Sec 4-7

Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

LEAST SQUARES APPROXIMATION

Another approach to approximating a function f (x)


on an interval a ≤ x ≤ b is to seek an approximation
p(x) with a small ‘average error’ over the interval of
approximation. A convenient definition of the average
error of the approximation is given by
" Z b #1
1 2
E(p; f ) ≡ [f (x) − p(x)]2 dx (1)
b−a a
This is also called the root-mean-square-error (de-
noted subsequently by RMSE ) in the approximation
of f (x) by p(x). Note first that choosing p(x) to
minimize E(p; f ) is equivalent to minimizing
Z b
[f (x) − p(x)]2 dx
a
thus dispensing with the square root and multiplying
fraction (although the minimums are generally differ-
ent). The minimizing of (1) is called the least squares
approximation problem.
Example. Let f (x) = ex, let p(x) = α0 + α1x, α0,
α1 unknown. Approximate f (x) over [−1, 1]. Choose
α0, α1 to minimize
Z 1
g(α0, α1) ≡ [ex − α0 − α1x]2 dx (2)
−1
Z 1 ( 2x 2 2 2 x
)
e + α0 + α1x − 2α0e
g(α0, α1) = x dx
−1 −2α1xe + 2α0α1x
Integrating,

g(α0, α1) = c1α20 + c2α21 + c3α0α1 + c4α0 + c5α1 + c6


with constants {c1, . . . , c6}, e.g.
³ ´
c1 = 2, 1
c6 = e − e−1 /2.
g is a quadratic polynomial in the two variables α0,
α1. To find its minimum, solve the system
∂g ∂g
= 0, =0
∂α0 ∂α1
It is simpler to return to (2) to differentiate, obtaining
Z 1
2 [ex − α0 − α1x] (−1) dx = 0
−1
Z 1
2 [ex − α0 − α1x] (−x) dx = 0
−1
This simplifies to
Z 1
2α0 = exdx = e − e−1
−1
Z 1
2
α1 = xexdx = 2e−1
3 −1

e − e−1 .
α0 = = 1.1752
2
.
α1 = 3e−1 = 1.1036
Using these values for α0 and α1, we denote the re-
sulting linear approximation by

1(x) = α0 + α1x
It is called the best linear approximation to ex in the
sense of least squares. For the error,
.
max |ex − 1(x)| = 0.439
−1≤x≤1
Errors in linear approximations of ex:

Approximation Max Error RMSE


Taylor t1(x) 0.718 0.246
Least squares 1(x) 0.439 0.162
Chebyshev c1(x) 0.372 0.184
Minimax m1(x) 0.279 0.190

x
y=e

y=l (x) 1
1

x
-1 1

The linear least squares approximation to ex


THE GENERAL CASE

Approximate f (x) on [a, b], and let n ≥ 0. Seek p(x)


to minimize the RMSE. Write

p(x) = α0 + α1x + · · · + αnxn

Z1 " #2
f (x) − α0 − α1x
g(α0, α1, . . . , αn) ≡ dx
− · · · − αnxn
−1
Find coefficients α0, α1, . . . , αn to minimize this in-
tegral. The integral g(α0, α1, . . . , αn) is a quadratic
polynomial in the n + 1 variables α0, α1, . . . , αn.

To minimize g(α0, α1, . . . , αn), invoke the conditions


∂g
= 0, i = 0, 1, . . . , n
∂αi
This yields a set of n+1 equations that must be satis-
fied by a minimizing set α0, α1, . . . , αn for g. Manip-
ulating this set of conditions leads to a simultaneous
linear system.
To better understand the form of the linear system,
consider the special case of [a, b] = [0, 1]. Differenti-
ating g with respect to each αi, we obtain
Z 1
2 [ex − α0 − · · · − αnxn] (−1) dx = 0
−1
Z 1
2 [ex − α0 − · · · − αnxn] (−x) dx = 0
−1
..
Z 1
2 [ex − α0 − · · · − αnxn] (−xn) dx = 0
−1
Then the linear system is
n
X Z 1
αj
= xif (x) dx, i = 0, 1, . . . , n
j=0 i + j + 1 0

We will study the solution of simultaneous linear sys-


tems in Chapter 6. There we will see that this linear
system is ‘ill-conditioned’ and is difficult to solve ac-
curately, even for moderately sized values of n such as
n = 5. As a consequence, this is not a good approach
to solving for a minimizer of g(α0, α1, . . . , αn).
LEGENDRE POLYNOMIALS

Define the Legendre polynomials as follows.

P0(x) = 1

1 dn h³ 2 ´ni
Pn(x) = n
· n x −1 , n = 1, 2, . . .
n!2 dx
For example,

P1(x) = x
1³ 2 ´
P2(x) = 3x − 1
2
1³ 3 ´
P3(x) = 5x − 3x
2
1³ 4 2
´
P4(x) = 35x − 30x + 3
8
The Legendre polynomials have many special proper-
ties, and they are widely used in numerical analysis
and applied mathematics.
y

x
-1 1

P (x)
1
P (x)
2
P (x)
3
P (x)
4

-1

Legendre polynomials of degrees 1, 2, 3, 4


PROPERTIES

Introduce the special notation


Z b
(f, g) = f (x)g(x) dx
a
for general functions f (x) and g(x).

• Degree and normalization:

deg Pn = n, Pn(1) = 1, n≥0

• Triple recursion relation: For n ≥ 1,


2n + 1 n
Pn+1(x) = xPn(x) − Pn−1(x)
n+1 n+1

• Orthogonality and size:



³ ´ 
0,
 i 6= j
Pi, Pj = 2

 , i=j
2j + 1
• Zeroes:
All zeroes of Pn(x) are located in [−1, 1] ;
all zeroes are simple roots of Pn(x)

• Basis: Every polynomial p(x) of degree ≤ n can


be written in the form
n
X
p(x) = β j Pj (x)
j=0
with the choice of β 0, β 1, . . . , β n uniquely deter-
mined from p(x):
³ ´
p, Pj
βj = ³ ´, j = 0, 1, . . . , n
Pj , Pj
FINDING THE LEAST SQUARES
APPROXIMATION

We solve the least squares approximation problem on


only the interval [−1, 1]. Approximation problems on
other intervals [a, b] can be accomplished using a lin-
ear change of variable.

We seek to find a polynomial p(x) of degree n that


minimizes
Z b
[f (x) − p(x)]2 dx
a
This is equivalent to minimizing

(f − p, f − p) (3)
We begin by writing p(x) in the form
n
X
p(x) = β j Pj (x)
j=0
n
X
p(x) = β j Pj (x)
j=0
Substitute into (3), obtaining
ge (β 0
, β 1, . . . , β n) ≡ (f − p, f − p)

n
X n
X
= f − β j Pj , f − β iPi
j=0 i=0
Expand this into the following:
n (f, P )2
X j
ge = (f, f ) −
j=0 (Pj , Pj )
n ³ " #2
X ´ (f, Pj )
+ Pj , Pj β j −
j=0 (Pj , Pj )
Looking at this carefully, we see that it is smallest
when
(f, Pj )
βj = , j = 0, 1, . . . , n
(Pj , Pj )
Looking at this carefully, we see that it is smallest
when
(f, Pj )
βj = , j = 0, 1, . . . , n
(Pj , Pj )
The minimum for this choice of coefficients is
n (f, P )2
X j
ge = (f, f ) −
j=0 (Pj , Pj )
We call
n (f, P )
X j
n(x) = Pj (x) (4)
j=0 (Pj , Pj )
the least squares approximation of degree n to f (x)
on [−1, 1].

If β n = 0, then its actual degree is less than n.


Example. Approximate f (x) = ex on [−1, 1]. We use
(4) with n = 3:
3
X (f, Pj )
3(x) = β j Pj (x), βj = (5)
j=0 (Pj , Pj )
The coefficients {β 0, β 1, β 2, β 3} are as follows.

j 0 1 2 3
βj 2.35040 0.73576 0.14313 0.02013

Using (5) and simplifying,


2 3
3(x) = .996294 + .997955x + .536722x + .176139x

The error in various cubic approximations:

Approximation Max Error RMSE


Taylor t3(x) .0516 .0145
Least squares 3(x) .0112 .00334
Chebyshev c3(x) .00666 .00384
Minimax m3(x) .00553 .00388
y

0.0112

x
-1 1

-0.00460

Error in the cubic least squares approximation to ex

You might also like