arxiv: v1 [cs.it] 20 Oct 2014

Size: px

Start display at page:

Download "arxiv: v1 [cs.it] 20 Oct 2014"

Sydney Robinson
5 years ago
Views:

1 FRAMES FOR SUBSPACES OF C N MATTHEW HIRN, DAVID WIDEMANN arxiv: v1 [cs.it] 20 Oct 2014 Abstract. We present a theory of finite frames for subspaces of C N. The definition of a subspace frame is given and results analogous to those from frame theory for C N are proven. 1. Introduction Frames have been used in many applications of signal processing. They give stable signal representations and allow modelling for noisy environments. Recent work with hyperspectral data has shown the need to consider subspaces of high dimensional spaces. In particular, ongoing work has led us to investigate frames for these subspaces. 2. Frames for C N Let Φ = {ϕ j } s CN, where s N. Φ is a finite frame for C N if there exist constants A,B > 0 such that (1) A f 2 f,ϕ j 2 B f 2, f C N. The numbers A,B are called the frame bounds. It is a well known fact that any spanning set is a frame for C N, while every frame is indeed a spanning set. A frame is tight if one can choose A = B in the definition, i.e., if (2) f,ϕ j 2 = A f 2, f C N. Finally, a frame is unit norm if (3) ϕ j = 1, j = 1,...,s. If Φ satisfies (1), (2), and (3), then we say Φ is a finite unit norm tight frame (FUNTF) for C N. In this case, the frame bounds satisfy A = B = s/n. In particular, ifφis afuntf with framebounds A = B = 1, then Φis anorthonormalbasis. Assume now that Φ = {ϕ j } s is a frame for CN. The analysis operator of Φ is defined as follows: L : C N C s, Lf := { f,ϕ j } s. Date: November 19, Key words and phrases. Frames, frame potential. 1

2 2 M. HIRN, D. WIDEMANN The synthesis operator is given by: L : C s C N, L {c j } s = c j ϕ j. One obtains the frame operator by composing L with L: S : C N C N, Sf = L Lf = f,ϕ j ϕ j Some important properties of S are the following: (i) S is invertible and self-adjoint. (ii) Every f C N can be represented as (4) f = f,s 1 ϕ j ϕ j = f,ϕ j S 1 ϕ j. (iii) Φ is a tight frame if and only if S = AI. Based on equation (4), one defines the dual frame of Φ as Φ = { ϕ j } s := {S 1 ϕ j } s ; the frame operator of Φ is S 1. If Φ is a tight frame for C N, then S 1 = 1 AI, and the representation formula is simple: f = 1 f,ϕ j ϕ j = 1 f,ϕ j ϕ j. A A 3. Frames for Subspaces of C N Let Φ = {ϕ j } s CN and let W be a subspace of C N of dimension r < N. We say Φ is a finite subspace frame for W if span(φ) = W. It is clear from this definition that there exist constants A,B > 0 such that (5) A f 2 f,ϕ j 2 B f 2, f W. Wenotethat ifwehadinsteadused(5) asourdefinition, then itwouldnotnecessarily imply that span(φ) = W but ratherthat span(φ) W. The unit norm property as well as the notion of a tight frame remain similar in this setting. More specifically, if we can take A = B in (5) then we call Φ a tight subspace frame. Finally, if Φ is a finite unit norm tight subspace frame, then we say Φ is a subspace FUNTF. We define L, L, and S exactly the same as in section 1, however we note that the properties of these maps change for subspace frames. In particular, we see: (a) L : C N C s is no longer injective, but rather ker(l) = (C N \W) {0}. (b) L : C s C N is no longer surjective, but rather image(l ) = W. (c) Based on (a) and (b), we see that S : C N C N is no longer invertible. Because of (c), none of properties (i) - (iii) from section 1 hold for subspace frames. The question then becomes: in what sense do subspace frames satisfy properties (i) - (iii) above? Theorems below show that subspace frames satisfy natural modifications of the above properties.

3 FRAMES FOR SUBSPACES OF C N 3 Let W on be a set of r orthonormal vectors such that span(w on ) = W. We will also consider W on as an N r matrix where the columns of this matrix are the vectors in the set W on. We define Φ W to be the r s matrix whose columns are the coordinates of Φ in W on ; that is: (6) Φ W := W on Φ, where we have implicitly used the matrix form of Φ, that is the N s matrix whose columns are the elements of Φ. The j th column of Φ W is the projected W-subspace coordinates of Φ. Proposition 1. The set Φ W consisting of the columns of the matrix Φ W is a frame for C r. Proof. Since span(w on ) = W, we have ker(won) W = {0}. Therefore, since span(φ) = W as well, we see that Won Φ has rank r. We denote the analysis, synthesis, and frame operators of Φ W by L W, L W, and S W, respectively. In terms of the analysis operator, L, for Φ, L W = LW on. By proposition 1 we see that S W will satisfy (i) - (iii). Theorem 1. Φ is a subspace FUNTF for W with frame bound A if and only if Φ W is a FUNTF for C r with frame bound A. Proof. We do the forward direction first: let g C r, then: Therefore we have: S W g,g = L W g,l W g = Φ W on g,φ W on g = W on g,ϕ j 2 = A W on g 2 = A W on g,w on g S W g,g A W on g,w on g = 0 = S W g,g A W on W ong,g = 0 = g,(s W AI)g = 0 = S W = AI For the reverse direction, let f W. There exists g C r such that W on g = f. Therefore, A f 2 = A f,f = A W on g,w on g = Ag,g = S W g,g = W on L LW on g,g = LW on g,lw on g = Lf,Lf = f,ϕ j 2

4 4 M. HIRN, D. WIDEMANN We define the dual frame of Φ W in the usual way, that is Φ W = S 1 W Φ W. We now define the dual subspace frame of Φ as follows: (7) Φ := Won ΦW = W on S 1 W W onφ. As the name implies, the set Φ = { ϕ j } s = {W ons 1 W W on ϕ j} s following properties: will have the Proposition 2. Φ is a subspace frame for W. Proof. This follows from proposition 1. Theorem 2. Every f W can be represented as f = f, ϕ j ϕ j = f,ϕ j ϕ j. Proof. The first representation formula is Φ Φ f = f for all f W. Letting f = W on g for some g C r, we have: (8) Φ Φ f = Φ(W on S 1 W W on Φ) f = ΦΦ W on (S 1 W ) W on(w on g) = SW on S 1 W g = SW on (W onsw on ) 1 g Since W on W on is the identity on W, (8) = W on W onsw on (W onsw on ) 1 g = W on g = f The second representation formula is ΦΦ f = f for all f W. ΦΦ f = (W on S 1 W W onφ)φ f = W on (W on SW on) 1 W on SW ong = W on g = f The following commutative diagram illustrates the above ideas:

5 FRAMES FOR SUBSPACES OF C N 5 Φ subspace frame for W C N (subspace FUNTF for W C N ) W ons 1 W W on Φ subspace frame for W C N (subspace FUNTF for W C N ) W on W on Φ W frame for C r (FUNTF for C r ) S 1 W ΦW frame for C r (FUNTF for C r ) 4. Frame Potential Define the frame potential of a finite unit norm frame Φ = {ϕ j } s for CN as: FP(Φ) := k=1 ϕ j,ϕ k 2. In [1] a characterization of FUNTFs is given in terms of the frame potential: Theorem 3 (Benedetto and Fickus 2002). For a given N and s, let S N 1 denote the unit sphere in C N and consider: Then: FP : S N 1 S }{{ N 1 [0, ). } s times (1) Every local minimizer of the frame potential is also a global minimizer. (2) If s N, the minimum value of the frame potential is s, and the minimizers are precisely the orthonormal sequences in C N. (3) If s N, the minimum value of the frame potential is s 2 /N, and the minimizer are precisely the FUNTFs for C N. The following theorem is a trivial generalization of theorem 3: Theorem 4. For a given s and N, let W be a subspace of C N of dimension r < N and consider the resctricted frame potential: Then: FP W : (S N 1 S }{{ N 1 ) W [0, ). } s times (1) Every local minimizer of the restricted frame potential is also a global minimizer. (2) If s r, the minimum value of the restricted frame potential is s, and the minimizers are precisely the orthonormal sequences in W. (3) If s r, the minimum value of the restricted frame potential is s 2 /r, and the minimizer are precisely the subspace FUNTFs for W.

6 6 M. HIRN, D. WIDEMANN Theorem 4 shows that the minimum value of the frame potential depends on the dimension of the subspace W. Proof. Let W on be a set of r orthonormal vectors such that span(w on ) = W and consider it as an N r matrix. If Φ = {ϕ j } s is a finite unit norm set of vectors in W, then the coordinates of Φ in W on are given by the r s matrix Φ W = Won Φ. In [1] it is shown that FP(Φ) = Tr(S 2 ), where S is the frame operator of Φ. Using the previous two statements we then have: FP W (Φ) = Tr(S 2 ) = Tr([(W on Φ W )(W on Φ W ) ] 2 ) = Tr([Φ W Φ W] 2 ) = Tr(S 2 W ) = FP(Φ W ) Since Φ W is a finite unit norm set of vectors in C r, we can apply theorem 3 to get (1) and (2). Combining theorem 3 along with theorem 1 gives (3). References [1] John J. Benedetto and Matt Fickus. Finite normalized tight frames. Adv. Comput. Math., 18: , [2] J. Kovacevic and A Chebira. Life beyond bases: The advent of frames. IEEE SP Mag., Feature article. Norbert Wiener Center, University of Maryland, College Park, MD 20742

Preconditioning techniques in frame theory and probabilistic frames

Preconditioning techniques in frame theory and probabilistic frames Department of Mathematics & Norbert Wiener Center University of Maryland, College Park AMS Short Course on Finite Frame Theory: A Complete