Dears, I checked with TeX/fs2000_corr_ME In fact the filter does the right thing, since the prior to measurement error correlation allows +/- 1, and we also get a warning: Warning: Your prior allows for correlations between measurement errors larger than +-1 and will not integrate to 1 due to truncation. Please change your prior When the multivariate filter stops, the correlation matrix implied by F is indeed [1 -1; -1 1] F is also not positive definite at that point, log(det(F)) would be a complex number, etc ... So, I think this behavior is correct and traps properly the issue around F. cheers M -- Marco Ratto, Financial and Economic Analysis Joint Research Centre European Commission, TP 361, 21027 ISPRA(VA), ITALY Tel: +39 0332 78 3794 Fax: +39 0332 78 5752, marco.ratto@jrc.ec.europa.eu http://www.macfinrobods.eu/ On 10/15/2015 8:30 AM, Marco Ratto wrote:
Then it might be that my modified tolerance make things worse [I thought to decrease singularity issues in multivariate filters ...]. I'll check as well Marco