Vous êtes sur la page 1sur 7

IV.

EXTREME VALUES OF FUNCTIONS OF SEVERAL REAL


VARIABLES

IV. EXTREME VALUES OF FUNCTIONS OF SEVERAL REAL VARIABLES

1. Stationary Points
Definition 1.1. Let D Rn and f : D R. The point a D is said to be:
(1) a local maximum if f (x) 6 f (a) for all points x sufficiently close to a;
(2) a local minimum if f (x) > f (a) for all points x sufficiently close to a;
(3) a global (or absolute) maximum if f (x) 6 f (a) for all points x D;
(4) a global (or absolute) minimum if f (x) > f (a) for all points x D;
(5) a local or global extremum if it is a local or global maximum or minimum.
Definition 1.2. Let D Rn and f : D R. The point a D is said to be a critical
or stationary point if f (a) = 0 and a singular point if f does not exist at a.
Fact 1.3. Let D Rn and f : D R. If f has a local or global extremum at the
point a D, then a must be either:
(1) a critical point of f , or
(2) a singular point of f , or
(3) a boundary point of D.
Fact 1.4. If f is a continuous function on a closed bounded set then f is bounded
and attains its bounds.
Definition 1.5. A critical point a which is neither a local maximum nor minimum
is called a saddle point.
Fact 1.6. A critical point a is a saddle point if and only if there are arbitrarily small
values of h for which f (a + h) f (a) takes both positive and negative values.
Definition 1.7. If f : R2 R is a function of two variables such that all second
order partial derivatives exist at the point (a, b), then the Hessian matrix of f at
(a, b) is the matrix

fxx fxy
H=
,
fyx fyy
where the derivatives are evaluated at (a, b).
If f : R3 R is a function of three variables such that all second order partial
derivatives exist at the point (a, b, c), then the Hessian of f at (a, b, c) is the matrix

fxx fxy fxz


H = fyx fyy fyz ,
fzx fzy fzz
where the derivatives are evaluated at (a, b, c).
Definition 1.8. Let A be an n n matrix and, for each 1 6 r 6 n, let Ar be the
r r matrix formed from the first r rows and r columns of A. The determinants
det(Ar ), 1 6 r 6 n, are called the leading minors of A.

IV. EXTREME VALUES OF FUNCTIONS OF SEVERAL REAL VARIABLES

Theorem 1.9 (The Leading Minor Test). Suppose that f : R2 R is a sufficiently


smooth function of two variables with a critical point at (a, b) and H the Hessian of
f at (a, b). If det(H) 6= 0, then (a, b) is:
2
(1) a local maximum if 0 > det(H1 ) = fxx and 0 < det(H) = fxx fyy fxy
;
2
(2) a local minimum if 0 < det(H1 ) = fxx and 0 < det(H) = fxx fyy fxy ;
(3) a saddle point if neither of the above hold.
where the partial derivatives are evaluated at (a, b).
Suppose that f : R3 R is a sufficiently smooth function of three variables with a
critical point at (a, b, c) and Hessian H at (a, b, c). If det(H) 6= 0, then (a, b, c) is:
(1) a local maximum if 0 > det(H1 ), 0 < det(H2 ) and 0 > det(H3 );
(2) a local minimum if 0 < det(H1 ), 0 < det(H2 ) and 0 < det(H3 );
(3) a saddle point if neither of the above hold.
where the partial derivatives are evaluated at (a, b, c).
In each case, if det H = 0, then (a, b) can be either a local extremum or a saddle
point
Proof. Proof not examinable.

Key Points.
A continuous function on a closed bounded set is bounded and achieves
its bounds.
To find the extreme values of a function on a closed bounded set it
is necessary to consider the value of the function at stationary points
(f = 0), singular points (f does not exist) and boundary points
(points on the edge of the set).
Stationary points can be classified as local maxima, local minima or
saddle points.
If The Leading Minor Test 1.9 is not applicable, the stationary point
must be classified by directly applying Definition 1.1 and Fact 1.6. For
example in the two variable case, if f has a stationary point at (a, b)
we consider the sign of
f (a + h, b + k) f (a, b)
for arbitrarily small, positive and negative values of h and k (that are
not both zero).

IV. EXTREME VALUES OF FUNCTIONS OF SEVERAL REAL VARIABLES

Example. Find and classify the stationary points of the following functions:
(1) f (x, y, z) = x4 + x2 y + y 2 + z 2 + xz + 1;
(2) f (x, y) = y 2 + (x + 1)2 y + (x + 1)4 .
Solution. (a) f (x, y, z) = x4 + x2 y + y 2 + z 2 + xz + 1, so
f (x, y) = (4x3 + 2xy + z)i + (x2 + 2y)j + (2z + x)k.
Critical points occur when f = 0, i.e. when
(1)

0 = 4x3 + 2xy + z

(2)

0 = x2 + 2y

(3)

0 = 2z + x

Using equations (2) and (3) to eliminate


y and z from (1),
we see that 4x3 x3 x/2 =
2
0 or x(6x 1) = 0, giving x= 0, x = 6/6and x = 6/6.
Hence wehave three
stationary points: (0, 0, 0), ( 6/6, 1/12, 6/12) and ( 6/6, 1/12, 6/12).
Since fxx = 12x2 + 2y, fxy = 2x, fxz = 1, fyy = 2, fyz = 0 and fzz = 2, the Hessian
matrix is

12x2 + 2y 2x 1
2x
2 0 .
H=
1
0 2

At ( 6/6, 1/12, 6/12),

11/6
6/3 1

H = 6/3
2
0 ,
1
0
2
which has leading minors 11/6 > 0,

11/6
6/3
det
= 11/3 6/9 = 3 > 0
6/3
2

and det H = 22/312/92 = 4 > 0. By the Leading Minor Test, then, ( 6/6, 1/12, 6/12)
is a localminimum.
At ( 6/6, 1/12, 6/12),

11/6

6/3
1

H = 6/3
2
0 ,
1
0
2
which has leading minors 11/6 > 0,

11/6
6/3
det
= 11/3 6/9 = 3 > 0,
6/3
2

and det H = 22/312/92 = 4 > 0. By the Leading Minor Test, ( 6/6, 1/12, 6/12)
is also a local minimum.

IV. EXTREME VALUES OF FUNCTIONS OF SEVERAL REAL VARIABLES

At (0, 0, 0) the Hessian is

0 0 1
H = 0 2 0 .
1 0 2
Since det(H) = 2, we can apply the leading minor test which tells us that this is a
saddle point since the first leading minor is 0. An alternative method is as follows.
In this case we consider the value of the expression
D = f (0, 0, 0) f (0 + h, 0 + k, 0 + l) = h4 + h2 k + k 2 + l2 + hl,
for arbitrarily small values of h, k and l. But for very small h, k and l, cubic
terms and above are negligible in comparison to quadratic and linear terms, so that
D k 2 + l2 + hl. If h, k and l are all positive, D > 0. However, if k = 0 and h < 0
and 0 < l < |h|, then D < 0. Hence close to (0, 0, 0), f both increases and decreases,
so (0, 0, 0) is a saddle point.
(2) f (x, y) = y 2 + (x + 1)2 y + (x + 1)4 so

f = 2(x + 1)y + 4(x + 1)3 i + 2y + (x + 1)2 j.


Stationary points occur when f = 0, i.e. at (1, 0).
Let us classify this stationary point without considering the Leading Minor Test
(in this case the Hessian has determinant 0 at (1, 0) so the test is not applicable).
Let
D = f (1, 0) f (1 + h, 0 + k) = k 2 + h2 k + h4 .
Completing the square we see that D = (k + h2 /2)2 + 3h4 /4. So for any arbitrarily
small values of h and k, that are not both 0, D > 0 and we see that f has a local
maximum at (1, 0).

IV. EXTREME VALUES OF FUNCTIONS OF SEVERAL REAL VARIABLES

2. Constrained Extrema and Lagrange Multipliers


Definition 2.1. Let f and g be functions of n variables. An extreme value of f (x)
subject to the condition g(x) = 0, is called a constrained extreme value and g(x) = 0
is called the constraint.
Definition 2.2. If f : Rn R is a function of n variables, the Lagrangian function
of f subject to the constraint g(x1 , . . . , xn ) = 0 is the function of n + 1 variables
L(x1 , . . . , xn , ) = f (x1 , . . . , xn ) + g(x1 , . . . , xn ),
where is known as the Lagrange multiplier.
The Lagrangian function of f subject to the k constraints gi (x1 , . . . , xn ) = 0, 1 6
i 6 k, is the function with k Lagrange multipliers, i , 1 6 i 6 k,
L(x1 , . . . , xn , ) =f (x1 , . . . , xn ) +

k
X

i gi (x1 , . . . , xn ).

i=1

Key Points.
To find the extreme values of f subject to the constraint g(x) = 0:
(1) calculate L, remembering that it is a function of the n + 1 variables x1 , . . . , xn and ;
(2) find values of x1 , . . . , xn such that L(x1 , . . . , xn , ) = 0 (you do
not have to explicitly find the corresponding values of ):
(3) evaluate f at these points to find the required extrema.
Note that the equation L = 0 is equivalent to the equations
L
0=
= fxi + gxi , 1 6 i 6 n,
and g(x1 , . . . , xn ) = 0.
xi
So, in the two variable case, we have Lagranian function L(x, y, ) =
f (x, y) + g(x, y) and are solving the equations:
0 = fx + gx ,

0 = fy + gy ,

and

g(x, y) = 0.

With more than one constraint we solve the equation


L(x1 , . . . , xn , 1 , . . . , k ) = 0.
Theorem 2.3 (Not Examinable). Let f : R2 R and P = (x0 , y0 ) be a point on the
curve C, with equation g(x, y) = 0, at which f restricted to C has a local extremum.
Suppose that both f and g have continuous partial derivatives near to P and that P
is not an end point of C and that g(x0 , y0 ) 6= 0. Then there is some 0 such that
(x0 , y0 , 0 ) is a critical point of the Lagrangian Function
L(x, y, ) = f (x, y) + g(x, y).
Proof. Sketch only. Since P is not an end point and g 6= 0, C has a tangent at P
with normal g. If f is not parallel to g at P , then it has non-zero projection
along this tangent at P . But then f increases and decreases away from P along C,
so P is not an extremum. Hence f and g are parallel and there is some such
that f = g and the result follows.

IV. EXTREME VALUES OF FUNCTIONS OF SEVERAL REAL VARIABLES

Example. Find the rectangular box with the largest volume that fits inside the
ellipsoid x2 /a2 + y 2 /b2 + z 2 /c2 = 1, given that it sides are parallel to the axes
Solution. Clearly the box will have the greatest volume if each of its corners touch
the ellipse. Let one corner of the box be corner (x, y, z) in the positive octant, then
the box has corners (x, y, z) and its volume is V = 8xyz.
We want to maximize V given that x2 /a2 + y 2 /b2 + z 2 /c2 1 = 0. (Note that since
the constraint surface is bounded a max/min does exist). The Lagrangian is
2

x
y2 z2
L(x, y, z, ) = 8xyz + 2 + 2 + 2 1
a
b
c
and this has critical points when L = 0, i.e. when
L
x
0=
= 8yz + 2 2 ,
x
a
L
y
= 8zx + 2 2 ,
0=
y
b
L
z
0=
= 8xy + 2 2 ,
z
c
2

2
x
L
y
z2
0=
=
+ 2 + 2 1 .

a2
b
c
(Note that L will always be the constraint equation.) As we want to maximize V
we can assume that xyz 6= 0 so that x, y, z 6= 0.) Hence, eliminating , we get
yz
zx
xy
= 4a2
= 4b2
= 4c2 ,
x
y
z
so that y 2 a2 = x2 b2 and z 2 b2 = y 2 c2 . But then x2 /a2 = y 2 /b2 = z 2 /c2 so
x2 y 2 z 2
x2
+
+
=
3
a2
b2
c2
a2

or x = a/ 3, which implies that y = b/ 3 and z= c/ 3 (they are


all positive by
a b
c
assumption). So L has only one stationary point , , , (for some value
3 3 3
of , which we could work out if we wanted to). Since it is the only stationary point
it must the required max and the max volume is
a b c
8abc
8 = .
3 3 3
3 3
1=

Chris Good, Nov 04

Vous aimerez peut-être aussi