This is the reason why Lucene scoring was a little bit off for searches with very short terms (if I recall correctly)

gistfile1.java

	private final Similarity similarity = new DefaultSimilarity()
	{
		private static final long serialVersionUID = 1L;

		@Override
		public byte encodeNormValue(final float f)
		{
			return SmallFloat.floatToByte52(f);
		}

		@Override
		public float decodeNormValue(final byte b)
		{
			return SmallFloat.byte52ToFloat(b);
		}
	};