Scribd
Upload
23 views7 pages

Unconstrained Minimization

This document discusses unconstrained minimization problems and algorithms for solving them. It introduces the optimality condition, descent methods for finding minimizing sequences through descent directions and step sizes, gradient descent using the negative gradient as the descent direction, and Newton's method using the inverse Hessian as the descent direction.

Uploaded by

tsilavinarakotomavo2002
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
23 views7 pages

Unconstrained Minimization

This document discusses unconstrained minimization problems and algorithms for solving them. It introduces the optimality condition, descent methods for finding minimizing sequences through descent directions and step sizes, gradient descent using the negative gradient as the descent direction, and Newton's method using the inverse Hessian as the descent direction.

Uploaded by

tsilavinarakotomavo2002
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 7

Unconstrained minimization

Mathématiques Informatique et Statistique Appliquées (MISA)


Université d’Antananarivo
Unconstrained minimization problems

minimize f (x)
where f : RN → R is convex and twice differentiable. We assume that the optimal
value p ? = inf x f (x) is attained (and finite).
Optimality condition
f is differentiable and convex, a point x ? is optimal iff ∇f (x ? ) = 0.

Sometimes, we can analytically solve the optimality equation but usually the problem
must be solved by an iterative algorithm.
Descent methods
We want to produce a minimizing sequence x (k) , k = 1, ..., where
x (k+1) = x (k) + t (k) ∆x (k) with f (x (k+1) ) < f (x (k) )
I ∆x (k) is the search direction
I t (k) > 0 is the step size

Descent direction
∆x (k) is a descent direction if ∇f (x (k) )> ∆x (k) < 0 (from convexity).
Algorithm 1: General descent method
Start at a point x ∈ dom f
repeat
Determine a descent direction ∆x
Choose a step size t > 0
Update x = x + t∆x
until stopping criterion is satisfied;
Choosing the step size - Line search

Exact line search


Choose t to minimize f along the ray {x + t∆x|t > 0}

t = arg min f (x + t∆x)


t>0

Backtracking line search


An inexact line search approach that chooses t just to reduce f enough. It depends on
two constants α ∈ (0, 0.5) and β ∈ (0, 1).
Algorithm 2: Backtracking line search
t=1
repeat
t = βt
until f (x + t∆x) < f (x) + αt∇f (x)> ∆x;
Gradient descent

It is a descent method with search direction ∆x = −∇f (x).


Algorithm 3: Gradient descent
Start at a point x ∈ dom f
repeat
∆x = −∇f (x)
Choose a step size t > 0
Update x = x + t∆x
until stopping criterion is satisfied;
The stopping criterion is usually of the form ||∇f (x)||2 < .
Newton’s method

It is a descent method with search direction ∆x = −(∇2 f (x))−1 ∇f (x).

We can say that v = ∆x


I minimizes the second order approximation

1
f (x + v ) ≈ f (x) + ∇f (x)> v + v > ∇2 f (x)v
2
I or equivalently solves the linearized optimality condition

∇f (x + v ) ≈ ∇f (x) + ∇2 f (x)v = 0
More

Explore further
I Steepest descent methods (generalizes gradient descent)
I Quasi-Newton methods (like BFGS, L-BFGS)
I Convergence analysis
I Self-concordance
I Gradient-free optimization

You might also like