CSE 102 – Tutorial 8 – Binary representation

After you are finished with the tutorial, please upload your submission to Moodle by the indicated deadline (end of the day Friday, May 10th).

Warmup: bit operations in Python

# To print the binary representation of a number, use the function `bin`...
print(bin(42))
# ...or the format spec {:b}.
print('0b{:b}'.format(42))

# Bitwise-or is done with |
x = 9
y = 10
z = x | y
print(f'{x} | {y} = {z}')
print(f'0b{x:b} | 0b{y:b} = 0b{z:b}')

# Bitwise-and is done with &
z = x & y
print(f'{x} & {y} = {z}')
print(f'0b{x:b} & 0b{y:b} = 0b{z:b}')

# Bitwise-xor is done with ^
z = x ^ y
print(f'{x} ^ {y} = {z}')
print(f'0b{x:b} ^ 0b{y:b} = 0b{z:b}')

# Bitwise-not is done with ~
z = ~ x
print(f'~ {x} = {z}')
print(f'~ 0b{x:b} = 0b{z:b}')

# Bitwise shift-left is << and shift-right is >>
z = y << 2
w = y >> 2
print(f'{y} << 2 = {z}')
print(f'0b{y:b} << 2 = 0b{z:b}')
print(f'{y} >> 2 = {w}')
print(f'0b{y:b} >> 2 = 0b{w:b}')

Generating pseudorandom numbers

Recall from Lecture 4 that a pseudorandom number generator (PRNG) is a deterministic algorithm generating a sequence of numbers (or other objects) whose properties approximate that of a random sequence. In this section, you will implement a simple kind of PRNG known as a linear feedback shift register (LFSR). To explain how this works, let’s consider the following illustration of an LFSR generating 16-bit numbers:

In the diagram, the bit vector may be interpreted as holding the current value of the sequence, namely 44257 (= 0b1010110011100001). Note that the positions are numbered from right to left, running from the least significant digit to the most significant digit. The indices 0, 2, 3, and 5 are called tap positions, and are used to compute the next value of the sequence as follows: shift all of the bits of the current value over to the right, while taking the sum modulo 2 (also known as the exclusive or, or xor) of the bits in the tap positions as the new value for the leftmost digit (index 15). In this example, the LFSR will produce 22128 (= 0b0101011001110000) as the next value of the sequence.

More formally, for any 16-bit number \(X\) and index \(0 \le i < 16\), let us write \(X[i]\) for the bit in the \(i\)th position of \(X\). Let \(I \subseteq \{0,\dots,15\}\) be any subset of tap positions. The LSFR based on \(I\) generates a sequence of 16-bit numbers \((X_n)_{n\in\mathbb{N}}\) defined as follows:

• \(X_0\) is some seed value
• \(X_{n+1}\) has:
  • \(X_{n+1}[15] = \sum_{i \in I} X_n[i]\) (modulo 2)
  • \(X_{n+1}[i] = X_n[i+1]\) for all \(0 \le i < 15\)

6. The operation of “tapping” the \(i\)th least significant bit of a 16-bit number can be naively implemented in terms of uint16_to_bitstring, although this is inefficient:

   def slowtap_uint16(x, i):
       return uint16_to_bitstring(x)[15-i]

   Complete the following by writing a better version of the tap operation, which computes the result directly using bitwise arithmetic:

   def tap_uint16(x, i):
       """ Return 1 or 0 depending on whether the ith-least significant bit
           of x is 1 or 0.
       """
       pass #FIXME

   You can test your function by comparing it to the result of slowtap_uint16 above.

7. Write a function that takes a 16-bit number and a list of tap positions, and returns the sum modulo 2 of all the values in the tapped positions.

   def polytap_uint16(x, I):
       """ Tap x at all the positions in I (which is a list of tap
           positions) and return the xor of all of these tapped values.
       """
       pass #FIXME

   For example, you should have polytap_uint16(44257, [0, 2, 3, 5]) = 0.

8. Finish the implementation of the LFSR function below. It should compute the next number in the sequence \(X_i\) defined above.

   def lfsr_uint16(x, I):
       """Return the successor of x in an LFSR based on tap positions I"""
       pass #FIXME

   You can test it using the following code:

   def test_lfsr_uint16(seed):
       I = [0, 2, 3, 5, 15]
       xs = [seed]
       for _ in range(10):
           xs.append(lfsr_uint16(xs[-1], I))
       return xs
   
   assert test_lfsr_uint16(42) == [42, 21, 10, 32773, 49154,
                                   57345, 28672, 14336, 7168,
                                   3584, 1792]

Huffman coding

Lossless data compression is the operation of transforming a sequence of bits into a shorter sequence of bits that encodes the same information – the initial sequence of bits being reobtained using a decompression algorithm. In short, it is a coding operation which shortens the size (transmission, storage) of the data at the cost of compression work.

Huffman coding is a variable length coding, developed by David Albert Huffman while a student at MIT. The algorithm has been published in 1952 in a paper entitled “A Method for the Construction of Minimum-Redundancy Codes”.

In this section, we are interested in compressing (character) strings. In a normal mode of operation, when using the extended ASCII encoding, strings are encoded using 8 bits per character. To get the binary representation of a string, we can use the .encode() method. For example, the following function takes a string, gets its ASCII encoding as an array of bytes (represented using the class bytearray in Python), and formats it so that it is human-readable: a string made of 0s and 1s, one character per bit, where groups of 8 bits are separated by a space (again, for readability).

def bin_repr_of_ascii_string(s):
    return ' '.join(format(x, '08b') for x in s.encode('ascii'))

For instance, on the string "abc", we obtain the following representation:

01100001 01100010 01100011

We see that the ASCII encoding of "abc" is 3 bytes long (as expected) and one can check, for example, that the byte 0b01100001 (97 in decimal notation) corresponds to the character a in the ASCII table.

(Note that in extended ASCII, since characters are encoded using 8-bit words, you are limited to 256 characters. Basically, the encoding space is barely large enough to encode latin characters without diacritic marks. As of today, there exist dozens of different encodings, some of them using multiple bytes per character, or even variable-width, i.e. the encoding of one character may vary in size. For this tutorial, we will stick to extended ASCII.)

Under Huffman coding, compression is achieved by replacing the most frequent characters with short codes (< 8 bits) and the least frequent characters with long codes (> 8 bits). In average, size reductions of between 20 and 90% are thus observed. It is a widely used compression method, often as a complement to other algorithms.

Huffman uses a particular approach to choose the representation of each coded symbol. The idea is that each code should not be a prefix of any other code. In other words, none of the codes chosen should be the beginning of another code. We will see that more in details later.

Description of the Huffman coding

The principle of Huffman coding is based on the creation of a tree structure called a Huffman tree. That tree is built as follows: suppose the sentence \(s\) to be compressed is

Forsan et haec olim meminisse juvabit

We first compute the probability of appearance, in the sentence, of each character. For example, in \(s\), the character c appears once, while the character a and the space character appear \(3\) and \(5\) times, respectively. Since \(s\) is of length \(37\), the inverse probabilities of appearance of c, a and the space character are resp. equal to \(\frac{1}{37}\), \(\frac{3}{37}\) and \(\frac{5}{37}\). We give below the probability of appearance for all the characters of the sentence:

Then, the tree structure is constructed as follows. For each character \(c\) with probability \(p\), we start by constructing a set of isolated leaf nodes with value \(c\) and associated probability \(p\). Then, we iteratively repeat the following process: we select two nodes with lowest associated probabilities and remove them from the set – let’s call them \(n_1\) and \(n_2\), with associated probabilities \(p_1\) and \(p_2\). We then create a new node with children \(n_1\) and \(n_2\) (the order is irrelevant here), with no data and with associated probability \(p_1 + p_2\), and add it to the set. Eventually, the set will only contain one node that forms the Huffman tree.

To obtain the Huffman code of a character \(c\), we walk down the tree from the root to the leaf that contains \(c\), adding a 0 (resp. a 1) to the code each time we take a left-edge (resp. a right-edge).

The construction of a Huffman tree for the probability table given above is shown below:

Here we have also indicated a path from the root down to a leaf node u, which goes right, then left twice, then right, and finally left. We thus obtain the code 10010 for u, the code 0101 for t, and so on. You can check that in general, the letters with a higher probability of appearance have shorter codes.

Last, one can encode an entire string by encoding its characters one by one, using their respective Huffman codes, and by concatenating all the obtained codes.

Implementation

We now have all the information for implementing the Huffman coding. In this part, we are not going to manipulate bits (and bit strings) but are going to represent codes with standard strings made of 0s and 1s. Note that this somewhat defeats the goal of Huffman coding since we are going to represent \(1\) bit with a character which in turn requires (at least) \(8\) bits to be stored — but still it will help us to get an understanding of the underlying principles of Huffman coding!

Binary Huffman Implementation (optional)

In the previous section, we saw the principle of Huffman encoding and developed basic encoding & decoding functionalities. However, to obtain a functional implementation, some work remains:

• first, for representing bit-strings, we used (character) strings made of 0s and 1s. That is, we used 1 character (i.e. 8 bits if we assume the ASCII encoding of characters) for encoding 1 bit of information! In that setting, the function huffman_encode is more an anti-data-compression implementation than a data-compression one.

• then, the functions huffman_encode & huffman_decode assume that you have the relevant Huffman tree at disposition. For huffman_encode, this is not a problem since you have the non-encoded message at disposition and the tree can be reconstructed from it. But for the huffman_decode function, you cannot assume that – the goal of the function is to decode the message without already having access to its decoding! In consequence, the Huffman tree must be part of the encoded message, i.e. we have to find a way to encode a Huffman tree using bit-strings. Note that encoding must be efficient (in term of space) because now, the encoding of the Huffman tree is going to be part of compressed data – the more space it takes, the less the compression ratio will be. (There are other solutions. E.g., the Huffman tree could be fixed a priori and be a pre-shared knowledge between the compressor and the decompressor. For example, fax machines¹ use a modified Huffman algorithm where the Huffman tree is fixed for all machines & transmissions.)

• last, for strings that are made from only one character (e.g. aaaaaa), the Huffman tree is made of a single node. In consequence, the repeated character is going to be encoded with a code of length \(0\). And thus, all strings of that form are going to be compressed as the empty bit-string and uncompressed as the empty string – you can check this with your implementation. We will have to detect such cases.

As a first step, we will adapt the function huffman_encode & huffman_decode to manipulate bitstrings. First, we need to modify huffman_codes:

We can now modify the huffman_encode function to return actual bit-strings. However, since the minimal unit of information a computer can manipulate is a byte, we are going to use byte arrays (or byte strings) to store our encoded messages. Python comes with two classes for manipulating byte arrays: bytes and bytearray. The type bytes (documentation) is very similar to the one for strings: it is an immutable array of bytes (i.e. integers in the range \([0 \ldots 255]\)) and you can iterate over it or randomly access its elements. The type bytearray (documentation), on the other end, represents mutable arrays of bytes. Its interface is close to the one of lists and uses the same syntax for accessing, adding or removing elements. Here are some examples of their usage:

# Create a immutable byte-string of length 4 containing the bytes:
#
# Decimal      Binary
# -------------------
#      97  0b01100001
#      98  0b01100010
#      99  0b01100011
#       1  0b00000001
#
b = bytes([97, 98, 99, 1])
print(b)

# Note that if you print `b`, Python is going to display this
#
#   b'abc\x01'
#
# First note the `b` in front of quote - it indicates that this
# is a byte string literal. Then, we can see that bytes are represented
# by characters! Indeed, Python displays bytes using the associated
# character in the ASCII encoding (when possible). If not, it displays
# \x?? where ?? is the value of the byte in hexadecimal notation (i.e.
# in base 16, where the digits are 0123456789abcdef) -- recall that 'a'
# (resp. 'b' & 'c') are encoded by 97 (resp. 98 & 99) in ASCII.

# We can now access different properties of b:
assert len(b) == 4     # b contains 4 bytes
assert b[0] == 97      # its first byte if 0
assert b[1:2] == b'bc'

# We can then create a mutable byte-array...
mb = bytearray([97, 98, 99, 1])
# or: mb = bytearray(b'abc\x01')

# ...and modify it:
mb[0] = 100
mb.append(101)

# ...and get its content:
print(len(mb))
print(bytes(mb))

Python also gives all the necessary functions for converting characters in bytes and bytes in characters:

# We can convert a string to its ASCII encoding...
s = 'abc'
e = s.encode('ascii')

# ...and convert it back to a string
s2 = e.decode('ascii')
assert s == s2

# Moreover, the function `chr(x)` gives the character whose
# ASCII code is `x` (as an integer). A contrario, the function
# `ord(c)` gives the ASCII code (as an integer) of the
# character `c`

We are now ready to write binary versions of huffman_encode and huffman_decode:

Ok, we are not far from the end. What remains is to (binary) encode the Huffman tree and to pack all things together. There are numerous ways of encoding a Huffman tree. Here is one possible way:

• the bit encoding of a leaf node is the bit 1 followed by the 8-bit long ASCII code of the character this node encodes (as returned by ord),

• the bit encoding of an internal node is the bit 0 followed by the encoding of its left children and its right children (in that order).

Finally, the binary encoding of a Huffman tree is simply the binary encoding of its root padded with 0s at the end if the encoding length (in bits) is not a multiple of 8.

We are now ready to write our final compression function. The encoding of a string s is the contenation of:

• its length (on 4 bytes, using little endian). In Python, you can obtain this encoding using int.to_bytes(x, 4, 'little') where x is the integer to encode (you can decode it using int.from_bytes(b, 'little')).

• the binary encoding of the Huffman tree t obtained from the frequency analysis of s.

• the binary encoding of s under the tree t.