The problem with digital watermarking is that it seeks to break mathematics and information theory. The trick is to keep in mind perceptual content and compressibility. The real meaningful information in a digital file that represents an analog source is whatever features can be perceived by a human (or perhaps in certain cases by a machine, but the issue is the same). Anything that cannot be perceived is noise, not meaningful content; not every bit in a digital representation of analog data is necessarily information in the proper sense. An ideal (lossy) compression algorithm for analog data (MP3 and Ogg Vorbis come close for sound; JPEG comes close for images; video compression techniques are still subject to large improvements) keeps every perceptible feature of the representation while discarding every non-perceptible feature. While you cannot know for certain the "ideal-ness" of a single digital representation, a smaller representation producing the same perceptible features is closer to this "ideal." A digital watermark is, by definition, a non-perceptible feature (otherwise the perceiver could simply remove it). In other words, the watermark adds entropy to the digital encoding, while doing nothing to add meaningful information to the representation.

SDMI is a good illustration. In developing a music format that includes copyright identification (digital) information, the RIAA has exactly two choices at a conceptual level: (1) They can increase the size of music files over the size of an ideally compressed format in order to include the copyright identification; (2) They can replace some of the analog information in the digital representation with copyright information (in other words, make the format sound worse to a discerning ear). The exact same tradeoff exists for watermarks in images and other analog sources. In practice, no digital watermarking format has ever stood up to any serious scrutiny, and watermarks have always proven to be relatively easy to remove once analyzed. In theory, there is an inherent conflict between the goals of maximum compression and inclusion of a watermark.