Solve constrained optimization problems with candidate points, gradient equations, and a contour-style visual.
Results are calculated automatically as you enter data.
| Point | f(x,y) | lambda / normalized residual | Classification |
|---|
▼ See explanations and tips below ▼
Constrained optimization is the process of finding the largest or smallest value of a function while following one or more rules. The function being optimized is called the objective function. The rule that limits the possible inputs is called a constraint.
In ordinary optimization, you might ask where a function \(f(x,y)\) is largest or smallest over all possible values of \(x\) and \(y\). In constrained optimization, you ask a narrower question: where is \(f(x,y)\) largest or smallest among only the points that satisfy an equation such as \(g(x,y)=c\)?
That difference matters because many real problems have limits. A design may need a fixed volume, a budget may be fixed, a point may need to stay on a curve, or two variables may need to satisfy a relationship. Lagrange multipliers provide a way to locate possible maxima and minima without solving the constraint for one variable first.
For functions of two variables with one equality constraint, the key idea is geometric: at a smooth constrained optimum, a level curve of the objective function touches the constraint curve. At that point, the two curves are tangent, and their gradient vectors are parallel.
Lagrange multipliers are useful because they turn a constrained optimization problem into a system of equations. Instead of guessing points on a curve, you set up equations that describe where an optimum could occur.
This is especially helpful in multivariable calculus because constraints often describe curves or surfaces. For example, you might want to:
The method does not automatically prove that a candidate is the absolute maximum or minimum in every situation. It gives candidates that must still be checked, compared, and interpreted in the context of the constraint and domain.
For one equality constraint in two variables, the standard setup is:
with the constraint:
In component form, this means:
Where:
The reason this works comes from level curves. The gradient of a function is perpendicular to its level curves. At a constrained optimum, the objective contour is tangent to the constraint curve. If two smooth curves are tangent, their perpendicular directions are parallel, so the gradients must point in the same or opposite directions. The multiplier \(\lambda\) expresses that parallel relationship.
After solving the equations, you evaluate \(f(x,y)\) at each detected candidate point. The smallest and largest displayed values are comparisons within that detected set, not automatic proofs of global extrema.
With two equality constraints in two variables, the feasible points are usually intersections of two curves:
In that setting, the practical task is to find points that satisfy both equations and then compare \(f(x,y)\) at those feasible points. If the two constraints intersect at several points, each intersection may need to be tested. If they do not intersect in the selected region, there may be no feasible candidate to compare.
In hand calculations, students often compute symbolic partial derivatives and solve exact equations. A numerical calculator works differently. It searches within a chosen square region for points that satisfy the relevant equations, estimates derivatives numerically, and compares the objective values at the detected candidates.
For example, a centered finite-difference estimate of the \(x\)-partial derivative has the form:
A similar formula estimates the \(y\)-partial derivative:
These formulas can be very useful, but they are approximations. They work best when the function is smooth near the point being tested. Sharp corners, discontinuities, undefined values, or very flat derivative behavior can make the numerical result harder to trust.
The calculator uses a scale-aware adaptive search. It samples progressively finer grids, adds starting points where constraint contours cross grid cells, estimates useful multiplier starts, and refines each candidate numerically. Equivalent equations multiplied by a very large or very small nonzero constant are normalized before convergence is tested.
Adaptive search can find substantially more branches than one fixed grid, but it is still not a formal proof that every possible point has been found. A stabilized result means two successive search passes produced the same candidate set. A numerical-limit result means the search changed through its final pass, so the displayed list may be incomplete.
Suppose you want to optimize:
subject to the ellipse:
Let:
The gradients are:
Set \(\nabla f=\lambda \nabla g\):
From these equations:
Substitute into the constraint:
So \(\lambda=\sqrt{2}\) or \(\lambda=-\sqrt{2}\). The candidate points are:
and
Evaluate \(f\) at both points:
The largest value is \(2\sqrt{2}\), and the smallest value is \(-2\sqrt{2}\).
Now suppose:
with two constraints:
The second constraint says \(x=y\). Substitute \(y=x\) into the first constraint:
So \(x=1\) or \(x=-1\). Since \(y=x\), the feasible points are:
Evaluate the objective function:
Both feasible points produce the same objective value. This is a useful reminder that a constrained problem can have ties, symmetric candidates, or several points with the same result.
Using the two-constraint example above, the feasible points are \((1,1)\) and \((-1,-1)\). If the selected search box is only \(0 \le x \le 2\) and \(0 \le y \le 2\), then \((-1,-1)\) is outside the region being searched.
In that case, a numerical search may report only \((1,1)\). That does not mean \((-1,-1)\) is mathematically impossible. It only means the point was outside the selected bounds. Widening the search box to include negative coordinates allows both feasible points to be considered.
The main result compares the smallest and largest candidate points detected inside the selected closed search box. “Detected” means the point passed the normalized numerical checks and lies within the selected bounds; it does not mean the result is a symbolic proof of a global maximum or minimum everywhere.
The candidate table gives more detail:
If the solver detects dependent constraints or a non-isolated stationary set, it stops with a degenerate-system diagnosis instead of displaying arbitrary samples as a complete candidate list. An isolated feasible point with a zero constraint gradient may still be displayed, but its local classification is inconclusive.
The contour visualization is a guide to the geometry. Objective contours show where \(f(x,y)\) has similar values, constraint curves show the feasible equations, and candidate markers show detected points. Optional gradient vectors can help illustrate the parallel-gradient idea in one-constraint problems.
Numerical values are rounded for display. Very small values may appear as \(0\), and very large or very small nonzero values may appear in exponential notation. Use the displayed result as an approximation unless you have independently verified an exact symbolic solution.
One common mistake is treating every displayed candidate as an exact answer. Numerical methods estimate derivatives and solve equations approximately. For homework, exams, or formal work, you may still need to show symbolic derivatives, exact algebra, and a justification that the chosen point is truly a maximum or minimum.
Another common mistake is using too narrow a search box. If an important candidate lies outside the selected minimum and maximum bounds, it cannot be found in that search. Choose bounds wide enough to include all relevant parts of the constraint curve.
Expression entry also matters. Explicit multiplication, such as \(2*x\), is the clearest notation; supported implicit products such as \(2x\) are also accepted. Use only the variables \(x\) and \(y\). Use a decimal point, not a decimal comma, for numeric constants and bounds.
It is also easy to confuse equality constraints with inequality constraints. A constraint such as \(x^2+y^2=1\) describes a curve. A constraint such as \(x^2+y^2 \le 1\) describes a filled region. These are different optimization problems, and inequality constraints require additional boundary and interior checks.
Finally, do not ignore smoothness. The Lagrange multiplier condition is based on gradients. If a function has a corner, jump, undefined point, or other nonsmooth behavior, the usual gradient-based condition may miss important behavior or become unreliable.
Use Lagrange multipliers when you have a smooth objective function and one or more equality constraints. In two-variable problems, this often means optimizing \(f(x,y)\) along a curve such as a line, circle, ellipse, or other level curve.
This approach is especially useful when:
For closed and bounded feasible sets, comparing all relevant candidates can help identify absolute extrema. For open, unbounded, or incomplete search regions, additional reasoning is needed.
Lagrange multipliers identify candidate points under smoothness and regularity assumptions. If \(\nabla g=\mathbf{0}\) at a feasible point, or if the constraint is not smooth there, the standard condition may not apply cleanly. Such points may need separate analysis.
For numerical results, keep these limits in mind:
For important academic, engineering, financial, scientific, or safety-related decisions, check the result with exact work, another numerical method, or a qualified professional when appropriate.
Supported expression names include common functions such as \(\sin\), \(\cos\), \(\tan\), \(\sqrt{\phantom{x}}\), \(\log\), \(\exp\), \(\min\), \(\max\), \(\text{abs}\), and constants such as \(\pi\) and \(e\) through the accepted interface names. Variables other than \(x\) and \(y\) are not part of the calculation.
In a one-constraint problem, \(\lambda\) is the scalar that makes the objective gradient parallel to the constraint gradient. It is part of the system used to locate candidate points. In this calculator, \(\lambda\) is shown for one-constraint mode, while two-constraint mode reports a solver residual instead.
The Lagrange equations can produce more than one candidate point. A point may be a maximum, a minimum, or neither, so the objective function must be evaluated at each candidate. The comparison step is what identifies the smallest and largest detected objective values.
The displayed values are numerical approximations. The calculator estimates gradients with finite differences and solves systems numerically, so results can be affected by rounding, search bounds, nonsmooth functions, and solver behavior. For exact answers, verify the result with symbolic calculus when possible.
There may be no feasible point inside the selected search box, the expressions may produce non-finite values, or the adaptive numerical search may not have resolved a candidate. “No candidate detected” is not proof that the feasible set is empty. Check the expressions, widen the bounds, and verify the constraint independently.
No. Inequality constraints such as \(g(x,y) \le c\) require a different analysis because the optimum may occur in the interior, on the boundary, or at special corner-like points. This calculator is designed for equality constraints.
With two equality constraints in two variables, the calculator focuses on finding feasible intersections of the two constraint curves and comparing \(f\) at those points. Because it solves the intersection equations directly in that mode, the table reports residual information rather than a Lagrange multiplier.
Books and Open Textbooks