# Problems With 'Central Differences'

Numerical Methods for Derivative Estimation

In the previous section, we explained the motivation of the
*Central Differences* method for estimating derivative values
from data, concentrating on the case of evaluating derivative
values for the locations where the function values are tabulated,
with particular emphasis on the special case
that derivative estimates are calculated at points where
the function value is tabulated.

For example, we found that a third order estimator formula for the derivative, as constructed using Central Difference methods, is the following.

y'_{k}≈ 1/60·(y_{k+3}- y_{k-3}) - 3/20·(y_{k+2}- y_{k-2}) + 3/4·(y_{k+1}- y_{k-1})

We will now put this formula to the test.

## Testing the central differences estimator

We will apply the selected derivative estimator formula to a data set generated by a known function, so we can (cheat and) analytically verify the accuracy of the calculated derivative approximations. Here is a plot of the arbitrary function values, with a another plot below it showing the derivative values estimated using the Central Differences estimator (blue), compared to the theoretical ideal values (green).

What is going on here? Well, the "data set" is not really mathematically perfect; it contains on the order of 1% random roundoff and measurement noise. You can see that the derivative estimator values are distributed randomly around the correct values, but the noise is blown all out of proportion. The key lesson here:

** Derivative estimators produced by central
difference methods are grossly sensitive to small data errors such
as measurement errors, truncation errors, and random noise.**

What we need to do is find a formula that has accuracy comparable to the central differences formulas, but without the ugly noise sensitivity. And to do that, we need to distinguish the desirable behaviors from the noise response. The next section will explore that topic.

Footnotes: