Normalization of the Polynomials

The columns of are orthonormal with respect to the inner product

as shown by the following statements:

reset fuzz;         /* print tiny numbers as zero */
w = j(ncol(x),1,1); /* default weight is all ones */
/* Verify orthonormal */
L = P`*diag(w)*P;
print L;

Some reference books on orthogonal polynomials do not normalize the columns of the matrix that represents the orthogonal polynomials. For example, a textbook might give the following as a fourth-degree polynomial evaluated on evenly spaced data:

textbookPoly = { 1 -2  2 -1  1,
                 1 -1 -1  2 -4,
                 1  0 -2  0  6,
                 1  1 -1 -2 -4,
                 1  2  2  1  1 };

To compare this representation to the normalized representation that the ORPOL function produces, use the following program:

/* Normalize the columns of textbook representation */
normalPoly = textbookPoly;
do i = 1 to ncol( normalPoly );
   v = normalPoly[,i];
   norm = sqrt(v[##]);
   normalPoly[,i] = v / norm;
end;

/* Compare the normalized matrix with ORPOL */
x = T(1:5); /* Any evenly spaced data gives the same answer */
imlPoly = orpol( x, 4 );

diff = imlPoly - normalPoly;
maxDiff = abs(diff)[<>];
reset fuzz; /* print tiny numbers as zero */
print maxDiff;

Polynomial Regression

A typical use for orthogonal polynomials is to fit a polynomial to a set of data. Given a set of points , , the classical theory of orthogonal polynomials says that the best approximating polynomial of degree is given by

where and where is the th column of the matrix returned by ORPOL. But the matrix is orthonormal with respect to the inner product, so for all . Thus you can easily compute a regression onto the span of polynomials.

In the following program, the weight vector is used to overweight or underweight particular data points. The researcher has reasons to doubt the accuracy of the first measurement. The last data point is also underweighted because it is a leverage point and is believed to be an outlier. The second data point was measured twice and is overweighted. (Rerunning the program with a weight vector of all ones and examining the new values of the fit variable is a good way to understand the effect of the weight vector.)

x = {0.1, 2,   3,  5,    8,   10, 20};
y = {0.5, 1, 0.1, -1, -0.5, -0.8, 0.1};

/* The second measurement was taken twice.
   The first and last data points are underweighted
   because of uncertainty in the measurements.  */
w = {0.5, 2,   1,  1,    1,    1, 0.2};
maxDegree = 4;
P = orpol(x,maxDegree,w);

/* The best fit by a polynomial of degree k is
   Sum c_i P_i  where c_i = <f,P_i> */
start InnerProduct(f,g,w);
   h = f#g#w;
   return (h[+]);
finish;

c = j(1,maxDegree+1);
do i = 1 to maxDegree+1;
   c[i] = InnerProduct(y, P[,i], w);
end;

FitResults = j(maxDegree+1,2);
do k = 1 to maxDegree+1;
   fit = P[,1:k] * c[1:k];
   resid = y - fit;
   FitResults[k,1] = k-1;       /* degree of polynomial */
   FitResults[k,2] = resid[##]; /* sum of square errors */
end;
print FitResults[colname={"Degree" "SSE"}];

Testing Linear Hypotheses

The ORPOL function can also be used to test linear hypotheses. Suppose you have an experimental design with factor levels. (The factor levels can be equally or unequally spaced.) At the th level, you record observations, . If , then the design is said to be balanced; otherwise it is unbalanced. You want to fit a polynomial model to the data and then ask how much variation in the data is explained by the linear component, how much variation is explained by the quadratic component after the linear component is taken into account, and so on for the cubic, quartic, and higher-level components.

To be completely concrete, suppose you have four factor levels (1, 4, 6, and 10) and that you record seven measurements at first level, two measurements at the second level, three measurements at the third level, and four measurements at the fourth level. This is an example of an unbalanced and unequally spaced factor-level design. The following program uses orthogonal polynomials to compute the Type I sum of squares for the linear hypothesis. (The program works equally well for balanced designs and for equally spaced factor levels.)

The following program calls the ORPOL function to generate the orthogonal polynomial matrix , and uses it to form the Type I hypothesis matrix . The program then uses the DESIGN function to generate , the design matrix associated with the experiment. The program then computes , the estimated parameters of the linear model. Since was expressed in terms of the orthogonal polynomial matrix , the computations involved in forming the Type I sum of squares are considerably simplified.

/* unequally spaced and unbalanced factor levels */
levels = { 1,1,1,1,1,1,1,
           4,4,
           6,6,6,
           10,10,10,10};

/* data for y. Make sure the data are sorted
   according to the factor levels */
y = {2.804823, 0.920085, 1.396577, -0.083318,
     3.238294, 0.375768, 1.513658,            /* level 1 */
     3.913391, 3.405821,                      /* level 4 */
     6.031891, 5.262201, 5.749861,            /* level 6 */
    10.685005, 9.195842, 9.255719,  9.204497 /* level 10 */
};

a      = {1,4,6,10}; /* spacing */
trials = {7,2,3,4};  /* sample sizes */
maxDegree = 3; /* model with Intercept,a,a##2,a##3 */

P = orpol(a,maxDegree,trials);

/* Test linear hypotheses:
   How much variation is explained by the
   i_th polynomial component after components
   0..(i-1) have been taken into account? */

/* the columns of L are the coefficients of the
   orthogonal polynomial contrasts */
L = diag(trials)*P;

/* form design matrix */
x = design(levels);

/* compute b, the estimated parameters of the
   linear model. b is the mean of the y values
   at each level.
   b = ginv(x'*x) * x` * y
   but since x is the output from DESIGN, then
   x`*x = diag(trials) and so
   ginv(x`*x) = diag(1/trials) */
b = diag(1/trials)*x`*y;

/* (L`*b)[i] is the best linear unbiased estimated
   (BLUE) of the corresponding orthogonal polynomial
   contrast */
blue = L`*b;

/* The variance of (L`*b) is
   var(L`*b) = L`*ginv(x`*x)*L
     =[P`*diag(trials)]*diag(1/trials)*[diag(trials)*P]
     = P`*diag(trials)*P
     = Identity         (by definition of P)

   Therefore the standardized square of
   (L`*b) is computed as
   SS1[i] = (blue[i]*blue[i])/var(L`*b)[i,i])
          = (blue[i])##2                            */

SS1 = blue # blue;
rowNames = {"Intercept" "Linear" "Quadratic" "Cubic"};
print SS1[rowname=rowNames format=11.7 label="Type I SS"];

Figure 23.222 indicates that most of the variation in the data can be explained by a first-degree polynomial.

Generating Families of Orthogonal Polynomials

There are classical families of orthogonal polynomials (for example, Legendre, Laguerre, Hermite, and Chebyshev) that arise in the study of differential equations and mathematical physics. These "named" families are orthogonal on particular intervals with respect to the inner product . The functions returned by the ORPOL function are different from these named families because the ORPOL function uses a different inner product. There are no built-in functions that can automatically generate these families; however, you can write a program to generate them.

Each named polynomial family , satisfies a three-term recurrence relation of the form

where the constants , , and are relatively simple functions of . To generate these "named" families, use the three-term recurrence relation for the family. The recurrence relations are found in references such as Abramowitz and Stegun (1972) or Thisted (1988).

For example, the so-called Legendre polynomials (represented for the polynomial of degree ) are defined on with the weight function . They are standardized by requiring that for all . Thus . The linear polynomial is orthogonal to so that

which implies . The standardization implies that . The remaining Legendre polynomials can be computed by looking up the three-term recurrence relation: , , and . The following program computes Legendre polynomials evaluated at a set of points:

maxDegree = 6;
/* evaluate polynomials at these points */
x = T( do(-1,1,0.05) );

/* define the standard Legendre Polynomials
   Using the 3-term recurrence with
   A[j]=0, B[j]=(2j-1)/j, and C[j]=(j-1)/j
   and the standardization P_j(1)=1
   which implies P_0(x)=1, P_1(x)=x. */
legendre = j(nrow(x), maxDegree+1);
legendre[,1] = 1; /* P_0 */
legendre[,2] = x; /* P_1 */

do j = 2 to maxDegree;
   legendre[,j+1] = (2*j-1)/j # x # legendre[,j] -
                      (j-1)/j # legendre[,j-1];
end;