math.wiki
Calculus › Partial derivatives › Extrema and the second derivative test
Sign in

Extrema and the second derivative test

A smooth function of one variable has a peak or a valley only where its derivative is zero, and the sign of the second derivative tells which. What replaces these tests for a function of several variables, whose graph can rise in some directions and fall in others?

Where you are 8 in Partial derivatives

Why this exists

In one variable, Critical points and local versus absolute extrema placed every interior extreme where f' = 0 or f' fails to exist, The first and second derivative tests classified such points by the sign of f'', and Optimization compared the critical points with the two ends of [a, b]. With two inputs the graph can rise in some directions and fall in others; the boundary of a region is a curve, not two points; and f'' becomes four second partials.

Here an interior extreme has zero gradient (Gradient and directional derivatives), the eigenvalues of the Hessian matrix classify it, and on a closed bounded region the extreme value theorem sends us to the critical points and the boundary.

Priya and Hiro price two flavors at their gelato shop (Example 2).

Definition

Let f : S \to \mathbb{R} with S \subseteq \mathbb{R}^n, and \mathbf{a} \in S. (The letter D is kept for the discriminant.)

  • f has a local maximum at \mathbf{a} if some open ball B about \mathbf{a} has f(\mathbf{x}) \le f(\mathbf{a}) for every \mathbf{x} \in B \cap S; a local minimum reverses the inequality. It is strict if the inequality is strict for \mathbf{x} \ne \mathbf{a}. An absolute maximum (minimum) on S asks the inequality for every \mathbf{x} \in S. Any of these is an extremum.
  • \mathbf{a} is an interior point of S if some open ball about it lies in S. A point is a boundary point of S if every open ball about it contains points of S and points not in S. S is closed if it contains all its boundary points, and bounded if it lies inside some ball.
  • A critical point of f is an interior point where \nabla f = \mathbf{0} or some first partial does not exist. A saddle point is a critical point \mathbf{a} with \nabla f(\mathbf{a}) = \mathbf{0} such that every open ball about \mathbf{a} contains points where f > f(\mathbf{a}) and points where f < f(\mathbf{a}).
  • f is of class C^2 on an open set U if its second partials exist and are continuous on U. Its Hessian matrix H_f(\mathbf{a}) is the n \times n matrix with (i, j) entry f_{x_i x_j}(\mathbf{a}), symmetric (The transpose and symmetric matrices) by Clairaut’s theorem (Partial derivatives). In two variables the discriminant is its determinant:
H_f = \begin{bmatrix} f_{xx} & f_{xy} \\ f_{yx} & f_{yy} \end{bmatrix}, \qquad D = f_{xx} f_{yy} - f_{xy}^2.
  • The quadratic form of the Hessian at \mathbf{a} is
Q_{\mathbf{a}}(\mathbf{h}) = \mathbf{h}^T H_f(\mathbf{a})\,\mathbf{h} = \sum_{i, j} f_{x_i x_j}(\mathbf{a})\,h_i h_j,

the column \mathbf{h} being the vector \langle h_1, \dots, h_n \rangle. In two variables Q(h, k) = f_{xx} h^2 + 2 f_{xy} hk + f_{yy} k^2.

Intuition

At a critical point (a, b) where f is differentiable, the tangent plane is horizontal (Tangent planes and linear approximation), so the next term of the Taylor formula (Result 2) decides:

f(a + h, b + k) \approx f(a, b) + \tfrac12 Q(h, k).

Write A = f_{xx}, B = f_{xy}, C = f_{yy} at (a, b). If A \ne 0, completing the square gives

A\,Q(h, k) = (Ah + Bk)^2 + (AC - B^2)k^2.

If D = AC - B^2 > 0, the right side is positive for every nonzero step, so Q has the sign of A in every direction: near (a, b) the graph is shaped like a bowl, up if A > 0 and down if A < 0, with nearly elliptical level curves (Figure 1). If D < 0, Q has the sign of A for k = 0 and the opposite sign along Ah + Bk = 0: a saddle. If D = 0, Q vanishes along a line of directions and higher terms decide.

Along the line \mathbf{a} + t\mathbf{u}, with \mathbf{u} a unit vector, g(t) = f(\mathbf{a} + t\mathbf{u}) has g'(0) = 0 and g''(0) = Q(\mathbf{u}). The test asks for more than a minimum along every line; Example 4 shows why. On a closed bounded region the boundary takes the place of the ends of an interval.

−12 9 (0, 0) (5, −3)
Figure 1Level curves of f(x, y) = x^2 + xy + y^2 - 6x on R = [0, 5] \times [-3, 0], at the levels \displaystyle -\frac{45}{4}, -9, -6, -2 and 3 from the inside out (Example 3). The absolute minimum -12 is at the interior critical point (4, -2) and the absolute maximum 9 at the corner (0, -3). The small dots are the critical numbers on the edges, where a level curve touches an edge.

Results

Result 1 (Fermat’s theorem). If f has a local extremum at an interior point \mathbf{a} of S and f_{x_i}(\mathbf{a}) exists, then f_{x_i}(\mathbf{a}) = 0. So every interior local extremum is a critical point.

Result 2 (second-order Taylor formula). Let f be of class C^2 on an open ball B about \mathbf{a}, and \mathbf{a} + \mathbf{h} \in B. Then for some c with 0 < c < 1,

f(\mathbf{a} + \mathbf{h}) = f(\mathbf{a}) + \nabla f(\mathbf{a}) \cdot \mathbf{h} + \tfrac12 Q_{\mathbf{a} + c\mathbf{h}}(\mathbf{h}).

If f is a polynomial of degree at most 2, its second partials are constant and the formula is exact with Q_{\mathbf{a}}, for every \mathbf{h}.

Result 3 (Hessian test). Let f be of class C^2 on an open ball about \mathbf{a}, with \nabla f(\mathbf{a}) = \mathbf{0}, and let \lambda_1, \dots, \lambda_n be the eigenvalues of H_f(\mathbf{a}) (real, as the matrix is symmetric).

  • (a) If every \lambda_i > 0, f has a strict local minimum at \mathbf{a}.
  • (b) If every \lambda_i < 0, f has a strict local maximum at \mathbf{a}.
  • (c) If some \lambda_i > 0 and some \lambda_j < 0, \mathbf{a} is a saddle point.
  • (d) Otherwise some \lambda_i = 0, and the test gives no conclusion.

Result 4 (second derivative test). Let f be of class C^2 on an open disk about (a, b), with \nabla f(a, b) = \mathbf{0} and D = f_{xx} f_{yy} - f_{xy}^2 at (a, b).

  • (a) If D > 0 and f_{xx}(a, b) > 0, f has a strict local minimum at (a, b).
  • (b) If D > 0 and f_{xx}(a, b) < 0, f has a strict local maximum at (a, b).
  • (c) If D < 0, (a, b) is a saddle point.
  • (d) If D = 0, the test gives no conclusion.

Result 5 (extreme value theorem). If S is nonempty, closed and bounded and f is continuous on S, then f has an absolute maximum and an absolute minimum on S, each taken at a critical point of f or at a boundary point of S.

Why it is true

Result 1. Take an open ball of radius r about \mathbf{a} inside S on which f \le f(\mathbf{a}) (for a maximum), and let g(t) = f(\mathbf{a} + t\mathbf{e}_i) for \lvert t \rvert < r, with \mathbf{e}_i the ith standard unit vector. Then g has a local maximum at 0, and g'(0) is the limit that defines f_{x_i}(\mathbf{a}), so it exists. Fermat’s theorem in one variable (Critical points and local versus absolute extrema) gives g'(0) = 0. A minimum is the same.

Result 2. Let \mathbf{h} \ne \mathbf{0} and r be the radius of B. Since \lVert \mathbf{h} \rVert < r, \mathbf{a} + t\mathbf{h} \in B for \lvert t \rvert < r / \lVert \mathbf{h} \rVert, an open interval containing [0, 1]; on it let g(t) = f(\mathbf{a} + t\mathbf{h}). Each f_{x_i} has continuous partials, so it is differentiable (Result 2 of Tangent planes and linear approximation, whose proof runs the same way in n variables), and so is f. The chain rule along a curve (Result 1 of Gradient and directional derivatives), applied to f and then to each f_{x_i}, gives

\begin{aligned} g'(t) &= \textstyle\sum_{i} f_{x_i}(\mathbf{a} + t\mathbf{h})\,h_i, \\ g''(t) &= \textstyle\sum_{i, j} f_{x_i x_j}(\mathbf{a} + t\mathbf{h})\,h_i h_j \\ &= Q_{\mathbf{a} + t\mathbf{h}}(\mathbf{h}). \end{aligned}

Taylor’s theorem with the Lagrange remainder (The Taylor remainder and the Lagrange error bound), of degree 1 on [0, 1], gives c \in (0, 1) with \displaystyle g(1) = g(0) + g'(0) + \frac12 g''(c): the formula.

Result 3. Let H = H_f(\mathbf{a}). The spectral theorem for real symmetric matrices (stated in Diagonalization) gives perpendicular unit vectors \mathbf{v}_1, \dots, \mathbf{v}_n with H\mathbf{v}_i = \lambda_i \mathbf{v}_i. For \mathbf{h} = \sum_i c_i \mathbf{v}_i, Q_{\mathbf{a}}(\mathbf{h}) = \mathbf{h} \cdot H\mathbf{h} = \sum_i \lambda_i c_i^2 and \lVert \mathbf{h} \rVert^2 = \sum_i c_i^2.

(a) Let m > 0 be the least eigenvalue, so Q_{\mathbf{a}}(\mathbf{h}) \ge m\lVert \mathbf{h} \rVert^2. Let \varepsilon(\mathbf{p}) = \sum_{i, j} \lvert f_{x_i x_j}(\mathbf{p}) - f_{x_i x_j}(\mathbf{a}) \rvert. Since \lvert h_i h_j \rvert \le \lVert \mathbf{h} \rVert^2, \lvert Q_{\mathbf{p}}(\mathbf{h}) - Q_{\mathbf{a}}(\mathbf{h}) \rvert \le \varepsilon(\mathbf{p})\lVert \mathbf{h} \rVert^2. By continuity of the second partials there is \delta > 0, below the radius of the ball, with \displaystyle \varepsilon(\mathbf{p}) < \frac{m}{2} when \lVert \mathbf{p} - \mathbf{a} \rVert < \delta. For 0 < \lVert \mathbf{h} \rVert < \delta, Result 2 with \nabla f(\mathbf{a}) = \mathbf{0} gives \mathbf{p} = \mathbf{a} + c\mathbf{h}, within \delta of \mathbf{a}, with

f(\mathbf{a} + \mathbf{h}) - f(\mathbf{a}) = \tfrac12 Q_{\mathbf{p}}(\mathbf{h}) \ge \tfrac{m}{4}\lVert \mathbf{h} \rVert^2 > 0.

(b) Apply (a) to -f.

(c) Let \mathbf{v} be a unit eigenvector for \lambda_i > 0 and g(t) = f(\mathbf{a} + t\mathbf{v}). By the proof of Result 2, g'(0) = \nabla f(\mathbf{a}) \cdot \mathbf{v} = 0 and g''(0) = Q_{\mathbf{a}}(\mathbf{v}) = \lambda_i > 0, with g'' continuous, so the one-variable second derivative test (The first and second derivative tests) gives f(\mathbf{a} + t\mathbf{v}) > f(\mathbf{a}) for small t \ne 0. An eigenvector for \lambda_j < 0 gives nearby points with f < f(\mathbf{a}).

(d) At the origin x^4 + y^4, -x^4 - y^4 and x^4 - y^4 all have zero gradient and zero Hessian, and they have a minimum, a maximum and a saddle point.

Result 4. The characteristic polynomial of the 2 \times 2 Hessian is \lambda^2 - (f_{xx} + f_{yy})\lambda + D, so \lambda_1 \lambda_2 = D and \lambda_1 + \lambda_2 = f_{xx} + f_{yy} (Eigenvalues and eigenvectors). If D < 0, the eigenvalues have opposite signs: Result 3(c). If D > 0, they share a sign; also f_{xx} f_{yy} > f_{xy}^2 \ge 0, so f_{xx} and f_{yy} share a sign, which is then that of \lambda_1 + \lambda_2: Result 3(a) or (b). If D = 0, an eigenvalue is 0, and the three functions of Result 3(d) have D = 0.

Result 5. The existence of the extremes is stated here without proof; it is proved in Real Analysis from the Bolzano-Weierstrass theorem, like the one-variable The extreme value theorem. If an extremum is taken at \mathbf{a} \in S that is not a boundary point, some open ball about \mathbf{a} does not contain points outside S (it does contain \mathbf{a} \in S), so \mathbf{a} is interior. There the absolute extremum is a local one, so by Result 1 every first partial that exists at \mathbf{a} is 0: \mathbf{a} is a critical point.

Worked examples

Example 1

Four critical points

Find and classify every critical point of f(x, y) = x^3 + 3xy^2 - 15x - 12y.

Solution

f is a polynomial, so of class C^2, and its critical points solve f_x = 3x^2 + 3y^2 - 15 = 0 and f_y = 6xy - 12 = 0: x^2 + y^2 = 5 and xy = 2. Then (x + y)^2 = 5 + 4 = 9 and (x - y)^2 = 5 - 4 = 1, so x + y = \pm 3 and x - y = \pm 1, giving (2, 1), (1, 2), (-1, -2) and (-2, -1). With f_{xx} = 6x, f_{xy} = 6y, f_{yy} = 6x, D = 36(x^2 - y^2), and Result 4 gives:

pointDf_{xx}ftype
(2, 1)10812-28min
(-2, -1)108-1228max
(1, 2)-1086-26saddle
(-1, -2)-108-626saddle

Neither extreme is absolute: f(x, 0) = x^3 - 15x takes every real value.

Example 2

Pricing two flavors of gelato

Priya and Hiro sell pistachio gelato at x dollars a pint and lemon at y dollars a pint. Their records fit a summer day’s sales by 113 - 10x + 4y pints of pistachio and 50 + 3x - 8y pints of lemon. A pint costs them \$4 to make for pistachio and \$3 for lemon. Find the prices that maximize the day’s profit, and prove that the maximum is absolute.

Solution

The profit in dollars is

\begin{aligned} P(x, y) &= (x - 4)(113 - 10x + 4y) \\ &\quad + (y - 3)(50 + 3x - 8y) \\ &= -10x^2 + 7xy - 8y^2 + 144x + 58y - 602. \end{aligned}

The critical points solve -20x + 7y + 144 = 0 and 7x - 16y + 58 = 0, a linear system with determinant 271 \ne 0, so the only one is (10, 8). There P_{xx} = -20, P_{xy} = 7, P_{yy} = -16 and D = 320 - 49 = 271 > 0: a strict local maximum. As P has degree 2, Result 2 is exact, and completing the square,

\begin{aligned} P(10 + h, 8 + k) &= 350 - 10h^2 + 7hk - 8k^2 \\ &= 350 - 10\left(h - \tfrac{7}{20}k\right)^2 - \tfrac{271}{40}k^2, \end{aligned}

which is below 350 unless h = k = 0. The absolute maximum is \$350 a day, at \$10 and \$8, selling 45 and 16 pints: 6 \cdot 45 + 5 \cdot 16 = 350.

Example 3

Absolute extremes on a rectangle

Find the absolute maximum and minimum of f(x, y) = x^2 + xy + y^2 - 6x on R = [0, 5] \times [-3, 0].

Solution

R is closed and bounded and f is continuous, so Result 5 applies. Inside, f_x = 2x + y - 6 and f_y = x + 2y vanish only at (4, -2), where f = 16 - 8 + 4 - 24 = -12. On each edge f is a function of one variable on a closed interval (Optimization):

edgef therecritical numbervalue
y = 0x^2 - 6xx = 3-9
y = -3x^2 - 9x + 9\displaystyle x = \frac92\displaystyle -\frac{45}{4}
x = 0y^2nonenone
x = 5y^2 + 5y - 5\displaystyle y = -\frac52\displaystyle -\frac{45}{4}

The corners give f(0, 0) = 0, f(5, 0) = -5, f(5, -3) = -11 and f(0, -3) = 9. Comparing, the absolute maximum is 9, at the corner (0, -3), and the absolute minimum is -12, at (4, -2) (Figure 1).

Example 4

Proof: a minimum on every line, but not a minimum

Let f(x, y) = (y - x^2)(y - 3x^2). Prove that along every line through the origin f has a strict local minimum at the origin, yet the origin is a saddle point. What does Result 4 say there?

Solution

On the line x = 0, f(0, t) = t^2. On the line y = mx,

f(t, mt) = t^2 (m - t)(m - 3t).

If m = 0 this is 3t^4. If m \ne 0 and \displaystyle 0 < \lvert t \rvert < \frac{\lvert m \rvert}{3}, both m - t and m - 3t have the sign of m. Either way f > 0 = f(0, 0) near the origin on the line, off the origin.

Next, f = y^2 - 4x^2 y + 3x^4 has \nabla f = \langle 12x^3 - 8xy, 2y - 4x^2 \rangle, zero at the origin. On the parabola y = 2x^2, between the parabolas where the factors vanish, f = (x^2)(-x^2) = -x^4 < 0 for x \ne 0, while f > 0 on the y-axis. Every disk about the origin meets both, so the origin is a saddle point. The region where f < 0 lies between two parabolas tangent to the x-axis, and every line leaves the origin outside it (Figure 2).

At the origin f_{xx} = 36x^2 - 8y = 0, f_{xy} = -8x = 0 and f_{yy} = 2, so D = 0 and Result 4 is silent, as it must be.

f < 0 f < 0 f > 0 f > 0
y = 3x^2
y = x^2
Figure 2The function f(x, y) = (y - x^2)(y - 3x^2) of Example 4 is negative exactly in the shaded region between the parabolas y = x^2 and y = 3x^2. Near the origin every line through it, such as the dashed lines y = 0.7x and y = -2.7x, stays outside that region, so f has a minimum at the origin along every line; a curve such as y = 2x^2 runs inside it.
Example 5

Proof: the least-squares line

Let (x_1, y_1), \dots, (x_n, y_n) be data points, n \ge 2, with the x_i not all equal. Prove that E(m, b) = \sum_{i} (y_i - m x_i - b)^2 has exactly one critical point, and that it is the absolute minimum of E.

Solution

Write S_x = \sum x_i, S_y = \sum y_i, S_{xx} = \sum x_i^2 and S_{xy} = \sum x_i y_i. Then

\begin{aligned} E_m &= 2(S_{xx} m + S_x b - S_{xy}), \\ E_b &= 2(S_x m + n b - S_y). \end{aligned}

The critical points solve a linear system with determinant \Delta = n S_{xx} - S_x^2. With \displaystyle \bar{x} = \frac{S_x}{n}, \displaystyle \sum (x_i - \bar{x})^2 = S_{xx} - 2\bar{x}S_x + n\bar{x}^2 = S_{xx} - \frac{S_x^2}{n}, so \Delta = n\sum (x_i - \bar{x})^2 > 0, since some x_i \ne \bar{x}. So there is exactly one critical point (m^*, b^*). The second partials are the constants 2S_{xx}, 2S_x, 2n, and E has degree 2, so Result 2 is exact:

\begin{aligned} &E(m^* + h, b^* + k) - E(m^*, b^*) \\ &\quad = S_{xx} h^2 + 2S_x hk + n k^2 \\ &\quad = \textstyle\sum (h x_i + k)^2. \end{aligned}

If this is 0, then h x_i + k = 0 for every i; h \ne 0 would make every \displaystyle x_i = -\frac{k}{h}, so h = 0 and then k = 0. Hence (m^*, b^*) is the strict absolute minimum: the line of The least squares criterion and its line, whose system is the normal equations of Least squares: the best approximate solution of Ax = b.

Boundaries

The test needs smoothness and a margin. Result 4 needs continuous second partials, and says nothing when D = 0 (Example 4, Exercise 5). Where a partial fails to exist, as for the minimum of \lvert x \rvert + \lvert y \rvert at the origin, the function must be examined directly.

One-variable shortcuts fail. On an interval, a lone critical point that is a local maximum is the absolute maximum (The first and second derivative tests), and two local minima have a local maximum between them. In two variables both fail (Exercises 13 and 14).

Closed and bounded are both needed. x^2 + y^2 on the open disk x^2 + y^2 < 1 has no maximum, and x on \mathbb{R}^2 has neither extreme.

Where this goes next

When the boundary is a level curve g(x, y) = k that is awkward to parametrize, Lagrange multipliers finds the extremes on it, where \nabla f is parallel to \nabla g.

Exercises

Classifying critical points

  1. Find and classify every critical point of f(x, y) = x^2 + 3xy + y^2 - 5x - 5y.
  2. Find and classify every critical point of f(x, y) = 2x^3 - 6xy + 3y^2.
  3. Find and classify every critical point of f(x, y) = xy\,e^{-(x^2 + y^2)/2}.
  4. Find the critical point of f(x, y, z) = x^2 + y^2 + z^2 + xy + yz + zx - x + z and classify it with the eigenvalues of the Hessian.
  5. Show that the origin is a critical point with D = 0 of both f(x, y) = (x - y)^2 + y^4 and g(x, y) = (x - y)^2 - y^4, and classify it for each.

Absolute extremes

  1. Find the absolute maximum and minimum of f(x, y) = x^2 + y^2 - 2x - 4y on the closed disk x^2 + y^2 \le 20.
  2. Find the absolute maximum and minimum of f(x, y) = xy(6 - x - y) on the closed triangle with vertices (0, 0), (6, 0) and (0, 6).

At the gelato shop and beyond

  1. With the model of Example 2, Hiro keeps pistachio at \$9 a pint. Which lemon price makes the most profit then, and how much daily profit does that give up?
  2. Find the least-squares line for the points (0, 1), (1, 2), (2, 4) and (3, 5), and the least value of E.
  3. Find the point of the plane x + 2y + z = 6 nearest the origin by minimizing x^2 + y^2 + (6 - x - 2y)^2, and prove that the minimum is absolute.

Proofs

  1. Let f be of class C^2 on an open disk about (a, b), with \nabla f(a, b) = \mathbf{0} and f_{xx}(a, b) f_{yy}(a, b) < 0. Prove that (a, b) is a saddle point by looking only along the two axis directions.
  2. Let f be of class C^2 on an open disk about (a, b), with a local minimum at (a, b). Prove that f_{xx}(a, b) \ge 0, f_{yy}(a, b) \ge 0 and D \ge 0 there.
  3. Prove that f(x, y) = 3xe^y - x^3 - e^{3y} has exactly one critical point, that it is a local maximum, and that it is not an absolute maximum.
  4. Prove that f(x, y) = (x^2 - 1)^2 + (x^2 y - x - 1)^2 has exactly two critical points and that both are absolute minima.
Answers
  1. A saddle point at (1, 1), where f = -5. 2x + 3y - 5 = 3x + 2y - 5 = 0 only at (1, 1), and D = 2 \cdot 2 - 3^2 = -5 < 0.
  2. A saddle point at (0, 0) and a local minimum -1 at (1, 1). f_x = 6x^2 - 6y and f_y = 6y - 6x give y = x = x^2. Here D = 12x \cdot 6 - 36 = 72x - 36: -36 at the origin, 36 with f_{xx} = 12 at (1, 1).
  3. A saddle point at (0, 0); local maxima e^{-1} at (1, 1) and (-1, -1); local minima -e^{-1} at (1, -1) and (-1, 1). With E = e^{-(x^2 + y^2)/2}, f_x = y(1 - x^2)E and f_y = x(1 - y^2)E: y = 0 forces x = 0, else x = \pm 1 and y = \pm 1. Also f_{xx} = y(x^3 - 3x)E, f_{yy} = x(y^3 - 3y)E, f_{xy} = (1 - x^2)(1 - y^2)E. At the origin D = -1. At (\pm 1, \pm 1), f_{xy} = 0 and f_{xx} = f_{yy} = -2xy\,e^{-1}, so D = 4e^{-2} > 0 and f_{xx} has the sign of -xy.
  4. (1, 0, -1), a strict local minimum with value -1, in fact absolute. The gradient \langle 2x + y + z - 1, x + 2y + z, x + y + 2z + 1 \rangle vanishes only there. The Hessian is I + J, J the matrix of ones, with eigenvalues 4 (for \langle 1, 1, 1 \rangle) and 1, 1 (for vectors whose entries sum to 0). As f has degree 2, Result 2 is exact, \displaystyle f = -1 + \frac12 Q(\mathbf{h}) with Q(\mathbf{h}) \ge \lVert \mathbf{h} \rVert^2.
  5. A strict local minimum for f and a saddle point for g. Both gradients vanish at the origin, and both functions have f_{xx} = 2, f_{xy} = -2, f_{yy} = 2 there: D = 4 - 4 = 0. f \ge 0, with f = 0 only where x = y and y = 0. g(t, 0) = t^2 > 0 and g(t, t) = -t^4 < 0 for t \ne 0.
  6. Absolute minimum -5 at (1, 2), absolute maximum 40 at (-2, -4). f = (x - 1)^2 + (y - 2)^2 - 5, with its critical point (1, 2) inside. On the circle f = 20 - \langle 2, 4 \rangle \cdot \langle x, y \rangle, and \displaystyle \lvert \langle 2, 4 \rangle \cdot \langle x, y \rangle \rvert \le \sqrt{20}\sqrt{20} = 20, with equality at \pm\langle 2, 4 \rangle; so f runs from 0 at (2, 4) to 40 at (-2, -4).
  7. Absolute maximum 8 at (2, 2); absolute minimum 0, on the whole boundary. f = 0 on the edges and f > 0 inside, where f_x = y(6 - 2x - y) and f_y = x(6 - x - 2y) vanish only at (2, 2).
  8. \displaystyle y = \frac{121}{16} \approx 7.56 dollars, for a profit of \displaystyle \frac{10929}{32} \approx 341.53 dollars, giving up about \$8.47. P(9, y) = -8y^2 + 121y - 116, a downward parabola with vertex at \displaystyle y = \frac{121}{16}.
  9. \displaystyle y = \frac75 x + \frac{9}{10}, with least value \displaystyle E = \frac15. Here n = 4, S_x = 6, S_y = 12, S_{xx} = 14, S_{xy} = 25, so 14m + 6b = 25 and 6m + 4b = 12. The residuals are 0.1, -0.3, 0.3, -0.1.
  10. (1, 2, 1), at distance \displaystyle \sqrt6. The gradient \langle 4x + 4y - 12, 4x + 10y - 24 \rangle vanishes only at (1, 2), where f = 6; then z = 1. The constant Hessian has f_{xx} = 4, f_{xy} = 4, f_{yy} = 10, so D = 24 > 0. As f has degree 2, \displaystyle f(1 + h, 2 + k) = 6 + \frac12 Q(h, k) exactly, and the completed square of the Intuition makes Q > 0 off (0, 0).
  11. Say f_{xx}(a, b) > 0 > f_{yy}(a, b); the other case is the same. g(t) = f(a + t, b) has g'(0) = f_x(a, b) = 0 and continuous g''(t) = f_{xx}(a + t, b) with g''(0) > 0, so the one-variable second derivative test gives f(a + t, b) > f(a, b) for small t \ne 0. Likewise f(a, b + t) < f(a, b) for small t \ne 0. Every disk about (a, b) contains points of both kinds.
  12. For a unit vector \mathbf{u} let g(t) = f(\mathbf{a} + t\mathbf{u}), \mathbf{a} = (a, b). By Result 1 and the proof of Result 2, g'(0) = 0 and g''(0) = Q_{\mathbf{a}}(\mathbf{u}). If g''(0) < 0, the second derivative test would give g(t) < g(0) for small t \ne 0, against the minimum. So Q_{\mathbf{a}}(\mathbf{u}) \ge 0, and as Q_{\mathbf{a}}(s\mathbf{u}) = s^2 Q_{\mathbf{a}}(\mathbf{u}), Q_{\mathbf{a}} \ge 0 everywhere. With A, B, C as in the Intuition, A = Q(1, 0) \ge 0 and C = Q(0, 1) \ge 0. If A > 0, Q(-B, A) = A(AC - B^2) \ge 0 gives D \ge 0. If A = 0, Q(s, 1) = 2Bs + C \ge 0 for every s forces B = 0, so D = 0.
  13. f_x = 3e^y - 3x^2 and f_y = 3e^y(x - e^{2y}), so x = e^{2y} and e^y = x^2 = e^{4y}: y = 0, x = 1. At (1, 0), f = 1, f_{xx} = -6x = -6, f_{xy} = 3e^y = 3, f_{yy} = 3xe^y - 9e^{3y} = -6, so D = 36 - 9 = 27 > 0 with f_{xx} < 0: a strict local maximum. But f(-3, 0) = -9 + 27 - 1 = 17 > 1.
  14. f_y = 2x^2(x^2 y - x - 1) and f_x = 4x(x^2 - 1) + 2(x^2 y - x - 1)(2xy - 1). If x = 0, f_x = 2 \ne 0. So f_y = 0 forces x^2 y = x + 1, and then f_x = 4x(x^2 - 1) = 0 gives x = \pm 1, with y = 2 for x = 1 and y = 0 for x = -1. At (1, 2) and (-1, 0), f = 0 \le f everywhere: both are absolute minima, with no third critical point between them.
math.wiki · CC-BY-SA-4.0