Introduction
I have become aware of a possible degradation in the Z-focus performance over the past year. It seems clearly a function of using the ambient temperature rather than the antenna leg temperature in order to correct the Z-focus. My likely response - to revert to using T_leg - is simple and uncontroversial, but I thought I'd solicit further opinion.

Thank you
Iain

I have just submitted a contribution to the paper going to the Board in which I state, in regard to the focus of the JCMT :

"Corrective algorithms with dependences upon elevation and temperature maintain focusses within 0.1mm rms in Z and 0.2mm in X & Y."

But something rang slightly untrue with the '0.1mm' bit. In fact the most recent analyses suggested that the Z-performance is not this good and is not as good as it used to be. The following table shows some characteristics of the SCUBA Z-axis data over the past 2 years (month-by month). Columns 3 & 4 shows the mean Z-focus (rather cosmetic in this context) and, of particular note here, the standard deviation of the data :

 ------------------------------------------------------------------
     1        2     3     4         5    6    7       8    9   10
                                   fit-vs-T_leg      fit-vs-T_amb
   date   No.obs.  mean  s.d.    slope sigma s.d.  slope sig  s.d.
 ------------------------------------------------------------------
 Jan 2002    56   -0.01  0.11    -0.009  2  0.10   -0.016  4  0.09
 Feb 2002    65    0.13  0.09     0.008  2  0.09   -0.004  1  0.09
 Mar 2002   123    0.05  0.12    -0.000  0  0.12   -0.014  4  0.12
 Apr 2002    63   -0.06  0.11    -0.005  1  0.11   -0.022  5  0.10
 May 2002    24   -0.06  0.11    -0.027  2  0.09   -0.052  7  0.06
 Jun 2002    37   -0.10  0.11     0.014  2  0.10   -0.006  1  0.11
 Jul 2002     0     -     -        -     -   -       -     -   -
 Aug 2002    50   -0.23  0.13     0.004  0  0.13    0.016  1  0.13
 Sep 2002    44   -0.24  0.12    -0.025  6  0.11   -0.048  5  0.10
 Oct 2002    96   -0.21  0.10     0.008  2  0.10   -0.004  1  0.10   
 ------------------------------------------------------------------
 Nov 2002     0     -     -        -     -   -       -     -   -
 Dec 2002a  122    0.05  0.16     0.016  4  0.15   -0.007  2  0.16
 Dec 2002b   71   -0.10  0.23     0.014  1  0.23   -0.039  7  0.18
 Jan 2003   103    0.02  0.15     0.017  5  0.13    0.002  0  0.14
 Feb 2003   151    0.02  0.16     0.017  4  0.16   -0.021  4  0.16
 Mar 2003   153    0.09  0.18     0.023  6  0.16   -0.006  1  0.18
 Apr 2003    95    0.09  0.17     0.025  4  0.14   -0.020  3  0.16
 May 2003   117    0.13  0.17     0.021  4  0.14   -0.016  2  0.17
 Jun 2003    98    0.13  0.18     0.023  3  0.17   -0.034  6  0.16
 Jul 2003   101    0.13  0.18     0.018  3  0.17   -0.039  6  0.16
 Aug 2002    86    0.17  0.18     0.021  3  0.17   -0.024  3  0.16
 Sep 2003a   57   -0.03  0.17     0.031  3  0.15   -0.034  3  0.16
 Sep 2003b   82    0.34  0.16     0.038  7  0.12   -0.010  1  0.16
 ------------------------------------------------------------------

* Dec 2002a & b cover dates Dec 01-22 & 23-31, resp. - (the latter
  data looked quite different, qualitatively, from the former).

* Sep 2003a & b cover dates Sep 01-13 & 13-30 resp.. - a step change
  is seen between the two subsets.

The s.d. of the data appear to have increased from ~0.11mm prior to Oct 2002 to ~0.16mm today :

(Data from the 3 years prior to Oct 2002 also support a steady rms value of ~0.11.)

The dividing line within the table marks the point (31 Oct 2002) at which the dependence of Z upon a 'temperature' was changed from T_leg to T_amb. Columns 5-7 and 8-10 show the results of fitting Z w.r.t. T_leg and T_amb. It is expected that one set of fits on each side of the line ought to be redundant, and , apart from in Sep 2002, the T_leg algorithm appeared convincingly useful prior to October 2002. However, the same can not be said for the T_amb algorithm afterwards, although no substantial and consistent improvement arises from fitting against either temperature - where 'substantial' means reproducing the ~0.11mm s.d. seen prior to Oct 2002.

The totalities of data on either side of the dividing line have been adjusted for the monthly mean zero-points and combined into datasets labelled '2002' and '2003' for convenience. Relevant dependencies are described below :

              '2002'  (N=558)
                                    s.d.         
     Z = ( 0.000 +- 0.001)*T_leg    0.11   range -6 to +12
     Z = (-0.008 +- 0.002)*T_amb    0.11         


              '2003'  (N=1042)     
                                    s.d.         
     Z = ( 0.016 +- 0.002)*T_leg    0.17   range -5 to +12
     Z = (-0.011 +- 0.002)*T_amb    0.18         

The T_leg algorithm used in '2002' was consistently successful; the T_amb algorithm used in '2003' a lot less so. At this point in my analysis (15 Oct 2003) it seemed that even accounting for the strong residual dependence of Z upon T_leg in '2003' would not reduce the s.d. to '2002' levels. Some other explanation still seemed necessary; something that may have occurred - perhaps coincidentally - at about the same time as the algorithm was changed. (So I looked for something).

SCUBA was not available between 04 Nov 2002 and 04 Dec 2002 - due to cryogenic problems - and there were substantial changes made to the cabin in the summer of 2002 installing the K-mirror for HARP, but it seemed prudent first to examine equivalent data from another FE :

                               RxA 
 ------------------------------------------------------------------
     1        2     3     4         5    6    7       8    9   10
                                   fit-vs-T_leg      fit-vs-T_amb
   date   No.obs.  mean  s.d.    slope sigma s.d.  slope sig  s.d.
 ------------------------------------------------------------------
  '2002'    314   -0.07  0.12    -0.008  4  0.12   -0.016  6  0.11
  '2003'    497    0.09  0.24     0.024  8  0.22   -0.014  4  0.23   

Again the performance in '2003' is noticeably worse than in '2002' as judged by the s.d., so the problem is 'telescope-wide'.

There was one other clue though, that, despite the rhetoric above, indicated that the T_amb-dependent algorithm is still the most likely suspect. Shown below are plots of the Z-focus -vs- (UT)hour of the day, for both '2002' and '2003' :

The '2003' data show distinct curvature indicating that the current T_amb algorithm does not provide a good 'temperature' for this purpose, especially at the times of day when the temperatures are changing quickly. Fitting a curve to the '2003' data reduces the s.d. to ~0.14mm : not as good yet as the '2002' value, but much better than the raw or fitted '2003' values.

This prompted Firmin Oliveira to suggest plotting the Z-focus residuals against the difference in temperature (T_leg - T_amb). The results (for '2002' and '2003') are dramatic :

Not only are the temperature differences surprisingly large (from about -10 to +5 degrees !) but the Z-error in 2003 correlates exceptionally strongly with the temperature 'error' (at the 30-sigma level) :

     Z = (0.058 +- 0.002 ) * (T_leg - T_amb)  -0.005    s.d.=0.12

The s.d. about the best fitting line is 0.12mm - essentially the '2002' value.
The use of T_amb is (surprisingly) inappropriate, and it seems essential to change the algorithm back to using T_leg and to collect data again. It should be obvious after about a month if the situation has improved.

The software was reverted/updated at 15:00 HST Thu 16 Oct 2003, with a starting value for DZT_COEFF of 0.70 (see mt_smudir:smu.ifl).