Professional Documents
Culture Documents
14 Fourier Integral Fourier Transform (Zill, Dennis G) ..
14 Fourier Integral Fourier Transform (Zill, Dennis G) ..
14 Fourier Integral Fourier Transform (Zill, Dennis G) ..
diffusion process in which the random movement of gas that c(x, 0) 0, x 0, and limx6T c(x, t) 0, t 0.
molecules transports the chemical away from its source. (b) Use a CAS to plot the graph of the solution in part (a) for
FIGURE 15.2.4 shows an ant emitting an alarm chemical into x 0 at the fixed times t 0.1, t 0.5, t 1, t 2,
the still air of a tunnel. If c(x, t) denotes the concentration t 5.
of the chemical x centimeters from the source at time t, then (c) For a fixed time t, show that H0Tc(x, t) dx Ak. Thus Ak
c(x, t) satisfies represents the total amount of chemical discharged.
2c c
k ,¬x 0, t 0,
x 2 t x
0
and k is a positive constant. The emission of pheromones as
FIGURE 15.2.4 Ants in Problem 33
a discrete pulse gives rise to a boundary condition of the
form
u
AG(t),
x x 0
where G(t) is the Dirac delta function.
From Fourier Series to Fourier Integral Suppose a function f is defined on (p, p).
If we use the integral definitions of the coefficients (9), (10), and (11) of Section 12.2 in (8) of that
section, then the Fourier series of f on the interval is
p p p
1 1 T nS nS nS nS
f (x)
2p
p
f (t) dt
p n3
1
ca
p
f (t) cos
p
tdtb cos
p
x a p
f (t) sin
p
t dtb sin
p
xd . (1)
p p p
1 1 T
f (x)
2S
a p
f (t) dtb 'D
S n3
1
ca p
f (t) cosDnt dtb cosDn x a
p
f (t)sinDnt dtb sinDn xd 'D. (2)
We now expand the interval (p, p) by letting p 6 T. Since p 6 T implies that D 6 0, the
T
limit of (2) has the form limD60 8 n 1F(Dn)D, which is suggestive of the definition of the
integral H0T F(D) dD. Thus if HT
T
f (t) dt exists, the limit of the first term in (2) is zero and the limit
of the sum becomes
T T T
1
f (x)
S
0
ca
T
f (t) cosDt dtb cosDx a T
f (t) sinDt dtb sinDxd dD. (3)
The result given in (3) is called the Fourier integral of f on the interval (T, T). As the fol-
lowing summary shows, the basic structure of the Fourier integral is reminiscent of that of a
Fourier series.
T
where A(D) T
f (x) cosDx dx (5)
T
B(D)
T
f (x) sinDx dx. (6)
y SOLUTION The function, whose graph is shown in FIGURE 15.3.1, satisfies the hypotheses of
Theorem 15.3.1. Hence from (5) and (6) we have at once
1
T
x
A(D)
T
f (x) cosDx dx
2 0 2 T
2
sin2D
cos Dx dx
0
D
T 2
1 cos2D
B(D)
T
f (x) sinDx dx sin Dx dx
0
D
.
T
*This means that the integral T
= f (x)= dx converges.
Cosine and Sine Integrals When f is an even function on the interval (T, T), then
the product f (x) cos Dx is also an even function, whereas f (x) sin Dx is an odd function. As a
consequence of property (g) of Theorem 12.3.1, B(D) 0, and so (4) becomes
T T
2
f (x)
S
0
a
0
f (t) cosDt dtb cosDx dD.
T
f (t) cosDt dt 2
0
f (t) cosDt dt.
Similarly, when f is an odd function on (T, T) the products f (x) cos Dx and f (x) sin Dx are odd
and even functions, respectively. Therefore A(D) 0 and
T T
2
f (x)
S
0
a
0
f (t) sinDt dtb sinDx dD.
T
where A(D) 0
f (x) cosDx dx. (9)
(ii) The Fourier integral of an odd function on the interval (T, T) is the sine integral
T
2
f (x)
S 0
B(D) sinDx dD, (10)
T
where B(D) 0
f (x) sinDx dx. (11)
x
A(D) 0
f (x) cosDx dx 0
f (x) cosDx dx a
f (x) cosDx dx
–a a a
sinaD
FIGURE 15.3.2 Function f in Example 2
cos Dx dx
0
D
,
T
2 sinaD cosDx
and so f (x)
S 0 D
dD. (12)
FIGURE 15.3.4 In Example 3, (a) is the FIGURE 15.3.4 shows the graphs of the functions and their extensions represented by the two
even extension of f ; (b) is the odd integrals.
extension of f
Complex Form The Fourier integral (4) also possesses an equivalent complex form, or
exponential form, that is analogous to the complex form of a Fourier series (see Section 12.4).
If (5) and (6) are substituted into (4), then
T T
1
f (x)
S
0 T
f (t)I cosDt cosDx sinDt sinDxJ dt dD
T T
1
S
0 T
f (t) cosD(t x) dt dD
T T
1
2S T T
f (t) cosD(t x) dt dD (15)
T T
1
2S T T
f (t)I cosD(t x) i sinD(t x)J dt dD (16)
T T
1
2S T T
f (t) e iD(t x) dt dD
T T
1
2S T
a
T
f (t) e iDt dtb eiDx dD. (17)
We note that (15) follows from the fact that the integrand is an even function of D. In (16) we
have simply added zero to the integrand,
T T
i
T T
f (t) sinD(t x) dt dD 0,
T
where C(D) T
f (x)e iDx dx. (19)
This latter form of the Fourier integral will be put to use in the next section when we return
y
1.5 to the solution of boundary-value problems.
1
0.5 Use of Computers The convergence of a Fourier integral can be examined in a manner
0 x
that is similar to graphing partial sums of a Fourier series. To illustrate, let’s use the results in
parts (a) and (b) of Example 3. By definition of an improper integral, the Fourier cosine integral
–0.5
representation of f (x) ex, x 0 in (13) can be written as f (x) limb 6T Fb(x), where
–1
b
2 cosDx
–3 –2 –1 0
(a) F20(x)
1 2 3
Fb(x)
S 0 1 D2
dD,
1.5
y and x is treated as a parameter. Similarly, the Fourier sine integral representation of f (x) ex,
1 x 0 in (14) can be written as f (x) limb 6TGb(x), where
0.5 b
2 D sinDx
0
–0.5
x Gb(x)
S 0 1 D2
dD.
–1
Because the Fourier integrals (13) and (14) converge, the graphs of the partial integrals Fb(x)
–3 –2 –1 0 1 2 3 and Gb(x) for a specified value of b 0 will be an approximation to the graph of f and its
(b) G20(x) even and odd extensions shown in Figure 15.3.4(a) and 15.3.4(b), respectively. The graphs of
Fb(x) and Gb(x) for b 20 given in FIGURE 15.3.5 were obtained using Mathematica and its
FIGURE 15.3.5 Graphs of partial integrals NIntegrate application. See Problem 21 in Exercises 15.3.
0, x0 e x, =x= 1 In Problems 17 and 18, solve the given integral equation for the
5. f (x) e x 6. f (x) e
e , x0 0, =x= 1 function f.
T
In Problems 7–12, represent the given function by an
appropriate cosine or sine integral.
17. 0
f (x) cos Dx dx eD
[Hint: D is a dummy variable of integration.] (b) As a consequence of part (a), f (x) lim Fb(x), where
b6T
T b
sinkx S sinD(x 1) sinD(x 1)
(b) Show in general that, for k 0, 0 x
dx .
2 Fb(x)
1
S 0 D
dD.
20. Use the complex form (19) to find the Fourier integral Show that the last integral can be written as
representation of f (x) e =x= . Show that the result is the same
1
as that obtained from (8) and (9). Fb(x) ISi(b (x 1)) Si(b (x 1))J,
S
Computer Lab Assignment where Si(x) is the sine integral function. See Problem 43
21. While the integral (12) can be graphed in the same manner in Exercises 2.3.
discussed on page 781 to obtain Figure 15.3.5, it can also be (c) Use a CAS and the sine integral form obtained in part (b)
expressed in terms of a special function that is built into to graph Fb(x) on the interval [3, 3] for b 4, 6, and 15.
a CAS. Then graph Fb(x) for larger values of b 0.
INTRODUCTION Up to now we have studied and used only one integral transform: the
Laplace transform. But in Section 15.3 we saw that the Fourier integral had three alternative
forms: the cosine integral, the sine integral, and the complex or exponential form. In the present
section we shall take these three forms of the Fourier integral and develop them into three new
integral transforms naturally called Fourier transforms. In addition, we shall expand on the
concept of a transform pair, that is, an integral transform and its inverse. We shall also see that
the inverse of an integral transform is itself another integral transform.
Transform Pairs The Laplace transform F(s) of a function f (t) is defined by an integral,
but up to now we have been using the symbolic representation f (t)
1{F(s)} to denote the
inverse Laplace transform of F(s). Actually, the inverse Laplace transform is also an integral
transform. If
T
5 f (t)6 0
estf (t) dt F(s), (1)
The last integral is called a contour integral; its evaluation requires the use of complex variables
and is beyond the scope of this discussion. The point here is this: Integral transforms appear in
transform pairs. If f (x) is transformed into F(D) by an integral transform
b
F(D) f (x) K(D, x) dx,
a
(3)
called the inverse transform. The functions K and H in the integrands of (3) and (4) are called
the kernels of their respective transforms. We identify K(s, t) est as the kernel of the Laplace
transform and H(s, t) est /2Si as the kernel of the inverse Laplace transform.
T
1
Inverse Fourier transform: @1 5F(D)6
2S T
F(D) e iDx dD f (x) (6)
T
(ii) Fourier sine transform: @s 5 f (x)6 0
f (x) sinDx dx F(D) (7)
Inverse Fourier T
2
sine transform: @1
s 5F(D)6
S 0
F(D) sinDx dD f (x) (8)
T
(iii) Fourier cosine transform: @c 5 f (x)6 0
f (x) cosDx dx F(D) (9)
Inverse Fourier T
2
cosine transform: @1
c 5F(D)6
S 0
F(D)cosDx dD f (x) (10)
Existence The conditions under which (5), (7), and (9) exist are more stringent than those
for the Laplace transform. For example, you should verify that @{1}, @ s{1}, and @ c{1} do not
exist. Sufficient conditions for existence are that f be absolutely integrable on the appropriate
interval and that f and f be piecewise continuous on every finite interval.
Operational Properties Since our immediate goal is to apply these new transforms to
boundary-value problems, we need to examine the transforms of derivatives.
Fourier Transform Suppose that f is continuous and absolutely integrable on the inter-
val (T, T) and f is piecewise continuous on every finite interval. If f (x) 6 0 as x 6 T, then
integration by parts gives
T
@5 f (x)6
T
f (x) e iDx dx
T T
f (x) e iDx d
T
iD T
f (x) e iDx dx
T
iD T
f (x) e iDx dx;
Similarly, under the added assumptions that f is continuous on (T, T), f (x) is piecewise
continuous on every finite interval, and f (x) 6 0 as x 6 T, we have
@{ f (x)} (iD)2 @{ f (x)} D2F(D). (12)
In general, under conditions analogous to those leading to (12), we have
Fourier Sine Transform Suppose that f and f are continuous, f is absolutely integrable
on the interval [0, T), and f is piecewise continuous on every finite interval. If f 6 0 and f 6 0
as x 6 T, then
T
@s 5 f (x)6
0
f (x) sinDx dx
T T
f (x) sinDxd
0
D
0
f (x)cosDx dx
T T
D cf (x)cosDx `
0
D 0
f (x)sinDx dxd
Fourier Cosine Transform Under the same assumptions that lead to (9), we find the
Fourier cosine transform of f (x) to be
The nature of the transform properties (12), (13), and (14) indicate, in contrast to the Laplace
transform, that Fourier transforms are suitable for problems in which the spatial variable x (or y)
is defined on an infinite or semi-infinite interval. But a natural question then arises:
How do we know which transform to use on a given boundary-value problem?
Clearly, to use the Fourier transform (5), the domain of the variable to eliminate must be
(T, T). To utilize a sine or cosine transform, the domain of at least one of the spatial variables
in the problem must be [0, T). However, the determining factor in choosing between the sine
transform (7) and the cosine transform (9) is the type of boundary condition specified at x 0
(or y 0), that is, whether u or its first partial derivative is given at this boundary.
In solving boundary-value problems using integral transforms most solutions are formal.
In the language of mathematics, this means assumptions about the solution u and its partial
These assumptions are derivatives go unstated. But one assumption should be kept in the back of your mind. In the
sometimes used during the examples that follow, it will be assumed without further mention that u and u> x (or u> y)
actual solution process. See
Problems 13, 14, and 26 in approach 0 as x 6T (or y 6T). These are not major restrictions since these conditions
Exercises 15.4. hold in most applications.
If we write
2u u d dU
@e f D2U(D, t)¬¬and¬¬@ e f @5u(x, t)6 ,
x 2 t dt dt
2u u
f @ek f @e f ,
u0 x 2 t
becomes the ordinary differential equation
dU dU
x kD2U(D, t) or kD2U(D, t) 0.
–1 1 dt dt
Solving the last equation by the method of Section 2.3 gives U(D, t) ce kD t. The initial
2
By Euler’s formula
e iD cos D i sin D
e iD cos D i sin D.
e iD e iD
Subtracting these two results and solving for sin D gives sin D . Hence we can
2i
sin D
rewrite the transform of the initial condition as U(D, 0) 2u0 . Applying this condition
D
sin D
to the solution U(D, t) ce kD t gives U(D, 0) c 2u0
2
and so
D
sinD kD2t
U(D, t) 2u0 e .
D
It then follows from the inverse Fourier transform (6) that
T
u0 sinD kD2t iDx
u(x, t)
S T D
e e dD.
This integral can be simplified somewhat by using Euler’s formula again as eiDx
cos Dx sin Dx and noting that
T
sinD kD2t
T D
e sin Dx dD 0
because the integrand is an odd function of D. Hence we finally have the solution
T
u0 sinD cosDx kD2t
u(x, t)
S T D
e dD. (15)
It is left to the reader to show that the solution (15) in Example 1 can be expressed in terms
of the error function. See Problem 23 in Exercises 15.4.
T
b
@c 5e bx6 0
e bx cos Dx dx
b 2 D2
. (17)
Another way to quickly obtain (and remember) these two results is to identify the two integrals
with the more familiar Laplace transform in (2) of Section 4.1. With the symbols x, b, and
D playing the part of t, s, and k, respectively, it follows that (16) and (17) are identical to (d)
and (e) in Theorem 4.1.1.
u
0,¬0 x S.
y y 0
2u 2u
In view of (14), @c e 2
f @ c e 2 f @ c{0}
x y
becomes
d 2U d 2U
D 2
U(x, D) u y(x, 0) 0¬¬or¬¬ D 2U 0.
dx 2 dx 2
Since the domain of x is a finite interval, we choose to write the solution of the ordinary
differential equation as
U(x, D) c1 cosh Dx c2 sinh Dx. (18)
Now @ c{u(0, y)} @ c{0} and @ c{u(S, y)} @ c{ey} are in turn equivalent to
1
U(0, D) 0 and U(S, D) . (19)
1 D2
Note that the value U(S, D) in (19) is (17) of Example 2 with b 1. When we apply the two
conditions in (19) to the solution (18) we obtain c1 0 and c2 1[(1 D 2) sinh D]. Therefore,
sinhDx
U(x, D) ,
(1 D2 ) sinhDS
and so from (10) we arrive at
T
2 sinhDx
u(x, y)
S
0 (1 D2 )sinhDS
cosDy dD. (20)
Had u(x, 0) been given in Example 3 rather than uy(x, 0), then the sine transform would have
been appropriate.
u
A,¬t 0.
x x 0 x
u = e –x
7. Solve Problem 5 if the end x 0 is insulated. FIGURE 15.4.2 Infinite plate in Problem 17
8. Find the temperature u(x, t) in a semi-infinite rod if u(0, t) 1, y
t 0, and u(x, 0) ex, x 0. 18.
2u u u=0
9. (a) a 2 , T x T, ¬t 0
x 2 t u = e –y
u 1
u(x, 0) f (x),¬ g(x), T x T
t t 0 u = 100 x
0 π
u = f (x)
(b) If g(x) 0, show that the solution of part (a) can be written
as u(x, t) 12 [ f (x at) f (x at)]. FIGURE 15.4.3 Semi-infinite plate in Problem 18
10. Find the displacement u(x, t) of a semi-infinite string if
u(0, t) 0, t0 19. Use the result @ 5ex >4p 6 2"Spep D to solve the
2 2 2 2
boundary-value problem
u
u(x, 0) xex,¬ 0, x 0.
t t 0 2u u
k , T x T,¬t 0
11. Solve the problem in Example 3 if the boundary conditions x 2 t
at x 0 and x S are reversed: u(x, 0) ex , T x T.
2
u(0, y) ey, u(S, y) 0, y 0. 20. If @{ f (x)} F(D) and @{g(x)} G(D), then the convolution
theorem for the Fourier transform is given by
12. Solve the problem in Example 3 if the boundary condition at
y 0 is u(x, 0) 1, 0
x
S. T
50, 0 x 1.
u(x, 0) e Problem 19 to show that a solution of the boundary-value
0, x 1. problem
14. Solve Problem 13 if the boundary condition at x 0 is 2u u
u(0, y) 0, y 0. k 2
, T x T,¬t 0
x t
2u 2u
15. 2
2 0, x 0, 0 y 2 u(x, 0) f (x), T x T
x y
T
u(0, y) 0, 0
y
2 1
u(x, 0) f (x), u(x, 2) 0, x 0
is u(x, t)
2"kSt
T
f (W) e(x W) >4kt dW.
2
3 f (x) G (x nT ). (2)
n T
T T
T n T
iDx
3 f (x) G (x nT )e dx. (3)
By the sifting property of the Dirac delta function (see the Remarks at the end of Section 4.5),
(3) is the same as
T
F(D) 3 f (nT )e iDnT. (4)
n T
The expression F(D) in (4) is called the discrete Fourier transform (DFT) of the function f.
We often write the coefficients f (nT ) in (4) as f (n) or fn. It is also worth noting that since eiDx is
periodic in D and eiDT ei(DT2S) ei(D2S/T)T, we only need to consider the function for D in
[0, 2S/T ]. Let N 2S/T. This places x in the interval [0, 2S]. So, because we sample over one
period, the sum in (4) is actually finite.
Now consider the function values f (x) at N equally spaced points, x nT, n 0, 1, 2, … ,
N 1, in the interval [0, 2S]; that is, f0, f1, f2, … , fN1. The (finite) discrete Fourier series
T
f (x) 8 n T cne inx using these N terms gives us
2S 2S
If we let Yn ei2S/n cos i sin and use the usual laws of exponents, this system of
n n
equations is the same as
f0 c0 c1 c2 S cN 1
f1 c0 c1YN c2Y2N S cN 1YNN 1
f2 c0 c1Y2N c2Y4N S cN 1Y2(N
N
1)
(5)
1)2
fN 1 c0 c1YNN 1 c2YN2(N 1) S cN 1Y(N
N .
*Note that the symbol T used here does not have the same meaning as in Section 12.4.
f0 1 1 1 S 1 c0
f1 1 YN Y2N YNN 1 c1
7 f2 ? 7 1 Y2N Y4N YN2(N 1)? 7 c2 ? . (6)
fN 1 1 YNN 1 Y2(N
N
1) S YN(N 1)2
cN 1
Let the N N matrix in (6) be denoted by the symbol FN. Given the inputs f0, f1, f2, … , fN1, is
there an easy way to find the Fourier coefficients c0, c1, c2, … , cN1? If F N is the matrix consist-
ing of the complex conjugates of the entries of FN and if I denotes the N N identity matrix,
then we have
1
FN FN FN FN N I¬¬and so¬¬F 1
N F .
N N
It follows from (6) and the last equation that
c0 f0
c1 f1
1
7 c2 ? FN 7 f2 ? .
N
cN 1 fN 1
Discrete Transform Pair Recall from Section 15.4 that in the Fourier transform pair
we use a function f (x) as input and compute the coefficients that give the amplitude for each
frequency k (ck in the case of periodic functions of period 2S) or we compute the coefficients
that give the amplitude for each frequency D (F(D) in the case of nonperiodic functions).
Also, given these frequencies and coefficients, we could reconstruct the original function f (x).
In the discrete case, we use a sample of N values of the function f (x) as input and compute the
coefficients that give the amplitude for each sampled frequency. Given these frequencies and
coefficients, it is possible to reconstruct the n sampled values of f (x). The transform pair, the
discrete Fourier transform pair, is given by
1
c F f¬¬and¬¬f FN c (7)
N N
c0 f0
c1 f1
where c 7 c2 ?¬¬and¬¬f 7 f2 ? .
cN 1 fN 1
Heat Equation and Discrete Fourier Series If the function f in the initial-value
problem
2u u
k ,¬T x T, t 0
x 2 t (8)
u(x, 0) f (x), T x T
is periodic with period 2S, the solution can be written in terms of a Fourier series for f (x). We
can also approximate this solution with a finite sum
n1
u(x, t) 3 ck(t) e ikx.
k0
If we examine both sides of the one-dimensional heat equation in (8), we see that
n 1 dc
u j ijx
3 e
t j 0 dt
n1
2u
and k k 2 ijx
3 cj (t)(i j) e ,
x 2 j 0
d 2e ijx
since (i j)2e ijx .
dx 2
Equating these last two expressions, we have the first-order DE
dcj
k j 2cj (t)¬with solution¬cj (t) cj (0) e k j t.
2
dt
n1
The final task is to find the values c j(0). However, recall that u(x, t) 8 k 0 ck (t)eikx and
u(x, 0) f (x), so cj(0) are the coefficients of the discrete Fourier series of f (x). Compare this
with Section 13.3.
Heat Equation and Discrete Fourier Transform The initial-value problem (8)
can be interpreted as the mathematical model for the temperature u(x, t) in an infinitely long bar.
In Section 15.4 we saw that we can solve (8) using the Fourier transform and that the solution
u(x, t) depends on the Fourier transform F(D) of f (x) (see pages 784–785). We can approximate
F(D) by taking a different look at the discrete Fourier transform.
First we approximate values of the transform by discretizing the integral @{ f (x)} F(D)
T iDx
HT f (x) e dx. Consider an interval [a, b]. Let f (x) be given at n equally spaced points
ba
xj a j, j 0, 1, 2, … , n 1.
n
b a n1 ba
3 f aa jb e iDxj
n j 0 n
n j 0 n
n j 0 n
2SM
If we now choose a convenient value for D, say with M an integer, we have
ba
2SM b a i 2SMa n 1 ba 2SjM
Fa b e b a 3 f aa jbe i n
ba n j 0 n
b a i 2SMa n 1 ba
e b a 3 f aa jb Y jM
n ,
n j 0 n
where, recall that Yn ei2S/n. This is a numerical approximation to the Fourier transform of f (x)
2SM
evaluated at points with M an integer.
ba
|F(α )| |F(α )|
2 2
1.75 1.75
1.5 1.5
1.25 1.25
1 1
0.75 0.75
0.5 0.5
0.25 0.25
α α
–10 –5 5 10 –10 –5 5 10
(a) (b)
FIGURE 15.5.3 In Example 2 (a) is the graph of |F(D)|; (b) is the discrete Fourier
transform of f
Aliasing A problem known as aliasing may appear whenever one is sampling data at
equally spaced intervals. If you have ever seen a motion picture where rotating wheels appear to
be rotating slowly (or even backwards!), you have experienced aliasing. The wheels may rotate
(a) y = sin 20π x; x range: [0, 1]; y range: [–1, 1] (b) y = sin 100π x; x range: [0, 1]; y range: [–1, 1]
at a high rate, but because the frames in a motion picture are “sampled” at equally spaced inter-
vals, we see a low rate of rotation.
Graphing calculators also suffer from aliasing due to the way they sample points to create
graphs. For example, plot the trigonometric function y sin 20Sx with frequency 10 on a Texas
Instruments TI-92 and you get the nice graph in FIGURE 15.5.4(a). At higher frequencies, say
y sin 100Sx with frequency 50, you get the correct amount of cycles, but the amplitudes of
the graph in Figure 15.5.4(b) are clearly not 1.
On a calculator such as the Texas Instruments TI-83, the graphs in FIGURE 15.5.5 show aliasing
much more clearly.
The problem lies in the fact that e2nSi cos 2nS i sin 2nS 1 for all integer values of n.
The discrete Fourier series cannot distinguish einx from 1 as these functions are equal at sampled
points x 2kS>n. The higher frequency is seen as the lower one. Consider the functions cos (Sn>2)
and cos (7Sn>2). If we sample at the points n 0, 1, 2, … , these two functions appear the same,
the lower frequency is assumed, and the amplitudes (Fourier coefficients) associated with the
higher frequencies are added in with the amplitude of the low frequency. If these Fourier coef-
ficients at large frequencies are small, however, we do not have a big problem. In the Sampling
Theorem below, we will see what can take care of this problem.
Signal Processing Beyond solving PDEs as we have done earlier, the ideas of this sec-
tion are useful in signal processing. Consider the functions we have been dealing with as signals
from a source. We would like to reconstruct a signal transmitted by sampling it at discrete points.
The problem of calculating an infinite number of Fourier coefficients and summing an infinite
series to reconstruct a signal (function) is not practical. A finite sum could be a decent approxi-
mation, but certain signals can be reconstructed by a finite number of samples.
Theorem 15.5.1 Sampling Theorem
If a signal f (x) is band-limited; that is, if the range of frequencies of the signal lie in a band
A
k
A, then the signal can be reconstructed by sampling two times for every cycle of
the highest frequency present; in fact,
T
nS sin(Ax nS)
f (x) 3 f a b .
n T A Ax nS
T
F(D) T
f (x)e iDx dx (9)
T A
1 1
f (x)
2S T
F(D)e iDx
dD
2S
A
F(D)eiDx dD. (10)
A
1
where cn
2A A
F(D)einSD>A dD. (12)
S nS
which by (12) is equal to cn. Substituting cn f a b into (11) yields
A A
T
S nS inSD>A
F(D) 3 f a b e .
n T A A
A T
1 S nS inSD>A iDx
f (x)
2S A
a 3
n T A
f a b e
A
be dD
A
1 T nS
2A n 3
T
f a b
A A
e inSD>A eiDx dD
A
1 T nS nS
3
2A n T
f a b
A A
e A
iD x
dD
1 T nS 1 nS nS
f a b ae iAa A xb
eiAa A xb
b
2A n 3
T A nS
ia xb
A
1 T nS 1
3 f a b 2isin (nS Ax)
2A n T A nS
ia xb
A
T
nS sin(nS Ax)
3 f a b
n T A nS Ax
T
nS sin(Ax nS)
3 f a b .
n T A Ax nS
Band-Limited Signals A signal that contains many frequencies can be filtered so that
only frequencies in an interval survive, and it becomes a band-limited signal. Consider the signal
f (x). Multiply the Fourier transform F(D) of f by a function G(D) that is 1 on the interval contain-
ing the frequencies D you wish to keep, and 0 elsewhere. This multiplication of two Fourier
transforms in the frequency domain is a convolution of f (x) and g(x) in the time domain. Recall
that Problem 20 in Exercises 15.4 states that
T
@ 1{F(D)G(D)}
T
f (W) g (x W) dW.
The integral on the right-hand side is called the convolution of f and g and is written f *g. The
last statement can be written more compactly as
@{ f *g} F(D)G(D).
sinAx
The analogous idea for Laplace transforms is in Section 4.4. The function g(x) has as
Sx
its Fourier transform the pulse function
1, A D A
G(D) e
0, elsewhere.
This implies that the function ( f * g)(x) is band-limited, with frequencies between A
and A.
Computing with the Fast Fourier Transform Return to the discrete Fourier trans-
form of f (x), where we have f sampled at n equally spaced points a distance of T apart, namely,
0, T, 2T, 3T, … , (n 1)T. (We used T S/n at the beginning of this section.) Substituting this,
the discrete Fourier transform
ba n j 0 n
n1
2Sk
becomes Fa b T 3 f ( jT )Ykjn,¬k 0, 1, 2, S , n 1.
nT j 0
n1
ck 3 fj Ykjn,¬k 0, 1, 2, S , n 1.
j 0
f0 1 1 1 S 1 c0
f1 1 Yn Y2n S Ynn 1 c1
7 f2 ? 7 1 Y2n Y4n S 2(n 1)
Yn ? 7 c2 ? ,
fn 1 1 n1
Yn Y2(n
n
1) S Yn(n 1)2
cn 1
or f Fnc. The key to the FFT is properties of Yn and matrix factorization. If n 2N, we can
write Fn in the following way (which we will not prove):
I 2N 1 D2N 1 F2N 1 0
F2N a ba b P, (13)
I 2N 1 D2N 1 0 F2N 1
where Ik is the k k identity matrix and P is the permutation matrix that rearranges the matrix c
so that the even subscripts are ordered on the top and the odd ones are ordered on the bottom.
The matrix D is a diagonal matrix defined by
1
Y 2N
D 2N 1 7 (Y2N)2 ?.
I
N1
1
(Y2N)2
Note that each of the F2N 1 matrices can, in turn, be factored. In the end, the matrix Fn with n2
nonzero entries is factored into the product of n simpler matrices at a great savings to the number
of computations needed on the computer.
EXAMPLE 3 The FF T
Let n 2 4 and let F4 be the matrix in Example 1:
2
1 1 1 1
1 i 1 i
F4 ± .
1 1 1 1
1 i 1 i
1 0 1 0 1 1 0 0 1 0 0 0
0 1 0 i 1 1 0 0 0 0 1 0
F4 ± ± ± . (14)
1 0 1 0 0 0 1 1 0 1 0 0
0 1 0 i 0 0 1 1 0 0 0 1
A B P
We have inserted dashed lines in the matrices marked A and B so that you can identify the
submatrices I2, D2, D2, and F2 by comparing (14) directly with (13). You are also encour-
3
5
aged to multiply out the right side of (14) and verify that you get F4. Now if c ± , then
8
20
1 0 1 0 1 1 0 0 3
0 1 0 i 1 1 0 0 8
± ± ±
1 0 1 0 0 0 1 1 5
0 1 0 i 0 0 1 1 20
1 0 1 0 11 36
0 1 0 i 5 5 15i
± ± ± f.
1 0 1 0 25 14
0 1 0 i 15 5 15i
1. Show that F 1
4 4 F4 .
1 7. Write the matrix F8 and then write it in factored form (13).
2. Prove the sifting property of the Dirac delta function: Verify that the product of the factors is F8. If instructed, use
a CAS to verify the result.
T
8. Let Yn ei2S/n cos (2S>n) i sin (2S>n). Since ei2Sk 1,
T
f (x) G(x a) dx f (a).
the numbers Ykn , k 0, 1, 2, … , n 1, all have the property
that (Ykn )n 1. Because of this, Ykn , k 0, 1, 2, … , n 1, are
[Hint: Consider the function called the nth roots of unity and are solutions of the equation
zn 1 0. Find the eighth roots of unity and plot them in the
1
, =x a= H xy-plane where a complex number is written z x iy. What
2H
GH(x a) | do you notice?
0, elsewhere.
Use the mean value theorem for integrals and then let 3 6 0.] Computer Lab Assignments
3. Find the Fourier transform of the Dirac delta function G(x). 9. Use a CAS to verify that the function f *g, where f (x) e5x
2
4. Show that the Dirac delta function is the identity under the sin2x
convolution operation; that is, show f *G G * f f. [Hint: and g(x) , is band-limited. If your CAS can handle
Sx
Use Fourier transforms and Problem 3.] it, plot the graphs of @{ f *g} and F(D)G(D) to verify the
5. Show that the derivative of the Dirac delta function G(x a) result.
has the property that it sifts out the derivative of a function f 10. If your CAS has a discrete Fourier transform command, choose
at a. [Hint: Use integration by parts.] any six points and compare the result obtained using this
6. Use a CAS to show that the Fourier transform of the function command with that obtained from c 16 F 6f.
sinAx
g(x) is the pulse function
Sx
1, A D A
G(D) e
0, elsewhere.