PID output not zero

i'm currently trying to tune the servos of my milling machine. I'm having a bit of trouble with the pid output. there is continuosly a small error which causes the pid and analog output to never be 0. Changing P, bias or deadband doesn't help. Having a low I value doesn't help, having a larger I value causes the pid output to make a slanted "staircase" line.
I also had trouble with the servo hunting between two encoder counts (linear scales), but this was solved by adjusting the zero offset on the servo drive. The servo drive is an Indramat TDM.

What am I doing wrong?

I think its pretty difficult to get really precise PID
tuning when the position resolution is so low.

Normally you would expect to have at least 10 counts of
position error so you can get close to linear control in
the error band.

1um resolution isn't good enough?

But if the PID loop has an error, it should try to correct it, right? Or can it be that the servo drive has a small deadband and doesn't respond to the low analog output voltage? Should I try with a much higher I value?

Sorry I thought those were imperial units.

You can use a larger I term but note that if you
use any I term you will always dither between
2 encoder counts, with higher I terms speeding
the response time to small errors and the frequency
of the dither. Any dead band will just make the dither
range larger.

So one option is to use a large I term and live
with a micron or two of dither.

Okay, I won't be using deadband. Haven't had any luck with that, neither on my lathe retrofit. It only offsets the error and hunting.

Do you know why the 0.002 error isn't corrected by the P gain?

I had the same problem. I solved by changing the deadband term in pid.c file. With this modification has the deadband a real unit. If the position +/- deadband -> PID force to zero.  My first idea was to round the encoder feedback. In wood are 5µ not necessary.
But after 4 months i can say this modification works. The following error is near the same like without deadband and in stop position no vibration. For test you could exchange the file pid.so in folder /usr/lib/linuxcnc/modules/
I spoke here with Andy about this. He mean lincurve could also usefull to affect the PID. But i was happy with this solution an not searching more. 

I change this lines in file pid.c
oldnew

Thanks for your suggestion, but that doesn't really solve the problem I think. I just sets the PID output to 0 while there is still a small error. The PID loop should correct the error and then go to zero.

Just did some more testing and tuning. I got a pretty nice f-error graph on all 3 axis. Up to yesterday I had only used the x-axis. It seems that all 3 axis have the 0.002mm offset/error. I tried with some I and D values, but those didn't improve anything.

I don't think this small error will cause any problems during machining, but still it looks weird. If anyone has a suggestion what might be the cause, let me know.

If you have enough I term, the error will get corrected but you will have dithering.
Actually it will get corrected even with a small I term but it may take a long time...
(minutes to hours)