Preserving Stochastic Order Under Maps and Addition

How Transformations Affect Order

Usual stochastic order compares distributions by how much probability they place above each threshold. This comparison is useful because it can survive operations applied to both random variables: a monotone transformation changes their scale or direction, while adding common independent noise blurs their values without changing their ordering.

Write $X\leq_{\mathrm{st}}Y$ when $X$ is no larger than $Y$ in the usual stochastic order. The equivalent CDF and survival-function definitions make the direction of the comparison explicit.

Definition: For real-valued random variables $X$ and $Y$, $X\leq_{\mathrm{st}}Y$ if and only if, for every $t\in\mathbb{R}$, $$ F_X(t)\geq F_Y(t), \qquad\text{equivalently}\qquad P(X>t)\leq P(Y>t). $$ Here $F_X(t)=P(X\leq t)$ is the CDF of $X$. The direction means that $Y$ has at least as much upper-tail probability as $X$ at every threshold.

The earlier material on quantile couplings gives a useful representation of this order. For a CDF $F$, define its generalized inverse by $Q_F(u)=\inf\{x:F(x)\geq u\}$, for $0<u<1$. If $U$ is uniform on $(0,1)$, then $Q_{F_X}(U)$ has the distribution of $X$. Moreover, when $X\leq_{\mathrm{st}}Y$, the quantiles satisfy $Q_{F_X}(u)\leq Q_{F_Y}(u)$ for every $u\in(0,1)$. Thus the variables can be represented on one probability space so that their order holds pointwise.

Nondecreasing Transformations Preserve Order

A nondecreasing function never turns a smaller input into a larger output than a larger input. Applying it to an ordered coupling therefore preserves the order of the resulting random variables. The function need not be strictly increasing: it may flatten intervals and give the two transformed variables equal values.

Theorem: If $X\leq_{\mathrm{st}}Y$ and $g:\mathbb{R}\to\mathbb{R}$ is measurable and nondecreasing, then $g(X)\leq_{\mathrm{st}}g(Y)$. Measurability ensures that the transformed quantities are random variables; a monotone real-valued function is Borel measurable.

Proof. Use a quantile coupling: let $U$ be uniform on $(0,1)$, and set $X^*=Q_{F_X}(U)$ and $Y^*=Q_{F_Y}(U)$. These have the respective distributions of $X$ and $Y$, and $X^*\leq Y^*$ almost surely. Since $g$ is nondecreasing, $g(X^*)\leq g(Y^*)$ almost surely. For every $t$, this implies $$ P(g(X^*)>t)\leq P(g(Y^*)>t). $$ Because the transformed variables have the same respective distributions as $g(X)$ and $g(Y)$, the survival-function criterion gives $g(X)\leq_{\mathrm{st}}g(Y)$. $\square$

Worked Example: Rescaling Bernoulli Variables

Let $X\sim\operatorname{Bernoulli}(0.2)$ and $Y\sim\operatorname{Bernoulli}(0.65)$. A Bernoulli variable with success probability $p$ is $1$ with probability $p$ and $0$ otherwise. Verify the order and compare the transformed variables under $g(x)=2x+1$.

For $t<0$, both CDFs are $0$; for $0\leq t<1$, $F_X(t)=0.8$ and $F_Y(t)=0.35$; and for $t\geq1$, both CDFs are $1$. Hence $F_X(t)\geq F_Y(t)$ for every $t$, so $X\leq_{\mathrm{st}}Y$.

The function $g(x)=2x+1$ is increasing. It maps $0$ to $1$ and $1$ to $3$. Therefore $g(X)$ equals $3$ with probability $0.2$, while $g(Y)$ equals $3$ with probability $0.65$. At thresholds $1\leq t<3$, their CDFs are $0.8$ and $0.35$, respectively; below $1$ both CDFs are $0$, and at or above $3$ both are $1$. Thus $g(X)\leq_{\mathrm{st}}g(Y)$, as the theorem predicts.

Nonincreasing Transformations Reverse Order

A nonincreasing function reverses pointwise inequalities: if $x\leq y$, then $h(x)\geq h(y)$. The same coupling argument therefore shows that applying such a function reverses usual stochastic order. This includes the transformation $h(x)=-x$, which changes the sign of every value.

Theorem: If $X\leq_{\mathrm{st}}Y$ and $h:\mathbb{R}\to\mathbb{R}$ is measurable and nonincreasing, then $h(Y)\leq_{\mathrm{st}}h(X)$.

Proof. In the quantile coupling above, $X^*\leq Y^*$ almost surely. Since $h$ is nonincreasing, $h(Y^*)\leq h(X^*)$ almost surely. The coupled variables have the distributions of $h(Y)$ and $h(X)$, respectively. Therefore $P(h(Y)>t)\leq P(h(X)>t)$ for every $t$, which is exactly $h(Y)\leq_{\mathrm{st}}h(X)$. $\square$

Worked Example: Reversing Order by Reflection

Let $X\sim\operatorname{Bernoulli}(0.25)$ and $Y\sim\operatorname{Bernoulli}(0.75)$. As in the previous example, the CDFs show that $X\leq_{\mathrm{st}}Y$. Apply the decreasing function $h(x)=2-x$.

The transformed variable $h(X)$ equals $1$ with probability $0.25$ and $2$ with probability $0.75$. The variable $h(Y)$ equals $1$ with probability $0.75$ and $2$ with probability $0.25$.

To check the reversed order directly, for $1\leq t<2$ we have $F_{h(Y)}(t)=0.75$ and $F_{h(X)}(t)=0.25$. Below $1$ both CDFs are $0$, and at or above $2$ both are $1$. Thus $F_{h(Y)}(t)\geq F_{h(X)}(t)$ for all $t$, so $h(Y)\leq_{\mathrm{st}}h(X)$. The mass at the smaller transformed value is larger for $h(Y)$, consistent with the reversal.

Adding a Common Independent Random Variable

A shared noise term can be added to both variables without changing their order, provided that the noise is independent of each variable. Independence allows us to construct an ordered pair and the noise together without disturbing either sum’s marginal distribution.

Theorem: Suppose $X\leq_{\mathrm{st}}Y$, and let $Z$ be real-valued and independent of $X$ and of $Y$. Then $$ X+Z\leq_{\mathrm{st}}Y+Z. $$ The original $X$ and $Y$ need not be independent of each other.

Proof. Construct the quantile-coupled variables $X^*=Q_{F_X}(U)$ and $Y^*=Q_{F_Y}(U)$ using a uniform $U$. Take $Z^*$ with the distribution of $Z$, independent of $U$. Then $Z^*$ is independent of both $X^*$ and $Y^*$, $X^*\leq Y^*$ almost surely, and each sum has the desired distribution: $X^*+Z^*$ has the distribution of $X+Z$, while $Y^*+Z^*$ has the distribution of $Y+Z$. Pointwise, $$ X^*+Z^*\leq Y^*+Z^*. $$ Consequently $P(X^*+Z^*>t)\leq P(Y^*+Z^*>t)$ for every $t$. The survival-function criterion proves $X+Z\leq_{\mathrm{st}}Y+Z$. $\square$

Worked Example: Adding the Same Independent Noise

Let $X\sim\operatorname{Bernoulli}(0.25)$, $Y\sim\operatorname{Bernoulli}(0.75)$, and $Z\sim\operatorname{Bernoulli}(0.5)$, with $Z$ independent of each of $X$ and $Y$. The Bernoulli probabilities give $X\leq_{\mathrm{st}}Y$. Calculate both sum distributions.

For $X+Z$, independence gives $P(X+Z=0)=0.75(0.5)=0.375$, $P(X+Z=1)=0.75(0.5)+0.25(0.5)=0.5$, and $P(X+Z=2)=0.25(0.5)=0.125$. The probabilities add to $0.375+0.5+0.125=1$.

For $Y+Z$, the corresponding probabilities are $P(Y+Z=0)=0.25(0.5)=0.125$, $P(Y+Z=1)=0.25(0.5)+0.75(0.5)=0.5$, and $P(Y+Z=2)=0.75(0.5)=0.375$. These also add to $1$.

At $t=0$, the sum CDFs are $0.375$ and $0.125$; at $t=1$, they are $0.375+0.5=0.875$ and $0.125+0.5=0.625$. Below $0$, both are $0$, and at or above $2$, both are $1$. Thus $F_{X+Z}(t)\geq F_{Y+Z}(t)$ at every threshold, verifying $X+Z\leq_{\mathrm{st}}Y+Z$.

Why the Conditions Matter

These preservation results depend on the operation, not merely on the fact that it is applied to both distributions. A function that is not monotone can send ordered inputs in conflicting directions. Likewise, adding random variables with the same marginal noise distribution is not enough if the dependence between the noise and each variable differs.

Worked Example: A Nonmonotone Map Can Destroy Comparability

Let $X=-2$ with probability $1$, and let $Y$ equal $-1$ or $3$ with probability $0.5$ each. Since $Y\geq -2$ in every outcome, $X\leq_{\mathrm{st}}Y$. Consider the nonmonotone map $g(x)=x^2$.

Here $g(X)=4$ with probability $1$, whereas $g(Y)$ equals $1$ or $9$, each with probability $0.5$. At $t=2$, $F_{g(X)}(2)=0$ and $F_{g(Y)}(2)=0.5$, so $g(X)\leq_{\mathrm{st}}g(Y)$ fails. The reverse order fails as well: at $t=4$, $F_{g(Y)}(4)=0.5$ but $F_{g(X)}(4)=1$, so $F_{g(Y)}(4)\geq F_{g(X)}(4)$ is false. Squaring is not nondecreasing on the whole real line, and the transformed variables are not ordered either way.

Independence in the addition theorem is also essential. For example, let $X\sim\operatorname{Bernoulli}(0.25)$ and $Y\sim\operatorname{Bernoulli}(0.75)$, so $X\leq_{\mathrm{st}}Y$. In the sum involving $X$, choose $Z=-2$ when $X=1$ and $Z=0$ otherwise. In the sum involving $Y$, choose a noise variable with the same marginal distribution, $P(Z=-2)=0.25$ and $P(Z=0)=0.75$, but set $Z=-2$ when $Y=0$ and $Z=0$ otherwise. In both constructions the noise has the same marginal distribution, but it is dependent on the variable being added to.

Then $X+Z$ is $-1$ with probability $0.25$ and $0$ with probability $0.75$, while $Y+Z$ is $-2$ with probability $0.25$ and $1$ with probability $0.75$. At $t=-1.5$, $P(X+Z>t)=1$ but $P(Y+Z>t)=0.75$. Thus $X+Z\leq_{\mathrm{st}}Y+Z$ fails. The shared noise distribution alone does not give the conclusion; the required independence prevents this kind of dependence-driven change.

Common Mistakes

Using the wrong direction for CDFs. For $X\leq_{\mathrm{st}}Y$, the CDF of $X$ is greater than or equal to the CDF of $Y$; its survival probabilities are smaller than or equal to those of $Y$.
Assuming any transformation preserves order. Preservation is guaranteed for nondecreasing maps, and reversal for nonincreasing maps. A nonmonotone map needs a separate argument.
Forgetting the domain of monotonicity. Squaring is nondecreasing on nonnegative inputs but not on all of $\mathbb{R}$. A support restriction can matter.
Ignoring dependence in a sum. The theorem requires the added variable to be independent of each summand being compared. Equal noise marginals do not suffice.
Expecting strict order to remain strict. A nondecreasing map can collapse different values to the same output, and adding noise can make distinct distributions harder to distinguish. The theorem guarantees weak order, not strict inequality.

Key takeaway: A nondecreasing map preserves usual stochastic order; a nonincreasing map reverses it. Adding the same independent real-valued random variable preserves it. Quantile coupling turns each claim into a pointwise inequality, while monotonicity and independence are the conditions that make the argument work.

Check Your Understanding

Use the CDF or survival-function definition and attend to the hypotheses in each question.

If $X\leq_{\mathrm{st}}Y$, what order relation follows for $3X-4$ and $3Y-4$? What theorem justifies it?
If $X\leq_{\mathrm{st}}Y$, what relation follows for $-X$ and $-Y$? State the direction carefully.
Why does the addition result require independence of the noise from each variable, rather than only a common noise marginal distribution?
Can a nondecreasing transformation make two distinct distributions equal? Explain briefly.
Give a reason that $X\leq_{\mathrm{st}}Y$ alone does not guarantee $X^2\leq_{\mathrm{st}}Y^2$ for real-valued variables.