Chapter 11S Poisson summation formula Masatsugu Sei Suzuki Department of Physics, SUNY at Bimghamton (Date: January 08, 2011)
Poisson summation formula Fourier series Convolution Random walk Diffusion Magnetization Gaussian distribution function
11S.1 Poisson summataion formula For appropriate functions f(x), the Poisson summation formula may be stated as
f (x n) 2 F(k m) , (1) n m where m and n are integers, and F(k) is the Fourier transform of f(x) and is defined by
1 F(k) F[ f (x)] eikx f (x)dx . 2
Note that the inverse Fourier transform is given by
1 f (x) eikx F(k)dk . 2
The factor of the right hand side of Eq.(1) arises from the definition of the Fourier transform. ((Proof)) The proof of Eq.(1) is given as follows.
1 f (x n) F(k)dk eikn n 2 n
1
We evaluate the factor
I eikn . n
It is evident that I is not equal to zero only when k = 2m (m; integer). Therefore I can be expressed by
I A (k 2m) , m where A is the normalization factor. Then
1 f (x n) F(k)dk[A (k 2m)] n 2 m A F(k) (k 2m)dk 2 m A F(k 2m) 2 m A F(k 2n) 2 n
The normalization factor, A, is readily shown to be 2 by considering the symmetrical case
n2 f (x n) e ,
1 2 F(k 2n) en 2 ______((Mathematica))
2
f x_ Exp x2 ; FourierTransform f x , x, k, FourierParameters 0, 1 2 k 4 2 ______Since
2 f (x n) A 2 . F(k 2n) or
I eikn 2 (k 2m) . (2) n m
Using this formula, we have
f (x n) 2 F(k 2n) . (Poisson sum formula) n n
11S.2 Summary When we put k 2x in I of Eq.(2)
e2ixn 2 (2x 2m) n m 2 (x m) 2 m (x m) m or
3
e2ixn (x m) . (3) n m
11S.3. Convolution of Dirac comb: another method in the derivation of Poisson sum formula The convolution of functions f(x) and g(x) is defined by
1 1 f * g f (x )g( )d f ( )g(x )d 2 2
The Fourier transform of the convolution is given by
F[ f * g] F[ f ]F[g].
Here we assume that
g(x) (x na) . (Dirac comb) n
The Fourier transform of g(x) is
1 1 G(k) F[g(x)] eikx (x na)dx eikna . 2 n 2 n
The convolution f*g is obtained as
1 1 f * g f (x ) ( na)d f (x na) . (4) 2 n 2 n
The Fourier transform of the convolution is
1 F[ f * g] F(k)G(k) F(k) eikna . 2 n
We use the Poisson summation formula;
4
ka eikna eikna ( m) n n m 2 a 2m [ (k )] m 2 a 2 2m (k ) a m a where
e2ixn (x m) . n m
ka with x . Then we get 2
1 1 F[ f * g] F(k) eikna F(k) eikna 2 n 2 n 1 2 2m (k )F(k) 2 a m a 1 2 2m 2m (k )F(k ) 2 a m a a
The inverse Fourier transform of F[ f * g] is obtained as
1 1 1 2 2m 2m f * g F[ f * g]eikxdk eikxdk (k )F(k ) 2 2 2 a m a a 1 2 2m F(k m) eikxdk (k ) a m a a 2m 1 2 i x F(k m)e a a m a where
5
2 2 1 i mx F(k m) e a f (x)dx . a 2
Finally we get
2m 1 1 2m i x f * g f (x na) F(k )e a . (5) 2 n a m a or
2m 2 2m i x f (x na) F(k )e a . n a m a
When a = 1, we get
f (x n) 2 F(k 2m)ei2mx . (6) n m
When x = 0,
f (x n) 2 F(k 2m) . (7) n m
This is the Poisson sum formula.
11S.3 Fourier transform of periodic function We consider a periodic function N(x);
N(x a) N(x) , where a is the periodicity. The function N(x) can be described by
2m 2 2m i x N(x) f (x na) F(k )e a n a m a iGx NGe G
6
Note that f(x) is defined only in the limited region (for example, -a/2≤x≤a/2). G is the reciprocal lattice defined by
2 G m . a
The Fourier coefficient NG is given by
2 N F(k G) G a . 1 1 a / 2 eiGx f (x)dx eiGx f (x)dx a a a / 2 where f(x) is just like a Gaussian distribution function around x = 0.
N x
x
a
Fig. Plot of N(x) as a function of x. a is the lattice constant of the one-dimensional chain.
((Example)) Suppose that f(x) is given by a Gaussian distribution,
7
1 x2 f (x) exp( ) . 2 2 2
Then we get
1 a / 2 1 1 1 2iG 2 1 2iG 2 N eiGx f (x)dx exp( G2 2 )[erf ( ) erf ( )] , G a a / 2 2 2 2 2 2 2 where erf(x) is the error function and is defined by
z 2 2 erf (z) et dt 0
2 2 Figure shows the intensity N vs n, where G n . G a 2 NG 1.0 0.8
0.6
0.4
0.2
n 2 4 6 8 10
2 2 Fig. a = 1. = 0.1. G n . The intensity N vs n (= integer). a G
11S.10 Fourier series
8
Suppose that the function N(x) is a periodic function of x with the periodicity a. Then we have
2m 2 2m i x N(x) f (x na) F(k )e a n a m a iGx NGe G
2 where G m (m: integer). a
((Example-1))
f(x) = x for |x| 1 1 2i(1)n F(k G n) xexp(inx)dx for n ≠ 0. 2 1 n 2 1 1 F(k G 0) xdx 0 for n = 0. 2 1 1 i(1)n 1 i(1)n N(x) eimx eimx n n n n n0 We make a plot of N(x) as a function of x for the summation for n = -nmax and nmax with nmax = 50. 9 N x 1.0 nmax = 50 0.5 x -3 -2 -1 1 2 3 -0.5 -1.0 Fig. nmax = 50. The Gibbs phenomenon is clearly seen. ((Example-2)) f(x) = 0 for -1 1 1 i(1)n[1 (1)n ] F(k G n) exp(inx)dx for n ≠ 0. 2 0 n 2 1 1 1 F(k G 0) dx , for n = 0. 2 0 2 1 1 i(1)n[1 (1)n ] N(x) eimx 2 n 2n n0 10 N x 1.0 0.8 nmax = 100 0.6 0.4 0.2 x -2 -1 1 2 Fig. nmax = 100. The Gibbs phenomenon is clearly seen. ______11S.11 Fourier transform of function having values at inetger x-value We consider a function defined by f (x) f (m) (x m) . (8) m y x Fig. Plot of f(x) which is described by a combination of the Dirac delta function with f(m) at x = m. The Fourier transform of f(x) is given by 11 1 F(k) f (x)eikxdx 2 1 f (m) (x m)eikxdx 2 m 1 f (m) (x m)eikxdx 2 m 1 eikm f (m) 2 m We note that 1 1 F(k 2 ) ei(k 2 )m f (m) eikm f (m) F(k) , 2 m 2 m In other words, F(k) is a periodic function of k with a periodicity 2. We can also show that 1 f (m) F(k)eikm dk , (9) 2 where eik (mm')dk 2 , (10) m,m' since 1 1 1 F(k)eikmdk f (m') eik (mm')dk 2 2 2 m' 1 f (m')2 m,m' 2 m' f (m) Note that the inverse Fourier transform of F(k) is obtained as 12 1 1 1 f (x) F(k)eikxdk dkeikx eikm f (m) 2 2 2 m 1 f (m) dkeik (xm) 2 m 1 f (m)2 (x m) 2 m f (m) (x m) m ((Note)) The proof where the Poisson summation formula is used. We have another method for the derivation of Eq.(9): 1 1 (2n1) f (x) F(k)eikx dk F(k)eikx dk 2 2 n (2n1) We put k' k 2n Then we have 1 1 f (x) F(k)eikx dk F(k'2n )ei(k '2n )xdk' 2 2 n 1 ei2nx F(k')eik 'x dk' n 2 1 (x m) F(k)eikx dk m 2 1 (x m) F(k)eikmdk 2 m (x m) f (m) m Here we use the Poisson summation formula 13 ei2xn (x m) . n m 11S.11 Random walk in the one-dimensional chain We consider the case when a particle moves on the one dimensional chain. The particle can be located only on a discrete position [x = m (m; integer)]. the particle starts from the origin x = 0. The particle can move from x = m to x = m+1, or from x = m to x = m-1 for each jump. We assume that the probability W(m, N) such that the particle reaches at the position x = m after jumps with N times. The probability W(m, N+1) can be described by m-1 m m+1 N N+1 N Fig. Random walk model. 1 1 W (m, N 1) W (m 1, N) W (m 1, N) , 2 2 where one is a jump from x = m-1 to x = m for the (N+1)-th jump, and the another is a jump from x = m+1 to x = m for the (N+1)-th jump. The probability for these jumps is 1/2. In order to find the expression of W (m, N) we define the Fourier transform 1 (k, N) eikmW (m, N) . 2 m Note that (k, N) is a periodic function of k with a periodicity 2, and 1 1 eikmW (m 1, N) eik eik (m1)W (m 1, N) 2 m 2 m , eik (k, N) 14 1 1 eikmW (m 1, N) eik eik (m1)W (m 1, N) 2 m 2 m . eik(k, N) Then we have 1 (k, N 1) eikmW (m, N 1) 2 m 1 (eik eik )(k, N) 2 (cosk)(k, N) or (k, N) (cosk)N (k, N 0) . We use the initial condition that the particle is located at x = 0 at the probability of 1; W (m, N 0) m,0 . or 1 ikm 1 ikm 1 (k, N 0) e W (m, N 0) e m,0 . 2 m 2 m 2 So we get 1 (k, N) (cosk)N . 2 15 F k,N 0.4 0.3 N=200 0.2 N=400 0.1 N=1000 k -0.3 -0.2 -0.1 0.1 0.2 0.3 Fig. Plot of (k, N) vs k, where N = 200, 400, 600, 800, and 1000. In the large limit of N, (k, N ) approaches a Gaussian distribution. The inverse Fourier transform of (k, N) is 1 1 1 W (m, N) (k, N)eikmdk dkeikm (cosk) N 2 2 2 1 dkeikm (cosk) N 2 Here we use the binomial theorem to get N 1 (cosk)N (eik eik )N 2 N 1 N N! eik (2l N ) 2 l 0 (N l)!l! Then we have 16 N 1 1 N N! W (m, N) dkeikm eik (2l N ) 2 2 l0 (N l)!l! N 1 N N! 1 dkeik (m2lN ) 2 l0 (N l)!l! 2 N 1 N N! 1 (2 ) m2lN ,0 2 l0 (N l)!l! 2 N 1 N N! m2lN ,0 2 l0 (N l)!l! N 1 N N! m2lN ,0 2 l0 (N l)!l! 1 N! 2 N N m N m ( )! ! 2 2 where N m m N 2l , or l . 2 This implies that for N = even, m should be even. In other words, W(N = even, m = odd) = 0. For N = odd, m should be odd. In other words, W(N = odd, m = even) = 0. 11S.13 Numerical calculation of W(m, N) Using the Mathematica, we calculate numerically the value of W (m, N) given by 1 1 W (m, N) dkeikm (cosk) N dk cos(km)(cosk) N , 2 0 for the cases with N = 12 (even) and N = 13 (odd). (i) N = 12. 17 W m,N 0.20 N=12 0.15 0.10 0.05 m -10 -5 5 10 This figure clearly shows that W(m, N = 12) = 0 for m = -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, and 11 (m = odd). (ii) N = 13. W m,N 0.20 N=13 0.15 0.10 0.05 m -10 -5 5 10 This figure clearly shows that W(m, N = 13) = 0 for m = -12, -10, -8, -6, -4, -2, 0, 2, 4, 6, 8, 10, and 12 (m = even). ((Mathematica)) 18 Clear "Gobal`" ; 1 J1 m_, N1_ : Cos k m Cos k N1 k 0 d12 Table m,J1 m,12 , m, 12, 12, 1 ; f1 ListPlot d12, PlotStyle Red, PointSize 0.02 , AxesLabel "m", "W m,N " ; f2 Graphics Text Style "N12", Black, 15 , 6, 0.18 ; Show f1, f2 W m,N 0.20 N=12 0.15 0.10 0.05 m -10 -5 5 10 d13 Table m,J1 m,13 , m, 13, 13, 1 ; g1 ListPlot d13, PlotStyle Red, PointSize 0.02 , AxesLabel "m", "W m,N " ; g2 Graphics Text Style "N13", Black, 15 , 6, 0.18 ; Show g1, g2 W m,N 0.20 N=13 0.15 0.10 0.05 m -10 -5 5 10 11S.14 Limit of long distance and long time Using the Stirling's formula, 19 1 1 ln(x!) ln(2 ) (x )ln(x) x for large x, 2 2 and the Mathematica, we have the series expansion to the order of (m/N)8, N m N m lnW (m, N) N ln(2) ln(N!) ln( )!ln( )! 2 2 2 4 1 N 1 m 1 m ln( ) (1 N) (3 N) 2 2 2 N 12 N 6 8 1 m 1 m (5 N) (7 N) ... 30 N 56 N 1 N 1 m2 ln( ) 2 2 2 N or 1 N 1 m2 W (m, N) exp[ ln( ) ] 2 2 2 N 1 2 N 1 m exp[ln( ) 2 ] 2 2 N 1 2 N 1 m ( ) 2 exp( ) 2 2 N 1 2 2 2 1 m exp( ) N 2 N ((Mathematica)) 20 Clear "Gobal`" ; 1 1 f1 x_ : Log 2 x Log x x; 2 2 N m N m g1N Log 2 f1 N f1 f1 ; 2 2 rule1 m N x ; g11 g1 . rule1; Series g11, x,0,8 Simplify Normal 1 1 1 N x2 3 N x4 2 12 1 1 1 N 5 N x6 7 N x8 Log 30 56 2 2 11S.15 Diffusion We assume that the distance between the nearest neighbor lattices is x and the time taken for each jump is t. Then we have t Nt , x mx . The probability of finding particle between x and x + dx is 1 2 x t dx 1 1 x 2 P(x,t)dx W ( , ) dx exp( ) x t 2x (x)2 4t x 2 4 t 2t 2t 1 x 2 dx exp( ) 4Dt 4Dt where D is the diffusion constant and is defined by (x)2 D . 2t 21 x t dx Here we note that the factor 2x of the W ( , ) arises from the fact that (i) for x t 2x every jump (N = even), the probability of finding the particle at the sites with even m is zero, and that (ii) for every jump (N = odd), the probability of finding the particle at the sites with odd m is zero. The distance for the jump is 2x, but not x. The final form of P(x, t) is obtained as 1 x2 P(x,t) exp( ) . 4Dt 4Dt We note that P(x, t) satisfies the diffusion equation given by P(x,t) 2P(x,t) D , t x2 with the initial condition P(x,t 0) (x) . 11.S16 Gaussian distribution in magnetization; analogy of random walk We consider a system consisting of N independent spins. Each spin has a magnetic moment . In the absence of an external magnetic field, each spin has the magnetic moment (±) along the z axis. We assume that the number of spins having the z component magnetic moment (+) is N and the number of spins having the z-component magnetic moment (-): 1 1 N (N n) , N (N n) 2 2 where N N N . Here we discuss the probability distribution of total magnetic moment, M, which is given by 22 M (N N ) n . The probability that the total magnetization has M = n is obtained as 1 N! 1 N! W (M ) 2 N N !N ! 2 N 1 1 [ (N n)]![( (N n)]! 2 2 Using the Stirling's formula, we have 1 N n2 lnW (M ) ln( ) 2 2 2N for n< 2 n2 W (M ) ( )1/ 2 exp( ). N 2N where M = n. The average of magnetization is equal to zero. M n nW (M )dn 0 . Since M 2 2 n 2 n2W (M )dn 2 2 N , the standard deviation is obtained as 2 2 M M M 2N . Since M 2N 1 2N 2N 2N 23 the relative width of the Gaussian distribution becomes sharp as N increases. We make a plot of W(M = n) as a function of N, where N = 100. W M=nm 0.08 0.06 0.04 0.02 n -40 -20 20 40 Fig. Plot of W(M) vs n. M = 2n. N = 100. ______REFERENCES C. Kittel, Elementary Statistical Physics (Dover Publications, Inc, Mineola, NY, 2004). K. Kitahara, Nonequilibrium Statistical Physics, 2nd edition (Iwanami, Tokyo, 1998). [in Japanese]. R. Reif, Statistical Thermal Physics, McGraw-Hill, New York, 1965). C. Kittel and H. Kroemer, Thermal Physics, 2nd edition (W.H. Freeman and Company, New York, 1980). 24