/// <summary>Creates a 'n x r' matrix B such that B * B' is a correlation matrix and 'near' to the specified symmetric, normalized matrix of dimension n. A rank reduction will apply if r is strict less than n. /// </summary> /// <param name="rawCorrelationMatrix">The symmetric, normalized matrix where to find the 'nearest' correlation matrix.</param> /// <param name="state">The state of the operation in its <see cref="PseudoSqrtMatrixDecomposer.State"/> representation (output).</param> /// <param name="triangularMatrixType">A value indicating which part of <paramref name="rawCorrelationMatrix"/> to take into account.</param> /// <param name="outputEntries">This argument will be used to store the matrix entries of the resulting matrix B, i.e. the return value array points to this array if != <c>null</c>; otherwise a memory allocation will be done.</param> /// <param name="worksspaceContainer">A specific <see cref="PseudoSqrtMatrixDecomposer.WorkspaceContainer"/> object to reduce memory allocation; ignored if <c>null</c>.</param> /// <returns>A <see cref="DenseMatrix"/> object that represents a matrix B such that B * B' is the 'nearest' correlation matrix with respect to <paramref name="rawCorrelationMatrix"/>.</returns> /// <remarks>In general the return object does <b>not</b> represents the pseudo-root of <paramref name="rawCorrelationMatrix"/>, i.e. output of the Cholesky decomposition. /// <para>The parameters <paramref name="outputEntries"/>, <paramref name="worksspaceContainer"/> allows to avoid memory allocation and to re-use arrays if the calculation of correlation matrices will be done often.</para></remarks> public override DenseMatrix Create(DenseMatrix rawCorrelationMatrix, out State state, double[] outputEntries = null, PseudoSqrtMatrixDecomposer.WorkspaceContainer worksspaceContainer = null, BLAS.TriangularMatrixType triangularMatrixType = BLAS.TriangularMatrixType.LowerTriangularMatrix) { if (rawCorrelationMatrix.IsQuadratic == false) { throw new ArgumentException("rawCorrelationMatrix"); } int n = rawCorrelationMatrix.RowCount; if ((outputEntries == null) || (outputEntries.Length < n * n)) { outputEntries = new double[n * n]; } var ws = worksspaceContainer as Workspace; if ((ws == null) || (ws.Dimension < n)) { ws = new Workspace(n); } int m; BLAS.Level1.dcopy(n * n, rawCorrelationMatrix.Data, ws.data); LAPACK.EigenValues.Symmetric.driver_dsyevr(LapackEigenvalues.SymmetricGeneralJob.All, n, ws.data, out m, ws.eigenValues, outputEntries, ws.isuppz, ws.work, ws.iwork); var originalEigenValueDataTable = InfoOutputDetailLevel.IsAtLeastAsComprehensiveAs(InfoOutputDetailLevel.High) ? CreateDataTableWithEigenvalues("Eigenvalues.Original", m, ws.eigenValues) : null; int rank = n; int minNumberOfEigenvaluesToSetZero = n - Math.Min(MaximalRank ?? n, n); int i = 0; while ((i < minNumberOfEigenvaluesToSetZero) || (ws.eigenValues[i] < 0.0)) { ws.eigenValues[i] = 0.0; i++; rank--; } var adjustedEigenValueDataTable = InfoOutputDetailLevel.IsAtLeastAsComprehensiveAs(InfoOutputDetailLevel.High) ? CreateDataTableWithEigenvalues("Eigenvalues.Adjusted", m, ws.eigenValues) : null; VectorUnit.Basics.Sqrt(n, ws.eigenValues); // calculate sqrt of eigenvalues only once, i.e. the array 'eigenValues' contains the sqrt of the eigenvalues! for (i = 0; i < n; i++) { var t_i = 0.0; for (int j = n - 1; j >= n - rank; j--) { t_i += outputEntries[i + j * n] * outputEntries[i + j * n] * ws.eigenValues[j] * ws.eigenValues[j]; outputEntries[i + j * n] *= ws.eigenValues[j]; } BLAS.Level1.dscal(rank, 1.0 / Math.Sqrt(t_i), outputEntries, -n, i + (n - 1) * n); // [i+j*n] *= 1/Math.Sqrt(tempValue) for j = n-1, ..., n-rank } /* The eigenvalues are in ascending order. Thefore the first (and not last) 'rank' columns of the eigenvectors are not required. Therefore we swap the relevant part */ BLAS.Level1.dscal(n * (n - rank), 0.0, outputEntries); BLAS.Level1.dswap(n * rank, outputEntries, 1, outputEntries, 1, n * (n - rank), 0); state = State.Create(rank, detailProperties: new[] { InfoOutputProperty.Create("Eigenvalues set to 0.0", n - rank) }, detailDataTables: new[] { originalEigenValueDataTable, adjustedEigenValueDataTable }, iterationsNeeded: 1, infoOutputDetailLevel: InfoOutputDetailLevel); return(new DenseMatrix(n, rank, outputEntries)); }
/// <summary>Creates a 'n x r' matrix B such that B * B' is a correlation matrix and 'near' to the specified symmetric, normalized matrix of dimension n. A rank reduction will apply if r is strict less than n. /// </summary> /// <param name="rawCorrelationMatrix">The symmetric, normalized matrix where to find the 'nearest' correlation matrix.</param> /// <param name="state">The state of the operation in its <see cref="PseudoSqrtMatrixDecomposer.State"/> representation (output).</param> /// <param name="triangularMatrixType">A value indicating which part of <paramref name="rawCorrelationMatrix"/> to take into account.</param> /// <param name="outputEntries">This argument will be used to store the matrix entries of the resulting matrix B, i.e. the return value array points to this array if != <c>null</c>; otherwise a memory allocation will be done.</param> /// <param name="worksspaceContainer">A specific <see cref="PseudoSqrtMatrixDecomposer.WorkspaceContainer"/> object to reduce memory allocation; ignored if <c>null</c>.</param> /// <returns>A <see cref="DenseMatrix"/> object that represents a matrix B such that B * B' is the 'nearest' correlation matrix with respect to <paramref name="rawCorrelationMatrix"/>.</returns> /// <remarks>In general the return object does <b>not</b> represents the pseudo-root of <paramref name="rawCorrelationMatrix"/>, i.e. output of the Cholesky decomposition. /// <para>The parameters <paramref name="outputEntries"/>, <paramref name="worksspaceContainer"/> allows to avoid memory allocation and to re-use arrays if the calculation of correlation matrices will be done often.</para></remarks> public override DenseMatrix Create(DenseMatrix rawCorrelationMatrix, out State state, double[] outputEntries = null, PseudoSqrtMatrixDecomposer.WorkspaceContainer worksspaceContainer = null, BLAS.TriangularMatrixType triangularMatrixType = BLAS.TriangularMatrixType.LowerTriangularMatrix) { if (rawCorrelationMatrix.IsQuadratic == false) { throw new ArgumentException("rawCorrelationMatrix"); } int n = rawCorrelationMatrix.RowCount; var ws = worksspaceContainer as Workspace; if ((ws == null) || (ws.Dimension < n)) { ws = new Workspace(n, this); } /* calculate an initial value for the optimizer: * (i) Apply the EZN algorithm for the calculation of a matrix B_0 such that B_0 * B_0^t is near to the (raw) correlation matrix * (ii) calculate angle parameters \theta such that B_0 = B(\theta). * */ State initState; var initialAngleParameterMatrix = GetAngleParameter(m_InitialDecomposer.Create(rawCorrelationMatrix, out initState, outputEntries, ws.InitalDecomposerWorkspace, triangularMatrixType), ws.ArgMinData); /* prepare and apply optimization algorithm: */ int rank = initState.Rank; if ((outputEntries == null) || (outputEntries.Length < n * n)) { outputEntries = new double[n * rank]; } var B = new DenseMatrix(n, rank, outputEntries, createDeepCopyOfArgument: false); var C = new DenseMatrix(n, n, ws.CorrelationMatrixData, createDeepCopyOfArgument: false); var optAlgorithm = m_MultiDimOptimizer.Create(n * (rank - 1)); optAlgorithm.SetFunction(theta => { GetParametricMatrix(theta, B); C.AddAssignment(B, B.T, beta: 0.0); // C = B * B^t VectorUnit.Basics.Sub(n * n, C.Data, rawCorrelationMatrix.Data, ws.TempDifferences); return(BLAS.Level1.dnrm2sq(n * n, ws.TempDifferences)); }); double minimum; var optState = optAlgorithm.FindMinimum(ws.ArgMinData, out minimum); state = State.Create(rank, optState.IterationsNeeded, InfoOutputDetailLevel, Tuple.Create <string, IInfoOutputQueriable>("Initial.State", initState), Tuple.Create <string, IInfoOutputQueriable>("Final.Optimizer", optState), Tuple.Create <string, IInfoOutputQueriable>("Initial.Parameters", initialAngleParameterMatrix), InfoOutputDetailLevel.IsAtLeastAsComprehensiveAs(InfoOutputDetailLevel.High) ? Tuple.Create <string, IInfoOutputQueriable>("Final.Parameters", new DenseMatrix(n, rank, ws.ArgMinData, createDeepCopyOfArgument: false)) : null ); GetParametricMatrix(ws.ArgMinData, B); // B should be already set to B(\theta^*), we just want to be sure that B is correct on exit return(B); }