[allenu]

I agree. My theory now it's due the fractal nature of the tasks in software development. To design at a high level and break out the work, you have to make some assumptions about the ground-level designs (i.e. how individual components talk to each other). As you start implementing those, you'll have often encounter unexpected constraints that require you to rework the high-level design. As you progress through implementation, you find more and more tasks that need to be done and eventually the number of tasks generated levels off and starts to decay and it's only then that you really have an idea of how long you need (assuming no large bugs surprise you at that point).

I find even heuristics like double your estimate or padding it aren't useful because there's so much variance.