1. TAYLOR ’ S FORMULA - 8.Taylor n dim
Chapter 8 TAYLOR’S FORMULA. EXTREMES We will extend in this chapter Taylor’s formula to scalar
p
functions with many variables of the form IR. We
f : A IR n+1
remind that in the case
p=1, if A is an interval, f C (A) and a,x
A,
then it exists between a and x so that: n
1 k n 1 1 (*)
f ( x ) d f ( x a ) d f ( x a ) a k k ! ( n 1 )!
We will show that even for
p1, in certain conditions, a similar
formula exists, so we can approximate in a neighbourhood
V the V a
function (of
f with Taylor’s polynomial function T n degree):
n n1 k .
f ( x ) T ( x ) d f ( x a ) n a k k !
One of the most important problems in technique, economy, and so on is the one of optimisation of diverse processes; the optimisation of a process relies in finding the extremes (minimum or maximum) of a function function that mathematically models the process in certain conditions imposed to the variables conditions that give the domain of definition of that function. This type of problems is handled by the theory of optimisation. We will approach only a few techniques of finding the extreme points (locale, sometimes even global) and their nature by using the formula (*).
1. TAYLOR’S FORMULA
p
Let IR be a real function of
f : A IR p real variables, p1.
Definition 1. If f is differentiable n times in the point a A
p
then the polynomial function : IR IR
T n
1 1 n x A
T ( x ) f ( a ) d f ( x a ) ... d f ( x a ), n a a
1 ! n ! is called Taylor’s polynomial function of degree n associated to the
function , then we can write f in the point a; denoting d f f ( a ) n
a
1 k .
T ( x ) d f ( x a ) n a k k !
The function:
R : A IR, R (x)=f(x) T (x), x A n n n
is called remainder of order
n, and the relationship f(x)= T n (x)+ R n (x), x A
we call it Taylor’s formula of order n, or the Taylor’s development of
the function f around the point a.
Remark. Using the operator of differentiation d we have
( k )
k
;
d f dx ... dx f ( a ) a 1 p
x x 1 p
if
a=(a 1 ,…,a p ), x=(x 1 ,…,x p ), then: ( k )
k
d f ( x a ) ( x a ) ... ( x a ) f ( a ) .
a 1 1 p p x x 1 p
In order to extend the formula (*) valid on the interval
A for the case
p1 we need two new notions: the one of segment and the other of
convex set.p
Definition 2. Let x,y IR . We call a segment which binds
the points x and y the set: p x t(y x) I R .
- The set p p x t(y x)
x , y t ,
1
- p
I R
I R
x , y t ,
1
is called an open segment. If for any ,
x,y AIR [x,y]
A, we say
that A is a convex set.
p
be a convex set,
Theorem 1(Taylor’s formula). Let AIR n+1
on Å . For any
f : A IR be a function of class C A and a A x A
exists (
a,x) such that
) ( )! 1 (
... ) ( 1 1 1 )) ( ( ) ( ... ) (
... )) ( ( ) ( ... )) ( ( ) ( ) ( ) ( 1 2 2 1 2 1 1 1 1 ''
.
) 1 ( 1 1 1 a x t a f
x
a x x a xp
p pF a x t p p p
a x t a x x f a x a x t a x f
a x t a x f a x a x t a x f
According to the rules of derivation of the functions of several variables and the preceding Remark we obtain:
F F F F (0) and .
F a x a x t p p p
, we need explicit values ) ( ),..., ( ), ( ' ) ) 1 ( ( ''
F a x t n p
p p
n) ( 1 1 1 ) ( a x t a f x
a x
x… )) ( ( ) ( ... ) ( ) (
.
) 2 ( 1 1 1 a x t a f
x
a x x a xp
p p)) ( ( ) ( ... ) (
a x t a x f a x a x t a x x f a x a x p p p p p p
)) ( ( ) ( ... )) ( ( ) ( ) ( ... 2 2 1 2 1 1
n n
F F F x f
1 ) ( !
1 ,…,x p
F is derivable n+1 times,
it results that
f is of class C n+1
. Because
F a x t a a x t a f a x t a f t
IR defined by: )) ( ),..., ( ( )) ( ( ) ( 1 1 1 p p p
F : [0,1]
function
) and the
), x=(x
F (n+1)
1
,…,a pProof. Let us consider a=(a
.
1 ) ( ) ( 1 2 a x f d n a x f d
n
a x f d a x f d a f x f n n a a a1
1 ) ( !
2
1 ... ) ( !
with the derivative
continuous. From Mac-Laurin’s formula it follows that for any
F n F
n
)! 1 (
n n
) 1 ( ( ) ) 1 ( (
1 ) ( )
1
!
1 ... ) ( '
!
1 ) (
Because ) (
t [0,1] it exists (0,1) so that:
.
n t
F
n t F t F t F t F n n n n 1 ) ( ) ( ) 1 ( 1 ) ( '' 2 F t
!
2 ) ( '
... ) ( !
) ( !
) ( )! 1 (
( n 1 )
( n 1 ) .
F ( t ) ( x a ) ... ( x a ) f ( a t ( x a )) 1
1 p p
x x 1
Therefore, '' 2 (n) n
F( ) f(a), F'( ) d f(x a a a
a), F ( ) d f(x a),...,F ( ) d f(x
a) ( n 1 ) n 1
and .
F ( ) d f ( x a ), where a - (x
a) (a, x)
As a consequence: n n
1 ( k ) ( n 1 1 ) k n
1 1 1 F (
1 ) f ( x ) F ( ) F ( ) d f ( x a ) d f ( x a )
a k k
k ! ( n 1 )! k ! ( n 1 )! and Taylor’s formula is proved.Remark 1. For n=0, from Taylor’s formula we can write p
f ,
f(x) f(a) d f(x
î k k
a) (î ) (x a ), î (a,x) k
x
1 kequality witch represents a generalization (for
p1) of Lagrange’s formula.
Remark 2. The remainder of order n from Taylor’s formula:
1 n 1 R ( x ) d f ( x a ), (a, x) n ( n 1 )! allows us to evaluate the error from the approximation of order n
n:
1 k ,
f(x) T (x) d f(x n a
a), x A k k!
in case that the partial derivatives of order
n+1 of the function f on A
(which interfere in the differential expression of order
n+1) are
bounded by the same constant
M0:
p n 1 n 1 n 1 M M p . f ( x ) T ( x ) R ( x ) x a x a n n i i
( n 1 )! ( n k 1 1 )!
Remark 3. If a=0=(0,...,0)A, Taylor’s formula: n ( k )
1
f ( x ) f ( ) x ... x f ( )
1 p
k k ! x x 1 1 p
( n 1 )
1
x ... x f ( x ) 1 p
( n 1 )! x x 1 p
is called Mac-Laurin’s formula.
Remark 4. In the case p=2 (using Newton’s binomial) the
Taylor development of the function
f around the point (a,b) A (or in powers of x a and y
b) becomes:
1 f f
f ( x , y ) f ( a , b ) ( a , b )( x a ) ( a , b )( y b )
1 ! x y 2 2 2
1 f f f 2 2 ( a , b )( x a ) 2 2 ( a , b )( x a )( y b ) ( a , b )( y b ) 2
2 ! x x y y
n n 1 f k n k k
, ... C ( a , b )( x a ) ( y b ) R ( x , y ) n n n k k
n ! x y k
where: n n 1 1 1 f k n 1 k k
R ( x , y ) C ( , )( x a ) ( y b ) , n n
1
n
1 k k ( n 1 )! x y k
. a -
( x a ), b (y
b), with (0,1) In this case the graphical representation of
f is a surface of
equation:
S : z = f(x,y), (x,y) A.
2
2 Let us suppose that and f C (A), d f 0 d f U is ( a , b ) ( a , b )neighbourhood of the point
(a,b)
A.
The approximation of order 0:
f(x,y) T (x,y) f(a,b), (x,y) U
means, from a geometrical point of view, that if
(x,y) U the point
M(x,y,f(x,y)) is replaced with the point M (x,y,f(a,b)) which belongs to
a part of the parallel plane with
xOy of equation:
S : z = T (a,b) = f(a,b), (x,y)
A,
approximation which, in general, is unsatisfying even if the neighbourhood
U is “small”. (see Fig.1.) z z=T (x,y) M (a,b,f(a,b)) M y
U (a,b,0) x Fig.1.
The approximation of order
1:
f(x,y) T (x,y) f(a,b) d f(x-a,y-b) , (x,y) U 1 (a,b)
is more precise, because it presumes the substitution of the point M with the point
M (x,y,T (x,y)) from the part of the plane tangent in the
1
1
point
P(a,b,f(a,b)) on the surface S:
S 1 : z = T 1 (x,y), (x,y) A.
(see Fig.2.).
The approximation of second order:
1 2
f(x,y) T (x,y) f(a,b) d f(x-a,y-b) d f(x-a,y-b) , (x,y) U 2 (a,b) (a,b)
2 is more precise than the preceding; in this case the point
M(x,y,f(x,y)) on the graphic S, (x,y) U is replaced with the point
M2 (x,y,T 2 (x,y)) belonging to the surface of equation: S : z = T (x,y), (x,y)
A,
2
2
surface tangent in
P(a,b,f(a,b)) on the surface S and which has as plane tangent in (see Fig.3.). P the plane S
1
Example 1. Expand the polynomial function f(x,y,z) = x
3
3
2
- +y
- +xyz z
- +xy+x z+1
in powers of x 1, y+1 and z.
M M 1 S 1 z y x z=T 1 (x,y) (a,b,f(a,b))
(a,b,0)
Fig.2.
S 2 z y x z=T 2 (x,y) (a,b,f(a,b))
(a,b,0) A Fig.3
Solution. We will use Taylor’s development of the function f
3
around the point . Because
a=(1, 1,1) IR f is a polynomial of third
4
degree
d f=0, so R
3 (x,y,z)=0 and Taylor’s formula is:
f ( x , y , z ) f ( 1 , 1 , 1 ) d f ( x ( 1 , 1 , 1 ) 1 , y 1 , z 1 ) 1 2
1
3
d f ( x ( 1 , 1 , 1 ) ( 1 , y 1 , z 1 ) d f ( x 1 , 1 , 1 ) 1 , y 1 , z 1 )2 2 2
6 But
,
df 2 3 x dx 3 y dy yzdx zxdy xydz 2 2 2 zdz ydx xdy dx dz d f
6 xdx 6 ydy ( zdy ydz ) dx ( zdx xdz ) dy 2 , ( ydx xdy ) dz 2 dz 2 dxdy and 3 3 3 .
d f
6 ( dx dy dxdydz ) Now we compute
,
f(1,1,1)=2, d f 2 2 ( 1 , 1 , 1 ) 2 2 dx 2 5 dy 4 dz
;
d f ( 1 , 1 , 1 ) 6 dx 6 dy 2 dz 4 dxdy 2 dydz 2 dzdx therefore: 2 2 2 f ( x , y , z )
2 2 ( x 1 ) 5 ( y 1 ) 4 ( z 1 ) 3 ( x 1 ) 3 ( y 1 ) ( z 1 ) 2 ( x 1 )( y 1 ) ( y 1 )( z 1 ) ( z 1 )( x 1 ) 3 3 . ( x 1 ) ( y 1 ) ( x 1 )( y 1 )( z 1 )
Remark. The polynomial f from the preceding example is a
vector from the linear space IR [
3 x,y,z] of the polynomials with three
variables of degree smaller or equal with three formulated in the
2
2
2
2
2
2
2
2
2
canonical base B ={1,
x, y, z, xy, yz, x , y , z , x y, x z, y x, y z, z x, z y, c
3
3
3
[
x , y , z , xyz} (dim IR x,y,z]=20). The development of f in powers of
3 x 1, y+1, z 1 means in fact formulating this vector in the base
2
2 B={
1, x 1, y+1, z 1, (x 1)(y+1), (y+1)(z 1), (x 1)(z 1), (x 1) , (y+1) ,
2
2
2
2
2
2 (z 1) , (x 1) (y+1), (x 1) (z 1), (y+1) (x 1), (y+1) (z 1), (z 1) (x 1),
2
3
3
3 (z 1) (y+1), (x 1) , (y+1) , (z 1) , (x 1) (y+1) (z 1)}.
Example 2. Using Taylor’s formula of order three compute
2.1 the approximate value of the number (0.9) . y
Solution. We need the value of the function in the f(x,y)=x
point
(x,y)=(0.9; 2.1); we will choose as point (a,b) around which we will develop the function
f in which we can compute precisely the
function z and its partial derivatives and also, as close as possible from ( 2 . 1 x,y). Let then be (a,b)=(1,2). Then:
1 2 ( . 9 ) f ( x , y ) T ( x , y ) f ( 3 ( 1 , 2 ) d f ( x a , y b ) d f ( x a , y b ) 1 , 2 ) ( 1 , 2 )
2
1 3
1
1
1
2 1
1
1 3
1
1 d f ( x a , y b ) 1 d f , d f , d f , ( 1 , 2 ) ( 1 , 2 ) ( 1 , 2 ) ( 1 , 2 )
6
10
10
2
10
10
6
10
10
As y 1 y ,
d f yx dx x ln xdy| ( 2 y 1 , 2 ) ( 1 y 1 , 2 ) 2 2 dx
d f x dxdy y ( y
1 ) x dx ( 1 , 2 ) y 1 y
1 y y
1 x ln x dx yx dx x ln xdy ln xdy x dxdy| 2 3 ( 1 , 2 ) 2
= , and ,
dxdy
2 dx dxdy 2 dx ( dx dy ) d f ( 1 , 2 ) 7 dx dy 2 . 1
2
7 it follows that: . ( . 9 ) 1 . 807 10 1000
Let us observe that the approximation of order 0 is
2.1
2.1
(0.9) T (x,y)=1, and the one of order two is (0.9) T
1 (x,y)=0,8. In
none of these cases we do not have an estimation of the error. If we want to compute an expression with a pre-established precision we must find an .
nIN so that R n 4 Example 3. Let us compute with two exact N 4 .
1 .
9 2 1 4 1 decimals. We consider the function which we
f ( x , y ) x y
compute around the point (4; 1). Because
N=f(x,y) where x=4.1 and y= 0.9, according to Taylor’s formula there exists (4; 4.1) and
(0.9; 1) so that: n
1 k
1
1
N f ( 4 ; 1 ) d f , R ( x , y ) , ( 4 ; 1 ) n k 1 k !10
10
1 n 1
1 1 where . We must find an nIN
R ( x , y ) d f , n ( , )
( n 1 )!
10 10
2
minimal so that , where (
R ( x , y ) n 10 x,y)=(4.1; 0.9). Because:
1 1
1
1
R ( x , y ) d f , f ' ( , ) f ' ( , ) ( , ) x y
10 10
10
10
1 f ' ( , ) f ' ( , ) , x y
10 1 1
1 2 4
1
1
1
and , f ' ( , ) f ' ( , ) x y 4 3
2
2
4
2
3 then and the approximation of order zero:
R ( x , y )
40 N
T (x,y)=f(4.1)=2 does not look very satisfying. Let us try an estimation as precise as possible for the reminder of first order.
Because
1 2
1 1 1 ''
1 2 '' 1 '' 2 2 R ( x , y ) d f , f ( , ) f ( , ) f ( , ) 1 ( , ) xy x 2 2 2 y
2
10
10
2
10
10
10
1 '' '' ''
f 2 ( , )
2 f ( , ) f xy 2 ( , )
x y
200 and 3 1 ''
1 2 4
1 ,
f x 2 ( , )
4
32 1 3 ''
1 2 4
1
1 4
10
5
, f ( , ) xy
8
8
2
9
24 '' 1 7 7
3 2 4
3 1 4 10
3
10
5
, f y 2 ( , )
2
16
16
2
9
32
9
24 1
1
10 5
1 It follows that , so we can affirm
R ( x , y ) 1
200
32
24 24 100 with certitude that the approximation of the first order:
1 1
1
1 1
N T ( x , y ) f ( 1 ( 4 ; 1 ) d f , 4 ; 1 ) 2 2 . 025 1 . 975
10 10 10
4 2 is an evaluation with two exact decimals of the number
N.
2. LOCAL EXTREMA
p
Let f : AIR IR, p1.
Definition 3. The point a A is called local extreme point
(relative) of function such
f if there exists a neighbourhood VV a
that „the increase” of function
f:
E(x)=f(x) f(a)
keeps a constant sigh on the set
A V; if E(x) 0, x A
V
we say that
a is a point of local maximum, (or relative) for f and we
write f =f(a); if
max
E(x) 0, x A
V
the point
a carries the name of local minimum (or relative) of the function f and we write f =f(a). min
If
V=A, and the point a is a local extreme (minimum or
maximum) we say that
a is a global extreme (or absolute) of the
function f.
We saw that, in the case of functions with a single real variable, Fermat’s theorem offers necessary conditions of extreme (if I is an extreme point of the function
a f : I IR IR and f is
differentiable at or, equivalently,
a, then f ' ( a ) d f=0). Fermat’s a theorem admits a generalization for functions with several variables.
Theorem 2(of Fermat - necessary conditions for local p extreme). Let f : AIR IR, p1. If a Å is an extreme point
of the function f and f is differentiable in a, then
f
. d f=0, or, equivalently, (a) , k1 , p
a
x
k
Proof. Because the point a belongs to the interior of the set
A, and it is a local extreme of the function f, it follows that there
exists
r0 such that E(x)= f(x) f(a) keeps a constant sign for all p
and the auxiliary
x S(a,r)
A. Let s be an arbitrary versor from IR
function IR , .
g : ( r , r ) g(t) f(a ts)
Then
g ( t ) g ( ) f ( a ts ) f ( a ) E ( a ts ) keeps a constant sign for
t ( r,r), because a ts S(a,r); this means
that the function
g has a local extreme in t=0. Because g is
differentiable in
t=0 (being a composite function of differentiable
f
functions), from Fermat’s theorem it results that ;
g ' ( ) ( a )
s
f p
therefore , for any versor ; particularly, assigning ( a ) s from IR
s p
(vectors of the canonical basis in IR ) we obtain
s=e k , k
1 , p
f
, or
(a) , k 1 , p d a f=0.
x k
p
Definition 4. Let Å . f : AIR IR, p1 and aIf
f is differentiable at a, and d a f=0, the point a is called stationary point of the function f.
If a is stationary point for f, or f is not differentiable in a, we say that
a is a critical point of the function f (for function f).
Remark 1. From Fermat’s theorem results that the interior
points which are local extremes of the differentiable function
f are
found, if existing, within the solutions of the system of p equations with
p unknowns.
f .
(x ,...x ) , k 1 p 1 , p
x k Remark 2. As in the case p=1, Fermat’s theorem gives
necessary, but not sufficient, conditions of existence of the local extreme points.
For example, the function
2 f : IR IR, f(x,y)=xy
has null partial derivatives in (0,0), but the origin is not a local extreme point of the function
f (f being a quadratic undefined form).
So (0,0) is a stationary point (therefore critical) for
f which is not a
local extreme. But, for the function
g : [0,)[0,) IR, g(x,y)=xy ,
the origin (0,0) is a critical point (
g is not differentiable in (0,0)) which is not stationary, but is a global extreme point.
Remark 3. We have seen that from single variable functions the second order’s differential sign in a stationary point gives us information over the nature of this point. In the case of several variable functions the second order differential (which is a quadratic form) sign in a stationary point will determine the nature of the respective point. We will use the following lemmas:
p Lemma 1. Let : IR IR be a quadratic form. p
(a) If is positively defined (i.e. \ {0})
(x)0, for any x IR 2 p there exists , for . m0 such that ( x ) m x x IR p
(b) If is undefined then there exist IR \ {0 such
s , s
1
2
- that (ts ) 0 and (ts ) 0, for any t IR .
1
2
p
Proof. (a). LetIR . Then S x | x
1 S is a closed and
bounded set, therefore a compact set, and is a continuous
function, therefore is bounded and reaches its edge on
S (theorem
16, cap.5). Let on
m be the minim of S; of course m0 because
is positively defined and:
(x) m, for any x S (1) p
Now, let \ {0}; then
x IR
1
1
1 and ;
x S x ( x )
2
x x x
p consequently, from (1) results that , for any .
( x ) m x x IR
p
(b) Let be the canonical basis from IR . As is of
B c
quadratic form, there exists an orthogonal basis has
B in which
canonical form. Let to
T=(b ij ) be the transformation matrix from B c B.
Then: 2 2 2 (2)
( x ,..., x ) y y ... y 1 p 1 1 2 2 p p
where p (3)
y b x , i i ij j 1 , 2 ,...,p j 1
and , ,..., are the eigenvalues of the attached matrix of the
1 2 p
quadratic form in canonical base. We can assume (with an eventual permutation of the indexes) that 0 and 0 (the form
1
2
being undefined). The systems:
(4)
y 1 i 1 , y , i 2 , 3 ,...,p
(5)
y 2 i 1 , y , i 1 , 3 , 4 ,...,p have, according to (3), the matrix
respectively, a unique solution. Let , respectively
s ( a ,..., a ) 1 1 p be the solutions of the systems (4), respectively (5). s ( b ,..., b ) 2 1 p
Then, according to (2) we obtain: and (6)
( s ) ( s ) 1 1 2
2
- But is a quadratic form, so for any
t IR 2 2 2 and .
( ts ) t ( s ) t ( ts ) t 1 1 1 2 2
2 Lemma 2. Let
f C (A) and a
A. Then there exists a
function
: AIR such that for any x
A: 2
1 2 ,
f ( x ) f ( a ) d f ( x a ) d f ( x a ) x a ( x )
a a2 where . lim ( x ) ( a ) x a
Proof. Let x A \ {a}. According to Taylor’s formula there
exists (
a,x) such that:
1 2 (1)
f ( x ) f ( a ) d f ( x a ) d f ( x a ) a
2 Let 2 2
f f ù
(2) ij
(x) (î ) (a), i ,j
1 , p '' x x x x i j i j
Because and
f C ( A ) a when x x x i j
a, from (2) results: ù
(3) lim (x) , i,j ij 1 , p x a Then: 2 2 '' '' p
d f ( x a ) d f ( x a ) f ( ) f ( a ) ( x a )( x a ) a x x x x i i j j i j i j
i , j
1
so, from (2) we obtain: 2 2 p
(4) d f ( x a ) d f ( x a ) ( x )( x a )( x a )
a ij i i j j
i , j
1 Let (a)0 and
p
2
,when
( x ) ( x )( x a )( x a ) x 2 ij i i j j a.