For any two functions $f_1 : [0,1] →\mathbb R$ and $f_2 : [0,1] →\mathbb R$, define the function $g : [0,1] →\mathbb R$ as $g(x) = \max(f_1(x),f_2(x))$ for all $x ∈ [0,1]$.

A. If $f_1$ and $f_2$ are linear, then $g$ is linear

B. If $f_1$ and $f_2$ are differentiable, then g is differentiable

C. If $f_1$ and $f_2$ are convex, then g is convex

D. None of the above

My attempt:- $g(x) = \max(f_1(x),f_2(x))=\frac{f_1(x)+f_2(x)}{2}+\frac{|f_1(x)-f_2(x)|}{2}$

A. $g(0)=\max\{f_1(0),f_2(0)\}=0$

$g(cx+y)=\max\{f_1(cx+y),f_2(cx+y)\}=\frac{f_1(cx+y)+f_2(cx+y)}{2}+\frac{|f_1(cx+y)-f_2(cx+y)|}{2}=\frac{cf_1(x)+f_1(y)+cf_2(x)+f_2(y)}{2}+\frac{|cf_1(x)+f_1(y)-(cf_2(x)+f_2(y))|}{2}=c\frac{f_1(x)+f_2(x)}{2}+\frac{f_1(y)+f_2(y)}{2}+\frac{|c(f_1(x)-f_2(x))+(f_1(y)-f_2(y))|}{2}$. $c(f_1(x)-f_2(x))$ and $(f_1(x)-f_2(x))$ lie in the same line. so, Equality holds in the triangular ineqality.
$c\frac{f_1(x)+f_2(x)}{2}+\frac{f_1(y)+f_2(y)}{2}+\frac{|c(f_1(x)-f_2(x))|}{2}+\frac{|(f_1(y)-f_2(y))|}{2}\implies$ $c\frac{f_1(x)+f_2(x)}{2}+\frac{f_1(y)+f_2(y)}{2}+|c|\frac{|(f_1(x)-f_2(x))|}{2}+\frac{|(f_1(y)-f_2(y))|}{2}$ It need not be linear. Since,If $c<0$, $|c|=-c$.

B. $\lim_{h\to 0}\frac{g(h)-g(0)}{h}=\lim_{h\to 0}\frac{f_1(h)+f_2(h)}{2h}+\frac{|f_1(h)-f_2(h)|}{2h}-(\frac{f_1(0)+f_2(0)}{2h}+\frac{|f_1(0)-f_2(0)|}{2h})=\lim_{h\to 0}\frac{f_1(h)-f_1(0)+f_2(h)-f_2(0)}{2h}+\frac{|f_1(h)-f_2(h)|}{2h}-\frac{|f_1(0)-f_2(0)|}{2h}$

How do I proceed further?

C. Let $x,y \in[0,1],g(tx+(1-t)y)=\max(f_1(tx+(1-t)y),f_2(tx+(1-t)y))=\frac{f_1(tx+(1-t)y)+f_2(tx+(1-t)y)}{2}+\frac{|f_1(tx+(1-t)y)-f_2(tx+(1-t)y)|}{2}\leq \frac{f_1(tx+(1-t)y)+f_2(tx+(1-t)y)}{2}+\frac{|f_1(tx+(1-t)y)|}{2}+\frac{f_2(tx+(1-t)y)|}{2}$

How do I proceed further?

I couldn't find any counterexample for $B$ and $C$. So, I tried to prove it. I am not able to complete the proof. Plese help me.

  • 1
    $\begingroup$ For B. you can take $f_{1}(x)=x$ and $f_{2}(x)=-x$, then $max(f_{1},f_{2}) = |x|$, which is not differentiable. $\endgroup$ – Gabriel Fernandes May 30 '18 at 18:19
  • $\begingroup$ Thank you @GabrielFernandes $\endgroup$ – Unknown x May 30 '18 at 18:20
  • 1
    $\begingroup$ @GabrielFernandes Except the domain is given as $[0,1]$ so you would have to take something more like $f_1(x) = x - \frac{1}{2}$ and $f_2(x) = \frac{1}{2} - x$. $\endgroup$ – Daniel Schepler May 30 '18 at 18:21
  • $\begingroup$ @ManeeshNarayanan sorry I've just seen that you are restricted to $[0,1]$. $\endgroup$ – Gabriel Fernandes May 30 '18 at 18:22
  • $\begingroup$ Maybe A. should be stated as: if $f_{1}:\mathbb{R} \rightarrow \mathbb{R}$ and $f_{2}:\mathbb{R} \rightarrow \mathbb{R}$ are both linear then there is $g:\mathbb{R} \rightarrow \mathbb{R}$ linear such that $g|[0,1] = max(f_{1}|[0,1],f_{2}|[0,1])$. (In this case A is true) $\endgroup$ – Gabriel Fernandes May 30 '18 at 18:57

For a counterexample to B, consider two linear functions, one sloping upward and one sloping downward, crossing in the middle, so that the max is V-shaped and thus has a point of nondifferentiability (this would also be a counterexample to A).

For C, since convexity is defined in terms if inequalities, it's probably just easier to use the $\max$ definition directly than to use your decomposition. $$tg(x)+(1-t)g(y) \geq tf_1(x)+(1-t)f_1(y) \geq f_1(tx+(1-t)y)$$ and $$tg(x)+(1-t)g(y) \geq tf_2(x)+(1-t)f_2(y) \geq f_2(tx+(1-t)y)$$ thus $$tg(x)+(1-t)g(y) \geq \max\{f_1(tx+(1-t)y),f_2(tx+(1-t)y)\} = g(tx+(1-t)y)$$


B. is false. Just take $f(x)=x$ and $g(x)=1-x$.

C. is true. It follows from the definition of convex function.

  • $\begingroup$ C $\frac{f_1(tx+(1-t)y)+f_2(tx+(1-t)y)}{2}+\frac{|f_1(tx+(1-t)y)-f_2(tx+(1-t)y)|}{2}\leq \frac{f_1(tx+(1-t)y)+f_2(tx+(1-t)y)}{2}+\frac{|f_1(tx+(1-t)y)|}{2}+\frac{f_2(tx+(1-t)y)|}{2}$ for this part,$ \frac{|f_j(tx+(1-t)y)|}{2},j=1,2$, $\endgroup$ – Unknown x May 31 '18 at 1:32
  • $\begingroup$ $\frac{|f_j(tx+(1-t)y)|}{2} \leq\frac{|tf_j(x)+(1-t)f_j(y)|}{2}$ since $f_j$ is convex. Then using the triangle inequality it follows. Sir, Am I correct? $\endgroup$ – Unknown x May 31 '18 at 1:34
  • $\begingroup$ @ManeeshNarayanan This is not correct. Your last inequality assumes that $|f_j|$ is convex. But it doesn't have to be. $\endgroup$ – José Carlos Santos May 31 '18 at 7:31
  • $\begingroup$ can you help me? how do I proceed further? $\endgroup$ – Unknown x May 31 '18 at 10:29
  • $\begingroup$ @ManeeshNarayanan You already have this answer. $\endgroup$ – José Carlos Santos May 31 '18 at 10:38

A. is not true. $f_1(x) = x$ and $f_2(x) = 1-x$.

B. is not true, see above, not differentiable at $x =\frac{1}{2}$.

C. is true though. Plug in definition of a convex function.

  • $\begingroup$ At first, I thought as you with respect to A. However, is $f_2$ linear? $\endgroup$ – José Carlos Santos May 30 '18 at 18:27
  • $\begingroup$ @JoséCarlosSantos ...yes? It's a linear function, even if it's not linear as an operator. $\endgroup$ – Y. Forman May 30 '18 at 18:28
  • $\begingroup$ How do you define “linear function”? $\endgroup$ – José Carlos Santos May 30 '18 at 18:29
  • $\begingroup$ A necessary (but not sufficient) condition for $f: [0,1] \mapsto \mathbb{R}$ to be linear is $f(\frac{1}{2}) = \frac{1}{2}[f(0) + f(1)]$ $\endgroup$ – Mike May 30 '18 at 18:33
  • 1
    $\begingroup$ Since $[0,1]$ isn't a vector space, "linear transformation" probably wouldn't even be applicable as a predicate on functions $[0, 1] \to \mathbb{R}$. $\endgroup$ – Daniel Schepler May 30 '18 at 18:51

Regarding $C.$

If $f_1(x) \ge f_2(x)$ and $f_1(y) \ge f_2(y)$

$\lambda g(x) + (1-\lambda)g(y) = \lambda f_1(x) + (1-\lambda)f_1(y) \ge f_2(x\lambda + (1-\lambda)y)$


$\lambda g(x) + (1-\lambda)g(y) \ge \lambda f_2(x) + (1-\lambda)f_2(y) \ge f_2(x\lambda + (1-\lambda)y)$

If $f_1(x) \ge f_2(x)$ and $f_2(y) \ge f_1(y)$

$\lambda g(x) + (1-\lambda)g(y) \ge \lambda f_1(x) + (1-\lambda)f_1(y)$ and $\lambda g(x) + (1-\lambda)g(y) \ge \lambda f_2(x) + (1-\lambda)f_2(y)$

But $g(x\lambda + (1-\lambda)y)$ equals either $f_1(x\lambda + (1-\lambda)y)$ or $f_2(x\lambda + (1-\lambda)y)$

$g(x\lambda + (1-\lambda)y) \ge g(x\lambda + (1-\lambda)y)$


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.