**
More on 3x+1 Completeness
**

Although there is no really hard evidence, data currently available suggests
that the completeness is limited to a value **below** the limit of
`ln(2) / ln(3) = 0.6309...`

There are two arguments to support this view. Both of them were originally
contributed by Eric Farin.

The first argument is based on current data. From the general formula

E.ln(2) = O.ln(3) + ln(N) + ln(Res)

through substitution of
E.ln(2) = E.C.ln(3) + ln(N) + ln(Res)

which on regrouping yields
ln(2) - C.ln(3) = ( ln(N) + ln(Res) ) / E

Therefore, if C is to get close to the theoretical limit we expect

Class records up to ± 2200 are either exact, or else correct within an order of magnitude (i.e. no more than a factor 6 too high). For higher class records current "best known candidates" are depicted. Beyond 2200 the values are therefore less and less certain, although empirical evidence suggests most of them are probably not too far from the exact values.

The curve apparently shows a tendency to converge to a value roughly somewhere
between 35 and 45. This appears to indicate that Gamma is bound, and therefore that
the Completeness will **not** get infinitesimally close to its theoretical limit.

The second argument is a statistical one. From every number N within a Delay Class x we can derive a number 2N in Delay Class x+1 plus possibly another one if N is of the form 6k+4, in which case 2k+1 has also Delay x+1. In the first case the number has one more even step, in the latter case it has one more odd step. We can therefore approach the number of members having O odd steps in any Delay Class D by

( D! / ( (D-O)! . O! ) ) . (1/3)^{O}

Since we know the completeness of a number once D and O are known we
can estimate how many numbers exist having a specific Completeness C.
Using the Stirling formula for
N! = sqrt(2.π.N) . (N/e)^{N}

we eventually find that the log of the number is approached by
C.ln(C) + (1-C).ln(1-C) + C.ln(3)

We can therefore state a number with a specific Completeness C can only exist
if the expression above is positive. Setting the expression equal to zero and
solving for C yields
Thus there appears to be some evidence to support

- Conjecture 3:
- A number C
_{max}exists such that for all N C(N) < C_{max}.

with C_{max}< ln(2) / ln(3).

If the constant mentioned above is indeed the correct value a simple calculation
shows that the maximum value for Gamma will be `41.677647656...`
The data on Gamma records show that values really close
to this constant have so far not been observed though.

Back to the general 3x+1 page.