Comparision of Speed and Memory Requirement with Standard FHT

Direct speed tests on a Sun have shown that the MFHT is about 8% faster than an FHT line fitter adapted from the plane fitter described in section 5.11.2.3. This improvement is almost entirely due to the hypersphere (circle) approximation to a hypercube (square) used in the FHT, which increases the total number of subdivisions. The speeds of individual subdivisions in the two versions are almost identical.

When memory is limited, the MFHT is preferable. Ignoring the arrays of flags and individual variables, the MFHT allocates space for one array, ${\bf I}^{\rm av}$, at each call to $divide\_rectangle$. On the contrary, the FHT requires four: ${\bf R}_{\bf b}$ for each value [-1,-1], [-1,1], [1,-1] and [1,1] of ${\bf b}$ (refer to procedure $divide\_cube$ in section 5.11.2.3). The FHT also uses global arrays, whereas the MFHT does not. The subdivision procedure in the FHT can be reordered so as to require only one array (which is used four times), but at the cost of introducing repeated tests into the code, slowing the algorithm by about 50%.