Professional Documents
Culture Documents
AA215A Lecture 1
Vector and Function Spaces
Antony Jameson
1
Lecture 1
and concave if
f (tx1 + (1 − t)x2 ) ≥ tf (x1 ) + (1 − t)f (x2 ), (1.2)
whenever 0 < t < 1. It is strictly convex or concave if equality in these expressions implies x1 = x2 .
If f is twice differentiable, it is convex if f 00 ≥ 0 and concave if f 00 ≤ 0, and strictly convex or
concave if f 00 > 0 or f 00 < 0.
Jensen’s theorem:
If f (x) is a concave function, Ã n !
Xn X
ti f (xi ) ≤ f ti xi , (1.3)
i=1 i=1
Pn
whenever t1 , ..., tn ∈ (0, 1) and i=1 ti = 1.
This may be proved by induction. It is true for n = 2 by the definition of a concave function.
Suppose n ≥ 3 and the assertion holds for smaller values of n. Then, for i = 2, ..., n, set t0i =
2
LECTURE 1. VECTOR AND FUNCTION SPACES 3
Pn 0
ti /(1 − t1 ), so that i=2 ti = 1. Then,
n
X n
X
ti f (x1 ) = t1 f (x1 ) + (1 − t1 ) t0i f (xi )
i=1 i=2
à n !
X
≤ t1 f (x1 ) + (1 − t1 )f t0i xi
i=2
à n
!
X
0
≤f t1 x1 + (1 − t1 ) ti xi
i=2
à n
!
X
=f ti xi
i=1
and hence
n
Y n
X
api i ≤ pi a i (1.5)
i=1 i=1
1
This is known as the generalized AM-GM inequality. Setting pi = n, recovers the basic AM-GM
inequality
à n !1 n
Y n
1X
ai ≤ ai (1.6)
n
i=1 i=1
ap bq
ab = xp11 xp22 ≤ p1 x1 + p2 x2 = + (1.8)
p q
Take vectors such that
n
X n
X
|ak |p = |bk |q = 1 (1.9)
k=1 k=1
LECTURE 1. VECTOR AND FUNCTION SPACES 4
Then ¯ n ¯
¯X ¯ X n n µ
X ¶
¯ ¯ |ak |p |bk |q 1 1
¯ a b
k k¯ ≤ |a b
k k | ≤ + = + =1 (1.10)
¯ ¯ p q p q
k=1 k=1 k=1
and consequently ¯ n ¯
n
X ¯X ¯ n
X
2 ¯ ¯ 2
ak + 2ρ ¯ ak bk ¯ + ρ b2k ≥ 0 (1.13)
¯ ¯
k=1 k=1 k=1
Then setting P
| nk=1 ak bk |
ρ = − Pn 2 (1.14)
k=1 bk
Pn 2
and multiplying by k=1 bk à !à ! ¯ n ¯
n
X n
X ¯X ¯
¯ ¯
a2k b2k ≥¯ ak bk ¯ (1.15)
¯ ¯
k=1 k=1 k=1
1. ||x|| ≥ 0
• ||x + y||∞ = max |xi + yi | ≤ max |xi | + max |yi | ≤ ||x||∞ + ||y||∞
i i i
Then
¡ ¢1
||x + y||2 = (x + y)T (x + y) 2
¡ T T T
¢ 12 ³ T T 1
T 1
T
´1
2
≤ x x + 2|y x| + y y ≤ x x + 2(x x) (y y) + y y
2 2
= ||x||2 + ||y||2
This need only be proved for M (x) = N∞ (x) since if M and N are both equivalent to N∞ then
they are also equivalent to each other. Suppose that
c1 N∞ ≤ M ≤ c2 N∞ (1.23)
and
d1 N∞ ≤ N ≤ d2 N∞ (1.24)
Then
c1 N ≤ d2 M (1.25)
and
d1 M ≤ c2 N (1.26)
where
c1 c2
N ≤M ≤ N (1.27)
d2 d1
Consider the unit cell S of vectors for which Ns (x) = 1. Let x0 and x1 be elements such that
so that
||AB|| ≤ ||A|| ||B|| (1.35)
LECTURE 1. VECTOR AND FUNCTION SPACES 7
since ¯ ¯ ¯ ¯ ¯ ¯
¯X ¯ ¯X ¯ ¯X ¯
||Ax||∞ ¯
= max ¯ ¯ ¯
aij xj ¯ ≤ max ¯ ¯ ¯
aij ¯ max |xj | ≤ max ¯ aij ¯¯ ||x||∞ (1.37)
i i j i
j j j
1.6.2 1 Norm
Corresponding to ||x||1 we have
X
||A||1 = max |aij | (max absolute column sum) (1.38)
j
i
since
à !
X ¯¯ X ¯
¯ XX X X
||Ax||1 = ¯ aij xj ¯¯ ≤ |aij ||xj | = |aij | |xj |
¯
i j i j j i
à ! à !
X X X
≤ max |aij | |xj | = max |aij | ||x||1 (1.39)
j j
i j i
with the max attained when x = ej , where j is the index of the max sum.
1.6.3 2 Norm
Corresponding to ||x||2 we have
||Ax||22 xH AH Ax
= (1.40)
||x||22 xH x
where AH is the Hermitian transpose of A. P
Let ui be an eigenvector of AH A with real eigenvalue σ 2 . Set x = αi ui , then
P 2 2
xH AH Ax α σ
H
= P i 2 i ≤ σ12 (1.41)
x x αi
where σ12 is the largest eigenvalue of AH A. The decomposition is possible because ui are indepen-
dent.
concepts of three dimensional Euclidean geometry carry over to function spaces. In particular we
can introduce norms to measure the size of a function, and an abstract definition of a generalized
inner product of two functions, corresponding to the scalar product of two vectors. This leads to
the concept of orthogonal functions. The introduction of axiomatic definition leads to proofs which
hold for a variety of different norms and inner products. In order to enable standard arguments of
real analysis to be carried over, function spaces are required to satisfy an axiom of completeness
that the limit of a sequence of elements in a given space is contained in the space. Spaces with an
inner product are known as Hilbert spaces in honor of their inventor.
1.7.2 Norms
Norms will be a measure of the size of a function, regarded as a vector. For example
µZ 1 ¶ 12
2
||f || = f dx (1.42)
0
is a generalization of
à 3
! 12
X
||f || = fi2 dx (1.43)
1
for Euclidean space.
To be more precise we require the norm of a function to satisfy the following axioms:
1. ||f || ≥ 0
µZ b ¶ 21
2
||f ||2 = f (x)dx Euclidean norm (1.48)
a
LECTURE 1. VECTOR AND FUNCTION SPACES 9
Z b
||f ||1 = |f (x)|dx (1.49)
a
These are special cases of
µZ b ¶ p1
p
||f ||p = |f (x)| dx for p ≥ 1 (1.50)
a
Note that, for example, we can regard ||f ||2 as the limit of
à n
!1
1X 2 2
||f || = fi dx (1.51)
n
0
Thus the ∞ and 1 norms satisfy the triangle inequality. They also already satisfy axioms (1), (2)
and (4).
1
||f + g||2 = (f + g, f + g) 2
1
≤ ((f, f ) + 2|(f, g)| + (g, g)) 2
³ 1 1
´1 (1.69)
2
≤ (f, f ) + 2(f, f ) 2 (g, g) 2 + (g, g)
= ||f || + ||g||
using the Schwartz inequality.
Proof:
implies
ci = 0 for i = 0, 1, ..., n (1.73)
Orthogonal functions are independent since then from the Pythagorean theorem
¯¯ n ¯¯2
¯¯X ¯¯ n
X
¯¯ ¯¯
¯¯ ci φi ¯¯ = c2i ||φi ||2 (1.74)
¯¯ ¯¯
i=0 i=0
Then if f is continuous
lim En (f ) = 0 (1.76)
n→∞
That is, a continuous function can be arbitrarily well approximated in a closed interval by a
polynomial of sufficiently high order.
The proof is by construction of the required pn (x) (Isaacson and Keller, p183).
Antony Jameson
2 Approximation Theory 2
2.1 Least Squares Approximation and Projection . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Alternative Proof of Least Squares Solution . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Bessel’s Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.4 Orthonormal Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.5 Construction of Orthogonal Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.6 Orthogonal Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.7 Zeros of Orthogonal Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.8 Legendre Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.9 Chebyshev Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.10 Best Approximation Polynomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.11 Best Interpolating Polynomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.12 Jacobi Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.13 Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.14 Error Estimate for Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.15 Orthogonality Relations for Discrete Fourier Series . . . . . . . . . . . . . . . . . . . 11
2.16 Trigonometric Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.17 Error Estimate for Trigonometric Interpolation . . . . . . . . . . . . . . . . . . . . . 13
2.18 Fourier Cosine Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.19 Cosine Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.20 Chebyshev Expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.21 Chebyshev Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.22 Portraits of Fourier, Legendre and Chebyshev . . . . . . . . . . . . . . . . . . . . . . 17
1
Lecture 2
Approximation Theory
in some norm. Ideally we might use the infinity norm, but it turns out it is computationally very
expensive to do this, requiring an iterative process. The best known method for finding the best
approximation in the infinity norm is the exchange algorithm (M.J.D.Powell) On the other hand
it is very easy to calculate the best approximation in the Euclidean norm as follows. Choose cj to
minimize the least squares integral
2
Z Xn
J(c0 , c1 , c2 , ...cn ) = cj φj − f dx (2.2)
j=0
Where
n
X
c∗j φj − f, φi = (f ∗ − f, φi ) = 0, for i = 0, 1, 2, ..., n (2.5)
j=0
or
n
X
aij c∗j = bi (2.6)
j=0
2
LECTURE 2. APPROXIMATION THEORY 3
where
aij = (φi , φj ), bi = (f, φi ) (2.7)
Notice that since f∗ is a linear combination of the φj these equations state that
(f − f ∗ , f ∗ ) = 0 (2.8)
Thus the least squares approximation f ∗ is orthogonal to the error f − f ∗ . This means that f ∗ is
the projection of f onto the space spanned by the basis functions φj , in the same way that a three
dimensional vector might be projected onto a plane.
If the φj form an orthonormal set,
aij = δij (2.9)
where δij = 1 if i = j and 0 if i 6= j, and we have directly
c∗i = (f, φi ) (2.10)
This is the case for orthogonal polynomials because of the Weiesstrass theorem.
LECTURE 2. APPROXIMATION THEORY 4
We require
(fi , fj ) = δij (2.29)
This gives
1
d0 = p (2.30)
(g0 , g0 )
Then 0 = (f0 , f1 ) = d1 ((f0 , g1 ) − c01 ) and in general
where for k = 0 we define Q0 (x) = 1. Then Qk (x) changes sign at the same points as Pn (x), and
Z b
Pn (x)Qk (x)w(x)dx 6= 0 (2.39)
a
But this is impossible since Pn (x) is orthogonal to all polynomials of degree < n.
L0 (x) = 1 (2.40)
1 dn £ ¤
Ln (x) = (x2 − 1) n
, for n = 1, 2, ... (2.41)
2n n! dxn
LECTURE 2. APPROXIMATION THEORY 6
L0 = 1
L1 (x) = x
1
L2 (x) = (3x2 − 1)
2
1 (2.42)
L3 (x) = (5x3 − 3x)
2
1
L4 (x) = (35x4 − 30x2 + 3)
8
1
L5 (x) = (63x5 − 70x3 + 15x)
8
where they can be successively generated for n ≥ 2 by the recurrence formula
2n + 1 n
Ln+1 = xLn (x) − Ln−1 (x) (2.43)
n+1 n+1
They are alternately odd and even and are normalized so that
we have
T0 (x) = 1 (2.53)
T1 (x) = x (2.54)
..
.
Tn+1 (x) = 2xTn (x) − Tn−1 (x) (2.55)
kπ
x0k = cos for k = 0, ..., n (2.58)
n
Property 4 (Continuous orthogonality):
They are orthogonal in the inner product
Z 1
dx
(f, g) = f (x)g(x) √ (2.59)
−1 1 − x2
since
Z π 0 if r 6= s
π
(Tr , Ts ) = cos rθ cos sθ dθ = 2 if r = s 6= 0 (2.60)
0
π if r = s = 0
Property 5 (Discrete orthogonality):
They are orthogonal in the discrete inner product
n
X
(f, g) = f (xj )g(xj ) (2.61)
j=0
2j + 1 π
arccos xj = θj = , j = 0, 1, ..., n (2.62)
2n + 1 2
LECTURE 2. APPROXIMATION THEORY 8
Then
n
X 0 if r 6= s
n+1
(Tr , Ts ) = cos rθj cos sθj = 2 if r = s 6= 0 (2.63)
j=0 n + 1 if r = s = 0
Property
¡ n−1 ¢ 6 (Minimax):
1/2 Tn (x) has smallest maximum norm in [−1, 1] of all polynomials with leading coefficient
unity.
Proof:
Suppose ||pn (x)||∞ were smaller. Now at the n + 1 extrema x0 , ..., xn of Tn (x), which all have the
same magnitude of unity,
1
pn (xn ) < Tn (xn ) (2.64)
2n−1
1
pn (xn−1 ) > Tn (xn−1 ) (2.65)
2n−1
..
.
¡ ¢
Thus pn (x) − 1/2n−1 Tn (x) changes sign n times in [−1, 1], but this is impossible since it is of
degree n − 1 and has n − 1 roots.
has same sign as f (x) − Pn (x). Thus it has opposite sign at n + 2 points, giving n + 1 sign changes.
But this is impossible since it has only n roots.
f n+1 (ξ)
f (x) − Pn (x) = wn (x) (2.67)
(n + 1)!
where
wn (x) = (x − x0 )(x − x1 ) ... (x − xn ) (2.68)
Thus we can minimize ||wn (x)||∞ in [−1, 1] by making xk the zeros of Tn+1 (x). Then
µ ¶n
1
wn (x) = Tn+1 (x) (2.69)
2
µ ¶n
1
||wn (x)||∞ = (2.70)
2
LECTURE 2. APPROXIMATION THEORY 9
Thus the family of Jacobi polynomials includes both the Legendre polynomials
Let us minimize ½Z ¾1
2π ¡ ¢2 2
||f − Sn ||2 = f (θ) − Sn (θ) dθ (2.75)
0
Let
J(a0 , a1 , ..., an , b1 , ..., bn ) = ||f − Sn ||2 (2.76)
The trigonometric functions satisfy the orthogonality relations
Z 2π ½
0 if r 6= s
cos(rθ) cos(sθ)dθ = (2.77)
0 π if r = s 6= 0
Z 2π ½
0 if r =
6 s
sin(rθ) sin(sθ)dθ = (2.78)
0 π if r = s 6= 0
Z 2π
sin(rθ) cos(sθ)dθ = 0 (2.79)
0
At the minimum allowing for these
Z 2π Z 2π
∂J
0= = −2 f (θ) cos(rθ)dθ + 2ar cos2 (rθ)dθ (2.80)
∂ar 0 0
whence Z 2π
1
ar = f (θ) cos(rθ)dθ (2.81)
π 0
and similarly Z 2π
1
br = f (θ) sin(rθ)dθ (2.82)
π 0
LECTURE 2. APPROXIMATION THEORY 10
and since the right hand side is independent of n the sum converges and
Also
lim ||f − Sn ||2 = 0 (2.85)
n→∞
If not, it would violate Weierstrass’s theorem translated to trigonometric functions (Isaacson and
Keller, p 230, p 198).
2M
|ar | ≤ (2.89)
rK
and similarly
2M
|br | ≤ (2.90)
rK
LECTURE 2. APPROXIMATION THEORY 11
Also ¯ m ¯
¯ X ¯
¯ ¯
|Sm (θ) − Sn (θ)| ≤ ¯ (ar cos rθ + br sin rθ)¯
¯ ¯
r=n+1
Xm
4M
≤
rK (2.91)
r=n+1
Z ∞
dξ
≤ 4M
n ξK
4M 1
=
K − 1 nK−1
where the integral contains the sum of the rectangles shown in the sketch.
Thus the partial sums Sn (θ) form a Cauchy sequence and converge uniformly to f (θ), and
4M 1
|f (θ) − Sn (θ)| ≤ K−1
(2.92)
K −1n
where
π
ω = ei N , ω 2N = 1 (2.94)
Then ½
1 − ω 2N r 0 if ω r 6= 1
S= = (2.95)
1 − ωr 2N if ω r = 1
LECTURE 2. APPROXIMATION THEORY 12
Therefore
2N
X −1 ½
irθj −isθj 0 if |r − s| 6= 0, 2N, 4N
e e = (2.96)
2N if |r − s| = 0, 2N, 4N
j=0
at the 2n + 1 points
j
θj = 2π (2.101)
2n
It is easy to verify that
2n−1
X 0 if r 6= s
cos rθj cos sθj = n if r = s 6= 0, n (2.102)
j=0 2n if r = s = 0, n
2n−1
X 0 if r 6= s
sin rθj sin sθj = n if r = s 6= 0, n (2.103)
j=0 0 if r = s = 0, n
LECTURE 2. APPROXIMATION THEORY 13
2n−1
X
sin rθj cos sθj = 0 (2.104)
j=0
It follows on multiplying through by cos rθj or sin rθj and summing over θj that
2n−1
1 X
Ar = f (θj ) cos rθj (2.105)
n
j=0
2n−1
1 X
Br = f (θj ) sin rθj (2.106)
n
j=0
for r = 1, ..., n. Note that writing
1
c0 = A0 (2.107)
2
1
cr = (Ar − iBr ) (2.108)
2
1
c−r = (Ar + iBr ) (2.109)
2
the sum can be expressed as
n−1
X n
X
Un (θ) = cr eirθ = ’c eirθ (2.110)
r
r=−n r=−n
where
2n−1
1 X
cr = f (θj )e−irθj (2.111)
2n
j=0
P
and ’ is defined as
n
X n
X
’a = 1
r ar − (an + a−n ) (2.112)
r=−n r=−n
2
Here
X ½
i(r−k)θj 2n if r − k = 2qn where q = 0, 1, ...
e = (2.116)
0 if r − k 6= 2qn
Hence
∞
X
Cr = cr + (c2kn+r + c2kn−r ) (2.117)
k=1
and similarly
∞
X
C−r = c−r + (c2kn+r + c2kn−r ) (2.118)
k=1
Accordingly
∞
X
Ar = Cr + C−r = ar + (a2kn+r + a2knr−r ) (2.119)
k=1
and similarly
∞
X
Br = i(Cr − C−r ) = br + (b2kn+r − b2kn−r ) (2.120)
k=1
Thus the difference between the interpolation and Fourier coefficients can be seen to be an aliasing
error in which all the higher harmonics are lumped into the base modes.
Now we can estimate |Ar − ar |, |Br − br | using
2M 2M
|ar | ≤ K
, |br | ≤ K . (2.121)
r r
We get
∞
X¯ ¯
¯ 1 1 ¯
|Ar − ar | ≤ ¯
2M ¯ + ¯
(2kn − r)K (2kn + r)K ¯
k=1
¯ ¯ (2.122)
2M X 1 ¯¯ ¯
∞
1 1 ¯
= ¯ ¡ ¢ + ¡ ¢ ¯
(2n)K kK ¯ 1 − r K 1 + r K¯
k=1 2kn 2kn
Since r ≤ n this is bounded by
∞ µ Z ∞ ¶ µ ¶
2M X 1 k 2(2K + 1)M dξ 2(2K + 1)M 1
(2 + 1) ≤ 1+ ≤ 1+ (2.123)
(2n)K kK 2K nK 1 ξK 2K nK K −1
k=1
where Z π
1
a0 = f (θ)dθ (2.129)
π 0
Z π
2
ar = f (θ) cos(rθ)dθ, r > 0 (2.130)
π 0
d
Hence since dθ cos(rθ) = 0 at θ = 0, π the function will not be well approximated at 0, π unless
0 0
f (θ) = f (π) = 0. The cosine series in fact implies an even continuation at θ = 0, π.
Now if all derivatives of f (θ) of order 0, 1, ..., K −1, are continuous, and f (p) (0) = f (p) (π) = 0 for
all odd p < K and f (K) (θ) is integrable, then integration by parts from the coefficients of f (K) (θ)
gives
M
|ar | ≤ k (2.131)
r
and let
Un (θj ) = f (θj ) for j = 0, 1, ...n (2.133)
where
2j + 1 π
θj = (2.134)
n+1 2
Then
n
X 0 if 0 ≤ r =
6 s≤n
n+1
cos rθj cos sθj = 2 if 0 ≤ r = s ≤ n (2.135)
j=0 n + 1 if 0 = r = s
Now if we multiply the first equation by cos sθj and sum, we find that
n
2 X
ar = f (θj ) cos rθj for 0 < r < n (2.136)
n+1
j=0
n
1 X
a0 = f (θj ) (2.137)
n+1
j=0
LECTURE 2. APPROXIMATION THEORY 16
approximate f (x) in the interval [−1, 1]. Then G(θ) = g(cos θ) is the Fourier cosine series for
F (θ) = f (cos θ) for 0 ≤ θ ≤ π, since
Tr (cos θ) = cos rθ (2.139)
∞
X
G(θ) = g(cos θ) = ar cos rθ (2.140)
r=0
Thus Z Z
π 1
1 1 dx
a0 = f (cos θ)dθ = f (x) √ (2.141)
π 0 π −1 1 − x2
and Z Z
π 1
2 2 dx
ar = f (cos θ) cos rθdθ = f (x)Tr (x) √ (2.142)
π 0 π −1 1 − x2
Now from the theory of Fourier cosine series if f (p) (x) is continuous for |x| ≤ 1 and p = 0, 1, ..., K −1,
and f (K) (x) is integrable
M
|ar | ≤ K (2.143)
r
¡ 1 ¢
Since |Tr (x)| ≤ 1 for |x| ≤ 1 the remainder after n terms is O nK−1 . Now
when p is odd, since then F (p) contains the term sin θ. As a result, the favorable error estimate
applies, provided that derivatives of f exist up to the required order.
where µ ¶
2j + 1 π
xj = cos θj = cos (2.150)
n+1 2
These are the zeros of Tn+1 (x). Now for any k < n we can find the discrete least squares ap-
proximation. When k = n, because of the above equation, the error is zero, so the least squares
approximation is the interpolation polynomial. Moreover, since Gn (x) is a polynomial of degree n,
this is the Lagrange interpolation polynomial at the zeros of Tn+1 (x).