ H_\vartheta H ϑ . If one first restricts to finite sequences, the corresponding functions can be approximated in L 2 L^2 L 2 by finite Fourier sums.
It remains to characterize the basis property. The determinant of the coefficient matrix is
det M ϑ = e i ϑ − 1. (142) \det M_\vartheta=e^{i\vartheta}-1. \tag*{(142)} det M ϑ = e i ϑ − 1. (142)
Thus
M ϑ is invertible ⟺ e i ϑ ≠ 1 (143) M_\vartheta\text{ is invertible}\Longleftrightarrow e^{i\vartheta}\ne1 \tag*{(143)} M ϑ is invertible ⟺ e i ϑ = 1 (143)
⟺ ϑ ∉ 2 π Z (144) \Longleftrightarrow\vartheta\notin2\pi\mathbb{Z} \tag*{(144)} ⟺ ϑ ∈ / 2 π Z (144)
⟺ α δ ∉ Z . (145) \Longleftrightarrow\alpha\delta\notin\mathbb{Z}. \tag*{(145)} ⟺ α δ ∈ / Z . (145)
This is equivalent to cos ( π α δ ) ≠ ± 1 \cos(\pi\alpha\delta)\ne\pm1 cos ( π α δ ) = ± 1 .
Assume first that α δ ∉ Z \alpha\delta\notin\mathbb{Z} α δ ∈ / Z . Given an arbitrary g ∈ L 2 ( E ) g\in L^2(E) g ∈ L 2 ( E ) , define
y ( x ) = ( g ( x ) g ( x + α ) ) , 0 ≤ x < 1. (146) y(x)=\begin{pmatrix}g(x)\\g(x+\alpha)\end{pmatrix},\qquad0\le x<1. \tag*{(146)} y ( x ) = ( g ( x ) g ( x + α ) ) , 0 ≤ x < 1. (146)
Then
z ( x ) = M ϑ − 1 y ( x ) (147) z(x)=M_\vartheta^{-1}y(x) \tag*{(147)} z ( x ) = M ϑ − 1 y ( x ) (147)
belongs to L 2 ( 0 , 1 ; C 2 ) L^2(0,1;\mathbb{C}^2) L 2 ( 0 , 1 ; C 2 ) . Write z = ( A , B δ ) T z=(A,B_\delta)^{\mathsf T} z = ( A , B δ ) T , define
B ( x ) = e − 2 π i δ x B δ ( x ) , B(x)=e^{-2\pi i\delta x}B_\delta(x), B ( x ) = e − 2 π i δ x B δ ( x ) ,
and take the ordinary Fourier coefficients of A A A and B B B . This produces a , b ∈ ℓ 2 ( Z ) a,b\in\ell^2(\mathbb{Z}) a , b ∈ ℓ 2 ( Z ) whose synthesis equals g g g . Hence the synthesis operator is onto; together with (127), this proves the Riesz-basis assertion.
Conversely, suppose α δ ∈ Z \alpha\delta\in\mathbb{Z} α δ ∈ Z . Then e i ϑ = 1 e^{i\vartheta}=1 e i ϑ = 1 and the two columns of M ϑ M_\vartheta M ϑ coincide. Choose any nonzero h ∈ L 2 ( 0 , 1 ) h\in L^2(0,1) h ∈ L 2 ( 0 , 1 ) and set
A = h , B δ = − h . (148) A=h,\qquad B_\delta=-h. \tag*{(148)} A = h , B δ = − h . (148)
Define B ( x ) = e − 2 π i δ x B δ ( x ) B(x)=e^{-2\pi i\delta x}B_\delta(x) B ( x ) = e − 2 π i δ x B δ ( x ) and let a , b ∈ ℓ 2 ( Z ) a,b\in\ell^2(\mathbb{Z}) a , b ∈ ℓ 2 ( Z ) be the Fourier coefficient sequences of A A A and B B B . Then
F ( x ) = A ( x ) + B δ ( x ) = 0 F(x)=A(x)+B_\delta(x)=0 F ( x ) = A ( x ) + B δ ( x ) = 0
on the first interval, while on the second interval
F ( x + α ) = A ( x ) + e i ϑ B δ ( x ) = h ( x ) − h ( x ) = 0. F(x+\alpha)=A(x)+e^{i\vartheta}B_\delta(x)=h(x)-h(x)=0. F ( x + α ) = A ( x ) + e i ϑ B δ ( x ) = h ( x ) − h ( x ) = 0.
The coefficient pair is nonzero. Finite truncations of the coefficient sequences approximate this nonzero kernel vector and show that no positive lower Riesz bound can hold. This proves the converse in (128).
Non-integer translation
Exponential bases on disconnected intervals have been constructed by several methods, including perturbations of integer frequencies subject to conditions depending on the gaps between the intervals; see Asipchuk and Drezels [7 ] . The related question of preserving exponential bases under modifications of the interval positions has been studied in Asipchuk [8 ] . Our problem is different in that the frequency set is fixed to the union of two complete shifted lattices, Z ∪ ( Z + δ ) \mathbb{Z} \cup(\mathbb{Z} + \delta) Z ∪ ( Z + δ ) , while the relative position of the two intervals is allowed to vary. We seek necessary and sufficient conditions for the Riesz-basis property, together with the optimal bounds. For a non-integer translation, folding modulo one produces two distinct matrix fibers, and stability requires both of them to be invertible.
Let 0 < δ < 1 0 < \delta< 1 0 < δ < 1 , let α ∈ R ∖ Z \alpha\in\mathbb{R} \setminus\mathbb{Z} α ∈ R ∖ Z , and assume that [ 0 , 1 ] [0,1] [ 0 , 1 ] and [ α , α + 1 ] [\alpha,\alpha+1] [ α , α + 1 ] are disjoint. Write
α = q + β , q = ⌊ α ⌋ , 0 < β < 1 , (149) \alpha= q + \beta,\qquad q = \lfloor\alpha\rfloor,\qquad0 < \beta< 1, \tag*{(149)} α = q + β , q = ⌊ α ⌋ , 0 < β < 1 , (149)
and set E = [ 0 , 1 ] ∪ [ α , α + 1 ] E = [0,1] \cup[\alpha,\alpha+1] E = [ 0 , 1 ] ∪ [ α , α + 1 ] . For finitely supported a = ( a n ) a = (a_n) a = ( a n ) and b = ( b n ) b = (b_n) b = ( b n ) define
F ( x ) = ∑ n ∈ Z a n e 2 π i n x + ∑ n ∈ Z b n e 2 π i ( n + δ ) x . (150) F(x) = \sum_{n\in\mathbb{Z}} a_n e^{2\pi i n x} + \sum_{n\in\mathbb{Z}} b_n e^{2\pi i(n+\delta)x}. \tag*{(150)} F ( x ) = n ∈ Z ∑ a n e 2 π in x + n ∈ Z ∑ b n e 2 π i ( n + δ ) x . (150)
Let
c α , δ = 2 max { ∣ cos ( π q δ ) ∣ , ∣ cos ( π ( q + 1 ) δ ) ∣ } . (151) c_{\alpha,\delta} = 2\max\{|\cos(\pi q\delta)|,|\cos(\pi(q+1)\delta)|\}. \tag*{(151)} c α , δ = 2 max { ∣ cos ( π q δ ) ∣ , ∣ cos ( π ( q + 1 ) δ ) ∣ } . (151)
Theorem 11.2 (Exact bounds and the non-integer criterion). *For every finitely supported a , b ∈ ℓ 2 ( Z ) a,b \in\ell^2(\mathbb{Z}) a , b ∈ ℓ 2 ( Z ) ,
( 2 − c α , δ ) ( ∥ a ∥ 2 + ∥ b ∥ 2 ) ≤ ∥ F ∥ L 2 ( E ) 2 ≤ ( 2 + c α , δ ) ( ∥ a ∥ 2 + ∥ b ∥ 2 ) . (152) (2-c_{\alpha,\delta})(\lVert a\rVert^2+\lVert b\rVert^2) \leq\lVert F\rVert^2_{L^2(E)} \leq(2+c_{\alpha,\delta})(\lVert a\rVert^2+\lVert b\rVert^2). \tag*{(152)} ( 2 − c α , δ ) (∥ a ∥ 2 + ∥ b ∥ 2 ) ≤ ∥ F ∥ L 2 ( E ) 2 ≤ ( 2 + c α , δ ) (∥ a ∥ 2 + ∥ b ∥ 2 ) . (152) The constants are optimal. Moreover,
{ e 2 π i n x } n ∈ Z ∪ { e 2 π i ( n + δ ) x } n ∈ Z is a Riesz basis of L 2 ( E ) ⟺ q δ ∉ Z and ( q + 1 ) δ ∉ Z . (153) \{e^{2\pi i n x}\}_{n\in\mathbb{Z}} \cup\{e^{2\pi i(n+\delta)x}\}_{n\in\mathbb{Z}}\text{ is a Riesz basis of }L^2(E)
\Longleftrightarrow q\delta\notin\mathbb{Z}\quad\text{and}\quad(q+1)\delta\notin\mathbb{Z}. \tag*{(153)} { e 2 π in x } n ∈ Z ∪ { e 2 π i ( n + δ ) x } n ∈ Z is a Riesz basis of L 2 ( E ) ⟺ q δ ∈ / Z and ( q + 1 ) δ ∈ / Z . (153)
Proof. Write
A ( x ) = ∑ n a n e 2 π i n x , B ( x ) = ∑ n b n e 2 π i n x , C ( x ) = e 2 π i δ x B ( x ) . (154) A(x)=\sum_n a_n e^{2\pi i n x},\qquad B(x)=\sum_n b_n e^{2\pi i n x},\qquad C(x)=e^{2\pi i\delta x}B(x). \tag*{(154)} A ( x ) = n ∑ a n e 2 π in x , B ( x ) = n ∑ b n e 2 π in x , C ( x ) = e 2 π i δ x B ( x ) . (154)
Then
F ( x ) = A ( x ) + C ( x ) . (155) F(x)=A(x)+C(x). \tag*{(155)} F ( x ) = A ( x ) + C ( x ) . (155)
Both A A A and B B B are one-periodic.
The second interval is
[ α , α + 1 ] = [ q + β , q + β + 1 ] . [\alpha,\alpha+1]=[q+\beta,q+\beta+1]. [ α , α + 1 ] = [ q + β , q + β + 1 ] .
Split it as
[ q + β , q + 1 ) ∪ [ q + 1 , q + β + 1 ] . (156) [q+\beta,q+1)\cup[q+1,q+\beta+1]. \tag*{(156)} [ q + β , q + 1 ) ∪ [ q + 1 , q + β + 1 ] . (156)
The first piece is parameterized by y = x + q y=x+q y = x + q , with β ≤ x < 1 \beta\leq x<1 β ≤ x < 1 , and the second by y = x + q + 1 y=x+q+1 y = x + q + 1 , with 0 ≤ x < β 0\leq x<\beta 0 ≤ x < β . Hence
∫ α α + 1 ∣ F ( y ) ∣ 2 d y = ∫ β 1 ∣ F ( x + q ) ∣ 2 d x + ∫ 0 β ∣ F ( x + q + 1 ) ∣ 2 d x . (157) \int_{\alpha}^{\alpha+1}|F(y)|^2\,dy
=
\int_{\beta}^{1}|F(x+q)|^2\,dx
+
\int_{0}^{\beta}|F(x+q+1)|^2\,dx. \tag*{(157)} ∫ α α + 1 ∣ F ( y ) ∣ 2 d y = ∫ β 1 ∣ F ( x + q ) ∣ 2 d x + ∫ 0 β ∣ F ( x + q + 1 ) ∣ 2 d x . (157)
For an arbitrary integer k k k , periodicity gives
A ( x + k ) = A ( x ) , C ( x + k ) = e 2 π i δ ( x + k ) B ( x + k ) = e 2 π i k δ e 2 π i δ x B ( x ) = e 2 π i k δ C ( x ) . A(x+k)=A(x),\qquad C(x+k)=e^{2\pi i\delta(x+k)}B(x+k)=e^{2\pi ik\delta}e^{2\pi i\delta x}B(x)=e^{2\pi ik\delta}C(x). A ( x + k ) = A ( x ) , C ( x + k ) = e 2 π i δ ( x + k ) B ( x + k ) = e 2 π ik δ e 2 π i δ x B ( x ) = e 2 π ik δ C ( x ) .
Therefore
F ( x + q + 1 ) = A ( x ) + e 2 π i ( q + 1 ) δ C ( x ) , 0 ≤ x < β , F ( x + q ) = A ( x ) + e 2 π i q δ C ( x ) , β ≤ x < 1. (158) \begin{aligned}
F(x+q+1)&=A(x)+e^{2\pi i(q+1)\delta}C(x), &&0\le x<\beta,\\
F(x+q)&=A(x)+e^{2\pi iq\delta}C(x), &&\beta\le x<1. \tag*{(158)}
\end{aligned} F ( x + q + 1 ) F ( x + q ) = A ( x ) + e 2 π i ( q + 1 ) δ C ( x ) , = A ( x ) + e 2 π i q δ C ( x ) , 0 ≤ x < β , β ≤ x < 1. (158)
Combining (155) , (158) , and (158), define
M k = ( 1 1 1 e 2 π i k δ ) . (159) M_k=\begin{pmatrix}1&1\\1&e^{2\pi ik\delta}\end{pmatrix}. \tag*{(159)} M k = ( 1 1 1 e 2 π ik δ ) . (159)
Then
\begin{pmatrix}F(x)\\F(x+q+1)\end{pmatrix}
=M_{q+1}\begin{pmatrix}A(x)\\C(x)\end{pmatrix},\qquad0\le x<\beta, \tag*{(160)}\end{pmatrix}
and
\begin{pmatrix}F(x)\\F(x+q)\end{pmatrix}
=M_q\begin{pmatrix}A(x)\\C(x)\end{pmatrix},\qquad\beta\le x<1. \tag*{(161)}\end{pmatrix}
Thus, with z ( x ) = ( A ( x ) , C ( x ) ) T z(x)=(A(x),C(x))^{T} z ( x ) = ( A ( x ) , C ( x ) ) T ,
∥ F ∥ L 2 ( E ) 2 = ∫ 0 β ∥ M q + 1 z ( x ) ∥ C 2 2 d x + ∫ β 1 ∥ M q z ( x ) ∥ C 2 2 d x . (162) \lVert F\rVert_{L^2(E)}^2=\int_0^\beta\lVert M_{q+1}z(x)\rVert_{\mathbb{C}^2}^2\,\mathrm{d}x+\int_\beta^1\lVert M_qz(x)\rVert_{\mathbb{C}^2}^2\,\mathrm{d}x. \tag*{(162)} ∥ F ∥ L 2 ( E ) 2 = ∫ 0 β ∥ M q + 1 z ( x ) ∥ C 2 2 d x + ∫ β 1 ∥ M q z ( x ) ∥ C 2 2 d x . (162)
For each integer k k k , the Gram matrix is
H k = M k ∗ M k = ( 2 1 + e 2 π i k δ 1 + e − 2 π i k δ 2 ) . (163) H_k=M_k^*M_k=\begin{pmatrix}2&1+e^{2\pi ik\delta}\\1+e^{-2\pi ik\delta}&2\end{pmatrix}. \tag*{(163)} H k = M k ∗ M k = ( 2 1 + e − 2 π ik δ 1 + e 2 π ik δ 2 ) . (163)
Repeating the determinant calculation in (135) –(136) , now with ϑ = 2 π k δ \vartheta=2\pi k\delta ϑ = 2 π k δ , gives
λ ± ( H k ) = 2 ± 2 ∣ cos ( π k δ ) ∣ . (164) \lambda_{\pm}(H_k)=2\pm2\lvert\cos(\pi k\delta)\rvert. \tag*{(164)} λ ± ( H k ) = 2 ± 2 ∣ cos ( π k δ )∣ . (164)
Consequently, on [ 0 , β ) [0,\beta) [ 0 , β ) ,
( 2 − 2 ∣ cos ( π ( q + 1 ) δ ) ∣ ) ∥ z ( x ) ∥ 2 ≤ ∥ M q + 1 z ( x ) ∥ 2 ≤ ( 2 + 2 ∣ cos ( π ( q + 1 ) δ ) ∣ ) ∥ z ( x ) ∥ 2 , (165) \left(2-2\lvert\cos(\pi(q+1)\delta)\rvert\right)\lVert z(x)\rVert^2
\le\lVert M_{q+1}z(x)\rVert^2
\le\left(2+2\lvert\cos(\pi(q+1)\delta)\rvert\right)\lVert z(x)\rVert^2, \tag*{(165)} ( 2 − 2 ∣ cos ( π ( q + 1 ) δ )∣ ) ∥ z ( x ) ∥ 2 ≤ ∥ M q + 1 z ( x ) ∥ 2 ≤ ( 2 + 2 ∣ cos ( π ( q + 1 ) δ )∣ ) ∥ z ( x ) ∥ 2 , (165)
whereas on [ β , 1 ) [\beta,1) [ β , 1 ) ,
( 2 − 2 ∣ cos ( π q δ ) ∣ ) ∥ z ( x ) ∥ 2 ≤ ∥ M q z ( x ) ∥ 2 ≤ ( 2 + 2 ∣ cos ( π q δ ) ∣ ) ∥ z ( x ) ∥ 2 . (166) \left(2-2\lvert\cos(\pi q\delta)\rvert\right)\lVert z(x)\rVert^2
\le\lVert M_qz(x)\rVert^2
\le\left(2+2\lvert\cos(\pi q\delta)\rvert\right)\lVert z(x)\rVert^2. \tag*{(166)} ( 2 − 2 ∣ cos ( π q δ )∣ ) ∥ z ( x ) ∥ 2 ≤ ∥ M q z ( x ) ∥ 2 ≤ ( 2 + 2 ∣ cos ( π q δ )∣ ) ∥ z ( x ) ∥ 2 . (166)
The global lower constant is the smaller of the two lower constants:
min { 2 − 2 ∣ cos ( π q δ ) ∣ , 2 − 2 ∣ cos ( π ( q + 1 ) δ ) ∣ } = 2 − 2 max { ∣ cos ( π q δ ) ∣ , ∣ cos ( π ( q + 1 ) δ ) ∣ } = 2 − c α , δ . \begin{aligned}
\min\{2-2\lvert\cos(\pi q\delta)\rvert,2-2\lvert\cos(\pi(q+1)\delta)\rvert\}
&=2-2\max\{\lvert\cos(\pi q\delta)\rvert,\lvert\cos(\pi(q+1)\delta)\rvert\}\\
&=2-c_{\alpha,\delta}.
\end{aligned} min { 2 − 2 ∣ cos ( π q δ )∣ , 2 − 2 ∣ cos ( π ( q + 1 ) δ )∣} = 2 − 2 max {∣ cos ( π q δ )∣ , ∣ cos ( π ( q + 1 ) δ )∣} = 2 − c α , δ .
Similarly, the global upper constant is
max { 2 + 2 ∣ cos ( π q δ ) ∣ , 2 + 2 ∣ cos ( π ( q + 1 ) δ ) ∣ } = 2 + 2 max { ∣ cos ( π q δ ) ∣ , ∣ cos ( π ( q + 1 ) δ ) ∣ } = 2 + c α , δ . \begin{aligned}
\max\{2+2\lvert\cos(\pi q\delta)\rvert,2+2\lvert\cos(\pi(q+1)\delta)\rvert\} \\
&= 2 + 2\max\{|\cos(\pi q\delta)|,|\cos(\pi(q+1)\delta)|\}=2+c_{\alpha,\delta}.
\end{aligned} max { 2 + 2 ∣ cos ( π q δ )∣ , 2 + 2 ∣ cos ( π ( q + 1 ) δ )∣} = 2 + 2 max { ∣ cos ( π q δ ) ∣ , ∣ cos ( π ( q + 1 ) δ ) ∣ } = 2 + c α , δ .
Integrating (165) –(166) and using
∫ 0 1 ∥ z ( x ) ∥ C 2 2 d x = ∫ 0 1 ( ∣ A ( x ) ∣ 2 + ∣ C ( x ) ∣ 2 ) d x = ∥ a ∥ 2 2 + ∥ b ∥ 2 2 \int_{0}^{1}\lVert z(x)\rVert_{\mathbb{C}^{2}}^{2}\,\mathrm{d}x=\int_{0}^{1}\left(|A(x)|^{2}+|C(x)|^{2}\right)\,\mathrm{d}x=\lVert a\rVert_{2}^{2}+\lVert b\rVert_{2}^{2} ∫ 0 1 ∥ z ( x ) ∥ C 2 2 d x = ∫ 0 1 ( ∣ A ( x ) ∣ 2 + ∣ C ( x ) ∣ 2 ) d x = ∥ a ∥ 2 2 + ∥ b ∥ 2 2
gives (152) .
The constants do not depend on β \beta β . Even if one of the two pieces has small but positive length, an L 2 L^{2} L 2 vector function may be supported entirely on that piece, and the matrix on that piece can determine the global lower or upper bound. Choosing an eigenvector direction on the relevant piece, followed by finite Fourier approximation, also proves optimality.
We now prove the basis criterion. First suppose
q δ ∉ Z and ( q + 1 ) δ ∉ Z . (167) q\delta\notin\mathbb{Z}\qquad\text{and}\qquad(q+1)\delta\notin\mathbb{Z}. \tag*{(167)} q δ ∈ / Z and ( q + 1 ) δ ∈ / Z . (167)
Then both matrices M q M_q M q and M q + 1 M_{q+1} M q + 1 are invertible. The piecewise matrix
M ( x ) = { M q + 1 , 0 ≤ x < β , M q , β ≤ x < 1 (168) M(x)=
\begin{cases}
M_{q+1}, & 0\le x<\beta,\\
M_q, & \beta\le x<1
\end{cases} \tag*{(168)} M ( x ) = { M q + 1 , M q , 0 ≤ x < β , β ≤ x < 1 (168)
has an essentially bounded inverse because it assumes only two invertible matrix values. Given any g ∈ L 2 ( E ) g\in L^{2}(E) g ∈ L 2 ( E ) , unfold it into
y ( x ) = { ( g ( x ) , g ( x + q + 1 ) ) T , 0 ≤ x < β , ( g ( x ) , g ( x + q ) ) T , β ≤ x < 1. (169) y(x)=
\begin{cases}
(g(x),g(x+q+1))^{T}, & 0\le x<\beta,\\
(g(x),g(x+q))^{T}, & \beta\le x<1.
\end{cases} \tag*{(169)} y ( x ) = { ( g ( x ) , g ( x + q + 1 ) ) T , ( g ( x ) , g ( x + q ) ) T , 0 ≤ x < β , β ≤ x < 1. (169)
Set z ( x ) = M ( x ) − 1 y ( x ) z(x)=M(x)^{-1}y(x) z ( x ) = M ( x ) − 1 y ( x ) . Then z = ( A , C ) T ∈ L 2 ( 0 , 1 ; C 2 ) z=(A,C)^{T}\in L^{2}(0,1;\mathbb{C}^{2}) z = ( A , C ) T ∈ L 2 ( 0 , 1 ; C 2 ) . Defining B ( x ) = e − 2 π i δ x C ( x ) B(x)=e^{-2\pi i\delta x}C(x) B ( x ) = e − 2 π i δ x C ( x ) and taking the Fourier coefficients of A A A and B B B gives sequences a , b ∈ ℓ 2 ( Z ) a,b\in\ell^{2}(\mathbb{Z}) a , b ∈ ℓ 2 ( Z ) whose synthesis equals g g g . Thus the synthesis map is onto, and (152) gives the two-sided estimate.
Conversely, suppose first that q δ ∈ Z q\delta\in\mathbb{Z} q δ ∈ Z . Then e 2 π i q δ = 1 e^{2\pi iq\delta}=1 e 2 π i q δ = 1 , and the matrix on [ β , 1 ) [\beta,1) [ β , 1 ) becomes
M q = ( 1 1 1 1 ) . (170) M_q=
\begin{pmatrix}
1 & 1\\
1 & 1
\end{pmatrix}. \tag*{(170)} M q = ( 1 1 1 1 ) . (170)
This matrix has the nonzero null vector ( 1 , − 1 ) T (1,-1)^{T} ( 1 , − 1 ) T . Choose a nonzero function
h ∈ L 2 ( 0 , 1 ) , supp h ⊆ [ β , 1 ) , (171) h\in L^{2}(0,1),\qquad\operatorname{supp}h\subseteq[\beta,1), \tag*{(171)} h ∈ L 2 ( 0 , 1 ) , supp h ⊆ [ β , 1 ) , (171)
and set
A ( x ) = h ( x ) , C ( x ) = − h ( x ) . (172) A(x)=h(x),\qquad C(x)=-h(x). \tag*{(172)} A ( x ) = h ( x ) , C ( x ) = − h ( x ) . (172)
On [ β , 1 ) [\beta,1) [ β , 1 ) ,
F ( x ) = A ( x ) + C ( x ) = h ( x ) − h ( x ) = 0 , F ( x + q ) = A ( x ) + e 2 π i q δ C ( x ) = h ( x ) − h ( x ) = 0. (173) \begin{aligned}
F(x)&=A(x)+C(x)=h(x)-h(x)=0,\\
F(x+q)&=A(x)+e^{2\pi iq\delta}C(x)=h(x)-h(x)=0. \tag*{(173)}
\end{aligned} F ( x ) F ( x + q ) = A ( x ) + C ( x ) = h ( x ) − h ( x ) = 0 , = A ( x ) + e 2 π i q δ C ( x ) = h ( x ) − h ( x ) = 0. (173)
On [ 0 , β ) [0,\beta) [ 0 , β ) , both A A A and C C C vanish by the support condition, so the two corresponding values of F F F also vanish. Thus F = 0 F=0 F = 0 almost everywhere on E E E .
To translate this construction into coefficient sequences, define again B ( x ) = e − 2 π i δ x C ( x ) B(x)=e^{-2\pi i\delta x}C(x) B ( x ) = e − 2 π i δ x C ( x ) . Since multiplication by e − 2 π i δ x e^{-2\pi i\delta x} e − 2 π i δ x is unitary on L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) , both A A A and B B B belong to L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) . Let
a n = ∫ 0 1 A ( x ) e − 2 π i n x d x , b n = ∫ 0 1 B ( x ) e − 2 π i n x d x . (174) a_n=\int_0^1 A(x)e^{-2\pi i n x}\,\mathrm{d}x,\qquad
b_n=\int_0^1 B(x)e^{-2\pi i n x}\,\mathrm{d}x. \tag*{(174)} a n = ∫ 0 1 A ( x ) e − 2 π in x d x , b n = ∫ 0 1 B ( x ) e − 2 π in x d x . (174)
Parseval gives a , b ∈ ℓ 2 ( Z ) a,b\in\ell^2(\mathbb{Z}) a , b ∈ ℓ 2 ( Z ) , and they are not both zero because h ≠ 0 h\ne0 h = 0 . Hence the synthesis operator has a nontrivial kernel. Finite truncations again show that no positive lower Riesz bound can hold.
If instead ( q + 1 ) δ ∈ Z (q+1)\delta\in\mathbb{Z} ( q + 1 ) δ ∈ Z , use exactly the same construction with a nonzero h h h supported in [ 0 , β ) [0,\beta) [ 0 , β ) . On that interval, M q + 1 M_{q+1} M q + 1 has identical columns, and the choice A = h A=h A = h , C = − h C=-h C = − h again produces a nonzero coefficient pair whose synthesis is zero. This completes the converse and proves (152).
Matrix fiberization on finite unions of intervals
The two-interval calculations of the preceding section are instances of a more general folding principle. We first extend this mechanism to finitely many translated unit intervals and then to arbitrary finite unions of bounded intervals, where the target fiber dimension may vary from cell to cell.
The fiberization of exponential systems over lattice multi-tiles and sub-multi-tiles is well established. In particular, Agora et al. [9 ] obtained Riesz-basis results for lattice multi-tiles using finite-dimensional fiber matrices, together with existence and converse multi-tiling theorems. Barbieri et al. [10 ] subsequently treated exponential frames on sub-multi-tiles and gave exact frame criteria and optimal bounds in terms of possibly rectangular fiber matrices. Related explicit Fourier-matrix criteria for unions of translated cubes were also studied in Asipchuk et al. (2025).
Our purpose in this section is not to introduce a new general fiberization principle, but to develop its explicit finite-cell realization for finite unions of intervals. In this geometry, the fiber changes are determined directly by the integer and fractional parts of the interval endpoints. This permits a concrete description of the local phase matrices, their exact spectral bounds, and their dependence on endpoint configurations.
We retain the proofs of the rank and spectral criteria in this specialized setting in order to make the subsequent calculations and applications self-contained. The resulting framework also provides the notation used for the arithmetic fiber-exchange arguments and the commensurate-lattice reduction below.
Finitely many translated unit intervals
Let
E = ⋃ j = 1 J [ α j , α j + 1 ) , (175) E=\bigcup_{j=1}^{J}[\alpha_j,\alpha_j+1), \tag*{(175)} E = j = 1 ⋃ J [ α j , α j + 1 ) , (175)
where the intervals are pairwise disjoint up to endpoints, and write
α j = q j + β j , q j ∈ Z , 0 ≤ β j < 1. (176) \alpha_j=q_j+\beta_j,\qquad q_j\in\mathbb{Z},\qquad0\le\beta_j<1. \tag*{(176)} α j = q j + β j , q j ∈ Z , 0 ≤ β j < 1. (176)
Let δ 1 , … , δ K \delta_1,\ldots,\delta_K δ 1 , … , δ K be pairwise incongruent modulo Z \mathbb{Z} Z , and define
Λ = ⋃ k = 1 K ( Z + δ k ) . (177) \Lambda=\bigcup_{k=1}^{K}(\mathbb{Z}+\delta_k). \tag*{(177)} Λ = k = 1 ⋃ K ( Z + δ k ) . (177)
For coefficient sequences a ( k ) = ( a n ( k ) ) n ∈ Z a^{(k)}=(a_n^{(k)})_{n\in\mathbb{Z}} a ( k ) = ( a n ( k ) ) n ∈ Z , set
( S a ) ( y ) = ∑ k = 1 K ∑ n ∈ Z a n ( k ) e 2 π i ( n + δ k ) y . (178) (Sa)(y)=\sum_{k=1}^{K}\sum_{n\in\mathbb{Z}}a_n^{(k)}e^{2\pi i(n+\delta_k)y}. \tag*{(178)} ( S a ) ( y ) = k = 1 ∑ K n ∈ Z ∑ a n ( k ) e 2 π i ( n + δ k ) y . (178)
For x ∈ [ 0 , 1 ) x\in[0,1) x ∈ [ 0 , 1 ) , the point of the j j j -th interval having fractional part x x x is
y j ( x ) = x + κ j ( x ) , κ j ( x ) = q j + 1 [ 0 , β j ) ( x ) . (179) y_j(x)=x+\kappa_j(x),\qquad\kappa_j(x)=q_j+\mathbf{1}_{[0,\beta_j)}(x). \tag*{(179)} y j ( x ) = x + κ j ( x ) , κ j ( x ) = q j + 1 [ 0 , β j ) ( x ) . (179)
For each frequency channel put
B k ( x ) = ∑ n ∈ Z a n ( k ) e 2 π i n x , A k ( x ) = e 2 π i δ k x B k ( x ) , (180) B_k(x)=\sum_{n\in\mathbb{Z}}a_n^{(k)}e^{2\pi i n x},\qquad A_k(x)=e^{2\pi i\delta_k x}B_k(x), \tag*{(180)} B k ( x ) = n ∈ Z ∑ a n ( k ) e 2 π in x , A k ( x ) = e 2 π i δ k x B k ( x ) , (180)
and define the J × K J\times K J × K matrix field
M ( x ) = ( e 2 π i δ k κ j ( x ) ) 1 ≤ j ≤ J 1 ≤ k ≤ K . (181) M(x)=\left(e^{2\pi i\delta_k\kappa_j(x)}\right)_{\substack{1\leq j\leq J\\1\leq k\leq K}}. \tag*{(181)} M ( x ) = ( e 2 π i δ k κ j ( x ) ) 1 ≤ j ≤ J 1 ≤ k ≤ K . (181)
Theorem 12.1 (Exact matrix-symbol fiberization for unit intervals). With the preceding notation,
∥ S a ∥ L 2 ( E ) 2 = ∫ 0 1 ∥ M ( x ) A ( x ) ∥ C J 2 d x , (182) \lVert Sa\rVert_{L^2(E)}^2=\int_0^1\lVert M(x)A(x)\rVert_{\mathbb{C}^J}^2\,dx, \tag*{(182)} ∥ S a ∥ L 2 ( E ) 2 = ∫ 0 1 ∥ M ( x ) A ( x ) ∥ C J 2 d x , (182) where A = ( A 1 , … , A K ) T A=(A_1,\ldots,A_K)^{\mathsf{T}} A = ( A 1 , … , A K ) T . Moreover,
∫ 0 1 ∥ A ( x ) ∥ C K 2 d x = ∑ k = 1 K ∥ a ( k ) ∥ ℓ 2 2 . (183) \int_0^1\lVert A(x)\rVert_{\mathbb{C}^K}^2\,dx=\sum_{k=1}^{K}\lVert a^{(k)}\rVert_{\ell^2}^2. \tag*{(183)} ∫ 0 1 ∥ A ( x ) ∥ C K 2 d x = k = 1 ∑ K ∥ a ( k ) ∥ ℓ 2 2 . (183) Every vector function in L 2 ( 0 , 1 ; C K ) L^2(0,1;\mathbb{C}^K) L 2 ( 0 , 1 ; C K ) arises from a unique coefficient family.
Proof. At y j ( x ) = x + κ j ( x ) y_j(x)=x+\kappa_j(x) y j ( x ) = x + κ j ( x ) , periodicity of B k B_k B k gives
( S a ) ( y j ( x ) ) = ∑ k = 1 K e 2 π i δ k ( x + κ j ( x ) ) B k ( x + κ j ( x ) ) = ∑ k = 1 K e 2 π i δ k κ j ( x ) A k ( x ) . \begin{aligned}
(Sa)(y_j(x))=\sum_{k=1}^{K}e^{2\pi i\delta_k(x+\kappa_j(x))}B_k(x+\kappa_j(x)) \\
&=\sum_{k=1}^{K}e^{2\pi i\delta_k\kappa_j(x)}A_k(x).
\end{aligned} ( S a ) ( y j ( x )) = k = 1 ∑ K e 2 π i δ k ( x + κ j ( x )) B k ( x + κ j ( x )) = k = 1 ∑ K e 2 π i δ k κ j ( x ) A k ( x ) .
Thus the vector of folded values is M ( x ) A ( x ) M(x)A(x) M ( x ) A ( x ) . Changing variables on each disjoint interval gives (182). Parseval and ∣ e 2 π i δ k x ∣ = 1 \lvert e^{2\pi i\delta_kx}\rvert=1 ∣ e 2 π i δ k x ∣ = 1 give (183).
Conversely, given A k ∈ L 2 ( 0 , 1 ) A_k\in L^2(0,1) A k ∈ L 2 ( 0 , 1 ) , define
B k ( x ) = e − 2 π i δ k x A k ( x ) B_k(x)=e^{-2\pi i\delta_kx}A_k(x) B k ( x ) = e − 2 π i δ k x A k ( x )
and take its ordinary Fourier coefficients. This produces the unique sequence a ( k ) ∈ ℓ 2 ( Z ) a^{(k)}\in\ell^2(\mathbb{Z}) a ( k ) ∈ ℓ 2 ( Z ) .
The dependence of M ( x ) M(x) M ( x ) on x x x is extremely simple. List the distinct points of
{ 0 , 1 , β 1 , … , β J } \{0,1,\beta_1,\ldots,\beta_J\} { 0 , 1 , β 1 , … , β J }
as
0 = γ 0 < γ 1 < ⋯ < γ R = 1 0=\gamma_0<\gamma_1<\cdots<\gamma_R=1 0 = γ 0 < γ 1 < ⋯ < γ R = 1
and put
I r = ( γ r − 1 , γ r ) , 1 ≤ r ≤ R . I_r=(\gamma_{r-1},\gamma_r),\qquad1\le r\le R. I r = ( γ r − 1 , γ r ) , 1 ≤ r ≤ R .
Each κ j \kappa_j κ j is constant on every I r I_r I r , and hence
M ( x ) = M r , x ∈ I r , (184) M(x)=M_r,\qquad x\in I_r, \tag*{(184)} M ( x ) = M r , x ∈ I r , (184)
for finitely many matrices M r ∈ C J × K M_r\in\mathbb{C}^{J\times K} M r ∈ C J × K .
Lemma 12.2 (Uniform bounds from finitely many matrices). Let
M : L 2 ( 0 , 1 ; C K ) ⟶ L 2 ( 0 , 1 ; C J ) , ( M F ) ( x ) = M ( x ) F ( x ) , \mathcal{M}:L^2(0,1;\mathbb{C}^K)\longrightarrow L^2(0,1;\mathbb{C}^J),\qquad(\mathcal{M}F)(x)=M(x)F(x), M : L 2 ( 0 , 1 ; C K ) ⟶ L 2 ( 0 , 1 ; C J ) , ( M F ) ( x ) = M ( x ) F ( x ) , where M ( x ) = M r M(x)=M_r M ( x ) = M r on the positive-length cells I r I_r I r . Then the optimal squared lower and upper synthesis constants are
A E = min r λ min ( M r ∗ M r ) , B E = max r λ max ( M r ∗ M r ) . (185) A_E=\min_r\lambda_{\min}(M_r^*M_r),\qquad B_E=\max_r\lambda_{\max}(M_r^*M_r). \tag*{(185)} A E = r min λ m i n ( M r ∗ M r ) , B E = r max λ m a x ( M r ∗ M r ) . (185) The operator M \mathcal{M} M is bounded below if and only if every M r M_r M r has full column rank K K K . The optimal squared lower constant for the adjoint is
A f r = min r λ min ( M r M r ∗ ) , (186) A_{\mathrm{fr}}=\min_r\lambda_{\min}(M_rM_r^*), \tag*{(186)} A fr = r min λ m i n ( M r M r ∗ ) , (186) and it is positive if and only if every M r M_r M r has full row rank J J J . The optimal upper constant for the adjoint is again B E B_E B E .
Proof. For F ∈ L 2 ( 0 , 1 ; C K ) F\in L^2(0,1;\mathbb{C}^K) F ∈ L 2 ( 0 , 1 ; C K ) ,
∥ M F ∥ 2 = ∑ r ∫ I r ⟨ M r ∗ M r F ( x ) , F ( x ) ⟩ d x . \lVert\mathcal{M}F\rVert^2=\sum_r\int_{I_r} \langle M_r^*M_rF(x),F(x)\rangle\,\mathrm{d}x. ∥ M F ∥ 2 = r ∑ ∫ I r ⟨ M r ∗ M r F ( x ) , F ( x )⟩ d x .
The finite-dimensional spectral theorem gives the inequalities with the constants in (185). If r 0 r_0 r 0 attains the minimum and u u u is a unit eigenvector of M r 0 ∗ M r 0 M_{r_0}^*M_{r_0} M r 0 ∗ M r 0 corresponding to A E A_E A E , then
F ( x ) = ∣ I r 0 ∣ − 1 / 2 1 I r 0 ( x ) u F(x)=|I_{r_0}|^{-1/2}1_{I_{r_0}}(x)u F ( x ) = ∣ I r 0 ∣ − 1/2 1 I r 0 ( x ) u
has norm one and satisfies ∥ M F ∥ 2 = A E \lVert\mathcal{M}F\rVert^2=A_E ∥ M F ∥ 2 = A E . The analogous construction on a maximizing cell proves optimality of B E B_E B E .
The matrix M r ∗ M r M_r^*M_r M r ∗ M r is positive definite exactly when M r M_r M r has full column rank. Since only finitely many matrices occur, positivity on every cell implies a uniform positive minimum. Applying the same argument to multiplication by M ( x ) ∗ M(x)^* M ( x ) ∗ gives (186) and the row-rank criterion.
Theorem 12.4 (Rank classification for translated unit intervals). The system generated by Λ \Lambda Λ on E E E has the following properties.
(i) It is a Riesz sequence if and only if J ≥ K J\ge K J ≥ K and every positive-cell matrix M r M_r M r has column rank K K K .
(ii) It is a frame for L 2 ( E ) L^2(E) L 2 ( E ) if and only if K ≥ J K\ge J K ≥ J and every positive-cell matrix M r M_r M r has row rank J J J .
(iii) It is a Riesz basis of L 2 ( E ) L^2(E) L 2 ( E ) if and only if J = K J=K J = K and every positive-cell matrix M r M_r M r is invertible.
The optimal squared synthesis bounds are the constants in (185) . In the frame case the optimal lower frame bound is (186) and the optimal upper frame bound is B E B_E B E . In the basis case,
ess sup x ∈ ( 0 , 1 ) ∥ M ( x ) − 1 ∥ = max r ∥ M r − 1 ∥ = A E − 1 / 2 . (187) \operatorname*{ess\,sup}_{x\in(0,1)}\lVert M(x)^{-1}\rVert=\max_r\lVert M_r^{-1}\rVert=A_E^{-1/2}. \tag*{(187)} x ∈ ( 0 , 1 ) ess sup ∥ M ( x ) − 1 ∥ = r max ∥ M r − 1 ∥ = A E − 1/2 . (187)
Proof. Under the unitary coefficient transformation of Theorem 12.1 , synthesis is multiplication by M ( x ) M(x) M ( x ) . A Riesz sequence is equivalent to a positive lower synthesis bound and hence, by Lemma 12.2 , to full column rank on every cell. A frame is equivalent to a positive lower bound for the adjoint, hence to full row rank. A Riesz basis requires both properties and therefore forces J = K J=K J = K and invertibility on every cell. Conversely, in the square case the finitely many inverses M r − 1 M_r^{-1} M r − 1 define a bounded inverse multiplier. Formula (187) is immediate from
∥ M r − 1 ∥ = λ min ( M r ∗ M r ) − 1 / 2 . \lVert M_r^{-1}\rVert=\lambda_{\min}(M_r^*M_r)^{-1/2}. ∥ M r − 1 ∥ = λ m i n ( M r ∗ M r ) − 1/2 .
Proposition 12.5 (Finite-cell reduction and chamber rigidity). For the unit-interval geometry, the matrix field assumes at most J + 1 J+1 J + 1 values. All optimal bounds and all rank criteria reduce to the positive-cell matrices M r M_r M r . Furthermore, if the integer parts q j q_j q j , the shifts δ k \delta_k δ k , and the weak ordering of the fractional parts β j \beta_j β j are fixed, then the complete list of cell matrices, the optimal bounds, and the Riesz-sequence/frame/basis status remain unchanged while the β j \beta_j β j vary without crossing a chamber boundary.
Proof. The vector
( 1 [ 0 , β 1 ) ( x ) , … , 1 [ 0 , β J ) ( x ) ) \left(\mathbf{1}_{[0,\beta_1)}(x),\ldots,\mathbf{1}_{[0,\beta_J)}(x)\right) ( 1 [ 0 , β 1 ) ( x ) , … , 1 [ 0 , β J ) ( x ) )
changes only when x x x crosses one of the fractional parts β j \beta_j β j . Thus the sequence of active subsets and hence the matrices M r M_r M r depend only on the order in which the fractional parts are crossed. The lengths of the cells do not enter the extremal eigenvalues, because the optimal test functions can be localized to any positive-measure cell.
Arithmetic progressions of frequency shifts
Suppose the shifted lattices form an arithmetic progression,
δ k = δ 0 + ( k − 1 ) τ , 1 ≤ k ≤ K . (188) \delta_k = \delta_0 + (k-1)\tau, \qquad1 \le k \le K. \tag*{(188)} δ k = δ 0 + ( k − 1 ) τ , 1 ≤ k ≤ K . (188)
On a fixed cell the integers κ j \kappa_j κ j are constant. Put
z j = e 2 π i τ κ j . z_j = e^{2\pi i\tau\kappa_j}. z j = e 2 π i τ κ j .
Up to unitary diagonal factors, the local matrix becomes
( 1 z 1 ⋯ z 1 K − 1 1 z 2 ⋯ z 2 K − 1 ⋮ ⋮ ⋱ ⋮ 1 z J ⋯ z J K − 1 ) . \begin{pmatrix}
1 & z_1 & \cdots& z_1^{K-1}\\
1 & z_2 & \cdots& z_2^{K-1}\\
\vdots& \vdots& \ddots& \vdots\\
1 & z_J & \cdots& z_J^{K-1}
\end{pmatrix}. 1 1 ⋮ 1 z 1 z 2 ⋮ z J ⋯ ⋯ ⋱ ⋯ z 1 K − 1 z 2 K − 1 ⋮ z J K − 1 .
Corollary 12.6 (Vandermonde basis criterion for unit intervals). Assume J = K J=K J = K and (188). The exponential family is a Riesz basis if and only if, on every positive-measure cell,
τ ( κ j − κ ℓ ) ∉ Z ( j ≠ ℓ ) . (189) \tau(\kappa_j-\kappa_\ell)\notin\mathbb{Z}\qquad(j\ne\ell). \tag*{(189)} τ ( κ j − κ ℓ ) ∈ / Z ( j = ℓ ) . (189)
Proof. The determinant of the local Vandermonde factor is
∏ 1 ≤ j < ℓ ≤ J ( z ℓ − z j ) , \prod_{1\le j<\ell\le J}(z_\ell-z_j), 1 ≤ j < ℓ ≤ J ∏ ( z ℓ − z j ) ,
which is nonzero exactly when the nodes are pairwise distinct. This is equivalent to (189). The result follows from Theorem 12.4 . □
For J = K = 2 J=K=2 J = K = 2 , δ 1 = 0 \delta_1=0 δ 1 = 0 , and δ 2 = δ \delta_2=\delta δ 2 = δ , this construction reproduces the constant matrix of § 11.1 and the two piecewise constant matrices of § 11.2 . Thus the earlier 2 × 2 2\times2 2 × 2 formulas are exactly the two-channel specialization of the general unit-interval theory.
Arbitrary finite unions and variable fiber dimensions
The common unit length is not essential to the folding argument. What changes for unequal lengths is the number of points of E E E lying above a given fractional coordinate. This multiplicity must therefore be built into the target fiber.
The role of integer-translate multiplicity in exponential systems has been studied previously. De Carli et al. [12 ] characterized frame and Riesz-sequence properties of the standard integer exponential system through the multiplicity function
Φ ( x ) = ∑ m ∈ Z 1 E ( x + m ) . \Phi(x)=\sum_{m\in\mathbb{Z}}1_E(x+m). Φ ( x ) = m ∈ Z ∑ 1 E ( x + m ) .
In the present setting, this multiplicity is q ( x ) = ∣ Ω x ∣ q(x)=|\Omega_x| q ( x ) = ∣ Ω x ∣ . For a single frequency lattice, the matrix formulation below reduces to the corresponding scalar fiber description. With several shifted lattices, the multiplicity function alone no longer determines stability: the ranks and singular values of the associated phase matrices must also be taken into account.
Let
E = ⋃ j = 1 J [ a j , b j ) , a j < b j , (190) E=\bigcup_{j=1}^{J}[a_j,b_j),\qquad a_j<b_j, \tag*{(190)} E = j = 1 ⋃ J [ a j , b j ) , a j < b j , (190)
be a finite union of bounded intervals that are pairwise disjoint up to endpoints. Fix K ≥ 1 K\ge1 K ≥ 1 and pairwise incongruent shifts δ 1 , … , δ K \delta_1,\ldots,\delta_K δ 1 , … , δ K modulo Z \mathbb{Z} Z . Let
Λ = ⋃ k = 1 K ( Z + δ k ) , (191) \Lambda=\bigcup_{k=1}^{K}(\mathbb{Z}+\delta_k), \tag*{(191)} Λ = k = 1 ⋃ K ( Z + δ k ) , (191)
and define synthesis by
( S a ) ( y ) = ∑ k = 1 K ∑ n ∈ Z a n ( k ) e 2 π i ( n + δ k ) y . (192) (Sa)(y)=\sum_{k=1}^{K}\sum_{n\in\mathbb{Z}}a_n^{(k)}e^{2\pi i(n+\delta_k)y}. \tag*{(192)} ( S a ) ( y ) = k = 1 ∑ K n ∈ Z ∑ a n ( k ) e 2 π i ( n + δ k ) y . (192)
For x ∈ [ 0 , 1 ) x\in[0,1) x ∈ [ 0 , 1 ) define the integer fiber
Ω x = { m ∈ Z : x + m ∈ E } , q ( x ) = ∣ Ω x ∣ . (193) \Omega_x=\{m\in\mathbb{Z}:x+m\in E\},\qquad q(x)=|\Omega_x|. \tag*{(193)} Ω x = { m ∈ Z : x + m ∈ E } , q ( x ) = ∣ Ω x ∣. (193)
Because E E E is bounded, every Ω x \Omega_x Ω x is contained in one fixed finite subset of Z \mathbb{Z} Z . Partition ( 0 , 1 ) (0,1) ( 0 , 1 ) at the distinct points
{ 0 , 1 } ∪ { { a j } , { b j } : 1 ≤ j ≤ J } , (194) \{0,1\}\cup\bigl\{\{a_j\},\{b_j\}:1\le j\le J\bigr\}, \tag*{(194)} { 0 , 1 } ∪ { { a j } , { b j } : 1 ≤ j ≤ J } , (194)
where { t } = t − ⌊ t ⌋ \{t\}=t-\lfloor t\rfloor { t } = t − ⌊ t ⌋ denotes fractional part. Let the resulting positive-length open cells be I 1 , … , I R I_1,\ldots,I_R I 1 , … , I R . Membership of x + m x+m x + m in E E E can change only at one of the partition points, so
Ω x = Ω r , q ( x ) = q r : = ∣ Ω r ∣ , x ∈ I r . (195) \Omega_x=\Omega_r,\qquad q(x)=q_r:=|\Omega_r|,\qquad x\in I_r. \tag*{(195)} Ω x = Ω r , q ( x ) = q r := ∣ Ω r ∣ , x ∈ I r . (195)
Some q r q_r q r may be zero.
Proposition 12.7. (Folding with variable fiber dimensions). With the convention C 0 = { 0 } \mathbb{C}^{0}=\{0\} C 0 = { 0 } , define
H E = ⨁ r = 1 R L 2 ( I r ; C q r ) , (196) \mathcal{H}_E=\bigoplus_{r=1}^{R}L^2(I_r;\mathbb{C}^{q_r}), \tag*{(196)} H E = r = 1 ⨁ R L 2 ( I r ; C q r ) , (196) and
( W f ) r ( x ) = ( f ( x + m ) ) m ∈ Ω r . (197) (Wf)_r(x)=\bigl(f(x+m)\bigr)_{m\in\Omega_r}. \tag*{(197)} ( W f ) r ( x ) = ( f ( x + m ) ) m ∈ Ω r . (197) Then
W : L 2 ( E ) ⟶ H E W:L^2(E)\longrightarrow\mathcal{H}_E W : L 2 ( E ) ⟶ H E is unitary.
Define
A k ( x ) = e 2 π i δ k x ∑ n ∈ Z a n ( k ) e 2 π i n x , A ( x ) = ( A 1 ( x ) , … , A K ( x ) ) T . (198) A_k(x)=e^{2\pi i\delta_kx}\sum_{n\in\mathbb{Z}}a_n^{(k)}e^{2\pi inx},\qquad A(x)=\bigl(A_1(x),\ldots,A_K(x)\bigr)^T. \tag*{(198)} A k ( x ) = e 2 π i δ k x n ∈ Z ∑ a n ( k ) e 2 π in x , A ( x ) = ( A 1 ( x ) , … , A K ( x ) ) T . (198) The coefficient map is unitary from ⨁ k = 1 K ℓ 2 ( Z ) \bigoplus_{k=1}^{K}\ell^2(\mathbb{Z}) ⨁ k = 1 K ℓ 2 ( Z ) onto L 2 ( 0 , 1 ; C K ) L^2(0,1;\mathbb{C}^K) L 2 ( 0 , 1 ; C K ) . Under these transformations,
( W S a ) r ( x ) = M r A ( x ) , M r = ( e 2 π i δ k m ) m ∈ Ω r 1 ≤ k ≤ K ∈ C q r × K . (199) (WSa)_r(x)=M_rA(x),\qquad M_r=\left(e^{2\pi i\delta_km}\right)_{\substack{m\in\Omega_r\\1\le k\le K}}\in\mathbb{C}^{q_r\times K}. \tag*{(199)} ( W S a ) r ( x ) = M r A ( x ) , M r = ( e 2 π i δ k m ) m ∈ Ω r 1 ≤ k ≤ K ∈ C q r × K . (199)
Consequently,
∥ S a ∥ L 2 ( E ) 2 = ∑ r = 1 R ∫ I r ∥ M r A ( x ) ∥ C q r 2 d x , (200) \|Sa\|_{L^2(E)}^2=\sum_{r=1}^{R}\int_{I_r}\|M_rA(x)\|_{\mathbb{C}^{q_r}}^2\,\mathrm{d}x, \tag*{(200)} ∥ S a ∥ L 2 ( E ) 2 = r = 1 ∑ R ∫ I r ∥ M r A ( x ) ∥ C q r 2 d x , (200)
while
∑ k = 1 K ∥ a ( k ) ∥ ℓ 2 2 = ∫ 0 1 ∥ A ( x ) ∥ C K 2 d x . (201) \sum_{k=1}^{K}\|a^{(k)}\|_{\ell^2}^2=\int_0^1\|A(x)\|_{\mathbb{C}^K}^2\,\mathrm{d}x. \tag*{(201)} k = 1 ∑ K ∥ a ( k ) ∥ ℓ 2 2 = ∫ 0 1 ∥ A ( x ) ∥ C K 2 d x . (201)
Proof. Decompose R \mathbb{R} R into the unit cells [ m , m + 1 ) [m,m+1) [ m , m + 1 ) and use y = x + m y=x+m y = x + m . Then
∫ E ∣ f ( y ) ∣ 2 d y = ∫ 0 1 ∑ m ∈ Ω x ∣ f ( x + m ) ∣ 2 d x = ∑ r = 1 R ∫ I r ∥ ( W f ) r ( x ) ∥ C q r 2 d x . \int_E |f(y)|^2\,\mathrm{d}y=\int_0^1\sum_{m\in\Omega_x}|f(x+m)|^2\,\mathrm{d}x
=\sum_{r=1}^{R}\int_{I_r}\|(Wf)_r(x)\|_{\mathbb{C}^{q_r}}^2\,\mathrm{d}x. ∫ E ∣ f ( y ) ∣ 2 d y = ∫ 0 1 m ∈ Ω x ∑ ∣ f ( x + m ) ∣ 2 d x = r = 1 ∑ R ∫ I r ∥ ( W f ) r ( x ) ∥ C q r 2 d x .
Thus W W W is an isometry. It is onto because a vector field ( F r ) r ∈ H E (F_r)_r\in\mathcal{H}_E ( F r ) r ∈ H E defines a function on E E E by
f ( x + m ) = ( F r ( x ) ) m , x ∈ I r , m ∈ Ω r . f(x+m)=(F_r(x))_m,\qquad x\in I_r,\quad m\in\Omega_r. f ( x + m ) = ( F r ( x ) ) m , x ∈ I r , m ∈ Ω r .
The representation y = x + m y=x+m y = x + m with x ∈ [ 0 , 1 ) x\in[0,1) x ∈ [ 0 , 1 ) is unique outside a null set of endpoints. Parseval shows that
a ( k ) ⟼ ∑ n a n ( k ) e 2 π i n x a^{(k)}\longmapsto\sum_{n}a_n^{(k)}e^{2\pi i n x} a ( k ) ⟼ n ∑ a n ( k ) e 2 π in x
is unitary, and multiplication by e 2 π i δ k x e^{2\pi i\delta_k x} e 2 π i δ k x is unitary, proving the coefficient statement. For a finite coefficient family and m ∈ Z m\in\mathbb{Z} m ∈ Z ,
( S a ) ( x + m ) = ∑ k = 1 K e 2 π i δ k m e 2 π i δ k x ∑ n ∈ Z a n ( k ) e 2 π i n ( x + m ) = ∑ k = 1 K e 2 π i δ k m A k ( x ) , (Sa)(x+m)=\sum_{k=1}^{K}e^{2\pi i\delta_km}e^{2\pi i\delta_kx}\sum_{n\in\mathbb{Z}}a_n^{(k)}e^{2\pi i n(x+m)}
=\sum_{k=1}^{K}e^{2\pi i\delta_km}A_k(x), ( S a ) ( x + m ) = k = 1 ∑ K e 2 π i δ k m e 2 π i δ k x n ∈ Z ∑ a n ( k ) e 2 π in ( x + m ) = k = 1 ∑ K e 2 π i δ k m A k ( x ) ,
using e 2 π i n m = 1 e^{2\pi i n m}=1 e 2 π inm = 1 . This is exactly (199) . Since only finitely many matrices occur, the multiplier is bounded, and density extends the identities to all square-summable coefficient families. □
Theorem 12.8 (Exact rank criteria for arbitrary finite unions). The exponential family associated with (191) on the set E E E in (190) satisfies the following equivalences.
(i) It is a Riesz sequence in L 2 ( E ) L^2(E) L 2 ( E ) if and only if every M r M_r M r has column rank K K K . Necessarily q r ≥ K q_r\ge K q r ≥ K on every cell.
(ii) It is a frame for L 2 ( E ) L^2(E) L 2 ( E ) if and only if every M r M_r M r with q r > 0 q_r>0 q r > 0 has row rank q r q_r q r . Necessarily q r ≤ K q_r\le K q r ≤ K on every nonempty cell.
(iii) It is a Riesz basis of L 2 ( E ) L^2(E) L 2 ( E ) if and only if
and M r M_r M r is invertible for every cell.
The optimal squared synthesis bounds are
A s e q = min r λ min ( M r ∗ M r ) , B = max r λ max ( M r ∗ M r ) , (202) A_{\mathrm{seq}}=\min_{r}\lambda_{\min}(M_r^*M_r),\qquad B=\max_{r}\lambda_{\max}(M_r^*M_r), \tag*{(202)} A seq = r min λ m i n ( M r ∗ M r ) , B = r max λ m a x ( M r ∗ M r ) , (202)
with the convention M r ∗ M r = 0 K × K M_r^*M_r=0_{K\times K} M r ∗ M r = 0 K × K if q r = 0 q_r=0 q r = 0 . In the frame case, the optimal lower frame bound is
A f r = min r : q r > 0 λ min ( M r M r ∗ ) , (203) A_{\mathrm{fr}}=\min_{r:q_r>0}\lambda_{\min}(M_rM_r^*), \tag*{(203)} A fr = r : q r > 0 min λ m i n ( M r M r ∗ ) , (203)
and the optimal upper frame bound is B B B .
Proof. By Proposition 12.7, synthesis acts on I r I_r I r as the finite matrix M r M_r M r , while the adjoint acts as M r ∗ M_r^* M r ∗ . A positive lower synthesis bound is equivalent to positivity of every λ min ( M r ∗ M r ) \lambda_{\min}(M_r^*M_r) λ m i n ( M r ∗ M r ) and therefore to full column rank on every cell. A lower frame bound is equivalent to positivity of λ min ( M r M r ∗ ) \lambda_{\min}(M_rM_r^*) λ m i n ( M r M r ∗ ) on every nonempty target cell and hence to full row rank.
The exact constants follow from the same localization argument as in Lemma 12.2: choose an extremizing cell and an extremal unit eigenvector, then multiply that vector by the normalized indicator of the cell. Because only finitely many cells occur, the minimum of the positive local singular values is positive whenever all required ranks are full.
A Riesz basis requires both the sequence and frame properties. Thus every cell must satisfy simultaneously q r ≥ K q_r\ge K q r ≥ K and q r ≤ K q_r\le K q r ≤ K , so q r = K q_r=K q r = K and M r M_r M r must be invertible. Conversely, if every q r = K q_r=K q r = K and every M r M_r M r is invertible, the cellwise inverse multiplier is bounded because there are only finitely many inverse matrices. □
Integrating gives the necessary measure identity
The converse is false: total measure K K K does not imply constant fiber multiplicity, and constant multiplicity by itself still has to be accompanied by invertibility of the prescribed phase matrices.
Proposition 12.10 (Chamber rigidity for arbitrary finite unions). Fix the integer parts of all endpoints a j , b j a_j,b_j a j , b j , the shifts δ k \delta_k δ k , and the combinatorial order of the fractional endpoint set in (194). As the endpoint fractional parts vary within a chamber without crossing one another or an integer boundary, the ordered list of integer fibers Ω r \Omega_r Ω r remains unchanged. Consequently, the matrices M r M_r M r , all exact bounds in Theorem 12.8, and the Riesz-sequence/frame/basis status remain unchanged; only the lengths of the cells vary.
Proof. For each fixed integer m m m , membership of x + m x+m x + m in any interval [ a j , b j ) [a_j,b_j) [ a j , b j ) can change only when x x x crosses the fractional part of an endpoint. If the integer parts and the order of these crossing events are fixed, the same membership changes occur in the same order. Hence the fiber sets Ω r \Omega_r Ω r and therefore the phase matrices M r M_r M r are unchanged. □
Examples with unequal interval lengths
Example 12.11 (Unequal lengths with an exact Riesz basis). Take
E = [ 0 , 1 / 2 ) ∪ [ 3 / 2 , 3 ) , δ 1 = 0 , δ 2 = 1 / 3. E=[0,1/2)\cup[3/2,3),\qquad\delta_1=0,\qquad\delta_2=1/3. E = [ 0 , 1/2 ) ∪ [ 3/2 , 3 ) , δ 1 = 0 , δ 2 = 1/3. The component lengths are 1 / 2 1/2 1/2 and 3 / 2 3/2 3/2 . The integer fibers are
Ω x = { { 0 , 2 } , 0 < x < 1 / 2 , { 1 , 2 } , 1 / 2 < x < 1. \Omega_x=
\begin{cases}
\{0,2\}, & 0<x<1/2,\\
\{1,2\}, & 1/2<x<1.
\end{cases} Ω x = { { 0 , 2 } , { 1 , 2 } , 0 < x < 1/2 , 1/2 < x < 1. Let ζ = e 2 π i / 3 \zeta=e^{2\pi i/3} ζ = e 2 π i /3 . The two local matrices are
M 1 = ( 1 1 1 ζ 2 ) , M 2 = ( 1 ζ 1 ζ 2 ) . M_1=\begin{pmatrix}1&1\\1&\zeta^2\end{pmatrix},\qquad
M_2=\begin{pmatrix}1&\zeta\\1&\zeta^2\end{pmatrix}. M 1 = ( 1 1 1 ζ 2 ) , M 2 = ( 1 1 ζ ζ 2 ) . Each Gram matrix has diagonal entries 2 and an off-diagonal entry of modulus 1. Hence each has eigenvalues 1 and 3. Therefore
{ e 2 π i n y : n ∈ Z } ∪ { e 2 π i ( n + 1 / 3 ) y : n ∈ Z } \{e^{2\pi iny}:n\in\mathbb{Z}\}\cup\{e^{2\pi i(n+1/3)y}:n\in\mathbb{Z}\} { e 2 π in y : n ∈ Z } ∪ { e 2 π i ( n + 1/3 ) y : n ∈ Z } is a Riesz basis of L 2 ( E ) L^2(E) L 2 ( E ) with exact squared bounds
A = 1 , B = 3. A=1,\qquad B=3. A = 1 , B = 3.
Example 12.12 (Total measure alone is insufficient). Let
E = [ 0 , 1 / 2 ) ∪ [ 1 , 5 / 2 ) . E=[0,1/2)\cup[1,5/2). E = [ 0 , 1/2 ) ∪ [ 1 , 5/2 ) . Again ∣ E ∣ = 2 |E|=2 ∣ E ∣ = 2 , but
q ( x ) = { 3 , 0 < x < 1 / 2 , 1 , 1 / 2 < x < 1. q(x)=
\begin{cases}
3, & 0<x<1/2,\\
1, & 1/2<x<1.
\end{cases} q ( x ) = { 3 , 1 , 0 < x < 1/2 , 1/2 < x < 1. For any choice of two shifted integer lattices, the matrix on the first cell has three rows and two columns and therefore cannot have full row rank; the matrix on the second cell has one row and two columns and therefore cannot have full column rank. The corresponding two-lattice exponential family is consequently neither a frame nor a Riesz sequence in L 2 ( E ) L^2(E) L 2 ( E ) .
Arithmetic-shift criteria in the variable-fiber setting
When the shifts have the arithmetic form (188) , each cell matrix in (199) has rows indexed by integers m ∈ Ω r m\in\Omega_r m ∈ Ω r and columns
e 2 π i ( δ 0 + ( k − 1 ) τ ) m = e 2 π i δ 0 m ( e 2 π i τ m ) k − 1 . e^{2\pi i(\delta_0+(k-1)\tau)m}=e^{2\pi i\delta_0m}\left(e^{2\pi i\tau m}\right)^{k-1}. e 2 π i ( δ 0 + ( k − 1 ) τ ) m = e 2 π i δ 0 m ( e 2 π i τ m ) k − 1 .
Thus, up to a unitary diagonal row factor, M r M_r M r is a rectangular Vandermonde matrix with nodes
z m = e 2 π i τ m , m ∈ Ω r . z_m=e^{2\pi i\tau m},\qquad m\in\Omega_r. z m = e 2 π i τ m , m ∈ Ω r .
In the basis case q r = K q_r=K q r = K , Theorem 12.8 immediately gives the following criterion.
Corollary 12.13 (Vandermonde criterion on arbitrary finite unions). Assume the shifts satisfy (188). The corresponding family is a Riesz basis of L 2 ( E ) L^2(E) L 2 ( E ) if and only if (i) q r = K q_r = K q r = K for every cell, and
(ii) for every cell and every two distinct integers m , m ′ ∈ Ω r m,m' \in\Omega_r m , m ′ ∈ Ω r ,
τ ( m − m ′ ) ∉ Z . (205) \tau(m-m') \notin\mathbb{Z}. \tag*{(205)} τ ( m − m ′ ) ∈ / Z . (205)
Proof. On a cell with q r = K q_r = K q r = K , the determinant is a nonzero unimodular factor times
∏ m , m ′ ∈ Ω r m < m ′ ( e 2 π i τ m ′ − e 2 π i τ m ) . \prod_{\substack{m,m' \in\Omega_r\\m<m'}}\left(e^{2\pi i\tau m'}-e^{2\pi i\tau m}\right). m , m ′ ∈ Ω r m < m ′ ∏ ( e 2 π i τ m ′ − e 2 π i τ m ) .
It is nonzero precisely under (205).
Applications of the finite-cell framework
The finite-cell formulation developed in Section 12 has two further applications. We first use periodicity of spectra for finite unions of intervals to obtain finite Hadamard constraints and arithmetic rigidity on the integer fibers. We then show that rationally commensurate multirate sampling systems reduce exactly to finitely many common-lattice channels governed by the same phase matrices.
Periodic spectra, Hadamard fibers, and spectral rigidity
We next connect the variable-fiber representation of Section 12 with spectral sets. Let Ω ⊂ R \Omega\subset\mathbb{R} Ω ⊂ R be a bounded measurable set of measure one. Recall that Λ ⊂ R \Lambda\subset\mathbb{R} Λ ⊂ R is a spectrum for Ω \Omega Ω if the exponentials
{ e 2 π i λ x : λ ∈ Λ } \{e^{2\pi i\lambda x}:\lambda\in\Lambda\} { e 2 π iλ x : λ ∈ Λ }
form an orthogonal basis of L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) . For finite unions of intervals, every spectrum is periodic Bose and Madan [13 ] . Recent work has established the spectral-implies-tiling direction of Fuglede’s conjecture for unions of three intervals Shi [14 ] , while the corresponding general problem for four or more intervals remains a natural next case.
Assume that a spectrum of Ω \Omega Ω has period d ∈ N d \in\mathbb{N} d ∈ N and is written as
Λ = ⋃ k = 1 d ( d Z + λ k ) , λ k ≢ λ ℓ ( m o d d ) ( k ≠ ℓ ) . (206) \Lambda= \bigcup_{k=1}^{d}(d\mathbb{Z}+\lambda_k), \qquad\lambda_k \not\equiv\lambda_\ell\pmod d \quad(k\ne\ell). \tag*{(206)} Λ = k = 1 ⋃ d ( d Z + λ k ) , λ k ≡ λ ℓ ( mod d ) ( k = ℓ ) . (206)
For a normalized spectral set, the number of representatives in a period equals d d d by the density theorem. Define
E = d Ω , δ k = λ k / d . E=d\Omega,\qquad\delta_k=\lambda_k/d. E = d Ω , δ k = λ k / d .
Then
Λ / d = ⋃ k = 1 d ( Z + δ k ) . \Lambda/d=\bigcup_{k=1}^{d}(\mathbb{Z}+\delta_k). Λ/ d = k = 1 ⋃ d ( Z + δ k ) .
The fact that a Riesz basis formed by finitely many translates of a lattice forces a corresponding multi-tiling condition is known; see Agora et al. [9 ] . In the orthonormal spectral setting considered here, the finite-cell representation additionally forces each local square phase matrix to be a scaled unitary matrix. We use these Hadamard identities to investigate the arithmetic constraints generated by transitions between adjacent fibers.
Proposition 13.1 (Finite Hadamard fibers associated with a periodic spectrum). Let Ω \Omega Ω be a finite union of bounded intervals of measure one and let (206) be a spectrum. For E = d Ω E=d\Omega E = d Ω , let
Ω x E = { m ∈ Z : x + m ∈ E } , 0 ≤ x < 1 , \Omega_x^E=\{m\in\mathbb{Z}:x+m\in E\},\qquad0\leq x<1, Ω x E = { m ∈ Z : x + m ∈ E } , 0 ≤ x < 1 , and let Ω r E \Omega_r^E Ω r E be its finitely many cell values as in Proposition 12.7 . Then
∣ Ω r E ∣ = d (207) \lvert\Omega_r^E\rvert=d \tag*{(207)} ∣ Ω r E ∣ = d (207) for every positive-measure cell, and the phase matrix
M r = ( e 2 π i δ k m ) m ∈ Ω r E 1 ≤ k ≤ d (208) M_r=\left(e^{2\pi i\delta_km}\right)_{\substack{m\in\Omega_r^E\\1\leq k\leq d}} \tag*{(208)} M r = ( e 2 π i δ k m ) m ∈ Ω r E 1 ≤ k ≤ d (208) satisfies
M r ∗ M r = d I d . (209) M_r^*M_r=dI_d. \tag*{(209)} M r ∗ M r = d I d . (209) In particular, d − 1 / 2 M r d^{-1/2}M_r d − 1/2 M r is unitary for every cell.
Proof. Under the change of variables y = d x y=dx y = d x , the spectral family on Ω \Omega Ω becomes, up to the common spatial scaling, the family with frequency set Λ / d \Lambda/d Λ/ d on E = d Ω E=d\Omega E = d Ω . Since the original exponentials are orthogonal and ∣ E ∣ = d \lvert E\rvert=d ∣ E ∣ = d , the scaled synthesis operator satisfies
∥ S a ∥ L 2 ( E ) 2 = d ∑ k = 1 d ∥ a ( k ) ∥ 2 2 . \lVert Sa\rVert_{L^2(E)}^2=d\sum_{k=1}^{d}\lVert a^{(k)}\rVert_2^2. ∥ S a ∥ L 2 ( E ) 2 = d k = 1 ∑ d ∥ a ( k ) ∥ 2 2 .
Apply Proposition 12.7 . The coefficient map onto L 2 ( 0 , 1 ; C d ) L^2(0,1;\mathbb{C}^d) L 2 ( 0 , 1 ; C d ) is unitary, while on a cell I r I_r I r synthesis is multiplication by M r M_r M r . Localizing an arbitrary vector function to I r I_r I r gives
∥ M r z ∥ C ∣ Ω r E ∣ 2 = d ∥ z ∥ C d 2 ( z ∈ C d ) . \lVert M_r z\rVert_{\mathbb{C}^{\lvert\Omega_r^E\rvert}}^2=d\lVert z\rVert_{\mathbb{C}^d}^2\qquad(z\in\mathbb{C}^d). ∥ M r z ∥ C ∣ Ω r E ∣ 2 = d ∥ z ∥ C d 2 ( z ∈ C d ) .
Hence M r ∗ M r = d I d M_r^*M_r=dI_d M r ∗ M r = d I d , so M r M_r M r has rank d d d and ∣ Ω r E ∣ ≥ d \lvert\Omega_r^E\rvert\geq d ∣ Ω r E ∣ ≥ d .
Since
∫ 0 1 ∣ Ω x E ∣ d x = ∣ E ∣ = d , \int_0^1\lvert\Omega_x^E\rvert\,dx=\lvert E\rvert=d, ∫ 0 1 ∣ Ω x E ∣ d x = ∣ E ∣ = d ,
and ∣ Ω x E ∣ ≥ d \lvert\Omega_x^E\rvert\geq d ∣ Ω x E ∣ ≥ d almost everywhere, equality must hold almost everywhere. This proves (207) and (209). □
The preceding proposition has an elementary but useful geometric consequence. Let c ∈ [ 0 , 1 ) c\in[0,1) c ∈ [ 0 , 1 ) be a fractional endpoint of E E E . As x x x crosses c c c , the cardinality of Ω x E \Omega_x^E Ω x E changes by the number of interval left endpoints having fractional part c c c minus the number of right endpoints having fractional part c c c . Since the cardinality is constant, these multiplicities must balance.
Lemma 13.2 (Endpoint balance). Under the assumptions of Proposition 13.1 , for every c ∈ [ 0 , 1 ) c\in[0,1) c ∈ [ 0 , 1 ) the number of left endpoints of E E E with fractional part c c c equals the number of right endpoints of E E E with fractional part c c c , counted with multiplicity.
Proof. The jump of x ↦ ∣ Ω x E ∣ x\mapsto\lvert\Omega_x^E\rvert x ↦ ∣ Ω x E ∣ at c c c is exactly the difference between these two multiplicities. Proposition 13.1 shows that this function equals d d d almost everywhere, hence every jump is zero. □
Thus adjacent fibers have equal cardinality and differ by exchanges. The next lemma is the basic rigidity mechanism.
Lemma 13.3 (Single-row exchange rigidity). Let Γ , Γ ′ ⊂ Z \Gamma,\Gamma' \subset\mathbb{Z} Γ , Γ ′ ⊂ Z have cardinality d d d and suppose
Γ ′ = ( Γ ∖ { a } ) ∪ { b } , a ≠ b . \Gamma' = (\Gamma\setminus\{a\}) \cup\{b\}, \qquad a \ne b. Γ ′ = ( Γ ∖ { a }) ∪ { b } , a = b . Let D = { δ 1 , … , δ d } ⊂ R / Z D = \{\delta_1,\ldots,\delta_d\} \subset\mathbb{R}/\mathbb{Z} D = { δ 1 , … , δ d } ⊂ R / Z , and assume that both phase matrices
M ( Γ , D ) = ( e 2 π i δ k m ) m ∈ Γ , k , M ( Γ ′ , D ) = ( e 2 π i δ k m ) m ∈ Γ ′ , k M(\Gamma,D) = \left(e^{2\pi i\delta_km}\right)_{m\in\Gamma,k}, \qquad M(\Gamma',D) = \left(e^{2\pi i\delta_km}\right)_{m\in\Gamma',k} M ( Γ , D ) = ( e 2 π i δ k m ) m ∈ Γ , k , M ( Γ ′ , D ) = ( e 2 π i δ k m ) m ∈ Γ ′ , k satisfy M ∗ M = d I d M^*M=dI_d M ∗ M = d I d . Then
( δ j − δ k ) ( b − a ) ∈ Z ( 1 ≤ j , k ≤ d ) . (210) (\delta_j-\delta_k)(b-a)\in\mathbb{Z} \qquad(1\le j,k\le d). \tag*{(210)} ( δ j − δ k ) ( b − a ) ∈ Z ( 1 ≤ j , k ≤ d ) . (210) Proof. For j ≠ k j\ne k j = k , orthogonality of columns gives
∑ m ∈ Γ e 2 π i ( δ j − δ k ) m = 0 , ∑ m ∈ Γ ′ e 2 π i ( δ j − δ k ) m = 0. \sum_{m\in\Gamma}e^{2\pi i(\delta_j-\delta_k)m}=0, \qquad\sum_{m\in\Gamma'}e^{2\pi i(\delta_j-\delta_k)m}=0. m ∈ Γ ∑ e 2 π i ( δ j − δ k ) m = 0 , m ∈ Γ ′ ∑ e 2 π i ( δ j − δ k ) m = 0. Subtracting the two equalities yields
e 2 π i ( δ j − δ k ) b = e 2 π i ( δ j − δ k ) a , e^{2\pi i(\delta_j-\delta_k)b}=e^{2\pi i(\delta_j-\delta_k)a}, e 2 π i ( δ j − δ k ) b = e 2 π i ( δ j − δ k ) a , which is equivalent to (210). The case j = k j=k j = k is trivial.
More generally, suppose an adjacent-cell transition removes A = { a 1 , … , a r } A=\{a_1,\ldots,a_r\} A = { a 1 , … , a r } and inserts B = { b 1 , … , b r } B=\{b_1,\ldots,b_r\} B = { b 1 , … , b r } . The same subtraction gives the following exact identity.
Lemma 13.4 (Rank-r r r exchange identity). Under the same Hadamard hypotheses, if
Γ ′ = ( Γ ∖ A ) ∪ B , ∣ A ∣ = ∣ B ∣ = r , \Gamma'=(\Gamma\setminus A)\cup B,\qquad|A|=|B|=r, Γ ′ = ( Γ ∖ A ) ∪ B , ∣ A ∣ = ∣ B ∣ = r , then, for every j ≠ k j\ne k j = k ,
∑ ν = 1 r e 2 π i ( δ j − δ k ) a ν = ∑ ν = 1 r e 2 π i ( δ j − δ k ) b ν . (211) \sum_{\nu=1}^{r}e^{2\pi i(\delta_j-\delta_k)a_\nu}
=
\sum_{\nu=1}^{r}e^{2\pi i(\delta_j-\delta_k)b_\nu}. \tag*{(211)} ν = 1 ∑ r e 2 π i ( δ j − δ k ) a ν = ν = 1 ∑ r e 2 π i ( δ j − δ k ) b ν . (211)
For two-row exchanges one obtains a useful dichotomy.
Lemma 13.5 (Two-row exchange dichotomy). Assume r = 2 r=2 r = 2 in Lemma 13.4 and fix θ = δ j − δ k \theta=\delta_j-\delta_k θ = δ j − δ k . Then either
{ e 2 π i θ a 1 , e 2 π i θ a 2 } = { e 2 π i θ b 1 , e 2 π i θ b 2 } , (212) \{e^{2\pi i\theta a_1},e^{2\pi i\theta a_2}\}
=
\{e^{2\pi i\theta b_1},e^{2\pi i\theta b_2}\}, \tag*{(212)} { e 2 π i θ a 1 , e 2 π i θ a 2 } = { e 2 π i θ b 1 , e 2 π i θ b 2 } , (212) or both pairs are antipodal:
e 2 π i θ ( a 1 − a 2 ) = − 1 , e 2 π i θ ( b 1 − b 2 ) = − 1. (213) e^{2\pi i\theta(a_1-a_2)}=-1,\qquad e^{2\pi i\theta(b_1-b_2)}=-1. \tag*{(213)} e 2 π i θ ( a 1 − a 2 ) = − 1 , e 2 π i θ ( b 1 − b 2 ) = − 1. (213) Proof. Put z ν = e 2 π i θ a ν z_\nu=e^{2\pi i\theta a_\nu} z ν = e 2 π i θ a ν and w ν = e 2 π i θ b ν w_\nu=e^{2\pi i\theta b_\nu} w ν = e 2 π i θ b ν . Equation (211) says z 1 + z 2 = w 1 + w 2 = : s z_1+z_2=w_1+w_2=:s z 1 + z 2 = w 1 + w 2 =: s . If s = 0 s=0 s = 0 , each pair consists of antipodal points and (213) follows. If s ≠ 0 s\ne0 s = 0 , then for any two unit-modulus numbers with sum s s s ,
z 1 z 2 = s s ‾ . z_1z_2=\frac{s}{\overline{s}}. z 1 z 2 = s s . Thus z 1 , z 2 z_1,z_2 z 1 , z 2 and w 1 , w 2 w_1,w_2 w 1 , w 2 are the roots of the same quadratic polynomial
t 2 − s t + s s ‾ , t^2-st+\frac{s}{\overline{s}}, t 2 − s t + s s , which proves (212).
The single-exchange lemma already yields a conditional tiling criterion. Consider the cyclic sequence of positive-measure cells and assume every transition is a single-row exchange. Write the corresponding integer displacements as
q 1 , … , q R , q r = b r − a r , q_1,\ldots,q_R,\qquad q_r=b_r-a_r, q 1 , … , q R , q r = b r − a r ,
and put
g = gcd ( ∣ q 1 ∣ , … , ∣ q R ∣ ) . (214) g=\gcd(\lvert q_1\rvert,\ldots,\lvert q_R\rvert). \tag*{(214)} g = g cd(∣ q 1 ∣ , … , ∣ q R ∣) . (214)
Theorem 13.6 (A fiber-transition criterion forcing lattice spectrality and tiling). Under the assumptions of Proposition 13.1, suppose every adjacent fiber transition is a single-row exchange. Then
D − D ⊂ 1 g Z / Z , D = { δ 1 , … , δ d } . (215) D-D\subset\frac{1}{g}\mathbb{Z}/\mathbb{Z},\qquad D=\{\delta_1,\ldots,\delta_d\}. \tag*{(215)} D − D ⊂ g 1 Z / Z , D = { δ 1 , … , δ d } . (215) In particular g ≥ d g\ge d g ≥ d . If g = d g=d g = d , then D D D is a translate of the full subgroup 1 d Z / Z \frac{1}{d}\mathbb{Z}/\mathbb{Z} d 1 Z / Z , the original spectrum Λ \Lambda Λ is a translate of Z \mathbb{Z} Z , and Ω \Omega Ω tiles R \mathbb{R} R by translations of Z \mathbb{Z} Z .
Proof. Lemma 13.3 gives
( δ j − δ k ) q r ∈ Z (\delta_j-\delta_k)q_r\in\mathbb{Z} ( δ j − δ k ) q r ∈ Z
for every r r r , j j j , k k k . Bezout’s identity applied to the integers q r q_r q r gives
g ( δ j − δ k ) ∈ Z , g(\delta_j-\delta_k)\in\mathbb{Z}, g ( δ j − δ k ) ∈ Z ,
which proves (215) . Hence, after subtracting one element of D D D , the set D D D is contained in the subgroup 1 g Z / Z \frac{1}{g}\mathbb{Z}/\mathbb{Z} g 1 Z / Z , which has g g g elements. Since D D D contains d d d distinct elements, g ≥ d g\ge d g ≥ d .
If g = d g=d g = d , cardinality forces
D = δ 0 + 1 d { 0 , 1 , … , d − 1 } ( m o d 1 ) . D=\delta_0+\frac{1}{d}\{0,1,\ldots,d-1\}\pmod1. D = δ 0 + d 1 { 0 , 1 , … , d − 1 } ( mod 1 ) .
Therefore
Λ / d = D + Z = δ 0 + 1 d Z , \Lambda/d=D+\mathbb{Z}=\delta_0+\frac{1}{d}\mathbb{Z}, Λ/ d = D + Z = δ 0 + d 1 Z ,
and hence
Λ = d δ 0 + Z . \Lambda=d\delta_0+\mathbb{Z}. Λ = d δ 0 + Z .
Modulation by e 2 π i d δ 0 x e^{2\pi i d\delta_0x} e 2 π i d δ 0 x shows that Z \mathbb{Z} Z itself is a spectrum for Ω \Omega Ω . Let
P ( x ) = ∑ n ∈ Z 1 Ω ( x + n ) . P(x)=\sum_{n\in\mathbb{Z}}\mathbf{1}_{\Omega}(x+n). P ( x ) = n ∈ Z ∑ 1 Ω ( x + n ) .
The nonzero Fourier coefficients of the one-periodic function P P P are
P ^ ( k ) = ∫ Ω e − 2 π i k x d x = 0 , k ∈ Z ∖ { 0 } , \widehat{P}(k)=\int_{\Omega}e^{-2\pi ikx}\,\mathrm{d}x=0,\qquad k\in\mathbb{Z}\setminus\{0\}, P ( k ) = ∫ Ω e − 2 π ik x d x = 0 , k ∈ Z ∖ { 0 } ,
by orthogonality, while P ^ ( 0 ) = ∣ Ω ∣ = 1 \widehat{P}(0)=\lvert\Omega\rvert=1 P ( 0 ) = ∣ Ω ∣ = 1 . Hence P = 1 P=1 P = 1 almost everywhere, so Ω \Omega Ω tiles R \mathbb{R} R by Z \mathbb{Z} Z . ∎
For a union of four intervals, Proposition 13.1 and Lemma 13.2 reduce the geometry to finitely many balanced endpoint transitions. Transitions of multiplicity one are completely controlled by Lemma 13.3, and transitions of multiplicity two satisfy Lemma 13.5. The remaining obstruction is therefore finite and algebraic: one must classify the simultaneous identities (211) for the finitely many differences D − D D-D D − D , including the higher-multiplicity transitions that can occur when several endpoints have the same fractional part. This provides a possible finite-matrix route to the four-interval case of the one-dimensional Fuglede problem, but no general classification is asserted here.
Commensurate multirate sampling
We finally show that the variable-fiber theorem applies without change to finite unions of sampling lattices having different but rationally commensurate densities. This gives a partial framework for the multirate reconstruction problem discussed, for example, by Frederick and Yacoubou Djima Frederick and Djima [15 ] , who ask for extensions to more than two sampling sets of different densities.
Let
Λ = ⋃ j = 1 J ( α j Z + β j ) , α j > 0 , (216) \Lambda= \bigcup_{j=1}^{J}(\alpha_j\mathbb{Z}+\beta_j), \qquad\alpha_j>0, \tag*{(216)} Λ = j = 1 ⋃ J ( α j Z + β j ) , α j > 0 , (216)
and assume
α j α ℓ ∈ Q ( 1 ≤ j , ℓ ≤ J ) . (217) \frac{\alpha_j}{\alpha_\ell}\in\mathbb{Q}\qquad(1\le j,\ell\le J). \tag*{(217)} α ℓ α j ∈ Q ( 1 ≤ j , ℓ ≤ J ) . (217)
Choose L > 0 L>0 L > 0 such that
q j : = L α j ∈ N ( 1 ≤ j ≤ J ) . (218) q_j:=\frac{L}{\alpha_j}\in\mathbb{N}\qquad(1\le j\le J). \tag*{(218)} q j := α j L ∈ N ( 1 ≤ j ≤ J ) . (218)
Then
α j Z + β j = ⋃ r = 0 q j − 1 ( L Z + β j + r α j ) . (219) \alpha_j\mathbb{Z}+\beta_j=\bigcup_{r=0}^{q_j-1}\left(L\mathbb{Z}+\beta_j+r\alpha_j\right). \tag*{(219)} α j Z + β j = r = 0 ⋃ q j − 1 ( L Z + β j + r α j ) . (219)
After merging repetitions modulo L L L , let
γ 1 , … , γ K \gamma_1,\ldots,\gamma_K γ 1 , … , γ K
be the distinct coset representatives and set
δ k = γ k / L ( m o d 1 ) . (220) \delta_k=\gamma_k/L\pmod1. \tag*{(220)} δ k = γ k / L ( mod 1 ) . (220)
Theorem 13.7 (Exact reduction of commensurate multirate lattices). Let E ⊂ R E\subset\mathbb{R} E ⊂ R be a finite union of bounded intervals and let Λ \Lambda Λ be given by (216)–(217). Put
E ~ = L E . \widetilde{E}=LE. E = L E . Then the exponential system indexed by Λ \Lambda Λ in L 2 ( E ) L^2(E) L 2 ( E ) has the same Riesz-sequence, frame, and Riesz-basis properties as the system indexed by
Λ ~ = ⋃ k = 1 K ( Z + δ k ) (221) \widetilde{\Lambda}=\bigcup_{k=1}^{K}(\mathbb{Z}+\delta_k) \tag*{(221)} Λ = k = 1 ⋃ K ( Z + δ k ) (221)
in L 2 ( E ~ ) L^2(\widetilde{E}) L 2 ( E ) .
Let Ω ~ r \widetilde{\Omega}_r Ω r be the finitely many integer fibers of E ~ \widetilde{E} E and define
M r = ( e 2 π i δ k m ) m ∈ Ω ~ r 1 ≤ k ≤ K . (222) M_r=\left(e^{2\pi i\delta_k m}\right)_{\substack{m\in\widetilde{\Omega}_r\\1\leq k\leq K}}. \tag*{(222)} M r = ( e 2 π i δ k m ) m ∈ Ω r 1 ≤ k ≤ K . (222)
Then:
(i) the original system is a Riesz sequence if and only if every M r M_r M r has column rank K K K ;
(ii) it is a frame if and only if every nonempty M r M_r M r has full row rank;
(iii) it is a Riesz basis if and only if ∣ Ω ~ r ∣ = K |\widetilde{\Omega}_r|=K ∣ Ω r ∣ = K and M r M_r M r is invertible on every cell.
If A ~ \widetilde{A} A and B ~ \widetilde{B} B denote the optimal squared bounds computed from the singular values of the matrices M r M_r M r as in Theorem 12.8 , then the corresponding squared synthesis bounds on the original domain E E E are
A = A ~ L , B = B ~ L . (223) A=\frac{\widetilde{A}}{L},\qquad B=\frac{\widetilde{B}}{L}. \tag*{(223)} A = L A , B = L B . (223)
Proof. Equation (219) rewrites Λ \Lambda Λ as a finite union of cosets of the common lattice L Z L\mathbb{Z} L Z . After repetitions are removed,
Λ = ⋃ k = 1 K ( L Z + γ k ) . \Lambda=\bigcup_{k=1}^{K}(L\mathbb{Z}+\gamma_k). Λ = k = 1 ⋃ K ( L Z + γ k ) .
With y = L x y=Lx y = Lx ,
e 2 π i ( L n + γ k ) x = e 2 π i ( n + δ k ) y . e^{2\pi i(Ln+\gamma_k)x}=e^{2\pi i(n+\delta_k)y}. e 2 π i ( L n + γ k ) x = e 2 π i ( n + δ k ) y .
Moreover,
∫ E ∣ f ( x ) ∣ 2 d x = 1 L ∫ E ~ ∣ f ( y / L ) ∣ 2 d y . \int_E |f(x)|^2\,dx=\frac{1}{L}\int_{\widetilde{E}}|f(y/L)|^2\,dy. ∫ E ∣ f ( x ) ∣ 2 d x = L 1 ∫ E ∣ f ( y / L ) ∣ 2 d y .
Thus the synthesis inequality on E E E is exactly the synthesis inequality for (221) on E ~ \widetilde{E} E , multiplied by 1 / L 1/L 1/ L . The rank criteria and the optimal scaled bounds now follow directly from Theorem 12.8 . □
The Fourier-dual relation between exponential frames and stable sampling is classical; see, for example, Ortega-Cerdà and Seip [16 ] ; Aldroubi and Gröchenig [17 ] . The reconstruction formula below concerns the inverse of the exponential synthesis representation. Its interpretation as a sampling formula for bandlimited signals requires the corresponding Fourier-dual identification. The reduction also gives an explicit reconstruction map whenever the system is a Riesz basis.
Proposition 13.8 (Fiberwise reconstruction in the commensurate multirate case). Assume the hypotheses of Theorem 13.7 and suppose the equivalent Riesz-basis conditions hold. Let f ∈ L 2 ( E ) f\in L^2(E) f ∈ L 2 ( E ) and put
f ~ ( y ) = f ( y / L ) , y ∈ E ~ . \widetilde{f}(y)=f(y/L),\qquad y\in\widetilde{E}. f ( y ) = f ( y / L ) , y ∈ E . On a cell I r I_r I r , define the folded data vector
F r ( x ) = ( f ~ ( x + m ) ) m ∈ Ω ~ r . F_r(x)=\left(\widetilde{f}(x+m)\right)_{m\in\widetilde{\Omega}_r}. F r ( x ) = ( f ( x + m ) ) m ∈ Ω r . Then the channel functions are recovered pointwise by
A ( x ) = M r − 1 F r ( x ) , x ∈ I r . (224) A(x)=M_r^{-1}F_r(x),\qquad x\in I_r. \tag*{(224)} A ( x ) = M r − 1 F r ( x ) , x ∈ I r . (224)
Writing
B k ( x ) = e − 2 π i δ k x A k ( x ) , B_k(x)=e^{-2\pi i\delta_k x}A_k(x), B k ( x ) = e − 2 π i δ k x A k ( x ) ,
the coefficient sequence in the k k k -th common-lattice channel is the ordinary Fourier coefficient sequence
a n ( k ) = ∫ 0 1 B k ( x ) e − 2 π i n x d x . (225) a_n^{(k)}=\int_0^1 B_k(x)e^{-2\pi i n x}\,dx. \tag*{(225)} a n ( k ) = ∫ 0 1 B k ( x ) e − 2 π in x d x . (225)
After identifying each common-lattice channel with the corresponding coset in (219) , these coefficients give the unique expansion of f f f in the original multirate exponential basis.
Proof. This is the inverse of the fiberization in Proposition 12.7 . In the Riesz-basis case every M r M_r M r is square and invertible, giving (224) . The modulation used in (198) is then removed, and ordinary Fourier inversion on [ 0 , 1 ) [0,1) [ 0 , 1 ) gives (225) .
Theorem 13.7 does not address incommensurate lattice steps or genuinely non-lattice sampling sets. It does, however, show that any finite collection of rationally commensurate sampling densities is already contained in the finite-cell matrix-symbol framework developed above. In this regime the existence question, the optimal condition number, and the reconstruction map all reduce to finitely many explicit matrices.
Discussion and open problems
The results above identify two regimes in which the infinite-dimensional structure can be reduced exactly to finite-dimensional data. For periodic local perturbations, the synthesis operator is governed by a single constant matrix. For finite unions of intervals and finitely many shifted lattices, folding modulo one produces finitely many phase matrices, possibly of different row dimensions. Outside these structured regimes, the relevant invariants remain genuinely infinite-dimensional.
Several problems therefore remain open. For a general nonperiodic binary pattern, it would be useful to characterize geometrically the exact invariant
ρ δ , s = ∥ P − T 2 δ P + ∥ . \rho_{\delta,s}=\lVert P_-T_{2\delta}P_+\rVert. ρ δ , s = ∥ P − T 2 δ P + ∥ .
At the critical value δ = 1 / 4 \delta=1/4 δ = 1/4 , the periodic approximation theorem reduces the worst-case lower bound to a sequence of finite extremal problems, but both the balanced two-block extremizer conjecture and the asymptotic decay of A p ( 1 / 4 ) A_p(1/4) A p ( 1/4 ) remain open. A related question is whether the half-shifted Hilbert model can be compared quantitatively with the periodized cross blocks.
Although support-independent operator-norm perturbation estimates are available, a positive near-critical stability criterion for arbitrary nonperiodic sign configurations remains open. In the finite-alphabet setting, the norm-matrix criterion is only sufficient in the nonperiodic case, and an exact spectral classification remains open.
On disconnected domains, the finite-cell theory suggests two natural extensions. The first is a classification of the higher-multiplicity endpoint changes that arise in the four-interval spectral problem. The second is the extension of the commensurate multirate reduction to incommensurate lattice steps and genuinely non-lattice sampling sets.
Conclusion
The main theme of the paper is that shifted-sinc operators provide a common language for several exponential-system problems, while the relevant finite-dimensional reductions depend strongly on the underlying structure. For a fixed binary local perturbation, the exact squared synthesis bounds are determined by the single compression norm
ρ δ , s = ∥ P − T 2 δ P + ∥ . \rho_{\delta,s} = \lVert P_{-}T_{2\delta}P_{+}\rVert. ρ δ , s = ∥ P − T 2 δ P + ∥ .
Periodicity turns this infinite-dimensional invariant into a constant finite matrix problem, and the worst-case binary bound can in turn be approximated by periodic patterns.
For disconnected spatial domains, the corresponding reduction is geometric rather than pattern-based. Folding modulo one yields finitely many phase matrices indexed by the integer fibers of the domain. Their singular values and ranks provide exact quantitative bounds and complete Riesz-sequence, frame, and basis criteria, including the case of unequal interval lengths.
These two reductions also lead to applications beyond the original models: periodic spectra produce finite Hadamard fibers and arithmetic rigidity, while rationally commensurate multirate systems reduce to common-lattice channels. The remaining difficulties are concentrated in the genuinely nonperiodic, incommensurate, or asymptotic regimes, where no single finite matrix captures the entire problem.
Acknowledgments
P. Vellucci is a member of the INdAM Research Group GNCS.