Global Maximum of $f(x,y)$ on a set $M$.

extreme-value-theoremmaxima-minimamultivariable-calculusreal-analysis

I want to investigate the function $f(x,y) = 4x^2 + 9y – \frac{1}{3}y^3$ on the set $M:= \{(x,y) \in \mathbb{R}^2 : y \geq |x|\}$. In particular, I want to

a) proof the existence of a global maximum on the set $M$,

b) find all points in $M$ where this maximal value is reached.

First of all I did draw a picture of the set $M$. (Since $M$ is not very hard to illustrate I do not post it here.)

I really struggle with a). The only real theorem I know is that every continuous function on a compact set has a Min/Max. Since $M$ is not compact, I don't see how I can proof a). I thought about quadratic expansion, but that did not work.

Considering b), I first took a look at the gradient

\begin{align}
\operatorname{grad} f(x,y) = \begin{pmatrix} 8x \\ 9 – y^2\end{pmatrix}
\end{align}

which is zero for $(0,3)$ and $(0,-3) \notin M$. Using the Hessian Matrix $\mathcal{H}_f$ and evaluating it at $(0,3)$ leads to the fact that $(0,3)$ seems to be a saddle point. Hence, there looks to be no inner point, which might be interesting for the question.

Looking at the boundaries of $M$, as Fra suggested in his answer, I looked at

$$f|_{\partial M}(x,y) = \begin{cases} g(x) := 4x^2 +9x – \frac{x^3}{3}, & \mbox{if } x \geq 0\\
h(x) := 4x^2 -9x + \frac{x^3}{3}, & \mbox{if } x < 0
\end{cases}$$

and found that $g$ has a maximum, namely $\max\{4x^2+9x – \frac{x^3}{3}\} = 162$ for $x = 9$. $h$ as well has a maximum, $\max\{4x^2-9x + \frac{x^3}{3}\} = 162$ for $x = -9$.

So my answer to b) is $\{(9,9), (-9,9)\}$ with value $162$.

Can someone provide a solution/explanation for a)? Furthermore I would be glad if my solution for b) can be verified.

Best Answer

Overview

In full generality, proving things like this can be very hard. But when you have a nice function and a nice region of $\mathbb{R}^{n}$, there's a standard toolbox that I think works well in practice. Basically:

  1. Split up the region if necessary to reduce the dimension to make it more convenient to work with.
  2. Then compactify the region by adding "points at infinity" in an appropriate way (which way depends a bit on the context) or using a trick with arctan to achieve the same result.
  3. If the function can be extended to a continuous function on the compact region, then an extremum must exist, and you can use Multivariable Calculus techniques (e.g. look for critical points, parametrize the boundary and/or use Lagrange multipliers) to find candidates. Then just compare values of all of the candidates.
  4. If the function can't be so extended, or it's too much effort, you may still be able to show that there is a bound on the function over a part of the region, and then cut the bad part(s) out of consideration so that you reduce to the above.

There are many different ways this might look like in practice, so I'll present various ways that come to mind for tackling the problem at hand, as examples. In actuality, almost all of this is overkill for the specific original problem.

Horizontal Strips

For this particular function, the domain is two-dimensional, but it might be easier to manage if we cut it into one-dimensional horizontal strips of the form $\left[-t,t\right]\times\left\{ t\right\} $ for $t\ge0$.

Initial Work

Note that each strip is compact, so there is a maximum on each strip. If we can find the maxima and compare them, we can solve both parts.

On a horizontal strip, the function reduces to $g_{t}(x)=4x^{2}+9t-\frac{1}{3}t^{3}$. Since we know this is a parabola, we can simply note that the maximum is attained at $x=\pm t$ and the value is $g_{t}(t)=4t^{2}+9t-\frac{1}{3}t^{3}$. (In general, we could differentiate with respect to $x$ to obtain $g_{t}'(x)=8x$ so that the points worth checking are the critical point $x=0$ and the endpoints $x=\pm t$.)

Now that we have the maximum on each strip, we must see if and where there is a height $y=t_{0}$ (where $t\in\left[0,\infty\right)$) where this maximum $h(t)=4t^{2}+9t-\frac{1}{3}t^{3}$ is maximized. Unfortunately, $\left[0,\infty\right)$ is not compact, so we can either prove a bound to cut off some of it, or compactify it.

Prove a Bound

Without using any geometric properties of cubics, we could examine the derivative $h'(t)=-t^{2}+8t+9$. The only positive zero is at $t=9$, after which it's negative (since it's continuous and $h'(10)=-11<0$). Therefore, $h(t)$ is decreasing for $t\ge9$. In general, if the behavior of the derivative $h'(t)$ were more complicated, we could then restrict our attention to the compact $[0,9]$. But since we can see that $h'(t)$ is positive on $[0,9]$, we immediately know that a maximum is attained at $t=9$.

Putting everything together, the maximum is $h(9)=\boxed{162\text{ at }(\pm9,9)}$.

Compactification

Note that ${\displaystyle \lim_{t\to\infty}}h(t)=-\infty$ since $-\frac{1}{3}t^{3}$ dominates for large $t$. And $h(0)=0$. If we add in $\pm\infty$ from the extended reals, then we can work properly with the compact interval $[0,\infty]$. The only other candidate for a maximum would be where $h'(t)=0$, at $t=9$. Since $h(9)=162$ is greater than the endpoint values of $0$ and $-\infty$, it must be the maximum.

Alternatively, since $\tan$ is increasing on $\left(-\pi/2,\pi/2\right)$, we can define $\widetilde{h}(t)=\arctan\left(4\tan^{2}t+9\tan t-\frac{1}{3}\tan^{3}t\right)$ on $t\in\left[0,\pi/2\right)$ which has a maximum at $t_{0}$ exactly if $h(t)$ has a maximum at $\tan t_{0}$. But since ${\displaystyle \lim_{x\to\pi/2^{-}}}\tan x=\infty$ and ${\displaystyle \lim_{x\to-\infty}}\arctan x=-\pi/2$, $\widetilde{h}$ can be extended to a continuous function $[0,\pi/2]\to\mathbb{R}$, no $\infty$ required. Then the values of $\widetilde{h}$ at the endpoints are $0$ and $-\pi/2$, both of which are less than the value of $\arctan(162)$ at the zero of $\widetilde{h}'$, $t=\arctan9$.

Either way, again we have that the maximum of the original functions is $\boxed{162\text{ at }(\pm9,9)}$.

Vertical Strips

Another approach would be to cut the region into 1-dimensional vertical strips of the form $\left\{ t\right\} \times\left[\left|t\right|,\infty\right)$ for some real $t$. On each strip, we the function reduces to $g_{t}(y)=4t^{2}+9y-\frac{1}{3}y^{3}$ for $y\ge|t|$.

Compactification

One of these strips can be compactified in much the same way as discussed earlier: either being comfortable with the extended reals, or replacing $g_{t}(y)$ with $\widetilde{g_{t}}(y)=\arctan\left(4t^{2}+9\tan y-\frac{1}{3}\tan^{3}y\right)$ for $y\ge\arctan\left|t\right|$. We have potential extrema at the endpoints, and where $g_{t}'=0$. Note that ${\displaystyle \lim_{y\to\infty}}g_{t}(y)=-\infty$ so that we have continuity (though there won't be a maximum there), and $g_{t}\left(|t|\right)=4t^{2}+9|t|-\frac{1}{3}|t|^{3}$. $g_{t}'(y)=9-y^{2}$ has a $0$ at $y=3$ (where $g_{t}(y)=4t^{2}+18$) if $|t|\le3$ and no zeros if $|t|>3$. Note that $4t^{2}+18\ge4t^{2}+|t|\left(9-\frac{1}{3}t^{2}\right)$ for $\left|t\right|\le3$, so the maximum of $g_{t}(y)$ is attained at $y=\begin{cases} 3 & \text{ if }|t|\le3\\ |t| & \text{ if }|t|\ge3 \end{cases}$. The values attained are $h(t)=\begin{cases} 4t^{2}+18 & \text{ if }|t|\le3\\ 4t^{2}+|t|\left(9-\frac{1}{3}t^{2}\right) & \text{ if }|t|\ge3 \end{cases}$.

Now that we have the maximum on each strip, we must see if and where there is a horizontal position $x=t_{0}$ where this maximum $h(t)$ is maximized. Since $4t^{2}+18$ is increasing as $|t|$ increases, it achieves its maximum on $\left[-3,3\right]$ at $t=\pm3$. Therefore, it suffices to consider $h(t)=4t^{2}+|t|\left(9-\frac{1}{3}t^{2}\right)$ on. Since $h$ is an even function, we really just have to worry about $h(t)=4t^{2}+9t-\frac{1}{3}t^{3}$ on $\left[3,\infty\right)$. Above, we found that this function's maximum on $\left[0,\infty\right)$ is attained at $9$. So using a similar approach (perhaps looking at $h(3)$ instead of $h(0)$), we would reach the same conclusion, and the maximum of the original function is still $\boxed{162\text{ at }(\pm9,9)}$.

Prove a Bound

Instead of compactifying the whole vertical strip, we could have noticed that $g_{t}'(y)<0$ for $y>3$, so that $g_{t}(y)\le g_{t}\left(\max\left(3,\left|t\right|\right)\right)$ and we only need to consider the compact part of the strip where $y\in\left[|t|,\max\left(3,\left|t\right|\right)\right]$. Maximizing $g_{t}$ on this interval and then maximizing the maxima is very similar to the above, now.

Diagonal Strips

Given the shape of the domain, one more approach to reduce the dimension would be to cut it into 1-dimensional diagonal strips of the form $\left\{ \left(s\cos t,s\sin t\right):s\in\left[0,\infty\right)\right\} $ for $t\in\left[\pi/4,3\pi/4\right]$. On each strip, we the function reduces to $g_{t}(s)=4\left(s\cos t\right)^{2}+9\left(s\sin t\right)-\frac{1}{3}\left(s\sin t\right)^{3}$ for $s\ge0$.

As before, we could compactify/notice that $g_{t}(s)$ is eventually very negative since $\sin t>0$ if $t\in\left[\pi/4,3\pi/4\right]$. So to find candidates for a maximum, we only need to examine $s=0$, (where $g_{t}(s)=0$) and where $g_{t}'(s)=0$. $g_{t}'(s)=-s^{2}\sin^{3}t+8s\cos^{2}t+9\sin t$. The only positive zero is at $s_{t}=\dfrac{-8\cos^{2}t-\sqrt{64\cos^{4}t+36\sin^{4}t}}{-2\sin^{3}t}=\dfrac{4\cos^{2}t+\sqrt{16\cos^{4}t+9\sin^{4}t}}{\sin^{3}t}.$ Setting $r=\sqrt{16\cos^{4}t+9\sin^{4}t}$ for convenience, we find that $g_{t}(s_{t})$ is the following nonnegative quantity (and hence this is the maximum): $$h(t)=4\left(\dfrac{4\cos^{2}t+r}{\sin^{3}t}\cos t\right)^{2}+9\left(\dfrac{4\cos^{2}t+r}{\sin^{3}t}\sin t\right)-\frac{1}{3}\left(\dfrac{4\cos^{2}t+r}{\sin^{3}t}\sin t\right)^{3}$$

$$ =\cdots $$ $$ =\dfrac{2}{3\sin^{2}t}\left(4\cos^{2}t+r\right)\left(9+8\dfrac{\cos^{4}t}{\sin^{4}t}+2\dfrac{\cos^{2}t}{\sin^{4}t}r\right) $$

Now, the domain of $t$ is the compact interval $\left[\pi/4,3\pi/4\right]$, so we can optimize $h(t)$ relatively easily. At the endpoints where $t=\pi/2\pm\pi/4$, we have $h(t)=162$ and $s_{t}=9\sqrt{2}$, corresponding to the points $\left(\pm9,9\right)$ in the plane. We also have $h'(t)$ equal to a very complicated expression. But $h''(t)$ is a littler simpler can can be seen to be positive on $\left[\pi/4,3\pi/4\right]$, so that $h'(t)$ has at most one zero. In fact, $h'(\pi/2)=0$, and we have $h(\pi/2)=18<162$. Once again, the maximum is $\boxed{162\text{ at }(\pm9,9)}$.

The Whole Region#

All of the previous methods focused on slicing up the two-dimensional region. But we can deal with both dimensions at once.

Prove a Bound

Note that if $|x|\le y$, then $x^{2}\le y^{2}$, so that $f(x,y)=4x^{2}+9y-\frac{1}{3}y^{3}\le4y^{2}+9y-\frac{1}{3}y^{3}$. Since ${\displaystyle \lim_{y\to\infty}}4y^{2}+9y-\frac{1}{3}y^{3}=-\infty$, it suffices to look at a compact region of the form, say, $\left|x\right|\le y\le6+3\sqrt{7}$ since $6+3\sqrt{7}$ is the largest zero of $4y^{2}+9y-\frac{1}{3}y^{3}$ (beyond which it must be negative), and we know that $f(0,1)=26/3>0$. Therefore, $f$ has an absolute maximum somewhere in that large triangle of height $6+3\sqrt{7}$, either at a critical point, or along the boundary. As noted in the question, $\left(0,3\right)$ is the only critical point in the interior of the triangle. We have $f\left(0,3\right)=18>0$, so no point on the top edge (where $f\left(x,y\right)\le0$) is relevant. It remains to examine the bottom edges for candidates. Since $f$ is symmetric about $x=0$, it suffices to check the right edge $\left\{ (x,x):x\in\left[0,6+3\sqrt{7}\right]\right\} $.

This can be done either by examining $f(x,x)$ directly, or by using Lagrange multipliers to handle $f(x,y)$ with the constraint $y=x$. (With a more complicated curve like a circle $x^{2}+y^{2}=9$, aside from Lagrange Multipliers, you could parametrize with a new parameter as in $f\left(3\cos t,3\sin t\right)$, or break it up into separate functions of $x$, as in $f\left(x,\sqrt{9-x^{2}}\right)$ and $f\left(x,-\sqrt{9-x^{2}}\right)$.) Either way, the boundary point $(0,0)$ of that boundary edge is a candidate for a maximum (the other boundary point was ruled out already).

Looking at $f(x,x)$ directly is similar to work we've done above and involves $\frac{\mathrm{d}}{\mathrm{d}x}f(x,x)=0$. Lagrange Multipliers would have us examine the gradient of $y-x$, and solve $8x=\lambda\left(-1\right)$, $9-y^{2}=\lambda\left(1\right)$, and $y-x=0$, which yields the same equation as $\frac{\mathrm{d}}{\mathrm{d}x}f(x,x)=0$. In any case, we find candidates $(\pm9,9)$.

Taking all the candidates together, we find that the maximum must be one of $f\left(0,3\right)=18$, $f(0,0)=0$, and $f(\pm9,9)=162$. Since $162>18>0$, we have the maximum is $\boxed{162\text{ at }(\pm9,9)}$.

Compactify

Because a bound was so easy to find, it's not worth compactifying in two dimensions, but I include this for completeness and a nice graphical perspective.

Because of the extra dimension, the easiest way to make sure we are adding on acceptable "points at infinity" is to use the arctan trick. Define $\widetilde{f}\left(x,y\right)=\arctan\left(4\tan^{2}x+9\tan y-\frac{1}{3}\tan^{3}y\right)$ for $|x|\le y<\pi/2$, so that $\widetilde{f}$ has a global maximum at $\left(x,y\right)$ if $f$ does at $\left(\tan x,\tan y\right)$.

We can extend the domain of $\widetilde{f}$ to include the top of the triangle at $y=\pi/2$, via $\widetilde{f}\left(x,y\right)=-\pi/2$ if $y=\pi/2$. It turns out that $\widetilde{f}$ is then continuous on the entire compact triangle, for instance via the multivariable squeeze theorem and $$ -\pi/2\le\widetilde{f}(x,y)\le\arctan\left(4\tan^{2}y+9\tan y-\frac{1}{3}\tan^{3}y\right)\underset{y\to\pi/2^{-}}{\to}-\pi/2 $$

Now we can do the same sort of argument as before, noting that at the critical point we have $\widetilde{f}(0,\arctan3)=\arctan(18)>-\pi/2$, so the top edge is not the source of a candidate for the maximum. We can switch back to $f$ or stay with $\widetilde{f}$ when finding candidates, and arrive at the same result as every other time for the original function: $\boxed{162\text{ at }(\pm9,9)}$.

Graphs

One benefit of $\widetilde{f}$ is that it essentially allows us to graph all of $f$ at once.

Here are two views of $\widetilde{f}$.

the lip of the function away from y equals pi over two a sheer cliff near y equals pi over two

To understand how steep that cliff is, here is $\widetilde{f}(t,t)=\arctan\left(4\tan^{2}t+9\tan t-\frac{1}{3}\tan^{3}t\right)$ zoomed in near its zero of $t\approx1.49917$

a decreasing sigmoid from one point two down to negative one point two on the interval one point four nine nine to one point four nine nine three

Related Question