--- title: Spectrally Optimal Dual Frames url: https://www.emergentmind.com/topics/spectrally-optimal-dual-frames type: topic --- # Spectrally Optimal Dual Frames Spectrally optimal dual frames are dual systems chosen so that reconstruction is extremal with respect to a spectral criterion. The literature considered here studies two closely related optimization problems: one minimizes the spectrum of the dual frame operator under submajorization and convex-potential constraints, and the other minimizes the spectral radius of erasure error operators in robust reconstruction. In both settings, the canonical dual \(F^\#=\{S_F^{-1}f_i\}\) is the reference dual, but its optimality and uniqueness depend on the ambient constraints, the geometry of the frame, and, for graph-generated frames, the connectedness of the underlying graph [1108.4412], [1405.4277], [2412.15709], [2406.00776]. ## 1. Finite-frame setting and canonical duality Let \(H\) be a finite-dimensional Hilbert space and \(F=\{f_i\}_{i=1}^N\subset H\) a frame. Its analysis operator is \[ T_F(f)=(\langle f,f_i\rangle)_{i=1}^N, \] its synthesis operator is \(T_F^*\), and its frame operator is \[ S_F=T_F^*T_F,\qquad S_F(f)=\sum_{i=1}^N \langle f,f_i\rangle f_i. \] Since \(F\) is a frame, \(S_F\) is positive and invertible. A sequence \(G=\{g_i\}_{i=1}^N\) is a dual of \(F\) if \[ f=\sum_{i=1}^N \langle f,f_i\rangle g_i =\sum_{i=1}^N \langle f,g_i\rangle f_i, \qquad \forall f\in H. \] The canonical dual is \[ F^\#=\{S_F^{-1}f_i\}_{i=1}^N. \] A general dual can be written as \[ g_i=S_F^{-1}f_i+u_i,\qquad i=1,\dots,N, \] where the perturbation satisfies \[ \sum_{i=1}^N \langle f,f_i\rangle u_i=0,\qquad \forall f\in H \] [1204.5062], [2411.00502]. This parametrization is the starting point for both spectral-design problems. In one direction, one studies the possible spectra of frame operators \(S_G=T_G^*T_G\) as \(G\) ranges over all duals. In the other, one studies erasure error operators built from the pair \((F,G)\) and asks which dual minimizes a worst-case or average spectral-radius criterion. ## 2. Submajorization, convex potentials, and water-filling constructions One major line of work defines spectral optimality through the eigenvalues of \(S_G\). For a fixed frame \(F\), the class \[ D_t(F)=\{G\in D(F): \operatorname{tr} S_G \ge t\} \] admits a precise additive model: every \(S_G\) can be written \[ S_G=S_{F^\#}+B,\qquad B\ge 0,\qquad \operatorname{rank} B\le N-\dim H. \] The optimization problem is then to find \(G_t\in D_t(F)\) such that \(S_{G_t}\) is minimal in submajorization among \(\{S_G:G\in D_t(F)\}\). The solution has a water-filling form: the smallest eigenvalues of \(S_{F^\#}\) are pushed up to a plateau level \(c\), determined by the trace constraint, while the larger eigenvalues remain fixed [1108.4412]. A related perturbative model imposes both a trace constraint and a proximity constraint to the canonical dual. For \[ \mathcal D_{t,\varepsilon}(F)=\Bigl\{\mathcal G\in\mathcal D(F): \sum_i \|g_i\|^2\ge t,\ \|T_{\mathcal G}-T_{F^\#}\|\le \varepsilon\Bigr\}, \] the frame operators satisfy \[ S_{\mathcal G}=S_{F^\#}+B \] with positivity, trace, operator-norm, and rank restrictions on \(B\). The optimal eigenvalue vector is again obtained by a truncated water-filling construction, and the resulting dual simultaneously minimizes every convex potential \[ P_h(\mathcal G)=\sum_{i=1}^d h(\lambda_i(S_{\mathcal G})) \] for increasing convex \(h\). The frame potential \(h(x)=x^2\) and the mean squared error \(h(x)=1/x\) are prominent special cases [1405.4277]. The same spectral principle extends to shift-generated oblique duals in finitely generated shift-invariant spaces. There the frame operators of all shift-generated \(V\)-duals form an affine model \[ U_V(A)=\{A+B:\ B\ge 0,\ B\ \text{shift-preserving},\ \operatorname{Ran} B\subseteq V,\ \operatorname{rk} B(x)\le n-d(x)\ \text{a.e.}\}, \] where \(A=S_{E(F)^\#}\). Fiberwise Fan–Pall interlacing describes exactly which measurable eigenvalue fields occur, and a non-commutative water-filling operator \(A_c\), with fiber eigenvalues \[ \lambda_j^*(x)=\max\{\lambda_j(A(x)),c\}, \] produces a unique optimal shift-generated dual \(E(G^{op})\) minimizing every non-decreasing convex potential under the prescribed norm budget \(w\) [1508.01739]. ## 3. Admissible spectra, tight duals, and conditioning A complementary spectral question asks which singular-value patterns are possible for dual synthesis operators. Let \(\Phi\) be the synthesis operator of a fixed frame, with singular values \[ \sigma_1\ge \cdots \ge \sigma_n>0, \] and let \(r=m-n\), where \(m\) is the frame length and \(n\) the ambient dimension. A sequence \[ \sigma_1'\ge \cdots \ge \sigma_n'>0 \] is the singular-value list of some dual frame if and only if it satisfies the interlacing inequalities \[ \sigma_i'\ge \frac{1}{\sigma_{n-i+1}},\qquad i=1,\dots,r, \] and \[ \frac{1}{\sigma_{n-i+1}}\le \sigma_i' \le \frac{1}{\sigma_{n-i+r+1}},\qquad i=r+1,\dots,n. \] Equivalently, the eigenvalues of \(S_\Psi\) are constrained relative to those of the canonical dual [1204.5062]. In this model, the canonical dual corresponds to the minimal admissible spectral profile: \[ \bigl(1/\sigma_n,\dots,1/\sigma_1\bigr). \] Other duals are obtained by adding free columns in an SVD-based parametrization, and this makes it possible to prescribe extremal spectra explicitly. In particular, tight duals are characterized completely. If \(m\ge 2n\), then for any \[ c\ge 1/\sigma_n \] there exists a \(c\)-tight dual. If \(m<2n\), tight duals exist only when the \(2n-m\) smallest singular values of \(\Phi\) coincide; when \(m=2n-1\), there is exactly one choice \(c=1/\sigma_n\) [1204.5062]. This spectral design viewpoint makes clear that the canonical dual is not generally optimal for conditioning. Tight duals have condition number \(1\), whereas the canonical dual may exhibit a larger singular-value spread when the original frame is ill-conditioned. A plausible implication is that “spectrally optimal” must always be interpreted relative to the chosen criterion: minimal singular-value spread, minimal convex potential, and minimal erasure error are distinct objectives, even though they often share the canonical dual in special regimes. ## 4. Erasure error operators and spectral-radius optimality A second major line of work defines spectral optimality through robustness to coefficient losses. For a frame \(F=\{f_i\}_{i=1}^N\subset \mathcal H_n\), a dual \(G=\{g_i\}_{i=1}^N\), and a diagonal erasure mask \(D\) with exactly \(m\) ones, the error operator is \[ E_{F,G,D}=T_G^*DT_F:\mathcal H_n\to \mathcal H_n, \] and its spectral radius is \[ \rho(E_{F,G,D})=\max\{|\lambda|:\lambda\in \sigma(E_{F,G,D})\}. \] The corresponding average error functional is \[ \mathrm{AE}_{\mathfrak R^{(m),p}}(F,G) = \Biggl( \frac{1}{\binom{N}{m}} \sum_{D\in \mathcal D^{(m)}} [\rho(E_{F,G,D})]^p \Biggr)^{1/p}. \] For one erasure, \[ \rho(E_{F,G,D_i})=|\langle f_i,g_i\rangle|, \qquad \mathrm{AE}_{\mathfrak R^{(1),p}}(F,G) = \Bigl(\tfrac1N\sum_{i=1}^N |\langle f_i,g_i\rangle|^p\Bigr)^{1/p}. \] If \(F\) is a 1-uniform frame, meaning \(\|f_i\|\) is constant in \(i\), then a dual \(G\) is 1-erasure spectrally optimal if and only if \[ \langle f_i,g_i\rangle=\frac{n}{N}\qquad \text{for all }i, \] so the canonical dual of a uniform tight frame is spectrally optimal for one erasure. A uniqueness result also holds under the hypotheses \[ H_1\cap H_2=\{0\}, \qquad \{f_i:i\in I_2\}\ \text{linearly independent}, \] where \[ L=\max_i \|S_F^{-1/2}f_i\|, \quad I_1=\{i:\|S_F^{-1/2}f_i\|=L\}, \quad I_2=\{1,\dots,N\}\setminus I_1, \] and \(H_j=\operatorname{span}\{f_i:i\in I_j\}\). Under these conditions the canonical dual is the unique minimizer of the 1-erasure average spectral-radius criterion [2508.07158]. When both the frame and its dual are allowed to vary, the terminology becomes more rigidly uniform. One source characterizes one-erasure spectral optimality by 1-uniformity of the dual pair and two-erasure spectral optimality by 2-uniformity. In the latter case the optimality condition requires \[ |\langle f_i,g_i\rangle|=\text{constant} \] and \[ \Re\bigl(\langle f_i,g_j\rangle\langle f_j,g_i\rangle\bigr)=\text{constant},\qquad i\neq j. \] The same work compares three measures of error—Frobenius norm, spectral radius, and numerical radius—and states that for Parseval or tight frames all three notions of one-erasure optimality agree, whereas in general they form strictly nested families: \[ \{\text{Frobenius}\}\subsetneq \{\text{Numerical}\}\subsetneq \{\text{Spectral}\} \] [2412.15709]. The cited literature also records a limitation: for \(m>1\), a complete characterization of spectral-radius optimizers for a fixed frame is not generally available, although lower bounds, uniform/Parseval cases, and several explicit constructions are known [2508.07158]. ## 5. Graph-generated frames and the role of connectedness For graph-generated frames, the spectral-radius problem becomes particularly explicit. Let \(G=(V,E)\) be a simple graph on \(n\) vertices with Laplacian \[ L(G)=D-A. \] Assume \(G\) has exactly \(k\) nonzero Laplacian eigenvalues \(\lambda_1,\dots,\lambda_k\), choose an orthonormal eigenbasis \[ L(G)=M\operatorname{diag}(\lambda_1,\dots,\lambda_k,0,\dots,0)M^*, \] let \(M_1\) be the \(n\times k\) matrix of the first \(k\) eigenvectors, and set \[ D_1=\operatorname{diag}(\sqrt{\lambda_1},\dots,\sqrt{\lambda_k}). \] The synthesis operator \[ B e_i = D_1 M_1^* e_i,\qquad i=1,\dots,n \] defines a frame \(\Phi=\{\phi_i=B e_i\}_{i=1}^n\) for \(\mathbb C^k\), with frame operator \[ S_\Phi=B B^*=\operatorname{diag}(\lambda_1,\dots,\lambda_k). \] Its canonical dual is \[ \psi_i^{(\mathrm{can})}=S_\Phi^{-1}\phi_i=D_1^{-1}M_1^*e_i \] [2406.00776]. In this setting, the structure of all duals is unusually simple. If the graph has \(m\) connected components, then every dual has the form \[ \psi_i=S_\Phi^{-1}\phi_i+\mu_{j(i)}, \] where \(j(i)\) labels the component containing \(\phi_i\), and the componentwise offsets satisfy the usual duality constraint [2406.00776]. The main outcomes are summarized below. | Graph regime | Canonical dual outcome | Uniqueness | |---|---|---| | Connected graph | Unique spectrally optimal dual for \(1\)-erasure and \(2\)-erasures; later extended to all \(r\)-erasures | Unique | | Disconnected graph | Canonical dual remains spectrally optimal for \(1\)-erasure, \(2\)-erasures, and later all \(r\)-erasures | Non-unique; infinitely many optimal offsets | For a connected graph on \(n\) vertices, the associated \(L_G(n,n-1)\)-frame satisfies \[ \rho^{(1)}_{\Phi,\Psi^{(\mathrm{can})}}=\frac{n-1}{n}, \qquad \rho^{(2)}_{\Phi,\Psi^{(\mathrm{can})}}=1, \] and for every erased pair \(\Lambda=\{i,j\}\), \[ \operatorname{spec} E_{\Phi,\Psi^{(\mathrm{can})},\Lambda} = \{1,(n-2)/n\}. \] In both cases the canonical dual is optimal, and any alternate dual preserving the same bound must have zero offset, hence uniqueness [2406.00776]. For disconnected graphs with component sizes \(n_1,\dots,n_k\), the canonical dual still attains \[ \rho^{(1)}_{\Phi,\Psi^{(\mathrm{can})}} = \max_j \frac{n_j-1}{n_j}, \qquad \rho^{(2)}_{\Phi,\Psi^{(\mathrm{can})}}=1, \] but nontrivial componentwise offsets can be inserted without increasing the worst-case value. The result is an infinite family of spectrally optimal duals [2406.00776]. A later extension treats arbitrary \(r\)-erasures. It proves that if \(\Phi\) and \(\Phi'\) are unitarily equivalent frames with canonical duals \(\Phi^\dagger\) and \((\Phi')^\dagger\), then \[ \rho^{(r)}_{\Phi,\Phi^\dagger} = \rho^{(r)}_{\Phi',(\Phi')^\dagger}. \] For a connected graph-generated frame, the canonical dual is the unique spectrally optimal dual for every \(r=1,\dots,n-1\), and for an erasure set \(U\) of size \(r\), \[ \operatorname{Spec}(E_U)=\{1\ \text{(mult.\ }r-1),\ 1-r/n,\ 0\ \text{(rest)}\} \] when \(r