# Bug in Transformed Multi-Normal distribution?

Why is the Mathematica resulting in the weird output whenever the number of variables in the transformed distribution becomes greater than the number of variables in the original multivariate Normaldistribution?

This is correct:

TransformedDistribution[{y1 + y2, y1 + y3, y1 + y4,
y2 + y3}, {y1, y2, y3, y4} \[Distributed]
MultinormalDistribution[{\[Mu]1, \[Mu]2, \[Mu]3, \[Mu]4},
IdentityMatrix[4]]]


results in correct output

MultinormalDistribution[{\[Mu]1 + \[Mu]2, \[Mu]1 + \[Mu]3, \[Mu]1 + \
\[Mu]4, \[Mu]2 + \[Mu]3}, {{2, 1, 1, 1}, {1, 2, 1, 1}, {1, 1, 2,
0}, {1, 1, 0, 2}}]


But this

TransformedDistribution[{y1 + y2, y1 + y3, y1 + y4, y2 + y3, y1 + y2 + y3 + y4}, {y1, y2, y3, y4} \[Distributed]    MultinormalDistribution[{\[Mu]1, \[Mu]2, \[Mu]3, \[Mu]4},     IdentityMatrix[4]]]


results in weird output

TransformedDistribution[{\[FormalX]1 + \[FormalX]2, \[FormalX]1 + \
\[FormalX]3, \[FormalX]1 + \[FormalX]4, \[FormalX]2 + \[FormalX]3, \
\[FormalX]1 + \[FormalX]2 + \[FormalX]3 + \[FormalX]4}, {\[FormalX]1, \
\[FormalX]2, \[FormalX]3, \[FormalX]4} \[Distributed]
MultinormalDistribution[{\[Mu]1, \[Mu]2, \[Mu]3, \[Mu]4}, {{1, 0, 0,
0}, {0, 1, 0, 0}, {0, 0, 1, 0}, {0, 0, 0, 1}}]]


Is this some Mathematica bug or bug in my understanding?

thanks

• The second distribution is correct. Mathematica doesn't know a shorter or simpler representation of the distribution so it is represented as the input distribution with formal variables. If that distribution is called dist then you can evaluate Mean[dist] or StandardDeviation[dist] as you would with any distribution.. May 11, 2022 at 5:31
• PDF[TransformedDistribution[{y1+y2,y1+y3,y1+y4,y2+y3,y1+y2+y3+y4},{y1,y2,y3,y4}\[Distributed]MultinormalDistribution[{\[Mu]1,\[Mu]2,\[Mu]3,\[Mu]4},IdentityMatrix[4]]],{t1,t2,t3,t4,t5}] fails in 13 on Windows 10. Every command has its limitations. May 11, 2022 at 5:59

In short, your examples are of multivariate normals with covariance matrices that are not positive definite. (They are symmetric just not symmetric positive definite.)

As @BobHanlon mentions, Mean[dist] and StandardDeviation[dist] work fine.

Another command that works fine is Covariance:

dist = TransformedDistribution[{y1 + y2, y1 + y3, y1 + y4, y2 + y3, y1 + y2 + y3 + y4},
{y1, y2, y3, y4} \[Distributed] MultinormalDistribution[{μ1, μ2, μ3, μ4}, IdentityMatrix[4]]]

(cov = Covariance[dist]) // MatrixForm


But cov is singular:

Det[cov]
(* 0 *)


So my guess is that is probably related to the issue you're finding.

Suppose we try to construct the multivariate normal from knowing that covariance matrix:

MultinormalDistribution[cov]


This fails with the error message

Now suppose we construct a multivariate normal distribution from that covariance matrix (which doesn't know that it was constructed from 4 independent normals) and specify the mean:

dist2 = MultinormalDistribution[{0, 0, 0, 0, 0}, cov]


No error messages are given. But try Mean:

dist2 = MultinormalDistribution[{0, 0, 0, 0, 0}, cov];
Mean[dist2]


So I think it's all about not having a symmetric positive definite matrix in your examples.