Tangent planes and linear approximation
Near a point, a smooth surface looks flat. Which plane is it, and how well do its heights stand in for the values of the function nearby?
Where you are 8 in Partial derivatives
Why this exists
In one variable, Linear approximation and differentials replaced a function near a by its tangent line. For z = f(x, y) the partial derivatives of Partial derivatives give two tangent lines through (a, b, f(a, b)), one along each trace, and two lines span a plane. Whether that plane deserves the name is a real question: partials see only the two lines through (a, b) parallel to the axes, and a function can have both at a point where it is not even continuous. What settles it is differentiability: the surface stays close to the plane in every direction at once, closer than the distance to the point.
Sofia builds guitars in her luthier’s workshop, with Diego as her apprentice. How far a small change in a string’s length and tension moves its pitch, and how accurately each must be measured, is a question about a tangent plane (Example 3).
Definition
Let f : D \to \mathbb{R} with D \subseteq \mathbb{R}^2, and let (a, b) be an interior point of D. For numbers A and B and a small step \langle h, k \rangle, let
f is differentiable at (a, b) if there are numbers A and B with
the change in f is linear in the step plus an error small compared with the step’s length. f is differentiable on an open set U if it is differentiable at every point of U.
Then A = f_x(a, b) and B = f_y(a, b) (Result 1), and the linearization of f at (a, b) is
The tangent plane to the graph z = f(x, y) at P = (a, b, f(a, b)) is the graph z = L(x, y), and f(x, y) \approx L(x, y) near (a, b) is the linear approximation. Where f is not differentiable the graph has no tangent plane, even if both partials exist.
Differentials. For z = f(x, y) and independent variables dx, dy, the total differential at (a, b) is dz = f_x(a, b)\,dx + f_y(a, b)\,dy, the rise along the tangent plane for the step \langle dx, dy \rangle. The actual change is \Delta z = f(a + dx, b + dy) - f(a, b).
Matrix form and more variables. The row Df(a, b) = \begin{bmatrix} f_x(a, b) & f_y(a, b) \end{bmatrix} is the derivative of f at (a, b), and Ah + Bk is Df(a, b) times the column \begin{bmatrix} h \\ k \end{bmatrix} (Multiplying a matrix by a vector), the same vector as \langle h, k \rangle. For n variables, replace Ah + Bk by \sum_i A_i h_i and \displaystyle \sqrt{h^2 + k^2} by \lVert \mathbf{h} \rVert; then A_i = \dfrac{\partial f}{\partial x_i}, and every result below holds with the same proof.
Intuition
Magnify the graph near P by \displaystyle \frac{1}{r}, so the disk of radius r about (a, b) fills the view. The gap E between surface and plane becomes \displaystyle \frac{E}{r}, which tends to 0 for a differentiable f: under enough magnification the surface cannot be told from its tangent plane. The cone \displaystyle z = \sqrt{x^2 + y^2} looks the same at every magnification of its vertex and has no tangent plane there.
The tangent plane contains the tangent lines of the two traces through P (Figure 1), with directions \langle 1, 0, f_x(a, b) \rangle and \langle 0, 1, f_y(a, b) \rangle. A step \langle h, k \rangle is h of the first plus k of the second, so the height rises by f_x h + f_y k, the differential.
Results
Result 1 (differentiable implies partials and continuity). If f is differentiable at (a, b) with numbers A and B, then f_x(a, b) and f_y(a, b) exist, A = f_x(a, b), B = f_y(a, b), and f is continuous at (a, b).
Result 2 (continuous partials give differentiability). If f_x and f_y exist at every point of an open disk about (a, b) and are continuous at (a, b), then f is differentiable at (a, b). So a function of class C^1 on an open set U is differentiable on U.
Result 3 (the tangent plane). If f is differentiable at (a, b), its tangent plane at P has normal \mathbf{n} = \langle f_x(a, b), f_y(a, b), -1 \rangle, so its equation is
and it contains the lines through P with directions \langle 1, 0, f_x(a, b) \rangle and \langle 0, 1, f_y(a, b) \rangle. It is the best linear approximation: of all non-vertical planes through P, it is the only one whose height differs from f by an amount small compared with \displaystyle \sqrt{h^2 + k^2}.
Result 4 (an error bound). Let Q be the closed rectangle with opposite corners (a, b) and (a + h, b + k). Suppose f_x and f_y exist on an open set containing Q, and f_{xx}, f_{yx}, f_{yy} exist on Q with absolute value at most M. Then
Result 5 (relative error). If f(x, y) = C x^p y^q with C \ne 0, x > 0 and y > 0, then
Why it is true
Result 1. A limit as (h, k) \to (0, 0) is also the limit along the h-axis (Limits and continuity in several variables). With k = 0, \displaystyle \frac{f(a + h, b) - f(a, b) - Ah}{\lvert h \rvert} \to 0, and multiplying by \displaystyle \frac{\lvert h \rvert}{h} = \pm 1 keeps the limit 0, so \displaystyle \frac{f(a + h, b) - f(a, b)}{h} \to A: f_x(a, b) = A. The k-axis gives f_y(a, b) = B. With \displaystyle r = \sqrt{h^2 + k^2},
and each term tends to 0, since \lvert h \rvert, \lvert k \rvert \le r: f is continuous.
Result 2. Let the disk have radius \rho and \displaystyle 0 < r = \sqrt{h^2 + k^2} < \rho. Every point of the rectangle with corners (a, b) and (a + h, b + k) is within r of (a, b), so lies in the disk. Step in x at height b + k, then in y at x = a:
The mean value theorem along a segment (Result 2 of Partial derivatives) makes the brackets h f_x(\xi, b + k) and k f_y(a, \eta), with \xi between a and a + h and \eta between b and b + k. With A = f_x(a, b) and B = f_y(a, b),
As \lvert h \rvert, \lvert k \rvert \le r, \displaystyle \frac{\lvert E \rvert}{r} is at most the sum of the brackets’ absolute values. The points (\xi, b + k) and (a, \eta) are within r of (a, b), so continuity of f_x and f_y there sends both to 0.
Result 3. Moving z to the left in z = L(x, y) gives the equation, whose coefficients are \mathbf{n} (Lines and planes in space). The two directions are perpendicular to \mathbf{n}; indeed (The cross product and the triple product)
A non-vertical plane through P is z = f(a, b) + A(x - a) + B(y - b), and its gap from the graph above (a + h, b + k) is E(h, k). If \displaystyle \frac{E}{r} \to 0, Result 1 forces A = f_x(a, b) and B = f_y(a, b).
Result 4. Step in x along y = b, then in y along x = a + h, both edges of Q. The mean value theorem gives \xi between a and a + h and \eta between b and b + k with
The mean value theorem applied to x \mapsto f_x(x, b), whose derivative f_{xx} is at most M in absolute value, bounds the first bracket by M\lvert h \rvert. For the second, pass through (a, \eta):
by the mean value theorem in x with derivative f_{yx}, then in y with f_{yy}. So \lvert E \rvert \le M h^2 + M\lvert k \rvert(\lvert h \rvert + \lvert k \rvert) \le M(\lvert h \rvert + \lvert k \rvert)^2.
Result 5. \displaystyle f_x = \frac{p f}{x} and \displaystyle f_y = \frac{q f}{y}; divide df = f_x\,dx + f_y\,dy by f.
Worked examples
A tangent plane to an ellipsoid
Let \displaystyle f(x, y) = \sqrt{18 - x^2 - 2y^2}, the upper half of x^2 + 2y^2 + z^2 = 18. Find the tangent plane at (1, 2) and a normal to it, and estimate f(1.1, 1.95).
Solution\displaystyle f(1, 2) = \sqrt{18 - 1 - 8} = 3. Where x^2 + 2y^2 < 18, the chain rule gives the continuous partials
so f is differentiable there (Result 2). At (1, 2), \displaystyle f_x = -\frac13 and \displaystyle f_y = -\frac43, so the tangent plane is \displaystyle z = 3 - \frac13(x - 1) - \frac43(y - 2), or, times 3,
By Result 3 a normal is \displaystyle \left\langle -\frac13, -\frac43, -1 \right\rangle, or \langle 1, 4, 3 \rangle after multiplying by -3. With dx = 0.1 and dy = -0.05,
The true value is \displaystyle \sqrt{9.185} \approx 3.0307. The estimate is high, as it must be for a cap that curves down away from its tangent plane.
Tangent planes parallel to a given plane
Find every point of the surface z = x^2 + xy - y^3 where the tangent plane is parallel to the plane 3x - y - z = 0, and the tangent plane there.
SolutionThe partials f_x = 2x + y and f_y = x - 3y^2 are continuous, so every point has a tangent plane, with normal \langle 2x + y, x - 3y^2, -1 \rangle (Result 3). Parallel planes have parallel normals; the given normal \langle 3, -1, -1 \rangle has the same third component, so the normals must be equal:
Substituting y = 3 - 2x gives x - 3(3 - 2x)^2 + 1 = 0, that is 12x^2 - 37x + 26 = 0, with discriminant 37^2 - 4 \cdot 12 \cdot 26 = 121. So \displaystyle x = \frac{37 \pm 11}{24}: x = 2 or \displaystyle x = \frac{13}{12}.
At x = 2, y = -1, z = 3, and the tangent plane z = 3 + 3(x - 2) - (y + 1) is z = 3x - y - 4. At \displaystyle x = \frac{13}{12}, \displaystyle y = \frac56 and \displaystyle z = \frac{169}{144} + \frac{65}{72} - \frac{125}{216} = \frac{647}{432}, with tangent plane \displaystyle z = 3x - y - \frac{397}{432}. Neither contains the origin, so neither is the given plane itself.
Tuning a string in the workshop
A string of length L meters with mass 0.0025 kilograms per meter, under tension T newtons, sounds at
hertz. Sofia’s string has L = 0.5 and T = 64. (a) Diego shortens it to 0.49 meters and raises the tension to 66 newtons; estimate the new pitch. (b) Sofia measures L to within 0.002 meters and T to within 0.5 newtons. Estimate the largest error in the computed pitch.
Solution\displaystyle f(0.5, 64) = \frac{10 \cdot 8}{0.5} = 160. The partials
are continuous for L, T > 0, with f_L(0.5, 64) = -320 hertz per meter and f_T(0.5, 64) = 1.25 hertz per newton.
(a) With dL = -0.01 and dT = 2, df = -320(-0.01) + 1.25(2) = 3.2 + 2.5 = 5.7, so the pitch is about 165.7 hertz; exactly, \displaystyle \frac{10\sqrt{66}}{0.49} \approx 165.80.
(b) By the triangle inequality,
about 1.3 hertz. Result 5 with f = 10\,T^{1/2}L^{-1} gives the same in relative terms: \displaystyle \frac{df}{f} = \frac12 \cdot \frac{dT}{T} - \frac{dL}{L}, at most \displaystyle \frac{0.5}{128} + \frac{0.002}{0.5} \approx 0.79\% of 160.
How far the plane can drift
Estimate e^{0.1}\cos 0.2 by the linearization of f(x, y) = e^x \cos y at (0, 0), and bound the error with Result 4.
Solutionf(0, 0) = 1, f_x = e^x\cos y and f_y = -e^x \sin y, so L(x, y) = 1 + x and the estimate is L(0.1, 0.2) = 1.1. The second partials f_{xx} = e^x \cos y, f_{yx} = -e^x \sin y and f_{yy} = -e^x \cos y are at most e^{0.1} < 1.11 in absolute value on Q = [0, 0.1] \times [0, 0.2], so with M = 1.11,
In fact e^{0.1}\cos 0.2 \approx 1.0831, an error of about 0.0169. Result 4 is a guarantee, not an estimate; for continuous second partials, Taylor’s theorem, proved with the The multivariable chain rule, halves the constant to \displaystyle \frac{M}{2}.
Proof: partials without a tangent plane
Let \displaystyle f(x, y) = \sqrt[3]{x^3 + y^3}, the real cube root, for which Partial derivatives found f_x(0, 0) = f_y(0, 0) = 1. Prove that f is continuous at the origin but not differentiable there.
Solutionf is the continuous cube root of a polynomial, so it is continuous. If f were differentiable at (0, 0), Result 1 would force A = B = 1 and \displaystyle E(h, k) = \sqrt[3]{h^3 + k^3} - h - k. On the diagonal h = k = t \ne 0,
whose absolute value is the constant \displaystyle \frac{2 - \sqrt[3]{2}}{\sqrt{2}} \approx 0.523, not 0: a contradiction. Like the cone, the graph is made of lines through the origin, since f(tx, ty) = t f(x, y).
Proof: a tangent plane without continuous partials
Let \displaystyle f(x, y) = (x^2 + y^2)\sin\dfrac{1}{\sqrt{x^2 + y^2}} for (x, y) \ne (0, 0) and f(0, 0) = 0. Prove that f is differentiable at the origin with tangent plane z = 0, but f_x is not continuous there.
SolutionWith \displaystyle r = \sqrt{h^2 + k^2} and A = B = 0, E(h, k) = f(h, k) and
so f is differentiable at (0, 0), with both partials 0 (Result 1) and tangent plane z = 0. Away from the origin the product and chain rules give
On the positive x-axis r = x, so \displaystyle f_x(x, 0) = 2x\sin\frac1x - \cos\frac1x, which is -1 at \displaystyle x = \frac{1}{2\pi n} and 1 at \displaystyle x = \frac{1}{(2n + 1)\pi} for every positive integer n. Both sequences tend to the origin, so f_x is not continuous there: Result 2 is sufficient, not necessary.
Boundaries
Partials are not enough. Example 5 has both partials at the origin, is continuous, and has no tangent plane; a function with both partials at a point of discontinuity (Exercise 19 of Partial derivatives) is not differentiable there either, by Result 1.
Continuous partials are more than enough. Example 6 is differentiable at a point where f_x is not continuous.
The estimate is local. Differentiability alone does not say how small a step must be; a guaranteed bound needs control of the second partials, as in Result 4. An error budget from differentials, as in Example 3, is a first-order estimate, not a bound. At the rim of the cap in Example 1 the tangent plane is vertical and f has no partials; such planes are found from a level surface in Gradient and directional derivatives.
Where this goes next
Differentiability is the hypothesis of the The multivariable chain rule, whose proof starts from the error E. In Gradient and directional derivatives the linear part becomes \nabla f \cdot \langle h, k \rangle, the rate of change in every direction. At a horizontal tangent plane, Extrema and the second derivative test adds the quadratic term. For maps from \mathbb{R}^n to \mathbb{R}^m the row Df becomes the Jacobian matrix, the linearization used in Nonlinear systems and linearization.
Exercises
Tangent planes and linearizations
- Find the tangent plane to z = x^3 y - 2y^2 at (1, 2).
- Find the tangent plane to z = y\ln x at (1, 3).
- Find the linearization of f(x, y) = e^{2x - y} at (1, 2), and use it to estimate f(1.03, 2.01).
- Use a linearization of \displaystyle \sqrt{x^2 + y^2} at (5, 12) to estimate \displaystyle \sqrt{(5.03)^2 + (11.96)^2}.
- Find every point of z = x^3 - 3xy + y^3 where the tangent plane is horizontal.
- Find the point of z = 2x^2 + y^2 where the tangent plane is parallel to the plane 4x - 6y - z = 1, and the tangent plane there.
In Sofia’s workshop
- A string with the pitch formula of Example 3 has L = 0.5 and T = 100. Diego moves the bridge so the length becomes 0.505 meters, and the tension drops to 97 newtons. Estimate the new pitch.
- For the same kind of string, the tension is known to within 2\% and the length to within 0.5\%. Estimate the largest relative error in the pitch, and the error in hertz at a pitch of 200 hertz.
- Diego weighs a maple fretboard blank at m = 120 grams, within 0.5 gram, and measures it as \ell = 50, w = 6 and t = 0.8 centimeters, within 0.1, 0.05 and 0.02 centimeters. Find its density \rho = \dfrac{m}{\ell w t}, estimate the largest error, and say which measurement matters most.
- A function g is differentiable at (3, -1) with g(3, -1) = 4, g_x(3, -1) = 2 and g_y(3, -1) = -5. Estimate g(2.9, -0.8).
- Estimate f(1.1, 0.9) for f(x, y) = x^2 y by the linearization at (1, 1), bound the error with Result 4, and compare with the true error.
Differentiable or not
- Use the definition to show that f(x, y) = x^2 + 3xy is differentiable at (1, 2), and find its tangent plane there.
- Let \displaystyle f(x, y) = \sqrt{\lvert xy \rvert}, whose partials at the origin are both 0. Is f differentiable at (0, 0)?
- Let f(x, y) = \dfrac{x^2 y^2}{x^2 + y^2} for (x, y) \ne (0, 0) and f(0, 0) = 0. Is f differentiable at (0, 0)? If so, find its tangent plane.
- Let f(x, y) = x\lvert y \rvert. Decide whether f is differentiable at (0, 0) and at (1, 0).
Proofs
- Let f be differentiable at (a, b) and let \langle p, q \rangle be a unit vector. Prove that as t \to 0 the quotient \displaystyle \frac{f(a + tp, b + tq) - f(a, b)}{t} tends to p f_x(a, b) + q f_y(a, b).
- Prove that if f and g are differentiable at (a, b), then so is fg, with (fg)_x(a, b) = f_x(a, b)g(a, b) + f(a, b)g_x(a, b).
- Prove that if f_x exists on an open disk about (a, b) and is continuous at (a, b), and f_y(a, b) exists, then f is differentiable at (a, b).
- Prove that if f is differentiable at (a, b) and has a local maximum there, then its tangent plane at (a, b, f(a, b)) is horizontal.
Answers
- z = 6x - 7y + 2. At (1, 2), f = -6, f_x = 3x^2 y = 6 and f_y = x^3 - 4y = -7, so z = -6 + 6(x - 1) - 7(y - 2).
- z = 3x - 3. At (1, 3), f = 0, \displaystyle f_x = \frac{y}{x} = 3 and f_y = \ln x = 0.
- L(x, y) = 1 + 2(x - 1) - (y - 2), and f(1.03, 2.01) \approx 1 + 0.06 - 0.01 = 1.05, since f_x = 2e^{2x - y} and f_y = -e^{2x - y} are 2 and -1 at (1, 2). The true value is e^{0.05} \approx 1.0513.
- About 12.9746. The partials at (5, 12) are \displaystyle \frac{5}{13} and \displaystyle \frac{12}{13}, so \displaystyle L = 13 + \frac{5(0.03) + 12(-0.04)}{13} = 13 - \frac{33}{1300}. The true value is about 12.9747.
- (0, 0, 0) and (1, 1, -1). A horizontal plane needs f_x = 3x^2 - 3y = 0 and f_y = -3x + 3y^2 = 0, so y = x^2 and x = y^2 = x^4, giving x = 0 or x = 1.
- (1, -3, 11), with tangent plane z = 4x - 6y - 11. The normal \langle 4x, 2y, -1 \rangle must equal \langle 4, -6, -1 \rangle, so x = 1, y = -3, and z = 11 + 4(x - 1) - 6(y + 3).
- About 195 hertz. f(0.5, 100) = 200, \displaystyle f_L = -\frac{100}{0.25} = -400 and \displaystyle f_T = \frac{5}{0.5 \cdot 10} = 1, so df = -400(0.005) + 1(-3) = -5. The exact value is about 195.03.
- 1.5\%, or 3 hertz. By Result 5, \displaystyle \frac{df}{f} = \frac12 \cdot \frac{dT}{T} - \frac{dL}{L}, at most \displaystyle \frac12(2\%) + 0.5\% in absolute value.
- \displaystyle \rho = \frac{120}{50 \cdot 6 \cdot 0.8} = 0.5 grams per cubic centimeter, within about 0.020. Result 5, with the same proof for four variables, gives \displaystyle \frac{d\rho}{\rho} = \frac{dm}{m} - \frac{d\ell}{\ell} - \frac{dw}{w} - \frac{dt}{t}, at most \displaystyle \frac{0.5}{120} + \frac{0.1}{50} + \frac{0.05}{6} + \frac{0.02}{0.8} \approx 0.0395 of 0.5. The thickness alone contributes 2.5\%: it matters most.
- g(2.9, -0.8) \approx 4 + 2(-0.1) - 5(0.2) = 2.8.
- L(1.1, 0.9) = 1.1, error at most 0.088, true error 0.011. L(x, y) = 1 + 2(x - 1) + (y - 1). On [1, 1.1] \times [0.9, 1], f_{xx} = 2y \le 2, f_{yx} = 2x \le 2.2 and f_{yy} = 0, so M = 2.2 works and the bound is 2.2(0.1 + 0.1)^2. In fact f(1.1, 0.9) = 1.089.
- z = 8x + 3y - 7. Expanding, f(1 + h, 2 + k) - f(1, 2) = 8h + 3k + h^2 + 3hk. With A = 8, B = 3, and \displaystyle \lvert hk \rvert \le \frac{r^2}{2}, \displaystyle \lvert E \rvert = \lvert h^2 + 3hk \rvert \le \frac52 r^2, so \displaystyle \frac{\lvert E \rvert}{r} \to 0. The plane is z = 7 + 8(x - 1) + 3(y - 2).
- No. By Result 1 the only candidates are A = B = 0, so E = f. On the diagonal, \displaystyle \frac{f(t, t)}{\sqrt{2}\,\lvert t \rvert} = \frac{\lvert t \rvert}{\sqrt2\,\lvert t \rvert} = \frac{1}{\sqrt2}, which does not tend to 0.
- Yes, with tangent plane z = 0. f(h, 0) = 0 = f(0, k) gives both partials 0, and \displaystyle x^2 y^2 \le \frac{(x^2 + y^2)^2}{4} gives \displaystyle \frac{\lvert E \rvert}{r} = \frac{\lvert f \rvert}{r} \le \frac{r}{4} \to 0.
- Differentiable at (0, 0), not at (1, 0). At the origin both partials are 0 and \displaystyle \lvert f(h, k) \rvert = \lvert h \rvert \lvert k \rvert \le \frac{r^2}{2}, so \displaystyle \frac{\lvert E \rvert}{r} \le \frac{r}{2} \to 0. At (1, 0) the quotient \displaystyle \frac{f(1, k) - f(1, 0)}{k} = \frac{\lvert k \rvert}{k} is 1 for k > 0 and -1 for k < 0, so f_y(1, 0) does not exist, and by Result 1 f is not differentiable there.
- Let A = f_x(a, b) and B = f_y(a, b) (Result 1). For t \ne 0 the step \langle tp, tq \rangle has length \lvert t \rvert, and f(a + tp, b + tq) - f(a, b) is t(pA + qB) + E(tp, tq), so the quotient is \displaystyle pA + qB + \frac{E(tp, tq)}{t}. As t \to 0, (tp, tq) \to (0, 0) without equaling it, so \displaystyle \frac{\lvert E(tp, tq) \rvert}{\lvert t \rvert} \to 0 by the definition, and the quotient tends to pA + qB.
- Write f_0 = f(a, b) and \alpha = f_x h + f_y k (partials at (a, b)), so f(a + h, b + k) = f_0 + \alpha + E_f; likewise g = g_0 + \beta + E_g. Multiplying, fg(a + h, b + k) - f_0 g_0 = (f_0 g_x + g_0 f_x)h + (f_0 g_y + g_0 f_y)k + R with R = f_0 E_g + g_0 E_f + (\alpha + E_f)(\beta + E_g). Now \displaystyle \frac{f_0 E_g}{r} and \displaystyle \frac{g_0 E_f}{r} tend to 0; \displaystyle \frac{\lvert \beta + E_g \rvert}{r} \le \lvert g_x \rvert + \lvert g_y \rvert + \frac{\lvert E_g \rvert}{r} is bounded near (0, 0), while \alpha + E_f = f(a + h, b + k) - f_0 \to 0 by continuity (Result 1). So \displaystyle \frac{R}{r} \to 0: fg is differentiable, and by Result 1 its x partial is the coefficient of h, f_x g + f g_x at (a, b).
- Let the disk have radius \rho and 0 < r < \rho. Write f(a + h, b + k) - f(a, b) as the sum of f(a + h, b + k) - f(a, b + k) and f(a, b + k) - f(a, b). The mean value theorem on the horizontal segment, which lies in the disk, makes the first h f_x(\xi, b + k) with \xi between a and a + h. Write the second as k f_y(a, b) + k\,\varepsilon(k), where \displaystyle \varepsilon(k) = \frac{f(a, b + k) - f(a, b)}{k} - f_y(a, b) \to 0 as k \to 0, and \varepsilon(0) = 0. Then E = h[f_x(\xi, b + k) - f_x(a, b)] + k\,\varepsilon(k), so \displaystyle \frac{\lvert E \rvert}{r} \le \lvert f_x(\xi, b + k) - f_x(a, b) \rvert + \lvert \varepsilon(k) \rvert. The first term tends to 0 by continuity of f_x at (a, b), as (\xi, b + k) is within r of (a, b); the second because \lvert k \rvert \le r.
- By Result 1, f_x(a, b) and f_y(a, b) exist. The function x \mapsto f(x, b) has a local maximum at the interior point a and a derivative there, so f_x(a, b) = 0 (Critical points and local versus absolute extrema); in the same way f_y(a, b) = 0. The tangent plane is then z = f(a, b), which is horizontal.