Interplay of asymmetric errors and correlations

[TobiBoer]

Dear experts,

Im trying to add a new measurement, it contains multiple observables with asymmetric lower and upper errors as well as a correlation matrix.
When I call up the left/right errors of the measurement in the code I noticed the following:
The errors are NOT asymmetric anymore - both are now set on the average of the initial errors, the central value stays the same.
When I remove the correlation matrix, this does not happen and the errors become asymmetric.

So is it either having correlated measurements OR having asysmmetric errors?
Is there a proper way using flavio to handle this without spoiling the statistics?

[DavidMStraub]

That's correct, currently whenever you provide a correlation matrix, the PDF will be a multivariate normal distribution, see the implementation:

flavio/flavio/measurements.py

Lines 83 to 90 in 940d118

	for sym_err in error_dict['symmetric_errors']:
	squared_error += sym_err**2
	for asym_err in error_dict['asymmetric_errors']:
	squared_error += asym_err[0]*asym_err[1]
	errors.append(sqrt(squared_error))
	correlation = _fix_correlation_matrix(m_data['correlation'], len(observables))
	covariance = np.outer(np.asarray(errors), np.asarray(errors))*correlation
	m.add_constraint(observables, probability.MultivariateNormalDistribution(central_values, covariance))

How would you suggest to mathematically define a correlated asymmetric distribution? (I'm really asking out of ignorance 🙂)