Lossless Compression Handbook

376 PART iV / STANDARDS

Since dynamic coding can tailor itself to the input data, it tends to be more efficient for larger

data sets. However, unlike the case with fixed codes, dynamic Huffman trees must be explicitly

included in the output stream (literal/length tree first, then distance tree). Thus, for smaller input

streams, where the overhead of the trees outweighs their improved compression efficiency, fixed

codes are better. Because deflate organizes its output into blocks--something at which we hinted

earlier, with the end-of-block code~it can alternate between fixed and dynamic Huffman codes

as necessary in order to optimize compression. ...