LeastSqImpl.mesa
Last Edited by: McCreight, April 23, 1985 6:31:34 pm PST
DIRECTORY
LeastSq, Random, Real;
LeastSqImpl: CEDAR PROGRAM IMPORTS Random, Real EXPORTS LeastSq =
BEGIN
NewPointSet: PUBLIC PROC [ dim: NAT ] RETURNS [ ps: LeastSq.PointSet ] = {
ps ← NEW[LeastSq.PointSetRec[dim]];
ps.count ← 0;
FOR row: NAT IN [0..dim) DO
ps[row] ← NEW[LeastSq.RowRec[dim]];
FOR col: NAT IN [0..dim) DO
ps[row][col] ← 0.0;
ENDLOOP;
ENDLOOP;
};
AddPoint: PUBLIC PROC [ ps: LeastSq.PointSet, pt: LeastSq.Row, weight: REAL ← 1.0 ] = {
IF pt.colDim # ps.rowDim THEN ERROR;
ps.count ← ps.count+1;
FOR row: NAT IN [0..ps.rowDim) DO
FOR col: NAT IN [0..row] DO
ps[row][col] ← ps[row][col]+weight*pt[row]*pt[col];
ENDLOOP;
ENDLOOP;
};
LeastSqFit: PUBLIC PROC [ ps: LeastSq.PointSet ] RETURNS [ eq: LeastSq.Row ] = {
At this point, ps[row][col] contains the sum over all input points of pt[row]*pt[col], where pt[k] =
1 if k=0,
an independent variable if k>0 and k<dim-1,
the dependent variable if k=dim-1.
We begin with a basis set of polynomials: 1, ind1, ind2, .... We use the Gram-Schmidt algorithm (see the textbook Numerical Analysis, second edition, by Johnson and Riess, Addison Wesley, 1982, pages 255 and following for details) to construct a new basis set of polynomials that is orthogonal with respect to the given set of independent variable pairs, and normal in magnitude (in that the Euclidean norm of each polynomial is one).
The best least square approximation (with respect to the original basis set) is then a linear combination of the orthonormal polynomials, where the coefficient of polynomial i is the inner product of the dependent values with the vector of values of that polynomial evaluated at the independent variables.
basis: LeastSq.PointSet = NEW[LeastSq.PointSetRec[ps.rowDim]];
FOR row: NAT IN [0..ps.rowDim) DO -- fill ps symmetrically
FOR col: NAT IN (row..ps.rowDim) DO
ps[row][col] ← ps[col][row];
ENDLOOP;
ENDLOOP;
FOR row: NAT IN [0..ps.rowDim) DO
innerProd: REAL;
col: NAT;
basis[row] ← NEW[LeastSq.RowRec[ps.rowDim]];
FOR col IN [0..ps.rowDim) DO basis[row][col] ← 0.0 ENDLOOP;
FOR prevrow: NAT IN [0..row) DO
innerProd ← 0.0;
FOR col IN [0..ps.rowDim) DO
innerProd ← innerProd+basis[prevrow][col]*ps[row][col];
ENDLOOP; -- col
FOR col IN [0..ps.rowDim) DO
basis[row][col] ← basis[row][col]-innerProd*basis[prevrow][col];
ENDLOOP; -- col
ENDLOOP; -- prevrow
basis[row][row] ← 1.0;
IF row<ps.rowDim-1 THEN -- normalize this new orthogonal polynomial
BEGIN
innerProd ← 0.0;
FOR col IN [0..ps.rowDim) DO
FOR col1: NAT IN [0..ps.rowDim) DO
innerProd ← innerProd+basis[row][col]*basis[row][col1]*ps[col][col1];
ENDLOOP; -- col1
ENDLOOP; -- col
IF innerProd>0.0 THEN {
normFact: REAL = 1.0/Real.SqRt[innerProd];
FOR col IN [0..ps.rowDim) DO
basis[row][col] ← normFact*basis[row][col];
ENDLOOP; -- col
};
END;
ENDLOOP; -- row
eq ← basis[ps.rowDim-1];
};
NewRow: PROC [ dim: NAT ] RETURNS [ pt: LeastSq.Row ] = {
pt ← NEW[LeastSq.RowRec[dim]]
};
Test: PROC [ eqIn: LeastSq.Row, n: INT, noise: REAL ← 0 ] RETURNS [ eqOut: LeastSq.Row ] =
BEGIN
ps: LeastSq.PointSet = NewPointSet[eqIn.colDim];
pt: LeastSq.Row = NEW[LeastSq.RowRec[eqIn.colDim]];
pt[0] ← 1.0;
FOR i: INT IN [0..n) DO
sum: REAL ← eqIn[0]+noise*Random.Choose[-10000, 10000];
FOR j: NAT IN [1..eqIn.colDim-1) DO
pt[j] ← Random.Choose[-10000, 10000];
sum ← sum+eqIn[j]*pt[j];
ENDLOOP;
pt[eqIn.colDim-1] ← -sum;
AddPoint[ps, pt];
ENDLOOP;
eqOut ← LeastSqFit[ps];
END;
END.