MAT360 Summary: Convergence of the Secant Method

Bob Hastings came by the office one day, after we'd discussed the convergence of the secant method, and asked me how we know that, when n is large,

displaymath333

as our book claimed. We sat down, and worked out the details, but he wanted to see it done more rigorously. So I wrote the following. I hope it helps!

Newton's method can be represented as

  equation327

where m is the slope of the tangent line at the point tex2html_wrap_inline393 . If we don't want to calculate derivatives, we can always approximate m as

displaymath334

a discrete approximation to the derivative. This is the secant method, which requires two approximations to start ( tex2html_wrap_inline397 and tex2html_wrap_inline399 ), and doesn't converge quadratically (although it does converge super-linearly, with exponent of the golden ratio tex2html_wrap_inline401 ). Here's how:

First, we need to show that

displaymath335

We'll use Taylor series expansions, and order arguments.

Starting from (1), we add and subtract a p to get

displaymath336

Having popped out the tex2html_wrap_inline405 term, it comes down to showing that

displaymath337

when n is large. Taylor says that

displaymath338

since f(p)=0, we can reexpress A as

displaymath339

having cancelled a common term of tex2html_wrap_inline405 in numerator and denominator.

Now Taylor chimes in again, with

displaymath340

so that

displaymath341

So we replace m by its value in A to obtain the following:

displaymath342

(notice that I don't need but one power of tex2html_wrap_inline419 in the denominator - don't carry more than you need!).

Now we expand both f' and f'' in Taylor series about p, and throw all the ``order stuff'' into one term while keeping all the good stuff in the first term, to get

displaymath343

displaymath344

The first term on the right-hand side above simplifies to

displaymath345

as one can check, whereas the second term goes to zero quadratically in terms like tex2html_wrap_inline427 , tex2html_wrap_inline405 , or tex2html_wrap_inline419 . Hence,

displaymath346

as tex2html_wrap_inline433 (and tex2html_wrap_inline435 ). Thus as our book claimed, when n is large we have that

displaymath335

where

displaymath348

Now, assuming that

displaymath349

then

displaymath350

Thus

displaymath351

and, by continuity,

displaymath352

But this is the same convergence criterion, meaning that secant converges of order tex2html_wrap_inline439 , and also of order tex2html_wrap_inline441 : hence

displaymath353

or

displaymath354

with solutions

displaymath355

and since tex2html_wrap_inline443 ,

displaymath356

Furthermore,

displaymath357

so

displaymath358

and

displaymath359

or

displaymath360

QED.



Tue Oct 11 23:05:15 EDT 2005