/* PSPP - a program for statistical analysis.
- Copyright (C) 2009 Free Software Foundation, Inc.
+ Copyright (C) 2009, 2010, 2011 Free Software Foundation, Inc.
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
#include <config.h>
-#include <libpspp/assertion.h>
-#include "covariance.h"
-#include <gl/xalloc.h>
-#include "moments.h"
+#include "math/covariance.h"
+
#include <gsl/gsl_matrix.h>
-#include <data/case.h>
-#include <data/variable.h>
-#include <libpspp/misc.h>
-#include "categoricals.h"
+
+#include "data/case.h"
+#include "data/variable.h"
+#include "libpspp/assertion.h"
+#include "libpspp/misc.h"
+#include "math/categoricals.h"
+#include "math/moments.h"
+
+#include "gl/xalloc.h"
#define n_MOMENTS (MOMENT_VARIANCE + 1)
{
/* The variables for which the covariance matrix is to be calculated. */
size_t n_vars;
- const struct variable **vars;
+ const struct variable *const *vars;
/* Categorical variables. */
struct categoricals *categoricals;
/* Create a covariance struct.
*/
struct covariance *
-covariance_1pass_create (size_t n_vars, const struct variable **vars,
+covariance_1pass_create (size_t n_vars, const struct variable *const *vars,
const struct variable *weight, enum mv_class exclude)
{
size_t i;
- struct covariance *cov = xmalloc (sizeof *cov);
+ struct covariance *cov = xzalloc (sizeof *cov);
cov->passes = 1;
cov->state = 0;
cov->n_cm = (n_vars * (n_vars - 1) ) / 2;
- cov->cm = xcalloc (sizeof *cov->cm, cov->n_cm);
+ if (cov->n_cm > 0)
+ cov->cm = xcalloc (sizeof *cov->cm, cov->n_cm);
+ cov->categoricals = NULL;
return cov;
}
until then.
*/
struct covariance *
-covariance_2pass_create (size_t n_vars, const struct variable **vars,
- size_t n_catvars, const struct variable **catvars,
+covariance_2pass_create (size_t n_vars, const struct variable *const *vars,
+ struct categoricals *cats,
const struct variable *wv, enum mv_class exclude)
{
size_t i;
cov->exclude = exclude;
- cov->n_cm = - 1;
+ cov->n_cm = -1;
cov->cm = NULL;
- cov->categoricals = categoricals_create (catvars, n_catvars, wv);
+ cov->categoricals = cats;
return cov;
}
cm_idx (const struct covariance *cov, int i, int j)
{
int as;
- const int n2j = cov->n_vars - 2 - j;
- const int nj = cov->n_vars - 2 ;
+ const int n2j = cov->dim - 2 - j;
+ const int nj = cov->dim - 2 ;
assert (i >= 0);
- assert (j < cov->n_vars);
+ assert (j < cov->dim);
if ( i == 0)
return -1;
- if (j >= cov->n_vars - 1)
+ if (j >= cov->dim - 1)
return -1;
if ( i <= j)
return i - 1 + as;
}
-static void
+
+/*
+ Returns true iff the variable corresponding to the Ith element of the covariance matrix
+ has a missing value for case C
+*/
+static bool
+is_missing (const struct covariance *cov, int i, const struct ccase *c)
+{
+ const struct variable *var = i < cov->n_vars ?
+ cov->vars[i] :
+ categoricals_get_variable_by_subscript (cov->categoricals, i - cov->n_vars);
+
+ const union value *val = case_data (c, var);
+
+ return var_is_value_missing (var, val, cov->exclude);
+}
+
+
+static double
+get_val (const struct covariance *cov, int i, const struct ccase *c)
+{
+ if ( i < cov->n_vars)
+ {
+ const struct variable *var = cov->vars[i];
+
+ const union value *val = case_data (c, var);
+
+ return val->f;
+ }
+
+ return categoricals_get_binary_by_subscript (cov->categoricals, i - cov->n_vars, c);
+}
+
+#if 0
+void
dump_matrix (const gsl_matrix *m)
{
size_t i, j;
printf ("\n");
}
}
+#endif
/* Call this function for every case in the data set */
void
cov->state = 1;
}
- categoricals_update (cov->categoricals, c);
+ if (cov->categoricals)
+ categoricals_update (cov->categoricals, c);
- for (i = 0 ; i < cov->n_vars; ++i)
+ for (i = 0 ; i < cov->dim; ++i)
{
- const union value *val1 = case_data (c, cov->vars[i]);
+ double v1 = get_val (cov, i, c);
- if ( var_is_value_missing (cov->vars[i], val1, cov->exclude))
+ if ( is_missing (cov, i, c))
continue;
- for (j = 0 ; j < cov->n_vars; ++j)
+ for (j = 0 ; j < cov->dim; ++j)
{
double pwr = 1.0;
- const union value *val2 = case_data (c, cov->vars[j]);
- if ( var_is_value_missing (cov->vars[j], val2, cov->exclude))
+ if ( is_missing (cov, j, c))
continue;
for (m = 0 ; m <= MOMENT_MEAN; ++m)
double *x = gsl_matrix_ptr (cov->moments[m], i, j);
*x += pwr * weight;
- pwr *= val1->f;
+ pwr *= v1;
}
}
}
if (! cov->pass_two_first_case_seen)
{
+ size_t m;
assert (cov->state == 1);
cov->state = 2;
- cov->dim = cov->n_vars + categoricals_total (cov->categoricals);
+ cov->dim = cov->n_vars;
+
+ if (cov->categoricals)
+ cov->dim += categoricals_total (cov->categoricals)
+ - categoricals_get_n_variables (cov->categoricals);
+
cov->n_cm = (cov->dim * (cov->dim - 1) ) / 2;
cov->cm = xcalloc (sizeof *cov->cm, cov->n_cm);
cov->moments[i] = resize_matrix (cov->moments[i], cov->dim);
}
+ if (cov->categoricals)
+ categoricals_done (cov->categoricals);
+
+ /* Populate the moments matrices with the categorical value elements */
+ for (i = cov->n_vars; i < cov->dim; ++i)
+ {
+ for (j = 0 ; j < cov->dim ; ++j) /* FIXME: This is WRONG !!! */
+ {
+ double w = categoricals_get_weight_by_subscript (cov->categoricals, i - cov->n_vars);
+
+ gsl_matrix_set (cov->moments[MOMENT_NONE], i, j, w);
+
+ w = categoricals_get_sum_by_subscript (cov->categoricals, i - cov->n_vars);
+
+ gsl_matrix_set (cov->moments[MOMENT_MEAN], i, j, w);
+ }
+ }
+
+ /* FIXME: This is WRONG!! It must be fixed to properly handle missing values. For
+ now it assumes there are none */
+ for (m = 0 ; m < n_MOMENTS; ++m)
+ {
+ for (i = 0 ; i < cov->dim ; ++i)
+ {
+ double x = gsl_matrix_get (cov->moments[m], i, cov->n_vars -1);
+ for (j = cov->n_vars; j < cov->dim; ++j)
+ {
+ gsl_matrix_set (cov->moments[m], i, j, x);
+ }
+ }
+ }
+
/* Divide the means by the number of samples */
- for (i = 0; i < cov->n_vars; ++i)
+ for (i = 0; i < cov->dim; ++i)
{
- for (j = 0; j < cov->n_vars; ++j)
+ for (j = 0; j < cov->dim; ++j)
{
double *x = gsl_matrix_ptr (cov->moments[MOMENT_MEAN], i, j);
*x /= gsl_matrix_get (cov->moments[MOMENT_NONE], i, j);
}
}
- for (i = 0 ; i < cov->n_vars; ++i)
+ for (i = 0 ; i < cov->dim; ++i)
{
- const union value *val1 = case_data (c, cov->vars[i]);
+ double v1 = get_val (cov, i, c);
- if ( var_is_value_missing (cov->vars[i], val1, cov->exclude))
+ if ( is_missing (cov, i, c))
continue;
- for (j = 0 ; j < cov->n_vars; ++j)
+ for (j = 0 ; j < cov->dim; ++j)
{
int idx;
double ss ;
- const union value *val2 = case_data (c, cov->vars[j]);
+ double v2 = get_val (cov, j, c);
- const double s = pow2 (val1->f - gsl_matrix_get (cov->moments[MOMENT_MEAN], i, j)) * weight;
+ const double s = pow2 (v1 - gsl_matrix_get (cov->moments[MOMENT_MEAN], i, j)) * weight;
- if ( var_is_value_missing (cov->vars[j], val2, cov->exclude))
+ if ( is_missing (cov, j, c))
continue;
{
}
ss =
- (val1->f - gsl_matrix_get (cov->moments[MOMENT_MEAN], i, j))
+ (v1 - gsl_matrix_get (cov->moments[MOMENT_MEAN], i, j))
*
- (val2->f - gsl_matrix_get (cov->moments[MOMENT_MEAN], i, j))
+ (v2 - gsl_matrix_get (cov->moments[MOMENT_MEAN], i, j))
* weight
;
{
cov->cm [idx] += ss;
}
-
}
}
cov->state = 1;
}
- for (i = 0 ; i < cov->n_vars; ++i)
+ for (i = 0 ; i < cov->dim; ++i)
{
const union value *val1 = case_data (c, cov->vars[i]);
- if ( var_is_value_missing (cov->vars[i], val1, cov->exclude))
+ if ( is_missing (cov, i, c))
continue;
- for (j = 0 ; j < cov->n_vars; ++j)
+ for (j = 0 ; j < cov->dim; ++j)
{
double pwr = 1.0;
int idx;
const union value *val2 = case_data (c, cov->vars[j]);
- if ( var_is_value_missing (cov->vars[j], val2, cov->exclude))
+ if ( is_missing (cov, j, c))
continue;
idx = cm_idx (cov, i, j);
cm_to_gsl (struct covariance *cov)
{
int i, j;
- gsl_matrix *m = gsl_matrix_calloc (cov->n_vars, cov->n_vars);
+ gsl_matrix *m = gsl_matrix_calloc (cov->dim, cov->dim);
/* Copy the non-diagonal elements from cov->cm */
- for ( j = 0 ; j < cov->n_vars - 1; ++j)
+ for ( j = 0 ; j < cov->dim - 1; ++j)
{
- for (i = j+1 ; i < cov->n_vars; ++i)
+ for (i = j+1 ; i < cov->dim; ++i)
{
double x = cov->cm [cm_idx (cov, i, j)];
gsl_matrix_set (m, i, j, x);
}
/* Copy the diagonal elements from cov->moments[2] */
- for (j = 0 ; j < cov->n_vars ; ++j)
+ for (j = 0 ; j < cov->dim ; ++j)
{
double sigma = gsl_matrix_get (cov->moments[2], j, j);
gsl_matrix_set (m, j, j, sigma);
}
-static const gsl_matrix *
+static gsl_matrix *
covariance_calculate_double_pass (struct covariance *cov)
{
size_t i, j;
- for (i = 0 ; i < cov->n_vars; ++i)
+ for (i = 0 ; i < cov->dim; ++i)
{
- for (j = 0 ; j < cov->n_vars; ++j)
+ for (j = 0 ; j < cov->dim; ++j)
{
int idx;
double *x = gsl_matrix_ptr (cov->moments[MOMENT_VARIANCE], i, j);
return cm_to_gsl (cov);
}
-static const gsl_matrix *
+static gsl_matrix *
covariance_calculate_single_pass (struct covariance *cov)
{
size_t i, j;
/* Divide the moments by the number of samples */
if ( m > 0)
{
- for (i = 0 ; i < cov->n_vars; ++i)
+ for (i = 0 ; i < cov->dim; ++i)
{
- for (j = 0 ; j < cov->n_vars; ++j)
+ for (j = 0 ; j < cov->dim; ++j)
{
double *x = gsl_matrix_ptr (cov->moments[m], i, j);
*x /= gsl_matrix_get (cov->moments[0], i, j);
}
/* Centre the moments */
- for ( j = 0 ; j < cov->n_vars - 1; ++j)
+ for ( j = 0 ; j < cov->dim - 1; ++j)
{
- for (i = j + 1 ; i < cov->n_vars; ++i)
+ for (i = j + 1 ; i < cov->dim; ++i)
{
double *x = &cov->cm [cm_idx (cov, i, j)];
}
+/* Return a pointer to gsl_matrix containing the pairwise covariances. The
+ caller owns the returned matrix and must free it when it is no longer
+ needed.
-/*
- Return a pointer to gsl_matrix containing the pairwise covariances.
- The matrix remains owned by the COV object, and must not be freed.
- Call this function only after all data have been accumulated.
-*/
-const gsl_matrix *
+ Call this function only after all data have been accumulated. */
+gsl_matrix *
covariance_calculate (struct covariance *cov)
{
- assert ( cov->state > 0 );
+ if ( cov->state <= 0 )
+ return NULL;
switch (cov->passes)
{
}
}
+/*
+ Covariance computed without dividing by the sample size.
+ */
+static gsl_matrix *
+covariance_calculate_double_pass_unnormalized (struct covariance *cov)
+{
+ size_t i, j;
+ for (i = 0 ; i < cov->dim; ++i)
+ {
+ for (j = 0 ; j < cov->dim; ++j)
+ {
+ int idx;
+ double *x = gsl_matrix_ptr (cov->moments[MOMENT_VARIANCE], i, j);
+
+ idx = cm_idx (cov, i, j);
+ if ( idx >= 0)
+ {
+ x = &cov->cm [idx];
+ }
+ }
+ }
+ return cm_to_gsl (cov);
+}
+
+static gsl_matrix *
+covariance_calculate_single_pass_unnormalized (struct covariance *cov)
+{
+ size_t i, j;
+
+ for (i = 0 ; i < cov->dim; ++i)
+ {
+ for (j = 0 ; j < cov->dim; ++j)
+ {
+ double *x = gsl_matrix_ptr (cov->moments[MOMENT_VARIANCE], i, j);
+ *x -= pow2 (gsl_matrix_get (cov->moments[MOMENT_MEAN], i, j))
+ / gsl_matrix_get (cov->moments[MOMENT_NONE], i, j);
+ }
+ }
+ for ( j = 0 ; j < cov->dim - 1; ++j)
+ {
+ for (i = j + 1 ; i < cov->dim; ++i)
+ {
+ double *x = &cov->cm [cm_idx (cov, i, j)];
+
+ *x -=
+ gsl_matrix_get (cov->moments[MOMENT_MEAN], i, j)
+ *
+ gsl_matrix_get (cov->moments[MOMENT_MEAN], j, i)
+ / gsl_matrix_get (cov->moments[MOMENT_NONE], i, j);
+ }
+ }
+
+ return cm_to_gsl (cov);
+}
+
+
+/* Return a pointer to gsl_matrix containing the pairwise covariances. The
+ caller owns the returned matrix and must free it when it is no longer
+ needed.
+
+ Call this function only after all data have been accumulated. */
+gsl_matrix *
+covariance_calculate_unnormalized (struct covariance *cov)
+{
+ if ( cov->state <= 0 )
+ return NULL;
+
+ switch (cov->passes)
+ {
+ case 1:
+ return covariance_calculate_single_pass_unnormalized (cov);
+ break;
+ case 2:
+ return covariance_calculate_double_pass_unnormalized (cov);
+ break;
+ default:
+ NOT_REACHED ();
+ }
+}
+
+/* Function to access the categoricals used by COV
+ The return value is owned by the COV
+*/
+const struct categoricals *
+covariance_get_categoricals (const struct covariance *cov)
+{
+ return cov->categoricals;
+}
/* Destroy the COV object */
covariance_destroy (struct covariance *cov)
{
size_t i;
- free (cov->vars);
+
categoricals_destroy (cov->categoricals);
for (i = 0; i < n_MOMENTS; ++i)
projects / pspp-builds.git / blobdiff