Compression - Fake advantage

Author

Fake advantage

and_polar

2005-11-10, 10:54 pm

I have spent last few w

s testing and comparing several arithmetic encoders and Huffman encoding for large alphabets 1000 to 32000 symbols. I read several on-line articles and books regarding arithmetic encoding as well. They all stated advantage of arithmetic encoding beyond the doubt. Unfortunately, testing results were opposite. In practice, for the large alphabets, Huffman encoding gives the result on a fraction of percent larger the entropy, something like 0.3%. The arithmetic encoders theoretically should beat this limit, however, I can’t imagine the case when anybody was interested in improvement of compression on 0.3%. More than that, as it appeared to be, arithmetic encoders are not better, because of approximation. They, typically, 0.5% off the limit and in addition to that at least 3.5 times slower than Huffman. I tested Huffman encoder on smaller alphabets (200 symbols) and found that it is typically 1.5% off the limit. I returned back to multiple articles and discovered that authors of arithmetic encoders always say that arithmetic encoders must have better compression, but they, however, did not mention how better.
There is a difference between researchers and used cars dealers. Researchers say themselves about di

vantages of the methods that they suggested. If the whole industry of arithmetic encoding is created to beat 1.5% in compression while making processing 3 times slower they should at least mention that in papers.
I’d be glad to hear other opinions, Bodden, Sayood, Said, may be some other engineers.

Thanks.

Sponsored Links