Sometimes, we have a magic moment when just flipping the sign of a number makes a big difference,,,

IMPORTANT
Due to the change explained below, you may have to reduce the PRCL gain by a factor of 2 to keep the UGF of that loop.
Please remember to try this, if you have a difficulty in locking the interferometer (especially at the 3f PRMI stage).
I cannot test this now, because the microseismic is too large and we cannot lock the interferometer.
Also the PRCL feedforward gain may have to be readjusted.

What I did

Terrence found that the sign of an element in the EUL2OSEM matrix of PRM was fishy. 
As shown below, the top left element is 1, while other L coefficients are -1.
However, judging from the P and Y coefficients, the coils should have the same actuation polarity.
Therefore, this element was probably wrong.

So I flipped the sign of the L->H1 coefficient.

FIrst of all, the strange coherence between P and Y OpLev signals when we engaged the damping loop was reduced (oplevcoh.pdf).
Although there still is remaining coherence, the improvement is clear.

Next, I compared the spectra and RMS of the L, P, Y OpLev signals before and after the sign flip (oplevlen.pdf, oplevpit.pdf and oplevyaw.pdf).
The references in the plots are when the L-H1 coefficient was 1. 

You can see that there is a large improvement in the RMS of the mirror motion sensed by the OpLevs.
For L, the RMS reduction was a factor of 2 or more.
For P, a factor of 3 to 4 improvement.
For Y, almost a factor of 10 improvement.

Now the sign of the coefficient is corrected. However, this means the actuator efficiency of PRM in L will be increased roughly by a factor of 2.