Transcript Slide 1

Local Linear Approximation
for
Functions of Several Variables
Functions of One Variable
• When we zoom in on a “sufficiently nice”
function of one variable, we see a straight line.
Functions of two Variables
Functions of two Variables
Functions of two Variables
Functions of two Variables
Functions of two Variables
When we zoom in on a “sufficiently nice” function
of two variables, we see a plane.
f ( x, y )  5 x sin( y ) 
y
cos( x)
2
Describing the Tangent Plane
• To describe a tangent line
we need a single number--the slope.
• What information do we
need to describe this
plane?
• Besides the point (a,b), we
need two numbers: the
partials of f in the x- and ydirections.
T( a ,b ) ( x, y ) 
f
( a, b)
y
(a,b) f
x
( a, b)
Equation?
f (a, b)
f (a, b)
( x  a) 
( y  b)  f ( a, b)
x
y
Describing the Tangent Plane
T( a ,b ) ( x, y ) 
f (a, b)
f (a, b)
( x  a) 
( y  b)  f ( a, b)
x
y
We can also write this equation in vector form.
Write x = (x,y), p = (a,b), and
f (p) 
Tp (x)  f (p)  (x  p)  f (p)
Dot product!
f (a, b) f (a, b)
,
x
y
Gradient Vector!
General Linear Approximations
In the expression Tp (x)  f (p) x  f (p) we can think of f (p)
As a scalar function on 2: x  f (p) x . This function is linear.
For a function general vector-valued function
F:
m

n
and for a point p in
we want a linear function Lp :
F(x)  Lp (x)  F(p)
m

n
m
that satisfies
for x "close" to p.
Why don’t we just
subsume F(p) into Lp?
Linear--- in the linear
algebraic sense.
General Linear Approximations
In the expression p (x)  f (p) x  f (p) we can think of f (p)
As a scalar function on 2: x  f (p) x . This function is linear.
For a function general vector-valued function
F:
m

n
and for a point p in
we want a linear function Lp :
F(x)  Lp (x)  F(p)
m

n
m
that satisfies
for x "close" to p.
Note that the expression
Lp (x-p) is not a product.
It is the function Lp acting
on the vector (x-p).
To understand
Differentiability
in
Vector Fields
We must understand Linear Vector Fields:
Linear Transformations from
m
to
n
Linear Functions
A function L is said to be linear provided that
L ( x  y )  L ( x)  L ( y )
and
L ( x )   L ( x )
Note that L(0) = 0, since L(x) = L (x+0) = L(x)+L(0).
For a function L: m →
are very prescriptive.
n,
these requirements
Linear Functions
It is not very difficult to show that if L: m →n is linear,
then L is of the form:
L(x)  L( x1 , x2 , x3 ,
, xm )
 a11 x1  a12 x2
a x  a x
22 2
 21 1
  a31 x1  a32 x2


 an1 x1  an 2 x2
 a13 x3
 a23 x3
 a33 x3



 an 3 x3

 a1m xm 
 a2 m xm 
 a3m xm 


 anm xm 
where the aij’s are real numbers for j = 1, 2, . . . m and
i =1, 2, . . ., n.
Linear Functions
Or to write this another way. . .
L(x)  L( x1 , x2 , x3 ,
 a11 a12
a
 21 a22
  a31 a32


 an1 an 2
 Αx
, xm )
a13
a23
a33
an 3
a1m   x1 
a2 m   x2 
a3m   x3 
 
 
anm   xm 
In other words, every linear function L acts just like leftmultiplication by a matrix. Though they are different, we cheerfully
confuse the function L with the matrix A that represents it! (We
feel free to use the same notation to denote them both except where
it is important to distinguish between the function and the matrix.)
One more idea . . .
 a11 x1
a x
 21 1
A(x)   a31 x1


 am1 x1

a12 x2

a13 x3



a22 x2
a32 x2


a23 x3
a33 x3


 am3 x3

 am 2 x2

a1n xn 
 a2 n xn 
 a3n xn 


 amn xn 
Suppose that A = (A1 , A2 , . . ., An) . Then for 1  j  n
Aj(x) = aj1 x1+aj 2 x2+. . . +aj n xn
What is the partial of Aj with respect to xi?
One more idea . . .
 a11 x1
a x
 21 1
A(x)   a31 x1


 am1 x1

a12 x2

a13 x3



a22 x2
a32 x2


a23 x3
a33 x3


 am3 x3

 am 2 x2

a1n xn 
 a2 n xn 
 a3n xn 


 amn xn 
Suppose that A = (A1 , A2 , . . ., An) . Then for 1  j  n
Aj(x) = aj1 x1+aj 2 x2+. . . +aj n xn
The partials of the Aj’s are the entries in the matrix
that represents A!
Local Linear Approximation
Lp (x)  Αx
where A is the m  n matrix
 a11
a
 21
A   a31


 an1
a12
a13
a22
a23
a32
a33
an 2
an 3
a1m 
a2 m 
a3m 


anm 
For x "close" to p we have
F(x)  Lp (x)  F(p)
For all x we have
F(x)  Lp (x)  F(p)  E(x)
where E(x) is the error
committed by
Lp in approximating F.
Local Linear Approximation
Suppose that F: m → n is
given by coordinate functions
F=(F1, F2 , . . ., Fn) and all the
partial derivatives of F exist at p
in m and are continuous at p
then . . .
there is a matrix Lp such that F
can be approximated locally near
p by the affine function
Lp (x  p)  F(p)
What can we say about
the relationship between
the matrix DF(p) and
the coordinate functions
F1, F2, F3, . . ., Fn ?
Quite a lot, actually. . .
Lp will be denoted by DF(p)
and will be called the
Derivative of F at p or the
Jacobian matrix of F at p.
A Deep Idea to take on Faith
I ask you to believe that for all i and j with 1 i  n
and 1  j  m
L j
xi
(p) 
Fj
xi
(p)
This should not be too hard. Why?
Think about tangent lines, think about tangent planes.
Considering now the matrix formulation, what is the
partial of Lj with respect to xi?
The Derivative of F at p
(sometimes called the Jacobian Matrix of F at p)
F1
F1
 F1
(
p
)
(
p
)
(p)
 x

x

x
2
3
 1
F3
 F2
F2
(
p
)
(
p
)
(p)
 x

x

x
2
3
 1
D F(p)  J F (p)   F3
F3
F3
(
p
)
(
p
)
(p)


x

x

x
2
3
 1


Notice
two
Fm common
Fm
 Fm the
(p)
(p)
(p)


x

x

x
nomeclatures
for the
2
3
 1
derivative of a vector –
valued function.
F1

(p) 
xn


F2
(p) 
xn


F3
(p) 
xn



Fm
(p) 

xn
In Summary. . .
If F is a “reasonably well
How should
behaved” vector
field we
around thethink
pointabout
p, wethis
function
E(x)?
can form the
Jacobian
Matrix DF(p).
For x "close" to p we have
F(x)  DF(p)(x  p)  F(p)
For all x, we have
F(x)=DF(p) (x-p)+F(p)+E(x)
where E(x) is the error committed by
DF(p) +F(p)
in approximating F(x).
E(x) for One-Variable Functions
E(x) measures the vertical
distance between f (x) and Lp(x)
But E(x)→0 is not enough, even
for functions of one variable!
( x, Lp ( x))
( x, Lp ( x))
E ( x)
E ( x)
( p, f ( p))
( x, f ( x))
What happens to E(x)
as x approaches p?
( p, f ( p))
( x, f ( x))
Definition of the Derivative
A vector-valued function F=(F1, F2 , . . ., Fn) is differentiable at
p in m provided that
1) All partial derivatives of the component functions
F1, F2 , . . ., Fn of F exist on an open set containing p, and
2) The error E(x) committed by DF(p)(x-p)+F(p) in
approximating F(x) satisfies
E(x)
lim
0
x p x  p
(Book writes this as
lim
xp
F(x)  DF(p)(x  p)  F(p)
xp
 0 .)