// tri-diagonalize a symmetric matrix. The matrix will be destroyed and
// the diagonal will be returned in d and off diagonal in e. It uses
// the Householder transformation
// WARNING: allocates space
void Matrix::tridiagonalize(double *&d, double *&e)
{
d = new double [maxc]; // the diagonal elements
e = new double [maxc]; // the off-diagonal elements
householder(m, maxc, d, e);
}
void main()
{
double sum,s,c,r;
int i,j,k,n;
scanf("%d",&n);
double w[MAX], v[MAX], q[MAX];
double tridiagonal[MAX][MAX];
tridiagonal[0][0]=1.0;
for (i = 1;i<=n;i++)
for (j=1;j<=n;j++)
scanf("%lf",&tridiagonal[i][j]);
for (k=1;k<=n-2;k++)
{
sum = 0;
for (j = k+1; j <= n; j++)
sum += tridiagonal[j][k] * tridiagonal[j][k];
s=sqrt(sum);
if ( tridiagonal[k+1][k] < 0 )
s=-s;
r = sqrt( 2.0 * tridiagonal[k+1][k] * s + 2.0 * s * s );
for (j = 1; j <= k; j++)
w[j] = 0.0;
w[k+1] = ( tridiagonal[k+1][k] + s ) / r;
for (j=k+2;j<=n;j++)
w[j] = tridiagonal[j][k] / r;
for (i=1;i<=k;i++)
v[i] = 0.0;
for (i = k+1; i <= n; i++)
{
sum = 0.0;
for (j = k+1; j <= n; j++)
sum += tridiagonal[i][j] * w[j];
v[i] = sum;
}
c = 0;
for (j = k+1; j <= n; j++)
c += w[j] * v[j];
for (j = 1; j <= k; j++)
q[j] = 0.0;
for (j = k+1; j <= n; j++)
q[j] = v[j] - c * w[j];
for (j=k+2;j<=n;j++)
tridiagonal[j][k] = tridiagonal[k][j] = 0.0;
tridiagonal[k+1][k] = tridiagonal[k][k+1] = -s;
for (j = k; j <= n; j++)
tridiagonal[j][j] -= 4.0 * q[j] * w[j];
for (i = k+1; i <= n; i++)
{
for (j = i+1; j <= n; j++)
{
tridiagonal[i][j] = tridiagonal[i][j] - 2.0 * w[i] * q[j] - 2.0 * q[i] * w[j];
tridiagonal[j][i] = tridiagonal[i][j];
}
}
}
print(tridiagonal, n);
householder(tridiagonal,n);
eigen(tridiagonal,n);
}
void test_householder()
{
for(int i = 0; i < g_repeat; i++) {
CALL_SUBTEST_1( householder(Matrix<double,2,2>()) );
CALL_SUBTEST_2( householder(Matrix<float,2,3>()) );
CALL_SUBTEST_3( householder(Matrix<double,3,5>()) );
CALL_SUBTEST_4( householder(Matrix<float,4,4>()) );
CALL_SUBTEST_5( householder(MatrixXd(internal::random<int>(1,EIGEN_TEST_MAX_SIZE),internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
CALL_SUBTEST_6( householder(MatrixXcf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE),internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
CALL_SUBTEST_7( householder(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE),internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
CALL_SUBTEST_8( householder(Matrix<double,1,1>()) );
}
}
void partlyKnownDifferencesInPhaseConstraints(int M, int K, cv::Mat& Q2)
{
//constraints equation: PartlyKnownDifferencesInPhase
cv::Mat ce = cv::Mat::eye(M, M, cv::DataType<double>::type);
for(int i = 1; i < K-1; ++i)
{
cv::vconcat(ce, cv::Mat::eye(M, M, cv::DataType<double>::type), ce);
}
cv::hconcat(ce, -1 * cv::Mat::eye((K-1)*M, (K-1)*M, cv::DataType<double>::type), ce);
ce.row(0) = cv::abs(ce.row(0));
ce.row(1) = cv::abs(ce.row(1));
ce.row(2) = cv::abs(ce.row(2));
cv::Mat Q, R;
householder(ce.t(), Q, R);
// extracts A columns, 1 (inclusive) to 3 (exclusive).
//Mat B = A(Range::all(), Range(1, 3));
//Select last Np colums of Q to become the constraints null space
int Np = (M*K) - ce.rows;
Q2 = Q(cv::Range::all(), cv::Range(Q.cols - Np, Q.cols ));
}
matrix tridiagonalize(matrix A){
/*
Tridiagonalize a symmetric matrix using Householder transformations.
Transformations are performed in place with Householder vectors and coefficients,
rather than explicitly forming matrices and explicitly computing their products.
Returns T = Q^t * A * Q
Source: Golub and Van Loan, "Matrix Computations" p. 415
Input:
Matrix A Matrix to tridiagonalize.
Output:
Matrix T (returned) Tridiagonal matrix similar to input matrix.
*/
double *v = (double *) malloc(A.n * sizeof(double)) ;
double *p = (double *) malloc(A.n * sizeof(double)) ;
double *w = (double *) malloc(A.n * sizeof(double)) ;
double beta ;
double temp ;
double t1, t2;
int currentVectorLength = A.n - 1 ;
int n = A.n ;
int iVector, jVector;
int i,j,k;
for(k=0; k < n-2; k++){
beta = householder(A, k+1, k, v) ;
matrixVectorMultiply(A, k+1, k+1, v, beta, p) ;
temp = - 0.5 * beta * innerProduct(p, v, currentVectorLength ) ;
vectorPlusConstantByVector(w, p, temp, v, currentVectorLength ) ;
temp = 0.0;
for(j=k+1; j<n; j++){
temp += A.values[j][k] * A.values[j][k] ;
}
temp = sqrt(temp) ;
A.values[k+1][k] = temp ;
A.values[k][k+1] = temp ;
for(i=k+1, iVector=0; i<n; i++, iVector++){
for(j=k+1, jVector=0; j<n; j++, jVector++){
t1 = v[i]*w[j] ;
t2 = v[j]*w[i] ;
A.values[i][j] -= (v[iVector]*w[jVector] + v[jVector]*w[iVector]) ;
}
}
currentVectorLength-- ;
}
// zero off diagonal multipliers
for(i=0; i<n-1; i++){
for(j=i+2; j<n; j++){
A.values[i][j] = 0.0 ;
A.values[j][i] = 0.0 ;
}
}
free(v) ;
free(p) ;
free(w) ;
return A;
}
/*
* the eigvalue, 有可能在数学上出问题（虚数） Error
*/
Matrix* calc_eig(Matrix* p)
{
Matrix *oldAns = ans, *oldAns2;
Matrix *temp = NULL,*temp2 = NULL;
Matrix *v = NULL;
double r1;
int i, j, k = 0;
if (p == NULL)
{
//Error
return NULL;
}
if (p->m != p->n)
{
//Error
return NULL;
}
if (!stor_createMatrix(&v, p->m, 1))
{
//Error
return NULL;
}
if (!stor_createMatrix(&temp, p->m, p->n))
{
//Error
return NULL;
}
if (!stor_createMatrix(&temp2, p->m, p->n))
{
//Error
return NULL;
}
if (!stor_assign(temp, p))
{
//Error
return NULL;
}
for (i = 0; i < p->n - 2; i++)
{
for (j = 0; j < p->m; j++)
{
*stor_entry(v, j, 0) = *stor_entry(temp, j, i);
}
if (!householder(v, i + 1))
{
//Error
return NULL;
}
oldAns2 = ans;
ans = NULL;
if (!calc_mul(calc_mul(oldAns2, temp), oldAns2)) //H*A*H
{
//Error
return NULL;
}
stor_freeMatrix(temp);
temp = ans;
ans = NULL;
stor_freeMatrix(oldAns2);
}
//temp is 拟三角阵
while (k<=500 && !isUpTrian(temp))//最多500次迭代
{
k++;
ans = NULL;
if (!(temp2 = calc_eye(p->n)))
{
//Error
return NULL;
}
ans = NULL;
for (i = 0; i < p->n - 1; i++)
{
if (!calc_eye(p->n))
{
//Error
return NULL;
}
r1 = sqrt(*stor_entry(temp, i, i) * *stor_entry(temp, i, i)
+ *stor_entry(temp, i + 1, i)* *stor_entry(temp, i + 1, i));
if (util_isZero(r1))
{
//Error
return NULL;
}
*stor_entry(ans, i, i) = *stor_entry(temp, i, i) / r1;
*stor_entry(ans, i + 1, i + 1) = *stor_entry(temp, i, i) / r1;
*stor_entry(ans, i + 1, i) = -*stor_entry(temp, i + 1, i) / r1;
*stor_entry(ans, i, i + 1) = *stor_entry(temp, i + 1, i) / r1;
oldAns2 = ans;
ans = NULL;
if (!calc_mul(temp2, calc_trans(oldAns2)))//Q
{
//Error
return NULL;
}
stor_freeMatrix(temp2);
temp2 = ans;
ans = NULL;
if (!calc_mul(oldAns2, temp))//R
{
//Error
return NULL;
}
stor_freeMatrix(temp);
temp = ans;
ans = NULL;
stor_freeMatrix(oldAns2);
}
if (!calc_mul(temp, temp2))
{
//Error
return NULL;
}
stor_freeMatrix(temp);
temp = ans;
}
//优化迭代， 好像可以证明RQ一定是拟上三角阵
for (i = 0; i < p->m; i++)
{
for (j = 0; j < p->n; j++)
{
if (i != j) *stor_entry(ans, i, j) = 0;
}
}
return ans;
}
/*----------------------------------------------------------------------------*/
void NOMAD::Directions::compute ( std::list<NOMAD::Direction> & dirs ,
NOMAD::poll_type poll ,
const NOMAD::OrthogonalMesh & mesh )
{
dirs.clear();
// categorical variables: do nothing:
if ( _is_categorical )
return;
// binary variables: we use special directions:
if ( _is_binary )
{
compute_binary_directions ( dirs );
return;
}
NOMAD::Double pow_gps , cst;
const NOMAD::Direction * bl;
NOMAD::Direction * pd;
int i , j , k , hat_i;
std::list<NOMAD::Direction>::iterator it2 , end2;
const NOMAD::Point mesh_indices=mesh.get_mesh_indices();
int mesh_index=int(mesh_indices[0].value()); // single mesh_index used only for isotropic mesh
/*-----------------------------------*/
/* loop on the types of directions */
/*-----------------------------------*/
const std::set<NOMAD::direction_type> & dir_types =
(poll == NOMAD::PRIMARY) ? _direction_types : _sec_poll_dir_types;
std::set<NOMAD::direction_type>::const_iterator it , end = dir_types.end() ;
for ( it = dir_types.begin() ; it != end ; ++it )
{
if ( *it == NOMAD::UNDEFINED_DIRECTION ||
*it == NOMAD::NO_DIRECTION ||
*it == NOMAD::MODEL_SEARCH_DIR )
continue;
/*--------------------------------------------------*/
/* Ortho-MADS */
/* ---> dirs are on a unit n-sphere by construction */
/*--------------------------------------------------*/
if ( NOMAD::dir_is_orthomads (*it) )
{
bool success_dir;
NOMAD::Direction dir ( _nc , 0.0 , *it );
NOMAD::Double alpha_t_l;
success_dir=compute_dir_on_unit_sphere ( dir );
if ( success_dir )
{
#ifdef DEBUG
if ( dynamic_cast<const NOMAD::XMesh*> ( &mesh ) )
{
_out << std::endl
<< NOMAD::open_block ( "compute Ortho-MADS directions with XMesh" )
<< "type = " << *it << std::endl;
NOMAD::Point mesh_indices=mesh.get_mesh_indices();
_out<< "Mesh indices = ( ";
mesh_indices.NOMAD::Point::display( _out, " " , -1 , -1 );
_out << " )" << std::endl;
_out<< "Unit sphere Direction = ( ";
dir.NOMAD::Point::display ( _out , " " , -1 , -1 );
_out << " )" << std::endl << NOMAD::close_block();
}
else if ( dynamic_cast<const NOMAD::SMesh*> ( &mesh ) )
{
_out << std::endl
<< NOMAD::open_block ( "compute Ortho-MADS directions with SMesh" )
<< "type = " << *it
<< " on isotropic mesh " << std::endl
<< "mesh index (lk) = " << mesh_index << std::endl
<< "alpha (tk,lk) = " << alpha_t_l << std::endl
<< "Unit sphere Direction = ( ";
dir.NOMAD::Point::display ( _out , " " , -1 , -1 );
_out << " )" << std::endl << NOMAD::close_block();
}
#endif
// Ortho-MADS 2n and n+1:
// ----------------------
if ( *it == NOMAD::ORTHO_2N || *it == NOMAD::ORTHO_NP1_QUAD || *it == NOMAD::ORTHO_NP1_NEG )
{
// creation of the 2n directions:
// For ORTHO_NP1 the reduction from 2N to N+1 is done in MADS::POLL
NOMAD::Direction ** H = new NOMAD::Direction * [2*_nc];
// Ordering D_k alternates Hk and -Hk instead of [H_k -H_k]
for ( i = 0 ; i < _nc ; ++i )
{
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , *it ) );
H[i ] = &(*(--dirs.end()));
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , *it ) );
H[i+_nc] = &(*(--dirs.end()));
}
// Householder transformations on the 2n directions on a unit n-sphere:
householder ( dir , true , H );
delete [] H;
}
// Ortho-MADS 1 or Ortho-MADS 2:
// -----------------------------
else
{
dirs.push_back ( dir );
if ( *it == NOMAD::ORTHO_2 )
dirs.push_back ( -dir );
}
}
}
/*--------------------------------*/
/* LT-MADS */
/* ---> dirs NOT on unit n-sphere */
/*--------------------------------*/
else if ( NOMAD::dir_is_ltmads (*it) )
{
if ( !_lt_initialized)
lt_mads_init();
bl = get_bl ( mesh , *it , hat_i );
// LT-MADS 1 or LT-MADS 2: -b(l) and/or b(l):
// ------------------------------------------
if ( *it == NOMAD::LT_1 || *it == NOMAD::LT_2 )
{
dirs.push_back ( - *bl );
if ( *it == NOMAD::LT_2 )
dirs.push_back ( *bl );
}
// LT-MADS 2n or LT-MADS n+1:
// --------------------------
else {
// row permutation vector:
int * row_permutation_vector = new int [_nc];
row_permutation_vector[hat_i] = hat_i;
NOMAD::Random_Pickup rp ( _nc );
for ( i = 0 ; i < _nc ; ++i )
if ( i != hat_i )
{
j = rp.pickup();
if ( j == hat_i )
j = rp.pickup();
row_permutation_vector[i] = j;
}
rp.reset();
for ( j = 0 ; j < _nc ; ++j )
{
i = rp.pickup();
if ( i != hat_i )
{
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , *it ) );
pd = &(*(--dirs.end()));
create_lt_direction ( mesh , *it , i , hat_i , pd );
permute_coords ( *pd , row_permutation_vector );
}
else
dirs.push_back ( *bl );
// completion to maximal basis:
if ( *it == NOMAD::LT_2N )
dirs.push_back ( NOMAD::Direction ( - *(--dirs.end()) , NOMAD::LT_2N ) );
}
delete [] row_permutation_vector;
// completion to minimal basis:
if ( *it == NOMAD::LT_NP1 )
{
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::LT_NP1 ) );
pd = &(*(--dirs.end()));
end2 = --dirs.end();
for ( it2 = dirs.begin() ; it2 != end2 ; ++it2 )
{
for ( i = 0 ; i < _nc ; ++i )
(*pd)[i] -= (*it2)[i];
}
}
}
}
/*--------------------------------*/
/* GPS */
/* ---> dirs NOT on unit n-sphere */
/*--------------------------------*/
else
{
// GPS for binary variables: should'nt be here:
if ( *it == NOMAD::GPS_BINARY )
{
#ifdef DEBUG
_out << std::endl << "Warning (" << "Directions.cpp" << ", " << __LINE__
<< "): tried to generate binary directions at the wrong place)"
<< std::endl << std::endl;
#endif
dirs.clear();
return;
}
// this value represents the non-zero values in GPS directions
// (it is tau^{|ell_k|/2}, and it is such that delta_k * pow_gps = Delta_k):
if ( !pow_gps.is_defined() )
pow_gps = pow ( mesh.get_update_basis() , abs(mesh_index) / 2.0 );
// GPS 2n, static:
// ---------------
if ( *it == NOMAD::GPS_2N_STATIC )
{
for ( i = 0 ; i < _nc ; ++i )
{
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::GPS_2N_STATIC ) );
pd = &(*(--dirs.end()));
(*pd)[i] = pow_gps;
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::GPS_2N_STATIC ) );
pd = &(*(--dirs.end()));
(*pd)[i] = -pow_gps;
}
}
// GPS 2n, random:
// ---------------
else if ( *it == NOMAD::GPS_2N_RAND )
{
int v , cnt;
std::list <int>::const_iterator end3;
std::list <int>::iterator it3;
std::list <int> rem_cols;
std::vector<int> rem_col_by_row ( _nc );
// creation of the 2n directions:
std::vector<NOMAD::Direction *> pdirs ( 2 * _nc );
for ( i = 0 ; i < _nc ; ++i )
{
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::GPS_2N_RAND ) );
pdirs[i] = &(*(--dirs.end()));
(*pdirs[i])[i] = pow_gps;
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::GPS_2N_RAND ) );
pdirs[i+_nc] = &(*(--dirs.end()));
(*pdirs[i+_nc])[i] = -pow_gps;
rem_cols.push_back(i);
rem_col_by_row[i] = i;
}
// random perturbations:
for ( i = 1 ; i < _nc ; ++i )
{
if ( rem_col_by_row[i] > 0 )
{
v = NOMAD::RNG::rand()%3 - 1; // v in { -1 , 0 , 1 }
if ( v )
{
// we decide a (i,j) couple:
k = NOMAD::RNG::rand()%(rem_col_by_row[i]);
cnt = 0;
end3 = rem_cols.end();
it3 = rem_cols.begin();
while ( cnt != k )
{
++it3;
++cnt;
}
j = *it3;
// the perturbation:
(*pdirs[i])[j] = (*pdirs[j+_nc])[i] = -v * pow_gps;
(*pdirs[j])[i] = (*pdirs[i+_nc])[j] = v * pow_gps;
// updates:
rem_cols.erase(it3);
it3 = rem_cols.begin();
while ( *it3 != i )
++it3;
rem_cols.erase(it3);
for ( k = i+1 ; k < _nc ; ++k )
rem_col_by_row[k] -= j<k ? 2 : 1;
}
}
}
}
// GPS n+1, static:
// ----------------
else if ( *it == NOMAD::GPS_NP1_STATIC )
{
// (n+1)^st direction:
dirs.push_back ( NOMAD::Direction ( _nc , -pow_gps , NOMAD::GPS_NP1_STATIC ) );
// first n directions:
for ( i = 0 ; i < _nc ; ++i )
{
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::GPS_NP1_STATIC ) );
pd = &(*(--dirs.end()));
(*pd)[i] = pow_gps;
}
}
// GPS n+1, random:
// ----------------
else if ( *it == NOMAD::GPS_NP1_RAND )
{
// (n+1)^st direction:
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::GPS_NP1_RAND ) );
NOMAD::Direction * pd1 = &(*(--dirs.end()));
NOMAD::Double d;
// first n directions:
for ( i = 0 ; i < _nc ; ++i )
{
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::GPS_NP1_RAND ) );
pd = &(*(--dirs.end()));
d = NOMAD::RNG::rand()%2 ? -pow_gps : pow_gps;
(*pd )[i] = d;
(*pd1)[i] = -d;
}
}
// GPS n+1, static, uniform angles:
// --------------------------------
else if ( *it == NOMAD::GPS_NP1_STATIC_UNIFORM )
{
cst = pow_gps * sqrt(static_cast<double>(_nc)*(_nc+1))/_nc;
// n first directions:
for ( j = _nc-1 ; j >= 0 ; --j )
{
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::GPS_NP1_STATIC_UNIFORM ) );
pd = &(*(--dirs.end()));
k = _nc-j-1;
for ( i = 0 ; i < k ; ++i )
(*pd)[i] = -cst / sqrt(static_cast<double>(_nc-i)*(_nc-i+1));
(*pd)[k] = (cst * (j+1)) / sqrt(static_cast<double>(j+1)*(j+2));
}
// (n+1)^st direction:
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::GPS_NP1_STATIC_UNIFORM ) );
pd = &(*(--dirs.end()));
for ( i = 0 ; i < _nc ; ++i )
(*pd)[i] = -cst / sqrt(static_cast<double>(_nc-i)*(_nc-i+1));
}
// GPS n+1, random, uniform angles:
// --------------------------------
// (we apply the procedure defined in
// "Pattern Search Methods for user-provided points:
// application to molecular geometry problems",
// by Alberto, Nogueira, Rocha and Vicente,
// SIOPT 14-4, 1216-1236, 2004, doi:10.1137/S1052623400377955)
else if ( *it == NOMAD::GPS_NP1_RAND_UNIFORM )
{
cst = pow_gps * sqrt(static_cast<double>(_nc)*(_nc+1))/_nc;
// n first directions:
for ( j = _nc-1 ; j >= 0 ; --j )
{
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::GPS_NP1_STATIC_UNIFORM ) );
pd = &(*(--dirs.end()));
k = _nc-j-1;
for ( i = 0 ; i < k ; ++i )
(*pd)[i] = -cst / sqrt(static_cast<double>(_nc-i)*(_nc-i+1));
(*pd)[k] = (cst * (j+1)) / sqrt(static_cast<double>(j+1)*(j+2));
}
// (n+1)^st direction:
dirs.push_back ( NOMAD::Direction ( _nc , 0.0 , NOMAD::GPS_NP1_STATIC_UNIFORM ) );
pd = &(*(--dirs.end()));
for ( i = 0 ; i < _nc ; ++i )
(*pd)[i] = -cst / sqrt(static_cast<double>(_nc-i)*(_nc-i+1));
// random perturbations:
NOMAD::Random_Pickup rp ( _nc );
std::vector<bool> done ( _nc );
bool chg_sgn;
std::list<NOMAD::Direction>::iterator it;
NOMAD::Double tmp;
end2 = dirs.end();
for ( i = 0 ; i < _nc ; ++i )
done[i] = false;
for ( i = 0 ; i < _nc ; ++i )
{
k = rp.pickup();
if ( i != k && !done[i] )
{
chg_sgn = ( NOMAD::RNG::rand()%2 != 0 );
for ( it = dirs.begin() ; it != end2 ; ++it )
{
tmp = (*it)[i];
(*it)[i] = ( chg_sgn ? -1.0 : 1.0 ) * (*it)[k];
(*it)[k] = tmp;
}
done[i] = done[k] = true;
}
else
if ( NOMAD::RNG::rand()%2 )
for ( it = dirs.begin() ; it != end2 ; ++it )
(*it)[i] = -(*it)[i];
}
}
}
} // end of loop on the types of directions
// The direction are unscaled on a unit n-sphere when necessary (that is directions that are not orthomads)
for ( it2 = dirs.begin() ; it2 != dirs.end() ; ++it2 )
if ( ! NOMAD::dir_is_orthomads ( it2->get_type() ) )
{
NOMAD::Double norm_dir=it2->norm();
for (int i=0 ; i < _nc ; ++i )
(*it2)[i] /= norm_dir;
}
}
