The gold standard of optimisation is finding a demonstrably good feasible solution to a clearly stated problem, then explaining what the result does and does not establish. “Best” needs an objective. “Possible” needs constraints. A trustworthy answer needs both, together with evidence about the method and assumptions.
Optimisation connects algebra, geometry, calculus, numerical methods, algorithms and decision making. It appears in design, scheduling, model fitting and resource allocation. This guide brings those relationships together and links the deeper eduKate mathematical articles. The Operations Research guide develops the wider application of analytical methods to operating decisions; this guide concentrates on optimisation itself.
Our running example is a fictional rectangular learning garden with twenty metres of boundary available. Maximising its area is a small problem that can be solved and checked exactly. Adding an access requirement changes the answer, showing why an elegant unconstrained optimum may be unsuitable for the actual task.
Find your route
Start with the objective and feasible set, follow the garden calculation, or choose a deeper mathematical method in the directory.
Full chapter index · Related eduKate articles · Learning that transfers into life
Chapter index
Define best and feasible · Chapters 1–3
Choose the mathematical structure · Chapters 4–7
Judge the computation · Chapters 8–10
Work through examples · Chapters 11–14
Apply and learn · Chapters 15–16
Optimisation compares allowable alternatives according to a defined criterion. Minimising completion time, maximising usable area and fitting a model to data are different problems. The method cannot decide which goal matters without a specification.
A lower number is not automatically better. A smaller prediction error can be achieved by a model that performs poorly on new data. A shorter schedule can omit necessary recovery time. The objective must represent the intended task.
Write the criterion in ordinary language before introducing notation. Identify the quantity, unit, direction of improvement and relevant horizon. “Maximise learning” is too broad unless a defensible measurement and decision context are supplied.
The garden exercise maximises the area of an ideal rectangle. It does not maximise educational value, accessibility or ecological quality. Those wider aims would need additional conditions or a different model.
An optimum can therefore be mathematically correct while answering an incomplete question. This is a limitation of the specification rather than a reason to distrust mathematics.
The useful first check is to ask what an exceptionally high objective score could conceal. If a plan could score well while violating the actual purpose, revise the model before searching harder for its optimum.
A feasible set contains decisions satisfying the model’s constraints. Constraints can describe resource limits, geometry, minimum requirements, logical relationships or variable domains.
For a rectangle with side lengths L and W and twenty metres of boundary, 2L + 2W = 20. Positive dimensions are needed for a nondegenerate rectangle. The model assumes the whole boundary length is available for the rectangle.
If one side must be at least six metres for the intended layout, add L ≥ 6. That requirement changes the feasible set. It should not be treated as an optional preference merely because the unconstrained area optimum is attractive.
Check whether each constraint is genuinely required. An unnecessary restriction can remove useful alternatives, while an omitted requirement can produce an unusable solution. The decision owner should recognise the meaning of every important condition.
Variable domains matter too. A length may be continuous in a teaching model, while available construction modules impose discrete choices in practice. The model should state which interpretation applies.
A strong explanation distinguishes the real requirement from its mathematical representation. If the representation is approximate, say how that approximation affects the conclusion. Feasibility is the foundation of optimisation, not a final administrative check after the search is complete.
A mathematical expression needs consistent quantities. In the garden, L and W are measured in metres, perimeter in metres and area in square metres. Adding an area directly to a length without a justified scaling would create an unclear objective.
Scaling can affect numerical computation even when it does not change the underlying mathematical problem. A model combining very large and very small coefficients may require appropriate treatment by the numerical method.
Normalisation can make quantities comparable, but it introduces choices. Dividing each measure by a baseline changes how trade-offs are expressed. The baseline should be explained rather than treated as neutral.
Before solving, calculate one simple feasible example. A rectangle with L = 6 and W = 4 uses twenty metres of boundary and has area 24 m². This checks the relationship between the verbal task and the formula.
After solving, estimate the answer’s scale. An area of 250 m² from the same simple twenty-metre rectangular boundary model should prompt a check.
A data dictionary helps: variable, meaning, unit, bounds and source. It is useful in a school solution and essential in a larger model. Another reader should be able to reconstruct what the symbols represent without guessing from the final number.
Linear optimisation has linear objectives and constraints. Nonlinear optimisation includes nonlinear relationships. This distinction affects the geometry, available methods and kinds of guarantees a result may support.
The garden objective A = LW is nonlinear in the two dimensions. Using W = 10 − L converts it to A(L) = 10L − L² over the relevant interval. The simpler one-variable expression exposes its shape.
Not every nonlinearity makes a problem intractable. Some nonlinear structures have strong mathematical properties and efficient methods. Conversely, an apparently simple expression can have several local optima or difficult constraints.
A solver should match the problem structure. A tool designed for linear relationships cannot be assumed to handle an arbitrary nonlinear formula merely because the formula was entered into a spreadsheet.
Google’s official OR-Tools introduction distinguishes several optimisation problem classes. Its role here is an implementation reference, not a validation of any particular real-world model.
For learners, classify the relationships before choosing the technique. Can the problem be simplified algebraically? Is the domain finite enough to enumerate? Does calculus apply? The choice should follow the mathematics and required evidence, rather than the attractiveness of a software demonstration.
A local optimum is at least as good as nearby feasible alternatives under the relevant definition. A global optimum is at least as good as every feasible alternative. A method finding one does not automatically establish the other.
Imagine a performance curve with two separated peaks of different heights. Starting near the lower peak, a method that improves through small local steps may stop there. Its stopping condition does not prove the higher peak is absent.
The garden objective is particularly friendly: its concave quadratic shape permits a direct global argument over a convex interval. That property should be stated rather than generalised to every nonlinear task.
A zero derivative is a candidate condition in appropriate smooth interior problems, not a universal certificate of maximum value. The point may be a minimum or another stationary point, and boundaries still need examination.
For a finite decision set, exhaustive comparison can establish a global result if every feasible choice is covered. For larger problems, mathematical bounds and structural properties can provide other forms of evidence.
The gold standard is precise wording. “A feasible solution was found,” “a local optimum was found” and “global optimality was proved” report different achievements. Each can be useful when it matches the decision’s needs and is not presented as something stronger.
A convex feasible set contains the line segment joining any two of its points. In convex minimisation, the objective is convex and the feasible region is convex. Under that structure, a local minimum is also global.
For maximisation, the corresponding favourable structure is a concave objective over a convex feasible region. The garden’s area function after substitution is concave, and its allowed lengths form an interval.
These properties do not remove the need to verify the model or interpret the result. They establish mathematical relationships within the specified problem.
Do not infer convexity from one smooth-looking plot. A sampled display can miss important features. A formal argument or suitable structural rules provide stronger evidence.
The existing Convex Analysis guide connects these ideas with deeper mathematics. Boyd and Vandenberghe’s Convex Optimization provides an authoritative open reference.
For a beginner, the practical payoff is understanding why some methods can offer strong guarantees while others need more cautious conclusions. The useful question is not merely whether the graph looks simple, but which property justifies the claimed optimum and whether the actual problem has that property.
Some choices are indivisible: assigning a room, selecting a project or choosing a whole number of components. A continuous relaxation may provide useful information while allowing decisions that cannot be implemented.
Rounding can violate constraints or produce a poor feasible choice. Its effect depends on the model. A successful rounding example does not establish a general rule.
For the garden, suppose side lengths must be whole metres. The unconstrained optimum of five by five remains allowed, so the discrete requirement happens not to change the best area. A different module size or minimum dimension could change the result.
When the feasible set is small, list the allowed decisions and compare them. This is an excellent learning method because it provides a transparent independent check.
In larger integer problems, appropriate algorithms may return a feasible solution with a bound on the unresolved optimum. Explain that distinction instead of reporting a search result as automatically proved optimal.
The Operations Research guide develops an integer allocation example with resource constraints. It complements this mathematical root by showing how discrete choices interact with an operating system.
The standard remains feasibility first, then quality, then evidence for the claimed guarantee. A plan that cannot be executed is not rescued by a high objective score.
A numerical algorithm produces approximations through a defined procedure. Its stopping rule might concern a gradient, change in objective, feasibility tolerance, computational limit or a bound on optimality. These criteria have different meanings.
A tiny change between iterations may reflect convergence, a flat region, poor scaling or an unsuitable step size. It is not a universal proof that no better solution exists.
Check constraint violations in the units that matter. A small relative residual may still be operationally important if the quantity has a strict requirement. Conversely, numerical rounding need not imply a material practical failure.
Record the solver, relevant settings and result status when they affect reproducibility. The correct level of detail depends on the audience, but another analyst should be able to understand how the result was obtained.
Use simple cases to verify implementation. Can the code solve the garden problem whose answer is known exactly? Does it respond sensibly when the minimum-length constraint changes?
The existing Numerical Analysis and Computational Mathematics guides provide deeper routes. Their numerical questions differ from the decision question of what should be optimised.
The gold standard report translates the computational status into an honest statement: what was found, how accurately constraints were satisfied and which guarantee remains unresolved. A green completion message alone is insufficient evidence.
An optimum depends on the parameters and assumptions supplied. Sensitivity analysis examines how the solution or objective responds when those inputs change.
For the garden, changing available boundary length changes the best dimensions. Adding a minimum dimension changes the feasible interval. A construction requirement can matter more than another decimal place in the calculus.
Separate uncertainty from a deliberate scenario. A planner may compare several possible resource levels, but should not imply that all are equally likely unless a probability model supports that claim.
Also distinguish sensitivity of the objective from sensitivity of the decision. Several very different decisions can have nearly equal objective values. In that situation, practical considerations may favour a simpler or more robust option.
A useful report identifies which inputs could reverse the recommendation. It then asks whether those inputs are known well enough for the intended decision.
The Probability and Statistics guides support uncertainty interpretation; the optimisation model compares choices under that information. Keeping the roles distinct prevents an estimated coefficient from silently becoming a guarantee.
For a learner, change one assumption and predict the direction of the effect before calculating. The explanation of that change often reveals more understanding than reproducing the original optimum.
Many real decisions involve several goals: time, resource use, accessibility, reliability or environmental effects. One solution may improve one goal while worsening another.
A solution is dominated if another feasible solution is no worse on all relevant objectives and strictly better on at least one. Non-dominated solutions form a useful comparison set, but choosing among them still requires judgment.
Consider fictional route plans measured by staff time and travel time, both to be minimised. Plan A uses two staff-hours and thirty travel minutes. Plan B uses three staff-hours and twenty travel minutes. Neither dominates the other.
Plan C uses four staff-hours and forty travel minutes. A and B both dominate C under these two measures. Removing C is justified within the stated comparison; choosing A or B requires a trade-off.
Weighted sums can express a preference, but the units and weights need meaning. Adding hours and minutes without a stated conversion of priorities hides the choice.
An objective can also omit a relevant value. If access or safety requirements matter, include them appropriately rather than expecting a performance score to represent them automatically.
The practical gold standard is a transparent comparison. Mathematics can identify inefficient alternatives and make trade-offs visible without pretending that every human value reduces to one self-evident number.
The fictional rectangle has boundary length twenty metres. Let L and W be positive side lengths. Then L + W = 10 and area A = LW = L(10 − L).
Complete the square: A = 25 − (L − 5)². The squared term is nonnegative, so A ≤ 25. Equality occurs when L = 5, giving W = 5.
This establishes a global maximum of 25 m² for the stated rectangular model. It does not merely identify a candidate point. The expression shows why every other permitted length has no greater area.
A calculus check gives A′(L) = 10 − 2L, which is zero at L = 5. The second derivative is −2, consistent with concavity. The algebraic proof already supplies a clear global argument.
Test a nearby feasible rectangle: L = 6, W = 4 gives 24 m². L = 7, W = 3 gives 21 m². The comparisons fit the proven shape.
State the model’s boundaries. The result assumes a rectangle and full use of the stated boundary. It does not compare every possible geometric shape or include gates, paths, irregular ground or construction allowances.
The lesson is that a good solution combines a result with a reason no better feasible alternative exists. That reason can be algebraic, geometric or computational, provided its scope matches the claim.
| Length L | Width W | Area |
|---|---|---|
| 5 m | 5 m | 25 m² |
| 6 m | 4 m | 24 m² |
| 7 m | 3 m | 21 m² |
Now require L ≥ 6 metres for the fictional layout. Retain the twenty-metre rectangular boundary and positive width. Since W = 10 − L, the feasible interval becomes 6 ≤ L < 10.
The unconstrained optimum L = 5 is outside that interval. It cannot be selected even though its area is larger.
Using A = 25 − (L − 5)², every feasible L is at least one metre above five. The squared term is therefore at least one, giving A ≤ 24. Equality occurs at L = 6 and W = 4.
The constrained maximum is 24 m². The additional requirement reduces the best area by one square metre, or 4% of the original maximum.
This is a result about the specified requirement. It does not establish whether the requirement itself is necessary or desirable. That question belongs to the real purpose and appropriate decision process.
A calculus explanation reaches the same conclusion: A′(L) is negative throughout L ≥ 6 within the feasible interval, so the area decreases as L increases. The lower boundary is best.
A student who stops after finding a zero derivative would miss the answer. Boundary checks and feasibility are part of the solution, not optional additions after the calculation.
| Model | Best dimensions | Maximum area |
|---|---|---|
| Positive rectangular sides, perimeter 20 m | 5 m × 5 m | 25 m² |
| Also require L ≥ 6 m | 6 m × 4 m | 24 m² |
Consider fictional measurements at three input values: (0,1), (1,2) and (2,2). Suppose a simple model predicts the same constant c for every input.
The sum of squared residuals is S(c) = (1 − c)² + (2 − c)² + (2 − c)². Differentiating gives S′(c) = 6c − 10, so the best constant is c = 5/3.
At that value, the residuals are −2/3, 1/3 and 1/3. Their squared sum is 4/9 + 1/9 + 1/9 = 2/3. The second derivative is six, confirming the unique minimum.
This small calculation illustrates optimisation in model fitting. The chosen loss determines what best means. Minimising absolute residuals would pose a different objective and can produce a different summary.
A more flexible model could fit the three observations more closely. That alone would not establish better performance on new data. Model purpose, complexity and validation matter.
The numbers are invented and have no empirical claim attached. Their value is to show how an optimisation objective connects a candidate model with observed discrepancies.
The broader lesson is to distinguish fit from usefulness. A solver can minimise the selected loss perfectly while the loss or model family poorly represents the task the user actually needs.
| Observed value | Residual | Squared residual |
|---|---|---|
| 1 | −2/3 | 4/9 |
| 2 | 1/3 | 1/9 |
| 2 | 1/3 | 1/9 |
| Total | 2/3 |
Recalculate the objective and important constraints from the proposed decision. This catches interpretation mistakes, incorrect units and output-processing errors that a successful solver run may not reveal.
Use a different representation when it is informative. Check a calculus result algebraically, compare a small integer solution by enumeration or test a geometric claim with boundary cases.
Verification should address the meaningful failure modes. Repeating the identical calculation through the same mistaken interpretation offers limited reassurance. Independence of reasoning matters more than the number of repeated checks.
Distinguish verifying the mathematical solution from validating the model’s relevance. The garden proof can be correct while the design omits a required path. Both questions need attention.
For numerical problems, inspect result status, tolerances and bounds. For an application, trace the output into the actual workflow and confirm that the constraints retain their intended meaning.
The Verification, Judgment and Operations Research guides provide connected practical routes. This subject guide explains the mathematical quality of the optimum; those guides help connect it with reliable use.
A strong final sentence states the solution, domain and guarantee together. “Five by five metres globally maximises area among ideal rectangles with twenty metres of boundary” is more informative than “the best answer is five.”
Begin with a problem whose alternatives can be understood and checked. Ask the learner to identify the decision variables, objective and constraints before choosing a method.
Use a sequence of representations: a verbal task, a sketch, an equation, a graph and a final interpretation. The relationships between those representations are part of the learning.
After the first solution, change one condition. The garden’s minimum length is a useful example because it forces the learner to revisit feasibility and boundaries rather than repeat differentiation mechanically.
Ask for a proof or an appropriate bound, not only a candidate answer. At an introductory level, completing the square can make the guarantee clearer than a more advanced method used without explanation.
Then introduce a different optimisation purpose, such as fitting a constant to observations. The learner should recognise the common structure while noticing that the variables and objective now represent different things.
The wider education-transfer guide explains why this matters. Learning becomes usable when a student can choose and adapt a method as the question changes.
A useful correction targets the exact error: an omitted constraint, a mistaken domain, an invalid rounding step or an unjustified claim of globality. More algebra will not repair every one of those difficulties.
Does optimisation always use calculus?
No. Algebra, geometry, enumeration, linear programming, integer methods and many other techniques can be appropriate. Choose according to the structure and required guarantee.
Is a stationary point always an optimum?
No. A zero derivative is a candidate under suitable conditions. Its nature, boundaries and feasibility still need examination.
Does global optimum mean the real decision is solved?
It establishes the best result for the stated model. The model’s objectives, data and constraints must also fit the intended use.
Can two solutions be equally optimal?
Yes. Some problems have several or infinitely many optima. Practical criteria may help choose among them without changing the original optimal objective value.
What should a beginner write?
State variables and units, objective, constraints and domain. Show the solution method, verify feasibility and explain the guarantee. Finish with a limitation and one changed-condition check.
Choose one completed optimisation problem and ask where its definition of best came from. That question connects mathematical technique with purposeful reasoning and leads naturally to the specialist guides below.
Continue through the eduKate subject guides
Use the mathematical map for deeper methods, Operations Research for applied decisions and the evidence guides for verification and interpretation.
Mathematical subject map and methods
Models, applications and decision evidence
- How Mathematical Optimisation Works | Objectives, Constraints, Trade-Offs, Feasible Regions and Best Decisions
- How Engineering Decision Analysis Works | Choosing Between Designs When No Option Wins Everything
- How Mathematical Modelling Works | From Reality and Assumptions to Equations, Validation, Prediction and Decisions
- How Optimisation Works | How Objectives, Constraints and Trade-Offs Shape Better Solutions
- The Gold Standard of Operations Research
- The Gold Standard of Probability
- The Gold Standard Of Judgment
Related subject guides
- The Gold Standard of Organic Chemistry: molecular structure and chemical evidence
- The Gold Standard of Palaeontology: ancient life, geological context and time
- The Gold Standard of Physical Geography: water, landforms and environmental processes
Sources and scope
Boyd and Vandenberghe’s open convex-optimisation text supports the stated convex and concave optimality relationships. Google’s OR-Tools documentation supports the distinction between optimisation problem classes. The fictional garden, route comparison and constant-fit calculation are original teaching examples.
This framework concerns mathematical and decision quality. It is not a guarantee that a model includes every relevant practical requirement or that a numerical method has proved global optimality unless that evidence is explicitly supplied.