mat concentrate_step(mat x, umat x_nonmiss, vec pu, vec pmd, uvec x_miss_group_match,
umat miss_group_unique, uvec miss_group_counts,
umat miss_group_obs_col, umat miss_group_mis_col, uvec miss_group_p, int miss_group_n,
int n, int n_half, int p, vec theta0, mat G, int d, int EM_maxits,
double* xh_mem, unsigned int* misgrpuqh_mem, unsigned int* msgrpcth_mem,
unsigned int* msgrpoch_mem, unsigned int* msgrpmch_mem, unsigned int* msgrpph_mem)
{
// information required for subsampling
mat x_half(xh_mem, n_half, p, false, true);
umat miss_grp_uq_half( misgrpuqh_mem, miss_group_n, p, false, true); miss_grp_uq_half.zeros();
uvec miss_grp_cts_half( msgrpcth_mem, miss_group_n, false, true); miss_grp_cts_half.zeros();
umat miss_grp_oc_half( msgrpoch_mem, miss_group_n, p, false, true); miss_grp_oc_half.zeros();
umat miss_grp_mc_half( msgrpmch_mem, miss_group_n, p, false, true); miss_grp_mc_half.zeros();
uvec miss_grp_p_half( msgrpph_mem, n_half, false, true);miss_grp_p_half.zeros();
int miss_grp_n_half;
// compute adjusted pmd
vec pmd_adj(n);
for(int i = 0; i < n; i++) pmd_adj(i) = R::pchisq(pmd(i), (double) pu(i), 1, 0);
double pmd_adj_med = median(pmd_adj);
uvec x_half_ind = find( pmd_adj <= pmd_adj_med, n_half);
// find new missing pattern based on the halved samples
int miss_new_add = -1;
for(int i = 0; i < n_half; i++){
int index = (int) x_half_ind(i);
x_half.row(i) = x.row( index );
// check whether we found a new missing pattern
urowvec check_new_miss(p);
if( i == 0 ){ check_new_miss.zeros();
} else{ check_new_miss = (x_nonmiss.row( index ) == miss_grp_uq_half.row( miss_new_add )); }
if( sum(check_new_miss)== (unsigned int) p){
// if found same missing pattern
miss_grp_cts_half( miss_new_add )++;
}else{
// if found a new missing pattern
miss_grp_uq_half.row( miss_new_add + 1) = x_nonmiss.row( index );
miss_grp_cts_half( miss_new_add + 1)++;
miss_grp_oc_half.row( miss_new_add + 1) = miss_group_obs_col.row( x_miss_group_match( index)-1 );
miss_grp_mc_half.row( miss_new_add + 1) = miss_group_mis_col.row( x_miss_group_match( index)-1 );
miss_grp_p_half( miss_new_add + 1) = miss_group_p( x_miss_group_match( index)-1);
miss_new_add++;
}
}
miss_grp_n_half = miss_new_add + 1;
mat res = CovEM(x_half, n_half, p, theta0, G, d, miss_grp_uq_half, miss_grp_cts_half,
miss_grp_oc_half, miss_grp_mc_half, miss_grp_p_half, miss_grp_n_half,
0.0001, EM_maxits);
res.shed_rows(0,1);
return res;
}
void subsampling(double* subsample_mem, unsigned int* subsamp_nonmis_mem,
mat x, umat x_nonmiss, int nSubsampleSize, int p, uvec subsample_id)
{
mat subsamp( subsample_mem, nSubsampleSize, p, false, true);
umat subsamp_nonmiss( subsamp_nonmis_mem, nSubsampleSize, p, false, true);
uvec subsample_id_shuff = shuffle(subsample_id);
for(int i = 0; i < nSubsampleSize; i++)
{
subsamp.row(i) = x.row( subsample_id_shuff(i) );
subsamp_nonmiss.row(i) = x_nonmiss.row( subsample_id_shuff(i) );
}
}