The Finite Balian-Low Conjecture Mark Lammers and Simon Stampe Index Terms—Finite Gabor Systems, Zak Transform, BalianLow Theorem 0 1 0 0 0 0 T = 0 0 1 0 Abstract—We present a conjecture for a finite version of the celebrated Balian-Low Theorem for Gabor systems in L2 (R). We proceed to prove a special case of the conjecture for C9 . 0 1 0 0 0 1 .. . ··· ··· ··· .. . ··· ··· 0 0 0 0 0 0 0 , 1 0 I. I NTRODUCTION The Balian-Low Theorem for L2 (R) states that if g(x) ∈ L (R) and the critically sampled Gabor system {g(x − n)e2πimx }n,m∈Z produces an orthonormal basis, then either d d k dx g(x)k2 = ∞ or k dw ĝ(w)k2 = ∞ [1], [2], [3]. We will generalize this to the finite case but first we point out some immediate hurdles. We will consider finite Gabor systems for CN so the window function of the finite Gabor system will be a vector g in CN . It is natural to use a finite difference operator, D, to play the role of the derivative and the discrete Fourier transform, F, to play the role of the continuous Fourier transform. However, neither kDgk2 or kDFgk2 will ever be infinite because we are in a finite vector space. To get around this we will instead consider time-frequency measure kDgk2 + kDFgk2 and consider it’s growth as the dimension of the vector space grows. This is motivated by the GaussHermite differential equation, i.e., the time-independent d2 f Schrödinger equation with potential x2 , G(f ) = − 2 + dx x2 f = λf. Numerous authors have used variations on the discrete versions of the Gauss-Hermite differential equation [4], [5], [6], [7], [8], [9] to produce discrete versions of the Hermite functions for the DFT, i.e., eigenvectors for a difference operator corresponding to a discrete version of the differential operator G. In this proposition for a finite version of the Balian-Low Theorem we will conjecture what the minimizer of the measure kDgk2 +kDFgk2 is for orthonormal Gabor systems, and that the minimizer of this finite timefrequency measure goes to infinity as the dimension N of CN goes to infinity for orthonormal Gabor systems. 2 II. P RELIMINARIES First we present notation to generalize Gabor systems to CN . We use matrix versions of translation and modulation, Mark Lammers is with the Department of Mathematics, University of North Carolina at Wilmington, Wilmington, NC 28403, USA. email:lammersm@uncw.edu Simon Stampe is a medical student at SUNY Upstate Medical University. email:sstampe@gmail.com c 978-1-4673-7353-1/15/$31.00 2015 IEEE 1 0 0 M = 0 0 0 e2πi/N 0 0 0 0 0 e 4πi/N ··· ··· 0 0 0 .. . ··· ··· ··· .. . 0 0 0 0 0 e2πi(N −2)/N 0 0 e2πi(N −1)/N Next we define the circulant difference operator to be D = I − T , where I is the identity matrix, and we let F be 2πi the usual discrete Fourier transform matrix, Fn,m = (e N )mn m, n = 0 · · · N − 1. For simplicity of this note we will only be considering finite Gabor systems of vector spaces CN where N = d2 for some integer d. This allows us to take the same number of translations as modulations in producing the orthonormal Gabor systems. For clarity we note that Fd , Td , Md Id are the DFT transform, translation, modulation, and the identity operators for Cd . Similarly letting N = d2 , we define FN , TN , MN IN as the DFT transform, translation, modulation, 2 and the identity operators for CN . Letting g ∈ Cd for 2 some integer d, we define a finite Gabor system for Cd as d−1 md nd {TN MN g}m,n=0 . Because of the way we eventually define a finite Zak transform we now introduce a tensor product notation. We let ⊗ represent the Kronecker tensor product between two matrices. That is, if A is an m × n matrix and B is p × q matrix A ⊗ B is the mp × nq block matrix given by a11 B . . . a1n B .. .. A ⊗ B = ... . . an1 B ... ann B Let us note a few important properties of the Kronecker product: 1) (A ⊗ B)(C ⊗ D) = AC ⊗ BD. 2) (A ⊗ B)−1 = A−1 ⊗ B −1 . 3) (A ⊗ B)∗ = A∗ ⊗ B ∗ . 4) A ⊗ (B + C) = A ⊗ B + A ⊗ C. Using this notation we can write our finite Gabor systems d−1 md as {TNnd MN g}m,n=0 = {Tdn ⊗ Mdm g}d−1 m,n=0 but the real . reason we introduce it is to develop a notation for the finite Zak transform. The Zak transform, X Z(f (x)) (t, v) = f (t − n)e−2πinv , n∈Z 2 is a unitary map from L (R) to L2 ([0, 1]2 ). It is a tool that is used prominently in studying Gabor systems and has numerous nice properties with regard to them. In particular, for f ∈ L2 (R), Z(f (x − n))(t, v) = e2πint Z(f (x)) (t, v) N is divisible by four yields one repeated eigenvalue but this can be addressed (see section B of [16] for example). It follows that eigenvectors of G are also eigenvectors of F since F commutes with G. We refer to the columns of H as Hermite vectors and H itself as the Hermite Matrix. The eigenvector for G corresponding to the smallest eigenvector will be denoted g0 and will be referred to as the Gaussian vector. As expected, the Gaussian vector is an eigenvector of the DFT matrix F with corresponding eigenvalue 1, i.e., Fg0 = g0 . and Z(e2πimx f (x))(t, v) = e2πimv Z(f (x)) (t, v) . Perhaps most importantly for this note is the fact that if a critically sampled Gabor system {g(x − n)e2πimx }n,m∈Z produces an orthonormal basis, then |Z(g(x))(t, v)| = 1. A finite version of the Zak transform appears in [10], [11]. This representation maps a vector in Cnm to a Cn × Cm matrix, a natural extension of the continuous version which maps from L2 (R) to L2 ([0, 1]2 ). We will be using a version introduced in [12] which is simply a unitary operator from CN to itself, where N = d2 for some integer d. Define ZN = Fd ⊗ Id , ZN : CN → CN to be the N-dimensional −1 finite Zak transform. Note with this notation ZN = Z ∗. Additionally, with this notation the properties of original Zak transform can be expressed as: Theorem II.1. [10], [11], [12] Let the operators be defined as above. Then ZN TNd = (Md ⊗ Id )ZN , ZN MN = (Td∗ ⊗ 1 d = (Id ⊗ Md )ZN . Mdd )ZN and ZN MN Corollary II.2. [10], [11], [12] The finite Gabor system {Tdn ⊗ Mdm g}d−1 m,n=0 forms an orthonormal basis iff |(ZN g)(j)|2 = N1 . Corollary II.3. [10], [11] The finite Gabor system {Tdn ⊗ Mdm g}d−1 m,n=0 is a basis iff |(ZN g)(j)| > 0. III. F INITE H ERMITE F UNCTIONS Letting ∆ = D∗ D, we now define the finite Gauss-Hermite difference operator as G = F ∗ ∆F + ∆. It can be shown that F ∗ ∆F is a diagonal matrix and therefore acts like a multiplication operator and that the Fourier matrix commutes with G , FGF ∗ = G see [5], [13], [14], [6], [8]. We should note that the sign on our discrete Laplacian is positive instead of negative as it is in the differential operator. This is referred to the quantum repulsive operator in [15]. Both operators yield a diagonalization in the finite case ([13] Sec 6.2) but defining G as we have has several advantages from our prospective. All the eigenvalues of G are positive, G is invertible and it provides a nice connection with the finite time-frequency measure hGf, f i = kDf k2 + kDFf k2 that we will be using in our Finite Balian-Low Conjecture. The eigenvalues of G are unique, at least when N is not divisible by 4, so the eigenvectors of G are orthogonal and G has an orthogonal diagonalization of the form G = HΛH ∗ . The case when Fig. 1. Gaussian for N=100 from the matrix operations above. IV. Z AK TRANSFORM OF THE G AUSSIAN IN THE FINITE DOMAIN . It is well known that in the L2 (R) setting the Zak transform of the Gaussian function has a zero and it is precisely this that allows one to eventually conclude that not only do critically sampled Gabor systems not form an orthonormal basis, they do not form a basis at all. Things in our finite setting are a bit different. While it can be shown that if d is even the finite Zak transform of the Gaussian vector does indeed have a zero, one can compute odd cases where it does not. This is already a big difference from the L2 (R) case but as the dimension grows in the odd case the Jacobi theta type function, i.e., the Zak transform of the Gaussian [17], gets closer to having a 0 component. Therefore, the associated Gabor system will become very poorly conditioned. In the next section we will want to throw away the modulus of the components of the vector and only keep the phase, so we visualize the Zak transform of the Gaussians as a complex vector instead of the usual surface plot that only represents the modulus. Plotting Zak transforms of the Gaussian vectors as 3 dimensional − parameterized curves → r (n) = hn, real(f (n)), imag(f (n))i, in Fig.2. we see the even case intersects the black line representing real and imaginary parts of zero while the odd case does not. V. T HE C ONJECTURE . 2 Definition V.1. Let f ∈ Cd be a vector with no zero entries. Define the pointwize normalization of f as f (k) . J is highly non linear but it preserves J (f )(k) = d|f (k)| the phase of each entry of the vector and kJ (f )k = 1. From Corollary II.2 is is clear that the inverse Zak transforms of pointwise normalized functions will produce We are now ready for our Finite Balian-Low Conjecture. d−1 Conjecture V.2. Let {Tdn ⊗ Mdm f }m,n=0 be a Gabor system for CN , N = d2 that forms an orthonormal basis. 1) As d → ∞, kDf k2 + kDFf k2 → ∞ . ∗ 2) If d is odd, (ZN g0 )(j) 6= 0 and ZN J (ZN g0 ) is a 2 2 minimizer of kDf k + kDFf k . Fig. 2. Zak transform of Gaussian for N=100 and N=81 case. Fig. 3. Zak transform of Gaussian for N = 289 and it’s pointwise normalization. ∗ finite Gabor systems {Tdn ⊗ Mdm (ZN J (f ))}d−1 m,n=0 that are orthonormal bases. In fact, this is just the matrix version of −1/2 (g) in creating a window for a Parseval frame given by Sg 2 the L (R) case where Sg is the frame operator for the Gabor system {g(x − na)e2πimbx }n,m∈Z . ∗ J (f ). We have already Let’s consider a special case of ZN observed that when d is odd ZN (g0 ) does not have to have ∗ a zero so below we consider the graph of ZN J (Zn g0 ) for N = 289 along with it’s Fourier transform. VI. S PECIAL C ASE OF BALIAN -L OW C ONJECTURE FOR EIGENVECTORS OF F WITH N = 9. In this section we will prove the conjecture is true for N = 9 and the window function of the orthonormal Gabor system {Tdn ⊗Mdm g}d−1 m,n=0 is an eigenvector of the Fourier transform. Assuming that the window function is an eigenvector of the Fourier transform, it is natural condition to consider given that the Gaussian vector is the minimizer of the time-frequency measure over all norm one vectors. Additionally, it is natural to assume that the minimum occurs when kDf k = kDFf k and any eigenvector of F satisfies this condition. For the sake of simplicity we let Z9 = Z and F9 = F for the rest of this section. Theorem VI.1. Let N = d2 = 9 and let {Tdn ⊗ Mdm g}d−1 m,n=0 be a Gabor system that forms an orthonormal basis. If g is a an eigenvector of F, then Z ∗ J (Zg0 ) is a minimizer of kDf k2 + kDFf k2 . Proof. First we compute J (Zg0 ) to get (J (Zg0 ))T = i 1h 1 1 1 1 eπi/9 e5πi/9 1 e−πi/9 e−5πi/9 . 3 Now we let g be an eigenvector of F and Θ = dZg. If the Gabor system with window g generates an orthonormal basis, the modulus of each component of Θ is one. To complete our proof it is enough to show that there is a minimizer of the time-frequency measure g satisfying the conditions above so that Zg = J (Zg0 ). We denote the components of Θ by ΘT = eiθ0 eiθ1 eiθ2 eiθ3 eiθ4 eiθ5 eiθ6 eiθ7 eiθ8 . Because g is an eigenvector of F and all eigenvalues are modulus 1, the time frequency measure becomes kDgk2 + kDF gk2 = 2kDgk2 2 = 4 − hZ(T + T ∗ )Z ∗ Θ, Θi. 9 −1 Fig. 4. ZN J (ZN g0 ) and it’s Fourier transform for N = 289. It should be noted that in the example above only the real parts of these vectors were plotted because the imaginary part was of magnitude less than 10−15 . It should also be noted that in this example −1 −1 max(FN ZN J (ZN g0 ) − ZN J (ZN g0 )) < 1.8 · 10−14 giving strong indication that this is also an eigenvector of the Fourier transform corresponding to eigenvalue 1. A straightforward computation results in hZ(T + T ∗ )Z ∗ Θ, Θi = ei(θ1 −θ0 ) + ei(θ2 −θ0 ) + ei(θ0 −θ1 ) + ei(θ2 −θ1 ) + ei(θ0 −θ2 ) + ei(θ1 −θ2 ) + ei(θ4 −θ3 ) + ei(θ0 −θ3 −2π/3) + ei(θ3 −θ4 ) + ei(θ5 −θ4 ) + ei(θ3 −θ5 +2π/3) + ei(θ4 −θ5 ) + ei(θ7 −θ6 ) + ei(θ8 −θ6 +2π/3) + ei(θ6 −θ7 ) + ei(θ8 −θ7 ) + ei(θ6 −2π/3−θ8 ) + ei(θ7 −θ8 ) . Now combining terms above we get: hZ(T + T ∗ )Z ∗ Θ, Θi = 2(cos(θ1 − θ0 ) + cos(θ0 − θ2 ) + cos(θ2 − θ1 ) + cos(θ4 − θ3 ) + cos(θ5 − θ3 − 2π/3) + cos(θ5 − θ4 ) + cos(θ7 − θ6 ) + cos(θ6 − θ8 − 2π/3) + cos(θ8 − θ7 )). (1) Let P = ZF Z ∗ then Θ = dZg is an eigenvector of P with the same eigenvalue since ZF g = P Zg implying that λZg = P Zg. Applying P to Θ we get eiθ0 eiθ3 eiθ6 eiθ2 i(θ −2π/9) 5 PΘ = ei(θ −4π/9) e 8 eiθ1 ei(θ4 −π/9) ei(θ7 −2π/9) = λΘ First we notice that the operator P fixes the first component. So eiθ0 = λeiθ0 , which implies the eigenvalue λ must be 1 since eiθ0 6= 0. Now matching up the terms we get the following relations: θ0 θ1 θ4 θ5 θ7 θ8 = θ0 = θ2 = θ3 = θ6 = θ4 = θ4 + 4π/9 = θ4 − 2π/9 = θ4 − 6π/9. Fig. 5. Plot of 4 − 49 [3 + 2 cos(u) + 2 cos(u + 2π/9) + 2 cos(4π/9)]. We observe that this is the same vector we obtained calculating J (Zg0 ) . Remark VI.2. We point out that the argument used in the proof above to show that the λ = 1, generalizes to all dimensions because in [11] they show the matrix P is the product of a permutation matrix and a unitary diagonal matrix that always keeps the first component fixed. That is, d−1 if {Tdn ⊗ Mdm g}m,n=0 is a Gabor system that forms an orthonormal basis and g is a an eigenvector of F, then the eigenvalue must be 1. VII. N UMERICAL R ESULTS Using MATLAB we calculate the time-frequency measure ∗ kDN gk2 + kDN FN gk2 for g = ZN J (Zg0 ) for odd dimensions of d. Using a logarithmic regression we fit these points to the curve 6.08 ln(d) + 0.63 with a coefficient of determination r2 = 0.9996 in Fig. 6 below. Substituting the equations above into equation 1 we get hZ(T + T ∗ )Z ∗ Θ, Θi = 2[2cos(θ1 − θ0 ) + 2cos(θ4 − θ1 ) + cos(θ4 − θ1 − 2π/9) + 2cos(4π/9) + 1], and so the measure becomes kDgk2 + kDF gk2 4 = 4 − [2 cos(θ1 − θ0 ) + 2 cos(θ4 − θ1 ) 9 + 2 cos(θ4 − θ1 − 2π/9) + 2 cos(4π/9) + 1]. Fig. 6. The time-frequency measure kDN gk2 + kDN FN gk2 of g = Z ∗ J (Zg0 ) as a function of d (black squares). Graph of 6.08 ln(d) + 0.63 (blue dashes). We observe that θ0 only appears once in the expression so the relation θ0 = θ1 must occur in any minimization. Letting u = θ4 − θ1 we get the single variable expression: R EFERENCES kDgk2 + kDF gk2 4 = 4 − [3 + 2 cos(u) + 2 cos(u − 2π/9) + 2 cos(4π/9)] 9 which has critical points when u = π9 + nπ and a minimum when u = π9 . Choosing θ1 = 0 gives us that θ4 = 0 and this gives us the minimizer we were seeking: Zg T = i 1h 1 1 1 1 eπi/9 e5πi/9 1 e−πi/9 e−5πi/9 . 3 [1] R. Balian, “Un principe d’incertitude fort en théorie du signal ou en mécanique quantique,” C. R. Acad. Sci. Paris Sér. II Méc. Phys. Chim. Sci. Univers Sci. Terre, vol. 292, no. 20, pp. 1357–1362, 1981. [2] F. Low, “Complete sets of wave packets,” in In C. Detar,editor, A passion for Pysics-Essay in honor of Geoffery Chew, pp. 17–22, World Scientific, Singapore, 1985. [3] J. J. Benedetto, C. Heil, and D. F. Walnut, “Differentiation and the balian-low theorem,” Journal of Fourier Analysis and Applications, vol. 1, no. 4, pp. 355–402, 1994. [4] F. A. Grünbaum, “The eigenvectors of the discrete Fourier transform: a version of the Hermite functions,” J. Math. Anal. Appl., vol. 88, no. 2, pp. 355–363, 1982. [5] B. Dickinson and K. Steiglitz, “Eigenvectors and functions of the discrete Fourier transform,” IEEE Trans. Acoust. Speech Signal Process., vol. 30, no. 1, pp. 25–31, 1982. [6] J. McClellan and T. Parks, “Eigenvalue and eigenvector decomposition of the discrete Fourier transform,” IEEE Trans. Audio Electroacoust., vol. AU-20, no. 1, pp. 66–74, 1972. [7] M. L. Mehta, “Eigenvalues and eigenvectors of the finite Fourier transform,” J. Math. Phys., vol. 28, no. 4, pp. 781–785, 1987. [8] B. Santhanam and T. Santhanam, “On discrete gauss hermite functions and eigenvectors of the discrete fourier transform,” Signal Processing, pp. 2738–2746, 2008. [9] G. Strang, “The discrete cosine transform,” SIAM Rev., vol. 41, no. 1, pp. 135–147 (electronic), 1999. [10] M. An, A. K. Brodzik, and R. Tolimieri, Ideal sequence design in timefrequency space. Applied and Numerical Harmonic Analysis, Boston, MA: Birkhäuser Boston Inc., 2009. [11] A. K. Brodzik and R. Tolimieri, “Gerchberg-papoulis algorithm and the finite zak transform,” Proceedings of SPIE, vol. 4119, pp. 1084–1093, 2000. [12] M. C. Lammers, “The finite fractional Zak transform,” IEEE Signal Processing Letters, vol. 21, pp. 1064–1067, 2014. [13] M. A. K. Haldun M. Ozaktas, Zeev Zalevsky, The Fractional Fourier Transform: with Applications in Optics and Signal Processing. Wiley, 2001. [14] M. Lammers and A. Maeser, “An uncertainty principle for finite frames,” J. Math. Anal. Appl., vol. 373, no. 1, pp. 242–247, 2011. [15] K. B. Wolf, Integral Transforms in Science and Engineering. Applied and Numerical Harmonic Analysis, New York: Plenum Press, 1979. [16] Ç. Candan, M. A. Kutay, and H. M. Ozaktas, “The discrete fractional Fourier transform,” IEEE Trans. Signal Process., vol. 48, no. 5, pp. 1329–1337, 2000. [17] C. Heil, “A discrete Zak transform,” tech. rep., The MITRE Corporation, 1989.