Differentiation with a Computer

• The derivative of a function, $f(x)$ at $x=a$, is defined as
  $$\frac{df}{dx} \vert_{x=a}=lim_{\Delta x \rightarrow 0 } \frac{ f(a+\Delta x)-f(a)}{\Delta x}$$

• This is called a forward-difference approximation.
• The limit could be approached from the opposite direction, giving a backward-difference approximation.
• Forward-difference approximation. A computer can calculate an approximation to the derivative, if a very small value is used for $\Delta x$.
  $$\frac{df}{dx}\vert_{x=a}=\frac{f(a+\Delta x)-f(a)}{\Delta x}$$

• Recalculating with smaller and smaller values of $x$ starting from an initial value.
• What happens if the value is not small enough?

• We should expect to find an optimal value for $x$.
• Because round-off errors in the numerator will become great as $x$ approaches zero.
• When we try this for
  $$f(x)=e^xsin(x)$$
  at $x = 1.9$. The analytical answer is 4.1653826.
• Starting with $\Delta x=0.05$ and halving $\Delta x$ each time. Table 7.1 gives the results.
• We find that the errors of the approximation decrease as $\Delta x$ is reduced until about $\Delta x = 0.05/128$.

Table 7.1: Forward-difference approximations for $f(x)=e^x sin(x)$.

• Notice that each successive error is about 1/2 of the previous error as $\Delta x$ is halved until $\Delta x$ gets quite small, at which time round off affects the ratio.
• At values for $\Delta x$ smaller than $0.05/128$, the error of the approximation increases due to round off.
• In effect, the best value for $\Delta x$ is when the effects of round-off and truncation errors are balanced.
• If a backward-difference approximation is used; similar results are obtained.
• Backward-difference approximation.
  $$\frac{df}{dx}\vert_{x=a}=\frac{f(a)-f(a-\Delta x)}{\Delta x}$$

With MATLAB. Analytical answer to the function of Table 7.1.

With MATLAB. Numerical answer to the function of Table 7.1.

• It is not by chance that the errors are about halved each time.
• Look at this Taylor series where we have used $h$ for $\Delta x$:
  $$f(x+h)=f(x)+f'(x)*h +f''(\xi)*h^2/2$$

• Where the last term is the error term. The value of $\xi$ is at some point between $x$ and $x + h$.
• If we solve this equation for $f'(x)$, we get

  $$f'(x)=\frac{f(x+h)-f(x)}{h}-f''(\xi)*\frac{h}{2}$$ (7.1)

  
  

• Which shows that the errors should be about proportional to $h$, precisely what Table 7.1 shows.
• If we repeat this but begin with the Taylor series for $f(x - h)$, it turns out that

  $$f'(x)=\frac{f(x)-f(x-h)}{h}+f''(\zeta)*\frac{h}{2}$$ (7.2)

  
  

• Where $\zeta$ is between $x$ and $x - h$.

• The two error terms of Eqs. 7.1 and 7.2 are not identical though both are $O(h)$.
• If we add Eqs. 7.1 and 7.2, then divide by 2, we get the central-difference approximation to the derivative:

  $$f'(x)=\frac{f(x+h)-f(x-h)}{2h}-f'''(\xi)*\frac{h^2}{6}$$ (7.3)

  
  

• We had to extend the two Taylor series by an additional term to get the error because the $f''(x)$ terms cancel.
• This shows that using a central-difference approximation is a much preferred way to estimate the derivative.
• Even though we use the same number of computations of the function at each step,
• we approach the answer much more rapidly.

With MATLAB,

Table 7.2 illustrates this, showing that errors decrease about four fold when $\Delta x$ is halved (as Eq. 7.3 predicts) and that a more accurate value is obtained.

Table 7.2: Central-difference approximations for $f(x)=e^x sin(x)$.