Suppose $f\in C([a,b])$ is twice continuously differentiable and $f''(x)>0$ on the interval. Show that the best linear approximation $p$ to $f$ has the slope $p'(x)=(f(b)-f(a))/(b-a)$.

To my understanding, due to the second derivative of $f$ being greater than 0 on the entire interval, then we know for sure $f$ itself has to be concave up on the entire interval. This and the fact that $f$ is concave up tells us that the maximum of $f$ must be in the interval, more specifically, at one of the endpoints. Also, there is a minimum somewhere else in the interval.

Now, by Taylor's theorem, we know there is a Taylor polynomial $p(x)=f(c)+f'(c)(x-c)$ at $c\in [a,b]$. Due to the linearity of $f$, we know it is a polynomial of degree $1$ and so $f$ must resemble $f(x)=mx+b$ where $m,b\in\mathbb{R}$.

But now I am stuck. Looking at $p'(x)=(f(b)-f(a))/(b-a)$, the right side seems to resemble the definition of the derivative of $f$ and intuitively, it makes sense to me that $p$ approximating $f$ will have this matching slope but I cannot seem to prove it. How to proceed?

  • 2
    $\begingroup$ What does best linear approximation mean precisely? $\endgroup$ – user99914 Apr 19 '18 at 1:34
  • $\begingroup$ The best approximation $p$ should satisfy the equioscillation condition, so that would mean I need to show that the best approximation is the one that minimizes the sup norm. $\endgroup$ – Samad Ahmed Apr 19 '18 at 1:59
  • $\begingroup$ This answer of mine might be relevant: math.stackexchange.com/questions/1784262/… $\endgroup$ – marty cohen Apr 19 '18 at 4:32

First we consider the case $f(a) = f(b)=0$. Since $f''>0$, we have $f(x) <0$ and $f(x)$ attains a minimum at some $c\in (a, b)$.

Then it is easy to check that $L(x) = \frac 12 f(c)$ (i.e. constant function) is the best linear approximation with $\|f - L\|_\infty = \frac 12 |f(c)|$ (you need only that $f(a) = f(b) = 0$ and $f(c)\le f(x)\le 0$ to show this).

In the general case, consider

$$ g(x) = f(x) - \frac{f(b)-f(a)}{b-a}(x-a) + f(a).$$

Then $g'' = f'' >0$, $g(a) = g(b) = 0$. Thus $g$ is best approximated by constant function, and thus $f$ is best approximated by linear functions with slope $$\frac{f(b) - f(a)}{b-a}.$$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.