Computational Methods and advanced Statistics Tools I. FOURIER SERIES, FOURIER TRANSFORM. 1. UNIFORM CONVERGENCE A series of functions is a series where the general term is a sequence of functions: ∞ X fn (x), x ∈ I n=1 where I is some common domain ⊂ R. The series is pointwise convergent to a sum S(x) in I if ∀x ∈ I, ∀ǫ > 0, ∃n̄(ǫ, x) > 0 : | n X fk (x) − S(x)| < ǫ, ∀n > n̄(ǫ, x). n X gk (x) − S(x)| < ǫ, ∀n > n̄(ǫ) k=1 The series is uniformly convergent to S(x) in I if ∀ǫ > 0, ∃n̄(ǫ) > 0 : ∀x ∈ I, | k=1 If the functions fn (x), n ∈ N , are continuous for x ∈ I, the point convergence is not sufficient to ensure that the sum S(x) is continuous. But if the series of continuous functions is uniformly convergent to S(x) for x ∈ I, then S(x) is a continuous function in I. In turn, a sufficient condition for the uniform convergence of a series of functions is the ”total convergence” criterion: the series of functions is uniformly convergent in I if the modulus of the general term is bounded by the general term of a convergent numerical series. In other words, if ∞ X |fn (x)| ≤ αn , ∀x ∈ I, with αn < +∞ n=1 then the series of functions P∞ n=1 fn (x) is uniformly convergent. When such a condition is satisfied one can prove that: i) the sum of the series of function S(x) is continuous in I; ii) the definite integral on [a, b] ⊂ I of the sum of the series is equal to the sum of the integrals of fn (x); 1 2 iii) if, moreover, the general term fn (x) is differentiable and, in the interval I, |fn′ (x)| is bounded by the general term of convergent numerical series, the sum S(x) of the series of functions is differentiable and the sum of the derivatives converges to S ′ (x). For example, consider the series of continuous functions +∞ X n=0 fn (x) = +∞ X 1 sin(nx) 3 n n=0 Since the general term is bounded by the general term of a convergent numeric series, X 1 1 1 |fn (x)| = 3 | sin(nx)| ≤ 3 , with < +∞ n n n3 n then the sum S(x) of the series is finite and continuous, ∀x ∈ R. Moreover, since the series of the derivatives enjoys the same property, X 1 1 1 < +∞ |fn′ (x)| = 3 |n cos(nx)| ≤ 2 , with 2 n n n n then S(x) admits a continuous derivative S ′ (x), which coincides with the sum of the series of the derivatives ∀x ∈ R. In such a case we say that the series can be derived term by term. Besides, for any finite a, b with a < b, Z bX XZ b fn (x)dx fn (x)dx = n a a n 3 for fn (x) = sin(nx)/n , as well as for fn (x) = cos(nx)/n2 . In such cases we say that the series can be integrated term by term. 2. FOURIER SERIES A function f is periodic if it is defined for all real x and there is a positive number T such that (1) f (x + T ) = f (x), ∀x According to this definition, a constant function is periodic with arbitrary period. Trigonometric functions, such as sin x, sin 2x, ..., sin nx, or cos x, ..., cos nx, are periodic with period 2π. Besides exponential functions of imaginary arguments, too, such as eix , e2ix , ..., einx , or e−ix , e−2ix , ..., e−inx , are periodic functions of period 2π. Indeed e2πi = 1. If T is a period for f then f (x + nT ) = f (x), ∀n ∈ N : therefore f is periodic with periods 2T, 3T, ..., too. 3 A trigonometric polynomial is a finite sum of the type k a◦ X + (an cos nx + bn sin nx) Sk (x) = 2 n=1 (2) Any trigonometric polynomial is periodic with period 2π. Any trigonometric polynomial can be written in exponential form: (3) Sk (x) = k X cn einx . n=−k Indeed, for n = 0, c◦ = a◦ /2. For each n ≥ 1, in the equality an cos nx + bn sin nx = cn einx + c−n e−inx we can use Euler’s formulae iθ e = cos θ + i sin θ e−iθ = cos θ − i sin θ ⇐⇒ cos θ = 21 (eiθ + e−iθ ) sin θ = 2i1 (eiθ − e−iθ ) It suffices to write cos nx, sin nx in terms of exponentials. Thus we obtain the coefficients cn in terms of the an , bn ’s: (4) c◦ = 1 1 a◦ , cn = (an − ibn ), n ≥ 1; c−n = (an + ibn ), n ≥ 1. 2 2 2 In particular, if the an , bn ’s are real, then c−n is the complex conjugate of cn : c−n = c̄n , ∀n 6= 0. THEOREM I.2A - Let f (x + 2π) = f (x), ∀x ∈ R, and let us assume that f (x) is the sum of a trigonometric series +∞ (5) a◦ X f (x) = + (an cos nx + bn sin nx) 2 n=1 or, in equivalent manner, (6) f (x) = +∞ X cn einx . n=−∞ Moreover, assume that the series is uniformly convergent for x ∈ R. Then the coefficients of the series in exponential form are Z 2π 1 f (x)e−inx dx, ∀n ∈ Z (7) cn = 2π 0 and the coefficients of the series in trigonometric form are (8) Z Z 1 2π 1 2π an = f (x) cos(nx)dx, n ∈ N◦ , bn = f (x) sin(nx)dx, n ∈ N. π 0 π 0 4 Proof To fix ideas consider the exponential form (??) and let us integrate both sides of (??). By virtue of the uniform convergence, we can integrate term by term: Z 2π X einx 2π f (x)dx = 2πc◦ + = 2πc◦ in 0 0 n6=0 so that c◦ is the mean of the periodic function. Multiplying both sides by e−imx , m 6= 0, and integrating we find: Z 2π X ei(n−m)x 2π −imx + 2πcm = 2πcm f (x)e dx = cn i(n − m) 0 0 n6=m R 2π so that cm = (2π)−1 0 f (x)e−imx dx as stated. By inversion of (??) a◦ (9) = c◦ ; an = cn + c−n , bn = i(cn − c−n ), ∀n ≥ 1, 2 and, by Euler’s formula, 1 1 cos x = (eix + e−ix ), sin x = (eix − e−ix ) 2 2i Thus, for n ≥ 1, Z 2π Z 2π 1 1 −inx inx f (x)(e + e )dx = f (x)2 · cos(nx)dx an = 2π 0 2π 0 and Z 2π Z 2π 1 1 −inx inx f (x)i · (e − e )dx = f (x)i · (−2i) sin(nx)dx. bn = 2π 0 2π 0 △ The above calculation assumes that the period is 2π. In fact any period T > 0 can be handled so as to obtain similar formulae for the coefficients cn , an , bn . Above all, since the integrals (??), (??) exist under wide conditions, we can associate Fourier coefficients to any periodic function in a large class: DEF. I.2B (Fourier series) Let f (x + T ) = f (x)∀x ∈ R, and let f be integrable in the interval [0, T ]. The Fourier coefficients and the formal Fourier series, in exponential form, associated with f are defined by Z +∞ X 2π 1 T nx −i 2π T , f (x) ∼ cn ei T nx (10) cn = f (x)e T 0 n=−∞ The Fourier coefficients and the formal Fourier series, in trigonometric form, associated with f are defined by Z Z 2 T 2 T 2π 2π (11) an = f (x) cos( nx) dx, bn = f (x) sin( nx) dx T 0 T T 0 T 5 (12) +∞ +∞ X 2π 2π a◦ X + an cos( nx) + bn sin( nx) f (x) ∼ 2 T T n=1 n=1 REMARK I.2C - a) An equivalent expression of coefficients makes use of integrals from −T /2 to T /2, since the length of the interval is T : Z 2 T /2 2π (13) an = f (x) cos( nx) dx, n ∈ N◦ T −T /2 T (14) 2 bn = T Z T /2 f (x) sin( −T /2 2π nx) dx, n ∈ N T Actually the same numbers are obtained by integrating on any interval [c, c + T ], with fixed c ∈ R. b) The symbol ∼ reminds that the trigonometric expansion is still formal: in fact nothing is said about convergence, and, if the series is convergent, the sum is not necessarily f (x). RT c) The constant term in (??), i.e. c◦ ≡ a◦ /2 = T1 0 f (x) dx, is the mean value of f in one period. 3. THEOREMS ON FOURIER SERIES THEOREM I.3A (Uniformly convergent expansion if the periodic function is regular enough) Let the T −periodic function f admit continuous r−th derivative in [0, T ] with r ≥ 1. Then the Fourier coefficients of the series in exponential form satisfy (15) |cn | ≤ K|n|−r , ∀n ∈ Z − {0}, where the constant K is independent of n. Similarly the Fourier coefficients an , bn of the series in trigonometric form (16) |an | ≤ K̃n−r , |bn | ≤ K̃n−r , ∀n ∈ N for some K̃ > 0. As a consequence, any periodic function f of class C r with r ≥ 2, is the sum of a uniformly convergent Fourier series, both in exponential form and in trigonometric form. Proof. - Let f ∈ C r ([0, T ]). The derivative of a periodic function is still periodic, so integrating by parts r times the expression (?? ) ! T Z T 2π 2π T T 1 f (x) e−i T nx − e−i T nx dx f ′ (x) cn = T −i2πn −i2πn 0 0 6 1 T =+ T i2πn Therefore Z T ′ f (x)e 0 1 dx = + T T i2πn r Z T 2π f (r) (x)e−i T nx dx. 0 r Z T 2π T |f (r) (x)||e−i T nx | dx 2π|n| 0 r Z T T K 1 |f (r) (x)| dx ≤ ≤ r T |n| 2π |n|r 0 1 |cn | ≤ T where nx −i 2π T r T K= max |f (r) (x)|. x∈[0,T ] 2π Since an = cn + c−n and bn = i(cn − c−n ), for the an ’s and the bn ’s it suffices to choose K̃ = 2K. Then the uniform convergence is obvious: indeed K , ∀x ∈ R |cn ||einx | = |cn | ≤ |n|r and K̃ , ∀x ∈ R |n|r In both cases the general term of the Fourier expansion is bounded by the general term of a convergent numerical series, that is a gerneralized harmonic series with exponent r > 1: X K |a◦ | X 2K̃ |c◦ | + < +∞, + < +∞. △ r r |n| 2 |n| n≥1 n6=0 |an cos nx + bn sin nx| ≤ |an | + |bn | ≤ 2 DEF. A function f (x) is said piecewise continuous in (a, b) ⊂ R if f is defined and continuous except in a finite number of points xj ∈ (a, b), j = 1, ..., N , where there exist the finite limits: f (xj + 0) = limx→xj + (x), f (xj − 0) = limx→xj − f (x). Such type of regularity is part of the Dirichlet condition, which is in the hypothesis of the convergence criterion: THEOREM I.3B (Dirichlet’s Criterion) Let the function f (x) be periodic (with period T ). Let both f (x) and f ′ (x) be piecewise continuous. Then the trigonometric series associated with f is convergent to f (x) in each point where f is continuous; while in each point x◦ where f is discontinuous, the series is convergent to the half-sum f (x◦ + 0) + f (x◦ − 0) . 2 7 Figure 1. Graph of the Fourier series calculated by the first 25 harmonics, i.e. for n from 0 to 12. EXAMPLE I.3C (Fourier expansion of a periodic step function) - Let 1 x ∈ [2nπ, (2n + 1)π] f (x) = 0 elsewhere. which satisfies the Dirichlet conditions, with period T = 2π. The Fourier coefficients a◦ , an , bn are directly calculated: Z π Z π 1 1 1 a◦ = f (x) dx = 1 dx = . 2 2π −π 2π 0 2 For n ∈ N, π Z Z 1 π 1 π 1 sin(nx) an = = 0, f (x) cos(nx) dx = cos(nx) dx = π −π π 0 π n 0 π Z Z 1 π 1 cos(nx) 1 π f (x) sin(nx) dx = sin(nx) dx = − = bn = π −π π 0 π n 0 1 0, n even n = [1 − (−1) ] = 2 , n odd . nπ nπ 2 Thus bn vanishes for even n, it equals nπ for odd n. The Fourier series is: ∞ ∞ 2 1 X 1 X + sin [(2n + 1)x] . [an cos(nx) + bn sin(nx)] = + 2 n=1 2 n=0 (2n + 1)π 8 REMARK I.3D - (On the computation of Fourier coefficients) a) In order to compute the coefficients of the Fourier series, several remarks are useful. A function f (x) is said to be even if f (−x) = f (x), ∀x; it is odd if f (−x) = −f (x), ∀x. If the function f (x) is even, then all the Fourier coefficients bn vanish: indeed, under a substitution y = −x, Z Z 2π 2π 2 T /2 2 T f (x) sin( nx) dx = f (x) sin( nx) dx = bn = T 0 T T −T /2 T 2 T Z T /2 2 2π f (−y) sin(− ny) dy = − T T −T /2 Z T /2 f (y) sin( −T /2 2π ny) dy = −bn T Thus 2bn = 0, so that bn = 0. Similarly, it turns out that a◦ = an = 0 if the function is odd. Therefore an even periodic function admits a series expansion with only cosine terms; while an odd periodic function admits a series expansion with only sine terms. b) The next remark regards an arbitrary function f (x), defined in some interval, for example in (0, π). We can extend it both to an even function and to an odd function. Indeed, we can define it in the interval (−π, 0) so that it becomes even: f (x), x ∈ (0, π) limx→0+ f (x), x = 0 f◦ (x) = f (−x), x ∈ (−π, 0) Another choice is to define an odd extension of f (x): x ∈ (0, π) f (x), 0, x=0 f1 (x) = −f (−x), x ∈ (−π, 0) Finally, one can extend the functions f◦ (x) and f1 (x) to the whole of R, so that the resulting function is periodic with period 2π. Assuming the Dirichlet conditions to hold, it follows that ∞ X 1 an cos(nx) f◦ (x) = a◦ + 2 n=1 (17) where 1 an = π Z π 2 f◦ (x) cos(nx) dx = π −π Z π f (x) cos(nx) dx 0 and (18) f1 (x) = ∞ X n=1 bn sin(nx) 9 where 1 bn = π Z π 2 f1 (x) sin(nx) dx = π −π Z π f (x) sin(nx) dx 0 Indeed f1 (x)sin(nx) is even, as a product of two odd functions. Now under a restriction of (??), (??) to the interval (0, π), it follows that a function with domain (0, π) can be expanded in a cosine-series as well as in a sine-series. 4. APPLICATION OF THE FOURIER SERIES TO DIFFERENTIAL EQUATIONS EX. I.4A - The damped harmonic oscillator with a periodic force Let us compute a particular solution of the damped harmonic oscillator with a forcing term ẍ + µẋ + ω 2 x = f (t) (19) when f (t) is a real valued periodic function with period T . Although not real-valued, it is convenient first to consider f (t) = ρ · eiΩt where Ω = 2π is the pulsation of the forcing term. In this case a T particular solution in the form const. · eiΩt is easily found: x∗ (t) = ρeiΩt ω 2 − Ω2 + iΩµ As a second step let us consider f (t) given by a finite Fourier expansion in exponential form: N X f (t) = cn eiΩnt n=−N where cn = c̄−n since f (t) is real valued. A particular solution with period T is given by ∗ x (t) = N X x∗n (t), n=−N ∗ x∗n (t) = cn eiΩnt ω 2 − n2 Ω2 + inΩµ Notice that x is real since the addends of the sum with opposite indices n, −n are complex conjugate. Moreover x∗n (t) is a particular solution of the differential equation ẍ + µẋ + ω 2 x = cn eiΩnt . 10 Thus we can immediately verify: ∗ ∗ 2 ∗ ẍ + µẋ + ω x = N X (ẍ∗n + µẋ∗n + ω 2 x∗n ) n=−N = N X cn eiΩnt = f (t). n=−N Finally, let f (t) admit a Fourier expansion with infinitely many nonzero coefficients: ∞ X f (t) = cn eiΩnt n=−∞ A particular solution is looked for in the form: ∞ X cn ∗ eiΩnt (20) x (t) = 2 2 2 + inΩµ ω − n Ω n=−∞ To this end we have to check the convergence of (??); then, provided its convergence, to check whether is it a solution of the differential equation. Now a sufficient condition to derive k times term by term is the convergence of the series of the k−th order derivatives: ∞ X cn (iΩn)k (21) eiΩnt 2 − n2 Ω2 + inΩµ ω n=−∞ uniformly with respect to t. On the other hand we know that f ∈ C r implies the bound on the Fourier coefficients |cn | ≤ cn−r . Therefore the general term of ??) is bounded by cΩk nk p ≤ Cnk−r−2 nr (ω 2 − n2 Ω2 )2 + n2 µ2 Ω2 for some constant C > 0 independent of n. So the series (??) is uniformly convergent with respect to t if r + 2 − k > 1; In particular, the series (??) is a solution of the differential equation if r + 2 − 2 > 1 (k = 2), i.e. if the function f (t) is at least in C 2 . Notice: for µ small enough, the particular solution x∗ (t) amplifies the harmonics n = ±[ω/Ω], (where [·] denotes the integral part), of the function f (t) for which ω 2 − n2 Ω2 attains its minimum, while the other harmonics are damped. EXAMPLE I.4B Let us consider the one-dimensional heat equation with boundary and initial conditions: (22) 2 ∂u = 2 ∂∂xu2 ∂t u(0, t) = u(3, t) = 0, ∀t ≥ 0 u(x, 0) = 5 sin(4πx) − 3 sin(8πx) + 2 sin(10πx), x ∈ [0, 3]. Besides, the solution is required to remain bounded ∀(x, t) ∈ [0, 3]×R+ . 11 Let us solve the equation by separation of variables. By such a method, we look for a solution (if possible) of the form u(x, t) = f (x)g(t) . Then the equation reads f (x)ġ(t) = 2f ′′ (x)g(t). After separating variables, ġ(t) f ′′ (x) = . 2· f (x) g(t) Since the left hand side depends only on x, the right hand side depends only on t, and the equation holds for all (x, t) where x, t are independent variables, f ′′ (x) ġ(t) ∃λ ∈ R : 2 = 2λ = . f (x) g(t) Therefore we obtain two ordinary differential equations depending on an arbitrary parameter λ: f ′′ = λf, ġ = 2λg with general solution (23) f (x) = c1 e √ λx + c2 e − √ λx g(t) = ce2λt , f (x) = c1 x + c2 , g(t) = c if λ 6= 0 if λ = 0. Now, the boundedness requirement ∀t ∈+ implies λ < 0. Therefore, setting λ = −ω 2 , by Euler’s formulae the general solution (??) can be written: f (x) = a cos(ωx) + b sin(ωx), 2 g(t) = ce−2ω t . Hence a solution of the heat equation turns out to be 2 u(x, t) = f (x)g(t) = ce−2ω t [a cos(ωx) + b sin(ωx)] = 2 = e−2ω t [A cos(ωx) + B sin(ωx)] with constants A, B, ω to be determined by using the boundary conditions in (??): 2 u(0, t) = 0 =⇒ A = 0, i.e. u(x, t) = e−2ω t B sin(ωx) u(3, t) = 0 =⇒ B sin(3ω) = 0 This last equality, in turn, implies either B = 0, arbitrary ω, or arbitrary B, ω = nπ/3, n ∈ . Now B = 0 is immediately rejected since it gives an identically vanishing solution. Therefore a nontrivial solution of the diffusion equation in (??) with zero boundary condition has the form: nπ 2 2 (24) u(x, t) = Be−2n π t/9 sin( x), n ∈ 3 12 Figure 2. Graph of the function u(x, t) for some values of t: t = 0 (continuous line), t = 0.001 and t = 0.01 By the superposition principle (which holds for linear differential equations), a linear combination of solutions is still a solution of the homogeneous equation; so for any choice of n1 , n2 , n3 ∈ the function n1 π n2 π n3 π 2 2 2 2 2 2 B1 e−2n1 π t/9 sin( x)+B2 e−2n2 π t/9 sin( x)+B3 e−2n3 π t/9 sin( x) 3 3 3 is still a solution of the equation in (??). By comparing this type of solution with the initial condition in (??) we find: B1 = 5, n1 π/3 = 4π, B2 = −3, n2 π/3 = 8π, B3 = 2, n3 π/3 = 10π i.e. n1 = 12, n2 = 24, n3 = 30. Thus the unique (bounded) solution of the diffusion problem (??) is 2 2 2 u(x, t) = 5e−32π t sin(4πx) − 3e−128π t sin(8πx) + 2e−200π t sin(10πx) EXAMPLE I.4C - We solve the same diffusion equation with a different initial value function: ∂2u ∂u = 2 2 ∂t ∂x u(0, t) = u(3, t) = 0, ∀t ≥ 0 u(x, 0) = 25x(3 − x), x ∈ [0, 3] Again by separation of variables we obtain a solution of the heat equation, but it is no more sufficient to make a finite superposition of such functions to solve the problem. We try by a superposition of infinitely many functions: ∞ X π 2 2 u(x, t) = Bn e−2n π t/9 sin(n x), 3 n=1 13 Figure 3. Graph of the function u(x, t), for (x, t) ∈ (0, 3) × (0, 0.01) where we require u(x, 0) = 25x(3 − x), that is 25x(3 − x) = ∞ X π Bn sin(n x), x ∈ (0, 3). 3 n=1 This amounts to expand in a sine Fourier series the function 25x(3−x). According to Remark I.3D, this is possible if we regard the function as a restriction of an odd periodic function defined for all x ∈, with period 6: 25x(3 − x), x ∈ (0, 3) 25x(3 + x), x ∈ (−3, 0) f1 (x) = ... and so on. Then we have the usual Fourier coefficients for an odd function of period T : Z Z 2 T /2 4 T /2 2π 2π Bn = f1 (x) sin(n x) dx = f1 (x) sin(n x) dx = T −T /2 T T 0 T Z 2 3 900 = 25x(3 − x) sin(nπx/3) dx = 3 3 [1 − (−1)n ] 3 0 nπ Finally ∞ 1800 X 2 2 (2m + 1)−3 e−2(2m+1) π t/9 sin((2m + 1)πx/3). u(x, t) = 3 π m=0 The graph of u(x, t) is first represented for some values of t (Fig. 4), then for (x, t) ∈ (0, 3) × (0, 1) (Fig. 5). 14 Figure 4. For initial shape u(x, 0) = 25x(3 − x), graph of the solution u(x, t) when t = 0, t = 0.01, t = 0.1 Figure 5. For initial shape u(x, 0) = 25x(3 − x), graph of the solution u(x, t), (x, t) ∈ (0, 3) × (0, 1) 5. FOURIER TRANSFORM DEFINITION I.5A (Direct Fourier transform) Let f = f (t) be a real valued (or a complex valued) function of a real variable. The Fourier transform of f is the function: Z +∞ ˆ (25) F (f ) = f (ω) = f (t)e−iωt dt, ω ∈ R −∞ 15 whenever such an integral exists ∀ω ∈ R. The operator F , which transforms f (t) into fˆ(ω), is said ”Fourier transform”. PROPOSITIONS I.5B Let f (t) be an absolutely integrable function in R, that is: Z +∞ |f (t)|dt < +∞ (26) −∞ Then i) the Fourier transform fˆ(ω), for ω ∈ R, exists; ii) the modulus of fˆ(ω) is an even function if f is real valued; iii) limω→±∞ fˆ(ω) = 0 (Riemann’s lemma); iv) fˆ is a continuous function ∀ω ∈ R. Let f (t), f ′ (t) be piecewise continuous in any bounded interval, then v) the inversion formula Z +∞ Z +R 1 1 iωt ˆ (27) f (t) = f (ω)e dω ≡ PV lim fˆ(ω)eiωt dω R→+∞ 2π 2π −∞ −R holds in all points t where f (t) is continuous, while the formula Z +∞ 1 1 fˆ(ω)eiωt dω [f (t◦ + 0) + f (t◦ − 0)] = PV (28) 2 2π −∞ holds in all discontinuity points t◦ . Remarks. (i) Since ω and t are real, |eiωt | = 1, and Z +∞ Z +∞ −iωt | f (t)e dt| ≤ |f (t)|dt < +∞ −∞ −∞ so the existence of fˆ is immediate if f is absolutely integrable. (ii) If f = f¯ (real valued function), the complex conjugate of fˆ(ω) is equal to the Fourier transform of f calculated in −ω: Z +∞ Z +∞ ¯ˆ −iωt ¯ f (ω) = ( f (t)e dt) = f (t)eiωt dt = fˆ(−ω). −∞ −∞ Now writing fˆ(ω) in exponential notation fˆ(ω) = A(ω)eiφ(ω) we notice that A(ω)e−iφ(ω) = A(−ω)eiφ(−ω) , so that A(ω) = A(−ω). The proof of (iii)-(v) is omitted. Only remark that the integral in (??) is computed as a Cauchy principal value. The Cauchy Principal R +R Value integral is convergent if the limit limR→+∞ −R (...) exists and is 16 finite. Recall that the usual generalized integral of a function g(t) is convergent if the double limit Z R g(t) dt lim R,S→+∞ −S exists and is finite with R, S diverging to +∞ independently of each other. Of course if the integral is convergent in the generalized sense, then it is convergent in the sense of the Cauchy principal value. But the opposite implication is not true in general. An example is the function g(t) = t, which is integrable in the Cauchy principal value sense: Z +R Z +∞ t dt = 0. P t dt = lim R→+∞ −∞ −R The same function is not integrable in the generalized sense, since many different the double limit prescription: for examR R values are attainedRby 2R ple −2R tdt → −∞, while −R tdt → +∞. A similar situation regards any odd, piecewise continuous function. Finally, in some textbooks the direct ad the inverse Fourier transforms are defined in √ a slightly different way, attributing to both the same coefficient 1/ 2π (instead of 1 and 1/2π, respectively). Actually any other choice is good, provided that the product of the two coefficients is 1/2π. △ PROPOSITION I.5C Symmetry property If fˆ is the Fourier transform of f , and if both satisfy the inversion conditions I.5B, (v), then (29) F [fˆ](ω) = 2πf (−ω) Proof R +∞ 1 fˆ(t)eiωt dt = f (ω), then Since the inversion formula ensures 2π −∞ Z +∞ Z +∞ 1 −iωt ˆ ˆ f (t)e dt = 2π F (f ) = fˆ(t)e−iωt dt = 2πf (−ω). △ 2π −∞ −∞ PROPOSITION I.5D (Computation of Fourier transforms) (30) (31) (32) 1) If pa (t) = 2) If f (t) = 1, 0, sin(ω◦ t) t 3) If f (t) = e −αt2 t ∈ [−a, a] |t| > a then then p̂a (ω) = 2 π, π/2, fˆ(ω) = 0, then fˆ(ω) = r π − ω2 e 4α . α sin ωa ω |ω| < ω◦ ω = ±ω◦ |ω| > ω◦ 17 Proof 1) Let pa (t) = I[−a,a] (t), for each a > 0. The Fourier transform (??) is nothing else than −iωt a Z a e 1 sin ωa −iωt e dt = p̂a (ω) = = (eiωa − e−iωa ) = 2 −iω −a iω ω −a for ω 6= 0, while p̂a (0) = 2a. respectively. Fig. 6 and Fig. 7 show pa (t) and p̂(ω), 2) Now consider sin(a t) . t By using (30), the given function f (t) can be seen as a Fourier transform: 1 sin(a t) = pˆa (t), f (t) = t 2 where pa is the indicator function of [−a, a]. By applying the Fourier transform to both sides and using the symmetry property I.5C, 1 1 fˆ(ω) ≡ F [p̂a ](ω) = 2π pa (−ω) = π · pa (−ω) 2 2 in all points where pa is continuous. The half-sum rule (??) holds in the discontinuity points, so the explicit result is: |ω| < a π, π/2, ω = ±a (33) fˆ(ω) = 0, |ω| > a f (t) = 3) Finally consider the Gauss integral and the Fourier transform of the Gaussian function. R +∞ 2 a) Let I = −∞ e−αx dx. By Fubini’s theorem Z +∞ 2 Z +∞ Z +∞ 2 2 −αx2 −αx2 I = e−αy dy e dx e dx = −∞ −∞ −∞ = Z +∞ Z −∞ +∞ e−α(x 2 +y 2 ) dxdy −∞ Under a change of variables from cartesian to polar coordinates, ( x = r cos θ , dxdy → rdrdθ y = r sin θ the integral becomes Z 2π Z +∞ π 2 2 2 e−αr rdrdθ = 2π[e−αr /(−2α)]+∞ = . I = 0 α 0 0 p So I = π/α. 18 Figure 6. Graph of pa (t). b) F [e −αt2 ]= Z e R −iωt −αt2 e dt = Z e R √ √ )2 −( αt+ 2iω α ω2 · e− 4α dt Now by Cauchy’s theorem of complex analysis, the integral performed √ (parallel to the real line in the complex plane) is on the line R + 2iω α p equal to the integral performed on the real line, which is απ : r Z √ 2 ω2 π − ω2 e 4α . = e−( αt) · e− 4α dt = α R Therefore, up to a coefficient, the Fourier transform of a Gauss function is still a Gauss function, with α replaced by −1/(4α). △ 6. PROPERTIES OF THE FOURIER TRANSFORM THEOREM I.6A (Linearity ) Let the functions f1 , f2 admit Fourier transform F (f1 ) = fˆ1 and F (f2 ) = fˆ2 , respectively. Then for arbitrary coefficients c1 , c2 ∈ there exists the Fourier transform of c1 f1 (t)+c2 f2 (t) and F (c1 f1 + c2 f2 ) = c1 fˆ1 + c2 fˆ2 . Proof. A straightforward consequence of the linearity property of the integral. F and F −1 are linear operators, too. △. 19 Figure 7. Graph of p̂a (ω). THEOREM I.6B (Frequency shifting) Let the function f (t) admit Fourier transform fˆ(ω). For any constant ω◦ ∈, F [eiω◦ t f (t)] = fˆ(ω − ω◦ ). Proof. F [e iω◦ t f] = Z +∞ −∞ f (t)e−i(ω−ω◦ )t dt = fˆ(ω − ω◦ ) △ EX. I.6C (Fourier transform of a truncated cosine function) Is it possible to transform, in a generalized sense, the cosine (sine) function, which does not admit the integral transform I.5A in the strict sense? An approximation to the right answer (I.8H iv) is clear if we consider a truncated cosine, supported in intervals (−a, a) for arbitrarily large a > 0 : f (t) = pa (t) · cos(ω◦ t), pa (t) = I(−a,a) (t) where pa (t) is the indicator of (−a, a) and ω◦ is nonzero and fixed. By Euler’s formula, linearity, frequency shifting and by the above examples I.5D, 1 fˆ(ω) = F [pa (t) cos(ω◦ t)] = F (pa (t)eiω◦ t ) + F (pa (t)e−iω◦ t ) 2 1 [p̂a (ω − ω◦ ) + p̂a (ω + ω◦ )] 2 sin[a(ω − ω◦ )] sin[a(ω + ω◦ )] + . △ = ω − ω◦ ω + ω◦ = 20 Figure 8. Graph of the Fourier transform of the modulated indicator function pa (t) cos(ω◦ t), with ω◦ = 7, a = 5. As a → +∞, Fig. 8 shows that the Fourier transform appears more and more concentrated on the frequency ω◦ (and, of course, −ω◦ since this Fourier transform is an even function). △ Let us prove that the Fourier operator transforms derivatives in the time representation into multiplication by corresponding monomials in the frequency representation, and vice-versa. By this property the Fourier is a powerful tool for differential equations. THEOREM I.6D (Transformation of derivatives into multiplications and vice-versa ) Let f ∈ C n () and let f (r) (t) be absolutely integrable on for each r ≤ n. Then F [f (n) ] = (iω)n fˆ(ω). Similarly, assuming that tn f (t) is absolutely integrable, dn fˆ(ω) F [(−it) f (t)] = . dω n Proof. The proof is based on integration by parts and the fact that n (34) lim f (r) (t) = 0, t→±∞ r = 0, 1, ..., n − 1 since f (r) (t), r = 0, 1, . . . , n is absolutely integrable on . We can realize this fact by writing Z t (r) (r) f (t) = f (0) + f (r+1) (τ )dτ. 0 21 Since f (r+1) (τ ) is absolutely integrable, then the following limit exists and is finite: Z +∞ (r) (r) f (r+1) (τ )dτ. L = lim f (t) = f (0) + t→+∞ 0 Now, either L = 0 or L 6= 0 : we prove that the second case is impossible. Ab absurdo, let L 6= 0. Then there is an interval [T, +∞) in which |f (r) (t)| > L/2, so that f (r) (t) cannot be absolutely integrable. Thus (??) is proved. Consider Z +∞ (n) F (f )(ω) = f (n) (t)e−iωt dt −∞ = f (n−1) +∞ (t)e−iωt −∞ + iω Z +∞ f (n−1) (t)e−iωt dt −∞ (n−1) = iω F (f )(ω) since limt→±∞ f (t) = 0. By iterating such relation the statement follows. Finally, by direct inspection the trasform of (−it)f (t) is equal to dfˆ(ω)/dω, and the iteration on n concludes the proof. △ (n−1) 7. CONVOLUTION PRODUCT DEFINITION I.7A -(Convolution product) The convolution product of two functions f1 (t), f2 (t) defined for t ∈, is the generalized integral (provided it exists) Z +∞ f1 (x)f2 (t − x)dx. (35) (f1 ∗ f2 )(t) = −∞ REMARK I.7B - For the convolution product the usual associative and distributive (with respect to sum) properties hold. Moreover, the commutative property can be easily proved: by a change of variable t − x = y, Z +∞ Z +∞ f1 ∗ f2 (t) = f1 (x)f2 (t − x)dx = f1 (t − y)f2 (y)dy = f2 ∗ f1 (t). −∞ −∞ THM. I.7C - (Convolution theorem in time domain, convolution theorem in frequency domain) If the functions f1 , f2 are sufficiently regular 1 then (36) F (f1 ∗ f2 )(ω) = fˆ1 (ω) · fˆ2 (ω) 1A generic assumption of regularity so as to ensure the existence of the Fourier transforms, the change in the order of integration, etc. 22 F −1 (fˆ1 ∗ fˆ2 )(t) = (f1 ∗ f2 )(t). (37) Scheme of proof. Let us consider the Fourier transform of a convolution product: Z +∞ Z +∞ −iωt f1 (x)f2 (t − x)dx e dt F (f1 ∗ f2 )(ω) = −∞ −∞ = Z +∞ −∞ Z +∞ e−iωt f1 (x)f2 (t − x)dt dx −∞ By the change of variables (t, x) → (y, x), where x = x and y = t − x, we obtain Z +∞ Z +∞ e−iω(x+y) f1 (x)f2 (y)dy dx F (f1 ∗ f2 )(ω) = −∞ = Z +∞ e−iωx f1 (x) dx −∞ and vice-versa F −1 Z −∞ +∞ −∞ e−iωy f2 (y) dy = fˆ1 (ω) · fˆ2 (ω) ˆ ˆ f1 · f2 (t) = (f1 ∗ f2 ) (t) △ EXAMPLE I.7D - (The diffusion equation on a line) Consider the heat equation with initial condidion f (x) for ∈: 2 ∂u = k ∂∂xu2 ∂t u(x, 0) = f (x), −∞ < x < +∞ |u(x, t)| < M, −∞ < x < +∞, t > 0 In this problem it is useful to consider the Fourier transform of both sides with respect to x, so that the second order derivative (with respect to x) is transformed into a multiplication by the square of the conjugate variable (Thm. I.6F). Denoting p the Fourier variable conjugate with x, we obtain: dF [u] = −kp2 · F [u] dt where the Fourier transform of u is denoted by F [u]. Here we deal with a first order ordinary differential equation, since the unknown function F [u] no more depends on x. The general solution of such ordinary equation is: (38) F [u(x, t)] = Ce−kp 2t Setting t = 0 in (??) we see that F [u(x, 0)] = F [f (x)] = C so that 2 F [u] = fˆ(p) · e−kp t . 23 Now we can apply the convolution theorem (Thm. 7.1): 2 u(x, t) = f (x) ∗ F −1 [e−kp t ] From (??) we know that r r π −p2 /4α α −αx2 −1 −p2 /(4α) −αx2 e , or F [e ]= e F [e ]= α π With α = 1/4kt, it implies F −1 [e −kp2 t ]= r 1 −x2 /4kt e . 4kπt Then the solution is the convolution product Z +∞ 1 1 2 −x2 /4kt u(x, t) = f (x) ∗ √ = f (z) √ e e−(x−z) /4kt dz. 4πkt 4πkt −∞ Notice that in this formula the well known ”heat kernel” appears: G(z; t) = √ 1 2 e−z /4kt 4πkt which coincides with a Gaussian probability density with mean zero and variance 2kt. So we have three facts: a) the diffusion equation on the line with initial condition u(x, 0) = f (x) is solvable with a known explicit kernel: Z +∞ f (z)G(x − z; t)dz. u(x, t) = −∞ b) The physical interpretation of u(x, t) as the temperature at time t suggests that a choice of a compact support initial temperature f (x) gives rise to a temperature which is at once everywhere nonzero (!): u(x, t) 6= 0 ∀x ∈, ∀t > 0. An instantaneous diffusion to all points of space takes place. Setting σ 2 = 2kt, the convolution of any function 2 2 f (x) with the probability density σ√12π e−z /2σ is convergent to f as σ → 0. 24 Figure 9. Graph of the solution of the heat equation, starting from an initial function u(x, 0) with compact support, the indicator of (0, 10). Its diffusion is represented at times t = 0.01, t = 1, t = 10. Figure 10. Graph of the solution of the heat equation, starting from an initial function u(x, 0) with compact support, the indicator of (0, 10). Evidence of its ”diffusion” is given at times t = 0.01, t = 1, t = 10. 8. TEST FUNCTIONS AND DIRAC’S DELTA Let ϕ be a function of one real variable. We recall that the support of a function is the closure of the set where the function is nonzero: 25 supp(ϕ) = {x ∈ R : ϕ(x) 6= 0}. Also, we recall that a compact subset of R is a closed and bounded set. A test function is an infinitely differentiable function with compact support in R. The set of test functions is denoted C◦∞ (R). REMARK I.8A - Such functions do exist: an example is exp(−1/(1 − |x|2 ), |x| < 1 ϕ(x) = 0, |x| ≥ 1. Although the derivatives of 1/(1 − |x|2 ) are divergent as x → 1− (as x− → −1+ ) , we see that all derivatives of ϕ exist and are zero in 1− (in −1+ ) because they contain the exponentially small factor exp(−1/(1 − |x|2 ). So ϕ ∈ C◦∞ (R) and supp(ϕ) = {x ∈ Rn : |x| ≤ 1} = B1 (0), i.e. the support is the unitary ball centered at the origin. Similar test functions can be written with support coinciding with a given arbitrary compact set of R. DEF. I.8B - (The space of test functions) In the set C◦∞ (R), of infinitely differentiable functions with compact support, a sequence {ϕn } is said to converge to ϕ ∈ C◦∞ if (i) there is a compact set K such that supp(ϕn ) ⊂ K, ∀n ∈ N (k) (ii) the sequence {ϕn }n∈N of the k−th order derivatives is uniformly convergent to ϕ(k) as n → ∞, for k = 0, 1, 2... The vector space C◦∞ (R), when endowed with such notion of convergence, is said ”the space of test functions”’ and is denoted by D(R). Notice that the space D(R) is not normalizable, i.e. it is impossible to introduce a norm compatible with the above type of convergence. DEF. I.8C - (The space of distributions) Let D(R) be the space of test functions on R. A map T : D(R) → R is said a linear functional if < T, (λ1 ϕ1 + λ2 ϕ2 ) > = λ1 < T, ϕ1 > +λ2 < T, ϕ2 > . T is continuous if ϕn → ϕ in D(R), =⇒ < T, ϕn >→< T, ϕ > in R. The space D∗ (R) of linear and continuous functionals on D(R) is said the ”space of distributions” and each element T ∈ D∗ (R) is said a ”distribution”. D∗ (R) is endowed with a topology by means of the notion of weak convergence: 26 EXAMPLE I.8D - (Regular distributions). Let f be a ”locally integrable”’ real function on R: Z 1 f ∈ Lloc (R), i.e. ∃ |f (x)| dx < +∞, ∀ compact K ⊂ R. K The map Tf : D(R) → R is defined by Z < Tf , ϕ >= f · ϕ dx, ϕ ∈ D(R). Ω Such a map is well defined since ϕ has compact support in R. Moreover Tf is linear and continuous: indeed, if ϕn → ϕ in D(R), there exists a compact K ⊂ R such that all ϕn vanish out of K. Since the convergence is uniform in K, Z Z | < Tf , ϕn > − < Tf , ϕ > | = | (f ϕn − f ϕ)dx| ≤ ǫ |f |dx R K if only n is larger than some nǫ . Therefore < Tf , ϕn >→< Tf , ϕ >, i.e. Tf is continuous. Therefore Tf is a distribution, Tf ∈ D∗ (R). Distributions of this kind are said ”regular”. All the remaining distributions, which do not admit a representation by means of L1loc functions, are said ”singular”. DEFINITION I.8E - Dirac’s delta - Let a ∈ R. Dirac’s delta, in the point a, is the linear functional by which the value of ϕ in a is associated to any test function ϕ ∈ D(R): < δa , ϕ >:= ϕ(a). REMARK I.8F - Such a linear functional on D(R) is a distribution in the sense of Def. I.8C. Indeed it is continuous: if ϕn → ϕ in D(R), a fortiori the point convergence holds, so that < δa , ϕn >≡ ϕn (a) → ϕ(a) ≡< δa , ϕ > as n → ∞ In case a = 0, it is simply said ”Dirac’s delta”: < δ, ϕ >= ϕ(0), ∀ϕ ∈ D(R) . It is a ”singular” distribution in Rthe sense that there exists no function f ∈ L1loc (R) such that < δ, ϕ >= R f (x)ϕ(x). Nevertheless it is usual to write δ(x) , just as if it were a function, in a conventional notation: Z ∞ Z ∞ δ(x − a)ϕ(x)dx = ϕ(a). δ(x)ϕ(x)dx = ϕ(0), −∞ −∞ This notation is currently accepted, but in the sense of definitions I.8C, I.8E. Such a notation can be understood by approximants, too: THM. I.8G - Sequence of approximants of Dirac’s delta - 27 Let fRn be a sequence of probability densities, i.e. fn ∈ L1loc , with fn ≥ 0, and R fn (x)dx = 1, ∀n ∈ N . Let, moreover, fn (x) → 0 ∀x 6= 0. Then Z lim fn (x)ϕ(x)dx = ϕ(0), as n → ∞. n→+∞ R The same is true for other sequences, such as: fn = i e−inx πx . Proof. For the proof we restrict to a simple example of approximants: n n , x ∈ [− n1 , n1 ] 2 fn (x) = I[−1/n,1/n] (x) = 0, elsewhere. 2 Indeed n min ϕ(x) = m n ≤ 1 1 2 [− n ,+ n ] Z 1 +n 1 −n ϕ(x)dx ≤ Mn = max ϕ(x) 1 1 [− n ,+ n ] where, by continuity in 0, lim mn = ϕ(0), n→∞ whence the formula. lim Mn = ϕ(0), n→+∞ △ In the proper sense (I.5A) we know that the Fourier transform of functions such as constants or cosine functions do not exist. But such transforms can make sense in the wider framework of distributions. To this end the formal use of usual properties (linearity, frequency shifting, etc. ) gives the right results: EXAMPLES I.8H i) F [δ] ≡ Z Z e−iωt δ(t)dt = 1, F −1 [1] = δ. R e−iωt dt ≡ F [1] = 2πδ(ω), F −1 [2πδ(ω)] = 1. R Z iω◦ t iii)F [e ] = e−i(ω−ω◦ )t dt ≡ F [1](ω − ω◦ ) = 2πδ(ω − ω◦ ) ii) R 1 1 iv)F [cos(ω◦ t)] = F (eiω◦ t ) + F (e−iω◦ t ) 2 2 = π[δ(ω − ω◦ ) + δ(ω + ω◦ )]. (Compare (iv) with the Fourier transform of the truncated cosine function sin[a(ω − ω◦ )] sin[a(ω + ω◦ )] + , F [pa (t) cos(ω◦ t)] = ω − ω◦ ω + ω◦ which tends to zero, as a → ∞, ∀ω 6= ±ω◦ ). 28 In order to understand how to generate a derivative of a distribution, let us consider the simple case of a differentiable function f , at least with continuous derivative in an open connected set R. We consider the distribution Tf ′ associated with f ′ (x). For any ϕ ∈ D(R), take a, b such that supp(ϕ) ⊂ [a, b] ⊂ R; integrating by parts we obtain: Z Z b ′ b < Tf ′ , ϕ >= f (x)ϕ(x) dx = [f (x)ϕ(x)]a − f (x)ϕ′ (x) dx R a ′ Taking into account that ϕ ∈ D(R), it follows that < Tf ′ , ϕ > = − < f, ϕ′ > is the distribution which is determined by f ′ . DEFINITION I.8I - The derivative of a distribution. Let r = 1, 2, 3, .... The linear map from D∗ (R) to D∗ (R) defined by T → Dr T, with < Dr T, ϕ, >= (−1)r < T, Dr ϕ > is said r−th derivative, and Dr T is the r-th order derivative of the distribution T. EXAMPLE I.8L- The derivatives of the Heaviside function, of δ and δ′. Let us consider the Heaviside function 1, x>0 H(x) = 0, elsewhere. The corrspoinding distribution < Tf , ϕ >= has the derivative: ′ < DTf , ϕ >= − < Tf , ϕ >= − Z ∞ ϕ(x) dx 0 Z ∞ ϕ′ (x) dx = ϕ(0) =< δ, ϕ > 0 Therefore the derivative of the Heaviside function is the Dirac’s delta distribution. The derivative of δ is given by: < Dδ, ϕ >= − < δ, ϕ′ >= −ϕ′ (0). The second derivative of delta is: < D2 δ, ϕ >= (−1)2 < δ, ϕ′′ >= ϕ′′ (0). and so on. EX. I.8M - Let f : R → R be continuous except in the points a1 < a2 < ... < an . For sake of simplicity, fix ideas with n = 2. Let f admit a continuous derivative in the intervals (−∞, a1 ), (a1 , a2 ), 29 + − (a2 , +∞). Let, moreover, exists the finite limits f (a− 1 ), f (a1 ), f (a2 ), f (a+ 2 ). f is locally integrable, so the distribution Tf makes sense: < Tf , ϕ >= Z f (x)ϕ(x) dx = R s Z X k=0 ak+1 f (x)ϕ(x) dx ak The derivative is: ′ < DTf , ϕ >= − < Tf , ϕ >= − = 1 −[f ϕ]a−∞ Z R−{a1 ,a2 } + Z Z a1 ′ −∞ ϕf − [f ϕ]aa21 + a1 ′ −∞ Z fϕ − Z a2 ′ a1 fϕ − a2 ′ a1 ϕf − [f ϕ]+∞ a2 + Z Z ∞ f ϕ′ a2 +∞ ϕf ′ = a2 − + − ϕf ′ dx + ϕ(a1 )[f (a+ 1 ) − f (a1 )] + ϕ(a2 )[f (a2 ) − f (a2 )]. Therefore n X d − Tf = Tf ′ + [f (a+ k ) − f (ak )]δak . dx k=1 In probability, for example, this means that the derivative of a discrete distribution function F (x) is a ”probability density” in the distributional sense: n X d F (x) = F ′ (x) · IR−{a1 ,...,an } + P (X = ak ) · δ(x − ak ). dx k=1 30 fˆ(ω) f (t) c1 f1 (t) + c2 f2 (t) c1 fˆ1 (ω) + c2 fˆ2 (ω) f ′ (t) iω fˆ(ω) f (n) (t) (iω)n fˆ(ω) tn f (t) in d dωf (ω) n f (t − t◦ ) e−iωt◦ fˆ(ω) eiω◦ t f (t) (f1 ∗ f2 )(t) n ˆ fˆ(ω − ω◦ ) fˆ1 (ω) · fˆ2 (ω) f1 · f2 1 ˆ (f 2π 1 ∗ fˆ2 )(ω) sin(ω◦ t)/t pa (t) cos(ω◦ t) fˆ(ω) = π · I(−ω◦ ,ω◦ ) (ω) p π −ω2 /(4α) e α δ(t) 1 1 2πδ(ω) δ(t − a) e−iωa e−iω◦ t 2πδ(ω + ω◦ ) pa (t) = I(−a,a) (t) 2 sin(aω)/ω e−αt 2 eiω◦ t cos(ω◦ t) H(t) = I(0,∞) (t) sin[a(ω−ω◦ )] ω−ω◦ + sin[a(ω+ω◦ )] ω+ω◦ 2πδ(ω − ω◦ ) π[δ(ω − ω◦ ) + δ(ω + ω◦ )] πδ(t) + 1/(iω)