1. 1 OR II GSLM 52800

2. 2 Outline  some terminology  differences between LP and NLP  basic questions in NLP  gradient and Hessian  quadratic form  contour, graph, and tangent plane

3. 3 Feasible Points, Solution Set, and Neighborhood  feasible point: a point that satisfies all the constraints  solution set (feasible set, feasible region): the collection of all feasible points  neighborhood of x 0 = {x| |x  x 0 | <  } for some pre-specified  feasible region the neighborhood of a point for a given  A C B D  only the neighborhood of D is completely feasible for this 

4. 4 Weak and Strong; Local and Global  local minima: x 1, any point in [s, t], x 3  strict (strong) local minima: x 1, x 3  weak local minima: any point in [s, t]  strict global minimum: x 1  weak local maxima: any point in [s, t] x f(x ) 12x3x3 x2x2 x1x1 t s

5. 5 Differences Between Linear and Non-Linear Programming  linear programming  there exists an optimal extreme point (a corner point)  direction of improvement keeps on being so unless hitting a constraint  a local optimum point is also globally optimal direction of improvement optimal point

6. 6 Differences Between Linear and Non-Linear Programming  none of these necessarily holds for a non- linear program x f(x ) 12x3x3 x2x2 x1x1 t s min x 2 + y 2, s.t. -2  x, y  2

7. 7 Basic Questions in Non-Linear Programming  main question: given an initial location x 0, how to get to a local minimum, or, better, a global minimum  (a) the direction of improvement?  (b) the necessary conditions of an optimal point?  (c) the sufficient conditions of an optimal point?  (d) any conditions to simplify the processes in (a), (b), and (c)?  (e) any algorithmic procedures to solve a NLP problem?

8. 8 Basic Questions in Non-Linear Programming  calculus required for (a) to (e)  direction of improvement of f = gradient of f  shaped by constraints  convexity for (d), and also (b) and (c)  identification of convexity: definiteness of matrices, especially for Hessians

9. 9 Gradient and Hessian  gradient of f:  f(x) =  in short  Hessian = f and g j usually assumed to be twice differentiable functions  Hessian is a symmetric matrix

10. 10 Gradient and Hessian  e j : (0, …, 0, 1, 0, …, 0) T, where “1” at the jth position   for small , f(x+  e j )  f(x) +   in general,  x = (  x 1, …,  x n ) T from x, f(x+  x)  f(x) +

11. 11 Example 1.6.1  (a). f(x) = x 2 ; f(3.5+  )  ? for small   (b). f(x, y) = x 2 + y 2, f((1, 1) + (  x,  y))  ? for small  x,  y  gradient  f : direction of steepest accent of the objective fucntion

12. 12 Example 1.6.2  find the Hessian of  (a). f(x, y) = x 2 + 7y 2  (b). f(x, y) = x 2 + 5xy + 7y 2  (c). f(x, y) = x 3 + 7y 2

13. 13 Quadratic Form  general form: x T Qx/2 + c T x + a, where x is an n-dimensional vector; Q an n  n square matrix; c and a are matrices of appropriate dimensions  how to derive the gradient and Hessian?  gradient  f(x) = Qx+c  Hessian H = Q

14. 14 Quadratic Form  relate the two forms x T Qx/2 + c T x + a and f(x, y) =  1 x 2 +  2 xy+  3 y 2 +  4 x+  5 y+  6  Example 1.6.3

15. 15 Example 1.6.4  Find the first two derivatives of the following f(x)  f(x) = x 2 for x  [-2, 2]  f(x) = -x 2 for x  [-2, 2]

16. 16 Contour and Graph (i.e., Surface) of Function f  Example 1.7.1: f(x 1, x 2 ) =

17. 17 Contour and Graph (i.e., Surface) of Function f  an n-dimensional function  a contour of f: a diagram f(x) = c in the n- dimensional space for a given value c  the graph (surface function) of f: the diagram z = f(x) in the (n+1)st dimensional space as x and z vary

18. 18 Contour and Graph (i.e., Surface) of Function f  how do the contours of the one-dimensional function f(x) = x 2 look like?

19. 19 An Important Property Between the Gradient and the Tangent Plane at a Contour  the gradient of f at point x 0 is orthogonal to the tangent of the contour f(x) = c at x 0  many optimization results are related to the above property

20. 20 Gradient of f at x 0 Being Orthogonal to the Tangent of the Contour f(x) = c at x 0  Example 1.7.3: f(x 1, x 2 ) = x 1 +2x 2  gradient at (4, 2)?  tangent of contour at (4, 2)?

21. 21 Gradient of f at x 0 Being Orthogonal to the Tangent of the Contour f(x) = c at x 0  Example 1.7.2: f(x 1, x 2 ) =  point (x 10, x 20 ) on a contour f(x 1, x 2 ) = c

22. 22 Tangent at a Contour and the Corresponding Tangent Plane at a Surface  the above two are related  for contour of f(x, y) = x 2 +y 2, the tangent at (x 0, y 0 )  (x-x 0, y- y 0 ) T (2x 0, 2y 0 ) = 0 two orthogonal vectors u and v: u T v = 0

23. 23 Tangent at a Contour and the Corresponding Tangent Plane at a Surface  the tangent place at (x 0, y 0 ) for the surface of f(x, y) = x 2 +y 2  the surface: z = x 2 +y 2  defining a contour at a higher dimension: F(x, y, z) = x 2 +y 2  z  tangent plane at (x 0, y 0, ) of the surface: what happens when z =