Hierarchical Emergent Constraints (HEC)

• Hierarchical Emergent Constraints (HEC) is a statistical framework that combines model and observational uncertainties to constrain climate projections via emergent relationships.
• The framework employs a hierarchical Bayesian model with closed-form Gaussian solutions to compute posterior means and variances for future climate states.
• Applications, such as the snow–albedo feedback, demonstrate how high signal-to-noise ratios and strong inter-state correlations can significantly reduce prediction uncertainty.

Hierarchical Emergent Constraints (HEC) provide a statistical framework for constraining projections of future climate variables by leveraging emergent relationships between current and future climate states, while formally accounting for both observational and model-based uncertainties. HEC extends traditional emergent constraint (EC) analysis by embedding it within a hierarchical Bayesian model that explicitly incorporates the error structure of observations and model ensembles, allowing for rigorous quantification and propagation of uncertainty. Closed-form solutions for the posterior mean and variance of future climate states can be obtained under Gaussian assumptions, with the result depending critically on both the signal-to-noise ratio (SNR) between model spread and observation error, and the correlation coefficient ρ between current and future climate states in the model ensemble (Bowman et al., 2018).

1. Hierarchical Statistical Model in HEC

HEC treats the current climate state ($x_t$), its observation ($y_t$), and the future climate state ($z_{t+\tau}$) as random variables with explicit statistical dependencies. The joint distribution is constructed such that, given the actual (latent) current state, future and observed states are conditionally independent:

$$[z_{t+\tau}, x_t | y_t] \propto [z_{t+\tau}|x_t]\,[x_t|y_t]$$

Inference proceeds via Bayes’ theorem as:

$[x_t|y_t] \propto [y_t|x_t][x_t]$

Marginalizing over $x_t$ yields the central HEC estimator:

$$[z_{t+\tau}|y_t]=\int [z_{t+\tau}|x_t]\,[x_t|y_t]\,dx_t$$

This hierarchical construction unifies model-driven structural uncertainty, observational uncertainty, and the emergent constraint in a single statistical model.

2. Gaussian Formulation and Closed-form Solutions

Assuming Gaussian statistics and additive errors, the observed variable is modeled as $y_t = x_t + n_t$, where $x_t \sim \mathcal{N}(\mu_x, \sigma_x^2)$, $n_t \sim \mathcal{N}(0, \sigma_n^2)$, and $y_t$0. The signal-to-noise ratio is defined as:

$y_t$1

Assuming $y_t$2 jointly Gaussian with correlation ρ, the key closed-form expressions are:

• Posterior of $y_t$3 given $y_t$4 (assimilating data):

$y_t$5

$y_t$6

with $y_t$7.

• Regression of future on current:

$y_t$8

$y_t$9

where $z_{t+\tau}$0.

• HEC posterior for $z_{t+\tau}$1 given $z_{t+\tau}$2:

$z_{t+\tau}$3

$z_{t+\tau}$4

with $z_{t+\tau}$5.

Normalized-anomaly forms relate posterior predictions of future state anomalies to observed anomalies in the current state via:

$z_{t+\tau}$6

and the fractional variance reduction is

$z_{t+\tau}$7

These formulae rigorously characterize how constraint strength depends on both the observational quality (SNR) and the inter-state correlation (ρ).

3. Interpretation of Parameters and Uncertainty Quantification

Parameter definitions within the HEC framework are as follows:

Symbol	Interpretation	Uncertainty Type
$z_{t+\tau}$8	Mean/variance of ensemble current-climate state ($z_{t+\tau}$9)	Model/structural
$$[z_{t+\tau}, x_t | y_t] \propto [z_{t+\tau}|x_t]\,[x_t|y_t]$$0	Mean/variance of future state ($$[z_{t+\tau}, x_t | y_t] \propto [z_{t+\tau}|x_t]\,[x_t|y_t]$$1)	Model/structural
$$[z_{t+\tau}, x_t | y_t] \propto [z_{t+\tau}|x_t]\,[x_t|y_t]$$2	Variance of data–model error (observational errors)	Observational
$$[z_{t+\tau}, x_t | y_t] \propto [z_{t+\tau}|x_t]\,[x_t|y_t]$$3	Ratio $$[z_{t+\tau}, x_t | y_t] \propto [z_{t+\tau}|x_t]\,[x_t|y_t]$$4	Quality of observation
$$[z_{t+\tau}, x_t | y_t] \propto [z_{t+\tau}|x_t]\,[x_t|y_t]$$5	Pearson correlation, $$[z_{t+\tau}, x_t | y_t] \propto [z_{t+\tau}|x_t]\,[x_t|y_t]$$6	Emergent relationship
$$[z_{t+\tau}, x_t | y_t] \propto [z_{t+\tau}|x_t]\,[x_t|y_t]$$7	Weighting of data in state estimation	Data assimilation

The uncertainty in $$[z_{t+\tau}, x_t | y_t] \propto [z_{t+\tau}|x_t]\,[x_t|y_t]$$8 after incorporating both model and observational sources is strictly less than the prior model spread when both $$[z_{t+\tau}, x_t | y_t] \propto [z_{t+\tau}|x_t]\,[x_t|y_t]$$9 and SNR are large. If either is small, the posterior variance is dominated by $[x_t|y_t] \propto [y_t|x_t][x_t]$0, the unconstrained model uncertainty.

4. Practical Application: Snow-Albedo Feedback

HEC has been explicitly applied to the snow–albedo feedback (SAF) in Northern Hemisphere land areas. In the context of Bowman et al. (2018), the variables are instantiated as:

• $[x_t|y_t] \propto [y_t|x_t][x_t]$1 SCSAT (seasonal-cycle snow–albedo temperature sensitivity)
• $[x_t|y_t] \propto [y_t|x_t][x_t]$2 CCSAT (climate-change snow–albedo temperature sensitivity)

Using CMIP5 ensemble data ($[x_t|y_t] \propto [y_t|x_t][x_t]$3):

• $[x_t|y_t] \propto [y_t|x_t][x_t]$4 % K⁻¹, $[x_t|y_t] \propto [y_t|x_t][x_t]$5 % K⁻¹
• $[x_t|y_t] \propto [y_t|x_t][x_t]$6 % K⁻¹, $[x_t|y_t] \propto [y_t|x_t][x_t]$7 % K⁻¹
• $[x_t|y_t] \propto [y_t|x_t][x_t]$8

Observational data (MODIS+ERA-Interim):

• $[x_t|y_t] \propto [y_t|x_t][x_t]$9 % K⁻¹, $x_t$0 % K⁻¹
• $x_t$1

Key posterior quantities:

• $x_t$2 % K⁻¹, $x_t$3 % K⁻¹
• $x_t$4 % K⁻¹, $x_t$5 % K⁻¹
• 95% prediction interval: $x_t$6 % K⁻¹

This calculation demonstrates how both high inter-state correlation (ρ) and strong observational constraint (high SNR) are necessary to sharply reduce uncertainty in future projections.

5. Sensitivity and Limitations of HEC

Neglecting either SNR or ρ has substantial impacts:

• Setting $x_t$7 eliminates the emergent constraint; $x_t$8, $x_t$9.
• Assuming $$[z_{t+\tau}|y_t]=\int [z_{t+\tau}|x_t]\,[x_t|y_t]\,dx_t$$0 (i.e., $$[z_{t+\tau}|y_t]=\int [z_{t+\tau}|x_t]\,[x_t|y_t]\,dx_t$$1): recovers the "classic" EC result: $$[z_{t+\tau}|y_t]=\int [z_{t+\tau}|x_t]\,[x_t|y_t]\,dx_t$$2, $$[z_{t+\tau}|y_t]=\int [z_{t+\tau}|x_t]\,[x_t|y_t]\,dx_t$$3. For finite $$[z_{t+\tau}|y_t]=\int [z_{t+\tau}|x_t]\,[x_t|y_t]\,dx_t$$4, the classic approach will over-weight the observations, underestimate uncertainty, and potentially bias the mean if $$[z_{t+\tau}|y_t]=\int [z_{t+\tau}|x_t]\,[x_t|y_t]\,dx_t$$5 and $$[z_{t+\tau}|y_t]=\int [z_{t+\tau}|x_t]\,[x_t|y_t]\,dx_t$$6 differ.

The fractional anomaly mapping ($$[z_{t+\tau}|y_t]=\int [z_{t+\tau}|x_t]\,[x_t|y_t]\,dx_t$$7) and variance reduction depend jointly and nonlinearly on both $$[z_{t+\tau}|y_t]=\int [z_{t+\tau}|x_t]\,[x_t|y_t]\,dx_t$$8 and SNR. Figures 4–5 in the reference detail these dependencies.

6. Generalizations and Extensions within the Earth System

HEC is not restricted to univariate or Gaussian settings. Extensions include:

• Multivariate HEC: $$[z_{t+\tau}|y_t]=\int [z_{t+\tau}|x_t]\,[x_t|y_t]\,dx_t$$9, $y_t = x_t + n_t$0, $y_t = x_t + n_t$1 as vectors; equations generalize to a matrix–Kalman update.
• Non-Gaussian priors/likelihoods: Inference via MCMC or variational Bayes replaces analytic solutions.
• Multiple observations: Multiple data streams $y_t = x_t + n_t$2 can be combined in the Bayes step.
• Spatio-temporal HEC: Embedding $y_t = x_t + n_t$3 and $y_t = x_t + n_t$4 as fields in dynamic (e.g., Kalman filter or particle filter) frameworks.
• Causality checks: Causal discovery tools (Pearl’s do-calculus, Sugihara’s convergent cross-mapping) provide safeguards that emergent constraints reflect genuine physical linkages rather than coincidental correlations.

A plausible implication is that HEC can be adapted to a broad range of Earth system feedbacks and processes, provided the required statistical dependencies, error characteristics, and causal relationships can be reasonably specified.

7. Summary and Implications

HEC unites observational uncertainty ($y_t = x_t + n_t$5), model-structural uncertainty ($y_t = x_t + n_t$6, $y_t = x_t + n_t$7), and emergent correlation (ρ) into closed-form expressions for the posterior mean and variance of future climate states given empirical observations. It formalizes emergent constraint approaches, ensuring both sources of uncertainty are rigorously propagated. Applications such as the snow–albedo feedback demonstrate that both strong emergent correlations and high signal-to-noise in observations are essential for meaningful reductions in predictive uncertainty. Neglecting these components risks over-confident or biased projections of future climate responses (Bowman et al., 2018).