Arithmetic coding

Dario Phong/Hugi

Introduction

Arithmetic coding is a widely used entropy coder. The only problem is its speed, but compression tends to be better than what Huffman can achieve. This article presents a basic arithmetic coding implementation. If you have never implemented an arithmetic coder, this is the article which suits your needs. Otherwise look elsewhere for better implementations.

Arithmetic coding

The idea behind arithmetic coding is to have a probability line, 0.0 to 1.0, and for each symbol assign a part of this line, a 'range', to it based on the symbol's probability. The higher the probability of a symbol appearing, the bigger the range it is given on the line, and likewise a symbol which has a smaller probability of appearing is given a smaller portion of the line. Once we have assigned the ranges along the probability line, we can start to encode symbols. Every symbol defines where the output floating-point number lands. Let's say we have:

Symbol Probability Range
a 2 [0.0 , 0.5)
b 1 [0.5 , 0.75)
c 1 [0.75 , 1.0)

Note that the "[" means that the number is also included, so the symbol "a" owns the number range from 0 up to (but not including) 0.5. This basically means 0.0 up to 0.4999999... Once we have these ranges along the 0.0 to 1.0 probability line we can then start to encode the symbols and compute our output number. The algorithm to compute the output number is:

- Low = 0

- High = 1

- Loop. For all the symbols:

- Range = High - Low

- High = Low + range * high_range of the symbol being coded

- Low = Low + range * low_range of the symbol being coded

Where Range keeps track of where the next range should be, and High and Low specify the output number.

And now let's see an example:

Symbol Range Low value High value
. . 0 1
b 1 0.5 0.75
a 0.25 0.5 0.625
c 0.125 0.59375 0.625
a 0.03125 0.59375 0.609375

The output number will be 0.59375. The way of decoding is first to see where the number lands, output the corresponding symbol, and then extract the range of this symbol from the floating point number. The algorithm for extracting the ranges is:

- Loop. For all the symbols:

- Range = high_range of the symbol - low_range of the symbol

- Number = number - low_range of the symbol

- Number = number / range

And this is how decoding is performed:

Symbol Range Number
b 0.25 0.59375
a 0.5 0.375
c 0.25 0.75
a 0.5 0

You may reserve a little range for an EoF symbol, but in the case of an entropy coder you'll not need it (the main compressor will know when to stop), and with a stand-alone codec you can pass to the decompressor the length of the file, so it knows when to stop. (I never liked having a special EoF.)

Implementation

As you can see from the example it is a must that the whole floating point number is passed to the decompressor. No rounding must be performed. But with today's FPU the highest possible precision is about 80 bits, so we can't work with the whole number. We'll need to redefine our range. Instead of 0.0 to 1.0 it will be 0000h to FFFFh, which in fact is the same. And we'll also reduce the probabilities so we don't need the whole part, only 16 bits. Don't you believe that it's the same? Let's have a look at some numbers:

0.000 0.250 0.500 0.750 1.000
0000h 4000h 8000h C000h FFFFh

If we take a number and divide it by the maximum (FFFFh) we will clearly see:

- 0000h: 0/65535 = 0.0

- 4000h: 16384/65535 = 0.25

- 8000h: 32768/65535 = 0.5

- C000h: 49152/65535 = 0.75

- FFFFh: 65535/65535 = 1.0

Ok? We'll also adjust the probabilities so the bits needed for operating with the number aren't above 16 bits. And now, once we have defined a new interval and are sure that we can work with only 16 bits, we can start to do it. The way we deal with the infinite number is to have only the 16 first bits loaded and when needed shift more onto it:

1100 0110 0001 000 0011 0100 0100 ...

We work only with those bytes, as new bits are needed they'll be shifted. The algorithm of arithmetic coding means that if ever the msb of both high and low match are equal, then they'll never change. This is how we can output the higher bits of the output infinite number and continue working with just 16 bits. However, this is not always the case.

Underflow

Underflow occurs when both High and Low get close to a number but their msbs don't match, for example High = 0.300001 and Low = 0.29997. If we ever have such numbers and they continue getting closer and closer, we won't be able to output the msb, and then in a few iterations our 16 bits will not be enough. What we have to do in this situation is to shift the second digit (in our implementation the second bit) and when finally both msbs are equal also output the discarded digits.

Gathering the probabilities

In this example we'll use a simple static order-0 model. You have an array initialized with 0 and you count the occurrences of every byte. And then once we have them we have to adjust them, so they don't make us need more than 16 bits in the calculations. If we want to accomplish that, the total of our probabilities should be below 16384 (2^14). To scale them we can divide all the probabilities by a factor until all of them fit in 8 bits. However, there's an easier (and faster) way of doing so: You get the maximum probability and divide it by 256. This is the factor that you'll use to scale the probabilities. Also, when dividing, if the result ever is 0 or below, change it to 1, so the symbol has a range. The next scaling deals with the maximum of 2^14, add together all the symbol probabilities, and then check if it's above 2^14. If it is then divide them by a factor (2 or 4), and the following assumptions will be true:

- All the probabilities are inside the range of 0-255. This helps saving the header with the probabilities.

- The addition of all of them doesn't get above 2^14, and thus we'll need only 16 bits for the computations.

Saving the probabilities

Our probabilities are one byte long, so we can save the whole array. At most this will be 256 bytes long, and it's only written once, so it will not hurt compression a lot. If you expect some symbols not to appear you could RLE-code it. If you expect some probabilities to have lower values than others, you can use a flag to say how many bits the next probability uses and then code one with 4 or 8 bits. In any case you should tune the parameters.

Assign ranges

For every symbol we have to define its high value and low value. They define the range. Doing this is rather simple: We use its probability.

Symbol Probability Low High 0-1 (x/4)
a 2 0 2 [ 0.0 , 0.5 )
b 1 2 3 [ 0.5 , 0.75 )
c 1 3 4 [ 0.75 , 1 )

What we'll use is high and low, and when computing the number we'll perform the division, to make it fit between 0 and 1. Anyway, if you have a look at high and low you'll notice that the low value of the current symbol is equal to the high value of the last symbol. We can use it to use half the memory. We only have to care about setting the -1 symbol with a high value of 0:

Symbol Probability High
-1 0 0
a 2 2
b 1 3
c 1 4

And thus when reading the high value of a symbol we read it in its position, and for the low value we read the entry "position-1". I think you don't need pseudo code for doing such a routine, you just have to assign to the high value the current probability of the symbol plus the last high value and set it up with the symbol "-1" with a high probability of 0. I.e.: When reading the range of the symbol "b", we read its high value at the current position (of the symbol in the table), "3", and for the low value, the previous: "2". And because our probabilities take one byte, the whole table will only take 256 bytes.

Pseudo code

And this is the pseudo code for the initialization:

- Get probabilities and scale them

- Save probabilities in the output file

- High = FFFFh (16 bits)

- Low = 0000h (16 bits)

- Underflow_bits = 0 (16 bits should be enough)

Where High and Low define where the output number falls, and where Underflow_bits is the bits which could have produced an underflow and thus they were shifted.

And the routine to encode a symbol:

- Range = ( high - low ) + 1

- High = low + ( ( range * high_values [ symbol ] ) / scale ) - 1

- Low = low + ( range * high_values [ symbol - 1 ] ) / scale

- Loop (will exit when no more bits can be output or shifted)

- Msb of high = msb of low?

- Yes

- Output msb of low

- Loop. While underflow_bits > 0 (Let's output underflow bits pending for output)

- Output Not ( msb of low )

- Go to shift

- No

- Second msb of low = 1 and Second msb of high = 0 ? (Check for underflow)

- Yes

- Underflow_bits += 1 (Here we shift to avoid underflow)

- Low = low & 3FFFh

- High = high | 4000h

- Go to shift

- No

- The routine for encoding a symbol ends here.

Shift:

- Shift low to the left one time (Now we have to put in low and high new bits)

- Shift high to the left one time, and or the lsb with the value 1

- Repeat to the first loop

Some explanations:

- Note that the formulae before the loop should be done with 32 bit precision. (dword, long)

- Msb of high means the following, with a 16 bits number like that: abbb bbbb bbbb bbbb, a is the msb bit of it.

- Not ( msb of low ) is the "bitwise complement operator" of C, like in Assembler "not ax". First you perform a "not" on low and then you output its msb bit.

- "&" means "bitwise and".

- "|" means "bitwise inclusive or" (also called "or").

- Range must be 32 bits long, because the formula needs this precision.

- Scale is the addition total of all the probabilities.

Once you have encoded all the symbols you have to flush the encoder (output the last bits): output the second msb of low and also underflow_bits+1 in the way you output underflow bits. Because our maximum number of bits is 16 you also have to output 16 bits (all of them 0) so the decoder will get enough bytes.

Decoding

The first thing to do when decoding is read the probabilities, because the encoder did the scaling, you just have to read them and do the ranges. The process will be the following: see in what symbol range our number falls into and extract the code of this symbol from the code. Before starting we have to init "code". This value will hold the bits from the input. Init it to the first 16 bits in the input. This is how it's done:

- Range = ( high - low ) + 1 (See where the number lands)

- Temp = ( ( code - low ) + 1 ) * scale) - 1 ) / range )

- See what symbols corresponds to temp.

- Range = ( high - low ) + 1 (Extract the symbol code)

- High = low + ( ( range * high_values [ symbol ] ) / scale ) - 1

- Low = low + ( range * high_values [ symbol - 1 ] ) / scale
(Note that those formulae are the same that the encoder uses)

- Loop.

- Msb of high = msb of low?

- Yes

- Go to shift

- No

- Second msb of low = 1 and Second msb of high = 0 ?

- Yes

- Code = code ^ 4000h

- Low = low & 3FFFh

- High = high | 4000h

- Go to shift

- No

- The routine for decoding a symbol ends here.

Shift:

- Shift low to the left one time. (Now we have to put in low, high and code new bits)

- Shift high to the left one time, and or the lsb with the value 1

- Shift code to the left one time, and or it the next bit in the input

- Repeat to the first loop.

When searching for the current number (temp) in the table we use a for loop, which, based on the fact that the probabilities are sorted from low to high, has to do one comparison in the current symbol, until it's in the range of the number.

Closing words

First of all thanks to Mark Nelson for some help. This is the first version of this article. I hope to be able to mend possible mistakes, which you should report if you find. Also any idea is welcome. There are faster implementations, but this was only an introduction, once you have a good encoder you only need a good model to have good compression, so do a little bit of research.


Dario Phong/Hugi