Shillue asked:
There are some parts of the ALMA LO reference system that will have 
common mode phase drift.  That is all 64 antennas will have the same 
drift.  This is harmless except for VLBI mode.  But the system will have 
the same drift in either case so we have to design for VLBI mode.  So do 
you have any guidance for what the phase drift spec is for VLBI?  We 
specify phase drift as some few tens of femtoseconds, which is an 
appropriately small fraction of a wavelength at 950 GHz.  This is also 
specified as being over a time period of 1000 seconds.  

I responded:
I've discussed this with several people, including Mark Claussen and
Ed Fomalont.  The ALMA Scientific Requirements state that VLBI should
be achievable with ALMA on either a single element or as a phased array;
there have been no restrictions placed on the frequency.  
I think we should indeed assume that any antenna trying
to do high frequency VLBI will be at some site similar to Chajnantor.
Table 1 of Memo 382 lists some sites, of these only Hawaii is viable
above 350 GHz.  I would guess there would be a facility at the South Pole at
some point; perhaps nearby on higher Andean peaks such as Sairecabur.   
We'd envision phasing up the 64 ALMA atennas, which would
require using fast switching and WVR phase correction almost certainly
on baselines longer than the shortest ones.  One the VLBaselines, 
the beam would be ten microarcsec or so so there would most likely be
no unresolved calibrators; self-calibration would have to be used. 
One then shifts to the program source, possibly using the self-calibrated
phase from the calibrator along with an atmospheric phase correction
as a reference for the phase when imaging the unknown source.  

So the drift should not amount to more than a turn in some phase reference
period, as you mention.  A time period of 1000 second is probably too
long at 950 GHz; I doubt even exceptional atmospheric conditions would
allow that.  Ed thinks that one turn over 100 seconds is probably more
appropriate; Bill thinks this is the right timescale but probably one
tenth turn is more appropriate.

Larry noted that of course I had assumed a continuous well behaved
variation.  "What's really needed is a spec for the time scale on which the rms
variation is <<1 rad, since that's the maximum coherent integating
time beyond which you'll need to be doing self-cal.  My guess is that
this is 3 to 30 sec at 300 GHz, limited by the atmosphere.  Your LO
system should be enough better than this to ensure that the atmosphere
really is the limiting thing."

I agree with this.