The idea of testing whether the past of one signal improves the prediction of another can be traced back to Wiener. It was Granger who formalized the concept. Granger causality has become one of the most widely used tools to infer directed interactions between the components of a system. The name however can be misleading: it detects predictive relations among the measured signals, not causal relations among the mechanisms that generate them. This article starts by reviewing recent advances in observation-space Granger analysis. It then follows a path to directed structure in latent dynamical models, and asks under which conditions that structure can be read causally.
Figures & tables
Fig. 1: The causal ladder of this tutorial. From left to right: (1) the Granger graph of the observed process; (2) the directed graph of the latent dynamics; (3) the latent graph once it is identifiable; and (4) its causal reading, obtained by interpreting the transition equations as structural mechanisms. The labels below the arrows indicate what is needed to move up one level.
Fig. 2: Running example: three-node dynamical system used throughout the paper. The latent respiratory process x1 drives both the future of heart rate ( x2 ) and blood pressure ( x3 ). Only y2 and y3 are measured. Observation-space analysis may therefore infer a spurious directed link between y2 and y3 due to the shared hidden driver x1 , motivating the transition to latent state-space inference.
Fig. 3: Four failure modes of VAR-based Granger causality. The first three concern the parametrization of the observed process. (a) Preprocessing : filtering, subsampling, or measurement noise in general convert a VAR( p ) into a VARMA( p′,q ) process. (b) Model truncation : finite-order VAR fits truncate the induced VAR (∞) representation, biasing Φk and distorting Gy . (c) VARMA opacity : in VARMA models, Granger non-causality is not encoded as sparsity in Φ or Θ , but as a rational identity involving both [ 32 ] . The fourth failure mode, (d) Latent confounding , is qualitatively different: hidden drivers alter the information set itself, inducing spurious edges in Gy that observation-space estimators assuming causal sufficiency cannot remove. Together, these effects motivate the transition to latent state-space modeling.
Fig. 4: Three common operations in general turn a VAR process into a VARMA: temporal subsampling, linear filtering, and additive measurement noise. Any finite-order VAR fit to the transformed series is misspecified.
Fig. 5: Observation-space sparsity versus latent-space sparsity. Data are generated from the sparse LG-SSM shown in panel (c): a 5-dimensional Markov-1 latent process with sparse transition matrix A (chain 1→2→3→4→5 plus self-loops) and sparse state-noise precision graph GP (edges 1−3 and 2−4 ), observed through identity H with additive measurement noise. The progression 56(OLS)→25(Lasso)→9+2(latent SSM) illustrates the parametrization gap between observation-space and latent-space representations. We use T=2⋅105 samples. (a) OLS VAR( p=3 ) fit on the observations, together with the residual precision graph. Only arrows arriving at time t are displayed. Because the noisy observation process is VARMA, the induced finite-order VAR representation becomes dense even though the latent dynamics are sparse. (b) Row-wise Lasso-VAR( p=3 ) on the same data. The ℓ1 penalty removes many weak coefficients, but several lag- 2 and lag- 3 links persist, reflecting truncation effects and the effective infinite-order dependence induced in observation space. (c) DGLASSO [ 16 ] recovers the true sparse dual graph (Gx,GP) in latent space.
GraphEM [ 15 ]
DGLASSO [ 16 ]
GraphGrad [ 45 ]
Model class
LG-SSM
LG-SSM
General (nonlinear) SSM
Target parameter(s)
A
A , P=Q−1
C (polynomial coeff.)
Graph type
Directed
Directed + undirected
Directed
Inference alg.
Kalman filter/smoother
Kalman filter/smoother
Differentiable particle filter
Optimization alg.
EM + proximal M-step
Block alternating MM
Proximal gradient descent
Sparsity penalty
ℓ1 on A
ℓ1 on A and P
ℓ1 on C
TABLE I: Comparison of graph-based SSM estimation methods.
VAR-GC
GraphEM
σR
att. coeff.
R
P
F1
R
P
F1
0.0
0.60
1.00
0.99
1.00
1.00
0.60
0.75
0.5
0.49 – 0.53
1.00
0.98
0.99
1.00
0.49
0.66
1.0
0.31 – 0.38
1.00
0.96
0.98
1.00
0.37
0.54
1.5
0.19 – 0.27
0.82
0.96
0.88
1.00
0.31
0.47
2.0
0.13 – 0.19
0.38
0.94
0.54
1.00
0.31
0.47
TABLE II: Measurement-noise experiment : additive observation noise induces false negatives in VAR-based Granger causality estimation. Five-node chain, T=500 , α=0.05 with Bonferroni correction. The attenuated coefficient denotes the population OLS value associated with each true edge under measurement noise. Results are averaged over 500 independent Monte Carlo runs. Recall (R), precision (P) and F1 are pooled over runs; a GraphEM edge is declared when ∣Aij∣>0.05 , i=j .
No constraint
dmax=1
Pattern
SNR
Pair
Full
PCMCI lin
PCMCI nl
LS-CaRe
GraphEM
Single
10 dB
0.63
0.50
0.13
0.16
1.00
0.64
20 dB
0.58
0.75
0.18
0.13
1.00
0.50
40 dB
0.44
0.62
0.14
0.12
0.90
0.53
Circular
10 dB
0.09
0.00
0.05
0.18
0.71
0.60
20 dB
0.17
0.13
0.18
0.15
0.88
0.57
TABLE III: Nonlinear-constraints experiment : F1 scores on FitzHugh–Nagumo electrogram data [ 28 ] . The best-performing method in every setting uses the spatial locality constraint dmax=1 (LS-CaRe, GraphEM). PCMCI lin : linear partial correlation and PCMCI nl : nonlinear. Bold: best method per row.
Fig. 6: Conceptual relation between SCMs, SDCMs, and state-space models. SCMs encode static causal mechanisms; SDCMs extend them to dynamical systems with feedback and interventions on trajectories; discretizing an SDCM yields the transition equations of an SSM, which can also be read structurally on its own.
Fig. 7: Hidden intermediate node experiment . (a) True latent graph Gx . Blue nodes are latent components indirectly measured through noisy observations; x3 (gray dashed) is completely hidden. (b) Observation-space Granger causality applied to the observed processes yt=(y1,y2,y4,y5)⊤ . Besides the true observable edges, three spurious direct edges appear (red) due to the hidden mediator x3 . (c) GraphEM reconstructs the latent transition graph using the same yt and recovers the true Gx . Results are from a single run at T=50,000 .
Víctor Elvira ( Senior Member, IEEE ) is a Distinguished Research Professor at the Basque Centre for Applied Mathematics (BCAM), Bilbao, Spain, and a Professor with a Personal Chair in Statistics and Data Science at the School of Mathematics, University of Edinburgh. He received his Ph.D. in statistical signal processing from the University of Cantabria (Spain) in 2011. His research interests are in statistical signal processing, computational statistics, and probabilistic machine learning, with a focus on state-space models, Monte Carlo methods, and graph-based formulations for time-series data. He is an ELLIS Fellow, Fulbright Fellow, Marie Curie Fellow, Leverhulme Fellow, Alan Turing Fellow, and recipient of the 2024 EURASIP Early Career Award. He is the Principal Investigator of an ARIA-funded project (2025–2028) on probabilistic forecasting of climate tipping points.