4
$\begingroup$

Does this seem like a reasonable definition of a uniform distribution?

Let \begin{align} &S \text{ be a sample space}; \tag{1} \\ &X:S\rightarrow \mathbb{R} \; \text{ be a random variable on } S. \tag{2} \\ \end{align} Then \begin{align} & X(S) \text{ is distributed uniformly } \\ \Leftrightarrow &\text{ every realisation of } X \text{ has pre-image of equal size}, \tag{3.1}\\ \Leftrightarrow &\forall \; x_1,x_2 \in X(S): \; |X^{-1}(\{x_1\})|=|X^{-1}(\{x_2\})|. \tag{3.2} \end{align}

$\endgroup$
  • $\begingroup$ I think this is correct if $X$ can only take a finite number of values. For a continuous random variable, the preimage of any value is of measure equal to zero, even if the random variable is not uniform. $\endgroup$ – Pohoua May 18 at 16:09
  • 5
    $\begingroup$ This definition even doesn't have a probability measure $P$ in it so it is invalid. $\endgroup$ – Zhanxiong May 18 at 16:20
  • 1
    $\begingroup$ Cross-posted at math.stackexchange.com/q/3680737/321264. $\endgroup$ – StubbornAtom May 18 at 17:33
5
$\begingroup$

Let me answer the implicit question: what is a uniform distribution?

Because $X$ is a random variable, $S$ really is the underlying set in a probability space $(S,\mathfrak F, \mathbb P).$

We say $X$ has a continuous uniform distribution when there exists a subset $A\subset \mathbb R$ such that, for every interval $(a,b]\subset \mathbb{R},$ $$\mathbb{P}(X\in (a,b]) \ \propto\ \lambda((a,b]\cap A)$$ where $\lambda$ is Lebesgue measure.

$X$ has a discrete uniform distribution when there exists $A\subset \mathbb{R}$ such that for every interval $(a,b],$ $$\mathbb{P}(X\in (a,b])\ \propto\ |(a,b] \cap A|$$ where $|\cdot |$ is the cardinality of a set.

The implicit normalizing constants (denominators) in these equations are $\lambda(A)$ in the first case and $|A|$ in the second, both of which (therefore) must be finite and nonzero. In particular,

  • in the continuous case, $X$ has a probability density function equal to $$f_X(x) = \frac{1}{\lambda(A)} \mathcal{I}_A(x)$$ (where $\mathcal{I}_A$ is the indicator function of $A$);

  • in the discrete case we may write $|A|=n$ (a nonzero natural number) and see that for any number $x\in \mathbb R,$ $\Pr(X=x) = 1/n$ when $x\in A$ and otherwise $\Pr(X=x)=0.$ This defines the probability mass function of $X.$

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Are there some fundamental properties of $\lambda$ and $|\cdot|$ as measures that make them suitable for uniformity, like translation invariance plus $\sigma$-finiteness, in a way that doesn't require a discrete/continuous distinction? $\endgroup$ – jld May 18 at 18:48
  • $\begingroup$ @jld That's an interesting question. It exposes the role of a group operation on the codomain of $X.$ $\endgroup$ – whuber May 18 at 20:12
  • $\begingroup$ Ah interesting, thank you $\endgroup$ – jld May 18 at 20:36

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.