Median Sets and Nonunique Medians

Why a Distribution Can Have More Than One Median

A median is often introduced as a single “middle” value. For some distributions there is exactly one such value, but for others every point in an interval satisfies the same middle-probability requirement. This occurs, for example, when probability is split between separated regions and a gap lies between them. Atoms can also affect where a median set begins or ends.

To describe these cases without ambiguity, define a median by the probability on each side, including the point itself. This definition works for discrete, continuous, and mixed distributions; it does not require a density or a unique solution to an equation involving the CDF.

Definition: For a real-valued random variable $X$, a number $m$ is a median if $$ P(X\leq m)\geq \frac12 \qquad\text{and}\qquad P(X\geq m)\geq \frac12. $$ The median set is the set of all numbers $m$ satisfying both inequalities.

Write $F(x)=P(X\leq x)$ for the CDF and $F(m-)=P(X<m)$ for its left limit at $m$. Since $P(X\geq m)=1-P(X<m)$, the definition is equivalently

m\text{ is a median} \quad\Longleftrightarrow\quad F(m-)\leq \frac12\leq F(m).

The left limit matters whenever there may be an atom at $m$: the jump size is $P(X=m)=F(m)-F(m-)$. A median does not have to satisfy $F(m)=1/2$. It is enough that the probability level $1/2$ lies between the CDF immediately before $m$ and its value at $m$.

The Median Set Is an Interval

The following result describes the entire set, not just one median. It also guarantees that a real-valued distribution always has at least one median.

Theorem: Let $X$ be real-valued with CDF $F$. Define $$ a=\inf\{x\in\mathbb R:F(x)\geq 1/2\}, \qquad b=\sup\{x\in\mathbb R:F(x-)\leq 1/2\}. $$ Then $a$ and $b$ are finite, $a\leq b$, and the median set is exactly the closed interval $[a,b]$. In particular, the median is unique if and only if $a=b$.

Proof. The CDF limits satisfy $F(x)\to 0$ as $x\to-\infty$ and $F(x)\to 1$ as $x\to\infty$. Thus the set defining $a$ is nonempty and bounded below, so $a$ is finite. The set defining $b$ is nonempty because $F(x-)\to 0$ as $x\to-\infty$. Choose $t$ with $F(t)>1/2$. For every $x>t$, $F(x-)\geq F(t)>1/2$, so the set defining $b$ is bounded above and $b$ is finite.

The set $\{x:F(x)\geq 1/2\}$ is upward closed because $F$ is nondecreasing. For every $\varepsilon>0$, its definition as an infimum implies $F(a+\varepsilon)\geq 1/2$. Right-continuity of $F$, applied as $\varepsilon\downarrow 0$, gives $F(a)\geq 1/2$. For every $x<a$, $F(x)<1/2$, and hence $F(a-)=\sup_{x<a}F(x)\leq 1/2$. Therefore $a$ is a median.

To see that $b$ is also in the set, take any $y<b$. By the definition of supremum there is some $x$ in the set defining $b$ with $y<x$. Monotonicity gives $F(y)\leq F(x-)\leq 1/2$. Taking the supremum over $y<b$ yields $F(b-)\leq 1/2$. Also $a\leq b$, since $a$ is itself in the set defining $b$.

Finally, the condition $F(m)\geq 1/2$ holds for every $m\geq a$, and fails for $m<a$. The condition $F(m-)\leq 1/2$ holds for every $m\leq b$, and fails for $m>b$, by monotonicity of the left limits. Both conditions therefore hold exactly when $a\leq m\leq b$. This proves the result. $\square$

The left endpoint $a$ is the generalized-inverse quantile at probability $1/2$. The whole median set can be larger: reporting just that quantile gives one median, but may omit other valid medians. The right endpoint $b$ records how far the probability-balance condition continues to hold.

Worked Examples

Worked Example: A Gap in the Support Gives an Interval of Medians

Suppose $X$ has probability $1/2$ uniformly distributed on $[-2,-1]$ and probability $1/2$ uniformly distributed on $[1,3]$. Find its median set.

For any $m\in[-1,1]$, the entire left component lies at or below $m$, while the right component lies at or above $m$, up to endpoints that have probability zero. Consequently, $$ P(X\leq m)=\frac12 \qquad\text{and}\qquad P(X\geq m)=\frac12. $$ Every point in $[-1,1]$ is therefore a median.

If $m<-1$, then $P(X\leq m)<1/2$, because it includes less than all of the left component (or none of it). If $m>1$, then $P(X\geq m)<1/2$, because it includes less than all of the right component (or none of it). Thus no point outside $[-1,1]$ is a median. The median set is exactly $[-1,1]$, including both endpoints.

Worked Example: Two Atoms at the Ends of a Gap

Let $P(X=0)=1/2$ and $P(X=4)=1/2$. Determine all medians.

At $m=0$, $P(X\leq 0)=1/2$ and $P(X\geq 0)=1$, so 0 is a median. For $0<m<4$, the point 0 is below $m$ and the point 4 is above it, giving $$ P(X\leq m)=\frac12 \qquad\text{and}\qquad P(X\geq m)=\frac12. $$ At $m=4$, $P(X\leq 4)=1$ and $P(X\geq 4)=1/2$, so 4 is also a median. If $m<0$, then $P(X\leq m)=0$; if $m>4$, then $P(X\geq m)=0$. The median set is $[0,4]$.

This example has no probability between its two possible values, but every point between them meets the definition. A median need not be a value that $X$ can actually take.

Worked Example: An Atom Can Give a Unique Median

Suppose $X$ has mass $0.4$ at $-1$, mass $0.3$ at $2$, and mass $0.3$ at $5$. Check whether the median is unique.

At $m=2$, $$ P(X\leq 2)=0.4+0.3=0.7, \qquad P(X\geq 2)=0.3+0.3=0.6. $$ Both probabilities are at least $1/2$, so 2 is a median. For any $m<2$, $P(X\leq m)\leq 0.4<1/2$. For any $m>2$, $P(X\geq m)\leq 0.3<1/2$. Thus 2 is the unique median.

At the atom, the CDF jumps from $F(2-)=0.4$ to $F(2)=0.7$, crossing $1/2$. That jump allows the median condition to hold even though neither side probability is exactly $1/2$.

Worked Example: An Atom Can Also Be an Endpoint of a Median Interval

Let $P(X=0)=0.5$, $P(X=1)=0.2$, and $P(X=2)=0.3$. Find the median set.

At $m=0$, $P(X\leq 0)=0.5$ and $P(X\geq 0)=1$, so 0 is a median. For $0<m<1$, $P(X\leq m)=0.5$ and $P(X\geq m)=0.5$, so every such $m$ is a median. At $m=1$, the probabilities are $P(X\leq 1)=0.7$ and $P(X\geq 1)=0.5$, so 1 is a median too.

For $1<m\leq 2$, $P(X\geq m)=0.3<0.5$, and for $m>2$ it is zero. If $m<0$, $P(X\leq m)=0$. The median set is exactly $[0,1]$. Here an atom at 1 ends the interval of medians; its location and mass matter, not merely the fact that an atom is present.

How Gaps and Atoms Affect the Set

A gap in the support can produce multiple medians when the cumulative probability on one side has reached $1/2$ and remains at that level across the gap. Every point in that stretch can divide the probability into at least $1/2$ on each side. The separated uniform components in the first example illustrate this: the CDF stays at $1/2$ between the components.

Atoms affect the two endpoints through jumps in the CDF. A jump that goes strictly from below $1/2$ to above $1/2$ gives a unique median at that atom, as in the third example. If the CDF reaches exactly $1/2$ and stays there before rising again, the median set can extend across the intervening region; an atom can lie at an endpoint of that interval. Having atoms alone does not determine whether the median is unique.

A useful summary is that the median set is nonunique precisely when its endpoints satisfy $a<b$. The distribution may still assign zero probability to most or all points in that interval. What matters is whether both probability inequalities continue to hold throughout it, not whether the interval itself contains likely outcomes.

Common Mistake

A common error is to define a median by the equation $F(m)=1/2$, then conclude that no median exists when the CDF jumps over $1/2$, or that there can be only one median when it stays flat at $1/2$. The equation is not the definition. With an atom, the correct condition is $F(m-)\leq 1/2\leq F(m)$; with a flat stretch, more than one point can satisfy the probability requirements.

Key takeaway: A median is any $m$ with at least half the probability at or below it and at least half at or above it. The complete set of such values is a closed interval, which may reduce to one point or contain many points.

Check Your Understanding

Use the probability definition and the CDF characterization to answer these questions.

For masses $0.5$ at 0 and $0.5$ at 6, what is the median set? Explain why its endpoints are included.
If $F(m-)=0.45$ and $F(m)=0.65$, is $m$ a median? What does the jump represent?
Can a median lie outside the support of $X$? Use a distribution with separated components to explain.
What condition on the endpoints $a$ and $b$ distinguishes a unique median from a nonunique median?
Why does the equation $F(m)=1/2$ fail as a general definition of a median?