Half a Bit
How can we store half a bit of information?
A 1-bit storage unit stores, by definition,
1 bit. Now consider a defective storage unit, where
reading it may cause a bit flip with some probability p.
If p = 0, we are back to the initial case: with no
defect.
If p = 1/2, the value we read is completely random.
It is therefore essentially impossible to store information, and
intuitively the stored information is exactly 0 bits.
But what if 0 < p < 1/2? For p very
close to 0, it makes sense for the information to be
close to 1 bit, but slightly less. Likewise, for
p close to 1/2, we will have little
information, but more than 0 bits. The information
must vary continuously from 1 bit to 0
bits as p varies from 0 to
1/2.
To find p such that we have 0.5
bits, note that after one observation there are two possible cases:
the bit has been flipped, with probability p, or it is
still the original bit, with probability 1 - p.
We use Shannon information theory. We compute the Shannon entropy:
H(p) = -p log2(p) - (1 - p) log2(1 - p)
At this point, the stored information is:
I(p) = 1 - H(p)
Numerically, if 1 - H(p) = 1/2, that is
H(p) = 1/2, the probability solving the equation is
approximately:
p = 0.110027... ≈ 11.0%