Poisson summation and the discrete Fourier transform John Kerl University of Arizona, Math 523B April 14, 2006

Transcription

1 Poisson summation and the discrete Fourier transform John Kerl University of Arizona, Math 523B April 4, 2006

2 In Euclidean space, given a vector,... 2

3 ... we can put down a coordinate frame (say an orthonormal one). We can decompose the vector into its components, or reconstruct the vector from its components

4 Decomposition (how much in each direction) in R n : let {u,..., u n } be an orthonormal basis. Our vector is v: v = n j= c j u j How do we find the coefficients? Since the basis vectors are orthonormal, u i, u j = δ ij. Just dot each basis vector with v: u i, v = u i, n j= c j u j = n j= c j u i, u j = c i. 4

5 Recall that in C n, we must conjugate one argument (by our convention, the first) in order to get positive definiteness of the inner product: v, w = n j= v j w j. 5

6 Consider the space of periodic continuous functions from the interval [0, ] (or alternatively from the circle T ) to the complex numbers. (Continuous on compact domain = both and 2.) Try to form an orthonormal basis using e ikx as basis functions. Recall that for functions from R to C, k can take any real value. Here, what values can k take? For example, look at k = 3 and k = 3.2: (2) so, periodicity requires k = 2πl/ for l Z. 6

7 Inner product of two functions f(x) and g(x) is f(x), g(x) = 0 f(x)g(x)dx. The e ikx functions are readily seen to be orthonormal; it may be shown that they span our space as well. Decomposition (how much energy at each frequency ) of a function in terms of the basis functions: ˆf(k) = e ikx, f(x) = 0 e ikx f(x)dx. This is a Fourier transform. Often called a Fourier series when the domain is T : it maps from 2 (T, dx) to l 2 (2πZ/). 7

8 The previous integral is over infinitely (uncountably many) x in T. Approximate it using a Riemann sum on a uniform mesh of N points: ˆf(k) = = 0 e ikx f(x)dx N N j= j= e ikj/n f(j/n) N = N e ikx f(x j ) x N j= e ikj/n f(j/n). This latter is called the discrete Fourier transform: ˆf D (k) = N N j= e ikj/n f(j/n). We have x = /N and k = 2π/ so x k = 2π/N. 8

9 What are the distinct values of k now? Central point of this paper (to be quantifed below in terms of Poisson summation): when you sample only at N points, you cannot distinguish a frequency of k from a frequency of N +k. For example, with N = 8, here are sin( 2πx/) and sin(9 2πx/): (2) 0 0 9

10 Now look just at the 8 sampled points. true frequency? What is the (2) 0 0 Result: There are N different x s with x j = j/n; there are N different k s with k = 2πl/. Both j and l are taken mod N. 0

11 Duality Functions from R (Fourier transform) to C: x R; k R. The dual of R is R. Functions from T (Fourier series) to C: dual of T is Z2π/. Functions from Z N ( N ) to C (discrete Fourier transform): dual of Z N /N is Z N 2π/. Pontryagin duality is an interesting subject on its own; we only touch on it here.

12 Poisson summation formula Intuition: As seen above, the N-point DFT folds energy from all frequencies which are congruent mod N into the same bin. The formula: ˆf D (p) = ˆf(k + p). k Z 2πN Example: Take N = 8. Then the discrete Fourier transform for frequency, in terms of the true transform values, is ˆf D () =... + ˆf( 7) + ˆf() + ˆf(9) + ˆf(7)

13 Proposition. et be a positive real number. et f(x) (T, dx) with Fourier transform ˆf (Z 2π, dk/2π). Further suppose that Then k Z 2πN x Z N N f(x) = N ˆf(k) < +. k Z 2πN ˆf(k). Proof. Mimic Faris proof for Poisson summation on the real line. See term paper for full details. 3

14 emma. A phase shift in the time domain is a translation in the frequency domain. Proof. 0 e ikx [ e ipx f(x) ] dx = 0 e i(k+p)x f(x)dx = ˆf(k+p). By the lemma, and applying the proposition not to f(x) but rather to e ipx f(x)/n, we have N x Z N N e ipx f(x) = ˆf D (p) = k Z 2πN This proves the Poisson summation formula. ˆf(k + p). 4

15 First-order error analysis Intuition: The error in the DFT is the amount of energy at frequencies outside the range k = 0,,..., N. By the Riemann-ebesgue lemma, for f, the Fourier coefficients ˆf(p) eventually go to zero. But how fast? In particular, we can construct a sinusoidal input with frequency an +, for any integer a. Then the N-point DFT will fold all the input s energy into ˆf D (). Yet, the true Fourier coefficient will have its energy at ˆf(aN +). This high-frequency signal is highly oscillatory, though, which is to say that its second derivative takes on large values. 5

16 Quantitatively, the error in the DFT with respect to the CFT is ˆf(p) ˆf D (p) = ˆf(p) k Z 2πN ˆf(k + p) = k Z 2πN,k 0 ˆf(k + p). 6

17 Proposition. The error in the DFT depends linearly on the upper bound on the second derivative of the input. Furthermore, the convergence of the DFT to the true Fourier transform is quadratic in the number of mesh points. Proof. et M 2 be a positive real number such that for all x T, f (x) M 2. To relate this bound on the second derivative to a bound on ˆf(k +p) for k 0, use the derivative property of the Fourier transform: f (k + p) = i(k + p) ˆf(k + p). 7

18 A straightforward computation involving the transform f (k + p) yields a harmonic series, which is divergent. Note, however, that the square harmonic series converges. Thus, a slight trick is called for. Since we are assuming f C(T ), we have f 2 (T, dx). Following the technique of theorem 23. of Faris notes, put = (k + p)/(k + p) and use the Cauchy-Schwarz inequality on l 2 to obtain k 0,k Z 2πN k 0,k Z 2πN ˆf(k + p) = (k + p) 2 k 0,k Z 2πN k 0,k Z 2πN k + p (k + p) ˆf(k + p) (k + p) ˆf(k + p) 2

19 The use of the inequality is justified since we take f C(T ), hence f 2 (T, dx), and therefore ˆf 2 (T, dk/2π). For the first term, since k Z 2πN, take k = l2πn/. Then k 0,k Z 2πN k 2 = l 0,l Z ( l2πn = 2 2π 2 N 2 π 2 6 = 2 2N 2. ) 2 = 22 4π 2 N 2 l= l 2

20 For the other term in the error estimate, k 0,k Z 2πN (k + p) ˆf(k + p) 2 = = k 0,k Z 2πN k 0,k Z 2πN f (k + p) 2 f (k + p) k + p 2. To relate f (k + p) to the known bound on f (x), we may use the following lemma.

21 emma. et g(x) C(T ). By the compactness of T, g(x) is bounded: say by a positive real constant M. Then for all q Z2π/, ĝ(q) is also bounded by M. Proof. Compute ĝ(q) = = 0 e iqx g(x)dx 0 e iqx g(x) dx = M Mdx = 0 = M. 0 e iqx g(x) dx 0 g(x) dx 8

22 Apply the lemma to the 2nd term in the error estimate: k 0,k Z 2πN = M 2 2 = M N 2. f (k + p) k + p k 0,k Z 2πN 2 k + p 2 k 0,k Z 2πN M 2 2 M 2 k + p k 0,k Z 2πN 2 k 2 Combining these results yields k 0,k Z 2πN ˆf(k + p) 2 M 2 2N 2. 9