• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1[section:lanczos The Lanczos Approximation]
2
3[h4 Motivation]
4
5['Why base gamma and gamma-like functions on the Lanczos approximation?]
6
7First of all I should make clear that for the gamma function
8over real numbers (as opposed to complex ones)
9the Lanczos approximation (See [@http://en.wikipedia.org/wiki/Lanczos_approximation Wikipedia or ]
10[@http://mathworld.wolfram.com/LanczosApproximation.html Mathworld])
11appears to offer no clear advantage over more traditional methods such as
12[@http://en.wikipedia.org/wiki/Stirling_approximation Stirling's approximation].
13__pugh carried out an extensive comparison of the various methods available
14and discovered that they were all very similar in terms of complexity
15and relative error.  However, the Lanczos approximation does have a couple of
16properties that make it worthy of further consideration:
17
18* The approximation has an easy to compute truncation error that holds for
19all /z > 0/.  In practice that means we can use the same approximation for all
20/z > 0/, and be certain that no matter how large or small /z/ is, the truncation
21error will /at worst/ be bounded by some finite value.
22* The approximation has a form that is particularly amenable to analytic
23manipulation, in particular ratios of gamma or gamma-like functions
24are particularly easy to compute without resorting to logarithms.
25
26It is the combination of these two properties that make the approximation
27attractive: Stirling's approximation is highly accurate for large z, and
28has some of the same analytic properties as the Lanczos approximation, but
29can't easily be used across the whole range of z.
30
31As the simplest example, consider the ratio of two gamma functions: one could
32compute the result via lgamma:
33
34   exp(lgamma(a) - lgamma(b));
35
36However, even if lgamma is uniformly accurate to 0.5ulp, the worst case
37relative error in the above can easily be shown to be:
38
39   Erel > a * log(a)/2 + b * log(b)/2
40
41For small /a/ and /b/ that's not a problem, but to put the relationship another
42way: ['each time a and b increase in magnitude by a factor of 10, at least one
43decimal digit of precision will be lost.]
44
45In contrast, by analytically combining like power
46terms in a ratio of Lanczos approximation's, these errors can be virtually eliminated
47for small /a/ and /b/, and kept under control for very large (or very small
48for that matter) /a/ and /b/.  Of course, computing large powers is itself a
49notoriously hard problem, but even so, analytic combinations of Lanczos
50approximations can make the difference between obtaining a valid result, or
51simply garbage.  Refer to the implementation notes for the __beta function for
52an example of this method in practice.  The incomplete
53[link math_toolkit.sf_gamma.igamma gamma_p gamma] and
54[link math_toolkit.sf_beta.ibeta_function beta] functions
55use similar analytic combinations of power terms, to combine gamma and beta
56functions divided by large powers into single (simpler) expressions.
57
58[h4 The Approximation]
59
60The Lanczos Approximation to the Gamma Function is given by:
61
62[equation lanczos0]
63
64Where S[sub g](z) is an infinite sum, that is convergent for all z > 0,
65and /g/ is an arbitrary parameter that controls the "shape" of the
66terms in the sum which is given by:
67
68[equation lanczos0a]
69
70With individual coefficients defined in closed form by:
71
72[equation lanczos0b]
73
74However, evaluation of the sum in that form can lead to numerical instability
75in the computation of the ratios of rising and falling factorials (effectively
76we're multiplying by a series of numbers very close to 1, so roundoff errors
77can accumulate quite rapidly).
78
79The Lanczos approximation is therefore often written in partial fraction form
80with the leading constants absorbed by the coefficients in the sum:
81
82[equation lanczos1]
83
84where:
85
86[equation lanczos2]
87
88Again parameter /g/ is an arbitrarily chosen constant, and /N/ is an arbitrarily chosen
89number of terms to evaluate in the "Lanczos sum" part.
90
91[note
92Some authors
93choose to define the sum from k=1 to N, and hence end up with N+1 coefficients.
94This happens to confuse both the following discussion and the code (since C++
95deals with half open array ranges, rather than the closed range of the sum).
96This convention is consistent with __godfrey, but not __pugh, so take care
97when referring to the literature in this field.]
98
99[h4 Computing the Coefficients]
100
101The coefficients C0..CN-1 need to be computed from /N/ and /g/
102at high precision, and then stored as part of the program.
103Calculation of the coefficients is performed via the method of __godfrey;
104let the constants be contained in a column vector P, then:
105
106P = D B C F
107
108where B is an NxN matrix:
109
110[equation lanczos4]
111
112D is an NxN matrix:
113
114[equation lanczos3]
115
116C is an NxN matrix:
117
118[equation lanczos5]
119
120and F is an N element column vector:
121
122[equation lanczos6]
123
124Note than the matrices B, D and C contain all integer terms and depend
125only on /N/, this product should be computed first, and then multiplied
126by /F/ as the last step.
127
128[h4 Choosing the Right Parameters]
129
130The trick is to choose
131/N/ and /g/ to give the desired level of accuracy: choosing a small value for
132/g/ leads to a strictly convergent series, but one which converges only slowly.
133Choosing a larger value of /g/ causes the terms in the series to be large
134and\/or divergent for about the first /g-1/ terms, and to then suddenly converge
135with a "crunch".
136
137__pugh has determined the optimal
138value of /g/ for /N/ in the range /1 <= N <= 60/: unfortunately in practice choosing
139these values leads to cancellation errors in the Lanczos sum as the largest
140term in the (alternating) series is approximately 1000 times larger than the result.
141These optimal values appear not to be useful in practice unless the evaluation
142can be done with a number of guard digits /and/ the coefficients are stored
143at higher precision than that desired in the result.  These values are best
144reserved for say, computing to float precision with double precision arithmetic.
145
146[table Optimal choices for N and g when computing with guard digits (source: Pugh)
147[[Significand Size] [N] [g][Max Error]]
148[[24] [6] [5.581][9.51e-12]]
149[[53][13][13.144565][9.2213e-23]]
150]
151
152The alternative described by __godfrey is to perform an exhaustive
153search of the /N/ and /g/ parameter space to determine the optimal combination for
154a given /p/ digit floating-point type.  Repeating this work found a good
155approximation for double precision arithmetic (close to the one __godfrey found),
156but failed to find really
157good approximations for 80 or 128-bit long doubles.  Further it was observed
158that the approximations obtained tended to optimised for the small values
159of z (1 < z < 200) used to test the implementation against the factorials.
160Computing ratios of gamma functions with large arguments were observed to
161suffer from error resulting from the truncation of the Lancozos series.
162
163__pugh identified all the locations where the theoretical error of the
164approximation were at a minimum, but unfortunately has published only the largest
165of these minima.  However, he makes the observation that the minima
166coincide closely with the location where the first neglected term (a[sub N]) in the
167Lanczos series S[sub g](z) changes sign.  These locations are quite easy to
168locate, albeit with considerable computer time.  These "sweet spots" need
169only be computed once, tabulated, and then searched when required for an
170approximation that delivers the required precision for some fixed precision
171type.
172
173Unfortunately, following this path failed to find a really good approximation
174for 128-bit long doubles, and those found for 64 and 80-bit reals required an
175excessive number of terms.  There are two competing issues here: high precision
176requires a large value of /g/, but avoiding cancellation errors in the evaluation
177requires a small /g/.
178
179At this point note that the Lanczos sum can be converted into rational form
180(a ratio of two polynomials, obtained from the partial-fraction form using
181polynomial arithmetic),
182and doing so changes the coefficients so that /they are all positive/.  That
183means that the sum in rational form can be evaluated without cancellation
184error, albeit with double the number of coefficients for a given N.  Repeating
185the search of the "sweet spots", this time evaluating the Lanczos sum in
186rational form, and testing only those "sweet spots" whose theoretical error
187is less than the machine epsilon for the type being tested, yielded good
188approximations for all the types tested.  The optimal values found were quite
189close to the best cases reported by __pugh (just slightly larger /N/ and slightly
190smaller /g/ for a given precision than __pugh reports), and even though converting
191to rational form doubles the number of stored coefficients, it should be
192noted that half of them are integers (and therefore require less storage space)
193and the approximations require a smaller /N/ than would otherwise be required,
194so fewer floating point operations may be required overall.
195
196The following table shows the optimal values for /N/ and /g/ when computing
197at fixed precision.  These should be taken as work in progress: there are no
198values for 106-bit significand machines (Darwin long doubles & NTL quad_float),
199and further optimisation of the values of /g/ may be possible.
200Errors given in the table
201are estimates of the error due to truncation of the Lanczos infinite series
202to /N/ terms.  They are calculated from the sum of the first five neglected
203terms - and are known to be rather pessimistic estimates - although it is noticeable
204that the best combinations of /N/ and /g/ occurred when the estimated truncation error
205almost exactly matches the machine epsilon for the type in question.
206
207[table Optimum value for N and g when computing at fixed precision
208[[Significand Size][Platform/Compiler Used][N][g][Max Truncation Error]]
209[[24][Win32, VC++ 7.1] [6] [1.428456135094165802001953125][9.41e-007]]
210[[53][Win32, VC++ 7.1] [13] [6.024680040776729583740234375][3.23e-016]]
211[[64][Suse Linux 9 IA64, gcc-3.3.3] [17] [12.2252227365970611572265625][2.34e-024]]
212[[116][HP Tru64 Unix 5.1B \/ Alpha, Compaq C++ V7.1-006] [24] [20.3209821879863739013671875][4.75e-035]]
213]
214
215Finally note that the Lanczos approximation can be written as follows
216by removing a factor of exp(g) from the denominator, and then dividing
217all the coefficients by exp(g):
218
219[equation lanczos7]
220
221This form is more convenient for calculating lgamma, but for the gamma
222function the division by /e/ turns a possibly exact quality into an
223inexact value: this reduces accuracy in the common case that
224the input is exact, and so isn't used for the gamma function.
225
226[h4 References]
227
228# [#godfrey]Paul Godfrey, [@http://my.fit.edu/~gabdo/gamma.txt "A note on the computation of the convergent
229Lanczos complex Gamma approximation"].
230# [#pugh]Glendon Ralph Pugh,
231[@http://bh0.physics.ubc.ca/People/matt/Doc/ThesesOthers/Phd/pugh.pdf
232"An Analysis of the Lanczos Gamma Approximation"],
233PhD Thesis November 2004.
234# Viktor T. Toth,
235[@http://www.rskey.org/gamma.htm "Calculators and the Gamma Function"].
236# Mathworld, [@http://mathworld.wolfram.com/LanczosApproximation.html
237The Lanczos Approximation].
238
239[endsect][/section:lanczos The Lanczos Approximation]
240
241[/
242  Copyright 2006 John Maddock and Paul A. Bristow.
243  Distributed under the Boost Software License, Version 1.0.
244  (See accompanying file LICENSE_1_0.txt or copy at
245  http://www.boost.org/LICENSE_1_0.txt).
246]
247