Multivariable Calculus and Constrained Optimisation

Introduction

Business decisions often depend on several variables simultaneously. A firm’s output may depend on labour and capital; utility may depend on quantities of several goods; revenue may depend on prices, advertising expenditure, and production volume; and resource-allocation decisions may be limited by budgets, capacity, or regulations. Multivariable calculus provides a systematic language for analysing these relationships.

This chapter develops partial derivatives, gradients, Hessian matrices, optimisation with several variables, and Lagrange multipliers. The emphasis is on interpretation: marginal effects, interactions between variables, curvature, optimal allocation, and the economic meaning of constraints.

Learning Objectives

By the end of this chapter, you should be able to:

1. Interpret functions of several variables in business contexts.
2. Calculate and interpret first- and second-order partial derivatives.
3. Use gradients to describe rates of change and identify stationary points.
4. Apply Hessian matrices to classify local maxima and minima.
5. Solve constrained optimisation problems using Lagrange multipliers.
6. Interpret shadow prices and conduct sensitivity analysis.
7. Evaluate the assumptions and limitations of calculus-based business models.

1. Functions of Several Variables

A multivariable function maps several inputs to one output. For example, a production function may be written as

where Q is output, L is labour input, and K is capital input. A common illustrative model is

The model assumes that output is determined by these two inputs and that the function is sufficiently smooth for differentiation. Its units depend on the units chosen for labour, capital, and output.

A function’s domain is the set of feasible input combinations. In production, normally L ≥ 0 and K ≥ 0. The range is the set of possible output values.

A point such as (L,K)=(25,16) identifies one operating position. Holding one input fixed while changing the other produces a one-variable slice of the multivariable function.

2. Partial Derivatives and Marginal Analysis

The partial derivative with respect to one variable measures the instantaneous change in the output when that variable changes while all other specified variables remain constant. For z=f(x,y), the partial derivatives are

and

The phrase “holding other variables constant” is essential. Partial derivatives are local measures and should not automatically be interpreted as causal effects outside the model’s assumptions.

Worked Example 1: Production Marginals

Let

Using the power rule,

and

At L=25 and K=16,

and

Thus, at this operating point, a small increase of one unit of labour, with capital fixed, increases predicted output by approximately 4 output units. A small increase of one unit of capital, with labour fixed, increases predicted output by approximately 6.25 output units. These are local approximations, not exact changes for large input increases.

The differential provides a first-order approximation:

where dL and dK are small changes.

3. Marginal Revenue, Cost, and Utility

If profit is

then

is the marginal effect of x on profit, holding y constant. An interior unconstrained optimum requires

provided the function is differentiable and the optimum lies inside the feasible region.

For a utility function U(x,y), U_x and U_y are marginal utilities. The marginal rate of substitution, when U_y ≠ 0, is

Along an indifference curve, utility is constant, so

and therefore

The negative sign indicates the trade-off required to maintain utility, while the ratio gives its magnitude.

4. Gradients and Directional Change

The gradient of f(x,y) is the vector

For n variables, ∇f contains all first-order partial derivatives. The gradient points in the direction of the greatest local increase in f. Its magnitude is the maximum rate of increase per unit distance:

If u is a unit direction vector, the directional derivative is

where “·” denotes the dot product. This is useful in analytics when a firm can adjust several inputs only along a prescribed direction, such as a fixed technology-migration plan.

A stationary point satisfies

Stationarity is necessary for an interior local optimum under differentiability, but it is not sufficient: a stationary point may be a maximum, minimum, or saddle point.

5. Second Derivatives and Interaction Effects

Second-order partial derivatives measure curvature and changing marginal effects:

The cross-partial derivative is

Under appropriate smoothness conditions, specifically continuity of the relevant second derivatives, f_xy=f_yx. A positive cross-partial indicates that increasing one variable raises the marginal effect of the other. This is often interpreted as complementarity. A negative cross-partial suggests substitutability in the local model.

For example, if π_xy>0, a larger value of y increases the marginal profitability of x, subject to the model’s assumptions.

6. The Hessian and Classification of Stationary Points

The Hessian matrix contains second-order partial derivatives:

For a twice-continuously differentiable function of two variables, define

At a stationary point:

This is a local test. It does not prove that the point is the global optimum unless additional conditions, such as concavity or convexity over the feasible domain, are established.

Worked Example 2: Profit Classification

Suppose

The first derivatives are

Setting both equal to zero gives

Solving yields x=50/3 and y=20/3. The second derivatives are

Thus

Because π_xx<0, the stationary point is a local maximum. The negative-definite Hessian also shows that the quadratic profit function is strictly concave, so this is the unique global maximum over an unconstrained domain. If quantities must be non-negative, the point remains feasible and is still optimal.

7. Unconstrained Optimisation Procedure

A practical procedure is:

1. Define the objective function and variables.
2. State the feasible domain and units.
3. Compute all first partial derivatives.
4. Solve the stationary equations.
5. Check boundary points when the domain has boundaries.
6. Use the Hessian or concavity/convexity to classify candidates.
7. Compare objective values when several candidates remain.
8. Interpret the result in business terms and test sensitivity.

Ignoring boundaries is a common error. For example, a mathematical stationary point with negative production or negative expenditure may be economically infeasible.

8. Constrained Optimisation and Lagrange Multipliers

Many business problems contain an equality constraint. Suppose an objective f(x,y) is optimised subject to

where c is a fixed resource level. The Lagrangian is

The sign convention is arbitrary if used consistently. The first-order conditions are

Equivalently,

This means that at an interior constrained optimum, the objective’s gradient is parallel to the constraint’s gradient. The objective cannot improve through a small movement that remains on the constraint surface.

The multiplier λ is often interpreted as a shadow price: the approximate change in the optimised objective resulting from a one-unit increase in the constraint value, under regularity conditions and a stable optimum.

Worked Example 3: Utility under a Budget

A consumer maximises

subject to

where prices are 2 and 4 monetary units and the budget is 40. Form the Lagrangian:

The first-order conditions are

From the first two equations, y=2λ and x=4λ, so x=2y. Substitution into the budget gives

so 8y=40, y=5, and x=10. Utility is therefore U=50. The consumer spends 20 on each good. The result is consistent with the tangency condition

because U_x=y=5, U_y=x=10, and p_x/p_y=2/4=0.5.

9. Resource Allocation and Shadow Prices

Consider a firm allocating a fixed budget between two activities. Let x and y represent activity levels, with benefit B(x,y). If resource costs are a and b, the constraint is

At an interior optimum, the marginal benefit per unit of resource is equalised:

This condition is equivalent to the Lagrange first-order conditions. If one activity has a greater marginal benefit per resource unit, reallocating resources toward it can improve the objective until equality is restored or a boundary is reached.

For multiple constraints, write

Each multiplier measures the marginal value of relaxing its associated constraint, subject to the usual regularity and local-optimum assumptions.

10. Inequality Constraints and Boundary Solutions

Real decisions commonly have constraints such as x≥0, capacity limits, or minimum service requirements. Equality-only Lagrange analysis may miss a boundary optimum.

For inequality-constrained problems, the Karush–Kuhn–Tucker framework extends Lagrange multipliers. For a maximisation problem with constraints written as g_i(x)≥0, nonnegative multipliers and complementary slackness are used. Conceptually, complementary slackness states that a constraint has a nonzero shadow value only when it binds; a slack constraint generally has zero marginal value in the local model.

At a boundary, the firm may prefer more of an activity but be prevented by a non-negativity, capacity, or regulatory limit. Analysts should therefore inspect both interior candidates and all economically relevant boundaries.

11. Illustrative Case: Product-Mix Allocation

Background. A fictional manufacturer produces products A and B. Let x and y denote weekly output in hundreds of units. The estimated contribution is

The factory has a labour constraint

and non-negativity restrictions x≥0 and y≥0. The coefficients are illustrative contribution units and should be replaced by validated operational estimates in practice.

Problem. Determine the product mix that maximises contribution subject to labour availability.

Analysis. The Lagrangian is

The first-order conditions are

From the first equation, λ=30-x. From the second, λ=80-4y. Equating these gives x-4y=-50, or x=4y-50. Substituting into the resource constraint:

Thus 9y=130, so y=130/9≈14.44 and x=70/9≈7.78. Both are non-negative and satisfy the constraint.

The objective is strictly concave because its Hessian is

which is negative definite. Therefore, on the feasible line, the interior solution is the unique global maximum. Contribution is approximately

Outcome. The firm should produce approximately 778 units of A and 1,444 units of B per week, subject to the interpretation that x and y are measured in hundreds. The multiplier is λ=30-70/9=200/9≈22.22, indicating that a small increase in the labour-resource total is estimated to increase optimal contribution by approximately 22.22 contribution units per additional resource unit, locally.

Learning points. The optimal mix is not determined by unit contribution alone. It reflects marginal contribution, diminishing returns, and resource consumption. The result should be stress-tested against uncertain demand, integer production requirements, setup costs, and possible capacity constraints.

12. Practical Modelling, Technology, and Governance

Spreadsheet and statistical software can calculate derivatives symbolically or numerically, solve nonlinear programs, and perform sensitivity analysis. Automatic differentiation is useful in modern optimisation and machine-learning systems because it evaluates gradients accurately through computational models. However, software does not validate economic assumptions.

Best practices include:

• define variables, units, and feasible ranges before calculating;
• distinguish estimated relationships from causal claims;
• validate functions with historical and operational data;
• inspect boundaries and constraint feasibility;
• test sensitivity to prices, costs, coefficients, and forecasts;
• document rounding, solver tolerances, and starting values;
• monitor whether optimisation creates unfair allocation or discriminatory outcomes;
• preserve audit trails when recommendations affect employees, customers, or access to essential services.

AI-assisted optimisation can scale scenario analysis, but governance is necessary. Training data may encode historical bias, and a mathematically optimal allocation may conflict with legal duties, safety requirements, sustainability targets, or stakeholder commitments. Human review and transparent constraint definitions remain essential.

13. Limitations and Common Errors

Multivariable calculus is powerful but local and model-dependent. Common limitations include:

1. Incorrect ceteris paribus interpretation. A partial derivative holds other model variables fixed, whereas real variables may change together.
2. Local rather than global conclusions. A stationary point is not automatically a global optimum.
3. Unrealistic smoothness. Prices, staffing levels, and production quantities may be discrete or involve fixed costs.
4. Parameter uncertainty. Estimated coefficients may be unstable or biased.
5. Ignored constraints. Capacity, legal, ethical, and operational restrictions can invalidate an unconstrained solution.
6. Extrapolation. A derivative near observed data may be unreliable far outside the data range.
7. Unit errors. Mixing dollars with thousands of dollars, or units with hundreds of units, can distort results.

Chapter Summary

Multivariable functions represent business outcomes depending on several inputs. Partial derivatives measure local marginal effects while holding other variables fixed. The gradient collects first derivatives and identifies the direction of greatest increase. Second derivatives and cross-partials describe curvature and interactions. The Hessian provides a systematic local classification of stationary points.

Unconstrained optimisation requires solving the gradient equations and checking feasibility and curvature. Lagrange multipliers solve equality-constrained problems by equating the objective gradient with a multiplier-scaled constraint gradient. Multipliers have an important managerial interpretation as local shadow prices. Inequality constraints require boundary analysis and, more generally, KKT conditions. In all applications, mathematical results must be combined with data validation, sensitivity analysis, operational judgement, and ethical governance.

Key Terms

Partial derivative: Derivative with respect to one variable while other variables are held constant.

Gradient: Vector of first-order partial derivatives.

Directional derivative: Rate of change in a specified direction.

Stationary point: Point where all first partial derivatives are zero.

Marginal product: Change in output from a small input change, holding other inputs fixed.

Cross-partial derivative: Derivative measuring how one variable changes the marginal effect of another.

Hessian: Matrix of second-order partial derivatives.

Concavity: Curvature property associated with maximisation and diminishing marginal effects.

Convexity: Curvature property associated with minimisation and certain feasible sets.

Saddle point: Stationary point that is neither a local maximum nor a local minimum.

Unconstrained optimisation: Optimisation without explicit equality or inequality restrictions.

Constraint: Mathematical representation of a limitation on feasible decisions.

Lagrangian: Objective combined with constraints and multiplier terms.

Lagrange multiplier: Variable measuring the local marginal value of relaxing a constraint.

Shadow price: Economic interpretation of a constraint multiplier.

Feasible set: All decisions satisfying the stated restrictions.

Complementary slackness: KKT condition linking a constraint’s slackness to its multiplier.

Sensitivity analysis: Study of how results change when inputs or parameters change.

Review Questions

Short Answer Questions

1. What does a partial derivative measure in a two-variable business model?
2. What information is contained in a gradient?
3. Why is a stationary point not necessarily a maximum?
4. State the two-variable Hessian classification conditions for a local maximum.
5. What is the managerial interpretation of a Lagrange multiplier?

Long Answer Questions

1. Explain how partial derivatives, cross-partials, and second derivatives can be used to analyse a production function.
2. Describe a complete procedure for solving an unconstrained two-variable profit-maximisation problem.
3. Derive the Lagrange first-order conditions for maximising a benefit function subject to one resource constraint.
4. Explain why boundary analysis is important in business optimisation.
5. Discuss the assumptions and limitations of using calculus-based optimisation in operational decision-making.

Analytical / Application Questions

1. A firm has profit π(x,y)=50x+40y-x²-y²-xy. Find the stationary point and classify it using the Hessian.
2. Maximise U(x,y)=xy subject to 3x+6y=60, with x,y≥0. Interpret the allocation.
3. A manager receives a shadow price of 18 for an additional unit of a scarce resource. Explain what this value means, when it is useful, and why it should not automatically be interpreted as a long-run causal effect.

Multiple Choice Questions with Answers

Question 1

For a differentiable function f(x,y), what does f_x measure?

A. The change in f when both variables increase equally

B. The change in f with y held constant

C. The maximum value of f

D. The curvature of the feasible set

Correct Answer: B. The change in f with y held constant

Question 2

If ∇f=(3,4), what is the magnitude of the gradient?

A. 1

B. 5

C. 7

D. 12

Correct Answer: B. 5

Question 3

For a stationary point of a two-variable function, D=f_xxf_yy-(f_xy)²<0 implies what classification?

A. Local maximum

B. Local minimum

C. Saddle point

D. Global maximum

Correct Answer: C. Saddle point

Question 4

In a Lagrange multiplier problem, the condition ∇f=λ∇g indicates that the two gradients are:

A. Perpendicular

B. Parallel

C. Equal to zero in every case

D. Unrelated

Correct Answer: B. Parallel

Question 5

A production function has Q_L=6 at a particular operating point. What is the local interpretation?

A. Six units of labour are required for one unit of output

B. One additional unit of labour approximately increases output by six units, holding other inputs fixed

C. Output must equal six

D. Labour cost is six monetary units

Correct Answer: B. One additional unit of labour approximately increases output by six units, holding other inputs fixed

Question 6

For U(x,y)=xy, what is U_x?

A. x

B. y

C. xy

D. 1/x

Correct Answer: B. y

Question 7

If a constraint is slack and its multiplier is zero, the usual local interpretation is that:

A. The constraint is binding and highly valuable

B. Relaxing it has no first-order effect on the optimum

C. The objective is necessarily maximised globally

D. The variables must be negative

Correct Answer: B. Relaxing it has no first-order effect on the optimum

Question 8

For the profit function π=40x+30y-x²-y²-xy, what is π_xy?

A. −2

B. −1

C. 1

D. 2

Correct Answer: B. −1

Question 9

Which condition is essential when interpreting a Lagrange multiplier as a shadow price?

A. The solution and model remain locally valid under a small constraint change

B. All variables must be integers

C. The objective must be linear

D. The constraint must be ignored

Correct Answer: A. The solution and model remain locally valid under a small constraint change

Question 10

If B_x/a>B_y/b in a resource-allocation problem, the marginal-benefit condition suggests allocating relatively more resource toward:

A. Activity x

B. Activity y

C. Neither activity under all circumstances

D. The activity with the larger cost only

Correct Answer: A. Activity x