16
$\begingroup$

How can I find $3$ positive numbers that have a sum of $1$ and the sum of their squares is minimum?

So far I have:

$$x+y+z=1 \qquad \implies \qquad z=1-(x+y)$$ So, $$f(x,y)=xyz=xy(1-x-y)$$

But I'm stuck from here. Hints?

$\endgroup$
  • 4
    $\begingroup$ Note: the header and text references the "sum of their squares" but the question itself appears to refer to the product $xyz$. $\endgroup$ – lulu Dec 12 '15 at 13:50
  • $\begingroup$ @lulu that was my attempt, it could very well be incorrect $\endgroup$ – hax0r_n_code Dec 12 '15 at 13:50
  • 4
    $\begingroup$ Got it. In looking at the posted solutions, be sure to check the boundary ($x,y,z≥0$). It is true that in this case Lagrange does identify the minimum, but a priori that is not obvious. $\endgroup$ – lulu Dec 12 '15 at 13:54

10 Answers 10

31
$\begingroup$

$x \mapsto x^2$ is a convex function. By Jensen's inequality,

$$x^2+y^2+z^2 = 3\left(\frac{x^2+y^2+z^2}{3}\right) \ge 3\left(\frac{x+y+z}{3}\right)^2 = 3\left(\frac{1}{3}\right)^2 = \frac13$$

Since the equality $x^2+y^2+z^2 = \frac13$ is achieved at $x = y = z = \frac13$, this is the solution you seek.

If you want a more elementary approach, you can use the fact

$$\begin{align} x^2 + y^2 + z^2 =& \left(x-\frac13+\frac13\right)^2 + \left(y-\frac13+\frac13\right)^2 + \left(z -\frac13 + \frac13\right)^2\\ =& \left(x-\frac13\right)^2 + \left(y-\frac13\right)^2 + \left(z -\frac13\right)^2\\ &+ \frac23\left[\left(x-\frac13\right) + \left(y-\frac13\right) + \left(z -\frac13\right)\right] + 3\left(\frac13\right)^2\\ =& \left(x-\frac13\right)^2 + \left(y-\frac13\right)^2 + \left(z -\frac13\right)^2 + \frac13 \end{align} $$ to arrive at same conclusion.

$\endgroup$
20
$\begingroup$

By Cauchy-Schwarz, for every $x,y,z\in\Bbb R$ we have $$x+y+z=\langle(1,1,1),(x,y,z)\rangle \leq \left\|(1,1,1)\right\|_2\cdot \|(x,y,z)\|_2=\sqrt{3}\cdot \sqrt{x^2+y^2+z^2}$$ Now, if $x+y+z=1$, then squaring both sides and dividing by $3$ gives $$ \frac{1}{3}\leq x^2+y^2+z^2 $$ This lower bound is then attained for $|x|=|y|=|z|=1/3$.

$\endgroup$
  • 3
    $\begingroup$ definitely the most elegant way. $\endgroup$ – Kuifje Dec 12 '15 at 17:38
16
$\begingroup$

By elementary geometry. You are looking for a point on the plane given by $x+y+z=1$ that is closest to the origin, which means that the vector $(x,y,z)$ from the origin to that point is a normal vector to the plane. This happens if $x=y=z$, which gives you three independent linear equations in $x,y,z$ to solve.

$\endgroup$
11
$\begingroup$

Lagrange multipliers? So $f(x,y,z,\lambda) = x^2 + y^2 + z^2 - \lambda(x+y+z-1)$.

So critical points obey: $2x - \lambda = 0, 2y - \lambda = 0, 2z - \lambda = 0, x+y+z-1 = 0$. So $x = y = z = \frac{1}{3}$, which gives sum of squares $\frac{1}{3}$.

$\endgroup$
  • $\begingroup$ Thank you, but I'm not understanding how you got $\frac{1}{3}$? $\endgroup$ – hax0r_n_code Dec 12 '15 at 13:53
  • 2
    $\begingroup$ @free_mind $(\frac{1}{3})^2 + (\frac{1}{3})^2 +(\frac{1}{3})^2 = \frac{3}{9} = \frac{1}{3}$. $\endgroup$ – Henno Brandsma Dec 12 '15 at 13:59
  • 9
    $\begingroup$ This is not a solution. You need to understand what happens in the critical points; what if they were local maxima, or saddle points? $\endgroup$ – Federico Poloni Dec 13 '15 at 7:36
6
$\begingroup$

$x+y+z = 1$ is a plane intersecting the positive quadrant in an equilateral triangle through $(0,0,1), (0,1,0), (1,0,0)$. Minimising the sum of squares of the numbers corresponds to finding the point in the triangle closest to the origin. By symmetry, $x=y=z=\dfrac{1}{3}$.

$\endgroup$
4
$\begingroup$

$$f(x,y)=x^2+y^2+(1-x-y)^2$$ $$\frac{\partial f}{\partial x}=2x+2(1-x-y)(-1)=0\tag 1$$ $$\frac{\partial f}{\partial y}=2y+2(1-x-y)(-1)=0\tag 2$$ by subtracting $$x=y$$ substitute in 1 to get $$x=\frac{1}{3}$$ so $$y=\frac{1}{3}$$ and then $$z=1-x-y=\frac{1}{3}$$ to check the critical point, we can reduce the original function, which has three variables, to one varible as follow $$f(x)=x^2+x^2+(1-x-x)^2=2x^2+(1-2x)^2$$ it is very easy to find the critical point if there is one variable

$\endgroup$
  • 1
    $\begingroup$ How would reducing to one variable help "to check the critical point"? I believe that it is not a substitute for a proper analysis. $\endgroup$ – Federico Poloni Dec 13 '15 at 7:37
  • $\begingroup$ I guess E.H.E means that after reducing to one variable, we have a very easily analyzed quadratic function in one variable. $\endgroup$ – justhalf Dec 14 '15 at 2:11
  • $\begingroup$ @justhalf Yes, but analyzing that function doesn't tell you whether the point is a minimum or a maximum (or something else) of the original function. $\endgroup$ – Federico Poloni Dec 14 '15 at 10:15
2
$\begingroup$

You want to minimize:

$$min ~ x^2 + y^2 +z^2$$

With restrictions:

$$x,y,z \ge 0$$ $$ x + y +z=1$$

Now to solve this problem you can apply the method of Lagrange multipliers and minimize:

$$x^2+y^2+z^2+ \lambda (x+y+z-1))$$

$\endgroup$
1
$\begingroup$

Restating:

Find x,y,z such that the squared distance from the origin $d^2 := x^2 + y^2 + z^2$ to the plane $g(x,y,z) = x+ y+ z - 1 = 0$ is minimal.

The gradient of g(x,y,z) is a vector normal to the surface at ( x,y,z).

A normal to the plane:

$\nabla g= (\frac{\partial g}{\partial x},\frac{\partial g}{\partial y}, \frac{\partial g}{\partial z}) = (1,1,1)$.

Consider a vector $ \vec{r} = a (1,1,1) $ from (0,0) to the plane.

The vector components satisfy $ x + y + z - 1 = 0, $

$a + a + a - 1 = 0 , $

$a = 1/3.$

$\vec{r} = a (1,1,1) = (1/3,1/3,1/3)$, i.e.

$ x = y = z = 1/3.$

The min value:

$d^2 = x^2 + y^2 + z^2 = 1/3 $.

$\endgroup$
0
$\begingroup$

Another way to see that the optimum is attained at a constant vector is via majorization and the Karamata inequality.

Credit for pointing this out goes to @zeb in this answer to a question of mine.

$\endgroup$
0
$\begingroup$

By C-S $$3(x^2+y^2+z^2)=(1+1+1)(x^2+y^2+z^2)\geq(x+y+z)^2=1.$$ Thus, $x^2+y^2+z^2\geq\frac{1}{3}$.

The equality occurs for $x=y=z=\frac{1}{3}$, which is also the answer.

$\endgroup$

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.