Approximation By Spline Functions

Def. A function S is called a spline of degree k if

• The domain of S is an interval [a, b].

• S, S 0 , . . . S (k−1) are continuous on [a, b].

• There are points ti (the knots of S) such that a = t0 < t1 < · · · < tn = b and such
that S is a polynomial of degree at most k on each [ti , ti+1 ].

When k = 1, the splines are called linear splines, when k = 2, the splines are called
quadratic splines, when k = 3, the splines are called cubic splines. Here we are mainly
interested in linear splines and cubic splines.
Interpolation by Linear Splines
Problem: Given n + 1 points (t0 , y0 ), (t1 , y1 ),. . . ,(tn , yn ), where without loss of generality
we assume t0 < t1 < · · · < tn . Seek a linear spline S(x) such that S(ti ) = yi for 0 ≤ i ≤ n
and ti are the knots of S(x).
Obviously we can write


 S0 (x), t0 ≤ x ≤ t1
 S1 (x), t1 ≤ x ≤ t2

S(x) = ..


 .
 S (x), t
n−1 n−1 ≤x≤t n

where Si (x) = yi + mi (x − ti ), ti ≤ x ≤ ti+1 , with slope mi = (yi+1 − yi )/(ti+1 − ti ). So

S(x) is a piecewise linear polynomial.
Algorithm for evaluating S(x) (given x, ti , yi and mi , i = 0, 1, . . . , n):

for i = 0 : n − 1
if x − ti+1 ≤ 0,
exit loop
end
end
S ← yi + mi (x − ti )

Remarks:

• When x < t0 , the algorithm gives S = y0 + m0 (x − t0 ); when x > tn , it gives

S = yn−1 + mn−1 (x − tn−1 ).

• A binary search can be used to find the desired interval which consists of x. Averagely,
this is more efficient.

1
Interpolation by Cubic Splines
For a linear spline, generally S 0 is not continuous, so its graph lacks of smoothness. For
a quadratic spline, generally S 00 is not continuous, so the curvature of its graph changes
abruptly at each knot. So in practice, the most frequently used splines are cubic splines.
Problem: Given n + 1 points (t0 , y0 ), (t1 , y1 ),. . . ,(tn , yn ), where without loss of generality
we assume t0 < t1 < · · · < tn . Seek a cubic spline S(x) such that S(ti ) = yi for 0 ≤ i ≤ n
and ti are the knots of S(x).
Obviously we can write


 S0 (x), t0 ≤ x ≤ t1
 S1 (x), t1 ≤ x ≤ t2

S(x) = ..


 .
n−1 ≤ x ≤ tn
 S (x), t
n−1

were Si is a cubic polynomial on [ti , ti+1 ].

Number of unknowns:
Each Si has 4 unknowns. So there are a total of 4n unknowns.
Number of conditions:
(k) (k)
S(ti ) = yi for i = 0, 1, . . . , n result in n + 1 conditions. Si−1 (ti ) = Si (ti ) for k = 0, 1, 2
and i = 1, . . . n − 1 lead to 3(n − 1) conditions. So there are a total of 4n − 2 conditions.
In order to get a unique solution, we need 2 more extra conditions. Here we impose the
following two conditions:
S 00 (t0 ) = S 00 (tn ) = 0.
The resulting spline function is called a natural cubic spline.
Constructing a Natural Cubic Spline
Let zi = S 00 (ti ) (0 ≤ i ≤ n). On [ti , ti+1 ], Si00 (x) is a linear polynomial and Si00 (ti ) = zi and
Si00 (ti+1 ) = zi+1 . So we can write
x − ti+1 x − ti
Si00 (x) = zi + zi+1 .
ti − ti+1 ti+1 − ti

Integrating Si00 (x) twice, we obtain

zi zi+1
Si (x) = (ti+1 − x)3 + (x − ti )3 + cx + d, (1)
6hi 6hi
where hi ≡ ti+1 − ti , and c and d are constants of integration. We impose the conditions
Si (ti ) = yi and Si (ti+1 ) = yi+1 . Then we have
zi
h3i + cti + d = yi ,
6hi
zi+1
h3i + cti+1 + d = yi+1 ,
6hi

2
This gives

c = (yi+1 − yi )/hi − hi (zi+1 − zi )/6

d = (yi ti+1 − yi+1 ti )/hi + hi (ti zi+1 − ti+1 zi )/6.

Now we have to determine the zi and zi+1 in Si (x). In order to do this, we impose conditions
0
Si−1 (ti ) = Si0 (ti ). From eqn (1) we obtain

Si0 (x) = −(ti+1 − x)2 zi /(2hi ) + (x − ti )2 zi+1 /(2hi )

+(yi+1 − yi )/hi − (zi+1 − zi )hi /6

So
1 1
Si0 (ti ) = − hi zi − hi zi+1 + bi .
3 6
Analogously we can derive

0 1 1
Si−1 (ti ) = hi−1 zi−1 + hi−1 zi + bi−1 .
6 3
0
Thus Si−1 (ti ) = Si0 (ti ) leads to

hi−1 zi−1 + 2(hi−1 + hi )zi + hi zi+1 = 6(bi − bi−1 ),

i = 1, 2, . . . , n − 1.

Notice z0 = zn = 0. So we have the following tridiagonal system

    
2(h0 + h1 ) h1 z1 6(b1 − b0 )
 h1 2(h1 + h2 ) h2   z2   6(b2 − b1 ) 
    

 • • •  •  = 
   • 

 • • •   •   • 
hn−2 2(hn−2 + hn−1 ) zn−1 6(bn−1 − bn−2 )

Note that the matrix is strictly diagonally dominant by columns. So this can be reliably
solved by GENP.

3
Algorithm for finding zi , i = 0, . . . , n (given ti , yi , i = 0, . . . , n):

for i = 0 : n − 1
hi ← ti+1 − ti
bi ← (yi+1 − yi )/hi
end
% Forward elimination
u1 ← 2(h0 + h1 )
v1 ← 6(b1 − b0 )
for i = 2 : n − 1
ui ← 2(hi−1 + hi ) − h2i−1 /ui−1
vi ← 6(bi − bi−1 ) − hi−1 vi−1 /ui−1
end
% Back substitution
zn ← 0
for i = n − 1 : −1 : 1
zi ← (vi − hi zi+1 )/ui
end
z0 ← 0

Evaluation of S(x)
 
zi
3 3 zi+1 yi+1 − yi hi
Si (x) = (ti+1 − x) + (x − ti ) + − (zi+1 − zi ) x
6hi 6hi hi 6
yi ti+1 − yi+1 ti hi
+ + (ti zi+1 − ti+1 zi ).
hi 6
This is not the best computational form. As we want to utilize nested multiplication, we
write
Si (x) = Ai + Bi (x − ti ) + Ci (x − ti )2 + Di (x − ti )3 .
000
Notice Ai = Si (ti ), Bi = Si0 (ti ), Ci = 21 Si00 (ti ), Di = 16 Si (ti ). Then we can obtain

Ai = yi
Bi = −hi zi+1 /6 − hi zi /3 + (yi+1 − yi )/hi
Ci = zi /2
Di = (zi+1 − zi )/(6hi )

Si (x) = Ai + (x − ti )(Bi + (x − ti )(ci + (x − ti )Di )).

4
Algorithm for evaluating S(x) (given x, ti , yi and zi for i = 0, 1, . . . , n):

for i = 0 : n − 1
if x − ti+1 ≤ 0
exit loop
end
end
h ← ti+1 − ti
B ← −hzi+1 /6 − hzi /3 + (yi+1 − yi )/h
D ← (zi+1 − zi )/(6h)
S ← yi + (x − ti )(B + (x − ti )(zi /2 + (x − ti )D))

Note. You can use a binary search to find the desired interval consisting of x.

MATLAB Spline Tools: spline