I'm trying to make the jump from the idea of a percentile, say, over the real number line (where the nth percentile is simply the position in which n% of data points are below it, and 100-n% are above it), to the idea of the area under a probability density function.

If I want to know the 50% percentile from a set of numbers, I'll find the point in which half the numbers are below, half the numbers are above. That's the 50% percentile, and I'm done.

If I want to know the 50% percentile from a distribution, say, a Z-score, I'll evaluate the cdf from 0 - 50, and I'm done. Am I saying this correct?

This feel right intuitively, but I need some discussion to hammer it home. Or, I could be completely off...


You are close but not exactly right. Remember that the area under a probability distribution has to sum to 1. The cumulative density function (CDF) is a function with values in [0,1] since CDF is defined as $$ F(a) = \int_{-\infty}^{a} f(x) dx $$ where f(x) is the probability density function. Then 50th percentile is the total probability of 50% of the samples which means the point where CDF reaches 0.5. Or in more general terms, the p'th percentile is the point where the CDF reaches p/100.

| cite | improve this answer | |
  • 3
    $\begingroup$ Perhaps it's worth pointing out how close the OP got - instead of "evaluating a CDF" they should be evaluating an inverse CDF. $\endgroup$ – Silverfish Jan 11 '15 at 8:18
  • 1
    $\begingroup$ so close yet so far... :) $\endgroup$ – tumultous_rooster Dec 2 '15 at 20:40
  • $\begingroup$ In general, the inverse of a CDF (in the usual sense, i.e., inverse of a function) may not exist. We should consider the so-called generalized inverse (or pseudo-inverse) of a CDF. $\endgroup$ – Danny Pak-Keung Chan Sep 21 '17 at 21:46

No. Essentially, calculating a percentile (or a p-quantile) is equivalent to finding the inverse of a CDF.

Note that the inverse, in the usual sense, of a CDF may not exist and the notion of generalized inverse should be introduced. To make discussion precise, we clarify all definitions.

Definition: A CDF is a function $F:[-\infty,\infty]\rightarrow[0,1]$ that satisfies the following conditions:

  1. (Increasing) For any $x,y\in[-\infty,\infty]$, if $x<y$, then $F(x)\leq F(y)$,

  2. (Right-continuity) For any $a\in\mathbb{R}$, we have that $F(a)=\lim_{x\rightarrow a+}F(x)$,

  3. $F(-\infty)=\lim_{x\rightarrow-\infty}F(x)=0$, and

  4. $F(\infty)=\lim_{x\rightarrow\infty}F(x)=1$.

We have at least two versions of generalized inverse of $F$, denoted by $Inv_{1}F$ and $Inv_{2}F$, which are defined as follows.

$Inv_{1}F:[0,1]\rightarrow[-\infty,\infty]$, defined by $Inv_{1}F(x)=\inf\{y\mid F(y)\geq x\},$

$Inv_{2}F:[0,1]\rightarrow[-\infty,\infty]$, defined by $Inv_{2}F(x)=\inf\{y\mid F(y)>x\}$.

Here, we adopt the convention that $\inf(\emptyset)=\infty$.

If I remember correctly, given $p\in[0,1]$, the $p$-quantile is simply defined as $Inv_{1}F(p)$.

Of course, if $F$ is strictly increasing and continuous, both versions of generalized inverse are the same and reduce to the usual inverse of function$F^{-1}:[0,1]\rightarrow[-\infty,\infty].$

For more information: https://people.math.ethz.ch/~embrecht/ftp/generalized_inverse.pdf

| cite | improve this answer | |

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.