Gram-Spectrum Method

• The Gram-Spectrum Method is an operator-theoretic framework that uses the spectral decomposition of the Gramian to identify maximal subspaces where genuine frames emerge.
• It replaces the global invertibility requirement with local spectral invertibility over compact intervals, ensuring reliable frame bounds on selected spectral windows.
• The method yields canonical analysis and synthesis operators and finds applications in reproducing-kernel Hilbert spaces and random field representations.

Searching arXiv for the cited papers to ground the response. The Gram–Spectrum Method is an operator-theoretic framework for extracting frame structure from a countable system of vectors whose Gramian need not be boundedly invertible on all of $\ell^2(S)$. In the formulation based on Jorgensen–Tian, one starts with a countable family $\{\varphi_s\}_{s\in S}\subset\mathscr H$ in a separable Hilbert space, forms the infinite Gramian $$G_S=(\langle \varphi_i,\varphi_j\rangle_{\mathscr H})_{i,j\in S}$$, and uses the spectral decomposition of the associated selfadjoint operator to isolate maximal spectral subspaces on which the induced vectors become genuine frames with frame bounds given by the endpoints of a chosen spectral interval (Jorgensen et al., 2015). The method replaces a global invertibility requirement by local spectral invertibility on windows $J=[a,b]\subset(0,\infty)$, and thereby yields canonical analysis and synthesis operators together with applications to reproducing-kernel Hilbert spaces and random fields (Jorgensen et al., 2015).

1. Operator-theoretic setup

Let

$$\varphi:S\longrightarrow \mathscr H,\qquad s\mapsto \varphi_s,$$

where $S$ is countable and $\mathscr H$ is a separable Hilbert space. The central object is the $S$–Gramian

$$G_S=\bigl(\langle \varphi_i,\varphi_j\rangle_{\mathscr H}\bigr)_{i,j\in S}.$$

Under the square-summability assumption

$$\forall\,j\in S:\quad \sum_{i\in S}\bigl|\langle\varphi_i,\varphi_j\rangle\bigr|^2<\infty,$$

this matrix defines a densely-defined, positive, symmetric operator

$\{\varphi_s\}_{s\in S}\subset\mathscr H$0

by

$\{\varphi_s\}_{s\in S}\subset\mathscr H$1

The method is organized around this operator rather than around a priori frame inequalities. In the classical frame setting, bounded invertibility of the Gram matrix on all of $\{\varphi_s\}_{s\in S}\subset\mathscr H$2 is used to conclude that $\{\varphi_s\}_{s\in S}\subset\mathscr H$3 is a frame. Here the starting point is weaker: $\{\varphi_s\}_{s\in S}\subset\mathscr H$4 is only required to be semibounded and, under the stated non-deficiency condition, essentially selfadjoint (Jorgensen et al., 2015).

This shift in viewpoint is decisive. Instead of asking whether the full system already satisfies global frame bounds, the method asks how the spectrum of the Gramian decomposes the ambient coefficient space into regions where such bounds hold. This suggests an intrinsic spectral classification of the frame behavior of the original family.

2. Selfadjoint extension and spectral resolution

By von Neumann’s theory, the closure of $\{\varphi_s\}_{s\in S}\subset\mathscr H$5 is selfadjoint if and only if its deficiency spaces vanish. Equivalently, one requires

$\{\varphi_s\}_{s\in S}\subset\mathscr H$6

When this condition holds, there is a unique projection-valued measure

$\{\varphi_s\}_{s\in S}\subset\mathscr H$7

such that

$\{\varphi_s\}_{s\in S}\subset\mathscr H$8

in the sense of the functional calculus for unbounded selfadjoint operators (Jorgensen et al., 2015).

The spectral theorem is therefore the analytic core of the Gram–Spectrum Method. The Gramian is not treated merely as a matrix of pairwise inner products, but as a positive selfadjoint operator with a full spectral resolution. Once the operator is placed in this functional-calculus setting, spectral projections can be used to isolate subspaces on which the Gramian has uniform lower and upper bounds.

A plausible implication is that the method is especially well adapted to systems whose Gramian is unbounded or poorly conditioned globally, because it identifies intervals of the spectrum where the geometry is nonetheless stable enough to support frame reconstruction.

3. Spectral windows and maximal frame subspaces

Fix a compact subinterval

$\{\varphi_s\}_{s\in S}\subset\mathscr H$9

The associated spectral subspace is

$$G_S=(\langle \varphi_i,\varphi_j\rangle_{\mathscr H})_{i,j\in S}$$0

which is closed in $$G_S=(\langle \varphi_i,\varphi_j\rangle_{\mathscr H})_{i,j\in S}$$1. Functional calculus gives the two-sided estimate

$$G_S=(\langle \varphi_i,\varphi_j\rangle_{\mathscr H})_{i,j\in S}$$2

These inequalities are the local substitute for global Gramian invertibility. On $$G_S=(\langle \varphi_i,\varphi_j\rangle_{\mathscr H})_{i,j\in S}$$3, the operator $$G_S=(\langle \varphi_i,\varphi_j\rangle_{\mathscr H})_{i,j\in S}$$4 is bounded below by $$G_S=(\langle \varphi_i,\varphi_j\rangle_{\mathscr H})_{i,j\in S}$$5 and above by $$G_S=(\langle \varphi_i,\varphi_j\rangle_{\mathscr H})_{i,j\in S}$$6, so the spectral window behaves as a region of uniform stability. The exposition describes $$G_S=(\langle \varphi_i,\varphi_j\rangle_{\mathscr H})_{i,j\in S}$$7 as the maximal subspace obtained by bundling together the spectral pieces corresponding to $$G_S=(\langle \varphi_i,\varphi_j\rangle_{\mathscr H})_{i,j\in S}$$8 (Jorgensen et al., 2015).

The designation “maximal” should be understood in the sense used there: the spectral interval determines the largest subspace cut out by the projection $$G_S=(\langle \varphi_i,\varphi_j\rangle_{\mathscr H})_{i,j\in S}$$9 on which the Gramian enjoys the bounds $J=[a,b]\subset(0,\infty)$0. This provides a continuum of frame-bearing subspaces indexed by compact intervals in the positive spectrum.

4. Analysis, synthesis, and the frame inequalities

The method also furnishes canonical analysis and synthesis operators on each spectral slice. By the general theory of direct integrals, one may write

$J=[a,b]\subset(0,\infty)$1

On a fixed spectral window $J=[a,b]\subset(0,\infty)$2, define the analysis operator

$J=[a,b]\subset(0,\infty)$3

and the synthesis operator

$J=[a,b]\subset(0,\infty)$4

where the integral is understood in the strong sense on $J=[a,b]\subset(0,\infty)$5 (Jorgensen et al., 2015).

If one sets

$J=[a,b]\subset(0,\infty)$6

then for every $J=[a,b]\subset(0,\infty)$7,

$J=[a,b]\subset(0,\infty)$8

The spectral estimate becomes

$J=[a,b]\subset(0,\infty)$9

so $$\varphi:S\longrightarrow \mathscr H,\qquad s\mapsto \varphi_s,$$0 is a frame for $$\varphi:S\longrightarrow \mathscr H,\qquad s\mapsto \varphi_s,$$1 with frame bounds $$\varphi:S\longrightarrow \mathscr H,\qquad s\mapsto \varphi_s,$$2 and $$\varphi:S\longrightarrow \mathscr H,\qquad s\mapsto \varphi_s,$$3 (Jorgensen et al., 2015).

In this formulation, frame vectors are not assumed at the outset but are created by spectral projection. The paper’s phrase “creating frame vectors in maximal subspaces” captures precisely this mechanism. The resulting family $$\varphi:S\longrightarrow \mathscr H,\qquad s\mapsto \varphi_s,$$4 is derived from the coordinate basis via the spectral projection, while the Gramian determines the reconstruction operator through the inverse spectral weighting $$\varphi:S\longrightarrow \mathscr H,\qquad s\mapsto \varphi_s,$$5 on $$\varphi:S\longrightarrow \mathscr H,\qquad s\mapsto \varphi_s,$$6.

5. Departure from classical global invertibility

A central feature of the method is the explicit removal of the classical requirement that the Gram matrix be boundedly invertible on all of $$\varphi:S\longrightarrow \mathscr H,\qquad s\mapsto \varphi_s,$$7. Instead, one uses the spectrum of $$\varphi:S\longrightarrow \mathscr H,\qquad s\mapsto \varphi_s,$$8 to decompose $$\varphi:S\longrightarrow \mathscr H,\qquad s\mapsto \varphi_s,$$9 into a continuum of subspaces $S$0, then groups those for $S$1 into $S$2; on that subspace, the Gramian is invertible with two-sided bounds $S$3 (Jorgensen et al., 2015).

This is the principal conceptual distinction between the Gram–Spectrum Method and standard frame criteria. The question is no longer whether the original family $S$4 is a frame globally, but on which spectral sectors it induces a frame after projection. The procedure is therefore constructive in a spectral sense: each admissible interval in the positive spectrum generates a corresponding frame system.

A common misconception would be to treat the method as merely a rephrasing of ordinary frame theory. The source material indicates a stronger claim: it produces new frames even when the original family fails globally to be a frame. That is not a cosmetic reformulation of the frame inequalities but an alternative route to them through the selfadjoint extension and spectral calculus of the Gramian (Jorgensen et al., 2015).

6. Applications and neighboring uses of Gram objects

Two applications are highlighted. In the reproducing-kernel Hilbert space setting, assume each $S$5 is a function on a set $S$6 and that for each $S$7 the coordinate vector $S$8 lies in $S$9. On a window $\mathscr H$0, the subspace $\mathscr H$1 becomes an RKHS with kernel

$\mathscr H$2

and every $\mathscr H$3 satisfies the reproducing property

$\mathscr H$4

(Jorgensen et al., 2015).

In the random-field setting, let $\mathscr H$5 be a probability space and $\mathscr H$6 a countable set. If $\mathscr H$7 satisfies

$\mathscr H$8

together with the same no-defect condition for the Gramian, then

$\mathscr H$9

is selfadjoint in $S$0. For each spectral window $S$1 one obtains the maximal subspace

$S$2

on which $S$3 is a genuine frame with bounds $S$4 (Jorgensen et al., 2015).

The term “Gram” also appears in a different line of research centered on positive semidefinite Gram matrices for sum-of-squares representations of polynomials. There, the relevant object is the Gram spectrahedron

$S$5

which parametrizes sum-of-squares decompositions and is studied in polynomial optimization and convex algebraic geometry (Chua et al., 2016). The overlap in terminology concerns the use of Gram matrices, but the underlying problems differ: the Gram–Spectrum Method analyzes spectral subspaces and frame generation in Hilbert spaces, whereas Gram spectrahedra organize sums-of-squares representations of polynomials (Chua et al., 2016).