【发布时间】：2021-11-22 20:42:51
【问题描述】：
我试图弄清楚为什么我的伪逆 C 代码结果与 MATLAB 结果不同。
这是伪逆的代码：http://www.mymathlib.com/c_source/matrices/linearsystems/singular_value_decomposition.c
#include <string.h> // required for memcpy()
#include <float.h> // required for DBL_EPSILON
#include <math.h> // required for fabs(), sqrt();
#define MAX_ITERATION_COUNT 30 // Maximum number of iterations
// Internally Defined Routines
static void Householders_Reduction_to_Bidiagonal_Form(double* A, int nrows,
int ncols, double* U, double* V, double* diagonal, double* superdiagonal );
static int Givens_Reduction_to_Diagonal_Form( int nrows, int ncols,
double* U, double* V, double* diagonal, double* superdiagonal );
static void Sort_by_Decreasing_Singular_Values(int nrows, int ncols,
double* singular_value, double* U, double* V);
////////////////////////////////////////////////////////////////////////////////
// int Singular_Value_Decomposition(double* A, int nrows, int ncols, //
// double* U, double* singular_values, double* V, double* dummy_array) //
// //
// Description: //
// This routine decomposes an m x n matrix A, with m >= n, into a product //
// of the three matrices U, D, and V', i.e. A = UDV', where U is an m x n //
// matrix whose columns are orthogonal, D is a n x n diagonal matrix, and //
// V is an n x n orthogonal matrix. V' denotes the transpose of V. If //
// m < n, then the procedure may be used for the matrix A'. The singular //
// values of A are the diagonal elements of the diagonal matrix D and //
// correspond to the positive square roots of the eigenvalues of the //
// matrix A'A. //
//
int Singular_Value_Decomposition(double* A, int nrows, int ncols, double* U,
double* singular_values, double* V, double* dummy_array)
{
Householders_Reduction_to_Bidiagonal_Form( A, nrows, ncols, U, V,
singular_values, dummy_array);
if (Givens_Reduction_to_Diagonal_Form( nrows, ncols, U, V,
singular_values, dummy_array ) < 0) return -1;
Sort_by_Decreasing_Singular_Values(nrows, ncols, singular_values, U, V);
return 0;
}
////////////////////////////////////////////////////////////////////////////////
// static void Householders_Reduction_to_Bidiagonal_Form(double* A, int nrows,//
// int ncols, double* U, double* V, double* diagonal, double* superdiagonal )//
// //
// Description: //
// This routine decomposes an m x n matrix A, with m >= n, into a product //
// of the three matrices U, B, and V', i.e. A = UBV', where U is an m x n //
// matrix whose columns are orthogonal, B is a n x n bidiagonal matrix, //
// and V is an n x n orthogonal matrix. V' denotes the transpose of V. //
// If m < n, then the procedure may be used for the matrix A'. The //
// //
/
////////////////////////////////////////////////////////////////////////////////
// //
static void Householders_Reduction_to_Bidiagonal_Form(double* A, int nrows,
int ncols, double* U, double* V, double* diagonal, double* superdiagonal )
{
int i,j,k,ip1;
double s, s2, si, scale;
double dum;
double *pu, *pui, *pv, *pvi;
double half_norm_squared;
// Copy A to U
memcpy(U,A, sizeof(double) * nrows * ncols);
//
diagonal[0] = 0.0;
s = 0.0;
scale = 0.0;
for ( i = 0, pui = U, ip1 = 1; i < ncols; pui += ncols, i++, ip1++ ) {
superdiagonal[i] = scale * s;
//
// Perform Householder transform on columns.
//
// Calculate the normed squared of the i-th column vector starting at
// row i.
//
for (j = i, pu = pui, scale = 0.0; j < nrows; j++, pu += ncols)
scale += fabs( *(pu + i) );
if (scale > 0.0) {
for (j = i, pu = pui, s2 = 0.0; j < nrows; j++, pu += ncols) {
*(pu + i) /= scale;
s2 += *(pu + i) * *(pu + i);
}
//
//
// Chose sign of s which maximizes the norm
//
s = ( *(pui + i) < 0.0 ) ? sqrt(s2) : -sqrt(s2);
//
// Calculate -2/u'u
//
half_norm_squared = *(pui + i) * s - s2;
//
// Transform remaining columns by the Householder transform.
//
*(pui + i) -= s;
for (j = ip1; j < ncols; j++) {
for (k = i, si = 0.0, pu = pui; k < nrows; k++, pu += ncols)
si += *(pu + i) * *(pu + j);
si /= half_norm_squared;
for (k = i, pu = pui; k < nrows; k++, pu += ncols) {
*(pu + j) += si * *(pu + i);
}
}
}
for (j = i, pu = pui; j < nrows; j++, pu += ncols) *(pu + i) *= scale;
diagonal[i] = s * scale;
//
// Perform Householder transform on rows.
//
// Calculate the normed squared of the i-th row vector starting at
// column i.
//
s = 0.0;
scale = 0.0;
if (i >= nrows || i == (ncols - 1) ) continue;
for (j = ip1; j < ncols; j++) scale += fabs ( *(pui + j) );
if ( scale > 0.0 ) {
for (j = ip1, s2 = 0.0; j < ncols; j++) {
*(pui + j) /= scale;
s2 += *(pui + j) * *(pui + j);
}
s = ( *(pui + ip1) < 0.0 ) ? sqrt(s2) : -sqrt(s2);
//
// Calculate -2/u'u
//
half_norm_squared = *(pui + ip1) * s - s2;
//
// Transform the rows by the Householder transform.
//
*(pui + ip1) -= s;
for (k = ip1; k < ncols; k++)
superdiagonal[k] = *(pui + k) / half_norm_squared;
if ( i < (nrows - 1) ) {
for (j = ip1, pu = pui + ncols; j < nrows; j++, pu += ncols) {
for (k = ip1, si = 0.0; k < ncols; k++)
si += *(pui + k) * *(pu + k);
for (k = ip1; k < ncols; k++) {
*(pu + k) += si * superdiagonal[k];
}
}
}
for (k = ip1; k < ncols; k++) *(pui + k) *= scale;
}
}
// Update V
pui = U + ncols * (ncols - 2);
pvi = V + ncols * (ncols - 1);
*(pvi + ncols - 1) = 1.0;
s = superdiagonal[ncols - 1];
pvi -= ncols;
for (i = ncols - 2, ip1 = ncols - 1; i >= 0; i--, pui -= ncols,
pvi -= ncols, ip1-- ) {
if ( s != 0.0 ) {
pv = pvi + ncols;
for (j = ip1; j < ncols; j++, pv += ncols)
*(pv + i) = ( *(pui + j) / *(pui + ip1) ) / s;
for (j = ip1; j < ncols; j++) {
si = 0.0;
for (k = ip1, pv = pvi + ncols; k < ncols; k++, pv += ncols)
si += *(pui + k) * *(pv + j);
for (k = ip1, pv = pvi + ncols; k < ncols; k++, pv += ncols)
*(pv + j) += si * *(pv + i);
}
}
pv = pvi + ncols;
for ( j = ip1; j < ncols; j++, pv += ncols ) {
*(pvi + j) = 0.0;
*(pv + i) = 0.0;
}
*(pvi + i) = 1.0;
s = superdiagonal[i];
}
// Update U
pui = U + ncols * (ncols - 1);
for (i = ncols - 1, ip1 = ncols; i >= 0; ip1 = i, i--, pui -= ncols ) {
s = diagonal[i];
for ( j = ip1; j < ncols; j++) *(pui + j) = 0.0;
if ( s != 0.0 ) {
for (j = ip1; j < ncols; j++) {
si = 0.0;
pu = pui + ncols;
for (k = ip1; k < nrows; k++, pu += ncols)
si += *(pu + i) * *(pu + j);
si = (si / *(pui + i) ) / s;
for (k = i, pu = pui; k < nrows; k++, pu += ncols)
*(pu + j) += si * *(pu + i);
}
for (j = i, pu = pui; j < nrows; j++, pu += ncols){
*(pu + i) /= s;
}
}
else
for (j = i, pu = pui; j < nrows; j++, pu += ncols) *(pu + i) = 0.0;
*(pui + i) += 1.0;
}
}
////////////////////////////////////////////////////////////////////////////////
// static int Givens_Reduction_to_Diagonal_Form( int nrows, int ncols, //
// double* U, double* V, double* diagonal, double* superdiagonal ) //
// //
// Description: //
// This routine decomposes a bidiagonal matrix given by the arrays //
// diagonal and superdiagonal into a product of three matrices U1, D and //
// V1', the matrix U1 premultiplies U and is returned in U, the matrix //
// V1 premultiplies V and is returned in V. The matrix D is a diagonal //
// matrix and replaces the array diagonal. //
// //
// The method used to annihilate the offdiagonal elements is a variant //
// of the QR transformation. The method consists of applying Givens //
// rotations to the right and the left of the current matrix until //
// the new off-diagonal elements are chased out of the matrix. //
// //
// The process is an iterative process which due to roundoff errors may //
// not converge within a predefined number of iterations. (This should //
// be unusual.) //
// //
// Arguments: //
// int nrows //
// The number of rows of the matrix U. //
// int ncols //
// The number of columns of the matrix U. //
// double* U //
// On input, a pointer to a matrix already initialized to a matrix //
// with mutually orthogonal columns. On output, the matrix with //
// mutually orthogonal columns. //
// double* V //
// On input, a pointer to a square matrix with the same number of rows //
// and columns as the columns of the matrix U, i.e. V[ncols][ncols]. //
// The matrix V is assumed to be initialized to an orthogonal matrix. //
// On output, V is an orthogonal matrix. //
// double* diagonal //
// On input, a pointer to an array of dimension ncols which initially //
// contains the diagonal of the bidiagonal matrix. On output, the //
// it contains the diagonal of the diagonal matrix. //
// double* superdiagonal //
// On input, a pointer to an array of dimension ncols which initially //
// the first component is zero and the successive components form the //
// superdiagonal of the bidiagonal matrix. //
// //
// Return Values: //
// 0 Success //
// -1 Failure - The procedure failed to terminate within //
// MAX_ITERATION_COUNT iterations. //
// //
// Example: //
// #define M //
// #define N //
// double U[M][N]; //
// double V[N][N]; //
// double diagonal[N]; //
// double superdiagonal[N]; //
// int err; //
// //
// (your code to initialize the matrices U, V, diagonal, and ) //
// ( superdiagonal. - Note this routine is not accessible from outside) //
// ( i.e. it is declared static.) //
// //
// err = Givens_Reduction_to_Diagonal_Form( M,N,(double*)U,(double*)V, //
// diagonal, superdiagonal ); //
// if ( err < 0 ) printf("Failed to converge\n"); //
// else { ... } //
// ... //
////////////////////////////////////////////////////////////////////////////////
// //
static int Givens_Reduction_to_Diagonal_Form( int nrows, int ncols,
double* U, double* V, double* diagonal, double* superdiagonal )
{
double epsilon;
double c, s;
double f,g,h;
double x,y,z;
double *pu, *pv;
int i,j,k,m;
int rotation_test;
int iteration_count;
for (i = 0, x = 0.0; i < ncols; i++) {
y = fabs(diagonal[i]) + fabs(superdiagonal[i]);
if ( x < y ) x = y;
}
epsilon = x * DBL_EPSILON;
for (k = ncols - 1; k >= 0; k--) {
iteration_count = 0;
while(1) {
rotation_test = 1;
for (m = k; m >= 0; m--) {
if (fabs(superdiagonal[m]) <= epsilon) {rotation_test = 0; break;}
if (fabs(diagonal[m-1]) <= epsilon) break;
}
if (rotation_test) {
c = 0.0;
s = 1.0;
for (i = m; i <= k; i++) {
f = s * superdiagonal[i];
superdiagonal[i] *= c;
if (fabs(f) <= epsilon) break;
g = diagonal[i];
h = sqrt(f*f + g*g);
diagonal[i] = h;
c = g / h;
s = -f / h;
for (j = 0, pu = U; j < nrows; j++, pu += ncols) {
y = *(pu + m - 1);
z = *(pu + i);
*(pu + m - 1 ) = y * c + z * s;
*(pu + i) = -y * s + z * c;
}
}
}
z = diagonal[k];
if (m == k ) {
if ( z < 0.0 ) {
diagonal[k] = -z;
for ( j = 0, pv = V; j < ncols; j++, pv += ncols)
*(pv + k) = - *(pv + k);
}
break;
}
else {
if ( iteration_count >= MAX_ITERATION_COUNT ) return -1;
iteration_count++;
x = diagonal[m];
y = diagonal[k-1];
g = superdiagonal[k-1];
h = superdiagonal[k];
f = ( (y - z) * ( y + z ) + (g - h) * (g + h) )/(2.0 * h * y);
g = sqrt( f * f + 1.0 );
if ( f < 0.0 ) g = -g;
f = ( (x - z) * (x + z) + h * (y / (f + g) - h) ) / x;
// Next QR Transformtion
c = 1.0;
s = 1.0;
for (i = m + 1; i <= k; i++) {
g = superdiagonal[i];
y = diagonal[i];
h = s * g;
g *= c;
z = sqrt( f * f + h * h );
superdiagonal[i-1] = z;
c = f / z;
s = h / z;
f = x * c + g * s;
g = -x * s + g * c;
h = y * s;
y *= c;
for (j = 0, pv = V; j < ncols; j++, pv += ncols) {
x = *(pv + i - 1);
z = *(pv + i);
*(pv + i - 1) = x * c + z * s;
*(pv + i) = -x * s + z * c;
}
z = sqrt( f * f + h * h );
diagonal[i - 1] = z;
if (z != 0.0) {
c = f / z;
s = h / z;
}
f = c * g + s * y;
x = -s * g + c * y;
for (j = 0, pu = U; j < nrows; j++, pu += ncols) {
y = *(pu + i - 1);
z = *(pu + i);
*(pu + i - 1) = c * y + s * z;
*(pu + i) = -s * y + c * z;
}
}
superdiagonal[m] = 0.0;
superdiagonal[k] = f;
diagonal[k] = x;
}
}
}
return 0;
}
////////////////////////////////////////////////////////////////////////////////
// static void Sort_by_Decreasing_Singular_Values(int nrows, int ncols, //
// double* singular_values, double* U, double* V) //
// //
// Description: //
// This routine sorts the singular values from largest to smallest //
// singular value and interchanges the columns of U and the columns of V //
// whenever a swap is made. I.e. if the i-th singular value is swapped //
// with the j-th singular value, then the i-th and j-th columns of U are //
// interchanged and the i-th and j-th columns of V are interchanged. //
// //
// Arguments: //
// int nrows //
// The number of rows of the matrix U. //
// int ncols //
// The number of columns of the matrix U. //
// double* singular_values //
// On input, a pointer to the array of singular values. On output, the//
// sorted array of singular values. //
// double* U //
// On input, a pointer to a matrix already initialized to a matrix //
// with mutually orthogonal columns. On output, the matrix with //
// mutually orthogonal possibly permuted columns. //
// double* V //
// On input, a pointer to a square matrix with the same number of rows //
// and columns as the columns of the matrix U, i.e. V[ncols][ncols]. //
// The matrix V is assumed to be initialized to an orthogonal matrix. //
// On output, V is an orthogonal matrix with possibly permuted columns.//
// //
// Return Values: //
// The function is of type void. //
// //
// Example: //
// #define M //
// #define N //
// double U[M][N]; //
// double V[N][N]; //
// double diagonal[N]; //
// //
// (your code to initialize the matrices U, V, and diagonal. ) //
// ( - Note this routine is not accessible from outside) //
// ( i.e. it is declared static.) //
// //
// Sort_by_Decreasing_Singular_Values(nrows, ncols, singular_values, //
// (double*) U, (double*) V); //
// ... //
////////////////////////////////////////////////////////////////////////////////
// //
static void Sort_by_Decreasing_Singular_Values(int nrows, int ncols,
double* singular_values, double* U, double* V)
{
int i,j,max_index;
double temp;
double *p1, *p2;
for (i = 0; i < ncols - 1; i++) {
max_index = i;
for (j = i + 1; j < ncols; j++)
if (singular_values[j] > singular_values[max_index] )
max_index = j;
if (max_index == i) continue;
temp = singular_values[i];
singular_values[i] = singular_values[max_index];
singular_values[max_index] = temp;
p1 = U + max_index;
p2 = U + i;
for (j = 0; j < nrows; j++, p1 += ncols, p2 += ncols) {
temp = *p1;
*p1 = *p2;
*p2 = temp;
}
p1 = V + max_index;
p2 = V + i;
for (j = 0; j < ncols; j++, p1 += ncols, p2 += ncols) {
temp = *p1;
*p1 = *p2;
*p2 = temp;
}
}
}
////////////////////////////////////////////////////////////////////////////////
// void Singular_Value_Decomposition_Inverse(double* U, double* D, double* V,//
// double tolerance, int nrows, int ncols, double *Astar) //
// //
// Description: //
// This routine calculates the pseudo-inverse of the matrix A = UDV'. //
// where U, D, V constitute the singular value decomposition of A. //
// Let Astar be the pseudo-inverse then Astar = V(1/D)U', where 1/D is //
// the pseudo-inverse of D, i.e. if D[i] > 0 then (1/D)[i] = 1/D[i] and //
// if D[i] = 0, then (1/D)[i] = 0. Because the singular values are //
// subject to round-off error. A tolerance is given so that if //
// D[i] < tolerance, D[i] is treated as if it were 0. //
// The default tolerance is D[0] * DBL_EPSILON * ncols, assuming that the //
// diagonal matrix of singular values is sorted from largest to smallest, //
// if the user specified tolerance is less than the default tolerance, //
// then the default tolerance is used. //
// //
// Arguments: //
// double* U //
// A matrix with mutually orthonormal columns. //
// double* D //
// A diagonal matrix with decreasing non-negative diagonal elements. //
// i.e. D[i] > D[j] if i < j and D[i] >= 0 for all i. //
// double* V //
// An orthogonal matrix. //
// double tolerance //
// An lower bound for non-zero singular values (provided tolerance > //
// ncols * DBL_EPSILON * D[0]). //
// int nrows //
// The number of rows of the matrix U and B. //
// int ncols //
// The number of columns of the matrix U. Also the number of rows and //
// columns of the matrices D and V. //
// double* Astar //
// On input, a pointer to the first element of an ncols x nrows matrix.//
// On output, the pseudo-inverse of UDV'. //
// //
// Return Values: //
// The function is of type void. //
// //
// Example: //
// #define M //
// #define N //
// double U[M][N]; //
// double V[N][N]; //
// double D[N]; //
// double Astar[N][M]; //
// double tolerance; //
// //
// (your code to initialize the matrices U,D,V) //
// //
// Singular_Value_Decomposition_Inverse((double*) U, D, (double*) V, //
// tolerance, M, N, (double*) Astar); //
// //
// printf(" The pseudo-inverse of A = UDV' is \n"); //
// ... //
////////////////////////////////////////////////////////////////////////////////
// //
void Singular_Value_Decomposition_Inverse(double* U, double* D, double* V,
double tolerance, int nrows, int ncols, double *Astar)
{
int i,j,k;
double *pu, *pv, *pa;
double dum;
dum = DBL_EPSILON * D[0] * (double) ncols;
if (tolerance < dum) tolerance = dum;
for ( i = 0, pv = V, pa = Astar; i < ncols; i++, pv += ncols)
for ( j = 0, pu = U; j < nrows; j++, pa++)
for (k = 0, *pa = 0.0; k < ncols; k++, pu++)
if (D[k] > tolerance) *pa += *(pv + k) * *pu / D[k];
}
我已将容差设置为 1e-16。
输入矩阵：
MatA[4][4] = {
{1e-15,2e-15,3e-15,4e-15},
{5e-15,10e-15,7e-15,8e-15},
{9e-15, 18e-15, 11e-15,12e-15},
{13e-15,26e-15,15e-15,16e-15}
};
C代码结果：
-7.3177e+13 -3.6957e+13 -7.3773e+11 3.5482e+13
-1.4635e+14 -7.3915e+13 -1.4755e+12 7.0964e+13
1.0264e+14 5.7015e+13 1.1387e+13 -3.4240e+13
1.9055e+14 1.0400e+14 1.7450e+13 -6.9101e+13
Matlab 结果：
1.0e+14 *
-0.7348 -0.3712 -0.0076 0.3561
-1.4697 -0.7424 -0.0152 0.7121
1.0227 0.5682 0.1136 -0.3409
1.9015 1.0379 0.1742 -0.6894
我不确定我在哪里失去了准确性。我能看到准确性生效的唯一地方是 DBL_Epsilon 和 Tolerance。我还将 DBL_Epsilon 的值设为 4.94065645841247E-32。不知道如何获得更接近 matlab 输出的值。
【问题讨论】：
-
请不要发布代码链接。将代码包含在帖子本身中。
-
... 并且代码应该在帖子中以minimal reproducible example 的形式呈现 - 所以不是所有代码的转储，而是足够了 i> 所以可以编译它并重现问题。
-
请注意，您可以避免输入矩阵中的
e-15（只需乘以e15结果），即使在实践中它不应该改变太多问题。一个有趣的测试是使用两个代码检查输入矩阵的奇异值。 -
嗯，我没有字符空间来发布代码以及我如何调用伪逆函数。但我只是简单地调用 Singular_Value_Decomposition(double* A, int nrows, int ncols, double* U, double* alien_values, double* V, double* dummy_array) 然后 Singular_Value_Decomposition_Inverse(double* U, double* D, double* V, double tolerance , int nrows, int ncols, double *Astar)
-
这是相当多的代码。我怀疑您看到的差异将是由于处理小的奇异值造成的。您可以通过打开
pinv函数来查看Matlab 的确切功能，这应该可以帮助您进行调试。
标签： c matlab linear-algebra svd matrix-inverse