==> arithmetic/conway.s <==
This sequence and its remarkable properties were discovered, apparently
independently, by Douglas Hofstadter (circa 1975), John Conway (in the
early 1980's), B.W. Connoly, and others. Since Conway discovered many of
the deeper properties, and is the one responsible for popularizing the
sequence, it seems appropriate to name the sequence after him.

Some properties of a(n):

--  a(n+1) - a(n) = 0 or 1

--  a(2^n) = 2^(n-1)

--  n/2 <= a(n) <= n

--  A(n)= 2a(n) - n    has zeros at the powers of 2 and 
     positive "humps" in between

--  A(2^n + t) = A(2^(n+1) - t)  for t between 0 and 2^n,
     i.e. the humps are symmetric

-- a(2n) <= 2a(n)

-- a(n)/n --> 1/2  as  n --> infinity

-- a(n) and A(n) are self-similar, in the sense that their values on the
   (N+1)'st hump can be obtained by a "magnification" process applied
   to the values on the N'th hump. They are *not* chaotic sequences,
   since their asymptotics and combinatorial structure are very regular
   and rigid.

In a lecture at Bell Labs in 1988, Conway asked for the rate at
which  a(n)/n converges to 1/2.  In particular, he asked for
N(epsilon), the last value of  n  such that  |a(n)/n - 1/2|
exceeds epsilon, in the case  epsilon=1/20.  This problem was
solved by Colin Mallows of Bell Labs: he found that  N(1/20)=1489
and N(1/40)=6083008742. These values are reported in his article
in the American Mathematical Monthly, January 1991, p. 5.

Conway's sequence has a deep combinatorial structure, and is intimately
connected with all of the following: variants of Pascal's triangle, the
Gaussian distribution, combinatorial operations on finite sets,
coin-pushing games, and Fibonacci and Catalan numbers.  All of this is
explained in my joint paper with Ravi Vakil; anyone who would like to
receive a copy in LaTex format should contact me at the e-mail address
listed below.

One byproduct of our work is an algorithm to determine  N(epsilon) for
given positive epsilon.  Here are some particular values:

 e      Last  n  such that   |a(n)/n - 1/2| > e
----    ----------------------------------------------------------
1/20    1489  (found by Mallows in 1988)
1/30    758765
1/40    6083008742  (found by Mallows in 1988)
1/50    809308036481621
1/60    1684539346496977501739
1/70    55738373698123373661810220400
1/80    15088841875190938484828948428612052839
1/90    127565909103887972767169084026274554426122918035
1/100   8826608001127077619581589939550531021943059906967127007025

These values were computed by the Mathematica program listed below; the
algorithm is explained in our paper.  To use the program, load it into
Mathematica and type   neps[t]  for some numerical value of  t  (which
should probably be smaller than 1/5 or so). The program will output  N(t),
e.g.  neps[1/20] = 1489. These numbers grow very fast: N(epsilon) will be
about 2^((0.138015/epsilon)^2), so  N(1/1000) will have about 5735 digits.
The program requires very little memory space, but its runtime appears to
grow rapidly as epsilon decreases (on a Sun 3, it took about one second to
find the first value listed, and several minutes to find the last).

neps[eps_] := Block[{W}, L = 1 + Floor[(0.138015/eps)^2]; e = eps*2; 
    W = processvector[L]; While[Length[W] > 0, 
     nmax = 1 + Last[W][[4]]; L++; W = processvector[L]]; nmax]
 
processvector[L_] := 
  Block[{V}, V = startvector[L]; While[notfound[V], V = newbody[V]]; V]
 
startvector[L_] := Select[initialvector[L], gt]
 
initialvector[L_] := 
  Table[{L, b, Binomial[L - 1, b - 1], 
    2^(L + 1) - Sum[Binomial[L, c], {c, b, L}]}, {b, 0, L - 1}]
 
c[0] = 0
 
c[n_] := c[n] = Sum[Binomial[2*a, a]/(a + 1), {a, 0, n - 1}]
 
gt[x_] := U[x] > e
 
U[x_] := (x[[3]] + M[x[[1]], x[[2]]])/(x[[4]] + incp[x[[1]], x[[2]]])
 
M[n_, n_] = -1
 
M[n_, k_] := Binomial[n - 1, K[n, k]] - Binomial[n - 1, k - 1] + c[K[n, k]]
 
K[n_, k_] := Min[k, n - k]
 
incp[n_, k_] := Max[M[n, k], 1]
 
notfound[V_] := 
  Length[V] > 0 && Last[V][[2]] > 0 && Last[V][[1]] > Last[V][[2]]
 
newbody[V_] := Join[Drop[V, -1], newtail[V]]
 
newtail[V_] := Select[{vleft[Last[V]], vright[Last[V]]}, gt]
 
vleft[x_] := {x[[1]] - 1, x[[2]] - 1, x[[3]], x[[4]]}
 
vright[x_] := {x[[1]] - 1, x[[2]], x[[3]] + S[x[[1]] - 1, x[[2]] - 1], 
   x[[4]] + Binomial[x[[1]] - 1, x[[2]] - 1]}
 
S[n_, k_] := Binomial[n - 1, k] - Binomial[n - 1, k - 1]


-Tal Kubo     kubo@math.harvard.edu  or   kubo@zariski.harvard.edu