Basically that's what I'm working on now
I'm not surprised. I had the feeling that the idea was around. That's why I started this thread.
The key is to get T(z) to be close to conformal. Otherwise you end up with stretching.
Conformal mapping in 3-D is very limited: scaling, translation, reflection, rotation, and sphere inversion. Tglad's Mandelbox uses all of these but rotations. Also, he made sure the transformations were continuous.
I agree with David Makin about this point. For example, z[n+1]=z[n]^p+c is non conformal (I don't know if it is in 2d but in highier dimensions it's not). I think that folding and stretching are fundamental. Some other properties are perhaps necessay. For example the asymptotic behaviour of the whole formula. In the mandelbox or escape time ifs it's z*scale^i, in mandelbrot like fractals it's z^(power^i).
This (along with this thread:
http://www.fractalforums.com/3d-fractal-generation/tglad's-mandelbox-and-using-the-delta-de-methods-for-rifs/) was the stating point of what i've called "Kaleidoscopic IFS". The key difference is the explicit folding. At first I wanted to make your algorithm GPU friendly.
I think your general strategy certainly makes sense
but I think we can go further by specifying what the choices of Tn should be. As msltoe said, the transformations should generally be conformal mappings. Also, I've noticed that the best results are usually caused by a conflict of coordinate systems--the Mandelbrot, for example, is largely the product of scaling in polar coordinates (z^2) and translating in rectilinear coordinates (+ c). By contrast, using JUST polar scaling or translation does not produce anything very interesting. So may I suggest this:
z: initial point;
for (i = 0; i < max_iterations and somenorm(z) < bailout; i++) {
transformation T = P(i);
coordinate_system C = Q(i);
z = T(z, C);
}
where T is some conformal transformation, but applied within the context of the given coordinate system.
I think that your suggestion is equivalent to mine: the coordinate system may be integrated into the transform.
You asked a good question about what the choices of Tn should be.
Other questions are about how to find a good distance estimate (
http://www.fractalforums.com/3d-fractal-generation/de-calculation-for-'compound'-fractals/) and smooth iteration.