Hi,

I'm not sure if this exactly what you are looking for, but you may have a look at one of the papers about "Multiple imputation by chained equations" (MICE).

There is also a very good R package available for this.

Thomas

You could use the expected values of the latent state estimates of a state-space model fit to the observed data. In essence the unobserved "process noise" are the random variables, if I understand your question correctly. How difficult it would be to do this depends on the model and your data. Look up Kalman filtering/smoothing for one of the simpler examples.

Thanks for your useful answers, Thomas and Leo. I have read some paper evaluating Kalman filtering, but in my search I was looking for something similar to Krigging but applied to temporal series. I have also looking for papers evaluating ARIMA functions, but I couldn't find anyone related to interpolation trough time in trace gases, or evapotranspiration for example. As for  the MICE procedure, it is new to me and I'll explore it as deeply as I can.  

You could also consider Fuzzy models, as Sugeno Fuzzy Inference Systems, to asses how much they improve other results. I can do it for you if you give me a dataset, I mean, temporal series with gaps (or without them, but pointing where do you expect the gaps could occur). You can find me in Facebook or mail. In Spanish, haha.

Seguro, Gustavo, gracias. Ni bien termine con algunas otras cosas que tambien me ocupan ahora te mando un mensaje y te cuento como es el tema.