Natural cubic splines

Arne Morten Kvarving

Department of Mathematical Sciences
Norwegian University of Science and Technology

October 21 2008

Motivation

We are given a large dataset, i.e. a function sampled in

many points.
We want to find an approximation in-between these points.
Until now we have seen one way to do this, namely high order

interpolation - we express the solution over the whole domain

as one polynomial of degree N for N + 1 data points.

a = t0

t1

t2

t3

tn1

b = tn

Motivation

Let us consider the function

f (x) =

1
.
1 + x2

Known as Runges example.

While what we illustrate with this function is valid in general,

this particular function is constructed to really amplify the

problem.

Motivation
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
5

Figure: Runges example plotted on a grid with 100 equidistantly spaced

grid points.

Motivation
8
exact
7

interpolated

6
5
4
3
2
1
0
1
5

Figure: Runges example interpolated using a 15th order polynomial

based on equidistant sample points.

Motivation

It turns out that high order interpolation using a global

polynomial often exhibit these oscillations hence it is

dangerous to use (in particular on equidistant grids).
Another strategy is to use piecewise interpolation. For

instance, piecewise linear interpolation.

s0(x)

y1

s1 (x)
sn1 (x)

y0

x0

x1

x2

xn1 xn

Motivation
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
5

Figure: Runges example interpolated using piecewise linear interpolation.

We have used 7 points to interpolate the function in order to ensure that
we can actually see the discontinuities on the plot.

A better strategy - spline interpolation

We would like to avoid the Runge phenomenon for large

datasets we cannot do higher order interpolation.

The solution to this is using piecewise polynomial

interpolation.
However piecewise linear is not a good choice as the regularity

of the solution is only C 0 .

These desires lead to splines and spline interpolation.
s2(x)

s1(x)

s0(x)

sn1(x)

a = t0

t1

t2

t3

tn1

b = tn

Splines - definition

A function S(x) is a spline of degree k on [a, b] if

S C k1 [a, b].
a = t0 < t1 < < tn = b and

t0 x t1
S0 (x),

S1 (x),
t1 x t2
S(x) =
.
..

Sn1 (x), tn1 x tn

where Si (x) Pk .

Cubic spline

a0 x 3 + b0 x 2 + c0 x + d0 ,
t0 x t1

S0 (x) =
..
S(x) =
.

S (x) = a x 3 + b x 2 + c x + d , t
n1
n1
n1
n1
n1
n1 x tn .
which satisfies

Si1 (xi ) = Si (xi )

0
(xi ) = Si0 (xi )
S(x) C 2 [t0 , tn ] : Si1
00
Si1
(xi )

Si00 (xi )

, i = 1, 2, , n 1.

Cubic spline - interpolation

Given (xi , yi )ni=0 . Task: Find S(x) such that it is a cubic spline
interpolant.
The requirement that it is to be a cubic spline gives us

3(n 1) equations.
In addition we require that

S(xi ) = yi ,

i = 0, , n

which gives n + 1 equations.

This means we have 4n 2 equations in total.
We have 4n degrees of freedom (ai , bi , ci , di )n1
i=0 .
Thus we have 2 degrees of freedom left.

Cubic spline - interpolation

We can use these to define different subtypes of cubic splines:

S 00 (t0 ) = S 00 (tn ) = 0 - natural cubic spline.
S 0 (t0 ), S 0 (tn ) given - clamped cubic spline.

)
S0000 (t1 ) = S1000 (t1 )
00
000
(tn1 ) = Sn1
(tn1 )
Sn2

- Not a knot condition (MATLAB)

Natural cubic splines

Task: Find S(x) such that it is a natural cubic spline.

Let ti = xi , i = 0, , n.
Let zi = S 00 (xi ), i = 0, , n. This means the condition that

it is a natural cubic spline is simply expressed as z0 = zn = 0.

Now, since S(x) is a third order polynomial we know that

S 00 (x) is a linear spline which interpolates (ti , zi ).

Hence one strategy is to first construct the linear spline

interpolant S 00 (x), and then integrate that twice to obtain

S(x).

Natural cubic splines

The linear spline is simply expressed as

Si00 (x) = zi

x ti
x ti+1
+ zi+1
.
ti ti+1
ti+1 ti

We introduce hi = ti+1 ti , i = 0, , n which leads to

S 00 (x) = zi+1

x ti
ti+1 x
+ zi
.
hi
hi

We now integrate twice

zi+1
zi
(x ti )3 +
(ti+1 x)3
6hi
6hi
+ Ci (x ti ) + Di (ti+1 x) .

Si (x) =

Natural cubic splines

Interpolation gives:

Si (ti ) = yi

zi 2
h + Di hi = yi , i = 0, , n.
6 i

Continuity yields:

Si (ti+1 ) = yi+1

zi+1 2
h + Ci hi = yi+1 .
6 i

Natural cubic splines

We insert these expressions to find the following form of the

system
zi+1
zi
(x ti )3 +
(ti+1 x)3
6hi
6hi


yi+1 zi+1
+

hi (x ti )
hi
6


yi
hi
zi (ti+1 x) .
+
hi
6

Si (x) =

We then take the derivative.

Natural cubic splines

The derivative reads

zi+1
zi
(x ti )2
(ti+1 x)2
2hi
2hi
hi
1
+ (yi+1 yi ) (zi+1 zi ) .
h
6
{z
}
|i

Si0 (x) =

bi

In our abscissas this gives

1
hi
1
Si0 (ti ) = zi hi + bi zi+1 + hi zi
2
6
6
zi+1
hi
1
0
Si (ti+1 ) =
hi + bi zi+1 + hi zi
2
6
6
1
1
Si1 (ti ) = zi hi+1 + hi1 zi1 + bi1
3
6
Si0 (ti ) = Si1 (ti )
6 (bi bi1 ) = hi1 z1 + 2 (hi1 + hi ) zi + hi zi+1 .

Natural cubic splines - algorithm

This means that we can find our solution using the following
procedure:
First do some precalculations

hi = ti+1 ti ,
i = 0, , n 1
1
bi = (yi+1 yi ) , i = 0, , n 1
hi
vi = 2 (hi1 + hi ) ,
i = 1, . . . , n 1
ui = 6 (bi bi1 ) ,
z0 = zn = 0

i = 1, , n 1

Natural cubic splines - algorithm

Then solve the tridiagonal system

v1 h1
h1 v2

h2

h2
v3
..
.

h3
..
.
..
.

..
..

.
hn2

z1
z2
z3
..
.

u1
u2
u3
..
.

=

.

hn2 zn2 un2

zn1
un1
vn1

Natural cubic splines - example

Given the dataset

i
xi
yi
hi = xi+1 xi
bi = h1i (yi+1 yi )
vi = 2 (hi1 + hi )
ui = 6 (bi bi1 )

0
0.9
1.3
0.4
0.5

1
1.3
1.5
0.6
0.5833
2.0
0.5

The linear system reads


   
2.0 0.4 z1
0.5
=
0.4 1.6 z2
0.4

2
1.9
1.85
0.2
1.25
1.6
4

3
2.1
2.1

Natural cubic splines - example

We find z0 = 0.5, z1 = 0.125. This gives us our spline

functions
S0 (x) = 0.208 (x 0.9)3 + 3.78 (x 0.9) + 3.25 (1.3 x)
S1 (x) = 0.035 (x 1.3)3 + 0.139 (1.9 x)3 + 0.664 0.62x
S2 (x) = 0.104 (x 1.9)3 + 10.5 (x 1.9) + 9.25 (2.1 x)

Gaussian elimination of tridiagonal systems

Assume we are given a general tridiagonal system

d1
a1

c1
d2
..
.

c2
..
.

..

an1

b1
b2
..
.

.
,

cn1 bn1
bn
dn

First elimination (second row) yields

d1
0

c1
d2
..
.

c2
..
.

..

an1

b1
b2
..
.

,
,

cn1 bn1
bn
dn

a1
d2 = d2 c1
d1
a1

b2 = b2 b1
d1

Gaussian elimination of tridiagonal systems

This means that the elimination stage is

for i =2, , n
m = ai1 /di1
di = di mci1
bi = bi mbi1
end
And the backward substitution reads

xn = bn /dn
for i =n 1, , 1


xi = bi ci xi+1 /di
end
where b1 = b1 .

Gaussian elimination of tridiagonal systems

i =
This will work out fine as long as d
6 0.
Assume that |di | > |ai1 | + |ci | - i.e. diagonal dominance.
For the eliminated system diagonal domiance means that

|di | < |ci |.

We now want to show that diagonal domiance of the original

system implies that the eliminated system is also diagonal

dominant.

Gaussian elimination of tridiagonal systems

i1 | > |ci1 |. This is obviously

We now assume that |d
satisfied for d1 = d1 .
ai1
|ai1 |
|ci1 |
ci1 | |di |

|di1 |
di1
> |ai1 |ci | |ai1 | = |ci |.

|di | = |di

Hence the diagonal domiance is preserved which means that

di 6= 0. The algorithm produces a unique solution.

Why cubic splines?

Now to motivate why we use cubic splines.

First, let us introduce a measure for the smoothness of a

function:
Z
(f ) =

(f 00 (x))2 dx.

(1)

We then have the following theorem

Theorem
Given interpolation data (ti , yi )ni=0 . Among all functions
f C 2 [a, b] which interpolates (ti , yi ), the natural cubic spline is
the smoothest, where smoothness is measured through (1).

Why cubic splines?

We need to prove that

(f ) (S) f C 2 [a, b].

Introduce

g (x) = S(x) f (x),

g (x) C 2 [a, b]
g (ti ) = 0, i = 0, , n.

Inserting this yields

2
S 00 (x) g 00 (x) dx
a
Z b
= (S) + (g ) 2
S 00 (x)g 00 (x) dx

(f ) =

Now since (g ) > 0, we have proved our result if we can show

that
Z
b

S 00 (x)g 00 (x) dx = 0.

Why cubic splines?

We have that

Z
a

b
S (x)g (x) dx = g (x)S (x)a
00

00

00

g 0 (x)S 000 (x) dx

First part on the right hand side is zero since z0 = zn = 0.

Second part we split in an integral over each subdomain
Z

000

g (x)S (x) dx =
a

n1 Z
X
i=0
n1
X
i=0
n1
X
i=0

t i+1

g 0 (x)S 000 (x) dx

ti

Z
6ai

ti+1

g 0 (x) dx

ti
t

6ai g (x)|ti+1
= 0.
i

Cubic spline result

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
5

Figure: Runges example interpolated using cubic spline interpolation

based on 15 equidistant samples.