Interpolation, Smoothing, Extrapolation: What Is A Good Estimate For y For X 4.5, or X 15 ?

Download as pdf or txt
Download as pdf or txt
You are on page 1of 22

Interpolation, Smoothing, Extrapolation

A typical numerical application is to find a smooth parametrization


of available data so that results at intermediate (or extended)
positions can be evaluated.

What is a good estimate for


y for x=4.5, or x=15 ?
Options: if have a model, y=f(x),
then fit the data and extract
model parameters. Model then
used to give values at other
points.
If no model available, then use a
smooth function to interpolate

Winter Semester 2006/7 Computational Physics I Lecture 3 1


Interpolation
Start with interpolation. Simplest - linear interpolation. Imagine
we have two values of x, xa and xb, and values of y at these points,
ya, yb. Then we interpolate (estimate the value of y at an
intermediate point) as follows:

(yb  ya )
y = ya + (x  xa ) b
(xb  xa )

Winter Semester 2006/7 Computational Physics I Lecture 3 2


Interpolation
(yb  ya )
Back to the initial plot: y = ya + (x  xa )
(xb  xa )

Not very satisfying. Our intuition


is that functions should be
smooth. Try reproducing with a
higher order polynomial. If we
have n+1 points, then we can
represent the data with a
polynomial of order n.

Winter Semester 2006/7 Computational Physics I Lecture 3 3


Interpolation

Fit with a 10th order polynomial.


We go through every data point
(11 free parameters, 11 data
points). This gives a smooth
representation of the data and
indicates that we are dealing with
an oscillating function. However,
extrapolation is dangerous !

Winter Semester 2006/7 Computational Physics I Lecture 3 4


Lagrange Polynomials

For n+1 points (xi,yi), with

i = 0,1,, n xi  x j i

there is a unique interpolating polynomial of degree n with


p(xi ) = yi i = 0,1,, n
Can construct this polynomial using the Lagrange polynomials,
defined as:
(x  x0 )(x  xi 1 )(x  xi +1 )(x  xn )
Li (x) =
(xi  x0 )(xi  xi 1 )(xi  xi +1 )(xi  xn )
Degree n (denominator is constant), and
Li (xk ) =  i, k

Winter Semester 2006/7 Computational Physics I Lecture 3 5


Lagrange Polynomials

The Lagrange Polynomials can be used to form the interpolating


polynomial:
n n n
x  xk
p(x) =  yi Li (x) =  yi 
i=0 i=0 k = 0, k i xi  x k

*
* Example: 10th order polynomial
Lagrange=0.
*
Do I=0,10
term=1.
Do k=0,10
If (k.ne.I) then
term=term*(x-x0(k))/(x0(I)-x0(k))
Endif
Enddo
Lagrange=Lagrange+Y0(I)*term
Enddo

Winter Semester 2006/7 Computational Physics I Lecture 3 6


Lagrange Polynomials

Error estimation of the interpolation/extrapolation:

Define err(x) = f (x)  p(x) where f (x) is original function,


p(x) is interpolating function
Choose x  xi for any i = 0,1,, n
Now define
n

 (x  x )
i
F(x) = f (x)  p(x)  ( f (x )  p(x )) i=0
n

 (x  x )
i=0
i

Now look at properties of F(x)

Winter Semester 2006/7 Computational Physics I Lecture 3 7


Inter- and Extrapolation Error

F(xi ) = 0 for all i = 0,1,, n and F(x ) = 0.


I.e., F(x) has n + 2 zeroes

Rolle's theorem: There exists a  between x, x0 , x1 ,, xn such that


F (n +1) ( ) = 0
so,
(n + 1)!
0= f (n +1)
( )  ( f (x )  p(x )) n (p (n +1)
=0 )
 (x  x )
i=0
i

but x is arbitrary, so
f (n +1) ( ) n
e(x ) = 
(n + 1)! i = 0
(x  xi ) for some  between x, x0 , x1 ,, xn

Winter Semester 2006/7 Computational Physics I Lecture 3 8


Inter- and Extrapolation Error

Suppose we are trying to interpolate a sine function (our example).


We have 11 data points (n=10). Then
11
(n +1) d sin x
f (x) = 11
=  cos x so f (n +1) ( )  1
dx
n

 (x  x ) i
(b  a)11
e(x )  i=0
< for interpolation
(n + 1)! 11!

For extrapolation, the error grows as the power (n+1)

Winter Semester 2006/7 Computational Physics I Lecture 3 9


Splines

Assume have data { x k , yk } with k = 0, n i.e., n + 1 points


(also known as knots)

Define a = x0 , b = xn and arrange so that x0 < x1 <  < xn 1 < xn

A spline is polynomial interpolation between the data points which


satisfies the following conditions:

1. S(x) = Sk (x) for xk  x  xk +1 k = 0,1,, n  1

2. S(xk ) = yk k = 0,1,, n

3. Sk (xk +1 ) = Sk +1 (xk +1 ) k = 0,1,, n  2 i.e., S(x) is continuous

Winter Semester 2006/7 Computational Physics I Lecture 3 10


Splines

Linear Spline:
yk +1  yk
Sk (x) = yk + (x  xk )
xk +1  xk
Quadratic Spline:

zk +1  zk
Sk (x) = yk + zk (x  xk ) + (x  xk )2
2(xk +1  xk )
z0 has to be fixed, for example from requiring Sk (a) = z0 = 0.
then,
y  yk
zk +1 = zk + 2 k +1
xk +1  xk

Winter Semester 2006/7 Computational Physics I Lecture 3 11


Splines
The cubic spline satisfies the following conditions:

1. S(x) = Sk (x) for xk  x  xk +1 k = 0,1,, n  1


Sk (x) = Sk,0 + Sk,1 (x  xk ) + Sk,2 (x  xk )2 + Sk, 3 (x  xk )3

2. S(xk ) = yk k = 0,1,, n

3. Sk (xk +1 ) = Sk +1 (xk +1 ) k = 0,1,, n  2 i.e., S(x) is continuous

4. Sk (xk +1 ) = Sk +1 (xk +1 ) k = 0,1,, n  2 i.e., S (x) is continuous

5. Sk(xk +1 ) = Sk+1 (xk +1 ) k = 0,1,, n  2 i.e., S (x) is continuous

Winter Semester 2006/7 Computational Physics I Lecture 3 12


Splines

Need at least 3rd order polynomial to satisfy the conditions.


Number of parameters is 4n. Fixing Sk(xk)=yk gives n+1
conditions. Fixing Sk(xk+1)=Sk+1(xk+1) gives an additional n-1
conditions. Matching the first and second derivative gives another
2n-2 conditions, for a total of 4n-2 conditions. Two more
conditions are needed to specify a unique cubic spline which
satisfies the conditions on the previous page:

S (a) = 0 S (b) = 0 Natural cubic spline

Can take other options for the boundary conditions

Winter Semester 2006/7 Computational Physics I Lecture 3 13


Splines
Linear spline

Cubic spline

Quadratic spline

Winter Semester 2006/7 Computational Physics I Lecture 3 14


Cubic Splines

The cubic spline is optimal in the following sense:


1. It is accurate to fourth order, and

5
f(x)-S(x)  max f ( 4 ) (x)  h 4 where h = max k xk +1  xk
384 a  x b

2. It is the minimum curvature function linking the set of data


points.
f (x)
Curvature is defined as  f (x)
(1 + f (x) ) 2 3/2

Any smooth interpolating


b b function must have
Cubic spline satisfies  [ S  (x) ]2
dx   [ f  (x) ]2
dx curvature at least as
a a large as a cubic spline

Winter Semester 2006/7 Computational Physics I Lecture 3 15


Cubic Splines

Proof of 2.
Start with algebraic identity F 2  S 2 = (F  S)2  2S(S  F)
Let F = f (x), S = S (x)
then
b b b b

 [ f  (x) ]2
dx   [ S  (x) ]2
dx =  [ f  (x)  S  (x) ]2
dx  2  S (x)[ S (x)  f (x)]dx
a a a a
b

 [ f (x)  S(x)] dx  0
2

a
b n 1 xk+1

 S(x)[ S(x)  f (x)]dx =   S(x)[ S(x)  f (x)]dx


a k = 0 xk

Now we use integration by parts to solve the integrals

Winter Semester 2006/7 Computational Physics I Lecture 3 16


Cubic Splines
b b
Recall:
 u(x)v(x)dx = u(x)v(x)   u (x)v(x)dx
b
a
a a

b 
n 1 xk+1

a S(x)( S(x)  f (x))dx =   S (x) ( S (x)  f (x)) xk   S (x) ( S (x)  f (x))dx
xk+1

k=0  xk

The first term is


n 1

 S(x)( S(x)  f (x)) x = S (b) ( S (b)  f (b))  S (a) ( S (a)  f (a))


xk+1

k
k=0

= 0 From the boundary conditions for natural cubic spline


S (x) is a constant (since we have a cubic) and can be taken out of the integral, so
b n 1 xk+1

 S(x)( S(x)  f (x))dx =  S  ( S(x)  f (x))dx


a k=0
k
xk
n 1
= Sk( S(x)  f (x)) x
xk+1

k
k=0

but since S(xk ) = f (xk ), = 0

Winter Semester 2006/7 Computational Physics I Lecture 3 17


Cubic Splines

b b

 [ f (x)] dx   [ S(x)] dx  0
2 2

a a

We have proven that a cubic spline has a smaller or equal


curvature than any function which fulfills the interpolation
requirements. This also includes the function we started with.

Physical interpretation: a clamped flexible rod picks the minimum


curvature to minimize energy - spline

Winter Semester 2006/7 Computational Physics I Lecture 3 18


Data Smoothing

If we have a large number of data points, interpolation with


polynomials, splines, etc is very costly in time and multiplies the
number of data. Smoothing (or data fitting) is a way of reducing.
In smoothing, we just want a parametrization which has no model
associated to it. In fitting, we have a model in mind and try to
extract the parameters.

Data fitting is a full semester topic of its own.

A few brief words on smoothing of a data set. The simplest


approach is to find a general function with free parameters which
can be adjusted to give the best representation of the data. The
parameters are optimized by minimizing chi squared:
 2
n
(yi  f (xi ;  ))
 =
2
2
i=0 w i

Winter Semester 2006/7 Computational Physics I Lecture 3 19


Data Smoothing
 2
n
(yi  f (xi ;  ))
 =
2
2
i=0 w i

 are the parameters of the function to be fit
yi are the measured points at values xi
wi is the weight given to point i

In our example, let’s take f (x; A,  ) = A cos(x +  )

And set wi = 1 i

Now we minimize 2 as a function of A and 

Winter Semester 2006/7 Computational Physics I Lecture 3 20


Data Smoothing

Best fit for A=1, =3/2

A cos(x +  ) = 1 cos(x + 3 / 2) = cos x cos 3 / 2  sin x sin 3 / 2 = sin x

Winter Semester 2006/7 Computational Physics I Lecture 3 21


Exercises
1. Calculate the Lagrange Polynomial for the following data:

x y
0 1

0.5 0.368

1. 0.135

2. 0.018

2. For the same data, find the natural cubic spline coefficients.
Plot the data, the lagrange polynomial and the cubic spline
interpolations.
3. Smooth the data in the table with the function f(x)=Aexp(-bx).
What did you get for A,b ?

Winter Semester 2006/7 Computational Physics I Lecture 3 22

You might also like