Bug in the evidence from merged runs
The evidence from merged runs is being over estimated due to this line.
To verify this, here is some example code:
In : from scipy.special import logsumexp In : a = np.array([11, 12, 13]) # Create fake Z's In : np.sqrt(np.mean(a**2)) # Actual RMS value calculated directly Out: 12.027745701779143 In : lna = np.log(a) # Take the natural log In : np.exp(0.5 * logsumexp(2 * lna, b=1. / len(a))) # This is how the RMS should be calculated Out: 12.027745701779143 In : np.exp(logsumexp(2 * lna, b=1. / len(a))) # This is how we do it currently Out: 144.66666666666666