Shannon–Fano–Elias coding

In information theory, Shannon–Fano–Elias coding is a precursor to arithmetic coding, in which probabilities are used to determine codewords.^[1]

Algorithm description

Given a discrete random variable X of ordered values to be encoded, let $p(x)$ be the probability for any x in X. Define a function

$\bar F(x) = \sum_{x_i < x}p(x_i) + \frac 12 p(x)$

Algorithm:

For each x in X,

Let Z be the binary expansion of $\bar F(x)$.

Choose the length of the encoding of x, $L(x)$, to be the integer $\left\lceil \log_2 \frac {1}{p(x)} \right\rceil + 1$

Choose the encoding of x, $code(x)$, be the first $L(x)$ most significant bits after the decimal point of Z.

Example

Let X = {A, B, C, D}, with probabilities p = {1/3, 1/4, 1/6, 1/4}.

For A

$\bar F(A) = \frac 12 p(A) = \frac 12 \cdot \frac 13 = 0.1666...$

In binary, Z(A) = 0.0010101010...

L(A) = $\left\lceil \log_2 \frac 1 \frac 1 3 \right\rceil + 1$ = 3

code(A) is 001

For B

$\bar F(B) = p(A) + \frac 12 p(B) = \frac 13 + \frac 12 \cdot \frac 14 = 0.4583333...$

In binary, Z(B) = 0.01110101010101...

L(B) = $\left\lceil \log_2 \frac 1 \frac 1 4 \right\rceil + 1$ = 3

code(B) is 011

For C

$\bar F(C) = p(A) + p(B) + \frac 12 p(C) = \frac 13 + \frac 14 + \frac 12 \cdot \frac 16 = 0.66666...$

In binary, Z(C) = 0.101010101010...

L(C) = $\left\lceil \log_2 \frac 1 \frac 1 6 \right\rceil + 1$ = 4

code(C) is 1010

For D

$\bar F(D) = p(A) + p(B) + p(C) + \frac 12 p(D) = \frac 13 + \frac 14 + \frac 16 + \frac 12 \cdot \frac 14 = 0.875$

In binary, Z(D) = 0.111

L(D) = $\left\lceil \log_2 \frac 1 \frac 1 4 \right\rceil + 1$ = 3

code(D) is 111

Algorithm analysis

Prefix code

Shannon–Fano–Elias coding produces a binary prefix code, allowing for direct decoding.

Let bcode(x) be the rational number formed by adding a decimal point before a binary code. For example, if code(C)=1010 then bcode(C) = 0.1010. For all x, if no y exists such that

$bcode(x) \le bcode(y) < bcode(x) + 2^{-L(x)}$

then all the codes form a prefix code.

By comparing F to the CDF of X, this property may be demonstrated graphically for Shannon–Fano–Elias coding.

By definition of L it follows that

$2^{-L(x)} \le \frac 12 p(x)$

And because the bits after L(y) are truncated from F(y) to form code(y), it follows that

$\bar F(y) - bcode(y) \le 2^{-L(y)}$

thus bcode(y) must be no less than CDF(x).

So the above graph demonstrates that the $bcode(y) - bcode(x) > p(x) \ge 2^{-L(x)}$, therefore the prefix property holds.

Code length

The average code length is $LC(X) = \sum_{x \epsilon X}p(x)L(x) = \sum_{x \epsilon X}p(x)(\left\lceil \log_2 \frac {1}{p(x)} \right\rceil + 1)$.
Thus for H(X), the Entropy of the random variable X,

$H(X) + 1 \le LC(X) < H(X) + 2$

Shannon Fano Elias codes from 1 to 2 extra bits per symbol from X than entropy, so the code is not used in practice.

References

1. T. M. Cover and Joy A. Thomas (2006). Elements of information theory (2nd ed.). John Wiley and Sons. pp. 127–128. ISBN 978-0-471-24195-9.