34 Georgia Tech ECE 6250 Fall 2019 Notes by J Romberg and M Davenport Last

# 34 georgia tech ece 6250 fall 2019 notes by j romberg

• Notes
• 16

This preview shows page 4 - 9 out of 16 pages.

34 Georgia Tech ECE 6250 Fall 2019; Notes by J. Romberg and M. Davenport. Last updated 2:06, November 18, 2019 Subscribe to view the full document.

By the chain rule, d d α f ( x k +1 ) = f ( x k +1 ) T d d α x k +1 = f ( x k +1 ) T r k . So we need to choose α k such that f ( x k +1 ) r k , or more concisely r k +1 r k ( r T k +1 r k = 0) . So let’s do this r T k +1 r k = 0 ( b - Hx k +1 ) T r k = 0 ( b - H ( x k + α k r k )) T r k = 0 ( b - Hx k ) T r k - α k r T k Hr k = 0 r T k r k - α k r T k Hr k = 0 and so the optimal step size is α k = r T k r k r T k Hr k . The steepest descent algorithm performs this iteration until k Hx k - b k 2 is below some tolerance δ : 35 Georgia Tech ECE 6250 Fall 2019; Notes by J. Romberg and M. Davenport. Last updated 2:06, November 18, 2019 Steepest Descent, version 1 Initialize: x 0 = some guess, k = 0, r 0 = b - Hx 0 . while not converged, k r k k 2 δ do α k = r T k r k / r T k Hr k x k +1 = x k + α k r k r k +1 = b - Hx k +1 k = k + 1 end while There is a nice trick that can save us one of two applications of H needed in each iteration above. Notice that r k +1 = b - Hx k +1 = b - H ( x k + α k r k ) = r k - α k Hr k . So we can save an application of H by updating the residual rather than recomputing it at each stage. Steepest Descent, more efficient version 2 Initialize: x 0 = some guess, k = 0, r 0 = b - Hx 0 . while not converged, k r k k 2 δ do q = Hr k α k = r T k r k / r T k q x k +1 = x k + α k r k r k +1 = r k - α k q k = k + 1 end while 36 Georgia Tech ECE 6250 Fall 2019; Notes by J. Romberg and M. Davenport. Last updated 2:06, November 18, 2019 Subscribe to view the full document.

The effectiveness of SD depends critically on how H is conditioned and the starting point. Consider the two examples on the next page. -4 -2 2 4 -4 -2 2 4 6 -4 -2 2 4 -4 -2 2 4 6 -4 -2 2 4 -4 -2 2 4 6 -4 -2 2 4 -4 -2 2 4 6 1 (c) 2 1 (d) 2 1 (a) 2 1 (b) 2 (from Shewchuk, “... without the agonizing pain”) When the conditioning of H is poor and we choose a bad starting point, convergence can take many iterations even in simple cases. 37 Georgia Tech ECE 6250 Fall 2019; Notes by J. Romberg and M. Davenport. Last updated 2:06, November 18, 2019 The method of conjugate gradients (CG) An excellent companion resource for this section is the manuscript: J. Shewchuk: “An introduction to the conjugate gradient method without the agonizing pain”. We can see from the example on the last page that steepest descent can be inefficient because it can move in essentially the same direction many times. CG avoids this by ensuring that each step is orthogonal (in an ap- propriate inner product) to all of the previous steps that have been taken. Miraculously, this can be done with very little overhead. Suppose for a moment that we pre-determine N step directions d 0 , . . . , d N - 1 that are orthogonal (but not necessarily normalized), d T j d i = 0 for i 6 = j . This means that { d k / k d k k 2 , k = 0 , . . . , N - 1 } is an orthobasis for R N . Then given a starting point x 0 , the initial error e 0 = x 0 - b x , where b x is the solution that satisfies H b x = b , can be expanded as e 0 = N - 1 X =0 c d k d k 2 , where c = d T e 0 k d k 2 . (3) Given step sizes α 0 , α 1 , . . . , the error after the k th step is e k = x k - b x = x k - 1 + α k - 1 d k - 1 - b x = e k - 1 + α k - 1 d k - 1 = e k - 2 + α k - 1 d k - 1 + α k - 2 d k - 2 . Subscribe to view the full document. • Fall '08
• Staff

### What students are saying

• As a current student on this bumpy collegiate pathway, I stumbled upon Course Hero, where I can find study resources for nearly all my courses, get online help from tutors 24/7, and even share my old projects, papers, and lecture notes with other students.

Kiran Temple University Fox School of Business ‘17, Course Hero Intern

• I cannot even describe how much Course Hero helped me this summer. It’s truly become something I can always rely on and help me. In the end, I was not only able to survive summer classes, but I was able to thrive thanks to Course Hero.

Dana University of Pennsylvania ‘17, Course Hero Intern

• The ability to access any university’s resources through Course Hero proved invaluable in my case. I was behind on Tulane coursework and actually used UCLA’s materials to help me move forward and get everything together on time.

Jill Tulane University ‘16, Course Hero Intern