Optimization and Lagrange Multipliers Explained: Choosing the Best Option Under a Constraint

Understanding optimization and Lagrange multipliers — simply put

Almost every question in economics is a question about doing the best you can with limited means. A household has a fixed income and wants the most satisfaction from it. A firm wants to produce a given output at the lowest cost. A government wants to meet a climate target without spending more than it must. Each of these is a constrained optimization problem: an objective to maximize or minimize, and a constraint that limits the choices available.

The Lagrange multiplier method is the standard tool for solving these problems. It folds the constraint into the objective, so that a problem with a restriction becomes an ordinary problem of finding where some derivatives equal zero. Its by-product, the multiplier itself, turns out to be one of the most useful numbers in economics: the value of relaxing the constraint by one unit.

How does it actually help?

Take a government that must cut emissions by a fixed total amount R and can split that cut across sectors such as power generation, industry, buildings and transport. Cutting emissions costs money, and each additional tonne cut in a sector costs more than the last. Let $a_i$ be the tonnes cut in sector i and let $C_i(a_i)$ be the cost of that cut. The government's problem is to choose the cuts that minimize total cost while still adding up to R:

The Lagrange method writes a single function, the Lagrangian, that combines the objective and the constraint:

Setting the derivative of the Lagrangian with respect to each $a_i$ equal to zero gives $C_i'(a_i)$ = $\lambda$ for every sector. $C_i'(a_i)$ is the marginal abatement cost, the cost of cutting one more tonne in sector i. The condition says the cheapest split is the one where every sector's last tonne costs the same, and that common cost is $\lambda$. If one sector's last tonne cost more than another's, moving a tonne of the cut from the expensive sector to the cheap one would lower the total bill without changing the total cut.

Why does this matter?

The equal-marginal-cost rule is the logic behind carbon pricing, emissions trading, and many environmental regulations. It also drives results well beyond climate policy. A consumer maximizing utility on a budget equates the satisfaction from the last euro spent on each good. A firm minimizing cost equates the output gained from the last euro spent on each input. Every one of these results comes from the same first-order conditions.

The multiplier matters too. $\lambda$ measures how much the minimum total cost rises if the required cut R rises by one tonne. That makes it the shadow price of the constraint: the price the constraint would have if it were traded in a market. Regulators use shadow prices to judge whether a target is cheap or expensive to tighten, and to compare targets across sectors.

What makes the method reliable?

The method finds the right answer when the objective and the constraint are smooth, when the constraint actually binds, and when the problem is shaped so that a point satisfying the first-order conditions is a minimum rather than a maximum or a saddle. With costs that rise ever more steeply as cuts deepen, those conditions hold. The full list, and what fails without each one, follows the worked example.

The takeaway

Constrained optimization turns "the best we can do with what we have" into equations. The Lagrange multiplier solves them and, in the same step, prices the constraint. The rest of this guide applies the method to Germany's 2030 climate target, then sets out the mechanics in full.

Solving it by hand: two sectors

Before the real data, the method can be worked through on numbers small enough to check by hand. Two sectors must cut R = 60 Mt between them. Sector 1 has cost slope $c_1 = 1$ and sector 2 has $c_2 = 5$, so their costs are $C_1 = \tfrac{1}{2}a_1^2$ and $C_2 = \tfrac{5}{2}a_2^2$, in millions of euros when the cuts are in Mt. The Lagrangian is:

Step 1: differentiate and set each derivative to zero. $\partial\mathcal{L}/\partial a_1 = a_1 - \lambda = 0$, $\partial\mathcal{L}/\partial a_2 = 5a_2 - \lambda = 0$, and $\partial\mathcal{L}/\partial\lambda = -(a_1 + a_2 - 60) = 0$, which is the constraint itself.

Step 2: write each cut in terms of the multiplier. $a_1 = \lambda$ and $a_2 = \lambda/5$.

Step 3: substitute into the constraint. $\lambda + \lambda/5 = 60$, so $1.2\lambda = 60$ and $\lambda = 50$. The cuts are $a_1 = 50$ Mt and $a_2 = 10$ Mt.

Step 4: check. The last tonne costs 1 × 50 = €50 in sector 1 and 5 × 10 = €50 in sector 2, so marginal costs are equal. Total cost is $\tfrac12 \times 50^2 + \tfrac52 \times 10^2 = 1{,}250 + 250$ = €1,500 million, which equals $\lambda R/2 = 50 \times 60/2$. This shortcut holds whenever every cost is quadratic: each sector's cost, $\tfrac{c_i}{2}a_i^2$, is half its cut times its marginal cost $\lambda$.

Step 5: second-order check. The bordered Hessian puts the constraint's derivatives (1, 1) in its border and the second derivatives of the Lagrangian, $\partial^2\mathcal{L}/\partial a_1^2 = 1$ and $\partial^2\mathcal{L}/\partial a_2^2 = 5$, inside, so its rows are (0, 1, 1), (1, 1, 0) and (1, 0, 5). Expanding along the first row, its determinant is 0 − 1 × (5 − 0) + 1 × (0 − 1) = −6. A negative determinant marks a constrained minimum, so the split is the cheapest one, not the most expensive.

A real example: splitting Germany's emissions cut

Germany's Federal Climate Change Act requires greenhouse gas emissions to fall by 65% from their 1990 level by 2030, with greenhouse gas neutrality by 2045. An amendment in force since 17 July 2024 made the cross-sector total, assessed over several years, the main yardstick for whether the country is on track. The Federal Government said the change would concentrate cuts where the savings potential is largest, so that the targets are met in an economically efficient way. Sector emissions are still reported, and each ministry must still contribute.

The German Environment Agency (UBA) put total emissions at 648.9 million tonnes of CO₂ equivalent (Mt) in 2025, about 48% below 1990. It calculated that from 2026 emissions would need to fall by about 42 Mt a year on average to reach the 2030 target. Five years at that pace is a cut of about R = 210 Mt.

The four largest sectors emitted 582.9 Mt between them in 2025. The illustrative model below assigns the whole 210 Mt cut to those four sectors and holds the remaining sectors, mainly agriculture and waste, unchanged; together they account for the other 66 Mt, the difference between the total and the four sectors.

Each sector's cost of abatement is assumed to rise with the square of the cut, $C_i(a_i) = \tfrac{c_i}{2}\,a_i^2$, so its marginal cost is $c_i a_i$. The cost slopes $c_i$ are illustrative assumptions, chosen so that marginal cost rises fastest in transport and buildings, where cuts depend on replacing vehicles and heating systems, and slowest in the energy sector, which covers power and heat generation. They are not official estimates.

Each cost slope is measured in euros per tonne for each million tonnes cut. With a slope of 2, the 10th million tonne cut in a sector costs €20 a tonne and the 50th costs €100 a tonne.

With this cost function the first-order condition $c_i a_i = \lambda$ gives $a_i = \lambda / c_i$. Adding up across sectors and using the constraint gives $\lambda = R \big/ \textstyle\sum_i (1/c_i)$.

With $\textstyle\sum_i (1/c_i) = 1 + 0.5 + 0.25 + 0.2 = 1.95$, the multiplier is $\lambda = 210/1.95$ ≈ €108 per tonne. Every sector cuts until its last tonne costs €108. The energy sector, the cheapest to decarbonize in this model, takes more than half of the cut; transport, the most expensive, takes about a tenth.

Reading the result

The alternative is to ask every sector for the same proportional cut. Cutting 36.0% everywhere meets the same 210 Mt total, but the marginal costs no longer match.

Under the equal-percentage split the last tonne cut in transport costs €264 while the last tonne in the energy sector costs €68. Moving a tonne of the cut from transport to the energy sector saves about €195. The total cost of the equal-percentage split is about €14.7 billion; the least-cost split costs about €11.3 billion. The €3.4 billion difference, 23% of the equal-percentage cost, is the saving from letting marginal costs equalize.

The multiplier has a second reading. In the least-cost solution total cost is $C^* = \lambda R/2$ ≈ €11.3 billion. Raising the required cut from 210 to 211 Mt raises C* by about €108 million, which is $\lambda$ multiplied by one million tonnes. That is the shadow price at work: the multiplier tells a policymaker what one more unit of ambition costs, without re-solving the whole problem.

Two cautions apply to these numbers. First, the size of the saving depends on the assumed cost slopes. With slopes that differ less across sectors (1, 1.5, 2 and 2.5) the least-cost split saves 6.6% rather than 23%; with the same slope in every sector it still saves 4.1%, because an equal-percentage rule asks the largest emitter for more tonnes than equal marginal costs would. Only the relative slopes matter for the percentage: doubling every slope doubles both costs and leaves the saving unchanged. The rule of equalizing marginal costs holds whatever the slopes are; how much it saves is an empirical question that needs estimated cost curves.

Second, the model is static: it compares costs in a single year at the end point. The 2024 amendment gives flexibility across years as well as across sectors, and choosing when to cut is a dynamic optimization problem that this one-period model does not capture.

The assumptions

The objective and constraint are differentiable. The method works by setting derivatives to zero, so the cost functions and the constraint must have well-defined slopes. If a cost function has a kink, for example a fixed charge that applies only once a technology is adopted, the first-order conditions can miss the true minimum.

The constraint qualification holds. The gradient of the constraint must not be zero at the solution. Geometrically, the method finds the point where the objective's level curve just touches the constraint, and that touching point is defined by the constraint's direction. If the constraint has no well-defined direction at the optimum, the multiplier may not exist even though a solution does.

The second-order condition holds. The first-order conditions identify every point where the objective is flat along the constraint, which includes maxima, minima and saddle points. For a minimum, the objective must curve upward along the constraint. Rising marginal costs, as in the example, guarantee this; with falling marginal costs, for instance strong learning effects, the same conditions would pick out the most expensive split instead of the cheapest.

The solution is interior, or inequality constraints are handled explicitly. The basic method assumes every choice variable can move freely in both directions at the optimum. If the least-cost cut for a sector exceeded its total emissions, the constraint $a_i \le e_i$ would bind, where $e_i$ is the sector's current emissions, and the answer would require the Karush-Kuhn-Tucker conditions set out below.

The formula

The derivations below go beyond what is needed for the intuition, but they are the version used in graduate courses and research.

One equality constraint. Let $x = (x_1, \dots, x_n)$ be the choice variables, f(x) the objective and g(x) = b the constraint, where b is a constant. The Lagrangian is:

The first-order conditions set every partial derivative of the Lagrangian to zero:

The vector condition has a geometric meaning. At the optimum the gradient of the objective is parallel to the gradient of the constraint, so the level curve of f is tangent to the constraint. If the two gradients pointed in different directions, there would be a direction along the constraint in which f still improves.

The multiplier as a shadow price. Let V(b) = f(x*(b)) be the optimal value as a function of the constraint level. Differentiating and using the first-order conditions gives the envelope theorem result:

In the emissions example, $V(R) = R^2 \big/ \big(2\textstyle\sum_i 1/c_i\big)$, so $dV/dR = R \big/ \textstyle\sum_i (1/c_i) = \lambda$, matching the €108 per tonne found above.

Second-order condition. Whether a point satisfying the first-order conditions is a maximum or a minimum depends on the bordered Hessian, the matrix of second derivatives of the Lagrangian bordered by the constraint's gradient. With two choice variables:

A determinant of $\bar{H}$ greater than zero identifies a local maximum and a determinant less than zero identifies a local minimum. With n variables and m constraints, the last n − m leading principal minors of $\bar{H}$ decide (a leading principal minor is the determinant of a top-left square block of the matrix; with three variables and one constraint, the last two are the 3×3 block and the full 4×4 determinant): for a minimum they all share the sign of $(-1)^m$; for a maximum they alternate in sign, with det $\bar{H}$ itself taking the sign of $(-1)^n$.

In the two-sector example, $\bar{H}$ has rows (0, 1, 1), (1, $c_1$, 0) and (1, 0, $c_2$). Its determinant is $-(c_1 + c_2) = -6$, which is negative, confirming a minimum.

Sign conventions differ between textbooks. Some write the Lagrangian as $f(x) + \lambda\,(b - g(x))$, which is identical to the form used here, and some as $f(x) + \lambda\,(g(x) - b)$, which flips the sign of λ. The optimal choice x* is the same under every convention; only the sign of the multiplier changes.

Several constraints. With m equality constraints $g_j(x) = b_j$, there is one multiplier per constraint:

Inequality constraints: the Karush-Kuhn-Tucker conditions. Many economic constraints are inequalities: spending cannot exceed income, abatement cannot exceed emissions. For the problem of maximizing f(x) subject to g(x) ≤ b, the Karush-Kuhn-Tucker (KKT) conditions are:

Complementary slackness says that either the constraint binds, or its multiplier is zero. A constraint that does not bind has a shadow price of zero: relaxing it changes nothing. In the emissions example, if a sector's least-cost cut exceeded its emissions, that sector would cut its emissions to zero, its upper limit would carry a positive multiplier, and the remaining sectors would share the rest of the cut at a higher common marginal cost.

When this isn't the right tool

Non-convex problems. The first-order conditions find local optima. When costs fall with scale, or choices are lumpy (build a plant or not), a local optimum need not be the global one, and integer or global optimization methods are needed.

Decisions over time. Climate targets, saving and investment unfold over many periods, and today's choice changes tomorrow's options. Static Lagrange problems treat each period separately; dynamic optimization, using optimal control or dynamic programming, handles the link between periods.

Uncertain costs. The cost slopes in the example were assumed. Real marginal abatement costs are estimated with error, and the least-cost split is only as good as those estimates. Comparing outcomes under several cost assumptions, or estimating cost curves with regression methods, is part of any serious application.

Who pays. The method minimizes the total cost and says nothing about who bears it. Two splits with the same total cost can have very different effects on households, regions and industries.

Practice problems

Problem 1. Two firms must together cut emissions by 30 tonnes. Firm 1's abatement cost is $C_1 = a_1^2$ and firm 2's is $C_2 = 2a_2^2$. Find the least-cost split, the multiplier and the total cost, and check that the multiplier is close to the extra cost of a 31st tonne.

Solution. The first-order conditions are $2a_1 = \lambda$ and $4a_2 = \lambda$, so $a_1 = 2a_2$. The constraint $a_1 + a_2 = 30$ gives $a_2 = 10$, $a_1 = 20$ and $\lambda = 40$. Total cost is 400 + 200 = 600. For a general total R the cuts are 2R/3 and R/3, so cost is $C^*(R) = 2R^2/3$ and $dC^*/dR = 4R/3 = 40$ at R = 30, matching λ. The exact extra cost of the 31st tonne, $C^*(31) - C^*(30) = 2(31^2 - 30^2)/3 \approx 40.7$, is close to $\lambda$ because $\lambda$ measures the cost of an arbitrarily small increase.

Problem 2. A consumer maximizes $u = x^{0.5} y^{0.5}$ subject to 2x + 4y = 80. Find the optimal bundle and the Lagrange multiplier, and interpret the multiplier.

Solution. The first-order conditions are $0.5\,x^{-0.5}y^{0.5} = 2\lambda$ and $0.5\,x^{0.5}y^{-0.5} = 4\lambda$. Dividing one by the other gives y/x = 1/2, so 2x = 4y. Substituting into the budget gives x = 20 and y = 10. Then $\lambda = 0.5\sqrt{10/20}/2 \approx 0.177$. The multiplier is the marginal utility of income: one more unit of income raises maximum utility by about 0.177. Check: 2 × 20 + 4 × 10 = 80, and maximum utility $\sqrt{200} \approx 14.14$ divided by income 80 is also 0.177, as it must be when utility has constant returns to scale.

Further reading

• Comparative Statics Explained — internal
• Linear Algebra for Economists Explained — internal
• Consumer Theory and Duality Explained — internal
• Producer Theory and Cost Functions Explained — internal
• General Equilibrium and the Welfare Theorems Explained — internal
• OLS Regression Explained — internal
• UBA and Federal Environment Ministry, greenhouse gas data for 2025 (14 March 2026) — external reference
• German Federal Government, the amended Federal Climate Change Act — external reference

Emissions for 2025 and the required average annual fall are from the German Environment Agency's March 2026 release; the 2030 target and the 2024 amendment are from the Federal Government. The sector cost slopes are illustrative assumptions, and the cost and multiplier figures are EconoLens calculations from that model, not official estimates.