Can objectives have different units?

Yes. Strict ordering compares each objective with itself across alternatives rather than adding incompatible quantities, though any tolerances still need units.

Does the approach always select one decision?

No. Several decisions can have identical final objective vectors.

Is the result automatically fair?

No. Fairness depends on the objectives, constraints, and chosen priority policy.

AI Glossary / Mathematics and optimization / Intermediate

Lexicographic optimization

Definition

Lexicographic optimization handles several objectives in a fixed priority order. It optimizes the highest-priority objective first, then optimizes the next among solutions that preserve the earlier optimum, and continues in that way. Under strict lexicographic preferences, no improvement in a lower-priority objective can compensate for worsening a higher-priority one.

What it means

The word lexicographic comes from dictionary ordering: the first differing position decides the comparison. In optimization, those positions are objectives such as missed commitments, cost, and emissions, ordered according to an explicit policy.
This approach is useful when priorities are genuinely hierarchical. It avoids asking for a single exchange rate between every pair of objectives, but it requires a strong commitment about what takes precedence. It is not the same as saying that one objective is merely somewhat more important than another.

How it works

Define the feasible decisions, objective functions, optimization directions, and priority order. For exact minimization, solve the first objective and retain only its optimal set. Solve the second over that set, preserve its optimum as well, and continue. A finite candidate list can instead be compared directly using ordered objective vectors.
Practical software uses numerical tolerances and may allow explicit degradation of earlier objectives. Those settings create a relaxed hierarchy and must be interpreted using the solver's documented semantics. If an earlier stage stops before its optimum is established, the final result should not be described as an exact lexicographic optimum without qualification.

Three practical examples

A scheduling model can minimize missed commitments first, then cost, then another declared tie-breaker.
A service-planning exercise can preserve a required priority measure while seeking the most economical solution among tied alternatives.
A design study can compare strictly ordered objectives and explain how the selected design changes when that priority order changes.

Think of it this way

Think of sorting candidates by three columns in order: examine column one first, use column two only to break a tie, and use column three only if both earlier columns tie. The analogy is exact for a finite list, but continuous optimization also requires solving subproblems and handling numerical accuracy.

Worked example

Consider an original planning problem with four feasible plans. Each has an ordered vector (missed commitments, cost units, emission units), and all three objectives are minimized. Plan A has (0,120,8), B has (0,100,12), C has (1,20,2), and D has (0,100,9). The priority order is commitments first, cost second, and emissions third.
At the first stage, the smallest number of missed commitments is zero. Plans A, B, and D remain; C is excluded despite its low cost and emissions. At the second stage, the smallest cost among those remaining is 100, leaving B and D. At the third stage, D's emission value 9 is better than B's 12. Plan D is the strict lexicographic choice.
Compare D with A directly. Their first entries tie at zero. D's cost 100 is better than A's 120, so D wins at the second position even though A's emissions are lower. Once the first difference has decided the ordering, later entries cannot reverse it.
Now change the priority order to cost first, missed commitments second, and emissions third. Plan C wins immediately because its cost 20 is the smallest. The feasible plans have not changed; the preference policy has. This is why priority order belongs in the model's documentation rather than being treated as an implementation detail.
A weighted sum expresses a different policy unless its weights are chosen with sufficient mathematical safeguards. For example, the score 50 times missed commitments plus cost gives A=120, B=100, C=70, and D=100. It chooses C, contradicting the original strict hierarchy. A large-looking weight is not automatically large enough relative to possible lower-priority improvements.
Finally, suppose the policy deliberately permits up to one additional missed commitment after determining that the best count is zero. The later cost stage can then consider all four plans and choose C. That can be a legitimate relaxed policy, but it is not the strict ordering originally stated. This simple explicit constraint relaxation is illustrative; a software package's tolerance parameter may have more specialized semantics.

Illustrative case study

In a fictional dispatch project, Juniper Lane Services agrees to prioritize missed commitments, then operating cost, then an environmental metric. Its analyst presents the four-plan comparison and asks the team to approve the ordering as a policy choice.
The team records a strict initial hierarchy and separately evaluates a relaxed scenario. It reports both objective vectors rather than describing one plan as best without context. The case is illustrative and does not imply that this priority order is appropriate for every organization.

Common misconceptions and mistakes

Do not confuse lexicographic optimization with lexicographic max-min optimization, which compares sorted outcomes rather than a permanently ordered list of named objectives. Do not silently replace strict priorities with arbitrary weights.
An optimized soft objective is also not a hard feasibility condition: the best available solution can still have a positive violation count if zero is impossible. Numerical tolerances, incomplete solves, and grouped equal-priority objectives can change interpretation. Check these details before claiming an exact hierarchy was enforced.

Business use

Business value comes from making decision rules explicit and explainable. A useful report gives the objective order, direction, optimum or best-known value at each stage, permitted degradation, and termination status. Sensitivity analysis should include alternative priorities when stakeholders disagree. If lower-priority gains should sometimes justify higher-priority losses, a tradeoff model or explicit tolerance policy may describe the decision better than a strict hierarchy.

Check your understanding

Why does plan A lose to D even though A has lower emissions?

Answer: They tie on missed commitments, and D wins on the higher-priority cost objective. Emissions are considered only after those earlier objectives tie.

Related terms

Multi-objective optimization
Weighted sum
Hierarchical objective
Priority order
Pareto efficiency
Objective tolerance
Feasible set
Leximin ordering
Goal programming

Learn next

Add plan E=(0,100,8) and repeat the three stages. Then add plan F=(0,100,8) with different operational details. Explain why the objective hierarchy chooses an outcome vector but may still need a declared tie-breaker to choose one plan.

Sources

Primary implementation source: Gurobi Optimizer Reference Manual, Multiple Objectives, including hierarchical priorities and model-dependent degradation settings: https://docs.gurobi.com/projects/optimizer/en/current/features/multiobjective.html . The four-plan calculations and explicit relaxation example are original.

Previous: Lexicographic max-min optimization Next: Lexicon