The L-moments are useful as robust summary statistics for various probability distributions, similar to the moments but only requiring the mean of the distribution to exist. Each L-moment is a linear sum of expected values of order statistics. I had the thought that we could perhaps improve on this if we use the median instead, giving us a set of "M-moments," let's call them, which exist for all probability distributions, even if the mean doesn't exist.

We would like to take the definition of L-moment and replace the expected value $\mathbf{E}X$ with the median $\textbf{M}X$. The L-moment is a sum of expected values of order statistics, so when moving to the median there are two slightly different ways to do this: we can take a sum of medians or a median of sums, which give very slightly different results. Both seem reasonable and I'm not sure if either is "right," but perhaps one is well-studied. I'll call the two versions $M_n$ and $N_n$: $$ \begin{align} M_1 &= \textbf{M}X & N_1 &= \textbf{M}X \\ M_2 &= (\textbf{M}X_{2:2} - \textbf{M}X_{1:2})/2 & N_2 &= \textbf{M}(X_{2:2} - X_{1:2})/2 \\ M_3 &= (\textbf{M}X_{3:3} - 2 \textbf{M}X_{2:3} + \textbf{M}X_{1:3})/3 & N_3 &= \textbf{M}(X_{3:3} - 2 X_{2:3} + \textbf{M}X_{1:3})/3 \\ M_4 &= (\textbf{M}X_{4:4} - 3 \textbf{M}X_{3:4} + 3 \textbf{M}X_{2:4} - \textbf{M}X_{1:4})/4 & N_4 &= \textbf{M}(X_{4:4} - 3 X_{3:4} + 3 X_{2:4} - X_{1:4})/4 \\ \cdots & & & \cdots \end{align} $$

where $X_{k:n}$ is the $k$'th order statistic from a random sample of size $n$ (treated as another random variable).

The question is: do these things have a name, and have they been studied? Do we know how robust, efficient, etc they are? It's such a simple and basic idea that it seems very likely this is well known, and it seems extremely useful that these exist for every probability distribution.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.