Aabcs
Aabcs
eff : IRE IR then the PARTIAL DERIVATIVE LetKEIN If . f :RV-IR and all its partial
y) f(x, y)
lim
is (y)
+t, -
+) y).
f(x,
my
+
INTERIOR POINT while that wr.t. is )=
say that f function of class C"
A point xEIR" is an we is a
Br (x) [S Let f : R-I have all its partial derivatives at If this is true forkEI then fis SMOOTH
of Siff J . .
(a) , ...
.
-o
.
Sinc = -
- +...
3
COS = - ..
6
A PARAMETRISATION of a curve C Th Let HER be open Iff has continuous partial
- E
.
1 +x = 1+ ...
log (1 + x) = x
I
:
I
-
,
fxZ
firs
fay
M SERV
set : f(x , ..., = 3 at a EU then f is CONTINUOUS at a
Hf( y z)
= =
n+ , ,
11
# sequence (n) in IR" converses A Quadratic Form Q() = nQh
=
to aeiff
lim k Pl -
= 0 .
Positive definite
n > 00
-
·
iff Q (h) > 0th Fo
Let AIRL be open then f : A- IR
: ·
Negative definite iff Q(h) < 0th0
is CONTINUOUS at oEIR iff · Positive semidefinite iff Q(h) 0th ,
- sequence (xn) A: ·
Negative semidefinite iff QCh10th
xn + xo = f(xn) + f(x0) If f : IR"-IR AND ITS PARTIAL DERIV ARE CONT ·
In definite iff Q(h) takes both +, -
If g(x) #O ,
then so is f/g . SPECIAL CASES :
·
det Q >0
,0 = POS DEF .
.
be symmetric and M (m , 1)
,
THE GRADIENT IS 1 TO LEVEL CURVES
=
h . ,
(mm) My= Ms S(
Thm Let +: I - 1R", r(t) = (r(t) , rn(t) Let f : IR"-IR be smooth , I , be its level curve
.,
Mr= M
. .
M
.
=
....,
·
then =
( ...,) Xf(r(t)) o (t) = 0 ·
If det MkJOKE POS DEF
Mao
· If det
/R-IR" then THE GRADIENT ISL TOLEVEL SURFACES
Let r: the velocity of V .
= NEG . DEF .
St (Sist)dt)
,
MINIMUM at a EIR" if J By(a)-ExEBylal f(x), fla) .
~ (tz) r(t , ) -
= =
.
NF1 to tan . . .
(b , (t))d
-
is the quadratic form f's 2D Taylor Polynom maximum minimum value Sato
of .
or on in the other case)
f(a+ h) = f(a) + ]Q(h) then JxER7XF(x(0) X Xg(x0) =
·
Then :
Staleya se
Then def - a is minimum
corollary (Multiple Constraints)
I
2
.
Q pos a :
= a is
Q neg def a maximum
·
[xiXgi(o)
LEAST SQUARES METHOD DIMENSION REDUCTION ALGORITHM
/S S ]
squares estimate off are the values of di
21
where s
. f(x, y)
for cyto
=
=
e
= g ,
(f(:) Yi)?
.
·
for 0
xy
=
[xi)S
·
u v
, ,
, ,
nv = +
f(xx + (1 xy) = xf(x) + (1 x)f(y)
- -
.
eigenvectors v
,, v.
Note V , +
V
·
A function is if it is convex
Su = V ; (v)
=
convex 1
Sv = <v
;
=
.
on
every line segment in .
· Let &R be open and convex , fineir Hence F(dv , + BV) < x + 32x2 =
.
·
If fis smooth and strictly convex then it EXP'Le (1 +* 1 + x ,
>1, |=n
Siss
Df (a) = 0 then a is a
convex
()y]
Fixed Step Size o
a ((b + c) ((aob) + (a -)
((f(x
· =
, Y) docdy =
·
Optimal Step Size CaoBb = [B(ao b)
1 Set 1.
SideJay
.
.
2 l := ExEI" : x =(n + tDf(xn) ,
tER3 =
.
3 Find minimum of on l
.
4 At that t , get <n+= xn + t0f(xn) PRINCIPALADES :
·
+
1. Seto choose fixed dE(0 , 1) => D All 1 ; ER
,
.
2 For t , choose t = 1 initially 2) ONB exist
.
3 Halve t until Armijo Condition 3) Diagonalisable
f(xn) f(xn - -
tnXf(x(n)) >, It 18 +(m)4