Weaving Properties of Hilbert Space Frames Peter G. Casazza Richard G. Lynch Director: The Frame Research Center Department of Mathematics University of Missouri Columbia, MO, 65211 Website: http://www.framerc.org Department of Mathematics University of Missouri Columbia, MO 65211 Email: rilynch37@gmail.com Abstract—We will prove some new results in the theory of Weaving Frames. Two frames {ϕi }i∈I and {ψi }i∈I in a Hilbert space H are woven if there are constants 0 < A ≤ B so that for every subset σ ⊂ I, the family {ϕi }i∈σ ∪ {ψi }i∈σc is a frame for H with frame bounds A, B. We begin by introducing the main results in weaving frames. We then prove some new basic properties. This is followed by showing a fundamental connection between frames and projections, providing intuition on woven frames. Finally, a weaving equivalent of an unconditional basis for weaving Riesz bases is considered. real Hilbert space and I can represent a finite or countably infinite index set. I. I NTRODUCTION where A and B are the lower Riesz bound and upper Riesz bound, respectively. This note focuses on an intriguing area of research called weaving frames [1]. Two frames {ϕi }i∈I and {ψi }i∈I for a Hilbert space H are woven if there are constants 0 < A ≤ B so that for every subset σ ⊂ I, the family {ϕi }i∈σ ∪ {ψi }i∈σc is a frame for H with frame bounds A and B. A potential application of weaving frames together is dealing with wireless sensor networks which may be subjected to distributed processing under different frames. The theory could also have use in the preprocessing of signals using Gabor frames. In this paper, we review fundamental properties in weaving frames from [1], followed by some new basic properties. A relation of frames to projections is then considered and gives a better understanding of what it really means for two frames to be woven. We conclude by showing a weaving equivalent of an unconditional basis. We note that there is another concept in the literature which uses multiple frames called Quilted Frames introduced by M. Dörfler [4], [5], and are seemingly unrelated to woven frames. Quilted Gabor frames are systems constructed from globally defined frames by restricting these to certain, possibly compact, regions in the time-frequency or time-scale plane. II. F RAME T HEORY P RELIMINARIES A brief introduction to frame theory is given in this section, which contains the necessary background for this paper. For a thorough approach to the basics of frame theory, see [2], [3]. Unless otherwise noted, H will denote either a finite or infinite The authors were supported by NSF 1307685; NSF ATD 1042701; NSF ATD 1321779; AFOSR DGE51: FA9550-11-1-0245. c 978-1-4673-7353-1/15/$31.00 2015 IEEE Definition 1. A family of vectors Φ = {ϕi }i∈I in H is said to be a Riesz basis if there are constants 0 < A ≤ B < ∞ so that for all {ci }i∈I ∈ `2 (I), 2 X X X 2 ci ϕi |ci |2 A |ci | ≤ ≤B i∈I i∈I i∈I Riesz bases have proved to be very useful in some applications in which the assumption of orthonormality is too extreme. Similar to an orthonormal basis, Riesz bases satisfy uniqueness of a decomposition as well as stability. There are times when assuming the sequence is a Riesz basis is even too strong. In these cases we work with frames, which are redundant families of vectors that have proper subsets spanning the space. Redundancy is the fundamental property of frames which makes them so useful in practice. Definition 2. A family of vectors Φ = {ϕi }i∈I in H is said to be a frame if there are constants 0 < A ≤ B < ∞ so that for all x ∈ H, X Akxk2 ≤ |hx, ϕi i|2 ≤ Bkxk2 , i∈I where A and B are a chosen lower frame bound and upper frame bound, respectively. If only B is assumed, then it is called a B-Bessel sequence. If A = B, it is said to be an A-tight frame and if A = B = 1, it is a Parseval frame. The values {hx, ϕi i}i∈I are called the frame coefficients of the vector x ∈ H with respect to the frame Φ. If Φ = {ϕi }i∈I is a sequence in H, then the analysis operator of Φ is the operator T : H → `2 (I) given by T x := {hx, ϕi i}i∈I and the associated synthesis operator is given by the adjoint operator T ∗ : `2 (I) → H and satisfies X T ∗ {ci }i∈I := ci ϕi . i∈I The frame operator S : H → H is defined by S := T ∗ T and satisfies X Sx = T ∗ T x = hx, ϕi iϕi i∈I for any x ∈ H. These three operators are well-defined when the sequence Φ is assumed to be at least a Bessel sequence. If Φ is a frame with upper and lower bounds A and B, respectively, then the frame operator is a positive, self-adjoint, and invertible operator that also satisfies for any x ∈ H, X hAx, xi ≤ hSx, xi = kT xk2 = |hx, ϕi i|2 ≤ hBx, xi, i∈I and hence operator inequality A · Id ≤ S ≤ B · Id holds. Also, note that {S −1/2 ϕi }i∈I is a Parseval frame, called the canonical Parseval frame of Φ. Finally, the norm of S is kSk = kT ∗ T k = kT k2 . III. W EAVING F RAMES P RELIMINARIES This section is dedicated to a brief introduction to weaving frames and is grounded in reviewing results from [1]. We begin with the formal definition in full generality. Definition 3. A finite family of frames {ϕij }M j=1,i∈I in H is said to be woven if there are universal constants A and B so M that for every partition {σj }M j=1 of I, the family {ϕij }j=1,i∈σj is a frame for H with lower and upper frame bounds A and B, respectively. Each family {ϕij }M j=1,i∈σj is called a weaving. The overall goal is to discover conditions for which a family of frames is woven. Proposition 3.1 from [1] shows one does not need to check for a universal upper bound, as it is always given by the sum of the upper frame bounds. A natural follow-up question is whether one must check for a universal lower bound. The answer is obviously no in the finite dimensional case since there are only finitely many ways to partition the index set and thus a universal bound is easily obtained. However, in the infinite dimensional setting it is not immediately clear that the lower frame bounds of the weavings do not tend to zero. To show that a univeral bound must be obtained, we introduce a weaker form of weaving. {ϕij }M j=1,i∈I Definition 4. A family of frames in H is said to be weakly woven if for every partition {σj }M j=1 of I, the family {ϕij }M is a frame for H. j=1,i∈σj One of the main theorems in [1] (Theorem 4.5) proves that weakly woven is equivalent to the frames being woven. Definition 6. If W1 and W2 are nontrivial subspaces of H, define dW1 (W2 ) := inf{kx − yk : x ∈ W1 , y ∈ SW2 } where SW2 = SH ∩ W2 and SH is the unit sphere in H. Similarly define dW2 (W1 ) by swapping the roles of W1 and W2 . The distance between W1 and W2 is defined as d(W1 , W2 ) := min{dW1 (W2 ), dW2 (W1 )}. The following theorem, Theorem 5.7 of [1], gives an important relationship in weaving Riesz bases with this distance. ∞ Theorem 7. If Φ = {ϕi }∞ i=1 and Ψ = {ψi }i=1 are Riesz bases in H, then the following are equivalent: (i) Φ and Ψ are woven. (ii) For any σ ⊂ N, d span{ϕi }i∈σ , span{ψi }i∈σc > 0. (iii) There is a constant D > 0 so that for any σ ⊂ N, d span{ϕi }i∈σ , span{ψi }i∈σc ≥ D. This will later be generalized to frames by using projections instead of distance. Perturbations were also considered in [1]. Intuitively, a frame and a small perturbation of itself should be woven as the next proposition confirms. Proposition 8. [1] If Φ = {ϕi }i∈I is a frame with bounds A and B, and F is an invertible operator satisfying A , B are woven. kId − F k2 < then {ϕi }i∈I and {F ϕi }i∈I This result extends easily to the case of finitely many operators with minor adjustments. Proposition 9. Let {ϕi }i∈I be a frame for H and {Fj }nj=1 invertible operators on H with r n X A kId − Fj k < . B j=1 Then {ϕi }i∈I and {Fj ϕi }nj=1,i∈I are woven. To conclude this section, we note that the whole theory of weaving frames was born out of the idea preprocessing of signals using Gabor frames. However, we were unable to answer the following problem, which still remains open. Here Tam is the shift operator by a factor of am and Mbn is modulation operator by a factor of bn. ∞ Theorem 5. Given two frames {ϕi }∞ i=1 and {ψi }i=1 for H the following are equivalent: (i) The two frames are woven. (ii) The two frames are weakly woven. Problem 10. Given a fixed lattice generated by a, b > 0 with 2 ab < 1 and rotated Gaussians Uj gαj , where gαi (x) = e−αi x , are the Gabor frames {Tam Mbn Uj gαj }M j=1,m,n∈Z woven? Therefore, it only needs to be checked that each weaving is a frame, possibly each having different lower frame bounds. Another direction is to study the special case when each frame is a Riesz basis. It turns out that weaving Riesz bases is classified by the following notion of distance. In this section, we prove some new basic properties in the theory of weaving frames. We begin by showing that an invertible operator applied to woven frames leaves them woven. However, applying an operator to only one of the IV. W EAVING F RAMES frames will not. We then show that multiplying the frame vectors of two woven frames by uniformly bounded constants gives woven frames. Afterwards, we see if the weaving property may be checked on smaller index sets and look at the repercussions of deleting frame vectors. Finally, we prove a relationship between the norms of the frame operators of the original frames and the frame operators of the weavings. Proposition 11. Suppose {ϕij }M j=1,i∈I is a woven family of frames for H with common frame bounds A and B. If F is an invertible operator on H, then {F ϕij }M j=1,i∈I is also woven with bounds AkF −1 k−2 and BkF k2 . In particular, the bounds do not change if F is unitary. Proof. It is a known fact that if a frame has bounds A and B, then applying an invertible operator F to it gives a frame with bounds AkF −1 k−2 and BkF k2 . Since the sequence {ϕij }M j=1,i∈σj is a frame with lower and upper bounds A and B, respectively, for any partition {σj }M j=1 of I, then the sequence {F ϕij }M is a frame with bounds AkF −1 k−2 j=1,i∈σj 2 M and BkF k . That is, {F ϕij }j=1,i∈I is woven with universal bounds AkF −1 k−2 and BkF k2 . Remark 12. Proposition 11 can be relaxed to a bounded operator F with closed range if F −1 is replaced with F † [3, Prop 5.3.1]. The next result gives that a frame and a nonidentical reordering of itself may not be woven. Proposition 13. If {ϕi }i∈I is a Riesz basis with Riesz bounds A, B and π is a permutation of I, then for every σ ⊂ I the family {ϕi }i∈σ ∪{ϕπ(i) }i∈σc is a frame sequence with bounds A and 2B. However, {ϕi }i∈I and {ϕπ(i) }i∈I are woven if and only if π = Id. Proof. For any x ∈ span {ϕi }i∈σ ∪ {ϕπ(i) }i∈σc and for any σ ⊂ I, X X X |hx, ϕi i|2 + |hx, ϕπ(i) i|2 ≥ |hx, ϕi i|2 i∈σ i∈σ c σ∪(σ c ∩π(σ c )) 2 ≥ Akxk , since any subsequence of a Riesz basis is a Riesz sequence with the same bounds. The upper frame bound is the sum of the upper frames bounds, which is 2B. Note that it is not B due to redundancy. The however part is now proven via contraposition. Assume π 6= Id so that π(i0 ) = j0 6= i0 for some i0 , j0 ∈ I. Let σ = I\{i0 }. Then {ϕi }i∈σ ∪ {ϕπ(i) }i∈σc = {ϕi }I\{i0 } ∪ {ϕj0 } which is the set in which ϕj0 appears twice, but ϕi0 does not appear at all and therefore the closure of the span is not the entire space. Since a frame is always woven with a copy of itself, the previous proposition implies that applying an operator to only one of the woven frames may not leave woven frames. The following proposition gives conditions on multiplying the frame vectors by individual constants and still be left with woven frames. Proposition 14. If {ϕi }i∈I and {ψi }i∈I are woven with universal lower and upper weaving bounds A and B, respectively, and 0 < C ≤ |ai |2 , |bi |2 ≤ D < ∞ are constants, then {ai ϕi }i∈I and {bi ψi }i∈I are also woven, but with lower and upper bounds AC and BD, respectively. In particular, 1 1 √ ϕi √ ψi and A A i∈I i∈I are woven with lower and upper bounds 1 and B/A, respectively. Proof. A simple calculation yields this. The next result gives that weavings may possibly be checked on smaller index sets than the original. Proposition 15. If J ⊂ I and {ϕi }i∈J and {ψi }i∈J are woven frames then {ϕi }i∈I and {ψi }i∈I are woven. Proof. For any σ ⊂ I and any x ∈ H we have X X Akxk2 ≤ |hx, ϕi i|2 + |hx, ψi i|2 i∈σ c ∩J i∈σ∩J ≤ X X 2 |hx, ϕi i| + |hx, ψi i|2 . i∈σ c i∈σ Since the upper bound is always given, this implies {ϕi }i∈I and {ψi }i∈I are woven. Hence, adding vectors {ϕ0j }j∈J and {ψj0 }j∈J (even letting = 0 for all j ∈ J) to two woven frames still leaves two woven frames. Therefore, asking for woven frames to have some “property” for all subsets is not achievable. It is possible to remove vectors from woven frames and still be left with woven frames as the next result shows. ϕ0j Proposition 16. Suppose {ϕi }i∈I and {ψi }i∈I are woven with universal constants A and B. If J ⊂ I and X |hx, ϕi i|2 ≤ Dkxk2 i∈J for some 0 < D < A and for all x ∈ H, then {ϕi }i∈I\J and {ψi }i∈I\J are also frames for H and are woven with universal lower and upper frame bounds A − D and B, respectively. Proof. The fact that B is an upper weaving bound is obvious. Suppose that σ ⊂ I\J. Then for all x ∈ H, X X |hx, ϕi i|2 + |hx, ψi i|2 i∈σ i∈(I\J)\σ = X |hx, ϕi i|2 − X i∈σ∪J + X i∈J 2 |hx, ψi i| i∈(I\J)\σ ≥ (A − D)kxk2 |hx, ϕi i|2 so that a lower weaving bound is A − D. Taking σ = J c and σ = ∅ gives that {ϕi }i∈I\J and {ψi }i∈I\J are frames, respectively. Proof. Let (Tj )σj be the analysis operator of Φj restricted to the sum over σj . For any x ∈ H, 2 X X X (Sj )σj x2 = hx, ϕij iϕij An immediate corollary considers the case of when the frame is woven with itself and hence vectors may possibly be removed to be left with a frame. j∈J j∈J j∈J Corollary 17. If {ϕi }i∈I is a frame with lower frame bound A and X |hx, ϕi i|2 ≤ Dkxk2 ≤ The following example is an application of Proposition 16. Example 18. Let {ei }∞ i=1 be an orthonormal basis for H. Consider the two frames 100 100 e2 , . . . , e i , ei , . . .} Φ = {ϕi }∞ i=1 = {e1 , 100e1 , e2 , 2 i Ψ = {ψi }∞ i=1 = {e1 , e1 , e2 , e2 , . . . , ei , ei , . . . , }. These two frames are (1, 101)-woven, and it is clear that the even terms in each frame may be removed to still leave two woven frames. However, J = 2N does not satisfy Proposition 16. On the other hand, > 0 can be chosen so that Φ = {ϕi, }∞ i=1 100 100 e2 , . . . , e i , ei , . . .} 2 i and Ψ are still woven with lower woven bound one, but X |hx, ϕi, i|2 ≤ Dkxk2 = {e1 , 100e1 , e2 , i∈2N for some 0 < D < 1 and thus the proposition applies showing that the even terms are not necessary to determine weaving. Finally, the next proposition gives a relationship between the norms of the frame operators of the original frames and the weavings. Proposition 19. Let {Φj }j∈J be a (finite or infinite) collection of frames Φj = {ϕij }i∈I for H with respective frame operators Sj . Suppose there are constants A, B > 0 so that for any partition {σj }j∈J of I, the collection Ψ = {ϕij : i ∈ σj , j ∈ J} is a frame for H with lower and upper frame bounds, A and B, respectively. That is, the frames {Φj }j∈J0 are woven for any finite J0 ⊂ J (by choosing σj = ∅ for all j ∈ J\J0 ). If SΨ represents the frame operator of Ψ, then for any x ∈ H, X (Sj )σj x2 ≤ B SΨ 2 , A j∈J where (Sj )σj denotes the frame operator Sj with sum restricted to σj . X j∈J B X |hx, ϕij i|2 i∈σj = BhSΨ x, xi. i∈J for some 0 < D < A and for all x ∈ H, then {ϕi }i∈J c is a frame with lower frame bound A − D. i∈σj X (Tj∗ )σj (Tj )σj x2 = Now using the inequality A·Id ≤ SΨ , where Id is the identity operator, gives X (Sj )σj x2 ≤ BhSΨ x, xi = BhS 1/2 x, S 1/2 xi Ψ Ψ j∈J ≤B 1 1/2 1/2 SΨ SΨ x, SΨ x A = B Sψ 2 A as desired. V. W EAVINGS AND P ROJECTIONS This section is devoted to developing an important characterization of frames with projections. An immediate corollary gives some intuition on what it means to weave two frames together. Theorem 20. Given vectors {ϕi }i∈I in H, the following are equivalent: (i) {ϕi }i∈I is a frame. (ii) For any σ ⊂ I, let Wσ = span{ϕi }i∈σ , Wσc = span{ϕi }i∈σc , and let P be the orthogonal projection of H onto Wσ⊥ . Then P |Wσc is onto Wσ⊥ and {P ϕi }i∈σc is a frame for Wσ⊥ . Proof. The fact that (ii) ⇒ (i) is immediate by considering σ = ∅. To prove (i) ⇒ (ii), note that {P ϕi }i∈I is a frame for Wσ⊥ . However, P ϕi = 0 for all i ∈ σ and thus {P ϕi }i∈σc is a frame for Wσ⊥ . Let T1 : `2 (σ c ) 7→ Wσc and T : `2 (σ c ) 7→ Wσ⊥ be the operators defined by T1 ei := ϕi and T ei := P ϕi for i ∈ σ c , where {ei }i∈σc is the standard orthonormal basis. It must be checked that P |Wσc is onto. However, note that T is onto since {P ϕi }i∈σc is a frame and that T = P |Wσc T1 . Hence, P |Wσc must also be onto, concluding the proof. Corollary 21. Given two frames Φ = {ϕi }i∈I and Ψ = {ψi }i∈I for H, the following are equivalent: (i) The frames Φ and Ψ are woven. (ii) For any σ ⊂ I, let Wσ = span{ϕi }i∈σ , Wσc = span{ψi }i∈σc , and let P be the orthogonal projection of H onto Wσ⊥ . Then P |Wσc is onto Wσ⊥ and {P ϕi }i∈σc is a frame for Wσ⊥ . VI. U NCONDITIONAL C ONSTANTS OF W EAVING R IESZ BASES In this last section, we consider the weaving equivalent of an unconditional basis for H. Recall that {xn }∞ n=1 is an unconditional sequence in H if and only if there is a constant C > 0 so that for all σ ⊂ HN and for all scalars {an }∞ n=1 we have X X X X ∞ a x a x + = C ≤ C a x a x i i i i i i i i i∈σ c i∈σ i∈σ i=1 ∞ Theorem 22. Let Φ = {ϕi }∞ i=1 and Ψ = {ψi }i=1 be Riesz basic sequences in H with frame bounds A1 , B1 and A2 , B2 respectively. The following are equivalent: (i) There are constants 0 < A ≤ B so that for every σ ⊂ N the family {ϕi }i∈σ ∪ {ψi }i∈σc is a Riesz basis with bounds A, B. That is, Φ and Ψ are woven. (ii) There is a constant 0 < C satisfying for all scalars {ai }∞ i=1 and all σ ⊂ N Hence, 1 C 2C +1 ! X X ai ψi ai ϕi + i∈σ c i∈σ ( ) X X C ≤ ai ψi max ai ϕi , C +1 i∈σ c i∈σ X X a i ϕi + ai ψi ≤ . i∈σ c i∈σ P (iv) ⇒ (ii): If i∈σ ai ϕi = 0 we are done. So assume not and by (iv) we have: X X 1 P E≤ a ϕ + a ψ i i i i . aϕ i∈σ i i i∈σ c i∈σ So (ii) holds. At this point we know that (ii) ⇔ (iii) ⇔ (iv). Now (i) ⇒ (ii): Given σ and {ai }∞ i=1 , by (i) we have 2 X X 2 ai ϕi |ai |2 ≤ B i∈σ i∈σ ! X X X ai ϕi + a i ψi ai ϕi ≤ C . (iii) There is a constant 0 < D satisfying for all scalars {ai }∞ i=1 and all σ ⊂ N X X ai ψi a i ϕi + D i∈σ c i∈σ X X a ϕ + a ψ ≤ i i i i . i∈σ c i∈σ (iv) There is a constant 0 < E satisfying scalars P for all {ai }∞ and all σ ⊂ N so that if a ϕ = 1 i=1 i∈σ i i then X X E≤ ai ϕi + a i ψi . i∈σ c i∈σ Proof. The implications (iii) ⇒ (ii), and (ii) ⇒ (iv) are clear. (ii) ⇒ (iii): Given the assumptions in (ii) we compute: X X X X a i ψi ≤ ai ϕi + ai ψi + a i ϕi i∈σ c i∈σ c i∈σ i∈σ X X ≤ ai ϕi + ai ψi i∈σ c i∈σ X X 1 ai ϕi + ai ψi + . C i∈σ i∈σ c So, X X X C ai ψi ≤ ai ϕi + ai ψi . C + 1 i∈σc i∈σ i∈σ c |ai |2 + i∈σ i∈σ c i∈σ i∈σ ≤ B2 X X |ai |2 i∈σ c X 2 X B2 a i ϕi + ai ψi . 2 A i∈σ i∈σ c ≤ Finally we prove (iii) ⇒ (i): For every sequence of scalars {ai }∞ i=1 and every σ ⊂ N we have: ∞ X i=1 |ai |2 = X |ai |2 + X |ai |2 i∈σ c i∈σ X X 1 1 ai ϕi + ai ψi ≤ A1 i∈σ A2 i∈σc ! X X 1 1 ai ψi ≤ max , ai ϕi + A1 A2 i∈σ c i∈σ X X 1 1 1 ≤ max , a ϕ + ai ψi i i D A1 A2 i∈σ i∈σ c proving the lower bound. The upper bound is obvious. This completes the proof of the theorem. R EFERENCES [1] T. Bemrose, P.G. Casazza, K. Gröchenig, M.C. Lammers, R.G. Lynch, Weaving Hilbert Space Frames, Preprint. [2] P.G. Casazza and G. Kutyniok, Finite frames: Theory and applications, Springer (2013). [3] O. Christensen, An Introduction to Frames and Riesz Bases, Birkhaüser, Boston (2003). [4] M. Dörfler, Quilted Gabor frames - a new concept for adaptive timefrequency representation, Adv. Appl. Math 47 No. 4 (2011) 668-687. [5] M.Dörfler, Frames adapted to a phase-space cover, Constr. Approx. 39 No. 3 (2014) 445-484.