fselector 0.1.0
Sign up to get free protection for your applications and to get access to all the features.
- data/LICENSE +21 -0
- data/README.md +195 -0
- data/lib/fselector.rb +41 -0
- data/lib/fselector/algo_continuous/PMetric.rb +51 -0
- data/lib/fselector/algo_continuous/ReliefF_c.rb +190 -0
- data/lib/fselector/algo_continuous/Relief_c.rb +150 -0
- data/lib/fselector/algo_continuous/TScore.rb +52 -0
- data/lib/fselector/algo_continuous/discretizer.rb +219 -0
- data/lib/fselector/algo_continuous/normalizer.rb +59 -0
- data/lib/fselector/algo_discrete/Accuracy.rb +35 -0
- data/lib/fselector/algo_discrete/AccuracyBalanced.rb +37 -0
- data/lib/fselector/algo_discrete/BiNormalSeparation.rb +45 -0
- data/lib/fselector/algo_discrete/ChiSquaredTest.rb +69 -0
- data/lib/fselector/algo_discrete/CorrelationCoefficient.rb +42 -0
- data/lib/fselector/algo_discrete/DocumentFrequency.rb +36 -0
- data/lib/fselector/algo_discrete/F1Measure.rb +41 -0
- data/lib/fselector/algo_discrete/FishersExactTest.rb +47 -0
- data/lib/fselector/algo_discrete/GMean.rb +37 -0
- data/lib/fselector/algo_discrete/GSSCoefficient.rb +43 -0
- data/lib/fselector/algo_discrete/GiniIndex.rb +44 -0
- data/lib/fselector/algo_discrete/InformationGain.rb +96 -0
- data/lib/fselector/algo_discrete/MatthewsCorrelationCoefficient.rb +45 -0
- data/lib/fselector/algo_discrete/McNemarsTest.rb +57 -0
- data/lib/fselector/algo_discrete/MutualInformation.rb +42 -0
- data/lib/fselector/algo_discrete/OddsRatio.rb +46 -0
- data/lib/fselector/algo_discrete/OddsRatioNumerator.rb +41 -0
- data/lib/fselector/algo_discrete/Power.rb +46 -0
- data/lib/fselector/algo_discrete/Precision.rb +31 -0
- data/lib/fselector/algo_discrete/ProbabilityRatio.rb +41 -0
- data/lib/fselector/algo_discrete/Random.rb +40 -0
- data/lib/fselector/algo_discrete/ReliefF_d.rb +173 -0
- data/lib/fselector/algo_discrete/Relief_d.rb +135 -0
- data/lib/fselector/algo_discrete/Sensitivity.rb +38 -0
- data/lib/fselector/algo_discrete/Specificity.rb +35 -0
- data/lib/fselector/base.rb +322 -0
- data/lib/fselector/base_continuous.rb +25 -0
- data/lib/fselector/base_discrete.rb +355 -0
- data/lib/fselector/ensemble.rb +181 -0
- data/lib/fselector/fileio.rb +455 -0
- data/lib/fselector/util.rb +707 -0
- metadata +86 -0
@@ -0,0 +1,707 @@
|
|
1
|
+
#
|
2
|
+
# add functions to Array class
|
3
|
+
#
|
4
|
+
class Array
|
5
|
+
# summation
|
6
|
+
# @return [Float] sum
|
7
|
+
def sum
|
8
|
+
self.inject(0.0) { |s, i| s+i }
|
9
|
+
end
|
10
|
+
|
11
|
+
|
12
|
+
# average (mean)
|
13
|
+
# @return [Float] average (mean)
|
14
|
+
def ave
|
15
|
+
self.sum / self.size
|
16
|
+
end
|
17
|
+
alias :mean :ave # make mean as an alias of ave
|
18
|
+
|
19
|
+
|
20
|
+
# variance
|
21
|
+
# @return [Float] variance
|
22
|
+
def var
|
23
|
+
u = self.ave
|
24
|
+
v2 = self.inject(0.0) { |v, i| v+(i-u)*(i-u) }
|
25
|
+
|
26
|
+
v2/(self.size-1)
|
27
|
+
end
|
28
|
+
|
29
|
+
|
30
|
+
# standard deviation
|
31
|
+
# @return [Float] standard deviation
|
32
|
+
def sd
|
33
|
+
Math.sqrt(self.var)
|
34
|
+
end
|
35
|
+
|
36
|
+
|
37
|
+
# scale to [min, max]
|
38
|
+
def to_scale(min=0.0, max=1.0)
|
39
|
+
if (min >= max)
|
40
|
+
abort "[#{__FILE__}@#{__LINE__}]: "+
|
41
|
+
"min must be smaller than max"
|
42
|
+
end
|
43
|
+
|
44
|
+
old_min = self.min
|
45
|
+
old_max = self.max
|
46
|
+
|
47
|
+
self.collect do |v|
|
48
|
+
if old_min == old_max
|
49
|
+
max
|
50
|
+
else
|
51
|
+
min + (v-old_min)*(max-min)/(old_max-old_min)
|
52
|
+
end
|
53
|
+
end
|
54
|
+
end
|
55
|
+
|
56
|
+
|
57
|
+
# convert to zscore
|
58
|
+
#
|
59
|
+
# ref: [Wikipedia](http://en.wikipedia.org/wiki/Standard_score)
|
60
|
+
def to_zscore
|
61
|
+
ave = self.ave
|
62
|
+
sd = self.sd
|
63
|
+
|
64
|
+
return self.collect { |v| (v-ave)/sd }
|
65
|
+
end
|
66
|
+
|
67
|
+
|
68
|
+
# to symbol
|
69
|
+
# @return [Array<Symbol>] converted symbols
|
70
|
+
def to_sym
|
71
|
+
self.collect { |x| x.to_sym }
|
72
|
+
end
|
73
|
+
|
74
|
+
|
75
|
+
end
|
76
|
+
|
77
|
+
|
78
|
+
#
|
79
|
+
# add functions to String class
|
80
|
+
#
|
81
|
+
class String
|
82
|
+
# comment line?
|
83
|
+
#
|
84
|
+
# @param [String] char line beginning char
|
85
|
+
def comment?(char='#')
|
86
|
+
return self =~ /^#{char}/
|
87
|
+
end
|
88
|
+
|
89
|
+
|
90
|
+
# blank line?
|
91
|
+
def blank?
|
92
|
+
return self =~ /^\s*$/
|
93
|
+
end
|
94
|
+
|
95
|
+
|
96
|
+
#
|
97
|
+
# Enhanced String.split with escape char, which means
|
98
|
+
# string included in a pair of escape char is considered as a whole
|
99
|
+
# even if it matches the split regular expression. this is especially
|
100
|
+
# useful to parse CSV file that contains comma in a doube-quoted string
|
101
|
+
# e.g. 'a,"b, c",d'.split_me(/,/, '"') => [a, 'b, c', d]
|
102
|
+
#
|
103
|
+
# @param [Regex] delim_regex regular expression for split
|
104
|
+
# @param [String] quote quote char such as ' and "
|
105
|
+
# @return [Array<String>]
|
106
|
+
#
|
107
|
+
def split_me(delim_regex, quote_char="'")
|
108
|
+
d, q = delim_regex, quote_char
|
109
|
+
if not self.count(q) % 2 == 0
|
110
|
+
$stderr.puts "unpaired char of #{q} found, return nil"
|
111
|
+
return nil
|
112
|
+
end
|
113
|
+
self.split(/#{d.source} (?=(?:[^#{q}]* #{q} [^#{q}]* #{q})* [^#{q}]*$) /x)
|
114
|
+
end
|
115
|
+
|
116
|
+
|
117
|
+
end
|
118
|
+
|
119
|
+
#puts "a, 'b,c, d' ,'e'".split_me(/,\s*/, "'")
|
120
|
+
#=>a
|
121
|
+
#=>_'b,c, d'_
|
122
|
+
#=>'e'
|
123
|
+
|
124
|
+
|
125
|
+
#
|
126
|
+
# adapted from the Ruby statistics libraries --
|
127
|
+
# [Rubystats](http://rubystats.rubyforge.org)
|
128
|
+
#
|
129
|
+
# - for Fisher's exact test (Rubystats::FishersExactTest.calculate())
|
130
|
+
# used by algo\_binary/FishersExactText.rb
|
131
|
+
# - for inverse cumulative normal distribution function (Rubystats::NormalDistribution.get\_icdf())
|
132
|
+
# used by algo\_binary/BiNormalSeparation.rb. note the original get\_icdf() function is a private
|
133
|
+
# one, so we have to open it up and that's why the codes here.
|
134
|
+
#
|
135
|
+
#
|
136
|
+
module Rubystats
|
137
|
+
MAX_VALUE = 1.2e290
|
138
|
+
SQRT2PI = 2.5066282746310005024157652848110452530069867406099
|
139
|
+
SQRT2 = 1.4142135623730950488016887242096980785696718753769
|
140
|
+
TWO_PI = 6.2831853071795864769252867665590057683943387987502
|
141
|
+
|
142
|
+
#
|
143
|
+
# Fisher's exact test calculator
|
144
|
+
#
|
145
|
+
class FishersExactTest
|
146
|
+
# new()
|
147
|
+
def initialize
|
148
|
+
@sn11 = 0.0
|
149
|
+
@sn1_ = 0.0
|
150
|
+
@sn_1 = 0.0
|
151
|
+
@sn = 0.0
|
152
|
+
@sprob = 0.0
|
153
|
+
|
154
|
+
@sleft = 0.0
|
155
|
+
@sright = 0.0
|
156
|
+
@sless = 0.0
|
157
|
+
@slarg = 0.0
|
158
|
+
|
159
|
+
@left = 0.0
|
160
|
+
@right = 0.0
|
161
|
+
@twotail = 0.0
|
162
|
+
end
|
163
|
+
|
164
|
+
|
165
|
+
# Fisher's exact test
|
166
|
+
def calculate(n11_,n12_,n21_,n22_)
|
167
|
+
n11_ *= -1 if n11_ < 0
|
168
|
+
n12_ *= -1 if n12_ < 0
|
169
|
+
n21_ *= -1 if n21_ < 0
|
170
|
+
n22_ *= -1 if n22_ < 0
|
171
|
+
n1_ = n11_ + n12_
|
172
|
+
n_1 = n11_ + n21_
|
173
|
+
n = n11_ + n12_ + n21_ + n22_
|
174
|
+
prob = exact(n11_,n1_,n_1,n)
|
175
|
+
left = @sless
|
176
|
+
right = @slarg
|
177
|
+
twotail = @sleft + @sright
|
178
|
+
twotail = 1 if twotail > 1
|
179
|
+
values_hash = { :left =>left, :right =>right, :twotail =>twotail }
|
180
|
+
return values_hash
|
181
|
+
end
|
182
|
+
|
183
|
+
private
|
184
|
+
|
185
|
+
# Reference: "Lanczos, C. 'A precision approximation
|
186
|
+
# of the gamma function', J. SIAM Numer. Anal., B, 1, 86-96, 1964."
|
187
|
+
# Translation of Alan Miller's FORTRAN-implementation
|
188
|
+
# See http://lib.stat.cmu.edu/apstat/245
|
189
|
+
def lngamm(z)
|
190
|
+
x = 0
|
191
|
+
x += 0.0000001659470187408462/(z+7)
|
192
|
+
x += 0.000009934937113930748 /(z+6)
|
193
|
+
x -= 0.1385710331296526 /(z+5)
|
194
|
+
x += 12.50734324009056 /(z+4)
|
195
|
+
x -= 176.6150291498386 /(z+3)
|
196
|
+
x += 771.3234287757674 /(z+2)
|
197
|
+
x -= 1259.139216722289 /(z+1)
|
198
|
+
x += 676.5203681218835 /(z)
|
199
|
+
x += 0.9999999999995183
|
200
|
+
|
201
|
+
return(Math.log(x)-5.58106146679532777-z+(z-0.5) * Math.log(z+6.5))
|
202
|
+
end
|
203
|
+
|
204
|
+
def lnfact(n)
|
205
|
+
if n <= 1
|
206
|
+
return 0
|
207
|
+
else
|
208
|
+
return lngamm(n+1)
|
209
|
+
end
|
210
|
+
end
|
211
|
+
|
212
|
+
def lnbico(n,k)
|
213
|
+
return lnfact(n) - lnfact(k) - lnfact(n-k)
|
214
|
+
end
|
215
|
+
|
216
|
+
def hyper_323(n11, n1_, n_1, n)
|
217
|
+
return Math.exp(lnbico(n1_, n11) + lnbico(n-n1_, n_1-n11) - lnbico(n, n_1))
|
218
|
+
end
|
219
|
+
|
220
|
+
def hyper(n11)
|
221
|
+
return hyper0(n11, 0, 0, 0)
|
222
|
+
end
|
223
|
+
|
224
|
+
def hyper0(n11i,n1_i,n_1i,ni)
|
225
|
+
if n1_i == 0 and n_1i ==0 and ni == 0
|
226
|
+
unless n11i % 10 == 0
|
227
|
+
if n11i == @sn11+1
|
228
|
+
@sprob *= ((@sn1_ - @sn11)/(n11i.to_f))*((@sn_1 - @sn11)/(n11i.to_f + @sn - @sn1_ - @sn_1))
|
229
|
+
@sn11 = n11i
|
230
|
+
return @sprob
|
231
|
+
end
|
232
|
+
if n11i == @sn11-1
|
233
|
+
@sprob *= ((@sn11)/(@sn1_-n11i.to_f))*((@sn11+@sn-@sn1_-@sn_1)/(@sn_1-n11i.to_f))
|
234
|
+
@sn11 = n11i
|
235
|
+
return @sprob
|
236
|
+
end
|
237
|
+
end
|
238
|
+
@sn11 = n11i
|
239
|
+
else
|
240
|
+
@sn11 = n11i
|
241
|
+
@sn1_ = n1_i
|
242
|
+
@sn_1 = n_1i
|
243
|
+
@sn = ni
|
244
|
+
end
|
245
|
+
@sprob = hyper_323(@sn11,@sn1_,@sn_1,@sn)
|
246
|
+
return @sprob
|
247
|
+
end
|
248
|
+
|
249
|
+
def exact(n11,n1_,n_1,n)
|
250
|
+
|
251
|
+
p = i = j = prob = 0.0
|
252
|
+
|
253
|
+
max = n1_
|
254
|
+
max = n_1 if n_1 < max
|
255
|
+
min = n1_ + n_1 - n
|
256
|
+
min = 0 if min < 0
|
257
|
+
|
258
|
+
if min == max
|
259
|
+
@sless = 1
|
260
|
+
@sright = 1
|
261
|
+
@sleft = 1
|
262
|
+
@slarg = 1
|
263
|
+
return 1
|
264
|
+
end
|
265
|
+
|
266
|
+
prob = hyper0(n11,n1_,n_1,n)
|
267
|
+
@sleft = 0
|
268
|
+
|
269
|
+
p = hyper(min)
|
270
|
+
i = min + 1
|
271
|
+
while p < (0.99999999 * prob)
|
272
|
+
@sleft += p
|
273
|
+
p = hyper(i)
|
274
|
+
i += 1
|
275
|
+
end
|
276
|
+
|
277
|
+
i -= 1
|
278
|
+
|
279
|
+
if p < (1.00000001*prob)
|
280
|
+
@sleft += p
|
281
|
+
else
|
282
|
+
i -= 1
|
283
|
+
end
|
284
|
+
|
285
|
+
@sright = 0
|
286
|
+
|
287
|
+
p = hyper(max)
|
288
|
+
j = max - 1
|
289
|
+
while p < (0.99999999 * prob)
|
290
|
+
@sright += p
|
291
|
+
p = hyper(j)
|
292
|
+
j -= 1
|
293
|
+
end
|
294
|
+
j += 1
|
295
|
+
|
296
|
+
if p < (1.00000001*prob)
|
297
|
+
@sright += p
|
298
|
+
else
|
299
|
+
j += 1
|
300
|
+
end
|
301
|
+
|
302
|
+
if (i - n11).abs < (j - n11).abs
|
303
|
+
@sless = @sleft
|
304
|
+
@slarg = 1 - @sleft + prob
|
305
|
+
else
|
306
|
+
@sless = 1 - @sright + prob
|
307
|
+
@slarg = @sright
|
308
|
+
end
|
309
|
+
return prob
|
310
|
+
end
|
311
|
+
|
312
|
+
|
313
|
+
end # class
|
314
|
+
|
315
|
+
#
|
316
|
+
# Normal distribution
|
317
|
+
#
|
318
|
+
class NormalDistribution
|
319
|
+
# Constructs a normal distribution (defaults to zero mean and
|
320
|
+
# unity variance)
|
321
|
+
def initialize(mu=0.0, sigma=1.0)
|
322
|
+
@mean = mu
|
323
|
+
if sigma <= 0.0
|
324
|
+
return "error"
|
325
|
+
end
|
326
|
+
@stdev = sigma
|
327
|
+
@variance = sigma**2
|
328
|
+
@pdf_denominator = SQRT2PI * Math.sqrt(@variance)
|
329
|
+
@cdf_denominator = SQRT2 * Math.sqrt(@variance)
|
330
|
+
end
|
331
|
+
|
332
|
+
|
333
|
+
# Obtain single PDF value
|
334
|
+
# Returns the probability that a stochastic variable x has the value X,
|
335
|
+
# i.e. P(x=X)
|
336
|
+
def get_pdf(x)
|
337
|
+
Math.exp( -((x-@mean)**2) / (2 * @variance)) / @pdf_denominator
|
338
|
+
end
|
339
|
+
|
340
|
+
|
341
|
+
# Obtain single CDF value
|
342
|
+
# Returns the probability that a stochastic variable x is less than X,
|
343
|
+
# i.e. P(x<X)
|
344
|
+
def get_cdf(x)
|
345
|
+
complementary_error( -(x - @mean) / @cdf_denominator) / 2
|
346
|
+
end
|
347
|
+
|
348
|
+
|
349
|
+
# Obtain single inverse CDF value.
|
350
|
+
# returns the value X for which P(x<X).
|
351
|
+
def get_icdf(p)
|
352
|
+
check_range(p)
|
353
|
+
if p == 0.0
|
354
|
+
return -MAX_VALUE
|
355
|
+
end
|
356
|
+
if p == 1.0
|
357
|
+
return MAX_VALUE
|
358
|
+
end
|
359
|
+
if p == 0.5
|
360
|
+
return @mean
|
361
|
+
end
|
362
|
+
|
363
|
+
mean_save = @mean
|
364
|
+
var_save = @variance
|
365
|
+
pdf_D_save = @pdf_denominator
|
366
|
+
cdf_D_save = @cdf_denominator
|
367
|
+
@mean = 0.0
|
368
|
+
@variance = 1.0
|
369
|
+
@pdf_denominator = Math.sqrt(TWO_PI)
|
370
|
+
@cdf_denominator = SQRT2
|
371
|
+
x = find_root(p, 0.0, -100.0, 100.0)
|
372
|
+
#scale back
|
373
|
+
@mean = mean_save
|
374
|
+
@variance = var_save
|
375
|
+
@pdf_denominator = pdf_D_save
|
376
|
+
@cdf_denominator = cdf_D_save
|
377
|
+
return x * Math.sqrt(@variance) + @mean
|
378
|
+
end
|
379
|
+
|
380
|
+
private
|
381
|
+
|
382
|
+
#check that variable is between lo and hi limits.
|
383
|
+
#lo default is 0.0 and hi default is 1.0
|
384
|
+
def check_range(x, lo=0.0, hi=1.0)
|
385
|
+
raise ArgumentError.new("x cannot be nil") if x.nil?
|
386
|
+
if x < lo or x > hi
|
387
|
+
raise ArgumentError.new("x must be less than lo (#{lo}) and greater than hi (#{hi})")
|
388
|
+
end
|
389
|
+
end
|
390
|
+
|
391
|
+
|
392
|
+
def find_root(prob, guess, x_lo, x_hi)
|
393
|
+
accuracy = 1.0e-10
|
394
|
+
max_iteration = 150
|
395
|
+
x = guess
|
396
|
+
x_new = guess
|
397
|
+
error = 0.0
|
398
|
+
_pdf = 0.0
|
399
|
+
dx = 1000.0
|
400
|
+
i = 0
|
401
|
+
while ( dx.abs > accuracy && (i += 1) < max_iteration )
|
402
|
+
#Apply Newton-Raphson step
|
403
|
+
error = cdf(x) - prob
|
404
|
+
if error < 0.0
|
405
|
+
x_lo = x
|
406
|
+
else
|
407
|
+
x_hi = x
|
408
|
+
end
|
409
|
+
_pdf = pdf(x)
|
410
|
+
if _pdf != 0.0
|
411
|
+
dx = error / _pdf
|
412
|
+
x_new = x -dx
|
413
|
+
end
|
414
|
+
# If the NR fails to converge (which for example may be the
|
415
|
+
# case if the initial guess is too rough) we apply a bisection
|
416
|
+
# step to determine a more narrow interval around the root.
|
417
|
+
if x_new < x_lo || x_new > x_hi || _pdf == 0.0
|
418
|
+
x_new = (x_lo + x_hi) / 2.0
|
419
|
+
dx = x_new - x
|
420
|
+
end
|
421
|
+
x = x_new
|
422
|
+
end
|
423
|
+
return x
|
424
|
+
end
|
425
|
+
|
426
|
+
|
427
|
+
#Probability density function
|
428
|
+
def pdf(x)
|
429
|
+
if x.class == Array
|
430
|
+
pdf_vals = []
|
431
|
+
for i in (0 ... x.length)
|
432
|
+
pdf_vals[i] = get_pdf(x[i])
|
433
|
+
end
|
434
|
+
return pdf_vals
|
435
|
+
else
|
436
|
+
return get_pdf(x)
|
437
|
+
end
|
438
|
+
end
|
439
|
+
|
440
|
+
|
441
|
+
#Cummulative distribution function
|
442
|
+
def cdf(x)
|
443
|
+
if x.class == Array
|
444
|
+
cdf_vals = []
|
445
|
+
for i in (0...x.size)
|
446
|
+
cdf_vals[i] = get_cdf(x[i])
|
447
|
+
end
|
448
|
+
return cdf_vals
|
449
|
+
else
|
450
|
+
return get_cdf(x)
|
451
|
+
end
|
452
|
+
end
|
453
|
+
|
454
|
+
|
455
|
+
|
456
|
+
# Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
|
457
|
+
#
|
458
|
+
# Developed at SunSoft, a Sun Microsystems, Inc. business.
|
459
|
+
# Permission to use, copy, modify, and distribute this
|
460
|
+
# software is freely granted, provided that this notice
|
461
|
+
# is preserved.
|
462
|
+
#
|
463
|
+
# x
|
464
|
+
# 2 |\
|
465
|
+
# erf(x) = --------- | exp(-t*t)dt
|
466
|
+
# sqrt(pi) \|
|
467
|
+
# 0
|
468
|
+
#
|
469
|
+
# erfc(x) = 1-erf(x)
|
470
|
+
# Note that
|
471
|
+
# erf(-x) = -erf(x)
|
472
|
+
# erfc(-x) = 2 - erfc(x)
|
473
|
+
#
|
474
|
+
# Method:
|
475
|
+
# 1. For |x| in [0, 0.84375]
|
476
|
+
# erf(x) = x + x*R(x^2)
|
477
|
+
# erfc(x) = 1 - erf(x) if x in [-.84375,0.25]
|
478
|
+
# = 0.5 + ((0.5-x)-x*R) if x in [0.25,0.84375]
|
479
|
+
# where R = P/Q where P is an odd poly of degree 8 and
|
480
|
+
# Q is an odd poly of degree 10.
|
481
|
+
# -57.90
|
482
|
+
# | R - (erf(x)-x)/x | <= 2
|
483
|
+
#
|
484
|
+
#
|
485
|
+
# Remark. The formula is derived by noting
|
486
|
+
# erf(x) = (2/sqrt(pi))*(x - x^3/3 + x^5/10 - x^7/42 + ....)
|
487
|
+
# and that
|
488
|
+
# 2/sqrt(pi) = 1.128379167095512573896158903121545171688
|
489
|
+
# is close to one. The interval is chosen because the fix
|
490
|
+
# point of erf(x) is near 0.6174 (i.e., erf(x)=x when x is
|
491
|
+
# near 0.6174), and by some experiment, 0.84375 is chosen to
|
492
|
+
# guarantee the error is less than one ulp for erf.
|
493
|
+
#
|
494
|
+
# 2. For |x| in [0.84375,1.25], let s = |x| - 1, and
|
495
|
+
# c = 0.84506291151 rounded to single (24 bits)
|
496
|
+
# erf(x) = sign(x) * (c + P1(s)/Q1(s))
|
497
|
+
# erfc(x) = (1-c) - P1(s)/Q1(s) if x > 0
|
498
|
+
# 1+(c+P1(s)/Q1(s)) if x < 0
|
499
|
+
# |P1/Q1 - (erf(|x|)-c)| <= 2**-59.06
|
500
|
+
# Remark: here we use the taylor series expansion at x=1.
|
501
|
+
# erf(1+s) = erf(1) + s*Poly(s)
|
502
|
+
# = 0.845.. + P1(s)/Q1(s)
|
503
|
+
# That is, we use rational approximation to approximate
|
504
|
+
# erf(1+s) - (c = (single)0.84506291151)
|
505
|
+
# Note that |P1/Q1|< 0.078 for x in [0.84375,1.25]
|
506
|
+
# where
|
507
|
+
# P1(s) = degree 6 poly in s
|
508
|
+
# Q1(s) = degree 6 poly in s
|
509
|
+
#
|
510
|
+
# 3. For x in [1.25,1/0.35(~2.857143)],
|
511
|
+
# erfc(x) = (1/x)*exp(-x*x-0.5625+R1/S1)
|
512
|
+
# erf(x) = 1 - erfc(x)
|
513
|
+
# where
|
514
|
+
# R1(z) = degree 7 poly in z, (z=1/x^2)
|
515
|
+
# S1(z) = degree 8 poly in z
|
516
|
+
#
|
517
|
+
# 4. For x in [1/0.35,28]
|
518
|
+
# erfc(x) = (1/x)*exp(-x*x-0.5625+R2/S2) if x > 0
|
519
|
+
# = 2.0 - (1/x)*exp(-x*x-0.5625+R2/S2) if -6<x<0
|
520
|
+
# = 2.0 - tiny (if x <= -6)
|
521
|
+
# erf(x) = sign(x)*(1.0 - erfc(x)) if x < 6, else
|
522
|
+
# erf(x) = sign(x)*(1.0 - tiny)
|
523
|
+
# where
|
524
|
+
# R2(z) = degree 6 poly in z, (z=1/x^2)
|
525
|
+
# S2(z) = degree 7 poly in z
|
526
|
+
#
|
527
|
+
# Note1:
|
528
|
+
# To compute exp(-x*x-0.5625+R/S), let s be a single
|
529
|
+
# PRECISION number and s := x then
|
530
|
+
# -x*x = -s*s + (s-x)*(s+x)
|
531
|
+
# exp(-x*x-0.5626+R/S) =
|
532
|
+
# exp(-s*s-0.5625)*exp((s-x)*(s+x)+R/S)
|
533
|
+
# Note2:
|
534
|
+
# Here 4 and 5 make use of the asymptotic series
|
535
|
+
# exp(-x*x)
|
536
|
+
# erfc(x) ~ ---------- * ( 1 + Poly(1/x^2) )
|
537
|
+
# x*sqrt(pi)
|
538
|
+
# We use rational approximation to approximate
|
539
|
+
# g(s)=f(1/x^2) = log(erfc(x)*x) - x*x + 0.5625
|
540
|
+
# Here is the error bound for R1/S1 and R2/S2
|
541
|
+
# |R1/S1 - f(x)| < 2**(-62.57)
|
542
|
+
# |R2/S2 - f(x)| < 2**(-61.52)
|
543
|
+
#
|
544
|
+
# 5. For inf > x >= 28
|
545
|
+
# erf(x) = sign(x) *(1 - tiny) (raise inexact)
|
546
|
+
# erfc(x) = tiny*tiny (raise underflow) if x > 0
|
547
|
+
# = 2 - tiny if x<0
|
548
|
+
#
|
549
|
+
# 7. Special case:
|
550
|
+
# erf(0) = 0, erf(inf) = 1, erf(-inf) = -1,
|
551
|
+
# erfc(0) = 1, erfc(inf) = 0, erfc(-inf) = 2,
|
552
|
+
# erfc/erf(NaN) is NaN
|
553
|
+
#
|
554
|
+
# $efx8 = 1.02703333676410069053e00
|
555
|
+
#
|
556
|
+
# Coefficients for approximation to erf on [0,0.84375]
|
557
|
+
#
|
558
|
+
|
559
|
+
# Error function.
|
560
|
+
# Based on C-code for the error function developed at Sun Microsystems.
|
561
|
+
# Author:: Jaco van Kooten
|
562
|
+
|
563
|
+
def error(x)
|
564
|
+
e_efx = 1.28379167095512586316e-01
|
565
|
+
|
566
|
+
ePp = [ 1.28379167095512558561e-01,
|
567
|
+
-3.25042107247001499370e-01,
|
568
|
+
-2.84817495755985104766e-02,
|
569
|
+
-5.77027029648944159157e-03,
|
570
|
+
-2.37630166566501626084e-05 ]
|
571
|
+
|
572
|
+
eQq = [ 3.97917223959155352819e-01,
|
573
|
+
6.50222499887672944485e-02,
|
574
|
+
5.08130628187576562776e-03,
|
575
|
+
1.32494738004321644526e-04,
|
576
|
+
-3.96022827877536812320e-06 ]
|
577
|
+
|
578
|
+
# Coefficients for approximation to erf in [0.84375,1.25]
|
579
|
+
ePa = [-2.36211856075265944077e-03,
|
580
|
+
4.14856118683748331666e-01,
|
581
|
+
-3.72207876035701323847e-01,
|
582
|
+
3.18346619901161753674e-01,
|
583
|
+
-1.10894694282396677476e-01,
|
584
|
+
3.54783043256182359371e-02,
|
585
|
+
-2.16637559486879084300e-03 ]
|
586
|
+
|
587
|
+
eQa = [ 1.06420880400844228286e-01,
|
588
|
+
5.40397917702171048937e-01,
|
589
|
+
7.18286544141962662868e-02,
|
590
|
+
1.26171219808761642112e-01,
|
591
|
+
1.36370839120290507362e-02,
|
592
|
+
1.19844998467991074170e-02 ]
|
593
|
+
|
594
|
+
e_erx = 8.45062911510467529297e-01
|
595
|
+
|
596
|
+
abs_x = (if x >= 0.0 then x else -x end)
|
597
|
+
# 0 < |x| < 0.84375
|
598
|
+
if abs_x < 0.84375
|
599
|
+
#|x| < 2**-28
|
600
|
+
if abs_x < 3.7252902984619141e-9
|
601
|
+
retval = abs_x + abs_x * e_efx
|
602
|
+
else
|
603
|
+
s = x * x
|
604
|
+
p = ePp[0] + s * (ePp[1] + s * (ePp[2] + s * (ePp[3] + s * ePp[4])))
|
605
|
+
|
606
|
+
q = 1.0 + s * (eQq[0] + s * (eQq[1] + s *
|
607
|
+
( eQq[2] + s * (eQq[3] + s * eQq[4]))))
|
608
|
+
retval = abs_x + abs_x * (p / q)
|
609
|
+
end
|
610
|
+
elsif abs_x < 1.25
|
611
|
+
s = abs_x - 1.0
|
612
|
+
p = ePa[0] + s * (ePa[1] + s *
|
613
|
+
(ePa[2] + s * (ePa[3] + s *
|
614
|
+
(ePa[4] + s * (ePa[5] + s * ePa[6])))))
|
615
|
+
|
616
|
+
q = 1.0 + s * (eQa[0] + s *
|
617
|
+
(eQa[1] + s * (eQa[2] + s *
|
618
|
+
(eQa[3] + s * (eQa[4] + s * eQa[5])))))
|
619
|
+
retval = e_erx + p / q
|
620
|
+
|
621
|
+
elsif abs_x >= 6.0
|
622
|
+
retval = 1.0
|
623
|
+
else
|
624
|
+
retval = 1.0 - complementary_error(abs_x)
|
625
|
+
end
|
626
|
+
return (if x >= 0.0 then retval else -retval end)
|
627
|
+
end
|
628
|
+
|
629
|
+
# Complementary error function.
|
630
|
+
# Based on C-code for the error function developed at Sun Microsystems.
|
631
|
+
# author Jaco van Kooten
|
632
|
+
|
633
|
+
def complementary_error(x)
|
634
|
+
# Coefficients for approximation of erfc in [1.25,1/.35]
|
635
|
+
|
636
|
+
eRa = [-9.86494403484714822705e-03,
|
637
|
+
-6.93858572707181764372e-01,
|
638
|
+
-1.05586262253232909814e01,
|
639
|
+
-6.23753324503260060396e01,
|
640
|
+
-1.62396669462573470355e02,
|
641
|
+
-1.84605092906711035994e02,
|
642
|
+
-8.12874355063065934246e01,
|
643
|
+
-9.81432934416914548592e00 ]
|
644
|
+
|
645
|
+
eSa = [ 1.96512716674392571292e01,
|
646
|
+
1.37657754143519042600e02,
|
647
|
+
4.34565877475229228821e02,
|
648
|
+
6.45387271733267880336e02,
|
649
|
+
4.29008140027567833386e02,
|
650
|
+
1.08635005541779435134e02,
|
651
|
+
6.57024977031928170135e00,
|
652
|
+
-6.04244152148580987438e-02 ]
|
653
|
+
|
654
|
+
# Coefficients for approximation to erfc in [1/.35,28]
|
655
|
+
|
656
|
+
eRb = [-9.86494292470009928597e-03,
|
657
|
+
-7.99283237680523006574e-01,
|
658
|
+
-1.77579549177547519889e01,
|
659
|
+
-1.60636384855821916062e02,
|
660
|
+
-6.37566443368389627722e02,
|
661
|
+
-1.02509513161107724954e03,
|
662
|
+
-4.83519191608651397019e02 ]
|
663
|
+
|
664
|
+
eSb = [ 3.03380607434824582924e01,
|
665
|
+
3.25792512996573918826e02,
|
666
|
+
1.53672958608443695994e03,
|
667
|
+
3.19985821950859553908e03,
|
668
|
+
2.55305040643316442583e03,
|
669
|
+
4.74528541206955367215e02,
|
670
|
+
-2.24409524465858183362e01 ]
|
671
|
+
|
672
|
+
abs_x = (if x >= 0.0 then x else -x end)
|
673
|
+
if abs_x < 1.25
|
674
|
+
retval = 1.0 - error(abs_x)
|
675
|
+
elsif abs_x > 28.0
|
676
|
+
retval = 0.0
|
677
|
+
|
678
|
+
# 1.25 < |x| < 28
|
679
|
+
else
|
680
|
+
s = 1.0/(abs_x * abs_x)
|
681
|
+
if abs_x < 2.8571428
|
682
|
+
r = eRa[0] + s * (eRa[1] + s *
|
683
|
+
(eRa[2] + s * (eRa[3] + s * (eRa[4] + s *
|
684
|
+
(eRa[5] + s *(eRa[6] + s * eRa[7])
|
685
|
+
)))))
|
686
|
+
|
687
|
+
s = 1.0 + s * (eSa[0] + s * (eSa[1] + s *
|
688
|
+
(eSa[2] + s * (eSa[3] + s * (eSa[4] + s *
|
689
|
+
(eSa[5] + s * (eSa[6] + s * eSa[7])))))))
|
690
|
+
|
691
|
+
else
|
692
|
+
r = eRb[0] + s * (eRb[1] + s *
|
693
|
+
(eRb[2] + s * (eRb[3] + s * (eRb[4] + s *
|
694
|
+
(eRb[5] + s * eRb[6])))))
|
695
|
+
|
696
|
+
s = 1.0 + s * (eSb[0] + s *
|
697
|
+
(eSb[1] + s * (eSb[2] + s * (eSb[3] + s *
|
698
|
+
(eSb[4] + s * (eSb[5] + s * eSb[6]))))))
|
699
|
+
end
|
700
|
+
retval = Math.exp(-x * x - 0.5625 + r/s) / abs_x
|
701
|
+
end
|
702
|
+
return ( if x >= 0.0 then retval else 2.0 - retval end )
|
703
|
+
end
|
704
|
+
|
705
|
+
end # class
|
706
|
+
|
707
|
+
end # module
|