I selected nine periods of ten minutes during last night's locks (see fig. 1). For each period I computed the transfer function from jitter (as measured by the diagnostic brteadboard Q1Y signal) to DARM. The second figure shows that the coupling indeed changes over time. 

I used the period martked as (4), which correspond to the largest coupling, fit a transfer function and used it in the time domain to perform jitter subtraction for all nine periods. The third figure shows the reduction in DARM noise: clearly the filter works well for period (4), but the performance is significantly degraded for the other periods.

In conclusion, if we want to subtract the jitter noise, we'll need some level of adaptive filtering.