tea leaf of estimation: why and how

[hyunjimoon]

1. concentrated
2. boundary hitting
3. a: point estimator vs b: sample-based estimator

• 1,2| 3a

• 1,2| 3b

• 1,~2| 3a

• 1,~2| 3b

cause: range [a,b] is hard prior for MCMC whereas entire space defining for optimization (subtractive vs additive)

1,3a: always

1,3b: blackhole in posterior space (too narrow variance; funnel)

2,3a:
2,3b:

hampering estimator's exploration. plotting joint posterior space and identifying wacky shape is actionable first step.

could introducing prior solve both 2|3a,2|3b? constrained to extend the boundary so that the estimation arrives as the global optimal wise score of optimal important because if there are many different parent values Sarah per use time series non-uniqueness would be less justifiable when different timeseries are projected from two different permitter values

[hyunjimoon]

BOUNDARY should be wider than prior

• aspire to not hit boundary for 3a, 3b
• there's no boundary for 3b, multiple boundaries at once.
• failure of nuclear power plant (no failure) - p = 0 (poisson distribution; probability can't