Sequential Nonlinear Orientation of Edges (SNOE)

• Sequential Nonlinear Orientation of Edges (SNOE) is a constraint-based algorithm that recovers directed acyclic graphs from data modeled by nonlinear additive noise.
• It leverages the pairwise additive noise model (PANM) criterion and a sequential edge orientation strategy with statistical guarantees to resolve undirected edges in CPDAGs.
• The method achieves high computational efficiency and robustness, outperforming traditional approaches on both synthetic and real-world datasets.

Sequential Nonlinear Orientation of Edges (SNOE) is a constraint-based algorithm for causal discovery that recovers directed acyclic graphs (DAGs) in settings governed by nonlinear additive noise models (ANMs). The procedure addresses the challenges of orienting undirected edges in completed partially directed acyclic graphs (CPDAGs) by exploiting a local identifiability criterion—specifically, the pairwise additive noise model (PANM)—and employs a sequential edge orientation strategy with statistical guarantees. SNOE achieves computational efficiency, robustness to model misspecification, and strong empirical accuracy across both synthetic and real datasets (Huang et al., 5 Jun 2025).

1. Structural Equation Models, CPDAGs, and Additive Noise Models

SNOE operates on random variables $X=\{X_1,\dots,X_p\}$ generated by a structural equation model (SEM) whose causal relations are encoded via a DAG $G_0=(V,E)$. Such models satisfy the factorization:

$$p(X_1,\ldots,X_p) = \prod_{i=1}^p p(X_i\mid X_{\text{pa}_{G}(i)})$$

where $\text{pa}_G(i)$ denotes the set of parents of node $i$. Markov equivalence of DAGs is characterized by shared skeletons and v-structures, leading to equivalence classes represented compactly by CPDAGs, where compelled edges are directed and reversible edges remain undirected.

Under the additive noise model (ANM), each node obeys $X_i = f_i(X_{\text{pa}_G(i)}) + \epsilon_i$ with independent noise terms ($\epsilon_i \perp X_{\text{pa}_G(i)}$). The restricted ANM setting further requires $f_i$ to be three times continuously differentiable, the noise densities to be non-vanishing and smoothly differentiable, and an additional differential-equation-based identifiability condition. Under faithfulness and causal minimality, the true DAG $G_0$ is identifiable from $p(X)$.

2. PANM Identifiability and Edge Orientation

The core theoretical innovation is the use of the pairwise additive noise model (PANM) criterion for orienting edges within a CPDAG. For an undirected pair $X$–$Y$ in a partially directed acyclic graph $G$, with conditioning sets $Z_1=\text{pa}_G(X)$, $Z_2=\text{pa}_G(Y)$, the data $(X, Y\,|\,Z_1, Z_2)$ is said to follow a PANM if either:

1. $X = f_X(Z_1) + \epsilon_X$, $\epsilon_X \perp Z_1$ and $Y = f_Y(X, Z_2) + \epsilon_Y$, $\epsilon_Y \perp (X, Z_2)$, or
2. The symmetric relation holds under $Y \to X$.

Under a restricted ANM, if $G$ admits a consistent extension to $G_0$ and $(X, Y\, |\, \text{pa}_G(X),\text{pa}_G(Y))$ adheres to a PANM, the causal direction $X\to Y$ or $Y\to X$ is generically identifiable. Algorithm 1, a sequential edge orientation routine at the population level, repeatedly finds PANM-admissible undirected edges, orients them, and applies Meek's orientation rules (including those for common children). The algorithm is guaranteed to recover the true $G_0$ under the restricted ANM assumptions (Huang et al., 5 Jun 2025).

3. SNOE Algorithm: Stages and Procedures

The practical SNOE algorithm executes in three principal stages:

Stage	Goal	Key Operations
Stage 1	Initial CPDAG construction	PC-stable (two thresholds), v-structures, Meek's rules, candidate edge selection
Stage 2	Sequential orientation (OrientEdges)	PANM-adherence scoring, likelihood-based edge orientation, update CPDAG (Meek + common-child rules)
Stage 3	Covariate pruning	Generalized additive modeling, test for insignificance, prune superfluous parents/neighbors

Stage 1: Constructs an initial CPDAG using PC-stable with two conditional independence thresholds ($\alpha_2 > \alpha_1$) to produce a sparse skeleton and separation sets. V-structures and Meek's rules yield the initial CPDAG, and candidate orientation edges are defined.

Stage 2: For each undirected edge $X$–$Y$, PANM-adherence scores

$$\hat{I}(X \to Y) = \max_{Z \in \text{pa}_G(X)} I(\hat{\epsilon}_X, Z)$$

and similar for $Y \to X$ are computed. Edges are ranked by the minimum adherence score. The highest-ranking edge undergoes a likelihood-ratio based orientation test (see Section 4), and upon orientation, the graph is updated according to local Meek's rules and the common-child rule. This is repeated until no clear PANM-admissible orientation can be made.

Stage 3: For every node, a generalized additive model (GAM) is fit, testing for the contribution of each parent/neighbor; edges deemed statistically insignificant are removed.

An optional fourth stage re-applies orientation with a stricter threshold $\alpha$ to further direct edges under assumed identifiability.

4. Statistical Tests and Consistency Guarantees

The statistical decision for edge orientation employs a likelihood-ratio test. Given nodes $X$–$Y$ with parental sets $Z_1$ and $Z_2$, the competing models are:

• Under $X \to Y$:

$$\hat{F}(X, Y\,|\,Z_1, Z_2) = \hat{p}(Y\,|\,X,Z_2)\cdot \hat{p}(X\,|\,Z_1)$$

• Under $Y \to X$:

$$\hat{G}(X, Y\,|\,Z_1, Z_2) = \hat{p}(X\,|\,Y,Z_1)\cdot \hat{p}(Y\,|\,Z_2)$$

Using two-fold sample splitting, independent log-likelihoods $\ell F_i$ and $\ell G_i$ are calculated. The test statistic is:

$T_n = \frac{LR_n}{\sqrt{n} \cdot \hat{s}}$

where $LR_n = \sum_{i=1}^n (\ell F_i - \ell G_i)$ and $$\hat{s}^2 = \operatorname{Var}_i(\ell F_i - \ell G_i)$$. The null $H_0\!:\! \mathbb{E}[\ell F - \ell G]=0$ admits standard normal asymptotics for $T_n$ as $n \to \infty$. Decisions are made according to $|T_n|$ versus the critical $z$-score.

The SNOE procedure is consistent: for estimators and independence tests satisfying appropriate regularity, and thresholds $\alpha_1,\alpha_2,\alpha \to 0$ as $n \to \infty$, the probability of exact recovery $P(\hat{G}_n = G_0) \to 1$.

5. Computational Complexity and Practical Aspects

The computational complexity of SNOE is composed of contributions from each stage:

• Stage 1 (PC-stable): $O(p^d)$, polynomial for sparse graphs ($d$ is maximum degree).
• Stage 2: For $m$ undirected edges, $O(m)$ local regressions and dependence-score calculations.
• Stage 3: $O(p)$ GAM fits (one per node).

In aggregate, SNOE scales as $O(p^{\text{maxDeg}} + m)$ and is substantially faster than global score-based or continuous-optimization approaches for large $p$.

Key practical choices include:

• Regression via generalized additive models (GAMs) with thin-plate splines; other nonparametric smoothers are also valid.
• Independence measures: normalized mutual information on discretized, debiased residuals, or kernel-based methods.
• Conditional independence testing: partial-correlation, RCoT (fast KCI), or GCM.
• Hyperparameters: empirical defaults are $\alpha_1 \approx 0.05$ (CPDAG), $\alpha_2 \approx 0.25$ (candidate edges), orientation test $\alpha \approx 0.05$ (down to $10^{-3}$ or $10^{-4}$ for pruning).
• Sample splitting versus two-fold cross-validation for the log-likelihood test, with CV yielding modestly higher power at slight computational cost.

SNOE demonstrates robustness to non-Gaussian noise and mild misspecification of SEM structure, performing correct PANM orientation even when residual-independence fails in incorrectly specified directions.

6. Empirical Performance and Benchmarks

SNOE's empirical validation was conducted on both synthetic and real-world datasets:

• Synthetic networks (Mehra, Alarm, Mildew, Water, Magic, $p=11$–$76$) drawn from a variety of SEM families, including linear-Gaussian, invertible nonlinear (cubic, exponential, inverse sine, piecewise), and non-invertible (Gaussian process draws).
  • For $n=1000$, averaged over $N=75$ repetitions, SNOE-SS and SNOE-CV achieve F1 $\approx 0.85$ (linear), $0.78$ (invertible nonlinear), $0.80$ (non-invertible), exceeding NOTEARS (F1 $=0.42$–$0.37$), DAGMA (F1 $=0.28$–$0.36$), SCORE (F1 $=0.30$–$0.32$), CAM (F1 $=0.75$–$0.62$).
  • Structural Hamming Distance (SHD) is 30–50% lower than competitors.
  • Runtimes: SNOE-SS $=$ 30–90 s per graph; NOTEARS/DAGMA $=$ 300–1800 s; CAM $=$ 100–900 s; SCORE $=$ 50–600 s.
• Non-Gaussian noise: Under $t_5$, Laplace, and Gumbel noise, SNOE's F1 remains within $0.02$ of the Gaussian case, outperforming all baselines.
• Real data: Sachs protein-signaling ($p=11$, $n=2603$) with true DAG possessing 17 directed edges.
  • SNOE-CV: F1 $=0.52$, SHD $=12$, TP $=7$, FP $=2$, FN $=9$, wrong-dir $=1$.
  • CAM (F1 $=0.39$, SHD $=19$), NOTEARS (F1 $=0.40$, SHD $=13$), SCORE (F1 $=0.29$).

7. Significance, Limitations, and Connections

SNOE synthesizes the PANM local identifiability criterion with a ranking of undirected edges by residual independence, an efficient likelihood-ratio test, and classic Meek's orientation rules. The result is a methodology that consistently recovers nonlinear causal DAGs from observational data with near-linear computational cost and theoretical guarantees.

This approach is robust to both non-Gaussianity and mild model misspecification, and consistently outperforms global score-based and continuous-optimization DAG learning algorithms in terms of accuracy and computational efficiency. Its design leverages local structure to sidestep prohibitive global search or optimization, making it highly suitable for large-scale and complex SEMs.

This suggests that SNOE marks a significant methodological advance in the constraint-based nonlinear causal discovery literature, particularly for domains where model faithfulness, identifiability, and computational tractability are critical (Huang et al., 5 Jun 2025).