You are on page 1of 29

The analysis of Runge Phenomenon

Bao Yan, Tianjian Xing

Dept. of Mathematics, SEU

THR. JUL-28-2011

(a) (b)

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 1 / 23
Outline

1 Introduction to Runge phenomenon

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 2 / 23
Outline

1 Introduction to Runge phenomenon

2 The analysis of Runge phenomenon


Theorem
The analysis
Nodes distributing analysis
Stability

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 2 / 23
Outline

1 Introduction to Runge phenomenon

2 The analysis of Runge phenomenon


Theorem
The analysis
Nodes distributing analysis
Stability

3 Conclusion
The reason of Runge phenomenon
The disadvantage of high order interpolation
How to avoid

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 2 / 23
Runge phenomenon
What is Runge phenomenon?
for example:
1
f (x) =
1 + 25x 2

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 3 / 23
interpolating polynomial to an underlying continuous function as the number of equally
spaced points (and hence the polynomial degree) increases. This phenomenon is illustrated
Runge phenomenon
by Runge’s function, shown graphically in Fig. 7.5, where we see that the interpolating
polynomials of increasing degree converge nicely to the function in the middle of the interval,
but diverge near the endpoints.
..... ....
.. ..
2.0 .. ...
.. .. 2 .. ...
... .. f (t) = 1/(1 + 25t )
..................................... ... ..
... .... ... ....
.
.. .. .. ..
.. .. .... .... .... .... .... p (t) .. ..
.. .. 5 .. ..
... ... ... ...
1.5 .. ...
..
.
..
............... p (t) 10
. .
.. ..
.. ..
... .. ..
.. ... ..
. .. . .
. .. ..
. .
. .. .. .
. . . .
. .. .
. .
.
1.0 .
. ..
..
. .
..
....
.
.
..
.. ............
...... . .
......
.
.
.
..
.
.
.
. .. ..
.... . .
..... . .
. .. ... . .
..... .. .
. . ....... .. .
..... .
. .
. .. ..... . .. .
. .. .
...... ....... . .
.
. .
. .
.
.
..
.. ..... .. .... .... .... .... .... .... .... ......
. ... .
.
..... .. ........... .
0.5 .
. ..
.. ...
.
... . .
.
.... ........
. .. .
......... ....
. . . .
.
.
.
.
.
. .. .. .... ........... ........ ... .... .
. .
.. ... .
. ........ . ..
............ .... ... .
. . .. . ..
..
.. .. . .. .
. .. . ...... ... ........................ ... .......... ...............
. .. . .
. .
. . .. .. . .
. ..
. ...................................... ..... ... ............... .. ......
..
...
.. .
....................... .
.
. .
. .. .......................................... .
.......................................... .
. .
..
... .
. .... . .
.. .
...
...
..
.. .... ......
......
...
...
..........................................
.
.
.
0.0 ... ....
.... .... .... ....
.
...
... ...
..
.
.... .... .... . .... ....
...
.
−1.0 ..
.. ...
.. −0.5 0.0 0.5 ..
.. ...
.
1.0
... ... ... ...
.... ....
Figure 7.5: Interpolation of Runge’s function at equally spaced points.
Figure: Interpolation of Runge’s function at equally spaced nodes

7.2.7 Placement of Interpolation Points


As we have just seen, equally spaced interpolation points may give unsatisfactory results,
especially near the ends of the interval.
2011 Summer School (CSE)
If, instead of being equally spaced,
The analysis of Runge Phenomenon
the points4 /are
THR. JUL-28-2011 23
Interpolation error analysis

Theorem
set f(x) in [a,b] include n+1 mutual different nodes:x0 , x1 , x2 , . . . , xn The
intervalxn at [a, b] has N order continuous derivative,and in (a,b) exists
within n+1 order derivative. Then for
any x ∈ [a, b],we have

f (n+1) (ξ)
Rn (x) = f (x) − pn (x) = Wn+1 (x)
(n + 1)!

where

ξ ∈ (a, b) Wn+1 (x) = (x − x0 )(x − x1 ) · · · (x − xn )

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 5 / 23
The analysis of Runge phenomenon

The property of the high order derivative(The high order derivative


of the function must rise very fast)
I If max |f (n+1) (x)| ≤ Mn+1 ,we have
a≤x≤b

Mn+1
|Rn (x)| ≤ |Wn+1 (x)|
(n + 1)!

I If max |f (n) (x)| ≤ M for all n>0. Then


a≤x≤b

M
max |Rn (x)| ≤ (b − a)n+1 −→ 0( as n → ∞ )
a≤x≤b (n + 1)!

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 6 / 23
The analysis of Runge phenomenon

For the function.


1
f (x) =
1 + 25x 2
under the calculation of mathematics. we have

max f (11) (x) = 1.77219 × 1015


−1≤x≤1

The value of high order derivative increased rapidly as the order goes
higher

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 7 / 23
The analysis of Runge phenomenon

So far, we know that the rapidly increased value of high order


derivative has significant influence on runge phenomenon .
Carefully observed,we can find that runge phenomenon has another
feature.
→ The nodes are equally spaced.

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 8 / 23
interpolating polynomial to an underlying continuous function as the number of equally
spaced points (and hence the polynomial degree) increases. This phenomenon is illustrated
Runge phenomenon
by Runge’s function, shown graphically in Fig. 7.5, where we see that the interpolating
polynomials of increasing degree converge nicely to the function in the middle of the interval,
but diverge near the endpoints.
..... ....
.. ..
2.0 .. ...
.. .. 2 .. ...
... .. f (t) = 1/(1 + 25t )
..................................... ... ..
... .... ... ....
.
.. .. .. ..
.. .. .... .... .... .... .... p (t) .. ..
.. .. 5 .. ..
... ... ... ...
1.5 .. ...
..
.
..
............... p (t) 10
. .
.. ..
.. ..
... .. ..
.. ... ..
. .. . .
. .. ..
. .
. .. .. .
. . . .
. .. .
. .
.
1.0 .
. ..
..
. .
..
....
.
.
..
.. ............
...... . .
......
.
.
.
..
.
.
.
. .. ..
.... . .
..... . .
. .. ... . .
..... .. .
. . ....... .. .
..... .
. .
. .. ..... . .. .
. .. .
...... ....... . .
.
. .
. .
.
.
..
.. ..... .. .... .... .... .... .... .... .... ......
. ... .
.
..... .. ........... .
0.5 .
. ..
.. ...
.
... . .
.
.... ........
. .. .
......... ....
. . . .
.
.
.
.
.
. .. .. .... ........... ........ ... .... .
. .
.. ... .
. ........ . ..
............ .... ... .
. . .. . ..
..
.. .. . .. .
. .. . ...... ... ........................ ... .......... ...............
. .. . .
. .
. . .. .. . .
. ..
. ...................................... ..... ... ............... .. ......
..
...
.. .
....................... .
.
. .
. .. .......................................... .
.......................................... .
. .
..
... .
. .... . .
.. .
...
...
..
.. .... ......
......
...
...
..........................................
.
.
.
0.0 ... ....
.... .... .... ....
.
...
... ...
..
.
.... .... .... . .... ....
...
.
−1.0 ..
.. ...
.. −0.5 0.0 0.5 ..
.. ...
.
1.0
... ... ... ...
.... ....
Figure 7.5: Interpolation of Runge’s function at equally spaced points.
Figure: Interpolation of Runge’s function at equally spaced nodes

7.2.7 Placement of Interpolation Points


As we have just seen, equally spaced interpolation points may give unsatisfactory results,
especially near the ends of the interval.
2011 Summer School (CSE)
If, instead of being equally spaced,
The analysis of Runge Phenomenon
the points9 /are
THR. JUL-28-2011 23
The analysis of Runge phenomenon

Is that an important reason? What if we take the unequally spaced


nodes?
There is an example of Chebyshev nodes

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 10 / 23
on, as illustrated for Runge’s function in Fig. 7.7. Of course, one may have no
cing the interpolation points, either because of existing measured data or bec
For the carefully selected nodes:
ular distribution (such as equally spaced) is required for other reasons.

2.0 ..................................... f (t) = 1/(1 + 25t2 )


.... .... .... .... .... p5 (t)
1.5 ............... p10 (t)

1.0 ................
.........
........ .......
...... ... ..
. ....... ... ..
... ..
.. ... ... ..
... ..... ... ..
... ..
.. ... .... .
. ...
. .......
.
......
. ......
0.5 .
. .....
.......... .... .... .... .... .... .... .... .... .... .... .............
.
..
... .
.
..
..
. ............
.... .... ............ ....... .... ....
.......... .... ..
.... ....... ... ......... .. ....
...
. .... ............................. ... .................. .... ..........
. .. ...... ........... .....
...
.... ... ..... ............. .. ... ..
. . ... . ..
.
..........................
...
.
...
.
..
. ... ..
...
.............................. ......... .... ........................................................................ .......
........................................ .......... .............................
.. ...
0.0 .
−1.0 −0.5 0.0 0.5 1.0
Figure 7.7: Interpolation of Runge’s function at the Chebyshev points.
Figure: Interpolation of Runge’s function at the Chebyshev points.

Piecewise
2011 Summer SchoolPolynomial
(CSE) Interpolation
The analysis of Runge Phenomenon THR. JUL-28-2011 11 / 23
Nodes distributing analysis

We denote by Pn g,the polynomial of order n that agrees with a given g


at the given n sides of the sequence {xi },we assume that the
interpolation site sequence lies in some interval [a,b],and we measure
the size of a continuous function f on [a,b] by

kf k = max |f (x)|
a≤x≤b

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 12 / 23
Nodes distributing analysis

Further,we recall from the Lagrange Polynomials


Y x − xj
li (x) =
xj − xj
j6=i

In terms of which
X X X
|(Pn g)(x)| = | g(xi )li | ≤ |g(xi )||li (x)| ≤ (max(g(xi )) |li (x)|)
i
i i i

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 13 / 23
Nodes distributing analysis

We introduce the so-called lebesgue functiom


n
X
λn (x) = |li (x)|
i=1

since max |g(xi )| ≤ kgk,we obtain the estimate


i

kPn gk ≤ kλn kkgk

Actually,there exist functions g(other than zero function) for which

kpn gk = kλn kkgk

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 14 / 23
Nodes distributing analysis

On the other hand, for uniformly spaced interpolation sites,we have

2n
kλn k ∼
en ln n
(as first proved by A.N.Turetskii,1940)

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 15 / 23
Nodes distributing analysis

Fortunately,this situation can be be improved if we carefully select the


nodes.
There is a proof for Chebyshev nodes

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 16 / 23
Chebyshev sites are good

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 17 / 23
Chebyshev sites are good

Computed by M.J.D.Powell, with λcn the corresponding Lebesgue


function ,we have
2
kλcn k ∼ ln n
Π

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 18 / 23
Stability

Let us consider a set of function values {f̃ (xi )} which is a perturbation


of the data f (xi ) relative to the nodes xi ,within i = 0, . . . , n, in an
interval [a,b]. The perturbation may be due,for instance,to the effect of
rounding errors,or may be cause by an error in the experimental
measure of the data.
Denoting by Πn f̃ the interpolating polynomial on the set of values
f̃ (xi ),we have

n
X
kΠn f − Πn f̃ k∞ = max | (f (xi ) − f̃ (xj ))lj (x)| (1)
a≤x≤b
j=0

≤ Λn (X ) max |f (xi ) − f̃ (xi )| (2)


i=0,...,n

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 19 / 23
As previous noticed,Λn (X ) grow as n→ ∞ and in particular, in the case
of Lagrange interpolation on equally spaced nodes, it can be proved
that(see[Nat65])

2n+1
Λn (X ) '
en log n
For the carefully selected nodes, it can be proved that

2
Λn (X ) ∼ log(n + 1), n = 0, 1 . . .
Π

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 20 / 23
The reason of Runge phenomenon

The high order derivative increased so fast


The nodes are equally spaced

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 21 / 23
The disadvantage of high order interpolation

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 22 / 23
The disadvantage of high order interpolation

The Nature of Non uniform convergence

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 22 / 23
The disadvantage of high order interpolation

The Nature of Non uniform convergence


Computational Complexity

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 22 / 23
The disadvantage of high order interpolation

The Nature of Non uniform convergence


Computational Complexity
Too much Swing and Oscillation
For these reasons,in many cases,high order interpolation is not
practical

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 22 / 23
How to avoid

Spline interpolation

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 23 / 23
How to avoid

Spline interpolation
Unequally spaced interpolation

2011 Summer School (CSE) The analysis of Runge Phenomenon THR. JUL-28-2011 23 / 23

You might also like