Professional Documents
Culture Documents
Fejer Theorem
Fejer Theorem
William Wu
wwu@ocf.berkeley.edu
June 1 2004
A Fourier series can be understood as the decomposition of a periodic function into its projections onto an orthonormal basis. More precisely, consider the vector space of continuous
functions from [, ] to R, on which we define the inner product between two functions f
and g as
Z
hf, gi =
f (x)g(x)dx.
X
(an cos(nt) + bn sin(nt))
a0 =
an =
bn =
R
1
2 f (x)dx
R
1
f (x) cos(nx)dx
R
1
f (x) sin(nx)dx.
for n > 0
These coefficients are the projections of f onto the orthonormal basis functions
1 cos(mx) sin(nx)
,
,
m, n Z.
William Wu
f(n)einx
f (x)einx dx.
inx
f
(x)
f
(n)e
dx = 0.
Z
lim
n=N
William Wu
for a fixed x in [, ], there are no guarantees on the difference between f (x) and the
series approximation at x.
A stronger and quite natural sense of convergence is pointwise convergence, in which
we demand that at each point x [, ], the series approximation converges to f (x).
Jordans Pointwise Convergence Theorem then states that if f is sectionally continuous and
0
x0 is such that the one-sided derivatives f 0 (x+
0 ) and f (x0 ) both exist, then the Fourier
P
inx
series n f (n)e 0 converges to f (x0 ). This theorem is often useful for proving pointwise
convergence, and its conditions often hold. However, sometimes pointwise convergence can
be an inappropriate notion of convergence. A canonical example is the sequence of functions
defined by gn (x) : x xn for x [0, 1]. Then (gn ) converges pointwise to a function h
that equals 0 for x [0, 1), but equals 1 for x = 1. Thus although (gn ) consists only of
continuous functions, oddly the limit function is discontinuous.
To avoid such problems, we desire the even stronger notion of uniform convergence, such
that the rate at which the series converges is identical for all points in [, ]. By adopting
the metric
d(f, g) = sup{|f (x) g(x)| : t [, ]}
over the space of continuous functions from [, ] to R, we can force convergence to imply
uniform convergence, simply by definition. This metric space is denoted by C([, ], R).
It can also be proven that C([, ], R) is a vector space, and thus the concept of series is
well-defined.
We are now primed to appreciate Fejers remarkable theorem.
Fej
ers Theorem: Let f : [, ] R be a continuous function with f () = f (). Then
the Fourier series of f (C,1)-converges to f in C([, ], R), where C([, ], R) is the
metric space of continuous functions from [, ] to R.
Without imposing any additional conditions on f aside from being continuous and periodic,
Fejers theorem shows that Fourier series can still achieve uniform convergence, granted that
we instead consider the arithmetic means of partial Fourier sums.
2
2.1
Proof
Fej
ers Kernel
Before proceeding further, we first prove some properties of Fejers kernel a trigonometric
polynomial that often appears in Fourier analysis. These properties will be useful in the
proof of Fejers theorem.
William Wu
Fejers kernel can be expressed in either of the following two equivalent ways:
Fn (x) =
(1)
Fn (x) =
1 X
Dk (x)
n+1
(2)
k=0
P
where Dk (x) is the Dirichlet kernel Dk (x) = km=k eimx . Depending on the circumstances,
one form of Fejers kernel can lend more clarity than the other. Conversion between the two
forms is just a tedious exercise in manipulating trigonometric identities. To avoid detracting
from the flow of our presentation, we will not present the proof of this conversion here.
However, the meticulous reader is welcome to read the proof in the Appendix section.
Lemma: The Fejer kernel has the following properties:
Z
1
Fn (x)dx = 1
2
Fn (x) 0
i
ii
(3)
(4)
n |x|
Fn (x)dx = 0
(5)
Proofs:
i
1
2
Fn (x)dx = 1
We appeal to the second form of Fejers kernel given by (2). Substituting the definition
of Dirichlets kernel yields:
Fn (x) =
n
k
1 X X imx
e .
n+1
k=0 m=k
1
2
Fn (x)dx =
=
1
2
1
n+1
"
#
n
k
1 X X imx
e
dx
n+1
n
X
k=0 m=k
k
X
k=0 m=k
1
2
imx
dx .
William Wu
1
eimx = 0. But when m = 0, 2
1
2
eimx = 1. Thus,
1 X
1 = 1.
n+1
Fn (x) =
k=0
ii Fn (x) 0
The non-negativity of the Fejer kernel follows immediately from the first form of the
Fejer kernel (1).
R
iii For each fixed > 0, limn |x| Fn (x)dx = 0.
We again use the Fejer kernels first form. For |t| we have
Thus
0 Fn (x)
1
1
,
2
n + 1 sin /2
2.2
1
sin2 x/2
1
.
sin2 /2
|x| .
Fej
ers Theorem
To discuss Cesaro convergence of Fourier series, we introduce notation for both the partial
Fourier sums, and the arithmetic means of those partial sums. Denote the nth partial sum
of the Fourier series by sn , and denote the corresponding nth Cesaro sum by n .
sn (x) =
n
X
f(k)eikx
(6)
k=n
n
n (x) =
1 X
sk (x)
n+1
(7)
k=0
Now we aim to rewrite these expressions in terms of the Fejer kernel. Substituting the
integral form of f(k) into (6) yields
sn (x) =
n
X
k=n
Z
n
X
1
ikt
f (t)e
dt eikx
2
k=n
Z
n
X
1
=
f (t)
eik(xt) dt
2
f(k)eikx =
k=n
William Wu
1
2
f (t)Dn (x t)dt.
f (x t)Dn (t)dt.
(8)
Z
n
1 X 1
f (x t)Dk (t)dt
n+1
2
k=0
"
#
Z
n
1 X
1
f (x t)
Dk (t) dt
2
n+1
k=0
Z
1
f (x t)Fn (t)dt
2
1 X
n (x) =
sk (x) =
n+1
k=0
=
=
1
n (x) f (x) =
2
(f (x t) f (x))Fn (t)dt.
1
2
|f (x t) f (x)| Fn (t)dt.
Continuous functions on [, ] are uniformly continuous. That is, given > 0, there exists
a > 0 such that |x y| implies |f (x) f (y)| . We now break our integral into two
integrals, with the limits of integration divided about and .
|n (x) f (x)| =
1
2
1
|f (x t) f (x)| Fn (t)dt +
2
|t|
Z
|t|
!
|f (x t) f (x)| Fn (t)dt .
William Wu
Z
|t|
Fn (t)dt
1
2
Fn (t)dt =
M
2M Fn (t)dtFn (t)dt =
|t|
Z
|t|
Fn (t)dt.
R
Finally by Lemma iii, there exists an N N such that for all n N , |t| Fn (t)dt .
Conclusively, for all n N , |f (x) n (x)| + = 2. This completes the proof.
In this section, we examine one of the very first applications of Fourier series. It dates from
Fouriers seminal 1807 paper The Analytical Theory of Heat, in which Fourier series are
used to solve the practical problem of heat flow in various metallic solids.
Imagine a wire of unit length that is twisted into a circle. Suppose this circle is heated by
some continuous initial temperature distribution f . As time passes, the heat redistributes
itself about the circle, moving from hotter areas to colder areas. After a long time, we would
expect the heat to be evenly distributed over the circle. But in the interim, we would like
an expression for the temperature as a function of both space and time.
It is convenient to think of this circle as the unit interval [0, 1] wrapped around on itself.
We then denote the temperature function by u(x, t), where x is the spatial variable lying
on the real line modulo 1, and t is time. The initial condition is then u(x, 0) = f (x). Note
that the circularity forces u(x, t) to be periodic in x with period 1, for any fixed t. Thus
u(x, t) has some Fourier series expansion
u(x, t) =
cn (t)e2inx
n=
William Wu
At this point we recall Newtons famous heat conduction equation, which approximates the
conduction of heat in solids. The equation is 2 uxx = ut , where 2 is the thermal diffusivity
constant; to simplify matters, we will let 2 = 12 . With the intent of applying this equation,
we first differentiate cn (t) with respect to t
Z
c0n (t)
=
0
ut (x, t)e2inx dx
=
0
1
uxx (x, t)e2inx dx.
2
Now we would like to remove the spatial derivatives from u. To do this we integrate by
parts twice, using the facts that e2in = 1, and u(0, t) = u(1, t) by periodicity of u. After
integrating by parts we have
Z
d2
1
u(x, t) 2 e2inx dx
dx
0 2
Z 1
1
=
u(x, t)(4 2 n2 )e2inx dx
0 2
Z 1
= (2 2 n2 )
u(x, t)e2inx dx
c0n (t)
= (2 2 n2 )cn (t)
To our approval, we discover that cn (t) obeys a canonical ordinary differential equation! Its
solution is of course
cn (t) = cn (0)e2
2 n2 t
Expressing cn (0) in integral form shows that cn (0) is simply the nth Fourier coefficient of
the initial distribution function f :
Z
cn (0) =
Z
u(x, 0)e2inx dx =
f (x)e2inx dx.
Denoting this coefficient by f(n), we can elegantly write the general solution of the heat
equation as
William Wu
u(x, t) =
2 2
f(n)e2 n t e2inx .
n=
Fn (x) =
or
n
Fn (x) =
1 X
Dk (x)
n+1
k=0
P
where Dk (x) is the Dirichlet kernel Dk (x) = km=k eimx . In this section we prove the
equivalence of these expressions. Namely, we will manipulate the second form of Fejers
kernel listed above into the first form.
We start with the following Lemma:
Lemma:
1+2
n
X
cos(kx) =
k=1
sin[(n + 12 )x)]
.
sin( x2 )
Proof: Recall the trigonometric product identity 2 cos(u) sin(v) = sin(u + v) sin(u v).
Setting u = kx and v = x2 , we then have
2 cos (kx) =
sin
k + 12 x sin k 21 x
sin x2
By substituting the above expression for 2 cos (kx), we then have a telescoping sum
k + 12 x sin k 21 x
x
2
cos(kx) =
sin
2
k=1
k=1
x
1
sin n + 2 x sin 2
=
sin x2
n
X
n
X
sin
William Wu
sin[(n + 12 )x)]
1
sin( x2 )
=
which yields the result.
Using this Lemma and De Moivres formula, we can now rewrite Dirichlets kernel as
Dk (x) =
n
X
ikx
=1+2
k=n
n
X
cos(kx) =
k=1
sin[(n + 21 )x)]
.
sin( x2 )
(n + 1)Fn (x) =
n
X
Dk (x)
k=0
n
X
sin[(n + 12 )x)]
sin( x2 )
k=0
)
( n
X
1
i(k+1/2)x
e
=
Im
sin(x/2)
k=0
(
)
i(n+1)x 1
1
e
=
Im eix/2
sin(x/2)
eix 1
)
(
1
ei(n+1)x 1
=
Im
sin(x/2)
eix/2 eix/2
=
=
=
1 cos[(n + 1)x]
2 sin2 (x/2)
sin2 [(n + 1)x/2]
.
sin2 [x/2]
References
[1] Jerrold E. Marsden and Michael E. Hoffman. Elementary Classical Analysis, (New York: Freeman, 2003).
[2] Thomas A. Garrity. All The Mathematics You Missed But Need To Know For Graduate School,
(New York: Cambridge, 2002).
[3] Brad E. Osgood. Lecture notes for EE261, (Stanford University, 2003).
10