108 Chapter 3.

Interpolation and Extrapolation

f (x, y, z). Multidimensional interpolation is often accomplished by a sequence of

one-dimensional interpolations. We discuss this in §3.6.

CITED REFERENCES AND FURTHER READING:

Abramowitz, M., and Stegun, I.A. 1964, Handbook of Mathematical Functions, Applied Mathe-

http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
readable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website
Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machine-
Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software.
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5)
matics Series, Volume 55 (Washington: National Bureau of Standards; reprinted 1968 by
Dover Publications, New York), §25.2.
Stoer, J., and Bulirsch, R. 1980, Introduction to Numerical Analysis (New York: Springer-Verlag),
Chapter 2.
Acton, F.S. 1970, Numerical Methods That Work; 1990, corrected edition (Washington: Mathe-
matical Association of America), Chapter 3.
Kahaner, D., Moler, C., and Nash, S. 1989, Numerical Methods and Software (Englewood Cliffs,
NJ: Prentice Hall), Chapter 4.
Johnson, L.W., and Riess, R.D. 1982, Numerical Analysis, 2nd ed. (Reading, MA: Addison-
Wesley), Chapter 5.
Ralston, A., and Rabinowitz, P. 1978, A First Course in Numerical Analysis, 2nd ed. (New York:
McGraw-Hill), Chapter 3.
Isaacson, E., and Keller, H.B. 1966, Analysis of Numerical Methods (New York: Wiley), Chapter 6.

3.1 Polynomial Interpolation and Extrapolation

Through any two points there is a unique line. Through any three points, a
unique quadratic. Et cetera. The interpolating polynomial of degree N − 1 through
the N points y1 = f (x1 ), y2 = f (x2 ), . . . , yN = f (xN ) is given explicitly by
Lagrange’s classical formula,

(x − x2 )(x − x3 )...(x − xN ) (x − x1 )(x − x3 )...(x − xN )

P (x) = y1 + y2
(x1 − x2 )(x1 − x3 )...(x1 − xN ) (x2 − x1 )(x2 − x3 )...(x2 − xN )
(x − x1 )(x − x2 )...(x − xN −1 )
+ ··· + yN
(xN − x1 )(xN − x2 )...(xN − xN −1 )
(3.1.1)
There are N terms, each a polynomial of degree N − 1 and each constructed to be
zero at all of the xi except one, at which it is constructed to be y i .
It is not terribly wrong to implement the Lagrange formula straightforwardly,
but it is not terribly right either. The resulting algorithm gives no error estimate, and
it is also somewhat awkward to program. A much better algorithm (for constructing
the same, unique, interpolating polynomial) is Neville’s algorithm, closely related to
and sometimes confused with Aitken’s algorithm, the latter now considered obsolete.
Let P1 be the value at x of the unique polynomial of degree zero (i.e.,
a constant) passing through the point (x 1 , y1 ); so P1 = y1 . Likewise define
P2 , P3 , . . . , PN . Now let P12 be the value at x of the unique polynomial of
degree one passing through both (x 1 , y1 ) and (x2 , y2 ). Likewise P23 , P34 , . . . ,
P(N −1)N . Similarly, for higher-order polynomials, up to P 123...N , which is the value
of the unique interpolating polynomial through all N points, i.e., the desired answer.
 3.1 Polynomial Interpolation and Extrapolation 109

The various P ’s form a “tableau” with “ancestors” on the left leading to a single
“descendant” at the extreme right. For example, with N = 4,

x1 : y1 = P1
P12
x2 : y2 = P2 P123
(3.1.2)

http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
readable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website
Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machine-
Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software.
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5)
P23 P1234
x3 : y3 = P3 P234
P34
x4 : y4 = P4

Neville’s algorithm is a recursive way of filling in the numbers in the tableau

a column at a time, from left to right. It is based on the relationship between a
“daughter” P and its two “parents,”

(x − xi+m )Pi(i+1)...(i+m−1) + (xi − x)P(i+1)(i+2)...(i+m)

Pi(i+1)...(i+m) =
xi − xi+m
(3.1.3)

This recurrence works because the two parents already agree at points x i+1 . . .
xi+m−1 .
An improvement on the recurrence (3.1.3) is to keep track of the small
differences between parents and daughters, namely to define (for m = 1, 2, . . . ,
N − 1),
Cm,i ≡ Pi...(i+m) − Pi...(i+m−1)
Dm,i ≡ Pi...(i+m) − P(i+1)...(i+m) . (3.1.4)

Then one can easily derive from (3.1.3) the relations

(xi+m+1 − x)(Cm,i+1 − Dm,i )

Dm+1,i =
xi − xi+m+1
(3.1.5)
(xi − x)(Cm,i+1 − Dm,i )
Cm+1,i =
xi − xi+m+1

At each level m, the C’s and D’s are the corrections that make the interpolation one
order higher. The final answer P 1...N is equal to the sum of any y i plus a set of C’s
and/or D’s that form a path through the family tree to the rightmost daughter.
Here is a routine for polynomial interpolation or extrapolation from N input
points. Note that the input arrays are assumed to be unit-offset. If you have
zero-offset arrays, remember to subtract 1 (see §1.2):

#include <math.h>
#include "nrutil.h"

void polint(float xa[], float ya[], int n, float x, float *y, float *dy)
Given arrays xa[1..n] and ya[1..n], and given a value x, this routine returns a value y, and
an error estimate dy. If P (x) is the polynomial of degree N − 1 such that P (xai ) = yai , i =
1, . . . , n, then the returned value y = P (x).
{
int i,m,ns=1;
float den,dif,dift,ho,hp,w;
110 Chapter 3. Interpolation and Extrapolation

float *c,*d;

dif=fabs(x-xa[1]);
c=vector(1,n);
d=vector(1,n);
for (i=1;i<=n;i++) { Here we find the index ns of the closest table entry,
if ( (dift=fabs(x-xa[i])) < dif) {
ns=i;

http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
readable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website
Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machine-
Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software.
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5)
dif=dift;
}
c[i]=ya[i]; and initialize the tableau of c’s and d’s.
d[i]=ya[i];
}
*y=ya[ns--]; This is the initial approximation to y.
for (m=1;m<n;m++) { For each column of the tableau,
for (i=1;i<=n-m;i++) { we loop over the current c’s and d’s and update
ho=xa[i]-x; them.
hp=xa[i+m]-x;
w=c[i+1]-d[i];
if ( (den=ho-hp) == 0.0) nrerror("Error in routine polint");
This error can occur only if two input xa’s are (to within roundoff) identical.
den=w/den;
d[i]=hp*den; Here the c’s and d’s are updated.
c[i]=ho*den;
}
*y += (*dy=(2*ns < (n-m) ? c[ns+1] : d[ns--]));
After each column in the tableau is completed, we decide which correction, c or d,
we want to add to our accumulating value of y, i.e., which path to take through the
tableau—forking up or down. We do this in such a way as to take the most “straight
line” route through the tableau to its apex, updating ns accordingly to keep track of
where we are. This route keeps the partial approximations centered (insofar as possible)
on the target x. The last dy added is thus the error indication.
}
free_vector(d,1,n);
free_vector(c,1,n);
}

Quite often you will want to call polint with the dummy arguments xa
and ya replaced by actual arrays with offsets. For example, the construction
polint(&xx[14],&yy[14],4,x,y,dy) performs 4-point interpolation on the tab-
ulated values xx[15..18], yy[15..18]. For more on this, see the end of §3.4.

CITED REFERENCES AND FURTHER READING:

Abramowitz, M., and Stegun, I.A. 1964, Handbook of Mathematical Functions, Applied Mathe-
matics Series, Volume 55 (Washington: National Bureau of Standards; reprinted 1968 by
Dover Publications, New York), §25.2.
Stoer, J., and Bulirsch, R. 1980, Introduction to Numerical Analysis (New York: Springer-Verlag),
§2.1.
Gear, C.W. 1971, Numerical Initial Value Problems in Ordinary Differential Equations (Englewood
Cliffs, NJ: Prentice-Hall), §6.1.
 3.2 Rational Function Interpolation and Extrapolation 111

3.2 Rational Function Interpolation and

Extrapolation

Some functions are not well approximated by polynomials, but are well
approximated by rational functions, that is quotients of polynomials. We de-

http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
readable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website
Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machine-
Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software.
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5)
note by Ri(i+1)...(i+m) a rational function passing through the m + 1 points
(xi , yi ) . . . (xi+m , yi+m ). More explicitly, suppose

Pµ (x) p0 + p1 x + · · · + pµ xµ
Ri(i+1)...(i+m) = = (3.2.1)
Qν (x) q0 + q1 x + · · · + qν xν

Since there are µ + ν + 1 unknown p’s and q’s (q 0 being arbitrary), we must have

m+1 = µ+ν +1 (3.2.2)

In specifying a rational function interpolating function, you must give the desired
order of both the numerator and the denominator.
Rational functions are sometimes superior to polynomials, roughly speaking,
because of their ability to model functions with poles, that is, zeros of the denominator
of equation (3.2.1). These poles might occur for real values of x, if the function
to be interpolated itself has poles. More often, the function f (x) is finite for all
finite real x, but has an analytic continuation with poles in the complex x-plane.
Such poles can themselves ruin a polynomial approximation, even one restricted to
real values of x, just as they can ruin the convergence of an infinite power series
in x. If you draw a circle in the complex plane around your m tabulated points,
then you should not expect polynomial interpolation to be good unless the nearest
pole is rather far outside the circle. A rational function approximation, by contrast,
will stay “good” as long as it has enough powers of x in its denominator to account
for (cancel) any nearby poles.
For the interpolation problem, a rational function is constructed so as to go
through a chosen set of tabulated functional values. However, we should also
mention in passing that rational function approximations can be used in analytic
work. One sometimes constructs a rational function approximation by the criterion
that the rational function of equation (3.2.1) itself have a power series expansion
that agrees with the first m + 1 terms of the power series expansion of the desired
function f (x). This is called P adé approximation, and is discussed in §5.12.

Bulirsch and Stoer found an algorithm of the Neville type which performs
rational function extrapolation on tabulated data. A tableau like that of equation
(3.1.2) is constructed column by column, leading to a result and an error estimate.
The Bulirsch-Stoer algorithm produces the so-called diagonal rational function, with
the degrees of numerator and denominator equal (if m is even) or with the degree
of the denominator larger by one (if m is odd, cf. equation 3.2.2 above). For the
derivation of the algorithm, refer to [1]. The algorithm is summarized by a recurrence