Fractional programming optimises ratios such as useful output per unit of energy, benefit per unit of resource, or cost per completed task. The gold standard is to define what the numerator and denominator actually measure, establish the denominator’s sign, preserve every feasibility constraint, and verify the final ratio in the original model.
A high ratio is not automatically a high total benefit. This guide develops that distinction through original worked examples, then derives threshold tests, a parametric solution method and a linear-fractional transformation. You will learn how to prove an optimum, diagnose misleading numerical results and recognise when changing the denominator changes the decision itself.
Start with the Gold Standard master, the Education guide to transferable understanding and the How X Works hub. The Parametric Optimization root supplies a related route into families of optimisation problems. Here the defining feature is the ratio and the conditions needed to manipulate it correctly.
Choose your fractional-programming route
Chapters 1–3: Define and solve a ratio
Chapters 4–6: Test targets and iterate
Chapters 7–9: Certify and transform
Chapters 10–12: Diagnose and compute
Chapter index
Define and solve a ratio · Chapters 1–3
1. Define the ratio before choosing an algorithm
Test targets and iterate · Chapters 4–6
4. Turn a target ratio into a feasibility question
Certify and transform · Chapters 7–9
7. Solve the continuous subproblem and check its residual
8. Transform a linear-fractional problem into a linear programme
Diagnose and compute · Chapters 10–12
10. Diagnose zero denominators and missing optima
Transfer and continue · Chapters 13–14
13. Transfer the calculation when overhead or service requirements change
14. Answer common questions and choose a useful next project
CHAPTER 1 OF 14 · Define and solve a ratio
1. Define the ratio before choosing an algorithm
Write a single-ratio maximisation problem as choosing x in a feasible set X to maximise R(x) = f(x)/g(x). The numerator f records the outcome of interest. The denominator g records the resource or reference quantity against which that outcome is measured. Throughout the main derivations below, assume g(x) > 0 for every feasible x.
The positivity assumption is substantive. Multiplying an inequality by a positive number preserves its direction, multiplying by a negative number reverses it, and dividing by zero is undefined. An algorithm that assumes positive denominators cannot safely be applied to a model whose denominator can cross zero merely because the starting point happens to be positive.
For an efficiency example, the numerator might be completed units of useful work and the denominator total energy consumed during the same operating period. Include the same system boundary in both quantities. Counting output for a whole facility while counting only one machine’s energy would produce a number, but not the facility efficiency claimed in its label.
A ratio needs units and a time boundary
If f is measured in completed tasks and g in kilowatt-hours, the ratio has units of tasks per kilowatt-hour. If the denominator changes to hours, it becomes throughput rather than energy efficiency. If both numerator and denominator are monetary amounts, the ratio can be dimensionless but still depend on what costs and benefits were included.
Write the units beside the formula and explain whether the data refer to one batch, one hour or one year. A fixed setup cost allocated per batch behaves differently from the same cost allocated across annual production. This is a modelling question before it is an algebra question.
Also state whether the goal is to maximise benefit per resource or minimise resource per benefit. Taking a reciprocal preserves the preferred decision only when the quantities are strictly positive and the objective direction is reversed appropriately. A zero-benefit option may be valid in the first formulation while making the reciprocal undefined.
CHAPTER 2 OF 14 · Define and solve a ratio
2. Separate efficiency from total contribution
Suppose three mutually exclusive operating plans have outcome and resource pairs A = (12, 2), B = (20, 5) and C = (21, 7). Their ratios are 6, 4 and 3. Plan A is the most efficient by this metric, even though it produces the smallest total outcome.
If the requirement is to produce at least 18 outcome units, A is infeasible. Among the remaining plans, B has the larger ratio and also uses less resource than C. The minimum-output constraint changes the relevant comparison. It is not something that should be added informally after announcing the unconstrained ratio winner.
Now suppose the objective were simply to maximise total outcome under a resource budget of 7. With these three indivisible plans, C would win. The data have not changed; the decision criterion has. “Best” is meaningful only after the objective and constraints are specified.
Why independent maximisation and minimisation fail
One cannot normally maximise f and minimise g separately and divide the resulting values. The largest numerator may occur at one feasible decision and the smallest denominator at another. Combining them creates an unattainable pair. In this example, 21/2 = 10.5 is not the performance of any available plan.
The numerator and denominator must be evaluated at the same x. This rule also applies when building a spreadsheet: references to different rows can silently create a fictional ratio. Keep the decision identifier, numerator, denominator and feasibility result together in every comparison.
Ratios can reward small operations that avoid expensive marginal output. That is not a defect if efficiency is the intended goal. It becomes a defect when the actual task includes service coverage, a production commitment or a minimum number of people helped. Represent those requirements explicitly instead of expecting the ratio to infer them.
For a related example of how omitted requirements distort a mathematically correct answer, see the existing diet optimisation and nutrient-constraints article. The useful connection is the discipline of defining feasibility; the present article’s calculations are generic resource models, not dietary advice.
CHAPTER 3 OF 14 · Define and solve a ratio
3. Solve a continuous efficiency model exactly
Consider the teaching model f(x) = 6x − x² and g(x) = 2 + x for 0 ≤ x ≤ 4. Think of x as an adjustable operating intensity, f as useful output with diminishing returns, and g as a fixed resource requirement of 2 plus a variable requirement x. The units have been normalised consistently for this illustration.
The denominator lies between 2 and 6, so it is strictly positive. The ratio is R(x) = (6x − x²)/(2 + x). The quotient rule gives R′(x) = (12 − 4x − x²)/(2 + x)². The numerator vanishes when x² + 4x − 12 = 0, giving roots 2 and −6.
Only x = 2 lies in the feasible interval. The derivative is positive before 2 and negative after 2, so the unique maximiser is x* = 2. Its output is 8, its resource use is 4 and its ratio is 2. The endpoints have ratios 0 and 8/6, both lower.
A shorter global certificate
We can prove the same result without relying on a plot or local derivative interpretation. Subtract twice the denominator from the numerator: f(x) − 2g(x) = 6x − x² − 4 − 2x = −(x − 2)². This expression is never positive and is zero exactly at x = 2.
Since g(x) is positive, the identity proves f(x)/g(x) ≤ 2 throughout the feasible interval. Equality is attained at the stated candidate, so 2 is the global optimum. The certificate contains both an upper bound valid for every feasible decision and a feasible decision attaining it.
Maximum total output occurs instead at x = 3, where f = 9. Its ratio is 9/5 = 1.8. The ratio optimiser accepts one less output unit in exchange for one less resource unit. Neither calculation contradicts the other: they solve different optimisation problems.
Keep this example available as a test case for software. A correct implementation should return an intensity near 2 for the ratio objective and near 3 for the numerator-only objective. If both return 3, inspect whether the denominator was actually included in the function passed to the solver.
CHAPTER 4 OF 14 · Test targets and iterate
4. Turn a target ratio into a feasibility question
For any proposed target λ and positive g(x), the statement f(x)/g(x) ≥ λ is equivalent to f(x) − λg(x) ≥ 0. This elementary multiplication is the bridge from a ratio to a family of non-fractional problems.
In the continuous example, a target of 2 requires −(x − 2)² ≥ 0, so only x = 2 meets it. A target above 2 cannot be reached. A smaller target admits a range of intensities rather than a single point.
For λ = 1.5, the inequality is −x² + 4.5x − 3 ≥ 0. Its boundary roots are (9 − √33)/4 and (9 + √33)/4, approximately 0.814 and 3.686. Since the quadratic opens downwards, every x between those roots reaches or exceeds the target. Both roots lie within the feasible interval [0, 4].
When target tests have useful convex structure
Suppose X is convex, f is concave, g is positive and convex, and λ ≥ 0. Then f − λg is concave. Its superlevel set where it is nonnegative is convex. A target-ratio feasibility test can therefore be a convex feasibility problem even when the ratio itself is not a concave function.
This observation does not say that every fractional programme is easy. The numerator, denominator, feasible set and target sign matter. A sum of several ratios is another structure. Integer decisions can destroy convexity. A denominator that changes sign invalidates the simple multiplication rule used here.
For a reader, the practical question is whether a proposed performance threshold is achievable under the original constraints. For an algorithm, repeated threshold tests can narrow an interval containing the best ratio. Both uses depend on a trustworthy feasibility decision, not merely on the numerical appearance of one attempted point.
If a local search fails to find a point reaching the target, that is not necessarily proof that the target is impossible. A global infeasibility certificate or an appropriate convex feasibility method carries a stronger claim than an unsuccessful search from one starting point.
CHAPTER 5 OF 14 · Test targets and iterate
5. Derive the zero of a parametric value function
Define H(λ) as the maximum, over x in X, of f(x) − λg(x). To make the next statements straightforward, assume X is nonempty and compact, f and g are continuous, and g is strictly positive on X. These assumptions ensure that the maxima exist and that g has a positive minimum.
Let r* be the maximum feasible ratio. At λ = r*, every feasible x satisfies f(x) − r*g(x) ≤ 0, and a ratio maximiser attains zero. Therefore H(r*) = 0. If λ < r*, evaluating at a ratio maximiser gives a positive value, so H(λ) > 0. If λ > r*, every candidate is strictly below zero, and compactness ensures H(λ) < 0.
Thus solving the ratio problem can be viewed as locating the zero of H. The parameter λ has the units of the ratio. The expression f − λg has the units of the numerator, because λ converts the resource term into comparable outcome units.
Why the value function decreases
Increasing λ penalises the denominator more heavily. If g(x) ≥ m > 0 for all feasible x, increasing λ by Δ > 0 reduces every candidate expression by at least Δm. Consequently H(λ + Δ) ≤ H(λ) − Δm. The value function is strictly decreasing under these assumptions.
This also explains why a small denominator can weaken numerical interpretation. If m is tiny, a small change in the difference objective may correspond to a much larger change in the ratio. A residual must be read alongside the denominator scale.
The primary paper by Shen and Yu, Fractional Programming for Communication Systems—Part I, discusses classical single-ratio methods and their distinction from multiple-ratio formulations. The derivation above is a self-contained way to understand the central parameter test before implementing any particular method.
CHAPTER 6 OF 14 · Test targets and iterate
6. Follow a complete finite Dinkelbach iteration
A Dinkelbach-style iteration for maximising one ratio solves the subproblem that maximises f(x) − λg(x), then replaces λ by the ratio at the selected solution. Exact global subproblem solutions and suitable assumptions underpin the classical guarantees. The following finite example makes every subproblem completely checkable.
Use the three plans A = (12, 2), B = (20, 5) and C = (21, 7). Start with λ = 0. The difference values are simply 12, 20 and 21, so C is selected. Update λ to 21/7 = 3.
At λ = 3, the difference values are A: 12 − 3 × 2 = 6; B: 20 − 3 × 5 = 5; C: 21 − 3 × 7 = 0. The best plan is A. Update λ to 12/2 = 6.
At λ = 6, the values are A: 0; B: −10; C: −21. Their maximum is zero, certifying that no available plan has a ratio above 6. Plan A attains that ratio, so the procedure stops with an exact optimum.
Why the update improves the ratio
If xλ solves the subproblem, then f(xλ)/g(xλ) − λ = H(λ)/g(xλ). A positive subproblem value means that the new ratio is larger than the old parameter. At the optimum, the best difference is zero and no further ratio improvement is possible.
The finite example terminates quickly because there are only three possible ratios and each subproblem is solved by exhaustive comparison. It does not justify promising the same iteration count for continuous or large discrete models. The expensive part may be solving each subproblem rather than updating the scalar parameter.
When an approximate solver is used, a candidate’s difference value is only a lower bound on H in this maximisation formulation. A near-zero value found locally does not prove that another feasible decision cannot give a positive difference. To certify the final ratio, obtain a valid upper bound on the global subproblem value or exploit structure that makes the subproblem globally solvable.
CHAPTER 7 OF 14 · Certify and transform
7. Solve the continuous subproblem and check its residual
For f(x) = 6x − x², g(x) = 2 + x and X = [0, 4], the subproblem is to maximise −x² + (6 − λ)x − 2λ. It is a concave quadratic. Its unconstrained maximiser is x = (6 − λ)/2; if that point lies outside [0, 4], the appropriate boundary gives the constrained maximum.
Starting at λ = 0 gives x = 3 and the updated ratio 9/5 = 1.8. At λ = 9/5, the subproblem maximiser is x = 21/10 = 2.1. The numerator there is 819/100, the denominator is 41/10, and the updated ratio is 819/410, approximately 1.997561.
At the next step the maximiser is x = (6 − 819/410)/2 = 1641/820, approximately 2.001220. The sequence is approaching the exact ratio optimum established earlier. The numerical approach is useful, but the identity f − 2g = −(x − 2)² remains the clearest exact certificate.
Translate a subproblem bound into a ratio bound
Suppose a feasible candidate has ratio λ and a global subproblem bound proves H(λ) ≤ ε. If g(x) ≥ m > 0 everywhere, then every feasible ratio satisfies f(x)/g(x) ≤ λ + ε/m. The candidate supplies a lower bound λ; the certified subproblem bound supplies an upper bound λ + ε/m.
In this example m = 2. If a global upper bound on H is 0.0002, the ratio gap is at most 0.0001. This conclusion requires an upper bound on the maximum difference, not just a small value at the latest iterate. The distinction is essential when the inner problem is nonconvex.
Also distinguish ratio error from decision error. Several decisions can have nearly equal ratios, especially near a flat maximum. A tight bound on objective value does not automatically imply that x is close to a unique optimiser. Report the decision, numerator, denominator, ratio and the type of bound separately.
CHAPTER 8 OF 14 · Certify and transform
8. Transform a linear-fractional problem into a linear programme
Consider maximising (3x + 2)/(x + 2) subject to 0 ≤ x ≤ 4. The numerator and denominator are affine, and the denominator is positive. Introduce t = 1/(x + 2) and y = tx. Then y + 2t = 1, y ≥ 0 and y ≤ 4t. The ratio objective becomes 3y + 2t.
All these expressions are linear in y and t. Conversely, any point satisfying these constraints with t > 0 yields x = y/t and recovers the original denominator relation. In this particular transformed feasible set, t cannot be zero: y ≤ 4t and y ≥ 0 would force y = 0, contradicting y + 2t = 1.
Use y = 1 − 2t in the objective to obtain 3 − 4t. The bound y ≤ 4t gives 1 − 2t ≤ 4t, hence t ≥ 1/6. Maximising the objective therefore chooses t = 1/6 and y = 2/3. Recover x = 4 and ratio 7/3.
Verify in the original variables
Substitution gives (3 × 4 + 2)/(4 + 2) = 14/6 = 7/3. A direct identity provides another check: (3x + 2)/(x + 2) = 3 − 4/(x + 2), which increases throughout [0, 4]. Both calculations identify the same endpoint.
This is a concrete instance of the Charnes–Cooper transformation. Boyd and Vandenberghe’s Convex Optimization lecture materials present linear-fractional programming and its linear-programming reformulation. The important implementation lesson is to transform every constraint, not merely the objective.
For a vector model with Ax ≤ b and x ≥ 0, the corresponding relation is Ay ≤ bt. A denominator dᵀx + β becomes the normalisation dᵀy + βt = 1. Recovery requires t > 0; treatment of t = 0 and unbounded domains needs care rather than an automatic division by whatever value a solver returns.
CHAPTER 9 OF 14 · Certify and transform
9. Understand mixtures as denominator-weighted averages
Suppose two feasible operating modes have positive denominators g₁ and g₂ and ratios r₁ = f₁/g₁ and r₂ = f₂/g₂. If their outcomes and resources combine linearly with proportions θ and 1 − θ, the combined ratio is [θg₁r₁ + (1 − θ)g₂r₂]/[θg₁ + (1 − θ)g₂].
This is a weighted average of the individual ratios, with weights proportional to denominator contribution. It therefore lies between r₁ and r₂. It is generally not the simple arithmetic average of the ratios, even when the operating proportions are equal.
Mix half of plan A = (12, 2) and half of plan B = (20, 5). The combined outcome is 16 and the resource is 3.5, giving ratio 32/7, approximately 4.571. The arithmetic mean of 6 and 4 is 5, which is not the combined efficiency. The higher-resource mode receives more weight in the aggregate ratio.
When the averaging argument does not apply
The derivation assumes linear combination of both totals. Shared setup costs, congestion, interference or nonlinear interactions can break that assumption. If running two activities together eliminates a duplicated fixed overhead, their combined ratio need not follow this simple formula. Model the shared overhead explicitly rather than treating it as if it were incurred twice.
The same issue appears in reporting. Average productivity across teams can mean the average of team ratios or total output divided by total hours. These measures answer different questions. The latter weights teams by hours, while the former weights each team equally regardless of size.
Before optimising an average, identify who or what receives equal weight. A mathematically correct optimisation can favour an unintended allocation when the aggregation rule is wrong. Store raw numerators and denominators alongside reported ratios so that another reader can reconstruct the intended total.
CHAPTER 10 OF 14 · Diagnose and compute
10. Diagnose zero denominators and missing optima
The function R(x) = 1/x on 0 < x ≤ 1 has no finite maximum. Values grow without bound as x approaches zero. A solver returning a tiny positive x and a huge ratio has not discovered a meaningful efficient design; it has exposed the open boundary and the denominator singularity in this model.
By contrast, R(x) = x/(1 + x) for x ≥ 0 is bounded above by 1 but never reaches 1 at a finite x. Its supremum is 1, while no maximising finite decision exists. A large-x numerical answer is an approximation to an unattained limit, not an exact optimiser.
These examples explain why existence assumptions matter. A continuous ratio on a nonempty compact feasible set with a strictly positive continuous denominator attains its maximum. Dropping compactness or allowing the denominator to approach zero can change that conclusion substantially.
Cancellation does not silently restore an excluded point
For x ≠ 0, the expression x/x equals 1. At x = 0, the original ratio is undefined. Algebraic cancellation describes the function on its original domain; it does not automatically add the missing point. A continuous extension can be defined deliberately, but that is an additional modelling choice that should be stated.
If a denominator can be negative, do not apply positive-denominator threshold tests across the entire domain. Separate sign regions if appropriate and derive the inequalities for each one. Near zero, rounding can even obscure which region a computed point occupies.
An arbitrary numerical floor can make software run while changing the mathematical problem. Replacing g by max(g, 0.001), for example, changes every ratio in the affected region. If a positive lower bound represents a real minimum operating cost, include it transparently. If it is only a numerical patch, report its effect and do not claim equivalence with the original formulation.
CHAPTER 11 OF 14 · Diagnose and compute
11. Keep single ratios, sums and uncertainty distinct
A ratio of totals and a sum of ratios are different objectives. Using the same two modes, the sum 12/2 + 20/5 is 10, while the ratio of totals (12 + 20)/(2 + 5) is 32/7. No algebra allows one expression to replace the other without changing the task.
Single-ratio transformations therefore should not be copied mechanically into a multiple-ratio model. A network may combine several link-specific efficiency or signal-quality terms, each with its own denominator. The Shen–Yu research paper linked earlier studies transformations for such communication problems and distinguishes stationary-point results from stronger global guarantees.
An uncertain denominator changes what an average means
Suppose two equally likely scenarios have outcome/resource pairs (10, 1) and (10, 9). The expected scenario ratio is (10 + 10/9)/2 = 50/9, approximately 5.556. The ratio of expected outcome to expected resource is 10/5 = 2. These are not interchangeable measures.
The first treats each scenario’s efficiency equally. The second compares expected totals. Neither is automatically the correct criterion for every decision. Specify whether performance is evaluated per scenario, over accumulated resource use, at a worst case or against a service guarantee.
If the denominator is estimated with error, a small denominator can amplify ratio uncertainty. For independent small perturbations, a first-order calculation gives ΔR approximately Δf/g − fΔg/g². This approximation is local and becomes unreliable near a zero denominator or for large errors. It nevertheless shows why denominator measurement deserves as much attention as numerator measurement.
For a robust decision, build uncertainty into the model and define the guarantee you actually need. Repeatedly evaluating one nominal optimum under a few scenarios is a useful diagnostic, but it is not by itself proof of optimality for a robust or stochastic formulation.
Begin with a written model containing the feasible set, objective direction, units and a verified denominator lower bound where available. Find at least one feasible decision and compute its ratio directly. That decision gives a baseline which a proposed improvement must actually beat under the same definitions.
Choose a method that matches the structure. A small finite set can be enumerated exactly. A one-variable model may admit a derivative or algebraic proof. A linear-fractional model may admit a linear-programming transformation. A general nonlinear ratio requires more care about global subproblem solutions and the strength of any certificate.
After solving, reconstruct x in the original variables. Recompute every constraint, f(x), g(x) and f(x)/g(x). Compare these with transformed variables and the solver’s objective. A transformed point with a tiny t, an infeasible recovered x or a denominator of the wrong sign must be investigated before reporting success.
Keep a feasible lower bound and a certified upper bound
For maximisation, the best verified feasible ratio is a lower bound on the true optimum. A valid proof that f − λg ≤ 0 everywhere supplies an upper bound λ. A global upper bound on H combined with a positive denominator lower bound supplies the error estimate developed earlier.
Do not reverse these roles. A solver’s promising candidate is evidence of achievable performance, not proof that better performance is impossible. Conversely, an upper bound may be unattainable and is not itself a deployable decision.
Use the exact examples in this guide as regression checks for a computational implementation. Test the three-plan optimum 6, the continuous optimum 2 and the linear-fractional optimum 7/3. Include a deliberately infeasible minimum-output case and a denominator-singularity case. The purpose is to test whether the implementation respects the model’s logic, not just whether it returns a number.
Finally, retain the raw data and the model version. A changed resource price, overhead allocation or output definition creates a different optimisation problem even if the code and article title remain the same. Reproducibility requires knowing which quantities were actually divided.
CHAPTER 13 OF 14 · Transfer and continue
13. Transfer the calculation when overhead or service requirements change
Replace the fixed overhead 2 in the continuous example by a positive parameter c. The ratio becomes R_c(x) = (6x − x²)/(c + x) on [0, 4]. Its derivative has numerator 6c − 2cx − x². Setting this to zero gives the positive critical point x_c = −c + √(c² + 6c).
For every c > 0, this point lies between 0 and 3. The derivative is positive before it and negative afterwards, so it is the unique ratio maximiser on [0, 4]. At c = 2, the formula returns 2. At c = 1, it returns √7 − 1, approximately 1.646. At c = 6, it returns √72 − 6, approximately 2.485.
Higher fixed overhead encourages a higher operating intensity in this model because useful output can spread that overhead across more activity. The statement is specific to these functions and constraints. It is not a universal rule about all machines, businesses or services.
Add a minimum-output requirement
Return to c = 2 but require f(x) ≥ 8.5. Completing the square gives f(x) = 9 − (x − 3)², so feasible intensities satisfy |x − 3| ≤ √0.5. Within [0, 4], this is the interval [3 − 1/√2, 3 + 1/√2].
The original ratio optimum x = 2 is now infeasible. Since R decreases for x > 2, the best feasible intensity is the left endpoint x = 3 − 1/√2, approximately 2.293. Its output is exactly 8.5 and its ratio is 8.5/(5 − 1/√2), approximately 1.980. A service requirement has lowered the maximum achievable efficiency while increasing delivered output.
Check the changed context yourself
If the minimum output rises to 9, only x = 3 remains feasible and the ratio is 1.8. If it rises above 9, no feasible intensity exists under the stated output curve. An algorithm should report the infeasibility rather than inventing a higher intensity that violates the curve or the interval.
This exercise is a transfer test: identify which constraint changed, recover the new feasible region, and then optimise over it. The Parametric Optimization guide develops the broader skill of tracking decisions as parameters and active requirements change.
CHAPTER 14 OF 14 · Transfer and continue
14. Answer common questions and choose a useful next project
Is fractional programming about fractional decision variables?
No. The defining feature is a ratio in the objective or related formulation, not the permission to choose noninteger quantities. A fractional programme can have continuous or discrete decisions. The three-plan example was entirely discrete, while the intensity examples used a continuous interval.
Why not optimise the numerator alone?
That answers a different question unless the denominator is a fixed positive constant. In the worked continuous model, maximum output occurs at intensity 3, whereas maximum efficiency occurs at 2. Report both quantities when readers need to understand the trade-off rather than presenting one as a substitute for the other.
Does a positive denominator make every ratio problem convex?
No. Positivity makes certain algebraic transformations valid. Convexity or quasiconvexity depends on additional structure. Even when target feasibility sets are convex, the original ratio need not be concave. Multiple ratios, discrete choices and nonconvex constraints can introduce further difficulty.
When is a small difference residual enough to stop?
When its interpretation is justified. A certified global bound H(λ) ≤ ε and a denominator bound g ≥ m > 0 give a ratio-gap bound ε/m. A small value at one local solution is weaker. State the tolerance, the bound direction and the assumptions instead of treating every small solver number as the same kind of evidence.
What should I learn first?
Begin with fractions, inequalities, units, feasible regions and the difference between a candidate and a proof of optimality. Then reproduce the square identity in Chapter 3, the finite iteration in Chapter 6 and the variable recovery in Chapter 8. Calculus helps with smooth examples but does not replace checking domain restrictions.
For a practical project, build a small table of feasible operating plans with raw outcome, resource and minimum-service columns. Compare the ratio winner with the total-output winner. Add one fixed overhead and observe when the ranking changes. Keep the example fictional unless you have reliable measured inputs and permission to use them.
Next implement the continuous model and compare numerical answers with the analytical optimum. Deliberately change c and the minimum-output requirement, then check the predictions in Chapter 13. The goal is to explain why the chosen decision changes, not merely to produce a colourful parameter sweep.
Use the linear-programming specialist guide for constraint modelling, the Parametric Optimization root for changing inputs, and the primary Shen–Yu paper and Boyd–Vandenberghe lecture materials for deeper mathematical routes. A dependable ratio calculation ends with a feasible decision, interpretable totals and evidence supporting the exact claim being made.