Content deleted Content added
cite Numerical Recipes |
m →Algorithm: fix "big K" notation |
||
| (8 intermediate revisions by 4 users not shown) | |||
Line 1:
In mathematics, '''Lentz's algorithm''' is an [[algorithm]] to evaluate [[generalized continued fraction|continued fraction]]s, and was originally devised to compute tables of spherical [[Bessel function]]s.<ref name=":0">{{Cite report|last=Lentz|first=W. J.|date=September 1973|title=A Method of Computing Spherical Bessel Functions of Complex Argument with Tables|url=https://apps.dtic.mil/sti/pdfs/AD0767223.pdf|publisher=Atmospheric Sciences Laboratory, US Army Electronics Command|___location=White Sands Missile Range, New Mexico|type=Research and Development Technical Report ECOM-5509 }}</ref><ref name="numerical-recipes-c++">{{cite book|title=Numerical Recipes in C++| pages=177–179|isbn= 0 521 75033 4}}</ref>
The version usually employed now is due to Thompson and Barnett.<ref name="
== History ==
The idea was introduced in 1973 by William J. Lentz<ref name=":0" /> and was simplified by him in 1982.<ref>{{Cite book|last=J.|first=Lentz, W.|url=http://worldcat.org/oclc/227549426|title=A Simplification of Lentz's Algorithm.|date=August 1982|publisher=Defense Technical Information Center|oclc=227549426}}</ref> Lentz suggested that calculating ratios of spherical Bessel functions of complex arguments can be difficult. He developed a new continued fraction technique for calculating the ratios of spherical Bessel functions of consecutive order. This method was an improvement compared to other methods because it started from the beginning of the continued fraction rather than the tail, had a built-in check for convergence, and was numerically stable. The original algorithm uses algebra to bypass a zero in either the numerator or denominator.<ref name="Lentz 668–671">{{Cite journal|last=Lentz|first=William J.|date=1976-03-01|title=Generating Bessel functions in Mie scattering calculations using continued fractions|url=http://dx.doi.org/10.1364/ao.15.000668|journal=Applied Optics|volume=15|issue=3|pages=668–671|doi=10.1364/ao.15.000668|pmid=20165036 |bibcode=1976ApOpt..15..668L |issn=0003-6935|url-access=subscription}}</ref> Simpler Improvements to overcome unwanted zero terms include an altered recurrence relation<ref>{{Cite journal|last1=Jaaskelainen|first1=T.|last2=Ruuskanen|first2=J.|date=1981-10-01|title=Note on Lentz's algorithm|url=http://dx.doi.org/10.1364/ao.20.003289|journal=Applied Optics|volume=20|issue=19|pages=3289–3290|doi=10.1364/ao.20.003289|pmid=20333144 |bibcode=1981ApOpt..20.3289J |issn=0003-6935|url-access=subscription}}</ref> suggested by Jaaskelainen and Ruuskanen in 1981 or a simple shift of the denominator by a very small number as suggested by Thompson and Barnett in 1986.<ref name="
== Initial work ==
Line 22:
etc., or using the [[Generalized continued fraction#Notation|big-K notation]], if
:<math>{f}_{n} = {b}_{0} + \underset{j = 1}\overset{n}\operatorname{K}\frac{{a}_{j}}{{b}_{j}
is the <math>n</math>th convergent to <math>f</math> then
Line 30:
where <math>{A}_{n}</math> and <math>{B}_{n}</math> are given by the Wallis-Euler recurrence relations
:<
\begin{align}
{A}_{-1} & = 1 & {B}_{-1} & = 0\\
{A}_{0} & = {b}_{0} & {B}_{0} & = 1\\
\end{align}
▲:<math>{A}_{n} = {b}_{n} {A}_{n - 1} + {a}_{n} {A}_{n - 2}</math>
Lentz's method defines
Line 48 ⟶ 44:
:<math>{D}_{n} = \frac{{B}_{n - 1}}{{B}_{n}}</math>
so that the <math>n</math>th convergent is <math>{f}_{n} = {C}_{n} {D}_{n} {f}_{n - 1}</math> with <math>{f}_{0} = \frac{{A}_{0}}{{B}_{0}} = {b}_{0}</math> and uses the recurrence relations
:<math>
\begin{align}
\end{align}
▲:<math>{C}_{0} = \frac{{A}_{0}}{{A}_{- 1}} = {b}_{0}</math>
</math>
When the product <math>{C}_{n} {D}_{n}</math> approaches unity with increasing <math>n</math>, it is hoped that <math>{f}_{n}</math> has converged to <math>f</math>.<ref name="Numerical-Recipes">{{Cite book |last1=Press |first1=W.H. |title=Numerical Recipes: The Art of Scientific Computing |last2=Teukolsky |first2=S.A. |last3=Vetterling |first3=W.T. |last4=Flannery |first4=B. P. |publisher=Cambridge University Press |year=2007 |edition=3rd |pages=207–208}}</ref>▼
Lentz's algorithm has the advantage of side-stepping an inconvenience of the Wallis-Euler relations, namely that the numerators <math>A_n</math> and denominators <math>B_n</math> are prone to grow or diminish very rapidly with increasing <math>n</math>. In direct numerical application of
▲:<math>{D}_{n} = \frac{1}{{b}_{n} + {a}_{n} {D}_{n - 1}}</math>
== Thompson and Barnett modification ==
▲When the product <math>{C}_{n} {D}_{n}</math> approaches unity with increasing <math>n</math>, it is hoped that <math>{f}_{n}</math> has converged to <math>f</math>.<ref>{{Cite book |last1=Press |first1=W.H. |title=Numerical Recipes: The Art of Scientific Computing |last2=Teukolsky |first2=S.A. |last3=Vetterling |first3=W.T. |last4=Flannery |first4=B. P. |publisher=Cambridge University Press |year=2007 |edition=3rd |pages=207–208}}</ref>
In Lentz's original algorithm, it can happen that <math>{C}_{n} = 0</math>, resulting in division by zero at the next step. The problem can be remedied simply by setting <math>{C}_{n} = \varepsilon</math> for some sufficiently small <math>\varepsilon</math>. This gives <math>{C}_{n + 1} = {b}_{n + 1} + \frac{{a}_{n + 1}}{\varepsilon} = \frac{{a}_{n + 1}}{\varepsilon}</math> to within floating-point precision, and the product <math>{C}_{n} {C}_{n + 1} = {a}_{n + 1}</math> irrespective of the precise value of ε. Accordingly, the value of <math>{f}_{0} = {C}_{0} = {b}_{0}</math> is also set to <math>\varepsilon</math> in the case of <math>{b}_{0} = 0</math>.
Similarly, if the denominator in <math>{D}_{n} = \frac{1}{{b}_{n} + {a}_{n} {D}_{n - 1}}</math> is zero, then setting <math>{D}_{n} = \frac{1}{\varepsilon}</math> for small enough <math>\varepsilon</math> gives <math>{D}_{n} {D}_{n + 1} = \frac{1}{{a}_{n + 1}}</math> irrespective of the value of <math>\varepsilon</math>.<ref name="Thompson-and-Barnett" /><ref name="Numerical-Recipes" />
▲Lentz's algorithm has the advantage of side-stepping an inconvenience of the Wallis-Euler relations, namely that the numerators <math>A_n</math> and denominators <math>B_n</math> are prone to grow or diminish very rapidly with increasing <math>n</math>. In direct numerical application of the the Wallis-Euler relations, this means that <math>A_{n-1}</math>, <math>A_{n-2}</math>, <math>B_{n-1}</math>, <math>B_{n-2}</math> must be periodically checked and rescaled to avoid floating-point overflow or underflow<ref name="numerical-recipes" />.
== Applications ==
Lentz's algorithm was used widely in the late twentieth century. It was suggested that it doesn't have any rigorous analysis of error propagation. However, a few empirical tests suggest that it's at least as good as the other methods.<ref>{{Cite book |last1=Press |first1=W.H. |title=Numerical Recipes in Fortran, The Art of Scientific Computing|last2=Teukolsky |first2=S.A. |last3=Vetterling |first3=W.T. |last4=Flannery |first4=B. P. |publisher=Cambridge University Press |year=1992 |edition=2nd |page=165}}</ref> As an example, it was applied to evaluate exponential integral functions. This application was then called modified Lentz algorithm.<ref>{{Cite journal|last1=Press|first1=William H.|last2=Teukolsky|first2=Saul A.|date=1988|title=Evaluating Continued Fractions and Computing Exponential Integrals|journal=Computers in Physics|volume=2|issue=5|pages=88|doi=10.1063/1.4822777|bibcode=1988ComPh...2...88P |issn=0894-1866|doi-access=free}}</ref> It's also stated that the Lentz algorithm is not applicable for every calculation, and convergence can be quite rapid for some continued fractions and vice versa for others.<ref>{{Cite journal|last1=Wand|first1=Matt P.|last2=Ormerod|first2=John T.|date=2012-09-18|title=Continued fraction enhancement of Bayesian computing|url=http://dx.doi.org/10.1002/sta4.4|journal=Stat|volume=1|issue=1|pages=31–41|doi=10.1002/sta4.4|pmid=22533111 |s2cid=119636237 |issn=2049-1573|url-access=subscription}}</ref>
==References==
| |||