//$$ newmat8.cpp Advanced LU transform, scalar functions // Copyright (C) 1991,2,3,4,8: R B Davies #define WANT_MATH #include "include.h" #include "newmat.h" #include "newmatrc.h" #include "precisio.h" #ifdef use_namespace namespace NEWMAT { #endif #ifdef DO_REPORT #define REPORT { static ExeCounter ExeCount(__LINE__,8); ++ExeCount; } #else #define REPORT {} #endif /************************** LU transformation ****************************/ void CroutMatrix::ludcmp() // LU decomposition from Golub & Van Loan, algorithm 3.4.1, (the "outer // product" version). // This replaces the code derived from Numerical Recipes in C in previous // versions of newmat and being row oriented runs much faster with large // matrices. { REPORT Tracer trace( "Crout(ludcmp)" ); sing = false; Real* akk = store; // runs down diagonal Real big = fabs(*akk); int mu = 0; Real* ai = akk; int k; for (k = 1; k < nrows; k++) { ai += nrows; const Real trybig = fabs(*ai); if (big < trybig) { big = trybig; mu = k; } } if (nrows) for (k = 0;;) { /* int mu1; { Real big = fabs(*akk); mu1 = k; Real* ai = akk; int i; for (i = k+1; i < nrows; i++) { ai += nrows; const Real trybig = fabs(*ai); if (big < trybig) { big = trybig; mu1 = i; } } } if (mu1 != mu) cout << k << " " << mu << " " << mu1 << endl; */ indx[k] = mu; if (mu != k) //row swap { Real* a1 = store + nrows * k; Real* a2 = store + nrows * mu; d = !d; int j = nrows; while (j--) { const Real temp = *a1; *a1++ = *a2; *a2++ = temp; } } Real diag = *akk; big = 0; mu = k + 1; if (diag != 0) { ai = akk; int i = nrows - k - 1; while (i--) { ai += nrows; Real* al = ai; Real mult = *al / diag; *al = mult; int l = nrows - k - 1; Real* aj = akk; // work out the next pivot as part of this loop // this saves a column operation if (l-- != 0) { *(++al) -= (mult * *(++aj)); const Real trybig = fabs(*al); if (big < trybig) { big = trybig; mu = nrows - i - 1; } while (l--) *(++al) -= (mult * *(++aj)); } } } else sing = true; if (++k == nrows) break; // so next line won't overflow akk += nrows + 1; } } void CroutMatrix::lubksb(Real* B, int mini) { REPORT // this has been adapted from Numerical Recipes in C. The code has been // substantially streamlined, so I do not think much of the original // copyright remains. However there is not much opportunity for // variation in the code, so it is still similar to the NR code. // I follow the NR code in skipping over initial zeros in the B vector. Tracer trace("Crout(lubksb)"); if (sing) Throw(SingularException(*this)); int i, j, ii = nrows; // ii initialised : B might be all zeros // scan for first non-zero in B for (i = 0; i < nrows; i++) { int ip = indx[i]; Real temp = B[ip]; B[ip] = B[i]; B[i] = temp; if (temp != 0.0) { ii = i; break; } } Real* bi; Real* ai; i = ii + 1; if (i < nrows) { bi = B + ii; ai = store + ii + i * nrows; for (;;) { int ip = indx[i]; Real sum = B[ip]; B[ip] = B[i]; Real* aij = ai; Real* bj = bi; j = i - ii; while (j--) sum -= *aij++ * *bj++; B[i] = sum; if (++i == nrows) break; ai += nrows; } } ai = store + nrows * nrows; for (i = nrows - 1; i >= mini; i--) { Real* bj = B+i; ai -= nrows; Real* ajx = ai+i; Real sum = *bj; Real diag = *ajx; j = nrows - i; while(--j) sum -= *(++ajx) * *(++bj); B[i] = sum / diag; } } /****************************** scalar functions ****************************/ inline Real square(Real x) { return x*x; } Real GeneralMatrix::SumSquare() const { REPORT Real sum = 0.0; int i = storage; Real* s = store; while (i--) sum += square(*s++); ((GeneralMatrix&)*this).tDelete(); return sum; } Real GeneralMatrix::SumAbsoluteValue() const { REPORT Real sum = 0.0; int i = storage; Real* s = store; while (i--) sum += fabs(*s++); ((GeneralMatrix&)*this).tDelete(); return sum; } Real GeneralMatrix::Sum() const { REPORT Real sum = 0.0; int i = storage; Real* s = store; while (i--) sum += *s++; ((GeneralMatrix&)*this).tDelete(); return sum; } // maxima and minima // There are three sets of routines // MaximumAbsoluteValue, MinimumAbsoluteValue, Maximum, Minimum // ... these find just the maxima and minima // MaximumAbsoluteValue1, MinimumAbsoluteValue1, Maximum1, Minimum1 // ... these find the maxima and minima and their locations in a // one dimensional object // MaximumAbsoluteValue2, MinimumAbsoluteValue2, Maximum2, Minimum2 // ... these find the maxima and minima and their locations in a // two dimensional object // If the matrix has no values throw an exception // If we do not want the location find the maximum or minimum on the // array stored by GeneralMatrix // This won't work for BandMatrices. We call ClearCorner for // MaximumAbsoluteValue but for the others use the AbsoluteMinimumValue2 // version and discard the location. // For one dimensional objects, when we want the location of the // maximum or minimum, work with the array stored by GeneralMatrix // For two dimensional objects where we want the location of the maximum or // minimum proceed as follows: // For rectangular matrices use the array stored by GeneralMatrix and // deduce the location from the location in the GeneralMatrix // For other two dimensional matrices use the Matrix Row routine to find the // maximum or minimum for each row. static void NullMatrixError(const GeneralMatrix* gm) { ((GeneralMatrix&)*gm).tDelete(); Throw(ProgramException("Maximum or minimum of null matrix")); } Real GeneralMatrix::MaximumAbsoluteValue() const { REPORT if (storage == 0) NullMatrixError(this); Real maxval = 0.0; int l = storage; Real* s = store; while (l--) { Real a = fabs(*s++); if (maxval < a) maxval = a; } ((GeneralMatrix&)*this).tDelete(); return maxval; } Real GeneralMatrix::MaximumAbsoluteValue1(int& i) const { REPORT if (storage == 0) NullMatrixError(this); Real maxval = 0.0; int l = storage; Real* s = store; int li = storage; while (l--) { Real a = fabs(*s++); if (maxval <= a) { maxval = a; li = l; } } i = storage - li; ((GeneralMatrix&)*this).tDelete(); return maxval; } Real GeneralMatrix::MinimumAbsoluteValue() const { REPORT if (storage == 0) NullMatrixError(this); int l = storage - 1; Real* s = store; Real minval = fabs(*s++); while (l--) { Real a = fabs(*s++); if (minval > a) minval = a; } ((GeneralMatrix&)*this).tDelete(); return minval; } Real GeneralMatrix::MinimumAbsoluteValue1(int& i) const { REPORT if (storage == 0) NullMatrixError(this); int l = storage - 1; Real* s = store; Real minval = fabs(*s++); int li = l; while (l--) { Real a = fabs(*s++); if (minval >= a) { minval = a; li = l; } } i = storage - li; ((GeneralMatrix&)*this).tDelete(); return minval; } Real GeneralMatrix::Maximum() const { REPORT if (storage == 0) NullMatrixError(this); int l = storage - 1; Real* s = store; Real maxval = *s++; while (l--) { Real a = *s++; if (maxval < a) maxval = a; } ((GeneralMatrix&)*this).tDelete(); return maxval; } Real GeneralMatrix::Maximum1(int& i) const { REPORT if (storage == 0) NullMatrixError(this); int l = storage - 1; Real* s = store; Real maxval = *s++; int li = l; while (l--) { Real a = *s++; if (maxval <= a) { maxval = a; li = l; } } i = storage - li; ((GeneralMatrix&)*this).tDelete(); return maxval; } Real GeneralMatrix::Minimum() const { REPORT if (storage == 0) NullMatrixError(this); int l = storage - 1; Real* s = store; Real minval = *s++; while (l--) { Real a = *s++; if (minval > a) minval = a; } ((GeneralMatrix&)*this).tDelete(); return minval; } Real GeneralMatrix::Minimum1(int& i) const { REPORT if (storage == 0) NullMatrixError(this); int l = storage - 1; Real* s = store; Real minval = *s++; int li = l; while (l--) { Real a = *s++; if (minval >= a) { minval = a; li = l; } } i = storage - li; ((GeneralMatrix&)*this).tDelete(); return minval; } Real GeneralMatrix::MaximumAbsoluteValue2(int& i, int& j) const { REPORT if (storage == 0) NullMatrixError(this); Real maxval = 0.0; int nr = Nrows(); MatrixRow mr((GeneralMatrix*)this, LoadOnEntry+DirectPart); for (int r = 1; r <= nr; r++) { int c; maxval = mr.MaximumAbsoluteValue1(maxval, c); if (c > 0) { i = r; j = c; } mr.Next(); } ((GeneralMatrix&)*this).tDelete(); return maxval; } Real GeneralMatrix::MinimumAbsoluteValue2(int& i, int& j) const { REPORT if (storage == 0) NullMatrixError(this); Real minval = FloatingPointPrecision::Maximum(); int nr = Nrows(); MatrixRow mr((GeneralMatrix*)this, LoadOnEntry+DirectPart); for (int r = 1; r <= nr; r++) { int c; minval = mr.MinimumAbsoluteValue1(minval, c); if (c > 0) { i = r; j = c; } mr.Next(); } ((GeneralMatrix&)*this).tDelete(); return minval; } Real GeneralMatrix::Maximum2(int& i, int& j) const { REPORT if (storage == 0) NullMatrixError(this); Real maxval = -FloatingPointPrecision::Maximum(); int nr = Nrows(); MatrixRow mr((GeneralMatrix*)this, LoadOnEntry+DirectPart); for (int r = 1; r <= nr; r++) { int c; maxval = mr.Maximum1(maxval, c); if (c > 0) { i = r; j = c; } mr.Next(); } ((GeneralMatrix&)*this).tDelete(); return maxval; } Real GeneralMatrix::Minimum2(int& i, int& j) const { REPORT if (storage == 0) NullMatrixError(this); Real minval = FloatingPointPrecision::Maximum(); int nr = Nrows(); MatrixRow mr((GeneralMatrix*)this, LoadOnEntry+DirectPart); for (int r = 1; r <= nr; r++) { int c; minval = mr.Minimum1(minval, c); if (c > 0) { i = r; j = c; } mr.Next(); } ((GeneralMatrix&)*this).tDelete(); return minval; } Real Matrix::MaximumAbsoluteValue2(int& i, int& j) const { REPORT int k; Real m = GeneralMatrix::MaximumAbsoluteValue1(k); k--; i = k / Ncols(); j = k - i * Ncols(); i++; j++; return m; } Real Matrix::MinimumAbsoluteValue2(int& i, int& j) const { REPORT int k; Real m = GeneralMatrix::MinimumAbsoluteValue1(k); k--; i = k / Ncols(); j = k - i * Ncols(); i++; j++; return m; } Real Matrix::Maximum2(int& i, int& j) const { REPORT int k; Real m = GeneralMatrix::Maximum1(k); k--; i = k / Ncols(); j = k - i * Ncols(); i++; j++; return m; } Real Matrix::Minimum2(int& i, int& j) const { REPORT int k; Real m = GeneralMatrix::Minimum1(k); k--; i = k / Ncols(); j = k - i * Ncols(); i++; j++; return m; } Real SymmetricMatrix::SumSquare() const { REPORT Real sum1 = 0.0; Real sum2 = 0.0; Real* s = store; int nr = nrows; for (int i = 0; iSumSquare(); return s; } Real BaseMatrix::NormFrobenius() const { REPORT return sqrt(SumSquare()); } Real BaseMatrix::SumAbsoluteValue() const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->SumAbsoluteValue(); return s; } Real BaseMatrix::Sum() const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->Sum(); return s; } Real BaseMatrix::MaximumAbsoluteValue() const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->MaximumAbsoluteValue(); return s; } Real BaseMatrix::MaximumAbsoluteValue1(int& i) const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->MaximumAbsoluteValue1(i); return s; } Real BaseMatrix::MaximumAbsoluteValue2(int& i, int& j) const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->MaximumAbsoluteValue2(i, j); return s; } Real BaseMatrix::MinimumAbsoluteValue() const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->MinimumAbsoluteValue(); return s; } Real BaseMatrix::MinimumAbsoluteValue1(int& i) const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->MinimumAbsoluteValue1(i); return s; } Real BaseMatrix::MinimumAbsoluteValue2(int& i, int& j) const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->MinimumAbsoluteValue2(i, j); return s; } Real BaseMatrix::Maximum() const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->Maximum(); return s; } Real BaseMatrix::Maximum1(int& i) const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->Maximum1(i); return s; } Real BaseMatrix::Maximum2(int& i, int& j) const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->Maximum2(i, j); return s; } Real BaseMatrix::Minimum() const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->Minimum(); return s; } Real BaseMatrix::Minimum1(int& i) const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->Minimum1(i); return s; } Real BaseMatrix::Minimum2(int& i, int& j) const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); Real s = gm->Minimum2(i, j); return s; } Real DotProduct(const Matrix& A, const Matrix& B) { REPORT int n = A.storage; if (n != B.storage) Throw(IncompatibleDimensionsException(A,B)); Real sum = 0.0; Real* a = A.store; Real* b = B.store; while (n--) sum += *a++ * *b++; return sum; } Real Matrix::Trace() const { REPORT Tracer trace("Trace"); int i = nrows; int d = i+1; if (i != ncols) Throw(NotSquareException(*this)); Real sum = 0.0; Real* s = store; // while (i--) { sum += *s; s += d; } if (i) for (;;) { sum += *s; if (!(--i)) break; s += d; } ((GeneralMatrix&)*this).tDelete(); return sum; } Real DiagonalMatrix::Trace() const { REPORT int i = nrows; Real sum = 0.0; Real* s = store; while (i--) sum += *s++; ((GeneralMatrix&)*this).tDelete(); return sum; } Real SymmetricMatrix::Trace() const { REPORT int i = nrows; Real sum = 0.0; Real* s = store; int j = 2; // while (i--) { sum += *s; s += j++; } if (i) for (;;) { sum += *s; if (!(--i)) break; s += j++; } ((GeneralMatrix&)*this).tDelete(); return sum; } Real LowerTriangularMatrix::Trace() const { REPORT int i = nrows; Real sum = 0.0; Real* s = store; int j = 2; // while (i--) { sum += *s; s += j++; } if (i) for (;;) { sum += *s; if (!(--i)) break; s += j++; } ((GeneralMatrix&)*this).tDelete(); return sum; } Real UpperTriangularMatrix::Trace() const { REPORT int i = nrows; Real sum = 0.0; Real* s = store; while (i) { sum += *s; s += i--; } // won t cause a problem ((GeneralMatrix&)*this).tDelete(); return sum; } Real BandMatrix::Trace() const { REPORT int i = nrows; int w = lower+upper+1; Real sum = 0.0; Real* s = store+lower; // while (i--) { sum += *s; s += w; } if (i) for (;;) { sum += *s; if (!(--i)) break; s += w; } ((GeneralMatrix&)*this).tDelete(); return sum; } Real SymmetricBandMatrix::Trace() const { REPORT int i = nrows; int w = lower+1; Real sum = 0.0; Real* s = store+lower; // while (i--) { sum += *s; s += w; } if (i) for (;;) { sum += *s; if (!(--i)) break; s += w; } ((GeneralMatrix&)*this).tDelete(); return sum; } Real IdentityMatrix::Trace() const { Real sum = *store * nrows; ((GeneralMatrix&)*this).tDelete(); return sum; } Real BaseMatrix::Trace() const { REPORT MatrixType Diag = MatrixType::Dg; Diag.SetDataLossOK(); GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(Diag); Real sum = gm->Trace(); return sum; } void LogAndSign::operator*=(Real x) { if (x > 0.0) { log_value += log(x); } else if (x < 0.0) { log_value += log(-x); sign = -sign; } else sign = 0; } void LogAndSign::PowEq(int k) { if (sign) { log_value *= k; if ( (k & 1) == 0 ) sign = 1; } } Real LogAndSign::Value() const { Tracer et("LogAndSign::Value"); if (log_value >= FloatingPointPrecision::LnMaximum()) Throw(OverflowException("Overflow in exponential")); return sign * exp(log_value); } LogAndSign::LogAndSign(Real f) { if (f == 0.0) { log_value = 0.0; sign = 0; return; } else if (f < 0.0) { sign = -1; f = -f; } else sign = 1; log_value = log(f); } LogAndSign DiagonalMatrix::LogDeterminant() const { REPORT int i = nrows; LogAndSign sum; Real* s = store; while (i--) sum *= *s++; ((GeneralMatrix&)*this).tDelete(); return sum; } LogAndSign LowerTriangularMatrix::LogDeterminant() const { REPORT int i = nrows; LogAndSign sum; Real* s = store; int j = 2; // while (i--) { sum *= *s; s += j++; } if (i) for(;;) { sum *= *s; if (!(--i)) break; s += j++; } ((GeneralMatrix&)*this).tDelete(); return sum; } LogAndSign UpperTriangularMatrix::LogDeterminant() const { REPORT int i = nrows; LogAndSign sum; Real* s = store; while (i) { sum *= *s; s += i--; } ((GeneralMatrix&)*this).tDelete(); return sum; } LogAndSign IdentityMatrix::LogDeterminant() const { REPORT int i = nrows; LogAndSign sum; if (i > 0) { sum = *store; sum.PowEq(i); } ((GeneralMatrix&)*this).tDelete(); return sum; } LogAndSign BaseMatrix::LogDeterminant() const { REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); LogAndSign sum = gm->LogDeterminant(); return sum; } LogAndSign GeneralMatrix::LogDeterminant() const { REPORT Tracer tr("LogDeterminant"); if (nrows != ncols) Throw(NotSquareException(*this)); CroutMatrix C(*this); return C.LogDeterminant(); } LogAndSign CroutMatrix::LogDeterminant() const { REPORT if (sing) return 0.0; int i = nrows; int dd = i+1; LogAndSign sum; Real* s = store; if (i) for(;;) { sum *= *s; if (!(--i)) break; s += dd; } if (!d) sum.ChangeSign(); return sum; } Real BaseMatrix::Determinant() const { REPORT Tracer tr("Determinant"); REPORT GeneralMatrix* gm = ((BaseMatrix&)*this).Evaluate(); LogAndSign ld = gm->LogDeterminant(); return ld.Value(); } LinearEquationSolver::LinearEquationSolver(const BaseMatrix& bm) { gm = ( ((BaseMatrix&)bm).Evaluate() )->MakeSolver(); if (gm==&bm) { REPORT gm = gm->Image(); } // want a copy if *gm is actually bm else { REPORT gm->Protect(); } } #ifdef use_namespace } #endif