Path: csiph.com!x330-a1.tempe.blueboxinc.net!newsfeed.hal-mli.net!feeder3.hal-mli.net!newsfeed.hal-mli.net!feeder1.hal-mli.net!news.stack.nl!newsfeed.xs4all.nl!newsfeed6.news.xs4all.nl!xs4all!post.news.xs4all.nl!not-for-mail
From: inq1ltd <inq1ltd@inqvista.com>
To: python-list@python.org
Subject: Re: Numeric root-finding in Python
Date: Sun, 12 Feb 2012 10:20:17 -0500
User-Agent: KMail/4.7.2 (Linux/2.6.37.6-0.11-desktop; KDE/4.7.2; i686; ; )
In-Reply-To: <4f375f0f$0$29986$c3e8da3$5496439d@news.astraweb.com>
References: <4f375f0f$0$29986$c3e8da3$5496439d@news.astraweb.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 7Bit
Content-Type: text/plain; charset="us-ascii"
Sender: inq1ltd <inq1ltd@inqvista.com>
Cc: Steven D'Aprano <steve+comp.lang.python@pearwood.info>
Precedence: list
Newsgroups: comp.lang.python
Message-ID: <mailman.5726.1329062119.27778.python-list@python.org>
Lines: 92
NNTP-Posting-Host: 2001:888:2000:d::a6
Xref: x330-a1.tempe.blueboxinc.net comp.lang.python:20275



I don't know the first thing about this math problem however,

if I were to code this I might try ;

   except ZeroDivisionError:
         assert w = -1

rather than;

   except ZeroDivisionError:
         assert w == -1


jimonlinux



On Sunday, February 12, 2012 06:41:20 AM Steven D'Aprano wrote:
> This is only peripherally a Python problem, but in case anyone has any
> good ideas I'm going to ask it.
> 
> I have a routine to calculate an approximation of Lambert's W function,
> and then apply a root-finding technique to improve the approximation.
> This mostly works well, but sometimes the root-finder gets stuck in a
> cycle.
> 
> Here's my function:
> 
> import math
> def improve(x, w, exp=math.exp):
>     """Use Halley's method to improve an estimate of W(x) given
>     an initial estimate w.
>     """
>     try:
>         for i in range(36):  # Max number of iterations.
>             ew = exp(w)
>             a = w*ew - x
>             b = ew*(w + 1)
>             err = -a/b  # Estimate of the error in the current w.
>             if abs(err) <= 1e-16:
>                 break
>             print '%d: w= %r err= %r' % (i, w, err)
>             # Make a better estimate.
>             c = (w + 2)*a/(2*w + 2)
>             delta = a/(b - c)
>             w -= delta
>         else:
>             raise RuntimeError('calculation failed to converge', err)
>     except ZeroDivisionError:
>         assert w == -1
>     return w
> 
> 
> Here's an example where improve() converges very quickly:
> 
> py> improve(-0.36, -1.222769842388856)
> 0: w= -1.222769842388856 err= -2.9158979924038895e-07
> 1: w= -1.2227701339785069 err= 8.4638038491998997e-16
> -1.222770133978506
> 
> That's what I expect: convergence in only a few iterations.
> 
> Here's an example where it gets stuck in a cycle, bouncing back and forth
> between two values:
> 
> py> improve(-0.36787344117144249, -1.0057222396915309)
> 0: w= -1.0057222396915309 err= 2.6521238905750239e-14
> 1: w= -1.0057222396915044 err= -2.6521238905872001e-14
> 2: w= -1.0057222396915309 err= 2.6521238905750239e-14
> 3: w= -1.0057222396915044 err= -2.6521238905872001e-14
> 4: w= -1.0057222396915309 err= 2.6521238905750239e-14
> 5: w= -1.0057222396915044 err= -2.6521238905872001e-14
> [...]
> 32: w= -1.0057222396915309 err= 2.6521238905750239e-14
> 33: w= -1.0057222396915044 err= -2.6521238905872001e-14
> 34: w= -1.0057222396915309 err= 2.6521238905750239e-14
> 35: w= -1.0057222396915044 err= -2.6521238905872001e-14
> Traceback (most recent call last):
>   File "<stdin>", line 1, in <module>
>   File "<stdin>", line 19, in improve
> RuntimeError: ('calculation failed to converge', -2.6521238905872001e-14)
> 
> (The correct value for w is approximately -1.00572223991.)
> 
> I know that Newton's method is subject to cycles, but I haven't found any
> discussion about Halley's method and cycles, nor do I know what the best
> approach for breaking them would be. None of the papers on calculating
> the Lambert W function that I have found mentions this.
> 
> Does anyone have any advice for solving this?