> the present must agree with boundary conditions in the future. This is true but it's not really what's meant by backwards causation here. The specific thing

> the present must agree with boundary conditions in the future. This is true but it's not really what's meant by backwards causation here. The specific thing you need is for what happens from times (-ε, ε) to depend on what will happen at times (1-ε, 1+ε) even after conditioning on the intervening times (ε, 1-ε). Action principles are still local: maximizing the integral of the Lagrangian L(x(t)) over time (by, for instance, taking the shortest path) is equivalent to satisfying dL/dx = d/dt dL/d(dx/dt). > by (virtually) taking all of them. This is a superficially appealing analogy but very misleading. Even if you take Feynman diagrams literally, which you shouldn't, the "path" integral runs over field configurations. The classical notion of a particle following some path simply breaks down here.