First, let me emphasize that I'm talking about changing THE past. Accounts with multiple timelines, say where travelling back in time causes a branching timeline, don't have this problem.
Here is the problem stripped down to its bare essentials: Suppose you have a pool table with two balls: a cue ball C and a nine ball N. You also have a time machine just big enough to fit C. You set up the time machine at time t1 so that if a ball appears in it, then the ball will roll out, hit N and knock N into a corner pocket, P, thus changing the past. Nothing happens. At time t2, you put the cue ball C into the time machine, fiddle the knobs to send it back to t1, and press the power button. C disappears. Did C go back in time and change the past? Well, no, because I've already stipulated that nothing happened at t1.
Well, let's change the story. Let's say that at t1, a ball appears in the time machine, rolls out, and knocks N into the corner pocket. Then at t2, you send C back to t1. Have you changed the past? No, because the past has not changed. It was always the case that a ball appeared and knocked N into the pocket, so once again, the past has not changed.
Where do we get the idea that the past can change? In time travel stories, it usually involves some time traveler who remembers an alternative past, and we assume for dramatic purposes that the individual is not hallucinating. So let's incorporate a form of memory into our experiment.
At time t1, nothing happens. At time t2, we write on C a message: "At time t1, no ball came out of the time machine". Then we put C in the time machine and send it back to t1 as before. What happened at t1? Nothing. Again, no change.
But let's say that at t1 C comes out of the time machine and hits the nine ball and sends it into the pocket. We pick C up and there, in our own handwriting is the message, "At time t1, no ball came out of the time machine". Ah hah! Proof that the past has changed, right? Not so fast. At what point in time did the past change? When you say something changed, you mean that at one time, say t_a, it was in one state, and at another time, say t_b, it was in another state. Presumably, t_b here is t1 since at t1 the past had already changed. When was t_a? There is no time at which t_a could have been.
You have to assume a second timeline. There are two ways I can think of to do this: first, there are two parallel timelines and t_a is t1 in the second timeline. What happened is not that the past changed, but that sending C "back in time" actually sent it to an earlier point in a different timeline. You have not changed the past; you have only influenced another timeline. This model is at least coherent, meaning that I can grasp what it means.
Alternatively, the second timeline can be a sort of meta-timeline in which our normal time is an object that can change. I think this is the sort of vague notion most people have in mind when they think of changing the past. In this model, the past is not something fixed; it is something that can change from the perspective of this meta-timeline. But this notion is bizarre. When C comes back with the writing on it, the time when the writing is no more. But then the cause of C having the writing on it is no more. So why does C still have the writing on it? The writing is uncaused. This meta timeline has to hold not only the timeline of changing timelines, but also has to be a repository of ghost causes from past states of the changing timeline.
I don't see how anyone can take this seriously once they follow out the implications. I think it's a consequence of the geometricization of time. It is useful in physics to view time geometrically, and once you are taught to do that, things like time travel seem to make sense because you are unconsciously treating time as a dimension of space where things can change, but time really isn't like that; it's only a way of visualizing time.