X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=src%2Fmath%2Fts%2Finnovations.c;h=ba2120fb4e6bc735f41d300b5aef406be5c2828d;hb=81579d9e9f994fb2908f50af41c3eb033d216e58;hp=43d86ec1677c8c9bb381f59b23109da8f09940f5;hpb=29e712969d0d5bc1ee20458309dee7df2c38f1ae;p=pspp-builds.git

diff --git a/src/math/ts/innovations.c b/src/math/ts/innovations.c
index 43d86ec1..ba2120fb 100644
--- a/src/math/ts/innovations.c
+++ b/src/math/ts/innovations.c
@@ -1,22 +1,19 @@
-/*
-  src/math/time-series/arma/innovations.c
-  
-  Copyright (C) 2006 Free Software Foundation, Inc. Written by Jason H. Stover.
-  
-  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 the Free
-  Software Foundation; either version 2 of the License, or (at your option)
-  any later version.
-  
-  This program is distributed in the hope that it will be useful, but WITHOUT
-  ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
-  FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
-  more details.
-  
-  You should have received a copy of the GNU General Public License along with
-  this program; if not, write to the Free Software Foundation, Inc., 51
-  Franklin Street, Fifth Floor, Boston, MA 02111-1307, USA.
- */
+/* PSPP - a program for statistical analysis.
+   Copyright (C) 2006, 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
+   the Free Software Foundation, either version 3 of the License, or
+   (at your option) any later version.
+
+   This program is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+   GNU General Public License for more details.
+
+   You should have received a copy of the GNU General Public License
+   along with this program.  If not, see <http://www.gnu.org/licenses/>. */
+
 /*
   Find preliminary ARMA coefficients via the innovations algorithm.
   Also compute the sample mean and covariance matrix for each series.
@@ -28,185 +25,329 @@
   0-387-97429-6. Sections 5.2, 8.3 and 8.4.
  */
 
+#include <config.h>
+
+#include "math/ts/innovations.h"
+
 #include <gsl/gsl_matrix.h>
 #include <gsl/gsl_vector.h>
 #include <math.h>
 #include <stdlib.h>
-#include <data/case.h>
-#include <data/casefile.h>
-#include <libpspp/alloc.h>
-#include <libpspp/compiler.h>
-#include <libpspp/message.h>
-#include <math/coefficient.h>
-#include <math/ts/innovations.h>
+
+#include "libpspp/compiler.h"
+#include "libpspp/misc.h"
+#include "math/coefficient.h"
+
+#include "gl/xalloc.h"
 
 static void
-get_mean_variance (size_t n_vars, const struct casefile *cf,
-		   struct innovations_estimate **est)
-		   
+get_mean (const gsl_matrix *data,
+	  struct innovations_estimate **est)
+
 {
-  struct casereader *r;
-  struct ccase c;
   size_t n;
+  size_t i;
   double d;
-  const union value *tmp;
+  double tmp;
 
-  for (n = 0; n < n_vars; n++)
+  for (n = 0; n < data->size2; n++)
     {
-      est[n]->n_obs = 2.0;
+      est[n]->n_obs = 0.0;
       est[n]->mean = 0.0;
-      est[n]->variance = 0.0;
     }
-  for (r = casefile_get_reader (cf); casereader_read (r, &c);
-       case_destroy (&c))
+  for (i = 0; i < data->size1; i++)
     {
-      for (n = 0; n < n_vars; n++)
+      for (n = 0; n < data->size2; n++)
 	{
-	  tmp = case_data (&c, est[n]->variable->fv);
-	  if (!mv_is_value_missing (&(est[n]->variable->miss), tmp))
+	  tmp = gsl_matrix_get (data, i, n);
+	  if (!isnan (tmp))
 	    {
-	      d = (tmp->f - est[n]->mean) / est[n]->n_obs;
-	      est[n]->mean += d;
-	      est[n]->variance += est[n]->n_obs * est[n]->n_obs * d * d;
 	      est[n]->n_obs += 1.0;
+	      d = (tmp - est[n]->mean) / est[n]->n_obs;
+	      est[n]->mean += d;
 	    }
 	}
     }
-  for (n = 0; n < n_vars; n++)
+}
+static void
+update_cov (struct innovations_estimate **est, gsl_vector_const_view x,
+	    gsl_vector_const_view y, size_t lag)
+{
+  size_t j;
+  double xj;
+  double yj;
+
+  for (j = 0; j < x.vector.size; j++)
     {
-      /* Maximum likelihood estimate of the variance. */
-      est[n]->variance /= est[n]->n_obs;
+      xj = gsl_vector_get (&x.vector, j);
+      yj = gsl_vector_get (&y.vector, j);
+      if (!isnan (xj))
+	{
+	  if (!isnan (yj))
+	    {
+	      xj -= est[j]->mean;
+	      yj -= est[j]->mean;
+	      *(est[j]->cov + lag) += xj * yj;
+	    }
+	}
     }
 }
-
-/*
-  Read the first MAX_LAG cases.
- */
-static bool
-innovations_init_cases (struct casereader *r, struct ccase **c, size_t max_lag)
+static int
+get_covariance (const gsl_matrix *data,
+		struct innovations_estimate **est, size_t max_lag)
 {
-  bool value = true;
-  size_t lag = 0;
+  size_t lag;
+  size_t j;
+  size_t i;
+  int rc = 1;
 
-  while (value && lag < max_lag)
+  assert (data != NULL);
+  assert (est != NULL);
+
+  for (j = 0; j < data->size2; j++)
+    {
+      for (lag = 0; lag <= max_lag; lag++)
+	{
+	  *(est[j]->cov + lag) = 0.0;
+	}
+    }
+  /*
+    The rows are in the outer loop because a gsl_matrix is stored in
+    row-major order.
+   */
+  for (i = 0; i < data->size1; i++)
+    {
+      for (lag = 0; lag <= max_lag && lag < data->size1 - i; lag++)
+	{
+	  update_cov (est, gsl_matrix_const_row (data, i),
+		      gsl_matrix_const_row (data, i + lag), lag);
+	}
+    }
+  for (j = 0; j < data->size2; j++)
     {
-      lag++;
-      value = casereader_read (r, c[lag]);
+      for (lag = 0; lag <= max_lag; lag++)
+	{
+	  *(est[j]->cov + lag) /= est[j]->n_obs;
+	}
     }
-  return value;
+
+  return rc;
 }
 
-/*
-  Read one case and update C, which contains the last MAX_LAG cases.
- */
-static bool
-innovations_update_cases (struct casereader *r, struct ccase **c, size_t max_lag)
+static double
+innovations_convolve (double *x, double *y, struct innovations_estimate *est,
+		      int i)
 {
-  size_t lag;
-  bool value = false;
-  
-  for (lag = 0; lag < max_lag - 1; lag++)
+  int k;
+  double result = 0.0;
+
+  assert (x != NULL && y != NULL);
+  assert (est != NULL);
+  assert (est->scale != NULL);
+  assert (i > 0);
+  for (k = 0; k < i; k++)
     {
-      c[lag] = c[lag+1];
+      result += x[k] * y[k] * est->scale[i-k-1];
     }
-  value = casereader_read (r, c[lag]);
-  return value;
+  return result;
 }
 static void
-get_covariance (size_t n_vars, const struct casefile *cf, 
-		struct innovations_estimate **est, size_t max_lag)
+innovations_update_scale (struct innovations_estimate *est, double *theta,
+			  size_t i)
 {
-  struct casereader *r;
-  struct ccase **c;
-  size_t lag;
-  size_t n;
-  bool read_case = false;
-  double d;
-  double x;
-  const union value *tmp;
-  const union value *tmp2;
+  double result = 0.0;
+  size_t j;
+  size_t k;
 
-  c = xnmalloc (max_lag, sizeof (*c));
-  
-  for (lag = 0; lag < max_lag; lag++)
+  if (i < (size_t) est->max_lag)
     {
-      c[lag] = xmalloc (sizeof *c[lag]);
+      result = est->cov[0];
+      for (j = 0; j < i; j++)
+	{
+	  k = i - j - 1;
+	  result -= pow2 (theta[k]) * est->scale[j];
+	}
+      est->scale[i] = result;
     }
+}
+static void
+init_theta (double **theta, size_t max_lag)
+{
+  size_t i;
+  size_t j;
 
-  r = casefile_get_reader (cf);
-  read_case = innovations_init_cases (r, c, max_lag);
+  for (i = 0; i < max_lag; i++)
+    {
+      for (j = 0; j <= i; j++)
+	{
+	  theta[i][j] = 0.0;
+	}
+    }
+}
+static void
+innovations_update_coeff (double **theta, struct innovations_estimate *est,
+			  size_t max_lag)
+{
+  size_t i;
+  size_t j;
+  size_t k;
 
-  while (read_case)
+  for (i = 0; i < max_lag; i++)
     {
-      for (n = 0; n < n_vars; n++)
+      theta[i][i] = est->cov[i+1] / est->scale[0];
+      for (j = 1; j <= i; j++)
 	{
-	  tmp2 = case_data (c[0], est[n]->variable->fv);
-	  if (!mv_is_value_missing (&est[n]->variable->miss, tmp2))
-	    {
-	      x = tmp2->f - est[n]->mean;
-	      for (lag = 1; lag <= max_lag; lag++)
-		{
-		  tmp = case_data (c[lag], est[n]->variable->fv);
-		  if (!mv_is_value_missing (&est[n]->variable->miss, tmp))
-		    {
-		      d = (tmp->f - est[n]->mean);
-		      *(est[n]->cov + lag) += d * x;
-		    }
-		}
-	    }
+	  k = i - j;
+	  theta[i][k] = (est->cov[k+1] -
+			 innovations_convolve (theta[i] + k + 1, theta[j - 1], est, j))
+	    / est->scale[j];
 	}
-      read_case = innovations_update_cases (r, c, max_lag);
+      innovations_update_scale (est, theta[i], i + 1);
+    }
+}
+static void
+get_coef (const gsl_matrix *data,
+	  struct innovations_estimate **est, size_t max_lag)
+{
+  size_t i;
+  size_t n;
+  double **theta;
+
+  theta = xnmalloc (max_lag, sizeof (*theta));
+  for (i = 0; i < max_lag; i++)
+    {
+      theta[i] = xnmalloc (max_lag, sizeof (**(theta + i)));
     }
-  for (lag = 0; lag <= max_lag; lag++)
+
+  for (n = 0; n < data->size2; n++)
     {
-      for (n = 0; n < n_vars; n++)
+      init_theta (theta, max_lag);
+      innovations_update_scale (est[n], theta[0], 0);
+      innovations_update_coeff (theta, est[n], max_lag);
+      /* Copy the final row of coefficients into EST->COEFF.*/
+      for (i = 0; i < max_lag; i++)
 	{
-	  *(est[n]->cov + lag) /= (est[n]->n_obs - lag);
+	  /*
+	    The order of storage here means that the best predicted value
+	    for the time series is computed as follows:
+
+	    Let X[m], X[m-1],... denote the original series.
+	    Let X_hat[0] denote the best predicted value of X[0],
+	    X_hat[1] denote the projection of X[1] onto the subspace
+	    spanned by {X[0] - X_hat[0]}. Let X_hat[m] denote the
+	    projection of X[m] onto the subspace spanned by {X[m-1] - X_hat[m-1],
+	    X[m-2] - X_hat[m-2],...,X[0] - X_hat[0]}.
+
+	    Then X_hat[m] = est->coeff[m-1] * (X[m-1] - X_hat[m-1])
+	                  + est->coeff[m-1] * (X[m-2] - X_hat[m-2])
+			  ...
+			  + est->coeff[m-max_lag] * (X[m - max_lag] - X_hat[m - max_lag])
+	   */
+	  pspp_coeff_set_estimate (est[n]->coeff[i], theta[max_lag - 1][i]);
 	}
     }
-  for (lag = 0; lag < max_lag; lag++)
+
+  for (i = 0; i < max_lag; i++)
     {
-      free (c[lag]);
+      free (theta[i]);
     }
-  free (c);
+  free (theta);
 }
 
-struct innovations_estimate ** pspp_innovations (const struct variable **vars, size_t *n_vars,
-						 size_t lag, const struct casefile *cf)
+static void
+innovations_struct_init (struct innovations_estimate *est,
+			 const struct design_matrix *dm,
+			 size_t lag)
+{
+  size_t j;
+
+  est->mean = 0.0;
+  /* COV[0] stores the lag 0 covariance (i.e., the variance), COV[1]
+     holds the lag-1 covariance, etc.
+   */
+  est->cov = xnmalloc (lag + 1, sizeof (*est->cov));
+  est->scale = xnmalloc (lag + 1, sizeof (*est->scale));
+  est->coeff = xnmalloc (lag, sizeof (*est->coeff)); /* No intercept. */
+
+  /*
+    The loop below is an unusual use of PSPP_COEFF_INIT(). In a
+    typical model, one column of a DESIGN_MATRIX has one
+    coefficient. But in a time-series model, one column has many
+    coefficients.
+   */
+  for (j = 0; j < lag; j++)
+    {
+      pspp_coeff_init (est->coeff + j, dm);
+    }
+  est->max_lag = (double) lag;
+}
+/*
+  The mean is subtracted from the original data before computing the
+  coefficients. The mean is NOT added back, so if you want to predict
+  a new value, you must add the mean to X_hat[m] to get the correct
+  value.
+ */
+static void
+subtract_mean (gsl_matrix *m, struct innovations_estimate **est)
 {
-  struct innovations_estimate **est;
   size_t i;
   size_t j;
+  double tmp;
 
-  est = xnmalloc (*n_vars, sizeof *est);
-  for (i = 0; i < *n_vars; i++)
+  for (i = 0; i < m->size1; i++)
     {
-      if (vars[i]->type == NUMERIC)
-	{
-	  est[i] = xmalloc (sizeof **est);
-	  est[i]->variable = vars[i];
-	  est[i]->mean = 0.0;
-	  est[i]->variance = 0.0;
-	  est[i]->cov = xnmalloc (lag, sizeof (est[i]->cov));
-	  est[i]->coeff = xnmalloc (lag, sizeof (*est[i]->coeff));
-	  for (j = 0; j < lag; j++)
-	    {
-	      est[i]->coeff[j] = xmalloc (sizeof (*(est[i]->coeff + j)));
-	    }
-	}
-      else
+      for (j = 0; j < m->size2; j++)
 	{
-	  *n_vars--;
-/* 	  msg (MW, _("Cannot compute autocovariance for a non-numeric variable %s"), */
-/* 		     var_to_string (vars[i])); */
+	  tmp = gsl_matrix_get (m, i, j) - est[j]->mean;
+	  gsl_matrix_set (m, i, j, tmp);
 	}
     }
+}
+struct innovations_estimate **
+pspp_innovations (const struct design_matrix *dm, size_t lag)
+{
+  struct innovations_estimate **est;
+  size_t i;
+
+  est = xnmalloc (dm->m->size2, sizeof *est);
+  for (i = 0; i < dm->m->size2; i++)
+    {
+      est[i] = xmalloc (sizeof *est[i]);
+/*       est[i]->variable = vars[i]; */
+      innovations_struct_init (est[i], dm, lag);
+    }
+
+  get_mean (dm->m, est);
+  subtract_mean (dm->m, est);
+  get_covariance (dm->m, est, lag);
+  get_coef (dm->m, est, lag);
 
-  /*
-    First data pass to get the mean and variance.
-   */
-  get_mean_variance (*n_vars, cf, est);
-  get_covariance (*n_vars, cf, est, lag);
-  
   return est;
 }
+
+static void
+pspp_innovations_free_one (struct innovations_estimate *est)
+{
+  size_t i;
+
+  assert (est != NULL);
+  for (i = 0; i < (size_t) est->max_lag; i++)
+    {
+      pspp_coeff_free (est->coeff[i]);
+    }
+  free (est->scale);
+  free (est->cov);
+  free (est);
+}
+
+void pspp_innovations_free (struct innovations_estimate **est, size_t n)
+{
+  size_t i;
+
+  assert (est != NULL);
+  for (i = 0; i < n; i++)
+    {
+      pspp_innovations_free_one (est[i]);
+    }
+  free (est);
+}