Why Marginalization Improves Inverse TEXAS Sampling#

This note explains why the inverse TEXAS models marginalize over the forward-calibration ensemble inside the likelihood (manuscript Appendix B, Eqs. B2–B3) instead of sampling one temperature per observation per calibration draw — the “full latent” formulation that was tried first and is now archived. The marginal form runs much faster and gives cleaner posteriors. The goal is to make this clear for a scientific audience, without assuming a background in Bayesian computation.


1. The Two Approaches#

(A) The “full latent” formulation (archived)#

  • For each data point n (e.g., a core-top sample), and each forward-model draw m (the ensemble), we introduced an explicit latent parameter:

    t_est[n, m]
    
  • That means:

    • Number of unknowns grows with both the number of samples (N) and the ensemble size (M).

    • For N ≈ 1,200 and M ≈ 100, that’s ~120,000 latent parameters.

  • The model then combines all of these latents into the likelihood.


(B) The marginalized formulation (shipped)#

  • Instead of sampling t_est[n, m] for every ensemble draw, we only keep a single latent parameter per sample:

    t_est[n]
    
  • We then compute the likelihood by marginalizing over the ensemble:

    log p(data[n] | t_est[n]) = log( 1/M * Σ p(data[n] | t_est[n], params_m) )
    
  • This uses the log_sum_exp trick in Stan, which is stable and efficient.

  • Result:

    • Only N ≈ 1,200 latent parameters (independent of M).

    • The ensemble variability is still accounted for, but without exploding the parameter space.


2. Side-by-Side Stan Sketch#

Here’s the essential difference, shown in simplified pseudo-Stan code:

Full Latent Version#

parameters {
  matrix[N, M] t_est;   // each sample × ensemble
}

model {
  for (m in 1:M) {
    vector[N] mu_scaledRI = logistic_fn(t_est[, m], params[m]);
    scaledRI ~ normal(mu_scaledRI, sigma[m]);
  }
}

Marginalized Version#

parameters {
  vector[N] t_est;      // only one per sample
}

model {
  for (n in 1:N) {
    vector[M] log_lik_m;
    for (m in 1:M) {
      real mu_scaledRI = logistic_fn(t_est[n], params[m]);
      log_lik_m[m] = normal_lpdf(scaledRI[n] | mu_scaledRI, sigma[m]);
    }
    // Marginalize using log-sum-exp
    target += log_sum_exp(log_lik_m) - log(M);
  }
}

Key change: Instead of sampling t_est[n, m] for all ensemble members, we sample only t_est[n] and analytically average over ensemble members.


3. Why Marginalization is Better#

3.1 Fewer Parameters#

  • Full latent model: ~120,000 parameters.

  • Marginal model: ~1,200 parameters.

  • Sampling speed scales badly with dimensionality. By cutting the number of unknowns 100-fold, the sampler spends far less time per iteration.

3.2 Better Posterior Geometry#

  • In the full latent model, t_est[n, m] are strongly correlated across m.

  • The sampler has to “wiggle” through a very tangled space.

  • Marginalization collapses these correlations analytically, so the sampler only needs to explore the temperature dimension.

3.3 Cleaner Effective Sample Size (ESS)#

  • ESS measures how many “independent” samples you get out of MCMC.

  • With fewer parameters and smoother geometry, marginal models give higher ESS per unit time.

  • This means your credible intervals stabilize faster.


4. Practical Impact#

  • Speed: Warm-up and sampling progress much faster, because each iteration moves N parameters instead of N × M.

  • Stability: Fewer divergences, better adaptation of step sizes, and cleaner R̂ diagnostics.

  • Scalability: Larger ensembles cost only likelihood evaluations, not parameters. The manuscript uses M = 300 for the global coretop reconstructions and M = 500 (the package default for a 4,000-draw posterior) for the downcore ones.


5. Analogy (for non-statisticians)#

Think of this like averaging weather forecasts:

  • Full latent model: You ask 100 forecasters for 1,200 cities, and try to track every single opinion separately. That’s 120,000 numbers bouncing around.

  • Marginalized model: You only track the city-level temperature (1,200 numbers), and average the forecasts mathematically in the likelihood. You still capture uncertainty from forecasters, but you don’t waste time simulating each one.


6. Key Takeaway#

Marginalization trades a slightly more complex likelihood calculation for a massive reduction in sampling complexity.

  • The science is the same (we still respect ensemble variability).

  • But computationally, the marginalized model is leaner, faster, and more robust.

This is why the three inverse Stan models shipped with TEXAS are all marginal, and why predict_T_from_proxyObs has no other path.

7. The deterministic point inverse, and when not to use it#

TEXAS.inverse_generalized_logistic_fixed_upper inverts the forward curve for a single parameter set. It is not a reconstruction.

For the published T₀-shift calibration the nonthermal predictors shift the curve’s location, so the point inverse is the thermal inverse evaluated at the shifted location:

T0_eff = t0 + gamma_G23·gdgt23ratio + gamma_NO3·log10(no3)   [NO₃ term only where 0 < no3 < no3_cutoff]
x      = T0_eff − ln(((1 − b) / (y − b))**v − 1) / k

i.e. call inverse_generalized_logistic_fixed_upper(y, t0=T0_eff, k=k, b=b, v=v) with the sample’s own predictor values folded into t0. (The companion inverse_generalized_logistic_fixed_upper_multivariate is the point inverse of the additive formulation of the initial submission — it subtracts β_G23·gdgt23ratio + β_NO3·log10(no3) from the response first — and is kept for the archived GHEA posteriors.) Predictors are inputs to be supplied, not quantities solved for: pass the same covariate values that applied to those observations.

The thermal inverse is defined only for b < y < 1. Values outside that range are physically unreachable for the given parameters and come back as np.nan, with a single RuntimeWarning if any occur. That NaN is the honest answer, and it is the first reason this function is not a reconstruction: the Bayesian path returns a posterior in exactly that region, driven by the prior, which is informative in a way a NaN is not.

The second reason is uncertainty. Feeding this function the posterior means of t0, k, b and v gives one number per sample and no interval — and the plug-in estimate is not the median of the marginal posterior, because the inverse is non-linear. For an uncertainty-aware reconstruction that marginalises over the full forward posterior, use TEXAS.predict_T_from_proxyObs, which is the Stan path this page describes.

Use the point inverse for what it is good at: drawing a single calibration curve on a figure, checking an algebraic identity against the forward function, or sanity-checking one parameter set by hand.