🌀 nonlinear sciences

Operator theoretic causality analysis of fluid flows using linearized dynamics

This paper introduces Linear Operator Causality Analysis (LOCA), an operator-theoretic framework that leverages linearized governing equations to quantify causal interactions in fluid flows, offering a physically interpretable alternative to data-driven methods while also providing a data-driven approximation for direct time-series analysis.

Original authors: Ankit Srivastava, Victor Jimenez-Fernandez, Louis Cattafesta, Scott Dawson

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ankit Srivastava, Victor Jimenez-Fernandez, Louis Cattafesta, Scott Dawson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a chaotic dance of fluid particles, like a swirling crowd at a concert. You want to know: Who is leading the dance, and who is just following? Did the person in the front row push the person behind them, or did they both just react to the same loud music?

For a long time, scientists tried to answer this by just watching the crowd and guessing who influenced whom based on patterns in the video. This paper introduces a new, sharper way to look at the dance, called LOCA (Linear Operator Causality Analysis). Instead of just staring at the video, LOCA looks at the rulebook that governs the dance.

The Rulebook vs. The Footage

Most old methods, like Granger causality or Transfer Entropy, are like detectives who only have a blurry security camera. They look at the footage and say, "Hey, whenever Person A moves, Person B moves a split second later, so A must have caused B." But here's the catch: if Person A and Person B are both reacting to a third person, Person C, the detective might get confused and think A caused B.

The authors argue that if you actually have the equations (the rulebook) that describe how the fluid moves, you don't need to guess. You can calculate exactly how a tiny nudge to one part of the fluid will ripple through the system. They call this "operator-theoretic causality." It's like knowing the exact physics of the dance floor: if you push a dancer here, you know exactly where they will bump into someone else, and when.

The "Time-Travel" Matrix

The paper uses a special math tool called a matrix exponential. Think of this as a "time-travel machine" for the fluid's rules.

Immediate Causality: If you nudge a fluid particle right now, does it instantly push its neighbor? The rulebook tells you this immediately.
Delayed Causality: What happens if you wait 5 seconds? The nudge might travel through a chain of other particles before reaching a distant one. The math shows that even if two particles don't touch directly, they can still be connected through a long chain of friends.
The Global View: The authors created a "super-metric" (called M) that sums up all possible connections over all time. It's like asking, "If I push this button, will it ever cause a reaction anywhere in the system, no matter how long I wait?"

What They Ruled Out (The "Don't Trust This" List)

The paper is very clear about what not to rely on if you want the full picture:

Don't trust short-sighted guesses: If you only look at a tiny slice of time, you might miss a connection that takes a while to develop. The paper shows that a connection might look invisible at 1 second but become huge at 20 seconds.
Don't trust "truncated" models blindly: If you try to simplify the fluid dance by ignoring the "quiet" dancers (low-energy modes), you might miss the most important leaders. The authors found that sometimes the "quiet" modes are actually the ones driving the big, loud movements later on.
Don't assume all methods agree: The paper proves that the "rulebook method" (LOCA) and the "camera method" (Granger) give different answers when the dancers are all correlated (moving together). The camera method might say "no influence" because it thinks the movement was predictable, while the rulebook method says "huge influence" because it knows a nudge would have caused a specific reaction.

The Experiments: Two Dance Floors

The team tested their idea on two very different scenarios:

1. The Smooth Slide (Couette Flow)
Imagine two giant plates sliding past each other with fluid in between. This is a smooth, predictable flow.

The Result: They compared their "rulebook" method against the "camera" method. When they looked at the raw data, the two methods disagreed. But when they transformed the data into a special set of "uncorrelated" steps (called POD modes), the two methods agreed perfectly. This confirmed their theory: the methods only agree when the variables aren't confusingly linked.
The Lesson: They also found that if you cut off the "quiet" modes (using only the top 5 or 10), you miss the most important causal links. You need to keep the "quiet" ones to see the whole picture.

2. The Chaotic Wake (Two Plates)
Next, they looked at a messy, chaotic flow behind two tilted plates. This is a turbulent, unpredictable mess where vortices (swirls) are constantly forming and breaking.

The Result: Since they didn't have the perfect rulebook for this messy flow, they used a data-driven trick (DMD) to estimate the rulebook from the video footage.
The Finding: Even in this chaos, their method found clear "leaders." They discovered that specific spots right behind the top plate were the main sources of causality, sending information downstream. The "camera" method (Granger) saw much weaker links because it was confused by the correlations, but the "rulebook" method (LOCA-DMD) saw the strong, structured flow of influence.

How Sure Are They?

The authors are very confident in the math. They have proved that their method is mathematically equivalent to a specific type of "Dynamic Causal Effect" and that it connects to fundamental concepts like controllability (can we steer the system?) and observability (can we see what's happening?).

However, for the chaotic wake flow, they are simulating the results. They didn't prove that LOCA works on every real-world fluid in existence; they showed it works beautifully in their computer models of a smooth slide and a chaotic wake. They suggest that this approach is a powerful new tool, but it relies on having either the equations or very good data to build those equations.

The Takeaway

If you want to understand why a fluid moves the way it does, don't just watch the movie and guess who started it. Look at the script. The authors show that by using the underlying physics (the script), you can find the true leaders of the dance, even when the crowd is moving in a chaotic, confusing way. They've built a tool that sees connections that other methods miss, especially when the system is complex and the dancers are all moving together.

Technical Summary: Operator Theoretic Causality Analysis of Fluid Flows Using Linearized Dynamics

Problem Statement
Fluid flows exhibit complex interactions across spatial and temporal scales, leading to instabilities and turbulence that are difficult to predict and control. Understanding the causal relationships governing these interactions is essential for advancing fluid dynamics. While numerous data-driven methods exist for inferring causality—such as Granger causality, transfer entropy, and various information-theoretic metrics—these approaches rely on statistical analysis of time-series data. A significant limitation of these data-driven techniques is that achieving statistical convergence often requires restricting analysis to a limited number of measurements or modes, potentially obscuring causal pathways inherent in high-dimensional systems. Furthermore, standard statistical methods may yield misleading inferences due to correlated variables or state truncation. There is a gap in the literature for analyzing causality directly from the known governing equations of fluid systems, particularly for linearized dynamics, to provide a physics-based, interpretable alternative.

Methodology
The authors propose Linear Operator Causality Analysis (LOCA), an operator-theoretic framework for quantifying causality in linear and linearized dynamical systems. The core premise is that for a system governed by $\dot{\mathbf{u}} = \mathbf{A}\mathbf{u}$ , the matrix exponential $e^{\mathbf{A}\tau}$ encodes the causal influence of one state component on another over a time horizon $\tau$ .

The methodology is structured around three primary definitions:

Immediate Causality: Defined by non-zero off-diagonal elements in the system matrix $\mathbf{A}$ (i.e., $A_{ij} \neq 0$ ), representing direct coupling.
Delayed Causality: Quantified by the partial derivative of the state at time $t+\tau$ with respect to a perturbation at time $t$ , given explicitly by the matrix exponential elements $[e^{\mathbf{A}\tau}]_{ij}$ .
Global Causality Bound: To address the time-dependence of $e^{\mathbf{A}\tau}$ and capture causal pathways over all time horizons, the authors introduce a cancellation-free matrix $\mathbf{M}_\alpha = (\alpha\mathbf{I} - |\mathbf{A}|)^{-1}$ , where $\alpha > \rho(|\mathbf{A}|)$ . This matrix serves as an upper bound on the exponentially discounted, time-integrated magnitude of the causal response. A zero entry in $\mathbf{M}_\alpha$ certifies that no causal influence exists between two modes for any time horizon.

The framework defines a modified $F$ -statistic, $F_{j \to i} = [e^{\mathbf{A}\tau}]_{ij}^2 \sigma_j^2 / \sigma_{\epsilon, i}^2$ , which incorporates state variances and error terms to quantify causal strength. This formulation is shown to be a specific realization of Dynamical Causal Effects (DCE).

To handle cases where the governing equations are unknown, the authors develop a data-driven counterpart, LOCA-DMD. This approach utilizes Dynamic Mode Decomposition (DMD) to estimate the system matrix $\mathbf{A}$ directly from time-series data, subsequently applying the LOCA framework to the estimated operator.

Key Contributions

Physics-Based Causality Metric: The paper establishes LOCA as a physics-based alternative to statistical causality measures, deriving causal relationships directly from linearized governing equations rather than inferring them from observational data alone.
Global Indicator: The introduction of $\mathbf{M}_\alpha$ provides a rigorous, time-independent certificate for causal disconnection, overcoming the limitations of fixed-time-lag analyses which can miss indirect causal pathways (e.g., in nilpotent systems).
Theoretical Connections: The work explicitly connects LOCA to:
- Dynamical Causal Effects (DCE): Positioning LOCA as the deterministic, local limit of DCE.
- Granger Causality and Transfer Entropy: Clarifying that LOCA and single-lag Granger causality coincide only under specific conditions (e.g., when source variables are uncorrelated and residuals are orthogonal). Generally, LOCA measures the response to an independently specified variation, whereas Granger measures the unique predictive contribution after conditioning on other variables.
- Graph Theory: Interpreting $\mathbf{M}_\alpha$ as a weighted transitive closure that accounts for indirect pathways without cancellation effects.
- Control Theory: Demonstrating that zero entries in $\mathbf{M}_\alpha$ imply rank deficiencies in controllability and observability matrices, offering guidance for sensor and actuator placement.
Data-Driven Extension: The development of LOCA-DMD allows the application of these operator-theoretic insights to systems where only time-series data is available.

Results
The methodology was demonstrated on two fluid flow examples:

Linearized Couette Flow:
- Comparisons between operator-based LOCA, LOCA-DMD, and single-lag Granger causality showed good agreement when applied to Proper Orthogonal Decomposition (POD) coefficients (which are uncorrelated).
- In physical coordinates (where variables are correlated), LOCA and Granger results diverged significantly. LOCA identified strong causal links (e.g., the lift-up mechanism) that Granger analysis underweighted due to the conditioning on correlated variables.
- The study highlighted that truncated POD models can miss dominant causal interactions involving lower-energy modes, whereas balanced truncation (which preserves controllability and observability) retained these pathways more effectively.
- The global causality metric $\mathbf{M}_\alpha$ correctly identified the block-triangular structure of the linearized Navier-Stokes operator, showing no causal influence from vorticity to wall-normal velocity.
Nonlinear Wake Flow (Two Collinear Plates):
- Applied to a chaotic wake flow using only data (LOCA-DMD and Granger).
- In POD space, LOCA-DMD and Granger results were similar, consistent with the uncorrelated nature of POD modes.
- In physical space (using velocity measurements at discrete points), LOCA-DMD revealed a more pronounced and spatially structured causal network compared to Granger causality. LOCA-DMD identified specific upstream locations (near vortex formation) as dominant causal sources influencing downstream regions over extended time horizons, whereas Granger values diminished significantly due to the conditioning on correlated velocity measurements.

Significance and Claims
The paper claims that LOCA provides a "physically-motivated and interpretable measure of causal interactions" that complements data-driven statistical techniques. By leveraging the linearized governing equations (or their data-driven estimates), LOCA mitigates common issues in data-driven analysis, such as misleading inferences caused by correlated variables or the loss of causal pathways due to state truncation.

The authors argue that LOCA offers a distinct perspective from Granger causality: it quantifies the system's response to a specific intervention (counterfactual causality) rather than just the improvement in prediction accuracy. The framework is presented not as a replacement for statistical methods but as a complementary tool that can guide data-driven analyses and provide deeper insight into the structural mechanisms of fluid flows, particularly regarding controllability, observability, and the identification of optimal causal pathways in high-dimensional systems. The work suggests that understanding these operator-based causal structures is vital for the prediction, control, and design of fluid systems.