Huffman Coding

Supposing you already read the story about Shannon-Fano Coding (and even probably solved the exercise) let us now learn the sequel of it.

One of the authors of that algorithm, Robert Shannon proposed the problem about searching for optimal variable-length code to his student David Huffman who at last came upon brilliant idea - to build the code-tree in "reverse" order - i.e. not by splitting the array (as in case of Shannon-Fano variant) but by joining elements into nodes.

Though the algorithm (as any based on character frequency) is not universally suitable, it is anyway often used. For example while browsing internet you see a lot of images in either JPEG or PNG format. Both of these formats use Huffman at some stage of processing.

Algorithm

Let us start with the example. Given the sample text DAVIDAHUFFMAN let us at first count the usage of letters and sort them for convenience according to counts:

char    count
A       3
D       2
F       2
H       1
I       1
M       1
N       1
U       1
V       1

Now let us at each step take two entries with the smallest counts and join them into a new element - supposing it to be a tree node, so that one entry goes to the right hand and another to the left.

At first we will join U and V. Total count for new node would be 1+1=2 so we reinsert the node into the array:

A=3  D=2  F=2  (U,V)=2  H=1  I=1  M=1  N=1

Two more steps and we join M with N and H with I:

A=3  D=2  F=2  (U,V)=2  (M,N)=2  (H,I)=2

The following step requires us to join pairs of letters M,N with H,I - their total count will be 4:

((M,N),(H,I))=4  A=3  D=2  F=2  (U,V)=2

Then we couple the single letter F with the pair U,V:

((M,N),(H,I))=4  (F,(U,V))=4  A=3  D=2

The following steps are similar - we join rightmost nodes and insert them again according to their total count:

(A,D)=5  ((M,N),(H,I))=4  (F,(U,V))=4           - A and D

(((M,N),(H,I)),(F,(U,V)))=8  (A,D)=5            - quadruple M-N-H-I with triple F-U-V

((((M,N),(H,I)),(F,(U,V))),(A,D))=13            - the whole tree is built

You probably noticed that when we join the two rightmost nodes we still preserv their order so that most frequent goes to the left branch and least frequent to the right.

Our resulting record resembles one used in Tree Builder exercise, so let us plot the tree:

                             |
                   +---------+--------+
                   |                  |
          +--------+-------+      +---+---+
          |                |      |       |
    +-----+-----+      +---+---+  A       D
    |           |      |       |
+---+---+   +---+---+  F   +---+---+
|       |   |       |      |       |
M       N   H       I      U       V

As earlier, bit-sequence for each letter is defined by the way from the root of the tree. For example, let us now use 1 for left branches and 0 for right ones:

char    code
A       01
D       00
F       101
M       1111
N       1110
H       1101
I       1100
U       1001
V       1000

The total message length will be 40 bits for 13 characters, i.e. about 3 bits per character. To calculate this we simply write down the letters, length of their codes and counts:

char    count   length
A         3       2
D         2       2
F         2       3
H         1       4
I         1       4
M         1       4
N         1       4
U         1       4
V         1       4

then multiply counts by lengths and at last summarize them all:

3*2 + 2*2 + 2*3 + 1*4 + 1*4 + 1*4 + 1*4 + 1*4 + 1*4 = 40

It was proved by Huffman that the code constructed by this algorithm is optimal - i.e. no any code using whole amount of bits for each character allow to compress text into smaller resulting bit-string if it relies only upon the frequencies of the characters.

Better results could still be achieved with Arithmetic Coding which in its core allows to use fractional amounts of bits.

Problem Statement

Given a fragment of text you will be asked to find the compression ratio provided by Huffman's algorithm:

             original size
ratio = -----------------------
            compressed size

I.e. in case with DAVIDAHUFFMAN original size is 13*8=104 bits while compressed size is only 40 bits, so the compression ratio is 2.6. We assume that original text was encoded using 8 bits per character.

*With real

Input data will contain a sample of text, consisting of letters, punctuation marks, spaces and probably digits.
Answer should contain a single real value - the compression ratio (with 1e-6 precision or better).

Example:

input data:
DAVIDAHUFFMAN

answer:
2.6

Note that real implementation of the algorithm also needs to store the code-table or counts-table in the resulting file - this slightly reduces the compression ratio especially if the original size is comparatively small.

Homework: would you like to compare compression ratio for Huffman algorithm with one of Shannon-Fano?