Let $(s_n) = \{s_1,s_2,\dots,s_n\}$ be a sequence of real numbers. Let us assume that we do not know the the value of any $s_i, \ i\in \{1,2,\dots,n\}$, but instead the sequence $(s_n)$ has been divided in some unknown way to $k$ non-overlapping subsequences and let the length of $j$th, $j\in \{1,2,\dots,k\}$ such a subsequence be $l_j$. We can assume that all lengths $l_j$ are greater than zero and thus smaller than $n$.

Assume we have been given only the set of means of these subsequences $(a_k) = \{a_1,a_2,\dots,a_k\}$, e.g. $a_1 = \frac{s_2+s_5+s_{n-2}}{3},$ thus $l_1 = 3.$

Question: is it possible to use only the values of $(a_k)$ and $n$ to find out the mean of $(s_n)$?

If $k = n$ or if $k=1$, the question is trivial. If $l_j$ are equal for all values of $j$, the mean of $(a_k)$ is equal to the mean of $(s_n)$.

Let us find all possible ways to represent $n$ as a sum of $k$ integers, call such a representation a set of weights. For example, if $n=5$ and $k=3$, $(1,2,2)$ is a set of weights. If we consider every possible way to multiply the values of $(a_k)$ with corresponding values in a set of weights, add them together and divide the result by $n$, at least one of these is the true mean of $(s_n)$, namely the one with the set of weights $(l_1,l_2,\dots,l_k)$. However, we can then only say that the true mean is in this list of possibilities found using this method.

  • $\begingroup$ It is just an example, poorly constructed, it seems. I just came up with some numbers and wanted to use the n, the number of numbers in the sequence, instead of any fixed number. $\endgroup$ – Valtteri Nov 30 '15 at 17:44
  • $\begingroup$ We cannot reconstruct the mean given the set of means of an unknown partition. $\endgroup$ – André Nicolas Nov 30 '15 at 17:51

No, without the weights it is impossible. If you have the weights, then the answer is:


But if you don't have the weights, then the averages no longer contain enough information. This is alluded to in the final paragraph. For example, consider the sequences $1, 1, 2, 2$ and $1, 1, 1, 2$. We could split the first into $1, 1$ and $2, 2$ and the second into $1, 1, 1$ and $2$. In both cases, we get $k=2$ and $(a_1, a_2)$ = $(1, 2)$. But clearly the averages of the two sequences differ, so $k$, $n$ and the vector of $a_i$ is clearly not enough information to determine the average of the initial sequence. Entropy has won.

  • $\begingroup$ Well, that's true, I said that much in the final paragraph, but we do not know the values of $l_i$. $\endgroup$ – Valtteri Nov 30 '15 at 17:49
  • $\begingroup$ Ah yeah I just got that. Then I'm pretty sure the answer is no. $\endgroup$ – Colm Bhandal Nov 30 '15 at 17:50
  • $\begingroup$ Seems so. Was hoping that maybe there would be some way to find it as a mean of weighted means of means or something similar. Or that maybe the true mean would be the expected value of some random process that uses the values of $(a_k)$. But I don't know. $\endgroup$ – Valtteri Nov 30 '15 at 17:52
  • $\begingroup$ Well, if all you're allowed use is the set $a_k$ and the number $n$, and not the lengths, then its impossible. As I demonstrate above with a counter-example, for a given $n$ and set $a_k$, there is more than one possibility for a sequence. So no matter what process you use, it's doomed if it is only allowed the information specified. $\endgroup$ – Colm Bhandal Nov 30 '15 at 17:58
  • $\begingroup$ Thanks. That basically answers my question. Still wondering what the best possible estimate for the true mean would be, given the means of the partition, but that wasn't my question. $\endgroup$ – Valtteri Nov 30 '15 at 18:01

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.