Reports until 18:57, Saturday 02 May 2015
H1 CAL (DetChar, INJ, ISC)
jeffrey.kissel@LIGO.ORG - posted 18:57, Saturday 02 May 2015 - last comment - 16:13, Monday 04 May 2015(18186)
DARM Coupled Cavity Pole is now at 355 [Hz], ESD Pole is actually 2.2 [kHz], OLGTF Model Frequency Dependence Matches Data from 10 [Hz] - 2 [kHz] to within 2.5% and 1 [deg], but CAL-CS & GDS/DMT/LAL h(t) Pipelines Does Not
J. Kissel

Having processed the last two DARM open loop gains -- the first measurements out to 5 [kHz], and the first measurements after we've tuned up the ASC loops to give us a consistent recycling gain of 35-40 -- I can now make the following statements:
- The DARM Coupled Cavity Pole (CCP) frequency is now consistently much closer to the "as designed" value. The measurements are consistent with a model of the DARM loop using a CCP of 355 [Hz].
- The ETMY ESD's driver pole frequency is 2.2 [kHz], not the colloquially thrown about 2 [kHz].
- I've added the violin modes to the model of the QUAD suspension, and they have no appreciable effect, but I'll leave them in for completeness.

On what these measurements and changes to the model mean for the uncertainty (precision and accuracy):
- The modeled unknown time delay remains at 0 +/- 5 [us].
- The frequency dependent uncertainty in the calibration model remains at +/- 2.5% in magnitude and 1 [deg], but I've data to back up that this extends out to the entire required* frequency band 10 and 2000 [Hz]. See attachment 2015-05-02_FittedCCP_H1DARMOLGTF.pdf (* OK, what *used* to be the requirement. The requirements are now extended albiet inflated out to 5 [kHz]; see T1300950)
- Over these last two measurements, the scale factor used to match the model against measurement has only varied by 3%. Indeed, if you cluster the eight measurements, grouping by CCP, then the standard deviations of the three, three, and two measurements are 7%, 41%, and 3%. However, the total standard deviation of all measurements is 22%, and the latter two measurements are only 1 day apart. From the data I have, I'm still not confident in decreasing the scale-factor uncertainty below 22%; perhaps PCAL can make a better statement on this (but I fear the current Mini-Run noise is too high for the low-frequency PCAL line). See 
2015-05-02_upto5kHzOnly_FittedCCP_H1DARMOLGTF.pdf
- The current calibration installed in the CAL-CS model is incorrect, both in frequency dependence and scale factor. 
    - Frequency dependence: The CAL-CS filters assume a DARM CCP of 389 [Hz] in the sensing path, and an ESD Driver Pole (ESDP) of 2.2 [kHz] in the actuation path. For the latest lock stretches, that inflates the uncertainty of CAL-DELTAL_EXTERNAL_DQ with a known, systematic error of 
      current / correct = ( 1 / (current CCP / correct CCP) + (current ESDP / correct ESDP) ) / (properly normalized)
                        = ( 1/zpk(-2*pi*389,-2*pi*355,355/389) + zpk(-2*pi*2e3,-2*pi*2.2e3,2.2e3/2e3) ) / 2
      which translates to as much as a 7% magnitude error at 1 [kHz], and 1.8 [deg] swing in phase surrounding 1 [kHz].
      This is demonstrated in 2015-05-02_H1CAL-CS_Systematic.pdf,
      This is also reflected in the statistical uncertainty estimate. All comparisons are shown if a 389 Hz CCP and 2 kHz ESDP were used in 2015-05-02_389HzCCP_2p0kHzESDPole_H1DARMOLGTF.pdf.
    - Scale Factor: this depends on whether we want to use 22% or 3% for our scale factor uncertainty. If 22%, then the last two lock stretches still fall within the 1.1e6 +/ 22% [ct/m]. But, if the scale factor is 1.2725e6 (the mean of the last two scale factors), then that indeed falls outside of 1.1e6 +/-3%
    - We *still* have not implemented any compensation for the analog or digital, AA or AI filters. 
- The GDS/DMT produced calibration is off just as much as the CAL-CS produced calibration, because GDS/DMT calibration is using the same filters.

------------------
All parameter files have been committed (with updated ESD driver poles, and fitted DARM coupled cavity poles) here:
/ligo/svncommon/CalSVN/aligocalibration/trunk/Runs/PreER7/H1/Scripts/
H1DARMparams_1109994128.m
H1DARMparams_1111998876.m
H1DARMparams_1112399129.m
H1DARMparams_1112933759.m
H1DARMparams_1112942996.m
H1DARMparams_1113119652.m
H1DARMparams_1114541595.m
H1DARMparams_1114634170.m

Which are processed by the DARM model function here:
/ligo/svncommon/CalSVN/aligocalibration/trunk/Runs/PreER7/H1/Scripts/H1DARMmodel_preER7.m

And then compared with the script here:
/ligo/svncommon/CalSVN/aligocalibration/trunk/Runs/PreER7/H1/Scripts/CompareDARMOLGTFs.m
Non-image files attached to this report
Comments related to this report
jeffrey.kissel@LIGO.ORG - 16:13, Monday 04 May 2015 (18213)DetChar, ISC
J. Kissel

At the requested of Gabriele, I've isolated/highlighted the change in measurements' cavity pole frequency in the attachment below. In it, I
(pg 1) have divided the open loop gain transfer function *measurement* by all components of the *model* for each measurement *except* for the frequency response of the IFO (which we have modeled as a single pole filter at the designated frequency). That includes
     - The entire actuation function, A
     - The entire control filter and gain, D
     - The sensing function's non-ifo frequency dependence from the AA (both digital and analog) filters and the uncompensated OMC DCPD whitening poles
     - The DC optical gain
(pg 2) Remind people that I've filtered out every data point but the most ridiculously coherent (coh = 0.99, nAvgs = 20), such that we can be certain that all frequency points used have sqrt( (1 - 0.99) / (2 * 20 * 0.99) ) = 1.6% uncertainty. Indeed, as shown by the historgram, and the vast majority have greater than 0.999 coherence, i.e. sqrt( (1 - 0.999) / (2 * 20 * 0.999) ) = 0.5% uncertainty.
Non-image files attached to this comment