//! Returns the pseudo square root of a real symmetric matrix
/*! Given a matrix \f$ M \f$, the result \f$ S \f$ is defined
* as the matrix such that \f$ S S^T = M. \f$
* If the matrix is not positive semi definite, it can
* return an approximation of the pseudo square root
* using a (user selected) salvaging algorithm.
*
* For more information see: "The most general methodology to create
* a valid correlation matrix for risk management and option pricing
* purposes", by R. Rebonato and P. Jдckel.
* The Journal of Risk, 2(2), Winter 1999/2000
* http://www.rebonato.com/correlationmatrix.pdf
*
* Revised and extended in "Monte Carlo Methods in Finance",
* by Peter Jдckel, Chapter 6.
*
* \pre the given matrix must be symmetric.
*
* \relates Matrix
*
* \warning Higham algorithm only works for correlation matrices.
*
* \test
* - the correctness of the results is tested by reproducing
* known good data.
* - the correctness of the results is tested by checking
* returned values against numerical calculations.
*/
public static Matrix pseudoSqrt(Matrix matrix, SalvagingAlgorithm sa)
{
int size = matrix.rows();
#if QL_EXTRA_SAFETY_CHECKS
checkSymmetry(matrix);
#else
if (size != matrix.columns())
{
throw new Exception("non square matrix: " + size + " rows, " + matrix.columns() + " columns");
}
#endif
// spectral (a.k.a Principal Component) analysis
SymmetricSchurDecomposition jd = new SymmetricSchurDecomposition(matrix);
Matrix diagonal = new Matrix(size, size, 0.0);
// salvaging algorithm
Matrix result = new Matrix(size, size);
bool negative;
switch (sa)
{
case SalvagingAlgorithm.None:
// eigenvalues are sorted in decreasing order
if (!(jd.eigenvalues()[size - 1] >= -1e-16))
{
throw new Exception("negative eigenvalue(s) (" + jd.eigenvalues()[size - 1] + ")");
}
result = MatrixUtilities.CholeskyDecomposition(matrix, true);
break;
case SalvagingAlgorithm.Spectral:
// negative eigenvalues set to zero
for (int i = 0; i < size; i++)
{
diagonal[i, i] = Math.Sqrt(Math.Max(jd.eigenvalues()[i], 0.0));
}
result = jd.eigenvectors() * diagonal;
normalizePseudoRoot(matrix, result);
break;
case SalvagingAlgorithm.Hypersphere:
// negative eigenvalues set to zero
negative = false;
for (int i = 0; i < size; ++i)
{
diagonal[i, i] = Math.Sqrt(Math.Max(jd.eigenvalues()[i], 0.0));
if (jd.eigenvalues()[i] < 0.0)
{
negative = true;
}
}
result = jd.eigenvectors() * diagonal;
normalizePseudoRoot(matrix, result);
if (negative)
{
result = hypersphereOptimize(matrix, result, false);
}
break;
case SalvagingAlgorithm.LowerDiagonal:
// negative eigenvalues set to zero
negative = false;
for (int i = 0; i < size; ++i)
{
diagonal[i, i] = Math.Sqrt(Math.Max(jd.eigenvalues()[i], 0.0));
if (jd.eigenvalues()[i] < 0.0)
{
negative = true;
}
}
result = jd.eigenvectors() * diagonal;
normalizePseudoRoot(matrix, result);
if (negative)
{
result = hypersphereOptimize(matrix, result, true);
}
break;
case SalvagingAlgorithm.Higham:
int maxIterations = 40;
double tol = 1e-6;
result = highamImplementation(matrix, maxIterations, tol);
result = MatrixUtilities.CholeskyDecomposition(result, true);
break;
default:
throw new Exception("unknown salvaging algorithm");
}
return(result);
}
// Optimization function for hypersphere and lower-diagonal algorithm
private static Matrix hypersphereOptimize(Matrix targetMatrix, Matrix currentRoot, bool lowerDiagonal)
{
int i, j, k, size = targetMatrix.rows();
Matrix result = new Matrix(currentRoot);
Vector variance = new Vector(size);
for (i = 0; i < size; i++)
{
variance[i] = Math.Sqrt(targetMatrix[i, i]);
}
if (lowerDiagonal)
{
Matrix approxMatrix = result * Matrix.transpose(result);
result = MatrixUtilities.CholeskyDecomposition(approxMatrix, true);
for (i = 0; i < size; i++)
{
for (j = 0; j < size; j++)
{
result[i, j] /= Math.Sqrt(approxMatrix[i, i]);
}
}
}
else
{
for (i = 0; i < size; i++)
{
for (j = 0; j < size; j++)
{
result[i, j] /= variance[i];
}
}
}
ConjugateGradient optimize = new ConjugateGradient();
EndCriteria endCriteria = new EndCriteria(100, 10, 1e-8, 1e-8, 1e-8);
HypersphereCostFunction costFunction = new HypersphereCostFunction(targetMatrix, variance, lowerDiagonal);
NoConstraint constraint = new NoConstraint();
// hypersphere vector optimization
if (lowerDiagonal)
{
Vector theta = new Vector(size * (size - 1) / 2);
const double eps = 1e-16;
for (i = 1; i < size; i++)
{
for (j = 0; j < i; j++)
{
theta[i * (i - 1) / 2 + j] = result[i, j];
if (theta[i * (i - 1) / 2 + j] > 1 - eps)
{
theta[i * (i - 1) / 2 + j] = 1 - eps;
}
if (theta[i * (i - 1) / 2 + j] < -1 + eps)
{
theta[i * (i - 1) / 2 + j] = -1 + eps;
}
for (k = 0; k < j; k++)
{
theta[i * (i - 1) / 2 + j] /= Math.Sin(theta[i * (i - 1) / 2 + k]);
if (theta[i * (i - 1) / 2 + j] > 1 - eps)
{
theta[i * (i - 1) / 2 + j] = 1 - eps;
}
if (theta[i * (i - 1) / 2 + j] < -1 + eps)
{
theta[i * (i - 1) / 2 + j] = -1 + eps;
}
}
theta[i * (i - 1) / 2 + j] = Math.Acos(theta[i * (i - 1) / 2 + j]);
if (j == i - 1)
{
if (result[i, i] < 0)
{
theta[i * (i - 1) / 2 + j] = -theta[i * (i - 1) / 2 + j];
}
}
}
}
Problem p = new Problem(costFunction, constraint, theta);
optimize.minimize(p, endCriteria);
theta = p.currentValue();
result.fill(1);
for (i = 0; i < size; i++)
{
for (k = 0; k < size; k++)
{
if (k > i)
{
result[i, k] = 0;
}
else
{
for (j = 0; j <= k; j++)
{
if (j == k && k != i)
{
result[i, k] *= Math.Cos(theta[i * (i - 1) / 2 + j]);
}
else if (j != i)
{
result[i, k] *= Math.Sin(theta[i * (i - 1) / 2 + j]);
}
}
}
}
}
}
else
{
Vector theta = new Vector(size * (size - 1));
const double eps = 1e-16;
for (i = 0; i < size; i++)
{
for (j = 0; j < size - 1; j++)
{
theta[j * size + i] = result[i, j];
if (theta[j * size + i] > 1 - eps)
{
theta[j * size + i] = 1 - eps;
}
if (theta[j * size + i] < -1 + eps)
{
theta[j * size + i] = -1 + eps;
}
for (k = 0; k < j; k++)
{
theta[j * size + i] /= Math.Sin(theta[k * size + i]);
if (theta[j * size + i] > 1 - eps)
{
theta[j * size + i] = 1 - eps;
}
if (theta[j * size + i] < -1 + eps)
{
theta[j * size + i] = -1 + eps;
}
}
theta[j * size + i] = Math.Acos(theta[j * size + i]);
if (j == size - 2)
{
if (result[i, j + 1] < 0)
{
theta[j * size + i] = -theta[j * size + i];
}
}
}
}
Problem p = new Problem(costFunction, constraint, theta);
optimize.minimize(p, endCriteria);
theta = p.currentValue();
result.fill(1);
for (i = 0; i < size; i++)
{
for (k = 0; k < size; k++)
{
for (j = 0; j <= k; j++)
{
if (j == k && k != size - 1)
{
result[i, k] *= Math.Cos(theta[j * size + i]);
}
else if (j != size - 1)
{
result[i, k] *= Math.Sin(theta[j * size + i]);
}
}
}
}
}
for (i = 0; i < size; i++)
{
for (j = 0; j < size; j++)
{
result[i, j] *= variance[i];
}
}
return(result);
}
