Hi Regina,
On Sun, 2010-10-24 at 18:13 +0200, Regina Henschel wrote:
> I'm currently working on LINEST and have attached a draft to issue
Cool ! this is an awesome patch :-)
> There is no mathematical problem, but I'm uncertain about coding style.
And your coding style looks beautiful, a great improvement over what
was there before.
> The algorithms work on matrices. They have a lot of parts which are
> nearly identical but the matrices are transposed. How to handle that?
Good question; and one to which I don't know the answer (sadly) -
whatever uses the least code, and yet performs well I suppose.
> And I do not know how much comments are needed for those, who have to
> maintain the code later, and if a separate documentation is needed.
Your level of commenting is great.
> What are your opinions?
First - I'm sorry it took so long to get back to you; sad. Secondly
I've turned your changes into a patch (which I append). I split your
changes out into a new module - so that this (nice new) piece can be
licensed under the LGPLv3+/MPL combination - if you're happy with that.
Finally - you updated the CheckMatrix signature, but I didn't see the
header change for that; it'd be great to expand on the diff to include
those bits & return it.
It'd be wonderful to have your changes included - though we have a
feature freeze in 2 days now ;-)
Many thanks & looking forward to working with you [ do you hang out on
IRC ? it'd be great to introduce yourself there #libreoffice on
irc.freenode.net ].
Thanks !
Michael.
PS. do you think a self contained regression test in sc/qa/unit/ would
be good for this ? or does that belong better in a spreadsheet we can
load and compare results in ? [ it would be good to get some of them
into qa/unit so we can load / calculate and check the answers during
build I suspect ].
--
[email protected] <><, Pseudo Engineer, itinerant idiot
diff --git a/sc/source/core/tool/interpr5.cxx b/sc/source/core/tool/interpr5.cxx
index ff72817..6fe6546 100644
--- a/sc/source/core/tool/interpr5.cxx
+++ b/sc/source/core/tool/interpr5.cxx
@@ -34,7 +34,6 @@
#include <rtl/math.hxx>
#include <rtl/logfile.hxx>
#include <string.h>
-#include <math.h>
#include <stdio.h>
#if OSL_DEBUG_LEVEL > 1
@@ -2060,302 +2059,6 @@ void ScInterpreter::ScRGP()
RTL_LOGFILE_CONTEXT_AUTHOR( aLogger, "sc", "er", "ScInterpreter::ScRGP" );
CalulateRGPRKP(FALSE);
}
-bool ScInterpreter::CheckMatrix(BOOL _bLOG,BOOL _bTrendGrowth,BYTE& nCase,SCSIZE& nCX,SCSIZE& nCY,SCSIZE& nRX,SCSIZE& nRY,SCSIZE& M,SCSIZE& N,ScMatrixRef& pMatX,ScMatrixRef& pMatY)
-{
- nCX = 0;
- nCY = 0;
- nRX = 0;
- nRY = 0;
- M = 0;
- N = 0;
- pMatY->GetDimensions(nCY, nRY);
- const SCSIZE nCountY = nCY * nRY;
- for ( SCSIZE i = 0; i < nCountY; i++ )
- {
- if (!pMatY->IsValue(i))
- {
- PushIllegalArgument();
- return false;
- }
- }
-
- if ( _bLOG )
- {
- ScMatrixRef pNewY = pMatY->CloneIfConst();
- for (SCSIZE nElem = 0; nElem < nCountY; nElem++)
- {
- const double fVal = pNewY->GetDouble(nElem);
- if (fVal <= 0.0)
- {
- PushIllegalArgument();
- return false;
- }
- else
- pNewY->PutDouble(log(fVal), nElem);
- }
- pMatY = pNewY;
- }
-
- if (pMatX)
- {
- pMatX->GetDimensions(nCX, nRX);
- const SCSIZE nCountX = nCX * nRX;
- for ( SCSIZE i = 0; i < nCountX; i++ )
- if (!pMatX->IsValue(i))
- {
- PushIllegalArgument();
- return false;
- }
- if (nCX == nCY && nRX == nRY)
- nCase = 1; // einfache Regression
- else if (nCY != 1 && nRY != 1)
- {
- PushIllegalArgument();
- return false;
- }
- else if (nCY == 1)
- {
- if (nRX != nRY)
- {
- PushIllegalArgument();
- return false;
- }
- else
- {
- nCase = 2; // zeilenweise
- N = nRY;
- M = nCX;
- }
- }
- else if (nCX != nCY)
- {
- PushIllegalArgument();
- return false;
- }
- else
- {
- nCase = 3; // spaltenweise
- N = nCY;
- M = nRX;
- }
- }
- else
- {
- pMatX = GetNewMat(nCY, nRY);
- if ( _bTrendGrowth )
- {
- nCX = nCY;
- nRX = nRY;
- }
- if (!pMatX)
- {
- PushIllegalArgument();
- return false;
- }
- for ( SCSIZE i = 1; i <= nCountY; i++ )
- pMatX->PutDouble((double)i, i-1);
- nCase = 1;
- }
- return true;
-}
-void ScInterpreter::CalulateRGPRKP(BOOL _bRKP)
-{
- RTL_LOGFILE_CONTEXT_AUTHOR( aLogger, "sc", "er", "ScInterpreter::CheckMatrix" );
- BYTE nParamCount = GetByte();
- if ( !MustHaveParamCount( nParamCount, 1, 4 ) )
- return;
- BOOL bConstant, bStats;
- if (nParamCount == 4)
- bStats = GetBool();
- else
- bStats = FALSE;
- if (nParamCount >= 3)
- bConstant = GetBool();
- else
- bConstant = TRUE;
- ScMatrixRef pMatX;
- ScMatrixRef pMatY;
- if (nParamCount >= 2)
- pMatX = GetMatrix();
- else
- pMatX = NULL;
- pMatY = GetMatrix();
- if (!pMatY)
- {
- PushIllegalParameter();
- return;
- } // if (!pMatY)
- BYTE nCase; // 1 = normal, 2,3 = mehrfach
- SCSIZE nCX, nCY;
- SCSIZE nRX, nRY;
- SCSIZE M = 0, N = 0;
- if ( !CheckMatrix(_bRKP,FALSE,nCase,nCX,nCY,nRX,nRY,M,N,pMatX,pMatY) )
- return;
-
- ScMatrixRef pResMat;
- if (nCase == 1)
- {
- if (!bStats)
- pResMat = GetNewMat(2,1);
- else
- pResMat = GetNewMat(2,5);
- if (!pResMat)
- {
- PushIllegalArgument();
- return;
- }
- double fCount = 0.0;
- double fSumX = 0.0;
- double fSumSqrX = 0.0;
- double fSumY = 0.0;
- double fSumSqrY = 0.0;
- double fSumXY = 0.0;
- double fValX, fValY;
- for (SCSIZE i = 0; i < nCY; i++)
- for (SCSIZE j = 0; j < nRY; j++)
- {
- fValX = pMatX->GetDouble(i,j);
- fValY = pMatY->GetDouble(i,j);
- fSumX += fValX;
- fSumSqrX += fValX * fValX;
- fSumY += fValY;
- fSumSqrY += fValY * fValY;
- fSumXY += fValX*fValY;
- fCount++;
- }
- if (fCount < 1.0)
- PushNoValue();
- else
- {
- double f1 = fCount*fSumXY-fSumX*fSumY;
- double fX = fCount*fSumSqrX-fSumX*fSumX;
- double b, m;
- if (bConstant)
- {
- b = fSumY/fCount - f1/fX*fSumX/fCount;
- m = f1/fX;
- }
- else
- {
- b = 0.0;
- m = fSumXY/fSumSqrX;
- }
- pResMat->PutDouble(_bRKP ? exp(m) : m, 0, 0);
- pResMat->PutDouble(_bRKP ? exp(b) : b, 1, 0);
- if (bStats)
- {
- double fY = fCount*fSumSqrY-fSumY*fSumY;
- double fSyx = fSumSqrY-b*fSumY-m*fSumXY;
- double fR2 = f1*f1/(fX*fY);
- pResMat->PutDouble (fR2, 0, 2);
- if (fCount < 3.0)
- {
- pResMat->PutString(ScGlobal::GetRscString(STR_NO_VALUE), 0, 1 );
- pResMat->PutString(ScGlobal::GetRscString(STR_NO_VALUE), 1, 1 );
- pResMat->PutString(ScGlobal::GetRscString(STR_NO_VALUE), 1, 2 );
- pResMat->PutString(ScGlobal::GetRscString(STR_NO_VALUE), 0, 3 );
- }
- else
- {
- pResMat->PutDouble(sqrt(fSyx*fCount/(fX*(fCount-2.0))), 0, 1);
- pResMat->PutDouble(sqrt(fSyx*fSumSqrX/fX/(fCount-2.0)), 1, 1);
- pResMat->PutDouble(
- sqrt((fCount*fSumSqrY - fSumY*fSumY - f1*f1/fX)/
- (fCount*(fCount-2.0))), 1, 2);
- if (fR2 == 1.0)
- pResMat->PutString(ScGlobal::GetRscString(STR_NO_VALUE), 0, 3 );
- else
- pResMat->PutDouble(fR2*(fCount-2.0)/(1.0-fR2), 0, 3);
- }
- pResMat->PutDouble(((double)(nCY*nRY))-2.0, 1, 3);
- pResMat->PutDouble(fY/fCount-fSyx, 0, 4);
- pResMat->PutDouble(fSyx, 1, 4);
- }
- }
- } // if (nCase == 1)
- if ( nCase != 1 )
- {
- SCSIZE i, j, k;
- if (!bStats)
- pResMat = GetNewMat(M+1,1);
- else
- pResMat = GetNewMat(M+1,5);
- if (!pResMat)
- {
- PushIllegalArgument();
- return;
- }
- ScMatrixRef pQ = GetNewMat(M+1, M+2);
- ScMatrixRef pE = GetNewMat(M+2, 1);
- ScMatrixRef pV = GetNewMat(M+1, 1);
- pE->PutDouble(0.0, M+1);
- pQ->FillDouble(0.0, 0, 0, M, M+1);
- if (nCase == 2)
- {
- for (k = 0; k < N; k++)
- {
- double Yk = pMatY->GetDouble(k);
- pE->PutDouble( pE->GetDouble(M+1)+Yk*Yk, M+1 );
- double sumYk = pQ->GetDouble(0, M+1) + Yk;
- pQ->PutDouble( sumYk, 0, M+1 );
- pE->PutDouble( sumYk, 0 );
- for (i = 0; i < M; i++)
- {
- double Xik = pMatX->GetDouble(i,k);
- double sumXik = pQ->GetDouble(0, i+1) + Xik;
- pQ->PutDouble( sumXik, 0, i+1);
- pQ->PutDouble( sumXik, i+1, 0);
- double sumXikYk = pQ->GetDouble(i+1, M+1) + Xik * Yk;
- pQ->PutDouble( sumXikYk, i+1, M+1);
- pE->PutDouble( sumXikYk, i+1);
- for (j = i; j < M; j++)
- {
- const double fVal = pMatX->GetDouble(j,k);
- double sumXikXjk = pQ->GetDouble(j+1, i+1) +
- Xik * fVal;
- pQ->PutDouble( sumXikXjk, j+1, i+1);
- pQ->PutDouble( sumXikXjk, i+1, j+1);
- }
- }
- }
- }
- else
- {
- for (k = 0; k < N; k++)
- {
- double Yk = pMatY->GetDouble(k);
- pE->PutDouble( pE->GetDouble(M+1)+Yk*Yk, M+1 );
- double sumYk = pQ->GetDouble(0, M+1) + Yk;
- pQ->PutDouble( sumYk, 0, M+1 );
- pE->PutDouble( sumYk, 0 );
- for (i = 0; i < M; i++)
- {
- double Xki = pMatX->GetDouble(k,i);
- double sumXki = pQ->GetDouble(0, i+1) + Xki;
- pQ->PutDouble( sumXki, 0, i+1);
- pQ->PutDouble( sumXki, i+1, 0);
- double sumXkiYk = pQ->GetDouble(i+1, M+1) + Xki * Yk;
- pQ->PutDouble( sumXkiYk, i+1, M+1);
- pE->PutDouble( sumXkiYk, i+1);
- for (j = i; j < M; j++)
- {
- const double fVal = pMatX->GetDouble(k,j);
- double sumXkiXkj = pQ->GetDouble(j+1, i+1) +
- Xki * fVal;
- pQ->PutDouble( sumXkiXkj, j+1, i+1);
- pQ->PutDouble( sumXkiXkj, i+1, j+1);
- }
- }
- }
- }
- if ( !Calculate4(_bRKP,pResMat,pQ,bConstant,N,M) )
- return;
-
- if (bStats)
- Calculate(pResMat,pE,pQ,pV,pMatX,bConstant,N,M,nCase);
- }
- PushMatrix(pResMat);
-}
void ScInterpreter::ScRKP()
{
diff --git a/sc/source/core/tool/interpr7.cxx b/sc/source/core/tool/interpr7.cxx
new file mode 100644
index 0000000..4b2c8b0
--- /dev/null
+++ b/sc/source/core/tool/interpr7.cxx
@@ -0,0 +1,1013 @@
+// MARKER(update_precomp.py): autogen include statement, do not remove
+#include "precompiled_sc.hxx"
+
+#include <rtl/math.hxx>
+#include "interpre.hxx"
+#include "globstr.hrc"
+
+// -----------------------------------------------------------------------------
+// Helper methods for LINEST/LOGEST and TREND/GROWTH
+// All matrices must already exist and have the needed size, no control tests
+// done. Those methodes, which names start with lcl_T, are adapted to case 3,
+// where Y (=observed values) is given as row.
+// Remember, ScMatrix matrices are zero based, index access (column,row).
+// -----------------------------------------------------------------------------
+
+namespace {
+// Multiply n x m Mat A with m x l Mat B to n x l Mat R
+void lcl_MFastMult(ScMatrixRef pA, ScMatrixRef pB, ScMatrixRef pR,
+ SCSIZE n, SCSIZE m, SCSIZE l)
+
+{
+ double sum;
+ for (SCSIZE row = 0; row < n; row++)
+ {
+ for (SCSIZE col = 0; col < l; col++)
+ { // result element(col, row) =sum[ (row of A) * (column of B)]
+ sum = 0.0;
+ for (SCSIZE k = 0; k < m; k++)
+ sum += pA->GetDouble(k,row) * pB->GetDouble(col,k);
+ pR->PutDouble(sum, col, row);
+ }
+ }
+}
+
+// <A;B> over all elements; uses the matrices as vectors of length M
+double lcl_GetSumProduct(ScMatrixRef pMatA, ScMatrixRef pMatB, SCSIZE nM)
+{
+ double fSum = 0.0;
+ for (SCSIZE i=0; i<nM; i++)
+ fSum += pMatA->GetDouble(i) * pMatB->GetDouble(i);
+ return fSum;
+}
+
+// Special version for use within QR decomposition.
+// Euclidean norm of column index C starting in row index R;
+// matrix A has count N rows.
+double lcl_GetColumnEuclideanNorm(ScMatrixRef pMatA, SCSIZE nC, SCSIZE nR, SCSIZE nN)
+{
+ double fNorm = 0.0;
+ for (SCSIZE row=nR; row<nN; row++)
+ fNorm += (pMatA->GetDouble(nC,row)) * (pMatA->GetDouble(nC,row));
+ return sqrt(fNorm);
+}
+
+// Euclidean norm of row index R starting in column index C;
+// matrix A has count N columns.
+double lcl_TGetColumnEuclideanNorm(ScMatrixRef pMatA, SCSIZE nR, SCSIZE nC, SCSIZE nN)
+{
+ double fNorm = 0.0;
+ for (SCSIZE col=nC; col<nN; col++)
+ fNorm += (pMatA->GetDouble(col,nR)) * (pMatA->GetDouble(col,nR));
+ return sqrt(fNorm);
+}
+
+// Special version for use within QR decomposition.
+// Maximum norm of column index C starting in row index R;
+// matrix A has count N rows.
+double lcl_GetColumnMaximumNorm(ScMatrixRef pMatA, SCSIZE nC, SCSIZE nR, SCSIZE nN)
+{
+ double fNorm = 0.0;
+ for (SCSIZE row=nR; row<nN; row++)
+ if (fNorm < fabs(pMatA->GetDouble(nC,row)))
+ fNorm = fabs(pMatA->GetDouble(nC,row));
+ return fNorm;
+}
+
+// Maximum norm of row index R starting in col index C;
+// matrix A has count N columns.
+double lcl_TGetColumnMaximumNorm(ScMatrixRef pMatA, SCSIZE nR, SCSIZE nC, SCSIZE nN)
+{
+ double fNorm = 0.0;
+ for (SCSIZE col=nC; col<nN; col++)
+ if (fNorm < fabs(pMatA->GetDouble(col,nR)))
+ fNorm = fabs(pMatA->GetDouble(col,nR));
+ return fNorm;
+}
+
+// Special version for use within QR decomposition.
+// <A(Ca);B(Cb)> starting in row index R;
+// Ca and Cb are indices of columns, matrices A and B have count N rows.
+double lcl_GetColumnSumProduct(ScMatrixRef pMatA, SCSIZE nCa,
+ ScMatrixRef pMatB, SCSIZE nCb, SCSIZE nR, SCSIZE nN)
+{
+ double fResult = 0.0;
+ for (SCSIZE row=nR; row<nN; row++)
+ fResult += pMatA->GetDouble(nCa,row) * pMatB->GetDouble(nCb,row);
+ return fResult;
+}
+
+// <A(Ra);B(Rb)> starting in column index C;
+// Ra and Rb are indices of rows, matrices A and B have count N columns.
+double lcl_TGetColumnSumProduct(ScMatrixRef pMatA, SCSIZE nRa,
+ ScMatrixRef pMatB, SCSIZE nRb, SCSIZE nC, SCSIZE nN)
+{
+ double fResult = 0.0;
+ for (SCSIZE col=nC; col<nN; col++)
+ fResult += pMatA->GetDouble(col,nRa) * pMatB->GetDouble(col,nRb);
+ return fResult;
+}
+
+double lcl_GetSign(double fValue)
+{
+ if (fValue < 0.0)
+ return -1.0;
+ else if (fValue > 0.0)
+ return 1.0;
+ else
+ return 0.0;
+}
+
+/* Calculates a QR decomposition with Householder reflection.
+ * For each NxK matrix A exists a decomposition A=Q*R with an orthogonal
+ * NxN matrix Q and a NxK matrix R.
+ * Q=H1*H2*...*Hk with Householder matrices H. Such a householder matrix can
+ * be build from a vector u by H=I-(2/u'u)*(u u'). This vectors u are returned
+ * in the columns of matrix A, overwriting the old content.
+ * The matrix R has a quadric upper part KxK with values in the upper right
+ * triangle and zeros in all other elements. Here the diagonal elements of R
+ * are stored in the vector R and the other upper right elements in the upper
+ * right of the matrix A.
+ * The function returns false, if calculation breaks. But because of round-off
+ * errors singularity is often not detected.
+ */
+bool lcl_CalculateQRdecomposition(ScMatrixRef pMatA,
+ ::std::vector< double>& pVecR, SCSIZE nK, SCSIZE nN)
+{
+ double fScale ;
+ double fEuclid ;
+ double fFactor ;
+ double fSignum ;
+ double fSum ;
+ // ScMatrix matrices are zero based, index access (column,row)
+ for (SCSIZE col = 0; col <nK; col++)
+ {
+ // calculate vector u of the householder transformation
+ fScale = lcl_GetColumnMaximumNorm(pMatA, col, col, nN);
+ if (fScale == 0.0)
+ {
+ // A is singular
+ return false;
+ }
+ for (SCSIZE row = col; row <nN; row++)
+ pMatA->PutDouble( pMatA->GetDouble(col,row)/fScale, col, row);
+
+ fEuclid = lcl_GetColumnEuclideanNorm(pMatA, col, col, nN);
+ fFactor = 1.0/fEuclid/(fEuclid + fabs(pMatA->GetDouble(col,col)));
+ fSignum = lcl_GetSign(pMatA->GetDouble(col,col));
+ pMatA->PutDouble( pMatA->GetDouble(col,col) + fSignum*fEuclid, col,col);
+ pVecR[col] = -fSignum * fScale * fEuclid;
+
+ // apply Householder transformation to A
+ for (SCSIZE c=col+1; c<nK; c++)
+ {
+ fSum =lcl_GetColumnSumProduct(pMatA, col, pMatA, c, col, nN);
+ for (SCSIZE row = col; row <nN; row++)
+ pMatA->PutDouble( pMatA->GetDouble(c,row)
+ - fSum * fFactor * pMatA->GetDouble(col,row), c, row);
+ }
+ }
+ return true;
+}
+
+// same with transposed matrix A, N is count of columns, K count of rows
+bool lcl_TCalculateQRdecomposition(ScMatrixRef pMatA,
+ ::std::vector< double>& pVecR, SCSIZE nK, SCSIZE nN)
+{
+ double fScale ;
+ double fEuclid ;
+ double fFactor ;
+ double fSignum ;
+ double fSum ;
+ // ScMatrix matrices are zero based, index access (column,row)
+ for (SCSIZE row = 0; row <nK; row++)
+ {
+ // calculate vector u of the householder transformation
+ fScale = lcl_TGetColumnMaximumNorm(pMatA, row, row, nN);
+ if (fScale == 0.0)
+ {
+ // A is singular
+ return false;
+ }
+ for (SCSIZE col = row; col <nN; col++)
+ pMatA->PutDouble( pMatA->GetDouble(col,row)/fScale, col, row);
+
+ fEuclid = lcl_TGetColumnEuclideanNorm(pMatA, row, row, nN);
+ fFactor = 1.0/fEuclid/(fEuclid + fabs(pMatA->GetDouble(row,row)));
+ fSignum = lcl_GetSign(pMatA->GetDouble(row,row));
+ pMatA->PutDouble( pMatA->GetDouble(row,row) + fSignum*fEuclid, row,row);
+ pVecR[row] = -fSignum * fScale * fEuclid;
+
+ // apply Householder transformation to A
+ for (SCSIZE r=row+1; r<nK; r++)
+ {
+ fSum =lcl_TGetColumnSumProduct(pMatA, row, pMatA, r, row, nN);
+ for (SCSIZE col = row; col <nN; col++)
+ pMatA->PutDouble( pMatA->GetDouble(col,r)
+ - fSum * fFactor * pMatA->GetDouble(col,row), col, r);
+ }
+ }
+ return true;
+}
+
+
+/* Applies a Householder transformation to a column vector Y with is given as
+ * Nx1 Matrix. The Vektor u, from which the Householder transformation is build,
+ * is the column part in matrix A, with column index C, starting with row
+ * index C. A is the result of the QR decomposition as obtained from
+ * lcl_CaluclateQRdecomposition.
+ */
+void lcl_ApplyHouseholderTransformation(ScMatrixRef pMatA, SCSIZE nC,
+ ScMatrixRef pMatY, SCSIZE nN)
+{
+ // ScMatrix matrices are zero based, index access (column,row)
+ double fDenominator = lcl_GetColumnSumProduct(pMatA, nC, pMatA, nC, nC, nN);
+ double fNumerator = lcl_GetColumnSumProduct(pMatA, nC, pMatY, 0, nC, nN);
+ double fFactor = 2.0 * (fNumerator/fDenominator);
+ for (SCSIZE row = nC; row < nN; row++)
+ pMatY->PutDouble(
+ pMatY->GetDouble(row) - fFactor * pMatA->GetDouble(nC,row), row);
+}
+
+// Same with transposed matrices A and Y.
+void lcl_TApplyHouseholderTransformation(ScMatrixRef pMatA, SCSIZE nR,
+ ScMatrixRef pMatY, SCSIZE nN)
+{
+ // ScMatrix matrices are zero based, index access (column,row)
+ double fDenominator = lcl_TGetColumnSumProduct(pMatA, nR, pMatA, nR, nR, nN);
+ double fNumerator = lcl_TGetColumnSumProduct(pMatA, nR, pMatY, 0, nR, nN);
+ double fFactor = 2.0 * (fNumerator/fDenominator);
+ for (SCSIZE col = nR; col < nN; col++)
+ pMatY->PutDouble(
+ pMatY->GetDouble(col) - fFactor * pMatA->GetDouble(col,nR), col);
+}
+
+/* Solve for X in R*X=S using back substitution. The solution X overwrites S.
+ * Uses R from the result of the QR decomposition of a NxK matrix A.
+ * S is a column vector given as matrix, with at least elements on index
+ * 0 to K-1; elements on index>=K are ignored. Vector R must not have zero
+ * elements, no check is done.
+ */
+void lcl_SolveWithUpperRightTriangle(ScMatrixRef pMatA,
+ ::std::vector< double>& pVecR, ScMatrixRef pMatS,
+ SCSIZE nK, bool bIsTransposed)
+{
+ // ScMatrix matrices are zero based, index access (column,row)
+ double fSum;
+ SCSIZE row;
+ // SCSIZE is never negative, therefore test with rowp1=row+1
+ for (SCSIZE rowp1 = nK; rowp1>0; rowp1--)
+ {
+ row = rowp1-1;
+ fSum = pMatS->GetDouble(row);
+ for (SCSIZE col = rowp1; col<nK ; col++)
+ if (bIsTransposed)
+ fSum -= pMatA->GetDouble(row,col) * pMatS->GetDouble(col);
+ else
+ fSum -= pMatA->GetDouble(col,row) * pMatS->GetDouble(col);
+ pMatS->PutDouble( fSum / pVecR[row] , row);
+ }
+}
+
+/* Solve for X in R' * X= T using forward substitution. The solution X
+ * overwrites T. Uses R from the result of the QR decomposition of a NxK
+ * matrix A. T is a column vectors given as matrix, with at least elements on
+ * index 0 to K-1; elements on index>=K are ignored. Vector R must not have
+ * zero elements, no check is done.
+ */
+void lcl_SolveWithLowerLeftTriangle(ScMatrixRef pMatA,
+ ::std::vector< double>& pVecR, ScMatrixRef pMatT,
+ SCSIZE nK, bool bIsTransposed)
+{
+ // ScMatrix matrices are zero based, index access (column,row)
+ double fSum;
+ for (SCSIZE row = 0; row < nK; row++)
+ {
+ fSum = pMatT -> GetDouble(row);
+ for (SCSIZE col=0; col < row; col++)
+ {
+ if (bIsTransposed)
+ fSum -= pMatA->GetDouble(col,row) * pMatT->GetDouble(col);
+ else
+ fSum -= pMatA->GetDouble(row,col) * pMatT->GetDouble(col);
+ }
+ pMatT->PutDouble( fSum / pVecR[row] , row);
+ }
+}
+
+/* Calculates Z = R * B
+ * R is given in matrix A and vector VecR as obtained from the QR
+ * decompostion in lcl_CalculateQRdecomposition. B and Z are column vectors
+ * given as matrix with at least index 0 to K-1; elements on index>=K are
+ * not used.
+ */
+void lcl_ApplyUpperRightTriangle(ScMatrixRef pMatA,
+ ::std::vector< double>& pVecR, ScMatrixRef pMatB,
+ ScMatrixRef pMatZ, SCSIZE nK, bool bIsTransposed)
+{
+ // ScMatrix matrices are zero based, index access (column,row)
+ double fSum;
+ for (SCSIZE row = 0; row < nK; row++)
+ {
+ fSum = pVecR[row] * pMatB->GetDouble(row);
+ for (SCSIZE col = row+1; col < nK; col++)
+ if (bIsTransposed)
+ fSum += pMatA->GetDouble(row,col) * pMatB->GetDouble(col);
+ else
+ fSum += pMatA->GetDouble(col,row) * pMatB->GetDouble(col);
+ pMatZ->PutDouble( fSum, row);
+ }
+}
+
+double lcl_GetMeanOverAll(ScMatrixRef pMat, SCSIZE nN)
+{
+ double fSum = 0.0;
+ for (SCSIZE i=0 ; i<nN; i++)
+ fSum += pMat->GetDouble(i);
+ return fSum/static_cast<double>(nN);
+}
+
+// Calculates means of the columns of matrix X. X is a RxC matrix;
+// ResMat is a 1xC matrix (=row).
+void lcl_CalculateColumnMeans(ScMatrixRef pX, ScMatrixRef pResMat,
+ SCSIZE nC, SCSIZE nR)
+{
+ double fSum = 0.0;
+ for (SCSIZE i=0; i < nC; i++)
+ {
+ fSum =0.0;
+ for (SCSIZE k=0; k < nR; k++)
+ fSum += pX->GetDouble(i,k); // GetDouble(Column,Row)
+ pResMat ->PutDouble( fSum/static_cast<double>(nR),i);
+ }
+}
+
+// Calculates means of the rows of matrix X. X is a RxC matrix;
+// ResMat is a Rx1 matrix (=column).
+void lcl_CalculateRowMeans(ScMatrixRef pX, ScMatrixRef pResMat,
+ SCSIZE nC, SCSIZE nR)
+{
+ double fSum = 0.0;
+ for (SCSIZE k=0; k < nR; k++)
+ {
+ fSum =0.0;
+ for (SCSIZE i=0; i < nC; i++)
+ fSum += pX->GetDouble(i,k); // GetDouble(Column,Row)
+ pResMat ->PutDouble( fSum/static_cast<double>(nC),k);
+ }
+}
+
+void lcl_CalculateColumnsDelta(ScMatrixRef pMat, ScMatrixRef pColumnMeans,
+ SCSIZE nC, SCSIZE nR)
+{
+ for (SCSIZE i = 0; i < nC; i++)
+ for (SCSIZE k = 0; k < nR; k++)
+ pMat->PutDouble( ::rtl::math::approxSub
+ (pMat->GetDouble(i,k) , pColumnMeans->GetDouble(i) ) , i, k);
+}
+
+void lcl_CalculateRowsDelta(ScMatrixRef pMat, ScMatrixRef pRowMeans,
+ SCSIZE nC, SCSIZE nR)
+{
+ for (SCSIZE k = 0; k < nR; k++)
+ for (SCSIZE i = 0; i < nC; i++)
+ pMat->PutDouble( ::rtl::math::approxSub
+ ( pMat->GetDouble(i,k) , pRowMeans->GetDouble(k) ) , i, k);
+}
+
+// Case1 = simple regression
+// MatX = X - MeanX, MatY = Y - MeanY, y - haty = (y - MeanY) - (haty - MeanY)
+// = (y-MeanY)-((slope*x+a)-(slope*MeanX+a)) = (y-MeanY)-slope*(x-MeanX)
+double lcl_GetSSresid(ScMatrixRef pMatX, ScMatrixRef pMatY, double fSlope,
+ SCSIZE nN)
+{
+ double fSum = 0.0;
+ double fTemp = 0.0;
+ for (SCSIZE i=0; i<nN; i++)
+ {
+ fTemp = pMatY->GetDouble(i) - fSlope * pMatX->GetDouble(i);
+ fSum += fTemp * fTemp;
+ }
+ return fSum;
+}
+} // private namespace
+
+void ScInterpreter::CalulateRGPRKP(BOOL _bRKP)
+{
+ BYTE nParamCount = GetByte();
+ if ( !MustHaveParamCount( nParamCount, 1, 4 ) )
+ return;
+ bool bConstant, bStats;
+
+// optional forth parameter
+ if (nParamCount == 4)
+ bStats = GetBool();
+ else
+ bStats = false;
+
+// The third parameter may not be missing in ODF, if the forth parameter
+// is present.
+ if (nParamCount >= 3)
+ {
+ if (IsMissing())
+ {
+ PushIllegalParameter();
+ return;
+ }
+ else
+ bConstant = GetBool();
+ }
+ else
+ bConstant = TRUE;
+
+ ScMatrixRef pMatX;
+ if (nParamCount >= 2)
+ {
+ if (IsMissing())
+ { //In ODF1.2 empty second parameter (which is two ;; ) is allowed
+ Pop();
+ pMatX = NULL;
+ }
+ else
+ {
+ pMatX = GetMatrix();
+ }
+ }
+ else
+ pMatX = NULL;
+
+ ScMatrixRef pMatY;
+ pMatY = GetMatrix();
+ if (!pMatY)
+ {
+ PushIllegalParameter();
+ return;
+ }
+
+ // 1 = simple; 2 = multiple with Y as column; 3 = multiple with Y as row
+ BYTE nCase;
+
+ SCSIZE nCX, nCY; // number of columns
+ SCSIZE nRX, nRY; //number of rows
+ SCSIZE K = 0, N = 0; // K=number of variables X, N=number of data samples
+
+ // Fill default values in matrix X, transform Y to log(Y) in case _bLOGEST,
+ // determine sizes of matrices X and Y, determine kind of regression, clone
+ // Y in case bLOGEST, if constant.
+ if ( !CheckMatrix(_bRKP,nCase,nCX,nCY,nRX,nRY,K,N,pMatX,pMatY) )
+ {
+ PushIllegalParameter();
+ return;
+ }
+
+ // Enough data samples?
+ if ( (bConstant && (N<K+1)) || (!bConstant && (N<K)) || (N<1) || (K<1) )
+ {
+ PushIllegalParameter();
+ return;
+ }
+
+ ScMatrixRef pResMat;
+ if (bStats)
+ pResMat = GetNewMat(K+1,5);
+ else
+ pResMat = GetNewMat(K+1,1);
+ if (!pResMat)
+ {
+ PushError(errStackOverflow);
+ return;
+ }
+ // Fill unused cells in pResMat; order (column,row)
+ if (bStats)
+ {
+ for (SCSIZE i=2; i<K+1; i++)
+ {
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), i, 2 );
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), i, 3 );
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), i, 4 );
+ }
+ }
+
+ // Uses sum(x-MeanX)^2 and not [sum x^2]-N * MeanX^2 in case bConstant.
+ // Clone constant matrices, so that Mat = Mat - Mean is possible.
+ double fMeanY = 0.0;
+ if (bConstant)
+ {
+ ScMatrixRef pNewX = pMatX->CloneIfConst();
+ ScMatrixRef pNewY = pMatY->CloneIfConst();
+ if (!pNewX || !pNewY)
+ {
+ PushError(errStackOverflow);
+ return;
+ }
+ pMatX = pNewX;
+ pMatY = pNewY;
+ // DeltaY is possible here; DeltaX depends on nCase, so later
+ fMeanY = lcl_GetMeanOverAll(pMatY, N);
+ for (SCSIZE i=0; i<N; i++)
+ {
+ pMatY->PutDouble( ::rtl::math::approxSub(pMatY->GetDouble(i),fMeanY), i );
+ }
+ }
+
+ if (nCase==1)
+ {
+ // calculate simple regression
+ double fMeanX = 0.0;
+ if (bConstant)
+ { // Mat = Mat - Mean
+ fMeanX = lcl_GetMeanOverAll(pMatX, N);
+ for (SCSIZE i=0; i<N; i++)
+ {
+ pMatX->PutDouble( ::rtl::math::approxSub(pMatX->GetDouble(i),fMeanX), i );
+ }
+ }
+ double fSumXY = lcl_GetSumProduct(pMatX,pMatY,N);
+ double fSumX2 = lcl_GetSumProduct(pMatX,pMatX,N);
+ if (fSumX2==0.0)
+ {
+ PushNoValue(); // all x-values are identical
+ return;
+ }
+ double fSlope = fSumXY / fSumX2;
+ double fIntercept = 0.0;
+ if (bConstant)
+ fIntercept = fMeanY - fSlope * fMeanX;
+ pResMat->PutDouble(_bRKP ? exp(fIntercept) : fIntercept, 1, 0); //order (column,row)
+ pResMat->PutDouble(_bRKP ? exp(fSlope) : fSlope, 0, 0);
+
+ if (bStats)
+ {
+ double fSSreg = fSlope * fSlope * fSumX2;
+ pResMat->PutDouble(fSSreg, 0, 4);
+
+ double fDegreesFreedom =static_cast<double>( (bConstant) ? N-2 : N-1 );
+ pResMat->PutDouble(fDegreesFreedom, 1, 3);
+
+ double fSSresid = lcl_GetSSresid(pMatX,pMatY,fSlope,N);
+ pResMat->PutDouble(fSSresid, 1, 4);
+
+ if (fDegreesFreedom == 0.0 || fSSresid == 0.0 || fSSreg == 0.0)
+ { // exact fit; test SSreg too, because SSresid might be
+ // unequal zero due to round of errors
+ pResMat->PutDouble(0.0, 1, 4); // SSresid
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), 0, 3); // F
+ pResMat->PutDouble(0.0, 1, 2); // RMSE
+ pResMat->PutDouble(0.0, 0, 1); // SigmaSlope
+ if (bConstant)
+ pResMat->PutDouble(0.0, 1, 1); //SigmaIntercept
+ else
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), 1, 1);
+ pResMat->PutDouble(1.0, 0, 2); // R^2
+ }
+ else
+ {
+ double fFstatistic = (fSSreg / static_cast<double>(K))
+ / (fSSresid / fDegreesFreedom);
+ pResMat->PutDouble(fFstatistic, 0, 3);
+
+ // standard error of estimate
+ double fRMSE = sqrt(fSSresid / fDegreesFreedom);
+ pResMat->PutDouble(fRMSE, 1, 2);
+
+ double fSigmaSlope = fRMSE / sqrt(fSumX2);
+ pResMat->PutDouble(fSigmaSlope, 0, 1);
+
+ if (bConstant)
+ {
+ double fSigmaIntercept = fRMSE
+ * sqrt(fMeanX*fMeanX/fSumX2 + 1.0/static_cast<double>(N));
+ pResMat->PutDouble(fSigmaIntercept, 1, 1);
+ }
+ else
+ {
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), 1, 1);
+ }
+
+ double fR2 = fSSreg / (fSSreg + fSSresid);
+ pResMat->PutDouble(fR2, 0, 2);
+ }
+ }
+ PushMatrix(pResMat);
+ }
+ else // calculate multiple regression;
+ {
+ // Uses a QR decomposition X = QR. The solution B = (X'X)^(-1) * X' * Y
+ // becomes B = R^(-1) * Q' * Y
+ if (nCase ==2) // Y is column
+ {
+ ::std::vector< double> aVecR(N); // for QR decomposition
+ // Enough memory for needed matrices?
+ ScMatrixRef pMeans = GetNewMat(K, 1); // mean of each column
+ ScMatrixRef pMatZ; // for Q' * Y , inter alia
+ if (bStats)
+ pMatZ = pMatY->Clone(); // Y is used in statistic, keep it
+ else
+ pMatZ = pMatY; // Y can be overwritten
+ ScMatrixRef pSlopes = GetNewMat(1,K); // from b1 to bK
+ if (!pMeans || !pMatZ || !pSlopes)
+ {
+ PushError(errStackOverflow);
+ return;
+ }
+ if (bConstant)
+ {
+ lcl_CalculateColumnMeans(pMatX, pMeans, K, N);
+ lcl_CalculateColumnsDelta(pMatX, pMeans, K, N);
+ }
+ if (!lcl_CalculateQRdecomposition(pMatX, aVecR, K, N))
+ {
+ PushNoValue();
+ return;
+ }
+ // Later on we will divide by elements of aVecR, so make sure
+ // that they aren't zero.
+ bool bIsSingular=false;
+ for (SCSIZE row=0; row < K && !bIsSingular; row++)
+ bIsSingular = bIsSingular || aVecR[row]==0.0;
+ if (bIsSingular)
+ {
+ PushNoValue();
+ return;
+ }
+ // Z = Q' Y;
+ for (SCSIZE col = 0; col < K; col++)
+ {
+ lcl_ApplyHouseholderTransformation(pMatX, col, pMatZ, N);
+ }
+ // B = R^(-1) * Q' * Y <=> B = R^(-1) * Z <=> R * B = Z
+ // result Z should have zeros for index>=K; if not, ignore values
+ for (SCSIZE col = 0; col < K ; col++)
+ {
+ pSlopes->PutDouble( pMatZ->GetDouble(col), col);
+ }
+ lcl_SolveWithUpperRightTriangle(pMatX, aVecR, pSlopes, K, false);
+ double fIntercept = 0.0;
+ if (bConstant)
+ fIntercept = fMeanY - lcl_GetSumProduct(pMeans,pSlopes,K);
+ // Fill first line in result matrix
+ pResMat->PutDouble(_bRKP ? exp(fIntercept) : fIntercept, K, 0 );
+ for (SCSIZE i = 0; i < K; i++)
+ pResMat->PutDouble(_bRKP ? exp(pSlopes->GetDouble(i))
+ : pSlopes->GetDouble(i) , K-1-i, 0);
+
+
+ if (bStats)
+ {
+ double fSSreg = 0.0;
+ double fSSresid = 0.0;
+ // re-use memory of Z;
+ pMatZ->FillDouble(0.0, 0, 0, 0, N-1);
+ // Z = R * Slopes
+ lcl_ApplyUpperRightTriangle(pMatX, aVecR, pSlopes, pMatZ, K, false);
+ // Z = Q * Z, that is Q * R * Slopes = X * Slopes
+ for (SCSIZE colp1 = K; colp1 > 0; colp1--)
+ {
+ lcl_ApplyHouseholderTransformation(pMatX, colp1-1, pMatZ,N);
+ }
+ fSSreg =lcl_GetSumProduct(pMatZ, pMatZ, N);
+ // re-use Y for residuals, Y = Y-Z
+ for (SCSIZE row = 0; row < N; row++)
+ pMatY->PutDouble(pMatY->GetDouble(row) - pMatZ->GetDouble(row), row);
+ fSSresid = lcl_GetSumProduct(pMatY, pMatY, N);
+ pResMat->PutDouble(fSSreg, 0, 4);
+ pResMat->PutDouble(fSSresid, 1, 4);
+
+ double fDegreesFreedom =static_cast<double>( (bConstant) ? N-K-1 : N-K );
+ pResMat->PutDouble(fDegreesFreedom, 1, 3);
+
+ if (fDegreesFreedom == 0.0 || fSSresid == 0.0 || fSSreg == 0.0)
+ { // exact fit; incl. observed values Y are identical
+ pResMat->PutDouble(0.0, 1, 4); // SSresid
+ // F = (SSreg/K) / (SSresid/df) = #DIV/0!
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), 0, 3); // F
+ // RMSE = sqrt(SSresid / df) = sqrt(0 / df) = 0
+ pResMat->PutDouble(0.0, 1, 2); // RMSE
+ // SigmaSlope[i] = RMSE * sqrt(matrix[i,i]) = 0 * sqrt(...) = 0
+ for (SCSIZE i=0; i<K; i++)
+ pResMat->PutDouble(0.0, K-1-i, 1);
+
+ // SigmaIntercept = RMSE * sqrt(...) = 0
+ if (bConstant)
+ pResMat->PutDouble(0.0, K, 1); //SigmaIntercept
+ else
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), K, 1);
+
+ // R^2 = SSreg / (SSreg + SSresid) = 1.0
+ pResMat->PutDouble(1.0, 0, 2); // R^2
+ }
+ else
+ {
+ double fFstatistic = (fSSreg / static_cast<double>(K))
+ / (fSSresid / fDegreesFreedom);
+ pResMat->PutDouble(fFstatistic, 0, 3);
+
+ // standard error of estimate = root mean SSE
+ double fRMSE = sqrt(fSSresid / fDegreesFreedom);
+ pResMat->PutDouble(fRMSE, 1, 2);
+
+ // standard error of slopes
+ // = RMSE * sqrt(diagonal element of (R' R)^(-1) )
+ // standard error of intercept
+ // = RMSE * sqrt( Xmean * (R' R)^(-1) * Xmean' + 1/N)
+ // (R' R)^(-1) = R^(-1) * (R')^(-1). Do not calculate it as
+ // a whole matrix, but iterate over unit vectors.
+ double fSigmaSlope = 0.0;
+ double fSigmaIntercept = 0.0;
+ for (SCSIZE col = 0; col < K; col++)
+ {
+ //re-use memory of MatZ
+ pMatZ->FillDouble(0.0,0,0,0,K-1); // Z = unit vector e
+ pMatZ->PutDouble(1.0, col);
+ //Solve R' * Z = e
+ lcl_SolveWithLowerLeftTriangle(pMatX, aVecR, pMatZ, K, false);
+ // Solve R * Znew = Zold
+ lcl_SolveWithUpperRightTriangle(pMatX, aVecR, pMatZ, K, false);
+ // now Z is column col in (R' R)^(-1)
+ fSigmaSlope = fRMSE * sqrt(pMatZ->GetDouble(col));
+ pResMat->PutDouble(fSigmaSlope, K-1-col, 1);
+ // (R' R) ^(-1) is symmetric
+ if (bConstant)
+ fSigmaIntercept += lcl_GetSumProduct(pMeans, pMatZ, K);
+ }
+ if (bConstant)
+ {
+ fSigmaIntercept = fRMSE
+ * sqrt(fSigmaIntercept + 1.0 / static_cast<double>(N));
+ pResMat->PutDouble(fSigmaIntercept, K, 1);
+ }
+ else
+ {
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), K, 1);
+ }
+
+ double fR2 = fSSreg / (fSSreg + fSSresid);
+ pResMat->PutDouble(fR2, 0, 2);
+ } // end not exact fit
+ } //end bStats
+ PushMatrix(pResMat);
+ } //end nCase == 2
+ else // nCase == 3, Y is row, all matrices are transposed
+ {
+ ::std::vector< double> aVecR(N); // for QR decomposition
+ // Enough memory for needed matrices?
+ ScMatrixRef pMeans = GetNewMat(1, K); // mean of each row
+ ScMatrixRef pMatZ; // for Q' * Y , inter alia
+ if (bStats)
+ pMatZ = pMatY->Clone(); // Y is used in statistic, keep it
+ else
+ pMatZ = pMatY; // Y can be overwritten
+ ScMatrixRef pSlopes = GetNewMat(K,1); // from b1 to bK
+ if (!pMeans || !pMatZ || !pSlopes)
+ {
+ PushError(errStackOverflow);
+ return;
+ }
+ if (bConstant)
+ {
+ lcl_CalculateRowMeans(pMatX, pMeans, N, K);
+ lcl_CalculateRowsDelta(pMatX, pMeans, N, K);
+ }
+
+ if (!lcl_TCalculateQRdecomposition(pMatX, aVecR, K, N))
+ {
+ PushNoValue();
+ return;
+ }
+
+ // Later on we will divide by elements of aVecR, so make sure
+ // that they aren't zero.
+ bool bIsSingular=false;
+ for (SCSIZE row=0; row < K && !bIsSingular; row++)
+ bIsSingular = bIsSingular || aVecR[row]==0.0;
+ if (bIsSingular)
+ {
+ PushNoValue();
+ return;
+ }
+ // Z = Q' Y
+ for (SCSIZE row = 0; row < K; row++)
+ {
+ lcl_TApplyHouseholderTransformation(pMatX, row, pMatZ, N);
+ }
+ // B = R^(-1) * Q' * Y <=> B = R^(-1) * Z <=> R * B = Z
+ // result Z should have zeros for index>=K; if not, ignore values
+ for (SCSIZE col = 0; col < K ; col++)
+ {
+ pSlopes->PutDouble( pMatZ->GetDouble(col), col);
+ }
+ lcl_SolveWithUpperRightTriangle(pMatX, aVecR, pSlopes, K, true);
+ double fIntercept = 0.0;
+ if (bConstant)
+ fIntercept = fMeanY - lcl_GetSumProduct(pMeans,pSlopes,K);
+ // Fill first line in result matrix
+ pResMat->PutDouble(_bRKP ? exp(fIntercept) : fIntercept, K, 0 );
+ for (SCSIZE i = 0; i < K; i++)
+ pResMat->PutDouble(_bRKP ? exp(pSlopes->GetDouble(i))
+ : pSlopes->GetDouble(i) , K-1-i, 0);
+
+
+ if (bStats)
+ {
+ double fSSreg = 0.0;
+ double fSSresid = 0.0;
+ // re-use memory of Z;
+ pMatZ->FillDouble(0.0, 0, 0, N-1, 0);
+ // Z = R * Slopes
+ lcl_ApplyUpperRightTriangle(pMatX, aVecR, pSlopes, pMatZ, K, true);
+ // Z = Q * Z, that is Q * R * Slopes = X * Slopes
+ for (SCSIZE rowp1 = K; rowp1 > 0; rowp1--)
+ {
+ lcl_TApplyHouseholderTransformation(pMatX, rowp1-1, pMatZ,N);
+ }
+ fSSreg =lcl_GetSumProduct(pMatZ, pMatZ, N);
+ // re-use Y for residuals, Y = Y-Z
+ for (SCSIZE col = 0; col < N; col++)
+ pMatY->PutDouble(pMatY->GetDouble(col) - pMatZ->GetDouble(col), col);
+ fSSresid = lcl_GetSumProduct(pMatY, pMatY, N);
+ pResMat->PutDouble(fSSreg, 0, 4);
+ pResMat->PutDouble(fSSresid, 1, 4);
+
+ double fDegreesFreedom =static_cast<double>( (bConstant) ? N-K-1 : N-K );
+ pResMat->PutDouble(fDegreesFreedom, 1, 3);
+
+ if (fDegreesFreedom == 0.0 || fSSresid == 0.0 || fSSreg == 0.0)
+ { // exact fit; incl. case observed values Y are identical
+ pResMat->PutDouble(0.0, 1, 4); // SSresid
+ // F = (SSreg/K) / (SSresid/df) = #DIV/0!
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), 0, 3); // F
+ // RMSE = sqrt(SSresid / df) = sqrt(0 / df) = 0
+ pResMat->PutDouble(0.0, 1, 2); // RMSE
+ // SigmaSlope[i] = RMSE * sqrt(matrix[i,i]) = 0 * sqrt(...) = 0
+ for (SCSIZE i=0; i<K; i++)
+ pResMat->PutDouble(0.0, K-1-i, 1);
+
+ // SigmaIntercept = RMSE * sqrt(...) = 0
+ if (bConstant)
+ pResMat->PutDouble(0.0, K, 1); //SigmaIntercept
+ else
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), K, 1);
+
+ // R^2 = SSreg / (SSreg + SSresid) = 1.0
+ pResMat->PutDouble(1.0, 0, 2); // R^2
+ }
+ else
+ {
+ double fFstatistic = (fSSreg / static_cast<double>(K))
+ / (fSSresid / fDegreesFreedom);
+ pResMat->PutDouble(fFstatistic, 0, 3);
+
+ // standard error of estimate = root mean SSE
+ double fRMSE = sqrt(fSSresid / fDegreesFreedom);
+ pResMat->PutDouble(fRMSE, 1, 2);
+
+ // standard error of slopes
+ // = RMSE * sqrt(diagonal element of (R' R)^(-1) )
+ // standard error of intercept
+ // = RMSE * sqrt( Xmean * (R' R)^(-1) * Xmean' + 1/N)
+ // (R' R)^(-1) = R^(-1) * (R')^(-1). Do not calculate it as
+ // a whole matrix, but iterate over unit vectors.
+ double fSigmaSlope = 0.0;
+ double fSigmaIntercept = 0.0;
+ for (SCSIZE row = 0; row < K; row++)
+ {
+ //re-use memory of MatZ
+ pMatZ->FillDouble(0.0,0,0,K-1,0); // Z = unit vector e
+ pMatZ->PutDouble(1.0, row);
+ //Solve R' * Z = e
+ lcl_SolveWithLowerLeftTriangle(pMatX, aVecR, pMatZ, K, true);
+ // Solve R * Znew = Zold
+ lcl_SolveWithUpperRightTriangle(pMatX, aVecR, pMatZ, K, true);
+ // now Z is column col in (R' R)^(-1)
+ fSigmaSlope = fRMSE * sqrt(pMatZ->GetDouble(row));
+ pResMat->PutDouble(fSigmaSlope, K-1-row, 1);
+ // (R' R) ^(-1) is symmetric
+ if (bConstant)
+ fSigmaIntercept += lcl_GetSumProduct(pMeans, pMatZ, K);
+ }
+ if (bConstant)
+ {
+ fSigmaIntercept = fRMSE
+ * sqrt(fSigmaIntercept + 1.0 / static_cast<double>(N));
+ pResMat->PutDouble(fSigmaIntercept, K, 1);
+ }
+ else
+ {
+ pResMat->PutString(ScGlobal::GetRscString(STR_NV_STR), K, 1);
+ }
+
+ double fR2 = fSSreg / (fSSreg + fSSresid);
+ pResMat->PutDouble(fR2, 0, 2);
+ } // end not exact fit
+ } //end bStats
+ PushMatrix(pResMat);
+ } //end nCase == 3
+ } // end multiple regression
+ return;
+}
+
+
+// Changed CheckMatrix, which allways returns correct values for M and N
+bool ScInterpreter::CheckMatrix(BOOL _bLOG,BYTE& nCase,SCSIZE& nCX,SCSIZE& nCY,SCSIZE& nRX,SCSIZE& nRY,SCSIZE& M,SCSIZE& N,ScMatrixRef& pMatX,ScMatrixRef& pMatY)
+{
+ nCX = 0;
+ nCY = 0;
+ nRX = 0;
+ nRY = 0;
+ M = 0;
+ N = 0;
+ pMatY->GetDimensions(nCY, nRY);
+ const SCSIZE nCountY = nCY * nRY;
+ for ( SCSIZE i = 0; i < nCountY; i++ )
+ {
+ if (!pMatY->IsValue(i))
+ {
+ PushIllegalArgument();
+ return false;
+ }
+ }
+
+
+ if ( _bLOG )
+ {
+ ScMatrixRef pNewY = pMatY->CloneIfConst();
+ for (SCSIZE nElem = 0; nElem < nCountY; nElem++)
+ {
+ const double fVal = pNewY->GetDouble(nElem);
+ if (fVal <= 0.0)
+ {
+ PushIllegalArgument();
+ return false;
+ }
+ else
+ pNewY->PutDouble(log(fVal), nElem);
+ }
+ pMatY = pNewY;
+ }
+
+
+ if (pMatX)
+ {
+ pMatX->GetDimensions(nCX, nRX);
+ const SCSIZE nCountX = nCX * nRX;
+ for ( SCSIZE i = 0; i < nCountX; i++ )
+ if (!pMatX->IsValue(i))
+ {
+ PushIllegalArgument();
+ return false;
+ }
+ if (nCX == nCY && nRX == nRY)
+ {
+ nCase = 1; // simple regression
+ M = 1;
+ N = nCountY;
+ }
+ else if (nCY != 1 && nRY != 1)
+ {
+ PushIllegalArgument();
+ return false;
+ }
+ else if (nCY == 1)
+ {
+ if (nRX != nRY)
+ {
+ PushIllegalArgument();
+ return false;
+ }
+ else
+ {
+ nCase = 2; // Y is column
+ N = nRY;
+ M = nCX;
+ }
+ }
+ else if (nCX != nCY)
+ {
+ PushIllegalArgument();
+ return false;
+ }
+ else
+ {
+ nCase = 3; // Y is row
+ N = nCY;
+ M = nRX;
+ }
+ }
+ else
+ {
+ pMatX = GetNewMat(nCY, nRY);
+ nCX = nCY;
+ nRX = nRY;
+ if (!pMatX)
+ {
+ PushIllegalArgument();
+ return false;
+ }
+ for ( SCSIZE i = 1; i <= nCountY; i++ )
+ pMatX->PutDouble((double)i, i-1);
+ nCase = 1;
+ N = nCountY;
+ M = 1;
+ }
+ return true;
+}
diff --git a/sc/source/core/tool/makefile.mk b/sc/source/core/tool/makefile.mk
index b1cad58..bd241b3 100644
--- a/sc/source/core/tool/makefile.mk
+++ b/sc/source/core/tool/makefile.mk
@@ -82,6 +82,7 @@ SLOFILES = \
$(SLO)$/interpr4.obj \
$(SLO)$/interpr5.obj \
$(SLO)$/interpr6.obj \
+ $(SLO)$/interpr7.obj \
$(SLO)$/lookupcache.obj \
$(SLO)$/navicfg.obj \
$(SLO)$/odffmap.obj \
_______________________________________________
LibreOffice mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/libreoffice
