First let us remind ourselves of the statement of the AM–GM inequality:

Theorem: (AM–GM Inequality) For any sequence $(x_n)$ of $N\geqslant 1$ non-negative real numbers, we have $$\frac1N\sum_k x_k \geqslant \left(\prod_k x_k\right)^{\frac1N}$$

It is well known that the sum operator ∑ can be generalised so that it operates on continuous functions rather than on discrete sequences. This generalisation is the integral operator ∫.

Likewise, we can generalise the product operator to operate on continuous functions. We begin with the following property of the discrete product: $$\log\prod_k A_k=\sum_k \log A_k$$ (assuming that $A_k>0$). Using this we can define the continuous product (also known as the geometric integral according to Wikipedia) as follows: $$\prod_a^b f(x)\ dx := \exp\int_a^b\log f(x)\ dx.$$

assuming that the integral converges. From these definitions we can seek to generalise the AM–GM inequality to continuous functions:

Proposition: (Continuous AM–GM Inequality) For any suitably well-behaved non-negative function $f$ defined on $[a,b]$ with $a<b$, we have: $$\frac1{b-a}\int_a^bf(x)\ dx \geqslant \left(\prod_a^b f(x)\ dx\right)^\frac1{b-a}$$ or in traditional notation: $$\frac1{b-a}\int_a^bf(x)\ dx \geqslant \exp\left(\frac1{b-a} \int_a^b \log f(x)\ dx\right)$$

As a simple illustration, consider the function $f(x)=1+\sin x$ and fix the lower bound $a=0$. On the following graph the $x$-axis represents the upper bound of integration $b$, and on the $y$-axis are represented the the arithmetic mean (blue) and the geometric mean (red) of $f$ on the interval $[a,b]$ where $a=0$ and $b=x$.

Clearly the claimed inequality seems to hold, and is in fact quite tight when $b\approx0$.

This inequality is of interest not least because it is readily abe to produce a number of non-trivial numerical inequalities pertaining to known constants. For instance, again with the example $f(x)=1+\sin x$, set $b=\pi/2$. Then assuming the proposition holds, we have

$$\frac{2}{\pi}\int_0^{\pi/2} 1 + \sin x\ dx \geqslant \exp \left( \frac{2}{\pi} \int_0^{\pi/2} \log(1+\sin x)\ dx \right)$$ Evaluating these integrals and rearranging, we obtain: $$G \leqslant \frac\pi4\log\left(2+\frac2\pi\right) \approx 0.9313$$ where $G\approx 0.9159$ is Catalan's constant.

Anyway I would like to ask, firstly:

Is the claimed inequality true, and if so what is the proof and on what class of functions is it applicable?

Secondly, and this is more of a fun little challenge:

Assuming its veracity, can this inequality be used to prove remarkable numerical inequalities, e.g. $\pi < 22/7$ or $e^\pi - \pi < 20$?

  • 13
    $\begingroup$ Have you heard about Jensen's inequality? $\endgroup$ – mickep Apr 23 '17 at 12:29
  • 2
    $\begingroup$ Very nice question! +2 if I could! But don't be so quick to accept an answer... This discourages other people from writing answers. Wait a few days... Especially given your "fun little challenge". I'd love to see someone solve it... $\endgroup$ – Pedro A Apr 23 '17 at 13:31
  • 2
    $\begingroup$ Thanks for the advice @Hamsteriffic, I've "un-accepted" the answer to see if that encourages someone else to write a more comprehensive answer. $\endgroup$ – user1892304 Apr 23 '17 at 13:49
  • 2
    $\begingroup$ Note that the continuous product should have $dx$ in superscript, since it comes from raising $f(x_i)$ to the power of $\Delta x$, similarly to how multiplying by $\Delta x$ in the Riemann sum gives multiplication by $dx$ in the final notation of integration. $\endgroup$ – Ruslan Apr 23 '17 at 20:28
  • 1
    $\begingroup$ Similarly, you can also consider $\operatorname{HM}(f,a,b)=\frac{b-a}{\int_a^b\frac{dx}{f(x)}} as generalized harmonic mean. $\endgroup$ – g.kov Apr 24 '17 at 17:26

I would write this inequality as $$\frac1{b-a}\int_a^b\exp(g(x))\,dx\ge\exp\left(\frac1{b-a}\int_a^b g(x)\,dx\right).$$ In this guise it is the case of Jensen's inequality for the convex function $\phi(t)=\exp(t)$.


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.