Bayesian Framework

Bayes' Theorem for Inverse Problems

\[p(\mathbf{x}|\mathbf{y}) = \frac{p(\mathbf{y}|\mathbf{x})p(\mathbf{x})}{p(\mathbf{y})} \propto p(\mathbf{y}|\mathbf{x})p(\mathbf{x})\]

Posterior ∝ Likelihood × Prior

σ² = 2.0
σₙ = 1.0 mGal
N = 50
--
MAP Estimate
--
Posterior Mean
--
Posterior Std
--
95% CI Width

Markov Chain Monte Carlo Sampling

Metropolis-Hastings Algorithm

Accept proposal \(\mathbf{x}'\) with probability \(\alpha = \min\left(1, \frac{p(\mathbf{x}'|\mathbf{y})q(\mathbf{x}|\mathbf{x}')}{p(\mathbf{x}|\mathbf{y})q(\mathbf{x}'|\mathbf{x})}\right)\)

σᵨ = 1.0
500 samples
Ready
0 samples
--
Accept Rate
--
ESS
--
R-hat
--
Sample Mean

Uncertainty Quantification

N = 60

Formal vs Calibrated Error

Formal errors from covariance matrix often underestimate true uncertainty. Calibration using external data or cross-validation provides more realistic estimates.

Spatial Resolution

Uncertainty varies spatially due to non-uniform data coverage. Areas with dense observations have lower uncertainty than data-sparse regions.

Degree-Dependent Error

Higher spherical harmonic degrees have larger uncertainties due to signal attenuation at satellite altitude and measurement noise amplification.

--
Mean Uncertainty
--
Max Uncertainty
--
Min Uncertainty
--
Reliable Degree

Bayesian Model Selection

Model Evidence (Marginal Likelihood)

\[p(\mathbf{y}|\mathcal{M}) = \int p(\mathbf{y}|\mathbf{x},\mathcal{M})p(\mathbf{x}|\mathcal{M})d\mathbf{x}\]

Bayes Factor: \(BF_{12} = \frac{p(\mathbf{y}|\mathcal{M}_1)}{p(\mathbf{y}|\mathcal{M}_2)}\)

Occam's Razor

Bayesian model selection naturally penalizes complexity. Models with more parameters must provide substantially better fit to overcome the Occam factor.

Evidence Interpretation

log BF > 1: Substantial evidence
log BF > 2.5: Strong evidence
log BF > 5: Decisive evidence

Cross-Validation

Leave-one-out cross-validation (LOO-CV) provides an alternative for model comparison when computing evidence is intractable.

--
Best Model
--
Log Evidence
--
Eff. Parameters
--
Posterior Prob.