Dynamic Nonlinear Granger Factor Models

• The framework extends classical Granger causality by incorporating nonlinear transitions and low-dimensional factor structures for dynamic causal inference.
• It employs advanced techniques like operator-valued kernels, Koopman-inspired autoencoders, and weighted factor graphs to capture state-dependent interdependencies.
• Empirical studies reveal enhanced forecasting accuracy and causal edge detection in diverse fields such as neuroscience, finance, and environmental sciences.

Dynamic nonlinear Granger-causal factor models are a class of statistical machine learning frameworks designed to discover, quantify, and forecast time-varying, non-additive dependencies among components of multivariate time series. These models generalize classical linear Granger-causal vector autoregressions by (i) incorporating nonlinear transition mechanisms, (ii) imposing explicit or implicit low-dimensional factor structures, and (iii) enabling time-varying or conditionally activated causal graphs. Theoretical motivation and empirical validation for these models arises in domains where complex, state-dependent interdependencies between variables are present, such as neuroscience, finance, gene regulatory networks, and environmental sciences (Gregorová et al., 2017, Adesunkanmi et al., 2024, Brown et al., 27 May 2025).

1. Foundations and Theoretical Rationale

Classical Granger causality formalizes a directed temporal dependency: variable $x_j$ is said to Granger-cause $x_i$ if, after conditioning on the entire past of $x_i$, incorporating past information from $x_j$ further improves prediction of $x_i$’s future (Gregorová et al., 2017). In practical settings, standard linear vector autoregressive models (VARs) impose severe limitations: they are inadequate for modeling non-Gaussian dynamics, nonlinear feedback, and systems with latent regimes or mixtures of mechanisms.

Dynamic nonlinear Granger-causal factor models address these challenges by simultaneously lifting the data into nonlinear (possibly overcomplete) feature spaces and by encoding contemporaneous or lagged interactions through flexible, often sparsity-promoting, parameterizations. Additionally, they allow the causal structure to evolve over time, either discretely (across behavioral or latent states) or smoothly (modulated by low-dimensional state variables) (Brown et al., 27 May 2025).

2. Formal Model Classes and Their Specification

Three dominant operationalizations of dynamic nonlinear Granger-causal factor models appear in recent research:

1. Operator-Valued Kernel Methods (Vector-Valued RKHS): Functions $f: \mathbb{R}^{mp} \to \mathbb{R}^m$ are learned in an RKHS with a matrix-valued kernel $H$, itself decomposed to encode partitioned nonlinearities by input and output series:

$$H(x, x') = \sum_{j=1}^m \sum_{i=1}^{s_j} k^{(ji)}(x^{(j)}, x'^{(j)}) \cdot L^{(ji)}$$

where each $k^{(ji)}$ is a dictionary input kernel and $L^{(ji)}$ are learned diagonal output kernels. Granger-causal relations are recovered by thresholding the diagonal entries of $x_i$0 (Gregorová et al., 2017).

2. Koopman-Inspired Nonlinear Latent Autoencoders (e.g., NKDCD): Nonlinear encoders $x_i$1 lift observed time series into a latent space with factorized coordinates, in which linear sparse dynamics (VARs) are then fit:

$x_i$2

Sparsity penalties (e.g., group lasso variants) on blocks of the matrices $x_i$3 identify nonlinear Granger-causal influences. The decoder $x_i$4 ensures expressivity for general nonlinear observation models (Adesunkanmi et al., 2024).

3. Conditionally-Weighted Mixture of Nonlinear Factor Graphs: The observed time series $x_i$5 is modeled as a dynamic convex combination of $x_i$6 static causal graphs/factors, each realized by a nonlinear autoregressive map $x_i$7:

$x_i$8

Dynamic graph $x_i$9 is formed as a time-dependent weighted sum of static adjacencies inferred from each factor. The state model $x_i$0 outputs the weights $x_i$1, capturing time-variation in causal structure (Brown et al., 27 May 2025).

3. Encoding and Inferring Nonlinear Granger Causality

These models generalize linear Granger causality by considering nonlinear (potentially nonparametric) maps for forecasting, with directed, lagged structure encoded through (i) partitioned kernel composition, (ii) block-sparse transition matrices, or (iii) edge-variable (or lag/factor) attributions in neural network parameters.

• In operator-valued kernel approaches, a causal link $x_i$2 is present if some $x_i$3 after solving a regularized risk problem.
• In neural latent models, $x_i$4 is identified if, for at least one lag, $x_i$5 where each $x_i$6 is a latent block.
• In factor mixture models, the effect of $x_i$7 on $x_i$8 at lag $x_i$9 in factor $x_j$0 is quantified by the magnitude $x_j$1; dynamic Granger graphs are recovered by thresholding weighted sums over lags and factors.

Model sparsity and identifiability are enforced by penalties: entry-wise $x_j$2, group $x_j$3, or structured lasso, ensuring that non-informative lags/series are pruned from the inferred graph (Gregorová et al., 2017, Adesunkanmi et al., 2024, Brown et al., 27 May 2025).

4. Estimation, Optimization, and Scalability

Estimation procedures for these models are unified by the need to jointly fit nonlinear function parameters and causal graph structure. Key algorithmic principles include:

• Proximal Gradient/ISTA: Employed for problems with separable convex regularizers; e.g., operator-valued kernel models (per-output, per-series subproblems decouple) and block-group lasso on latent VAR weights (Gregorová et al., 2017, Adesunkanmi et al., 2024).
• Alternating (Block) Minimization: Under more complex lasso penalties (e.g., group, hierarchical), alternate between variable blocks (e.g., coefficients and kernel/output matrices). May not guarantee global optimum but converges to a stationary point.
• End-to-End Gradient Learning: In deep autoencoders or neural mixture models, all parameters—including nonlinear encoders/decoders, VAR weights, and state models—are optimized jointly via gradient descent, occasionally interleaved with proximal shrinkage steps (Adesunkanmi et al., 2024, Brown et al., 27 May 2025).
• Parallelization: The decoupling of per-output subproblems or factorwise updates enables substantial parallelism, improving scalability to large $x_j$4 ($x_j$5) or $x_j$6.
• Complexity: Main computational cost arises from Gram matrix manipulation (kernel methods, $x_j$7 per output) or block-wise proximal updates in large neural models ($x_j$8 for $x_j$9 variables, $x_i$0-dimensional embedding, $x_i$1 lags) (Gregorová et al., 2017, Adesunkanmi et al., 2024).

5. Empirical Performance and Benchmarking

Dynamic nonlinear Granger-causal factor models demonstrate superior empirical performance on synthetic and real-world datasets with nonlinear, nonstationary, or state-dependent dynamics:

• On synthetic datasets with multi-block or dynamic structure, operator-valued kernel methods recover the correct sparse causal block structure and achieve 5-10% lower hold-out MSE versus linear competitors (Gregorová et al., 2017).
• Koopman-inspired autoencoder models (NKDCD) consistently outperform linear VARs and componentwise neural models in AUROC/AUPR on Lorenz-96, fMRI, and gene network benchmarks—offering 5–20% lifts in detection metrics (Adesunkanmi et al., 2024).
• Dynamic factor-graph mixtures (REDCLIFF-S) yield 22–28% higher average $x_i$2 for dynamic edge discovery compared to best static causal discovery baselines; in multistate DREAM4 gene networks, ROC-AUC increases from 0.41–0.45 (static baselines) to $x_i$30.49 (Brown et al., 27 May 2025).

Empirical findings indicate that the factorized and dynamic nonlinear models are particularly effective in state-dependent or multi-regime systems, where classic static models are insufficient.

6. Scientific Applications and Case Studies

The applicability of these models extends to a diverse range of scientific and engineering domains:

• Neuroscience: In rodent local field potential recordings, dynamic nonlinear factor models identify state-dependent changes in cross-area causality (e.g., under stress), recovering edges that coincide with established anatomical pathways and suggesting new hypotheses about hippocampal–thalamic interactions (Brown et al., 27 May 2025).
• Economics and Finance: NKDCD recovers known factor structures (e.g., Fama–French 3-factor model), and matches linear VARs in linear settings while outperforming them on nonlinear flows (Adesunkanmi et al., 2024).
• Complex Networks: In gene regulatory data, dynamic nonlinear Granger-causal factor models substantially increase the reliability of dynamic interaction graph identification under heterogeneous or noisy mixtures (Adesunkanmi et al., 2024, Brown et al., 27 May 2025).
• Environmental/Epidemiological Systems: Operator-RKHS and neural Granger factor models enhance predictive accuracy and interpretability in non-Gaussian, nonlinear spatiotemporal systems, as demonstrated in river flow studies (Gregorová et al., 2017).

7. Limitations, Open Challenges, and Future Directions

Limitations of current dynamic nonlinear Granger-causal factor models include:

• The necessity of extensive hyperparameter tuning for embedding dimensions, lag orders, regularization strengths, and architecture size (Adesunkanmi et al., 2024, Brown et al., 27 May 2025).
• Computational complexity scaling poorly with the number of variables and factors due to the size of parameterized latent blocks or kernel dictionaries.
• Model selection—specifically, identifying the appropriate number of factors ($x_i$4), latent dimension ($x_i$5), or dynamic modes—remains an open problem.
• Selection of threshold $x_i$6 for presence/absence of Granger-causal edges is dataset-specific and may require post hoc calibration.

A plausible implication is that, as high-dimensional, highly nonlinear, and nonstationary time series become more common across scientific domains, research will increasingly focus on models that combine expressive, dynamic factorization with explicit causal interpretability, leveraging advances in scalable optimization and principled regularization.

---

Key References:

• "Forecasting and Granger Modelling with Non-linear Dynamical Dependencies" (Gregorová et al., 2017)
• "NeuroKoopman Dynamic Causal Discovery" (Adesunkanmi et al., 2024)
• "Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series" (Brown et al., 27 May 2025)