Hrm, no axes on the plot?

Glancing at the appnote, they show an R || C equivalent, floating; presumably there is also some common or normal mode impedance as well, which we should like to know. This is very much like an ADC's equivalent circuit, which you can also look at for ideas.

Treating it for now as a single port (just assuming that CM magically works out, so we need only concern ourselves with DM), we have a gain-bandwidth relationship:

BW = 1 / (2 pi (R || Rs) C)

Max gain: Rs = R

If we reduce Rs, we increase BW, but we must step down our signal source to a lower impedance, thus sacrificing gain -- proportionally. (If we increase Rs, BW does not increase much -- it's hard limited by R -- while gain drops.)

We can peak the capacitance, using some combination of filter networks, extending bandwidth modestly. The simplest case being a series inductor, having a value around X_L = R^2 C (the exact value depending on which type of filter we want; m-derived filters are very common here), and affecting a nice 30% or so extra bandwidth. More advanced cases use tapped ("tee") coils, and can use several additional poles (the returns for which diminish rapidly, so that we aren't very much concerned about 3+ pole types), pushing 100, even 200% additional bandwidth. Classic Tektronix scopes did this endlessly; a good reference is Starič and Margan, *Wideband Amplifiers* (excerpts of which the author has available).

Note we don't need to match Rs = R. If Rs > R, we simply lose gain (as reflected power), and bandwidth stays at its minimum. If Rs < R, bandwidth goes up, at the expense of gain.

If there's a source termination requirement (i.e., a ~constant 50 ohm or whatever input resistance), we need a termination resistor. If Rs = R, we're done; if Rs < R, simply terminate with the balance, Rt = Rs R / (R - Rs). (The exact location of the terminator can be anywhere along the peaking network, depending on what's being optimized: flatness of SWR, gain, phase, etc. More generally, we put in whatever (potentially lossy) equalizing networks are needed to achieve the desired result.)

Finally, to go from something unbalanced like 50 ohm BNC or SMA or whatever, to the differential input: use a balun transformer, preferably a very small one given the high frequency range. This probably won't extend very low, but an LF cutoff of some 10s of MHz would seem to be more than enough given the frequencies of interest, and should be no problem finding suitable parts. If LF to DC operation is desired, very fast fully-differential op-amps could be used.

Tim