Why not instead design the filter for the impedance it does see? You

suggested a pi network for coupling the load to the PA. Be aware that

you will have real trouble getting a lossless linear passive

reciprocal network to present, say, the desired 4000 ohm load to the

plate of a beam power tube and at the same time transform the 10k ohms

you see looking back at that plate to 50 ohms. It ain't going to

happen. If you add an L, or change the pi values, or whatever, to get

the 10k transformed to 50, then the 50 will present a 10k load to the

plate, rather far from the optimum. Same with Joerg's example of a

low-impedance output, say less than an ohm, and your 50 ohm load

transformed to the 12.5 ohm optimal load he suggested. Depending on

the network you use, the 1 ohm will transform to something different

from 50 ohms. A transformer would transform in constant impedance

ratio; a quarter wave transmission line would transform in a

reciprocal fashion: 25 ohm line to transform 12.5 to 50 would

transform 1 to 625. A pi behaves more like a quarter wave

transmission line than like a transformer, but not the same as a line

in general.

Three ways to solve your dilemma: (a) design whatever follows, be it

a filter or another stage or whatever, to operate properly with the

source impedance it WILL see; (b) use feedback in the amplifier to

adjust the source impedance to what you want (which gets at least

tricky at RF); and (c) add resistive (dissipative) loading. For

example, for (c) applied to the hypothetical valve mentioned above

with an optimum plate load of 4k ohms and an effective plate

resistance of 10k ohms, put 13.333k ohms shunt from plate to ground

(with DC blocking). Then design a pi network to match between 50 ohms

and 10k||13.333k. That means the 50 ohm load will "see" a 50 ohm

source impedance, and will present 10k||13.333k to the plate; but

there's also 13.333k plate load, so the net plate load will be 4k, as

desired. But that wastes 30% of the available power in the 13.333k

resistance. Much better to use (a) or (b) or a combination.

I thought I had posted a reply this afternoon about this, with a

reference to a nice article on reciprocal networks, but I see it

didn't make it. You might find the section beginning at "Properties

of reciprocal and non-reciprocal networks" at

http://www.microwaves101.com/encyclopedia/Network_theory.cfm to be

helpful in understanding why with "normal" networks you can't solve

your problem as stated in the basenote of this thread. You'll see

there that you may be able to solve your problem using a network with

nonisotropic material in it...