Applications in Engineering
Interpolating polynomials are not very good at approximating discontinuous
functions—the error analysis specifically requires a certain amount
of continuity. Unfortunately, most sampled data is associated with random
noise which is, by its nature, non-differentiable. As a consequence,
using polynomial interpolation will result in significant errors espcially
as the number of points goes up.
Useful background for this topic includes:
- Walter Rudin, Principles of Mathematical Analysis, 3rd Ed.,
McGraw Hill, 1976.
In this topic, we will look at the effects on polynomial interpolation
by various features which may cause significant sources of error.
The resulting lack-of-accuracy is often termed polynomial wiggle.
The Approximation Theorem of Weierstrauss
Werestrauss's approximation theorem states that for any continuous
function defined on a closed interval [a, b],
there exists a sequence of polynomials which converge
uniformly to the function. While this does not guarantee that
there exists interpolating polynomials (the proof uses Bernstein
polynomials which are fixed at the end points), it at least
provides a basis for approximating functions using
interpolating polynomials. We will see, however, that there are
Polynomials are very poor at interpolating discontinuities: any
reasonable error analysis cannot be performed due the lack of continuity.
To demonstrate this, Figures 1 and 2 show the interpolation of the
unit step function with 5 and 7 points, respectively.
Figure 1. Interpolating the unit step function with 5 points.
Figure 2. Interpolating the unit step function with 7 points.
A classic example of a function which cannot be easily interpolated
is the Gaussian function e-x2. The
coefficients of the Taylor series of this function do not drop off as
quickly as they do with other more familiar functions such as the
simple exponential or other trigonometric functions. Consequently,
interpolating polynomials produce less than desirable results unless
the points are very close to each other. Figure 3 shows the Gaussian
function interpolated with 5 and 7 points.
Figure 3. Interpolating the Gaussian function with 5 points.
Figure 4. Interpolating the Gaussian function with 7 points.
Errors in Points
The exponential function is sufficiently well behaved that interpolating
polynomials provide a reasonable fit, as is demonstrated in Figures 5 and
6 which plot the function and its approximation, and the error of the
approximation for a polynomial of degree 7 interpolating the eight
points 0, 1, 2, ..., 7.
Figure 5. Interpolating the exponential function with 8 points.
Figure 6. The difference between the interpolating polynomial and the exponential function on [0, 7].
The points being interpolated are
1, 0.3679, 0.1353, 0.04979, 0.01832, 0.006738, 0.002479, 0.0009119
and the demonstrate the fit, Figure 7 focuses on tail of the plot.
Figure 7. A zoom of the ordinate axis to show the closeness.
1.035, 0.3410, 0.1422, 0.04621, 0.01776, 0.006648, 0.002676, 0.0009092
are the eight above values together with an introduced relative
error of no greater than 10% for each point (the errors were randomly generated
by Maple). The small errors result in a significant change to the
interpolating polynomial as is shown in Figure 8 which shows the polynomial
interpolating the points which have noise associated with them.
Figure 8. A zoom of the exponential function, a polynomial interpolating
the exact values, and the polynomial interpolating noisy values.
Error and Spacing
The last example demonstrated how a small amount of noise can result
in a significant reduction in the accuracy of the interpolating polynomial;
however, with lower degree polynomials this is of less significance.
When interpolating two points, the error is propagated linear, as is
shown in Figure 9 which shows a family of interpolating polynomials
around the points (1, 5±0.1) and (2, 3±0.1).
Figure 9. Interpolating lines between points (1, 5±0.1) and (2, 3±0.1).
If three points are interpolated with a quadratic and the points are
evenly spaced, errors in the y values do not propagate too significantly, as is shown in Figure 10. In this case, the interpolating polynomial is
p(x) = 0.5 x2 − 3.5 x + 8.
Figure 10. Interpolating quadratic polynomials between points (1, 5±0.1), (2, 3±0.1), and (3, 2±0.1).
Figure 11 takes the same polynomial found to fit the points
(1, 5), (2, 3), and (3, 2); evaluates it at 2.5; and then interpolates
the three points (1, 5±0.1), (2.5, 2.375±0.1), and (3, 2±0.1). As can be seen, there is a much greater variation in the possible number
of interpolating polynomials.
Figure 11. Interpolating quadratic polynomials between points (1, 5±0.1), (2, 3±0.1), and (3, 2±0.1).
The effect becomes even more obvious when choosing the points
(1, 5±0.1), (2.9, 2.055±0.1), (3, 2±0.1) as
is shown in Figure 12.
Figure 12. Interpolating quadratic polynomials between points (1, 5±0.1), (2, 3±0.1), and (3, 2±0.1).
This can be seen by looking at the error for interpolating polynomials
(as described in the topic on the Vandermonde Matrix); however, these
images are presented here to emphasize the problems: For data with
noise, the points should not be too close, otherwise the error will
be significantly magnified even for lower-order interpolating polynomials.
The previous section demonstrates that spacing must be taken into
consideration; however, Runge's phenomenon shows that equal spacing is
not appropriate, either. The interpolating polynomials of the
function f(x) = 1/(1 + 25x2) which sample
the function at equally spaced points on the interval [-1, 1]
do not converge to the function
but instead oscillate more-and-more wildly towards the boundaries. The
interpolating polynomials of the aforementioned function are shown
in Figure 13.
See the discussion on error analysis for one method to correct this.
Figure 13. Interpolations of equally spaced points to f(x) = 1/(1 + 25x2) (click for more iterations).
Interpolating oscillating functions can also lead to erroneous
interpolating functions. Figures 14, 15, and 16 demonstrate how
interpolating seven points -2, -1, 0, 1, ..., 4 on the functions
sin(x), sin(2x), and sin(4x) can yeild either
a very good approximation or a very poor approximation.
Figure 14. Interpolating seven points on sin(x).
Figure 15. Interpolating seven points on sin(2x).
Figure 16. Interpolating seven points on sin(4x).
If the data comes from a system which is known to be sinusoidal
with a known frequency ω, it would be better to interpolate
be better to use an interpolating function of the form
f(x) = c1sin(ωx) + c2cos(ωx).
Polynomial interpolation should not be used if the points come
from a function which is discontinuous and oscillating functions
should be sampled at a rate at least four times the frequency
(at least for polynomial interpolating).
Caution should be exercised
when interpolating data with noise: the greater the noise, the lower
the degree of any interpolating polynomial and care should be taken
that the x-values are not too close but that equally spaced x-values
may also cause problems for very high degree interpolating polynomials.
This topic has discussed a number of the problems which are associated
with polynomial interpolation. These are given the rather humorous
name polynomial wiggle which suggests the effects of the
One observation which may be made by looking at a polynomial with
equally spaced roots, as shown in
Figure 17. The error of an interpolating polynomial on nine points
depends on this function and it is clear that the error are more
magnified at the ends and the error is suppressed near the middle.
Figure 17. The error of an interpolating polynomial with 9 equally
One possible solution is to more densely pack the interpolating points
closer to the boundaries. While the theory is beyond the scope of this
text (for now), it can be shown that if the points are chosen according
to the roots of a Chebyshev polynomial T9(x), the
error is reduced uniformly, as is shown in Figure 18.
Figure 18. The error of an interpolating polynomial with 9 points
spaced according to the roots of the Chebyshev polynomial T9(x).
Applications to Engineering
The following Maple commands find the roots of the nth
> ChebyshevT( 5, x );
> expand( % );
16 x - 20 x + 5 x
> fsolve( % );
-0.9510565163, -0.5877852523, 0., 0.5877852523, 0.9510565163
> ChebyshevT( 7, x );
> expand( % );
7 5 3
64 x - 112 x + 56 x - 7 x
> fsolve( % );
-0.9749279122, -0.7818314825, -0.4338837391, 0., 0.4338837391, 0.7818314825, 0.9749279122