Mathematicians look at the world a bit oddly. We are interested in
obscure things, we derive great satisfaction from being certain about
things others might take for granted or never think about, and
frequently we can apply these to make interesting statements about the
"real world" (much to our embarrassment). I hope that you will agree
that this is one of those interesting statements.
Theorem (Curt McMullen,
1994):
Backgammon ends with probability 1.

That is, if one uses random dice (even mildly biased ones), and any
legal playing strategy (even trying to lose), then the probability that
the game has ended by the nth move gets arbitrarily close to 1 as n
increases. There is some move such that the chance the game has ended
by that move is at least 99%, at least 99.999%, etc. so the chance that
a game continues forever is 0.
This result, like the bulk of mathematics, involves little
computation. I haven't seen the details written up anywhere yet, but
this result should be accessible to any serious backgammon player. I'll
use 3 steps to demonstrate the theorem.
Step 1: Let us imagine a variation: Player A calls (chooses) the
dice, and Player B has to make a legal move. It is enough to show that
this modified game is a win for Player A as long as the initial
position is legal.
Step 2: In this modified game, from any legal position Player A
may choose the dice so that not both players are on the
bar. (This can be done in 20 rolls.)
Step 3: From any position in which not both players are on the
bar, Player A may end the game by calling 24 enough times in a
row. (9000 is enough times.)
Why does it suffice to show that calling the dice can end the game? Let us suppose that from any position, some fixed number n of carefully chosen rolls will end the game. (We can take n to be the maximum number of rolls needed over the finite set of possible legal backgammon positions, if we had a value for each position.) There is a chance of at least that the dice will behave as though called by Player A for n rolls. That's not much, but suppose it doesn't happen. Then after the first n rolls, if the game is still going on, there will again be at least a chance that the dice will behave as though called by Player A. For the game to continue, it must keep avoiding these chances. That will happen for a while, but with probability 1, eventually the probability event will occur, and the dice will behave as though possessed. If the lottery is fair and you keep playing, eventually you will win, and will even win infinitely often.
Ok, so how might a player calling the dice end the game? First, Player
A can get at least one player off of the bar. That's easy enough to do
if the position is legal; it can't be that both players are shut
out. So if a player is not shut out, give them doubles of a number
which is open. If this gets all of that player's checkers off the bar,
then we can move on to the next step. If not, then 4 checkers came off
the bar, and at most one was hit. So there are now at least 3 fewer
checkers on the bar. One can give a better estimate, but since there
are 30 checkers total then after at most 10 exchanges (20 rolls) at
least one player will have no checkers on the bar.
Finally, if at least one player is not on the bar, then repeatedly
calling 24 will end the game. Part of Curt McMullen's idea was that in
this case at least one player must be able to play part of a 24. If
you have made points 2 and 4 pips ahead of a checker of mine, then you
have a 2 to play, from the point 4 ahead to the point 2 ahead. You
might not be able to play this if you are on the bar, but then either
you can move off the bar or I have made points 2 and 4 pips ahead of
you, my 2 and 4 points. So the only possible way that we would both be
stuck under this deluge of 24's is if we are both on the bar with our
2 and 4 points made. That can't happen from a position in which one
player is not on the bar by a sequence of 24 rolls, since the only way
for both players to be on the bar is if one hits from the bar, and if
you hit me from the bar with a 24 it must be that my 2 or 4 point
isn't made. So although it might be the case that both players are on
the bar, under successive 24's starting from a position in which at
most one player is on the bar at least one player can move.
Ok, but what if the players just keep sending each other back? Now the
second part of Curt McMullen's idea comes in: When you get hit, your
checker goes to the bar, which is your 25 point. Henceforth, that
checker will always be on an odd point if the dice always show
24. Your odd points are your opponent's even points, so once a checker
of yours is hit, it can't hit a checker on one of your opponent's odd
points. Even though the pip count might oscillate, in some sense there
is progress being made; a checker on an even point must advance
eventually and cannot be sent back to an even point. This suggests
using a modified pipcount which decreases on each exchange
whether or not hits are made.
Define the modified pipcount to be the pipcount of checkers on
odd points plus 12.5 times the pipcount of checkers on even
points.
For example, in the above position, the modified pipcount for white is
6 odd pips plus 12.5 times 8 even pips = 106. For red, there are 51
odd pips and 12.5 times 6 even pips, so the modified pipcount is
126. The total modified pipcount is 106 + 126 = 232.
The modified pipcount decreases with every 2 or 4 played:
 If no checker is hit, clearly the pipcount decreases by at least
2, since either the odd pips or even pips decrease by 2 or 4.
 If you hit by moving a checker on an odd point of yours, it must
hit a checker on an even point of your opponent. Your odd pip total
decreases by at least 2. The hit checker contributed at
least 25 pips (2 × 12.5) to your opponent's modified pipcount, and
now contributes exactly 25 (on the bar), so your opponent's
modified pipcount stays the same or decreases. So the total
modified pipcount decreases by at least 2.
 If you hit by moving a checker on an even point of yours, it must
hit a checker on an odd point of your opponent. This decreases your
modified pipcount by at least 25 (2 even pips times 12.5). Your
opponent's modified pipcount increases by at most 22 (from 3 to
25), so the total decreases by at least 3.
Finally, the modified pipcount of a checker is greatest when it is on
the 24 point, where its value is 24 times 12.5 = 300. There are 30
checkers, so the maximum possible modified pipcount is 300 times 30 =
9000. With each exchange, the modified pipcount decreases by at least
2, so after 9000 24's in a row (4500 exchanges) the modified pipcount
will have been reduced to 0, and the game will be over.
So from any legal position, there is at least a chance that
the game will end within the next 9020 rolls, so backgammon
ends with probability 1.
Of course, one can tighten the estimates somewhat, and more intelligent
choices by player A would end the game much sooner. From the starting
position, 8 55's followed by 11 66's would end the game. One could
use 36 instead of 24. The proof can't be tightened too much since
it is possible for the game to last over 500 rolls of 24. This method
of proof also provides no practical limit on the length of a game,
since the universe will run down before the halflife of a backgammon
game that ends with probability every 9020 rolls. Further,
any digital random number generator will eventually cycle, and it is
possible that a game played using such a generator could cycle,
too.
On the other hand, this result means that there should be no draws in
backgammon. If there is a cap on the cube or in match play, perfect
play exists, and no matter how wild the position appears, there is some
equity one may theoretically assign to the position. I don't know about
you, but I'll sleep better at night knowing this.
