Mathematical optimisation is the study of choosing the best feasible option according to an explicitly defined objective. It begins when several answers are possible, resources are limited, and “better” needs a meaning precise enough to calculate and defend.
A calculation can tell us the cost of one timetable. Optimisation asks which timetable should be chosen. A formula can describe the energy used by one journey. Optimisation compares allowable journeys. An equation can fit a line through observations. Optimisation explains what “best fit” means when no line passes through every point.
The crucial word is feasible. An arrangement that requires unavailable teachers, impossible travel times or negative quantities is not a better solution. It is outside the problem. The other crucial phrase is according to an objective. The cheapest arrangement is not automatically the fairest, safest or most educationally valuable arrangement. Mathematics can make a choice precise without deciding every human value that belongs in it.
This guide develops optimisation from an understandable planning example into constraints, proof of optimality, gradients, convexity, duality, integer decisions and decisions under uncertainty. The examples are constructed teaching models, not measured claims about eduKate classes, learning gains or operational performance.
Read through the mathematics: Mathematics Learning Hub → How Mathematics Works → Mathematical Optimisation. Helpful foundations are Algebra, Linear Algebra and Calculus.
Reading map: Define the decision · Build and solve a model · Prove the answer · Understand convexity · Choose an algorithm · Handle uncertainty · Verify and interpret · Questions and learning checks.
1. The decision must be defined before the method
An optimisation problem has decision variables, an objective and a feasible set. In a minimisation problem, we seek a feasible vector x whose objective f(x) is no larger than the objective at any other feasible vector. The vector may contain one number, thousands of allocations or an entire sequence of future actions. Maximisation reverses the comparison. The formal framework is developed in Boyd and Vandenberghe’s Convex Optimization.
Suppose a learning centre is planning a workshop day. “Improve the workshop” is not yet an optimisation problem. Do we choose session lengths, the number of groups, room assignments, the sequence of activities or the students attending each session? Do we minimise waiting, maximise coverage, balance teacher workload or protect a minimum amount of individual feedback? These are different decisions even when they concern the same day.
A useful first sentence is: “Choose these variables, within these limits, to improve this quantity.” Writing that sentence exposes ambiguity before it becomes an elaborate calculation. The sentence should also identify who has authority over each variable. A planner can allocate available hours; the planner cannot simply choose an arbitrary level of student understanding as though understanding were an adjustable switch.
2. Variables are choices; parameters describe the situation
Let x represent hours allocated to one activity. If the planner can select x, it is a decision variable. Let T represent the available time. If T is fixed for this planning run, it is a parameter. A model may later be rerun with a larger T, but changing an input between runs is different from optimising it within a run.
This distinction prevents hidden promises. A model that chooses both lesson quality and cost freely can appear to deliver everything because it has omitted the relationship between them. A model that chooses delivery dates without travel constraints has not solved scheduling. It has written wishes as variables.
The domain also belongs to the variable definition. Hours may be divisible. People generally are not. A room-selection variable might be binary: 1 means use the room and 0 means do not use it. A proportion lies between 0 and 1. A temperature can be negative on some measurement scales, whereas a count of supplied chairs cannot. The permissible values are not small implementation details; they determine what sort of problem we are solving.
A good model therefore records the meaning, unit and domain of every variable alongside its symbol. This makes the mathematics readable to someone other than the person who wrote it.
3. Constraints turn wishes into an admissible region
A constraint states a condition that a proposed solution must satisfy. A capacity limit may be an inequality such as x+y≤6. A balance condition may be an equality such as incoming quantity minus outgoing quantity equals the change in stock. A logical condition can state that an activity is allowed only when its prerequisite has been completed.
The intersection of all these conditions is the feasible set. Satisfying each constraint separately in different plans does not help: one plan must satisfy them together. This is why locally sensible arrangements can become globally impossible. Each teacher may have an acceptable schedule in isolation, yet the combined schedules may require two classes to use the same room simultaneously.
Consider the constraints x≥4 and x≤3. There is no feasible x. An optimiser should not return a compromise such as 3.5 and label the problem solved. It should expose infeasibility. Resolving that conflict requires changing the problem, for example by adding capacity or revising a requirement. The mathematical result may therefore be an explanation of impossibility rather than a recommended schedule.
4. A complete small model: allocating workshop time
Imagine two workshop activities, A and B. Let x and y be their allocated hours. Activity A uses two preparation units per teaching hour; B uses one. There are six teaching hours and eight preparation units available. At least one hour must go to each activity. For this demonstration only, assign planning scores of three points per hour of A and two points per hour of B.
The score is a hypothetical comparison device. It is not a claim that teaching outcomes actually grow linearly or that activity A produces a measured fifty per cent improvement over B. Keeping that boundary visible lets us use the model to learn optimisation without disguising assumptions as evidence.
Maximise S = 3x + 2y subject to x + y ≤ 6 teaching-hour limit 2x + y ≤ 8 preparation-unit limit x ≥ 1, y ≥ 1 minimum coverage
The candidate x=2, y=4 uses all six teaching hours and all eight preparation units. Its score is 14. Another candidate, x=1, y=5, also uses six teaching hours, but its score is 13. Choosing the higher-scoring activity for every hour would violate the coverage and preparation limits. The resource interactions, rather than one activity’s individual score, control the answer.
5. Looking at the feasible corners
Because the example has two variables, its feasible set can be drawn in a plane. Each linear inequality selects one side of a line. The combined feasible set is a polygon. Its corners provide a compact set of candidates in this bounded linear example.
| Allocation (x, y) | Teaching hours | Preparation units | Score |
|---|---|---|---|
| (1, 1) | 2 | 3 | 5 |
| (1, 5) | 6 | 7 | 13 |
| (2, 4) | 6 | 8 | 14 |
| (3.5, 1) | 4.5 | 8 | 12.5 |
The half-hour allocation in the last row is allowed because x and y currently represent divisible hours. If activities had to occupy whole-hour slots, we would add integrality requirements. Fortunately, the best candidate here already uses whole numbers, and the proof in the next section will cover both versions.
Drawing corners helps us discover the candidate, but discovery and verification are separate jobs. For a larger model, a solver may discover the answer without an understandable picture. We still need a way to explain why the proposed allocation is genuinely best rather than simply better than a few alternatives we happened to inspect.
6. A proof of optimality is stronger than a good-looking answer
For every feasible allocation in the workshop model, the objective can be rewritten as 3x+2y=(x+y)+(2x+y). The first bracket is at most 6 and the second at most 8. Therefore every feasible allocation has a score no greater than 14.
We have already found a feasible allocation with score 14. It reaches a bound that no feasible allocation can exceed. That proves optimality. The argument does not depend on checking every possible pair of real numbers, and it does not depend on trusting a computer’s claim.
This is a small example of an optimality certificate. It contains two complementary objects: a feasible construction and a bound. One shows that the score can be achieved; the other shows that nothing better is possible under the model. Their agreement closes the argument.
Notice what the certificate does not prove. It does not prove that the scores represent actual learning. It does not prove that the eight preparation units are estimated accurately. It proves a precise statement inside the declared assumptions. That is exactly its value: the mathematical guarantee has a clear boundary, so it can be inspected rather than inflated.
7. Optimisation is not the same as prediction
A predictive model asks what is likely to happen. An optimisation model asks what should be selected under its objective and constraints. Predictions may supply inputs to an optimiser, but the tasks remain different.
For example, a forecast might estimate attendance at each workshop. The optimisation then chooses room assignments using those estimates. If attendance is underestimated, the mathematical schedule can be optimal for the forecast and unsuitable for the actual crowd. Improving the optimisation algorithm would not necessarily repair that error. The prediction or uncertainty treatment needs attention.
The same distinction applies to learning plans. A diagnostic estimate of a student’s weak topics is not itself a study allocation. Choosing an allocation requires priorities, available time and beliefs about what practice changes. A beautifully solved allocation based on an inaccurate diagnosis remains vulnerable.
Keeping prediction and optimisation separate allows error to be located. Was the forecast wrong? Was the objective wrong? Were constraints missing? Was the algorithm inaccurate? These questions identify repairable mechanisms instead of treating every disappointing outcome as a mysterious failure of mathematics.
8. “Best” can hide a value judgement
Suppose Plan A costs less but leaves one group waiting much longer. Plan B costs more but shares waiting time more evenly. There may be no single plan that is best on every dimension. Writing cost as the objective does not make waiting unimportant; it merely removes waiting from the comparison unless another part of the model represents it.
One approach is to minimise cost subject to a maximum waiting-time requirement. Another is to minimise a weighted combination of cost and waiting. A third is to present several non-dominated alternatives for a human decision. Each approach answers a different question.
The units matter. Adding dollars directly to minutes without a conversion or explicit weighting hides an arbitrary choice. A change from minutes to seconds would multiply one numerical term by sixty and could change the recommendation, although the real situation had not changed. The model must define why one unit of one quantity trades against another.
This is an ethical and mathematical discipline at the same time. State the objective, name excluded considerations, and do not let an impressive output number speak on behalf of values the model never included.
9. Pareto efficiency identifies trade-offs, not a winner
An allocation is Pareto efficient when no feasible alternative improves at least one objective without worsening another. The efficient set identifies serious trade-off candidates. It does not automatically identify which of them people should prefer.
Imagine three transport plans represented by cost and journey time, with both quantities to be minimised. A costs 10 units and takes 20 minutes. B costs 12 units and takes 15 minutes. C costs 13 units and takes 22 minutes. C is dominated by A: A costs less and arrives sooner. Neither A nor B dominates the other.
Eliminating C improves the discussion without pretending that mathematics alone decides between A and B. A traveller facing a strict deadline may choose B. Someone with a strict budget may choose A. The disagreement is not necessarily a calculation error; it may reflect different admissible constraints or priorities.
The deeper lesson is that optimisation can clarify a decision even when it does not reduce the decision to one universal ranking. Sometimes the most useful result is a map of the price paid for each improvement.
10. Why convexity changes what can be guaranteed
A set is convex when the straight segment between any two of its points stays inside the set. A function is convex when its value along such a segment is no greater than the corresponding weighted average of the endpoint values. For minimisation, this geometry prevents a locally best feasible point from hiding a better point elsewhere in the same convex feasible set.
We can see the reason directly. Suppose a local minimum x had a better feasible point y. Take a point only a tiny fraction of the way from x towards y. Convexity keeps it feasible, and the convex-function inequality makes its objective lower than at x. That contradicts local minimality. The conclusion follows from the segment structure, not from a solver being unusually clever.
This does not imply that every convex problem has a solution, a unique solution or an effortless computation. A minimum might not be attained; the feasible set might be empty; numerical conditioning might still be difficult. Convexity gives powerful structure, not permission to ignore the rest of the problem.
For a formal course route covering convex sets, linear and quadratic programmes, duality and algorithms, see Madeleine Udell’s convex optimisation syllabus.
11. A stationary point is only a candidate
In differentiable unconstrained optimisation, an interior local minimum must have zero gradient. But the converse is false. The function f(x)=x³ has derivative zero at x=0, yet values increase through that point rather than forming a minimum or maximum. Solving f′(x)=0 locates candidates; it does not finish the reasoning.
For f(x)=(x−3)²+2, the square is nonnegative, so the global minimum is 2 at x=3. Differentiation gives f′(x)=2(x−3), agreeing with the direct argument. Now impose 0≤x≤2. The unconstrained minimiser is forbidden. On that interval, the square shrinks as x approaches 2, so the constrained minimum is attained at the boundary x=2, with value 3.
This simple change explains many exam and modelling errors. A method that considers only zero derivatives may miss the boundary where the best allowable answer actually sits. Constraints change not only the possible answers but also the logic used to certify them.
12. Lagrange multipliers explain constrained balance
Consider minimising x²+y² while requiring x+y=6. Substitution gives y=6−x, so the objective becomes x²+(6−x)²=2(x−3)²+18. Its minimum is 18 at x=y=3. The equality constraint converts a two-variable problem into a one-dimensional family.
The multiplier approach writes L=x²+y²+λ(x+y−6). Stationarity gives 2x+λ=0 and 2y+λ=0, together with x+y=6. The first two equations imply x=y; the constraint then gives x=y=3. The multiplier organises the balance between improving the objective and remaining on the constraint surface.
With inequalities, Karush–Kuhn–Tucker conditions combine feasibility, multiplier sign conditions, stationarity and complementary slackness. For differentiable convex problems, satisfying these conditions certifies optimality; appropriate constraint qualifications are needed when asserting necessity. The precise conditions are stated in Convex Optimization, section 5.5.
The teaching priority is to understand what the equations protect. Multipliers are not numbers to calculate merely because the problem looks advanced. They represent how constraints participate in preventing further improvement.
13. Duality and the value of an extra unit
Return to the workshop certificate. We assigned a weight of one to each capacity constraint. Their weighted total produced the score bound 14. Those weights can be read as a small dual construction: resource limits create a bound on what the original allocation can achieve.
Now increase the teaching-time limit from 6 to 6+δ while keeping preparation capacity at 8. Where both capacity limits remain active, solve x+y=6+δ and 2x+y=8. This gives x=2−δ and y=4+2δ. The resulting score is 14+δ. For a sufficiently small positive δ, all minimum-coverage constraints still hold.
The model therefore assigns a local marginal value of one score point to an extra teaching hour in this range. But the formula cannot be extended indefinitely. Once δ reaches 1, x reaches its minimum of 1. Beyond that, another constraint changes the structure. An extra resource’s value depends on which other limits are binding.
This is why sensitivity analysis is more useful than reading a multiplier as an eternal price. It answers a local “what changes next?” question. It does not say the same trade-off survives every expansion of the organisation.
14. An algorithm creates a sequence, not an entitlement
An optimisation algorithm is a rule for generating candidate solutions. Gradient descent moves against a gradient. Interior-point methods follow paths through a constrained problem. Branch-and-bound separates discrete possibilities and compares bounds. A heuristic searches for useful answers without necessarily proving global optimality.
Different methods use different information. Derivative-based methods need reliable derivatives or approximations to them. Some methods exploit convexity. Discrete search methods exploit logical restrictions and bounds. A method suited to a smooth least-squares fit is not automatically suited to assigning indivisible staff members to shifts.
Mathematical modelling systems can enforce composition rules that help identify convex structure before numerical solution. This is the purpose of disciplined convex programming. By contrast, convex–concave programming is used as a structured heuristic for certain nonconvex problems; a successful run should not be misreported as a universal global guarantee.
The practical question is therefore not “Which algorithm is strongest?” It is “Which mathematical structure does this problem possess, and which method can use that structure while reporting an honest result?”
15. A gradient-descent example we can audit completely
Use f(x)=(x−3)²+2 again. Gradient descent with step size α updates x by subtracting α times 2(x−3). With α=1/4, the update becomes xₙ₊₁=xₙ/2+3/2. Starting from x₀=0 gives 0, 1.5, 2.25, 2.625 and so on.
Subtract the true minimiser 3 from both sides: xₙ₊₁−3=(xₙ−3)/2. Each step halves the error exactly. We therefore have a proof of convergence for this example, not merely a sequence that looks encouraging.
Change the step size to α=1. The update becomes xₙ₊₁=6−xₙ. Starting from 0 produces 0,6,0,6 indefinitely. The gradient is correct, the objective is convex, and the algorithm still fails to settle because the update is too aggressive. More generally, the error multiplier is 1−2α, so this quadratic iteration contracts precisely when 0<α<1.
That calculation creates a direct bridge to Dynamical Systems. An optimiser is itself a time-evolving system. Its stability matters independently of the location of the optimum.
16. Integer choices cannot be repaired by casual rounding
Suppose x and y indicate whether two alternatives are selected, with x,y in {0,1}, and exactly one must be chosen. The constraint is x+y=1. A relaxed model allowing fractional choices might return x=y=0.5. Rounding both upward selects two alternatives and violates the rule. Rounding both downward selects none and also violates it.
Rounding is a transformation of the candidate solution. It must be checked against every constraint just like any other transformation. A small numerical change may create a large structural change when variables represent whole vehicles, people, classes or machines.
Relaxations remain useful. For a minimisation problem, enlarging the feasible set cannot increase the best objective value, so a continuous relaxation can provide a lower bound for an integer problem. A feasible integer construction supplies an upper bound. Their gap measures how much improvement may remain possible.
The important distinction is between using a relaxation to guide or bound a search and pretending its fractional result is already a valid real-world plan. The first is mathematical leverage. The second quietly changes the task.
17. Infeasible, unbounded and unfinished are different outcomes
An infeasible problem has no admissible solution. An unbounded maximisation problem allows its objective to grow without a finite upper limit. An unfinished computation may have found a useful candidate without proving whether it is optimal. None of these statements means the same thing.
Imagine maximising x with only x≥0. Increasing x always improves the objective, so there is no finite maximum. If x represents production, the missing upper bound may indicate an omitted resource constraint. By contrast, x≥4 and x≤3 creates infeasibility. Adding a production cap does not repair that contradiction.
A numerical solver also has tolerances, iteration limits and possible numerical difficulties. An algorithm stopping does not itself distinguish all these cases. Research on infeasibility detection and certificates illustrates why reliable termination criteria are a mathematical problem in their own right.
A responsible result should preserve the status returned by the method. “Feasible solution found” must not become “global optimum proved” during the journey from a solver log into a management summary.
18. Uncertain inputs change the decision problem
Suppose preparation demand is uncertain. Activity A might require two units per hour in a typical situation but three when additional support is needed. The earlier allocation x=2,y=4 then needs ten units rather than eight. A plan optimised for the typical case is not automatically feasible in the more demanding case.
A robust approach can require feasibility across an explicitly chosen set of scenarios. Replacing 2x+y≤8 with 3x+y≤8 protects the example against that particular worst-case coefficient. It may sacrifice the nominal score. That sacrifice is not an algorithmic defect; it is the price of the stronger requirement.
A stochastic approach instead assigns probabilities to scenarios and optimises an expected objective, perhaps with additional risk constraints. The quality of that answer depends on the probabilities and consequences represented. A rare failure with a severe consequence should not disappear merely because its contribution to an average looks small.
Both approaches require declared assumptions. “Robust” does not mean safe against every imaginable event. “Expected best” does not mean best in every realised outcome. The uncertainty set or probability model is part of the specification.
19. Sensitivity reveals whether the answer is fragile
A model can produce a highly precise answer from imprecise inputs. That precision may be misleading. Suppose two candidate plans have scores 14.00 and 13.99, but the underlying score coefficients are estimates uncertain by several tenths. Calling the first plan decisively superior overstates the evidence.
One useful check is to vary plausible inputs and see whether the recommendation changes. Another is to identify constraints that are almost violated. A third is to compare several near-optimal plans and ask which offers simpler implementation, more spare capacity or easier recovery.
In the workshop example, using all eight preparation units leaves no modelled reserve. That can be entirely appropriate in a deterministic exercise. In an uncertain operation, the same fact should prompt a question about buffers. The mathematical optimum is often on a boundary precisely because all improvement opportunities have been used up.
The correct response is not to distrust optimisation. It is to optimise the right problem, with uncertainty, recovery capacity and acceptable constraint violation represented where they matter.
20. Repeated decisions require feedback
A one-time plan chooses an allocation from the information available now. A repeated decision process observes outcomes, updates the state and makes another choice. The objective may include future consequences rather than only the next period.
Consider a simple study-planning model. Choosing the topic with the highest immediate predicted score gain every day might neglect prerequisites that enable later progress. A longer-horizon model would represent how today’s action changes tomorrow’s available capabilities. The teaching example illustrates the structure; it does not claim that learning can be fully captured by one numerical state.
In control applications, optimisation can be embedded inside repeated state measurement and replanning. Stanford’s overview of convex optimisation includes resource allocation, model fitting and real-time control among its applications.
The connection to dynamics is essential. A sequence of individually attractive actions can destabilise the larger process when effects are delayed or constraints couple time periods. A plan should therefore specify not only what is selected but when the next observation arrives and what would trigger revision.
21. What changes when other decision-makers respond?
An optimiser often treats the environment as fixed while selecting a plan. But another person, organisation or software agent may change its behaviour in response. Then the outcome depends on several interacting choices rather than on one planner’s variables alone.
Imagine assigning traffic to a supposedly faster route. Once many travellers receive the same recommendation, that route may cease to be faster. A calculation that treats travel time as fixed has omitted a feedback relationship. Depending on the modelling question, the repair may require congestion functions, a central allocation model or a game among route users.
This leads to Game Theory. Each participant may optimise, but their joint outcome need not optimise any shared objective. An equilibrium and a social optimum are different mathematical objects.
Information also affects what can be chosen. A planner who cannot observe demand, capacity or preferences solves a different problem from one who can. The related Information Theory article explains how uncertainty and communication limits can be quantified without confusing them with the value judgement inside the objective.
22. A solution needs a mathematical audit and a world audit
The mathematical audit checks that the returned variables satisfy the declared constraints and achieve the reported objective. Recalculate the score independently. Check units, signs, bounds, equalities and integer restrictions. Inspect numerical tolerances and any available optimality gap. For a small model, substitute the answer by hand. For a large model, preserve a separate checking procedure.
The world audit asks whether the declared problem still matches the real decision. Are the people available? Are the rooms usable? Are the measurements current? Does a supposedly divisible resource actually arrive in fixed packages? Are there consequences not represented by the objective?
These audits should not be collapsed. A correct model with a faulty computation needs algorithmic repair. A perfect computation of an unsuitable model needs modelling repair. An implementable plan can also become unsuitable after conditions change, which calls for monitoring and revision.
The Numerical Analysis route is useful for the computational audit. The wider How Mathematics Works root keeps the larger relationship between representation, valid transformation and interpretation visible.
23. A learning route from school mathematics to optimisation
The earliest preparation is not advanced notation. It is learning to distinguish a quantity from its unit, an unknown from a given value, and a requirement from a preference. Word problems involving limited money, time or capacity already contain the ingredients of constrained choice.
Algebra supplies equations and inequalities. Coordinate geometry makes a feasible region visible. Calculus supplies local rates and stationary-point reasoning. Linear algebra packages many interacting variables. Probability represents uncertain inputs, while numerical analysis explains how algorithms can produce reliable approximate answers.
A productive exercise is to solve one small model in three ways: by direct reasoning, by a graph and by a general method. Agreement reveals the invariant structure. Disagreement exposes either a calculation error or a hidden difference in assumptions.
For the workshop example, the best explanatory moment is the certificate 3x+2y≤14. The student sees that mathematics can rule out infinitely many better alternatives with a short argument. The goal is not to memorise an optimisation vocabulary list. It is to recognise how a choice becomes defensible.
24. Questions that reveal real understanding
Does optimal mean perfect?
No. It means no admissible alternative improves the specified objective under the stated model. A plan can be optimal for cost while unacceptable for another consideration that was omitted. The useful follow-up is not “Is this optimal?” alone, but “Optimal for which objective, under which constraints, using which information?”
Can there be several best solutions?
Yes. If several feasible points have the same best objective value, the problem has multiple optima. This may create flexibility rather than a defect. A secondary criterion can select among them, or the planner can choose the easiest to implement while preserving the primary objective.
Why not test every possible choice?
For a tiny finite problem, complete enumeration may be ideal. But continuous problems have infinitely many candidates, and discrete problems can have extremely large collections. Twenty binary choices already produce 2²⁰ configurations. Structure, bounds and appropriate algorithms reduce the search without pretending that computational cost does not exist.
What does a solver’s numerical tolerance mean?
It specifies how closely a numerical computation must satisfy its stopping or feasibility tests. The significance depends on units and scale. A deviation of 0.001 might be harmless in one model and important in another. Tolerance is a computational contract that must be interpreted in the application, not an automatic real-world permission.
What should a learner check in the workshop example?
First verify that (2,4) satisfies every constraint. Then calculate its score. Finally reconstruct the upper-bound argument from the two capacity inequalities. As a variation, replace the six-hour limit by seven hours: (1,6) becomes feasible and reaches the new bound of 15. Explain why continuing that same adjustment indefinitely would violate the minimum allocation to A.
25. The deeper mechanism: choice, evidence and return
Optimisation connects several kinds of reasoning. Representation turns a situation into variables and constraints. Comparison turns preferences into an objective. Geometry reveals possible movement. Algorithms search. Bounds and certificates establish what has been achieved. Interpretation returns the mathematical answer to the decision that created the problem.
Leaving out any layer creates a characteristic failure. No representation means vague advice. No constraints means impossible recommendations. No explicit objective means concealed priorities. No certificate means an uncertain quality claim. No interpretation means a correct number with no dependable use.
The workshop model shows all these layers in a small enough form to inspect. We defined two choices, named two resource limits, found a feasible allocation, proved its score could not be beaten, changed a parameter and watched another constraint become important. The same habits remain valuable when the notation grows into thousands of variables.
Mathematical optimisation does not replace judgement. It makes the chosen objective, the feasible alternatives and the justification for a decision inspectable. Its strongest contribution is not simply producing a better number. It is showing what “better” means, what it costs, and what evidence supports the claim.
Sources and further mathematical study
Formal foundations: Boyd and Vandenberghe, Convex Optimization. Modelling and algorithms: Grant, Boyd and Ye, Disciplined Convex Programming; Shen and colleagues, Disciplined Convex-Concave Programming. Reliability of solver outcomes: Banjac and colleagues, Infeasibility Detection. Applications and teaching progression: Boyd’s overview and Udell’s course outline. The numerical workshop and gradient-descent calculations above are teaching derivations, not reported empirical results.
How Mathematics Works | Batch 07
Mathematical Optimisation chooses under constraints. Game Theory studies choices that respond to other choices. Information Theory measures uncertainty and communication limits. Dynamical Systems studies how states evolve and whether repeated updates remain stable.
Return to the Mathematics Learning Hub or continue through the How Mathematics Works root.