Most image rotation algorithms now try to be "lossless", in the
sense that they do not actually decompress, rotate, and recompress
the image, but simply perform a "transpose" operation on the compressed
data, and on the 8x8 pixel macroblocks that comprise the compressed
data.

This operation can only be done perfectly if the pixel dimensions of
the image are divisible by eight in both directions.  If this is not
the case, then a few rows or columns of data must be discarded from
the edge(s) to make the image conform.

This will result in a small change in image size, but, once done,
the image dimensions will now be a multiple of 8 pixels and no
further loss should occur if additional lossless rotations are
performed.

The reason for preferring this approach is that JPEG is a lossy
compression algorithm, and, though decompressing and recompressing
an image without alteration using exactly the same compression
parameters does not cause information loss, any change in either
the image or the compression (particularly increasing the compression)
will, in general, cause some loss of information.  This loss is
usually practically invisible under normal conditions, but it is
a good reason to avoid sequential decompression, modification, and
recompression cycles on images.

If you need to make multiple, separate modifications to an image,
try to make them either all in one session, with only one decompression
at the beginning and a final re-compression at the end, or use a
lossless compression for storing the intermediate results, like TIFF.

Also, it is good practice never to save a modified image "over" the
original image, since you may wish to return to the original to re-do
some processing.

-michael