bandpass filter

Well, yeah, an LC tank does that. What you REALLY want, though, is both zero shift AND zero derivative of phase shift (i.e. zero phase shift and local minimum or maximum phase at the center of the passband).

The curve doesn't have an extremum, just kisses zero-slope at one point. It's good, but there's lots of the passband, much less of the zero-phase part. And, it depends on matching components. If it were only resistors required for trim, like a state-variable filter, you'd be better served, by the usual custom-filter suppliers.

For instance, here's a nice acausal all-pass, both passive and NIC-based versions, compared with a causal one.

Looks superficially fine in an AC sim, but try running .tran 15 and see what happens. (The positive phase slope means that the poles of the transfer function are in the unstable half-plane.)

Cheers

Phil Hobbs

(Eternal-september lets me post attachments, but Supernews apparently filters them, so here's the same post without the attachment.)

For instance, here's a nice acausal all-pass, both passive and NIC-based versions, compared with a causal one.

Looks superficially fine in an AC sim, but try running .tran 15 and see what happens. (The positive phase slope means that the poles of the transfer function are in the unstable half-plane.)

Cheers

Phil Hobbs

It sounds like you've gotten an answer from Lasse in another reply.

I dunno if there's a name for it.

Whatever you actually do, I think the best you can hope for is a filter that doesn't change phase _much_, and that _is_ subject to component variations -- so it'll still be your task to make sure that all your errors due to component variations, temperature, and whatever stack up to something that's inside your tolerance band.

Le Wed, 07 Aug 2013 13:57:42 -0400, Phil Hobbs a écrit:

Hmmm, not much time to think more about it (going for a few days hiking), but IIRC I've seen that done... well on some limited bandwidth.

The trick was to Padé expand exp(tau.s) (the time machine) around your useful center frequency to 2nd, third or whatever order you want, then use that transfer function to synthesize your time machine filter. IIRC I toyed with that in spice and it worked reasonably well.

To JL, set tau to you BPF group propagation time and I guess you'll have something interesting (the math is left to you...)

It appears to in an AC analysis, but any transfer function with the wrong phase slope is going to be unstable, because the poles are in the wrong half plane. See e.g. the LTspice file I posted earlier today.

Have a great walk!

Cheers

Phil Hobbs

A network is causal if the output happens as a result of and *after* the input. All real boxes are causal. An ideal lowpass filter is not causal; its impulse response has voltages that happen before the input. It could be used to predict the future and game the stock market.

So you can't build an ideal lowpass filter. As you build better and better (sharper cutoff) lowpass filters, you inevitably add more and more time delay. So you get an impulse response that looks like the response of the ideal filter, but delayed. The delay prevents the causality catastrophe.

A lot of design concepts can be shot down quickly, without doing a lot of math, because they violate a basic conservation principle, like conservation of energy, or causality, or something like that.

Our customer has a master timing signal they they want us to tap off of at a number of sites. It would probably be 40 KHz, probably a square wave. Our boxes would be high impedance loop-thru taps on the cable, and we have to do fanout and PLL stuff to the signal to clock and trigger other time-coherent instruments. I want a bandpass on the front end to take out ground loops, wideband noise, cable reflections, whatever... just extract the fundamental. I'd like the filter to have near zero phase change for modest changes in his frequency or our temperature.

I really don't want to do this, but this is a good customer that we don't say "no" to. I got interested in minimizing the phase shift in the filter. If we stick in an ordinary bpf, we'd likely never get caught.

No. Whatever the phase shift is through the filter, I don't want it to change.

What have you just defined by the above statement?

You've defined a PLL with an Analog phase detector.

Depending on the type of your noise, you might want a LOW-Q BP in front of the PLL. ...Jim Thompson

I'm not dropping lines. I get your sales pitch on every post.

I don't consider a Q of 50 to be reasonable unless you plan on trimming. If the filter is active, you will probably have poor PSRR at that frequency.

I've done a lot of SCF back in the day and would always try to get the filtering to a minimum. People think more is more, but in reality, you start adding more stages, you get more noise, THD, etc. When I got to do my own Bell 212 "compatible" filter as part of a single chip modem, I totally undid all the f***ed up stuff Bell put in the spec, and was able to remove a third of the filter stages. [Not to mention tweaking the group delay to work with the standard phone line test filters, which Bell didn't do.]

Less is more if you expect the product to be manufacturable. If you aren't selling a few million, then maybe a Q of 50 is OK.

We will certainly have a PLL in our box, but I don't want to connect the phase detector to the raw cable signal. That would introduce all sorts of phase errors as a function of amplitude, waveform, cable reflections.

I want that filter to clean up the signal, extract the fundamental, and not contribute phase wander of its own.

He wants zero phase shift versus frequency.

Of course, that isn't possible.

But in this case, he doesn't want linear phase, but actually no phase shift. Now once he realizes that can't be done and states an acceptable margin, then this is worth discussion. But I assure you with a Q of 50, the filter will be touchy.

You do realize an all pass filter has phase shift. That is why you use them.

You typical bandpass filter will have a lot of delay at the edges. You insert all pass filters to add delay in the middle of the band. That is how you design a linear phase bandpass filters. But delay is d phi/ d t. Delay means phase shift.