Response Supports And Valid Values

Why Response Support Matters

Before choosing a probability model for a response, identify which values the response can take and which values are supported by the proposed distribution. These are related but not identical questions. A response may be recorded on a broad allowable scale, while a particular model assigns probability only to part of that scale. Moreover, a value can belong to the support even when the probability of observing that exact value is zero.

For example, a count cannot be negative, and a binary outcome can only be zero or one. For a continuous measurement, however, observing any one exact value may have probability zero even though values near it are possible. The support gives a precise way to distinguish such cases.

Definition of Support

Definition: Let $X$ be a real-valued random variable with distribution $F$. Its topological support, written $\operatorname{supp}(X)$, is the set of all $x\in\mathbb{R}$ such that every open neighborhood of $x$ has positive probability: $$ \operatorname{supp}(X) = \left\{x\in\mathbb{R}: P\bigl(|X-x|<\varepsilon\bigr)>0 \text{ for every }\varepsilon>0 \right\}. $$

In words, a value is in the support if the response has a positive chance of landing arbitrarily close to it. The definition concerns neighborhoods, not necessarily the probability of the single value itself. Thus, for a continuous response, a point can have probability zero and still be in the support.

It is useful to separate the allowable response values from the support. Allowable values describe the response’s scale or outcome space; for example, counts take values in $\{0,1,2,\ldots\}$. The support is determined by the distribution of $X$. A particular count model could assign positive probability to every nonnegative integer, or it could concentrate its probability on only some of them. The support records what that distribution allows locally.

Worked Examples: Discrete Responses

Worked Example: A Binary Response

Let $Y$ be a binary response with $P(Y=0)=0.7$ and $P(Y=1)=0.3$. Find its support and distinguish it from its allowable response values.

The allowable values are $\{0,1\}$. Each has positive probability, so every neighborhood of 0 contains the outcome 0 with positive probability, and every neighborhood of 1 contains the outcome 1 with positive probability. Therefore both values are in the support.

If $x\notin\{0,1\}$, its distance from the finite set $\{0,1\}$ is positive. Choose a neighborhood around $x$ smaller than that distance. It contains neither possible outcome, so its probability is zero. Hence $$ \operatorname{supp}(Y)=\{0,1\}. $$

The allowable response values remain $\{0,1\}$ even if a different model sets $P(Y=1)=0$. In that case the support under that particular distribution is only $\{0\}$. Support can therefore depend on the model’s probability assignments.

Worked Example: A Count Response

Let $Y$ take the nonnegative integer values $k=0,1,2,\ldots$, with $P(Y=k)=2^{-(k+1)}$. These probabilities sum to one because $\sum_{k=0}^{\infty}2^{-(k+1)}=1$. Find the support.

Every nonnegative integer $k$ has positive probability $2^{-(k+1)}$. Any neighborhood of $k$ contains $k$, so it has positive probability. Thus every nonnegative integer belongs to the support.

If $x$ is not a nonnegative integer, there is a small enough neighborhood of $x$ that contains no nonnegative integer. Since $Y$ can take only nonnegative integer values, that neighborhood has probability zero. Consequently, $$ \operatorname{supp}(Y)=\{0,1,2,\ldots\}. $$

This support is appropriate for a count response whose model assigns positive probability to every nonnegative integer. The fact that counts are allowable values by itself would not guarantee that every count belongs to the support of every possible count distribution.

Two Properties of Support

The neighborhood definition has two useful consequences. First, the support is closed and has probability one. Second, for a purely discrete distribution, the support is the closure of the values with positive point probability—not necessarily just those values themselves.

Theorem: For any real-valued random variable $X$, $\operatorname{supp}(X)$ is a closed subset of $\mathbb{R}$ and $P\bigl(X\in\operatorname{supp}(X)\bigr)=1$. It is also contained in every closed set $C$ satisfying $P(X\in C)=1$.

Proof. Write $S=\operatorname{supp}(X)$. If $x\notin S$, the definition gives an $\varepsilon>0$ for which $P(|X-x|<\varepsilon)=0$. Every point sufficiently close to $x$, for instance every $y$ with $|y-x|<\varepsilon/2$, has a neighborhood contained in $(x-\varepsilon,x+\varepsilon)$. That neighborhood has probability zero, so $y\notin S$. Thus the complement of $S$ is open and $S$ is closed.

To show that $P(X\in S)=1$, consider all open intervals with rational endpoints that have probability zero under $X$. There are at most countably many such intervals. If $x\notin S$, some open interval around $x$ has probability zero. By the density of the rationals, there is an interval with rational endpoints that contains $x$ and is contained in that zero-probability interval. Therefore $X\notin S$ is contained in the union of the events that $X$ falls in one of these countably many zero-probability rational intervals. A countable union of events of probability zero has probability zero, so $P(X\notin S)=0$.

Finally, let $C$ be closed with $P(X\in C)=1$. If $x\notin C$, closedness gives a neighborhood of $x$ disjoint from $C$; its probability under $X$ is zero. Hence $x\notin S$, which proves $S\subseteq C$. This establishes the stated properties. $\square$

The theorem says that support is the smallest closed set that contains the response with probability one. “Smallest” here means contained in every other closed set with probability one; it does not mean that every support point must itself be observed with positive probability.

Theorem: Suppose $X$ has a purely discrete distribution, concentrated on its point atoms $A=\{x:P(X=x)>0\}$. Then $$ \operatorname{supp}(X)=\overline{A}, $$ where $\overline{A}$ is the closure of $A$.

Proof. If $x\in\overline{A}$, every neighborhood of $x$ contains some $a\in A$. The event $X=a$ is contained in the event that $X$ falls in that neighborhood, and $P(X=a)>0$. Thus every neighborhood of $x$ has positive probability, so $x\in\operatorname{supp}(X)$.

If $x\notin\overline{A}$, there is a neighborhood of $x$ disjoint from $A$. Since the distribution is concentrated on $A$, the probability that $X$ falls in that neighborhood is zero. Hence $x\notin\operatorname{supp}(X)$. Both inclusions hold, proving the result. $\square$

If $A$ has no finite accumulation point, it is closed in $\mathbb{R}$, and in that case the support equals $A$. If its atoms accumulate at a finite point, that limit point also belongs to the support.

Worked Example: A Zero-Probability Limit Point

Let $P(Y=1/n)=2^{-n}$ for $n=1,2,\ldots$. The probabilities sum to one, since $\sum_{n=1}^{\infty}2^{-n}=1$. Identify the support, paying particular attention to zero.

The atom set is $A=\{1,1/2,1/3,\ldots\}$, and the discrete-support result gives $\operatorname{supp}(Y)=\overline{A}$. The sequence $1/n$ converges to zero, so $0\in\overline{A}$. More directly, for every $\varepsilon>0$, choose $n>1/\varepsilon$. Then $0<1/n<\varepsilon$, and $P(|Y-0|<\varepsilon)\geq P(Y=1/n)=2^{-n}>0$.

Although $P(Y=0)=0$, zero is in the support. The only finite limit point added to $A$ is zero, so $$ \operatorname{supp}(Y)=\{0\}\cup\{1,1/2,1/3,\ldots\}. $$ This is why “the support is the set of values with positive probability” is not generally correct for discrete distributions.

Bounded and Positive Continuous Responses

For a continuous response, support is usually described by intervals rather than by individual point probabilities. A continuous measurement may have probability zero at every exact value. The question is whether every sufficiently small interval around a point has positive probability.

Worked Example: A Bounded Continuous Response

Suppose $Y$ is a continuous measurement satisfying $P(2\leq Y\leq5)=1$, and suppose every interval $(u,v)$ with $2<u<v<5$ has positive probability. Find its support.

For any $x\in(2,5)$, every neighborhood of $x$ contains a smaller interval lying inside $(2,5)$. That smaller interval has positive probability, so $x$ is in the support. Every neighborhood of 2 contains an interval $(2,v)$ for some $v\in(2,5)$, which has positive probability; the same reasoning applies to 5. Thus both endpoints are also in the support.

If $x<2$, choose a neighborhood that lies entirely below 2; it has probability zero because $Y\in[2,5]$ with probability one. If $x>5$, a neighborhood lying entirely above 5 likewise has probability zero. Therefore $\operatorname{supp}(Y)=[2,5]$.

The endpoints belong to the support even if $P(Y=2)=P(Y=5)=0$. The support identifies the closed range approached with positive probability, not just values that occur as exact observations.

Worked Example: A Positive Continuous Response

Suppose a positive measurement $Y$ satisfies $P(Y>0)=1$, and every interval $(u,v)$ with $0<u<v$ has positive probability. Find its support.

Every $x>0$ is in the support because each neighborhood of $x$ contains an interval $(u,v)$ with $0<u<v$, and that interval has positive probability. Zero is also in the support: for every $\varepsilon>0$, the interval $(\varepsilon/3,2\varepsilon/3)$ lies inside $(0,\varepsilon)$ and has positive probability. Every neighborhood of zero therefore has positive probability, although $P(Y=0)=0$.

For $x<0$, choose a neighborhood entirely below zero. It has probability zero because $Y>0$ with probability one. It follows that $\operatorname{supp}(Y)=[0,\infty)$.

This example highlights a useful distinction: the response is strictly positive with probability one, yet zero is a support point. The support is closed, so a boundary approached by positive values can belong to it without being an observed value.

Support Need Not Fill the Allowable Range

A bounded response does not automatically have every point between its smallest and largest possible values in its support. For example, suppose a continuous measurement takes values in $[0,1]\cup[3,4]$, with positive probability in every open interval contained in either component. Its support is $[0,1]\cup[3,4]$, not the full interval $[0,4]$. A point such as 2 is within the overall bounds, but a sufficiently small neighborhood around 2 has probability zero.

This distinction is important when deciding whether a response model is appropriate. First check whether the model’s support is consistent with the response scale and the data-generating situation. Then check whether its support matches the values the proposed distribution can actually generate. A model that allows negative values may be unsuitable for a response that must be nonnegative, even if its mean is positive.

Common Mistakes

Equating support with positive point probabilities. That works only when the atom set is closed. In general, a discrete distribution’s support is the closure of its atoms; an accumulation point can have zero point probability and still be in the support.
Removing a boundary because its exact probability is zero. Support depends on probabilities of neighborhoods. Endpoints of continuous intervals and zero for a strictly positive response can be support points.
Assuming allowable values determine support. A response’s scale describes what values are permissible in principle. A particular model may assign probability only to a subset of them.
Assuming boundedness means a full interval of support. A bounded response may have gaps in the values its distribution supports.

Key takeaway: A response value is in the topological support when every neighborhood of it has positive probability. Support is closed and has probability one; for a purely discrete response it is the closure of the positive-probability atoms.

Check Your Understanding

Use the neighborhood definition of support to answer each question.

A binary response has positive probability at both 0 and 1. What is its support? How would the support change if $P(Y=1)=0$?
A count response has positive probability at every nonnegative integer and nowhere else. State its allowable values and its support.
For the distribution with $P(Y=1/n)=2^{-n}$, explain why zero belongs to the support although $P(Y=0)=0$.
A continuous response lies in $[a,b]$ with probability one and has positive probability in every interval inside $(a,b)$. What is its support, and why can the endpoints be included when their point probabilities are zero?
Give a reason why a bounded response could have support that is not a single interval.