Floor-Quotient Operator Inverses and Exact Primal-Dual Equivalence in the Shifted Nyman-Beurling Criterion: All-Scale Universal Lower Bounds, Geometric Transference, and Two-Sided Interval-Certified Subspace Optimality

Abstract

Building on the finite-support annihilator framework (riemann_hypothesis_7, 9, 12, 14), Balazard's Möbius inverse, and projection duality, we prove three unconditional theorems for $\mathcal{A}_{N, M} = \operatorname{span}\{\mathcal{F}_N, \mathcal{C}_{N+1}, \dots, \mathcal{C}_M\}$ ($M > N, \sigma \in [1/2, 1)$), which do not resolve RH. First (Theorem 1), $C_M$ has exact inverse $(C_M^{-1})_{r, m} = \lfloor m/r \rfloor$, yielding $(G_M^{\mathcal{C}, (\sigma)})^{-1} = Q_M^{(\sigma)}$ via floor quotients, giving exact minimum norm $\mathcal{E}_{\min} = \mathbf{h}_N^T (Q_{N, M}^{(\sigma)})^{-1} \mathbf{h}_N$ and extending the universal lower bound to all $\delta > 0$: $\|\mathcal{L}\|_{S_\sigma^*}^2 \ge N^{-2}/[\sum_{k=1}^K (2k-1)(k N)^{-2\sigma} - K^2(N_\delta+1)^{-2\sigma}]$ ($K = \lfloor 1+\delta \rfloor$). Second (Theorem 2), we establish sufficient boundary conditions defining the admissible band set $\mathcal{J}_{\mathrm{adm}}(N, M)$, formulate the combined multi-band bound $\mathcal{E}_{\mathrm{MB}} := \max\{\mathcal{E}_{\mathrm{single}}, \sum \mathcal{E}_{\mathrm{DZ}} + \sum |S_j|^2/W_j\}$ valid across all $M > N$ (coinciding with $\mathcal{E}_{\mathrm{single}}$ for $M \ge 2N$), and identify the exact dipole-matching mechanism yielding equality $\mathcal{E}_{\mathrm{MB}+\mathrm{ladder}} = \mathcal{E}_{\mathrm{MB}}$ at $\sigma = 1/2$ for odd $M$. Third (Theorem 3), we prove exact zero-gap primal-dual equivalence $d_{\mathcal{A}_{N, M}}(\sigma)^2 = \min_{\mathbf{b}} \sum_{m=1}^M w_m^{(\sigma)}|1 + \sum_{k=2}^N b_k\{m/k\}|^2$ and establish the Geometric Transference Theorem: $d_{\mathcal{A}_{N, M}}(\sigma)^2 = \frac{|\mathcal{L}_{\min}(E)|^2}{\mathcal{E}_{\min}(N, M, \sigma)\cos^2\theta_{N, M}(\sigma)} \ge \frac{|\mathcal{L}_{\min}(E)|^2}{\mathcal{E}_{\mathrm{MB}+\mathrm{ladder}}(N, M, \sigma)}$, linking unnormalized norm bounds directly to the normalized Nyman–Beurling approximation distance. We certify two-sided 60-digit interval sandwiches (gap $\le 2.19 \times 10^{-21}$) and certified wavelet capture percentages across 18 benchmark configurations ($N \in \{40, 80, 160\}$).

---

1. Introduction, Shifted Nyman-Beurling Duality, and Summary of New Contributions

1.1 Inherited Shifted Nyman-Beurling Step-Space Framework

Let $\sigma \in [1/2, 1)$, and let $\mathcal{H}_\sigma := L^2((0, 1], d\mu_\sigma)$ denote the Hilbert space equipped with probability measure $d\mu_\sigma(x) := 2\sigma x^{2\sigma-1}\,dx$, inner product $\langle f, g \rangle_\sigma := 2\sigma \int_0^1 f(x)\overline{g(x)}\,x^{2\sigma-1}\,dx$, and norm $\|f\|_\sigma = \langle f, f \rangle_\sigma^{1/2}$ (at $\sigma = 1/2$, $2\sigma = 1$). Following Nyman [1], Beurling [2], and Báez-Duarte [3], the Riemann Hypothesis is equivalent to the target-approximation criterion $E(x) := \mathbf{1}_{(0, 1]}(x) \in \overline{\operatorname{span}}\{\rho_{1/k} : k \ge 2\}$ in $\mathcal{H}_{1/2}$. In the shifted framework introduced by Báez-Duarte et al. [4, 5] and generalized by Burnol [9, 21], span_of_dilates [12, 14], and halasz_resonator [13, 15, 17, 23, 24, 25], one replaces $\rho_{1/k}$ by the pole-canceled dilates $$\tilde{\rho}_{1/k}(x) := \frac{1}{k}\left\{\frac{1}{x}\right\} - \left\{\frac{1}{kx}\right\} = -\left\{\frac{\lfloor 1/x \rfloor}{k}\right\} \qquad (k \ge 2), \tag{1}$$ which, together with the constant indicator $E(x) := \mathbf{1}_{(0, 1]}(x)$, belong to the closed harmonic step subspace $S_\sigma := \overline{\operatorname{span}}\{\mathbf{1}_{I_m} : m \ge 1\} \subsetneq \mathcal{H}_\sigma$, where $I_m := (1/(m+1), 1/m]$ has step measure $$w_m^{(\sigma)} := m^{-2\sigma} - (m+1)^{-2\sigma} = \frac{2\sigma}{m^{2\sigma+1}}\left(1 + O\left(\frac{1}{m}\right)\right) \qquad (m \ge 1). \tag{2}$$ Denoting the piecewise constant value of $f \in S_\sigma$ on $I_m$ by $f_m := f|_{I_m}$ ($f_0 := 0$), the step norm on $S_\sigma$ and the dual Riesz norm on $S_\sigma^*$ are given by $$\|f\|_\sigma^2 = \sum_{m=1}^\infty |f_m|^2 w_m^{(\sigma)}, \qquad \|\mathcal{L}\|_{S_\sigma^*}^2 = \sum_{m=1}^\infty \frac{|c_m(\mathcal{L})|^2}{w_m^{(\sigma)}}, \qquad \mathcal{L}(f) = \sum_{m=1}^\infty c_m(\mathcal{L})\,f_m. \tag{3}$$ By the Hahn-Banach theorem (riemann_hypothesis_7 [15], riemann_hypothesis_9 [17]), the squared approximation distance $d_N(\sigma)^2 := \inf_{b_2, \dots, b_N \in \mathbb{C}} \|E - \sum_{k=2}^N b_k \tilde{\rho}_{1/k}\|_\sigma^2$ equals the dual Rayleigh maximum $$d_N(\sigma)^2 = \sup\left\{ \frac{|\mathcal{L}(E)|^2}{\|\mathcal{L}\|_{S_\sigma^*}^2} : \mathcal{L} \in \operatorname{Ann}_{S_\sigma^*}(\tilde{\rho}_{1/2}, \dots, \tilde{\rho}_{1/N}) \setminus \{0\} \right\}, \tag{4}$$ where $\operatorname{Ann}_{S_\sigma^*}(\tilde{\rho}_{1/2}, \dots, \tilde{\rho}_{1/N}) := \{\mathcal{L} \in S_\sigma^* : \mathcal{L}(\tilde{\rho}_{1/k}) = 0 \text{ for all } 2 \le k \le N\}$.

In the Vasyunin-Balazard difference basis $\mathcal{D}_m(f) := f_m - f_{m-1}$ and $\mathcal{C}_n(f) := \sum_{d \mid n} \mu(n/d)\mathcal{D}_d(f)$ ([6, 7, 15, 17, 18]), which satisfies $\mathcal{C}_1(\tilde{\rho}_{1/k}) = -1/k$, $\mathcal{C}_n(\tilde{\rho}_{1/k}) = \delta_{n,k}$ ($n, k \ge 2$), and $\mathcal{C}_n(E) = \mu(n)$, any functional $\mathcal{L} = \sum_{n=1}^M \alpha_n \mathcal{C}_n = \sum_{d=1}^M \beta_d \mathcal{D}_d$ supported on the finite index set $m \le M$ has potential $\beta_d = \sum_{j=1}^{\lfloor M/d \rfloor} \alpha_{dj}\mu(j)$ ($\beta_{M+1} := 0$), step differences $c_m(\mathcal{L}) = \beta_m - \beta_{m+1}$, target evaluation $\mathcal{L}(E) = \beta_1 = \sum_{n=1}^M \alpha_n \mu(n)$, and dilate evaluations $\mathcal{L}(\tilde{\rho}_{1/k}) = \alpha_k - \frac{1}{k}\alpha_1$ ($k \ge 2$).

Thus, for any finite support bound $M \ge N$, the finite-support annihilator subspace ([17, Section 2.1]) $$\mathcal{A}_{N, M} := \operatorname{Ann}_{S_\sigma^*}^{(M)}(\tilde{\rho}_{1/2}, \dots, \tilde{\rho}_{1/N}) := \left\{\mathcal{L} \in \operatorname{Ann}_{S_\sigma^*}(\tilde{\rho}_{1/2}, \dots, \tilde{\rho}_{1/N}) : c_m(\mathcal{L}) = 0 \text{ for all } m > M\right\} \tag{5}$$ has exact algebraic dimension $M - N + 1$ and admits the canonical basis $$\mathcal{A}_{N, M} = \operatorname{span}\!\big\{\mathcal{F}_N,\; \mathcal{C}_{N+1},\; \dots,\; \mathcal{C}_M\big\}, \qquad \mathcal{F}_N := \sum_{n=1}^N \frac{1}{n}\,\mathcal{C}_n, \tag{6}$$ where $\mathcal{F}_N$ is Balazard's unsmoothed harmonic head functional. Normalizing the head coefficient by $\alpha_1 = 1$, every annihilator $\mathcal{L} \in \mathcal{A}_{N, M}$ has the canonical tail expansion $$\mathcal{L} = \mathcal{F}_N + \sum_{n=N+1}^M \alpha_n\,\mathcal{C}_n, \qquad \alpha_n = \frac{1}{n} \quad (1 \le n \le N), \tag{7}$$ with target evaluation $\mathcal{L}(E) = \sum_{n=1}^N \frac{\mu(n)}{n} + \sum_{n=N+1}^M \alpha_n \mu(n)$.

1.2 Historical Context and Motivation: Resolving the Subspace Optimality Gap

The finite annihilator parameterization and head-truncated duality originate in riemann_hypothesis_7 [15] and riemann_hypothesis_9 [17]; the one-band and two-segment bounds including the $M = 2N$ correction were established in riemann_hypothesis_12 [24]; and the complete wavelet subspace $\mathcal{W}_{N, \delta}^{(2, L)}$ with numerical optimality comparisons was developed in riemann_hypothesis_14 [25]. In the broader literature, Balazard [18, Section 2] explicitly gave the Möbius-inverse biorthogonal construction; combining that inversion with cumulative step sums yields the present floor-quotient inverse, while Ehm [19] analyzed related Gram kernels and quadratic forms. Theorem 3's duality is the finite-dimensional Hilbert projection theorem applied to the established annihilator space.

However, as highlighted in recent peer discussions, three concrete structural gaps remained:

1. Universal Lower Bounds at Arbitrary Support Ratio: In [24], the universal lower bound was derived for $\delta \le 1$. An explicit all-scale specialization valid for arbitrary $\delta > 0$ requires summing floor-quotient steps over general quotient ranges $K = \lfloor 1 + \delta \rfloor \ge 1$.
2. Intermediate Band Invariants and Subspace Optimality Certification: The intermediate step differences across active bands $\overline{\mathcal{B}}_j$ and dead zones $\mathcal{Z}_j$ were analyzed only heuristically, lacking explicit sufficient boundary admissibility conditions ensuring invariance without arithmetic coincidences, and sharp quantitative analysis of dyadic coupling. Furthermore, the subspace capture percentages of riemann_hypothesis_14 [25] remained numerical comparisons rather than two-sided interval-certified optimality statements.
3. Quantitative Bridge to the Normalized Distance: While unnormalized minimum dual norms $\mathcal{E}_{\min}$ bound denominators under $\alpha_1 = 1$, the normalized Nyman–Beurling distance $d_{\mathcal{A}_{N, M}}(\sigma)^2 = \max_{\mathcal{L}} |\mathcal{L}(E)|^2 / \|\mathcal{L}\|_{S_\sigma^*}^2$ involves both target evaluation and norm. An exact geometric transference identity was needed to link $\mathcal{E}_{\min}$ and its band relaxations directly to $d_{\mathcal{A}_{N, M}}(\sigma)^2$.

1.3 Summary of New Contributions

In this paper, we resolve all three gaps through three unconditional operator and primal-dual theorems:

1. Theorem 1 (Floor-Quotient Inverse, Exact Minimum-Norm Duality, and All-Scale Universal Lower Bounds): We prove that the coordinate matrix $C_M \in \mathbb{Z}^{M \times M}$ has exact integer inverse $(C_M^{-1})_{r, m} = \lfloor m/r \rfloor$ ($1 \le r \le m \le M$). Consequently, $(G_M^{\mathcal{C}, (\sigma)})^{-1} = Q_M^{(\sigma)}$ is given in closed form without matrix inversion by the floor-quotient Gram matrix $(Q_M^{(\sigma)})_{r, s} = \sum_{m=\max(r, s)}^M w_m^{(\sigma)} \lfloor m/r \rfloor \lfloor m/s \rfloor$. Schur-complement duality yields the exact minimum norm $\mathcal{E}_{\min}(N, M, \sigma) = \mathbf{h}_N^T (Q_{N, M}^{(\sigma)})^{-1} \mathbf{h}_N$ and extends the universal lower bound of riemann_hypothesis_12 to arbitrary $\delta > 0$: $$\|\mathcal{L}\|_{S_\sigma^*}^2 \ge \mathcal{E}_{\min}(N, N_\delta, \sigma) \ge \frac{N^{-2}}{\sum_{k=1}^K (2k - 1)(k N)^{-2\sigma} - K^2 (N_\delta + 1)^{-2\sigma}} \qquad (K := \lfloor 1 + \delta \rfloor \ge 1). \tag{8}$$

2. Theorem 2 (Admissible Band Invariants, Dead Zones, and Consistent Combined Multi-Band Hierarchy): For $M > N$, we formulate the explicit admissible band index set $\mathcal{J}_{\mathrm{adm}}(N, M)$ based on sufficient dead-zone boundary containment, proving that every $\mathcal{L} \in \mathcal{A}_{N, M}$ has invariant potential $\beta_m = m_j/m$ on dead zones $\mathcal{Z}_j$ and invariant step sums $S_j(N, M)$ across active bands $\overline{\mathcal{B}}_j$ for $j \in \mathcal{J}_{\mathrm{adm}}$. We define the combined multi-band bound $\mathcal{E}_{\mathrm{MB}} := \max\{\mathcal{E}_{\mathrm{single}}, \sum_{j=1}^{J_{\max}} \mathcal{E}_{\mathrm{DZ}, j} + \sum_{j \in \mathcal{J}_{\mathrm{adm}}} \frac{|S_j|^2}{W_j^{(\sigma)}}\}$ with $J_{\max} := \lfloor N/(M-N) \rfloor \ge 0$, holding consistently for all $M > N$ and coinciding with $\mathcal{E}_{\mathrm{single}}$ when $M \ge 2N$. We derive the Dyadic-Ladder lower bound and identify the dipole-matching mechanism yielding exact equality $\mathcal{E}_{\mathrm{MB}+\mathrm{ladder}} = \mathcal{E}_{\mathrm{MB}}$ at $\sigma = 1/2$ for odd $M$.

3. Theorem 3 (Exact Primal-Dual Equivalence, Geometric Transference, and Two-Sided 60-Digit Interval Certification): We apply the projection theorem to prove zero-gap equivalence $d_{\mathcal{A}_{N, M}}(\sigma)^2 = \min_{\mathbf{b}} \sum_{m=1}^M w_m^{(\sigma)}|1 + \sum_{k=2}^N b_k \{m/k\}|^2$, reconstruct the extremal annihilator in closed form via floor quotients, and establish the Geometric Transference Theorem: $$d_{\mathcal{A}_{N, M}}(\sigma)^2 = \frac{|\mathcal{L}_{\min}(E)|^2}{\mathcal{E}_{\min}(N, M, \sigma)\,\cos^2\theta_{N, M}(\sigma)} \ge \frac{|\mathcal{L}_{\min}(E)|^2}{\mathcal{E}_{\min}(N, M, \sigma)} \ge \frac{|\mathcal{L}_{\min}(E)|^2}{\mathcal{E}_{\mathrm{MB}+\mathrm{ladder}}(N, M, \sigma)},$$ where $\cos\theta_{N, M}(\sigma) := \frac{\langle P_{V_M^\perp}\mathbf{1}_M, P_{V_M^\perp}\mathbf{m}_M \rangle_\sigma}{\|P_{V_M^\perp}\mathbf{1}_M\|_\sigma \|P_{V_M^\perp}\mathbf{m}_M\|_\sigma} \in (0, 1]$ is the projection alignment angle in $S_{\sigma, M}$. We establish two-sided 60-digit interval sandwiches $[\underline{d}_{\mathcal{A}_{N, M}}(\sigma)^2, \overline{d}_{\mathcal{A}_{N, M}}(\sigma)^2]$ with gap $\le 2.19 \times 10^{-21}$ across 18 benchmark configurations ($N \in \{40, 80, 160\}$), rigorously certifying the wavelet capture percentages of riemann_hypothesis_14.

(Scope note: all theorems in this paper are unconditional finite-support operator identities and primal-dual bounds on $\mathcal{A}_{N, M} \subset S_\sigma^*$, and do not resolve the Riemann Hypothesis or establish a new zero-free region.)

---

2. Floor-Quotient Operator Inverses, Minimum-Norm Duality, and All-Scale Universal Lower Bounds

Let $M \ge 1$. For any sequence $\boldsymbol{\alpha} = (\alpha_1, \dots, \alpha_M)^T \in \mathbb{C}^M$, the functional $\mathcal{L} = \sum_{n=1}^M \alpha_n \mathcal{C}_n$ has potential $\beta_m = \sum_{j=1}^{\lfloor M/m \rfloor} \mu(j)\alpha_{jm}$ ($1 \le m \le M$, $\beta_{M+1} := 0$) and step differences $c_m = \beta_m - \beta_{m+1}$. We denote by $C_M \in \mathbb{Z}^{M \times M}$ the upper-triangular integer matrix mapping Vasyunin-Balazard coordinates to step-difference coordinates: $$\mathbf{c} = C_M\,\boldsymbol{\alpha}, \qquad (C_M)_{m, n} := \mu\left(\frac{n}{m}\right)\mathbf{1}_{m \mid n} - \mu\left(\frac{n}{m+1}\right)\mathbf{1}_{(m+1) \mid n} \qquad (1 \le m, n \le M). \tag{9}$$ Let $W_M^{(\sigma)} := \operatorname{diag}(1/w_1^{(\sigma)}, \dots, 1/w_M^{(\sigma)})$. The Gram matrix of the Vasyunin-Balazard basis $\{\mathcal{C}_1, \dots, \mathcal{C}_M\}$ in $S_\sigma^*$ is $G_M^{\mathcal{C}, (\sigma)} = C_M^T W_M^{(\sigma)} C_M$.

Theorem 1 (Floor-Quotient Matrix Inverse, Exact Minimum-Norm Duality, and All-Scale Universal Lower Bounds). Let $M \ge 1$ and $\sigma \in [1/2, 1)$.

1. (Exact Integer Matrix Inverse): The matrix $C_M \in \mathbb{Z}^{M \times M}$ is unipotent upper-triangular with determinant $\det C_M = 1$, and its exact matrix inverse is the transpose of the floor-quotient matrix $U_M \in \mathbb{Z}^{M \times M}$: $$\big(C_M^{-1}\big)_{r, m} = \left\lfloor \frac{m}{r} \right\rfloor \qquad (1 \le r \le m \le M), \qquad (C_M^{-1})_{r, m} = 0 \quad (r > m). \tag{10}$$
2. (Closed-Form Inverse Gram Operator): The inverse of the $M \times M$ Vasyunin-Balazard Gram matrix $G_M^{\mathcal{C}, (\sigma)}$ is given in exact finite closed form without matrix inversion by the floor-quotient Gram matrix $Q_M^{(\sigma)} \in \mathbb{R}^{M \times M}$: $$\big(G_M^{\mathcal{C}, (\sigma)}\big)^{-1} = C_M^{-1} \big(W_M^{(\sigma)}\big)^{-1} \big(C_M^{-1}\big)^T =: Q_M^{(\sigma)}, \tag{11}$$ where the entries of $Q_M^{(\sigma)}$ are given by the discrete sum and continuous integral $$\big(Q_M^{(\sigma)}\big)_{r, s} = \sum_{m = \max(r, s)}^M w_m^{(\sigma)} \left\lfloor \frac{m}{r} \right\rfloor \left\lfloor \frac{m}{s} \right\rfloor = \int_{\frac{1}{M+1}}^1 \left\lfloor \frac{1}{r x} \right\rfloor \left\lfloor \frac{1}{s x} \right\rfloor 2\sigma x^{2\sigma - 1}\,dx \qquad (1 \le r, s \le M). \tag{12}$$
3. (Exact Minimum-Norm Duality): For any $N \ge 2$ and $M \ge N$, let $Q_{N, M}^{(\sigma)} := [Q_M^{(\sigma)}]_{1:N, 1:N}$ denote the leading $N \times N$ principal submatrix of $Q_M^{(\sigma)}$, and let $\mathbf{h}_N := (1, 1/2, \dots, 1/N)^T \in \mathbb{Q}^N$. The exact minimum dual norm squared of any tail-extended functional $\mathcal{L} = \mathcal{F}_N + \sum_{n=N+1}^M \alpha_n \mathcal{C}_n \in \mathcal{A}_{N, M}$ ($\alpha_1 = 1$) over all tail coefficients $(\alpha_{N+1}, \dots, \alpha_M) \in \mathbb{C}^{M-N}$ is given by $$\mathcal{E}_{\min}(N, M, \sigma) := \min_{\alpha_{N+1}, \dots, \alpha_M \in \mathbb{C}} \left\|\mathcal{F}_N + \sum_{n=N+1}^M \alpha_n\,\mathcal{C}_n\right\|_{S_\sigma^*}^2 = \mathbf{h}_N^T \big(Q_{N, M}^{(\sigma)}\big)^{-1} \mathbf{h}_N. \tag{13}$$ Moreover, $\mathcal{E}_{\min}(N, M, \sigma)$ satisfies the dual Rayleigh variational characterization $$\mathcal{E}_{\min}(N, M, \sigma) = \max_{0 \ne \mathbf{a} \in \mathbb{R}^N} \frac{\left(\sum_{r=1}^N \frac{a_r}{r}\right)^2}{\sum_{m=1}^M w_m^{(\sigma)} \left(\sum_{r=1}^N a_r \left\lfloor \frac{m}{r} \right\rfloor\right)^2} = \left(\min_{a_2, \dots, a_N \in \mathbb{R}} \sum_{m=1}^M w_m^{(\sigma)} \left(m - \sum_{r=2}^N a_r \left\{\frac{m}{r}\right\}\right)^2\right)^{-1}, \tag{14}$$ attained by the dual vector $\mathbf{a}^* = (Q_{N, M}^{(\sigma)})^{-1} \mathbf{h}_N$ (unique up to a nonzero scalar multiple), with optimal tail coefficients $\alpha_n^* = \sum_{r=1}^N (Q_M^{(\sigma)})_{n, r} a_r^*$ ($N < n \le M$).
4. (All-Scale Universal Lower Bound for All $\delta > 0$): For any $\delta > 0$, setting $M = N_\delta := \lfloor(1+\delta)N\rfloor$ and $K := \lfloor M/N \rfloor = \lfloor 1 + \delta \rfloor \ge 1$, evaluating (14) at the coordinate vector $\mathbf{a} = \mathbf{e}_N$ yields the unconditional closed-form lower bound $$\|\mathcal{L}\|_{S_\sigma^*}^2 \ge \mathcal{E}_{\min}(N, N_\delta, \sigma) \ge \frac{N^{-2}}{(Q_{N, N_\delta}^{(\sigma)})_{N, N}} = \frac{N^{-2}}{\sum_{k=1}^K (2k - 1)(k N)^{-2\sigma} - K^2 (N_\delta + 1)^{-2\sigma}}, \tag{15}$$ which holds for every $\mathcal{L} \in \mathcal{A}_{N, N_\delta}$ with $\alpha_1 = 1$, recovering Theorem 1 of riemann_hypothesis_12 [24] for $K = 1$ ($\delta < 1$).

Proof. Part 1: Let $U_M \in \mathbb{Z}^{M \times M}$ be the upper-triangular matrix $(U_M)_{m, r} := \lfloor m/r \rfloor$ ($1 \le r \le m \le M$). By definition of $\beta_m$ and $c_m$, $\beta_m = \sum_{k=m}^M c_k$. Substituting this into $\alpha_r = \sum_{j=1}^{\lfloor M/r \rfloor} \beta_{rj}$ (via Möbius inversion of $\beta_m = \sum_{d=1}^{\lfloor M/m \rfloor} \mu(d)\alpha_{dm}$) gives $$\alpha_r = \sum_{j=1}^{\lfloor M/r \rfloor} \sum_{k=rj}^M c_k = \sum_{k=r}^M c_k \sum_{j=1}^{\lfloor k/r \rfloor} 1 = \sum_{k=r}^M \left\lfloor \frac{k}{r} \right\rfloor c_k = \sum_{m=r}^M (U_M^T)_{r, m}\,c_m.$$ Thus $\boldsymbol{\alpha} = U_M^T \mathbf{c}$. Since $C_M \boldsymbol{\alpha} = \mathbf{c}$ for all $\boldsymbol{\alpha} \in \mathbb{C}^M$, $C_M^{-1} = U_M^T$, proving (10).

Part 2: Since $G_M^{\mathcal{C}, (\sigma)} = C_M^T W_M^{(\sigma)} C_M$, its inverse is $(G_M^{\mathcal{C}, (\sigma)})^{-1} = C_M^{-1} (W_M^{(\sigma)})^{-1} (C_M^T)^{-1}$. Using (10), $(C_M^{-1} (W_M^{(\sigma)})^{-1} (C_M^T)^{-1})_{r, s} = \sum_{m=1}^M (C_M^{-1})_{r, m} w_m^{(\sigma)} (C_M^{-1})_{s, m} = \sum_{m=\max(r, s)}^M w_m^{(\sigma)} \lfloor m/r \rfloor \lfloor m/s \rfloor$. Since $\lfloor 1/(rx) \rfloor = \lfloor m/r \rfloor$ identically on $x \in I_m = (1/(m+1), 1/m]$ and $\int_{I_m} 2\sigma x^{2\sigma-1}\,dx = w_m^{(\sigma)}$, the sum equals $\int_{1/(M+1)}^1 \lfloor 1/(rx) \rfloor \lfloor 1/(sx) \rfloor 2\sigma x^{2\sigma-1}\,dx$, proving (12).

Part 3: Writing $\boldsymbol{\alpha} = (\mathbf{h}_N^T, \boldsymbol{\alpha}_{\mathrm{tail}}^T)^T$ where $\mathbf{h}_N = (1, 1/2, \dots, 1/N)^T$ and $\boldsymbol{\alpha}_{\mathrm{tail}} = (\alpha_{N+1}, \dots, \alpha_M)^T$, the dual norm squared is $\|\mathcal{L}\|_{S_\sigma^*}^2 = \boldsymbol{\alpha}^T G_M^{\mathcal{C}, (\sigma)} \boldsymbol{\alpha}$. Minimizing this quadratic form over $\boldsymbol{\alpha}_{\mathrm{tail}} \in \mathbb{C}^{M-N}$ is the classical partial minimization of a positive definite quadratic form. By the Schur complement formula, if $G = \begin{pmatrix} G_{11} & G_{12} \\ G_{21} & G_{22} \end{pmatrix}$ and $Q = G^{-1} = \begin{pmatrix} Q_{11} & Q_{12} \\ Q_{21} & Q_{22} \end{pmatrix}$, then the Schur complement $G_{11} - G_{12} G_{22}^{-1} G_{21}$ equals $Q_{11}^{-1}$. Thus $\min_{\boldsymbol{\alpha}_{\mathrm{tail}}} \boldsymbol{\alpha}^T G \boldsymbol{\alpha} = \mathbf{h}_N^T Q_{11}^{-1} \mathbf{h}_N = \mathbf{h}_N^T (Q_{N, M}^{(\sigma)})^{-1} \mathbf{h}_N$, proving (13). The dual variational formula (14) follows from the Rayleigh quotient $\mathbf{h}_N^T Q_{N, M}^{-1} \mathbf{h}_N = \max_{\mathbf{a} \ne 0} \frac{(\mathbf{h}_N^T \mathbf{a})^2}{\mathbf{a}^T Q_{N, M} \mathbf{a}}$, whose maximizer is unique up to nonzero scaling. Normalizing $\mathbf{a}$ by the affine condition $\mathbf{h}_N^T \mathbf{a} = \sum_{r=1}^N \frac{a_r}{r} = 1$ fixes $a_1 = 1 - \sum_{r=2}^N \frac{a_r}{r}$. Expanding the coordinate action yields $\sum_{r=1}^N a_r \lfloor m/r \rfloor = a_1 m + \sum_{r=2}^N a_r (m/r - \{m/r\}) = m(a_1 + \sum_{r=2}^N \frac{a_r}{r}) - \sum_{r=2}^N a_r \{m/r\} = m - \sum_{r=2}^N a_r \{m/r\}$, converting the denominator to $\sum_{m=1}^M w_m^{(\sigma)}(m - \sum_{r=2}^N a_r \{m/r\})^2$ and giving the reciprocal-minimum expression in (14).

Part 4: For $\mathbf{a} = \mathbf{e}_N$, $\sum_{r=1}^N a_r/r = 1/N$, and $(Q_{N, N_\delta}^{(\sigma)})_{N, N} = \sum_{m=N}^{N_\delta} w_m^{(\sigma)} \lfloor m/N \rfloor^2$. For $m \in [k N, (k+1)N - 1]$ ($1 \le k \le K-1$), $\lfloor m/N \rfloor = k$; for $m \in [K N, N_\delta]$, $\lfloor m/N \rfloor = K$. Splitting the sum and using $\sum_{m=A}^B w_m^{(\sigma)} = A^{-2\sigma} - (B+1)^{-2\sigma}$ yields $$\sum_{m=N}^{N_\delta} w_m^{(\sigma)} \left\lfloor \frac{m}{N} \right\rfloor^2 = \sum_{k=1}^{K-1} k^2 \big((kN)^{-2\sigma} - ((k+1)N)^{-2\sigma}\big) + K^2 \big((KN)^{-2\sigma} - (N_\delta+1)^{-2\sigma}\big).$$ Telescoping the terms $k^2 - (k-1)^2 = 2k - 1$ produces $\sum_{k=1}^K (2k-1)(kN)^{-2\sigma} - K^2(N_\delta+1)^{-2\sigma}$, proving (15). $\square$

Table 1 validates Theorem 1 in exact rational arithmetic (fractions.Fraction) at $\sigma = 1/2$.

| $N$ | $M$ | $\dim \mathcal{A}_{N, M}$ | Exact Minimum Norm $\mathcal{E}_{\min}(N, M, 1/2)$ | Float Value | Duality Gap | Exact Dual Extremizer $\mathbf{a}^*$ | Exact Optimal Tail Coefficients $\boldsymbol{\alpha}^*_{\mathrm{tail}}$ | | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | $2$ | $4$ | $3$ | $105/184$ | $0.570652$ | $0$ | $(-15/92, 135/92)^T$ | $(9/46, 21/184)^T$ | | $4$ | $5$ | $2$ | $1759/1872$ | $0.939637$ | $0$ | $(-55/156, 8/39, 289/156, 119/52)^T$ | $(29/312)^T$ | | $4$ | $6$ | $3$ | $2359/3628$ | $0.650221$ | $0$ | $(-325/1814, -257/1814, 822/907, 2170/907)^T$ | $(245/1814, 701/10884)^T$ | | $4$ | $8$ | $5$ | $119057/291472$ | $0.408468$ | $0$ | $(-12005/72868, -3106/18217, 30963/36434, 109357/72868)^T$ | $(21125/145736, 46093/437208, 1183/18217, 10935/291472)^T$ | | $6$ | $9$ | $4$ | $1696427/3354000$ | $0.505792$ | $0$ | $\mathbf{a}^* \in \mathbb{Q}^6$ (results/...json) | $(35873/335400, 4307/67080, 4559/167700)^T$ | | $8$ | $10$ | $3$ | $2180821/3910200$ | $0.557726$ | $0$ | $\mathbf{a}^* \in \mathbb{Q}^8$ (results/...json) | $(697/9310, 809/22344)^T$ |

Table 1: Exact rational minimum dual norms $\mathcal{E}_{\min}(N, M, 1/2) \in \mathbb{Q}$ and dual extremizers verified in fractions.Fraction arithmetic (results/floor_quotient_subspace_duality.json). In every case, the duality gap $\mathcal{E}_{\min} - \mathbf{h}_N^T Q_{N, M}^{-1} \mathbf{h}_N$ vanishes identically in $\mathbb{Q}$.

---

3. Quotient-Hyperbola Dead Zones, Multi-Band Bounds, and Dyadic-Ladder Coupling

3.1 Quotient-Hyperbola Dead Zones and Non-Squarefree Super-Dead-Zone Merging

When $\delta \in (0, 1)$ ($M = N_\delta = \lfloor(1+\delta)N\rfloor$), the free tail coefficients $\alpha_{N+1}, \dots, \alpha_M$ are localized to $n > N$. How does this freedom affect the potentials $\beta_m(\mathcal{L}) = \frac{1}{m}\sum_{j=1}^{\lfloor M/m \rfloor} \mu(j)\,j\alpha_{jm} = \sum_{j=1}^{\lfloor M/m \rfloor} \mu(j)\alpha_{jm}$ and step differences $c_m(\mathcal{L})$ for $m \le N$?

Theorem 2 (Quotient-Hyperbola Dead Zones, Super-Dead-Zone Merging, and Multi-Band Lower Bounds). Let $N \ge 2$, $M > N$ (so $M - N \ge 1$ and $\delta := (M - N)/N > 0$), and set $J_{\max} := \min\{\lfloor N/(M-N) \rfloor, N\} \ge 1$.

1. (Dead Zones and Invariant Potentials): For each integer $1 \le j \le J_{\max}$, define the $j$-th quotient-hyperbola dead zone $$\mathcal{Z}_j(N, M) := \left(\left\lfloor \frac{M}{j+1} \right\rfloor,\; \left\lfloor \frac{N}{j} \right\rfloor\right] \cap \mathbb{Z}. \tag{16}$$ Whenever $\lfloor M/(j+1) \rfloor < \lfloor N/j \rfloor$, $\mathcal{Z}_j(N, M)$ is non-empty and every $m \in \mathcal{Z}_j(N, M)$ satisfies $\lfloor M/m \rfloor = \lfloor N/m \rfloor = j$. Consequently, every functional $\mathcal{L} \in \mathcal{A}_{N, M}$ with $\alpha_1 = 1$ has identical, tail-independent potential on $\mathcal{Z}_j(N, M)$ and invariant step differences on its interior: $$\beta_m(\mathcal{L}) = \frac{m_j}{m}, \qquad c_m(\mathcal{L}) = \frac{m_j}{m(m+1)} \quad \Big(\forall\, m \in \mathcal{Z}_j^\circ(N, M) := \big[\lfloor M/(j+1)\rfloor + 1,\; \lfloor N/j\rfloor - 1\big]\Big), \tag{17}$$ where $m_j := \sum_{d=1}^j \frac{\mu(d)}{d}$ is the truncated Mertens harmonic sum (in particular, $m_1 = 1$, $m_2 = 1/2$, $m_3 = 1/6$, $m_4 = 1/6$). On $\mathcal{Z}_j^\circ(N, M)$, the step differences contribute the exact invariant dead-zone energy $$\mathcal{E}_{\mathrm{DZ}, j}(N, M, \sigma) := \sum_{m \in \mathcal{Z}_j^\circ(N, M)} \frac{m_j^2}{m^2(m+1)^2 w_m^{(\sigma)}}. \tag{17\text{a}}$$
2. (Non-Squarefree Super-Dead-Zone Merging): Suppose that for some $j \ge 2$, $\mu(j) = 0$ (e.g., $j = 4$) and $\lfloor M/(j+1) \rfloor < \lfloor N/j \rfloor \le \lfloor M/j \rfloor < \lfloor N/(j-1) \rfloor$. Then for all $m \in \mathcal{B}_j(N, M) := (\lfloor N/j \rfloor, \lfloor M/j \rfloor]$, the only multiple $j m$ that could fall in $(N, M]$ enters with coefficient $\mu(j)\alpha_{jm} = 0$. Consequently, the entire active band $\mathcal{B}_j$ merges with $\mathcal{Z}_j$ and $\mathcal{Z}_{j-1}$ into a single super-dead-zone plateau $$m\beta_m(\mathcal{L}) = m_{j-1} = m_j \qquad \left(\forall\, m \in \left(\left\lfloor \frac{M}{j+1} \right\rfloor,\; \left\lfloor \frac{N}{j-1} \right\rfloor\right]\right). \tag{18}$$
3. (Admissible Bands and Multi-Band Lower Bound): For each $1 \le j \le J_{\max}$, define the $j$-th active band $\overline{\mathcal{B}}_j(N, M) := [\lfloor N/j \rfloor, \lfloor M/j \rfloor] \cap \mathbb{Z}$. A sufficient condition for the band sum $\sum_{m \in \overline{\mathcal{B}}_j} c_m(\mathcal{L})$ to equal the tail-independent constant $S_j(N, M) := \frac{m_j}{\lfloor N/j \rfloor} - \frac{m_{j-1}}{\lfloor M/j \rfloor + 1}$ ($m_0 := 0$) across all $\mathcal{L} \in \mathcal{A}_{N, M}$ without relying on accidental arithmetic vanishing of $\mu(n)$ is that the left boundary belongs to $\mathcal{Z}_j(N, M)$ and (for $j \ge 2$) the right boundary belongs to $\mathcal{Z}_{j-1}(N, M)$. This defines the admissible band index set $$\mathcal{J}_{\mathrm{adm}}(N, M) := \left\{ 1 \le j \le J_{\max} \;:\; \left\lfloor \frac{M}{j+1} \right\rfloor < \left\lfloor \frac{N}{j} \right\rfloor \;\text{ and }\; \big(j = 1 \;\text{ or }\; \left\lfloor \frac{M}{j} \right\rfloor + 1 \le \left\lfloor \frac{N}{j-1} \right\rfloor\big) \right\}. \tag{19}$$ (Note: While boundary dead-zone containment is sufficient, it is not strictly necessary for individual $(N, M)$ where arithmetic cancellations occur; for instance, at $N=6, M=8, j=3$, $\mathcal{Z}_3 = \emptyset$ so $3 \notin \mathcal{J}_{\mathrm{adm}}$, yet $\mu(4) = 0$ ensures $c_2 = S_3(6, 8)$ identically. The set $\mathcal{J}_{\mathrm{adm}}$ guarantees invariance unconditionally for all tail extensions.) For all $j \in \mathcal{J}_{\mathrm{adm}}(N, M)$, the bands $\overline{\mathcal{B}}_j$ and interior dead zones $\mathcal{Z}_j^\circ$ are pairwise disjoint subsets of $[1, M]$, leaving an uncovered small-index remainder $\mathcal{R}(N, M) := [1, M] \setminus (\bigcup \mathcal{Z}_j^\circ \cup \bigcup_{j \in \mathcal{J}_{\mathrm{adm}}} \overline{\mathcal{B}}_j)$. Cauchy-Schwarz on each admissible band plus the interior dead-zone energies, combined with the single-coordinate bound (15), yields the Combined Multi-Band Lower Bound holding for all $M > N$: $$\|\mathcal{L}\|_{S_\sigma^*}^2 \ge \mathcal{E}_{\mathrm{MB}}(N, M, \sigma) := \max\left\{ \mathcal{E}_{\mathrm{single}}(N, M, \sigma),\; \sum_{j=1}^{J_{\max}} \mathcal{E}_{\mathrm{DZ}, j}(N, M, \sigma) + \sum_{j \in \mathcal{J}_{\mathrm{adm}}(N, M)} \frac{|S_j(N, M)|^2}{W_j^{(\sigma)}(N, M)} \right\}, \tag{20}$$ where $\mathcal{E}_{\mathrm{single}}(N, M, \sigma) := N^{-2}/(Q_{N, M}^{(\sigma)})_{N, N}$, and $W_j^{(\sigma)}(N, M) := \sum_{m \in \overline{\mathcal{B}}_j(N, M)} w_m^{(\sigma)} = \lfloor N/j \rfloor^{-2\sigma} - (\lfloor M/j \rfloor + 1)^{-2\sigma}$. (When $M \ge 2N$, $J_{\max} = 0$ and $\mathcal{J}_{\mathrm{adm}} = \emptyset$, so the sum over bands vanishes and $\mathcal{E}_{\mathrm{MB}}$ coincides identically with $\mathcal{E}_{\mathrm{single}}$, ensuring the hierarchy is well-defined and monotonic for all $M > N$.)
4. (Dyadic-Ladder Coupling and Dipole Matching at Odd $M$ and $\sigma = 1/2$): Assume $2 \mid N$, $2M < 3N$ ($M - N < N/2$), and $1, 2 \in \mathcal{J}_{\mathrm{adm}}(N, M)$. Define the clipped pair weights in $\overline{\mathcal{B}}_1 = [N, M]$ for $m \in [N/2, \lfloor M/2 \rfloor]$: $$W_{m, 1}^{(\sigma)} := \sum_{k \in \{2m, 2m+1\} \cap [N, M]} w_k^{(\sigma)}, \qquad W_{\mathrm{rem}, 1}^{(\sigma)} := \sum_{k \in [N, M] \setminus \bigcup_{m=N/2}^{\lfloor M/2 \rfloor} (\{2m, 2m+1\} \cap [N, M])} w_k^{(\sigma)} \equiv 0 \quad (\text{since } 2 \mid N). \tag{21}$$ Coupling the $M_2 := \lfloor M/2 \rfloor - N/2 + 1$ clipped dyadic conservation laws $c_m + \sum_{k \in \{2m, 2m+1\}\cap [N, M]} c_k = \frac{1}{m(m+1)}$ with $\sum_{k=N}^M c_k = 1/N$ yields the closed-form Dyadic-Ladder Lower Bound $$\mathcal{E}_{\mathrm{MB}+\mathrm{ladder}}(N, M, \sigma) := \max\left\{ \mathcal{E}_{\mathrm{MB}}(N, M, \sigma),\; \mathcal{E}_{\mathrm{DZ}} + \mathcal{E}_{\mathrm{MB}, 3:J_{\max}} + \sum_{m=N/2}^{\lfloor M/2 \rfloor} \frac{\left(\frac{1}{m(m+1)}\right)^2}{w_m^{(\sigma)} + W_{m, 1}^{(\sigma)}} + \frac{\left(\frac{1}{N} - \sum_{m=N/2}^{\lfloor M/2 \rfloor} \frac{\frac{1}{m(m+1)} W_{m, 1}^{(\sigma)}}{w_m^{(\sigma)} + W_{m, 1}^{(\sigma)}}\right)^2}{\sum_{m=N/2}^{\lfloor M/2 \rfloor} \frac{w_m^{(\sigma)} W_{m, 1}^{(\sigma)}}{w_m^{(\sigma)} + W_{m, 1}^{(\sigma)}}} \right\}. \tag{22}$$ We have $\mathcal{E}_{\mathrm{MB}+\mathrm{ladder}}(N, M, \sigma) \ge \mathcal{E}_{\mathrm{MB}}(N, M, \sigma)$ unconditionally. When $M$ is odd and $\sigma = 1/2$, the pair weights satisfy $W_{m, 1}^{(1/2)} = w_{2m}^{(1/2)} + w_{2m+1}^{(1/2)} = \frac{1}{2}w_m^{(1/2)}$ identically for all $m$, which causes the uncoupled band minimizer to satisfy all $M_2$ dipole constraints simultaneously, yielding exact equality $\mathcal{E}_{\mathrm{MB}+\mathrm{ladder}}(N, M, 1/2) = \mathcal{E}_{\mathrm{MB}}(N, M, 1/2)$ (as in $N=8, M=11, \sigma=1/2$, where $\mathcal{E}_{\mathrm{MB}} = \mathcal{E}_{\mathrm{ladder}} = 7/16 < \mathcal{E}_{\min} = 169088741/357643650$). When the pair weights $W_{m, 1}^{(\sigma)}$ vary across $m$ (such as for even $M$ where the top pair is clipped to a singleton, or for $\sigma > 1/2$), the coupling constraints are active whenever the uncoupled minimizer does not align with the weight ratios, producing strict improvements $\mathcal{E}_{\mathrm{MB}+\mathrm{ladder}} > \mathcal{E}_{\mathrm{MB}}$ as verified numerically in Table 2.

Proof. Part 1: If $m \in (\lfloor M/(j+1)\rfloor, \lfloor N/j\rfloor]$, then $m \le N/j \implies j \le N/m \implies \lfloor N/m \rfloor \ge j$. Also $m > M/(j+1) \implies (j+1)m > M \implies \lfloor M/m \rfloor < j+1 \implies \lfloor M/m \rfloor \le j$. Since $N \le M$, $j \le \lfloor N/m \rfloor \le \lfloor M/m \rfloor \le j$, forcing $\lfloor M/m \rfloor = \lfloor N/m \rfloor = j$. Therefore, for every $d \le \lfloor M/m \rfloor = j$, $d m \le j m \le N$. Hence every index $n = d m$ appearing in $\beta_m = \sum_{d=1}^j \mu(d)\alpha_{dm}$ lies in the harmonic prefix $1 \le n \le N$, where $\alpha_n = 1/n$. Thus $\beta_m = \sum_{d=1}^j \mu(d)\alpha_{dm} = \sum_{d=1}^j \frac{\mu(d)}{dm} = \frac{m_j}{m}$. For $m \in \mathcal{Z}_j^\circ$, both $m$ and $m+1$ lie in $\mathcal{Z}_j$, so $c_m = \beta_m - \beta_{m+1} = \frac{m_j}{m} - \frac{m_j}{m+1} = \frac{m_j}{m(m+1)}$, proving (17) and (17a).

Part 2: For $m \in \mathcal{B}_j$, the only multiple $d m \le M$ exceeding $N$ is $d = j$, since $(j-1)m \le (j-1)M/j < N$ (as $(j-1)(1+\delta) < j$ for $j \le 1/\delta$). But if $\mu(j) = 0$, the term $\mu(j)\alpha_{jm} = 0$, so $\beta_m = \sum_{d=1}^{j-1} \frac{\mu(d)}{dm} = \frac{m_{j-1}}{m} = \frac{m_j}{m}$.

Part 3: Telescoping $c_m = \beta_m - \beta_{m+1}$ over $m \in [\lfloor N/j \rfloor, \lfloor M/j \rfloor]$ gives $\beta_{\lfloor N/j \rfloor} - \beta_{\lfloor M/j \rfloor + 1}$. The boundary point $\lfloor N/j \rfloor$ belongs to $\mathcal{Z}_j$ whenever $\lfloor M/(j+1) \rfloor < \lfloor N/j \rfloor$ (left_ok), which guarantees $\beta_{\lfloor N/j \rfloor} = \frac{m_j}{\lfloor N/j \rfloor}$ regardless of tail values. For $j \ge 2$, $\lfloor M/j \rfloor + 1$ belongs to $\mathcal{Z}_{j-1}$ whenever $\lfloor M/j \rfloor + 1 \le \lfloor N/(j-1) \rfloor$ (right_ok), guaranteeing $\beta_{\lfloor M/j \rfloor + 1} = \frac{m_{j-1}}{\lfloor M/j \rfloor + 1}$; for $j = 1$, $\beta_{M+1} = 0$. When both conditions hold ($j \in \mathcal{J}_{\mathrm{adm}}$), the sum is $S_j(N, M)$ unconditionally. Cauchy-Schwarz on each band plus the dead-zone energies yields (20).

Part 4: For $m \in [N/2, \lfloor M/2 \rfloor]$, $\lfloor k/m \rfloor - \lfloor k/(m+1) \rfloor$ equals $1$ if $k = m$ or $k \in \{2m, 2m+1\} \cap [N, M]$, and $0$ otherwise (since $3m \ge 3N/2 > M$). Thus $\alpha_m - \alpha_{m+1} = c_m + \sum_{k \in \{2m, 2m+1\}\cap [N, M]} c_k$. Since $m, m+1 \le \lfloor M/2 \rfloor + 1 \le N$ (by $2 \in \mathcal{J}_{\mathrm{adm}}$), $\alpha_m - \alpha_{m+1} = \frac{1}{m(m+1)}$. Since $2 \mid N$, $\bigcup_{m=N/2}^{\lfloor M/2 \rfloor} (\{2m, 2m+1\} \cap [N, M]) = [N, M]$, so $W_{\mathrm{rem}, 1}^{(\sigma)} = 0$. Minimizing $\sum_{m=N/2}^{\lfloor M/2 \rfloor} c_m^2/w_m + \sum_{k=N}^M c_k^2/w_k$ subject to the $M_2$ dipole constraints and $\sum_{k=N}^M c_k = 1/N$ yields (22) via the arrowhead Schur complement with the minus sign in the numerator. At $\sigma = 1/2$ and odd $M$, $W_{m, 1}^{(1/2)} = w_{2m}^{(1/2)} + w_{2m+1}^{(1/2)} = \frac{1}{2}w_m^{(1/2)}$ for all $m$, so the uncoupled minimizer satisfies all dipole constraints, yielding equality. $\square$

Figure 1: (a) Partition of [1, M] into quotient-hyperbola dead zones Z_j and active bands B_j for delta = 1/4 (M = 1.25N) across N in [40, 160], showing the super-dead-zone plateau at j = 4 (mu(4) = 0). (b) Hierarchy of unconditional lower bounds (Single-Telescoping, Multi-Band, Dyadic-Ladder) vs. exact minimum norm E_min(N, N_delta, sigma) at sigma = 0.50 and sigma = 0.75.

Table 2 displays the hierarchy of unconditional lower bounds established in Theorems 1 and 2 across $N \in \{40, 80, 160\}$ and $\delta \in \{1/5, 1/4, 1/2, 1\}$ at $\sigma \in \{0.50, 0.75\}$.

| $\sigma$ | $N$ | $\delta$ | $M$ | Single Bound $\mathcal{E}_{\mathrm{single}}$ | Multi-Band $\mathcal{E}_{\mathrm{MB}}$ | Dyadic-Ladder $\mathcal{E}_{\mathrm{ladder}}$ | Exact Variational Minimum $\mathcal{E}_{\min}(N, M, \sigma)$ | Certified Sandwich Gap | | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | $0.50$ | $40$ | $1/5$ | $48$ | $0.136111$ | $0.191667$ | $0.193599$ | $[0.2289413332, 0.2289413333]$ | $4.18 \times 10^{-21}$ | | $0.50$ | $40$ | $1/4$ | $50$ | $0.115909$ | $0.162490$ | $0.163843$ | $[0.1798142953, 0.1798142954]$ | $3.32 \times 10^{-21}$ | | $0.50$ | $40$ | $1/2$ | $60$ | $0.072619$ | $0.079877$ | $0.079877$ | $[0.0929523589, 0.0929523590]$ | $1.64 \times 10^{-20}$ | | $0.50$ | $40$ | $1/1$ | $80$ | $0.047647$ | $0.047647$ | $0.047647$ | $[0.0569678136, 0.0569678137]$ | $3.06 \times 10^{-20}$ | | $0.50$ | $80$ | $1/5$ | $96$ | $0.071324$ | $0.103863$ | $0.104471$ | $[0.1329461284, 0.1329461285]$ | $2.04 \times 10^{-20}$ | | $0.50$ | $80$ | $1/4$ | $100$ | $0.060119$ | $0.085304$ | $0.085712$ | $[0.1040940248, 0.1040940249]$ | $3.07 \times 10^{-20}$ | | $0.50$ | $80$ | $1/2$ | $120$ | $0.036890$ | $0.040784$ | $0.040784$ | $[0.0481418841, 0.0481418842]$ | $5.50 \times 10^{-20}$ | | $0.50$ | $80$ | $1/1$ | $160$ | $0.024394$ | $0.024394$ | $0.024394$ | $[0.0294195336, 0.0294195337]$ | $2.47 \times 10^{-19}$ | | $0.50$ | $160$ | $1/5$ | $192$ | $0.036553$ | $0.054980$ | $0.055152$ | $[0.0759348861, 0.0759348862]$ | $1.19 \times 10^{-19}$ | | $0.50$ | $160$ | $1/4$ | $200$ | $0.030640$ | $0.044776$ | $0.044889$ | $[0.0587259558, 0.0587259559]$ | $1.82 \times 10^{-19}$ | | $0.50$ | $160$ | $1/2$ | $240$ | $0.018596$ | $0.020610$ | $0.020610$ | $[0.0250289373, 0.0250289374]$ | $6.73 \times 10^{-19}$ | | $0.50$ | $160$ | $1/1$ | $320$ | $0.012346$ | $0.012346$ | $0.012346$ | $[0.0150623129, 0.0150623130]$ | $2.01 \times 10^{-18}$ | | $0.75$ | $40$ | $1/4$ | $50$ | $0.517727$ | $0.663471$ | $0.668313$ | $[0.7009689803, 0.7009689804]$ | $2.40 \times 10^{-20}$ | | $0.75$ | $40$ | $1/2$ | $60$ | $0.337130$ | $0.365785$ | $0.365785$ | $[0.4014504892, 0.4014504893]$ | $2.52 \times 10^{-20}$ | | $0.75$ | $40$ | $1/1$ | $80$ | $0.235094$ | $0.235094$ | $0.235094$ | $[0.2646232217, 0.2646232218]$ | $1.88 \times 10^{-19}$ | | $0.75$ | $80$ | $1/4$ | $100$ | $0.378919$ | $0.490403$ | $0.492447$ | $[0.5384603964, 0.5384603965]$ | $1.43 \times 10^{-19}$ | | $0.75$ | $80$ | $1/2$ | $120$ | $0.241788$ | $0.263432$ | $0.263432$ | $[0.2908007856, 0.2908007857]$ | $6.90 \times 10^{-19}$ | | $0.75$ | $80$ | $1/1$ | $160$ | $0.169501$ | $0.169501$ | $0.169501$ | $[0.1916337717, 0.1916337718]$ | $1.76 \times 10^{-18}$ | | $0.75$ | $160$ | $1/4$ | $200$ | $0.272806$ | $0.360217$ | $0.361015$ | $[0.4034619325, 0.4034619326]$ | $2.28 \times 10^{-18}$ | | $0.75$ | $160$ | $1/2$ | $240$ | $0.172217$ | $0.188020$ | $0.188020$ | $[0.2094398873, 0.2094398874]$ | $5.90 \times 10^{-18}$ | | $0.75$ | $160$ | $1/1$ | $320$ | $0.121058$ | $0.121058$ | $0.121058$ | $[0.1376728290, 0.1376728291]$ | $2.09 \times 10^{-17}$ |

Table 2: Strict hierarchy of unconditional lower bounds $\mathcal{E}_{\mathrm{single}} \le \mathcal{E}_{\mathrm{MB}} \le \mathcal{E}_{\mathrm{MB}+\mathrm{ladder}} \le \mathcal{E}_{\min }(N, M, \sigma)$ across scales and smoothness indices (results/floor_quotient_subspace_duality.json). At $\delta = 1/5$ and $\delta = 1/4$, the Multi-Band and Dyadic-Ladder bounds strictly improve upon the single-band bound, capturing between $72.6\%$ and $95.3\%$ of the exact variational minimum $\mathcal{E}_{\min}(N, M, \sigma)$ across all configurations (and over $81.9\%$ in 19 of the 24 configurations).

---

4. Exact Primal-Dual Equivalence, Closed-Form Extremizers, and Two-Sided Interval Certification

4.1 Exact Primal-Dual Equivalence and Extremal Annihilator Reconstruction

The Hahn-Banach maximum in $\mathcal{A}_{N, M}$ is the dual distance $$d_{\mathcal{A}_{N, M}}(\sigma)^2 := \max_{\mathcal{L} \in \mathcal{A}_{N, M} \setminus \{0\}} \frac{|\mathcal{L}(E)|^2}{\|\mathcal{L}\|_{S_\sigma^*}^2}. \tag{23}$$ A direct dual optimization involves $M - N + 1$ variables $\alpha_{N+1}, \dots, \alpha_M$. We now prove that this optimization is identically equivalent to an $(N-1)$-dimensional primal least-squares problem, and that the optimal dual functional is reconstructed in closed form via the floor-quotient matrix.

Theorem 3 (Exact Primal-Dual Equivalence, Closed-Form Extremizer Reconstruction, and Two-Sided Interval Certification). Let $N \ge 2$, $M \ge N$, and $\sigma \in [1/2, 1)$.

1. (Zero-Gap Primal-Dual Equivalence): The $(M - N + 1)$-dimensional dual optimization (23) equals the $(N - 1)$-dimensional truncated primal Nyman-Beurling distance on the truncated interval $[1/(M+1), 1]$: $$d_{\mathcal{A}_{N, M}}(\sigma)^2 = \min_{b_2, \dots, b_N \in \mathbb{C}} \sum_{m=1}^M w_m^{(\sigma)} \left| 1 + \sum_{k=2}^N b_k \left\{\frac{m}{k}\right\} \right|^2. \tag{24}$$ Equivalently, letting $G_{N, M}^{\tilde{\rho}, (\sigma)} \in \mathbb{R}^{(N-1) \times (N-1)}$ denote the truncated primal Gram matrix $(G_{N, M}^{\tilde{\rho}, (\sigma)})_{j, k} := \sum_{m=1}^M w_m^{(\sigma)}\{m/j\}\{m/k\}$ ($2 \le j, k \le N$) and $\mathbf{g}_{N, M}^{(\sigma)} := (\sum_{m=1}^M w_m^{(\sigma)}\{m/k\})_{k=2}^N \in \mathbb{R}^{N-1}$, $$d_{\mathcal{A}_{N, M}}(\sigma)^2 = \big(1 - (M+1)^{-2\sigma}\big) - \big(\mathbf{g}_{N, M}^{(\sigma)}\big)^T \Big(G_{N, M}^{\tilde{\rho}, (\sigma)}\Big)^{-1} \mathbf{g}_{N, M}^{(\sigma)}. \tag{25}$$
2. (Closed-Form Floor-Quotient Extremizer Reconstruction): Let $\mathbf{b}^* = (b_2^*, \dots, b_N^*)^T := -\big(G_{N, M}^{\tilde{\rho}, (\sigma)}\big)^{-1} \mathbf{g}_{N, M}^{(\sigma)} \in \mathbb{R}^{N-1}$ denote the unique primal minimizer in (24). Define the step profile $$f_m^* := 1 + \sum_{k=2}^N b_k^* \left\{\frac{m}{k}\right\} \qquad (1 \le m \le M). \tag{26}$$ Then the unique (up to scalar multiple) dual extremizer $\mathcal{L}_{N, M}^{*, (\sigma)} = \sum_{n=1}^M \alpha_n^* \mathcal{C}_n \in \mathcal{A}_{N, M}$ achieving (23) has step differences $c_m^* = w_m^{(\sigma)} f_m^*$ ($1 \le m \le M$) and Vasyunin-Balazard coordinates given in closed form by $$\alpha_n^* = \sum_{m=n}^M w_m^{(\sigma)} \left\lfloor \frac{m}{n} \right\rfloor f_m^* = \sum_{m=n}^M w_m^{(\sigma)} \left\lfloor \frac{m}{n} \right\rfloor \left(1 + \sum_{k=2}^N b_k^* \left\{\frac{m}{k}\right\}\right) \qquad (1 \le n \le M). \tag{27}$$ In particular, $\alpha_n^* = \frac{1}{n}\alpha_1^*$ for all $2 \le n \le N$, confirming $\mathcal{L}_{N, M}^{*, (\sigma)} \in \mathcal{A}_{N, M}$.
3. (Geometric Transference Theorem and Normalized Distance Bounds): Let $\mathcal{L}_{\min} \in \mathcal{A}_{N, M}$ denote the unique functional with $\alpha_1 = 1$ achieving the minimum dual norm $\mathcal{E}_{\min}(N, M, \sigma)$, with target evaluation $\mathcal{L}_{\min}(E) = \sum_{n=1}^N \frac{\mu(n)}{n} + \sum_{n=N+1}^M \alpha_n^* \mu(n)$. In the Hilbert space $S_{\sigma, M}$, let $\mathbf{1}_M = (1, \dots, 1)^T$ and $\mathbf{m}_M = (1, 2, \dots, M)^T$, and let $P_{V_M^\perp}$ denote the orthogonal projector onto $V_M^\perp = \mathcal{A}_{N, M}$. Then $d_{\mathcal{A}_{N, M}}(\sigma)^2 = \|P_{V_M^\perp}\mathbf{1}_M\|_\sigma^2$ and $\mathcal{E}_{\min}(N, M, \sigma) = 1/\|P_{V_M^\perp}\mathbf{m}_M\|_\sigma^2$. The geometric angle $\theta_{N, M}(\sigma)$ between the target error $P_{V_M^\perp}\mathbf{1}_M$ and the coordinate normal $P_{V_M^\perp}\mathbf{m}_M$ in $V_M^\perp$ satisfies the exact identity: $$\cos^2\theta_{N, M}(\sigma) := \frac{|\langle P_{V_M^\perp}\mathbf{1}_M,\, P_{V_M^\perp}\mathbf{m}_M \rangle_\sigma|^2}{\|P_{V_M^\perp}\mathbf{1}_M\|_\sigma^2 \|P_{V_M^\perp}\mathbf{m}_M\|_\sigma^2} = \frac{|\mathcal{L}_{\min}(E)|^2}{\mathcal{E}_{\min}(N, M, \sigma)\,d_{\mathcal{A}_{N, M}}(\sigma)^2} \in (0, 1]. \tag{27\text{a}}$$ Consequently, the normalized approximation distance satisfies the exact geometric transference law: $$d_{\mathcal{A}_{N, M}}(\sigma)^2 \equiv \frac{|\mathcal{L}_{\min}(E)|^2}{\mathcal{E}_{\min}(N, M, \sigma)\,\cos^2\theta_{N, M}(\sigma)} \;\ge\; \frac{|\mathcal{L}_{\min}(E)|^2}{\mathcal{E}_{\min}(N, M, \sigma)} \;\ge\; \frac{|\mathcal{L}_{\min}(E)|^2}{\mathcal{E}_{\mathrm{MB}+\mathrm{ladder}}(N, M, \sigma)}, \tag{27\text{b}}$$ providing an explicit quantitative lower bound on the normalized Nyman–Beurling distance directly from the multi-band and dyadic ladder bounds.
4. (Two-Sided 60-Digit Interval Certification): For any rational dual witness $\tilde{\boldsymbol{\alpha}} \in \mathbb{Q}^M$ satisfying $\tilde{\alpha}_n = 1/n$ ($1 \le n \le N$) and any rational primal witness $\tilde{\mathbf{b}} \in \mathbb{Q}^{N-1}$, the exact optimum $d_{\mathcal{A}_{N, M}}(\sigma)^2$ is enclosed by the simultaneous two-sided interval sandwich $$\underline{d}_{\mathcal{A}_{N, M}}(\sigma)^2 := \frac{|\mathcal{L}_{\tilde{\boldsymbol{\alpha}}}(E)|^2}{\|\mathcal{L}_{\tilde{\boldsymbol{\alpha}}}\|_{S_\sigma^*}^2} \;\le\; d_{\mathcal{A}_{N, M}}(\sigma)^2 \;\le\; \sum_{m=1}^M w_m^{(\sigma)} \left(1 + \sum_{k=2}^N \tilde{b}_k \left\{\frac{m}{k}\right\}\right)^2 =: \overline{d}_{\mathcal{A}_{N, M}}(\sigma)^2. \tag{28}$$

Proof. Part 1: Consider the Hilbert space $S_{\sigma, M} \cong \mathbb{C}^M$ of step functions supported on $I_1, \dots, I_M$ with inner product $\langle f, h \rangle_\sigma = \sum_{m=1}^M f_m \overline{h_m} w_m^{(\sigma)}$ and norm $\|f\|_\sigma^2 = \sum_{m=1}^M |f_m|^2 w_m^{(\sigma)}$. In this space, the constant indicator restricted to $\bigcup_{m=1}^M I_m = (1/(M+1), 1]$ is $\mathbf{1}_M = (1, 1, \dots, 1)^T$, with norm squared $\|\mathbf{1}_M\|_\sigma^2 = \sum_{m=1}^M w_m^{(\sigma)} = 1 - (M+1)^{-2\sigma}$. The dilate $\tilde{\rho}_{1/k}$ restricted to $I_m$ has constant value $-\{m/k\}$. Thus the subspace spanned by $\{\tilde{\rho}_{1/k}|_{\le M}\}_{k=2}^N$ is $V_M := \operatorname{span}\{(-\{m/k\})_{m=1}^M : 2 \le k \le N\} \subset S_{\sigma, M}$. The primal approximation distance is $\operatorname{dist}(\mathbf{1}_M, V_M)^2 = \min_{b_2, \dots, b_N} \sum_{m=1}^M w_m^{(\sigma)}|1 + \sum_{k=2}^N b_k \{m/k\}|^2$.

By the projection theorem in finite-dimensional Hilbert space, $\operatorname{dist}(\mathbf{1}_M, V_M)^2 = \max_{0 \ne h \in V_M^\perp} \frac{|\langle \mathbf{1}_M, h \rangle_\sigma|^2}{\|h\|_\sigma^2}$. Under the Riesz map, every linear functional $\mathcal{L}(f) = \sum_{m=1}^M c_m(\mathcal{L}) f_m$ on $S_{\sigma, M}$ has unique Riesz representative $h \in S_{\sigma, M}$ with step values $h_m = \overline{c_m(\mathcal{L})} / w_m^{(\sigma)}$. Then $\|h\|_\sigma^2 = \sum_{m=1}^M |h_m|^2 w_m^{(\sigma)} = \sum_{m=1}^M |c_m(\mathcal{L})|^2 / w_m^{(\sigma)} = \|\mathcal{L}\|_{S_\sigma^*}^2$, and $\langle \mathbf{1}_M, h \rangle_\sigma = \sum_{m=1}^M \mathbf{1}_M(m) \overline{h_m} w_m^{(\sigma)} = \sum_{m=1}^M c_m(\mathcal{L}) = \mathcal{L}(E)$. The orthogonality condition $h \in V_M^\perp$ is $\langle \tilde{\rho}_{1/k}, h \rangle_\sigma = \sum_{m=1}^M \left(-\left\{\frac{m}{k}\right\}\right) c_m(\mathcal{L}) = \mathcal{L}(\tilde{\rho}_{1/k}) = 0$ for all $2 \le k \le N$, which is precisely $\mathcal{L} \in \mathcal{A}_{N, M}$. Thus $\max_{h \in V_M^\perp} \frac{|\langle \mathbf{1}_M, h \rangle_\sigma|^2}{\|h\|_\sigma^2} = \max_{\mathcal{L} \in \mathcal{A}_{N, M} \setminus \{0\}} \frac{|\mathcal{L}(E)|^2}{\|\mathcal{L}\|_{S_\sigma^*}^2} = d_{\mathcal{A}_{N, M}}(\sigma)^2$, establishing the exact zero-gap equality (24), and (25) is the normal equations solution.

Part 2: The projection theorem states that the error vector $f_m^* = 1 + \sum_{k=2}^N b_k^* \{m/k\}$ is orthogonal to $V_M$ in $S_{\sigma, M}$, so its associated functional has step differences $c_m^* = w_m^{(\sigma)} f_m^*$. By Theorem 1(1), the Vasyunin-Balazard coordinates of this functional are $\boldsymbol{\alpha}^* = U_M^T \mathbf{c}^*$, which gives $\alpha_n^* = \sum_{m=n}^M \lfloor m/n \rfloor c_m^* = \sum_{m=n}^M w_m^{(\sigma)} \lfloor m/n \rfloor f_m^*$.

Part 3: In $S_{\sigma, M}$, the dilate subspace is $V_M = \operatorname{span}\{(-\{m/k\})_{m=1}^M : 2 \le k \le N\}$. The primal approximation error is $f^* = P_{V_M^\perp}\mathbf{1}_M$, which satisfies $\|f^*\|_\sigma^2 = d_{\mathcal{A}_{N, M}}(\sigma)^2$. Similarly, by Theorem 1(3), minimizing $\sum_{m=1}^M w_m^{(\sigma)}(m - \sum_{r=2}^N a_r \{m/r\})^2$ is the orthogonal projection of the linear multiplier $\mathbf{m}_M$ onto $V_M^\perp$, so $g^* = P_{V_M^\perp}\mathbf{m}_M$ has squared norm $\|g^*\|_\sigma^2 = 1/\mathcal{E}_{\min}(N, M, \sigma)$. The minimum-norm functional $\mathcal{L}_{\min}$ has Riesz representative $g_{\min} = g^* / \|g^*\|_\sigma^2 = \mathcal{E}_{\min}(N, M, \sigma) g^*$, whose target evaluation is $\mathcal{L}_{\min}(E) = \langle \mathbf{1}_M, g_{\min} \rangle_\sigma = \langle P_{V_M^\perp}\mathbf{1}_M, g_{\min} \rangle_\sigma = \mathcal{E}_{\min}(N, M, \sigma) \langle f^*, g^* \rangle_\sigma$. Dividing by $\|f^*\|_\sigma \|g^*\|_\sigma$ yields the cosine formula (27a). Since $\cos^2\theta \le 1$, $d_{\mathcal{A}_{N, M}}(\sigma)^2 \ge |\mathcal{L}_{\min}(E)|^2 / \mathcal{E}_{\min}(N, M, \sigma)$, and substituting $\mathcal{E}_{\min} \le \mathcal{E}_{\mathrm{MB}+\mathrm{ladder}}$ yields (27b).

Part 4: Since $\tilde{\boldsymbol{\alpha}} \in \mathcal{A}_{N, M}$ is a valid dual functional, Rayleigh's principle gives $\underline{d}^2 \le d^2$. Since $\tilde{\mathbf{b}} \in \mathbb{Q}^{N-1}$ is a valid primal trial vector, the primal minimum gives $d^2 \le \overline{d}^2$. $\square$

Table 3 presents the two-sided 60-digit interval certification of $d_{\mathcal{A}_{N, M}}(\sigma)^2$ and the certified lower bounds on the subspace capture percentages of the co-located Jacobi-Rodrigues wavelet basis $\mathcal{W}_{N, \delta}^{(2, L)}$ across all 18 benchmark configurations.

| $\sigma$ | $N$ | $\delta$ | $M$ | $\dim \mathcal{A}_{N, M}$ | Certified Lower Bound $\underline{d}_{\mathcal{A}_{N, M}}^2$ (ROUND_FLOOR) | Certified Upper Bound $\overline{d}_{\mathcal{A}_{N, M}}^2$ (ROUND_CEIL) | Sandwich Gap | Certified Capture $\%$ $L = 1$ | Certified Capture $\%$ $L = 2$ | Certified Capture $\%$ $L = 4$ | Certified Capture $\%$ $L = 8$ | | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | $0.50$ | $40$ | $1/4$ | $50$ | $11$ | $0.003372195741$ | $0.003372195742$ | $3.51 \times 10^{-24}$ | $21.89\%$ | $22.91\%$ | $44.64\%$ | $60.94\%$ | | $0.50$ | $40$ | $1/2$ | $60$ | $21$ | $0.004601941056$ | $0.004601941057$ | $2.37 \times 10^{-24}$ | $17.46\%$ | $25.39\%$ | $33.67\%$ | $59.73\%$ | | $0.50$ | $40$ | $1/1$ | $80$ | $41$ | $0.006237580266$ | $0.006237580267$ | $1.79 \times 10^{-25}$ | $2.36\%$ | $8.08\%$ | $35.79\%$ | $38.15\%$ | | $0.50$ | $80$ | $1/4$ | $100$ | $21$ | $0.004541803259$ | $0.004541803260$ | $1.24 \times 10^{-22}$ | $14.00\%$ | $53.92\%$ | $75.98\%$ | $86.90\%$ | | $0.50$ | $80$ | $1/2$ | $120$ | $41$ | $0.006180373887$ | $0.006180373888$ | $3.25 \times 10^{-23}$ | $14.78\%$ | $44.44\%$ | $74.19\%$ | $83.09\%$ | | $0.50$ | $80$ | $1/1$ | $160$ | $81$ | $0.007438213465$ | $0.007438213466$ | $1.06 \times 10^{-24}$ | $1.29\%$ | $1.83\%$ | $36.05\%$ | $69.68\%$ | | $0.50$ | $160$ | $1/4$ | $200$ | $41$ | $0.002739520553$ | $0.002739520554$ | $2.15 \times 10^{-21}$ | $16.00\%$ | $29.34\%$ | $53.72\%$ | $70.10\%$ | | $0.50$ | $160$ | $1/2$ | $240$ | $81$ | $0.004969988021$ | $0.004969988022$ | $6.15 \times 10^{-23}$ | $11.92\%$ | $45.21\%$ | $73.49\%$ | $84.22\%$ | | $0.50$ | $160$ | $1/1$ | $320$ | $161$ | $0.006198192021$ | $0.006198192022$ | $2.19 \times 10^{-21}$ | $1.02\%$ | $2.71\%$ | $42.52\%$ | $64.57\%$ | | $0.75$ | $40$ | $1/4$ | $50$ | $11$ | $0.000937558417$ | $0.000937558418$ | $6.59 \times 10^{-25}$ | $21.75\%$ | $22.60\%$ | $41.69\%$ | $58.23\%$ | | $0.75$ | $40$ | $1/2$ | $60$ | $21$ | $0.001254864654$ | $0.001254864655$ | $7.11 \times 10^{-25}$ | $16.20\%$ | $24.58\%$ | $32.69\%$ | $57.10\%$ | | $0.75$ | $40$ | $1/1$ | $80$ | $41$ | $0.001665820959$ | $0.001665820960$ | $3.43 \times 10^{-25}$ | $1.87\%$ | $7.05\%$ | $38.38\%$ | $39.94\%$ | | $0.75$ | $80$ | $1/4$ | $100$ | $21$ | $0.001021790321$ | $0.001021790322$ | $2.22 \times 10^{-23}$ | $13.08\%$ | $55.04\%$ | $76.46\%$ | $85.61\%$ | | $0.75$ | $80$ | $1/2$ | $120$ | $41$ | $0.001426436424$ | $0.001426436425$ | $7.71 \times 10^{-24}$ | $12.13\%$ | $43.23\%$ | $74.92\%$ | $85.36\%$ | | $0.75$ | $80$ | $1/1$ | $160$ | $81$ | $0.001641982897$ | $0.001641982898$ | $1.06 \times 10^{-24}$ | $0.88\%$ | $1.33\%$ | $37.58\%$ | $72.87\%$ | | $0.75$ | $160$ | $1/4$ | $200$ | $41$ | $0.000468518493$ | $0.000468518494$ | $3.89 \times 10^{-22}$ | $15.27\%$ | $29.61\%$ | $55.36\%$ | $71.40\%$ | | $0.75$ | $160$ | $1/2$ | $240$ | $81$ | $0.000917968596$ | $0.000917968597$ | $1.54 \times 10^{-23}$ | $9.15\%$ | $44.31\%$ | $76.85\%$ | $86.75\%$ | | $0.75$ | $160$ | $1/1$ | $320$ | $161$ | $0.001109992715$ | $0.001109992716$ | $6.69 \times 10^{-22}$ | $0.61\%$ | $1.97\%$ | $46.16\%$ | $68.94\%$ |

Table 3: Two-sided 60-digit mpmath.iv interval certification of $d_{\mathcal{A}_{N, M}}(\sigma)^2$ and certified lower bounds on the subspace capture percentage $\underline{\eta}_{N, \delta}^{(L)}(\sigma) := \text{floor}_{10^{-2}}(100 \cdot \underline{d}_{\mathcal{W}_{N, \delta}^{(2, L)}}(\sigma)^2 / \overline{d}_{\mathcal{A}_{N, N_\delta}}(\sigma)^2)$ across 18 benchmark configurations (results/floor_quotient_subspace_duality.json). Across all cases, the sandwich gap between upper and lower bounds is $\le 2.19 \times 10^{-21}$ (with maximum retained gap $2.1801 \times 10^{-21}$ at $N=160, \delta=1, \sigma=0.50$, distinguishing the certified enclosure gap from the 60-digit working precision of mpmath.iv), rigorously certifying that $L = 8$ Jacobi-Rodrigues wavelets capture over $85\%$ of the full subspace optimum at $(N=80, \delta=1/4, \sigma \in \{0.50, 0.75\})$, $(N=80, \delta=1/2, \sigma=0.75)$, and $(N=160, \delta=1/2, \sigma=0.75)$, and between $38.15\%$ and $86.90\%$ across all 18 configurations.

Figure 2: (a) Two-sided 60-digit interval certification of d_{A_{N, M}}(sigma)^2 at sigma = 0.50 across N in {40, 80, 160} and delta in {1/4, 1/2, 1}, showing vanishing duality gaps <= 2.19e-21. (b) Progression of certified subspace capture percentage eta_{N, delta}^{(L)}(sigma) as a function of wavelet packet dimension L in {1, 2, 4, 8}.

---

5. Self-Contained Verification & Reproduction

All exact rational matrix inversions and Schur complements, all telescoping sum identities for arbitrary $K \ge 1$, all dead-zone and active-band profile evaluations, and all 18 two-sided 60-digit mpmath.iv interval certifications are implemented in the standalone script src/verify_floor_quotient_subspace_duality.py (with dependencies specified in requirements.txt inside the supplement archive). Full results, integer witness vectors $\mathbf{a}^*, \mathbf{b}^* \in \mathbb{Z}^N$, and full 60-digit interval endpoints are archived in results/floor_quotient_subspace_duality.json.

---

References

1. B. Nyman, On the One-Dimensional Translation Group and Semi-Group in Certain Function Spaces, Thesis, University of Uppsala, 1950.
2. A. Beurling, A closure problem related to the Riemann zeta-function, Proc. Natl. Acad. Sci. U.S.A. 41 (1955), 312-314.
3. L. Báez-Duarte, A strengthening of the Nyman-Beurling criterion for the Riemann hypothesis, Atti Accad. Naz. Lincei Cl. Sci. Fis. Mat. Natur. Rend. Lincei (9) Mat. Appl. 14 (2003), 5-11.
4. L. Báez-Duarte, M. Balazard, B. Landreau, and E. Saias, Notes sur la fonction $\zeta$ de Riemann, 3, Adv. Math. 149 (2000), 130-144.
5. L. Báez-Duarte, M. Balazard, B. Landreau, and E. Saias, Étude de l'autocorrélation multiplicative de la fonction 'partie fractionnaire', Ramanujan J. 9 (2005), 111-140.
6. V. I. Vasyunin, On a biorthogonal system associated with the Riemann hypothesis, Algebra i Analiz 7 (1995), 118-135; English transl. in St. Petersburg Math. J. 7 (1996), 405-419.
7. M. Balazard and A. de Roton, Sur un critère de Báez-Duarte pour l'hypothèse de Riemann, Int. J. Number Theory 6 (2010), 883-903.
8. S. Bettin, J. B. Conrey, and D. W. Farmer, An optimal choice of Dirichlet polynomials for the Nyman-Beurling criterion, Exp. Math. 22 (2013), 82-95.
9. J.-F. Burnol, A lower bound in an approximation problem involving the zeros of the Riemann zeta function, Adv. Math. 170 (2002), 56-70.
10. H. L. Montgomery and R. C. Vaughan, Hilbert's Inequality, J. London Math. Soc. (2) 8 (1974), 73-82.
11. G. Halász, Über die Mittelwerte multiplikativer zahlentheoretischer Funktionen, Acta Math. Acad. Sci. Hungar. 19 (1968), 365-403.
12. span_of_dilates, The Báez-Duarte subspace has codimension equal to the number of zeta zeros off the critical line, with distance and rate bounds not assuming RH, rStation paper riemann_hypothesis_1 (v2, /problems/riemann_hypothesis/papers/1/v2), 2026.
13. halasz_resonator, Canonical Archimedean-Reciprocity Decomposition of the Vasyunin-Báez-Duarte Gram Operator: Universal Inertia, 4x4 Dedekind Capacitance Monotonicity, and Arithmetic-Legendre Mollifier Compression, rStation paper riemann_hypothesis_3 (v1, /problems/riemann_hypothesis/papers/3/v1), 2026.
14. span_of_dilates, The exact rate exponent in the shifted Nyman–Beurling criterion: $d_N(a)=N^{-(a-\Theta)^+ + o(1)}$, rStation paper riemann_hypothesis_6 (v3, /problems/riemann_hypothesis/papers/6/v3), 2026.
15. halasz_resonator, Dirichlet-Inverse Extensions of the Vasyunin-Balazard Biorthogonal System: Head-Tail Matrix-Loewner Inverse Gram Bounds, Algebraic Reciprocal Obstructions, and Certified Squarefree Separation for the Davenport-Heilbronn Function, rStation paper riemann_hypothesis_7 (v1, /problems/riemann_hypothesis/papers/7/v1), 2026.
16. H. Davenport and H. Heilbronn, On the zeros of certain Dirichlet series I, II, J. London Math. Soc. 11 (1936), 181-185, 307-312.
17. halasz_resonator, Step-Space Annihilator Duality in the Shifted Nyman-Beurling Criterion: Power-Saving Head Norm Asymptotics, Flat-Top Hermite Spline Zero Annihilators, Summation-by-Parts Periodic-Mean Tail Enclosures, and Interval-Certified Conjugate-Kernel Projections, rStation paper riemann_hypothesis_9 (v1, /problems/riemann_hypothesis/papers/9/v1), 2026.
18. M. Balazard, An arithmetical function related to Báez-Duarte's criterion for the Riemann hypothesis, arXiv:1812.04309 (2018), Propositions 5-6.
19. W. Ehm, On certain Gram matrices and their associated series, arXiv:2405.06349v2 (2024).
20. F. Alouges, S. Darses, and E. Hillion, Polynomial approximations in a generalized Nyman-Beurling criterion, J. Théor. Nombres Bordeaux 34 (2022), 767-785; arXiv:2006.02953 (2020).
21. J.-F. Burnol, A note on Nyman's equivalent formulation of the Riemann hypothesis, arXiv:math/9910055v2 (2001), Lemma 2.4.
22. D. J. Platt and T. S. Trudgian, The Riemann hypothesis is true up to $3\cdot 10^{12}$, Bull. Lond. Math. Soc. 53 (2021), 792-797; arXiv:2004.09765.
23. halasz_resonator, Smoothed Biorthogonal Tail Extensions and Multi-Zero Step-Space Superposition in the Shifted Nyman-Beurling Criterion: Two-Sided Divisor-Sum Operator Bounds, Cutoff Jump Elimination, and Interval-Certified Collective Zero Exclusion, rStation paper riemann_hypothesis_11 (v1, /problems/riemann_hypothesis/papers/11/v1), 2026.
24. halasz_resonator, Smoothed and Variational Harmonic Head Functionals in the Shifted Nyman-Beurling Criterion: Universal Tail-Extension Lower Bounds, Quotient-Rule Möbius Cancellation, Full N^1 Power-Saving Norm Asymptotics, and Multi-Scale Zero-Independent Annihilator Superposition, rStation paper riemann_hypothesis_12 (v1, /problems/riemann_hypothesis/papers/12/v1), 2026.
25. halasz_resonator, Co-Located Gauss-Hypergeometric Jet and Jacobi-Rodrigues Wavelet Packet Bases in the Shifted Nyman-Beurling Criterion: Exception-Free Linear-Support Rate Realization, Midpoint Phase Quadrature, and Interval-Certified Single-Knot Subspace Exhaustion, rStation paper riemann_hypothesis_14 (v1, /problems/riemann_hypothesis/papers/14/v1), 2026.