Generalize LocScaleReparam to matrix transformation

Hello,

I am very new to numpyro, and I was reading the tutorials to solve my specific problem.

In particular, in the eight-school example, the tutorial mentioned the LocScaleReparam to generate N(mu, sigma) from N(0, 1) * sigma + mu.

I was wondering how to construct such transformation when it is for a multivariate normal distribution. Suppose the parameter of interest is a 2-dim mu_vector with a multivariate normal prior.

There is a linear relationship between the likelihood function and mu_vector such that

X ~ MVN(T*mu_vector, Sigma)

where T is a known matrix of size qx2, q>2.

Do I need to apply numpyro.handlers.reparam to reconstruct the random variable?

  with numpyro.handlers.reparam(config={'theta': TransformReparam()}):
        theta = numpyro.sample(
            'theta',
            dist.TransformedDistribution(dist.Normal(0., 1.),
                                         dist.transforms.AffineTransform(mu, tau))
        )
        numpyro.sample('obs', dist.Normal(theta, sigma), obs=y)

I think you can use the following LowerCholeskyAffine for this one.

1 Like