Does first-order dominance require risk aversion?

No. It covers all nondecreasing expected-utility functions, subject to the necessary expectations.

Can two options be incomparable?

Yes. Crossing cumulative curves can prevent first-order ordering.

Does second-order dominance mean equal means are always required?

No. The equal-mean example is a useful special case, not the general definition.

AI Glossary / Graphs and symbolic AI / Intermediate

Stochastic dominance

Definition

Stochastic dominance is a way to compare probability distributions of outcomes under a stated class of preferences. First-order stochastic dominance identifies a distribution that is at least as desirable for every decision maker whose utility is nondecreasing in the outcome, assuming the relevant expectations exist. It compares the whole distribution, not just its average.

What it means

Suppose larger outcomes are better. A first-order dominant option gives no greater probability of falling at or below any chosen threshold. This is stronger than saying that its mean is larger.
Different orders of dominance correspond to different preference restrictions. Second-order dominance considers nondecreasing concave utility functions, which represent a preference for more together with risk aversion in the expected-utility framework. The word “dominance” therefore needs its order and outcome convention. For costs or delays, where smaller values are preferred, the direction of the comparison must be handled accordingly.

How it works

For first-order dominance, compare cumulative distribution functions. If F_A(x) is the probability that A's outcome is at most x, then A weakly first-order dominates B when F_A(x) is no greater than F_B(x) for every threshold x.
For discrete distributions, inspect the intervals created by all possible outcome values, including the jumps. Crossing cumulative curves generally prevent first-order ordering. Second-order comparisons require a different criterion or a direct argument for the relevant concave-utility class; simply comparing variances is insufficient. Also distinguish a theoretical distribution from an estimate based on limited data, where apparent dominance may be uncertain.

Three practical examples

An AI planning experiment can compare distributions of achieved reward rather than relying only on average reward.
A simulated service-design study can compare completion-time distributions, taking care that lower times are preferable.
A decision-support lesson can show when a ranking holds across a broad class of preferences and when specific tradeoffs must be supplied.

Think of it this way

Think of checking two options against every possible performance threshold rather than selecting one convenient cutoff. An option that meets the relevant comparison at every threshold supports a stronger statement. The analogy is limited because stochastic dominance concerns probability distributions and preference assumptions, not a promise that every individual observed result will be better.

Worked example

Consider original reward distributions A and B, where higher scores are preferred. A gives 2 or 6 with equal probability. B gives 1 or 5 with equal probability.
Below 1, both cumulative probabilities are zero. From 1 up to but excluding 2, A's cumulative probability is zero while B's is one half. From 2 up to but excluding 5, both are one half. From 5 up to but excluding 6, A's is one half and B's is one. At 6 and above, both are one.
Thus F_A(x) never exceeds F_B(x), so A first-order dominates B. Their means are 4 and 3 respectively, consistent with the dominance result. But the full threshold comparison, rather than the mean difference alone, establishes dominance.
Now draw A and B independently once. There is probability one quarter that A yields 2 while B yields 5. In that paired observation A is lower. Dominance is a statement about their distributions; it does not imply that every independently paired realization has the same ordering.
Next compare A with C, which yields 0 or 10 with equal probability. C's mean is 5, greater than A's mean of 4. However, at threshold 1, F_C is one half and F_A is zero, while at threshold 7, F_C is one half and F_A is one. The curves cross. Neither first-order dominates the other, despite C's higher mean.
Finally, compare a certain reward of 5 with the equal-probability reward 0 or 10. Both have mean 5. For every concave utility function u, concavity gives u(5) at least as large as the average of u(0) and u(10). The certain reward is therefore weakly preferred throughout the nondecreasing concave-utility class. This is a simple second-order dominance example, even though the cumulative curves cross and first-order dominance does not hold.

Illustrative case study

In a fictional simulation study, a team compares two planning policies across a defined set of randomized scenarios. One policy has a higher mean reward, but the plotted cumulative distributions cross.
The team avoids labeling it universally preferable and instead examines the outcome ranges stakeholders care about. It also checks uncertainty in the estimated curves before drawing conclusions. This fictional case illustrates distribution-aware evaluation, not evidence that a real policy dominates another.

Common misconceptions and mistakes

A higher average does not establish first-order dominance, and lower variance alone does not establish second-order dominance. A dominant distribution also need not win every independent paired draw.
Do not leave the preference direction implicit when comparing costs. Do not treat a small empirical sample as a known population distribution. Finally, dominance may leave many options incomparable. That is a meaningful result: choosing between them can require more specific preferences, constraints, or information rather than a forced universal ranking.

Business use

The concept is useful when teams want a comparison that is less dependent on one particular utility function. It can clarify when an option has broad distributional advantages and when the choice rests on stakeholder tradeoffs.
Define the outcome, units, scenario population, and dominance order before analysis. Keep other decision dimensions visible: a reward distribution alone may omit implementation cost or feasibility. Report crossing curves and uncertainty honestly. A decision-support system should explain the assumptions supporting a ranking rather than presenting the word dominant as an unconditional endorsement.

Check your understanding

Why does C's mean of five fail to prove that it first-order dominates A?

Answer: C has more probability at very low outcomes, while it also offers a larger upper outcome. The cumulative curves cross.

Related terms

Expected utility
Cumulative distribution function
First-order dominance
Second-order dominance
Risk aversion
Mean-preserving spread
Distributional evaluation

Learn next

Next, add one to every outcome of B. Its distribution becomes exactly A. Explain why this particular transformation provides an intuitive coupling with an improvement on every matched outcome, while independent draws still need not preserve that ordering.

Sources

Primary source: MIT OpenCourseWare, Microeconomic Theory III, Chapter 4, Stochastic Dominance: https://ocw.mit.edu/courses/14-123-microeconomic-theory-iii-spring-2015/875150f8deb05a910756a02a4f9f78a5_MIT14_123S15_Chap4.pdf . The reward distributions, threshold comparisons, and calculations are original.

Previous: Stochastic block model Next: Stochastic drift