Module

for

Steepest Descent or Gradient Method

 

Gradient and Newton's Methods

    Now we turn to the minimization of a function
[Graphics:Images/GradientSearchMod_gr_1.gif] of  n  variables, where  [Graphics:Images/GradientSearchMod_gr_2.gif]  and the partial derivatives of [Graphics:Images/GradientSearchMod_gr_3.gif] are accessible.  

 

Steepest Descent or Gradient Method

Definition (Gradient).  Let [Graphics:Images/GradientSearchMod_gr_4.gif] be a function of [Graphics:Images/GradientSearchMod_gr_5.gif] such that  [Graphics:Images/GradientSearchMod_gr_6.gif]  exists for  [Graphics:Images/GradientSearchMod_gr_7.gif].  The gradient of [Graphics:Images/GradientSearchMod_gr_8.gif], denoted by  [Graphics:Images/GradientSearchMod_gr_9.gif],  is the vector  

(1)        
[Graphics:Images/GradientSearchMod_gr_10.gif].  

Illustrations

 

    Recall that the gradient vector in (1) points locally in the direction of the greatest rate of increase of [Graphics:Images/GradientSearchMod_gr_17.gif].  Hence  [Graphics:Images/GradientSearchMod_gr_18.gif]  points locally in the direction of greatest decrease [Graphics:Images/GradientSearchMod_gr_19.gif].  Start at the point  [Graphics:Images/GradientSearchMod_gr_20.gif]  and search along the line through  [Graphics:Images/GradientSearchMod_gr_21.gif] in the direction  [Graphics:Images/GradientSearchMod_gr_22.gif].  You will arrive at a point  [Graphics:Images/GradientSearchMod_gr_23.gif],  where a local minimum occurs when the point [Graphics:Images/GradientSearchMod_gr_24.gif]  is constrained to lie on the line  [Graphics:Images/GradientSearchMod_gr_25.gif].  Since partial derivatives are accessible, the minimization process can be executed using either the quadratic or cubic approximation method.

    Next we compute  [Graphics:Images/GradientSearchMod_gr_26.gif]  and move in the search direction  [Graphics:Images/GradientSearchMod_gr_27.gif].  You will come to  [Graphics:Images/GradientSearchMod_gr_28.gif],  where a local minimum occurs when [Graphics:Images/GradientSearchMod_gr_29.gif]  is constrained to lie on the line  [Graphics:Images/GradientSearchMod_gr_30.gif].  Iteration will produce a sequence,  [Graphics:Images/GradientSearchMod_gr_31.gif],  of points with the property

        
[Graphics:Images/GradientSearchMod_gr_32.gif]

If  
[Graphics:Images/GradientSearchMod_gr_33.gif]  then [Graphics:Images/GradientSearchMod_gr_34.gif]  will be a local minimum [Graphics:Images/GradientSearchMod_gr_35.gif].  

 

Outline of the Gradient Method

    Suppose that  
[Graphics:Images/GradientSearchMod_gr_36.gif]  has been obtained.  
    
(i)    Evaluate the gradient vector  [Graphics:Images/GradientSearchMod_gr_37.gif].  

(ii)    Compute the search direction  [Graphics:Images/GradientSearchMod_gr_38.gif].  

(iii)    Perform a single parameter minimization of  [Graphics:Images/GradientSearchMod_gr_39.gif]  on the interval  [Graphics:Images/GradientSearchMod_gr_40.gif],  where b is large.  
    This will produce a value  
[Graphics:Images/GradientSearchMod_gr_41.gif]  where a local minimum for  [Graphics:Images/GradientSearchMod_gr_42.gif].  The relation  [Graphics:Images/GradientSearchMod_gr_43.gif]  
    shows that this is a minimum for  
[Graphics:Images/GradientSearchMod_gr_44.gif]  along the search line  [Graphics:Images/GradientSearchMod_gr_45.gif].
                            
(iv)    Construct the next point  [Graphics:Images/GradientSearchMod_gr_46.gif].

(v)    Perform the termination test for minimization, i.e.
    are the function values  
[Graphics:Images/GradientSearchMod_gr_47.gif]  and  [Graphics:Images/GradientSearchMod_gr_48.gif]  sufficiently close and the distance[Graphics:Images/GradientSearchMod_gr_49.gif]small enough ?

Repeat the process.

Proof  Gradient Search  Gradient Search  

 

Algorithm (Steepest Descent  or Gradient Method).  To numerically approximate a local minimum of  [Graphics:Images/GradientSearchMod_gr_50.gif],  where  f  is a continuous function of  n  real variables and  [Graphics:Images/GradientSearchMod_gr_51.gif]  by starting with one point  [Graphics:Images/GradientSearchMod_gr_52.gif]  and using the gradient method.

Computer Programs  Gradient Search  Gradient Search  

Program (Steepest Descent  or Gradient Method).  To numerically approximate a local minimum of  [Graphics:Images/GradientSearchMod_gr_53.gif],  where  f  is a continuous function of  n  real variables and  [Graphics:Images/GradientSearchMod_gr_54.gif]  by starting with one point  [Graphics:Images/GradientSearchMod_gr_55.gif]  and using the gradient method.  If partial derivatives are accessible, the minimization process can be executed using either the quadratic or cubic approximation method.  For illustration purposes we use the quadratic approximation method for finding a minimum.  

[Graphics:Images/GradientSearchMod_gr_56.gif]

Example 1.  Use the gradient search method to find the minimum of   [Graphics:Images/GradientSearchMod_gr_57.gif].  
Solution 1.

 

Example 2.  Use the gradient method to find the minimum of      [Graphics:Images/GradientSearchMod_gr_83.gif]
Rosenbrock's parabolic valley, circa 1960.
Solution 2.

 

Exercise 3.  Use the gradient method to find the minimum of  [Graphics:Images/GradientSearchMod_gr_102.gif].  
Looking at your graphs, estimate the location of the local minima.  
Solution 3.

 

Example 4.  Use the gradient search method to find the minimum of   [Graphics:Images/GradientSearchMod_gr_135.gif].  
Solution 4.

 

Various Scenarios and Animations for the Gradient Search Method.

Program (Steepest Descent  or Gradient Method).  To numerically approximate a local minimum of  f(X),  where  f  is a continuous function of  n  real variables and  [Graphics:Images/GradientSearchMod_gr_164.gif]  by starting with one point  [Graphics:Images/GradientSearchMod_gr_165.gif]  and using the gradient method.  Including graphics commands to draw the iterations used in finding the solution.

[Graphics:Images/GradientSearchMod_gr_166.gif]

Example 5.  Use the gradient search method to find the minimum of   [Graphics:Images/GradientSearchMod_gr_167.gif].  
Solution 5.

 

Animations (Gradient Search  Gradient Search).  Internet hyperlinks to animations.  

 

Research Experience for Undergraduates

Gradient Search  Gradient Search  Internet hyperlinks to web sites and a bibliography of articles.  

 

Download this Mathematica Notebook Gradient Search

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(c) John H. Mathews 2004