← Iris

Interpolation
Points N: 5
Max error (uniform): 0.000
Max error (Chebyshev): 0.000
Max interpolation error vs N
Number of points (N) 5
Runge function
Uniform interpolation
Chebyshev interpolation
Dragged point

In 1901, Carl Runge demonstrated that polynomial interpolation through equally-spaced points on f(x) = 1/(1 + 25x²) does not converge to the function as the number of points increases. Instead, the interpolation error grows exponentially near the endpoints of the interval.

The root cause is that equally-spaced nodes cluster too few points near the edges of [−1, 1], where the polynomial needs the most control. The Lebesgue constant — which measures how much an interpolation can amplify small perturbations — grows exponentially for equally-spaced points (roughly as 2n/n) but only logarithmically for Chebyshev points.

Chebyshev nodes xk = cos((2k−1)π/2n) are the zeros of Chebyshev polynomials of the first kind. They cluster densely near the endpoints, providing exactly the edge control that equally-spaced points lack. For any continuous function on [−1, 1], Chebyshev interpolation converges if the function has bounded variation, and converges exponentially fast if the function is analytic.

This has profound practical implications: if you must interpolate data, use Chebyshev-spaced nodes (or better yet, Chebyshev series). Equally-spaced interpolation is safe only for low-degree polynomials or on short intervals. Splines — piecewise low-degree polynomials — offer another practical escape from Runge's phenomenon.

Try dragging the interpolation points to see how the polynomial responds. Moving a point near the edge has a much larger effect than moving one near the center. This is the Lebesgue function in action.