Samplers accept invalid proposals

[chaozg]

Description

With a bounded prior, invalid proposals are accepted after a MH step. This happens to all samplers involving a MH step.

As seen in the following plot, an invalid proposal, negative here, is accepted after a MH step, which is not supposed to happen, and then the sampler gets stuck there.

Example to reproduce

import cuqi
beta = cuqi.distribution.Beta(0.5, 0.5)
simpler = cuqi.sampler.MALA(beta, x0=0.1, scale=0.1)
simples = simpler.sample(1000, 0)
plt.plot(simples.samples.T[:10])

DoD

• [x] Identity the reason behind:
  • [x] the validity of proposals is not checked at a MH step
• [x] solution agreed by team and issue merged
  • [x] validate proposal with isnan and isinf at MH steps in relevant samplers, i.e., MH, CWMH, MALA, NUTS
• [x] proper test functions are created

[chaozg]

@nabriis , I just want to draw your attention to this issue since you were not at the standup when I briefly reported it.

[chaozg]

Hi @jakobsj could you review the DoD here?

[jakobsj]

DoD approved!