1. Particles can go offscreen -> useless to update them
They can come back onscreen as well. So you'd better update them.
2. Particles can have alpha <= 0 -> useless to update them
Particles should not reach an alpha of zero until their lifetime is complete. I make mine fade out over the last 25% of their life (and fade in over the first 10%).
The cost of updating a single particle should be negligible. The fillrate of drawing all those blended layers will become a bottleneck long before the CPU processing does.
You should think of a particle emitter as a single dynamic mesh, not as a collection of parts.