The reduction method means reducing a big cube (NxNxN for N>3) to a 3x3x3 cube by first solving the centers (the central (N-2)x(N-2)x(N-2) square on each face) and the edges (the inner N-2 pieces along each edge of the cube). You are then essentially left with a 3x3x3 cube that you can try to solve by only turning the outer layers (which won't break the centers and edges you solved in the first stage).
The problem with this is that you may end up with a 3x3x3 cube that is not solvable. For instance, you can get a state where the entire cube is solved, except for two edges that need to swap locations. This isn't possible. In group theoretical language, only even permutations are possible. You can swap two _pairs_ of edges, but not just two edges.
When you end up in such an unsolvable 3x3x3 cube, you have to temporarily turn the inner layers of the cube and break apart the centers and edges you built in the first step, and then reassemble them again to a solvable 3x3x3 cube.
The problem with this is that you may end up with a 3x3x3 cube that is not solvable. For instance, you can get a state where the entire cube is solved, except for two edges that need to swap locations. This isn't possible. In group theoretical language, only even permutations are possible. You can swap two _pairs_ of edges, but not just two edges.
When you end up in such an unsolvable 3x3x3 cube, you have to temporarily turn the inner layers of the cube and break apart the centers and edges you built in the first step, and then reassemble them again to a solvable 3x3x3 cube.