์ œ์•ฝ ์กฐ๊ฑด ์•„๋ž˜์—์„œ ํ•จ์ˆ˜์˜ ๊ทน๊ฐ’์„ ์ฐพ๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด. Constraint Curve์™€ Level Curve๊ฐ€ ์ ‘ํ•  ๋•Œ ์ตœ๋Œ€/์ตœ์†Œ๋ฅผ ์ด๋ฃฌ๋‹ค. ์ œ์•ฝ ์กฐ๊ฑด์ด ๋‘ ๊ฐœ์ผ ๋•Œ๋Š” Lagrange Multiplier๊ฐ€ 2๊ฐœ ํ•„์š”ํ•จ. โœŒ๏ธ

9 minute read

๋ณต์ˆ˜์ „๊ณตํ•˜๊ณ  ์žˆ๋Š” ์ˆ˜ํ•™๊ณผ์˜ ์กธ์—…์‹œํ—˜์„ ์œ„ํ•ด ํ•™๋ถ€ ์ˆ˜ํ•™ ๊ณผ๋ชฉ๋“ค์„ ๋‹ค์‹œ ๊ณต๋ถ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฏธ์ ๋ถ„ํ•™ ํฌ์ŠคํŠธ ์ „์ฒด ๋ณด๊ธฐ

A powerful method for finding extreme values of constrained functions

โ€œ๋ผ๊ทธ๋ž‘์ฃผ ์Šน์ˆ˜๋ฒ•(Lagrange Multiplier)โ€๋ผ๋Š” ๋ฐฉ๋ฒ•๋ก ์€ ์ œํ•œ ์กฐ๊ฑด์ด ์žˆ๋Š” ์ƒํ™ฉ์—์„œ ํ•จ์ˆ˜์˜ ๊ทน๊ฐ’(extreme value)๋ฅผ ๊ฐ•๋ ฅํ•œ ๋ฐฉ๋ฒ•์ด๋‹ค. ์ง€๊ธˆ๊นŒ์ง€ ์•ž์žฅ์—์„œ ๋ฐฐ์šด ๋ฏธ์ ๋ถ„ํ•™2 ๋‚ด์šฉ์„ ์ „๋ถ€ ๋ฐ”ํƒ•์œผ๋กœ ํ•˜๋Š” ์ค‘์š”ํ•œ ์‘์šฉ ์‚ฌ๋ก€ ์ค‘ ํ•˜๋‚˜๋‹ค.

The Lagrange Multiplier

picture from wikimedia

๊ฑฐ๋‘์ ˆ๋ฏธํ•˜๊ณ  ๋ฐ”๋กœ ์–ด๋–ป๊ฒŒ ํ•˜๋Š” ๊ฑด์ง€ ๋ฐ”๋กœ ์‚ดํŽด๋ณด์ž.

์ด์ฐจ์› ํ‰๋ฉด ์ƒ์— ํ•จ์ˆ˜ $z = f(x, y)$๊ฐ€ ์žˆ๋‹ค. ์ด ํ•จ์ˆ˜์˜ ์ตœ๋Œ€/์ตœ์†Œ ๊ฐ’์„ ๊ตฌํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ด๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์ตœ๋Œ€/์ตœ์†Œ ๊ฐ’์„ ํ•จ์ˆ˜์˜ ์ •์˜์—ญ ์ „์ฒด๊ฐ€ ์•„๋‹ˆ๋ผ ํŠน์ • ๊ตญ์†Œ ๋ฒ”์œ„์— ์•ˆ์—์„œ์˜ ์ตœ๋Œ€/์ตœ์†Œ ๊ฐ’์„ ์ฐพ๊ณ  ์‹ถ๋‹ค. ์ด๋•Œ, ๊ตญ์†Œ ๋ฒ”์œ„๋Š” ์˜์—ญ(Region)์ด ์•„๋‹ˆ๋ผ $g(x, y) = k$๋ผ๋Š” ๊ณก์„ ์œผ๋กœ ์ฃผ์–ด์ง„ ์ƒํ™ฉ์ด๋‹ค. ์ด ๊ณก์„  ์œ„์—์„œ ํ•จ์ˆ˜ $f(x, y)$์˜ ์ตœ๋Œ€/์ตœ์†Œ ๊ฐ’์„ ๊ตฌํ•ด์•ผ ํ•˜๋ฉฐ, $g(x, y) = k$๋ฅผ ์ œ์•ฝ์กฐ๊ฑด(Constraint)๋ผ๊ณ  ํ•œ๋‹ค.

๊ณก์„  $g(x, y) = k$ ์œ„์— ์กด์žฌํ•˜๋Š” ํ•จ์ˆ˜ $f(x, y)$์˜ ์ตœ๋Œ€/์ตœ์†Œ ๊ฐ’์˜ ์œ„์น˜๊ฐ€ $(a, b)$๋ผ๊ณ  ํ•˜์ž. ๊ทธ๋Ÿฌ๋ฉด, ์œ„์น˜ $(a, b)$๋Š” ์•„๋ž˜์˜ ์•„๋ž˜์˜ ๋“ฑ์‹์„ ๋งŒ์กฑํ•œ๋‹ค.

\[g(a, b) = k\]

์ œ์•ฝ์กฐ๊ฑด์„ ์ด๋ฃจ๋Š” ๊ณก์„  $g(x, y) = k$์— ์žˆ์œผ๋‹ˆ ์ด๊ฑด ๋‹น์—ฐํ•˜๋‹ค.

\[\nabla f = \lambda \nabla g\]

๊ฐ‘์ž๊ธฐ Gradient Vector๊ฐ€ ๋‚˜์™”๋‹ค!! ์œ„์˜ ์‹์€ ์ตœ๋Œ€/์ตœ์†Œ ์œ„์น˜ $(a, b)$์—์„œ ํ•จ์ˆ˜ $f(x, y)$์™€ $g(x, y)$์˜ Gradient Vector๊ฐ€ ์„œ๋กœ ๊ฐ™์€(parallel) ๋ฐฉํ–ฅ์„ ๋ฐ”๋ผ๋ณธ๋‹ค๋Š” ๊ฒƒ์„ ๋งํ•œ๋‹ค. ํ‰ํ–‰์„ ํ‘œํ˜„ํ•˜๊ธฐ ์œ„ํ•ด $\lambda \in \mathbb{R}$๋กœ ํ‘œํ˜„ํ•œ ๊ฒƒ.

์œ„์™€ ๊ฐ™์€ Gradient Vector๊ฐ€ ํ‰ํ–‰ํ•œ ์ƒํ™ฉ์ด ๋‚˜์˜ค๋Š” ์ด์œ ๋Š” ์ œ์•ฝ์กฐ๊ฑด $g(x, y) = k$ ๊ณก์„ ๊ณผ ํ•จ์ˆ˜ $f(x, y)$์˜ Level Curve $f(x, y) = z_0$๊ฐ€ ์„œ๋กœ ์ ‘ํ•˜๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.

์ฒ˜์Œ์—๋Š” Constraint Curve์™€ Level Curve๊ฐ€ ์ ‘ํ•œ๋‹ค๋Š” ์กฐ๊ฑด์ด ์ดํ•ด๊ฐ€ ์•ˆ ๋˜์—ˆ๋‹ค. โ€œ๊ผญ ์ ‘ํ•ด์•ผ๋งŒ ์ตœ๋Œ€/์ตœ์†Œ๊ฐ€ ์žˆ๋Š”๊ฑด๊ฐ€?โ€๋ผ๋Š” ์ƒ๊ฐ์ด ๋“ค์—ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์ฐพ์•„๋ณด๋‹ˆ stackexchange ์‚ฌ์ดํŠธ์—์„œ ์š”๋Ÿฐ ๋‹ต๋ณ€์„ ์ฐพ๊ณ  ๋“œ๋””์–ด ์ดํ•ด๊ฐ€ ์ข€ ๋˜์—ˆ๋‹ค.

For $f(x, y) = d$, you increment value $d$ until you touch $๐‘”(x, y)=c$. In the moment of contact you take a minimum. If you go on, just before $f(x, y) = d$ leaves the contact, you take the maximum.

์ฆ‰, $f(x, y) = d_{min}$ํ•˜๋Š” ์ ์—์„œ๋ถ€ํ„ฐ ์ ์  ํ•จ์ˆซ๊ฐ’์„ ํ‚ค์šฐ๋ฉฐ Level Curve๋ฅผ ํ™•์žฅํ•˜๋‹ค๊ฐ€ $g(x, y) = k$์™€ ์ ‘ํ•˜๋Š” ๊ทธ ์ˆœ๊ฐ„์ด minimum ์ˆœ๊ฐ„์ด๋‹ค. ์—ฌ๊ธฐ์„œ Level Curve์˜ ๊ฐ’์„ ๋” ๋Š˜๋ฆฌ๋ฉด ์ œ์•ฝ์กฐ๊ฑด์€ ๋งŒ์กฑํ•˜์ง€๋งŒ ์ ‘ํ•˜๋˜ ์ˆœ๊ฐ„๋ณด๋‹ค๋Š” ํ•จ์ˆซ๊ฐ’์ด ์ปค์ ธ๋ฒ„๋ฆฐ๋‹ค.

picture from wikimedia

๊ทธ๋ฆผ์„ ๊ณ๋“ค์—ฌ ํ•จ๊ป˜ ์ดํ•ดํ•ด๋ณด์ž.


\[\nabla f = \lambda \nabla g\]

Gradient Vector๋กœ ์ด๋ค„์ง„ ์‹์€ ๋ฒกํ„ฐ ๋“ฑ์‹์ด๋‹ค. ๋”ฐ๋ผ์„œ ์ด๊ฒƒ์„ ํ’€๊ธฐ ์œ„ํ•ด์„  ์„ฑ๋ถ„๋ณ„๋กœ ๋“ฑ์‹์„ ์„ธ์›Œ์„œ ํ’€๋ฉด ๋œ๋‹ค. ์ฒ˜์Œ์˜ ์ œ์•ฝ ์กฐ๊ฑด ์‹๊นŒ์ง€ ๊ฐ™์ด ์ ์œผ๋ฉด ์•„๋ž˜์™€ ๊ฐ™๋‹ค.

\[\begin{aligned} f_x &= \lambda g_x \\ f_y &= \lambda g_y \\ g(x, y) &= k \end{aligned}\]

์œ„์˜ ์—ฐ๋ฆฝ๋ฐฉ์ •์‹์„ ํ’€๋ฉด ์ œ์•ฝ์กฐ๊ฑด ์œ„์—์„œ์˜ ์ตœ๋Œ€/์ตœ์†Œ๊ฐ’๊ณผ ๊ทธ ์œ„์น˜๋ฅผ ์•Œ ์ˆ˜ ์žˆ๋‹ค. ์œ„์˜ ์—ฐ๋ฆฝ์‹์—์„œ ๊ตฌํ•ด์•ผ ํ•  ๋ฏธ์ง€์ˆ˜๋Š” ์ตœ๋Œ€์ตœ์†Œ๊ฐ’์˜ ์œ„์น˜ $(x, y)$ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ multiplier์ธ $\lambda$์˜ ๊ฐ’๋„ ๋ฏธ์ง€์ˆ˜๋กœ์„œ ๊ทธ ๊ฐ’์„ ์ฐพ์•„์•ผ ํ•œ๋‹ค. ์ด๋•Œ $\lambda$๋Š” ๊ทธ ๊ฐ’์„ ์ฐพ์•„๋„ ๋ณ„ ์˜๋ฏธ๋Š” ์—†์ง€๋งŒ ๋ฐฉ์ •์‹์„ ํ’€๋‹ค๋ณด๋ฉด ๊ทธ ๊ฐ’์„ ๋ฐ˜๋“œ์‹œ ์ฐพ์•„์•ผ ํ•œ๋‹ค๋Š” ๊ฑธ ๊นจ๋‹ซ๊ฒŒ ๋œ๋‹ค. (์˜ˆ์ œ์—์„œ ๋Š๋ผ๊ฒŒ ๋  ๊ฒƒ.)

Constrained Maxima/Minima

Find the point $(x, y, z)$ on the plane $2x + y - z = 5$ that is closest to the origin.

ํ‰๋ฉด $g(x, y, z) = 5$ ์œ„์— ์žˆ์œผ๋ฉด์„œ๋„ distance ํ•จ์ˆ˜ $d(x, y, z) = x^2 + y^2 + z^2$์˜ ๊ฐ’์„ ์ตœ์†Œํ™” ํ•˜๋Š” ์  $p(x, y, z)$๋ฅผ ์ฐพ์•„์•ผ ํ•œ๋‹ค. Lagrange Method๋ฅผ ์‚ฌ์šฉํ•ด Gradient์— ๋Œ€ํ•œ ์‹ $\nabla d = \lambda \nabla g$์„ ์„ธ์šฐ๋ฉด

\[\begin{aligned} 2x &= 2 \lambda \\ 2y &= \lambda \\ 2z &= - \lambda \\ \end{aligned}\]

์ด๊ฒƒ์„ $g(x, y, z) = 2x + y - z = 5$์— ๋Œ€์ž…ํ•˜๋ฉด, $2 \lambda + \lambda/2 + \lambda/2 = 5$๊ฐ€ ๋˜๊ณ , $\lambda = 5/3$์ด ๋œ๋‹ค.

์ด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋‹ค์‹œ $(x, y, z)$์— ๋Œ€์ž…ํ•˜๋ฉด, closest point $(x, y, z) = (5/3, 5/6, -5/6)$๊ฐ€ ๋œ๋‹ค. $\blacksquare$

Lagrange Method with Two Constraints

Thomas Calculus 13th ed.

์–ด๋–ค ๊ฒฝ์šฐ๋Š” ์ œ์•ฝ ์กฐ๊ฑด์ด 2๊ฐœ ์กด์žฌํ•  ์ˆ˜๋„ ์žˆ๋‹ค.

\[g_1(x, y, z) = 0 \text{ and } g_2(x, y, z) = 0\]

์ด๋•Œ ๋‘ ์ œ์•ฝ์กฐ๊ฑด์€ ๋‘˜๋‹ค ๋ฏธ๋ถ„๊ฐ€๋Šฅ ํ•˜๋ฉฐ, Gradient Vector๊ฐ€ ์„œ๋กœ ํ‰ํ–‰ํ•˜์ง€ ์•Š์•„์•ผ ํ•œ๋‹ค.

๋‘ ์ œ์•ฝ์กฐ๊ฑด $g_1 = 0$์™€ $g_2 = 0$๊ฐ€ ์„œ๋กœ ๊ต์ฐจํ•˜์—ฌ ์ƒ๊ธด ๊ณก์„  $C$๋ฅผ ์ƒ๊ฐํ•ด๋ณด์ž. ์šฐ๋ฆฌ๋Š” ์ด ๊ณก์„  $C$ ์œ„์—์„œ ํ•จ์ˆ˜ $f(x, y, z)$์˜ ๊ทน๋Œ€/๊ทน์†Œ ๊ฐ’์„ ์ฐพ์•„์•ผ ํ•œ๋‹ค. ๊ทธ๋ฆฌ๊ณ , ๊ณก์„  $C$๋Š” ํ•จ์ˆ˜ $f(x, y, z)$์™€ ์ ‘ํ•˜๋Š” ์ง€์ ์—์„œ ๊ทน๋Œ€/๊ทน์†Œ ๊ฐ’์„ ๊ฐ–๋Š”๋‹ค. ์ด๊ฒƒ์€ ์ œ์•ฝ์กฐ๊ฑด์ด ํ•˜๋‚˜ ์˜€์„ ๋•Œ์™€ ๋น„์Šทํ•œ ํŒจํ„ด์ด๋‹ค.

\[C \perp \nabla f\]

๋˜, ๊ณก์„  $C$๋Š” $\nabla g_1$, $\nabla g_2$์™€๋„ ์ˆ˜์ง ๊ด€๊ณ„๋ฅผ ์ด๋ฃฌ๋‹ค. ์ด๊ฒƒ์€ ๊ณก์„  $C$๊ฐ€ ์ œ์•ฝ์กฐ๊ฑด์„ ์ด๋ฃจ๋Š” ํ‰๋ฉด ์œ„์— ์กด์žฌํ•˜๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.

\[C \perp \nabla g_1 \text{ and } C \perp \nabla g_2\]

์œ„์˜ ๋‘ ์‚ฌ์‹ค์„ ๋ฐ”ํƒ•์œผ๋กœ $\nabla f$๊ฐ€ ์„œ๋กœ ํ‰ํ–‰ํ•˜์ง€ ์•Š์€ ๋‘ ๋ฒกํ„ฐ $\nabla g_1$, $\nabla g_2$๊ฐ€ ์ด๋ฃจ๋Š” ํ‰๋ฉด ์œ„์— ์žˆ์Œ์„ ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ์•„๋ž˜์™€ ๊ฐ™์€ ์ผ์ฐจ ๊ฒฐํ•ฉ ์‹์ด ์œ ๋„๋œ๋‹ค.

\[\nabla f = \lambda \nabla g_1 + \mu \nabla g_2\]


์˜ˆ์ œ๋ฅผ ํ†ตํ•ด ์ข€๋” ์ตํ˜€๋ณด์ž.

The plane $x + y + z = 1$ cuts the cylinder $x^2 + y^2 = 1$ in an ellipse. Find the points on the ellipse that lie closest to and farthest from the origin.

Thomas Calculus 13th ed. - Example Problem

์ œ์•ฝ ์กฐ๊ฑด๊ณผ ๊ฑฐ๋ฆฌ ํ•จ์ˆ˜๋ฅผ ์ •์˜ํ•˜์ž.

\[\begin{aligned} g_1 &= x + y + z - 1 = 0 \\ g_2 &= x^2 + y^2 - 1 = 0 \\ d &= x^2 + y^2 + z^2 \end{aligned}\]

Lagrange Method์— ๋”ฐ๋ผ Gradient์— ๋Œ€ํ•œ ์„ฑ๋ถ„์œผ๋กœ ๋“ฑ์‹๋“ค์„ ์„ธ์šฐ์ž.

\[\begin{aligned} \lambda + 2x \mu &= 2x \\ \lambda + 2y \mu &= 2y \\ \lambda &= 2z \\ \end{aligned}\]

์œ„์˜ ๋“ฑ์‹์— $\lambda = 2z$๋ฅผ ๋Œ€์ž…ํ•˜๊ณ  ์ •๋ฆฌํ•˜๋ฉด

\[\begin{aligned} z &= x(1 - \mu) \\ z &= y(1 - \mu) \\ \end{aligned}\]

์œ„์˜ ๋“ฑ์‹์€ 2๊ฐ€์ง€ ๊ฒฝ์šฐ์—์„œ ์„ฑ๋ฆฝํ•˜๊ฒŒ ๋˜๋Š”๋ฐ,

  1. $z = 0$์ด๊ณ , $\mu = 1$
  2. $x = y$์ด๊ณ , $\mu \ne 1$

1๋ฒˆ์˜ ๊ฒฝ์šฐ๋Š” $\lambda = 2z = 0$์ด ๋˜๊ณ , ์ง€์ ์€ $(1, 0, 0)$ ๋˜๋Š” $(0, 1, 0)$์ด ๋œ๋‹ค. ๊ทธ๋ฆผ์—์„œ $x$์ถ•, $y$์ถ• ์œ„์— ์žˆ๋Š” ์ ๋“ค์ด๋ฉฐ, closest point๋ฅผ ์ด๋ฃฌ๋‹ค.

2๋ฒˆ์˜ ๊ฒฝ์šฐ๋Š” $x = y = \pm \frac{1}{\sqrt{2}}$๊ฐ€ ๋˜๋ฉฐ, $z = 1 \mp \sqrt{2}$๊ฐ€ ๋œ๋‹ค. ๊ทธ๋ฆผ์—์„œ $P_1$, $P_2$๊ฐ€ ๋ฐ”๋กœ ๊ทธ ์ ๋“ค์ด๋ฉฐ, farthest point๋ฅผ ์ด๋ฃฌ๋‹ค.

Categories:

Updated: