Designing LC filters for amplifiers?

"A passive filter cannot shape response and also provide impedance matching."

The line above is the one I question. Clearly, you can impedance match with passive filters. The effect of source and load impedance on out of band performance is a different question.

I have been corrected. I was not thinking in terms of filters that dump unwanted energy into dummy loads, but minimal loss sort of filters (i.e. simple resonators). I was referring to the fact that simple resonators do not maintain a constant (and therefore source or load matching) impedance as the frequency swings from passband to stop band. Evidently, filter designers are a lot more sophisticated than I am.

Sometimes you have to assert something dumb to learn better.

As you note, non-dissipative filters have poor out-of-band return loss, but filter designers also have control over the phase of that return loss. That is, it may be a bad thing to present the driving source with a short-circuit out-of-band, so you design the filter to go to an open; or it could be just the opposite and bad news if you let the filter go to an open out of band, so you design it to go to a short.

There is also a place for "singly-terminated" filters, designed to operate with a source that's a high or low source impedance, but terminate into, e.g., 50 ohms, or vice-versa.

Indeed.

Cheers, Tom

Thanks for all the responses! I am a bit more confused now, however, because I had always thought the exact same thing as John P. did: i.e., that passive LC filters *cannot* match two different impedance's; they can only be designed to function properly between two different impedances, if so desired. Are we now saying that that concept is completely, totally dead wrong when designing pure (no resistive element) LC filters??

Thanks again,

-Bill

as

Sure, passive parts match impedances all the time. Consider simply adding a transformer to a filter. Consider a bandpass filter composed of parallel tuned tanks which are coupled, either capacitively or magnetically; the tank at either end may be tapped inductively or capacitively to provide an impedance match -- that works with a single tank as well as with a multi-tank filter. Consider that a typical tuned transmitter output network (e.g. PI network) is used primarily for impedance matching, but is also a filter.

What you cannot do with non-dissipative passive components (inductances and capacitances) is match impedances at DC. If that's not obvious, consider that capacitances are open circuits at DC, and inductances are short circuits. So low-pass filters can't be used to match impedances, at least not at the low frequency end, though like a typical PI network, they can effect a match at a specific frequency or over some range. Similarly for a high-pass filter. You'll see impedance matching included most commonly in band-pass filters. You'll also find filters that have internal impedance transformations to allow use of more practical component values, even though they're designed to be equally terminated at each end. If you read the filter design notes included in the help that comes with ELSIE and with the AADE Filter Designer you'll find lots more about these things.

You might stumble into a bandpass filter when you start by wanting to do an impedance matching job, in fact. For example, if you use a PI network to match between 5000 ohms and 50 ohms, it won't have a very broad frequency response. But if you cascade a set of L networks designed to do 5000:1581, 1581:500, 500:158, and 158:50, you'll discover that you have made a bandpass filter.

Cheers, Tom

Bill,

A no resitive can only Match and provide a filtering finction between two differint load and sources ove a defined frequency range, impedance range, closeness of the "match'. but make noe mistake when you are operatig in band, a 50 ohm source can be mathced to say a 200 (or simialairly 12.5 ohm load).. now once you have two differning imepdance te maount and extent of the math is a fixed product, lke he gain bandwdith product, you get a matching bnandwidth impedance again product, so the more the mis match the narrower the bandwidth.., etc..

Marc Popek

PS if you can operate a smith chart and you do a few (thousand :-) ) mathcng circles, wit the constant Q" overlay, you instantly get a feel for the the constant between mismatch and bandwdith.

My job for the day is to tune a lowpass filter for a 32 MHz DDS waveform synthesizer. Ideally it will be flat to 32 MHz, supress images well (32 is half of Nyquist), fix the sinx/x rolloff, fix any residual amplifier rolloffs, still have a decent step response, be made from available parts, and not take the rest of the week to do.

Wish me luck.

John

I think what needs to be "investigated" is the sensitivity to source/ load impedance when out of band. I recall when designing a duplexer that zeros in the stop band helped greatly when the two band split filters were connected to each other.

My disclaimer on this topic is 99% of the LC filter design I have done was to create leapfrog, i.e. active filters. Making real as in physical LC filters is quite different.

Is it a homework question?

Vladimir Vassilevsky

DSP and Mixed Signal Design Consultant

But if you add more filter/matching sections, you can adjust the bandwidth for a given impedance matching ratio. For example, a simple L section to match between 5k ohms and 50 ohms (as in my other reply to Bill...) gives you about 20% 3dB bandwidth. If you use two sections designed for 5k to 500 and 500 to 50, you get about 45% 3dB bandwidth (and a much flatter "top"). Depending on the coil (and capacitor) unloaded Q, you may find that you get lower loss with more parts as well, since each of two sections operates at a much lower loaded Q than the single section to do the same job. (This topic comes up occasionally over on rec.radio.amateur.antenna.) More sections can give you additional bandwidth, if that's what you need. Viewed as a filter, more sections can give you flatter band pass and steeper cutoff, while still providing an impedance transformation as well.

Cheers, Tom

No, we're really designing a 32 MHz 8-channel DDS arb.

It's a 32 MHz version of this:

I started with a textbook 0.5 dB Cheby lowpass, 34 MHz roughly so that

32 would be pretty much flat. Poked in standard values of Ls and Cs and simulated/tweaked in LT Spice until it looked OK.

Now I've got the parts soldered down in a real DDS channel and I'm plotting frequency response. It droops towards the high end, probably sinc plus output amps and a bit of dac rolloff, but tweaking the drive impedance seems to perk it up nicely. The trick is to get a good compromise value, averaging out the various lumps and bumps.

The Cheby rings a little for step outputs, but that's life. If I use a Gaussian, or even a transitional Gaussian, the rolloff will be rotten and the images will be really bad. Clock is 128 MHz, so the bad image is at 96.

But tuning a surfmount filter is truly tedious. Change a part: reload firmware and FPGA from background debugger: start it up. Poke a dozen frequencies or so: plot: repeat.

There's a place for one extra cap, across an inductor, for an elliptical notch. Maybe I'll try that too.

John

I see. Although it is possible to do the analog filter like that, I would approach this task little differently: make an analog filter as simple and as low Q as it is possible to have just enough of suppression of the aliases. This will make the filter relatively insensitive to the tolerance of the components, stray inductances and capacitances. Hopefully this will avoid any need of tuning in production. A signal would be corrected in the digital domain. This will require some extra dynamic range from the DAC. I think somewhat 1..2dB would be enough for the worst case, and that doesn't really make much of difference.

Vladimir Vassilevsky DSP and Mixed Signal Consultant

It's a 5th order filter because that's what it takes to supress the aliases! With a 32 MHz signal, the bad alias is at 96 MHz, only a factor of 3:1.

We certainly won't tune these filters in production. With 5% Ls and 2% C's, that shouldn't be necessary. The tuning is a one-time, engineering problem.

I sure wish these 0805 parts came with tiny screwdriver adjustments on top.

Use digital filters to fix the saggies of the analog filters? That would take scores of multipliers, and we're just about out of multipliers.

John

approach

Q

Chebyshev 5th has rather high Q poles, which makes it darn sensitive. Are you sure it will be enough consistency so it won't be a need to adjust anything in production, temperature range, etc?

really

Yes, fix the passband flatness and the pulse response by a digital filter.

The filtering has to be done only once, when preparing the waveform and storing it to memory. This doesn't have to be in a real time, so a CPU can do that.

Vladimir Vassilevsky DSP and Mixed Signal Consultant

Not sure yet. I'll know after some more work. And the final board specs can be tweaked, within reason. One never really knows how the fine print will affect sales.

I can evaluate the sensitivity in simulation, even though I can't really evaluate the frequency response! At least to the flatness I'd like.

I don't think that would work here. Once a waveform is loaded, the rate that it's played is programmable - it *is* a DDS synthesizer - so the filtering would have to be change as the rate changes. That can't be pre-loaded.

I do appreciate the ideas.

John

If you're using ferrite inductors, you can tune them with a magnet.

Cheers,

Phil Hobbs

Or a dremel.

John

Yeah, but when you take the magnet away, they basically recover.

Cheers,

Phil Hobbs