Inverse Square Weighted Interpolation
Suppose we wish to interpolate a surface through a set of non-uniformly
distributed control points z = z(x, y). Of course, there are
infinitely many surfaces passing through any finite set of control
points, so the result is not uniquely determined unless we specify
some additional information, which serves as a criterion for "goodness
of fit", constraining the overall shape of the surface.
One common technique is multiple linear regression. This requires us
to specify the "form" of the surface. For example, we could specify
z(x,y) = A + Bx + Cy + Dxy + Ex^2 + Fy^2 + Gx^2y^2
and then use linear regression to determine the coefficients A,B,..,G
such that the sum of squares of the z errors at the control points is
minimized. To ensure that this surface actually passes through each of
the N control points, we would need to choose a "form" with N terms.
For example, the above form will fit any seven control points exactly.
Another interesting method, and one that automatically passes through
each control point, is the "inverse square weighting" method. Given
the values of f(xi,yi) for N points (x1,y1), (x2,y2),...,(xN,yN), the
interpolated function for every other point (x,y) is
f(x1,y1) f(x2,y2) f(xN,yN)
-------- + -------- + ... + --------
R1^2 R2^2 RN^2
f(x,y) = --------------------------------------
1 1 1
---- + ---- + ... + ----
R1^2 R2^2 RN^2
where (Ri)^2 = (x-xi)^2 + (y-yi)^2. This can obviously be extended
to any number of dimensions. For points (x,y) far from any of the
control points, the value of f approaches the average of the f values
for the control points.
To illustrate, suppose we are given two control points f(0,0)=1 and
f(1,0)=-1. The surface generated by these two points is shown below.
For another example, suppose we are given the four points
f(0,0) = -4 f(1,0) = 5
f(0,1) = 5 f(1,1) = -1
The surface generated by these four points is shown below.
Return to MathPages Main Menu