Does dominance require being strictly better on every objective?

No. Under the convention used here, being no worse on all and strictly better on at least one is enough.

Can two alternatives be incomparable?

Yes. One may improve one objective while worsening another.

Can a dominated option become relevant later?

Yes. Changed constraints, new objectives, revised measurements or unavailable alternatives can change the feasible comparison.

AI Glossary / Graphs and symbolic AI / Intermediate

Pareto dominated

Definition

An option is Pareto dominated when another feasible option is at least as good on every specified objective and strictly better on at least one. The comparison respects whether each objective should be minimized or maximized.

What it means

Many decisions involve several objectives, such as reducing cost while improving accuracy. Pareto dominance identifies a comparison that does not require choosing an exchange rate between those objectives: one option offers an improvement without worsening any of the others.

For example, if two feasible systems have the same measured accuracy but one costs less, the more expensive system is dominated on those two objectives. If the cheaper system is also less accurate, neither automatically dominates the other.

The qualification “specified objectives” matters. An omitted requirement, uncertain measurement or difference in feasibility can change the conclusion. Dominance is a precise relationship within a stated decision model, not a declaration that an option is inferior in every possible respect.

How it works

List the feasible alternatives and the objectives used to compare them. Specify direction: lower latency may be preferable, while higher accuracy is preferable. Compare one alternative against another objective by objective.

For all-minimization objectives, B dominates A when every B value is no greater than A's corresponding value and at least one is smaller. For mixed directions, apply the appropriate inequality separately or transform the objectives consistently.

If each alternative is better on a different objective, they are incomparable under this dominance relation. Identical objective vectors do not dominate one another under the convention used here because neither has a strict improvement.

A filtering procedure removes dominated alternatives and retains a nondominated set. That set is relative to the alternatives examined. Missing feasible alternatives can still dominate something in the retained set.

Three practical examples

A team removes a model configuration that is both slower and less accurate than another feasible configuration on the same benchmark.
A planner excludes a route that consumes more energy and takes longer than an available route satisfying the same constraints.
A product group compares designs using cost and measured reliability, retaining options that offer genuine tradeoffs.

Think of it this way

Think of choosing between two identical-size boxes for a shipment. If one meets the same requirements and is both lighter and cheaper, those two criteria give no reason to prefer the other.

The analogy stops when the boxes differ in an unmeasured quality such as durability. Likewise, a dominance claim can be correct for recorded objectives yet inadequate for the real decision if the model omits something important.

Worked example

Consider four invented model configurations evaluated under identical conditions. Their pairs are (latency in milliseconds, accuracy percentage): A=(40,90), B=(30,92), C=(20,88), and D=(30,92). Lower latency and higher accuracy are preferred; assume all four meet the same feasibility requirements.

B dominates A because 30 is less than 40 and 92 exceeds 90. D also dominates A. B and D have identical objective values, so neither strictly improves on the other and neither dominates the other under the stated convention.

C is faster than B but less accurate. Therefore C and B are incomparable. After removing A, the nondominated configurations are B, C and D, representing two distinct objective vectors.

If B and D violate a required memory limit while A and C satisfy it, they cannot serve as feasible dominators in the constrained problem. The numbers are illustrative point values, not measured evidence about real systems.

Illustrative case study

The fictional Willow Models team creates a cost-and-accuracy comparison and removes several dominated configurations. Before choosing a remaining option, reviewers notice that one configuration's accuracy was measured on a different dataset.

Willow reruns comparable evaluations and records the uncertainty around the estimates. It also checks mandatory operating constraints before applying the dominance filter.

The fictional exercise shows why a clear inequality is not enough by itself. Comparable evidence, complete objectives and feasible alternatives are needed for the mathematical relationship to support a useful decision.

Common misconceptions and mistakes

Lower cost alone means dominance. The alternative must also be no worse on every other specified objective.
Equal results imply dominance. The convention here requires a strict improvement somewhere.
A nondominated option is the unique best choice. Several incomparable tradeoffs can remain.
A better average score proves dominance. An aggregate can hide deterioration on an individual objective.
Dominance survives any omitted criterion. Adding a relevant objective can change the relationship.

Business use

Dominance filtering can simplify discussions before stakeholders debate tradeoffs. It separates alternatives that are worse across the stated comparison from those requiring an actual preference decision.

Document objective definitions, measurement conditions, feasibility rules and the handling of uncertainty. A tiny difference between noisy benchmark estimates should not automatically be presented as a confidently established improvement. Teams may use uncertainty-aware comparisons or practical tolerances, but these should be named because they modify the decision rule.

After filtering, select among remaining alternatives using explicit priorities or constraints. Pareto dominance does not determine how much accuracy is worth a particular cost increase.

Check your understanding

Why does C=(20 ms,88% accuracy) fail to dominate B=(30 ms,92% accuracy)?

Answer: C improves latency but worsens accuracy. Since it is not at least as good on every objective, it does not dominate B; B does not dominate C either.

Related terms

Pareto optimal
Objective function
Pareto frontier
multiobjective optimization
feasible solution
tradeoff
nondominated set
scalarization

Learn next

Create a table of five alternatives with two objectives and explicit feasibility constraints. Mark every pairwise dominance relationship and retain the nondominated alternatives. Then add a third objective and check which conclusions remain valid.

Sources

Boyd and Vandenberghe discuss vector optimization and Pareto optimality: https://web.stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf

BYU's engineering design optimization textbook search excerpt corroborates the nondominance distinction; the full file was unavailable during review: https://flowlab.groups.et.byu.net/mdobook.pdf

The latency-and-accuracy table and Willow Models case are original examples. Terminology for weak and strict dominance varies; this entry states its inequality convention explicitly.

Previous: Pareto analysis Next: Pareto optimal