[tahoe-dev] Estimating reliability

[Brian Warner]

> I'll read and reply to the rest of your post when I'm more fully
> awake, but as I suggested earlier, it's better to simply avoid
> dealing with numbers that are close to 1.  Restate the problem so
> that you're looking at numbers close to 0, because then your
> calculations can use the full width of the mantissa. When the numbers
> are close to 1 you use up most -- or all! -- of the mantissa bits
> just holding that big sequence of 9s (actually 1s, but you get my
> meaning).

Right. Sometimes the numbers are close to zero, like the probability that
we'll drop from 10 shares to 2 shares in a month. Sometimes they are close to
one, line the probability that we'll keep all 10 shares in a month. Much of
the time we'll be multiplying these two sorts of numbers together, at which
point we have to pick one representation or the other. That's why I'm not
convinced that we can easily apply the restate-the-problem approach you
suggest, at least not without thinking very hard about it.

I'd trust Mathematica to give useful results without a lot of cleverness on
my part, but I'm less convinced that standard python will do the same, and
that's where I'm likely to prototype this. I'm glad you've got a copy of
Mathematica to play with, so we'll be able to see where python's floats break
down.

thanks,
 -Brian