Biased coin: p = 0.9, 0.1 — Shannon entropy

intropredictability costs information

Answer

0.4690 bits

Why this example is worth doing

A heavily biased coin carries less than half a bit per flip, because most outcomes are unsurprising. The page plots entropy against p across the whole range so the maximum at one half and the zeros at both ends are visible as a curve, which conveys the shape of the function far better than any single number. It also gives each outcome's surprisal, showing that the rare event carries more information and occurs less often.

Try your own input in the Shannon entropy. Bits per symbol for any distribution, with the surprisal of each symbol shown.

How the answer is reached

Shannon entropy of 2 symbols

2 symbols, 10 total occurrences. Divide by the total to get probabilities, then take −Σ p·log₂ p.

The surprisal column is the ideal code length; Huffman rounds it to whole bits. — Shannon entropy of 2 symbols
SymbolpSurprisal −log₂ p (bits)Contribution p·(−log₂ p)
H0.9000000.1520030.136803
T0.1000003.3219280.332193

H0.468996 bits/symbolthe probability-weighted mean of the surprisal column

Total information4.689956 bits10 symbols × H

Maximum possible H1.000000 bits/symbollog₂(2), reached only by the uniform distribution

H / log₂(m)0.468996how close to uniform this source is

Warning:

A symbol of probability 0 contributes exactly 0: the convention 0·log₂0 = 0 is a definition (the limit as p → 0), not an approximation. Evaluating it instead returns NaN, which is the usual bug on this page.

Source: C. E. Shannon, Bell System Technical Journal 27:379–423 (1948), Theorem 2

Compare with

Open this example in the Shannon entropy

The field arrives filled in with this example’s input.

Note:

Notation this page assumes

  • Bit strings are written most significant bit first, and bit 0 is the least significant bit.
  • A width is stated explicitly wherever it changes the answer; nothing is silently sign-extended or truncated.
  • Entropy is in bits per symbol — logarithms base 2 — and 0·log 0 is taken as 0.

Sources

  • Shannon, “A Mathematical Theory of Communication” (1948)