tierms_phase deverr

[telegraphic]

Looking at the code for tierms_phase, it looks like the deverr output is always zero. I think line 464 should be

deverrs.append(tie / math.sqrt(ncount))

instead of

deverrs.append(dev / math.sqrt(ncount))

As the dev variable is never used. The new method would be

def tierms_phase(phase, rate, taus):
    """ TIE rms """
    rate = float(rate)
    (m, taus_used) = tau_m(phase, rate, taus)
    count = len(phase)
    devs = []
    deverrs = []
    ns = []
    for mj in m:
        ncount = 0
        tie = []
        for i in range(count - mj):
            phases = [phase[i], phase[i + mj]]  # pair of phases at distance mj from eachother
            tie.append(max(phases) - min(phases))  # phase error
            ncount += 1
        # RMS of tie vector
        tie = [pow(x, 2) for x in tie]  # square
        tie = numpy.mean(tie)  # mean
        tie = math.sqrt(tie)  # root
        devs.append(tie)
        deverrs.append(tie / math.sqrt(ncount))
        ns.append(ncount)

    return remove_small_ns(taus_used, devs, deverrs, ns)

[fmeynadier]

Looking into this long standing issue...

Code has changed a lot since this report, notably because of switch to numpy-style.

The corresponding line now reads (allantools.py, 1030):

 deverrs[idx] = 0 / np.sqrt(ncount) # TODO! I THINK THIS IS WRONG!

Logically I think it should be "tie / np.sqrt(ncount)", as suggested. However Stable32 does not issue any confidence interval for TIE rms stats, maybe there is a good reason for that: do you think it may be misleading to use this zeroth-order estimator ?

Here are the choices:

• Leave it as it is
• Put np.nan instead to make sure nobody uses deverrs for tierms
• put "tie / np.sqrt(ncount)"
• a more sophisticated approach

In any case I can provide a test using a short data sample + expected TIE rms results from stable32 (I think right now there is no test checking that the results are OK, just that the function runs and returns a result).