Let's assume you have a set of S values, F1, F2, F3, ..., Fi, ... Fs such that you know the minimum and maximum possible values of Fi, but don't actually know any of the values Fi themselves. Let's call them $B_l$ for lower bound and $B_u$ for upper bound.

Now, assume I know the average of the set as well, $A = \frac{1}{S} \sum{F_i}$.

However, I'm interested in figuring out both the minimum and maximum of $\sum{F_i^q}$ as a function of A, Bl, Bu, and q.

For example, let's say I have 7 values. Their average is 6. These values I know fall between 0 and 20, but I don't actually know the values themselves, nor their actual min and max.

While I'm familiar with finding the min/max of a continuous function, I'm not totally where to even start with this one, and would appreciate any thoughts or suggested solution to the approach.

  • $\begingroup$ do you want to find the min and max of the sum as a function of $q$? And what is $q$? $\endgroup$
    – fonfonx
    Dec 21 '17 at 14:29
  • $\begingroup$ if min(F) = m, max(F) = M, let X = max(|m|, |M|); then you can get some form of a bound involving X, q and S, but that would be a crude bound (although would be tight in certain settings) $\endgroup$
    – E-A
    Dec 21 '17 at 14:31
  • $\begingroup$ Sorry, I clarified that while the values are drawn from a range, I don't know the true min and max of this particular set. $\endgroup$
    – jebyrnes
    Dec 21 '17 at 14:32
  • $\begingroup$ But the range is bounded? And do you have any knowledge on q at all? $\endgroup$
    – E-A
    Dec 21 '17 at 14:33
  • $\begingroup$ There's always Lagrange Multipliers. $\endgroup$ Dec 21 '17 at 14:37

Assuming $q \ge 1$ and non-negative $F_i$ i.e. $B_l \ge 0\,$, the generalized mean inequality gives:

$$ A = \frac{1}{S}\sum F_i \;\le\; \sqrt[q]{\frac{1}{S}\sum {F_i}^q} \;\le\; \max \{F_i\} \;\le\; B_u \\[5px] \iff\quad A^q \;\le\; \frac{1}{S}\sum {F_i}^q \;\le\; B_u^{\,q} $$

You may amend the RHS inequality somewhat by noting that $\,\max\{F_i\}\le S \cdot A - (S-1)\cdot B_l\,$, but other than that there is not much room to improve, since the LHS inequality becomes an equality for $\,F_i = F_j\,$, and the RHS one for $\,F_i = B_u\,$, so both bounds can in fact be attained.


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.