The impedance, by the way, is given by desired AC regulation -- that is, output impedance.
So, if we're making a 3.3V 1A regulator and we want it within 10% for all loads (including full-step, or peak-amplitude sine (at any frequency), load currents), we need a variation of < 0.33V, or an output impedance under 0.33V/1A = 0.33Ω.
So we would choose a filter impedance of this order.
When the filter is poorly damped, there will be an impedance peak, probably in the transition band (i.e., around cutoff), which gives a peak resistance Q times higher than this -- thus we want a low Q, i.e., well damped filter. (There will probably be valleys corresponding to Zo / Q as well.)
As for the filter prototype, a singly-terminated type is often the way to go. If we have a typical power supply with modest size switching choke, and large filter caps, then the added filter can treat those large caps as shorted terminations. The branch facing the low impedance will always be a series inductor. Then we need to apply a resistive termination on the other end of the filter (which can be L or C depending on order; we'll probably choose even order, because capacitors are usually cheaper or smaller than inductors), and also ensure that our load impedance is never lower than that termination (else we'd short it out, and screw up the filter's response).
So with that covered --
Then to impedance match the output as seen by the filter - we need a RC notch-like filter - creating some 50Ohm impedance at 5kHz (and above). And with R chosen low enough to allow desired DC current.
Not a notch, though an RLC series resonant tank can be used in some cases (like, if you need a constant resistance filter:
http://jeroen.web.cern.ch/jeroen/reports/crfilter.pdf ).
Typically, R = Zo, and C >= 2.5 Co, where Co is the total capacitance used in the filter. This gives an RC cutoff a fair factor below the filter's cutoff frequency Fc, ensuring resistance is dominant in the transition band (near Fc).
If we didn't have to worry about DC, R alone would suffice, but of course we wouldn't want to burn the supply's entire load in one dumb termination resistor (or actually, quite a lot more than full load, given the earlier suggestion!), so we use a coupling capacitor to block DC to it.
And for the input? Also the same 50Ohm at the filter freq of 5kHz?
Ideally yes, but since there are one-port-terminated filter prototypes, we can get away with just the one.
(And just to flesh that out -- the capacitor would be on the order of C >= 2.5 / (2 pi F Zo) = 1.6uF. Which, for a 50 ohm filter, yeah, that sounds about right; for a real power supply, obviously just sub in whatever impedance is needed.)
Also, if we placed termination between the regulator's big filter cap, and the filter, we wouldn't be able to use a shunt R+C, it would be shorted out by the big cap. Instead we would need a series R||L.
Which, in case we have a load that may have very low impedance -- this is exactly what we would use at the output as well. Or if we can't make assumptions either way, use both! (R+C and R||L networks are often seen in audio amps, for this reason -- who knows what kind of whacky impedances a variously-resonant speaker might present, nevermind piezo tweeters!)
Tim