0
$\begingroup$

Suppose I have a blackbox generating values $x_n$. If I take mean($x_n^2$) it flies all over the place as new values are added -- a supermassive 48-digit value could suddenly ramp up a previously stable-looking mean by a factor of 100. However mean(log($x_n$)) converges to L. how can I estimate mean($x_n^2$) from observing log($x_n$)?

I imagine I need to extract more properties of the distribution of log($x_n$), maybe the variance?

A simple analogy might be squares of numbers: if I take the numbers [1,3,4,2,2,12,5,1,8,2], the mean is 4. but if I now square each number: [1,9,16,4,4,144,25,1,64,4], the mean is now 209. So what transformation on 4 was required to achieve 209?

Can anyone recommend a direction of research?

PS The blackbox generator is as follows: take a random permutation of 1000 elements & return the order (i.e. number of cycles required before the original order is restored).

$\endgroup$
  • $\begingroup$ It sounds like you are limited by the number of values of $x_n$ you can observe and that your distribution has a very fat tail. However, no amount of knowledge about $\log(x_n)$ will beat the mean estimate of $\frac{1}{n}\sum_{i=1}^nx_i^2$. Your only option is to increase $n$. $\endgroup$ – Alex R. Aug 10 '17 at 17:26
  • $\begingroup$ If one indeed needs to operate on so widely dispersed numbers, log-transform might actually be the best starting step, purely because of computational reasons - i.e. precision loss when adding floating-point numbers. Maybe it could be avoided by assuming some distribution for $x$, i.e. lognormal or exponential? $\endgroup$ – juod Aug 10 '17 at 17:44
  • 5
    $\begingroup$ What else do you know about the distribution? Consider the following simple example- Suppose $Z\sim N(0,1)$ is a standard normal so $X := e^{\sigma Z}$ is log normal. Then $\mathbb{E}[\ln(X)] = 0$ but $\mathbb{E}[X^2] = \mathbb{E}[e^{2\sigma Z}] = e^{2\sigma^2}$ can be any positive number you like by varying $\sigma$. This shows there is no function such that $\mathbb{E}[X^2] = f(\mathbb{E}[\ln(X)])$ in general. So you need some extra information to have any hope .. $\endgroup$ – P.Windridge Aug 10 '17 at 17:56
  • $\begingroup$ Thanks for the feedback. I've added the generating function at the bottom of the post. $\endgroup$ – P i Aug 10 '17 at 18:06
  • $\begingroup$ You need more information to pin down the precise mean. But if it helps, Jensen's Inequality tells us that $e^{2L}$ is a lower bound for $E(x^2)$. $\endgroup$ – knrumsey Aug 10 '17 at 18:07
1
$\begingroup$

Just an idea. There is no simple transformation from $\DeclareMathOperator{\E}{\mathbb{E}}\E \log X$ to the expectation or variance of $X$. There cannot be, because this depends heavily on the distribution long out into the tails. If Your distribution is lognormal, however, there is a relation, you can find it at Wikipedia.

But you can try to use empirical transforms, see for instance How does saddlepoint approximation work? for examples.

The moment generating function (mgf) of $X$ is $M(t)=\E e^{t X}$. The Mellin transform of (almost surely positive) $X$ is $m(s)=\E X^s$. That is equal to the mgf of $\log X$, see The distribution of the product of a multivariate normal and a lognormal distribution. So you can estimate the mgf of $\log X$, and observe that $$ \mathbb{V}\text{ar} X=m(2) - m(1)^2 \quad \text{and} \\ \E X = m(1) $$ and you have your estimates.

$\endgroup$

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.