Lecture 15: Huffman Coding
CLRS- 16.3
Outline of this Lecture
Codes and Compression.
Huffman coding.
Correctness of the Huffman coding algorithm.
1
Suppose that we have a character data file
that we wish to store . The file contains only 6 char-
acters, appearing with the following frequencies:
Frequency in ’000s
A binary code encodes each character as a binary
string or codeword. We would like to find a binary
code that encodes the file using as few bits as possi-
ble, ie., compresses it as much as possible.
2
In a fixed-length code each codeword has the same
length. In a variable-length code codewords may have
different lengths. Here are examples of fixed and vari-
able legth codes for our problem (note that a fixed-
length code must have at least 3 bits per codeword).
Freq in ’000s
a fixed-length
a variable-length
The fixed length-code requires bits to store
the file. The variable-length code uses only
bits,
saving a lot of space! Can we do better?
Note: In what follows a code will be a set of codewords, e.g.,
and
3
Encoding
Given a code (corresponding to some alphabet ) and
a message it is easy to encode the message. Just
replace the characters by the codewords.
Example:
If the code is
then bad is encoded into 010011
If the code is
then bad is encoded into 1100111
4
Decoding
Given an encoded message, decoding is the process
of turning it back into the original message.
A message is uniquely decodable (vis-a-vis a particu-
lar code) if it can only be decoded in one way.
For example relative to 010011 is uniquely de-
codable to bad.
Relative to 1100111 is uniquely decodable to bad.
But, relative to 1101111 is not uniquely decipher-
able since it could have encoded either bad or acad.
In fact, one can show that every message encoded
using and are uniquely decipherable. The unique
decipherability property is needed in order for a code
to be useful.
5
Prefix-Codes
Fixed-length codes are always uniquely decipherable
(why).
We saw before that these do not always give the best
compression so we prefer to use variable length codes.
Prefix Code: A code is called a prefix (free) code if
no codeword is a prefix of another one.
Example: is
a prefix code.
Important Fact: Every message encoded by a prefix
free code is uniquely decipherable. Since no code-
word is a prefix of any other we can always find the
first codeword in a message, peel it off, and continue
decoding. Example:
We are therefore interested in finding good (best com-
pression) prefix-free codes.
6
Fixed-Length versus Variable-Length Codes
Problem: Suppose we want to store messages made
up of 4 characters with frequencies , , ,
(percents) respectively. What are the fixed-length
codes and prefix-free codes that use the least space?
7
Fixed-Length versus Variable-Length Prefix Codes
Solution:
characters
frequency
fixed-length code
prefix code
To store 100 of these characters,
(1) the fixed-length code requires bits,
(2) the prefix code uses only
a saving.
Remark: We will see later that this is the optimum
(lowest cost) prefix code.
8
Optimum Source Coding Problem
The problem: Given an alphabet
with frequency distribution find a binary prefix
code for that minimizes the number of bits
needed to encode a message of charac-
ters, where is the codeword for encoding , and
is the length of the codeword .
Remark: Huffman developed a nice greedy algorithm
for solving this problem and producing a minimum-
cost (optimum) prefix code. The code that it produces
is called a Huffman code .
9
n4/100
0 1
n3/55 e/45
0 1
n2/35 n1/20
0 1 0 1
a/20 b/15 c/5 d/15
Correspondence between Binary Trees and prefix codes.
1-1 correspondence between leaves and characters.
Label of leaf is frequency of character.
Left edge is labeled ; right edge is labeled 1
Path from root to leaf is codeword associated with
character.
10
n4/100
0 1
n3/55 e/45
0 1
n2/35 n1/20
0 1 0 1
a/20 b/15 c/5 d/15
Note that , the depth of leaf in tree is equal
to the length, of the codeword in code
associated with that leaf. So,
The sum is the weighted external
pathlength of tree .
The Huffman encoding problem is equivalent to the
minimum-weight external pathlength problem: given
weights , find tree with leaves
labeled that has minimum weighted exter-
nal path length.
11
Huffman Coding
Step 1: Pick two letters from alphabet with the
smallest frequencies and create a subtree that has
these two characters as leaves. (greedy idea)
Label the root of this subtree as .
Step 2: Set frequency .
Remove and add creating new alphabet
.
Note that .
Repeat this procedure, called merge, with new alpha-
bet until an alphabet with only one symbol is left.
The resulting tree is the Huffman code.
12
Example of Huffman Coding
Let be the
alphabet and its frequency distribution.
In the first step Huffman coding merges and .
a/20 b/15 n1/20 e/45
0 1
c/5 d/15
Alphabet is now .
13
Example of Huffman Coding – Continued
Alphabet is now .
Algorithm merges and
(could also have merged and ).
n2/35 n1/20 e/45
0 1 0 1
a/20 b/15 c/5 d/15
New alphabet is .
14
Example of Huffman Coding – Continued
Alphabet is .
Algorithm merges and .
55 e/45
0 1
n2/35 n1/20
0 1 0 1
a/20 b/15 c/5 d/15
New alphabet is .
15
Example of Huffman Coding – Continued
Current alphabet is .
Algorithm merges and and finishes.
n4/100
0 1
n3/55 e/45
0 1
n2/35 n1/20
0 1 0 1
a/20 b/15 c/5 d/15
16
Example of Huffman Coding – Continued
Huffman code is obtained from the Huffman tree.
n4/100
0 1
n3/55 e/45
0 1
n2/35 n1/20
0 1 0 1
a/20 b/15 c/5 d/15
Huffman code is
, , , , .
This is the optimum (minimum-cost) prefix code for
this distribution.
17
Huffman Coding Algorithm
Given an alphabet with frequency distribution
. The binary Huffman tree is constructed using
a priority queue, , of nodes, with labels (frequencies)
as keys.
Huffman( )
;
; the future leaves
for to Why ?
new node;
Extract-Min( );
Extract-Min( );
;
Insert( );
return Extract-Min( ) root of the tree
Running time is , as each priority queue
operation takes time .
18
An Observation: Full Trees
Lemma: The tree for any optimal prefix code must
be “full”, meaning that every internal node has exactly
two children.
Proof: If some internal node had only one child then
we could simply get rid of this node and replace it with
its unique child. This would decrease the total cost of
the encoding.
19
Huffman Codes are Optimal
Lemma: Consider the two letters, and with the smallest fre-
quencies. There is an optimal code tree in which these two let-
ters are sibling leaves in the tree in the lowest level.
Proof: Let be an optimum prefi x code tree, and let and
be two siblings at the maximum depth of the tree (must exist
because is full). Assume without loss of generality that
and (if this is not true, then rename these
characters). Since and have the two smallest frequencies it
follows that (they may be equal) and
(may be equal). Because and are at the deepest level of the
tree we know that and .
Now switch the positions of and in the tree resulting in a differ-
ent tree and see how the cost changes. Since is optimum,
Therefore, , that is, is an optimum tree. By
switching with we get a new tree which by a similar argu-
ment is optimum. The fi nal tree satisfi es the statement of the
claim.
20
Huffman Codes are Optimal
We have just shown there is an optimum tree agrees
with our first greedy choice, i.e., and are siblings,
and are in the lowest level.
Similarly to the proof we seen early for the fractional
knapsack problem, we still need to show the optimal
substructure property of Huffman coding problem.
Lemma: Let be a full binary tree representing an
optimal prefix code over an alphabet , where fre-
quency is defined for each character . Con-
sider any two characters and that appear as sib-
ling leaves in , and let be their parent. Then,
considering as a character with frequency
, the tree represents
an optimal prefix code for the alphabet
.
21
Huffman Codes are Optimal
Proof: An exercise. (Hint : First write down the cost
relation between , and . We then show is an op-
timal prefix code tree for by contradiction (by mak-
ing use of the assumption that is an optimal tree for
.))
By combining the results of the lemma, it follows that
the Huffman codes are optimal.
22