Causally Hierarchical Latent Dynamics (CHiLD)

• CHiLD is a hierarchical generative framework that models latent causal dynamics in time series using multi-layer latent variables.
• It employs a structured SEM with block-sparse temporal and hierarchical dependencies, guaranteeing identifiability from a minimal context of 2L+1 observations.
• Empirical evaluations on synthetic and real-world datasets demonstrate its superior performance in latent recovery and unbiased causal effect estimation.

The Causally Hierarchical Latent Dynamic (CHiLD) framework is a graphical and generative modeling approach for hierarchical temporal causal representation learning, designed to recover and analyze latent causal temporal dependencies across multiple layers of abstraction from time series data. CHiLD formalizes identifiability guarantees, model structure, inference methodology, and evaluation procedures within the context of hierarchical latent dynamics and spatio-temporal causal modeling, and provides theoretical results and empirical demonstrations in both synthetic and real-world domains (Li et al., 21 Oct 2025, Li et al., 25 Nov 2025).

1. Theoretical Foundations and Identifiability

The identifiability result at the core of CHiLD is the block-wise identifiability theorem for hierarchical latent processes. Given time series data generated as

$x_t = g(z_t^{(1)}, \epsilon_t^{(0)}),$

$$z_{t,i}^{(l)} = f_i^{(l)}(\mathrm{Pa}_d(z_{t,i}^{(l)}), \mathrm{Pa}_h(z_{t,i}^{(l)}), \epsilon_{t,i}^{(l)})$$

for $l=1,\dots,L-1$, and

$$z_{t,i}^{(L)} = f_i^{(L)}(\mathrm{Pa}_d(z_{t,i}^{(L)}), \epsilon_{t,i}^{(L)}),$$

where each noise term $\epsilon_{t,i}^{(l)}$ is temporally and spatially independent, and functions $g$ and $f_i^{(l)}$ may be nonlinear and stochastic, CHiLD guarantees that under regularity, independence, and injectivity conditions, the joint law of the hierarchical latent variables $z_t = (z_t^{(1)}, \dots, z_t^{(L)})$ is uniquely recoverable (up to invertible transforms within each layer) from the joint distribution of $2L+1$ adjacent observations $x_{t-L},\dots,x_{t+L}$.

Minimum context length for identifiability is $2L+1$: fewer contextual points collapse injectivity and prevent unique recovery of latent states (Li et al., 21 Oct 2025).

2. Hierarchical Latent Causal Model Specification

The structural equation model (SEM) underpinning CHiLD comprises $L$ hierarchical layers of latent variables. For each time $t$ and layer $l$, the dynamic is

• $z_t^{(l)}$ receives time-delayed parents $$\mathrm{Pa}_d(z_{t,i}^{(l)}) \subseteq z_{t-1}^{(l)}$$ and hierarchical parents $\mathrm{Pa}_h(z_{t,i}^{(l)}) \subseteq z_t^{(l+1)}$ (above layer)
• The strictly block-sparse adjacency structure implies no direct arcs from lower delayed layers to high layers (no skipping)
• The observed variable $x_t$ is generated from the first latent layer only, through stochastic function $g$ and independent noise $\epsilon_t^{(0)}$

Conditional independence constraints, such as

$$z_{t,i}^{(l)} \perp z_{t-1,j}^{(l+1)} \mid z_t \setminus \{z_{t,i}^{(l)},z_{t-1,j}^{(l+1)}\},$$

establish identifiability and model sparsity (Li et al., 21 Oct 2025).

3. Spatio-Temporal Generalization and Collapse Theorem

CHiLD is formally a special case of the Spatio-Temporal Hierarchical Causal Model (ST-HCM) (Li et al., 25 Nov 2025). In the generalization, each unit (e.g., spatial region) contains subunits (e.g., sensors) and is indexed over time. The model collapses the infinite-dimensional subunit space to finite-dimensional Q-variables via sufficient statistics. The summary directed acyclic graph (DAG) consists of

• $U_i$: unit-level, time-invariant latent confounders
• $Q_{i,t}^X, Q_{i,t}^A, Q_{i,t}^Y$: covariate, treatment, and outcome Q-variables
• Observed $X_{i,t}, A_{i,t}, Y_{i,t}$ (generated from Q-nodes)

Edge interpretations include direct effects, temporal auto-dependence, dynamic confounding, and spatial spillover via neighbors. The full joint distribution factorizes as

$$p(U_{1:n}, X_{1:n,1:T}, A_{1:n,1:T}, Y_{1:n,1:T}) = \prod_{i=1}^n p(U_i) \prod_{t=1}^T p(X_{i,t}|U_i) p(A_{i,t}|U_i, X_{i,\le t}, A_{i,< t}) p(Y_{i,t}|U_i, X_{i,\le t}, A_{i,\le t}, Y_{i,<t})$$

The spatio-temporal collapse theorem asserts that as the number of subunits per unit increases ($m \to \infty$), the KL divergence between the full subunit-marginal law and the collapsed Q-variable model converges to zero, establishing equivalence (Li et al., 25 Nov 2025).

4. Generative Model and Inference Algorithms

CHiLD implements a VAE-style generative model for time series:

$$p_\theta(z_{1:T}, x_{1:T}) = p_\theta(z_1) \prod_{t=2}^T p_\theta(z_t | z_{t-1}) \prod_{t=1}^T p_\theta(x_t | z_t)$$

Layer-wise priors $p_\theta(z_t | z_{t-1})$ are constructed with cascades of normalizing flows, inverting SEM equations and enforcing independent noise via the block-triangular Jacobian formula. For layer $l$, a multi-layer perceptron (MLP) parameterizes the noise and latent transitions.

Variational inference is conducted with a contextual encoder: a temporal convolutional network (TCN) mapping context window $x_{t-L:t+L}$ to encoded latent states. The ELBO is

$$\mathcal{L}(\theta, \phi) = \sum_{t=1}^T \mathbb{E}_{q_\phi}[ \log p_\theta(x_t | z_t)] - \sum_{t=1}^T \mathbb{E}_{q_\phi}[ \log q_\phi(z_t | c_t) - \log p_\theta(z_t | z_{t-1}) ]$$

Training involves reconstruction loss (MSE or Gaussian), analytic KL via change-of-variables, and KL annealing. Hyperparameters and architectural details follow explicitly as noted in the source (Li et al., 21 Oct 2025).

5. Causal Identification Strategies

CHiLD enables causal identification of interventions in hierarchical time series and spatio-temporal domains. For interventional distributions under a do-operation:

$p(Y_{i, t} | \mathrm{do}(A_{i, 1:t} = a_{1:t}))$

CHiLD supports two strategies:

• Sequential Adjustability: Backdoor path blocking via full spatio-temporal history allows identification by G-computation, integrating over histories and latent confounders.
• Instrumental Variable (IV) Identification: For valid instruments exogenous to outcomes, one can invert Fredholm integral equations to recover causal kernels mapping treatments to outcomes.

Estimation proceeds either via variational EM (mean-field posteriors over latent confounders) or two-stage plug-in algorithms (e.g., Linear Mixed Models or Gradient Boosting Machines for confounder estimation, followed by recursive G-computation for potential outcome estimation) (Li et al., 25 Nov 2025).

6. Empirical Evaluation and Applications

Synthetic data experiments demonstrate that CHiLD achieves state-of-the-art mean correlation coefficients (MCC) for latent recovery, outperforming comparison baselines (β-VAE, FactorVAE, SlowVAE, TCL, PCL, iVAE, TDRL, CaRiNG, IDOL), particularly for $L > 1$ layered settings (Li et al., 21 Oct 2025).

Empirical evaluations on climate, human motion, stock indices, fMRI, and MuJoCo datasets validate generative fidelity (lowest Context-FID, highest correlation scores) and preservation of temporal-hierarchical dependencies. Latent interpolation yields smooth, semantically meaningful variations matching abstract hierarchical control (e.g., global style, fine-grained dynamics).

Spatio-temporal applications, such as traffic sensor networks, show that aggregating models or ignoring spatial dynamics leads to biased effect estimates under confounding and spillover; CHiLD recovers unbiased and consistent estimates, with robustness to time-varying drift or spatial ordering violations (Li et al., 25 Nov 2025).

7. Significance, Limitations, and Extensions

CHiLD provides a principled graphical and generative framework for hierarchical causal dynamics, establishes rigorous identifiability results under mild conditions, and demonstrates consistent empirical performance. The collapse theorem furnishes foundational justification for summary modeling of complex hierarchical dynamics. A plausible implication is that CHiLD is extensible to arbitrary spatio-temporal domains with unobserved confounders and arbitrary hierarchy depth.

Potential limitations include the requirement for $2L+1$ context length for identifiability, and assumptions on regularity, independence, and injectivity in generative mechanisms. Extensions to higher-order Markov dependencies and alternative parameterizations appear straightforward within the outlined architecture (Li et al., 21 Oct 2025, Li et al., 25 Nov 2025).