Yes social networks are sparse. With 300 million Americans and the capacity to know perhaps only a few thousand, yes, things will be sparse. But that doesn't mean you cannot construct a spanning tree of that sparse graph. My PC isn't directly connected to Amazon or Netflix or really anyone of interest. But in a couple of dozen hops I'm connected to just about anyone in the world. Sparse doesn't mean unconnected.
The "badness" of the quadratic is that the "two hops" restriction SEEMS like a real restriction when in reality, it's not. It's bad in that the "two" in "two hops" sounds linear but the effects are quadratic. The quadratic-ness isn't inherently bad, I would never argue that. But the deception is.
Sweep(X) = 0.6 * X * 0.6 * X = 0.36 * X^2
I transposed the "percent of overlap" with "percent not overlapped" and viola problems.
If there's a 40% overlap from one person's social network to another that means that 60% of their social graph isn't included in the first person's and thus it's new for the sake of the two-hops strategy.
NumTargets(X,overlap) = 300mm / Sweep(X,overlap) = 300mm / ((1-overlap)*X)^2
NumTargets(100,99%) = 300mm
NumTargets(1000, 99%) = 3mm
NumTargets(100, 5%) = 33241
NumTargets(1000, 5%) = 332
So it's actually even worse then even I initially expected because the reach of the two hops method is doubly-quadratic. It's quadratic in terms of how much overlap there is from one person's social/business network to the next AND it's quadratic in how many connections each person has.
We're piling assumptions on top of assumptions here but lets suppose that at least 100k is the minimum number of targets we're comfortable with the NSA having to snoop on before they get everyone's data. It's not the right number but it's illustrative.
According to the spreadsheet I just put together we can satisfy this in a number of ways:
overlap = 99%, X just about all realistic numbers
overlap = 90%, X ~ 600 or less
overlap = 80%, X ~ 280 or less
overlap = 60%, X ~ 120 or less
overlap = 40%, X ~ 90 or less
overlap = 20%, X ~ 70 or less
overlap = 5%, X ~ 60 or less
There's no region that's got clearly reasonable assumptions where the NSA can't with a little bit of work get all the call records on all Americans again.
In other words, if it took overlap = 5% and X = 20,000 (to me not that realistic) in order to make it "easy" for the NSA to collect everyone's call records I might be less incensed. But the way this scales there's no reasonable assumptions that make it incredibly difficult for the NSA to vacuum up data on everyone.
That being said, it's generally believed to be (or measured) that the social network of the world is a sparse graph. So the fact that the sweep function is quadratic in the degree of a vertex is irrelevant. Using your notation, X is bounded by a constant for all intents and purposes.
In other words, there's nothing inherently bad about the quadratic behavior, and "knowing math" doesn't give you much except perhaps a false sense of intuition. The arithmetic (and your ad-hoc estimates) just works out in favor of your argument. Which is not all that surprising conditioned on the fact that you posted an argument :)