Here's my solution to an old qualifier problem. Would you tackle it differently? Is there a flaw in my work?

Suppose that $\alpha_1, \dotsc, \alpha_n$ are positive numbers such that $$\frac1{\alpha_1}+ \dotsb + \frac1{\alpha_n}<1.$$

Show that $$\int_{1}^\infty \dotsb\int_{1}^\infty \frac{dx_1 \dotsb dx_n}{x_1^{\alpha_1}+\dotsb + x_n^{\alpha_n}}<\infty.$$

My answer: $x_1^{\alpha_1}+\dotsb + x_n^{\alpha_n} \overset{\text{(AM $\geq$ GM)}}{\geq} n\cdot \sqrt[n]{x_1^{\alpha_1}\dotsb x_n^{\alpha_n}}$

$$\Rightarrow \int_{1}^\infty \dotsb\int_{1}^\infty \frac{dx_1 \dotsb dx_n}{x_1^{\alpha_1}+\dotsb + x_n^{\alpha_n}} \leq \frac1n \int_{1}^\infty \dotsb\int_{1}^\infty \frac{dx_1 \dotsb dx_n}{x_1^{\alpha_1/n}\dotsb x_n^{\alpha_n/n}},$$

at which point if we switch to hyperspherical coordinates

$$\leq \frac1n \int_{S^{n-1}} \int_{1}^\infty \frac{dx_1 \dotsb dx_n}{r^{\frac{\alpha_1+\dotsb + \alpha_n}{n}}|f(\sigma)|}r^{n-1} dr\,d\sigma,\tag{1}$$

where $\sigma$ is the position on the unit $(n-1)$ sphere in $\mathbb{R}^n$ and $|f(\sigma)|\leq1$. Note I have chosen a crude bound by integrating over the entire area outside the unit ball. Now

$$\frac1{\alpha_1}+ \dotsb + \frac1{\alpha_n}<1 \\\Rightarrow \frac{n}{\frac1{\alpha_1}+ \dotsb + \frac1{\alpha_n}}>n\\\overset{\text{AM $\geq$ HM}}{\Rightarrow}\frac{\alpha_1 + \dotsb+\alpha_n}{n}>n$$

so the integral (1) is integrating a power of $r$ which is $>1$ in the denominator, so it converges.

  • $\begingroup$ By Fubini's theorem, the integral related to the GM will diverge when $\alpha_i\le n$ for some $i$. $\endgroup$ – 23rd Oct 17 '13 at 20:06
  • $\begingroup$ Shoot, you're right! Where is the flaw in my logic? $\endgroup$ – Eric Auld Oct 17 '13 at 20:10
  • $\begingroup$ The flaw is you didn't obtain any control of the integral of $\frac{1}{|f(\sigma)|}$ over $S^{n-1}$. $\endgroup$ – 23rd Oct 17 '13 at 21:00
  • $\begingroup$ Of course! ---- $\endgroup$ – Eric Auld Oct 18 '13 at 2:05

The AM-GM could be modified as follows.

Denote $$p:=\left(\frac{1}{\alpha_1}+\cdots+\frac{1}{\alpha_n}\right)^{-1}>1,$$ and denote $$y_i=\alpha_i\log x_i,\quad\lambda_i=\frac{p}{\alpha_i}\in (0,1), \quad i=1,\cdots,n.$$ Since $\sum_{i=1}^n\lambda_i=1$, by Jensen's inequality for the exponential function,

$$\sum_{i=1}^n x_i^{\alpha_i}\ge \sum_{i=1}^n \lambda_i e^{y_i}\ge \exp\left(\sum_{i=1}^n \lambda_i y_i\right)=\prod_{i=1}^n x_i^p.\tag{1}$$

By $(1)$ and Fubini's theorem, the original integral is bounded by $$\left(\int_1^\infty\frac{dx}{x^p}\right)^n=\frac{1}{(p-1)^n}.$$

  • $\begingroup$ Wow, that is a very useful result. Thanks! $\endgroup$ – Eric Auld Oct 18 '13 at 14:23
  • $\begingroup$ @EricAuld: You are welcome! $\endgroup$ – 23rd Oct 18 '13 at 15:59
  • $\begingroup$ Where can I find out more about methods like the one above for $\sum_1^n x_i^{\alpha_i} \geq \prod_1^nx_i^p$? I don't think I would have come up with that on my own. $\endgroup$ – Eric Auld Oct 18 '13 at 16:38
  • $\begingroup$ @EricAuld: I don't know, but if you are familiar with Jensen's inequality, then you can apply it to different convex functions to produce kinds of inequalities like the one you mentioned as you wish. $\endgroup$ – 23rd Oct 18 '13 at 16:55

As already said on the page, the AM-GM inequality yields a sufficient condition too strong to be necessary. Instead one can rely directly on the change of variable $x_k^{\alpha_k}=ru_k^2$ with $r$ in $(1,\infty)$ and $u=(u_k)$ in the unit sphere $\sum\limits_ku_k^2=1$. Then the Jacobian of the transformation $(x_k)\to(r,u)$ is complicated to write down completely but it has the form $\prod\limits_k\mathrm dx_k=r^{\beta-1}\mathrm dr\mathrm d\mu(u)$ with $\beta=\sum\limits_k\frac1{\alpha_k}$, where $\mu$ is some finite measure on the unit sphere. Then $\sum\limits_kx_k^{\alpha_k}=r$ hence the integral to be considered converges if and only if $$ \int^\infty\frac{r^{\beta-1}\mathrm dr}r $$ converges, that is, if and only if $\beta\lt1$.

  • $\begingroup$ Thank you for helping me understand the problem better. This is very simple. Could you perhaps help me understand why $\prod_k \mathrm{d}x_k = r^{\beta - 1}dr d\mu(u)$? $\endgroup$ – Eric Auld Oct 18 '13 at 15:09
  • $\begingroup$ Well, you could try to compute the Jacobian when $x_k=r^{1/\alpha_k}\varphi_k(u)$ for each $k$. $\endgroup$ – Did Oct 18 '13 at 15:18
  • $\begingroup$ Ok, I get $$\frac{\partial\mathbf{x}}{\partial(r,\mathbf{u})} = \left( \begin{matrix} \frac{1}{\alpha_1}r^{1/\alpha_1 -1} & \vec{\frac{\partial x_1}{\partial \mathbf{u}}} \\ \vdots & \vdots \\ \frac{1}{\alpha_n}r^{1/\alpha_n -1} & \vec{\frac{\partial x_n}{\partial \mathbf{u}}}\end{matrix} \right),$$but how would I compute the determinant of something like that? (Or at least show it is of the form $r^{\beta -1}$? Thanks for your help $\endgroup$ – Eric Auld Oct 18 '13 at 15:33
  • $\begingroup$ Actually the Jacobian matrix is an $n\times n$ matrix whose first column is what you wrote and whose other columns are all multiples of the column of $r^{1/\alpha_k}$. Hence when one develops the determinant, every product multiply every factor $r^{1/\alpha_k}$ except one which is $r^{1/\alpha_k-1}$. Thus, every term is a factor of $r^{\beta-1}$. $\endgroup$ – Did Oct 18 '13 at 21:17
  • $\begingroup$ How do we know that $(x_k) \mapsto (r,u)$ is a diffeomorphism onto its image? $\endgroup$ – Eric Auld Nov 21 '13 at 21:59

$\newcommand{\angles}[1]{\left\langle #1 \right\rangle}% \newcommand{\braces}[1]{\left\lbrace #1 \right\rbrace}% \newcommand{\bracks}[1]{\left\lbrack #1 \right\rbrack}% \newcommand{\dd}{{\rm d}}% \newcommand{\ds}[1]{\displaystyle{#1}}% \newcommand{\equalby}[1]{{#1 \atop {= \atop \vphantom{\huge A}}}}% \newcommand{\expo}[1]{{\rm e}^{#1}}% \newcommand{\ic}{{\rm i}}% \newcommand{\imp}{\Longrightarrow}% \newcommand{\pars}[1]{\left( #1 \right)}% \newcommand{\pp}{{\cal P}}% \newcommand{\sgn}{{\rm sgn}}% \newcommand{\ul}[1]{\underline{#1}}% \newcommand{\verts}[1]{\left\vert #1 \right\vert}$ $\ds{% {\cal J} \equiv \int_{1}^{\infty}\cdots\int_{1}^{\infty} {\dd x_{1}\ldots\dd x_{n} \over {x_{1}^{\alpha_{1}} + \cdots + x_{n}^{\alpha_{n}}}}\ <\ \infty:\ {\large ?}}$

\begin{align} {\cal J} &= \int_{1}^{\infty}\cdots\int_{1}^{\infty}\bracks{% \int_{0}^{\infty}\expo{-\pars{x_{1}^{\alpha_{1}} + \cdots + x_{n}^{\alpha_{n}}}\mu}\,\dd\mu} \dd x_{1}\ldots\dd x_{n} \\[3mm]&= \int_{0}^{\infty}\dd\mu\prod_{i = 1}^{n}\int_{1}^{\infty} \expo{-\pars{x^{\alpha_{i}}}\mu}\,\,\dd x = \int_{0}^{\infty}\dd\mu\prod_{i = 1}^{n} {1 \over \alpha_{i}\,\mu^{1/\alpha_{i}}} \int_{\mu}^{\infty}\expo{-x}\,x^{\pars{1/\alpha_{i}} - 1}\,\,\dd x \\[3mm]&= \int_{0}^{\infty}\prod_{i = 1}^{n} {\Gamma\pars{1/\alpha_{i}\,,\,\mu} \over \alpha_{i}\,\mu^{1/\alpha_{i}}} \,\dd\mu = \int_{0}^{\infty} {1 \over \mu^{1/\alpha_{1} + \cdots + 1/\alpha_{n}}} \prod_{i = 1}^{n} {\Gamma\pars{1/\alpha_{i}\,,\,\mu} \over \alpha_{i}\,} \,\dd\mu \end{align}

$\Gamma\pars{z,a}$ is the $\it\mbox{Incomplete Gamma function}\ \pars{~\Re\, z > 0~}$ and $\Gamma\pars{z} \equiv \Gamma\pars{z,0}$ is the $\it\mbox{Gamma function}$.

  1. $\large\mu\ \gtrsim\ 0$: $$ {1 \over \mu^{1/\alpha_{1} + \cdots + 1/\alpha_{n}}} \prod_{i = 1}^{n} {\Gamma\pars{1/\alpha_{i}\,,\,\mu} \over \alpha_{i}\,}\ \sim\ {1 \over \mu^{\Lambda}} \prod_{i = 1}^{n} {\Gamma\pars{1/\alpha_{i}} \over \alpha_{i}\,}\,, \quad \Lambda \equiv {1 \over \alpha_{1}} + \cdots + {1 \over \alpha_{n}} < 1 $$ Since $0 < \Lambda < 1$, the integral ~ $\mu^{1 - \Lambda}$ when $\mu \gtrsim 0$
  2. $\large \mu\ \gg\ 1:$

    In this case, $\ds{% \Gamma\pars{{1 \over \alpha_{i}},\mu} \sim \mu^{\pars{1/\alpha_{i}} - 1}\,\expo{-\mu}}$ $$ {1 \over \mu^{1/\alpha_{1} + \cdots + 1/\alpha_{n}}} \prod_{i = 1}^{n} {\Gamma\pars{1/\alpha_{i}\,,\,\mu} \over \alpha_{i}\,}\ \sim\ {\expo{-n\mu} \over \mu^{n}} \prod_{i = 1}^{n}{1 \over \alpha_{i}\,}\,, $$

So far, we can conclude that the condition $\ds{\sum_{i = 1}{1 \over \alpha_{i}} < 1}$ guarantees the convergence of the integral.


IMHO the main flaw in your logic is that you have generated a sufficient but not a necessary condition for the original integral to converge. You write: $$ \int_1^\infty \cdots \int_1^\infty \frac{dx_1 \cdots dx_n}{x_1^{\alpha_1}+\cdots + x_n^{\alpha_n}} \leq \frac{1}{n} \int_{1}^\infty \cdots \int_{1}^\infty \frac{dx_1 \cdots dx_n}{x_1^{\alpha_1/n}\cdots x_n^{\alpha_n/n}} $$ But: $$ \frac{1}{n} \int_{1}^\infty \cdots \int_{1}^\infty \frac{dx_1 \cdots dx_n}{x_1^{\alpha_1/n}\cdots x_n^{\alpha_n/n}} = \frac{1}{n} \int_{1}^\infty \frac{dx_1}{x_1^{\alpha_1/n}} \cdots \int_{1}^\infty \frac{dx_n}{x_n^{\alpha_n/n}} $$ And: $$ \int_{1}^\infty \frac{dx_i}{x_i^{\alpha_i/n}} = \left[ \frac{x_i^{1-\alpha_i/n}}{1-\alpha_i/n} \right]_1^\infty = \left\{ \begin{array}{ll} \frac{1}{1-\alpha_i/n} & \mbox{if} \quad \alpha_i > n \\ \infty & \mbox{if} \quad \alpha_i \le n \end{array} \right. $$ So instead of $1/\alpha_1 + \cdots + 1/\alpha_n < 1$ we have $(1/\alpha_1 < 1/n) \wedge \cdots \wedge (1/\alpha_n < 1/n)$, which is a much stronger condition.


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.