The storage requirements for the four representations are wildly different in many respects. For purposes of comparison, we use M=360 for the map size and N=4694 for the number of visibilities. The first half of Table 2 shows the storage space (in floating point numbers) as a function of N and M. The second half shows the corresponding actual numbers.
Storage Requirements for Each Approach (All 9 Collimators)
Rectangular | Direct | HXT Style | Polar | |
Visibilities: | 2 N +1 | 2 N+1 | N | 2 N + 1 |
Intermediate: | 4 M N | M2 N | 2MN | |
Map pixels: | M2 | M2 | M2 | |
Visibilities: | 37 KB | 37 KB | 18.7 KB | 37 KB |
Intermediate: | 34 MB | 54 MB | 2433 MB | 13 MB |
Map pixels: | 34 MB | 0.5 MB | 0.5 MB | 0.5 MB |
The number of equivalent floating-point operations for one transform from visibility space to map space is readily estimated. The table shows the ``brute force'' numbers for the direct and polar methods as well as the ``recurrence'' numbers described in Sections 3 and 5. The former uses the relative timing (in IDL 5.2) for Bessel functions, FFT and exponentials: beselj(d,100) : FFT(d,1) : exp(2*pi*i*d*x) : d*x = 450 : 54 : 49 : 1. When arrays of exponentials or arrays of Bessel functions are calculated using recurrence, both the exponential and Bessel function computation times can be reduced to about 12 multiplications.
Computational Requirements for Each Approach (All 9 Collimators)
Rectangular | Direct | HXT Style | Polar | |
BRUTE FORCE: | ||||
Equiv. FP ops: | 14*N2 | 50N M2 | 50N M2 | 500 N M |
Equiv. FP ops: | 3.e+08 | 3.e+10 | 3.e+10 | 8.e+08 |
RECURRENCE: | ||||
Equiv. FP ops: | - | ? * N M | - | 20 N M |
Equiv. FP ops: | - | 9.e+09 | - | 4.e+07 |
These numbers should be considered accurate to about a factor of 2.
Clearly the HXT-style representation is a major loser in
computational speed, and with the use of recurrence, the polar
representation is the winner.
Since the HXT-style representation scales as M2, smaller maps could be made faster. A map, for example, would be made in only slightly more time than in the rectangular representation. But the storage requirements (67 MB) would still be larger than in the other methods.