Introduction
Redundancy is one of the fundamental features that distinguishes frames from bases. While a basis provides a unique linear representation of each vector, a frame allows multiple representations and, consequently, considerable freedom in the choice of reconstruction operators. This flexibility becomes particularly significant when additional geometric constraints are imposed on the frame vectors or their duals. Among such constraints, positive spanning occupies a distinguished position: it requires that every vector in the underlying space be generated using only nonnegative coefficients. The purpose of this paper is to investigate how this classical geometric property interacts with the algebraic and operator-theoretic structure of finite frames, with particular emphasis on equivalence, duality, redundancy, and robustness.
The theory of positive spanning sets has a long history, originating in the study of positive linear dependence. In his foundational work, Davis [6 ] developed fundamental properties of positive spanning families and positive bases, including geometric characterizations and cardinality restrictions for minimal positive spanning systems. These ideas subsequently appeared in the geometric study of convex polytopes and Gale diagrams [10 ] . Positive spanning sets have also become important in derivative-free optimization, where they provide collections of search directions capable of detecting descent without explicit derivative information; see [7 , 8 ] . Dynamical sampling provides another important connection to frame theory through operator-generated systems; see [11 ] for a comprehensive survey and [4 ] for recent constructions of positive frames with optimal redundancy under erasures. Related geometric configurations arise in redundant control systems, where suitable collections of control directions are needed to generate prescribed forces and motions; see, for example, [9 , 12 , 13 , 14 ] . These developments demonstrate the usefulness of positive spanning in several different settings. However, their principal emphasis is on positive dependence, minimality, geometric configurations, or applications, rather than on the structure of dual frames and the operator-theoretic consequences of positivity.
In finite-dimensional frame theory, every spanning family is a frame, but not every frame is positively spanning. Moreover, positivity introduces a requirement that differs fundamentally from ordinary linear spanning. A family F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N in an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n is called a positive frame if
span + ( F ) = { ∑ i = 1 N c i f i : c i ≥ 0 } = H n . \operatorname{span}_{+}(F) = \left\{ \sum_{i=1}^{N} c_i f_i : c_i \ge0 \right\} = \mathcal{H}_n. span + ( F ) = { i = 1 ∑ N c i f i : c i ≥ 0 } = H n .
The nonnegativity in this definition concerns the coefficients in a representation, not the coordinates of the frame vectors themselves. It is also important to distinguish positive spanning from reconstruction by a fixed linear coefficient map. Although a positive frame allows each vector to be expressed using nonnegative coefficients, those coefficients may depend nonlinearly on the vector. In contrast, a dual frame supplies a prescribed linear reconstruction formula. Understanding the relationship between these two forms of representation is one of the motivations for the present work.
The classical geometric theory provides a natural starting point. For a finite spanning family, positive spanning is equivalent to the existence of a strictly positive linear dependence among its vectors, or, equivalently, to the origin lying in the interior of their convex hull. In the language of frame operators, this condition becomes
F is positive ⟺ ker T F ∗ ∩ R + + N ≠ ∅ . F \text{ is positive} \quad\Longleftrightarrow\quad\ker T_F^* \cap\mathbb{R}_{++}^{N} \ne\varnothing. F is positive ⟺ ker T F ∗ ∩ R ++ N = ∅ .
where T F ∗ T_F^* T F ∗ denotes the synthesis operator. Although this equivalence follows from classical positive-spanning theory, its operator formulation suggests a broader perspective: the geometry of a positive frame can be investigated through its space of linear relations. In particular, it raises the question of how much of the structure of a positive frame is determined by its synthesis kernel.
Our first objective is therefore to study positive frames under invertible linear equivalence. Since two ordered frames are equivalent precisely when their synthesis operators have the same kernel, the classification problem can be reformulated in terms of subspaces of R N \mathbb{R}^{N} R N that intersect the strictly positive orthant. We establish a parametrization of equivalence classes of positive frames through these relation spaces. For frames with excess one, the parameter space reduces to the relative interior of a standard simplex, with each point representing a unique normalized strictly positive dependence. For arbitrary excess, we obtain a quotient description in terms of linearly independent strictly positive relations, together with an equivalent Grassmannian formulation. We also examine the role of orthogonal projections in constructing positive frames. These results provide a geometric classification that depends on the relation space rather than on the frame operator alone.
A more substantial question concerns duality. Given a frame F F F , one may ask whether its reconstruction freedom can be used to select a dual frame that positively spans the ambient space. This is not the same as asking whether F F F itself is positive: the original frame and its duals can have very different positive-spanning properties. We prove that a finite frame admits a positive dual if and only if it is redundant. Thus the existence of a positive dual is governed entirely by the excess of the original frame. Our proof is constructive. Starting from the canonical dual, we use a strictly positive vector outside the analysis range to produce a rank-one perturbation satisfying both the reconstruction identity and a prescribed strictly positive linear relation. This gives an explicit mechanism for introducing positivity into a dual system, even when the original frame is not positively spanning.
Beyond existence, we investigate the geometry of the family D + ( F ) \mathcal{D}_{+}(F) D + ( F ) of all positive duals of a fixed frame. We establish its relative openness within the affine space of dual frames and examine the perturbations that preserve positivity and reconstruction simultaneously. In particular, we show that D + ( F ) \mathcal{D}_{+}(F) D + ( F ) has no extreme points: every positive dual can be expressed as the midpoint of two distinct positive duals. We also study dual frames satisfying prescribed positive relations and the more restrictive problem of obtaining a dual that is a positive basis. The latter leads to conditions involving complements of the analysis range and prescribed coordinate sign patterns. Together, these results describe how positive spanning restricts, but does not eliminate, the freedom available in dual-frame reconstruction.
The geometry becomes more delicate when strict positivity is relaxed. Motivated by this distinction, we consider semi-positive frames, which admit nontrivial nonnegative linear relations with at least one vanishing coefficient. We investigate their approximation by positive frames and the corresponding boundary behavior of positive duals. This provides a connection between strictly positive relations and degenerate nonnegative dependencies, and clarifies how positivity may be lost under limiting processes. The resulting picture is useful for understanding the topology of families of positive frames and their dual systems.
We further examine positive frames arising from finite group representations. Group-generated families possess additional symmetry, and their linear relations reflect the structure of the underlying representation. We investigate the relationship between balancedness, linear connectedness, and positive spanning, using the averaging operator associated with a finite group. In particular, the presence or absence of invariant vectors provides information about balanced orbit families and their positivity. This places the geometric theory of positive frames in a structured setting where algebraic symmetry can be used to identify positive dependencies.
Finally, we study positive dependence from the viewpoint of convex geometry and robustness. Minimal positive dependencies and the extreme points of the polytope of normalized nonnegative relations provide information about the internal structure of positive frames. We also consider the effect of deleting frame vectors. Unlike ordinary spanning, positive spanning requires every nonzero direction to have a strictly positive inner product with at least one frame vector. This leads to the directional quantity
d + ( F ) = min ∥ x ∥ = 1 # { i : ⟨ x , f i ⟩ > 0 } . d_{+}(F)=\min_{\lVert x\rVert=1}\#\{i:\langle x,f_i\rangle>0\}. d + ( F ) = ∥ x ∥ = 1 min # { i : ⟨ x , f i ⟩ > 0 } .
We show that a positive frame remains positively spanning after the deletion of any s s s vectors if and only if d + ( F ) ≥ s + 1 d_{+}(F)\ge s+1 d + ( F ) ≥ s + 1 . We also investigate related questions of exactness, cardinality, and approximation by positive Parseval frames. These results connect the combinatorial structure of positive dependencies with the stability of positive spanning under perturbations and erasures.
The paper is organized as follows. Section 2 recalls the necessary background on finite frames, positive spanning sets, and convex geometry. Section 3 develops geometric characterizations of positive frames, their equivalence classes through synthesis kernels, simplex and Grassmannian parametrizations, and the structure of positive bases. Section 4 investigates positive dual frames, including their existence, construction, and geometric properties. Section 5 examines the convex geometry of positive duals, their boundary behavior, and the approximation of semi-positive frames. Section 6 studies positive frames generated by finite group representations, with particular emphasis on balancedness and linear connectedness. Section 7 explores the convex geometry of nonnegative relations, positive-spanning robustness under erasures, exactness, cardinality bounds, and positive Parseval approximation. Finally, Section 8 presents concluding remarks and directions for future research.
Preliminaries
Throughout this paper, H n \mathcal{H}_n H n denotes an n n n -dimensional real Hilbert space, where n ≥ 1 n\ge1 n ≥ 1 . All vector spaces and operators are considered over R \mathbb{R} R . We work with finite indexed families F = { f i } i = 1 N ⊂ H n F=\{f_i\}_{i=1}^{N}\subset\mathcal{H}_n F = { f i } i = 1 N ⊂ H n , where repetitions and zero vectors are permitted unless otherwise specified. We recall the basic notions of finite frame theory, positive spanning, and convex geometry that will be used throughout the paper.
Finite frames and duality
A family F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N is called a frame for H n \mathcal{H}_n H n if there exist constants 0 < A ≤ B < ∞ 0<A\le B<\infty 0 < A ≤ B < ∞ such that
A ∥ x ∥ 2 ≤ ∑ i = 1 N ∣ ⟨ x , f i ⟩ ∣ 2 ≤ B ∥ x ∥ 2 , x ∈ H n . A\lVert x\rVert^2\le\sum_{i=1}^{N}\lvert\langle x,f_i\rangle\rvert^2\le B\lVert x\rVert^2,\qquad x\in\mathcal{H}_n. A ∥ x ∥ 2 ≤ i = 1 ∑ N ∣⟨ x , f i ⟩ ∣ 2 ≤ B ∥ x ∥ 2 , x ∈ H n .
In finite dimensions, this is equivalent to span ( F ) = H n \operatorname{span}(F)=\mathcal{H}_n span ( F ) = H n . The analysis operator T F : H n → R N T_F:\mathcal{H}_n\to\mathbb{R}^{N} T F : H n → R N and the synthesis operator T F ∗ : R N → H n T_F^{*}:\mathbb{R}^{N}\to\mathcal{H}_n T F ∗ : R N → H n are defined by
T F x = ∑ i = 1 N ⟨ x , f i ⟩ e i , T F ∗ c = ∑ i = 1 N c i f i , T_Fx=\sum_{i=1}^{N}\langle x,f_i\rangle e_i,\qquad T_F^{*}c=\sum_{i=1}^{N}c_if_i, T F x = i = 1 ∑ N ⟨ x , f i ⟩ e i , T F ∗ c = i = 1 ∑ N c i f i ,
where { e i } i = 1 N \{e_i\}_{i=1}^{N} { e i } i = 1 N is the standard orthonormal basis of R N \mathbb{R}^{N} R N and c = ( c i ) i = 1 N c=(c_i)_{i=1}^{N} c = ( c i ) i = 1 N . The associated frame operator is S F = T F ∗ T F S_F=T_F^{*}T_F S F = T F ∗ T F , which is positive, self-adjoint, and invertible.
A frame is called tight if S F = A I H n S_F=AI_{\mathcal{H}_n} S F = A I H n for some A > 0 A>0 A > 0 , and Parseval if S F = I H n S_F=I_{\mathcal{H}_n} S F = I H n . The canonical dual of F F F is F ~ = { S F − 1 f i } i = 1 N \widetilde{F}=\{S_F^{-1}f_i\}_{i=1}^{N} F = { S F − 1 f i } i = 1 N .
A frame G = { g i } i = 1 N G=\{g_i\}_{i=1}^{N} G = { g i } i = 1 N is called a dual frame of F F F if
x = ∑ i = 1 N ⟨ x , f i ⟩ g i , x ∈ H n , (1) x=\sum_{i=1}^{N}\langle x,f_i\rangle g_i,\qquad x\in\mathcal{H}_n, \tag*{(1)} x = i = 1 ∑ N ⟨ x , f i ⟩ g i , x ∈ H n , (1)
or equivalently, T G ∗ T F = I H n T_G^*T_F=I_{\mathcal{H}_n} T G ∗ T F = I H n . We denote by D ( F ) \mathcal{D}(F) D ( F ) the collection of all dual frames of F F F .
Recall that every dual frame of F F F has the representation
g i = S F − 1 f i + h i , 1 ≤ i ≤ N , g_i=S_F^{-1}f_i+h_i,\qquad1\leq i\leq N, g i = S F − 1 f i + h i , 1 ≤ i ≤ N ,
where H = { h i } i = 1 N H=\{h_i\}_{i=1}^N H = { h i } i = 1 N satisfies T H ∗ T F = 0 T_H^*T_F=0 T H ∗ T F = 0 . Consequently,
D ( F ) = { F ~ + H : T H ∗ T F = 0 } . \mathcal{D}(F)=\left\{\widetilde{F}+H:T_H^*T_F=0\right\}. D ( F ) = { F + H : T H ∗ T F = 0 } .
Thus D ( F ) \mathcal{D}(F) D ( F ) is an affine subspace of H n N \mathcal{H}_n^N H n N . The relation space of F F F is K F = ker T F ∗ K_F=\ker T_F^* K F = ker T F ∗ . Since T F ∗ T_F^* T F ∗ is surjective, the rank-nullity theorem gives
dim K F = N − n , K F = ( Ran T F ) ⊥ . \dim K_F=N-n,\qquad K_F=(\operatorname{Ran}T_F)^\perp. dim K F = N − n , K F = ( Ran T F ) ⊥ .
The integer N − n N-n N − n is called the excess of F F F . In particular, F F F is redundant if and only if K F ≠ { 0 } K_F\ne\{0\} K F = { 0 } .
We shall frequently use the orthogonal projection onto Ran T F \operatorname{Ran}T_F Ran T F , given by
P F = T F S F − 1 T F ∗ . P_F=T_FS_F^{-1}T_F^*. P F = T F S F − 1 T F ∗ .
Consequently, I R N − P F I_{\mathbb{R}^N}-P_F I R N − P F is the orthogonal projection onto K F K_F K F . These identities provide a convenient connection between the reconstruction operators and the linear relations among frame vectors.
Positive spanning sets and positive bases
For a finite family F = { f i } i = 1 N ⊂ H n F=\{f_i\}_{i=1}^N\subset\mathcal{H}_n F = { f i } i = 1 N ⊂ H n , its positive span is defined by
span + ( F ) = { ∑ i = 1 N c i f i : c i ≥ 0 } . \operatorname{span}_{+}(F)=\left\{\sum_{i=1}^N c_i f_i:c_i\geq0\right\}. span + ( F ) = { i = 1 ∑ N c i f i : c i ≥ 0 } .
A frame F F F is called a positive frame if span + ( F ) = H n \operatorname{span}_{+}(F)=\mathcal{H}_n span + ( F ) = H n . A dual frame G G G of F F F is called a positive dual frame if G G G is itself a positive frame. We write
D + ( F ) = { G ∈ D ( F ) : span + ( G ) = H n } . \mathcal{D}_{+}(F)=\{G\in\mathcal{D}(F):\operatorname{span}_{+}(G)=\mathcal{H}_n\}. D + ( F ) = { G ∈ D ( F ) : span + ( G ) = H n } .
A family F F F is said to admit a nonnegative relation if there exist coefficients c i ≥ 0 c_i\geq0 c i ≥ 0 , not all zero, such that ∑ i = 1 N c i f i = 0 \sum_{i=1}^N c_i f_i=0 ∑ i = 1 N c i f i = 0 . If every coefficient is strictly positive, the relation is called a strictly positive relation . For convenience, we write
R ≥ 0 N = { c ∈ R N : c i ≥ 0 } , R + + N = { c ∈ R N : c i > 0 } . \mathbb{R}_{\geq0}^N=\{c\in\mathbb{R}^N:c_i\geq0\},\qquad
\mathbb{R}_{++}^N=\{c\in\mathbb{R}^N:c_i>0\}. R ≥ 0 N = { c ∈ R N : c i ≥ 0 } , R ++ N = { c ∈ R N : c i > 0 } .
Thus nonnegative and strictly positive relations correspond, respectively, to nonzero vectors in K F ∩ R ≥ 0 N K_F\cap\mathbb{R}_{\geq0}^N K F ∩ R ≥ 0 N and vectors in K F ∩ R + + N K_F\cap\mathbb{R}_{++}^N K F ∩ R ++ N .
The theory of positive linear dependence was developed systematically by Davis [6 ] . We recall a classical result that will be used in our analysis. A positive frame F F F is called a positive basis if no proper subfamily of F F F positively spans H n \mathcal{H}_n H n . Thus a positive basis is a minimal positive spanning family, which need not be a basis in the usual linear-algebraic sense. The following cardinality restriction is another classical result of Davis.
Theorem 2.1 (Davis [6 ] ). Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^N F = { f i } i = 1 N be a positive basis for H n \mathcal{H}_n H n . Then
n + 1 ≤ N ≤ 2 n . n+1\leq N\leq2n. n + 1 ≤ N ≤ 2 n .
Positive bases consisting of n + 1 n+1 n + 1 and 2 n 2n 2 n vectors are called minimal positive bases and maximal positive bases , respectively. These configurations also arise in derivative-free optimization, where positive spanning sets are used to construct suitable search directions; see [7 , 8 , 20 ] . The following standard examples illustrate the distinction between positive frames and positive bases.
Example 2.2. Let { e i } i = 1 n \{e_i\}_{i=1}^{n} { e i } i = 1 n be the standard orthonormal basis of R n \mathbb{R}^{n} R n .
(i) The family
F 1 = { e 1 , … , e n , − ∑ i = 1 n e i } F_1=\left\{e_1,\ldots,e_n,-\sum_{i=1}^{n}e_i\right\} F 1 = { e 1 , … , e n , − i = 1 ∑ n e i } is a minimal positive basis consisting of n + 1 n+1 n + 1 vectors.
(ii) The family
F 2 = { e 1 , − e 1 , … , e n , − e n } F_2=\{e_1,-e_1,\ldots,e_n,-e_n\} F 2 = { e 1 , − e 1 , … , e n , − e n } is a maximal positive basis consisting of 2 n 2n 2 n vectors.
(iii) The family
F 3 = { ( 1 , 0 ) , ( 0 , 1 ) , ( − 1 , − 1 ) , ( 1 , 1 ) } F_3=\{(1,0),(0,1),(-1,-1),(1,1)\} F 3 = {( 1 , 0 ) , ( 0 , 1 ) , ( − 1 , − 1 ) , ( 1 , 1 )} is a positive frame for R 2 \mathbb{R}^{2} R 2 , but not a positive basis, since its first three vectors already positively span R 2 \mathbb{R}^{2} R 2 .
Convex geometry and positive relations
For a finite family F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N , its convex hull is
conv ( F ) = { ∑ i = 1 N α i f i : α i ≥ 0 , ∑ i = 1 N α i = 1 } . \operatorname{conv}(F)=\left\{\sum_{i=1}^{N}\alpha_i f_i:\ \alpha_i\geq0,\ \sum_{i=1}^{N}\alpha_i=1\right\}. conv ( F ) = { i = 1 ∑ N α i f i : α i ≥ 0 , i = 1 ∑ N α i = 1 } .
We use int ( C ) \operatorname{int}(C) int ( C ) , relint ( C ) \operatorname{relint}(C) relint ( C ) , and Ext ( C ) \operatorname{Ext}(C) Ext ( C ) to denote the interior, relative interior, and set of extreme points of a convex set C C C , respectively.
For a cone C ⊆ H n C\subseteq\mathcal{H}_n C ⊆ H n , its negative polar cone is defined by
C ∘ = { x ∈ H n : ⟨ x , y ⟩ ≤ 0 for all y ∈ C } . C^{\circ}=\{x\in\mathcal{H}_n:\langle x,y\rangle\leq0\text{ for all }y\in C\}. C ∘ = { x ∈ H n : ⟨ x , y ⟩ ≤ 0 for all y ∈ C } .
In particular, for the finitely generated cone C = span + ( F ) C=\operatorname{span}_{+}(F) C = span + ( F ) , we have
C ∘ = { x ∈ H n : ⟨ x , f i ⟩ ≤ 0 for all i } . C^{\circ}=\{x\in\mathcal{H}_n:\langle x,f_i\rangle\leq0\text{ for all }i\}. C ∘ = { x ∈ H n : ⟨ x , f i ⟩ ≤ 0 for all i } .
Since finitely generated cones are closed and convex, the bipolar theorem yields
span + ( F ) = H n ⟺ C ∘ = { 0 } . \operatorname{span}_{+}(F)=\mathcal{H}_n\quad\Longleftrightarrow\quad C^{\circ}=\{0\}. span + ( F ) = H n ⟺ C ∘ = { 0 } .
This gives a convex-analytic interpretation of [6 ] .
Another useful classical fact is the supporting hyperplane theorem: if a convex set has nonempty interior and the origin does not belong to that interior, then there exists a nonzero linear functional separating the origin from the interior. In the present finite-dimensional setting, such functionals can be represented by inner products.
We shall also use the following standard consequence of the separation theorem.
Lemma 2.3. Let F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a spanning family for H n \mathcal{H}_n H n . Then 0 ∈ int conv ( F ) 0 \in\operatorname{int}\operatorname{conv}(F) 0 ∈ int conv ( F ) if and only if, for every nonzero x ∈ H n x \in\mathcal{H}_n x ∈ H n , there exist indices i , j i,j i , j such that
⟨ x , f i ⟩ > 0 , ⟨ x , f j ⟩ < 0. \langle x,f_i\rangle> 0,\qquad\langle x,f_j\rangle< 0. ⟨ x , f i ⟩ > 0 , ⟨ x , f j ⟩ < 0. Proof. Suppose that 0 ∈ int conv ( F ) 0 \in\operatorname{int}\operatorname{conv}(F) 0 ∈ int conv ( F ) . For every nonzero x ∈ H n x \in\mathcal{H}_n x ∈ H n , there exists ε > 0 \varepsilon> 0 ε > 0 such that ± ε x ∈ conv ( F ) \pm\varepsilon x \in\operatorname{conv}(F) ± ε x ∈ conv ( F ) . Taking inner products with x x x shows that some frame vector has positive inner product with x x x , while another has negative inner product with x x x .
Conversely, suppose the stated sign condition holds. If the origin were not in the interior of conv ( F ) \operatorname{conv}(F) conv ( F ) , the separation theorem would provide a nonzero x ∈ H n x \in\mathcal{H}_n x ∈ H n such that ⟨ x , f i ⟩ ≥ 0 \langle x,f_i\rangle\geq0 ⟨ x , f i ⟩ ≥ 0 for every i i i , contradicting the assumed condition.
The preceding lemma is a convenient form of the classical positive-spanning criterion. In Section 3, we shall relate these geometric properties to strictly positive vectors in the synthesis kernel.
For later use, we also introduce the set of normalized nonnegative relations
Ω F = { c ∈ K F ∩ R ≥ 0 N : ∑ i = 1 N c i = 1 } . \Omega_F = \left\{c \in K_F \cap\mathbb{R}_{\geq0}^{N} : \sum_{i=1}^{N} c_i = 1\right\}. Ω F = { c ∈ K F ∩ R ≥ 0 N : i = 1 ∑ N c i = 1 } .
This is a compact convex polytope, possibly empty. Its elements encode nonnegative linear dependencies among the frame vectors. We shall investigate its geometric structure and extreme points in a later section.
Finally, we recall that a convex subset C C C of an affine space has a relative interior taken with respect to aff ( C ) \operatorname{aff}(C) aff ( C ) . The distinction between ordinary and relative interiors will be important when studying families of dual frames, since D ( F ) \mathcal{D}(F) D ( F ) is generally a proper affine subspace of H n N \mathcal{H}_n^{N} H n N .
Geometry and Classification of Positive Frames
Throughout this section, let H n \mathcal{H}_n H n be an n n n -dimensional real Hilbert space and let F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for H n \mathcal{H}_n H n . In this section, we investigate the geometric and operator-theoretic structure of positive frames, with particular emphasis on their characterization and classification under invertible linear transformations. We begin by relating positive spanning to strictly positive linear dependencies and the interior geometry of convex hulls. Building on these classical characterizations, we develop a relation-space approach to the classification of positive frames. In particular, we show that their equivalence classes can be completely described through the kernels of their synthesis operators. This leads to a parametrization by the relative interior of a standard simplex in the excess-one case and, more generally, by suitable subsets of Grassmannians. We also investigate constructions arising from orthogonal projections and demonstrate that positive Parseval frames with identical frame operators need not be equivalent. These results provide a geometric description of positive frames in which the synthesis kernel, rather than the frame operator alone, plays the central role.
Geometric Characterizations of Positive Frames
We begin by recalling equivalent geometric characterizations of positive frames in terms of strictly positive linear dependencies and the interior of their convex hulls. We then interpret these conditions through the kernel of the synthesis operator and establish some related properties that will be useful in the subsequent classification of positive frames.
The following result provides equivalent characterizations of positive frames in terms of strictly positive linear dependence and the geometry of the convex hull.
Proposition 3.1. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for H n \mathcal{H}_n H n . Then the following statements are equivalent:
( i ) (i) ( i ) F F F is a positive frame.
( i i ) (ii) ( ii ) There exist scalars λ 1 , … , λ N > 0 \lambda_1,\ldots,\lambda_N>0 λ 1 , … , λ N > 0 such that ∑ i = 1 N λ i f i = 0 \sum_{i=1}^{N}\lambda_i f_i=0 ∑ i = 1 N λ i f i = 0 .
( i i i ) (iii) ( iii ) 0 ∈ int ( conv ( F ) ) 0\in\operatorname{int}(\operatorname{conv}(F)) 0 ∈ int ( conv ( F )) .
Proof. We prove the implications ( i ) ⇒ ( i i ) ⇒ ( i i i ) ⇒ ( i ) (i)\Rightarrow(ii)\Rightarrow(iii)\Rightarrow(i) ( i ) ⇒ ( ii ) ⇒ ( iii ) ⇒ ( i ) .
( i ) ⇒ ( i i ) (i)\Rightarrow(ii) ( i ) ⇒ ( ii ) . Suppose that F F F is a positive frame. Since − f i ∈ span + ( F ) -f_i\in\operatorname{span}_{+}(F) − f i ∈ span + ( F ) for every 1 ≤ i ≤ N 1\leq i\leq N 1 ≤ i ≤ N , there exist a i j ≥ 0 a_{ij}\geq0 a ij ≥ 0 such that
− f i = ∑ j = 1 N a i j f j . -f_i=\sum_{j=1}^{N}a_{ij}f_j. − f i = j = 1 ∑ N a ij f j .
Summing over i i i and rearranging yields
∑ j = 1 N ( 1 + ∑ i = 1 N a i j ) f j = 0. \sum_{j=1}^{N}\left(1+\sum_{i=1}^{N}a_{ij}\right)f_j=0. j = 1 ∑ N ( 1 + i = 1 ∑ N a ij ) f j = 0.
Setting λ j = 1 + ∑ i = 1 N a i j > 0 \lambda_j=1+\sum_{i=1}^{N}a_{ij}>0 λ j = 1 + ∑ i = 1 N a ij > 0 establishes (ii).
( i i ) ⇒ ( i i i ) (ii)\Rightarrow(iii) ( ii ) ⇒ ( iii ) . Suppose that ∑ i = 1 N λ i f i = 0 \sum_{i=1}^{N}\lambda_i f_i=0 ∑ i = 1 N λ i f i = 0 for some λ i > 0 \lambda_i>0 λ i > 0 . Normalizing the coefficients, we obtain
0 = ∑ i = 1 N α i f i , α i = λ i ∑ j = 1 N λ j > 0 , ∑ i = 1 N α i = 1. 0=\sum_{i=1}^{N}\alpha_i f_i,\qquad\alpha_i=\frac{\lambda_i}{\sum_{j=1}^{N}\lambda_j}>0,\qquad\sum_{i=1}^{N}\alpha_i=1. 0 = i = 1 ∑ N α i f i , α i = ∑ j = 1 N λ j λ i > 0 , i = 1 ∑ N α i = 1.
Thus 0 ∈ conv ( F ) 0\in\operatorname{conv}(F) 0 ∈ conv ( F ) . Assume that 0 0 0 does not belong to the interior of conv ( F ) \operatorname{conv}(F) conv ( F ) . By the supporting hyperplane theorem, there exists a nonzero x ∈ H n x\in\mathcal{H}_n x ∈ H n such that ⟨ x , f i ⟩ ≥ 0 \langle x,f_i\rangle\geq0 ⟨ x , f i ⟩ ≥ 0 for every i i i . Consequently,
0 = ⟨ x , ∑ i = 1 N λ i f i ⟩ = ∑ i = 1 N λ i ⟨ x , f i ⟩ . 0=\left\langle x,\sum_{i=1}^{N}\lambda_i f_i\right\rangle=\sum_{i=1}^{N}\lambda_i\langle x,f_i\rangle. 0 = ⟨ x , i = 1 ∑ N λ i f i ⟩ = i = 1 ∑ N λ i ⟨ x , f i ⟩ .
Since every λ i > 0 \lambda_i>0 λ i > 0 , it follows that ⟨ x , f i ⟩ = 0 \langle x,f_i\rangle=0 ⟨ x , f i ⟩ = 0 for all i i i . As F F F spans H n \mathcal{H}_n H n , we obtain x = 0 x=0 x = 0 , a contradiction. Hence 0 ∈ int ( conv ( F ) ) 0\in\operatorname{int}(\operatorname{conv}(F)) 0 ∈ int ( conv ( F )) .
( i i i ) ⇒ ( i ) (iii)\Rightarrow(i) ( iii ) ⇒ ( i ) . Suppose that 0 ∈ int ( conv ( F ) ) 0\in\operatorname{int}(\operatorname{conv}(F)) 0 ∈ int ( conv ( F )) . Then there exists ε > 0 \varepsilon>0 ε > 0 such that B ( 0 , ε ) ⊆ conv ( F ) B(0,\varepsilon)\subseteq\operatorname{conv}(F) B ( 0 , ε ) ⊆ conv ( F ) . For any nonzero x ∈ H n x\in\mathcal{H}_n x ∈ H n , we have
ε x 2 ∥ x ∥ ∈ conv ( F ) . \frac{\varepsilon x}{2\lVert x\rVert}\in\operatorname{conv}(F). 2 ∥ x ∥ ε x ∈ conv ( F ) .
Therefore, there exist c i ≥ 0 c_i\geq0 c i ≥ 0 with ∑ i = 1 N c i = 1 \sum_{i=1}^{N}c_i=1 ∑ i = 1 N c i = 1 such that
ε x 2 ∥ x ∥ = ∑ i = 1 N c i f i . \frac{\varepsilon x}{2\lVert x\rVert}=\sum_{i=1}^{N}c_i f_i. 2 ∥ x ∥ ε x = i = 1 ∑ N c i f i .
It follows that
x = ∑ i = 1 N 2 ∥ x ∥ c i ε f i ∈ span + ( F ) . x=\sum_{i=1}^{N}\frac{2\lVert x\rVert c_i}{\varepsilon}f_i\in\operatorname{span}_{+}(F). x = i = 1 ∑ N ε 2 ∥ x ∥ c i f i ∈ span + ( F ) .
Since 0 0 0 also belongs to the positive span, we conclude that span + ( F ) = H n \operatorname{span}_{+}(F)=\mathcal{H}_n span + ( F ) = H n . ∎
The preceding theorem also admits a useful operator-theoretic formulation.
Remark 3.2. Let R + + N = { c = ( c i ) i = 1 N ∈ R N : c i > 0 for all i } \mathbb{R}_{++}^{N}=\{c=(c_i)_{i=1}^{N}\in\mathbb{R}^{N}:c_i>0\text{ for all }i\} R ++ N = { c = ( c i ) i = 1 N ∈ R N : c i > 0 for all i } . Then Proposition 3.1 implies that
F F F is a positive frame ⟺ ker T F ∗ ∩ R + + N ≠ ∅ \Longleftrightarrow\ker T_{F}^{*}\cap\mathbb{R}_{++}^{N}\ne\varnothing ⟺ ker T F ∗ ∩ R ++ N = ∅ .
Thus the positivity of a frame is completely determined by the existence of a strictly positive vector in its relation space ker T F ∗ \ker T_{F}^{*} ker T F ∗ .
Example 3.3. Consider the family F = { f i } i = 1 5 F=\{f_i\}_{i=1}^{5} F = { f i } i = 1 5 in R 3 \mathbb{R}^{3} R 3 , where
f 1 = ( 1 , 0 , 0 ) , f 2 = ( 0 , 1 , 0 ) , f 3 = ( 0 , 0 , 1 ) , f 4 = ( − 1 , − 1 , 0 ) , f 5 = ( − 1 , 0 , − 1 ) f_1=(1,0,0),\quad f_2=(0,1,0),\quad f_3=(0,0,1),\quad f_4=(-1,-1,0),\quad f_5=(-1,0,-1) f 1 = ( 1 , 0 , 0 ) , f 2 = ( 0 , 1 , 0 ) , f 3 = ( 0 , 0 , 1 ) , f 4 = ( − 1 , − 1 , 0 ) , f 5 = ( − 1 , 0 , − 1 ) .
Since f 1 , f 2 , f 3 f_1,f_2,f_3 f 1 , f 2 , f 3 form the standard basis of R 3 \mathbb{R}^{3} R 3 , the family F F F is a frame with excess two. Its synthesis operator is represented by
T F ∗ = ( 1 0 0 − 1 − 1 0 1 0 − 1 0 0 0 1 0 − 1 ) . T_F^*=\begin{pmatrix}1&0&0&-1&-1\\0&1&0&-1&0\\0&0&1&0&-1\end{pmatrix}. T F ∗ = 1 0 0 0 1 0 0 0 1 − 1 − 1 0 − 1 0 − 1 . A direct computation yields
ker T F ∗ = { ( s + t , s , t , s , t ) : s , t ∈ R } . \ker T_F^*=\{(s+t,s,t,s,t):s,t\in\mathbb{R}\}. ker T F ∗ = {( s + t , s , t , s , t ) : s , t ∈ R } . In particular, ( 2 , 1 , 1 , 1 , 1 ) ∈ ker T F ∗ (2,1,1,1,1)\in\ker T_F^* ( 2 , 1 , 1 , 1 , 1 ) ∈ ker T F ∗ , and hence F F F is a positive frame by Proposition 3.1. Furthermore, the normalized strictly positive relations are precisely
{ ( 1 3 , s , 1 3 − s , s , 1 3 − s ) : 0 < s < 1 3 } . \left\{\left(\frac{1}{3},s,\frac{1}{3}-s,s,\frac{1}{3}-s\right):0<s<\frac{1}{3}\right\}. { ( 3 1 , s , 3 1 − s , s , 3 1 − s ) : 0 < s < 3 1 } . Thus, unlike the excess-one case, the normalized positive relation is not unique. Notice also that
f 1 + f 2 + f 4 = 0 , f 1 + f 3 + f 5 = 0. f_1+f_2+f_4=0,\qquad f_1+f_3+f_5=0. f 1 + f 2 + f 4 = 0 , f 1 + f 3 + f 5 = 0. These relations correspond to two distinct positively dependent subfamilies, neither of which positively spans R 3 \mathbb{R}^{3} R 3 . However, their union positively spans R 3 \mathbb{R}^{3} R 3 . Finally, the subfamily obtained by deleting f 2 f_2 f 2 still spans R 3 \mathbb{R}^{3} R 3 , but it is not positive. Indeed, every relation supported on the remaining indices must satisfy s = 0 s=0 s = 0 , and therefore cannot have all coefficients strictly positive.
We observe that the convex hull of a positive frame satisfies a useful scaling property. Indeed, let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a positive frame for H n \mathcal{H}_n H n . Since 0 ∈ int conv ( F ) 0\in\operatorname{int}\operatorname{conv}(F) 0 ∈ int conv ( F ) , there exists r > 0 r>0 r > 0 such that B ( 0 , r ) ⊆ conv ( F ) B(0,r)\subseteq\operatorname{conv}(F) B ( 0 , r ) ⊆ conv ( F ) . Now, for any x ∈ H n x\in\mathcal{H}_n x ∈ H n , choose t > ∥ x ∥ / r t>\lVert x\rVert/r t > ∥ x ∥ / r . Then
x t ∈ B ( 0 , r ) ⊆ conv ( F ) , \frac{x}{t}\in B(0,r)\subseteq\operatorname{conv}(F), t x ∈ B ( 0 , r ) ⊆ conv ( F ) ,
which implies that x ∈ t conv ( F ) x\in t\operatorname{conv}(F) x ∈ t conv ( F ) . Thus, every vector in H n \mathcal{H}_n H n belongs to a suitable positive scaling of the convex hull of F F F .
Proposition 3.4. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a positive frame for H n \mathcal{H}_n H n , and let { e i } i = 1 N \{e_i\}_{i=1}^{N} { e i } i = 1 N be the standard orthonormal basis of R N \mathbb{R}^{N} R N . Suppose that ∑ i = 1 N λ i f i = 0 \sum_{i=1}^{N}\lambda_i f_i=0 ∑ i = 1 N λ i f i = 0 , λ i > 0 \lambda_i>0 λ i > 0 , and set u = ∑ i = 1 N λ i e i u=\sum_{i=1}^{N}\lambda_i e_i u = ∑ i = 1 N λ i e i . Then, for every x ∈ H n x\in\mathcal{H}_n x ∈ H n , there exists t ≥ 0 t\geq0 t ≥ 0 such that
T F x + t u ∈ span + { e 1 , … , e N } . T_Fx+tu\in\operatorname{span}_{+}\{e_1,\ldots,e_N\}. T F x + t u ∈ span + { e 1 , … , e N } .
Proof. Let x ∈ H n x \in\mathcal{H}_n x ∈ H n be arbitrary. We then have
T F x + t u = ∑ i = 1 N ( ⟨ x , f i ⟩ + t λ i ) e i . T_Fx+tu=\sum_{i=1}^{N}(\langle x,f_i\rangle+t\lambda_i)e_i. T F x + t u = i = 1 ∑ N (⟨ x , f i ⟩ + t λ i ) e i .
Since λ i > 0 \lambda_i>0 λ i > 0 for every i i i , we may choose
t ≥ max { 0 , max 1 ≤ i ≤ N − ⟨ x , f i ⟩ λ i } . t\geq\max\left\{0,\max_{1\leq i\leq N}\frac{-\langle x,f_i\rangle}{\lambda_i}\right\}. t ≥ max { 0 , 1 ≤ i ≤ N max λ i − ⟨ x , f i ⟩ } .
Then ⟨ x , f i ⟩ + t λ i ≥ 0 \langle x,f_i\rangle+t\lambda_i\geq0 ⟨ x , f i ⟩ + t λ i ≥ 0 for every i i i . Consequently, T F x + t u ∈ span + { e 1 , … , e N } T_Fx+tu\in\operatorname{span}_{+}\{e_1,\ldots,e_N\} T F x + t u ∈ span + { e 1 , … , e N } , as required. □ \square □
Equivalence Classes of Positive Frames
In the preceding section, we characterized positive frames through strictly positive linear relations among their vectors. We now investigate the structure of positive frames under invertible linear transformations. Our main objective is to describe their equivalence classes in terms of the kernels of their synthesis operators. In particular, we show that these classes admit a natural parametrization by certain subspaces of R N \mathbb{R}^{N} R N intersecting the strictly positive orthant. The case of frames with excess one is considered separately, where the corresponding parameter space has a particularly simple geometric description.
Two frames F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N and G = { g i } i = 1 N G=\{g_i\}_{i=1}^{N} G = { g i } i = 1 N for H n \mathcal{H}_n H n are said to be equivalent if there exists an invertible operator U : H n ⟶ H n U:\mathcal{H}_n\longrightarrow\mathcal{H}_n U : H n ⟶ H n such that g i = U f i g_i=Uf_i g i = U f i , i = 1 , … , N i=1,\ldots,N i = 1 , … , N . In this case we write F ∼ G F\sim G F ∼ G . For finite frames, equivalence is completely determined by the space of linear relations. More precisely,
F ∼ G ⟺ ker T F ∗ = ker T G ∗ . F\sim G \quad\Longleftrightarrow\quad\ker T_F^*=\ker T_G^*. F ∼ G ⟺ ker T F ∗ = ker T G ∗ .
Indeed, g i = U f i g_i=Uf_i g i = U f i with U U U invertible implies T G ∗ = U T F ∗ T_G^*=UT_F^* T G ∗ = U T F ∗ , so the kernels agree. Conversely, if the kernels agree, the rule U ( T F ∗ c ) = T G ∗ c U(T_F^*c)=T_G^*c U ( T F ∗ c ) = T G ∗ c is well defined and is a linear bijection, with U f i = g i Uf_i=g_i U f i = g i . Thus equivalence preserves positivity by Theorem 3.1.
This observation, together with the existence of strictly positive relations, gives a natural quotient parametrization of positive frames. We first consider positive frames with excess one. Let
R + n + 1 = { λ = ( λ i ) i = 1 n + 1 ∈ R n + 1 : λ i ≥ 0 } \mathbb{R}_{+}^{n+1}=\{\lambda=(\lambda_i)_{i=1}^{n+1}\in\mathbb{R}^{n+1}:\lambda_i\geq0\} R + n + 1 = { λ = ( λ i ) i = 1 n + 1 ∈ R n + 1 : λ i ≥ 0 }
and
R + + n + 1 = { λ = ( λ i ) i = 1 n + 1 ∈ R n + 1 : λ i > 0 } \mathbb{R}_{++}^{n+1}=\{\lambda=(\lambda_i)_{i=1}^{n+1}\in\mathbb{R}^{n+1}:\lambda_i>0\} R ++ n + 1 = { λ = ( λ i ) i = 1 n + 1 ∈ R n + 1 : λ i > 0 }
denote the nonnegative and strictly positive orthants of R n + 1 \mathbb{R}^{n+1} R n + 1 , respectively. Define
ω = conv { e 1 , … , e n + 1 } = { λ ∈ R + n + 1 : ∑ i = 1 n + 1 λ i = 1 } . \omega=\operatorname{conv}\{e_1,\ldots,e_{n+1}\}=\left\{\lambda\in\mathbb{R}_{+}^{n+1}:\sum_{i=1}^{n+1}\lambda_i=1\right\}. ω = conv { e 1 , … , e n + 1 } = { λ ∈ R + n + 1 : i = 1 ∑ n + 1 λ i = 1 } .
where { e i } i = 1 n + 1 \{e_i\}_{i=1}^{n+1} { e i } i = 1 n + 1 is the standard basis of R n + 1 \mathbb{R}^{n+1} R n + 1 . Its relative interior is given by
ω ˚ = { λ ∈ R + + n + 1 : ∑ i = 1 n + 1 λ i = 1 } . \mathring{\omega}=\left\{\lambda\in\mathbb{R}_{++}^{n+1}:\sum_{i=1}^{n+1}\lambda_i=1\right\}. ω ˚ = { λ ∈ R ++ n + 1 : i = 1 ∑ n + 1 λ i = 1 } .
The following proposition characterizes the equivalence classes of positive frames with excess one in terms of the relative interior of the standard simplex.
Proposition 3.5. The equivalence classes of ordered positive frames consisting of n + 1 n+1 n + 1 vectors in an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n are in one-to-one correspondence with the points of ω ˚ \mathring{\omega} ω ˚ .
Proof. Let F = ( f i ) i = 1 n + 1 F=(f_i)_{i=1}^{n+1} F = ( f i ) i = 1 n + 1 be positive. Since T F ∗ T_F^* T F ∗ is surjective, its kernel has dimension one. By Theorem 3.1, this kernel contains a strictly positive vector a = ( a i ) a=(a_i) a = ( a i ) . Thus
ker T F ∗ = span { a } , λ i = a i ∑ j = 1 n + 1 a j \ker T_F^*=\operatorname{span}\{a\},\qquad\lambda_i=\frac{a_i}{\sum_{j=1}^{n+1}a_j} ker T F ∗ = span { a } , λ i = ∑ j = 1 n + 1 a j a i
defines an element λ ∈ ω ˚ \lambda\in\mathring{\omega} λ ∈ ω ˚ with ∑ i λ i f i = 0 \sum_i\lambda_i f_i=0 ∑ i λ i f i = 0 . This normalized positive relation is unique: any other such vector μ \mu μ lies in the same one-dimensional kernel, so μ = t λ \mu=t\lambda μ = t λ for some t > 0 t>0 t > 0 , and summing the coordinates gives t = 1 t=1 t = 1 .
Define Φ ( [ F ] ) = λ \Phi([F])=\lambda Φ ([ F ]) = λ . Equivalent frames have the same synthesis kernel, so this map is well defined. If Φ ( [ F ] ) = Φ ( [ G ] ) \Phi([F])=\Phi([G]) Φ ([ F ]) = Φ ([ G ]) , then ker T F ∗ = span { λ } = ker T G ∗ \ker T_F^*=\operatorname{span}\{\lambda\}=\ker T_G^* ker T F ∗ = span { λ } = ker T G ∗ ; the kernel criterion above implies F ∼ G F\sim G F ∼ G . Hence Φ \Phi Φ is injective.
To prove surjectivity, fix λ ∈ ω ˚ \lambda\in\mathring{\omega} λ ∈ ω ˚ , put K = span { λ } K=\operatorname{span}\{\lambda\} K = span { λ } , and let P P P be the orthogonal projection onto K ⊥ ⊂ R n + 1 K^\perp\subset\mathbb{R}^{n+1} K ⊥ ⊂ R n + 1 . Define f i = P e i f_i=Pe_i f i = P e i . The family ( P e i ) (Pe_i) ( P e i ) spans the n n n -dimensional real Hilbert space K ⊥ K^\perp K ⊥ , and
∑ i = 1 n + 1 λ i P e i = P λ = 0. \sum_{i=1}^{n+1}\lambda_iPe_i=P\lambda=0. i = 1 ∑ n + 1 λ i P e i = P λ = 0.
It is therefore positive. Its synthesis operator is P P P , regarded as a map into K ⊥ K^\perp K ⊥ , with kernel K K K . Consequently its normalized positive relation is precisely λ \lambda λ . Finally, choose an isometric isomorphism from K ⊥ K^\perp K ⊥ onto H n \mathcal{H}_n H n to obtain a representative in the fixed ambient space. This proves surjectivity and the asserted bijection. □
Remark 3.6. For each λ ∈ ω ˚ \lambda\in\mathring{\omega} λ ∈ ω ˚ , one may obtain a concrete representative of the corresponding equivalence class without using a quotient space. Consider
λ ⊥ = { x ∈ R n + 1 : ⟨ x , λ ⟩ = 0 } , \lambda^\perp=\left\{x\in\mathbb{R}^{n+1}:\langle x,\lambda\rangle=0\right\}, λ ⊥ = { x ∈ R n + 1 : ⟨ x , λ ⟩ = 0 } ,
which is an n n n -dimensional subspace of R n + 1 \mathbb{R}^{n+1} R n + 1 . Let P λ ⊥ P_{\lambda^\perp} P λ ⊥ denote the orthogonal projection onto λ ⊥ \lambda^\perp λ ⊥ , and define f i λ = P λ ⊥ e i f_i^\lambda=P_{\lambda^\perp}e_i f i λ = P λ ⊥ e i , i = 1 , … , n + 1 i=1,\ldots,n+1 i = 1 , … , n + 1 . Then
∑ i = 1 n + 1 λ i f i λ = P λ ⊥ ( ∑ i = 1 n + 1 λ i e i ) = P λ ⊥ λ = 0. \sum_{i=1}^{n+1}\lambda_i f_i^\lambda=P_{\lambda^\perp}\left(\sum_{i=1}^{n+1}\lambda_i e_i\right)=P_{\lambda^\perp}\lambda=0. i = 1 ∑ n + 1 λ i f i λ = P λ ⊥ ( i = 1 ∑ n + 1 λ i e i ) = P λ ⊥ λ = 0.
Thus
{ P λ ⊥ e i } i = 1 n + 1 \{P_{\lambda^\perp}e_i\}_{i=1}^{n+1} { P λ ⊥ e i } i = 1 n + 1
is a positive frame for λ ⊥ \lambda^\perp λ ⊥ and provides a canonical representative of the equivalence class corresponding to λ ∈ ω ˚ \lambda\in\mathring{\omega} λ ∈ ω ˚ .
Remark 3.7. The preceding proposition concerns ordered frames . If permutations of the frame vectors are also included in the equivalence relation, then the corresponding parameter space is the quotient ω ˚ / S n + 1 \mathring{\omega}/S_{n+1} ω ˚ / S n + 1 , where the symmetric group S n + 1 S_{n+1} S n + 1 acts on ω ˚ \mathring{\omega} ω ˚ by permuting coordinates.
For the general parametrization, let k = N − n k=N-n k = N − n and use the strictly positive orthant R + + N \mathbb{R}_{++}^N R ++ N . Define
A k = { ( λ ( 1 ) , … , λ ( k ) ) ∈ ( R + + N ) k : λ ( 1 ) , … , λ ( k ) are linearly independent } . \mathcal{A}_k=\left\{(\lambda^{(1)},\ldots,\lambda^{(k)})\in(\mathbb{R}_{++}^N)^k:\lambda^{(1)},\ldots,\lambda^{(k)}\text{ are linearly independent}\right\}. A k = { ( λ ( 1 ) , … , λ ( k ) ) ∈ ( R ++ N ) k : λ ( 1 ) , … , λ ( k ) are linearly independent } .
On A k \mathcal{A}_k A k , define the relation ≈ \approx ≈ by
( λ ( 1 ) , … , λ ( k ) ) ≈ ( μ ( 1 ) , … , μ ( k ) ) (\lambda^{(1)},\ldots,\lambda^{(k)})\approx(\mu^{(1)},\ldots,\mu^{(k)}) ( λ ( 1 ) , … , λ ( k ) ) ≈ ( μ ( 1 ) , … , μ ( k ) )
if and only if span { λ ( 1 ) , … , λ ( k ) } = span { μ ( 1 ) , … , μ ( k ) } \operatorname{span}\{\lambda^{(1)},\ldots,\lambda^{(k)}\}=\operatorname{span}\{\mu^{(1)},\ldots,\mu^{(k)}\} span { λ ( 1 ) , … , λ ( k ) } = span { μ ( 1 ) , … , μ ( k ) } .
We first consider the construction of positive frames through orthogonal projections. Let { e i } i = 1 N \{e_i\}_{i=1}^{N} { e i } i = 1 N be an orthonormal basis of R N \mathbb{R}^{N} R N , and let x = ∑ i = 1 N λ i e i x=\sum_{i=1}^{N}\lambda_i e_i x = ∑ i = 1 N λ i e i be a nonzero vector with λ i ≥ 0 \lambda_i\geq0 λ i ≥ 0 for every i i i . Denote by P P P the orthogonal projection onto x ⊥ x^\perp x ⊥ . We now characterize when the family { P e i } i = 1 N \{Pe_i\}_{i=1}^{N} { P e i } i = 1 N forms a positive frame for x ⊥ x^\perp x ⊥ in terms of the coefficients of x x x .
Proposition 3.8. The family { P e i } i = 1 N \{Pe_i\}_{i=1}^{N} { P e i } i = 1 N is a positive frame for x ⊥ x^\perp x ⊥ if and only if λ i > 0 \lambda_i>0 λ i > 0 for every i i i .
Proof. Since P P P is the orthogonal projection onto x ⊥ x^\perp x ⊥ , the family { P e i } i = 1 N \{Pe_i\}_{i=1}^{N} { P e i } i = 1 N spans x ⊥ x^\perp x ⊥ . Suppose first that λ i > 0 \lambda_i>0 λ i > 0 for every i i i . Since P x = 0 Px=0 P x = 0 , we have
∑ i = 1 N λ i P e i = P ( ∑ i = 1 N λ i e i ) = P x = 0. \sum_{i=1}^{N}\lambda_iPe_i=P\left(\sum_{i=1}^{N}\lambda_i e_i\right)=Px=0. i = 1 ∑ N λ i P e i = P ( i = 1 ∑ N λ i e i ) = P x = 0.
Thus { P e i } i = 1 N \{Pe_i\}_{i=1}^{N} { P e i } i = 1 N admits a strictly positive linear relation. Since it spans x ⊥ x^\perp x ⊥ , it is a positive frame. Conversely, suppose that { P e i } i = 1 N \{Pe_i\}_{i=1}^{N} { P e i } i = 1 N is a positive frame for x ⊥ x^\perp x ⊥ . Then there exist c i > 0 c_i>0 c i > 0 such that
∑ i = 1 N c i P e i = 0. \sum_{i=1}^{N}c_iPe_i=0. i = 1 ∑ N c i P e i = 0.
Consequently,
P ( ∑ i = 1 N c i e i ) = 0 , P\left(\sum_{i=1}^{N}c_i e_i\right)=0, P ( i = 1 ∑ N c i e i ) = 0 ,
which implies ∑ i = 1 N c i e i ∈ ker P = span { x } \sum_{i=1}^{N}c_i e_i\in\ker P=\operatorname{span}\{x\} ∑ i = 1 N c i e i ∈ ker P = span { x } . Hence there exists t ≠ 0 t\neq0 t = 0 such that
∑ i = 1 N c i e i = t ∑ i = 1 N λ i e i . \sum_{i=1}^{N}c_i e_i=t\sum_{i=1}^{N}\lambda_i e_i. i = 1 ∑ N c i e i = t i = 1 ∑ N λ i e i .
By the linear independence of { e i } i = 1 N \{e_i\}_{i=1}^{N} { e i } i = 1 N , we obtain c i = t λ i c_i=t\lambda_i c i = t λ i for every i i i . Since c i > 0 c_i>0 c i > 0 and λ i ≥ 0 \lambda_i\geq0 λ i ≥ 0 , it follows that λ i > 0 \lambda_i>0 λ i > 0 for every i i i . □ \square □
Example 3.9. We construct positive Parseval frames for R 3 \mathbb{R}^{3} R 3 that are not equivalent despite having the same frame operator. We also illustrate that identical vector norms and the Parseval property do not guarantee positive spanning.
Let
u = 1 2 ( 1 , 1 , 1 , 1 ) T , v = 1 2 ( 1 , 1 , − 1 , − 1 ) T , u=\frac{1}{2}(1,1,1,1)^{T},\qquad v=\frac{1}{2}(1,1,-1,-1)^{T}, u = 2 1 ( 1 , 1 , 1 , 1 ) T , v = 2 1 ( 1 , 1 , − 1 , − 1 ) T , and denote by P u P_u P u and P v P_v P v the orthogonal projections onto u ⊥ u^\perp u ⊥ and v ⊥ v^\perp v ⊥ , respectively. Consider the families
F = { P u e i } i = 1 4 , G = { P v e i } i = 1 4 . F=\{P_ue_i\}_{i=1}^{4},\qquad G=\{P_ve_i\}_{i=1}^{4}. F = { P u e i } i = 1 4 , G = { P v e i } i = 1 4 . Both families are Parseval frames for their respective three-dimensional subspaces. Moreover,
∥ P u e i ∥ 2 = ∥ P v e i ∥ 2 = 3 4 , 1 ≤ i ≤ 4. \lVert P_ue_i\rVert^2=\lVert P_ve_i\rVert^2=\frac{3}{4},\qquad1\leq i\leq4. ∥ P u e i ∥ 2 = ∥ P v e i ∥ 2 = 4 3 , 1 ≤ i ≤ 4. Since u u u has strictly positive coordinates, F F F is a positive frame. However, G G G is not positive because its synthesis kernel is span { v } \operatorname{span}\{v\} span { v } , which contains no strictly positive vector.
Now let
w = 1 10 ( 1 , 1 , 2 , 2 ) T w=\frac{1}{\sqrt{10}}(1,1,2,2)^T w = 10 1 ( 1 , 1 , 2 , 2 ) T
and consider the positive Parseval frame H = { P w ⊥ e i } i = 1 4 H=\{P_{w^\perp}e_i\}_{i=1}^4 H = { P w ⊥ e i } i = 1 4 for w ⊥ w^\perp w ⊥ . Its synthesis kernel is span { w } \operatorname{span}\{w\} span { w } , so H H H is positive.
Although F F F and H H H are both positive Parseval frames with four vectors in three-dimensional Hilbert spaces, they are not equivalent. Indeed,
ker T F ∗ = span { u } , ker T H ∗ = span { w } , \ker T_F^*=\operatorname{span}\{u\},\qquad\ker T_H^*=\operatorname{span}\{w\}, ker T F ∗ = span { u } , ker T H ∗ = span { w } ,
and these subspaces are distinct. Furthermore, their normalized strictly positive relations are λ F = ( 1 4 , 1 4 , 1 4 , 1 4 ) \lambda_F=\left(\frac{1}{4},\frac{1}{4},\frac{1}{4},\frac{1}{4}\right) λ F = ( 4 1 , 4 1 , 4 1 , 4 1 ) and λ H = ( 1 6 , 1 6 , 1 3 , 1 3 ) \lambda_H=\left(\frac{1}{6},\frac{1}{6},\frac{1}{3},\frac{1}{3}\right) λ H = ( 6 1 , 6 1 , 3 1 , 3 1 ) , respectively. Thus Proposition 3.5 distinguishes their equivalence classes, even though both frames are Parseval. This example demonstrates that the equivalence classification of positive frames is governed by their relation spaces rather than by the Parseval property alone.
Higher Excess and Grassmannian Parametrization
We now extend the preceding classification to positive frames with arbitrary excess. Using strictly positive linear relations, we establish a parametrization of their equivalence classes through the kernels of their synthesis operators. This leads to a geometric description in terms of suitable subsets of Grassmannians, providing a unified classification of positive frames beyond the excess-one case.
Theorem 3.10. Let N > n N>n N > n and k = N − n k=N-n k = N − n . Then the equivalence classes of positive frames with N N N elements in H n \mathcal{H}_n H n admit the quotient parametrization
{ [ F ] : F is a positive frame in H n } ≅ A k / ≈ . \{[F]:F\text{ is a positive frame in }\mathcal{H}_n\}\cong\mathcal{A}_k/\mathord{\approx}. {[ F ] : F is a positive frame in H n } ≅ A k / ≈ . Here ≈ \approx ≈ identifies tuples having the same linear span.
Proof. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^N F = { f i } i = 1 N be a positive frame for H n \mathcal{H}_n H n . Since T F ∗ : R N ⟶ H n T_F^*:\mathbb{R}^N\longrightarrow\mathcal{H}_n T F ∗ : R N ⟶ H n is surjective, the rank-nullity theorem gives dim ker T F ∗ = N − n = k \dim\ker T_F^*=N-n=k dim ker T F ∗ = N − n = k . Since F F F is positive, Theorem 3.1 implies that there exists λ ∈ ker T F ∗ ∩ R + + N \lambda\in\ker T_F^*\cap\mathbb{R}_{++}^N λ ∈ ker T F ∗ ∩ R ++ N .
We first show that ker T F ∗ \ker T_F^* ker T F ∗ possesses a basis consisting entirely of strictly positive vectors. Choose vectors η ( 2 ) , … , η ( k ) \eta^{(2)},\ldots,\eta^{(k)} η ( 2 ) , … , η ( k ) such that λ , η ( 2 ) , … , η ( k ) \lambda,\eta^{(2)},\ldots,\eta^{(k)} λ , η ( 2 ) , … , η ( k ) is a basis for ker T F ∗ \ker T_F^* ker T F ∗ . Set λ ( 1 ) = λ \lambda^{(1)}=\lambda λ ( 1 ) = λ . For every j = 2 , … , k j=2,\ldots,k j = 2 , … , k , choose ε j > 0 \varepsilon_j>0 ε j > 0 sufficiently small and define λ ( j ) = λ + ε j η ( j ) \lambda^{(j)}=\lambda+\varepsilon_j\eta^{(j)} λ ( j ) = λ + ε j η ( j ) . Since every coordinate of λ \lambda λ is strictly positive and there are only finitely many coordinates, ε j \varepsilon_j ε j may be chosen small enough that λ ( j ) ∈ R + + N \lambda^{(j)}\in\mathbb{R}_{++}^N λ ( j ) ∈ R ++ N .
We claim that λ ( 1 ) , … , λ ( k ) \lambda^{(1)},\ldots,\lambda^{(k)} λ ( 1 ) , … , λ ( k ) are linearly independent. Indeed, suppose
∑ j = 1 k a j λ ( j ) = 0. \sum_{j=1}^{k}a_j\lambda^{(j)}=0. j = 1 ∑ k a j λ ( j ) = 0.
Using λ ( 1 ) = λ \lambda^{(1)}=\lambda λ ( 1 ) = λ , λ ( j ) = λ + ε j η ( j ) \lambda^{(j)}=\lambda+\varepsilon_j\eta^{(j)} λ ( j ) = λ + ε j η ( j ) (j ≥ 2 j\ge2 j ≥ 2 ), we obtain
( ∑ j = 1 k a j ) λ + ∑ j = 2 k a j ε j η ( j ) = 0. (2) \left(\sum_{j=1}^{k}a_j\right)\lambda+\sum_{j=2}^{k}a_j\varepsilon_j\eta^{(j)}=0. \tag*{(2)} ( j = 1 ∑ k a j ) λ + j = 2 ∑ k a j ε j η ( j ) = 0. (2)
Since λ , η ( 2 ) , … , η ( k ) \lambda,\eta^{(2)},\ldots,\eta^{(k)} λ , η ( 2 ) , … , η ( k ) are linearly independent, it follows that a j ε j = 0 a_j\varepsilon_j=0 a j ε j = 0 , j = 2 , … , k j=2,\ldots,k j = 2 , … , k . Since ε j > 0 \varepsilon_j>0 ε j > 0 , we have a j = 0 a_j=0 a j = 0 , j = 2 , … , k j=2,\ldots,k j = 2 , … , k . It then follows that a 1 = 0 a_1=0 a 1 = 0 . Hence the vectors are linearly independent.
Consequently, span { λ ( 1 ) , … , λ ( k ) } = ker T F ∗ \operatorname{span}\{\lambda^{(1)},\ldots,\lambda^{(k)}\}=\ker T_F^* span { λ ( 1 ) , … , λ ( k ) } = ker T F ∗ . Thus every positive frame determines an element of A k / ≈ \mathcal{A}_k/\approx A k / ≈ .
Now let F F F and G G G be positive frames. Then F ∼ G F\sim G F ∼ G if and only if ker T F ∗ = ker T G ∗ \ker T_F^*=\ker T_G^* ker T F ∗ = ker T G ∗ . Therefore, if { λ ( 1 ) , … , λ ( k ) } \{\lambda^{(1)},\ldots,\lambda^{(k)}\} { λ ( 1 ) , … , λ ( k ) } and { μ ( 1 ) , … , μ ( k ) } \{\mu^{(1)},\ldots,\mu^{(k)}\} { μ ( 1 ) , … , μ ( k ) } are strictly positive bases for ker T F ∗ \ker T_F^* ker T F ∗ and ker T G ∗ \ker T_G^* ker T G ∗ , respectively, then F ∼ G F\sim G F ∼ G if and only if span { λ ( 1 ) , … , λ ( k ) } = span { μ ( 1 ) , … , μ ( k ) } \operatorname{span}\{\lambda^{(1)},\ldots,\lambda^{(k)}\}=\operatorname{span}\{\mu^{(1)},\ldots,\mu^{(k)}\} span { λ ( 1 ) , … , λ ( k ) } = span { μ ( 1 ) , … , μ ( k ) } . Hence the quotient removes precisely the choice of basis and retains only the relation space.
Conversely, let ( λ ( 1 ) , … , λ ( k ) ) ∈ A k (\lambda^{(1)},\ldots,\lambda^{(k)})\in\mathcal{A}_k ( λ ( 1 ) , … , λ ( k ) ) ∈ A k and set K = span { λ ( 1 ) , … , λ ( k ) } K=\operatorname{span}\{\lambda^{(1)},\ldots,\lambda^{(k)}\} K = span { λ ( 1 ) , … , λ ( k ) } . Then dim K = k = N − n \dim K=k=N-n dim K = k = N − n , so dim K ⊥ = n \dim K^\perp=n dim K ⊥ = n . Let P K ⊥ P_{K^\perp} P K ⊥ be the orthogonal projection of R N \mathbb{R}^N R N onto K ⊥ K^\perp K ⊥ , and choose a linear isomorphism U : K ⊥ → H n U:K^\perp\to\mathcal{H}_n U : K ⊥ → H n . Define
f i = U P K ⊥ e i , i = 1 , … , N . f_i=UP_{K^\perp}e_i,\qquad i=1,\ldots,N. f i = U P K ⊥ e i , i = 1 , … , N .
Then T F ∗ = U P K ⊥ T_F^*=UP_{K^\perp} T F ∗ = U P K ⊥ , which is surjective, and ker T F ∗ = K \ker T_F^*=K ker T F ∗ = K . Since λ ( 1 ) ∈ K ∩ R + + N \lambda^{(1)}\in K\cap\mathbb{R}_{++}^N λ ( 1 ) ∈ K ∩ R ++ N , we obtain ∑ i = 1 N λ i ( 1 ) f i = 0 \sum_{i=1}^N\lambda_i^{(1)}f_i=0 ∑ i = 1 N λ i ( 1 ) f i = 0 . Consequently, F F F is a positive frame, and its synthesis kernel is precisely K K K . Therefore every element of A k / ≈ \mathcal{A}_k/\approx A k / ≈ determines an equivalence class of positive frames, and the correspondence is bijective. □
The preceding quotient description can also be expressed more intrinsically in terms of the Grassmannian.
Example 3.12. We construct two inequivalent positive Parseval frames for R 3 \mathbb{R}^3 R 3 having the same number of vectors. Consider
u = 1 5 ( 1 , 1 , 1 , 1 , 1 ) T u=\frac{1}{\sqrt{5}}(1,1,1,1,1)^T u = 5 1 ( 1 , 1 , 1 , 1 , 1 ) T and
v = 1 10 ( 1 , − 1 , 0 , 2 , − 2 ) T , w = 1 10 ( 2 , − 2 , 0 , 1 , − 1 ) T . v=\frac{1}{\sqrt{10}}(1,-1,0,2,-2)^T,\qquad w=\frac{1}{\sqrt{10}}(2,-2,0,1,-1)^T. v = 10 1 ( 1 , − 1 , 0 , 2 , − 2 ) T , w = 10 1 ( 2 , − 2 , 0 , 1 , − 1 ) T . Observe that { u , v } \{u,v\} { u , v } and { u , w } \{u,w\} { u , w } are orthonormal sets in R 5 \mathbb{R}^5 R 5 . Set
K = span { u , v } , L = span { u , w } . K=\operatorname{span}\{u,v\},\qquad L=\operatorname{span}\{u,w\}. K = span { u , v } , L = span { u , w } . Let A , B ∈ R 5 × 3 A,B\in\mathbb{R}^{5\times3} A , B ∈ R 5 × 3 be matrices whose columns form orthonormal bases of K ⊥ K^\perp K ⊥ and L ⊥ L^\perp L ⊥ , respectively. Define F = { f i } i = 1 5 F=\{f_i\}_{i=1}^5 F = { f i } i = 1 5 and G = { g i } i = 1 5 G=\{g_i\}_{i=1}^5 G = { g i } i = 1 5 by taking their frame vectors to be the rows of A A A and B B B , respectively. Since A T A = B T B = I 3 A^TA=B^TB=I_3 A T A = B T B = I 3 , both F F F and G G G are Parseval frames. Furthermore,
ker T F ∗ = K , ker T G ∗ = L . \ker T_F^*=K,\qquad\ker T_G^*=L. ker T F ∗ = K , ker T G ∗ = L .
As u u u is strictly positive and belongs to both kernels, Theorem 3.1 implies that both frames are positive.
On the other hand, K ≠ L K \ne L K = L , since v v v and w w w are not scalar multiples and both are orthogonal to u u u . Hence
ker T F ∗ ≠ ker T G ∗ . \ker T_F^* \ne\ker T_G^*. ker T F ∗ = ker T G ∗ .
By the characterization of frame equivalence in terms of synthesis kernels, F F F and G G G are not equivalent, despite satisfying
S F = S G = I 3 . S_F=S_G=I_3. S F = S G = I 3 .
Thus the frame operator alone does not determine the equivalence class of a positive frame.
Positive Bases and Minimal Positive Dependence
We now examine the minimality properties of positive spanning families and their connections with positive linear dependence. In particular, we consider positive bases, which are positive frames for which no proper subfamily retains the positive spanning property. These notions provide a natural setting for investigating the geometric structure of positive frames and the role of their linear relations.
Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^N F = { f i } i = 1 N be a finite family in H n \mathcal{H}_n H n . Recall that F F F is called a positive basis if span + ( F ) = H n \operatorname{span}_+(F)=\mathcal{H}_n span + ( F ) = H n and no proper subfamily of F F F positively spans H n \mathcal{H}_n H n . A subfamily { f i } i ∈ I \{f_i\}_{i\in I} { f i } i ∈ I is said to be minimally positively dependent if there exist scalars λ i > 0 \lambda_i>0 λ i > 0 , i ∈ I i\in I i ∈ I , such that
∑ i ∈ I λ i f i = 0 , \sum_{i\in I}\lambda_i f_i=0, i ∈ I ∑ λ i f i = 0 ,
but no proper subfamily admits a nontrivial nonnegative linear relation. Thus, minimal positive dependence describes a positive relation that cannot be supported on a smaller collection of vectors.
For a coefficient vector λ = ( λ i ) i = 1 N \lambda=(\lambda_i)_{i=1}^N λ = ( λ i ) i = 1 N , we write
supp ( λ ) = { i : λ i ≠ 0 } . \operatorname{supp}(\lambda)=\{i:\lambda_i\ne0\}. supp ( λ ) = { i : λ i = 0 } .
We shall use these notions to examine positive bases through their geometric properties and the supports of nonnegative linear relations.
Proposition 3.13. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^N F = { f i } i = 1 N be a frame for R n \mathbb{R}^n R n . Then F F F is a positive basis if and only if the following conditions hold:
(i) For every 0 ≠ b ∈ R n 0\ne b\in\mathbb{R}^n 0 = b ∈ R n , there exists an index i i i such that ⟨ b , f i ⟩ > 0 \langle b,f_i\rangle>0 ⟨ b , f i ⟩ > 0 .
(ii) For every i ∈ { 1 , … , N } i\in\{1,\ldots,N\} i ∈ { 1 , … , N } , there exists b i ∈ R n b_i\in\mathbb{R}^n b i ∈ R n such that
⟨ b i , f i ⟩ > 0 and ⟨ b i , f j ⟩ ≤ 0 for all j ≠ i . \langle b_i,f_i\rangle>0 \quad\text{and}\quad\langle b_i,f_j\rangle\le0 \quad\text{for all }j\ne i. ⟨ b i , f i ⟩ > 0 and ⟨ b i , f j ⟩ ≤ 0 for all j = i .
Proof. Suppose first that F F F is a positive basis. Then there exist strictly positive scalars λ 1 , … , λ N \lambda_1,\ldots,\lambda_N λ 1 , … , λ N such that
∑ i = 1 N λ i f i = 0. \sum_{i=1}^N\lambda_i f_i=0. i = 1 ∑ N λ i f i = 0.
Let 0 ≠ b ∈ R n 0\ne b\in\mathbb{R}^n 0 = b ∈ R n . If ⟨ b , f i ⟩ ≤ 0 \langle b,f_i\rangle\le0 ⟨ b , f i ⟩ ≤ 0 for every i i i , then the positive spanning property of F F F implies that b = 0 b=0 b = 0 , a contradiction. Therefore, there exists an index i i i such that ⟨ b , f i ⟩ > 0 \langle b,f_i\rangle>0 ⟨ b , f i ⟩ > 0 , proving (i). To establish (ii), fix i ∈ { 1 , … , N } i \in\{1,\ldots,N\} i ∈ { 1 , … , N } . Since F F F is a positive basis, we have
f i ∉ pos { f j : j ≠ i } . f_i \notin\operatorname{pos}\{f_j:j\ne i\}. f i ∈ / pos { f j : j = i } .
Let
C i = pos { f j : j ≠ i } . C_i=\operatorname{pos}\{f_j:j\ne i\}. C i = pos { f j : j = i } .
Since C i C_i C i is a finitely generated convex cone, it is closed and convex. Hence, there exists z ∈ C i z\in C_i z ∈ C i such that
∥ f i − z ∥ = dist ( f i , C i ) . \lVert f_i-z\rVert=\operatorname{dist}(f_i,C_i). ∥ f i − z ∥ = dist ( f i , C i ) .
Set b i = f i − z ≠ 0 b_i=f_i-z\ne0 b i = f i − z = 0 .
By the characterization of the metric projection onto a closed convex set,
⟨ b i , y − z ⟩ ≤ 0 , y ∈ C i . \langle b_i,y-z\rangle\le0,\qquad y\in C_i. ⟨ b i , y − z ⟩ ≤ 0 , y ∈ C i .
Since C i C_i C i is a convex cone, for every s ≥ 0 s\ge0 s ≥ 0 we have s z ∈ C i sz\in C_i sz ∈ C i . Taking y = s z y=sz y = sz gives
( s − 1 ) ⟨ b i , z ⟩ ≤ 0. (s-1)\langle b_i,z\rangle\le0. ( s − 1 ) ⟨ b i , z ⟩ ≤ 0.
Choosing s = 0 s=0 s = 0 and s = 2 s=2 s = 2 , we obtain ⟨ b i , z ⟩ = 0 \langle b_i,z\rangle=0 ⟨ b i , z ⟩ = 0 . Consequently,
⟨ b i , y ⟩ ≤ 0 , y ∈ C i . \langle b_i,y\rangle\le0,\qquad y\in C_i. ⟨ b i , y ⟩ ≤ 0 , y ∈ C i .
In particular,
⟨ b i , f j ⟩ ≤ 0 for all j ≠ i . \langle b_i,f_j\rangle\le0\qquad\text{for all }j\ne i. ⟨ b i , f j ⟩ ≤ 0 for all j = i .
Moreover,
⟨ b i , f i ⟩ = ⟨ b i , b i + z ⟩ = ∥ b i ∥ 2 + ⟨ b i , z ⟩ = ∥ b i ∥ 2 > 0. \begin{aligned}
\langle b_i,f_i\rangle&=\langle b_i,b_i+z\rangle\\
&=\lVert b_i\rVert^2+\langle b_i,z\rangle\\
&=\lVert b_i\rVert^2>0.
\end{aligned} ⟨ b i , f i ⟩ = ⟨ b i , b i + z ⟩ = ∥ b i ∥ 2 + ⟨ b i , z ⟩ = ∥ b i ∥ 2 > 0.
Thus condition (ii) holds.
Conversely, suppose that (i) and (ii) are satisfied. We first show that F F F positively spans R n \mathbb{R}^n R n . Assume, to the contrary, that pos ( F ) ≠ R n \operatorname{pos}(F)\ne\mathbb{R}^n pos ( F ) = R n . Since pos ( F ) \operatorname{pos}(F) pos ( F ) is a closed convex cone, the separation theorem implies that there exists 0 ≠ b ∈ R n 0\ne b\in\mathbb{R}^n 0 = b ∈ R n such that
⟨ b , f i ⟩ ≤ 0 for every i . \langle b,f_i\rangle\le0\qquad\text{for every }i. ⟨ b , f i ⟩ ≤ 0 for every i .
This contradicts (i). Hence,
pos ( F ) = R n . \operatorname{pos}(F)=\mathbb{R}^n. pos ( F ) = R n .
It remains to prove that F F F is exact. Suppose otherwise. Then there exists an index i i i such that
f i ∈ pos { f j : j ≠ i } . f_i\in\operatorname{pos}\{f_j:j\ne i\}. f i ∈ pos { f j : j = i } .
Thus, for some coefficients λ j ≥ 0 \lambda_j\ge0 λ j ≥ 0 ,
f i = ∑ j ≠ i λ j f j . f_i=\sum_{j\ne i}\lambda_jf_j. f i = j = i ∑ λ j f j .
By condition (ii), there exists b i ∈ R n b_i\in\mathbb{R}^n b i ∈ R n satisfying ⟨ b i , f i ⟩ > 0 \langle b_i,f_i\rangle>0 ⟨ b i , f i ⟩ > 0 and ⟨ b i , f j ⟩ ≤ 0 \langle b_i,f_j\rangle\le0 ⟨ b i , f j ⟩ ≤ 0 for every j ≠ i j\ne i j = i . Therefore,
⟨ b i , f i ⟩ = ∑ j ≠ i λ j ⟨ b i , f j ⟩ ≤ 0 , \langle b_i,f_i\rangle=\sum_{j\ne i}\lambda_j\langle b_i,f_j\rangle\le0, ⟨ b i , f i ⟩ = j = i ∑ λ j ⟨ b i , f j ⟩ ≤ 0 ,
which is a contradiction. Hence no vector can be removed from F F F while preserving positive spanning. Therefore, F F F is a positive basis.
Proposition 3.14. Let F I = { f i : i ∈ I } F_I = \{f_i : i \in I\} F I = { f i : i ∈ I } be a minimal positively dependent family of nonzero vectors in R n \mathbb{R}^n R n . Then
∣ I ∣ = dim S I + 1. |I| = \dim S_I + 1. ∣ I ∣ = dim S I + 1. Proof. Set d = dim S I d = \dim S_I d = dim S I . Since F I F_I F I admits a nontrivial linear dependence, we have
∣ I ∣ ≥ d + 1. |I| \ge d + 1. ∣ I ∣ ≥ d + 1. We claim that every proper subfamily of F I F_I F I is linearly independent. Suppose otherwise. Then there exists a proper subset J ⊊ I J \subsetneq I J ⊊ I and scalars a i a_i a i , not all zero, such that
∑ i ∈ J a i f i = 0. \sum_{i \in J} a_i f_i = 0. i ∈ J ∑ a i f i = 0. Extending these coefficients by setting a i = 0 a_i = 0 a i = 0 for i ∈ I ∖ J i \in I \setminus J i ∈ I ∖ J , we obtain
∑ i ∈ I a i f i = 0. \sum_{i \in I} a_i f_i = 0. i ∈ I ∑ a i f i = 0. Replacing a a a by − a -a − a , if necessary, we may assume that a i > 0 a_i > 0 a i > 0 for at least one i ∈ I i \in I i ∈ I . By hypothesis, there exist c i > 0 c_i > 0 c i > 0 , i ∈ I i \in I i ∈ I , satisfying
∑ i ∈ I c i f i = 0. \sum_{i \in I} c_i f_i = 0. i ∈ I ∑ c i f i = 0. Consequently, for every t ∈ R t \in\mathbb{R} t ∈ R ,
∑ i ∈ I ( c i − t a i ) f i = 0. \sum_{i \in I} (c_i - t a_i) f_i = 0. i ∈ I ∑ ( c i − t a i ) f i = 0. Choose
t 0 = min i ∈ I a i > 0 c i a i > 0. t_0 = \min_{\substack{i \in I \\ a_i > 0}} \frac{c_i}{a_i} > 0. t 0 = i ∈ I a i > 0 min a i c i > 0. Then c i − t 0 a i ≥ 0 c_i - t_0a_i \ge0 c i − t 0 a i ≥ 0 for every i ∈ I i \in I i ∈ I , and equality holds for at least one index i 0 i_0 i 0 . Moreover, since a i = 0 a_i = 0 a i = 0 for i ∈ I ∖ J i \in I \setminus J i ∈ I ∖ J , we have c i − t 0 a i = c i > 0 c_i - t_0a_i = c_i > 0 c i − t 0 a i = c i > 0 for those indices. Thus the resulting relation is nontrivial and has nonnegative coefficients, with at least one coefficient equal to zero. This contradicts the minimal positive dependence of F I F_I F I .
Therefore every proper subfamily of F I F_I F I is linearly independent. In particular, removing any one vector from F I F_I F I leaves a linearly independent family of cardinality ∣ I ∣ − 1 |I| - 1 ∣ I ∣ − 1 . Hence
∣ I ∣ − 1 ≤ d . |I| - 1 \le d. ∣ I ∣ − 1 ≤ d . Combining this with ∣ I ∣ ≥ d + 1 |I| \ge d + 1 ∣ I ∣ ≥ d + 1 , we obtain ∣ I ∣ = d + 1 |I| = d + 1 ∣ I ∣ = d + 1 , as required.
We next consider the existence of dual frames that are positive bases. Although every redundant frame admits a positive dual, such a dual need not be minimal with respect to positive spanning. It is therefore natural to ask under what conditions a frame admits a dual that is also a positive basis. The following theorem provides a characterization in terms of complementary subspaces of the range of the analysis operator and suitable coordinate sign conditions.
Theorem 3.15. Let F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n , and let R = ran T F ⊆ R N R = \operatorname{ran} T_F \subseteq\mathbb{R}^{N} R = ran T F ⊆ R N , k = N − n k = N - n k = N − n . Then F F F has a dual frame that is a positive basis for H n \mathcal{H}_n H n if and only if there exists a k k k -dimensional subspace W ⊆ R N W \subseteq\mathbb{R}^{N} W ⊆ R N such that
(i) R N = R ⊕ W \mathbb{R}^{N}=R\oplus W R N = R ⊕ W ;
(ii) W ∩ R + + N ≠ ∅ W\cap\mathbb{R}_{++}^{N}\ne\varnothing W ∩ R ++ N = ∅ ;
(iii) for every i ∈ { 1 , … , N } i\in\{1,\ldots,N\} i ∈ { 1 , … , N } , there exists y ( i ) ∈ W ⊥ y^{(i)}\in W^{\perp} y ( i ) ∈ W ⊥ such that
y i ( i ) > 0 and y j ( i ) ≤ 0 for all j ≠ i . y_i^{(i)}>0\quad\text{and}\quad y_j^{(i)}\le0\quad\text{for all }j\ne i. y i ( i ) > 0 and y j ( i ) ≤ 0 for all j = i .
Proof. Suppose first that F F F has a dual frame G = { g i } i = 1 N G=\{g_i\}_{i=1}^{N} G = { g i } i = 1 N that is a positive basis for H n \mathcal{H}_n H n . Set W = ker T G ∗ W=\ker T_G^* W = ker T G ∗ . Since G G G is a dual of F F F ,
T G ∗ T F = I H n . T_G^*T_F=I_{\mathcal{H}_n}. T G ∗ T F = I H n .
Thus T G ∗ T_G^* T G ∗ is surjective and dim W = N − n = k \dim W=N-n=k dim W = N − n = k . Moreover, if x ∈ R ∩ W x\in R\cap W x ∈ R ∩ W , then x = T F h x=T_Fh x = T F h for some h ∈ H h\in\mathcal{H} h ∈ H and T G ∗ x = 0 T_G^*x=0 T G ∗ x = 0 . Hence
h = T G ∗ T F h = T G ∗ x = 0 , h=T_G^*T_Fh=T_G^*x=0, h = T G ∗ T F h = T G ∗ x = 0 ,
so x = 0 x=0 x = 0 . Therefore R ∩ W = { 0 } R\cap W=\{0\} R ∩ W = { 0 } . Since dim R = n \dim R=n dim R = n and dim W = N − n \dim W=N-n dim W = N − n , we conclude that
R N = R ⊕ W , \mathbb{R}^{N}=R\oplus W, R N = R ⊕ W ,
establishing (i).
Because G G G is a positive basis, it positively spans H n \mathcal{H}_n H n . Consequently, there exist strictly positive scalars λ 1 , … , λ N \lambda_1,\ldots,\lambda_N λ 1 , … , λ N such that
∑ i = 1 N λ i g i = 0. \sum_{i=1}^{N}\lambda_i g_i=0. i = 1 ∑ N λ i g i = 0.
Let λ = ( λ 1 , … , λ N ) T \lambda=(\lambda_1,\ldots,\lambda_N)^T λ = ( λ 1 , … , λ N ) T . Then
T G ∗ λ = ∑ i = 1 N λ i g i = 0 , T_G^*\lambda=\sum_{i=1}^{N}\lambda_i g_i=0, T G ∗ λ = i = 1 ∑ N λ i g i = 0 ,
so λ ∈ W ∩ R + + N \lambda\in W\cap\mathbb{R}_{++}^{N} λ ∈ W ∩ R ++ N . This proves (ii).
Finally, since G G G is a positive basis, for each i i i there exists b i ∈ H n b_i\in\mathcal{H}_n b i ∈ H n such that
⟨ b i , g i ⟩ > 0 , ⟨ b i , g j ⟩ ≤ 0 ( j ≠ i ) . \langle b_i,g_i\rangle>0,\qquad\langle b_i,g_j\rangle\le0\quad(j\ne i). ⟨ b i , g i ⟩ > 0 , ⟨ b i , g j ⟩ ≤ 0 ( j = i ) .
Set y ( i ) = T G b i y^{(i)}=T_Gb_i y ( i ) = T G b i . Since T G = ( T G ∗ ) ∗ T_G=(T_G^*)^* T G = ( T G ∗ ) ∗ , we have
y ( i ) ∈ ran T G = ( ker T G ∗ ) ⊥ = W ⊥ . y^{(i)}\in\operatorname{ran}T_G=(\ker T_G^*)^\perp=W^\perp. y ( i ) ∈ ran T G = ( ker T G ∗ ) ⊥ = W ⊥ .
Its coordinates satisfy
y i ( i ) = ⟨ b i , g i ⟩ > 0 , y j ( i ) = ⟨ b i , g j ⟩ ≤ 0 ( j ≠ i ) . y_i^{(i)}=\langle b_i,g_i\rangle>0,\qquad y_j^{(i)}=\langle b_i,g_j\rangle\le0\quad(j\ne i). y i ( i ) = ⟨ b i , g i ⟩ > 0 , y j ( i ) = ⟨ b i , g j ⟩ ≤ 0 ( j = i ) .
Thus (iii) holds.
Conversely, suppose that a k k k -dimensional subspace W ⊆ R N W\subseteq\mathbb{R}^{N} W ⊆ R N satisfies (i)–(iii). Since R N = R ⊕ W \mathbb{R}^{N}=R\oplus W R N = R ⊕ W , every z ∈ R N z\in\mathbb{R}^{N} z ∈ R N has a unique representation z = T F h + w z=T_Fh+w z = T F h + w , with h ∈ H h\in\mathcal{H} h ∈ H and w ∈ W w\in W w ∈ W . Define L : R N → H n L:\mathbb{R}^{N}\to\mathcal{H}_n L : R N → H n by
L ( T F h + w ) = h . L(T_Fh+w)=h. L ( T F h + w ) = h .
Then L T F = I H n LT_F=I_{\mathcal{H}_n} L T F = I H n and ker L = W \ker L=W ker L = W . Let g i = L e i g_i = Le_i g i = L e i , where { e i } i = 1 N \{e_i\}_{i=1}^{N} { e i } i = 1 N is the standard basis of R N \mathbb{R}^{N} R N , and put G = { g i } i = 1 N G = \{g_i\}_{i=1}^{N} G = { g i } i = 1 N . The synthesis operator of G G G is L L L , so
T G ∗ T F = L T F = I H n . T_G^*T_F = LT_F = I_{\mathcal{H}_n}. T G ∗ T F = L T F = I H n .
Hence G G G is a dual frame of F F F .
By (ii), choose λ ∈ W ∩ R > 0 N \lambda\in W \cap\mathbb{R}_{>0}^{N} λ ∈ W ∩ R > 0 N . Since ker L = W \ker L = W ker L = W , we have
0 = L λ = ∑ i = 1 N λ i g i , λ i > 0. 0 = L\lambda= \sum_{i=1}^{N} \lambda_i g_i, \qquad\lambda_i > 0. 0 = L λ = i = 1 ∑ N λ i g i , λ i > 0.
As G G G spans H n \mathcal{H}_n H n , this strictly positive linear dependence implies that G G G positively spans H n \mathcal{H}_n H n . It remains to show that G G G is exact. For each i i i , condition (iii) provides y ( i ) ∈ W ⊥ y^{(i)} \in W^\perp y ( i ) ∈ W ⊥ . Since
W ⊥ = ( ker L ) ⊥ = ran L ∗ , W^\perp= (\ker L)^\perp= \operatorname{ran} L^*, W ⊥ = ( ker L ) ⊥ = ran L ∗ ,
there exists b i ∈ H b_i \in\mathcal{H} b i ∈ H such that y ( i ) = L ∗ b i y^{(i)} = L^*b_i y ( i ) = L ∗ b i . Therefore,
y j ( i ) = ⟨ L ∗ b i , e j ⟩ = ⟨ b i , L e j ⟩ = ⟨ b i , g j ⟩ . y_j^{(i)} = \langle L^*b_i,e_j\rangle= \langle b_i,Le_j\rangle= \langle b_i,g_j\rangle. y j ( i ) = ⟨ L ∗ b i , e j ⟩ = ⟨ b i , L e j ⟩ = ⟨ b i , g j ⟩ .
Condition (iii) consequently gives
⟨ b i , g i ⟩ > 0 , ⟨ b i , g j ⟩ ≤ 0 ( j ≠ i ) . \langle b_i,g_i\rangle> 0, \qquad\langle b_i,g_j\rangle\leq0 \quad(j \ne i). ⟨ b i , g i ⟩ > 0 , ⟨ b i , g j ⟩ ≤ 0 ( j = i ) .
If g i g_i g i belonged to pos { g j : j ≠ i } \operatorname{pos}\{g_j:j \ne i\} pos { g j : j = i } , we could write g i = ∑ j ≠ i a j g j g_i = \sum_{j\ne i} a_jg_j g i = ∑ j = i a j g j with a j ≥ 0 a_j \geq0 a j ≥ 0 . Taking inner products with b i b_i b i would yield
⟨ b i , g i ⟩ = ∑ j ≠ i a j ⟨ b i , g j ⟩ ≤ 0 , \langle b_i,g_i\rangle= \sum_{j\ne i} a_j\langle b_i,g_j\rangle\leq0, ⟨ b i , g i ⟩ = j = i ∑ a j ⟨ b i , g j ⟩ ≤ 0 ,
contradicting ⟨ b i , g i ⟩ > 0 \langle b_i,g_i\rangle> 0 ⟨ b i , g i ⟩ > 0 . Thus no g i g_i g i can be removed while preserving positive spanning. Hence G G G is a positive basis, and the theorem follows.
Existence and Geometry of Positive Dual Frames
In this section, we investigate the existence and geometric structure of positive dual frames in finite-dimensional real Hilbert spaces. We begin by establishing a necessary and sufficient condition for a frame to admit a positive dual and provide an explicit construction. We then examine the geometric properties of the collection of positive duals, including its affine and topological structure, and investigate dual frames satisfying prescribed positive relations. These results highlight the role of redundancy in the existence and structure of positive dual frames.
We now investigate whether a dual frame can be chosen to be positive and show that the existence of such a dual is characterized precisely by the redundancy of the original frame. Let F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a finite family of vectors in an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n . Then F F F is called a positive basis for H n \mathcal{H}_n H n if
span + ( F ) = H n \operatorname{span}_{+}(F) = \mathcal{H}_n span + ( F ) = H n
and no proper subfamily of F F F positively spans H n \mathcal{H}_n H n . Unlike an ordinary basis, a positive basis may contain more than n n n vectors. In fact, every positive basis for H n \mathcal{H}_n H n contains between n + 1 n+1 n + 1 and 2 n 2n 2 n vectors [6 ] . Positive bases with n + 1 n+1 n + 1 and 2 n 2n 2 n elements are called minimal and maximal positive bases, respectively.
We first establish that the collection of positive dual frames is relatively open in the set of all dual frames. For a finite family F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N in H n \mathcal{H}_n H n , define
ρ ( G ) = min ∥ x ∥ = 1 max 1 ≤ i ≤ N ⟨ x , f i ⟩ . \rho(G)=\min_{\lVert x\rVert=1}\max_{1\leq i\leq N}\langle x,f_i\rangle. ρ ( G ) = ∥ x ∥ = 1 min 1 ≤ i ≤ N max ⟨ x , f i ⟩ .
Lemma 4.1. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a positive frame for H n \mathcal{H}_n H n if and only if ρ ( F ) > 0 \rho(F)>0 ρ ( F ) > 0 .
Proof. Suppose that G G G positively spans H n \mathcal{H}_n H n . By the positive-spanning criterion, for every nonzero x ∈ H n x\in\mathcal{H}_n x ∈ H n , there exists i ∈ { 1 , … , N } i\in\{1,\ldots,N\} i ∈ { 1 , … , N } such that ⟨ x , f i ⟩ > 0 \langle x,f_i\rangle>0 ⟨ x , f i ⟩ > 0 . Consequently, max 1 ≤ i ≤ N ⟨ x , f i ⟩ > 0 \max_{1\leq i\leq N}\langle x,f_i\rangle>0 max 1 ≤ i ≤ N ⟨ x , f i ⟩ > 0 , for every unit vector x ∈ H n x\in\mathcal{H}_n x ∈ H n . Since the function
x ⟼ max 1 ≤ i ≤ N ⟨ x , f i ⟩ x\longmapsto\max_{1\leq i\leq N}\langle x,f_i\rangle x ⟼ 1 ≤ i ≤ N max ⟨ x , f i ⟩ is continuous on the compact unit sphere, it attains its minimum at some unit vector x 0 x_0 x 0 . Therefore, ρ ( F ) = max 1 ≤ i ≤ N ⟨ x 0 , f i ⟩ > 0 \rho(F)=\max_{1\leq i\leq N}\langle x_0,f_i\rangle>0 ρ ( F ) = max 1 ≤ i ≤ N ⟨ x 0 , f i ⟩ > 0 . Conversely, suppose that ρ ( F ) > 0 \rho(F)>0 ρ ( F ) > 0 . Then, for every unit vector x ∈ H n x\in\mathcal{H}_n x ∈ H n , max 1 ≤ i ≤ N ⟨ x , f i ⟩ ≥ ρ ( F ) > 0 \max_{1\leq i\leq N}\langle x,f_i\rangle\geq\rho(F)>0 max 1 ≤ i ≤ N ⟨ x , f i ⟩ ≥ ρ ( F ) > 0 . Thus, for every nonzero x ∈ H n x\in\mathcal{H}_n x ∈ H n , there exists i ∈ { 1 , … , N } i\in\{1,\ldots,N\} i ∈ { 1 , … , N } such that ⟨ x , f i ⟩ > 0 \langle x,f_i\rangle>0 ⟨ x , f i ⟩ > 0 . By the positive-spanning criterion, F F F positively spans H n \mathcal{H}_n H n .
Existence and Explicit Construction of Positive Duals
We now address the existence of positive dual frames for a given frame. While every frame admits a dual, the existence of a positive dual imposes an additional geometric requirement. In particular, the redundancy of the original frame plays a decisive role in this question. The following theorem provides a complete characterization, showing that the existence of a positive dual depends only on the excess of the frame.
Theorem 4.2. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n . Then F F F admits a positive dual frame if and only if N > n N>n N > n .
Proof. Suppose first that F F F admits a positive dual G = { g i } i = 1 N G=\{g_i\}_{i=1}^{N} G = { g i } i = 1 N . Since G G G is a positive frame, Theorem 3.1 implies that there exist c 1 , … , c N > 0 c_1,\ldots,c_N>0 c 1 , … , c N > 0 , such that
∑ i = 1 N c i g i = 0. \sum_{i=1}^{N}c_i g_i=0. i = 1 ∑ N c i g i = 0. Therefore G G G is linearly dependent. Since G G G is a frame for the n n n -dimensional space H n \mathcal{H}_n H n , it follows that N > n N>n N > n .
Conversely, suppose that N > n N>n N > n . Since F F F is a frame, dim Ran ( T F ) = n < N \dim\operatorname{Ran}(T_F)=n<N dim Ran ( T F ) = n < N . Thus Ran ( T F ) \operatorname{Ran}(T_F) Ran ( T F ) is a proper linear subspace of R N \mathbb{R}^{N} R N .
We claim that there exists c ∈ R + + N c\in\mathbb{R}_{++}^{N} c ∈ R ++ N such that c ∉ Ran ( T F ) c\notin\operatorname{Ran}(T_F) c ∈ / Ran ( T F ) . Indeed, R + + N \mathbb{R}_{++}^{N} R ++ N is a nonempty open subset of R N \mathbb{R}^{N} R N , whereas every proper linear subspace of R N \mathbb{R}^{N} R N has empty interior. Therefore R + + N \mathbb{R}_{++}^{N} R ++ N cannot be contained in Ran ( T F ) \operatorname{Ran}(T_F) Ran ( T F ) . Hence we may choose c = ( c 1 , … , c N ) ∈ R + + N ∖ Ran ( T F ) c=(c_1,\ldots,c_N)\in\mathbb{R}_{++}^{N}\setminus\operatorname{Ran}(T_F) c = ( c 1 , … , c N ) ∈ R ++ N ∖ Ran ( T F ) .
Let P F = T F S F − 1 T F ∗ P_F=T_FS_F^{-1}T_F^* P F = T F S F − 1 T F ∗ . Since F F F is a frame, P F P_F P F is the orthogonal projection of R N \mathbb{R}^{N} R N onto Ran ( T F ) \operatorname{Ran}(T_F) Ran ( T F ) . Define z = ( I − P F ) c z=(I-P_F)c z = ( I − P F ) c . Since c ∉ Ran ( T F ) c\notin\operatorname{Ran}(T_F) c ∈ / Ran ( T F ) , we have z ≠ 0 z\ne0 z = 0 . Moreover, z ∈ Ran ( T F ) ⊥ z\in\operatorname{Ran}(T_F)^\perp z ∈ Ran ( T F ) ⊥ . Since Ran ( T F ) ⊥ = ker T F ∗ \operatorname{Ran}(T_F)^\perp=\ker T_F^* Ran ( T F ) ⊥ = ker T F ∗ , we obtain z ∈ ker T F ∗ z\in\ker T_F^* z ∈ ker T F ∗ . In particular, T F ∗ z = 0 T_F^*z=0 T F ∗ z = 0 . Next, define u = S F − 1 T F ∗ c u=S_F^{-1}T_F^*c u = S F − 1 T F ∗ c . Equivalently,
u = S F − 1 ( ∑ i = 1 N c i f i ) = ∑ i = 1 N c i S F − 1 f i . u=S_F^{-1}\left(\sum_{i=1}^{N}c_i f_i\right)=\sum_{i=1}^{N}c_iS_F^{-1}f_i. u = S F − 1 ( i = 1 ∑ N c i f i ) = i = 1 ∑ N c i S F − 1 f i .
For each i = 1 , … , N i=1,\ldots,N i = 1 , … , N , define
h i = − z i ∥ z i ∥ 2 u h_i=-\frac{z_i}{\lVert z_i\rVert^2}u h i = − ∥ z i ∥ 2 z i u
and set g i = S F − 1 f i + h i g_i=S_F^{-1}f_i+h_i g i = S F − 1 f i + h i . Thus
g i = S F − 1 f i − z i ∥ z i ∥ 2 u . g_i=S_F^{-1}f_i-\frac{z_i}{\lVert z_i\rVert^2}u. g i = S F − 1 f i − ∥ z i ∥ 2 z i u .
We first prove that G = { g i } i = 1 N G=\{g_i\}_{i=1}^N G = { g i } i = 1 N is a dual frame of F F F . Let x ∈ H n x\in\mathcal{H}_n x ∈ H n . Then
∑ i = 1 N ⟨ x , h i ⟩ f i = ∑ i = 1 N ⟨ x , − z i ∥ z i ∥ 2 u ⟩ f i = − ⟨ x , u ⟩ ∥ z ∥ 2 ∑ i = 1 N z i f i = − ⟨ x , u ⟩ ∥ z ∥ 2 T F ∗ z = 0. \begin{aligned}
\sum_{i=1}^N\langle x,h_i\rangle f_i
&=\sum_{i=1}^N\left\langle x,-\frac{z_i}{\lVert z_i\rVert^2}u\right\rangle f_i\\
&=-\frac{\langle x,u\rangle}{\lVert z\rVert^2}\sum_{i=1}^N z_i f_i\\
&=-\frac{\langle x,u\rangle}{\lVert z\rVert^2}T_F^*z\\
&=0.
\end{aligned} i = 1 ∑ N ⟨ x , h i ⟩ f i = i = 1 ∑ N ⟨ x , − ∥ z i ∥ 2 z i u ⟩ f i = − ∥ z ∥ 2 ⟨ x , u ⟩ i = 1 ∑ N z i f i = − ∥ z ∥ 2 ⟨ x , u ⟩ T F ∗ z = 0.
Therefore,
∑ i = 1 N ⟨ x , g i ⟩ f i = ∑ i = 1 N ⟨ x , S F − 1 f i ⟩ f i + ∑ i = 1 N ⟨ x , h i ⟩ f i = x . \sum_{i=1}^N\langle x,g_i\rangle f_i=\sum_{i=1}^N\langle x,S_F^{-1}f_i\rangle f_i+\sum_{i=1}^N\langle x,h_i\rangle f_i=x. i = 1 ∑ N ⟨ x , g i ⟩ f i = i = 1 ∑ N ⟨ x , S F − 1 f i ⟩ f i + i = 1 ∑ N ⟨ x , h i ⟩ f i = x .
Hence G G G is a dual frame of F F F .
It remains to prove that G G G is positive. Since u = ∑ i = 1 N c i S F − 1 f i u=\sum_{i=1}^N c_iS_F^{-1}f_i u = ∑ i = 1 N c i S F − 1 f i , we have
∑ i = 1 N c i g i = ∑ i = 1 N c i S F − 1 f i − ∑ i = 1 N c i z i ∥ z ∥ 2 u = u − 1 ∥ z ∥ 2 ( ∑ i = 1 N c i z i ) u = u − ⟨ c , z ⟩ ∥ z ∥ 2 u . \begin{aligned}
\sum_{i=1}^N c_i g_i
&=\sum_{i=1}^N c_iS_F^{-1}f_i-\sum_{i=1}^N c_i\frac{z_i}{\lVert z\rVert^2}u\\
&=u-\frac{1}{\lVert z\rVert^2}\left(\sum_{i=1}^N c_i z_i\right)u\\
&=u-\frac{\langle c,z\rangle}{\lVert z\rVert^2}u.
\end{aligned} i = 1 ∑ N c i g i = i = 1 ∑ N c i S F − 1 f i − i = 1 ∑ N c i ∥ z ∥ 2 z i u = u − ∥ z ∥ 2 1 ( i = 1 ∑ N c i z i ) u = u − ∥ z ∥ 2 ⟨ c , z ⟩ u .
Now c = P F c + ( I − P F ) c = P F c + z c=P_Fc+(I-P_F)c=P_Fc+z c = P F c + ( I − P F ) c = P F c + z . Since P F c ∈ Ran ( T F ) P_Fc\in\operatorname{Ran}(T_F) P F c ∈ Ran ( T F ) and z ∈ Ran ( T F ) ⊥ z\in\operatorname{Ran}(T_F)^\perp z ∈ Ran ( T F ) ⊥ , we have ⟨ P F c , z ⟩ = 0 \langle P_Fc,z\rangle=0 ⟨ P F c , z ⟩ = 0 . Consequently,
⟨ c , z ⟩ = ⟨ P F c + z , z ⟩ = ⟨ P F c , z ⟩ + ⟨ z , z ⟩ = ∥ z ∥ 2 . \begin{aligned}
\langle c,z\rangle
&=\langle P_Fc+z,z\rangle\\
&=\langle P_Fc,z\rangle+\langle z,z\rangle\\
&=\lVert z\rVert^2.
\end{aligned} ⟨ c , z ⟩ = ⟨ P F c + z , z ⟩ = ⟨ P F c , z ⟩ + ⟨ z , z ⟩ = ∥ z ∥ 2 .
It follows that
∑ i = 1 N c i g i = u − ∥ z ∥ 2 ∥ z ∥ 2 u = 0. \sum_{i=1}^N c_i g_i=u-\frac{\lVert z\rVert^2}{\lVert z\rVert^2}u=0. i = 1 ∑ N c i g i = u − ∥ z ∥ 2 ∥ z ∥ 2 u = 0.
Since c i > 0 c_i>0 c i > 0 , i = 1 , … , N i=1,\ldots,N i = 1 , … , N , we have obtained a strictly positive linear dependence
∑ i = 1 N c i g i = 0. \sum_{i=1}^N c_i g_i=0. i = 1 ∑ N c i g i = 0.
By Theorem 3.1, G G G is a positive frame. Therefore G G G is a positive dual frame of F F F .
The proof of Theorem 4.2 gives an explicit construction of a positive dual. We record this observation for later use.
Remark 4.3. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a redundant frame for H n \mathcal{H}_n H n . For any c ∈ R + + N ∖ Ran ( T F ) c\in\mathbb{R}_{++}^{N}\setminus\operatorname{Ran}(T_F) c ∈ R ++ N ∖ Ran ( T F ) , put P F = T F S − 1 T F ∗ P_F=T_FS^{-1}T_F^* P F = T F S − 1 T F ∗ , z c = ( I − P F ) c z_c=(I-P_F)c z c = ( I − P F ) c , and u c = S − 1 T F ∗ c u_c=S^{-1}T_F^*c u c = S − 1 T F ∗ c . Then G c = { g i ( c ) } i = 1 N G_c=\{g_i^{(c)}\}_{i=1}^{N} G c = { g i ( c ) } i = 1 N , where
g i ( c ) = S − 1 f i − ( z c ) i ∥ z c ∥ 2 u c , i = 1 , … , N , g_i^{(c)}=S^{-1}f_i-\frac{(z_c)_i}{\lVert z_c\rVert^2}u_c,\qquad i=1,\ldots,N, g i ( c ) = S − 1 f i − ∥ z c ∥ 2 ( z c ) i u c , i = 1 , … , N ,
is a positive dual frame of F F F . Moreover,
∑ i = 1 N c i g i ( c ) = 0. \sum_{i=1}^{N}c_i g_i^{(c)}=0. i = 1 ∑ N c i g i ( c ) = 0.
Thus every strictly positive vector c ∈ R + + N ∖ Ran ( T F ) c\in\mathbb{R}_{++}^{N}\setminus\operatorname{Ran}(T_F) c ∈ R ++ N ∖ Ran ( T F ) gives an explicit positive dual of F F F .
Example 4.4. Consider the frame
F = { e 1 , e 2 , e 1 + e 2 } F=\{e_1,e_2,e_1+e_2\} F = { e 1 , e 2 , e 1 + e 2 }
for R 2 \mathbb{R}^{2} R 2 . Clearly, F F F is not a positive frame, since every vector in its positive cone has nonnegative coordinates. Every dual frame of F F F is of the form
G a , b = { ( 1 − a ) e 1 − b e 2 , − a e 1 + ( 1 − b ) e 2 , a e 1 + b e 2 } , G_{a,b}=\{(1-a)e_1-be_2,-ae_1+(1-b)e_2,ae_1+be_2\}, G a , b = {( 1 − a ) e 1 − b e 2 , − a e 1 + ( 1 − b ) e 2 , a e 1 + b e 2 } ,
where a , b ∈ R a,b\in\mathbb{R} a , b ∈ R . Indeed, the duality conditions are g 1 + g 3 = e 1 g_1+g_3=e_1 g 1 + g 3 = e 1 and g 2 + g 3 = e 2 g_2+g_3=e_2 g 2 + g 3 = e 2 .
A direct calculation shows that
a g 1 + b g 2 + ( a + b − 1 ) g 3 = 0. ag_1+bg_2+(a+b-1)g_3=0. a g 1 + b g 2 + ( a + b − 1 ) g 3 = 0.
Since G a , b G_{a,b} G a , b spans R 2 \mathbb{R}^{2} R 2 , it is a positive frame if and only if its one-dimensional synthesis kernel contains a strictly positive vector. Consequently,
G a , b ∈ D + ( F ) ⟺ a > 0 , b > 0 , a + b > 1. G_{a,b}\in D_+(F)\quad\Longleftrightarrow\quad a>0,\quad b>0,\quad a+b>1. G a , b ∈ D + ( F ) ⟺ a > 0 , b > 0 , a + b > 1.
Thus, the positive duals of F F F are parametrized by an unbounded open region of R 2 \mathbb{R}^{2} R 2 , even though F F F itself is not positive.
Dual Frames with Prescribed Positive Relations
We next examine the geometric structure of the collection of positive dual frames. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n . For a frame F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N in H n \mathcal{H}_n H n , let
D + ( F ) = { G = { g i } i = 1 N : G is a positive dual frame of F } . D_+(F)=\{G=\{g_i\}_{i=1}^{N}:G\text{ is a positive dual frame of }F\}. D + ( F ) = { G = { g i } i = 1 N : G is a positive dual frame of F } .
For each a ∈ R + + N a\in\mathbb{R}_{++}^{N} a ∈ R ++ N , define
D a ( F ) = { G ∈ D ( F ) : T G ∗ a = 0 } . D_a(F)=\{G\in D(F):T_G^*a=0\}. D a ( F ) = { G ∈ D ( F ) : T G ∗ a = 0 } .
Since a dual frame is positive precisely when its synthesis kernel contains a strictly positive vector, we have
D + ( F ) = ⋃ a ∈ R + + N D a ( F ) . D_+(F)=\bigcup_{a\in\mathbb{R}_{++}^{N}}D_a(F). D + ( F ) = a ∈ R ++ N ⋃ D a ( F ) .
Moreover, each nonempty D a ( F ) D_a(F) D a ( F ) is an affine subspace of D ( F ) D(F) D ( F ) and is therefore convex and path-connected. The following proposition characterizes the nonempty members of this family and determines their dimensions.
Proposition 4.5. Let F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for R n \mathbb{R}^n R n and let c ∈ R + + N c \in\mathbb{R}_{++}^{N} c ∈ R ++ N . Then D c ( F ) ≠ ∅ \mathcal{D}_c(F) \ne\varnothing D c ( F ) = ∅ if and only if c ∉ Ran ( T F ) c \notin\operatorname{Ran}(T_F) c ∈ / Ran ( T F ) . Moreover, whenever D c ( F ) \mathcal{D}_c(F) D c ( F ) is nonempty, it is an affine subspace of D ( F ) \mathcal{D}(F) D ( F ) of dimension
dim D c ( F ) = n ( N − n − 1 ) . \dim\mathcal{D}_c(F) = n(N-n-1). dim D c ( F ) = n ( N − n − 1 ) . Proof. Suppose first that D c ( F ) \mathcal{D}_c(F) D c ( F ) is nonempty, and let G ∈ D c ( F ) G \in\mathcal{D}_c(F) G ∈ D c ( F ) . If c ∈ Ran ( T F ) c \in\operatorname{Ran}(T_F) c ∈ Ran ( T F ) , then there exists x ∈ R n x \in\mathbb{R}^n x ∈ R n such that c = T F x c = T_Fx c = T F x . Since G G G is a dual frame of F F F , we obtain
0 = T G ∗ c = T G ∗ T F x = x , 0 = T_G^*c = T_G^*T_Fx = x, 0 = T G ∗ c = T G ∗ T F x = x , which contradicts c > 0 c > 0 c > 0 . Therefore, c ∉ Ran ( T F ) c \notin\operatorname{Ran}(T_F) c ∈ / Ran ( T F ) .
Conversely, suppose that c ∉ Ran ( T F ) c \notin\operatorname{Ran}(T_F) c ∈ / Ran ( T F ) . Let S F = T F ∗ T F S_F = T_F^*T_F S F = T F ∗ T F be the frame operator of F F F , and consider the orthogonal projection
P F = T F S F − 1 T F ∗ P_F = T_FS_F^{-1}T_F^* P F = T F S F − 1 T F ∗ onto Ran ( T F ) \operatorname{Ran}(T_F) Ran ( T F ) . Set z = ( I − P F ) c z = (I-P_F)c z = ( I − P F ) c . Since c ∉ Ran ( T F ) c \notin\operatorname{Ran}(T_F) c ∈ / Ran ( T F ) , we have z ≠ 0 z \ne0 z = 0 . Moreover, z ∈ ker T F ∗ z \in\operatorname{ker}T_F^* z ∈ ker T F ∗ , and hence
⟨ c , z ⟩ = ⟨ P F c + z , z ⟩ = ∥ z ∥ 2 > 0. \langle c,z\rangle= \langle P_Fc+z,z\rangle= \lVert z\rVert^2 > 0. ⟨ c , z ⟩ = ⟨ P F c + z , z ⟩ = ∥ z ∥ 2 > 0. Define u = S F − 1 T F ∗ c u = S_F^{-1}T_F^*c u = S F − 1 T F ∗ c and consider the linear operator L c : R N → R n L_c:\mathbb{R}^{N}\to\mathbb{R}^{n} L c : R N → R n given by
L c v = S F − 1 T F ∗ v − ⟨ v , z ⟩ ∥ z ∥ 2 u , v ∈ R N . L_cv = S_F^{-1}T_F^*v-\frac{\langle v,z\rangle}{\lVert z\rVert^2}u,\qquad v\in\mathbb{R}^{N}. L c v = S F − 1 T F ∗ v − ∥ z ∥ 2 ⟨ v , z ⟩ u , v ∈ R N . Since T F ∗ z = 0 T_F^*z=0 T F ∗ z = 0 , for every x ∈ R n x\in\mathbb{R}^{n} x ∈ R n we have
L c T F x = S F − 1 T F ∗ T F x − ⟨ T F x , z ⟩ ∥ z ∥ 2 u = x . L_cT_Fx = S_F^{-1}T_F^*T_Fx-\frac{\langle T_Fx,z\rangle}{\lVert z\rVert^2}u=x. L c T F x = S F − 1 T F ∗ T F x − ∥ z ∥ 2 ⟨ T F x , z ⟩ u = x . Thus L c T F = I R n L_cT_F=I_{\mathbb{R}^n} L c T F = I R n . Furthermore,
L c c = S F − 1 T F ∗ c − ⟨ c , z ⟩ ∥ z ∥ 2 u = u − u = 0. L_cc=S_F^{-1}T_F^*c-\frac{\langle c,z\rangle}{\lVert z\rVert^2}u=u-u=0. L c c = S F − 1 T F ∗ c − ∥ z ∥ 2 ⟨ c , z ⟩ u = u − u = 0. Let { e i } i = 1 N \{e_i\}_{i=1}^{N} { e i } i = 1 N be the standard orthonormal basis of R N \mathbb{R}^{N} R N , and define g i = L c e i g_i=L_ce_i g i = L c e i for i = 1 , … , N i=1,\ldots,N i = 1 , … , N . Then the family G = { g i } i = 1 N G=\{g_i\}_{i=1}^{N} G = { g i } i = 1 N satisfies T G ∗ = L c T_G^*=L_c T G ∗ = L c . Consequently,
T G ∗ T F = I R n and T G ∗ c = 0. T_G^*T_F=I_{\mathbb{R}^n}\qquad\text{and}\qquad T_G^*c=0. T G ∗ T F = I R n and T G ∗ c = 0. Therefore, G ∈ D c ( F ) G\in\mathcal{D}_c(F) G ∈ D c ( F ) , proving that D c ( F ) \mathcal{D}_c(F) D c ( F ) is nonempty.
We now determine its affine structure. Fix L c L_c L c as constructed above and let
M = Ran ( T F ) + span { c } . M=\operatorname{Ran}(T_F)+\operatorname{span}\{c\}. M = Ran ( T F ) + span { c } . Since c ∉ Ran ( T F ) c\notin\operatorname{Ran}(T_F) c ∈ / Ran ( T F ) and rank ( T F ) = n \operatorname{rank}(T_F)=n rank ( T F ) = n , we have dim M = n + 1 \dim M=n+1 dim M = n + 1 .
If L : R N → R n L:\mathbb{R}^{N}\to\mathbb{R}^{n} L : R N → R n is the synthesis operator of an element of D c ( F ) \mathcal{D}_c(F) D c ( F ) , then
L T F = I R n , L c = 0. LT_F=I_{\mathbb{R}^n},\qquad Lc=0. L T F = I R n , L c = 0. Hence A = L − L c A=L-L_c A = L − L c satisfies A T F = 0 AT_F=0 A T F = 0 and A c = 0 Ac=0 A c = 0 , so that A ∣ M = 0 A|_M=0 A ∣ M = 0 .
Conversely, if A ∈ L ( R N , R n ) A\in\mathcal{L}(\mathbb{R}^{N},\mathbb{R}^{n}) A ∈ L ( R N , R n ) satisfies A ∣ M = 0 A|_M=0 A ∣ M = 0 , then L = L c + A L=L_c+A L = L c + A satisfies
L T F = I R n , L c = 0. LT_F=I_{\mathbb{R}^n},\qquad Lc=0. L T F = I R n , L c = 0.
Therefore,
D c ( F ) ≅ L c + { A ∈ L ( R N , R n ) : A ∣ M = 0 } . \mathcal{D}_c(F) \cong L_c+\left\{A\in\mathcal{L}(\mathbb{R}^N,\mathbb{R}^n):A|_M=0\right\}. D c ( F ) ≅ L c + { A ∈ L ( R N , R n ) : A ∣ M = 0 } .
Since every such operator is uniquely determined by its restriction to M ⊥ M^\perp M ⊥ , the translation space is isomorphic to L ( M ⊥ , R n ) \mathcal{L}(M^\perp,\mathbb{R}^n) L ( M ⊥ , R n ) . As dim M ⊥ = N − n − 1 \dim M^\perp=N-n-1 dim M ⊥ = N − n − 1 , it follows that
dim D c ( F ) = n ( N − n − 1 ) . \dim\mathcal{D}_c(F)=n(N-n-1). dim D c ( F ) = n ( N − n − 1 ) .
This completes the proof.
An immediate consequence of the preceding proposition is that, for frames with excess one, a prescribed strictly positive linear relation determines a unique dual frame.
Corollary 4.6. Let F = { f i } i = 1 n + 1 F=\{f_i\}_{i=1}^{n+1} F = { f i } i = 1 n + 1 be a frame for an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n , and let a ∈ R + + n + 1 a\in\mathbb{R}_{++}^{n+1} a ∈ R ++ n + 1 . Then there exists a unique dual frame G G G of F F F satisfying
if and only if a ∉ Ran T F a\notin\operatorname{Ran}T_F a ∈ / Ran T F . In particular, such a dual frame is positive.
Proof. By Proposition 4.5, C a ( F ) C_a(F) C a ( F ) is nonempty if and only if a ∉ Ran T F a\notin\operatorname{Ran}T_F a ∈ / Ran T F . In this case,
dim C a ( F ) = n ( ( n + 1 ) − n − 1 ) = 0. \dim C_a(F)=n((n+1)-n-1)=0. dim C a ( F ) = n (( n + 1 ) − n − 1 ) = 0.
Hence C a ( F ) C_a(F) C a ( F ) consists of exactly one dual frame G G G . Since a a a is strictly positive and T G ∗ a = 0 T_G^*a=0 T G ∗ a = 0 , the frame G G G is positive.
Topology and Geometry of Positive Duals
We next investigate the geometric structure of the collection of positive dual frames. Recall that an element G ∈ D + ( F ) G\in\mathcal{D}_+(F) G ∈ D + ( F ) is called an extreme point of D + ( F ) \mathcal{D}_+(F) D + ( F ) if
G = t G 1 + ( 1 − t ) G 2 , 0 < t < 1 , G=tG_1+(1-t)G_2,\qquad0<t<1, G = t G 1 + ( 1 − t ) G 2 , 0 < t < 1 ,
for some G 1 , G 2 ∈ D + ( F ) G_1,G_2\in\mathcal{D}_+(F) G 1 , G 2 ∈ D + ( F ) implies G 1 = G 2 = G G_1=G_2=G G 1 = G 2 = G , where addition and scalar multiplication are understood componentwise.
The following theorem shows that, whenever F F F is redundant, no positive dual frame is an extreme point of D + ( F ) \mathcal{D}_+(F) D + ( F ) . Thus every positive dual admits a nontrivial convex decomposition into two distinct positive dual frames.
Theorem 4.7. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^N F = { f i } i = 1 N be a frame for an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n , where N > n N>n N > n . Then the set D + ( F ) \mathcal{D}_+(F) D + ( F ) of all positive dual frames of F F F has no extreme points.
Proof. If D + ( F ) = ∅ \mathcal{D}_+(F)=\varnothing D + ( F ) = ∅ , the conclusion is immediate. Otherwise, let G = { g i } i = 1 N ∈ D + ( F ) G=\{g_i\}_{i=1}^N\in\mathcal{D}_+(F) G = { g i } i = 1 N ∈ D + ( F ) be arbitrary. Since G G G is a dual frame of F F F , we have
∑ i = 1 N ⟨ x , f i ⟩ g i = x , x ∈ H . \sum_{i=1}^{N}\langle x,f_i\rangle g_i=x,\qquad x\in\mathcal{H}. i = 1 ∑ N ⟨ x , f i ⟩ g i = x , x ∈ H .
Equivalently,
T G ∗ T F = I H , (3) T_G^*T_F=I_{\mathcal{H}}, \tag*{(3)} T G ∗ T F = I H , (3)
where T F T_F T F and T G ∗ T_G^* T G ∗ denote the analysis and synthesis operators, respectively. Since N > n N>n N > n , the analysis operator T F : H n → R N T_F:\mathcal{H}_n\to\mathbb{R}^N T F : H n → R N has a nontrivial orthogonal complement to its range. Hence,
ker ( T F ∗ ) ≠ { 0 } . \operatorname{ker}(T_F^*)\ne\{0\}. ker ( T F ∗ ) = { 0 } .
Choose a nonzero vector d = ( d i ) i = 1 N ∈ ker ( T F ∗ ) d=(d_i)_{i=1}^N\in\operatorname{ker}(T_F^*) d = ( d i ) i = 1 N ∈ ker ( T F ∗ ) . Thus,
∑ i = 1 N d i f i = 0. \sum_{i=1}^{N}d_i f_i=0. i = 1 ∑ N d i f i = 0.
Let u ∈ H u\in\mathcal{H} u ∈ H be any nonzero vector and define H = { h i } i = 1 N H=\{h_i\}_{i=1}^N H = { h i } i = 1 N by
h i = d i u , 1 ≤ i ≤ N . h_i=d_i u,\qquad1\le i\le N. h i = d i u , 1 ≤ i ≤ N .
Then, for every x ∈ H x\in\mathcal{H} x ∈ H ,
∑ i = 1 N ⟨ x , f i ⟩ h i = ∑ i = 1 N ⟨ x , f i ⟩ d i u = ⟨ x , ∑ i = 1 N d i f i ⟩ u = 0. \begin{aligned}
\sum_{i=1}^{N}\langle x,f_i\rangle h_i=\sum_{i=1}^{N}\langle x,f_i\rangle d_i u \\
&=\left\langle x,\sum_{i=1}^{N}d_i f_i\right\rangle u=0.
\end{aligned} i = 1 ∑ N ⟨ x , f i ⟩ h i = i = 1 ∑ N ⟨ x , f i ⟩ d i u = ⟨ x , i = 1 ∑ N d i f i ⟩ u = 0.
Consequently, T H ∗ T F = 0 T_H^*T_F=0 T H ∗ T F = 0 .
For each t ∈ R t\in\mathbb{R} t ∈ R , define
G t = { g i + t h i } i = 1 N . G_t=\{g_i+th_i\}_{i=1}^N. G t = { g i + t h i } i = 1 N .
Since
T G t ∗ T F = T G ∗ T F + t T H ∗ T F = I H n , T_{G_t}^*T_F=T_G^*T_F+tT_H^*T_F=I_{\mathcal{H}_n}, T G t ∗ T F = T G ∗ T F + t T H ∗ T F = I H n ,
it follows that G t G_t G t is a dual frame of F F F for every t ∈ R t\in\mathbb{R} t ∈ R .
We now show that G t G_t G t remains a positive frame for all sufficiently small ∣ t ∣ |t| ∣ t ∣ . Since G G G is a positive frame, for every unit vector x ∈ H n x\in\mathcal{H}_n x ∈ H n ,
max 1 ≤ i ≤ N ⟨ x , g i ⟩ > 0. \max_{1\le i\le N}\langle x,g_i\rangle>0. 1 ≤ i ≤ N max ⟨ x , g i ⟩ > 0.
Consider the function φ : S H n → R \varphi:S_{\mathcal{H}_n}\to\mathbb{R} φ : S H n → R defined by
φ ( x ) = max 1 ≤ i ≤ N ⟨ x , g i ⟩ , \varphi(x)=\max_{1\le i\le N}\langle x,g_i\rangle, φ ( x ) = 1 ≤ i ≤ N max ⟨ x , g i ⟩ ,
where S H n = { x ∈ H n : ∥ x ∥ = 1 } S_{\mathcal{H}_n}=\{x\in\mathcal{H}_n:\|x\|=1\} S H n = { x ∈ H n : ∥ x ∥ = 1 } . As φ \varphi φ is continuous and S H n S_{\mathcal{H}_n} S H n is compact, there exists δ > 0 \delta>0 δ > 0 such that
δ = min ∥ x ∥ = 1 max 1 ≤ i ≤ N ⟨ x , g i ⟩ > 0. \delta=\min_{\|x\|=1}\max_{1\le i\le N}\langle x,g_i\rangle>0. δ = ∥ x ∥ = 1 min 1 ≤ i ≤ N max ⟨ x , g i ⟩ > 0.
Set
M = max 1 ≤ i ≤ N ∥ h i ∥ > 0. M=\max_{1\le i\le N}\|h_i\|>0. M = 1 ≤ i ≤ N max ∥ h i ∥ > 0.
For every unit vector x ∈ H x\in\mathcal{H} x ∈ H , choose an index i x i_x i x satisfying
⟨ x , g i x ⟩ = max 1 ≤ i ≤ N ⟨ x , g i ⟩ . \langle x,g_{i_x}\rangle=\max_{1\le i\le N}\langle x,g_i\rangle. ⟨ x , g i x ⟩ = 1 ≤ i ≤ N max ⟨ x , g i ⟩ .
Then
⟨ x , g i x + t h i x ⟩ = ⟨ x , g i x ⟩ + t ⟨ x , h i x ⟩ ≥ δ − ∣ t ∣ ∣ ⟨ x , h i x ⟩ ∣ ≥ δ − ∣ t ∣ ∥ h i x ∥ ≥ δ − ∣ t ∣ M . \begin{aligned}
\langle x,g_{i_x}+t h_{i_x}\rangle=\langle x,g_{i_x}\rangle+t\langle x,h_{i_x}\rangle\\
&\ge\delta-|t|\lvert\langle x,h_{i_x}\rangle\rvert\\
&\ge\delta-|t|\lVert h_{i_x}\rVert\\
&\ge\delta-|t|M.
\end{aligned} ⟨ x , g i x + t h i x ⟩ = ⟨ x , g i x ⟩ + t ⟨ x , h i x ⟩ ≥ δ − ∣ t ∣ ∣⟨ x , h i x ⟩∣ ≥ δ − ∣ t ∣ ∥ h i x ∥ ≥ δ − ∣ t ∣ M .
Choose ε = δ 2 M > 0 \varepsilon=\frac{\delta}{2M}>0 ε = 2 M δ > 0 . If ∣ t ∣ < ε |t|<\varepsilon ∣ t ∣ < ε , then
⟨ x , g i x + t h i x ⟩ > δ 2 > 0 \langle x,g_{i_x}+t h_{i_x}\rangle>\frac{\delta}{2}>0 ⟨ x , g i x + t h i x ⟩ > 2 δ > 0
for every unit vector x ∈ H n x\in\mathcal{H}_n x ∈ H n . Therefore,
max 1 ≤ i ≤ N ⟨ x , g i + t h i ⟩ > 0 , x ∈ S H . \max_{1\le i\le N}\langle x,g_i+t h_i\rangle>0,\qquad x\in S_{\mathcal{H}}. 1 ≤ i ≤ N max ⟨ x , g i + t h i ⟩ > 0 , x ∈ S H .
By the characterization of positive spanning sets, G t G_t G t is a positive frame. Hence,
G t ∈ D + ( F ) whenever ∣ t ∣ < ε . G_t\in\mathcal{D}_+(\mathcal{F})\qquad\text{whenever }|t|<\varepsilon. G t ∈ D + ( F ) whenever ∣ t ∣ < ε .
Now choose any t 0 t_0 t 0 satisfying 0 < t 0 < ε 0<t_0<\varepsilon 0 < t 0 < ε and define
G + = G + t 0 H , G − = G − t 0 H . G^+=G+t_0H,\qquad G^-=G-t_0H. G + = G + t 0 H , G − = G − t 0 H .
Then G + , G − ∈ D + ( F ) G^+,G^-\in\mathcal{D}_+(\mathcal{F}) G + , G − ∈ D + ( F ) . Moreover, since H ≠ 0 H\ne0 H = 0 , we have G + ≠ G − G^+\ne G^- G + = G − , while G = 1 2 G + + 1 2 G − G=\frac{1}{2}G^++\frac{1}{2}G^- G = 2 1 G + + 2 1 G − . Thus G G G is not an extreme point of D + ( F ) \mathcal{D}_+(\mathcal{F}) D + ( F ) . Since G ∈ D + ( F ) G\in\mathcal{D}_+(\mathcal{F}) G ∈ D + ( F ) was arbitrary, we conclude that Ext ( D + ( F ) ) = ∅ \operatorname{Ext}(\mathcal{D}_+(\mathcal{F}))=\varnothing Ext ( D + ( F )) = ∅ .
Example 4.8. Let { e i } i = 1 n \{e_i\}_{i=1}^n { e i } i = 1 n be the standard orthonormal basis of R n \mathbb{R}^n R n , where n ≥ 2 n\ge2 n ≥ 2 , and consider the positive frame
F = { e 1 , − e 1 , … , e n , − e n } . \mathcal{F}=\{e_1,-e_1,\ldots,e_n,-e_n\}. F = { e 1 , − e 1 , … , e n , − e n } .
For any invertible matrix A ∈ G L ( n , R ) A\in GL(n,\mathbb{R}) A ∈ G L ( n , R ) , define
G A = { 1 2 A e 1 , − 1 2 A e 1 , … , 1 2 A e n , − 1 2 A e n } . G_A=\left\{\frac{1}{2}Ae_1,-\frac{1}{2}Ae_1,\ldots,\frac{1}{2}Ae_n,-\frac{1}{2}Ae_n\right\}. G A = { 2 1 A e 1 , − 2 1 A e 1 , … , 2 1 A e n , − 2 1 A e n } .
Since
∑ i = 1 n ( ⟨ x , e i ⟩ 1 2 A e i + ⟨ x , − e i ⟩ ( − 1 2 A e i ) ) = A x , \sum_{i=1}^{n}\left(\langle x,e_i\rangle\frac{1}{2}Ae_i+\langle x,-e_i\rangle\left(-\frac{1}{2}Ae_i\right)\right)=Ax, i = 1 ∑ n ( ⟨ x , e i ⟩ 2 1 A e i + ⟨ x , − e i ⟩ ( − 2 1 A e i ) ) = A x ,
the family G A G_A G A is a dual frame of F \mathcal{F} F if and only if A = I n A=I_n A = I n . Thus, to obtain distinct dual frames, we instead consider matrices B ∈ R n × n B\in\mathbb{R}^{n\times n} B ∈ R n × n and set
G B = { 1 2 e i + B e i , − 1 2 e i + B e i } i = 1 n . G_B=\left\{\frac{1}{2}e_i+Be_i,-\frac{1}{2}e_i+Be_i\right\}_{i=1}^{n}. G B = { 2 1 e i + B e i , − 2 1 e i + B e i } i = 1 n .
For every B B B , the family G B G_B G B is a dual frame of F \mathcal{F} F , since
∑ i = 1 n ( ⟨ x , e i ⟩ ( 1 2 e i + B e i ) + ⟨ x , − e i ⟩ ( − 1 2 e i + B e i ) ) = x . \sum_{i=1}^{n}\left(\langle x,e_i\rangle\left(\frac{1}{2}e_i+Be_i\right)+\langle x,-e_i\rangle\left(-\frac{1}{2}e_i+Be_i\right)\right)=x. i = 1 ∑ n ( ⟨ x , e i ⟩ ( 2 1 e i + B e i ) + ⟨ x , − e i ⟩ ( − 2 1 e i + B e i ) ) = x .
Now take B 1 = I n B_1=I_n B 1 = I n and B 2 = − I n B_2=-I_n B 2 = − I n . The corresponding dual frames are
G B 1 = { 3 2 e i , 1 2 e i } i = 1 n , G B 2 = { − 1 2 e i , − 3 2 e i } i = 1 n . G_{B_1}=\left\{\frac{3}{2}e_i,\frac{1}{2}e_i\right\}_{i=1}^{n},\qquad G_{B_2}=\left\{-\frac{1}{2}e_i,-\frac{3}{2}e_i\right\}_{i=1}^{n}. G B 1 = { 2 3 e i , 2 1 e i } i = 1 n , G B 2 = { − 2 1 e i , − 2 3 e i } i = 1 n .
Neither of these families is a positive frame. Instead, choose B 1 = 1 4 I n B_1=\frac{1}{4}I_n B 1 = 4 1 I n and B 2 = − 1 4 I n B_2=-\frac{1}{4}I_n B 2 = − 4 1 I n . Then
G B 1 = { 3 4 e i , − 1 4 e i } i = 1 n , G B 2 = { 1 4 e i , − 3 4 e i } i = 1 n . G_{B_1}=\left\{\frac{3}{4}e_i,-\frac{1}{4}e_i\right\}_{i=1}^{n},\qquad G_{B_2}=\left\{\frac{1}{4}e_i,-\frac{3}{4}e_i\right\}_{i=1}^{n}. G B 1 = { 4 3 e i , − 4 1 e i } i = 1 n , G B 2 = { 4 1 e i , − 4 3 e i } i = 1 n .
Both are positive dual frames of F F F . Their midpoint is
1 2 ( G B 1 + G B 2 ) = { 1 2 e i , − 1 2 e i } i = 1 n , \frac{1}{2}(G_{B_1}+G_{B_2})=\left\{\frac{1}{2}e_i,-\frac{1}{2}e_i\right\}_{i=1}^{n}, 2 1 ( G B 1 + G B 2 ) = { 2 1 e i , − 2 1 e i } i = 1 n ,
which is also a positive frame. Hence this particular pair does not witness nonconvexity.
We next examine the topological structure of the collection of positive dual frames. Since positivity is characterized by the existence of a strictly positive linear relation, it is natural to ask whether this property is preserved under small perturbations that maintain the duality condition. The following proposition shows that positivity is stable within the affine space of dual frames.
Proposition 4.9. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n . Then the set D + ( F ) \mathcal{D}_{+}(F) D + ( F ) of all positive dual frames of F F F is relatively open in the set D ( F ) \mathcal{D}(F) D ( F ) of all dual frames of F F F .
Proof. Let G = { g i } i = 1 N ∈ D + ( F ) G=\{g_i\}_{i=1}^{N}\in\mathcal{D}_{+}(F) G = { g i } i = 1 N ∈ D + ( F ) be arbitrary. Since G G G is a positive frame, for every unit vector x ∈ H n x\in\mathcal{H}_n x ∈ H n , we have
max 1 ≤ i ≤ N ⟨ x , g i ⟩ > 0. \max_{1\leq i\leq N}\langle x,g_i\rangle>0. 1 ≤ i ≤ N max ⟨ x , g i ⟩ > 0.
Consider the function φ : S H n → R \varphi:S_{\mathcal{H}_n}\to\mathbb{R} φ : S H n → R defined by φ ( x ) = max 1 ≤ i ≤ N ⟨ x , g i ⟩ \varphi(x)=\max_{1\leq i\leq N}\langle x,g_i\rangle φ ( x ) = max 1 ≤ i ≤ N ⟨ x , g i ⟩ , where S H n = { x ∈ H : ∥ x ∥ = 1 } S_{\mathcal{H}_n}=\{x\in\mathcal{H}:\|x\|=1\} S H n = { x ∈ H : ∥ x ∥ = 1 } . Since φ \varphi φ is continuous and S H n S_{\mathcal{H}_n} S H n is compact, there exists δ > 0 \delta>0 δ > 0 such that
δ = min ∥ x ∥ = 1 max 1 ≤ i ≤ N ⟨ x , g i ⟩ > 0. \delta=\min_{\|x\|=1}\max_{1\leq i\leq N}\langle x,g_i\rangle>0. δ = ∥ x ∥ = 1 min 1 ≤ i ≤ N max ⟨ x , g i ⟩ > 0.
Let ε = δ / 2 \varepsilon=\delta/2 ε = δ /2 and consider any H = { h i } i = 1 N ∈ D ( F ) H=\{h_i\}_{i=1}^{N}\in\mathcal{D}(F) H = { h i } i = 1 N ∈ D ( F ) satisfying
max 1 ≤ i ≤ N ∥ g i − h i ∥ < ε . \max_{1\leq i\leq N}\|g_i-h_i\|<\varepsilon. 1 ≤ i ≤ N max ∥ g i − h i ∥ < ε .
For every unit vector x ∈ H n x\in\mathcal{H}_n x ∈ H n , choose an index i x i_x i x such that ⟨ x , g i x ⟩ = max 1 ≤ i ≤ N ⟨ x , g i ⟩ \langle x,g_{i_x}\rangle=\max_{1\leq i\leq N}\langle x,g_i\rangle ⟨ x , g i x ⟩ = max 1 ≤ i ≤ N ⟨ x , g i ⟩ . Then, by the Cauchy–Schwarz inequality,
max 1 ≤ i ≤ N ⟨ x , h i ⟩ ≥ ⟨ x , h i x ⟩ = ⟨ x , g i x ⟩ + ⟨ x , h i x − g i x ⟩ ≥ δ − ∥ h i x − g i x ∥ > δ − ε = δ 2 > 0. \begin{aligned}
\max_{1\leq i\leq N}\langle x,h_i\rangle
&\geq\langle x,h_{i_x}\rangle\\
&=\langle x,g_{i_x}\rangle+\langle x,h_{i_x}-g_{i_x}\rangle\\
&\geq\delta-\|h_{i_x}-g_{i_x}\|\\
&>\delta-\varepsilon=\frac{\delta}{2}>0.
\end{aligned} 1 ≤ i ≤ N max ⟨ x , h i ⟩ ≥ ⟨ x , h i x ⟩ = ⟨ x , g i x ⟩ + ⟨ x , h i x − g i x ⟩ ≥ δ − ∥ h i x − g i x ∥ > δ − ε = 2 δ > 0.
Since this holds for every unit vector x ∈ H n x\in\mathcal{H}_n x ∈ H n , the family H H H positively spans H n \mathcal{H}_n H n . Moreover, since H ∈ D ( F ) H\in\mathcal{D}(F) H ∈ D ( F ) , we have T H ∗ T F = I H n T_H^*T_F=I_{\mathcal{H}_n} T H ∗ T F = I H n . Consequently, H ∈ D + ( F ) H\in\mathcal{D}_{+}(F) H ∈ D + ( F ) .
Thus,
{ H ∈ D ( F ) : max 1 ≤ i ≤ N ∥ g i − h i ∥ < ε } ⊆ D + ( F ) . \left\{H \in\mathcal{D}(F): \max_{1\leq i\leq N}\|g_i-h_i\|<\varepsilon\right\}\subseteq\mathcal{D}_{+}(F). { H ∈ D ( F ) : 1 ≤ i ≤ N max ∥ g i − h i ∥ < ε } ⊆ D + ( F ) .
Since G ∈ D + ( F ) G\in\mathcal{D}_{+}(F) G ∈ D + ( F ) was arbitrary, we conclude that D + ( F ) \mathcal{D}_{+}(F) D + ( F ) is relatively open in D ( F ) \mathcal{D}(F) D ( F ) .
The minimal nonnegative linear dependencies among frame vectors provide useful information about the geometric structure of positive frames. In particular, we show that positive spanning can be characterized through the supports of extremal nonnegative relations. To investigate this connection, let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for R n \mathbb{R}^{n} R n and consider the normalized positive-dependence set
Ω F = { λ = ( λ i ) i = 1 N ∈ R ≥ 0 N : ∑ i = 1 N λ i f i = 0 , ∑ i = 1 N λ i = 1 } . \Omega_F=\left\{\lambda=(\lambda_i)_{i=1}^{N}\in\mathbb{R}_{\geq0}^{N}:\sum_{i=1}^{N}\lambda_i f_i=0,\quad\sum_{i=1}^{N}\lambda_i=1\right\}. Ω F = { λ = ( λ i ) i = 1 N ∈ R ≥ 0 N : i = 1 ∑ N λ i f i = 0 , i = 1 ∑ N λ i = 1 } .
For λ ∈ Ω F \lambda\in\Omega_F λ ∈ Ω F , write supp ( λ ) = { i : λ i > 0 } \operatorname{supp}(\lambda)=\{i:\lambda_i>0\} supp ( λ ) = { i : λ i > 0 } . Since Ω F \Omega_F Ω F is a compact convex polytope, its extreme points encode minimal nonnegative dependencies. The following proposition establishes that F F F is a positive frame if and only if the supports of these extreme points collectively cover all the frame indices.
Proposition 4.10. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for R n \mathbb{R}^{n} R n . Then the following statements are equivalent:
(i) F F F is a positive frame.
(ii) For every i ∈ { 1 , … , N } i\in\{1,\ldots,N\} i ∈ { 1 , … , N } , there exists λ ( i ) ∈ Ext ( Ω F ) \lambda^{(i)}\in\operatorname{Ext}(\Omega_F) λ ( i ) ∈ Ext ( Ω F ) such that λ i ( i ) > 0 \lambda_i^{(i)}>0 λ i ( i ) > 0 .
(iii)
⋃ λ ∈ Ext ( Ω F ) supp ( λ ) = { 1 , … , N } . \bigcup_{\lambda\in\operatorname{Ext}(\Omega_F)}\operatorname{supp}(\lambda)=\{1,\ldots,N\}. λ ∈ Ext ( Ω F ) ⋃ supp ( λ ) = { 1 , … , N } .
Proof. (i)⇒ \Rightarrow ⇒ (ii). Since F F F is a positive frame, there exist c i > 0 c_i>0 c i > 0 , 1 ≤ i ≤ N 1\leq i\leq N 1 ≤ i ≤ N , such that ∑ i = 1 N c i f i = 0 \sum_{i=1}^{N}c_i f_i=0 ∑ i = 1 N c i f i = 0 . Normalizing the coefficients gives
λ i = c i ∑ j = 1 N c j , 1 ≤ i ≤ N , \lambda_i=\frac{c_i}{\sum_{j=1}^{N}c_j},\qquad1\leq i\leq N, λ i = ∑ j = 1 N c j c i , 1 ≤ i ≤ N ,
so that λ ∈ Ω F \lambda\in\Omega_F λ ∈ Ω F and supp ( λ ) = { 1 , … , N } \operatorname{supp}(\lambda)=\{1,\ldots,N\} supp ( λ ) = { 1 , … , N } .
Since Ω F \Omega_F Ω F is a compact convex polytope, it is the convex hull of its extreme points. Thus, there exist λ ( 1 ) , … , λ ( m ) ∈ Ext ( Ω F ) \lambda^{(1)},\ldots,\lambda^{(m)}\in\operatorname{Ext}(\Omega_F) λ ( 1 ) , … , λ ( m ) ∈ Ext ( Ω F ) and t 1 , … , t m ≥ 0 t_1,\ldots,t_m\geq0 t 1 , … , t m ≥ 0 , with ∑ j = 1 m t j = 1 \sum_{j=1}^{m}t_j=1 ∑ j = 1 m t j = 1 , such that
λ = ∑ j = 1 m t j λ ( j ) . \lambda=\sum_{j=1}^{m}t_j\lambda^{(j)}. λ = j = 1 ∑ m t j λ ( j ) .
We claim that for every i ∈ { 1 , … , N } i\in\{1,\ldots,N\} i ∈ { 1 , … , N } , at least one extreme point has a strictly positive i i i th coordinate. Otherwise, for some i 0 i_0 i 0 , we would have λ i 0 ( j ) = 0 \lambda_{i_0}^{(j)}=0 λ i 0 ( j ) = 0 for all j j j , which implies
λ i 0 = ∑ j = 1 m t j λ i 0 ( j ) = 0 , \lambda_{i_0}=\sum_{j=1}^{m}t_j\lambda_{i_0}^{(j)}=0, λ i 0 = j = 1 ∑ m t j λ i 0 ( j ) = 0 ,
contradicting λ i 0 > 0 \lambda_{i_0}>0 λ i 0 > 0 . Hence (ii) holds.
(ii)⇒ \Rightarrow ⇒ (iii). For every i ∈ { 1 , … , N } i\in\{1,\ldots,N\} i ∈ { 1 , … , N } , condition (ii) provides an extreme point λ ( i ) \lambda^{(i)} λ ( i ) with λ i ( i ) > 0 \lambda_i^{(i)}>0 λ i ( i ) > 0 . Thus, i ∈ supp ( λ ( i ) ) i\in\operatorname{supp}(\lambda^{(i)}) i ∈ supp ( λ ( i ) ) , and consequently
{ 1 , … , N } ⊆ ⋃ λ ∈ Ext ( Ω F ) supp ( λ ) . \{1,\ldots,N\}\subseteq\bigcup_{\lambda\in\operatorname{Ext}(\Omega_F)}\operatorname{supp}(\lambda). { 1 , … , N } ⊆ λ ∈ Ext ( Ω F ) ⋃ supp ( λ ) .
The reverse inclusion is immediate, so (iii) follows.
(iii)⇒ \Rightarrow ⇒ (i). For each i ∈ { 1 , … , N } i \in\{1,\ldots,N\} i ∈ { 1 , … , N } , choose λ ( i ) ∈ Ext ( Ω F ) \lambda^{(i)} \in\operatorname{Ext}(\Omega_F) λ ( i ) ∈ Ext ( Ω F ) such that λ i ( i ) > 0 \lambda_i^{(i)} > 0 λ i ( i ) > 0 . Define
λ = 1 N ∑ i = 1 N λ ( i ) . \lambda= \frac{1}{N}\sum_{i=1}^{N}\lambda^{(i)}. λ = N 1 i = 1 ∑ N λ ( i ) .
By convexity, λ ∈ Ω F \lambda\in\Omega_F λ ∈ Ω F . Moreover, for every j ∈ { 1 , … , N } j \in\{1,\ldots,N\} j ∈ { 1 , … , N } ,
λ j = 1 N ∑ i = 1 N λ j ( i ) ≥ 1 N λ j ( j ) > 0. \lambda_j = \frac{1}{N}\sum_{i=1}^{N}\lambda_j^{(i)} \geq\frac{1}{N}\lambda_j^{(j)} > 0. λ j = N 1 i = 1 ∑ N λ j ( i ) ≥ N 1 λ j ( j ) > 0.
Therefore, λ ∈ ker T F ∗ ∩ R + + N \lambda\in\ker T_F^* \cap\mathbb{R}_{++}^{N} λ ∈ ker T F ∗ ∩ R ++ N . Since F F F spans R n \mathbb{R}^n R n , the existence of this strictly positive dependence implies that F F F is a positive frame.
The preceding result suggests that the boundary of the set of positive dual frames can be studied through the interaction between synthesis kernels and the positive orthant. In particular, positivity fails precisely when the kernel contains no strictly positive vector. This leads to the following characterization.
We next examine the boundary structure of the set of positive dual frames. Recall that D ( F ) \mathcal{D}(F) D ( F ) denotes the collection of all dual frames of F F F , equipped with the topology induced by the Euclidean norm on H N \mathcal{H}^N H N . For a subset A ⊆ D ( F ) \mathcal{A} \subseteq\mathcal{D}(F) A ⊆ D ( F ) , we denote its relative closure and relative boundary in D ( F ) \mathcal{D}(F) D ( F ) by A ‾ D ( F ) \overline{\mathcal{A}}^{\mathcal{D}(F)} A D ( F ) and ∂ D ( F ) A \partial_{\mathcal{D}(F)}\mathcal{A} ∂ D ( F ) A , respectively. Thus,
∂ D ( F ) A = A ‾ D ( F ) ∖ int D ( F ) ( A ) , \partial_{\mathcal{D}(F)}\mathcal{A}
=
\overline{\mathcal{A}}^{\mathcal{D}(F)}
\setminus\operatorname{int}_{\mathcal{D}(F)}(\mathcal{A}), ∂ D ( F ) A = A D ( F ) ∖ int D ( F ) ( A ) ,
where int D ( F ) ( A ) \operatorname{int}_{\mathcal{D}(F)}(\mathcal{A}) int D ( F ) ( A ) denotes the relative interior of A \mathcal{A} A .
Since D + ( F ) \mathcal{D}_+(F) D + ( F ) is relatively open in D ( F ) \mathcal{D}(F) D ( F ) , its boundary can be described in terms of the absence of strictly positive linear relations among the vectors of a dual frame. This observation yields the following characterization.
Proposition 4.11. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n . Then
∂ D ( F ) D + ( F ) = { G ∈ D + ( F ) ‾ D ( F ) : ker T G ∗ ∩ R + + N = ∅ } . \partial_{\mathcal{D}(F)}\mathcal{D}_+(F)
=
\left\{
G \in\overline{\mathcal{D}_+(F)}^{\mathcal{D}(F)}
:
\ker T_G^* \cap\mathbb{R}_{++}^{N} = \varnothing
\right\}. ∂ D ( F ) D + ( F ) = { G ∈ D + ( F ) D ( F ) : ker T G ∗ ∩ R ++ N = ∅ } .
Proof. Since D + ( F ) \mathcal{D}_+(F) D + ( F ) is relatively open in D ( F ) \mathcal{D}(F) D ( F ) , we have
∂ D ( F ) D + ( F ) = D + ( F ) ‾ D ( F ) ∖ D + ( F ) . \partial_{\mathcal{D}(F)}\mathcal{D}_+(F)
=
\overline{\mathcal{D}_+(F)}^{\mathcal{D}(F)}
\setminus\mathcal{D}_+(F). ∂ D ( F ) D + ( F ) = D + ( F ) D ( F ) ∖ D + ( F ) .
On the other hand, by the characterization of positive frames, a dual frame G G G is positive if and only if
ker T G ∗ ∩ R + + N ≠ ∅ . \ker T_G^* \cap\mathbb{R}_{++}^{N} \ne\varnothing. ker T G ∗ ∩ R ++ N = ∅ .
Consequently,
G ∉ D + ( F ) ⟺ ker T G ∗ ∩ R + + N = ∅ , G \notin\mathcal{D}_+(F)
\quad\Longleftrightarrow\quad
\ker T_G^* \cap\mathbb{R}_{++}^{N} = \varnothing, G ∈ / D + ( F ) ⟺ ker T G ∗ ∩ R ++ N = ∅ ,
for every G ∈ D ( F ) G \in\mathcal{D}(F) G ∈ D ( F ) . Combining these observations gives
∂ D ( F ) D + ( F ) = { G ∈ D + ( F ) ‾ D ( F ) : ker T G ∗ ∩ R + + N = ∅ } , \partial_{\mathcal{D}(F)}\mathcal{D}_+(F)
=
\left\{
G \in\overline{\mathcal{D}_+(F)}^{\mathcal{D}(F)}
:
\ker T_G^* \cap\mathbb{R}_{++}^{N} = \varnothing
\right\}, ∂ D ( F ) D + ( F ) = { G ∈ D + ( F ) D ( F ) : ker T G ∗ ∩ R ++ N = ∅ } ,
as required.
We next investigate the geometry of positive dual frames through the nonnegative linear relations among their elements. Let F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n , and let W F = Ran T F W_F = \operatorname{Ran} T_F W F = Ran T F . For each dual frame G ∈ D ( F ) G \in\mathcal{D}(F) G ∈ D ( F ) , define
K G + = ker T G ∗ ∩ R + N . K_G^+ = \ker T_G^* \cap\mathbb{R}_+^N. K G + = ker T G ∗ ∩ R + N .
Thus, K G + K_G^+ K G + consists of all nonnegative linear relations among the elements of G G G . In particular, G G G is positive if and only if K G + ∩ R + + N ≠ ∅ K_G^+ \cap\mathbb{R}_{++}^N \ne\varnothing K G + ∩ R ++ N = ∅ . To quantify the behavior of such relations, we associate with each G ∈ D ( F ) G \in\mathcal{D}(F) G ∈ D ( F ) the quantity
δ ( G ) = min a ∈ R + N ∥ a ∥ = 1 ∥ T G ∗ a ∥ . \delta(G) = \min_{\substack{a \in\mathbb{R}_+^N \\ \lVert a \rVert= 1}} \lVert T_G^* a \rVert. δ ( G ) = a ∈ R + N ∥ a ∥ = 1 min ∥ T G ∗ a ∥ .
Since the nonnegative unit sphere is compact, this minimum is attained. Moreover, δ ( G ) = 0 \delta(G) = 0 δ ( G ) = 0 precisely when K G + K_G^+ K G + contains a nonzero vector. Consequently, δ ( G ) > 0 \delta(G) > 0 δ ( G ) > 0 implies that G G G admits no nontrivial nonnegative linear relation and, in particular, is not a positive frame. The following theorem establishes a continuity property of this quantity and describes its relationship with positive dual frames.
Proposition 4.12. Let F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for an n n n -dimensional real Hilbert space H n \mathcal{H}_n H n . Then the following statements hold:
The function δ : D ( F ) → [ 0 , ∞ ) \delta: \mathcal{D}(F) \to[0,\infty) δ : D ( F ) → [ 0 , ∞ ) is Lipschitz continuous with respect to the operator norm on synthesis operators.
If δ ( G ) > 0 \delta(G) > 0 δ ( G ) > 0 , then G G G is not a positive dual frame of F F F .
If G ∈ D + ( F ) G \in\mathcal{D}_+(F) G ∈ D + ( F ) , then δ ( G ) = 0 \delta(G) = 0 δ ( G ) = 0 .
Proof. Let G , H ∈ D ( F ) G, H \in\mathcal{D}(F) G , H ∈ D ( F ) . For every a ∈ R + N a \in\mathbb{R}_+^N a ∈ R + N with $\lVert a \rVert = 1,
∣ ∥ T G ∗ a ∥ − ∥ T H ∗ a ∥ ∣ ≤ ∥ T G ∗ − T H ∗ ∥ . \left|\lVert T_G^*a\rVert-\lVert T_H^*a\rVert\right| \leq\lVert T_G^*-T_H^*\rVert. ∣ ∥ T G ∗ a ∥ − ∥ T H ∗ a ∥ ∣ ≤ ∥ T G ∗ − T H ∗ ∥ .
Taking minima over the nonnegative unit sphere gives
∣ δ ( G ) − δ ( H ) ∣ ≤ ∥ T G ∗ − T H ∗ ∥ , |\delta(G)-\delta(H)| \leq\lVert T_G^*-T_H^*\rVert, ∣ δ ( G ) − δ ( H ) ∣ ≤ ∥ T G ∗ − T H ∗ ∥ ,
which proves the first assertion.
If G G G is positive, there exists a ∈ R + + N a \in\mathbb{R}_{++}^N a ∈ R ++ N such that T G ∗ a = 0 T_G^*a = 0 T G ∗ a = 0 . Normalizing a a a yields δ ( G ) = 0 \delta(G) = 0 δ ( G ) = 0 . The remaining assertions follow immediately. □ \square □
Semi-Positive Frames and Approximation by Positive Frames
We now investigate a broader class of frames that admit nontrivial nonnegative linear relations. Although such frames need not be positive, they possess a natural connection with positive frames through approximation. In particular, we show that every semi-positive frame can be approximated by a sequence of positive frames. This provides a useful description of the closure of the collection of positive frames.
Definition 5.1. Let F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for H n \mathcal{H}_n H n . We say that F F F is a semi-positive frame if there exists a nonzero vector c = ( c i ) i = 1 N ∈ R ≥ 0 N c = (c_i)_{i=1}^{N} \in\mathbb{R}_{\geq0}^{N} c = ( c i ) i = 1 N ∈ R ≥ 0 N with at least one zero coordinate such that T F ∗ c = 0 T_F^*c = 0 T F ∗ c = 0 .
The following proposition establishes an approximation property of semi-positive frames.
Proposition 5.2. Let F = ( f i ) i = 1 N F=(f_i)_{i=1}^N F = ( f i ) i = 1 N be a frame for H n \mathcal{H}_n H n , where N > n N>n N > n . If F F F is semi-positive, then F F F belongs to the closure of the collection of positive frames with N N N elements in H n \mathcal{H}_n H n .
Proof. Since F F F is semi-positive, there exists a nonzero vector c ∈ R ≥ 0 N c\in\mathbb{R}_{\geq0}^N c ∈ R ≥ 0 N such that T F ∗ c = 0 T_F^*c=0 T F ∗ c = 0 .
Choose d ∈ R + + N d\in\mathbb{R}_{++}^N d ∈ R ++ N and define
c ( m ) = c + 1 m d , m ∈ N . c^{(m)}=c+\frac{1}{m}d,\qquad m\in\mathbb{N}. c ( m ) = c + m 1 d , m ∈ N . Clearly, c ( m ) ∈ R + + N c^{(m)}\in\mathbb{R}_{++}^N c ( m ) ∈ R ++ N and c ( m ) → c ≠ 0 c^{(m)}\to c\ne0 c ( m ) → c = 0 .
For each m ∈ N m\in\mathbb{N} m ∈ N , define
A m = T F ∗ − ( T F ∗ c ( m ) ) ( c ( m ) ) T ∥ c ( m ) ∥ 2 . A_m=T_F^*-\frac{(T_F^*c^{(m)})(c^{(m)})^T}{\lVert c^{(m)}\rVert^2}. A m = T F ∗ − ∥ c ( m ) ∥ 2 ( T F ∗ c ( m ) ) ( c ( m ) ) T . Then
A m c ( m ) = T F ∗ c ( m ) − T F ∗ c ( m ) ∥ c ( m ) ∥ 2 ∥ c ( m ) ∥ 2 = 0. A_mc^{(m)}=T_F^*c^{(m)}-\frac{T_F^*c^{(m)}}{\lVert c^{(m)}\rVert^2}\lVert c^{(m)}\rVert^2=0. A m c ( m ) = T F ∗ c ( m ) − ∥ c ( m ) ∥ 2 T F ∗ c ( m ) ∥ c ( m ) ∥ 2 = 0. Moreover, since T F ∗ c = 0 T_F^*c=0 T F ∗ c = 0 , we have
T F ∗ c ( m ) = 1 m T F ∗ d ⟶ 0. T_F^*c^{(m)}=\frac{1}{m}T_F^*d\longrightarrow0. T F ∗ c ( m ) = m 1 T F ∗ d ⟶ 0. As c ( m ) → c ≠ 0 c^{(m)}\to c\ne0 c ( m ) → c = 0 , it follows that A m → T F ∗ A_m\to T_F^* A m → T F ∗ in operator norm.
Since T F ∗ T_F^* T F ∗ is surjective, it has full row rank. By the openness of the set of full-row-rank matrices, A m A_m A m is also surjective for all sufficiently large m m m .
For such m m m , let F m = ( f i ( m ) ) i = 1 N F_m=(f_i^{(m)})_{i=1}^N F m = ( f i ( m ) ) i = 1 N be the family whose synthesis operator is A m A_m A m . Then F m F_m F m is a frame for H n \mathcal{H}_n H n . Furthermore,
T F m ∗ c ( m ) = A m c ( m ) = 0 , c ( m ) ∈ R + + N . T_{F_m}^*c^{(m)}=A_mc^{(m)}=0,\qquad c^{(m)}\in\mathbb{R}_{++}^N. T F m ∗ c ( m ) = A m c ( m ) = 0 , c ( m ) ∈ R ++ N . Thus F m F_m F m is a positive frame. Finally, A m → T F ∗ A_m\to T_F^* A m → T F ∗ implies that f i ( m ) → f i f_i^{(m)}\to f_i f i ( m ) → f i for every i = 1 , … , N i=1,\ldots,N i = 1 , … , N . Therefore, F F F belongs to the closure of the collection of positive frames with N N N elements in H n \mathcal{H}_n H n .
The following result describes the relative boundary of the collection of positive dual frames in terms of nonnegative linear relations.
Proposition 5.3. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^N F = { f i } i = 1 N be a redundant frame for H n \mathcal{H}_n H n . Then
∂ D ( F ) D + ( F ) = { G ∈ D ( F ) : G is semi-positive } . \partial_{\mathcal{D}(F)}\mathcal{D}_+(F)=\{G\in\mathcal{D}(F):G\text{ is semi-positive}\}. ∂ D ( F ) D + ( F ) = { G ∈ D ( F ) : G is semi-positive } . In particular, every dual frame belonging to the relative boundary of D + ( F ) \mathcal{D}_+(F) D + ( F ) is semi-positive.
Proof. We first show that
D + ( F ) ‾ D ( F ) = { G ∈ D ( F ) : ker T G ∗ ∩ R ≥ 0 N ≠ { 0 } } . \overline{\mathcal{D}_+(F)}^{\mathcal{D}(F)}=\{G\in\mathcal{D}(F):\ker T_G^*\cap\mathbb{R}_{\geq0}^N\ne\{0\}\}. D + ( F ) D ( F ) = { G ∈ D ( F ) : ker T G ∗ ∩ R ≥ 0 N = { 0 }} . Let G G G belong to the closure of D + ( F ) \mathcal{D}_+(F) D + ( F ) in D ( F ) \mathcal{D}(F) D ( F ) . Then there exists a sequence G m ∈ D + ( F ) G_m\in\mathcal{D}_+(F) G m ∈ D + ( F ) such that G m → G G_m\to G G m → G . Since each G m G_m G m is positive, there exists c ( m ) ∈ R + + N c^{(m)}\in\mathbb{R}_{++}^N c ( m ) ∈ R ++ N satisfying
T G m ∗ c ( m ) = 0 , ∥ c ( m ) ∥ = 1. T_{G_m}^*c^{(m)}=0,\qquad\lVert c^{(m)}\rVert=1. T G m ∗ c ( m ) = 0 , ∥ c ( m ) ∥ = 1.
By compactness, passing to a subsequence if necessary, we may assume that c ( m ) → c c^{(m)} \to c c ( m ) → c for some c ∈ R ≥ 0 N c \in\mathbb{R}^{N}_{\geq0} c ∈ R ≥ 0 N with ∥ c ∥ = 1 \lVert c\rVert= 1 ∥ c ∥ = 1 . Since T G m ∗ → T G ∗ T_{G_m}^{*} \to T_{G}^{*} T G m ∗ → T G ∗ , we obtain
T G ∗ c = 0. T_{G}^{*}c = 0. T G ∗ c = 0.
Thus G G G admits a nontrivial nonnegative linear relation.
Conversely, suppose that G ∈ D ( F ) G \in\mathcal{D}(F) G ∈ D ( F ) and there exists c ∈ R ≥ 0 N ∖ { 0 } c \in\mathbb{R}^{N}_{\geq0} \setminus\{0\} c ∈ R ≥ 0 N ∖ { 0 } such that T G ∗ c = 0 T_{G}^{*}c = 0 T G ∗ c = 0 . Since F F F is redundant, we may choose d ∈ R + + N d \in\mathbb{R}^{N}_{++} d ∈ R ++ N such that d ∉ Ran T F d \notin\operatorname{Ran} T_{F} d ∈ / Ran T F . For each m ∈ N m \in\mathbb{N} m ∈ N , set
c ( m ) = c + 1 m d . c^{(m)} = c + \frac{1}{m}d. c ( m ) = c + m 1 d .
Then c ( m ) ∈ R + + N c^{(m)} \in\mathbb{R}^{N}_{++} c ( m ) ∈ R ++ N and c ( m ) → c c^{(m)} \to c c ( m ) → c .
Let P F = T F S F − 1 T F ∗ P_{F}=T_{F}S_{F}^{-1}T_{F}^{*} P F = T F S F − 1 T F ∗ and define
z m = ( I − P F ) c ( m ) . z_m=(I-P_F)c^{(m)}. z m = ( I − P F ) c ( m ) .
Since T G ∗ c = 0 T_{G}^{*}c=0 T G ∗ c = 0 and T G ∗ T F = I H n T_{G}^{*}T_{F}=I_{H_n} T G ∗ T F = I H n , we have c ∉ Ran T F c\notin\operatorname{Ran}T_F c ∈ / Ran T F . Consequently, z = ( I − P F ) c ≠ 0 z=(I-P_F)c\ne0 z = ( I − P F ) c = 0 , and hence z m ≠ 0 z_m\ne0 z m = 0 for sufficiently large m m m .
Define
L m = T G ∗ − ( T G ∗ c ( m ) ) z m T ∥ z m ∥ 2 . L_m=T_G^*-\frac{(T_G^*c^{(m)})z_m^T}{\lVert z_m\rVert^2}. L m = T G ∗ − ∥ z m ∥ 2 ( T G ∗ c ( m ) ) z m T .
Since z m ∈ ker T F ∗ z_m\in\ker T_F^* z m ∈ ker T F ∗ , we have
L m T F = T G ∗ T F = I H n . L_mT_F=T_G^*T_F=I_{H_n}. L m T F = T G ∗ T F = I H n .
Moreover,
L m c ( m ) = T G ∗ c ( m ) − ⟨ z m , c ( m ) ⟩ ∥ z m ∥ 2 T G ∗ c ( m ) = 0 , L_mc^{(m)}=T_G^*c^{(m)}-\frac{\langle z_m,c^{(m)}\rangle}{\lVert z_m\rVert^2}T_G^*c^{(m)}=0, L m c ( m ) = T G ∗ c ( m ) − ∥ z m ∥ 2 ⟨ z m , c ( m ) ⟩ T G ∗ c ( m ) = 0 ,
because ⟨ z m , c ( m ) ⟩ = ∥ z m ∥ 2 \langle z_m,c^{(m)}\rangle=\lVert z_m\rVert^2 ⟨ z m , c ( m ) ⟩ = ∥ z m ∥ 2 . Let G m G_m G m be the family whose synthesis operator is L m L_m L m . Then G m G_m G m is a dual frame of F F F , and the strictly positive relation L m c ( m ) = 0 L_mc^{(m)}=0 L m c ( m ) = 0 implies that G m G_m G m is a positive frame. Thus G m ∈ D + ( F ) G_m\in\mathcal{D}_+(F) G m ∈ D + ( F ) . Furthermore, since T G ∗ c = 0 T_G^*c=0 T G ∗ c = 0 ,
T G ∗ c ( m ) = 1 m T G ∗ d ⟶ 0. T_G^*c^{(m)}=\frac{1}{m}T_G^*d\longrightarrow0. T G ∗ c ( m ) = m 1 T G ∗ d ⟶ 0.
As z m → z ≠ 0 z_m\to z\ne0 z m → z = 0 , it follows that L m → T G ∗ L_m\to T_G^* L m → T G ∗ . Therefore, G m → G G_m\to G G m → G , proving the asserted closure characterization.
Since D + ( F ) \mathcal{D}_+(F) D + ( F ) is relatively open in D ( F ) \mathcal{D}(F) D ( F ) , its relative boundary is
∂ D ( F ) D + ( F ) = D + ( F ) ‾ D ( F ) ∖ D + ( F ) . \partial_{\mathcal{D}(F)}\mathcal{D}_+(F)=\overline{\mathcal{D}_+(F)}^{\mathcal{D}(F)}\setminus\mathcal{D}_+(F). ∂ D ( F ) D + ( F ) = D + ( F ) D ( F ) ∖ D + ( F ) .
Hence every boundary point admits a nontrivial nonnegative relation but no strictly positive relation. Therefore, every such dual frame is semi-positive but not positive. Conversely, any semi-positive dual frame that is not positive belongs to the closure established above, and hence to the relative boundary. This completes the proof.
We now consider the special case of frames with excess one. In this setting, the positive dual frames admit an explicit parametrization in terms of strictly positive relations. Let us define
Δ ∘ = { c ∈ R + + n + 1 : ∑ i = 1 n + 1 c i = 1 } . \Delta^\circ=\left\{c\in\mathbb{R}^{n+1}_{++}:\sum_{i=1}^{n+1}c_i=1\right\}. Δ ∘ = { c ∈ R ++ n + 1 : i = 1 ∑ n + 1 c i = 1 } .
Proposition 5.4. Let F = { f i } i = 1 n + 1 F=\{f_i\}_{i=1}^{n+1} F = { f i } i = 1 n + 1 be a frame for H n \mathcal{H}_n H n . Then the collection of positive dual frames of F F F is in one-to-one correspondence with Δ ∘ ∖ Ran T F \Delta^\circ\setminus\operatorname{Ran}T_F Δ ∘ ∖ Ran T F . Moreover, D + ( F ) \mathcal{D}_+(F) D + ( F ) has either one or two connected components, each of which is contractible.
Proof. Let G ∈ D + ( F ) G\in\mathcal{D}_+(F) G ∈ D + ( F ) . Since G G G has n + 1 n+1 n + 1 vectors, dim ker T G ∗ = 1 \dim\ker T_G^*=1 dim ker T G ∗ = 1 . Positivity implies that this kernel is generated by a strictly positive vector. Normalizing its coordinates gives a unique c ∈ Δ ∘ c\in\Delta^\circ c ∈ Δ ∘ such that T G ∗ c = 0 T_G^*c=0 T G ∗ c = 0 . Since T G ∗ T F = I H n T_G^*T_F=I_{\mathcal{H}_n} T G ∗ T F = I H n , we necessarily have c ∉ Ran T F c\notin\operatorname{Ran}T_F c ∈ / Ran T F .
Conversely, let c ∈ Δ ∘ ∖ Ran T F c\in\Delta^\circ\setminus\operatorname{Ran}T_F c ∈ Δ ∘ ∖ Ran T F . Since
R n + 1 = Ran T F ⊕ span { c } , \mathbb{R}^{n+1}=\operatorname{Ran}T_F\oplus\operatorname{span}\{c\}, R n + 1 = Ran T F ⊕ span { c } ,
there exists a unique linear map L c : R n + 1 → H n L_c:\mathbb{R}^{n+1}\to\mathcal{H}_n L c : R n + 1 → H n satisfying
L c T F = I H n , L c c = 0. L_cT_F=I_{\mathcal{H}_n},\qquad L_cc=0. L c T F = I H n , L c c = 0.
The family G c = { L c e i } i = 1 n + 1 G_c=\{L_ce_i\}_{i=1}^{n+1} G c = { L c e i } i = 1 n + 1 is therefore a positive dual frame. Uniqueness shows that this correspondence is bijective.
Furthermore, the correspondence c ↦ G c c\mapsto G_c c ↦ G c and its inverse are continuous. Thus it is a homeomorphism.
Since Ran T F \operatorname{Ran}T_F Ran T F is a hyperplane in R n + 1 \mathbb{R}^{n+1} R n + 1 , there exists a nonzero a ∈ R n + 1 a\in\mathbb{R}^{n+1} a ∈ R n + 1 such that
Ran T F = { c : ⟨ a , c ⟩ = 0 } . \operatorname{Ran}T_F=\{c:\langle a,c\rangle=0\}. Ran T F = { c : ⟨ a , c ⟩ = 0 } .
Consequently,
Δ ∘ ∖ Ran T F = { c ∈ Δ ∘ : ⟨ a , c ⟩ > 0 } ∪ { c ∈ Δ ∘ : ⟨ a , c ⟩ < 0 } . \Delta^\circ\setminus\operatorname{Ran}T_F=\{c\in\Delta^\circ:\langle a,c\rangle>0\}\cup\{c\in\Delta^\circ:\langle a,c\rangle<0\}. Δ ∘ ∖ Ran T F = { c ∈ Δ ∘ : ⟨ a , c ⟩ > 0 } ∪ { c ∈ Δ ∘ : ⟨ a , c ⟩ < 0 } .
Each nonempty set on the right is convex and hence contractible. Therefore, D + ( F ) \mathcal{D}_+(F) D + ( F ) has at most two connected components, and every component is contractible. □ \square □
We now describe the collection of all semi-positive dual frames of a given frame. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for H n \mathcal{H}_n H n , where N > n N>n N > n , and let P F = T F S F − 1 T F ∗ P_F=T_FS_F^{-1}T_F^* P F = T F S F − 1 T F ∗ denote the orthogonal projection onto Ran T F \operatorname{Ran}T_F Ran T F . Consider the set
C F = { c ∈ R ≥ 0 N ∖ { 0 } : c i = 0 for some i , c ∉ Ran T F } . C_F=\{c\in\mathbb{R}_{\geq0}^{N}\setminus\{0\}:c_i=0\text{ for some }i,\ c\notin\operatorname{Ran}T_F\}. C F = { c ∈ R ≥ 0 N ∖ { 0 } : c i = 0 for some i , c ∈ / Ran T F } .
For each c ∈ C F c\in C_F c ∈ C F , set z c = ( I − P F ) c z_c=(I-P_F)c z c = ( I − P F ) c and u c = S F − 1 T F ∗ c u_c=S_F^{-1}T_F^*c u c = S F − 1 T F ∗ c , and define L c : R N → H n L_c:\mathbb{R}^{N}\to\mathcal{H}_n L c : R N → H n by
L c v = S F − 1 T F ∗ v − ⟨ v , z c ⟩ ∥ z c ∥ 2 u c , v ∈ R N . L_cv=S_F^{-1}T_F^*v-\frac{\langle v,z_c\rangle}{\lVert z_c\rVert^2}u_c,\qquad v\in\mathbb{R}^{N}. L c v = S F − 1 T F ∗ v − ∥ z c ∥ 2 ⟨ v , z c ⟩ u c , v ∈ R N .
Since z c ≠ 0 z_c\ne0 z c = 0 , this operator is well-defined. Moreover, a direct calculation shows that L c T F = I H n L_cT_F=I_{\mathcal{H}_n} L c T F = I H n and L c c = 0 L_cc=0 L c c = 0 . Consequently, every operator of the form L c + A L_c+A L c + A , where A T F = 0 AT_F=0 A T F = 0 and A c = 0 Ac=0 A c = 0 , is the synthesis operator of a semi-positive dual frame of F F F . Conversely, if G G G is any semi-positive dual frame of F F F , then there exists c ∈ C F c\in C_F c ∈ C F such that T G ∗ c = 0 T_G^*c=0 T G ∗ c = 0 . In this case, A = T G ∗ − L c A=T_G^*-L_c A = T G ∗ − L c satisfies A T F = 0 AT_F=0 A T F = 0 and A c = 0 Ac=0 A c = 0 .
Thus, the collection of all semi-positive dual frames of F F F admits the representation
D s p ( F ) = ⋃ c ∈ C F { G : T G ∗ = L c + A , A T F = 0 , A c = 0 } . \mathcal{D}_{\mathrm{sp}}(F)=\bigcup_{c\in C_F}\{G:T_G^*=L_c+A,\quad AT_F=0,\quad Ac=0\}. D sp ( F ) = c ∈ C F ⋃ { G : T G ∗ = L c + A , A T F = 0 , A c = 0 } .
We next investigate the extreme-point structure of the collection of positive dual frames. The following result shows that no positive dual frame can be an extreme point of this collection.
Proposition 5.5. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for H n \mathcal{H}_n H n , where N > n N>n N > n . Then every positive dual frame of F F F can be expressed as a nontrivial convex combination of two distinct positive dual frames.
Proof. Let G = { g i } i = 1 N ∈ D + ( F ) G=\{g_i\}_{i=1}^{N}\in\mathcal{D}_{+}(F) G = { g i } i = 1 N ∈ D + ( F ) . Since N > n N>n N > n , there exists a nonzero vector d = ( d i ) i = 1 N ∈ ker T F ∗ d=(d_i)_{i=1}^{N}\in\ker T_F^* d = ( d i ) i = 1 N ∈ ker T F ∗ . Choose a nonzero vector u ∈ H n u\in\mathcal{H}_n u ∈ H n and define H = { h i } i = 1 N H=\{h_i\}_{i=1}^{N} H = { h i } i = 1 N by h i = d i u h_i=d_i u h i = d i u .
For each t ∈ R t\in\mathbb{R} t ∈ R , consider
G t = { g i + t h i } i = 1 N . G_t=\{g_i+th_i\}_{i=1}^{N}. G t = { g i + t h i } i = 1 N . Since
T H ∗ T F x = ∑ i = 1 N ⟨ x , f i ⟩ h i = ⟨ x , ∑ i = 1 N d i f i ⟩ u = 0 , T_H^*T_Fx=\sum_{i=1}^{N}\langle x,f_i\rangle h_i=\left\langle x,\sum_{i=1}^{N}d_i f_i\right\rangle u=0, T H ∗ T F x = i = 1 ∑ N ⟨ x , f i ⟩ h i = ⟨ x , i = 1 ∑ N d i f i ⟩ u = 0 , we obtain
T G t ∗ T F = T G ∗ T F + t T H ∗ T F = I H n . T_{G_t}^*T_F=T_G^*T_F+tT_H^*T_F=I_{\mathcal{H}_n}. T G t ∗ T F = T G ∗ T F + t T H ∗ T F = I H n . Thus G t G_t G t is a dual frame of F F F for every t ∈ R t\in\mathbb{R} t ∈ R .
Since G G G is positive, there exists c = ( c i ) i = 1 N ∈ R + + N c=(c_i)_{i=1}^{N}\in\mathbb{R}_{++}^{N} c = ( c i ) i = 1 N ∈ R ++ N such that
T G ∗ c = ∑ i = 1 N c i g i = 0. T_G^*c=\sum_{i=1}^{N}c_i g_i=0. T G ∗ c = i = 1 ∑ N c i g i = 0. Let P G P_G P G denote the orthogonal projection onto Ran T G \operatorname{Ran}T_G Ran T G . For sufficiently small ∣ t ∣ |t| ∣ t ∣ , the operators T G t ∗ T_{G_t}^* T G t ∗ have full rank, and their orthogonal kernel projections I − P G t I-P_{G_t} I − P G t converge to I − P G I-P_G I − P G . Consequently,
c t = ( I − P G t ) c ⟶ ( I − P G ) c = c . c_t=(I-P_{G_t})c\longrightarrow(I-P_G)c=c. c t = ( I − P G t ) c ⟶ ( I − P G ) c = c . Since c c c is strictly positive, c t ∈ R + + N c_t\in\mathbb{R}_{++}^{N} c t ∈ R ++ N for all sufficiently small ∣ t ∣ |t| ∣ t ∣ . Moreover,
T G t ∗ c t = 0. T_{G_t}^*c_t=0. T G t ∗ c t = 0. Therefore, G t G_t G t is a positive dual frame of F F F whenever ∣ t ∣ |t| ∣ t ∣ is sufficiently small. Choose t > 0 t>0 t > 0 sufficiently small. Since H ≠ 0 H\ne0 H = 0 , the positive dual frames G t G_t G t and G − t G_{-t} G − t are distinct, and
G = 1 2 G t + 1 2 G − t . G=\frac{1}{2}G_t+\frac{1}{2}G_{-t}. G = 2 1 G t + 2 1 G − t . Hence G G G is not an extreme point of D + ( F ) \mathcal{D}_{+}(F) D + ( F ) . Since G G G was arbitrary, the proof is complete.
We next establish a geometric property of the collection of positive dual frames. In particular, we show that every positive dual frame of a positive frame can be represented as a convex combination of two relative boundary dual frames.
Lemma 5.6. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a positive frame for H n \mathcal{H}_n H n , and let G = { g i } i = 1 N ∈ D + ( F ) G=\{g_i\}_{i=1}^{N}\in\mathcal{D}_{+}(F) G = { g i } i = 1 N ∈ D + ( F ) . Choose d = ( d i ) i = 1 N ∈ ker T F ∗ d=(d_i)_{i=1}^{N}\in\ker T_F^* d = ( d i ) i = 1 N ∈ ker T F ∗ with d i > 0 d_i>0 d i > 0 for every i i i , and fix u ∈ H n ∖ { 0 } u\in\mathcal{H}_n\setminus\{0\} u ∈ H n ∖ { 0 } . Set H = { d i u } i = 1 N H=\{d_i u\}_{i=1}^{N} H = { d i u } i = 1 N and
A = { s ∈ R : G + s H ∈ D + ( F ) } . A=\{s\in\mathbb{R}:G+sH\in\mathcal{D}_{+}(F)\}. A = { s ∈ R : G + sH ∈ D + ( F )} . Then A A A is a bounded open subset of R \mathbb{R} R containing 0 0 0 .
Proof. Since d ∈ ker T F ∗ d \in\ker T_F^* d ∈ ker T F ∗ , we have
T H ∗ T F x = ∑ i = 1 N ⟨ x , f i ⟩ d i u = ⟨ x , ∑ i = 1 N d i f i ⟩ u = 0 T_H^*T_Fx=\sum_{i=1}^{N}\langle x,f_i\rangle d_i u=\left\langle x,\sum_{i=1}^{N}d_i f_i\right\rangle u=0 T H ∗ T F x = i = 1 ∑ N ⟨ x , f i ⟩ d i u = ⟨ x , i = 1 ∑ N d i f i ⟩ u = 0
for every x ∈ H n x \in\mathcal{H}_n x ∈ H n . Consequently,
T G + s H ∗ T F = T G ∗ T F + s T H ∗ T F = I H n , T_{G+sH}^*T_F=T_G^*T_F+sT_H^*T_F=I_{\mathcal{H}_n}, T G + sH ∗ T F = T G ∗ T F + s T H ∗ T F = I H n ,
so G + s H G+sH G + sH is a dual frame of F F F for every s ∈ R s \in\mathbb{R} s ∈ R .
To establish boundedness, let x = u / ∥ u ∥ x=u/\|u\| x = u /∥ u ∥ and define
α i = ⟨ x , g i ⟩ d i ∥ u ∥ , m = min 1 ≤ i ≤ N α i , M = max 1 ≤ i ≤ N α i . \alpha_i=\frac{\langle x,g_i\rangle}{d_i\|u\|},\qquad m=\min_{1\leq i\leq N}\alpha_i,\qquad M=\max_{1\leq i\leq N}\alpha_i. α i = d i ∥ u ∥ ⟨ x , g i ⟩ , m = 1 ≤ i ≤ N min α i , M = 1 ≤ i ≤ N max α i .
For every i i i and s ∈ R s \in\mathbb{R} s ∈ R ,
⟨ x , g i + s d i u ⟩ = d i ∥ u ∥ ( α i + s ) . \langle x,g_i+sd_i u\rangle=d_i\|u\|(\alpha_i+s). ⟨ x , g i + s d i u ⟩ = d i ∥ u ∥ ( α i + s ) .
If s ≤ − M s\leq-M s ≤ − M , then ⟨ x , g i + s d i u ⟩ ≤ 0 \langle x,g_i+sd_i u\rangle\leq0 ⟨ x , g i + s d i u ⟩ ≤ 0 for every i i i . By the positive-spanning criterion, G + s H G+sH G + sH is not positive.
Similarly, if s ≥ − m s\geq-m s ≥ − m , then ⟨ − x , g i + s d i u ⟩ ≤ 0 \langle-x,g_i+sd_i u\rangle\leq0 ⟨ − x , g i + s d i u ⟩ ≤ 0 for every i i i , and hence G + s H G+sH G + sH is not positive. Therefore,
A ⊆ ( − M , − m ) , A\subseteq(-M,-m), A ⊆ ( − M , − m ) ,
which proves that A A A is bounded.
Since positive spanning is an open condition on finite families of vectors, A A A is open. Finally, 0 ∈ A 0\in A 0 ∈ A because G ∈ D + ( F ) G\in\mathcal{D}_+(F) G ∈ D + ( F ) .
The preceding lemma allows us to represent every positive dual frame as a convex combination of two relative boundary dual frames.
Theorem 5.7. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a positive frame for H n \mathcal{H}_n H n . Then every positive dual frame of F F F can be expressed as a convex combination of two distinct dual frames belonging to the relative boundary of D + ( F ) \mathcal{D}_+(F) D + ( F ) in D ( F ) \mathcal{D}(F) D ( F ) .
Proof. Let G ∈ D + ( F ) G\in\mathcal{D}_+(F) G ∈ D + ( F ) . Since F F F is positive, there exists d = ( d i ) i = 1 N ∈ ker T F ∗ d=(d_i)_{i=1}^{N}\in\ker T_F^* d = ( d i ) i = 1 N ∈ ker T F ∗ with d i > 0 d_i>0 d i > 0 for every i i i . Fix u ∈ H n ∖ { 0 } u\in\mathcal{H}_n\setminus\{0\} u ∈ H n ∖ { 0 } and set H = { d i u } i = 1 N H=\{d_i u\}_{i=1}^{N} H = { d i u } i = 1 N . Consider
A = { s ∈ R : G + s H ∈ D + ( F ) } . A=\{s\in\mathbb{R}:G+sH\in\mathcal{D}_+(F)\}. A = { s ∈ R : G + sH ∈ D + ( F )} . By Lemma 5.6, A A A is bounded, open, and contains 0 0 0 .
Let ( a , b ) (a,b) ( a , b ) be the connected component of A A A containing 0 0 0 . Since A A A is bounded and open, we have
− ∞ < a < 0 < b < ∞ . -\infty<a<0<b<\infty. − ∞ < a < 0 < b < ∞. Moreover, G + s H ∈ D + ( F ) G+sH\in\mathcal{D}_+(F) G + sH ∈ D + ( F ) for every s ∈ ( a , b ) s\in(a,b) s ∈ ( a , b ) .
Define
G 1 = G + a H , G 2 = G + b H . G_1=G+aH,\qquad G_2=G+bH. G 1 = G + a H , G 2 = G + b H . Since G + s H G+sH G + sH is a dual frame of F F F for every s ∈ R s\in\mathbb{R} s ∈ R , both G 1 G_1 G 1 and G 2 G_2 G 2 belong to D ( F ) \mathcal{D}(F) D ( F ) .
By the maximality of ( a , b ) (a,b) ( a , b ) and the openness of A A A , neither a a a nor b b b belongs to A A A . Thus G 1 , G 2 ∉ D + ( F ) G_1,G_2 \notin D_+(F) G 1 , G 2 ∈ / D + ( F ) . On the other hand,
G + s H ⟶ G 1 as s → a + , G+sH \longrightarrow G_1 \quad\text{as } s \to a^+, G + sH ⟶ G 1 as s → a + ,
and
G + s H ⟶ G 2 as s → b − . G+sH \longrightarrow G_2 \quad\text{as } s \to b^-. G + sH ⟶ G 2 as s → b − .
Consequently,
G 1 , G 2 ∈ ∂ D ( F ) D + ( F ) . G_1,G_2 \in\partial_{D(F)}D_+(F). G 1 , G 2 ∈ ∂ D ( F ) D + ( F ) .
Finally, since a < 0 < b a<0<b a < 0 < b , we have
G = b b − a G 1 + − a b − a G 2 . G=\frac{b}{b-a}G_1+\frac{-a}{b-a}G_2. G = b − a b G 1 + b − a − a G 2 .
Both coefficients are strictly positive and their sum is one. Moreover, G 1 ≠ G 2 G_1 \ne G_2 G 1 = G 2 , since a ≠ b a \ne b a = b and H ≠ 0 H \ne0 H = 0 . Therefore, G G G is a strict convex combination of two distinct relative boundary dual frames.
Remark 5.8. The converse of Theorem 5.7 does not hold in general. In particular, a convex combination of two relative boundary dual frames need not be positive.
To see this, consider the positive frame F = { e 1 , − e 1 , e 2 , − e 2 } F=\{e_1,-e_1,e_2,-e_2\} F = { e 1 , − e 1 , e 2 , − e 2 } for R 2 \mathbb{R}^2 R 2 , where { e 1 , e 2 } \{e_1,e_2\} { e 1 , e 2 } is the standard orthonormal basis. Let
G 1 = { e 1 , 0 , e 2 , 0 } , G 2 = { 0 , − e 1 , e 2 , 0 } . G_1=\{e_1,0,e_2,0\}, \qquad G_2=\{0,-e_1,e_2,0\}. G 1 = { e 1 , 0 , e 2 , 0 } , G 2 = { 0 , − e 1 , e 2 , 0 } .
Clearly, both G 1 G_1 G 1 and G 2 G_2 G 2 are dual frames of F F F , but neither is positive. Moreover, both belong to the relative boundary of D + ( F ) D_+(F) D + ( F ) in D ( F ) D(F) D ( F ) , as they can be approximated by positive dual frames.
For every t ∈ ( 0 , 1 ) t\in(0,1) t ∈ ( 0 , 1 ) , we have
t G 1 + ( 1 − t ) G 2 = { t e 1 , − ( 1 − t ) e 1 , e 2 , 0 } . tG_1+(1-t)G_2=\{te_1,-(1-t)e_1,e_2,0\}. t G 1 + ( 1 − t ) G 2 = { t e 1 , − ( 1 − t ) e 1 , e 2 , 0 } .
Since all vectors in this family have nonnegative second coordinates, the family does not positively span R 2 \mathbb{R}^2 R 2 . Consequently,
t G 1 + ( 1 − t ) G 2 ∉ D + ( F ) , 0 < t < 1. tG_1+(1-t)G_2\notin D_+(F), \qquad0<t<1. t G 1 + ( 1 − t ) G 2 ∈ / D + ( F ) , 0 < t < 1.
Thus, although every positive dual frame of a positive frame is a convex combination of two suitably chosen relative boundary dual frames, the same conclusion does not hold for arbitrary pairs of boundary duals.
Example 5.9. Let { e 1 , e 2 } \{e_1,e_2\} { e 1 , e 2 } be the standard orthonormal basis of R 2 \mathbb{R}^2 R 2 , and consider the frame
F = { e 1 , e 2 , e 1 } . F=\{e_1,e_2,e_1\}. F = { e 1 , e 2 , e 1 } . Clearly, F F F is not a positive frame. Every dual frame of F F F can be written as
G ( u ) = { e 1 − u , e 2 , u } , u = ( a , b ) ∈ R 2 . G(u)=\{e_1-u,e_2,u\}, \qquad u=(a,b)\in\mathbb{R}^2. G ( u ) = { e 1 − u , e 2 , u } , u = ( a , b ) ∈ R 2 . Indeed, the duality condition is equivalent to g 1 + g 3 = e 1 g_1+g_3=e_1 g 1 + g 3 = e 1 and g 2 = e 2 g_2=e_2 g 2 = e 2 .
We claim that G ( u ) G(u) G ( u ) is a positive dual if and only if
( a < 0 , b < 0 ) or ( a > 1 , b > 0 ) . (a < 0, b < 0) \quad\text{or} \quad(a > 1, b > 0). ( a < 0 , b < 0 ) or ( a > 1 , b > 0 ) .
To verify this, observe that a strictly positive relation among the vectors of G ( u ) G(u) G ( u ) is equivalent to
α ( e 1 − u ) + β e 2 + γ u = 0 , α , β , γ > 0. \alpha(e_1-u)+\beta e_2+\gamma u=0,\qquad\alpha,\beta,\gamma>0. α ( e 1 − u ) + β e 2 + γ u = 0 , α , β , γ > 0.
Comparing coordinates yields
α + ( γ − α ) a = 0 , β + ( γ − α ) b = 0. \alpha+(\gamma-\alpha)a=0,\qquad\beta+(\gamma-\alpha)b=0. α + ( γ − α ) a = 0 , β + ( γ − α ) b = 0.
These equations admit strictly positive solutions precisely in the two regions described above. Now suppose that G ( u ) G(u) G ( u ) is positive. If a < 0 a<0 a < 0 and b < 0 b<0 b < 0 , set
u 1 = ( 2 a , 0 ) , u 2 = ( 0 , 2 b ) . u_1=(2a,0),\qquad u_2=(0,2b). u 1 = ( 2 a , 0 ) , u 2 = ( 0 , 2 b ) .
Then
G ( u ) = 1 2 G ( u 1 ) + 1 2 G ( u 2 ) . G(u)=\frac{1}{2}G(u_1)+\frac{1}{2}G(u_2). G ( u ) = 2 1 G ( u 1 ) + 2 1 G ( u 2 ) .
Both G ( u 1 ) G(u_1) G ( u 1 ) and G ( u 2 ) G(u_2) G ( u 2 ) belong to the relative boundary of D + ( F ) \mathcal{D}_{+}(F) D + ( F ) in D ( F ) \mathcal{D}(F) D ( F ) . Similarly, if a > 1 a>1 a > 1 and b > 0 b>0 b > 0 , set
u 1 = ( 1 , 2 b ) , u 2 = ( 2 a − 1 , 0 ) . u_1=(1,2b),\qquad u_2=(2a-1,0). u 1 = ( 1 , 2 b ) , u 2 = ( 2 a − 1 , 0 ) .
Again,
G ( u ) = 1 2 G ( u 1 ) + 1 2 G ( u 2 ) , G(u)=\frac{1}{2}G(u_1)+\frac{1}{2}G(u_2), G ( u ) = 2 1 G ( u 1 ) + 2 1 G ( u 2 ) ,
where G ( u 1 ) G(u_1) G ( u 1 ) and G ( u 2 ) G(u_2) G ( u 2 ) are distinct boundary dual frames. Thus every positive dual frame of F F F is a convex combination of two distinct dual frames belonging to the relative boundary of D + ( F ) \mathcal{D}_{+}(F) D + ( F ) , even though F F F itself is not positive.
The boundary behavior of positive dual frames raises a natural question concerning the preservation of positivity under convex combinations. Although convex combinations of dual frames preserve the reconstruction identity, they need not preserve positive spanning. The following example illustrates this phenomenon by showing that the midpoint of two boundary points of the set of positive duals may fail to be positive.
Example 5.10. Consider the frame F = { ( 1 , 0 ) , ( 1 , 1 ) , ( 0 , 1 ) } F=\{(1,0),(1,1),(0,1)\} F = {( 1 , 0 ) , ( 1 , 1 ) , ( 0 , 1 )} for R 2 \mathbb{R}^{2} R 2 . For 0 ≤ t < 1 0\leq t<1 0 ≤ t < 1 , define G t = { g 1 ( t ) , g 2 ( t ) , g 3 ( t ) } G_t=\{g_1^{(t)},g_2^{(t)},g_3^{(t)}\} G t = { g 1 ( t ) , g 2 ( t ) , g 3 ( t ) } , where
g 1 ( t ) = 1 2 − t ( 1 − t , − 1 ) , g_1^{(t)}=\frac{1}{2-t}(1-t,-1), g 1 ( t ) = 2 − t 1 ( 1 − t , − 1 ) , g 2 ( t ) = 1 2 − t ( 1 , 1 ) , g_2^{(t)}=\frac{1}{2-t}(1,1), g 2 ( t ) = 2 − t 1 ( 1 , 1 ) , g 3 ( t ) = 1 2 − t ( − 1 , 1 − t ) . g_3^{(t)}=\frac{1}{2-t}(-1,1-t). g 3 ( t ) = 2 − t 1 ( − 1 , 1 − t ) . It is straightforward to verify that G t G_t G t is a dual frame of F F F and
g 1 ( t ) + t g 2 ( t ) + g 3 ( t ) = 0. g_1^{(t)}+tg_2^{(t)}+g_3^{(t)}=0. g 1 ( t ) + t g 2 ( t ) + g 3 ( t ) = 0.
Thus G t G_t G t is positive for every 0 < t < 1 0 < t < 1 0 < t < 1 . However, G t G_t G t converges as t → 0 + t \to0^+ t → 0 + to the semi-positive dual frame
G 0 = { ( 1 2 , − 1 2 ) , ( 1 2 , 1 2 ) , ( − 1 2 , 1 2 ) } , G_0 = \left\{\left(\frac{1}{2},-\frac{1}{2}\right),\left(\frac{1}{2},\frac{1}{2}\right),\left(-\frac{1}{2},\frac{1}{2}\right)\right\}, G 0 = { ( 2 1 , − 2 1 ) , ( 2 1 , 2 1 ) , ( − 2 1 , 2 1 ) } ,
which is not positive since ker T G 0 ∗ = span { ( 1 , 0 , 1 ) } \ker T_{G_0}^* = \operatorname{span}\{(1,0,1)\} ker T G 0 ∗ = span {( 1 , 0 , 1 )} .
Now consider another dual frame
H = { ( 0 , 0 ) , ( 1 , 0 ) , ( − 1 , 1 ) } . H = \{(0,0),(1,0),(-1,1)\}. H = {( 0 , 0 ) , ( 1 , 0 ) , ( − 1 , 1 )} .
Since ker T H ∗ = span { ( 1 , 0 , 0 ) } \ker T_H^* = \operatorname{span}\{(1,0,0)\} ker T H ∗ = span {( 1 , 0 , 0 )} , the frame H H H is also semi-positive but not positive. Both G 0 G_0 G 0 and H H H belong to the relative boundary of D + ( F ) \mathcal{D}_+(\mathcal{F}) D + ( F ) . Their midpoint M = 1 2 ( G 0 + H ) M = \frac{1}{2}(G_0 + H) M = 2 1 ( G 0 + H ) satisfies
M = { ( 1 4 , − 1 4 ) , ( 3 4 , 1 4 ) , ( − 3 4 , 3 4 ) } . M = \left\{\left(\frac{1}{4},-\frac{1}{4}\right),\left(\frac{3}{4},\frac{1}{4}\right),\left(-\frac{3}{4},\frac{3}{4}\right)\right\}. M = { ( 4 1 , − 4 1 ) , ( 4 3 , 4 1 ) , ( − 4 3 , 4 3 ) } .
A direct computation gives
ker T M ∗ = span { ( 3 , 0 , 1 ) } . \ker T_M^* = \operatorname{span}\{(3,0,1)\}. ker T M ∗ = span {( 3 , 0 , 1 )} .
Hence M M M is not positive. This shows that positive dual frames may converge to non-positive boundary points and that convex combinations of such boundary points need not be positive.
Positive Frames Arising from Finite Group Actions
We now investigate positive frames arising from finite group representations, with particular emphasis on the relationship between linear connectedness and balancedness. The symmetry induced by a group action provides additional information about the linear relations among the frame vectors. In particular, the relation space of a group-generated frame is invariant under the natural permutation action of the group. This observation allows us to establish a connection between linear connectedness and the balancedness of frames with excess one.
Group Averaging, Balancedness, and Positivity
We investigate the relationship between balancedness and positivity for frames generated by finite group representations. Let G G G be a finite group and let π : G → O ( R n ) \pi: G \to O(\mathbb{R}^n) π : G → O ( R n ) be an orthogonal representation. Suppose that π \pi π admits a decomposition into pairwise inequivalent irreducible representations of the form
π ≃ π 1 ( m 1 ) ⊕ ⋯ ⊕ π k ( m k ) , π i ≄ π j ( i ≠ j ) . \pi\simeq\pi_1^{(m_1)} \oplus\cdots\oplus\pi_k^{(m_k)}, \qquad\pi_i \not\simeq\pi_j \quad(i \ne j). π ≃ π 1 ( m 1 ) ⊕ ⋯ ⊕ π k ( m k ) , π i ≃ π j ( i = j ) .
For each ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n , consider the group-generated family F ξ = { π ( g ) ξ : g ∈ G } \mathcal{F}_{\xi} = \{\pi(g)\xi: g \in G\} F ξ = { π ( g ) ξ : g ∈ G } , indexed by the elements of G G G . Recall that F ξ \mathcal{F}_{\xi} F ξ is balanced if ∑ g ∈ G π ( g ) ξ = 0 \sum_{g\in G}\pi(g)\xi= 0 ∑ g ∈ G π ( g ) ξ = 0 . The averaging operator associated with π \pi π is closely related to the subspace of vectors fixed by the group action. In particular, the absence of the trivial representation determines when every orbit family is balanced. Moreover, whenever such an orbit family spans R n \mathbb{R}^n R n , its balancedness implies positivity. The following proposition establishes these equivalences.
Proposition 6.1. Under the preceding assumptions, the following statements are equivalent:
(i) The trivial representation does not occur among π 1 , … , π k \pi_1,\ldots,\pi_k π 1 , … , π k .
(ii) ∑ g ∈ G π ( g ) = 0 \sum_{g\in G}\pi(g)=0 ∑ g ∈ G π ( g ) = 0 .
(iii) F ξ F_\xi F ξ is balanced for every ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n .
(iv) F ξ F_\xi F ξ is a positive frame whenever it spans R n \mathbb{R}^n R n .
Proof. Define the averaging operator
P = 1 ∣ G ∣ ∑ g ∈ G π ( g ) . P=\frac{1}{|G|}\sum_{g\in G}\pi(g). P = ∣ G ∣ 1 g ∈ G ∑ π ( g ) .
Let
I ( π ) = { x ∈ R n : π ( g ) x = x for all g ∈ G } \mathcal{I}(\pi)=\{x\in\mathbb{R}^n:\pi(g)x=x\text{ for all }g\in G\} I ( π ) = { x ∈ R n : π ( g ) x = x for all g ∈ G }
denote the subspace of invariant vectors. For every h ∈ G h\in G h ∈ G , we have
π ( h ) P = 1 ∣ G ∣ ∑ g ∈ G π ( h g ) = P . \pi(h)P=\frac{1}{|G|}\sum_{g\in G}\pi(hg)=P. π ( h ) P = ∣ G ∣ 1 g ∈ G ∑ π ( h g ) = P .
Consequently, ran P ⊆ I ( π ) \operatorname{ran}P\subseteq\mathcal{I}(\pi) ran P ⊆ I ( π ) . Conversely, if x ∈ I ( π ) x\in\mathcal{I}(\pi) x ∈ I ( π ) , then
P x = 1 ∣ G ∣ ∑ g ∈ G π ( g ) x = x . Px=\frac{1}{|G|}\sum_{g\in G}\pi(g)x=x. P x = ∣ G ∣ 1 g ∈ G ∑ π ( g ) x = x .
Thus I ( π ) ⊆ ran P \mathcal{I}(\pi)\subseteq\operatorname{ran}P I ( π ) ⊆ ran P , and hence
ran P = I ( π ) . \operatorname{ran}P=\mathcal{I}(\pi). ran P = I ( π ) .
Since π \pi π is orthogonal, P P P is the orthogonal projection onto I ( π ) \mathcal{I}(\pi) I ( π ) .
We first prove (i) ⇔ \Leftrightarrow ⇔ (ii). The trivial representation occurs in π \pi π if and only if I ( π ) ≠ { 0 } \mathcal{I}(\pi)\ne\{0\} I ( π ) = { 0 } . Therefore,
(i) ⟺ I ( π ) = { 0 } ⟺ P = 0 ⟺ ∑ g ∈ G π ( g ) = 0 , \begin{aligned}
\text{(i)} &\Longleftrightarrow\mathcal{I}(\pi)=\{0\}\\
&\Longleftrightarrow P=0\\
&\Longleftrightarrow\sum_{g\in G}\pi(g)=0,
\end{aligned} (i) ⟺ I ( π ) = { 0 } ⟺ P = 0 ⟺ g ∈ G ∑ π ( g ) = 0 ,
which proves the equivalence.
Next, (ii) ⇔ \Leftrightarrow ⇔ (iii) follows immediately from
∑ g ∈ G π ( g ) ξ = ( ∑ g ∈ G π ( g ) ) ξ . \sum_{g\in G}\pi(g)\xi=\left(\sum_{g\in G}\pi(g)\right)\xi. g ∈ G ∑ π ( g ) ξ = g ∈ G ∑ π ( g ) ξ .
Indeed, the right-hand side vanishes for every ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n if and only if ∑ g ∈ G π ( g ) = 0 \sum_{g\in G}\pi(g)=0 ∑ g ∈ G π ( g ) = 0 .
We now prove (iii) ⇒ \Rightarrow ⇒ (iv). Let ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n be such that F ξ F_\xi F ξ spans R n \mathbb{R}^n R n . By (iii),
∑ g ∈ G π ( g ) ξ = 0. \sum_{g\in G}\pi(g)\xi=0. g ∈ G ∑ π ( g ) ξ = 0.
Thus F ξ F_\xi F ξ admits a linear dependence with all coefficients strictly positive. Since F ξ F_\xi F ξ spans R n \mathbb{R}^n R n , it follows that F ξ F_\xi F ξ positively spans R n \mathbb{R}^n R n . Hence F ξ F_\xi F ξ is a positive frame.
Finally, suppose (iv) holds. We show that I ( π ) = { 0 } \mathcal{I}(\pi)=\{0\} I ( π ) = { 0 } . Assume, to the contrary, that I ( π ) ≠ { 0 } \mathcal{I}(\pi)\ne\{0\} I ( π ) = { 0 } . Then P ≠ 0 P\ne0 P = 0 . If there exists a frame vector ξ \xi ξ , then F ξ F_\xi F ξ is positive by (iv). Consequently, there exist coefficients c g > 0 c_g > 0 c g > 0 , g ∈ G g \in G g ∈ G , such that
∑ g ∈ G c g π ( g ) ξ = 0. \sum_{g\in G} c_g\pi(g)\xi= 0. g ∈ G ∑ c g π ( g ) ξ = 0.
Applying P P P and using P π ( g ) = P P\pi(g) = P P π ( g ) = P for every g ∈ G g \in G g ∈ G , we obtain
0 = ∑ g ∈ G c g P π ( g ) ξ = ( ∑ g ∈ G c g ) P ξ . 0 = \sum_{g\in G} c_gP\pi(g)\xi= \left(\sum_{g\in G}c_g\right)P\xi. 0 = g ∈ G ∑ c g P π ( g ) ξ = g ∈ G ∑ c g P ξ .
Since ∑ g ∈ G c g > 0 \sum_{g\in G} c_g > 0 ∑ g ∈ G c g > 0 , it follows that P ξ = 0 P\xi= 0 P ξ = 0 . However, P P P commutes with every π ( g ) \pi(g) π ( g ) , so
P π ( g ) ξ = π ( g ) P ξ = 0 ( g ∈ G ) . P\pi(g)\xi= \pi(g)P\xi= 0 \qquad(g \in G). P π ( g ) ξ = π ( g ) P ξ = 0 ( g ∈ G ) .
Since F ξ F_\xi F ξ spans R n \mathbb{R}^n R n , we conclude that P = 0 P = 0 P = 0 , a contradiction. Therefore I ( π ) = { 0 } I(\pi) = \{0\} I ( π ) = { 0 } , which proves (i) and completes the proof.
Linear Connectedness and Excess-One Orbit Frames
We now investigate the relationship between linear connectedness and balancedness for frames generated by finite group representations. The notion of linear connectedness was studied by Pehlivan, Han, and Mohapatra [21 ] in connection with spectrally optimal dual frames for erasures. We first recall the relevant definitions and a characterization from their work.
Let F = { f i } i = 1 N F = \{f_i\}_{i=1}^N F = { f i } i = 1 N be a finite family of nonzero vectors in H n \mathcal{H}_n H n . Two vectors f i f_i f i and f j f_j f j are said to be linearly F F F -connected if there exist vectors f k 1 , … , f k m f_{k_1}, \ldots, f_{k_m} f k 1 , … , f k m from F F F such that { f j , f k 1 , … , f k m } \{f_j, f_{k_1}, \ldots, f_{k_m}\} { f j , f k 1 , … , f k m } is linearly independent and
f i = c f j + ∑ ℓ = 1 m c ℓ f k ℓ , f_i = cf_j + \sum_{\ell=1}^{m}c_\ell f_{k_\ell}, f i = c f j + ℓ = 1 ∑ m c ℓ f k ℓ ,
where c , c 1 , … , c m c, c_1, \ldots, c_m c , c 1 , … , c m are all nonzero. The family F F F is called linearly connected if every two of its vectors are linearly F F F -connected.
For Λ ⊆ { 1 , … , N } \Lambda\subseteq\{1, \ldots, N\} Λ ⊆ { 1 , … , N } , write H Λ = span { f i : i ∈ Λ } \mathcal{H}_\Lambda= \operatorname{span}\{f_i : i \in\Lambda\} H Λ = span { f i : i ∈ Λ } . The family F F F is said to have the intersection dependent property if
H Λ ∩ H Λ c ≠ { 0 } \mathcal{H}_\Lambda\cap\mathcal{H}_{\Lambda^c} \ne\{0\} H Λ ∩ H Λ c = { 0 }
for every nonempty proper subset Λ ⊂ { 1 , … , N } \Lambda\subset\{1, \ldots, N\} Λ ⊂ { 1 , … , N } . Furthermore, F F F is called k k k -independent if every k k k -element subfamily is linearly independent. The following characterization was established in [21 Theorem 2.1] .
Theorem 6.2. [21 ] Let F = { f i } i = 1 N F = \{f_i\}_{i=1}^N F = { f i } i = 1 N be a finite family of nonzero vectors in H n \mathcal{H}_n H n , and let d = dim span ( F ) d = \dim\operatorname{span}(F) d = dim span ( F ) . Then the following are equivalent:
(i) F F F is linearly connected.
(ii) F F F has the intersection dependent property.
(iii) F F F contains a d d d -independent subfamily consisting of at least d + 1 d + 1 d + 1 vectors.
We now consider the special case of group-generated frames with excess one. In this setting, linear connectedness imposes a particularly strong restriction on the unique linear dependence among the orbit vectors. The following proposition shows that this dependence must be balanced.
Proposition 6.3. Let G G G be a finite group and let π : G → G L ( H n ) \pi: G \to GL(\mathcal{H}_n) π : G → G L ( H n ) be a representation of G G G on an n n n -dimensional Hilbert space H n \mathcal{H}_n H n . Suppose that F = { π ( g ) ξ : g ∈ G } F = \{\pi(g)\xi: g \in G\} F = { π ( g ) ξ : g ∈ G } is a linearly connected frame for H n \mathcal{H}_n H n with ∣ G ∣ = n + 1 |G| = n + 1 ∣ G ∣ = n + 1 . Then F F F is balanced, that is,
∑ g ∈ G π ( g ) ξ = 0. \sum_{g\in G}\pi(g)\xi=0. g ∈ G ∑ π ( g ) ξ = 0.
Proof. Since F F F is a frame for H n \mathcal{H}_n H n , we have rank ( T F ∗ ) = n \operatorname{rank}(T_F^*)=n rank ( T F ∗ ) = n . As ∣ G ∣ = n + 1 |G|=n+1 ∣ G ∣ = n + 1 , the rank-nullity theorem gives dim ker ( T F ∗ ) = 1 \dim\ker(T_F^*)=1 dim ker ( T F ∗ ) = 1 .
For each h ∈ G h\in G h ∈ G , define L h : C ∣ G ∣ → C ∣ G ∣ L_h:\mathbb{C}^{|G|}\to\mathbb{C}^{|G|} L h : C ∣ G ∣ → C ∣ G ∣ by
( L h c ) g = c h − 1 g , g ∈ G . (L_hc)_g=c_{h^{-1}g},\qquad g\in G. ( L h c ) g = c h − 1 g , g ∈ G .
Then, for every c = ( c g ) g ∈ G c=(c_g)_{g\in G} c = ( c g ) g ∈ G ,
T F ∗ ( L h c ) = ∑ g ∈ G ( L h c ) g π ( g ) ξ = ∑ g ∈ G c h − 1 g π ( g ) ξ = ∑ r ∈ G c r π ( h r ) ξ = π ( h ) ∑ r ∈ G c r π ( r ) ξ = π ( h ) T F ∗ c . \begin{aligned}
T_F^*(L_hc)&=\sum_{g\in G}(L_hc)_g\pi(g)\xi\\
&=\sum_{g\in G}c_{h^{-1}g}\pi(g)\xi\\
&=\sum_{r\in G}c_r\pi(hr)\xi\\
&=\pi(h)\sum_{r\in G}c_r\pi(r)\xi\\
&=\pi(h)T_F^*c.
\end{aligned} T F ∗ ( L h c ) = g ∈ G ∑ ( L h c ) g π ( g ) ξ = g ∈ G ∑ c h − 1 g π ( g ) ξ = r ∈ G ∑ c r π ( h r ) ξ = π ( h ) r ∈ G ∑ c r π ( r ) ξ = π ( h ) T F ∗ c .
Consequently,
T F ∗ L h = π ( h ) T F ∗ . T_F^*L_h=\pi(h)T_F^*. T F ∗ L h = π ( h ) T F ∗ .
If c ∈ ker ( T F ∗ ) c\in\ker(T_F^*) c ∈ ker ( T F ∗ ) , then
T F ∗ ( L h c ) = π ( h ) T F ∗ c = 0. T_F^*(L_hc)=\pi(h)T_F^*c=0. T F ∗ ( L h c ) = π ( h ) T F ∗ c = 0.
Thus,
L h ( ker T F ∗ ) ⊆ ker T F ∗ . L_h(\ker T_F^*)\subseteq\ker T_F^*. L h ( ker T F ∗ ) ⊆ ker T F ∗ .
Since L h L_h L h is invertible, we obtain
L h ( ker T F ∗ ) = ker T F ∗ . L_h(\ker T_F^*)=\ker T_F^*. L h ( ker T F ∗ ) = ker T F ∗ .
Now let 0 ≠ c = ( c g ) g ∈ G ∈ ker ( T F ∗ ) 0\ne c=(c_g)_{g\in G}\in\ker(T_F^*) 0 = c = ( c g ) g ∈ G ∈ ker ( T F ∗ ) . Since the kernel is one-dimensional, for every h ∈ G h\in G h ∈ G there exists a nonzero scalar α h \alpha_h α h such that
L h c = α h c . L_hc=\alpha_hc. L h c = α h c .
Equivalently,
c h − 1 g = α h c g , g , h ∈ G . c_{h^{-1}g}=\alpha_hc_g,\qquad g,h\in G. c h − 1 g = α h c g , g , h ∈ G .
Since F F F is linearly connected and dim ker ( T F ∗ ) = 1 \dim\ker(T_F^*)=1 dim ker ( T F ∗ ) = 1 , the coordinates of c c c are equal. Hence there exists a ≠ 0 a\ne0 a = 0 such that
where 1 = ( 1 , … , 1 ) T \mathbf{1}=(1,\ldots,1)^T 1 = ( 1 , … , 1 ) T . Therefore,
0 = T F ∗ c = a ∑ g ∈ G π ( g ) ξ . 0=T_F^*c=a\sum_{g\in G}\pi(g)\xi. 0 = T F ∗ c = a g ∈ G ∑ π ( g ) ξ .
Since a ≠ 0 a \ne0 a = 0 , it follows that
∑ g ∈ G π ( g ) ξ = 0. \sum_{g\in G}\pi(g)\xi=0. g ∈ G ∑ π ( g ) ξ = 0.
Hence F F F is balanced.
The group action imposes additional restrictions on the linear relations among orbit vectors. In the excess-one case, these relations are completely determined by a real character of the underlying group.
Proposition 6.4. Let G G G be a finite group and let π : G → G L ( H n ) \pi:G\to GL(\mathcal H_n) π : G → G L ( H n ) be a representation on a real Hilbert space. Suppose that F = { π ( g ) ξ : g ∈ G } F=\{\pi(g)\xi:g\in G\} F = { π ( g ) ξ : g ∈ G } is a frame for H n \mathcal H_n H n with ∣ G ∣ = n + 1 |G|=n+1 ∣ G ∣ = n + 1 . Then there exists a homomorphism χ : G → { − 1 , 1 } \chi:G\to\{-1,1\} χ : G → { − 1 , 1 } such that
ker T F ∗ = span { ( χ ( g ) ) g ∈ G } . \ker T_F^*=\operatorname{span}\{(\chi(g))_{g\in G}\}. ker T F ∗ = span {( χ ( g ) ) g ∈ G } . In particular, F F F is linearly connected.
Proof. Since ∣ G ∣ = n + 1 |G|=n+1 ∣ G ∣ = n + 1 , the synthesis kernel K = ker T F ∗ K=\ker T_F^* K = ker T F ∗ is one-dimensional. For each h ∈ G h\in G h ∈ G , let L h L_h L h denote the permutation operator on R G \mathbb R^G R G defined by L h e g = e h g L_he_g=e_{hg} L h e g = e h g . Then
T F ∗ L h = π ( h ) T F ∗ , T_F^*L_h=\pi(h)T_F^*, T F ∗ L h = π ( h ) T F ∗ ,
and consequently L h K = K L_hK=K L h K = K .
Choose a nonzero c ∈ K c\in K c ∈ K . Since K K K is one-dimensional, there exists a scalar χ ( h ) ≠ 0 \chi(h)\ne0 χ ( h ) = 0 such that L h c = χ ( h ) c L_hc=\chi(h)c L h c = χ ( h ) c . The identity L h k = L h L k L_{hk}=L_hL_k L hk = L h L k implies that χ \chi χ is a homomorphism. Since G G G is finite and χ \chi χ is real-valued, χ ( h ) ∈ { − 1 , 1 } \chi(h)\in\{-1,1\} χ ( h ) ∈ { − 1 , 1 } for every h ∈ G h\in G h ∈ G .
Moreover, the transitivity of the permutation action implies that no coordinate of c c c can vanish. Using the coordinate corresponding to the identity element, we obtain c g = c e χ ( g − 1 ) = c e χ ( g ) c_g=c_e\chi(g^{-1})=c_e\chi(g) c g = c e χ ( g − 1 ) = c e χ ( g ) . Thus
K = span { ( χ ( g ) ) g ∈ G } . K=\operatorname{span}\{(\chi(g))_{g\in G}\}. K = span {( χ ( g ) ) g ∈ G } .
Since the unique linear dependence has no vanishing coefficient, every n n n -element subfamily of F F F is linearly independent. Hence F F F is linearly connected by [21 Theorem 2.1] .
The preceding proposition yields a characterization of balancedness and positive spanning for excess-one orbit frames.
Corollary 6.5. Under the hypotheses of Proposition 6.4, the following statements are equivalent:
(i) F F F is balanced.
(ii) F F F is a positive frame.
(iii) The character χ \chi χ is trivial.
(iv) H n \mathcal H_n H n contains no nonzero vector fixed by every π ( g ) \pi(g) π ( g ) , g ∈ G g\in G g ∈ G .
Proof. By Proposition 6.4, the synthesis kernel is generated by ( χ ( g ) ) g ∈ G (\chi(g))_{g\in G} ( χ ( g ) ) g ∈ G . Therefore, the all-ones vector belongs to ker T F ∗ \ker T_F^* ker T F ∗ if and only if χ \chi χ is trivial. This proves (i)⇔ \Leftrightarrow ⇔ (iii).
Since χ ( g ) ∈ { − 1 , 1 } \chi(g)\in\{-1,1\} χ ( g ) ∈ { − 1 , 1 } , the kernel contains a strictly positive vector if and only if χ \chi χ is trivial. By Proposition 3.1, (ii)⇔ \Leftrightarrow ⇔ (iii).
Finally, the averaging operator
P = 1 ∣ G ∣ ∑ g ∈ G π ( g ) P=\frac{1}{|G|}\sum_{g\in G}\pi(g) P = ∣ G ∣ 1 g ∈ G ∑ π ( g )
is a projection onto the common fixed-point subspace of the representation. Since ξ \xi ξ is cyclic, the range of P P P is generated by P ξ P\xi P ξ . Consequently, the fixed-point subspace is trivial if and only if P ξ = 0 P\xi=0 P ξ = 0 , which is equivalent to balancedness.
An immediate consequence is that excess-one orbit frames generated by groups of odd order are necessarily positive.
Corollary 6.6. Let G G G be a finite group of odd order and let F = { π ( g ) ξ : g ∈ G } F=\{\pi(g)\xi:g\in G\} F = { π ( g ) ξ : g ∈ G } be a frame for H n \mathcal{H}_n H n with ∣ G ∣ = n + 1 |G|=n+1 ∣ G ∣ = n + 1 . Then F F F is balanced and positive. In particular, F F F is a minimal positive basis for H n \mathcal{H}_n H n .
Example 6.7. Let G = C 4 = ⟨ a : a 4 = e ⟩ G=C_4=\langle a:a^4=e\rangle G = C 4 = ⟨ a : a 4 = e ⟩ and consider the representation π : G → G L ( R 3 ) \pi:G\to GL(\mathbb{R}^3) π : G → G L ( R 3 ) determined by
π ( a ) = ( 1 0 0 0 0 − 1 0 1 0 ) \pi(a)=\begin{pmatrix}1&0&0\\0&0&-1\\0&1&0\end{pmatrix} π ( a ) = 1 0 0 0 0 1 0 − 1 0 For ξ = ( 1 , 1 , 0 ) \xi=(1,1,0) ξ = ( 1 , 1 , 0 ) , the corresponding orbit frame is
F = { ( 1 , 1 , 0 ) , ( 1 , 0 , 1 ) , ( 1 , − 1 , 0 ) , ( 1 , 0 , − 1 ) } . F=\{(1,1,0),(1,0,1),(1,-1,0),(1,0,-1)\}. F = {( 1 , 1 , 0 ) , ( 1 , 0 , 1 ) , ( 1 , − 1 , 0 ) , ( 1 , 0 , − 1 )} . A direct computation gives
ker T F ∗ = span { ( 1 , − 1 , 1 , − 1 ) } . \ker T_F^*=\operatorname{span}\{(1,-1,1,-1)\}. ker T F ∗ = span {( 1 , − 1 , 1 , − 1 )} . Thus every three vectors of F F F are linearly independent, and hence F F F is linearly connected. However,
∑ g ∈ G π ( g ) ξ = ( 4 , 0 , 0 ) ≠ 0. \sum_{g\in G}\pi(g)\xi=(4,0,0)\ne0. g ∈ G ∑ π ( g ) ξ = ( 4 , 0 , 0 ) = 0. Therefore F F F is neither balanced nor positive. The corresponding character is χ ( a j ) = ( − 1 ) j \chi(a^j)=(-1)^j χ ( a j ) = ( − 1 ) j , illustrating the obstruction to balancedness described in Theorem 6.5.
Positive Dependence, Convex Geometry, and Erasure Robustness
We begin by recalling the notion of minimal positive dependence in finite-dimensional real Hilbert spaces. Let I I I be a finite index set and let F I = { f i : i ∈ I } F_I=\{f_i:i\in I\} F I = { f i : i ∈ I } be a family of nonzero vectors in H n \mathcal{H}_n H n , with S I = span { f i : i ∈ I } S_I=\operatorname{span}\{f_i:i\in I\} S I = span { f i : i ∈ I } . The family F I F_I F I is said to be positively dependent if there exist strictly positive scalars c i c_i c i , i ∈ I i\in I i ∈ I , such that ∑ i ∈ I c i f i = 0 \sum_{i\in I}c_i f_i=0 ∑ i ∈ I c i f i = 0 . It is called minimally positively dependent if it is positively dependent, but no proper nonempty subfamily admits a nontrivial nonnegative linear dependence, that is, a relation with all coefficients nonnegative and at least one coefficient strictly positive.
Convex Geometry of Positive Relations
Suppose that F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N is an ( ℓ − 1 ) (\ell-1) ( ℓ − 1 ) -redundant and ℓ \ell ℓ -exact positive frame for R n \mathbb{R}^n R n , where ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 . By the known lower bound for redundant positive frames, we have N ≥ n + 2 ℓ − 1 N\ge n+2\ell-1 N ≥ n + 2 ℓ − 1 . On the other hand, every positive frame contains a positive basis B B B with l B ∣ ≤ 2 n lB\rvert\le2n l B ∣ ≤ 2 n . Since F F F is ℓ \ell ℓ -exact, there can be at most ℓ − 1 \ell-1 ℓ − 1 vectors outside B B B . Consequently,
n + 2 ℓ − 1 ≤ N ≤ 2 n + ℓ − 1. n+2\ell-1\le N\le2n+\ell-1. n + 2 ℓ − 1 ≤ N ≤ 2 n + ℓ − 1.
In particular, ℓ ≤ n \ell\le n ℓ ≤ n .
For ℓ = 1 \ell= 1 ℓ = 1 , both bounds are attained. Indeed, the family
F 1 = { e 1 , … , e n , − ∑ i = 1 n e i } F_1 = \left\{e_1,\ldots,e_n,-\sum_{i=1}^{n}e_i\right\} F 1 = { e 1 , … , e n , − i = 1 ∑ n e i }
is an exact positive frame for R n \mathbb{R}^n R n with n + 1 n+1 n + 1 elements, attaining the lower bound. Similarly,
F 2 = { ± e i : 1 ≤ i ≤ n } F_2 = \{\pm e_i: 1 \le i \le n\} F 2 = { ± e i : 1 ≤ i ≤ n }
is an exact positive frame with 2 n 2n 2 n elements, attaining the upper bound. Here { e i } i = 1 n \{e_i\}_{i=1}^{n} { e i } i = 1 n denotes the standard orthonormal basis of R n \mathbb{R}^n R n .
Let F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a finite family of vectors in a real Hilbert space H n \mathcal{H}_n H n . Consider the convex set
Ω = { λ = ( λ 1 , … , λ N ) ∈ [ 0 , ∞ ) N : ∑ i = 1 N λ i = 1 , ∑ i = 1 N λ i f i = 0 } . \Omega= \left\{\lambda= (\lambda_1,\ldots,\lambda_N) \in[0,\infty)^N : \sum_{i=1}^{N}\lambda_i = 1,\quad\sum_{i=1}^{N}\lambda_i f_i = 0\right\}. Ω = { λ = ( λ 1 , … , λ N ) ∈ [ 0 , ∞ ) N : i = 1 ∑ N λ i = 1 , i = 1 ∑ N λ i f i = 0 } .
For each λ ∈ Ω \lambda\in\Omega λ ∈ Ω , define I ( λ ) = { i : λ i > 0 } I(\lambda) = \{i : \lambda_i > 0\} I ( λ ) = { i : λ i > 0 } . The following proposition characterizes the extreme points of Ω \Omega Ω in terms of linear independence of the corresponding augmented vectors.
Proposition 7.1. Let λ ∈ Ω \lambda\in\Omega λ ∈ Ω . Then λ \lambda λ is an extreme point of Ω \Omega Ω if and only if the family
{ ( f i , 1 ) : i ∈ I ( λ ) } \{(f_i,1) : i \in I(\lambda)\} {( f i , 1 ) : i ∈ I ( λ )} is linearly independent in H n ⊕ R \mathcal{H}_n \oplus\mathbb{R} H n ⊕ R .
Proof. Suppose first that { ( f i , 1 ) : i ∈ I ( λ ) } \{(f_i,1) : i \in I(\lambda)\} {( f i , 1 ) : i ∈ I ( λ )} is linearly independent. Assume that
λ = t μ + ( 1 − t ) ν , 0 < t < 1 , \lambda= t\mu+ (1-t)\nu,\qquad0 < t < 1, λ = t μ + ( 1 − t ) ν , 0 < t < 1 , for some μ , ν ∈ Ω \mu,\nu\in\Omega μ , ν ∈ Ω .
If j ∉ I ( λ ) j \notin I(\lambda) j ∈ / I ( λ ) , then λ j = 0 \lambda_j = 0 λ j = 0 . Since μ j , ν j ≥ 0 \mu_j,\nu_j \ge0 μ j , ν j ≥ 0 , we have
0 = t μ j + ( 1 − t ) ν j , 0 = t\mu_j + (1-t)\nu_j, 0 = t μ j + ( 1 − t ) ν j , which implies μ j = ν j = 0 \mu_j = \nu_j = 0 μ j = ν j = 0 .
Consequently, using μ , ν ∈ Ω \mu,\nu\in\Omega μ , ν ∈ Ω , we obtain
∑ i ∈ I ( λ ) ( μ i − ν i ) f i = 0 and ∑ i ∈ I ( λ ) ( μ i − ν i ) = 0. \sum_{i\in I(\lambda)}(\mu_i-\nu_i)f_i = 0
\quad\text{and}\quad
\sum_{i\in I(\lambda)}(\mu_i-\nu_i) = 0. i ∈ I ( λ ) ∑ ( μ i − ν i ) f i = 0 and i ∈ I ( λ ) ∑ ( μ i − ν i ) = 0. Equivalently,
∑ i ∈ I ( λ ) ( μ i − ν i ) ( f i , 1 ) = 0. \sum_{i\in I(\lambda)}(\mu_i-\nu_i)(f_i,1) = 0. i ∈ I ( λ ) ∑ ( μ i − ν i ) ( f i , 1 ) = 0. By linear independence, μ i = ν i \mu_i = \nu_i μ i = ν i for every i ∈ I ( λ ) i \in I(\lambda) i ∈ I ( λ ) . Together with the equality outside I ( λ ) I(\lambda) I ( λ ) , this gives μ = ν = λ \mu= \nu= \lambda μ = ν = λ . Hence λ \lambda λ is an extreme point of Ω \Omega Ω .
Conversely, suppose that λ \lambda λ is an extreme point of Ω \Omega Ω . Assume, to the contrary, that { ( f i , 1 ) : i ∈ I ( λ ) } \{(f_i,1) : i \in I(\lambda)\} {( f i , 1 ) : i ∈ I ( λ )} is linearly dependent. Then there exist scalars a i a_i a i , not all zero, such that
∑ i ∈ I ( λ ) a i f i = 0 and ∑ i ∈ I ( λ ) a i = 0. \sum_{i\in I(\lambda)}a_i f_i = 0
\quad\text{and}\quad
\sum_{i\in I(\lambda)}a_i = 0. i ∈ I ( λ ) ∑ a i f i = 0 and i ∈ I ( λ ) ∑ a i = 0.
Extend a = ( a 1 , … , a N ) a=(a_1,\ldots,a_N) a = ( a 1 , … , a N ) by setting a i = 0 a_i=0 a i = 0 for i ∉ I ( λ ) i\notin I(\lambda) i ∈ / I ( λ ) .
Since λ i > 0 \lambda_i>0 λ i > 0 for every i ∈ I ( λ ) i\in I(\lambda) i ∈ I ( λ ) , we can choose ε > 0 \varepsilon>0 ε > 0 sufficiently small so that
λ i ± ε a i ≥ 0 ( 1 ≤ i ≤ N ) . \lambda_i\pm\varepsilon a_i\ge0\qquad(1\le i\le N). λ i ± ε a i ≥ 0 ( 1 ≤ i ≤ N ) .
Define
λ + = λ + ε a , λ − = λ − ε a . \lambda^+=\lambda+\varepsilon a,\qquad\lambda^-=\lambda-\varepsilon a. λ + = λ + ε a , λ − = λ − ε a .
Then λ + , λ − ∈ [ 0 , ∞ ) N \lambda^+,\lambda^-\in[0,\infty)^N λ + , λ − ∈ [ 0 , ∞ ) N . Moreover,
∑ i = 1 N λ i ± = ∑ i = 1 N λ i ± ε ∑ i = 1 N a i = 1 \sum_{i=1}^{N}\lambda_i^\pm=\sum_{i=1}^{N}\lambda_i\pm\varepsilon\sum_{i=1}^{N}a_i=1 i = 1 ∑ N λ i ± = i = 1 ∑ N λ i ± ε i = 1 ∑ N a i = 1
and
∑ i = 1 N λ i ± f i = ∑ i = 1 N λ i f i ± ε ∑ i = 1 N a i f i = 0. \sum_{i=1}^{N}\lambda_i^\pm f_i=\sum_{i=1}^{N}\lambda_i f_i\pm\varepsilon\sum_{i=1}^{N}a_i f_i=0. i = 1 ∑ N λ i ± f i = i = 1 ∑ N λ i f i ± ε i = 1 ∑ N a i f i = 0.
Thus λ + , λ − ∈ Ω \lambda^+,\lambda^-\in\Omega λ + , λ − ∈ Ω .
Since a ≠ 0 a\ne0 a = 0 , we have λ + ≠ λ − \lambda^+\ne\lambda^- λ + = λ − , while
λ = 1 2 λ + + 1 2 λ − . \lambda=\frac{1}{2}\lambda^++\frac{1}{2}\lambda^-. λ = 2 1 λ + + 2 1 λ − .
This contradicts the assumption that λ \lambda λ is an extreme point of Ω \Omega Ω . Therefore { ( f i , 1 ) : i ∈ I ( λ ) } \{(f_i,1):i\in I(\lambda)\} {( f i , 1 ) : i ∈ I ( λ )} is linearly independent.
Corollary 7.2. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a finite family in an n n n -dimensional real Hilbert space H \mathcal{H} H . If λ \lambda λ is an extreme point of Ω \Omega Ω , then
∣ I ( λ ) ∣ ≤ n + 1. |I(\lambda)|\le n+1. ∣ I ( λ ) ∣ ≤ n + 1. Proof. By the preceding proposition, the family { ( f i , 1 ) : i ∈ I ( λ ) } \{(f_i,1):i\in I(\lambda)\} {( f i , 1 ) : i ∈ I ( λ )} is linearly independent in H ⊕ R \mathcal{H}\oplus\mathbb{R} H ⊕ R . Since dim ( H ⊕ R ) = n + 1 \dim(\mathcal{H}\oplus\mathbb{R})=n+1 dim ( H ⊕ R ) = n + 1 , the conclusion follows.
Corollary 7.3. Let λ \lambda λ be an extreme point of Ω \Omega Ω . Then λ \lambda λ is the unique element of Ω \Omega Ω whose support is contained in I ( λ ) I(\lambda) I ( λ ) .
Proof. Suppose μ ∈ Ω \mu\in\Omega μ ∈ Ω satisfies I ( μ ) ⊆ I ( λ ) I(\mu)\subseteq I(\lambda) I ( μ ) ⊆ I ( λ ) . Since λ , μ ∈ Ω \lambda,\mu\in\Omega λ , μ ∈ Ω , we have
∑ i ∈ I ( λ ) ( λ i − μ i ) f i = 0 and ∑ i ∈ I ( λ ) ( λ i − μ i ) = 0. \sum_{i\in I(\lambda)}(\lambda_i-\mu_i)f_i=0
\qquad\text{and}\qquad
\sum_{i\in I(\lambda)}(\lambda_i-\mu_i)=0. i ∈ I ( λ ) ∑ ( λ i − μ i ) f i = 0 and i ∈ I ( λ ) ∑ ( λ i − μ i ) = 0. Consequently,
∑ i ∈ I ( λ ) ( λ i − μ i ) ( f i , 1 ) = 0. \sum_{i\in I(\lambda)}(\lambda_i-\mu_i)(f_i,1)=0. i ∈ I ( λ ) ∑ ( λ i − μ i ) ( f i , 1 ) = 0. By the preceding proposition, { ( f i , 1 ) : i ∈ I ( λ ) } \{(f_i,1):i\in I(\lambda)\} {( f i , 1 ) : i ∈ I ( λ )} is linearly independent. Hence λ i = μ i \lambda_i=\mu_i λ i = μ i for every i ∈ I ( λ ) i\in I(\lambda) i ∈ I ( λ ) , and therefore λ = μ \lambda=\mu λ = μ .
Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a finite frame consisting of pairwise distinct vectors in a real Hilbert space H n \mathcal{H}_n H n , and let T F T_F T F and T F ∗ T_F^* T F ∗ denote its analysis and synthesis operators, respectively. The following result characterizes the extreme points of conv ( F ) \operatorname{conv}(F) conv ( F ) in terms of the kernel of the synthesis operator.
Proposition 7.4. For each i ∈ { 1 , … , N } i \in\{1,\ldots,N\} i ∈ { 1 , … , N } , f i ∈ Ext ( conv ( F ) ) f_i \in\operatorname{Ext}(\operatorname{conv}(F)) f i ∈ Ext ( conv ( F )) if and only if there does not exist d = ( d 1 , … , d N ) ∈ ker T F ∗ d=(d_1,\ldots,d_N)\in\ker T_F^* d = ( d 1 , … , d N ) ∈ ker T F ∗ such that
d i < 0 , d j ≥ 0 ( j ≠ i ) , ∑ j = 1 N d j = 0. d_i<0,\qquad d_j\geq0\quad(j\neq i),\qquad\sum_{j=1}^{N}d_j=0. d i < 0 , d j ≥ 0 ( j = i ) , j = 1 ∑ N d j = 0.
Proof. Suppose that f i ∉ Ext ( conv ( F ) ) f_i\notin\operatorname{Ext}(\operatorname{conv}(F)) f i ∈ / Ext ( conv ( F )) . Since the vectors of F F F are pairwise distinct, there exist scalars λ j ≥ 0 \lambda_j\geq0 λ j ≥ 0 , j ≠ i j\neq i j = i , such that
f i = ∑ j ≠ i λ j f j , ∑ j ≠ i λ j = 1. f_i=\sum_{j\neq i}\lambda_j f_j,\qquad\sum_{j\neq i}\lambda_j=1. f i = j = i ∑ λ j f j , j = i ∑ λ j = 1.
Define d = ( d 1 , … , d N ) d=(d_1,\ldots,d_N) d = ( d 1 , … , d N ) by d i = − 1 d_i=-1 d i = − 1 and d j = λ j d_j=\lambda_j d j = λ j for j ≠ i j\neq i j = i . Then
T F ∗ d = − f i + ∑ j ≠ i λ j f j = 0. T_F^*d=-f_i+\sum_{j\neq i}\lambda_j f_j=0. T F ∗ d = − f i + j = i ∑ λ j f j = 0.
Thus d ∈ ker T F ∗ d\in\ker T_F^* d ∈ ker T F ∗ , with d i < 0 d_i<0 d i < 0 , d j ≥ 0 d_j\geq0 d j ≥ 0 for every j ≠ i j\neq i j = i , and ∑ j = 1 N d j = 0 \sum_{j=1}^{N}d_j=0 ∑ j = 1 N d j = 0 .
Conversely, suppose that there exists d ∈ ker T F ∗ d\in\ker T_F^* d ∈ ker T F ∗ satisfying
d i < 0 , d j ≥ 0 ( j ≠ i ) , ∑ j = 1 N d j = 0. d_i<0,\qquad d_j\geq0\quad(j\neq i),\qquad\sum_{j=1}^{N}d_j=0. d i < 0 , d j ≥ 0 ( j = i ) , j = 1 ∑ N d j = 0.
Since T F ∗ d = 0 T_F^*d=0 T F ∗ d = 0 , we have
d i f i + ∑ j ≠ i d j f j = 0. d_i f_i+\sum_{j\neq i}d_j f_j=0. d i f i + j = i ∑ d j f j = 0.
Consequently,
f i = ∑ j ≠ i d j − d i f j , ∑ j ≠ i d j − d i = 1. f_i=\sum_{j\neq i}\frac{d_j}{-d_i}f_j,\qquad\sum_{j\neq i}\frac{d_j}{-d_i}=1. f i = j = i ∑ − d i d j f j , j = i ∑ − d i d j = 1.
Since all coefficients are nonnegative, f i ∈ conv { f j : j ≠ i } f_i\in\operatorname{conv}\{f_j:j\neq i\} f i ∈ conv { f j : j = i } . As the frame vectors are pairwise distinct, f i f_i f i cannot be an extreme point of conv ( F ) \operatorname{conv}(F) conv ( F ) .
Positive Spanning and Erasure Robustness
We next introduce a quantity that measures the robustness of a positive frame under arbitrary vector erasures. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a finite family of vectors in R n \mathbb{R}^{n} R n . For each 0 ≠ x ∈ R n 0\neq x\in\mathbb{R}^{n} 0 = x ∈ R n , define
N F + ( x ) = # { i ∈ { 1 , … , N } : ⟨ x , f i ⟩ > 0 } , N_F^+(x)=\#\{i\in\{1,\ldots,N\}:\langle x,f_i\rangle>0\}, N F + ( x ) = # { i ∈ { 1 , … , N } : ⟨ x , f i ⟩ > 0 } ,
where # \# # denotes the cardinality of a set. We further define
d + ( F ) = min x ∈ R n ∖ { 0 } N F + ( x ) . d_+(F)=\min_{x\in\mathbb{R}^{n}\setminus\{0\}}N_F^+(x). d + ( F ) = x ∈ R n ∖ { 0 } min N F + ( x ) .
Equivalently, by homogeneity,
d + ( F ) = min ∥ x ∥ = 1 N F + ( x ) . d_+(F)=\min_{\|x\|=1}N_F^+(x). d + ( F ) = ∥ x ∥ = 1 min N F + ( x ) .
Recall that F F F is a positive frame for R n \mathbb{R}^{n} R n if pos ( F ) = R n \operatorname{pos}(F)=\mathbb{R}^{n} pos ( F ) = R n , where
pos ( F ) = { ∑ i = 1 N a i f i : a i ≥ 0 } . \operatorname{pos}(F)=\left\{\sum_{i=1}^{N}a_i f_i:a_i\ge0\right\}. pos ( F ) = { i = 1 ∑ N a i f i : a i ≥ 0 } .
By the positive spanning criterion, this is equivalent to the requirement that, for every 0 ≠ x ∈ R n 0\ne x\in\mathbb{R}^{n} 0 = x ∈ R n , there exists an index i i i such that ⟨ x , f i ⟩ > 0 \langle x,f_i\rangle>0 ⟨ x , f i ⟩ > 0 .
For a nonnegative integer s s s , we say that F F F is s s s -redundant as a positive frame if the deletion of any collection of at most s s s vectors leaves a positive frame for R n \mathbb{R}^{n} R n .
Proposition 7.5. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a finite family of vectors in R n \mathbb{R}^{n} R n , and let s ∈ { 0 , 1 , … , N } s\in\{0,1,\ldots,N\} s ∈ { 0 , 1 , … , N } . Then F F F remains a positive frame after the deletion of any collection of at most s s s vectors if and only if
d + ( F ) ≥ s + 1. d_{+}(F)\ge s+1. d + ( F ) ≥ s + 1. Proof. Suppose first that d + ( F ) ≥ s + 1 d_{+}(F)\ge s+1 d + ( F ) ≥ s + 1 . Let J ⊆ { 1 , … , N } J\subseteq\{1,\ldots,N\} J ⊆ { 1 , … , N } be arbitrary with ∣ J ∣ ≤ s |J|\le s ∣ J ∣ ≤ s , and consider the remaining family
F J c = { f i : i ∉ J } . F_{J^{c}}=\{f_i:i\notin J\}. F J c = { f i : i ∈ / J } . For every 0 ≠ x ∈ R n 0\ne x\in\mathbb{R}^{n} 0 = x ∈ R n , there are at least s + 1 s+1 s + 1 indices i i i such that ⟨ x , f i ⟩ > 0 \langle x,f_i\rangle>0 ⟨ x , f i ⟩ > 0 . Since at most s s s indices have been removed, there exists at least one index i ∉ J i\notin J i ∈ / J satisfying
⟨ x , f i ⟩ > 0. \langle x,f_i\rangle>0. ⟨ x , f i ⟩ > 0. By the positive spanning criterion, F J c F_{J^{c}} F J c positively spans R n \mathbb{R}^{n} R n . As J J J was arbitrary, F F F remains positive after any deletion of at most s s s vectors.
Conversely, suppose that F F F remains a positive frame after the deletion of any collection of at most s s s vectors. Assume, to the contrary, that d + ( F ) ≤ s d_{+}(F)\le s d + ( F ) ≤ s . Then there exists 0 ≠ x ∈ R n 0\ne x\in\mathbb{R}^{n} 0 = x ∈ R n such that
# { i : ⟨ x , f i ⟩ > 0 } ≤ s . \#\{i:\langle x,f_i\rangle>0\}\le s. # { i : ⟨ x , f i ⟩ > 0 } ≤ s . Let
J = { i : ⟨ x , f i ⟩ > 0 } . J=\{i:\langle x,f_i\rangle>0\}. J = { i : ⟨ x , f i ⟩ > 0 } . Clearly, ∣ J ∣ ≤ s |J|\le s ∣ J ∣ ≤ s . Upon deleting the vectors indexed by J J J , we obtain
⟨ x , f i ⟩ ≤ 0 for every i ∉ J . \langle x,f_i\rangle\le0\qquad\text{for every }i\notin J. ⟨ x , f i ⟩ ≤ 0 for every i ∈ / J . Thus the remaining family fails the positive spanning criterion and hence is not a positive frame. This contradicts our assumption. Therefore d + ( F ) ≥ s + 1 d_{+}(F)\ge s+1 d + ( F ) ≥ s + 1 , completing the proof. □ \square □
Exactness and Cardinality Bounds
We now investigate the relationship between positive spanning, exactness, and the cardinality of positive frames. In particular, we examine how the removal of frame vectors affects the positive spanning property and characterize positive frames that lose this property upon the deletion of a prescribed number of vectors. We also derive cardinality bounds that reflect the interplay between redundancy and exactness, providing further insight into the combinatorial structure of positive frames.
Proposition 7.6. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for H n \mathcal{H}_n H n , and let s s s be a nonnegative integer with N − s ≥ n + 1 N-s\ge n+1 N − s ≥ n + 1 . Then, for G = { g i } i = 1 N ∈ D ( F ) G=\{g_i\}_{i=1}^{N}\in\mathcal{D}(F) G = { g i } i = 1 N ∈ D ( F ) , the following statements are equivalent:
(i) G G G is an s s s -redundant positive frame.
(ii) For every nonzero x ∈ H n x \in\mathcal{H}_n x ∈ H n ,
∣ { i : ⟨ x , g i ⟩ > 0 } ∣ ≥ s + 1. \left|\{i : \langle x,g_i\rangle> 0\}\right| \ge s+1. ∣ { i : ⟨ x , g i ⟩ > 0 } ∣ ≥ s + 1.
Proof. (i) ⇒ \Rightarrow ⇒ (ii). Suppose that G G G is an s s s -redundant positive frame. Assume, to the contrary, that there exists a nonzero x ∈ H n x \in\mathcal{H}_n x ∈ H n such that
∣ { i : ⟨ x , g i ⟩ > 0 } ∣ ≤ s . \left|\{i : \langle x,g_i\rangle> 0\}\right| \le s. ∣ { i : ⟨ x , g i ⟩ > 0 } ∣ ≤ s .
Set
I + ( x ) = { i : ⟨ x , g i ⟩ > 0 } . I_+(x)=\{i : \langle x,g_i\rangle> 0\}. I + ( x ) = { i : ⟨ x , g i ⟩ > 0 } .
Choose a subset Λ ⊆ { 1 , … , N } \Lambda\subseteq\{1,\ldots,N\} Λ ⊆ { 1 , … , N } with ∣ Λ ∣ = s |\Lambda|=s ∣Λ∣ = s and I + ( x ) ⊆ Λ I_+(x)\subseteq\Lambda I + ( x ) ⊆ Λ . Then
⟨ x , g i ⟩ ≤ 0 for every i ∈ Λ c . \langle x,g_i\rangle\le0 \qquad\text{for every } i\in\Lambda^c. ⟨ x , g i ⟩ ≤ 0 for every i ∈ Λ c .
By the positive-spanning criterion, the subfamily { g i } i ∈ Λ c \{g_i\}_{i\in\Lambda^c} { g i } i ∈ Λ c is not a positive frame, contradicting the s s s -redundancy of G G G . Hence (ii) holds.
(ii) ⇒ \Rightarrow ⇒ (i). Suppose that (ii) holds, and let Λ ⊆ { 1 , … , N } \Lambda\subseteq\{1,\ldots,N\} Λ ⊆ { 1 , … , N } be arbitrary with ∣ Λ ∣ = s |\Lambda|=s ∣Λ∣ = s . For every nonzero x ∈ H n x\in\mathcal{H}_n x ∈ H n , there are at least s + 1 s+1 s + 1 indices i i i such that ⟨ x , g i ⟩ > 0 \langle x,g_i\rangle>0 ⟨ x , g i ⟩ > 0 . Since ∣ Λ ∣ = s |\Lambda|=s ∣Λ∣ = s , at least one such index belongs to Λ c \Lambda^c Λ c . Consequently,
max i ∈ Λ c ⟨ x , g i ⟩ > 0 \max_{i\in\Lambda^c}\langle x,g_i\rangle>0 i ∈ Λ c max ⟨ x , g i ⟩ > 0
for every nonzero x ∈ H n x\in\mathcal{H}_n x ∈ H n . By the positive-spanning criterion, { g i } i ∈ Λ c \{g_i\}_{i\in\Lambda^c} { g i } i ∈ Λ c positively spans H n \mathcal{H}_n H n . Since Λ \Lambda Λ was arbitrary, G G G is an s s s -redundant positive frame.
The preceding proposition also provides a characterization of positive frames that are f f f -redundant and ( f + 1 ) (f+1) ( f + 1 ) -exact. Here, ( f + 1 ) (f+1) ( f + 1 ) -exactness means that deleting any f + 1 f+1 f + 1 vectors destroys the positive spanning property.
Combining the preceding characterization of exactness with Proposition 7.5 , we obtain the following criterion for positive frames that are simultaneously f f f -redundant and ( f + 1 ) (f+1) ( f + 1 ) -exact.
Proposition 7.7. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a positive frame for R n \mathbb{R}^n R n , and let 1 ≤ ℓ ≤ N 1\le\ell\le N 1 ≤ ℓ ≤ N . Then the following statements are equivalent:
(i) F F F is ℓ \ell ℓ -exact.
(ii) For every J ⊆ { 1 , … , N } J\subseteq\{1,\ldots,N\} J ⊆ { 1 , … , N } with ∣ J ∣ = ℓ |J|=\ell ∣ J ∣ = ℓ , there exists x J ≠ 0 x_J\ne0 x J = 0 such that
⟨ x J , f i ⟩ ≤ 0 ( i ∉ J ) . \langle x_J,f_i\rangle\le0 \qquad(i\notin J). ⟨ x J , f i ⟩ ≤ 0 ( i ∈ / J ) . (iii) For every J ⊆ { 1 , … , N } J\subseteq\{1,\ldots,N\} J ⊆ { 1 , … , N } with ∣ J ∣ = ℓ |J|=\ell ∣ J ∣ = ℓ , there exists x J ≠ 0 x_J\ne0 x J = 0 such that
P F ( x J ) ⊆ J . P_F(x_J)\subseteq J. P F ( x J ) ⊆ J .
Corollary 7.8. Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a positive frame for R n \mathbb{R}^n R n , and let 0 ≤ f < N 0\le f<N 0 ≤ f < N . Then F F F is f f f -redundant and ( f + 1 ) (f+1) ( f + 1 ) -exact if and only if d + ( F ) = f + 1 d_+(F)=f+1 d + ( F ) = f + 1 and, for every J ⊆ { 1 , … , N } J\subseteq\{1,\ldots,N\} J ⊆ { 1 , … , N } with ∣ J ∣ = f + 1 |J|=f+1 ∣ J ∣ = f + 1 , there exists x J ≠ 0 x_J\ne0 x J = 0 such that P F ( x J ) = J P_F(x_J)=J P F ( x J ) = J .
The preceding characterization yields an explicit expression for the maximum number of erasures under which a finite positive frame remains positive.
Corollary 7.9. Let F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a finite positive frame for a real Hilbert space H \mathcal{H} H . Then the maximum number of arbitrary erasures that preserve positivity is
r + ( F ) = d + ( F ) − 1. r_{+}(F) = d_{+}(F) - 1. r + ( F ) = d + ( F ) − 1. Proof. By Proposition 7.5, F F F remains positive after any deletion of at most s s s vectors precisely when s < d + ( F ) s < d_{+}(F) s < d + ( F ) . Therefore, the largest such integer is d + ( F ) − 1 d_{+}(F) - 1 d + ( F ) − 1 .
Approximation by Positive Parseval Frames
Let H n \mathcal{H}_n H n be an n n n -dimensional real Hilbert space, and let N > n N > n N > n be fixed. Consider the set
K = { G = { g i } i = 1 N : G is a positive Parseval frame for H n } . \mathcal{K} = \{G = \{g_i\}_{i=1}^{N} : G\text{ is a positive Parseval frame for }\mathcal{H}_n\}. K = { G = { g i } i = 1 N : G is a positive Parseval frame for H n } .
We equip H n N \mathcal{H}_n^N H n N with the norm ∥ F ∥ 2 = ( ∑ i = 1 N ∥ f i ∥ 2 ) 1 / 2 \|F\|_2 = \left(\sum_{i=1}^{N}\|f_i\|^2\right)^{1/2} ∥ F ∥ 2 = ( ∑ i = 1 N ∥ f i ∥ 2 ) 1/2 .
We first observe that K \mathcal{K} K is bounded but neither convex nor closed. Indeed, since every G = { g i } i = 1 N ∈ K G = \{g_i\}_{i=1}^{N} \in\mathcal{K} G = { g i } i = 1 N ∈ K is Parseval, ∑ i = 1 N ∥ g i ∥ 2 = tr ( S G ) = n \sum_{i=1}^{N}\|g_i\|^2 = \operatorname{tr}(S_G) = n ∑ i = 1 N ∥ g i ∥ 2 = tr ( S G ) = n . Thus ∥ G ∥ 2 = n \|G\|_2 = \sqrt{n} ∥ G ∥ 2 = n for every G ∈ K G \in\mathcal{K} G ∈ K , which shows that K \mathcal{K} K is bounded. On the other hand, if G ∈ K G \in\mathcal{K} G ∈ K , then − G ∈ K -G \in\mathcal{K} − G ∈ K , whereas 1 2 G + 1 2 ( − G ) = 0 ∉ K \frac{1}{2}G + \frac{1}{2}(-G) = 0 \notin\mathcal{K} 2 1 G + 2 1 ( − G ) = 0 ∈ / K . Hence K \mathcal{K} K is not convex.
Furthermore, K \mathcal{K} K is not closed. To see this, let { e 1 , … , e n } \{e_1,\ldots,e_n\} { e 1 , … , e n } be an orthonormal basis of H n \mathcal{H}_n H n and put N = n + 1 N = n+1 N = n + 1 . For 0 < t < 1 0 < t < 1 0 < t < 1 , define
v i ( t ) = e i , 1 ≤ i ≤ n , v n + 1 ( t ) = − t ∑ i = 1 n e i . v_i(t) = e_i,\quad1 \leq i \leq n,\qquad v_{n+1}(t) = -t\sum_{i=1}^{n}e_i. v i ( t ) = e i , 1 ≤ i ≤ n , v n + 1 ( t ) = − t i = 1 ∑ n e i .
Then V t = { v i ( t ) } i = 1 n + 1 V_t = \{v_i(t)\}_{i=1}^{n+1} V t = { v i ( t ) } i = 1 n + 1 is a positive frame for every t > 0 t > 0 t > 0 . Let S t S_t S t denote its frame operator and define
G t = { S t − 1 / 2 v i ( t ) } i = 1 n + 1 . G_t = \{S_t^{-1/2}v_i(t)\}_{i=1}^{n+1}. G t = { S t − 1/2 v i ( t ) } i = 1 n + 1 .
Since S t − 1 / 2 S_t^{-1/2} S t − 1/2 is invertible, each G t G_t G t is a positive Parseval frame. Moreover, S t → I H n S_t \to I_{\mathcal{H}_n} S t → I H n as t → 0 + t \to0^{+} t → 0 + , and hence
G t ⟶ { e 1 , … , e n , 0 } , G_t \longrightarrow\{e_1,\ldots,e_n,0\}, G t ⟶ { e 1 , … , e n , 0 } ,
which is Parseval but not positive. Thus K \mathcal{K} K is not closed. For N > n + 1 N > n+1 N > n + 1 , the same conclusion follows by adjoining N − n − 1 N-n-1 N − n − 1 zero vectors to each frame in this construction.
We now determine the distance from an arbitrary positive frame to the set K \mathcal{K} K .
Proposition 7.10. For every positive frame F = { f i } i = 1 N F = \{f_i\}_{i=1}^{N} F = { f i } i = 1 N with frame operator S F S_F S F , we have
dist ( F , K ) = ∥ S F 1 / 2 − I H n ∥ H S = ( ∑ j = 1 n ( λ j − 1 ) 2 ) 1 / 2 , \operatorname{dist}(F,\mathcal{K}) = \|S_F^{1/2}-I_{\mathcal{H}_n}\|_{\mathrm{HS}} = \left(\sum_{j=1}^{n}(\sqrt{\lambda_j}-1)^2\right)^{1/2}, dist ( F , K ) = ∥ S F 1/2 − I H n ∥ HS = ( j = 1 ∑ n ( λ j − 1 ) 2 ) 1/2 ,
where λ 1 , … , λ n \lambda_1,\ldots,\lambda_n λ 1 , … , λ n are the eigenvalues of S F S_F S F . Furthermore, the distance is attained at the canonical Parseval frame { S F − 1 / 2 f i } i = 1 N \{S_F^{-1/2}f_i\}_{i=1}^{N} { S F − 1/2 f i } i = 1 N .
Proof. Let A = Θ F ∗ A = \Theta_F^{*} A = Θ F ∗ be the synthesis operator of F F F , so that A A ∗ = S F AA^{*} = S_F A A ∗ = S F . Define U = S F − 1 / 2 A U = S_F^{-1/2}A U = S F − 1/2 A . Then
U U ∗ = S F − 1 / 2 A A ∗ S F − 1 / 2 = I H n . (4) UU^{*} = S_F^{-1/2}AA^{*}S_F^{-1/2} = I_{\mathcal{H}_n}. \tag*{(4)} U U ∗ = S F − 1/2 A A ∗ S F − 1/2 = I H n . (4)
Since F F F is positive and S F − 1 / 2 S_F^{-1/2} S F − 1/2 is invertible, the family { S F − 1 / 2 f i } i = 1 N \{S_F^{-1/2}f_i\}_{i=1}^{N} { S F − 1/2 f i } i = 1 N is a positive Parseval frame. Thus U U U is the synthesis operator of an element of K \mathcal{K} K .
Let G = { g i } i = 1 N ∈ K G=\{g_i\}_{i=1}^{N}\in\mathcal{K} G = { g i } i = 1 N ∈ K be arbitrary, and let B = Θ G ∗ B=\Theta_G^* B = Θ G ∗ be its synthesis operator. Since G G G is Parseval, B B ∗ = I H BB^*=I_{\mathcal{H}} B B ∗ = I H . Using A = S F 1 / 2 U A=S_F^{1/2}U A = S F 1/2 U , we obtain
∥ A − B ∥ H S 2 = tr ( A A ∗ ) + tr ( B B ∗ ) − 2 tr ( A B ∗ ) = tr ( S F ) + n − 2 tr ( S F 1 / 2 U B ∗ ) . \|A-B\|_{\mathrm{HS}}^2=\operatorname{tr}(AA^*)+\operatorname{tr}(BB^*)-2\operatorname{tr}(AB^*)
=\operatorname{tr}(S_F)+n-2\operatorname{tr}(S_F^{1/2}UB^*). ∥ A − B ∥ HS 2 = tr ( A A ∗ ) + tr ( B B ∗ ) − 2 tr ( A B ∗ ) = tr ( S F ) + n − 2 tr ( S F 1/2 U B ∗ ) .
Since U U ∗ = B B ∗ = I H UU^*=BB^*=I_{\mathcal{H}} U U ∗ = B B ∗ = I H , we have ∥ U ∥ o p = ∥ B ∗ ∥ o p = 1 \|U\|_{\mathrm{op}}=\|B^*\|_{\mathrm{op}}=1 ∥ U ∥ op = ∥ B ∗ ∥ op = 1 . Consequently,
∥ U B ∗ ∥ o p ≤ ∥ U ∥ o p ∥ B ∗ ∥ o p = 1. \|UB^*\|_{\mathrm{op}}\leq\|U\|_{\mathrm{op}}\|B^*\|_{\mathrm{op}}=1. ∥ U B ∗ ∥ op ≤ ∥ U ∥ op ∥ B ∗ ∥ op = 1.
As S F 1 / 2 S_F^{1/2} S F 1/2 is positive definite, it follows that
tr ( S F 1 / 2 U B ∗ ) ≤ tr ( S F 1 / 2 ) . \operatorname{tr}(S_F^{1/2}UB^*)\leq\operatorname{tr}(S_F^{1/2}). tr ( S F 1/2 U B ∗ ) ≤ tr ( S F 1/2 ) .
Therefore,
∥ A − B ∥ H S 2 ≥ tr ( S F ) + n − 2 tr ( S F 1 / 2 ) = ∥ S F 1 / 2 − I H n ∥ H S 2 . \|A-B\|_{\mathrm{HS}}^2\geq\operatorname{tr}(S_F)+n-2\operatorname{tr}(S_F^{1/2})=\|S_F^{1/2}-I_{\mathcal{H}_n}\|_{\mathrm{HS}}^2. ∥ A − B ∥ HS 2 ≥ tr ( S F ) + n − 2 tr ( S F 1/2 ) = ∥ S F 1/2 − I H n ∥ HS 2 .
Equality holds when B = U B=U B = U , corresponding to the canonical Parseval frame { S F − 1 / 2 f i } i = 1 N \{S_F^{-1/2}f_i\}_{i=1}^{N} { S F − 1/2 f i } i = 1 N . Hence
dist ( F , K ) = ∥ S F 1 / 2 − I H n ∥ H S . \operatorname{dist}(F,\mathcal{K})=\|S_F^{1/2}-I_{\mathcal{H}_n}\|_{\mathrm{HS}}. dist ( F , K ) = ∥ S F 1/2 − I H n ∥ HS .
Finally, if λ 1 , … , λ n \lambda_1,\ldots,\lambda_n λ 1 , … , λ n are the eigenvalues of S F S_F S F , then
∥ S F 1 / 2 − I H n ∥ H S 2 = ∑ j = 1 n ( λ j − 1 ) 2 , \|S_F^{1/2}-I_{\mathcal{H}_n}\|_{\mathrm{HS}}^2=\sum_{j=1}^{n}(\sqrt{\lambda_j}-1)^2, ∥ S F 1/2 − I H n ∥ HS 2 = j = 1 ∑ n ( λ j − 1 ) 2 ,
which completes the proof.
We now consider the distance from an arbitrary frame, not necessarily positive, to the set K \mathcal{K} K . Let F = { f i } i = 1 N F=\{f_i\}_{i=1}^{N} F = { f i } i = 1 N be a frame for H n \mathcal{H}_n H n , with analysis operator T F T_F T F and frame operator S F = T F ∗ T F S_F=T_F^*T_F S F = T F ∗ T F . Recall that a frame G = { g i } i = 1 N G=\{g_i\}_{i=1}^{N} G = { g i } i = 1 N is positive if and only if there exists a vector c ∈ R > 0 N c\in\mathbb{R}_{>0}^{N} c ∈ R > 0 N such that T G ∗ c = 0 T_G^*c=0 T G ∗ c = 0 .
For each c ∈ R N c\in\mathbb{R}^{N} c ∈ R N with ∥ c ∥ = 1 \|c\|=1 ∥ c ∥ = 1 , let P c = I N − c c ∗ P_c=I_N-cc^* P c = I N − c c ∗ denote the orthogonal projection onto c ⊥ c^\perp c ⊥ . Since every positive Parseval frame G G G satisfies T G ∗ T G = I H n T_G^*T_G=I_{\mathcal{H}_n} T G ∗ T G = I H n and T G ∗ c = 0 T_G^*c=0 T G ∗ c = 0 for some strictly positive unit vector c c c , we have
∥ T F ∗ − T G ∗ ∥ H S 2 = tr ( S F ) + n − 2 tr ( T F ∗ T G ) = tr ( S F ) + n − 2 tr ( T F ∗ P c T G ) . \|T_F^*-T_G^*\|_{\mathrm{HS}}^2=\operatorname{tr}(S_F)+n-2\operatorname{tr}(T_F^*T_G)=\operatorname{tr}(S_F)+n-2\operatorname{tr}(T_F^*P_cT_G). ∥ T F ∗ − T G ∗ ∥ HS 2 = tr ( S F ) + n − 2 tr ( T F ∗ T G ) = tr ( S F ) + n − 2 tr ( T F ∗ P c T G ) .
For a fixed c > 0 c>0 c > 0 with ∥ c ∥ = 1 \|c\|=1 ∥ c ∥ = 1 , the rectangular orthogonal Procrustes formula gives
sup T G ∗ T G = I H n T G ∗ c = 0 tr ( T F ∗ P c T G ) = tr ( ( T F ∗ P c T F ) 1 / 2 ) . \sup_{\substack{T_G^*T_G=I_{\mathcal{H}_n}\\T_G^*c=0}}\operatorname{tr}(T_F^*P_cT_G)
=\operatorname{tr}\left((T_F^*P_cT_F)^{1/2}\right). T G ∗ T G = I H n T G ∗ c = 0 sup tr ( T F ∗ P c T G ) = tr ( ( T F ∗ P c T F ) 1/2 ) .
Consequently, taking the infimum over all positive Parseval frames and using the continuity of the trace of the positive square root, we obtain
dist ( F , K ) 2 = tr ( S F ) + n − 2 max c ∈ R > 0 N ∥ c ∥ = 1 tr ( ( T F ∗ P c T F ) 1 / 2 ) . \operatorname{dist}(F,\mathcal{K})^2=\operatorname{tr}(S_F)+n-2\max_{\substack{c\in\mathbb{R}_{>0}^{N}\\\|c\|=1}}\operatorname{tr}\left((T_F^*P_cT_F)^{1/2}\right). dist ( F , K ) 2 = tr ( S F ) + n − 2 c ∈ R > 0 N ∥ c ∥ = 1 max tr ( ( T F ∗ P c T F ) 1/2 ) .
Since T F ∗ P c T F = S F − ( T F ∗ c ) ⊗ ( T F ∗ c ) T_F^*P_cT_F=S_F-(T_F^*c)\otimes(T_F^*c) T F ∗ P c T F = S F − ( T F ∗ c ) ⊗ ( T F ∗ c ) , this can equivalently be written as
dist ( F , K ) 2 = tr ( S F ) + n − 2 max c ∈ R ≥ 0 N ∥ c ∥ = 1 tr ( [ S F − ( T F ∗ c ) ⊗ ( T F ∗ c ) ] 1 / 2 ) . \operatorname{dist}(F,\mathcal{K})^2=\operatorname{tr}(S_F)+n-2\max_{\substack{c\in\mathbb{R}_{\geq0}^{N}\\ \lVert c\rVert=1}}\operatorname{tr}\left(\left[S_F-(T_F^*c)\otimes(T_F^*c)\right]^{1/2}\right). dist ( F , K ) 2 = tr ( S F ) + n − 2 c ∈ R ≥ 0 N ∥ c ∥ = 1 max tr ( [ S F − ( T F ∗ c ) ⊗ ( T F ∗ c ) ] 1/2 ) .
Furthermore, since 0 ≤ P c ≤ I N 0\leq P_c\leq I_N 0 ≤ P c ≤ I N , we have T F ∗ P c T F ≤ S F T_F^*P_cT_F\leq S_F T F ∗ P c T F ≤ S F . By the operator monotonicity of the square root,
tr ( ( T F ∗ P c T F ) 1 / 2 ) ≤ tr ( S F 1 / 2 ) . \operatorname{tr}\left((T_F^*P_cT_F)^{1/2}\right)\leq\operatorname{tr}(S_F^{1/2}). tr ( ( T F ∗ P c T F ) 1/2 ) ≤ tr ( S F 1/2 ) .
It follows that
dist ( F , K ) ≥ ∥ S F 1 / 2 − I H n ∥ H S . \operatorname{dist}(F,\mathcal{K})\geq\lVert S_F^{1/2}-I_{\mathcal{H}_n}\rVert_{\mathrm{HS}}. dist ( F , K ) ≥ ∥ S F 1/2 − I H n ∥ HS .
Moreover, equality holds if and only if there exists a nonzero vector c ∈ R ≥ 0 N c\in\mathbb{R}_{\geq0}^{N} c ∈ R ≥ 0 N such that T F ∗ c = 0 T_F^*c=0 T F ∗ c = 0 . In particular, the distance formula in Proposition 7.10 remains valid for every semi-positive frame admitting a nontrivial nonnegative dependence. On the other hand, if ker T F ∗ ∩ R ≥ 0 N = { 0 } \ker T_F^*\cap\mathbb{R}_{\geq0}^{N}=\{0\} ker T F ∗ ∩ R ≥ 0 N = { 0 } , then
dist ( F , K ) > ∥ S F 1 / 2 − I H n ∥ H S . \operatorname{dist}(F,\mathcal{K})>\lVert S_F^{1/2}-I_{\mathcal{H}_n}\rVert_{\mathrm{HS}}. dist ( F , K ) > ∥ S F 1/2 − I H n ∥ HS .
To illustrate this distinction, consider the frame F = { 1 / 2 , 1 / 2 } F=\{1/\sqrt{2},1/\sqrt{2}\} F = { 1/ 2 , 1/ 2 } for H 1 = R \mathcal{H}_1=\mathbb{R} H 1 = R . Clearly, S F = I H 1 S_F=I_{\mathcal{H}_1} S F = I H 1 , so F F F is Parseval. However, F F F is not positive, since both frame vectors are strictly positive. Every positive Parseval frame with two vectors has one positive and one negative entry. Up to the choice of signs, such a frame can be written as
G θ = { cos θ , − sin θ } , 0 < θ < π 2 . G_\theta=\{\cos\theta,-\sin\theta\},\qquad0<\theta<\frac{\pi}{2}. G θ = { cos θ , − sin θ } , 0 < θ < 2 π .
Therefore,
∥ F − G θ ∥ 2 2 = ( 1 2 − cos θ ) 2 + ( 1 2 + sin θ ) 2 = 2 − 2 ( cos θ − sin θ ) . \begin{aligned}
\lVert F-G_\theta\rVert_2^2=\left(\frac{1}{\sqrt{2}}-\cos\theta\right)^2+\left(\frac{1}{\sqrt{2}}+\sin\theta\right)^2 \\
&=2-\sqrt{2}(\cos\theta-\sin\theta).
\end{aligned} ∥ F − G θ ∥ 2 2 = ( 2 1 − cos θ ) 2 + ( 2 1 + sin θ ) 2 = 2 − 2 ( cos θ − sin θ ) .
Taking the infimum over 0 < θ < π / 2 0<\theta<\pi/2 0 < θ < π /2 gives
dist ( F , K ) = 2 − 2 > 0. \operatorname{dist}(F,\mathcal{K})=\sqrt{2-\sqrt{2}}>0. dist ( F , K ) = 2 − 2 > 0.
The infimum is approached as θ → 0 + \theta\to0^+ θ → 0 + , but it is not attained, since the limiting Parseval frame { 1 , 0 } \{1,0\} { 1 , 0 } is not positive. Thus, even when a frame is Parseval, its distance to the set of positive Parseval frames need not vanish.
Concluding Remarks
In this paper, we have investigated the geometric and operator-theoretic structure of positive frames and their duals in finite-dimensional real Hilbert spaces. Our approach emphasizes the role of positive linear dependence and the synthesis kernel in understanding positive spanning, frame equivalence, and redundancy. In particular, we have characterized equivalence classes of positive frames through their relation spaces, obtaining parametrizations in terms of the relative interior of a standard simplex and suitable subsets of Grassmannians.
A central part of our investigation concerns the existence and structure of positive dual frames. We have established that a finite frame admits a positive dual precisely when it is redundant and provided an explicit construction using a rank-one perturbation of the canonical dual. We have also examined the geometry of the collection of positive duals, including its relative openness and the absence of extreme points, and obtained criteria for the existence of dual frames satisfying prescribed positive relations. Furthermore, we have considered positive-basis duals and the boundary behavior of semi-positive frames.
We have also studied positive frames generated by finite group representations, highlighting the connections between balancedness, linear connectedness, and positive spanning. In addition, our investigation of normalized nonnegative relation spaces provides a convex-geometric description of positive dependencies. We have examined the robustness of positive spanning under vector erasures and its relationship with exactness and cardinality restrictions, together with approximation properties of positive Parseval frames.
Several questions remain open for further investigation. One natural direction is to extend the present framework to infinite-dimensional Hilbert spaces, where positive spanning and the existence of strictly positive relations require additional care. Another interesting problem is to investigate positive dual frames under further constraints, such as prescribed norms, sparsity, or optimal reconstruction error in the presence of erasures. It would also be worthwhile to explore positive frames generated by more general group actions and to understand how the geometry of their relation spaces interacts with the underlying representation. We hope that the results developed here provide a useful starting point for these investigations and further connections between frame theory, convex geometry, and constrained reconstruction.
Acknowledgment
The first author is grateful to the Mohapatra Family Foundation and the College of Graduate Studies at the University of Central Florida for supporting his postdoctoral fellowship during the course of this research. The second author acknowledges partial support from a Simons Travel Award.
Declaration
Ethical approval: Not applicable.
Data Availability Statement: Not Applicable.
Conflict of Interest: Not Applicable.
Clinical Trial: Not applicable.