🤖 machine learning

A Mechanistic Analysis of Transformers for Dynamical Systems

This paper analyzes single-layer Transformers from a dynamical systems perspective, revealing that while softmax attention restricts linear system modeling through oversmoothing, it functions as an effective adaptive delay-embedding mechanism for reconstructing states in nonlinear, partially observable systems.

Original authors: Gregory Duthé, Nikolaos Evangelou, Wei Liu, Ioannis G. Kevrekidis, Eleni Chatzi

Published 2026-08-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gregory Duthé, Nikolaos Evangelou, Wei Liu, Ioannis G. Kevrekidis, Eleni Chatzi

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 trying to predict the future. Maybe you're guessing where a bouncing ball will land next, or trying to forecast the weather. For a long time, scientists have had a set of reliable tools for this, like looking at how a system has moved in the past few seconds to guess where it goes next. These tools are like a strict recipe: if you know the rules of physics, you can write down a math equation that tells you exactly what happens. But recently, a new kind of "super-brain" called a Transformer has become famous for predicting things, from writing stories to guessing stock prices. These Transformers are like a magical black box that looks at a whole history of events at once and makes a guess. The big question is: does this black box actually understand the physics of how things move, or is it just a really good guesser that might fail when things get tricky? This paper dives into the engine room of these Transformers to see if they really know what they are doing when they try to predict the motion of physical systems.

The researchers, led by Gregory Duthé and his team, decided to stop treating the Transformer as a mysterious black box and instead take it apart to see how it works on simple, moving systems. They focused on a specific type of Transformer that is very thin and simple—just one layer deep—to make it easier to understand. They asked: "If we feed this machine data about a swinging pendulum or a flowing river, what is it actually learning?"

Here is what they found, and it's a mix of "it works great" and "it hits a hard wall."

First, they looked at simple, straight-line movements (linear systems), like a spring bouncing up and down. They discovered that the Transformer has a weird blind spot. The way the Transformer pays attention to the past is like a chef who can only add ingredients, never take them away. It can mix past moments together, but it can't subtract them. If a system needs to subtract a past moment to create a wobble (like a spring that bounces back and forth), a single-headed Transformer gets confused. It tries to smooth everything out, making the bouncy spring look like a sluggish, slow-moving blob. It's like trying to draw a zigzag line using only a brush that can only move forward; you end up with a straight line. The paper shows that if you give the Transformer two "heads" (two different ways of paying attention) instead of one, it can fix this and finally draw the zigzag correctly.

Next, they moved to more complicated, wiggly movements (nonlinear systems), like a chaotic pendulum or fluid swirling around a cylinder. Here, the Transformer shines, but only under specific conditions. When the researchers gave the Transformer the full picture of the system (knowing both the position and speed of the object), it didn't do much better than a simple calculator. It was like giving a detective a full map when they only needed a compass.

However, when the researchers hid part of the information—giving the Transformer only the position but hiding the speed—the Transformer became a superhero. It started acting like a time-traveling detective. By looking at a sequence of past positions, it could figure out the hidden speed and reconstruct the full state of the system. It was essentially building a "delayed memory" to fill in the missing pieces, just like how you can guess a car's speed by looking at where it was a few seconds ago. The paper suggests that this is how the Transformer works best: by stitching together a history of observations to create a complete picture of the present.

But there's a catch. The researchers found that the Transformer needs enough "room" in its brain to do this. If they forced it to compress all that complex history into a tiny space, the picture got messy and folded in on itself, like trying to stuff a large map into a tiny envelope. The paper shows that if the system is complex, the Transformer needs a larger internal space to keep the different possibilities separate. Also, they found that if the system changes based on a hidden setting (like the speed of the wind), the Transformer might get confused unless you explicitly tell it what that setting is. Without that extra clue, it might mix up two different scenarios that look similar in the short term but behave very differently later.

In the end, the paper doesn't say Transformers are broken, but it does say they aren't magic. They are powerful tools that can learn to predict physical systems, but they have specific rules they must follow. They struggle with simple subtraction in their memory, they need enough space to unfold complex shapes, and they sometimes need a little help from us to understand the hidden settings of the world they are watching. By understanding these rules, scientists can stop treating these models as mysterious black boxes and start building them to be more reliable for predicting the real, moving world around us.

Technical Summary: A Mechanistic Analysis of Transformers for Dynamical Systems

Problem Statement

While Transformer architectures have achieved significant empirical success in time-series forecasting and are increasingly applied to dynamical systems (from fluid mechanics to neuroscience), their internal mechanisms remain poorly understood from a dynamical systems perspective. Unlike classical autoregressive (AR) and state-space models, which possess well-established theoretical foundations regarding stability, observability, and identifiability, Transformers are often treated as black boxes.

This gap is critical as attention-based models are deployed for general-purpose or zero-shot forecasting across diverse dynamical regimes. Specifically, it is unclear what classes of dynamical behavior Transformers can faithfully represent, under what conditions they succeed or fail, and how their internal computations relate to established concepts in nonlinear dynamics, such as delay-coordinate embeddings and attractor reconstruction. Existing literature has explored connections between attention and numerical integration or operator learning, but a direct mechanistic link between single-layer Transformer operations and classical dynamical system formulations (e.g., AR models, Takens' theorem) has not been firmly established.

Methodology

The authors adopt a "bottom-up" mechanistic interpretability approach, restricting their analysis to single-layer, single-head, decoder-only Transformers. This minimal architecture allows for the isolation of specific components: the causal self-attention mechanism and the feed-forward network (MLP).

The study investigates two primary operational regimes:

Linear Systems: The authors analyze attention as a linear, time-varying autoregressive operator. By removing the residual connection and the MLP (attention-only ablation), they derive closed-form expressions for the effective AR coefficients induced by the attention mechanism. They compare these against analytically derived AR coefficients from physical systems (e.g., Single-DOF and 2-DOF oscillators).
Nonlinear Systems: The authors examine scenarios where the full state is not observable (partial observability). Here, the Transformer is viewed as a mechanism for adaptive delay-coordinate embedding. The attention layer aggregates historical observations to reconstruct an implicit state, which is then processed by the MLP to approximate the nonlinear flow map.

The study utilizes three representative dynamical systems:

Linear: Single-DOF and 2-DOF structural oscillators.
Nonlinear ODE: The Van der Pol oscillator.
Nonlinear PDEs: The Chafee-Infante reaction-diffusion equation and 2D Navier-Stokes flow past a cylinder.

The analysis focuses on how the model handles partial observability, the role of temporal context (number of input tokens $n$ ), the impact of positional encoding (PE), and the dimensionality of the latent space ( $d_{lat}$ ) relative to the intrinsic dimension of the attractor ( $d_A$ ).

Key Contributions and Results

1. Linear Systems: The Convexity Constraint

In the single-head, attention-only setting, the authors demonstrate that the softmax normalization imposes a convexity constraint on the learned autoregressive operators.

Mechanism: The attention weights $\alpha_{t,\ell}$ are strictly non-negative and sum to unity. Consequently, the effective lag operators are non-negative scalar multiples of a single matrix.
Failure Mode: For linear systems requiring mixed-sign autoregressive coefficients (e.g., lightly damped oscillators where the AR coefficients must have opposite signs to represent phase inversion), a single attention head cannot faithfully represent the dynamics. This leads to "oversmoothing" and a failure to recover resonance peaks or coherent modal decay ridges.
Success Mode: When the target dynamics admit an AR representation with same-sign coefficients, the attention mechanism successfully captures the oscillatory behavior and damping characteristics.
Scalability: In multi-degree-of-freedom systems, attention can recover multiple interacting modes if the temporal context ( $n$ ) is sufficiently long to compensate for missing spatial observations, effectively acting as a delay-coordinate embedding. However, the fundamental sign constraint remains a limitation for single-head architectures.

2. Nonlinear Systems: Delay-Embedding and State Reconstruction

For nonlinear systems, the authors find that the Transformer's utility is most pronounced under partial observability.

Full Observability: When the full state vector is available, a simple feed-forward MLP (learning the time-one map) performs comparably to the Transformer. Attention provides no significant advantage because the system is already Markovian.
Partial Observability: When only a subset of variables is observed (e.g., position but not velocity), the Transformer significantly outperforms the MLP. The attention mechanism acts as an adaptive delay-embedding, aggregating past observations to disambiguate states that share the same instantaneous value but differ in phase.
Latent Dimensionality and Unfolding: The quality of state reconstruction depends critically on the latent dimension ( $d_{lat}$ $d_{l a t}$ ).
- In the Chafee-Infante system (intrinsic dimension $d_A=2$ ), setting $d_{lat}=2$ caused the latent representation to collapse into a narrow curve, failing to "unfold" the manifold due to the system's symmetry and the constraints of the architecture.
- Increasing $d_{lat}$ to 3 allowed the attention mechanism to first unfold the delay coordinates into a richer intermediate representation, which the MLP could then project onto the intrinsic 2D manifold. This suggests that $d_{lat}$ must exceed $d_A$ to avoid premature compression and folding.
Parameter Conditioning: In the Navier-Stokes cylinder flow experiment, models trained without explicit Reynolds number ($Re$) input learned a latent space where limit cycles for different $Re$ overlapped. This non-injectivity led to poor reconstruction of the full flow field, particularly in the wake region. Providing $Re$ as an explicit input disentangled the latent space, separating limit cycles along a continuous parameter-dependent manifold and enabling accurate full-field reconstruction.

Significance and Claims

The paper claims to bridge the gap between empirical Transformer performance and classical dynamical systems theory by providing a mechanistic explanation of why and when these models succeed or fail.

Interpretability: The authors argue that Transformers should not be viewed merely as generic sequence-to-sequence regressors but as data-adaptive state-space models. The attention mechanism constructs an implicit state from past observations, while the MLP approximates the local flow map.
Theoretical Limits: The study identifies fundamental representational limitations, specifically the inability of single-head attention to model mixed-sign dependencies in linear systems due to softmax convexity. This suggests that multi-head architectures or alternative normalization schemes may be necessary for certain physical regimes.
Design Guidelines: The results offer practical guidelines for designing attention-based models for dynamical systems:
- Ensure the temporal context length ( $n$ ) satisfies $n \ge 2d_A + 1$ (Takens' theorem) for partial observability.
- Ensure the latent dimension ( $d_{lat}$ ) is sufficiently large to unfold the attractor before projection, particularly for symmetric manifolds.
- Explicitly condition on system parameters (e.g., Reynolds number) if the latent space fails to uniquely separate different dynamical regimes.

The authors conclude that while Transformers can act as general temporal integrators, their inductive biases (specifically regarding spectral filtering and manifold topology) must be carefully aligned with the physical systems they model. They caution against treating them as universal "black-box" approximators without understanding these underlying mechanistic constraints.