Documentation request: What exactly happens inside arima_xgboost and prophet_xgboost?

[leonhGeis]

Recently, I am experimenting with the boosted arima function, which gives good results, compared to classical arima approaches. Nevertheless, I need to exclude this function from my model framework, as I do not exactly know, how the combination of both algorithms works.
How does the improvement take place? The source code implies a combination of arima predictions with xgboost predictions.