signal.normalize's BadCoefficients warning is too scary

[e-q]

It's easy to design digital filters with leading zeros and turn them into LTI objects. For example:

sys=sig.dlti(sig.firwin(101,0.1),1)

This raises the warning: BadCoefficients: Badly conditioned filter coefficients (numerator): the results may be meaningless

I feel this is somewhat overstated, since any FIR filter designed by firwin with a window that goes to zero at the ends (as many of the do) will have a leading zero, yet will do fine as a filter.

This can show up even just from calls to signal.decimate. One can get this warning from running:

 x=sig.decimate(np.random.randn(1000), 10, ftype='fir', n=200)

This has led colleagues of mine to believe that something inside the decimate function is fundamentally broken.

I propose we tone done the warning message, just letting the user know that some elements were trimmed. firwin could maybe also be edited to use something like win[1:-1], which would avoid raising a warning altogether.

[rgommers]

+1

[petermcculloch]

Ideally, a flag to prevent this behavior would be useful (allow_zero_padding?) in certain circumstances.

As an example, I was working on a project to analyze a classic reverb algorithm. Representing the building blocks of the reverb as dlti filters was straightforward until I got to a straight delay. The closest I could get was by placing a very small initial value in the transfer function's numerator.