Chapter 9 — Interpolation — Lagrange, Hermite, Splines
Reading between the lines of a table: finite differences and the shift operator, the interpolating polynomial in Lagrange and Newton form with its error term, divided differences, Hermite interpolation with derivative data, cubic splines — and Runge's phenomenon, which shows why more equally spaced nodes can make a polynomial interpolant worse.
Before computers, every scientist owned books of tables — logarithms, sines, Bessel functions, star positions — and the first skill taught with a table was how to read between its lines. Henry Briggs built his logarithm tables in the 1620s by adding up differences; Newton worked out his interpolation formulas in the 1670s for exactly this job; Lagrange's formula of 1795 (found earlier by Waring) writes the interpolating polynomial down at once; and Hermite extended the idea to data that prescribe slopes as well as values. The tables are gone, but the problem is everywhere: a sensor is calibrated at a handful of temperatures and must be read at all the others, a weather model knows the pressure on a grid and needs it at an aircraft's position, a font stores a letter as a few points and must draw a smooth outline through them.
Drawing a smooth curve through given points is also where interpolation meets its most surprising failure. In 1901 Runge showed that interpolating the harmless function $1/(1+25x^2)$ at more and more equally spaced points on $[-1,1]$ gives polynomials that oscillate ever more wildly near the ends. There are two cures. One is to choose the nodes cleverly — the Chebyshev nodes, crowded towards the ends, make the maximum of the nodal polynomial in the error term as small as possible. The other is to stop using a single polynomial: shipwrights and draughtsmen had long bent thin flexible strips, called splines, through fixed points, and in 1946 Schoenberg gave the mathematical version — piecewise cubics joined with continuous first and second derivatives, found from a tridiagonal linear system. This chapter builds all of this from one theorem, the existence, uniqueness and error formula of the interpolating polynomial; and Chapters 10 and 11 then differentiate and integrate interpolating polynomials to obtain quadrature rules and methods for differential equations.
- use the operators $\Delta,\nabla,\delta,\mu,E$ and the relations between them, compute differences of polynomials ($\Delta^nx^n=n!\,h^n$, $\Delta^{n+1}p=0$), and locate a wrong entry in a table from its difference columns;
- prove that the interpolating polynomial exists and is unique, write it in Lagrange form, and bound its error with $\dfrac{f^{(n+1)}(\xi)}{(n+1)!}\,\omega_{n+1}(x)$ — including the estimate $\tfrac18M_2h^2$ for linear interpolation;
- build divided-difference tables, use their symmetry and their link with derivatives and forward differences, and interpolate with Newton's divided-difference, forward and backward formulas, choosing the right formula for the data;
- construct the Hermite interpolant from values and slopes, state its degree $2n+1$ and its error term, and recognise when interpolation with derivative data has no solution or many;
- set up and solve the moment equations of natural and clamped cubic splines, test whether a given piecewise cubic is a spline, and state the minimum-curvature property;
- explain Runge's phenomenon and why Chebyshev nodes or splines cure it, with the bound $\max|f-p_n|\le\dfrac{M_{n+1}}{2^n(n+1)!}$ on $[-1,1]$.
How this is examined
Interpolation is a reliable source of short Part B questions in the numerical part of CSIR NET Unit 3. Expect to be asked for a forward or backward difference of a polynomial or of a tabulated function, or for an operator identity between $\Delta$, $\nabla$, $\delta$ and $E$; for the value at a new point of the polynomial that interpolates given data — often found fastest by the uniqueness argument rather than by a formula; for a divided difference or an entry of a divided-difference table; for a Lagrange basis polynomial or an interpolation error bound; for the degree or the error term of a Hermite interpolant; and for the constants that make a given piecewise cubic a (natural) cubic spline. Part C, where more than one option may be correct, tests statements: whether divided differences are symmetric in their arguments, whether the interpolating polynomials converge as the number of nodes grows, whether Chebyshev nodes always give convergence, whether a natural spline reproduces cubics, whether interpolation with derivative data always has a unique solution. The wrong options are nearly always one of the counterexamples in the ⚠️ box of §9.1. GATE MA asks very similar questions, often with numerical answers.
Sections in this chapter
From the CSIR NET Unit 1 course you need the algebra of polynomials (a nonzero polynomial of degree $\le n$ has at most $n$ roots), the binomial theorem, Rolle's theorem applied repeatedly and Taylor's theorem with the Lagrange remainder — the error term of every interpolation formula in this chapter is proved with them. From this course, the linear systems of §8.1: the moments of a cubic spline solve a tridiagonal, strictly diagonally dominant system, which is why the spline exists and is unique, and why it can be computed in $O(n)$ operations. Inverse interpolation gives one more way of estimating a root, to be compared with the iterations of §7.1. If this is your first visit, the welcome page explains how each section is organised.