1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446
|
\defmodule {gofs}
This module provides tools for computing goodness-of-fit test statistics
for testing the hypothesis $\cH_0$ that a sample of $N$ observations
$V_1,\dots,V_N$ comes from a given univariate probability
distribution $F$.
These test statistics generally measure, in different ways, the
distance between $F$ and the {\em empirical distribution function\/}
(EDF) $\hat F_N$ of $V_1,\dots,V_N$.
They are also called EDF test statistics.
The observations $V_i$ are usually transformed into $U_i = F(V_i)$,
which always satisfy $0\le U_i\le 1$, and which
follow the $U(0,1)$ distribution under $\cH_0$.
These observations are also usually sorted.
Here, $U_{(1)}, \dots, U_{(N)}$ stand for $N$ observations
$U_1,\dots,U_N$ sorted by increasing order, where $0\le U_i\le 1$.
Procedures for applying certain types of transformations to the
observations $V_i$ or $U_i$ are also provided.
This includes the transformation $U_i = F(V_i)$, as well as
the power ratio and iterated spacing transformations \cite{tSTE86a}.
\bigskip\hrule\medskip
\code\hide
/* gofs.h for ANSI C */
#ifndef GOFS_H
#define GOFS_H
\endhide
#include <testu01/bitset.h> /* From the library mylib */
#include <testu01/fmass.h>
#include <testu01/fdist.h>
#include <testu01/wdist.h>
\endcode
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\guisec{Environment variables}
\code
extern double gofs_MinExpected;
\endcode
\tab Used for the chi-square tests.
When a chi-square test statistic is computed, the expected number
of observations in each class should be large enough if we want
the chi-square test statistic to follow approximately the
chi-square distribution. Larger expected numbers are usually
required when these numbers differ between classes \cite{tREA88a}.
The function {\tt gofs\_MergeClasses} can be used to regroup classes
in order to make sure that the expected number in each class is
at least {\tt gofs\_MinExpected}.
The default value of this variable is 10.0.
\iffalse %%%
This is for testu01:
For some tests, the software will merge
classes in such a way that this is always so. For others, an error
message will be printed if this not the case (see the {\it restrictions}
that apply for the different tests).
\fi %%
\hpierre {Et si on mettait cette variable en param\`etre \`a
{\tt MergeClasses} \`a la place? Peser le pour et le contre.
En fait, dans plusieurs situations, la valeur 10.0 est beaucoup
trop \'elev\'ee. }
\hrichard {Cette variable est vraiment d'environnement. Depuis des
ann\'ees, on ne l'a jamais vari\'e. Et les fonctions qui l'utilisent
sont appel\'ees des douzaines de fois partout. Ce type de variables
devrait demeurer environnement.}
\endtab
\ifdetailed %%%%
\code
extern double gofs_EpsilonAD;
\endcode
\tab When computing the Anderson-Darling statistic $A_N^2$,
all observations $U_i$ are projected to the interval
$[\epsilon,\,1-\epsilon]$ for some $\epsilon > 0$, in order to
avoid numerical overflow when taking the logarithm of $U_i$ or
$1-U_i$. This variable gives the value of $\epsilon$;
its default value is {\tt DBL\_EPSILON/2.0}.
{\tt DBL\_EPSILON} (from {\tt float.h}) is usually $2^{-52}$.
\hpierre {Autre choix possible: cacher cela dans le .c.
Mais il ne semble pas y avoir d'avantage \`a faire cela,
tandis que le laisser ici peut permettre aux ``experts'' de faire
\'eventuellement des exp\'eriences avec le choix de $\epsilon$. }
\endtab
\fi %%%% detailed
\hrichard {Il faut mettre ``{\tt extern}'' parce que sinon
il y aurait autant d'instances de cette variable que d'inclusions
de ce fichier *.h dans le programme, ce qui provoquerait une erreur
avec plusieurs compilateurs.
Parmi toutes les conventions des diff\'erents compilateurs pour les
variables externes, celle-ci est la plus propre et est compatible
ANSI C. Parmi les 5 conventions utilis\'ees couramment, c'est celle
recommand\'ee par Harbison+Steele p. 94.}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\guisec{Transforming the observations}
\code
void gofs_ContUnifTransform (double V[], long N, wdist_CFUNC F,
double par[], double U[]);
\endcode
\tab Applies the transformation $U_i = F(V_i)$ to the values in
{\tt V[1..N]}, where $F$ is a {\em continuous\/} distribution function
given by {\tt F} and with parameters in {\tt par},
and puts the result in {\tt U[1..N]}.
If {\tt V} contains random variables from the distribution function
{\tt F}, then {\tt U} will contain uniform random variables over $(0,1)$.
\endtab
\code
void gofs_DiscUnifTransform (double V[], long N, wdist_DFUNC F,
fmass_INFO W, double U[]);
\endcode
\tab Applies the transformation $U_i = F(V_i)$ to the values in
{\tt V[1..N]}, where $F$ is a {\em discrete\/} distribution function
specified by {\tt F} and the previously-created structure {\tt W},
and puts the result in {\tt U[1..N]}.
Note: If {\tt V[1..N]} are the values of random variables with
distribution function {\tt F}, then {\tt U[1..N]} will contain
the values of {\em discrete\/} random variables distributed over the
set of values taken by {\tt F},
not uniform random variables over $(0,1)$.
\hrichard {Devrait probablement \^etre \'elimin\'e. Utilis\'e 1 fois dans
smultin, mais pourrait \^etre remplac\'e.}
\hpierre {Tu as probablement raison.}
\endtab
\code
void gofs_DiffD (double U[], double D[], long N1, long N2,
double a, double b);
\endcode
\tab Assumes that the real-valued observations {\tt U[N1..N2]}
are already sorted in increasing order and computes the differences
between the successive observations.
The difference {\tt U[i+1] - U[i]} is put in {\tt D[i]} for
{\tt N1 <= i < N2}, whereas {\tt U[N1] - a} is put into {\tt D[N1-1]}
and {\tt b - U[N2]} is put into {\tt D[N2]}.
The sizes of the arrays {\tt U} and {\tt D} must be at least {\tt N2+1}.
\hpierre {ATTENTION: J'ai chang\'e cette proc\'edure et la suivante
pour les rendre plus g\'en\'erales et surtout plus {\em semblables}.
Un appel \`a l'ancien {\tt DiffD (U, D, N)} doit se traduire par
{\tt DiffD (U, D, 1, N, 0.0, 1.0)}, tandis qu'un appel
\`a l'ancien {\tt DiffL (U, D, N1, N2, L)} doit se traduire par
{\tt DiffD (U, D, N1, N2, 0, L+U[N1])}. }
\endtab
\code
void gofs_DiffL (long U[], long D[], long N1, long N2, long a, long b);
#ifdef USE_LONGLONG
void gofs_DiffLL (longlong U[], longlong D[], long N1, long N2,
longlong a, longlong b);
void gofs_DiffULL (ulonglong U[], ulonglong D[], long N1, long N2,
ulonglong a, ulonglong b);
#endif
\endcode
\tab Same as {\tt gofs\_DiffD}, but for integer-valued observations.
\endtab
\code
void gofs_IterateSpacings (double V[], double S[], long N);
\endcode
\tab Applies one iteration of the {\em iterated spacings\/}
transformation \cite{rKNU98a,tSTE86a}.
Assumes that {\tt S[0...N]} contains the {\em spacings\/}
between $N$ real numbers $U_1,\dots,U_N$ in the interval $[0,1]$.
These spacings are defined by
$$ S_i = U_{(i+1)} - U_{(i)}, \qquad 0\le i\le N, $$
where $U_{(0)}=0$, $U_{(N+1)}=1$, and
$U_{(1)},\dots,U_{(N)}$, are the $U_i$ sorted in increasing order.
% These $U_i$ do not need to be in the array {\tt V}.
These spacings may have been obtained by calling {\tt gofs\_DiffD}.
This procedure transforms the spacings into new
spacings, by a variant of the method described
in section 11 of \cite {rMAR85a} and also by Stephens \cite{tSTE86a}:
% See also Knuth (1998), 3th edition.
it sorts $S_0,\dots,S_N$ to obtain
$S_{(0)} \le S_{(1)} \le S_{(2)} \le \cdots \le S_{(N)}$,
computes the weighted differences
\begin {eqnarray*}
S_{0} &=& (N+1) S_{(0)}, \\
S_{1} &=& N (S_{(1)}-S_{(0)}), \\
S_{2} &=& (N-1) (S_{(2)}-S_{(1)}),\\
& \vdots& \\
S_{N} &=& S_{(N)}-S_{(N-1)},
\end {eqnarray*}
and computes $V_i = S_0 + S_1 + \cdots + S_{i-1}$ for $1\le i\le N$.
It then returns $S_0,\dots,S_N$ in {\tt S[0..N]} and
$V_1,\dots,V_N$ in {\tt V[1..N]}.
Under the assumption that the $U_i$ are i.i.d.\ $U(0,1)$, the new
$S_i$ can be considered as a new set of spacings having the same
distribution as the original spacings, and the $V_i$ are a new sample
of i.i.d.\ $U(0,1)$ random variables, sorted by increasing order.
This transformation is useful to detect {\em clustering\/} in a data
set: A pair of observations that are close to each other is transformed
into an observation close to zero. A data set with unusually clustered
observations is thus transformed to a data set with an
accumulation of observations near zero, which is easily detected by
the Anderson-Darling GOF test.
\hrichard {Utilis\'e dans \tt snpair}
\endtab
\code
void gofs_PowerRatios (double U[], long N);
\endcode
\tab Applies the {\em power ratios\/} transformation $W$ described
in section 8.4 of Stephens \cite{tSTE86a}.
Assume that {\tt U[1...N]} contains $N$ real numbers
$U_{(1)},\dots,U_{(N)}$ from the interval $[0,1]$,
already sorted in increasing order, and computes the transformations:
$$ U'_i = (U_{(i)} / U_{(i+1)})^i, \qquad i=1,\dots,N,$$
with $U_{(N+1)} = 1$.
These $U'_i$ are sorted in increasing order and put back in
{\tt U[1...N]}.
If the $U_{(i)}$ are i.i.d.\ $U(0,1)$ sorted by increasing order,
then the $U'_i$ are also i.i.d.\ $U(0,1)$.
This transformation is useful to detect clustering, as explained in
{\tt gofs\_IterateSpacings}, except that here a pair of
observations close to each other is transformed
into an observation close to 1.
An accumulation of observations near 1 is also easily detected by
the Anderson-Darling GOF test.
\hrichard {Utilis\'e dans \tt snpair}
\endtab
\code
void gofs_MergeClasses (double NbExp[], long Loc[],
long *smin, long *smax, long *NbClasses);
\endcode
\tab This function is convenient for regrouping classes before
applying a chi-square test,
in the case where the expected number of observations in some of
the classes may be too small.
It merges classes of observations so that the expected
number of observations in each class is at least
{\tt gofs\_MinExpected}. Initially, the expected numbers in each class
are in {\tt NbExp[*smin...*smax]}.
When the function returns, if ${\tt Loc}[s] = j$, this means that class
$s$ has been merged with class $j$.
In this case, all observations that previously belonged to class $s$
are redirected to class $j$,
% i.e. considered as if they belong to class $j$,
and {\tt NbExp$[s]$} has been added to {\tt NbExp$[j]$}
and then set to zero.
{\tt NbClasses} gives the final number of classes,
{\tt smin} contains the new index of the lowest class,
and {\tt smax} the new index of the highest class.
\endtab
\code
void gofs_WriteClasses (double NbExp[], long Loc[],
long smin, long smax, long NbClasses);
\endcode
\tab Prints the classes before or after their regrouping by
{\tt gofs\_MergeClasses}.
The parameters are the same as for the latter function.
If {\tt NbClasses > 0}, assumes that {\tt gofs\_MergeClasses}
has already been called to regroup classes
% (otherwise {\tt Loc} will be undefined!)}
and prints the classes after the regrouping.
If {\tt NbClasses <= 0}, prints only the classes before any regrouping.
\endtab
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\guisec{Computing EDF test statistics}
\code
double gofs_Chi2 (double NbExp[], long Count[], long smin, long smax);
\endcode
\tab Computes and returns the chi-square statistic for the
observations $o_i$ in {\tt Count[smin...smax]}, for which the
corresponding expected values $e_i$ are in {\tt NbExp[smin...smax]}.
Assuming that $i$ goes from 1 to $k$, where $k =$ {\tt smax-smin+1}
is the number of classes, the chi-square statistic is defined as
\eq
X^2 = \sum_{i=1}^k \frac{(o_i - e_i)^2}{e_i}. \eqlabel{eq:chi-square}
\endeq
Under the hypothesis that the $e_i$ are the correct expectations and
if these $e_i$ are large enough, $X^2$ follows approximately the
chi-square distribution with $k-1$ degrees of freedom.
If some of the $e_i$ are too small, one can use {\tt gofs\_MergeClasses}
to regroup classes.
\endtab
\code
double gofs_Chi2Equal (double NbExp, long Count[], long smin, long smax);
\endcode
\tab Similar to {\tt gofs\_Chi2}, except that the expected
number of observations per class is assumed to be the same for
all classes, and equal to {\tt NbExp}.
\endtab
\code
long gofs_Scan (double U[], long N, double d);
\endcode
\tab Computes and returns the scan statistic $S_N(d)$,
defined in (\ref{eq:scan}).
The $N$ observations in the array {\tt U[1..N]} must be real numbers
in the interval $[0,1]$, sorted in increasing order.
(See {\tt fbar\_Scan} for the distribution function of $S_N(d)$).
\endtab
\code
double gofs_CramerMises (double U[], long N);
\endcode
\tab Computes and returns the Cram\'er-von Mises statistic $W_N^2$
(see \cite{tDUR73a,tSTE70a,tSTE86b}), defined by
\begin {equation}
W_N^2 = {1\over 12N} +
\sum_{j=1}^N \left(U_{(j)} - {(j-0.5) \over N}\right)^2,
\eqlabel {eq:CraMis}
\end {equation}
assuming that {\tt U[1...N]} contains $U_{(1)},\dots,U_{(N)}$
sorted in increasing order.
\endtab
\code
double gofs_WatsonG (double U[], long N);
\endcode
\tab Computes and returns the Watson statistic $G_N$
(see \cite{tWAT76a,tDAR83a}), defined by
\begin {eqnarray}
G_N &=& \sqrt{N} \max_{\rule{0pt}{7pt} 1\le j \le N} \left\{ j/N -
U_{(j)} + \overline U_N - 1/2 \right\}
\eqlabel {eq:WatsonG} \\[6pt]
&=& \sqrt{N}\left (D_N^+ + \overline U_N - 1/2\right), \nonumber
\end {eqnarray}
where $\overline U_N$ is the average of the observations $U_{(j)}$,
assuming that {\tt U[1...N]} contains the sorted $U_{(1)},\dots,U_{(N)}$.
\endtab
\code
double gofs_WatsonU (double U[], long N);
\endcode
\tab Computes and returns the Watson statistic $U_N^2$
(see \cite{tDUR73a,tSTE70a,tSTE86b}), defined by
\begin {eqnarray}
W_N^2 &=& {1\over 12N} +
\sum_{j=1}^N \left\{U_{(j)} - {(j- 0.5)\over N}\right\}^2, \\
U_N^2 &=& W_N^2 - N\left (\overline U_N - 1/2\right)^2.
\eqlabel {eq:WatsonU}
\end {eqnarray}
where $\overline U_N$ is the average of the observations $U_{(j)}$,
assuming that {\tt U[1...N]} contains the sorted $U_{(1)},\dots,U_{(N)}$.
\endtab
\code
double gofs_AndersonDarling (double U[], long N);
\endcode
\tab Computes and returns the Anderson-Darling statistic $A_N^2$
(see \cite{tLEW61a,tSTE86b,tAND52a}), defined by
\begin {eqnarray*}
A_N^2 &=& -N -{1\over N} \sum_{j=1}^N \left\{ (2j-1)\ln(U_{(j)})
+ (2N+1-2j) \ln(1-U_{(j)}) \right\}, \eqlabel {eq:Andar}
\end {eqnarray*}
assuming that {\tt U[1...N]} contains $U_{(1)},\dots,U_{(N)}$.
\ifdetailed %%%%
This function uses the environment variable {\tt gofs\_EpsilonAD}.
\fi %%%% detailed
\endtab
\code
void gofs_KS (double U[], long N, double *DP, double *DM, double *D);
\endcode
\tab Computes the Kolmogorov-Smirnov (KS) test statistics
$D_N^+$, $D_N^-$, and $D_N$, defined by
\begin {eqnarray}
D_N^+ &=& \max_{1\le j\le N} \left(j/N - U_{(j)}\right),
\eqlabel{eq:DNp} \\
D_N^- &=& \max_{1\le j\le N} \left(U_{(j)} - (j-1)/N\right),
\eqlabel{eq:DNm} \\
D_N &=& \max\ (D_N^+, D_N^-). \eqlabel{eq:DN}
\end {eqnarray}
and return their values in {\tt DP, DM}, and {\tt D}, respectively.
These statistics compare the empirical distribution of
$U_{(1)},\dots,U_{(N)}$, which are assumed to be in {\tt U[1...N]},
with the uniform distribution.
\hrichard {Pourquoi avoir enlev\'e les calculs des EDF de ce fichier et
l'avoir mis dans gofw? On calcule d\'ej\`a toutes les stats EDF
explicitement.}
\hpierre {Simplement pour \'eviter d'introduire {\tt TestType},
{\tt TestArray}, etc. dans ce module, et pouvoir tout cacher
cela ensemble \`a la fin de {\tt gofw}. Ces choses sont commodes
pour Testu01, mais trop sp\'ecialis\'ees et pas trop int\'eressantes
pour la plupart des gens. }
\endtab
\code
void gofs_KSJumpOne (double U[], long N, double a, double *DP, double *DM);
\endcode
\tab Compute the KS statistics $D_N^+(a)$ and $D_N^-(a)$ defined in the
description of the function
{\tt fdist\_KSPlusJumpOne}, assuming that $F$ is the
uniform distribution over $[0,1]$ and that
$U_{(1)},\dots,U_{(N)}$ are in {\tt U[1...N]}.
Returns the values in {\tt DP} and {\tt DM}.
\endtab
\hide%%%%%
\code
#if 0
void gofs_KSJumpsMany (double X[], int N, wdist_CFUNC F, double W[],
double *DP, double *DM, int Detail);
#endif
\endcode
\tab We assume that $X[1...N]$ is already sorted and contains the
$N$ empirical observations. We obtain the values of the distribution
$U = F(W, X[i])$ where $F$ is the theoretical
distribution which may be discontinuous, and $W$ are the parameters
of $F$. If {\tt Detail} $> 0$, the computed values of the
distribution will be printed.
Returns the values of the KS statistics in {\tt DP} and {\tt DM}.
\endtab
\endhide%%%%%%%%%%%%%%
\code
\hide
#endif
\endhide
\endcode
|