  
  [1X3 [33X[0;0YResults[133X[101X
  
  [33X[0;0YThe  [5XDifSets[105X  Package  was  designed with the goal of finding all difference
  sets  up  to  equivalence  in  groups  of  order 64 and 96, a goal which was
  accomplished.  Overall,  the algorithm has successfully computed results for
  1006  of the 1032 groups of order less than 100. Full results, which include
  timings,  number  of  sets, and the sets themselves can be found in the [10Xdata[110X
  subdirectory  of the package, which is organized by group order and contains
  a  single  [10X.txt[110X file for each computed group. A list of all timings can also
  be  found  in  the file [10Xgroups.csv[110X in the [10Xdata[110X directory, and the difference
  sets themselves can be loaded using the function [2XLoadDifferenceSets[102X ([14X2.6-2[114X).
  All  computations were performed using [5XGAP[105X 4.9.1 on a 4.00GHz i7-6700K using
  8GB  of  RAM.  Here  we  give  a  basic  overview of results and comments on
  timings.  Throughout this chapter we will refer to the group returned by the
  [5XGAP[105X function [10XSmallGroup(v, n)[110X as [10X[v, n][110X.[133X
  
  
  [1X3.1 [33X[0;0YOrder 16 and 36[133X[101X
  
  [33X[0;0YDifference  sets  in  groups  of  order  16 and 36 form the first nontrivial
  examples of the Hadamard parameters, and exhaustive enumerations are already
  well  known.  Still,  computation of these sets gives a useful benchmark and
  check of accuracy.[133X
  
  [33X[0;0YAlmost  all  groups  in these orders take less than a second. The group [10X[36,
  9][110X,  however,  takes several orders of magnitude longer than other groups of
  order  36.  This is because [10X[36, 9][110X does not have small normal subgroups (in
  particular,  its  smallest  nontrivial  normal  subgroup  has  order 9), and
  refining  across  a  large  gap  in  sizes,  expecially  near the end of the
  algorithm, requires checking significantly more preimages.[133X
  
         Group     Difference Sets   Time (seconds)   
        ────────   ───────────────   ──────────────  
        [10X[16, 1][110X           0              0.030        
        [10X[16, 2][110X           3              0.103        
        [10X[16, 3][110X           4              0.100        
        [10X[16, 4][110X           3              0.100        
        [10X[16, 5][110X           2              0.061        
        [10X[16, 6][110X           2              0.071        
        [10X[16, 7][110X           0              0.072        
        [10X[16, 8][110X           2              0.070        
        [10X[16, 9][110X           2              0.082        
        [10X[16, 10][110X          2              0.187        
        [10X[16, 11][110X          2              0.121        
        [10X[16, 12][110X          2              0.195        
        [10X[16, 13][110X          2              0.117        
        [10X[16, 14][110X          1              0.059        
  
         Group     Difference Sets   Time (seconds)   
        ────────   ───────────────   ──────────────  
        [10X[36, 1][110X           0              0.335        
        [10X[36, 2][110X           0              0.201        
        [10X[36, 3][110X           0              0.407        
        [10X[36, 4][110X           0              0.322        
        [10X[36, 5][110X           0              0.218        
        [10X[36, 6][110X           6              0.412        
        [10X[36, 7][110X           1              0.795        
        [10X[36, 8][110X           4              0.340        
        [10X[36, 9][110X           5             340.989       
        [10X[36, 10][110X          6              1.137        
        [10X[36, 11][110X          3              0.699        
        [10X[36, 12][110X          6              0.417        
        [10X[36, 13][110X          1              0.801        
        [10X[36, 14][110X          3              0.434        
  
  
  [1X3.2 [33X[0;0YOrder 64 and 96[133X[101X
  
  [33X[0;0YDifference  sets in groups of order 64 also satisfy the Hadamard parameters,
  while   difference  sets  in  groups  of  order  96  satisfy  the  McFarland
  parameters.  Since  there  are many groups of both orders, here we just give
  some  examples  and  summaries.  In  particular,  the  tables below list the
  fastest, slowest, and median five groups of each order, sorted by time.[133X
  
  [33X[0;0YGroups  of  order  64  are  [22Xp[122X-groups,  and  thus  always  have enough normal
  subgroups  to  form  long refining series. This means the refining steps are
  relatively  efficient  for  all  groups  in  this order. The main difference
  between  groups  is  the size of the automorphism group, and, in particular,
  four  of  the  five groups taking the largest amount of time are four of the
  five  groups  with  the  largest  automorphism  groups  in  this  order. The
  additional  group  in the top five, [10X[64, 235][110X, has a relatively large number
  of  difference  sets,  but  is  otherwise  unremarkable. In general, smaller
  numbers of difference sets correspond to faster times, and in fact the eight
  groups  with  no difference sets were computed the fastest, beating the next
  fastest  groups by an order of magnitude. Overall, the mean computation time
  for a group of order 64 was 3988.476 seconds, with a median time of 1493.175
  seconds.  This  means that the total computer time to compute all difference
  sets in groups of order 64 was roughly 12 days.[133X
  
  [33X[0;0YIn  groups  of  order  96  we  do  not  always  have large numbers of normal
  subgroups,   and,  as  with  [10X[36,  9][110X,  this  can  substantially  slow  down
  computation.  In  fact,  the five groups taking the longest computation time
  are  five  of  the six groups with fewest normal subgroups in this order. We
  are  helped,  however,  by  the  fact that the only valid choice of [22Xk[122X is 20,
  which  is  relatively  small  and  thus  does  not  lead to large numbers of
  preimages even across large gaps in the refining series. Many groups in this
  order  have no difference sets, but even for these groups computation can be
  slow.  While the fastest groups contain no difference sets, many groups with
  no  difference  sets  actually  take  much  longer than other groups that do
  contain  difference  sets. Overall, the mean computation time for a group of
  order  96  was  24447.991  seconds, with a median time of 11278.765 seconds.
  This  means  that  the total computer time to compute all difference sets in
  groups of order 96 was roughly 65 days.[133X
  
          Group     Difference Sets   Time (seconds)   
        ─────────   ───────────────   ──────────────  
        [10X[64, 52][110X           0              3.451        
        [10X[64, 54][110X           0              3.463        
        [10X[64, 47][110X           0              3.594        
        [10X[64, 186][110X          0              3.940        
         [10X[64, 1][110X           0              3.950        
        [10X[64, 166][110X        2312            1424.692      
        [10X[64, 134][110X         342            1439.484      
        [10X[64, 135][110X         540            1493.175      
         [10X[64, 7][110X         1320            1515.710      
        [10X[64, 160][110X        3192            1518.693      
        [10X[64, 192][110X         222           21131.394      
        [10X[64, 267][110X          4            23662.500      
        [10X[64, 235][110X        4317           24566.186      
        [10X[64, 260][110X         30            144338.020     
        [10X[64, 262][110X         148           229488.988     
  
          Group     Difference Sets   Time (seconds)   
        ─────────   ───────────────   ──────────────  
         [10X[96, 2][110X           0              8.731        
        [10X[96, 59][110X           0              8.791        
        [10X[96, 189][110X          0              29.378       
        [10X[96, 66][110X           0              29.777       
        [10X[96, 46][110X           0              44.478       
        [10X[96, 209][110X          4            10809.673      
        [10X[96, 133][110X         16            11198.052      
        [10X[96, 224][110X          0            11278.765      
        [10X[96, 89][110X           0            11349.466      
        [10X[96, 102][110X          0            11415.688      
        [10X[96, 227][110X         42            308246.830     
        [10X[96, 64][110X          14            310447.407     
        [10X[96, 70][110X          28            514559.313     
        [10X[96, 72][110X           2            515196.547     
        [10X[96, 71][110X           8            871439.024     
  
  
  [1X3.3 [33X[0;0YComments[133X[101X
  
  [33X[0;0YOverall,  the  algorithm  spends  almost  all  of  its  time performing four
  operations:    refining    sums    to   sums   in   several   stages   using
  [2XSomeRefinedDifferenceSums[102X  ([14X2.3-8[114X), refining sums to sets in the final stage
  using [2XSomeRefinedDifferenceSets[102X ([14X2.3-4[114X), removing equivalent difference sums
  in  several  stages  using  [2XEquivalentFreeListOfDifferenceSums[102X  ([14X2.4-3[114X), and
  removing   equivalent   difference   sets   in   the   final   stage   using
  [2XSmallestEquivalentFreeListOfDifferenceSets[102X  ([14X2.4-6[114X).  On  typical  groups of
  order  16  and  order  36 (i.e., not [10X[36, 9][110X), each of these four operations
  takes  roughly  the same time. On groups of order 64, some testing indicates
  that  one or two orders of magnitude more time are spent in the final stage,
  when    the    algorithm    uses   [2XSomeRefinedDifferenceSets[102X   ([14X2.3-4[114X)   and
  [2XSmallestEquivalentFreeListOfDifferenceSets[102X   ([14X2.4-6[114X).  This  discrepency  is
  likely  to  remain  or  increase  for  larger order groups, as the number of
  preimages  to  check  increases exponentially with the number of cosets. For
  the  tested  groups  of order 64, roughly 60% of the time in the final stage
  was spent refining, with the remaining 40% spent removing equivalent sets.[133X
  
  [33X[0;0YLarge automorphism groups make removing equivalents time-consuming and large
  jumps  in  the size of the normal subgroups used, especially near the end of
  the  algorithm, make refining difficult. So, in general, the algorithm seems
  to  work well when the group has a small automorphism group and many (small)
  normal  subgroups. In addition, the algorithm does better when the values of
  [22Xk[122X  that  need  to  be  checked  are small, as this limits both the number of
  preimages  to check as well as the amount of time required for checking sets
  and  equivalences.  It  is  also generally faster when the final result is a
  smaller number of difference sets.[133X
  
  [33X[0;0YThere  are  twenty-six  groups of order less than 100 in which the algorithm
  was  not able to complete a search. Fourteen of these groups are prime order
  cyclic.  As simple groups, these groups have no normal subgroups and thus no
  possibility  for  refining,  which  means  the  algorithm  must search every
  possible  subset  of  size [22Xk[122X to find all difference sets of size [22Xk[122X. Even for
  groups  of relatively small order, such as order 31, this is infeasible, and
  with  current  implementation  will overflow memory before even starting the
  search  (one of these groups, [10X[37, 1][110X is actually feasible to search without
  this  implementation issue, but the others have too many sets to check). The
  remaining  groups  have  either too few normal subgroups, large jumps in the
  refining  series,  large  possible  values  of  [22Xk[122X, or a combination of these
  problems.[133X
  
  [33X[0;0YThe  next  natural  cases  for exhaustive search are groups of order 100 and
  order   144,   which  give  the  next  Hadamard  parameters.  Unfortunately,
  preliminary  testing  indicates that this algorithm is not likely to be able
  to  compute  all  difference  sets  for these groups. For example, a typical
  difference  sum in [10X[100, 9][110X is [10X[5, 4, 3, 3, 0, 3, 2, 3, 2, 2, 2, 2, 2, 1, 2,
  1,  1,  2, 2, 3][110X, which has roughly [22X6 × 10^16[122X preimage sets to check. In the
  search  for  difference  sets in [10X[36, 9][110X the single difference sum [10X[6, 3, 3,
  3][110X, with around [22X3 × 10^7[122X preimages, takes around 300 seconds to search. Thus
  even  if  we  could check sets in [10X[100, 9][110X as fast as in [10X[36, 9][110X, the search
  would  take roughly 20000 years. Some testing suggests that coding pieces of
  the algorithm in C could give one or two orders of magnitude of speedup, but
  even  further speedup is required to make the search feasible, so some other
  improvements, either in theory or implementation, are needed as well.[133X
  
