This action might not be possible to undo. Are you sure you want to continue?

Fourier Transforms

Fourier series and their ilk are designed to solve boundary value problems on bounded

intervals. The extension of Fourier methods to the entire real line leads naturally to

the Fourier transform, an extremely powerful mathematical tool for the analysis of non-

periodic functions. The Fourier transform is of fundamental importance in a broad range

of applications, including both ordinary and partial diﬀerential equations, quantum me-

chanics, signal processing, control theory, and probability, to name but a few.

In this chapter, we begin by motivating the construction by investigating how Fourier

series behave as the length of the interval goes to inﬁnity. The resulting Fourier transform

maps a function deﬁned on physical space to a function deﬁned on the space of frequencies,

whose values quantify the amount of each periodic frequency residing in the function.

The inverse Fourier transform reconstructs the original function from its transform. The

integrals deﬁning the Fourier transform and its inverse are remarkably alike, and this

symmetry is exploited when assembling tables of Fourier transforms.

One of the most important properties of the Fourier transform is that it converts

calculus: diﬀerentiation and integration — into algebra: multiplication and division. This

underlies its application to boundary value problems for linear ordinary diﬀerential equa-

tions and, in the following chapter, boundary value and initial-boundary value problems

for partial diﬀerential equations. In engineering applications, the Fourier transform is

sometimes overshadowed by the Laplace transform, which is a particular subcase. The

Laplace transform is better suited to solving initial value problems, [23, 40], but will not

be developed in this text.

The Fourier transform is, like Fourier series, completely compatible with the calculus

of generalized functions. The ﬁnal section contains a brief introduction to the analytical

foundations of the subject, including the basics of Hilbert space. However, a full, rigorous

development requires advanced tools from mathematical analysis, and the interested reader

is therefore referred to more advanced texts, e.g., [44, 75, 130].

7.1. The Fourier Transform.

We begin by motivating the Fourier transform as a limiting case of Fourier series.

Although the rigorous details are rather exacting, the underlying idea is not so diﬃcult. Let

f(x) be a reasonably nice function deﬁned for all −∞< x < ∞. The goal is to construct a

Fourier expansion for f(x) in terms of basic trigonometric functions. One evident approach

is to construct its Fourier series on progressively longer and longer intervals, and then

take the limit as the lengths go to inﬁnity. This limiting process converts the Fourier

1/24/10 304 c (2010 Peter J. Olver

sums into integrals, and the resulting representation of a function is renamed the Fourier

transform. Since we are dealing with an inﬁnite interval, there are no longer any periodicity

requirements on the function f(x). Moreover, the frequencies represented in the Fourier

transform are no longer constrained by the length of the interval, and so we are eﬀectively

decomposing a quite general, non-periodic function into a continuous superposition of

trigonometric functions of all possible frequencies.

Let us present the details of this construction in a more concrete form. The computa-

tions will be signiﬁcantly simpler if we work with the complex version of the Fourier series

from the outset. Our starting point is the rescaled Fourier series (3.84) on a symmetric

interval [ −ℓ , ℓ ] of length 2 ℓ, which we rewrite in the adapted form

f(x) ∼

∞

¸

ν =−∞

π

2

´

f

ℓ

(k

ν

)

ℓ

e

i k

ν

x

. (7.1)

The sum is over the discrete collection of frequencies

k

ν

=

πν

ℓ

, ν = 0, ±1, ±2, . . . , (7.2)

corresponding to those trigonometric functions that have period 2 ℓ. For reasons that will

soon become apparent, the Fourier coeﬃcients of f are now denoted as

c

ν

= ' f ; e

i k

ν

x

` =

1

2 ℓ

ℓ

−ℓ

f(x) e

−i k

ν

x

dx =

π

2

´

f

ℓ

(k

ν

)

ℓ

, (7.3)

so that

´

f

ℓ

(k

ν

) =

1

√

2π

ℓ

−ℓ

f(x) e

−i k

ν

x

dx. (7.4)

This reformulation of the basic Fourier series formula allows us to smoothly pass to the

limit when the interval’s length ℓ →∞.

On an interval of length 2 ℓ, the frequencies (7.2) required to represent a function in

Fourier series form are equally distributed, with interfrequency spacing

∆k = k

ν+1

−k

ν

=

π

ℓ

. (7.5)

As ℓ → ∞, the spacing ∆k → 0, and so the relevant frequencies become more and more

densely packed in the space of all possible frequencies: −∞ < k < ∞. In the limit, we

anticipate that all possible frequencies will be represented. Indeed, letting k

ν

= k be

arbitrary in (7.4), and sending ℓ →∞, results in the inﬁnite integral

´

f(k) =

1

√

2π

∞

−∞

f(x) e

−i kx

dx (7.6)

known as the Fourier transform of the function f(x). If f(x) is a reasonably nice function,

e.g., piecewise continuous and decaying to 0 reasonably quickly as [ x [ → ∞, its Fourier

transform

´

f(k) is deﬁned for all possible frequencies −∞ < k < ∞. This formula will

sometimes conveniently be abbreviated as

´

f(k) = T[ f(x)], (7.7)

1/24/10 305 c (2010 Peter J. Olver

where T is the Fourier transform operator.

To reconstruct the function from its Fourier transform, we employ a similar limiting

procedure on the Fourier series (7.1), which we ﬁrst rewrite in a more suggestive form,

f(x) ∼

1

√

2π

∞

¸

ν =−∞

´

f

ℓ

(k

ν

) e

i k

ν

x

∆k, (7.8)

using (7.5). For each ﬁxed value of x, the right hand side has the form of a Riemann sum,

[8, 112], approximating the integral

1

√

2π

∞

−∞

´

f

ℓ

(k) e

i kx

dk.

As ℓ →∞, the functions (7.4) converge to the Fourier transform:

´

f

ℓ

(k) →

´

f(k); moreover,

the interfrequency spacing ∆k = π/ℓ → 0, and so one expects the Riemann sums to

converge to the limiting integral

f(x) ∼

1

√

2π

∞

−∞

´

f(k) e

i kx

dk. (7.9)

The resulting formula serves to deﬁne the inverse Fourier transform, which is used to re-

cover the original signal from its Fourier transform. In this manner, the Fourier series has

become a Fourier integral that reconstructs the function f(x) as a (continuous) superposi-

tion of complex exponentials e

i kx

of all possible frequencies, with

´

f(k)/

√

2π prescribing

the amount of contribution of the complex exponential of frequency k. In abbreviated

form, formula (7.9) can be written

f(x) = T

−1

[

´

f(k)], (7.10)

thus deﬁning the inverse of the Fourier transform operator (7.7).

It is worth pointing out that both the Fourier transform (7.7) and its inverse (7.10)

deﬁne linear maps on function space. This means that the Fourier transform of the sum

of two functions is the sum of their individual transforms, while multiplying a function by

a constant multiplies its Fourier transform by the same factor:

T[ f(x) + g(x)] = T[ f(x)] +T[ g(x)] =

´

f(k) +´g(k),

T[ cf(x)] = c T[ f(x)] = c

´

f(k).

(7.11)

A similar statement holds for the inverse Fourier transform T

−1

.

Recapitulating, by letting the length of the interval go to ∞, the discrete Fourier series

has become a continuous Fourier integral, while the Fourier coeﬃcients, which were deﬁned

only at a discrete collection of possible frequencies, have become a complete function

´

f(k)

deﬁned on all of frequency space k ∈ R. The reconstruction of f(x) from its Fourier

transform

´

f(k) via (7.9) can be rigorously justiﬁed under suitable hypotheses. For example,

if f(x) is piecewise C

1

on all of R and decays reasonably rapidly, f(x) → 0 as [ x [ → ∞,

in order that its Fourier integral (7.6) converges absolutely, then it can be proved, [44],

that the inverse Fourier integral (7.9) will converge to f(x) at all points of continuity, and

1/24/10 306 c (2010 Peter J. Olver

-2 -1 1 2

-0.2

0.2

0.4

0.6

0.8

1

1.2

-3 -2 -1 1 2 3

-0.5

0.5

1

1.5

2

-2 -1 1 2

-0.2

0.2

0.4

0.6

0.8

1

1.2

-2 -1 1 2

-0.2

0.2

0.4

0.6

0.8

1

1.2

Figure 7.1. Fourier Transform of a Rectangular Pulse.

to the midpoint

1

2

(f(x

−

) + f(x

+

)) at jump discontinuities — just like a Fourier series. In

particular, its Fourier transform

´

f(k) → 0 must also decay as [ k [ → 0, implying that (as

with Fourier series) the very high frequency modes make negligible contributions to the

reconstruction of the signal. A more precise result will be formulated in Theorem 7.13

below.

Example 7.1. The Fourier transform of the rectangular pulse (or box function

†

)

f(x) = σ(x + a) −σ(x −a) =

1, −a < x < a,

0, [ x [ > a.

(7.12)

of width 2a is easily computed:

´

f(k) =

1

√

2π

a

−a

e

−i kx

dx =

e

i ka

−e

−i ka

√

2π i k

=

2

π

sin ak

k

. (7.13)

On the other hand, the reconstruction of the pulse via the inverse transform (7.9) tells us

that

1

π

∞

−∞

e

i kx

sin ak

k

dk = f(x) =

1, −a < x < a,

1

2

, x = ±a,

0, [ x [ > a.

(7.14)

Note the convergence to the middle of the jump discontinuities at x = ±a. Splitting this

complex integral into its real and imaginary parts, we deduce a pair of striking trigono-

†

σ(x) is the step function (3.46).

1/24/10 307 c (2010 Peter J. Olver

metric integral identities

1

π

∞

−∞

cos kx sin ak

k

dk =

1, −a < x < a,

1

2

, x = ±a,

0, [ x [ > a,

1

π

∞

−∞

sin kx sinak

k

dk = 0.

(7.15)

Just as many Fourier series yield nontrivial summation formulae, the reconstruction

of a function from its Fourier transform often leads to nontrivial integration formulas.

One cannot compute the integral (7.14) by the Fundamental Theorem of Calculus, since

there is no elementary function whose derivative equals the integrand

†

. Moreover, it is not

even clear that the integral converges; indeed, the amplitude of the oscillatory integrand

decays like 1/[ k [, but the latter function does not have a convergent integral, and so the

usual comparison test for inﬁnite integrals, [8], fails to apply. The convergence of the

integral is marginal at best: the trigonometric oscillations somehow overcome the slow

rate of decay of 1/k and thereby induce the (conditional) convergence of the integral. In

Figure 7.1 we display the box function with a = 1, its Fourier transform, along with a

reconstruction obtained by numerically integrating (7.15). Since we are dealing with an

inﬁnite integral, we must break oﬀ the numerical integrator by restricting it to a ﬁnite

interval. The ﬁrst graph is obtained by integrating from −5 ≤ k ≤ 5 while the second is

from −10 ≤ k ≤ 10. The non-uniform convergence of the integral leads to the appearance

of a Gibbs phenomenon at the two discontinuities, similar to what we observed in the

non-uniform convergence of a Fourier series.

Example 7.2. Consider an exponentially decaying right-handed pulse

‡

f

r

(x) =

e

−ax

, x > 0,

0, x < 0,

(7.16)

where a > 0. We compute its Fourier transform directly from the deﬁnition:

´

f(k) =

1

√

2π

∞

0

e

−ax

e

−i kx

dx = −

1

√

2π

e

−(a+i k)x

a + i k

∞

x=0

=

1

√

2π (a + i k)

.

As in the preceding example, the inverse Fourier transform produces a nontrivial complex

integral identity:

1

2π

∞

−∞

e

i kx

a + i k

dk =

e

−ax

, x > 0,

1

2

, x = 0,

0, x < 0.

(7.17)

†

One can use Euler’s formula (3.59) to reduce (7.14) to a version of the exponential inte-

gral

Z

(e

αk

/k) dk, but it can be proved, [25], that neither integral can be written in terms of

elementary functions.

‡

Note that we can’t Fourier transform the entire exponential function e

−ax

because it does

not go to zero at both ±∞, which is required for the integral (7.6) to converge.

1/24/10 308 c (2010 Peter J. Olver

-3 -2 -1 1 2 3

-0.2

0.2

0.4

0.6

0.8

1

Right Pulse f

r

(x)

-3 -2 -1 1 2 3

-0.2

0.2

0.4

0.6

0.8

1

Left Pulse f

l

(x)

-3 -2 -1 1 2 3

-0.2

0.2

0.4

0.6

0.8

1

Even Pulse f

e

(x)

-3 -2 -1 1 2 3

-1

-0.5

0.5

1

Odd Pulse f

o

(x)

Figure 7.2. Exponential Pulses.

Similarly, a pulse that decays to the left,

f

l

(x) =

e

ax

, x < 0,

0, x > 0,

(7.18)

where a > 0 is still positive, has Fourier transform

´

f(k) =

1

√

2π (a − i k)

. (7.19)

This also follows from the general fact that the Fourier transform of f(−x) is

´

f(−k); see

Exercise . The even exponentially decaying pulse

f

e

(x) = e

−a| x |

(7.20)

is merely the sum of left and right pulses: f

e

= f

r

+ f

l

. Thus, by linearity,

´

f

e

(k) =

1

√

2π (a + i k)

+

1

√

2π (a − i k)

=

2

π

a

k

2

+ a

2

, (7.21)

The resulting Fourier transform is real and even because f

e

(x) is a real even function;

see Exercise . The inverse Fourier transform (7.9) produces another nontrivial integral

identity:

e

−a| x |

=

1

π

∞

−∞

a e

i kx

k

2

+ a

2

dk =

a

π

∞

−∞

cos kx

k

2

+ a

2

dk. (7.22)

(The imaginary part of the integral vanishes because the integrand is odd.) On the other

hand, the odd exponentially decaying pulse,

f

o

(x) = (sign x) e

−a| x |

=

e

−ax

, x > 0,

−e

ax

, x < 0,

(7.23)

1/24/10 309 c (2010 Peter J. Olver

is the diﬀerence of the right and left pulses, f

o

= f

r

− f

l

, and has purely imaginary and

odd Fourier transform

´

f

o

(k) =

1

√

2π (a + i k)

−

1

√

2π (a − i k)

= −i

2

π

k

k

2

+ a

2

. (7.24)

The inverse transform is

(signx) e

−a| x|

= −

i

π

∞

−∞

k e

i kx

k

2

+ a

2

dk =

1

π

∞

−∞

k sin kx

k

2

+ a

2

dk. (7.25)

As a ﬁnal example, consider the rational function

f(x) =

1

x

2

+ a

2

, where a > 0. (7.26)

Its Fourier transform requires integrating

´

f(k) =

1

√

2π

∞

−∞

e

−i kx

x

2

+ a

2

dx. (7.27)

The indeﬁnite integral (anti-derivative) does not appear in basic integration tables, and,

in fact, cannot be done in terms of elementary functions. However, we have just managed

to evaluate this particular integral! Look at (7.22). If we change x to k and k to −x,

then we exactly recover the integral (7.27) up to a factor of a

**2/π. We conclude that the
**

Fourier transform of (7.26) is

´

f(k) =

π

2

e

−a| k |

a

. (7.28)

This last example is indicative of an important general fact. The reader has no doubt

already noted the remarkable similarity between the Fourier transform (7.6) and its inverse

(7.9). Indeed, the only diﬀerence is that the former has a minus sign in the exponential.

This implies the following Symmetry Principle relating the direct and inverse Fourier

transforms.

Theorem 7.3. If the Fourier transform of the function f(x) is

´

f(k), then the Fourier

transform of

´

f(x) is f(−k).

The Symmetry Principle allows us to reduce the tabulation of Fourier transforms by

half. For instance, referring back to Example 7.1, we deduce that the Fourier transform of

the function

f(x) =

2

π

sin ax

x

is

´

f(k) = σ(−k + a) −σ(−k −a) = σ(k + a) −σ(k −a) =

1, −a < k < a,

1

2

, k = ±a,

0, [ k [ > a.

(7.29)

Note that, by linearity, we can divide both f(x) and

´

f(k) by

**2/π to deduce the Fourier
**

transform of

sin ax

x

.

1/24/10 310 c (2010 Peter J. Olver

Warning: Some authors omit the

√

2π factor in the deﬁnition (7.6) of the Fourier

transform

´

f(k). This alternative convention does have a slight advantage of eliminating

many

√

2π factors in the Fourier transforms of elementary functions. However, this ne-

cessitates an extra such factor in the reconstruction formula (7.9), which is achieved by

replacing

√

2π by 2π. A signiﬁcant disadvantage is that the resulting formulae for the

Fourier transform and its inverse are not as close, and so the Symmetry Principle of The-

orem 7.3 requires some modiﬁcation. (On the other hand, convolution — to be discussed

below — is a little easier without the extra factor.) Yet another, more recent convention

can be found in Exercise . When consulting any particular reference, the reader always

needs to check which Fourier transform convention is being used.

All of the functions in Example 7.2 required a > 0 for the Fourier integrals to converge.

The functions that emerge in the limit as a goes to 0 are of fundamental importance. Let

us start with the odd exponential pulse (7.23). When a →0, the function f

o

(x) converges

to the sign function

f(x) = sign x = σ(x) −σ(−x) =

+1, x > 0,

−1, x < 0.

(7.30)

Taking the limit of the Fourier transform (7.24) leads to

´

f(k) = −i

2

π

1

k

. (7.31)

The nonintegrable singularity of

´

f(k) at k = 0 is indicative of the fact that the sign func-

tion does not decay as [ x [ →∞. In this case, neither the Fourier transform integral nor its

inverse are well-deﬁned as standard (Riemann, or even Lebesgue, [112]) integrals. Never-

theless, it is possible to rigorously justify these results within the framework of generalized

functions.

More interesting are the even pulse functions f

e

(x), which, in the limit a →0, become

the constant function

f(x) ≡ 1. (7.32)

The limit of the Fourier transform (7.21) is

lim

a→0

2

π

2a

k

2

+ a

2

=

0, k = 0,

∞, k = 0.

(7.33)

This limiting behavior should remind the reader of our construction (5.10) of the delta

function as the limit of the functions

δ(x) = lim

n→∞

n

π (1 + n

2

x

2

)

= lim

a →0

a

π (a

2

+ x

2

)

.

Comparing with (7.33), we conclude that the Fourier transform of the constant function

(7.32) is a multiple of the delta function in the frequency variable:

´

f(k) =

√

2π δ(k). (7.34)

1/24/10 311 c (2010 Peter J. Olver

The direct transform integral

δ(k) =

1

2π

∞

−∞

e

−i kx

dx

is, strictly speaking, not deﬁned because the inﬁnite integrals of the oscillatory sine and

cosine functions don’t converge! However, this identity can be validly interpreted within

the framework of weak convergence and generalized functions. On the other hand, the

inverse transform formula (7.9) yields

∞

−∞

δ(k) e

i kx

dk = e

i k0

= 1,

which is in accord with the basic deﬁnition (5.16) of the delta function. As in the preceding

case, the delta function singularity at k = 0 manifests the lack of decay of the constant

function.

Conversely, the delta function δ(x) has constant Fourier transform

´

δ(k) =

1

√

2π

∞

−∞

δ(x) e

−i kx

dx =

e

−i k0

√

2π

≡

1

√

2π

, (7.35)

a result that also follows from the Symmerty Principle of Theorem 7.3. To determine the

Fourier transform of a delta spike δ

ξ

(x) = δ(x − ξ) concentrated at position x = ξ, we

compute

´

δ

ξ

(k) =

1

√

2π

∞

−∞

δ(x −ξ) e

−i kx

dx =

e

−i kξ

√

2π

. (7.36)

The result is a pure exponential in frequency space. Applying the inverse Fourier transform

(7.9) leads, formally, to the remarkable identity

δ

ξ

(x) = δ(x −ξ) =

1

2π

∞

−∞

e

−i k(x−ξ)

dk =

1

2π

' e

i kξ

; e

i kx

` , (7.37)

where ' ; ` denotes the L

2

Hermitian inner product on R. Since the delta function vanishes

for x = ξ, this identity is telling us that complex exponentials of diﬀering frequencies are

mutually orthogonal. However, this statement must be taken with a grain of salt, since

the integral does not converge in any standard sense (either Riemann or Lebesgue). But

it is possible to make sense of this identity within the language of generalized functions.

Indeed, multiplying both sides by f(x), and then integrating with respect to x, we ﬁnd

f(ξ) =

1

2π

∞

−∞

∞

−∞

f(x) e

−i k(x−ξ)

dxdk. (7.38)

This is a perfectly valid formula, being a restatement (or, rather, combination) of the

basic formulae (7.6) and (7.9) connecting the direct and inverse Fourier transforms of the

function f(x).

Vice versa, the Symmetry Principle tells us that the Fourier transform of a pure

exponential e

i κx

will be a shifted delta spike

√

2π δ(k − κ), concentrated at frequency

k = κ. Both results are particular cases of the general Shift Theorem, whose proof is left

as an exercise for the reader.

1/24/10 312 c (2010 Peter J. Olver

Short Table of Fourier Transforms

f(x)

´

f(k)

1

√

2π δ(k)

δ(x)

1

√

2π

σ(x)

π

2

δ(k) −

i

√

2π k

sign x −i

2

π

1

k

σ(x + a) −σ(x −a)

2

π

sin ak

k

e

−ax

σ(x)

1

√

2π (a + i k)

e

ax

(1 −σ(x))

1

√

2π (a − i k)

e

−a| x|

2

π

a

k

2

+ a

2

e

−ax

2 e

−k

2

/4a

√

2a

tan

−1

x

π

3/2

√

2

δ(k) − i

π

2

e

−| k |

k

f(cx + d)

e

i k d/c

[ c [

´

f

k

c

f(x)

´

f(−k)

´

f(x) f(−k)

f

′

(x) i k

´

f(k)

xf(x) i

´

f

′

(k)

f ∗ g(x)

√

2π

´

f(k) ´g(k)

Note: The parameters a, c, d are real, with a > 0 and c = 0.

1/24/10 313 c (2010 Peter J. Olver

Theorem 7.4. If f(x) has Fourier transform

´

f(k), then the Fourier transform of the

shifted function f(x − ξ) is e

−i kξ

´

f(k). Similarly, the transform of the product function

e

i κx

f(x), for real κ, is the shifted transform

´

f(k −κ).

Since the Fourier transform uniquely associates a function

´

f(k) on frequency space

with each (reasonable) function f(x) on physical space, one can characterize functions by

their transforms. Practical applications rely on tables (or, even better, computer algebra

systems such as Mathematica or Maple) that recognize a wide variety of transforms of

basic functions of importance in applications. The accompanying table lists some of the

most important examples of functions and their Fourier transforms. Note that, by applying

the Symmetry Principle of Theorem 7.3, each tabular entry can be used to deduce two

diﬀerent Fourier transforms. A more extensive collection of Fourier transforms can be

found in [100].

7.2. Derivatives and Integrals.

One of the most remarkable and important properties of the Fourier transform is that

it converts calculus into algebra! More speciﬁcally, the two basic operations in calculus —

diﬀerentiation and integration of functions — are realized as algebraic operations on their

Fourier transforms. (The downside is that algebraic operations on functions become more

complicated in the transform domain.)

Diﬀerentiation

Let us begin with derivatives. If we diﬀerentiate

†

the basic inverse Fourier transform

formula

f(x) ∼

1

√

2π

∞

−∞

´

f(k) e

i kx

dk.

with respect to x, we obtain

f

′

(x) ∼

1

√

2π

∞

−∞

i k

´

f(k) e

i kx

dk. (7.39)

The resulting integral is itself in the form of an inverse Fourier transform, namely of i k

´

f(k)

which immediately implies the following key result.

Proposition 7.5. The Fourier transform of the derivative f

′

(x) of a function is

obtained by multiplication of its Fourier transform by i k:

T[ f

′

(x)] = i k

´

f(k). (7.40)

Similarly, the Fourier transform of x f(x) is obtained by diﬀerentiating the Fourier trans-

form of f(x):

T[ x f(x)] = i

d

´

f

dk

. (7.41)

†

We are assuming the integrand is suﬃciently nice so that we can bring the derivative under

the integral sign; see [44, 143] for a fully rigorous justiﬁcation.

1/24/10 314 c (2010 Peter J. Olver

The second statement follows from the ﬁrst by use of the Symmetry Principle of

Theorem 7.3.

Example 7.6. The derivative of the even exponential pulse f

e

(x) = e

−a| x|

is a

multiple of the odd exponential pulse f

o

(x) = (sign x) e

−a| x |

:

f

′

e

(x) = −a (signx) e

−a| x |

= −af

o

(x).

Proposition 7.5 says that their Fourier transforms are related by

i k

´

f

e

(k) = i

2

π

ka

k

2

+ a

2

= −a

´

f

o

(k),

as previously noted in (7.21, 24). On the other hand, the odd exponential pulse has a jump

discontinuity of magnitude 2 at x = 0, and so its derivative contains a delta function:

f

′

o

(x) = −a e

−a| x|

+ 2 δ(x) = −af

e

(x) + 2 δ(x).

This is reﬂected in the relation between their Fourier transforms. If we multiply (7.24) by

i k we obtain

i k

´

f

o

(k) =

2

π

k

2

k

2

+ a

2

=

2

π

−

2

π

a

2

k

2

+ a

2

= 2

´

δ(k) −a

´

f

e

(k).

The Fourier transform, just like Fourier series, is completely compatible with the calculus

of generalized functions.

Higher order derivatives are handled by iterating the ﬁrst order formula (7.40).

Corollary 7.7. The Fourier transform of f

(n)

(x) is ( i k)

n

´

f(k).

This result has an important consequence: the smoothness of f(x) is manifested in the

rate of decay of its Fourier transform

´

f(k). We already noted that the Fourier transform

of a (nice) function must decay to zero at large frequencies:

´

f(k) → 0 as [ k [ → ∞.

(This result can be viewed as the Fourier transform version of the Riemann–Lebesgue

Lemma 3.41.) If the n

th

derivative f

(n)

(x) is also a reasonable function, then its Fourier

transform

¯

f

(n)

(k) = ( i k)

n

´

f(k) must go to zero as [ k [ →∞. This requires that

´

f(k) go to

zero more rapidly than [ k [

−n

. Thus, the smoother f(x), the more rapid the decay of its

Fourier transform. As a general rule of thumb, local features of f(x), such as smoothness,

are manifested by global features of

´

f(k), such as decay for large [ k [. The Symmetry

Principle implies that reverse is also true: global features of f(x) correspond to local

features of

´

f(k). For instance, smoothness of

´

f(k) implies decay of f(x) as x →±∞. This

local-global duality is one of the major themes of Fourier theory.

Integration

Integration is the inverse operation to diﬀerentiation, and so should correspond to

division by i k in frequency space. As with Fourier series, this is not completely correct;

there is an extra constant involved, which contributes an additional delta function in

frequency space.

1/24/10 315 c (2010 Peter J. Olver

Proposition 7.8. If f(x) has Fourier transform

´

f(k), then the Fourier transform of

its integral g(x) =

x

−∞

f(y) dy is

´ g(k) = −

i

k

´

f(k) + π

´

f(0) δ(k). (7.42)

Proof : First notice that

lim

x →−∞

g(x) = 0, lim

x→+∞

g(x) =

∞

−∞

f(x) dx =

√

2π

´

f(0).

Therefore, by subtracting a suitable multiple of the step function from the integral, the

resulting function

h(x) = g(x) −

√

2π

´

f(0) σ(x)

decays to 0 at both ±∞. Consulting our table of Fourier transforms, we ﬁnd

´

h(k) = ´ g(k) −π

´

f(0) δ(k) +

i

k

´

f(0) . (7.43)

On the other hand,

h

′

(x) = f(x) −

√

2π

´

f(0) δ(x).

Since h(x) → 0 as [ x [ → ∞, we can apply our diﬀerentiation rule (7.40), and conclude

that

i k

´

h(k) =

´

f(k) −

´

f(0). (7.44)

Combining (7.43) and (7.44) establishes the desired formula (7.42). Q.E.D.

Example 7.9. The Fourier transform of the inverse tangent function

f(x) = tan

−1

x =

x

0

dy

1 + y

2

=

x

−∞

dy

1 + y

2

−

π

2

can be computed by combining Proposition 7.8 with (7.28):

´

f(k) = −i

π

2

e

−| k |

k

+

π

3/2

√

2

δ(k).

7.3. Green’s Functions and Convolution.

The fact that the Fourier transform converts diﬀerentiation in the physical domain

into multiplication in the frequency domain is one of its most compelling features. A par-

ticularly important consequence is that it eﬀectively transforms diﬀerential equations into

algebraic equations, and thereby opens the door to their solution by elementary algebra!

One begins by applying the Fourier transform to both sides of the diﬀerential equation

under consideration. Solving the resulting algebraic equation will produce a formula for

the Fourier transform of the desired solution, which can then be immediately reconstructed

via the inverse Fourier transform. In the following chapter, we will use these techniques to

solve partial diﬀerential equations.

1/24/10 316 c (2010 Peter J. Olver

Solution of Boundary Value Problems

The Fourier transform is particularly well adapted to boundary value problems on the

entire real line. In place of the boundary conditions used on ﬁnite intervals, we look for

solutions that decay to zero suﬃciently rapidly as [ x [ →∞ — in order that their Fourier

transform be well-deﬁned (in the context of ordinary functions). In quantum mechanics,

[83, 89], these solutions are known as the bound states of the system, and correspond to

subatomic particles that are trapped or localized in a region of space by some sort of force

ﬁeld. For example, the electrons in an atom are bound states localized by the electrostatic

attraction of the nucleus.

As a speciﬁc example, consider the boundary value problem

−

d

2

u

dx

2

+ ω

2

u = h(x), −∞< x < ∞,

(7.45)

where ω > 0 is a positive constant. In lieu of boundary conditions, we require that the

solution u(x) →0 as [ x [ →∞.

We will solve this problem by applying the Fourier transform to both sides of the

diﬀerential equation. Taking Corollary 7.7 into account, the result is a linear algebraic

equation

k

2

´ u(k) + ω

2

´ u(k) =

´

h(k)

relating the Fourier transforms of u and h. Unlike the diﬀerential equation, the transformed

equation can be immediately solved for

´ u(k) =

´

h(k)

k

2

+ ω

2

. (7.46)

Therefore, we can reconstruct the solution by applying the inverse Fourier transform for-

mula (7.9):

u(x) =

1

√

2π

∞

−∞

´

h(k) e

i kx

k

2

+ ω

2

dk. (7.47)

For example, if the forcing function is an even exponential pulse,

h(x) = e

−| x |

with

´

h(k) =

2

π

1

k

2

+ 1

,

then (7.47) writes the solution as a Fourier integral:

u(x) =

1

π

∞

−∞

e

i kx

(k

2

+ ω

2

)(k

2

+ 1)

dk =

1

π

∞

−∞

cos kx

(k

2

+ ω

2

)(k

2

+ 1)

dk ,

noting that the imaginary part of the complex integral vanishes because the integrand is

an odd function. (Indeed, if the forcing function is real, the solution must also be real.)

The Fourier integral can be explicitly evaluated by using partial fractions to rewrite

´ u(k) =

2

π

1

(k

2

+ ω

2

)(k

2

+ 1)

=

2

π

1

ω

2

−1

1

k

2

+ 1

−

1

k

2

+ ω

2

, ω

2

= 1.

1/24/10 317 c (2010 Peter J. Olver

Thus, according to our Fourier Transform Table, the solution to this boundary value

problem is

u(x) =

e

−| x|

−

1

ω

e

−ω | x |

ω

2

−1

when ω

2

= 1. (7.48)

The reader may wish to verify that this function is indeed a solution, meaning that it is

twice continuously diﬀerentiable (which is not so immediately apparent from the formula),

decays to 0 as [ x [ →∞, and satisﬁes the diﬀerential equation everywhere. The “resonant”

case ω

2

= 1 is left as an exercise.

Remark: The method of partial fractions that you learned in ﬁrst year calculus is

often an eﬀective tool for evaluating (inverse) Fourier transforms of rational functions.

A particularly important case is when the forcing function

h(x) = δ

ξ

(x) = δ(x −ξ)

represents a unit impulse concentrated at x = ξ. The resulting square-integrable solution

is the Green’s function G(x; ξ) for the boundary value problem. According to (7.46), its

Fourier transform with respect to x is

´

G(k; ξ) =

1

√

2π

e

−i kξ

k

2

+ ω

2

,

which is the product of an exponential factor e

−i kξ

, representing the Fourier transform

of δ

ξ

(x), times a multiple of the Fourier transform of the even exponential pulse e

−ω | x |

.

We apply Theorem 7.4, and conclude that the Green’s function for this boundary value

problem is an exponential pulse centered at ξ, namely

G(x; ξ) =

1

2 ω

e

−ω | x−ξ |

. (7.49)

Observe that, as with other self-adjoint boundary value problems, the Green’s function

is symmetric under interchange of x and ξ. As a function of x, it satisﬁes the homogeneous

diﬀerential equation −u

′′

+ ω

2

u = 0, except at the point x = ξ when its derivative has

a jump discontinuity of unit magnitude. It also decays as [ x [ → ∞, as required by

the boundary conditions. The Green’s function superposition principle tells us that the

solution to the inhomogeneous boundary value problem (7.45) under a general forcing can

be represented in the integral form

u(x) =

∞

−∞

G(x; ξ) h(ξ) dξ =

1

2ω

∞

−∞

e

−ω| x−ξ |

h(ξ) dξ. (7.50)

The reader may enjoy recovering the particular exponential solution (7.48) from this inte-

gral formula.

Convolution

In our solution to the boundary value problem (7.45), we ended up deriving a formula

for its Fourier transform (7.46) as the product of two known Fourier transforms. The ﬁnal

1/24/10 318 c (2010 Peter J. Olver

Green’s function formula (7.50), obtained by applying the inverse Fourier transform, is

indicative of a general property. Its right hand side has the form of a convolution product

between functions.

Deﬁnition 7.10. The convolution of scalar functions f(x) and g(x) is the scalar

function h = f ∗ g deﬁned by the formula

h(x) = f(x) ∗ g(x) =

∞

−∞

f(x −ξ) g(ξ) dξ. (7.51)

We record the basic properties of the convolution product, leaving their veriﬁcation

as exercises for the reader. All of these assume that the implied convolution integrals

converge.

(a) Symmetry: f ∗ g = g ∗ f,

(b) Bilinearity:

f ∗ (ag + bh) = a(f ∗ g) + b(f ∗ h),

(af + bg) ∗ h = a(f ∗ h) + b(g ∗ h),

a, b ∈ C,

(c) Associativity: f ∗ (g ∗ h) = (f ∗ g) ∗ h,

(d) Zero function: f ∗ 0 = 0,

(e) Delta function: f ∗ δ = f.

One tricky feature is that the constant function 1 is not a unit for the convolution

product; indeed,

f ∗ 1 =

∞

−∞

f(ξ) dξ

is a constant function — the total integral of f — not the original function f(x). In fact,

according to the ﬁnal property, the delta function plays the role of the “convolution unit”:

f(x) ∗ δ(x) =

∞

−∞

f(x −ξ) δ(ξ) dξ = f(x).

In particular, our solution (7.49) has the form of a convolution product between an

even exponential pulse g(x) = (2ω)

−1

e

−ω| x |

and the forcing function:

u(x) = g(x) ∗ h(x).

On the other hand, its Fourier transform (7.46) is, up to a factor, the ordinary multiplica-

tive product

´ u(k) =

√

2π ´ g(k)

´

h(k)

of the Fourier transforms of g and h. In fact, this is a general property of the Fourier trans-

form: convolution in the physical domain corresponds to multiplication in the frequency

domain, and conversely.

1/24/10 319 c (2010 Peter J. Olver

Theorem 7.11. The Fourier transform of the convolution h(x) = f(x) ∗ g(x) of two

functions is a multiple of the product of their Fourier transforms:

´

h(k) =

√

2π

´

f(k) ´g(k). (7.52)

Vice versa, the Fourier transform of their product h(x) = f(x) g(x) is, up to multiple, the

convolution of their Fourier transforms:

´

h(k) =

1

√

2π

´

f(k) ∗ ´ g(k). (7.53)

Proof : Combining the deﬁnition of the Fourier transform with the convolution for-

mula (7.51), we ﬁnd

´

h(k) =

1

√

2π

∞

−∞

h(x) e

−i kx

dx =

1

√

2π

∞

−∞

∞

−∞

f(x −ξ) g(ξ) e

−i kx

dxdξ,

where we are assuming that the integrands are suﬃciently nice to allow us to interchange

the order of integration, [8]. Applying the change of variables η = x − ξ in the inner

integral produces

´

h(k) =

1

√

2π

∞

−∞

∞

−∞

f(η) g(ξ) e

−i k(ξ+η)

dξ dη

=

√

2π

1

√

2π

∞

−∞

f(η) e

−i kη

dη

1

√

2π

∞

−∞

g(ξ) e

−i kξ

dξ

=

√

2π

´

f(k) ´g(k),

proving (7.52). The second formula can be proved in a similar fashion, or by simply noting

that it follows directly from the Symmetry Principle of Theorem 7.3. Q.E.D.

Example 7.12. We already know, (7.29), that the Fourier transform of

f(x) =

sin x

x

is the box function

´

f(k) =

π

2

σ(k + 1) −σ(k −1)

=

π

2

, −1 < k < 1,

0, [ k [ > 1.

We also know that the Fourier transform of

g(x) =

1

x

is ´ g(k) = −i

π

2

signk.

Therefore, the Fourier transform of their product

h(x) = f(x) g(x) =

sin x

x

2

1/24/10 320 c (2010 Peter J. Olver

-3 -2 -1 1 2 3

-1

-0.5

0.5

1

Figure 7.3. The Fourier transform of

sin x

x

2

.

can be obtained by convolution:

´

h(k) =

1

√

2π

´

f(k) ∗ ´ g(k) =

1

√

2π

∞

−∞

´

f(κ) ´g(k −κ) dκ

= −i

π

8

1

−1

sign(k −κ) dκ =

i

π

2

k < −1,

−i

π

2

k, −1 < k < 1,

−i

π

2

k > 1.

A graph of

´

h(k) appears in Figure 7.3.

7.4. The Fourier Transform on Hilbert Space.

While we do not have the analytical tools to embark on a fully rigorous treatment

of the mathematical theory underlying the Fourier transform, it is worth outlining a few

of the more important features. We have already noted that the Fourier transform, when

deﬁned, is a linear map, taking functions f(x) on physical space to functions

´

f(k) on

frequency space. A critical question is precisely which function space should the theory be

applied to. Not every function admits a Fourier transform in the classical sense

†

— the

Fourier integral (7.6) is required to converge, and this places restrictions on the function

and its asymptotics at large distances.

It turns out the proper setting for the rigorous theory is the Hilbert space of complex-

valued square-integrable functions — the same inﬁnite-dimensional vector space that lies

at the heart of modern quantum mechanics. In Section 3.5, we already introduced the

Hilbert space L

2

[ a, b] on a ﬁnite interval; here we adapt Deﬁnition 3.32 to the entire real

line. Thus, the Hilbert space L

2

= L

2

(R) is the inﬁnite-dimensional vector space consisting

†

We leave aside the more advanced issues involving generalized functions in this subsection.

1/24/10 321 c (2010 Peter J. Olver

of all complex-valued functions f(x) which are deﬁned for all x ∈ R and have ﬁnite L

2

norm:

| f |

2

=

∞

−∞

[ f(x) [

2

dx < ∞. (7.54)

For example, any piecewise continuous function that satisﬁes the decay criterion

[ f(x) [ ≤

M

[ x [

1/2+δ

, for all suﬃciently large [ x [ ≫0, (7.55)

for some M > 0 and δ > 0, belongs to L

2

. However, as in Section 3.5, Hilbert space

contains many more functions, and the precise deﬁnitions and identiﬁcation of its elements

is quite subtle. The Hermitian inner product on the complex Hilbert space L

2

is prescribed

in the usual manner,

' f ; g ` =

∞

−∞

f(x) g(x) dx, (7.56)

so that | f |

2

= ' f ; f `. The Cauchy–Schwarz inequality

[ ' f ; g ` [ ≤ | f | | g | (7.57)

ensures that the inner product integral is ﬁnite whenever f, g ∈ L

2

.

Let us state the fundamental result governing the eﬀect of the Fourier transform

on functions in Hilbert space. It can be regarded as a direct analog of the Pointwise

Convergence Theorem 3.8 for Fourier series. A rigorous proof of this result can be found

in [44, 130].

Theorem 7.13. If f(x) ∈ L

2

is square-integrable, then its Fourier transform

´

f(k) ∈

L

2

is a well-deﬁned, square-integrable function of the frequency variable k. If f(x) is

continuously diﬀerentiable at a point x, then its inverse Fourier transform (7.9) equals its

value f(x). More generally, if the right and left hand limits f(x

−

), f(x

+

), and also f

′

(x

−

),

f

′

(x

+

) exist, then the inverse Fourier transform integral converges to the average value

1

2

f(x

−

) + f(x

+

)

.

Thus, the Fourier transform

´

f = T[ f ] deﬁnes a linear transformation from L

2

func-

tions of x to L

2

functions of k. In fact, the Fourier transform preserves inner products.

This important result is known as Parseval’s formula, whose Fourier series counterpart

appeared in (3.119).

Theorem 7.14. If

´

f(k) = T[ f(x)] and ´ g(k) = T[ g(x)], then ' f ; g ` = '

´

f ; ´g `, i.e.,

∞

−∞

f(x) g(x) dx =

∞

−∞

´

f(k) ´ g(k) dk. (7.58)

Proof : Let us sketch a formal proof that serves to motivate why this result is valid.

We use the deﬁnition (7.6) of the Fourier transform to evaluate

∞

−∞

´

f(k) ´ g(k) dk =

∞

−∞

1

√

2π

∞

−∞

f(x) e

−i kx

dx

1

√

2π

∞

−∞

g(y) e

+i ky

dy

dk

=

∞

−∞

∞

−∞

f(x) g(y)

1

2π

∞

−∞

e

−i k(x−y)

dk

dxdy.

1/24/10 322 c (2010 Peter J. Olver

Now according to (7.37), the inner k integral can be replaced by a delta function δ(x −y),

and hence

∞

−∞

´

f(k) ´ g(k) dk =

∞

−∞

∞

−∞

f(x) g(y) δ(x −y) dxdy =

∞

−∞

f(x) g(x) dx.

This completes our “proof”; see [44, 130] for a rigorous version. Q.E.D.

In particular, orthogonal functions, satisfying ' f ; g ` = 0, will have orthogonal Fourier

transforms, '

´

f ; ´g ` = 0. Choosing f = g in Parseval’s formula (7.58) results in the

Plancherel formula

| f |

2

= |

´

f |

2

, or, explicitly,

∞

−∞

[ f(x) [

2

dx =

∞

−∞

[

´

f(k) [

2

dk. (7.59)

Thus, the Fourier transform T: L

2

→L

2

deﬁnes a unitary or norm-preserving linear trans-

formation on Hilbert space, mapping L

2

functions of the physical variable x to L

2

functions

of the frequency variable k.

Quantum Mechanics and the Uncertainty Principle

In its popularized form, the Heisenberg Uncertainty Principle is a by now familiar

philosophical concept. It was ﬁrst formulated in the 1920’s by the German physicist

Werner Heisenberg, one of the founders of modern quantum mechanics, and states that,

in a physical system, certain quantities cannot be simultaneously measured with complete

accuracy. For instance, the more precisely one measures the position of a particle, the

less accuracy there will be in the measurement of its momentum; vice versa, the greater

the accuracy in the momentum, the less certainty in its position. A similar uncertainty

couples energy and time. Experimental veriﬁcation of the uncertainty principle can be

found even in fairly simple situations. Consider a light beam passing through a small hole.

The position of the photons is constrained by the hole; the eﬀect of their momenta is in

the pattern of light diﬀused on a screen placed beyond the hole. The smaller the hole, the

more constrained the position, and the wider the image on the screen, meaning the less

certainty there is in the observed momentum.

This is not the place to discuss the philosophical and experimental consequences of

Heisenberg’s principle. What we will show is that the Uncertainty Principle is, in fact, a

mathematical property of the Fourier transform! In quantum theory, each of the paired

quantities, e.g., position and momentum, are interrelated by the Fourier transform. Indeed,

Proposition 7.5 says that the Fourier transform of the diﬀerentiation operator represent-

ing momentum is a multiplication operator representing position and vice versa. This

Fourier transform-based duality between position and momentum, or multiplication and

diﬀerentiation, lies at the heart of the Uncertainty Principle.

In quantum mechanics, the wave functions of a quantum system are characterized as

the elements of unit norm, | ϕ| = 1, belonging to the underlying state space, which, in a

one-dimensional model of a single particle, is the Hilbert space L

2

= L

2

(R) consisting of

square integrable, complex valued functions of x. As we already noted in Section 3.5, the

squared modulus of the wave function, [ ϕ(x) [

2

, represents the probability density of the

1/24/10 323 c (2010 Peter J. Olver

particle being found at position x. Consequently, the mean value of any function f(x) of

the position variable is given by its integral against the system’s probability density, and

denoted by

' f(x) ` =

∞

−∞

f(x) [ ϕ(x) [

2

dx. (7.60)

In particular,

' x ` =

∞

−∞

x [ ϕ(x) [

2

dx (7.61)

is the mean or average measured position of the particle, while ∆x, deﬁned by

(∆x)

2

= '

x −' x `

2

` = ' x

2

` −' x `

2

(7.62)

is the variance, that is, the statistical deviation of the particle’s measured position from

the mean. We note that

' x

2

` =

∞

−∞

x

2

[ ϕ(x) [

2

dx = | x ϕ(x) |

2

. (7.63)

On the other hand, the momentum variable p is related to the Fourier transform

frequency via the de Broglie relation p = k, where

=

h

2π

≈ 1.055 10

−34

Joule seconds (7.64)

is Planck’s constant, whose value governs the quantization of physical quantities. There-

fore, the mean value of any function of momentum g(p) is given by its integral against the

squared modulus of the Fourier transformed wave function:

' g(p) ` =

∞

−∞

g( k) [ ´ ϕ(k) [

2

dk. (7.65)

In particular, the mean of the momentum measurements of the particle is given by

' p ` =

∞

−∞

k [ ´ ϕ(k) [

2

dk = −i

∞

−∞

ϕ

′

(x) ϕ(x) dx = −i ' ϕ

′

; ϕ`, (7.66)

where we used Parseval’s formula formula (7.58) to convert to an integral over position,

and (7.40) to infer that k ´ ϕ(k) is the Fourier transform of −i ϕ

′

(x). Similarly,

(∆p)

2

= '

p −' p `

2

` = ' p

2

` −' p `

2

(7.67)

is the squared variance in the momentum, where, by a similar computation,

' p

2

` =

2

∞

−∞

k

2

[ ´ ϕ(k) [

2

dk = −

2

∞

−∞

ϕ(x) ϕ

′′

(x) dx

=

2

∞

−∞

[ ϕ

′

(x) [

2

dx =

2

| ϕ

′

(x) |

2

.

(7.68)

With this interpretation, the Uncertainty Principle can be stated as follows.

1/24/10 324 c (2010 Peter J. Olver

Theorem 7.15. If ϕ(x) is a wave function, so | ϕ| = 1, then the variances in

position and momentum satisfy the inequality

∆x ∆p ≥

1

2

. (7.69)

The smaller the variance of a quantity such as position or momentum, the more

accurate will be its measurement. Thus, the Heisenberg inequality (7.69) quantiﬁes the

statement that the more accurately we are able to measure the momentum p, the less

accurate will be any measurement of its position x, and vice versa. For more details and

physical consequences, you should consult an introductory text on mathematical quantum

mechanics, e.g., [83, 89].

Proof : For any value of the real parameter t,

0 ≤ | t x ϕ(x) + ϕ

′

(x) |

2

= t

2

| x ϕ(x) |

2

+ t

' ϕ

′

(x) ; x ϕ(x) ` +' x ϕ(x) ; ϕ

′

(x) `

+| ϕ

′

(x) |

2

.

(7.70)

The middle term can be evaluated as follows:

' ϕ

′

(x) ; x ϕ(x) ` +' x ϕ(x) ; ϕ

′

(x) ` =

∞

−∞

x ϕ

′

(x) ϕ(x) + x ϕ(x) ϕ

′

(x)

dx

=

∞

−∞

x

d

dx

[ ϕ(x) [

2

dx = −

∞

−∞

[ ϕ(x) [

2

dx = −1,

where we employed integration by parts, noting that the boundary terms vanish since

ϕ(x) →0 as [ x [ →∞. Thus, in view of (7.63, 68), (7.70) reads

' x

2

` t

2

−t +

' p

2

`

2

≥ 0

for all real t. The minimum value of the left hand side of this inequality occurs at t

⋆

=

1/(2 ' x

2

`), where its value is

' p

2

`

2

−

1

4 ' x

2

`

≥ 0 and hence ' x

2

` ' p

2

` ≥

1

4

2

.

To obtain the Uncertainty Relation, one performs the selfsame calculation, but with x−' x `

replacing x and p −' p ` replacing p. The result is

'

x −' x `

2

` t

2

−t +

'

p −' p `

2

`

2

= (∆x)

2

t

2

−t +

(∆p)

2

2

≥ 0. (7.71)

Substituting t = 1/(2 (∆x)

2

) produces the Heisenberg inequality (7.69). Q.E.D.

1/24/10 325 c (2010 Peter J. Olver

sums into integrals, and the resulting representation of a function is renamed the Fourier transform. Since we are dealing with an inﬁnite interval, there are no longer any periodicity requirements on the function f (x). Moreover, the frequencies represented in the Fourier transform are no longer constrained by the length of the interval, and so we are eﬀectively decomposing a quite general, non-periodic function into a continuous superposition of trigonometric functions of all possible frequencies. Let us present the details of this construction in a more concrete form. The computations will be signiﬁcantly simpler if we work with the complex version of the Fourier series from the outset. Our starting point is the rescaled Fourier series (3.84) on a symmetric interval [ − ℓ , ℓ ] of length 2 ℓ, which we rewrite in the adapted form

∞

f (x) ∼

ν = −∞

π f ℓ (kν ) i kν x e . 2 ℓ

(7.1)

The sum is over the discrete collection of frequencies πν , ν = 0, ±1, ±2, . . . , kν = (7.2) ℓ corresponding to those trigonometric functions that have period 2 ℓ. For reasons that will soon become apparent, the Fourier coeﬃcients of f are now denoted as cν = f ; e i kν x = so that 1 2ℓ

ℓ

f (x) e− i kν x dx =

−ℓ

π f ℓ (kν ) , 2 ℓ

(7.3)

ℓ 1 f (x) e− i kν x dx. (7.4) f ℓ (kν ) = √ 2 π −ℓ This reformulation of the basic Fourier series formula allows us to smoothly pass to the limit when the interval’s length ℓ → ∞. On an interval of length 2 ℓ, the frequencies (7.2) required to represent a function in Fourier series form are equally distributed, with interfrequency spacing π ∆k = kν+1 − kν = . (7.5) ℓ As ℓ → ∞, the spacing ∆k → 0, and so the relevant frequencies become more and more densely packed in the space of all possible frequencies: −∞ < k < ∞. In the limit, we anticipate that all possible frequencies will be represented. Indeed, letting kν = k be arbitrary in (7.4), and sending ℓ → ∞, results in the inﬁnite integral

1 f (k) = √ 2π

∞

f (x) e− i k x dx

−∞

(7.6)

known as the Fourier transform of the function f (x). If f (x) is a reasonably nice function, e.g., piecewise continuous and decaying to 0 reasonably quickly as | x | → ∞, its Fourier transform f (k) is deﬁned for all possible frequencies −∞ < k < ∞. This formula will sometimes conveniently be abbreviated as f (k) = F [ f (x) ], 1/24/10 305

c 2010

(7.7)

Peter J. Olver

approximating the integral 1 √ 2π ∞ −∞ f ℓ (k) e i k x dk. the Fourier series has become a Fourier integral that reconstructs the function f (x) as a (continuous) superposi√ tion of complex exponentials e i k x of all possible frequencies.9) can be written f (x) = F −1 [ f (k) ]. moreover.11) A similar statement holds for the inverse Fourier transform F −1 . It is worth pointing out that both the Fourier transform (7. which we ﬁrst rewrite in a more suggestive form. Olver . the interfrequency spacing ∆k = π/ℓ → 0.8) using (7. f (x) → 0 as | x | → ∞. Recapitulating.10) deﬁne linear maps on function space.5). 112]. while the Fourier coeﬃcients.4) converge to the Fourier transform: f ℓ (k) → f (k). −∞ (7. the right hand side has the form of a Riemann sum.1). by letting the length of the interval go to ∞. F [ c f (x) ] = c F [ f (x) ] = c f (k). 1 f (x) ∼ √ 2π ∞ f ℓ (kν ) e i kν x ∆k. (7. the functions (7. In abbreviated form. ν = −∞ (7.9) can be rigorously justiﬁed under suitable hypotheses. and so one expects the Riemann sums to converge to the limiting integral 1 f (x) ∼ √ 2π ∞ f (k) e i k x dk. The reconstruction of f (x) from its Fourier transform f (k) via (7.9) will converge to f (x) at all points of continuity. while multiplying a function by a constant multiplies its Fourier transform by the same factor: F [ f (x) + g(x) ] = F [ f (x) ] + F [ g(x) ] = f (k) + g(k). To reconstruct the function from its Fourier transform. if f (x) is piecewise C1 on all of R and decays reasonably rapidly. in order that its Fourier integral (7. This means that the Fourier transform of the sum of two functions is the sum of their individual transforms. For example. As ℓ → ∞. In this manner. the discrete Fourier series has become a continuous Fourier integral. then it can be proved. that the inverse Fourier integral (7. (7. with f (k)/ 2 π prescribing the amount of contribution of the complex exponential of frequency k. formula (7. and 1/24/10 306 c 2010 Peter J.9) The resulting formula serves to deﬁne the inverse Fourier transform.where F is the Fourier transform operator . [44].7) and its inverse (7.7). which is used to recover the original signal from its Fourier transform. have become a complete function f (k) deﬁned on all of frequency space k ∈ R. [8.6) converges absolutely.10) thus deﬁning the inverse of the Fourier transform operator (7. which were deﬁned only at a discrete collection of possible frequencies. we employ a similar limiting procedure on the Fourier series (7. For each ﬁxed value of x.

x = ± a.5 1. Olver . its Fourier transform f (k) → 0 must also decay as | k | → 0.13 below. dk = f (x) = 2 π −∞ k 0.2 1 0.2 1.5 1 0.1. (7. π k (7. 0. | x | > a. the reconstruction of the pulse via the that 1.9) tells us −a < x < a. Example 7. A more precise result will be formulated in Theorem 7.2 -2 -1 -0. we deduce a pair of striking trigono- On the other hand. implying that (as with Fourier series) the very high frequency modes make negligible contributions to the reconstruction of the signal.4 0.5 1 2 -3 -2 -1 -0. ∞ ikx e sin a k 1 1 .12) e− i k x dx = −a e i ka − e− i ka √ = 2π i k 2 sin a k .6 0. (7.2 1.14) † σ(x) is the step function (3. The Fourier transform of the rectangular pulse (or box function† ) f (x) = σ(x + a) − σ(x − a) = of width 2 a is easily computed: 1 f (k) = √ 2π a 1. In particular.8 0.1. 1 to the midpoint 2 (f (x− ) + f (x+ )) at jump discontinuities — just like a Fourier series.2 1 2 -2 -1 -0.2 1 2 0.8 0.2 1 2 3 1 2 Figure 7. inverse transform (7.46).2 1 0. | x | > a. 1/24/10 307 c 2010 Peter J.2 -2 -1 -0.4 0.6 0.6 0. Fourier Transform of a Rectangular Pulse. − a < x < a.8 0.4 0.13) Note the convergence to the middle of the jump discontinuities at x = ± a.1. Splitting this complex integral into its real and imaginary parts.

Consider an exponentially decaying right-handed pulse‡ fr (x) = e− a x .14) to a version of the exponential integral (e /k) dk. | x | > a. x < 0. 0. dk = 2 k 0. can use Euler’s formula (3. fails to apply. Example 7.2. the inverse Fourier transform produces a nontrivial complex integral identity: e− a x . but it can be proved. Since we are dealing with an inﬁnite integral. along with a reconstruction obtained by numerically integrating (7. similar to what we observed in the non-uniform convergence of a Fourier series. the reconstruction of a function from its Fourier transform often leads to nontrivial integration formulas. αk † One Z Note that we can’t Fourier transform the entire exponential function e− a x because it does not go to zero at both ± ∞. the amplitude of the oscillatory integrand decays like 1/| k |. indeed. that neither integral can be written in terms of elementary functions. cos k x sin a k 1 .6) to converge. x < 0. Moreover.1 we display the box function with a = 1.16) where a > 0. Olver . The convergence of the integral is marginal at best: the trigonometric oscillations somehow overcome the slow rate of decay of 1/k and thereby induce the (conditional) convergence of the integral. 2 2 π −∞ a + i k 0. The ﬁrst graph is obtained by integrating from −5 ≤ k ≤ 5 while the second is from −10 ≤ k ≤ 10. it is not even clear that the integral converges. We compute its Fourier transform directly from the deﬁnition: 1 f (k) = √ 2π ∞ e 0 −ax − i kx e 1 e−(a+ i k)x dx = − √ 2π a + i k =√ 1 . we must break oﬀ the numerical integrator by restricting it to a ﬁnite interval. x = ± a. and so the usual comparison test for inﬁnite integrals. ∞ eikx 1 1 (7. x > 0. −a < x < a. x = 0. 1 π ∞ −∞ sin k x sin a k dk = 0. ∞ x=0 1. x > 0. its Fourier transform.metric integral identities 1 π ∞ −∞ (7. 2 π (a + i k) As in the preceding example. In Figure 7. but the latter function does not have a convergent integral. The non-uniform convergence of the integral leads to the appearance of a Gibbs phenomenon at the two discontinuities. k (7.59) to reduce (7. [8].15) Just as many Fourier series yield nontrivial summation formulae.15). One cannot compute the integral (7. since there is no elementary function whose derivative equals the integrand† .17) dk = .14) by the Fundamental Theorem of Calculus. which is required for the integral (7. [ 25 ]. ‡ 1/24/10 308 c 2010 Peter J.

6 0. x > 0.9) produces another nontrivial integral identity: ∞ ∞ 1 a eikx cos k x a e− a | x | = dk = dk. The inverse Fourier transform (7. has Fourier transform f (k) = √ 1 . The even exponentially decaying pulse fe (x) = e− a | x | is merely the sum of left and right pulses: fe = fr + fl .4 0. x > 0.2.2 1 -0.22) π −∞ k 2 + a2 π −∞ k 2 + a2 (The imaginary part of the integral vanishes because the integrand is odd.2 1 2 3 -3 -2 -1 1 0.8 0. a pulse that decays to the left.6 0. Olver . the odd exponentially decaying pulse.4 0. x < 0. 0.2 -3 -2 Left Pulse fl (x) 1 0. where a > 0 is still positive.6 0. (7.5 1 2 3 -3 -2 -1 -0. 2 π (a − i k) fl (x) = ea x .23) Peter J. − ea x .) On the other hand.8 0. x < 0. c 2010 (7. see Exercise . (7.2 1 2 3 -1 Even Pulse fe (x) Figure 7.1 0. by linearity. Similarly. fo (x) = (sign x) e− a | x | = 1/24/10 309 e− a x . 2 + a2 π k (7.19) This also follows from the general fact that the Fourier transform of f (− x) is f (− k).4 0. Thus.21) (7. Odd Pulse fo (x) Exponential Pulses.18) (7.8 0.2 2 3 Right Pulse fr (x) 1 0. see Exercise . f e (k) = √ 1 1 + √ = 2 π (a + i k) 2 π (a − i k) 2 a .5 -1 -0.20) The resulting Fourier transform is real and even because fe (x) is a real even function.2 -3 -2 -1 -0.

Theorem 7. 1.22). | k | > a.27) up to a factor of a 2/π.is the diﬀerence of the right and left pulses. 2 + a2 π k (7.28) f (k) = 2 a This last example is indicative of an important general fact. Indeed. by linearity.27) The indeﬁnite integral (anti-derivative) does not appear in basic integration tables. and has purely imaginary and odd Fourier transform f o (k) = √ The inverse transform is 1 1 − √ = −i 2 π (a + i k) 2 π (a − i k) i π ∞ −∞ k 2 . x2 + a 2 (7. f (k) = σ(− k + a) − σ(− k − a) = σ(k + a) − σ(k − a) = (7. we have just managed to evaluate this particular integral! Look at (7.3.9). in fact. However. then we exactly recover the integral (7.6) and its inverse (7. and. Note that. k 2 + a2 (7. The reader has no doubt already noted the remarkable similarity between the Fourier transform (7. + a2 ∞ −∞ where a > 0.25) As a ﬁnal example. The Symmetry Principle allows us to reduce the tabulation of Fourier transforms by half. we deduce that the Fourier transform of the function 2 sin a x f (x) = π x is − a < k < a. (7. This implies the following Symmetry Principle relating the direct and inverse Fourier transforms. If we change x to k and k to − x.26) is π e− a | k | . k = ± a. the only diﬀerence is that the former has a minus sign in the exponential.1. 1 . referring back to Example 7. We conclude that the Fourier transform of (7.26) Its Fourier transform requires integrating 1 f (k) = √ 2π e− i k x dx. Olver . If the Fourier transform of the function f (x) is f (k). consider the rational function f (x) = x2 1 . we can divide both f (x) and f (k) by sin a x .29) 2 0. For instance. fo = fr − fl . cannot be done in terms of elementary functions. transform of x 1/24/10 310 2/π to deduce the Fourier c 2010 Peter J. (7. then the Fourier transform of f (x) is f (− k).24) (sign x) e− a | x | = − k eikx 1 dk = k 2 + a2 π ∞ −∞ k sin k x dk.

−1. which. this necessitates √ extra such factor in the reconstruction formula (7.33). in the limit a → 0. More interesting are the even pulse functions fe (x). When a → 0.23). A signiﬁcant disadvantage is that the resulting formulae for the Fourier transform and its inverse are not as close. (On the other hand. π k (7. (7. k = 0. convolution — to be discussed below — is a little easier without the extra factor.21) is lim 2 2a = 2 + a2 π k 0.9). (7. or even Lebesgue.34) 1/24/10 311 c 2010 Peter J.24) leads to f (k) = − i 2 1 . (7.3 requires some modiﬁcation. more recent convention can be found in Exercise .) Yet another. δ(x) = lim 2 x2 ) 2 + x2 ) n → ∞ π (1 + n a → 0 π (a Comparing with (7. it is possible to rigorously justify these results within the framework of generalized functions.10) of the delta function as the limit of the functions a n = lim .32) is a multiple of the delta function in the frequency variable: √ f (k) = 2 π δ(k). (7. the function fo (x) converges to the sign function f (x) = sign x = σ(x) − σ(− x) = +1. we conclude that the Fourier transform of the constant function (7. become the constant function f (x) ≡ 1.32) The limit of the Fourier transform (7. the reader always needs to check which Fourier transform convention is being used.30) Taking the limit of the Fourier transform (7. Let us start with the odd exponential pulse (7. [112]) integrals. This alternative convention does have a slight advantage of eliminating √ many 2 π factors in the Fourier transforms of elementary functions.31) The nonintegrable singularity of f (k) at k = 0 is indicative of the fact that the sign function does not decay as | x | → ∞. The functions that emerge in the limit as a goes to 0 are of fundamental importance. Olver . k = 0. In this case. When consulting any particular reference. x > 0.33) a→0 This limiting behavior should remind the reader of our construction (5. and so the Symmetry Principle of Theorem 7. However. ∞. which is achieved by an replacing 2 π by 2 π.√ Warning: Some authors omit the 2 π factor in the deﬁnition (7. neither the Fourier transform integral nor its inverse are well-deﬁned as standard (Riemann.2 required a > 0 for the Fourier integrals to converge.6) of the Fourier transform f (k). x < 0. All of the functions in Example 7. Nevertheless.

Indeed. whose proof is left as an exercise for the reader. On the other hand. 1/24/10 312 c 2010 Peter J. Both results are particular cases of the general Shift Theorem. Conversely. the delta function singularity at k = 0 manifests the lack of decay of the constant function. the Symmetry Principle tells √ that the Fourier transform of a pure us i κx exponential e will be a shifted delta spike 2 π δ(k − κ). Since the delta function vanishes for x = ξ. · denotes the L2 Hermitian inner product on R.16) of the delta function. (7. not deﬁned because the inﬁnite integrals of the oscillatory sine and cosine functions don’t converge! However. we compute ∞ e− i k ξ 1 δ(x − ξ) e− i k x dx = √ .9) yields ∞ δ(k) e i k x dk = e i k 0 = 1. Olver . 2π 2π −∞ ∞ (7. As in the preceding case. combination) of the basic formulae (7.6) and (7. (7. 2π (7. the inverse transform formula (7.38) This is a perfectly valid formula. formally. this identity can be validly interpreted within the framework of weak convergence and generalized functions. the delta function δ(x) has constant Fourier transform 1 δ(k) = √ 2π e− i k 0 1 δ(x) e− i k x dx = √ ≡√ . To determine the Fourier transform of a delta spike δξ (x) = δ(x − ξ) concentrated at position x = ξ. Applying the inverse Fourier transform (7. this statement must be taken with a grain of salt.36) δξ (k) = √ 2 π −∞ 2π The result is a pure exponential in frequency space. to the remarkable identity δξ (x) = δ(x − ξ) = 1 2π ∞ e− i k(x−ξ) dk = −∞ 1 eikξ .3. But it is possible to make sense of this identity within the language of generalized functions. multiplying both sides by f (x).35) a result that also follows from the Symmerty Principle of Theorem 7. strictly speaking. this identity is telling us that complex exponentials of diﬀering frequencies are mutually orthogonal. However. −∞ which is in accord with the basic deﬁnition (5. since the integral does not converge in any standard sense (either Riemann or Lebesgue).The direct transform integral 1 δ(k) = 2π ∞ e− i k x dx −∞ is. rather. we ﬁnd 1 f (ξ) = 2π ∞ −∞ ∞ −∞ f (x) e− i k(x−ξ) dx dk. Vice versa. eikx . concentrated at frequency k = κ.37) where · . and then integrating with respect to x.9) connecting the direct and inverse Fourier transforms of the function f (x). being a restatement (or.9) leads.

1/24/10 313 c 2010 Peter J. with a > 0 and c = 0.Short Table of Fourier Transforms f (x) 1 δ(x) σ(x) sign x σ(x + a) − σ(x − a) e− a x σ(x) ea x (1 − σ(x)) e− a | x | e − a x2 f (k) √ 2 π δ(k) 1 √ 2π π i δ(k) − √ 2 2π k −i 2 1 π k 2 sin a k π k 1 √ 2 π (a + i k) √ 1 2 π (a − i k) 2 a 2 + a2 π k e− k /4 a √ 2a π 3/2 √ δ(k) − i 2 e i k d/c f |c| f (− k) f (− k) i k f (k) i f ′ (k) √ 2 π f (k) g(k) π e− | k | 2 k k c 2 tan−1 x f (c x + d) f (x) f (x) f ′ (x) x f (x) f ∗ g(x) Note: The parameters a. Olver . c. d are real.

1/24/10 314 c 2010 Peter J. A more extensive collection of Fourier transforms can be found in [100]. Since the Fourier transform uniquely associates a function f (k) on frequency space with each (reasonable) function f (x) on physical space. Practical applications rely on tables (or. Derivatives and Integrals.) Diﬀerentiation Let us begin with derivatives. then the Fourier transform of the shifted function f (x − ξ) is e− i k ξ f (k). the two basic operations in calculus — diﬀerentiation and integration of functions — are realized as algebraic operations on their Fourier transforms. −∞ (7. one can characterize functions by their transforms. (The downside is that algebraic operations on functions become more complicated in the transform domain.41) dk † We are assuming the integrand is suﬃciently nice so that we can bring the derivative under the integral sign.2. the Fourier transform of x f (x) is obtained by diﬀerentiating the Fourier transform of f (x): df F [ x f (x) ] = i .40) Similarly.3. each tabular entry can be used to deduce two diﬀerent Fourier transforms. even better. we obtain 1 f ′ (x) ∼ √ 2π ∞ i k f (k) e i k x dk. (7. The Fourier transform of the derivative f ′ (x) of a function is obtained by multiplication of its Fourier transform by i k: F [ f ′ (x) ] = i k f (k). 7. Olver . the transform of the product function e i κ x f (x).5.39) The resulting integral is itself in the form of an inverse Fourier transform. by applying the Symmetry Principle of Theorem 7.4. Note that. If we diﬀerentiate† the basic inverse Fourier transform formula ∞ 1 √ f (k) e i k x dk. One of the most remarkable and important properties of the Fourier transform is that it converts calculus into algebra! More speciﬁcally. computer algebra systems such as Mathematica or Maple) that recognize a wide variety of transforms of basic functions of importance in applications. for real κ. If f (x) has Fourier transform f (k). see [ 44. The accompanying table lists some of the most important examples of functions and their Fourier transforms. 143 ] for a fully rigorous justiﬁcation. (7.Theorem 7. Proposition 7. namely of i k f (k) which immediately implies the following key result. f (x) ∼ 2 π −∞ with respect to x. is the shifted transform f (k − κ). Similarly.

such as smoothness. the more rapid the decay of its Fourier transform. the odd exponential pulse has a jump discontinuity of magnitude 2 at x = 0.The second statement follows from the ﬁrst by use of the Symmetry Principle of Theorem 7.40). then its Fourier transform f (n) (k) = ( i k)n f (k) must go to zero as | k | → ∞. just like Fourier series. which contributes an additional delta function in frequency space. (This result can be viewed as the Fourier transform version of the Riemann–Lebesgue Lemma 3. smoothness of f (k) implies decay of f (x) as x → ± ∞. We already noted that the Fourier transform of a (nice) function must decay to zero at large frequencies: f (k) → 0 as | k | → ∞. there is an extra constant involved. As with Fourier series.5 says that their Fourier transforms are related by i k f e (k) = i 2 ka = − a f o (k). local features of f (x). The Fourier transform of f (n) (x) is ( i k)n f (k). As a general rule of thumb. The Symmetry Principle implies that reverse is also true: global features of f (x) correspond to local features of f (k).7. The derivative of the even exponential pulse fe (x) = e− a | x | is a multiple of the odd exponential pulse fo (x) = (sign x) e− a | x | : fe′ (x) = − a (sign x) e− a | x | = − afo (x). this is not completely correct.21. Example 7. π k 2 + a2 as previously noted in (7. If we multiply (7. and so its derivative contains a delta function: fo′ (x) = − a e− a | x | + 2 δ(x) = − a fe (x) + 2 δ(x). such as decay for large | k |.3. the smoother f (x). Higher order derivatives are handled by iterating the ﬁrst order formula (7. 1/24/10 315 c 2010 Peter J.24) by i k we obtain i k f o (k) = 2 k2 = π k 2 + a2 2 − π 2 a2 = 2 δ(k) − a f e (k). This is reﬂected in the relation between their Fourier transforms.41. Integration Integration is the inverse operation to diﬀerentiation. Proposition 7. Corollary 7. This requires that f (k) go to zero more rapidly than | k |− n . and so should correspond to division by i k in frequency space.6. π k 2 + a2 The Fourier transform. Olver . Thus. This local-global duality is one of the major themes of Fourier theory. For instance.) If the nth derivative f (n) (x) is also a reasonable function. On the other hand. 24). is completely compatible with the calculus of generalized functions. are manifested by global features of f (k). This result has an important consequence: the smoothness of f (x) is manifested in the rate of decay of its Fourier transform f (k).

E. The Fourier transform of the inverse tangent function x Q. In the following chapter.9. Since h(x) → 0 as | x | → ∞. by subtracting a suitable multiple of the step function from the integral.Proposition 7.43) 2 π f (0) δ(x).43) and (7. If f (x) has Fourier transform f (k). we will use these techniques to solve partial diﬀerential equations. The fact that the Fourier transform converts diﬀerentiation in the physical domain into multiplication in the frequency domain is one of its most compelling features.44) establishes the desired formula (7. A particularly important consequence is that it eﬀectively transforms diﬀerential equations into algebraic equations. we can apply our diﬀerentiation rule (7. 2 k 2 7.3. Example 7.8 with (7. and thereby opens the door to their solution by elementary algebra! One begins by applying the Fourier transform to both sides of the diﬀerential equation under consideration. (7.44) Combining (7. 1/24/10 316 c 2010 Peter J. then the Fourier transform of x its integral g(x) = −∞ f (y) dy is g(k) = − i f (k) + π f (0) δ(k). and conclude that i k h(k) = f (k) − f (0). Solving the resulting algebraic equation will produce a formula for the Fourier transform of the desired solution. f (x) = tan−1 x = 0 dy = 1 + y2 x −∞ dy π − 2 1+y 2 can be computed by combining Proposition 7.40). Green’s Functions and Convolution. h′ (x) = f (x) − √ i f (0) . which can then be immediately reconstructed via the inverse Fourier transform. the resulting function √ h(x) = g(x) − 2 π f (0) σ(x) decays to 0 at both ±∞. Olver . x → +∞ lim g(x) = −∞ f (x) dx = √ 2 π f (0).8.42). k ∞ (7. Therefore.42) Proof : First notice that x → −∞ lim g(x) = 0. Consulting our table of Fourier transforms. we ﬁnd h(k) = g(k) − π f (0) δ(k) + On the other hand.28): f (k) = − i π e− | k | π 3/2 + √ δ(k). k (7.D.

if the forcing function is real. + ω2 (7. In lieu of boundary conditions.9): ∞ h(k) e i k x 1 dk. Taking Corollary 7. Peter J. In quantum mechanics. Olver c 2010 . dx2 where ω > 0 is a positive constant. For example. ω 2 = 1.Solution of Boundary Value Problems The Fourier transform is particularly well adapted to boundary value problems on the entire real line.) The Fourier integral can be explicitly evaluated by using partial fractions to rewrite u(k) = 1/24/10 2 1 = 2 + ω 2 )(k 2 + 1) π (k 2 1 2−1 π ω 317 k2 1 1 − 2 + 1 k + ω2 . 89]. 2+1 π k then (7. the transformed equation can be immediately solved for u(k) = k2 h(k) . − ∞ < x < ∞. [83. consider the boundary value problem d2 u (7.45) + ω 2 u = h(x). if the forcing function is an even exponential pulse. h(x) = e− | x | with h(k) = 1 2 . As a speciﬁc example. In place of the boundary conditions used on ﬁnite intervals. (Indeed. the result is a linear algebraic equation k 2 u(k) + ω 2 u(k) = h(k) − relating the Fourier transforms of u and h.46) Therefore. We will solve this problem by applying the Fourier transform to both sides of the diﬀerential equation. and correspond to subatomic particles that are trapped or localized in a region of space by some sort of force ﬁeld. we can reconstruct the solution by applying the inverse Fourier transform formula (7.47) u(x) = √ 2 π −∞ k 2 + ω 2 For example.7 into account. we require that the solution u(x) → 0 as | x | → ∞. the solution must also be real. + ω 2 )(k 2 + 1) noting that the imaginary part of the complex integral vanishes because the integrand is an odd function. (7. Unlike the diﬀerential equation. we look for solutions that decay to zero suﬃciently rapidly as | x | → ∞ — in order that their Fourier transform be well-deﬁned (in the context of ordinary functions). the electrons in an atom are bound states localized by the electrostatic attraction of the nucleus.47) writes the solution as a Fourier integral: u(x) = 1 π ∞ −∞ ei kx 1 dk = 2 + ω 2 )(k 2 + 1) (k π ∞ −∞ (k 2 cos k x dk . these solutions are known as the bound states of the system.

according to our Fourier Transform Table. As a function of x. and conclude that the Green’s function for this boundary value problem is an exponential pulse centered at ξ.49) Observe that. The resulting square-integrable solution is the Green’s function G(x. The Green’s function superposition principle tells us that the solution to the inhomogeneous boundary value problem (7. (7. decays to 0 as | x | → ∞. as with other self-adjoint boundary value problems. it satisﬁes the homogeneous diﬀerential equation − u′′ + ω 2 u = 0. The “resonant” case ω 2 = 1 is left as an exercise. We apply Theorem 7. A particularly important case is when the forcing function h(x) = δξ (x) = δ(x − ξ) represents a unit impulse concentrated at x = ξ.45). we ended up deriving a formula for its Fourier transform (7. It also decays as | x | → ∞.45) under a general forcing can be represented in the integral form ∞ u(x) = −∞ G(x. representing the Fourier transform of δξ (x).4. 2 π k2 + ω 2 which is the product of an exponential factor e− i k ξ . the solution to this boundary value problem is 1 e− | x | − e− ω | x | ω when ω 2 = 1. Olver . meaning that it is twice continuously diﬀerentiable (which is not so immediately apparent from the formula). ξ) for the boundary value problem. 2ω (7. and satisﬁes the diﬀerential equation everywhere. According to (7. except at the point x = ξ when its derivative has a jump discontinuity of unit magnitude. times a multiple of the Fourier transform of the even exponential pulse e− ω | x | . ξ) = √ . Remark : The method of partial fractions that you learned in ﬁrst year calculus is often an eﬀective tool for evaluating (inverse) Fourier transforms of rational functions.50) The reader may enjoy recovering the particular exponential solution (7. −∞ (7. The ﬁnal 1/24/10 318 c 2010 Peter J.46) as the product of two known Fourier transforms. Convolution In our solution to the boundary value problem (7. namely G(x.48) from this integral formula. its Fourier transform with respect to x is 1 e− i k ξ G(k. ξ) h(ξ) dξ = 1 2ω ∞ e− ω| x−ξ | h(ξ) dξ. the Green’s function is symmetric under interchange of x and ξ. as required by the boundary conditions.48) u(x) = ω2 − 1 The reader may wish to verify that this function is indeed a solution. ξ) = 1 − ω | x−ξ | e .46).Thus.

(b) Bilinearity: a. b ∈ C. the delta function plays the role of the “convolution unit”: ∞ f (x) ∗ δ(x) = −∞ f (x − ξ) δ(ξ) dξ = f (x). is indicative of a general property. In fact. obtained by applying the inverse Fourier transform.50).49) has the form of a convolution product between an even exponential pulse g(x) = (2 ω)−1 e− ω| x | and the forcing function: u(x) = g(x) ∗ h(x). (e) Delta function: f ∗ δ = f.Green’s function formula (7. 1/24/10 319 c 2010 Peter J. leaving their veriﬁcation as exercises for the reader. (d) Zero function: f ∗ 0 = 0. (a f + b g) ∗ h = a (f ∗ h) + b (g ∗ h). its Fourier transform (7. and conversely. (a) Symmetry: f ∗ g = g ∗ f. Deﬁnition 7.10.51) We record the basic properties of the convolution product. our solution (7. the ordinary multiplicative product √ u(k) = 2 π g(k) h(k) of the Fourier transforms of g and h. Olver . In fact. (c) Associativity: f ∗ (g ∗ h) = (f ∗ g) ∗ h. indeed. f ∗ (a g + b h) = a (f ∗ g) + b (f ∗ h). All of these assume that the implied convolution integrals converge. Its right hand side has the form of a convolution product between functions. One tricky feature is that the constant function 1 is not a unit for the convolution product. The convolution of scalar functions f (x) and g(x) is the scalar function h = f ∗ g deﬁned by the formula ∞ h(x) = f (x) ∗ g(x) = −∞ f (x − ξ) g(ξ) dξ. (7.46) is. In particular. up to a factor. ∞ f ∗1= f (ξ) dξ −∞ is a constant function — the total integral of f — not the original function f (x). On the other hand. this is a general property of the Fourier transform: convolution in the physical domain corresponds to multiplication in the frequency domain. according to the ﬁnal property.

Example 7.12.51). h(k) = √ 2π (7.52) Vice versa. Olver . | k | > 1. [8].3. Applying the change of variables η = x − ξ in the inner integral produces 1 h(k) = √ 2π √ = 2π ∞ −∞ ∞ f (η) g(ξ) e− i k(ξ+η) dξ dη 1 √ 2π −∞ ∞ f (η) e −∞ −ikη dη 1 √ 2π ∞ g(ξ) e− i k ξ dξ −∞ = √ 2 π f (k) g(k). (7. we ﬁnd 1 h(k) = √ 2π 1 h(x) e− i k x dx = √ 2π −∞ ∞ ∞ −∞ ∞ −∞ f (x − ξ) g(ξ) e− i k x dx dξ. We already know.Theorem 7. (7. proving (7.52). the Fourier transform of their product h(x) = f (x) g(x) = 1/24/10 320 sin x x2 c 2010 Peter J.D. 2 0. up to multiple. π sign k. the Fourier transform of their product h(x) = f (x) g(x) is. that the Fourier transform of f (x) = is the box function f (k) = sin x x π . or by simply noting that it follows directly from the Symmetry Principle of Theorem 7. the convolution of their Fourier transforms: 1 f (k) ∗ g(k).53) Proof : Combining the deﬁnition of the Fourier transform with the convolution formula (7. We also know that the Fourier transform of g(x) = 1 x is π σ(k + 1) − σ(k − 1) = 2 −1 < k < 1.11. Q.E. The second formula can be proved in a similar fashion. where we are assuming that the integrands are suﬃciently nice to allow us to interchange the order of integration. The Fourier transform of the convolution h(x) = f (x) ∗ g(x) of two functions is a multiple of the product of their Fourier transforms: √ h(k) = 2 π f (k) g(k). 2 g(k) = − i Therefore.29).

In Section 3. Not every function admits a Fourier transform in the classical sense† — the Fourier integral (7. taking functions f (x) on physical space to functions f (k) on frequency space. −1 < k < 1.6) is required to converge. 2 −1 −i π 2 f (κ) g(k − κ) dκ k < −1. 1/24/10 321 c 2010 Peter J. it is worth outlining a few of the more important features. The Fourier transform of sin x .5 -1 1 2 3 Figure 7. when deﬁned. b ] on a ﬁnite interval. here we adapt Deﬁnition 3.4. 7.3.5. and this places restrictions on the function and its asymptotics at large distances. k > 1. Thus.5 -3 -2 -1 -0. x2 can be obtained by convolution: 1 1 f (k) ∗ g(k) = √ h(k) = √ 2π 2π ∞ −∞ = −i π 8 A graph of h(k) appears in Figure 7.3. i π 2 1 π sign(k − κ) dκ = −i k. the Hilbert space L2 = L2 (R) is the inﬁnite-dimensional vector space consisting † We leave aside the more advanced issues involving generalized functions in this subsection. While we do not have the analytical tools to embark on a fully rigorous treatment of the mathematical theory underlying the Fourier transform. It turns out the proper setting for the rigorous theory is the Hilbert space of complexvalued square-integrable functions — the same inﬁnite-dimensional vector space that lies at the heart of modern quantum mechanics. is a linear map. We have already noted that the Fourier transform. we already introduced the Hilbert space L2 [ a.32 to the entire real line. The Fourier Transform on Hilbert Space. Olver .1 0. A critical question is precisely which function space should the theory be applied to.

then the inverse Fourier transform integral converges to the average value 1 − + 2 f (x ) + f (x ) . g . i. c 2010 Peter J. (7. 130]. Theorem 7. More generally. ∞ ∞ f (x) g(x) dx = −∞ −∞ f (k) g(k) dk. (7. and the precise deﬁnitions and identiﬁcation of its elements is quite subtle. any piecewise continuous function that satisﬁes the decay criterion | f (x) | ≤ M . if the right and left hand limits f (x− ).56) so that f 2 = f . If f (x) ∈ L2 is square-integrable.14. f .g | ≤ f g (7. f (x+ ).9) equals its value f (x). The Hermitian inner product on the complex Hilbert space L2 is prescribed in the usual manner.119). Theorem 7. g ∈ L2 .5. the Fourier transform preserves inner products.6) of the Fourier transform to evaluate ∞ ∞ f (k) g(k) dk = −∞ −∞ ∞ 1 √ 2π ∞ −∞ ∞ f (x) e− i k x dx −∞ 1 √ 2π ∞ g(y) e+ i k y dy −∞ dk = −∞ f (x) g(y) 322 1 2π ∞ e− i k(x−y) dk −∞ dx dy. (7.57) ensures that the inner product integral is ﬁnite whenever f. We use the deﬁnition (7. (7. then its Fourier transform f (k) ∈ L2 is a well-deﬁned. belongs to L2 . A rigorous proof of this result can be found in [44. f ′ (x+ ) exist.13. The Cauchy–Schwarz inequality | f . In fact. the Fourier transform f = F [ f ] deﬁnes a linear transformation from L2 functions of x to L2 functions of k. and also f ′ (x− ). whose Fourier series counterpart appeared in (3. then f .. | x |1/2+δ for all suﬃciently large | x | ≫ 0. If f (x) is continuously diﬀerentiable at a point x.58) Proof : Let us sketch a formal proof that serves to motivate why this result is valid.e. then its inverse Fourier transform (7. Let us state the fundamental result governing the eﬀect of the Fourier transform on functions in Hilbert space. as in Section 3. It can be regarded as a direct analog of the Pointwise Convergence Theorem 3. Hilbert space contains many more functions. This important result is known as Parseval’s formula. ∞ f .of all complex-valued functions f (x) which are deﬁned for all x ∈ R and have ﬁnite L2 norm: ∞ f 2 = −∞ | f (x) |2 dx < ∞.8 for Fourier series. Thus. g = f .55) for some M > 0 and δ > 0. However. If f (k) = F [ f (x) ] and g(k) = F [ g(x) ]. Olver 1/24/10 .g = −∞ f (x) g(x) dx.54) For example. square-integrable function of the frequency variable k.

the squared modulus of the wave function. Indeed. one of the founders of modern quantum mechanics. the Heisenberg Uncertainty Principle is a by now familiar philosophical concept. orthogonal functions. the more precisely one measures the position of a particle. in fact. the greater the accuracy in the momentum. g = 0. are interrelated by the Fourier transform. As we already noted in Section 3. satisfying f . a mathematical property of the Fourier transform! In quantum theory. The smaller the hole. Choosing f = g in Parseval’s formula (7. belonging to the underlying state space.E. in a one-dimensional model of a single particle. mapping L2 functions of the physical variable x to L2 functions of the frequency variable k. explicitly. and the wider the image on the screen.5. −∞ This completes our “proof”. see [44. in a physical system. For instance.59) Thus. Quantum Mechanics and the Uncertainty Principle In its popularized form. and states that. What we will show is that the Uncertainty Principle is. Olver . complex valued functions of x. (7. vice versa. is the Hilbert space L2 = L2 (R) consisting of square integrable. This Fourier transform-based duality between position and momentum. meaning the less certainty there is in the observed momentum.Now according to (7. the more constrained the position. the wave functions of a quantum system are characterized as the elements of unit norm. This is not the place to discuss the philosophical and experimental consequences of Heisenberg’s principle.5 says that the Fourier transform of the diﬀerentiation operator representing momentum is a multiplication operator representing position and vice versa. or. which. | ϕ(x) |2 .D. the less accuracy there will be in the measurement of its momentum.. Q. the eﬀect of their momenta is in the pattern of light diﬀused on a screen placed beyond the hole. represents the probability density of the 1/24/10 323 c 2010 Peter J. certain quantities cannot be simultaneously measured with complete accuracy. Experimental veriﬁcation of the uncertainty principle can be found even in fairly simple situations. ϕ = 1.37). each of the paired quantities. position and momentum. −∞ | f (x) |2 dx = −∞ | f (k) |2 dk. In quantum mechanics. The position of the photons is constrained by the hole. the Fourier transform F : L2 → L2 deﬁnes a unitary or norm-preserving linear transformation on Hilbert space. In particular. the inner k integral can be replaced by a delta function δ(x − y). e. A similar uncertainty couples energy and time. and hence ∞ ∞ ∞ −∞ ∞ f (k) g(k) dk = −∞ −∞ f (x) g(y) δ(x − y) dx dy = f (x) g(x) dx. the less certainty in its position. f . g = 0. Consider a light beam passing through a small hole. 130] for a rigorous version. Proposition 7. It was ﬁrst formulated in the 1920’s by the German physicist Werner Heisenberg. or multiplication and diﬀerentiation.g.58) results in the Plancherel formula ∞ ∞ f 2 = f 2 . will have orthogonal Fourier transforms. lies at the heart of the Uncertainty Principle.

Consequently. x | ϕ(x) |2 dx (7.40) to infer that k ϕ(k) is the Fourier transform of − i ϕ′ (x). the statistical deviation of the particle’s measured position from the mean. x = ∞ (7.60) In particular.particle being found at position x. We note that ∞ x2 = −∞ x2 | ϕ(x) |2 dx = x ϕ(x) 2 .58) to convert to an integral over position. (7. Therefore.62) is the variance.63) On the other hand. p 2 = = 2 ∞ −∞ ∞ k | ϕ(k) | dk = − | ϕ (x) | dx = ′ 2 2 2 2 2 ϕ(x) ϕ′′ (x) dx −∞ (7. while ∆x.61) −∞ is the mean or average measured position of the particle. the mean value of any function of momentum g(p) is given by its integral against the squared modulus of the Fourier transformed wave function: = ∞ g(p) = −∞ g( k) | ϕ(k) |2 dk.64) 2π is Planck’s constant. and denoted by ∞ f (x) = −∞ f (x) | ϕ(x) |2 dx.055 × 10−34 Joule seconds (7. and (7. (∆p)2 = p− p 2 = p2 − p ∞ 2 (7.67) is the squared variance in the momentum. Similarly. the mean value of any function f (x) of the position variable is given by its integral against the system’s probability density. whose value governs the quantization of physical quantities. 1/24/10 324 c 2010 Peter J. where h ≈ 1.66) where we used Parseval’s formula formula (7. by a similar computation. (7.65) In particular. Olver . where. the momentum variable p is related to the Fourier transform frequency via the de Broglie relation p = k. ϕ .68) . deﬁned by (∆x)2 = x− x 2 = x2 − x 2 (7. the Uncertainty Principle can be stated as follows. the mean of the momentum measurements of the particle is given by ∞ p = −∞ k | ϕ(k) | dk = − i 2 −∞ ϕ′ (x) ϕ(x) dx = − i ϕ′ . that is. 2 −∞ ϕ (x) ′ 2 With this interpretation. ∞ (7.

1/24/10 325 c 2010 Peter J. For more details and physical consequences. Thus. where we employed integration by parts.69) quantiﬁes the statement that the more accurately we are able to measure the momentum p. ϕ′ (x) ∞ + ϕ′ (x) 2 . Thus.E. noting that the boundary terms vanish since ϕ(x) → 0 as | x | → ∞.D. ϕ (x) = ∞ ′ ′ x ϕ′ (x) ϕ(x) + x ϕ(x) ϕ′ (x) dx −∞ ∞ −∞ = −∞ x d | ϕ(x) |2 dx = − dx | ϕ(x) |2 dx = −1.63. the Heisenberg inequality (7. where its value is p2 2 − 1 ≥0 4 x2 and hence x2 p2 ≥ 1 4 2 .69) . The minimum value of the left hand side of this inequality occurs at t⋆ = 1/(2 x2 ). To obtain the Uncertainty Relation. but with x− x replacing x and p − p replacing p.Theorem 7.. then the variances in (7. [83. The result is x− x 2 t −t+ 2 p− p 2 2 = (∆x)2 t2 − t + (∆p)2 2 ≥ 0. (7. you should consult an introductory text on mathematical quantum mechanics. (7. 89].g. e.15. 68). The smaller the variance of a quantity such as position or momentum. If ϕ(x) is a wave function.69). and vice versa.70) The middle term can be evaluated as follows: ϕ (x) . (7. the less accurate will be any measurement of its position x. x ϕ(x) + x ϕ(x) .71) Q. one performs the selfsame calculation. Olver . Proof : For any value of the real parameter t. 0 ≤ t x ϕ(x) + ϕ′ (x) = t2 x ϕ(x) 2 2 +t ϕ′ (x) . the more accurate will be its measurement. in view of (7.70) reads x2 t 2 − t + p2 2 ≥0 for all real t. x ϕ(x) + x ϕ(x) . so position and momentum satisfy the inequality ∆x ∆p ≥ 1 2 ϕ = 1. Substituting t = 1/(2 (∆x)2 ) produces the Heisenberg inequality (7.

- D1142.1210479-1
- Defectologia - Soldadura
- 000043
- Reglamento Unico de Transporte de Gas Natural
- NTC3353 - SIDERURGIA
- 000043
- DOT Pipeline Repairs
- Dummy
- Regresion Lineal Multiple 3
- Regresion Lineal Multiple 3
- Cartas psicrométricas a diferentes altitudes
- Calculo Integral - Granville (Solucionario)
- Química e Investigación Criminal
- Electricidad Basica en Reparacion de Automoviles - Cesvimap
- Fractales - Manuel
- Excel Desde Cero
- Introducción a la teoría de números, ejemplos y algoritmos
- CALCULO VARIACIONAL
- QM

- ft
- m485
- lesson plan.doc
- Mathematica - Fourier Series.pdf
- Definition of Fourier Series and Typical Examples
- Fourier Transform
- Fourier Integral
- FourierNotes
- MA 588 Fourier Transforms
- 620-332
- fourier
- [A._P._Prudnikov,_K.A._Skórnik]_Operational_calculus
- Fourier transform (2).docx
- mathfs
- 776 Notes - Fft Methods
- m483
- HarHarmonic Analysismonic Analysis Lecture3
- EVALUATION OF FOURIER
- 00600_FourierTransform
- Signal Processing
- Representation Theorems in Hardy Spaces
- Reading Fourier
- Definition of Fourier Series and Typical Examples
- Bsc1,Bsco1,Bse1,Bset1,Bsm1 Mathematics III
- Fourier3 Computational Methods
- Adjoint Operators of Two Dimensional Fractional Fourier Mellin Transform
- Advanced Mechatronics Systems - Laboratory - Mechatronics Control Systems - 1st Ed
- Convergence
- The Fourier Series
- SS-R09-NOV2013
- Fourier Transform

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd