Most "shader effects" like filtering and normal mapping have very little in common that could be considered building blocks between effects. 2D filtering and normal mapping aren't any more difficult or advanced than each other, for example, and the same is true of many effects with only a few notable exceptions (the various flavors of shadow mapping come to mind). They're just different. That's why you'll see so many cookbook-style books when it comes to discussing actual graphical effects programming (as opposed to graphics programming at a fundamental level).
Most in-world effects (i.e., not post-processing ones) are rooted at some level in the physical properties of light and the simulation or approximation thereof. To that end, reading books that focus on ray-tracing as a rendering method may help you better understand the core theories involved (Physically Based Rendering is also an excellent read).
Once you understand that, the "scattered theory" present in books like GPU Gems will relate primarily to the specific ways in which the GPU can be manipulated to approximate the real-world physical scenarios desired, and will seem less disjointed.
Similarly, post-processing shaders tend to draw from signal and image processing theory. I think I recall having this book as a textbook in a class on the subject, but don't recall too much about it.
In general, though, you aren't going to find too much that caters to that kind of approach because the theory that binds together all the various shader effects that are in vogue right now is structured like a very, very shallow tree rather than a very deep one -- there are very few dependencies that cater to "bottom up" approaches once you get beyond the fundamentals of graphics programming theory itself (how to structure the scene, the transformation pipeline and associated linear algebra, rasterization, et cetera).