Sign Identifiability of Causal Effects in Stationary Stochastic Dynamical Systems

Imagine you are a detective trying to figure out how a complex machine works, but you can't open the box. You can only watch the machine run and record the noise it makes.

This paper is about solving a specific type of detective puzzle: ** figuring out the direction of cause-and-effect in a system that never stops moving, using only a "snapshot" of its behavior.**

Here is the breakdown of the paper using simple analogies.

1. The Setting: The Endless Dance Floor

The authors are studying Stationary Stochastic Dynamical Systems. That's a fancy way of saying: "A system that is constantly moving, influenced by random noise, but has settled into a steady rhythm."

The Analogy: Imagine a crowded dance floor where people are constantly bumping into each other, pushing, and pulling. Even though everyone is moving randomly, the crowd as a whole has a steady "vibe" or pattern.
The Goal: We want to know: If I push Person A, does Person B move forward or backward? (i.e., What is the sign of the causal effect?)

2. The Problem: The "Scale" Mystery

In the past, detectives (scientists) had a strict rule: to solve the puzzle, they needed to know exactly how "loud" the random noise was (the diffusion matrix).

The Analogy: Imagine trying to figure out who pushed whom, but you don't know if the dance floor is made of ice (slippery) or rubber (grippy). If you don't know the surface, you can't tell if a small push caused a big slide or a tiny slide.
The Paper's Innovation: The authors realized that the system has a special property called Scale Invariance. It's like zooming in or out on a map. If you double the size of the map, the directions (North vs. South) stay the same, even if the distances change.
The Breakthrough: They decided to stop trying to measure the exact strength of the push (which changes with the scale) and instead focus only on the direction (the sign). Is it a push (+) or a pull (-)?

3. The New Detective Tool: "Edge-Sign Identifiability"

The authors introduced a new concept called Edge-Sign Identifiability.

The Question: "Given the pattern of movement we see on the dance floor, can we be 100% sure that Person A pushes Person B, or could it be that A pulls B?"
The Three Outcomes:
1. Identifiable: The evidence is clear. The pattern only makes sense if A pushes B. (The sign is +).
2. Non-Identifiable: The evidence is confusing. The pattern looks exactly the same whether A pushes B or A pulls B. You can't tell the difference.
3. Partially Identifiable: This is the most interesting new discovery. Sometimes, the evidence is clear for some dance patterns but confusing for others. It's like a riddle that is easy to solve if the room is quiet, but impossible if the music is loud.

4. The "Graph" Map

The authors use a map (a graph) to draw the connections between people.

Nodes: The people (variables).
Arrows: The pushes/pulls (causal effects).
Cycles: Unlike old theories that assumed a straight line (A causes B causes C), real life has loops. A pushes B, B pushes C, and C pushes A back. The paper handles these loops naturally.

5. The "Magic" of the Math

The paper provides a set of rules (criteria) to look at the "snapshot" of the dance floor (the covariance matrix) and decide:

Can we determine the direction?
If yes, what is the formula to calculate it?

Example: The Instrumental Variable (The "Third Wheel")
Imagine you want to know if Coffee (A) causes Energy (B). But maybe Sleep (C) causes both.

Old way: You needed to know exactly how much "noise" (randomness) was in the system.
New way: The authors show that if you have a "Third Wheel" variable (like a specific type of coffee bean that affects Energy but not Sleep), you can look at the correlations and mathematically prove the direction of the arrow, even without knowing the exact noise levels.

6. The "Confounding" Surprise

One of the coolest findings is about Confounding (when a hidden third factor influences two people).

The Finding: In many cases, you can't tell the direction. But the authors found that in some specific situations, you can tell the direction, and in others, you can't.
The "Partial" Regime: They proved that "Partial Identifiability" is a real, common state. It's not just "we know" or "we don't know." It's "we know sometimes." This is a huge step forward because it tells scientists exactly when they can trust their conclusions and when they need more data.

Summary: Why Should You Care?

This paper is like upgrading a detective's toolkit.

Before: "I can only solve the case if I know the exact weight of every person in the room."
Now: "I can solve the case by just looking at who moves which way, even if I don't know their weights. And if the evidence is blurry, I can tell you exactly why it's blurry and when it becomes clear."

This is crucial for fields like biology (understanding gene networks), economics (market trends), and climate science, where we often only have "snapshots" of data and need to know if A causes B or B causes A, without needing perfect, impossible-to-get measurements of the underlying noise.

Here is a detailed technical summary of the paper "Sign Identifiability of Causal Effects in Stationary Stochastic Dynamical Systems" by Gijs van Seeventer and Saber Salehkaleybar.

1. Problem Statement

The paper addresses the problem of causal inference in continuous-time, linear, stationary stochastic dynamical systems, specifically modeled by Ornstein-Uhlenbeck (OU) processes.

Context: Unlike traditional Structural Causal Models (SCMs) which often assume acyclic graphs and full time-trajectory data, this work focuses on systems where only stationary observational data (samples from the steady-state distribution) are available. These systems naturally allow for cycles and self-loops.
The Core Challenge: In standard causal discovery for OU processes, the diffusion matrix ( $D$ ) is often assumed to be known or fixed. However, OU processes possess an intrinsic scale invariance: the drift matrix ( $A$ ) and diffusion matrix ( $D$ ) can be simultaneously rescaled without changing the stationary covariance matrix ( $\Sigma$ ). Fixing $D$ arbitrarily ignores this physical property and imposes an artificial constraint.
The Goal: The authors aim to determine if the sign of a causal effect (an entry in the drift matrix $A_{ij}$ ) can be uniquely identified from the observational covariance matrix $\Sigma$ , given a known causal graph structure, without assuming the diffusion matrix is known.

2. Methodology

2.1 Mathematical Framework

The system is modeled by the linear SDE:
$dX(t) = (AX(t) - b)dt + CdW(t)$
where $A$ is the drift matrix, $C$ is the diffusion coefficient, and $W(t)$ is a Wiener process. The stationary distribution is Gaussian with mean $b$ and covariance $\Sigma$ . The relationship between parameters is governed by the Lyapunov equation:
$A\Sigma + \Sigma A^T = -D$
where $D = CC^T$ . The authors assume $D$ is a diagonal positive definite matrix (uncorrelated noise), though they discuss extensions to general $D$ .

2.2 Key Concept: Edge-Sign Identifiability

Due to scale invariance, the magnitude of $A$ cannot be recovered from $\Sigma$ alone (only $A$ up to a global scaling factor). Therefore, the authors shift focus from recovering the exact value of $A_{ij}$ to recovering its sign ( $+, -, 0$ ).

They define three categories of identifiability for an edge $e$ in a graph $G$ :

Identifiable: For all valid covariance matrices $\Sigma$ compatible with the graph, the sign of the edge is unique.
Non-identifiable: There exist valid $\Sigma$ compatible with both positive and negative signs for the edge.
Partially Identifiable: There exist some $\Sigma$ where the sign is unique, and others where it is ambiguous. This is a novel regime identified in this work.

2.3 Theoretical Tools

M-Faithfulness: A condition ensuring that the marginal independences in $\Sigma$ exactly match the ancestral relationships in the graph $G$ .
Signature Sets ( $M^k_{G,e}$ ): Sets of covariance matrices that can be generated by a drift matrix where the edge $e$ has sign $k \in \{+, -, 0\}$ .
The $M^0$ Criterion (Theorem 4.2): A central theoretical result stating that an edge is non-identifiable for a specific $\Sigma$ if and only if $\Sigma$ belongs to the set $M^0_{G,e}$ (i.e., $\Sigma$ can be generated by a model where the edge weight is zero).
Graphical Criterion (Theorem 4.4): A condition to determine identifiability based purely on graph topology. An edge is identifiable if removing it changes the set of marginal independences (ancestral intersections) implied by the graph.

3. Key Contributions

Relaxation of Diffusion Assumptions: The paper removes the strong assumption that the diffusion matrix $D$ is known. By respecting scale invariance, the authors derive identifiability results that are more robust and physically realistic for stationary processes.
Introduction of Edge-Sign Identifiability: They formalize the concept of identifying the direction of causality rather than the magnitude, which is the only recoverable information under scale invariance.
Discovery of Partial Identifiability: The authors prove that for certain graph structures (specifically confounding and cyclic structures), the sign is not universally identifiable or non-identifiable. Instead, it depends on the specific values of the covariance matrix. They show that this "partial" regime has positive Lebesgue measure, meaning it is a common occurrence, not a rare edge case.
General and Specific Criteria:
- Derived a general graphical criterion for identifiability.
- Provided explicit algebraic expressions for the sign of causal effects in terms of $\Sigma$ for specific structures (e.g., Instrumental Variables, Chains).
- Analyzed the impact of latent variables, showing that they can destroy sign identifiability in structures that are identifiable in the fully observed case (e.g., simple cause-effect), but preserve it in others (e.g., Instrumental Variables).

4. Key Results

The paper analyzes nine specific causal structures (Fig. 1 in the paper), including classical settings (Cause-Effect, Chain, Confounding, Instrumental Variable) and novel cyclic settings.

Fully Observable Case (No Latent Variables):
- Identifiable: Cause-Effect (1a), Chain (1b), Instrumental Variable (1e), and Cycle with IV (1f). For these, explicit formulas for $\text{sign}(\alpha)$ in terms of $\Sigma$ entries are derived (e.g., for IV, $\text{sign}(\alpha) = \text{sign}(\sigma_{zy})/\text{sign}(\sigma_{zx})$ ).
- Partially Identifiable: Confounding (1c) and Cycle of length 3 (1d). In these cases, the sign is identifiable only if the covariance matrix satisfies specific algebraic inequalities. Numerical experiments confirm that roughly 44% to 64% of sampled covariance matrices in these regimes yield a unique sign.
- Non-identifiable: None in the fully observable case for the tested graphs (except where the edge is structurally redundant).
Latent Variable Case:
- When the confounding variable $H$ is hidden, the Cause-Effect (1a) and Confounding (1c) structures become non-identifiable.
- However, the Instrumental Variable (1e) and Cycle with IV (1f) structures remain identifiable even with latent variables. This highlights the robustness of IV settings in stationary SDEs.
Numerical Validation:
- Experiments on 1,000 random samples for each graph structure confirmed the theoretical predictions.
- Partially identifiable graphs showed empirical fractions of identifiability strictly between 0 and 1, validating the existence of the intermediate regime.

5. Significance and Implications

Theoretical Advancement: This work bridges the gap between causal discovery in acyclic SCMs and stationary stochastic processes. It provides the first rigorous characterization of sign identifiability without fixing the diffusion matrix, acknowledging the scale invariance inherent in physical systems.
Practical Relevance: In many real-world applications (e.g., systems biology, economics), full time-trajectories are unavailable, and noise structures are unknown. The ability to determine the direction of a causal link (positive or negative feedback) from steady-state data is highly valuable.
Nuance in Causal Inference: The discovery of partial identifiability challenges the binary view of identifiability. It suggests that for certain complex systems, causal direction can be determined for some datasets but not others, necessitating a data-dependent assessment rather than a purely structural one.
Robustness of IV: The results reinforce the utility of Instrumental Variables in stationary dynamical systems, showing they can recover causal signs even in the presence of unobserved confounders, provided the specific cyclic structure is maintained.

In summary, the paper establishes a new framework for causal inference in stationary dynamical systems, proving that while exact magnitudes are unidentifiable due to scale invariance, the signs of causal effects are often recoverable, with a rich landscape of fully identifiable, partially identifiable, and non-identifiable regimes depending on the graph topology and the presence of latent variables.