A description of the non_linear.ini config file is given here.

[MultiNest]

n_live_points

    The number of live points used by MultiNest to sample non-linear parameter space. More points provides a more
    thorough sampling of parameter space, at the expense of taking longer to run. The number of live points required
    for accurate sampling depends on the complexity of parameter space, below is a rough guideline of how many to use.

    for ~5 non-linear parameters -> 20-40 live points
    for ~10 non-linear parameters -> 30-50 live points
    for ~20 non-linear parameters -> 50-100 live points
    for 25+ non-linear parameters -> 75+ live points

sampling_efficiency

    The ratio of accepted to total samples MultiNest targets. A higher efficiency will converges on the high log_likelihood
    regions of parameter space faster at the risk of missing the global maxima solution.

    By default I would recommend a value of 0.8 (without constant efficiency mode) and 0.3 (with constant efficiency
    mode). Reduce to lower values if the inferred solution does not look accurate.

const_efficiency_mode

    The sampling efficiency determines the acceptance rate MultiNest targets. However, if MultiNest cannot map out
    parameter-space accurately it reduce the acceptance rate. Constant efficiency mode forces MultiNest to maintain
    the sampling efficiency acceptance rate. This can dramatically reduce run-times but increases the risk of missing
    the global maximum log likelihood solution.

    There are two circumstances where I find turning this mode on can lead to fast results without missing the global
    maximum solution:

    1) When the dimensionality of parameter space is low, e.g. < ~7 parameters.
    2) When the priors on the majority of model parameters are initialized using Gaussian priors centred on an accurate
       model from a previous search.

    In your configs I would recommend this mode is switched off, switching it on for specific searchs in custom pipeliines.

evidence_tolerance

    MultiNest will stop sampling when it estimates that continuing sampling will not increase the log evidence more than
    the evidence_tolerance value. Thus, the higher the evidence_tolerance the sooner MultiNest will stop running. Higher
    tolerances provide more accurate parameter errors.

    Pipelines include the log evidence tolerance as an input paramemter so that early searchs stop quickly, given that a
    precise estimate of parameter errors is not yet desired. Values of 100.0 / 1000.0 achieve this. For later searchs
    when error estimates are wanted use 0.5 or less.

multimodal

    Whether MultiNest uses multi-modal sampling, whereby the parameter space search will 'split' into multiple modes
    if it detects there are multiple peaks in log_likelihood space.

    For lens modeling there are rarely multiple solutions, thus this mode generally does not do anything. When multiple
    arise one typically has much higher log_likelihood values than the other, thus multimodal mode leads to a decrease in
    run time sampling both modes. Thus, I recommend you keep this setting off.

importance_nested_sampling

    Importance nested sampling mode uses information from the rejected points to improve the non-linear search. For
    lens modeling this appears to slow down the analysis with no clear benefit, so I recommend keeping it turned off.
