* Define some concept of how good you can expect things to get
* Find some way of improving what you've got
* Prove that you get to within some distance of "the best"
* Prove that it terminates in polynomial time.
I've not see the details of the proofs - they're likely to be mildly difficult, with a very nasty details. But I don't think they're especially deep. The really deep bit was getting the exact formulation right to make it work.
Things like the spectral theorems, diagonalization, eigenvalues and eigenvectors, these are all deeper results that have far-reaching implications. For example, the reason that when you tap the rim of a coffee-cup you get two different notes, blends of those notes, and no other notes is to do with eigenvectors.
Similarly for why Fourier transforms work (in part) the way they do. The reasons that the Sine and Cosine functions form an infinite-dimensional basis for the linear space of non-pathological functions, for example, let us then deal with frequency space as a dual space of function space.
These are deeper.
I can only study this stuff in my spare time so my knowledge is very patchy. But when I see stuff like eigenvectors I associate them towards a much more powerful concept of invariance - which I see everywhere. Recursion, programming, Machine Learning, dynamical systems, physics. And I feel on the edge of an epiphany but do not have sufficient knowledge to fully grasp it.
But it makes me think that all these things that appear over and over again in different guises, the subject is taught far too inefficiently. I feel there is a much simpler way to teach maths and have it flow and be holistic rather than a replay of history: disjointed and disconnected.
So then my next question is, relative to the whole of linear algebra, how advanced/deep/engaged-with-the-field is the LLL algorithm? Is this first-course-in-linear-algebra stuff? Is this "barely linear algebra" stuff?