Multivariable extrema and constrained optimisation
Introduced| English | Meaning |
|---|---|
| saddle/ˈsædl/ | A stationary point with values both above and below its value arbitrarily nearby. |
| Lagrange multipliers/ˈlæɡreɪndʒ ˌmʌltɪˈplaɪəz/ | A method for finding regular constrained-extremum candidates by making the objective gradient a multiple of the constraint gradient. |
| critical point/ˈkrɪtɪkl pɔɪnt/ | A point in the domain where the gradient is zero or fails to exist. |
A flat point can be anything
- The surfaces $x^2+y^2$, $-x^2-y^2$ and $x^2-y^2$ all have a horizontal tangent plane at the origin. Yet the origin is respectively a minimum, a maximum and a saddle 鞍点. A zero gradient finds a candidate; it does not classify it.
- We will use second derivatives, direct inequalities and Lagrange multipliers 拉格朗日乘数 to decide which values are actually attained. The allowed domain matters: an ambient saddle can become a minimum after imposing a constraint.
Local, global and strict are different claims
- A local minimum has no smaller value at allowed points sufficiently near it; a global minimum has no smaller value anywhere in the domain. A strict minimum also excludes equal values at distinct nearby points. Reverse the inequalities for maxima. For example, $x^2$ on the plane has non-strict minima along the entire $y$-axis.
- An interior critical point 临界点 is a point in the domain where the gradient is zero or fails to exist. A stationary point specifically has gradient zero. The function $\sqrt{x^2+y^2}$ has a strict global minimum at the origin, where its gradient does not exist. Such singular candidates need direct inspection.
Find candidates before classifying them
- At a differentiable interior local extremum, Fermat's condition gives $f_x=f_y=0$. Differentiability and interiority matter: a boundary extremum need not have zero gradient. For $f=x^2+xy+y^2$, the equations $2x+y=0$ and $x+2y=0$ give only the origin.
- A positive-looking Hessian at a nonstationary point gives no extremum test. For $f=x+x^2+y^2$, the origin has positive second derivatives but $f_x(0,0)=1$. Values on the $x$-axis are both above and below $f(0,0)$ arbitrarily nearby, so the origin is not an extremum.
For $f=x+x^2+y^2$, positive definiteness of the Hessian at the origin proves a local minimum there.
Difficulty 1/5 · 100 coins + a small bonus on first completion today
The gradient at the origin is (1,0), so the point is not stationary. On the x-axis, x+x² takes both signs arbitrarily near zero. The stationary-point hypothesis is essential.
The Hessian test needs smoothness and stationarity
- Assume $f$ has continuous second partial derivatives on a neighbourhood of a stationary point. At that point, write $A=f_{xx}$, $B=f_{xy}$ and $C=f_{yy}$; the Hessian is the symmetric matrix with rows $(A,B)$ and $(B,C)$. Its determinant is $D=AC-B^2$.
- If $D>0$, $A>0$ gives a strict local minimum and $A<0$ gives a strict local maximum. If $D<0$, the quadratic term takes both signs and the point is a saddle. If $D=0$, this test is inconclusive. For $x^2+xy+y^2$, $A=C=2$, $B=1$ and $D=3$, giving a strict minimum.
For $f = x^2 + xy + y^2$, compute $D = f_{xx}f_{yy} - f_{xy}^2$ at the origin.
Difficulty 2/5 · 140 coins + a small bonus on first completion today
$2 \cdot 2 - 1$; with $f_{xx} > 0$, a minimum.
Mixed terms can overturn the pure partials
- For $f=x^2+4xy+y^2$, $A=C=2$ and $B=4$, so $D=-12$. Along $y=x$ the value is $6x^2>0$ for $x\ne0$; along $y=-x$ it is $-2x^2<0$. These two paths verify the saddle even though both pure second partials are positive.
- For the original exercise $f=3x^2-6xy+y^2$, the gradient vanishes at the origin and $D=6\cdot2-(-6)^2=-24$. Along $y=0$ values are positive; along $y=3x$ they are $-6x^2$. Again the mixed term matters; inspecting only $f_{xx}$ and $f_{yy}$ would misclassify the point.
The origin is a critical point of $f = 3x^2 - 6xy + y^2$. It is
Difficulty 1/5 · 100 coins + a small bonus on first completion today
$D = 6 \cdot 2 - 36 < 0$.
Classify the origin for $f=x^2+4xy+y^2$. Verify stationarity, compute all Hessian entries and $D$, then confirm the verdict with two paths. Explain why the pure second partials alone are insufficient.
Compare your argument with the model and checklist.
A zero determinant calls for another argument
- At the origin, $x^4+y^4$ has a strict global minimum by nonnegativity, while $-x^4-y^4$ has a strict global maximum. Both have zero Hessian there. The function $x^4-y^4$ instead takes opposite signs along the coordinate axes, so it has a saddle with the same determinant $D=0$.
- A neighbourhood inequality can prove a minimum or maximum. Two paths with opposite signs can disprove either kind of extremum. But checking a collection of paths and finding only positive values does not control every nearby point; even all straight lines can miss a curved region of negative values.
If $D = 0$ at a critical point, the second-derivative test is inconclusive.
Difficulty 1/5 · 100 coins + a small bonus on first completion today
True. $x^4 + y^4$ (minimum), $-x^4 - y^4$ (maximum) and $x^4 - y^4$ (saddle) all have $D = 0$ at the origin.
Every straight line can miss the problem
- Let $F(x,y)=(y-x^2)(y-2x^2)$. On a fixed line $y=mx$ with $m\ne0$, this becomes $x^2(m-x)(m-2x)>0$ for sufficiently small nonzero $x$. On $y=0$ it is $2x^4>0$, and on the vertical line $x=0$ it is $y^2>0$. Each line separately has a strict minimum at the origin.
- The origin is stationary with $D=0$. On the curved path $y=3x^2/2$, however, $F=-x^4/4<0$ for $x\ne0$. Positive values also occur arbitrarily near the origin, so it is a saddle. The small neighbourhood allowed by a line check can depend on its slope; it does not give one neighbourhood covering every direction and curved path.
What is at the origin?
Classify critical points, including the D = 0 case.
If a smooth function has a strict minimum at the origin on every straight line through it, then it has a local minimum in the plane.
Difficulty 1/5 · 100 coins + a small bonus on first completion today
For F=(y-x²)(y-2x²), every fixed line has a strict minimum at zero, but on y=3x²/2 the value is -x⁴/4. No single neighbourhood is controlled by all separate line checks.
For $G(x,y)=(y-2x^2)(y-3x^2)$, show that each straight-line restriction through the origin has a strict minimum there, yet the origin is not a local minimum in the plane.
Compare your argument with the model and checklist.
Regular constraints lead to multiplier candidates
- Suppose $f$ and $g$ have continuous first partial derivatives near a constrained local extremum on $g(x,y)=c$, and $\nabla g\ne0$ there. The necessary condition is $\nabla f=\lambda\nabla g$, together with $g=c$. The multiplier $\lambda$ is an unknown; the extremum value is $f(x,y)$.
- At a regular constrained extremum, the derivative of $f$ along the constraint's tangent is zero. Thus $\nabla f$ is parallel to the normal $\nabla g$, including the possibility $\nabla f=0$. These equations produce candidates; feasibility, singular points, restricted endpoints and global behaviour still need checks.
Singular feasible points must not disappear
- Take $f(x,y)=x$ subject to $g(x,y)=x^2+y^2=0$. The only feasible point is the origin, so it is both a constrained global minimum and maximum, with value zero. But $\nabla g=0$ there.
- The multiplier equation would demand $(1,0)=\lambda(0,0)$, which has no solution. This does not disprove the extrema; the regularity assumption failed. Add singular feasible points to the candidate list and analyse them directly rather than dividing by a gradient component that may vanish.
If the Lagrange equations have no solution, a constrained extremum cannot exist.
Difficulty 1/5 · 100 coins + a small bonus on first completion today
For f=x under x²+y²=0, the origin is the only feasible point and is both a global minimum and maximum. The constraint gradient vanishes, so the regular multiplier theorem does not apply.
Prove the original line-constrained minimum
- Minimise $f=x^2+y^2$ subject to $x+y=6$. Since $\nabla g=(1,1)\ne0$, the multiplier equations give $2x=\lambda=2y$, hence $x=y=3$ and value $18$. Substitution gives $f(x,6-x)=2(x-3)^2+18$, hence $f\ge18$ on the whole line.
- There is no maximum because the feasible points $(3+t,3-t)$ have values $18+2t^2\to\infty$. The unbounded line alone would not prove this: $-(x^2+y^2)$ on the same line has a global maximum $-18$. Inspect the objective's actual growth along feasible points.
A function on an unbounded constraint set cannot attain a global maximum.
Difficulty 1/5 · 100 coins + a small bonus on first completion today
The function -(x²+y²) on x+y=6 has a global maximum -18 at (3,3), despite the unbounded line. One must inspect the objective, not only the domain.
A concave restriction gives a maximum
- For $f=xy$ subject to $x+y=10$, regular multipliers give $y=\lambda=x$, hence $(5,5)$. On the constraint, $f=x(10-x)=25-(x-5)^2$. Therefore the global maximum is $25$, attained only at $(5,5)$.
- There is no minimum on the whole line, since $(5+t,5-t)$ gives $25-t^2\to-\infty$. If additional bounds such as $x,y\ge0$ were imposed, the feasible set would instead be a closed segment; its endpoints would become part of the search. State the whole constraint set before claiming a global result.
Maximise $xy$ subject to $x + y = 10$.
Difficulty 2/5 · 140 coins + a small bonus on first completion today
Substitute y = 10 − x: xy = 10x − x², maximised at x = 5 (Lagrange gives y = λ = x). Then xy = 5 · 5 = 25.
Compact domains guarantee existence, not location
- A continuous function on a nonempty closed bounded subset of Euclidean space attains a global minimum and maximum. For a region with piecewise smooth boundary, inspect stationary and singular interior candidates, restrict to each boundary piece, and include corners and endpoints. Compare the actual function values.
- The theorem does not locate the extrema, nor does it say they must lie at corners. Without compactness, extrema may still exist, but existence is no longer guaranteed by this theorem. Without continuity, compactness alone is insufficient; both hypotheses do work.
Work every edge of the square
- For $f=x^2-y^2$ on $[-1,1]^2$, the only stationary interior point is the saddle $(0,0)$ with value zero. On the edges $x=\pm1$, $f=1-y^2$, so the largest edge value is $1$ at $y=0$. On $y=\pm1$, $f=x^2-1$, so the smallest edge value is $-1$ at $x=0$.
- All four corners have value zero. Since $0\le x^2,y^2\le1$, the global bounds are $-1\le f\le1$. The maximum is attained at $(\pm1,0)$ and the minimum at $(0,\pm1)$. Corners alone would miss both; an interior saddle does not settle the global problem.
What is the maximum of $x^2 - y^2$ on the square $[-1, 1] \times [-1, 1]$?
Difficulty 2/5 · 140 coins + a small bonus on first completion today
At (±1, 0), on the boundary.
What is the global minimum of $x^2-y^2$ on $[-1,1]^2$?
Difficulty 2/5 · 140 coins + a small bonus on first completion today
The bound x²-y² ≥ -1 is attained at (0,1) and (0,-1). Corners have value zero and would miss this minimum.
A circle constraint makes the candidate comparison finite
- Maximise and minimise $f=x+2y$ on $x^2+y^2=1$. The circle is nonempty and compact, and its constraint gradient never vanishes. Multipliers give $1=2\lambda x$ and $2=2\lambda y$. Thus $\lambda\ne0$, $y=2x$, and the feasible candidates are $(1,2)/\sqrt5$ and $-(1,2)/\sqrt5$.
- Their values are $\sqrt5$ and $-\sqrt5$. Since extrema exist and all candidates have been considered, these are the global maximum and minimum. The inequality $|x+2y|\le\sqrt5\sqrt{x^2+y^2}$ independently confirms the bounds and equality points.
What is the global maximum of $x+2y$ on $x^2+y^2=1$? Give a decimal to three places.
Difficulty 2/5 · 140 coins + a small bonus on first completion today
The maximum is √5 ≈ 2.236 at (1/√5, 2/√5). Cauchy–Schwarz gives |x+2y| ≤ √5 on the unit circle.
The unit circle can be written as $x^2+y^2=1$ or as $(x^2+y^2-1)^2=0$. For $f=x+2y$, explain why the squared representation breaks the regular multiplier equations, then recover the extrema using the regular representation.
Compare your argument with the model and checklist.
Constrained and ambient classifications can differ
- The origin is an ambient saddle of $f=x^2-y^2$. Restricted to $y=0$, the function is $x^2$ and has a strict constrained minimum. Restricted to $x=0$, it is $-y^2$ and has a strict constrained maximum. Each verdict compares only feasible nearby points.
- A full Hessian test in the plane answers the unconstrained question. For a constrained question, inspect the restriction or use an appropriate constrained test; do not reuse the ambient saddle label as the final verdict. Check the domain, regularity, candidate equations and attained values together.
Which extremum conclusions follow?
Check stationarity, regularity and feasible boundaries.
At the origin, how is $f=x^2-y^2$ classified when the allowed points satisfy $y=0$?
Difficulty 1/5 · 100 coins + a small bonus on first completion today
On the constraint the restriction is x², so the origin is a strict constrained minimum. The ambient saddle classification compares points that are not feasible here.
Find all global extrema of $f=x^2-y^2$ on $[-2,1]\times[-1,3]$. Inspect interior candidates, every edge and the corners; give both values and all attaining points.
Compare your argument with the model and checklist.
Recap
- saddle: A stationary point with values both above and below its value arbitrarily nearby.
- Lagrange multipliers: A method for finding regular constrained-extremum candidates by making the objective gradient a multiple of the constraint gradient.
- critical point: An interior point of a function's domain where its gradient is zero or fails to exist.
- Use the Hessian test only under its hypotheses, treat $D=0$ directly, and check every feasible boundary or singular candidate before deciding global extrema.