最大信息系数（maximal information coefficient，MIC）核心程序

/*
 * Libmine core library.
 *
 * This code is written by Davide Albanese <davide.albanese@gmail.com>
 * and Michele Filosi <filosi@fbk.eu>.
 *
 * Copyright (C) 2012-2016 Davide Albanese, Copyright (C) 2012 Michele
 * Filosi, Copyright (C) 2012 Fondazione Bruno Kessler.
 *
 * References:
 *   * Original MINE paper: DOI: 10.1126/science.1205438;
 *   * Minepy paper: DOI: 10.1093/bioinformatics/bts707D;
 *   * MIC_e and TIC: DOI: arXiv:1505.02213 and DOI: arXiv:1505.02214;
 *   * GMIC: DOI: arXiv:1308.5712.
 *
 * 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/>.
 */

#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>
#include <float.h>
#include "mine.h"

char *libmine_version = LIBMINE_VERSION;

#define MAX(a, b) (((a) > (b) ? (a) : (b)))
#define MIN(a, b) (((a) < (b) ? (a) : (b)))

#define TRUE 1
#define FALSE 0


void quicksort(double *a, int *idx, int l, int u)
{
  int i, m, idx_temp;
  double a_temp;

  if (l >= u)
    return;

  m = l;
  for (i=l+1; i<=u; i++)
    {
      if (a[i] < a[l])
        {
          ++m;

          idx_temp = idx[m];
          idx[m] = idx[i];
          idx[i] = idx_temp;

          a_temp = a[m];
          a[m] = a[i];
          a[i] = a_temp;
        }
    }

  idx_temp = idx[l];
  idx[l] = idx[m];
  idx[m] = idx_temp;

  a_temp = a[l];
  a[l] = a[m];
  a[m] = a_temp;

  quicksort(a, idx, l, m-1);
  quicksort(a, idx, m+1, u);
}


int *argsort(double *a, int n)
{
  double *a_cpy;
  int i, *idx;

  a_cpy = (double *) malloc(n * sizeof(double));
  if (a_cpy == NULL)
    return NULL;

  idx = (int *) malloc(n * sizeof(int));
  if (idx == NULL)
    {
      free(a_cpy);
      return NULL;
    }

  /* fill a_cpy */
  memcpy(a_cpy, a, n * sizeof(double));

  /* fill idx */
  for (i=0; i<n; i++)
    idx[i] = i;

  quicksort(a_cpy, idx, 0, n-1);
  free(a_cpy);

  return idx;
}


/*
 * Returns the entropy induced by the points on the partition Q.
 * See section 3.2.1, page 10, SOM.
 *
 * Parameters
 *   cumhist : cumulative histogram matrix along P_map
 *   cumhist_log : log(cumhist)
 *   q : number of rows of cumhist (number of partitions in Q_map)
 *   p : number of cols of cumhist (number of partitions in P_map)
 *   n : total number of points
 */
double hq(int **cumhist, double **cumhist_log, int q, int p, int n)
{
  int i;
  double total, total_log, prob, prob_log, H = 0.0;

  total = (double) n;
  total_log = log(total);

  for (i=0; i<q; i++)
    {
      prob = (double) cumhist[i][p-1] / total;
      if (prob != 0)
        {
          prob_log = cumhist_log[i][p-1] - total_log;
          H -= prob * prob_log;
       }
    }

  return H;
}

/*
 * Returns the entropy induced by the points on the partition
 * <c_0, c_s, c_t>. See line 5 of Algorithm 2, SOM.
 *
 * Parameters
 *   c : c_1, ..., c_p
 *   c_log : log(c)
 *   s : s in c_s
 *   t : t in c_t
 */
double hp3(int *c, double *c_log, int s, int t)
{
  int sum;
  double total, total_log, prob, prob_log, H = 0.0;

  if (s == t)
    return 0.0;

  total = (double) c[t-1];
  total_log = log(total);

  prob = (double) c[s-1] / total;
  if (prob != 0)
    {
      prob_log = c_log[s-1] - total_log;
      H -= prob * prob_log;
    }

  sum = c[t-1] - c[s-1];
  prob = (double) sum / total;
  if (sum != 0)
    {
      prob_log = log((double) sum) - total_log;
      H -= prob * prob_log;
    }

  return H;
}


/*
 * Returns the entropy induced by the points on the partition
 * <c_0, c_s, c_t>, Q. See line 5 of Algorithm 2 in SOM.
 *
 * Parameters
 *   cumhist : cumulative histogram matrix along P_map
 *   cumhist_log : log(cumhist)
 *   c : c_1, ..., c_p
 *   q : number of rows of cumhist (number of partitions in Q_map)
 *   p : number of cols of cumhist (number of partitions in P_map)
 *   s : s in c_s
 *   t : t in c_t
 */
double hp3q(int **cumhist, double **cumhist_log, int *c, int q, int p, int s, int t)

{
  int i, sum;
  double total, total_log, prob, prob_log, H = 0.0;

  total = (double) c[t-1];
  total_log = log(total);

  for (i=0; i<q; i++)
    {
      prob = (double) cumhist[i][s-1] / total;
      if (prob != 0)
        {
          prob_log = cumhist_log[i][s-1] - total_log;
          H -= prob * prob_log;
        }

      sum = cumhist[i][t-1] - cumhist[i][s-1];
      prob = (double) sum / total;
      if (prob != 0)
        {
          prob_log = log((double) sum) - total_log;
          H -= prob * prob_log;
        }
    }

  return H;
}

/*
 * Returns the entropy induced by the points on the partition <c_s, c_t>
 * and Q. See line 13 of Algorithm 2, SOM.
 *
 * Parameters
 *   cumhist : cumulative histogram matrix along P_map
 *   c : c_1, ..., c_p
 *   q : number of rows of cumhist (number of partitions
 *       in Q_map)
 *   p : number of cols of cumhist (number of partitions
 *       in P_map)
 *   s : s in c_s
 *   t : t in c_t
 */
double hp2q(int **cumhist, int *c, int q, int p, int s, int t)
{
  int i, sum;
  double total, total_log, prob, prob_log, H = 0.0 ;

  if (s == t)
    return 0.0;

  total = (double) (c[t-1] - c[s-1]);
  total_log = log(total);

  for (i=0; i<q; i++)
    {
      sum = cumhist[i][t-1] - cumhist[i][s-1];
      prob = (double) sum / total;
      if (prob != 0)
        {
          prob_log = log((double) sum) - total_log;
          H -= prob * prob_log;
        }
    }

  return H;
}

/*
 * Returns the map Q: D -> {0, ...,q-1}.
 * See Algorithm 3 in SOM.
 *
 * Parameters
 *   dy (IN): y-data sorted in increasing order
 *   n (IN): number of elements of dy
 *   y (IN): an integer greater than 1
 *   Q_map (OUT) : the map Q. Q_map must be a preallocated vector of
 *                 size n
 *   q (OUT) : number of partitions in Q_map. q can be < y
 *
 * Returns
 *   0
 */
int EquipartitionYAxis(double *dy, int n, int y, int *Q_map, int *q)
{
  int i, j, s, h, curr;
  double temp1, temp2;

  double rowsize = (double) n / (double) y;

  i = 0;
  h = 0;
  curr = 0;

  while (i < n)
    {
      s = 1;
      for (j=i+1; j<n; j++)
        {
          if (dy[i] == dy[j])
            ++s;
          else
            break;
        }

      temp1 = fabs((double) h + (double) s - rowsize);
      temp2 = fabs((double) h - rowsize);
      if ((h != 0) && (temp1 >= temp2))
        {
          ++curr;
          h = 0;
          temp1 = (double) n - (double) i;
          temp2 = (double) y - (double) curr;
          rowsize = temp1 / temp2;
        }

      for (j=0; j<s; j++)
         Q_map[i+j] = curr;

      i += s;
      h += s;
    }

  *q = curr + 1;

  return 0;
}

/*
 * Returns the map P: D -> {0, ...,p-1}.
 *
 * Parameters
 *   dx (IN) : x-data sorted in increasing order
 *   n (IN) : number of elements of dx
 *   Q_map (IN) : the map Q computed by EquipartitionYAxis sorted in
 *                increasing order by dx-values
 *   P_map (OUT) : the map P. P_map must be a preallocated vector
 *                 of size n
 *   p (OUT) : number of partitions in P_map
 *
 * Returns
 *   0 on success, 1 if an error occurs
 */
int GetClumpsPartition(double *dx, int n, int *Q_map, int *P_map, int *p)
{
  int i, j, flag, c, s;
  int *Q_tilde;

  i = 0;
  c = -1;

  Q_tilde = (int *) malloc (n * sizeof(int));
  if (Q_tilde == NULL)
    return 1;

  memcpy(Q_tilde, Q_map, n*sizeof(int));

  while (i < n)
    {
      s = 1;
      flag = FALSE;
      for (j=i+1; j<n; j++)
        {
          if (dx[i] == dx[j])
            {
              if (Q_tilde[i] != Q_tilde[j])
                flag = TRUE;
              ++s;
            }
          else
            break;
        }

      if ((s > 1) && (flag == TRUE))
        {
          for (j=0; j<s; j++)
            Q_tilde[i+j] = c;
          --c;
        }

      i += s;
    }

  i = 0;
  P_map[0] = 0;
  for (j=1; j<n; j++)
    {
      if (Q_tilde[j] != Q_tilde[j-1])
        ++i;
      P_map[j] = i;
    }

  *p = i + 1;
  free(Q_tilde);

  return 0;
}

/*
 * Returns the map P: D -> {0, ...,p-1}.
 *
 * Parameters
 *   dx (IN) : x-data sorted in increasing order
 *   n (IN) : number of elements of dx
 *   k_hat (IN) : maximum number of clumps
 *   Q_map (IN) : the map Q computed by EquipartitionYAxis sorted in
 *                increasing order by dx-values
 *   P_map (OUT) : the map P. P_map must be a preallocated vector
 *                 of size n
 *   p (OUT) : number of partitions in P_map
 *
 * Returns
 *   0 on success, 1 if an error occurs
 */
int GetSuperclumpsPartition(double *dx, int n, int k_hat, int *Q_map,
                            int *P_map, int *p)
{
  int i, ret;
  double *dp;

  /* clumps */
  ret = GetClumpsPartition(dx, n, Q_map, P_map, p);
  if (ret)
    return 1;

  /* superclumps */
  if (*p > k_hat)
    {
      dp = (double *) malloc (n * sizeof(double));
      if (dp == NULL)
        return 1;

      for (i=0; i<n; i++)
        dp[i] = (double) P_map[i];

      EquipartitionYAxis(dp, n, k_hat, P_map, p);

      free(dp);
    }

  return 0;
}


/* Returns (c_1, ..., c_k) */
int *compute_c(int *P_map, int p, int n)
{
  int i;
  int *c;


  c = (int *) malloc (p * sizeof(int));
  if (c == NULL)
    return NULL;

  for (i=0; i<p; i++)
    c[i] = 0;

  for (i=0; i<n; i++)
    c[P_map[i]]++;

  for (i=1; i<p; i++)
    c[i] += c[i-1];

  return c;
}


double *compute_c_log(int *c, int p)
{
  int i;
  double *c_log;

  c_log = (double *) malloc (p * sizeof(double));
  if (c_log == NULL)
    return NULL;

  for (i=0; i<p; i++)
    if (c[i] != 0)
      c_log[i] = log((double) c[i]);
    else
      c_log[i] = 0;

  return c_log;
}


/* Returns the cumulative histogram matrix along P_map */
int **compute_cumhist(int *Q_map, int q, int *P_map, int p, int n)
{
  int i, j;
  int **cumhist;


  cumhist = (int **) malloc (q * sizeof(int *));
  if (cumhist == NULL)
    return NULL;

  for (i=0; i<q; i++)
    {
      cumhist[i] = (int *) malloc (p * sizeof(int));
      if (cumhist[i] == NULL)
        {
          for (j=0; j<i; j++)
            free(cumhist[j]);

          free(cumhist);
          return NULL;
        }

      for (j=0; j<p; j++)
        cumhist[i][j] = 0;
    }

  for (i=0; i<n; i++)
    cumhist[Q_map[i]][P_map[i]]++;

  for (i=0; i<q; i++)
    for (j=1; j<p; j++)
      cumhist[i][j] += cumhist[i][j-1];

  return cumhist;
}


double ** compute_cumhist_log(int **cumhist, int q, int p)
{
  int i, j;
  double **cumhist_log;

  cumhist_log = (double **) malloc (q * sizeof(double *));
  if (cumhist_log == NULL)
    return NULL;

  for (i=0; i<q; i++)
    {
      cumhist_log[i] = (double *) malloc (p * sizeof(double));
      if (cumhist_log[i] == NULL)
        {
          for (j=0; j<i; j++)
            free(cumhist_log[j]);

          free(cumhist_log);
          return NULL;
        }

      for (j=0; j<p; j++)
        if (cumhist[i][j] != 0)
          cumhist_log[i][j] = log((double) cumhist[i][j]);
        else
          cumhist_log[i][j] = 0;
    }

    return cumhist_log;
}


/* Initializes the I matrix */
double **init_I(int p, int x)
{
  int i, j;
  double **I;


  I = (double **) malloc ((p+1) * sizeof(double *));
  if (I == NULL)
    return NULL;

  for (i=0; i<=p; i++)
    {
      I[i] = (double *) malloc ((x+1) * sizeof(double));
      if (I[i] == NULL)
        {
          for (j=0; j<i; j++)
            free(I[j]);
          free(I);
          return NULL;
        }

      for (j=0; j<=x; j++)
        I[i][j] = 0.0;
    }

  return I;
}


/* Computes the HP2Q matrix */
double **compute_HP2Q(int **cumhist, int*c, int q, int p)
{
  int i, j, s, t;
  double **HP2Q;


  HP2Q = (double **) malloc ((p+1) * sizeof(double *));
  if (HP2Q == NULL)
    return NULL;

  for (i=0; i<=p; i++)
    {
      HP2Q[i] = (double *) malloc ((p+1) * sizeof(double));
      if (HP2Q[i] == NULL)
        {
          for (j=0; j<i; j++)
            free(HP2Q[j]);
          free(HP2Q);
          return NULL;
        }
    }

  for (t=3; t<=p; t++)
    for (s=2; s<=t; s++)
      HP2Q[s][t] = hp2q(cumhist, c, q, p, s, t);

  return HP2Q;
}


/*
 * Returns the normalized MI scores.
 *
 * Parameters
 *   dx (IN) : x-data sorted in increasing order by dx-values
 *   dy (IN) : y-data sorted in increasing order by dx-values
 *   n (IN) : number of elements in dx and dy
 *   Q_map (IN) : the map Q computed by EquipartitionYAxis() sorted in
 *                increasing order by dx-values
 *   q (IN) : number of partitions in Q_map
 *   P_map (IN) : the map P computed by GetSuperclumpsPartition() sorted
 *                in increasing order by Dx-values
 *   p (IN) : number of partitions in P_map
 *   x (IN) : maximum grid size on dx-values
 *   score (OUT) : mutual information scores. score must be a
 *                 preallocated array of dimension x-1
 * Returns
 *   0 on success, 1 if an error occurs
 */
int OptimizeXAxis(double *dx, double *dy, int n, int *Q_map, int q,
      int *P_map, int p, int x, double *score)
{
  int i, s, t, l;
  int *c;
  int **cumhist;
  double **I, **HP2Q;
  double F, F_max, HQ, ct, cs;
  double **cumhist_log, *c_log;

  /* return score=0 if p=1 */
  if (p == 1)
    {
      for (i=0; i<x-1; i++)
        score[i] = 0.0;
      return 0;
    }

  /* compute c */
  c = compute_c(P_map, p, n);
  if (c == NULL)
    goto error_c;

  /* precompute log(c) (log(c)=0 when c=0) */
  c_log = compute_c_log(c, p);
  if (c_log == NULL)
    goto error_c_log;

  /* compute the cumulative histogram matrix along P_map */
  cumhist = compute_cumhist(Q_map, q, P_map, p, n);
  if (cumhist == NULL)
    goto error_cumhist;

  /* precompute log(cumhist) (log(cumhist)=0 when cumhist=0) */
  cumhist_log = compute_cumhist_log(cumhist, q, p);
  if (cumhist == NULL)
    goto error_cumhist_log;

  /* I matrix initialization */
  I = init_I(p, x);
  if (I == NULL)
    goto error_I;

  /* Precomputes the HP2Q matrix */
  HP2Q = compute_HP2Q(cumhist, c, q, p);
  if (HP2Q == NULL)
    goto error_HP2Q;

  /* compute H(Q) */
  HQ = hq(cumhist, cumhist_log, q, p, n);

  /* Find the optimal partitions of size 2, Algorithm 2 in SOM, lines 3-8 */
  for (t=2; t<=p; t++)
    {
      F_max = -DBL_MAX;
      for (s=1; s<=t; s++)
        {
          F = hp3(c, c_log, s, t) - hp3q(cumhist, cumhist_log, c, q, p, s, t);
          if (F > F_max)
            {
              I[t][2] = HQ + F;
              F_max = F;
            }
        }
    }

  /*
   * Inductively build the rest of the table of optimal partitions,
   * Algorithm 2 in SOM, lines 10-17
   */
  for (l=3; l<=x; l++)
    {
      for (t=l; t<=p; t++)
        {
          ct = (double) c[t-1];
          F_max = -DBL_MAX;
          for (s=l-1; s<=t; s++)
            {
              cs = (double) c[s-1];
              F = ((cs/ct) * (I[s][l-1]-HQ)) - (((ct-cs)/ct) * HP2Q[s][t]);

              if (F > F_max)
                {
                  I[t][l] = HQ + F;
                  F_max = F;
                }
            }
        }
    }

  /* Algorithm 2 in SOM, line 19 */
  for (i=p+1; i<=x; i++)
    I[p][i] = I[p][p];

  /* score */
  for (i=2; i<=x; i++)
    score[i-2] = I[p][i] / MIN(log(i), log(q));

  /* start frees */
  for (i=0; i<=p; i++)
    free(HP2Q[i]);
  free(HP2Q);

  for (i=0; i<=p; i++)
    free(I[i]);
  free(I);

  for (i=0; i<q; i++)
    free(cumhist_log[i]);
  free(cumhist_log);

  for (i=0; i<q; i++)
    free(cumhist[i]);
  free(cumhist);

  free(c_log);
  free (c);
  /* end frees */

  return 0;

  /* gotos */
  error_HP2Q:
    for (i=0; i<=p; i++)
      free(I[i]);
    free(I);
  error_I:
    for (i=0; i<q; i++)
      free(cumhist_log[i]);
    free(cumhist_log);
  error_cumhist_log:
    for (i=0; i<q; i++)
      free(cumhist[i]);
    free(cumhist);
  error_cumhist:
    free(c_log);
  error_c_log:
    free(c);
  error_c:
    return 1;
}


/*
 * Returns an initialized mine_score structure. Returns NULL if an error
 * occurs.
 */
mine_score *init_score(mine_problem *prob, mine_parameter *param)
{
  int i, j;
  double B;
  mine_score *score;

  if ((param->alpha > 0.0) && (param->alpha <= 1.0))
    B = MAX(pow(prob->n, param->alpha), 4);
  else if (param->alpha >= 4)
    B = MIN(param->alpha, prob->n);
  else
    goto error_score;

  score = (mine_score *) malloc (sizeof(mine_score));
  if (score == NULL)
    goto error_score;

  score->n = MAX((int) floor(B/2.0), 2) - 1;
  score->m = (int *) malloc(score->n * sizeof(int));
  if (score->m == NULL)
    goto error_score_m;

  for (i=0; i<score->n; i++)
    score->m[i] = (int) floor((double) B / (double) (i+2)) - 1;

  score->M = (double **) malloc (score->n * sizeof(double *));
  if (score->M == NULL)
    goto error_score_M;

  for (i=0; i<score->n; i++)
    {
      score->M[i] = (double *) malloc ((score->m[i]) * sizeof(double));
      if (score->M[i] == NULL)
        {
          for (j=0; j<i; j++)
            free(score->M[j]);
          goto error_score_M_i;
        }
    }

  return score;

  error_score_M_i:
    free(score->M);
  error_score_M:
    free(score->m);
  error_score_m:
    free(score);
  error_score:
    return NULL;
}


/* See mine.h */
mine_score *mine_compute_score(mine_problem *prob, mine_parameter *param)
{
  int i, j, k, p, q, ret;
  double *xx, *yy, *xy, *yx, *M_temp;
  int *ix, *iy;
  int *Q_map_temp, *Q_map, *P_map;
  mine_score *score;

  score = init_score(prob, param);
  if (score == NULL)
    goto error_score;

  xx = (double *) malloc (prob->n * sizeof(double));
  if (xx == NULL)
    goto error_xx;

  yy = (double *) malloc (prob->n * sizeof(double));
  if (yy == NULL)
    goto error_yy;

  xy = (double *) malloc (prob->n * sizeof(double));
  if (xy == NULL)
    goto error_xy;

  yx = (double *) malloc (prob->n * sizeof(double));
  if (yx == NULL)
    goto error_yx;

  Q_map_temp = (int *) malloc (prob->n * sizeof(int));
  if (Q_map_temp == NULL)
    goto error_Q_temp;

  Q_map = (int *) malloc (prob->n * sizeof(int));
  if (Q_map == NULL)
    goto error_Q;

  P_map = (int *) malloc (prob->n * sizeof(int));
  if (P_map == NULL)
    goto error_P;

  ix = argsort(prob->x, prob->n);
  if (ix == NULL)
    goto error_ix;

  iy = argsort(prob->y, prob->n);
  if (iy == NULL)
    goto error_iy;

  M_temp = (double *)malloc ((score->m[0]) * sizeof(double));
  if (M_temp == NULL)
    goto error_M_temp;

  /* build xx, yy, xy, yx */
  for (i=0; i<prob->n; i++)
    {
      xx[i] = prob->x[ix[i]];
      yy[i] = prob->y[iy[i]];
      xy[i] = prob->x[iy[i]];
      yx[i] = prob->y[ix[i]];
    }

  /* x vs. y */
  for (i=0; i<score->n; i++)
    {
      k = MAX((int) (param->c * (score->m[i]+1)), 1);

      ret = EquipartitionYAxis(yy, prob->n, i+2, Q_map, &q);
      if (ret)
        goto error_0;

      /* sort Q by x */
      for (j=0; j<prob->n; j++)
        Q_map_temp[iy[j]] = Q_map[j];
      for (j=0; j<prob->n; j++)
        Q_map[j] = Q_map_temp[ix[j]];

      ret = GetSuperclumpsPartition(xx, prob->n, k, Q_map, P_map, &p);
      if (ret)
        goto error_0;

      if (param->est == EST_MIC_APPROX)
        ret = OptimizeXAxis(xx, yx, prob->n, Q_map, q, P_map, p, score->m[i]+1,
                            score->M[i]);
      else /* EST_MIC_E */
        ret = OptimizeXAxis(xx, yx, prob->n, Q_map, q, P_map, p,
                            MIN(i+2, score->m[i]+1), score->M[i]);
      if (ret)
        goto error_0;
    }

  /* y vs. x */
  for (i=0; i<score->n; i++)
    {
      k = MAX((int) (param->c * (score->m[i]+1)), 1);

      ret = EquipartitionYAxis(xx, prob->n, i+2, Q_map, &q);
      if (ret)
        goto error_0;

      /* sort Q by y */
      for (j=0; j<prob->n; j++)
        Q_map_temp[ix[j]] = Q_map[j];
      for (j=0; j<prob->n; j++)
        Q_map[j] = Q_map_temp[iy[j]];

      ret = GetSuperclumpsPartition(yy, prob->n, k, Q_map, P_map, &p);
      if (ret)
        goto error_0;

      if (param->est == EST_MIC_APPROX)
        ret = OptimizeXAxis(yy, xy, prob->n, Q_map, q, P_map, p, score->m[i]+1,
                            M_temp);
      else /* EST_MIC_E */
        ret = OptimizeXAxis(yy, xy, prob->n, Q_map, q, P_map, p,
                            MIN(i+2, score->m[i]+1), M_temp);

      if (ret)
        goto error_0;

      if (param->est == EST_MIC_APPROX)
        for (j=0; j<score->m[i]; j++)
          score->M[j][i] = MAX(M_temp[j], score->M[j][i]);
      else /* EST_MIC_E */
        for (j=0; j<MIN(i+1, score->m[i]); j++)
          score->M[j][i] = M_temp[j];
    }

  free(M_temp);
  free(iy);
  free(ix);
  free(P_map);
  free(Q_map);
  free(Q_map_temp);
  free(yx);
  free(xy);
  free(yy);
  free(xx);

  return score;

  error_0:
    free(M_temp);
  error_M_temp:
    free(iy);
  error_iy:
    free(ix);
  error_ix:
    free(P_map);
  error_P:
    free(Q_map);
  error_Q:
    free(Q_map_temp);
  error_Q_temp:
    free(yx);
  error_yx:
    free(xy);
  error_xy:
    free(yy);
  error_yy:
    free(xx);
  error_xx:
    for (i=0; i<score->n; i++)
      free(score->M[i]);
    free(score->M);
    free(score->m);
    free(score);
  error_score:
    return NULL;
}


/* See mine.h */
char *mine_check_parameter(mine_parameter *param)
{
  if (((param->alpha <= 0.0) || (param->alpha > 1.0)) && (param->alpha < 4.0))
    return "alpha must be in (0.0, 1.0] or >= 4.0";

  if (param->c <= 0.0)
    return "c must be > 0.0";

  if ((param->est != EST_MIC_APPROX) && (param->est != EST_MIC_E))
    return "unknown estimator";

  return NULL;
}


/* See mine.h */
double mine_mic(mine_score *score)
{
  int i, j;
  double score_max = 0.0;

  for (i=0; i<score->n; i++)
    for (j=0; j<score->m[i]; j++)
      if (score->M[i][j] > score_max)
        score_max = score->M[i][j];

  return score_max;
}


/* See mine.h */
double mine_mas(mine_score *score)
{
  int i, j;
  double score_curr;
  double score_max = 0.0;

  for (i=0; i<score->n; i++)
    for (j=0; j<score->m[i]; j++)
      {
        score_curr = fabs(score->M[i][j] - score->M[j][i]);
        if (score_curr > score_max)
          score_max = score_curr;
      }

  return score_max;
}


/* See mine.h */
double mine_mev(mine_score *score)
{
  int i, j;
  double score_max = 0.0;

  for (i=0; i<score->n; i++)
    for (j=0; j<score->m[i]; j++)
      if (((j==0) || (i==0)) && score->M[i][j] > score_max)
        score_max = score->M[i][j];

  return score_max;
}


/* See mine.h */
double mine_mcn(mine_score *score, double eps)
{
  int i, j;
  double log_xy;
  double score_min = DBL_MAX;
  double delta = 0.0001; /* avoids overestimation of mcn */
  double mic = mine_mic(score);

  for (i=0; i<score->n; i++)
    for (j=0; j<score->m[i]; j++)
      {
        log_xy = log((i+2) * (j+2)) / log(2.0);
        if (((score->M[i][j]+delta) >= ((1.0 - eps) * mic))
             && (log_xy < score_min))
          score_min = log_xy;
      }

  return score_min;
}


/* See mine.h */
double mine_mcn_general(mine_score *score)
{
  int i, j;
  double log_xy;
  double score_min = DBL_MAX;
  double delta = 0.0001; /* avoids overestimation of mcn */
  double mic = mine_mic(score);

  for (i=0; i<score->n; i++)
    for (j=0; j<score->m[i]; j++)
      {
        log_xy = log((i+2) * (j+2)) / log(2.0);
        if (((score->M[i][j]+delta) >= (mic * mic)) && (log_xy < score_min))
          score_min = log_xy;
      }

  return score_min;
}

/* See mine.h */
double mine_tic(mine_score *score, int norm)
{
  int i, j, k=0;
  double tic = 0.0;

  for (i=0; i<score->n; i++)
    for (j=0; j<score->m[i]; j++)
      {
        tic += score->M[i][j];
        k++;
      }

  if (norm)
    tic /= k;

  return tic;
}

/* See mine.h */
double mine_gmic(mine_score *score, double p)
{
  int i, j, k, Z, B;
  mine_score *score_sub, *C_star;
  double gmic;

  /* alloc score_sub */
  score_sub = (mine_score *) malloc (sizeof(mine_score));

  /* alloc C_star */
  C_star = (mine_score *) malloc (sizeof(mine_score));
  C_star->m = (int *) malloc(score->n * sizeof(int));
  C_star->M = (double **) malloc (score->n * sizeof(double *));
  for (i=0; i<score->n; i++)
    C_star->M[i] = (double *) malloc ((score->m[i]) * sizeof(double));

  /* prepare score_sub */
  score_sub->M = score->M;

  /* prepare C_star */
  C_star->n = score->n;
  for (i=0; i<C_star->n; i++)
    C_star->m[i] = score->m[i];

  /* compute C_star */
  for (i=0; i<score->n; i++)
    for (j=0; j<score->m[i]; j++)
      {
        B = (i+2) * (j+2);
        score_sub->n = MAX((int) floor(B/2.0), 2) - 1;
        score_sub->m = (int *) malloc(score_sub->n * sizeof(int));
        for (k=0; k<score_sub->n; k++)
          score_sub->m[k] = (int) floor((double) B / (double) (k+2)) - 1;

        C_star->M[i][j] = mine_mic(score_sub);
        free(score_sub->m);
      }

  /* p=0 -> geometric mean */
  if (p == 0.0)
    {
      Z = 0;
      gmic = 1.0;
      for (i=0; i<C_star->n; i++)
        for (j=0; j<C_star->m[i]; j++)
          {
            gmic *= C_star->M[i][j];
            Z++;
          }
      gmic = pow(gmic, (double) Z);
    }
  /* p!=0 -> generalized mean */
  else
    {
      Z = 0;
      gmic = 0.0;
      for (i=0; i<C_star->n; i++)
        for (j=0; j<C_star->m[i]; j++)
          {
            gmic += pow(C_star->M[i][j], p);
            Z++;
          }
      gmic /= (double) Z;
      gmic = pow(gmic, 1.0/p);
    }

  free(score_sub);
  if (C_star->n != 0)
    {
      free(C_star->m);
      for (i=0; i<C_star->n; i++)
         free(C_star->M[i]);
      free(C_star->M);
    }
  free(C_star);

  return gmic;
}

/* See mine.h */
void mine_free_score(mine_score **score)
{
  int i;
  mine_score *score_ptr = *score;

  if (score_ptr != NULL)
    {
      if (score_ptr->n != 0)
        {
          free(score_ptr->m);
          for (i=0; i<score_ptr->n; i++)
            free(score_ptr->M[i]);
          free(score_ptr->M);
        }

      free(score_ptr);
      score_ptr = NULL;
    }
}


mine_pstats *mine_compute_pstats(mine_matrix *X, mine_parameter *param)
{
  int i, j, k;
  mine_problem prob;
  mine_score *score;
  mine_pstats *stats;

  /* Allocate memory for stats */
  stats = (mine_pstats *) malloc(sizeof(mine_pstats));
  stats->n = (X->n * (X->n-1)) / 2;
  stats->mic = (double *) malloc(stats->n * sizeof(double));
  stats->tic = (double *) malloc(stats->n * sizeof(double));

  k = 0;
  prob.n = X->m;
  for (i=0; i<X->n-1; i++)
    {
      prob.x = &X->data[i*X->m];
      for (j=i+1; j<X->n; j++)
        {
          prob.y = &X->data[j*X->m];
          score = mine_compute_score(&prob, param);
          stats->mic[k] = mine_mic(score);
          stats->tic[k] = mine_tic(score, TRUE);
          mine_free_score(&score);
          k++;
        }
    }

  return stats;
}


mine_cstats *mine_compute_cstats(mine_matrix *X, mine_matrix *Y,
                                 mine_parameter *param)
{
  int i, j, k;
  mine_cstats *stats;
  mine_problem prob;
  mine_score *score;


  if (X->m != Y->m)
    return NULL;

  /* Allocate memory for stats */
  stats = (mine_cstats *) malloc(sizeof(mine_cstats));
  stats->n = X->n;
  stats->m = Y->n;
  stats->mic = (double *) malloc((stats->n * stats->m) * sizeof(double));
  stats->tic = (double *) malloc((stats->n * stats->m) * sizeof(double));

  k = 0;
  prob.n = X->m;
  for (i=0; i<X->n; i++)
    {
      prob.x = &X->data[i*X->m];

      for (j=0; j<Y->n; j++)
        {
          prob.y = &Y->data[j*Y->m];
          score = mine_compute_score(&prob, param);
          stats->mic[k] = mine_mic(score);
          stats->tic[k] = mine_tic(score, TRUE);
          mine_free_score(&score);
          k++;
        }
    }

  return stats;
}