<div>I suggested using, as an interpolating function, A*e^(-kx). But, more constants would pass the function through more data points, for a more accurate interpolation. So I'd suggest adding a constant:</div><div> </div>
<div>A*e^(-kx)+B. To pass the function through 3 data points.</div><div> </div><div>Warren was using one exponential function for the entire range, instead of using a different one to interpolate, with a few points, for each interval between two consecutive integers. He said that that exponential wouldn't really be an exact measure of the frequencies of states of various sizes. And, in fact, I don't know if exponentials could give a good fit at all, for intervals in some parts of the overall range of interest.</div>
<div> </div><div>So maybe a polynomial interpolation would be needed for some regions. If so, it might even be desirable everywhere. As before, it would be a different interpolating polynomial for each interval between consecutive integers.</div>
<div> </div><div>When I said to sum P(s) between S(a) and S(b), I didn't say anything about how that would be done. </div><div> </div><div>P(s) is the inverse of the interpolating function. As I said, the inverse of an exponential is a logarithmic function. The inverse of some polynomial functions can't be written. I believe that's true of every polynomial of degree higher than two. </div>
<div> </div><div>One could use a quadratic, of course.</div><div> </div><div>But, evaluating a definite integral of a function whose inverse is easily antidifferentiated, isn't a problem. So, as long as the interpolating function itself is easily antidifferentiated, there won't be a problem.</div>
<div> </div><div>Aside from that, the logarithmic functions resulting from the exponential interpolating functions that I suggested can be antidifferntiated by a substitution and an integration by parts.</div><div> </div>
<div>When the Weighted-Webster procedure that I described is applied to the S(x) interpolating function, kx, which would be consistent with a flat probability distribution, the roundoff point, R, is a + 1/2. In other words, Webster's method. </div>
<div> </div><div>Mike Ossipoff</div><div> </div>