The following problem has occurred in a study of energy dissipation in a chain of coupled, damped oscillators.

The problem:

Let me define specific rational functions $f$, $g$, and $h$ from $\mathbb R_+^n$ into $\mathbb R_+$ ($n \geq 2$) by the following expressions: \begin{align*} f(x) & = 1 + \frac{x_2}{x_1} + \frac{x_2 x_4}{x_1 x_3} + \cdots = 1 + \frac{x_2}{x_1} \left( 1 + \frac{x_4}{x_3} \left( 1 + \frac{x_6}{x_5} \cdots \right) \right) \ , \tag*{(1a)} \\ g(x) & = 1 + \frac{x_3}{x_2} + \frac{x_3 x_5}{x_2 x_4} + \cdots = 1 + \frac{x_3}{x_2} \left( 1 + \frac{x_5}{x_4} \left( 1 + \frac{x_7}{x_6} \cdots \right) \right) \ , \tag*{(1b)} \\ \\ h(x) & = x_1 f(x)\, g(x) \\ & = x_1 + \cdots + x_n + \frac{x_1 x_3}{x_2} + \frac{x_2 x_4}{x_3} + \cdots + \frac{x_{n-2} x_n}{x_{n-1}} \\ & \qquad + \frac{x_1 x_3 x_5}{x_2 x_4} + \frac{x_2 x_4 x_6}{x_3 x_5} + \cdots + \frac{x_{n-4} x_{n-2} x_n}{x_{n-3} x_{n-1}} \\ & \qquad + \frac{x_1 x_3 x_5 x_7}{x_2 x_4 x_6} + \cdots \quad \text{ etc. } \ . \tag*{(1c)} \end{align*} The dots in (1a-c) indicate a continuation of the respective patterns, with as many terms as possible as long as the indices $i$ on all involved $x_i$ remain within $\{1,...,n\}$.

I am interested in a meaningful "minimization" of the function $h$. From its definition above it is obvious that $h$ is both strictly positive and a homogeneous function of degree 1. For any $t\in\mathbb R_+$ then $h(t\,x) = t\,h(x)$, which shows that $h$ has arbitrarily small values close to the origin of $\mathbb R^n$, and so it cannot have a global minimum in $\mathbb R_+^n$. An obvious way around this problem is to impose an additional constraint on $x$. Because of the homogeneity of $h$, the natural choice is a constraint of the form \begin{align*} \prod_{i=1}^n x_i = c \ , \tag*{(2)} \end{align*} with some constant $c>0$. Without loss of generality, we can choose any fixed value, for instance $c=1$; function values for any $c>0$ are (by homogeneity of $h$) related to those for $c=1$ by $h(t\,x) = t\,h(x)$, with $t=c^{1/n}$.

We know that, on the hypersurface $\mathbb H_c^n = \{ x \in \mathbb R_+^n \! : \prod_{i=1}^n x_i = c \}$ defined by the constraint (2), $h$ must have a minimum since it is continuous (even $C^\infty$) in $\mathbb R_+^n$ and (Proposition 1:) if $x_i \to \infty$ for some $i$, or if $x \in \mathbb H_c^n$ approaches the boundary of $\mathbb R_+^n$, then $h(x) \to \infty$. Proposition 1 is easily proved as follows. Noting that $h(x) > x_1 + \cdots + x_n$, the first part is obvious. For the second part, let us assume that $x_i < \epsilon$ for some $i$. Then due to the constraint (2) there is at least one $j$ such that $x_j > (c/\epsilon)^{\frac1{n-1}}$, whence $h(x) > (c/\epsilon)^{\frac1{n-1}}$. Consequently, if $\epsilon \to 0$ then $h(x) \to \infty$.

The question now is, at which point(s) $x_{\rm min} \in \mathbb H_c^n$ is this minimum located, and what is the value $h_{\rm min} = h(x_{\rm min})$?

Any minimum of $h$ on $\mathbb H_c^n$ must also be a stationary point of $h$ on $\mathbb H_c^n$ (i.e., the gradient of $h$ in $H_c^n$ must vanish at that point). I believe that for any given $n$ there is only one stationary point of $h$ on $\mathbb H_c^n$, which then also must be its (unique) minimum. Unfortunately, I cannot prove this yet, but I can prove the existence of a remarkably simple stationary point:

Proposition 2: Let us choose $c = \frac14$ in the constraint (2). The point $x$ with components \begin{align*} x_1 = x_n = \frac12 , \quad x_2 = \cdots = x_{n-1} = 1 \tag*{(3)} \end{align*} is a stationary point of $h$ on $\mathbb H_{1/4}^n$, with corresponding stationary value $h(x) = \frac{n^2}4$.

This extremely simple result suggests that there may be a simple reason why it is true, but so far I have only a (relatively) complicated proof which I won't write down here.


  1. Why is the form (3) of the stationary point this simple?
  2. Is (3) the only stationary point (and thus the global minimum) of $h$ under the constraint (2) with $c = \frac14$?

1 Answer 1


I'm not sure about stationary points, but the global minimum is certainly there. Let's do it for $n=5$. Write $$ f(x)=1+\frac{x_2}{2x_1}+\frac{x_2}{2x_1}+\frac{x_2x_4}{2x_1x_3}+\frac{x_2x_4}{2x_1x_3}\,, \\ g(x)=\frac 12+\frac 12+\frac{x_3}{2x_2}+\frac{x_3}{2x_2}+\frac{x_3x_5}{x_2x_4} $$ Now use Cauchy-Schwarz ($(\sum a)(\sum b)\ge (\sum\sqrt{ab})^2$) honestly coupling the terms as presented and then apply AM-GM.

  • $\begingroup$ This is a truly beautiful argument! Thank you so much! How on earth did you find it? Now that you've pointed it out to me I don't understand how I could have missed it in the first place. ;-) $\endgroup$ Commented Sep 13, 2016 at 13:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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