If the bitrate is 4Gb/s then a sample rate of >= 4 GSa/s will just be able to resolve it, given ideal thresholds and all. (No need to invoke brick wall filters or anything.) The fundamental frequency of alternating bits at 4Gb/s is 2GHz, so Nyquist is already taken care of. So to speak.
But talking about Nyquist in regards to a non-band-limited signal is rather perverse anyway.
Since we're talking information, the information-theoretic twist (nyTwist, you might say??!?...) is that, you can only observe *information* with a fundamental bandwidth <= half the sample rate. Which is a baud rate (i.e., symbol rate -- only the bit rate if the symbols are binary) equal to the sample rate.
Now, if you want to do signal quality analysis instead, you need to observe harmonics, which means a rated bandwidth well into the 10s of GHz.
A bandwidth as low as 1.5GHz might be enough to observe an eye diagram within spec, but it wouldn't be representative of the actual situation (preventing any conclusions regarding BER and susceptibility).
Tim