FAQs

FAQs#

When should I use allow_multi_valued_likelihood?#

allow_multi_valued_likelihood is a flag that can be set to True to allow sampling with likelihood functions that are not deterministic, i.e. they do not return the same value when called twice with the same point.

By default, allow_multi_valued_likelihood is set to False. This means that nessai will raise an error if the likelihood function returns different values for the same point:

Repeated calls to the log-likelihood with the same parameters returned different values.

Setting this flag to True can be useful when using likelihood functions that include, for example, a Monte Carlo integration step. However, it is important to note that this can lead to incorrect results if the scale of the variance on the likelihood for a single sample is comparable to the width of the likelihood distribution for the posterior samples.

Warning

This flag should be used with caution and only when necessary. It can lead to very inefficient sampling and results that are entirely incorrect.

Why am I getting warnings about repeated log-likelihood values?#

I you see the following warning:

Initial live points contain repeated log-likelihood values!
This will likely lead to issues in the sampling process.

then it is likely that the log-likelihood function is returning the same value for multiple points. This is not allowed in the standard nested sampling algorithm and can lead to incorrect results.

One common cause of this error is the use of numpy.nan_to_num to convert -numpy.inf to a finite value. nessai is designed to handle -numpy.inf values and treat them as invalid regions of the parameter space. It is therefore not recommended to use numpy.nan_to_num in the log-likelihood function.

Large dynamic range#

If you see an error such as:

Rescaling is not invertible for x! This may be due to the large dynamic range (log10 dynamic range=18.0, min=1.0377468396342845e-09, max=950413414.495942).

This may an indication that the dynamic range of the parameter is too large for the default floating point precession (float64). In this case, it may be beneficial to apply a log transformation to the parameter before applying a reparameterisation such as scaleandshift or rescaletobounds.

This can be done using the reparameterisations keyword argument when calling the sampler. For example:

fs = FlowSampler(
    model=model,
    ...,
    reparameterisations={
        'log-standardise': ["x"],
    }
)

The list should contain the name of the parameter(s) which large dynamic ranges. This will apply a log transformation to the x parameter, before standardising the samples. See Configuring reparameterisations for more details.

Contents