on Mon Jul 02 2007, "John Maddock" <john-AT-johnmaddock.co.uk> wrote:

> The main index page you refer to there scrolls so slowly on my
> system as to be unusable - the sample page
> http://randysimons.com/overige/multicolumn/ is much better in this
> respect, but the amount of scrolling up and down to read the columns
> is IMO intolerable.  How would this cope with a page like this:
> http://freespace.virgin.net/boost.regex/toolkit/html/math_toolkit/backgrounders/remez.html
> that which as well as being much longer than the average reference
> page has more than it's fair share of media objects?

One would columnize entire sections, or tuples of sections, before
moving on to a separate area.  So for example (your page uses a
subheading that matches the main heading, which is a bit weird. I also
used elipses liberally because you'd be able to fit lots more
horizontally in a column than I can here with a fixed-width font.

+------------------------------+------------------------------+------------------------------+
|** The Remez Method **        | We want to find the "best"   | In the 
following discussion  |
|                              |rational approximation, where |we'll use a 
concrete example  |
|The Remez algorithm is a      |"best" is defined to be the   |to illustrate 
the Remez       |
|methodology for locating the  |approximation that has the    |method: an 
approximation to   |
|minimax rational approximation|least deviation from f(x). We |the function ex 
over the range|
|to a function. This short     |can measure the deviation by  |[-1, 1].         
             |
|article gives a brief overview|way of an error function:     |                 
             |
|of the method, but it should  |                              |Before we can 
begin the Remez |
|not be regarded as a thorough |Eabs(x) = f(x) - R(x)         |method, we must 
obtain an     |
|theoretical treatment, for    |                              |initial value 
for the location|
|that you should consult your  |which is expressed in terms of|of the extrema 
of the error   |
|favorite textbook.            |absolute error, but we can    |function. We 
could "guess"    |
|                              |equally use relative error:   |these, but a 
much closer first|
|Imagine that you want to      |...                           |approximation 
can be obtained |
|approximate some function f(x)|                              |by first 
constructing an      |
|by way of a rational function | Unfortunately we don't know  |interpolated 
polynomial       |
|R(x), where R(x) may be either|where the extrema of the error|approximation to 
f(x).        |
|a polynomial P(x) or a ratio  |function are located!         |                 
             |
|of two polynomials P(x)/Q(x)  |                              |In order to 
obtain the N+1    |
|(a rational                   | ** The Remez Method **       |coefficients of 
the           |
|function). Initially we'll    |                              |interpolated 
polynomial we    |
|concentrate on the polynomial |The Remez method is an        |need N+1 points 
(x0...xN):    |
|case, as it's by far the      |iterative technique which,    |with our 
interpolated form    |
|easier to deal with, later    |given a broad range of        |passing through 
each of those |
|we'll extend to the full      |assumptions, will converge on |points that 
yields N+1        |
|rational function case.       |the extrema of the error      |simultaneous 
equations:       |
|                              |function, and therefore the   |                 
             |
|                              |minimax solution.             |                 
             |
+------------------------------+------------------------------+------------------------------+
| f(xi) = P(xi) = c0+ c1xi... +|                                                
             |
|cNxiN                         |                                                
             |
|                              |                                                
             |
|Which can be solved for the   |                                                
             |
|coefficients c0...cNin P(x).  |                                                
             |
|                              |                                                
             |
|Obviously this is not a       |                                                
             |
|minimax solution, indeed our  |                   Initial Interpolated 
Approximation        |
|only guarantee is that f(x)   |                                                
             |
|and P(x) touch at N+1         |                                                
             |
|locations, away from those    |                                                
             |
|points the error may be       |                                                
             |
|arbitrarily large. However, we|                                                
             |
|would clearly like this       |                                                
             |
|initial approximation to be as|                                                
             |
|close to f(x) as possible, 
and+------------------------------+------------------------------+
|it turns out that using the   | Which has a peak relative    | ** Remez Step 1 
***          |
|zeros of an orthogonal        |error of 1.2x10-3.            | ...             
             |
|polynomial as the initial     |                              |                 
             |
|interpolation points is a good|While this is a pretty good   |                 
             |
|choice. In our example we'll  |approximation already, judging|                 
             |
|use the zeros of a Chebyshev  |by the shape of the error     |                 
             |
|polynomial as these are       |function we can clearly do    |                 
             |
|particularly easy to          |better. Before starting on the|                 
             |
|calculate, interpolating for a|Remez method propper, we have |                 
             |
|polynomial of degree 4, and   |one more step to perform:     |                 
             |
|measuring relative error we   |locate all the extrema of the |                 
             |
|get the following error       |error function, and store     |                 
             |
|function:                     |these locations as our initial|                 
             |
|                              |Chebyshev control points.     |                 
             |
|                              | ...                          |                 
             |
+------------------------------+------------------------------+------------------------------+

-- 
Dave Abrahams
Boost Consulting
http://www.boost-consulting.com

The Astoria Seminar ==> http://www.astoriaseminar.com


-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/
_______________________________________________
Boost-docs mailing list
[email protected]
Unsubscribe and other administrative requests: 
https://lists.sourceforge.net/lists/listinfo/boost-docs

Reply via email to