Vortex Stretching in the Navier-Stokes Equations and Information Dissipation in Diffusion Models: A Reformulation from a Partial Differential Equation Viewpoint

This paper proposes a novel inverse-time PDE framework for Navier-Stokes vortex stretching that integrates score-based diffusion models to learn Lagrangian particle trajectories, revealing that information about initial positions dissipates rapidly in compressive directions while being preserved in stretching directions.

The Gist

The Big Picture: Un-mixing the Milk

Imagine you have a glass of milk and you drop a drop of red food coloring into it. If you stir it, the red color spreads out, swirls, and eventually mixes completely with the white milk. This is forward time: things get messy, spread out, and lose their original shape. In physics, this is called "diffusion."

Now, imagine you want to do the opposite: you want to look at the mixed-up pink milk and figure out exactly where the drop of red was before you stirred it. This is the inverse problem. In the real world, this is usually impossible because the information about the original drop has been "scrambled" and lost forever.

This paper asks: Is there a way to "un-stir" the milk? Specifically, the author is looking at how tiny whirlpools (vortices) in fluids behave when we try to run the movie backward.

The Problem: The "Backward Blur"

The author, Tsuyoshi Yoneda, explains that if you try to mathematically run the equations of fluid motion backward, you hit a wall. It's like trying to play a video of a shattered vase reassembling itself, but the laws of physics say the pieces should keep flying apart. The math becomes "ill-posed," meaning it breaks down and gives nonsense results.

However, the author noticed something cool: The math used to describe how fluids mix (Navier-Stokes equations) looks very similar to the math used in modern AI image generators (Diffusion Models).

• AI Image Generators: These AI tools learn by taking a clear picture, adding random noise until it's just static, and then learning how to remove that noise to get the picture back.
• The Connection: The author realized that the "noise" in AI is mathematically similar to the "viscosity" (thickness/friction) in fluids.

The Solution: The "Score" Function

To fix the broken backward math, the author borrowed a trick from AI called the Score Function.

Think of the Score Function as a GPS for a lost particle.

• Forward Time: A particle moves randomly, like a drunk person stumbling in the fog. It spreads out.
• Backward Time: We want to guide that particle back to where it started. The "Score" is a signal that tells the particle, "Hey, you are currently at position X, but the most likely place you came from is slightly to the left."

The author's big idea was to absorb the messy, broken math (the "backward blur") into this GPS signal. Instead of fighting the math, they let the AI learn the GPS signal (the "score") directly from data.

The Experiment: Stretching and Squeezing

The author set up a simulation of a specific type of fluid flow called a Burgers vortex. Imagine a piece of dough being pulled apart in one direction (stretching) while being squashed in the other (compressing).

They used a neural network (a type of AI) to learn the "GPS signal" needed to reverse this process. They tracked thousands of tiny particles as they moved forward, and then tried to use the AI to pull them back to their starting points.

The Results: What Was Lost and What Was Saved?

The experiment revealed a fascinating difference between the two directions of the flow:

1. The Squeezing Direction (Compression):

   • Analogy: Imagine squeezing a sponge. The water is forced out, and the sponge gets smaller.
   • Result: When the fluid is squeezed, the information about where the particles started is rapidly lost. Even with the AI's help, it was very hard to guess where the particles came from. The "GPS" signal was too weak to recover the past. The paper calls this "information dissipation."

2. The Stretching Direction:

   • Analogy: Imagine pulling a piece of taffy. It gets long and thin, but the ends stay distinct.
   • Result: In the direction where the fluid is being stretched, the information about the starting position was well preserved. The AI could successfully pull the particles back to their original spots.

The Conclusion

The paper concludes that in turbulent fluids, information is not lost equally in all directions.

• If a fluid is being squashed, the history of the particles is erased quickly and permanently.
• If a fluid is being stretched, the history remains visible and can be reconstructed.

The author suggests that this "information dissipation" is a fundamental part of how turbulence organizes itself. By using AI to learn the "score" (the GPS signal), we can finally see exactly how much of the past survives the chaos of the present, depending on whether the fluid is being stretched or squeezed.

In short: The paper uses AI techniques to reverse-engineer fluid motion. It found that while you can often "un-stretch" a fluid to see where it came from, you generally cannot "un-squash" it because the information gets destroyed in the process.

Technical Summary

Technical Summary: Vortex Stretching in the Navier–Stokes Equations and Information Dissipation in Diffusion Models

Problem Statement
The generation and stretching of vortices in three-dimensional turbulence are fundamental to fluid dynamics, yet the inverse problem—reconstructing initial vorticity distributions from coherent vortex structures observed at later times—remains poorly understood due to the intrinsic irreversibility of the Navier–Stokes equations. While forward-time dynamics involving vortex amplification by strain fields and viscous smoothing are well-established, the backward-time dynamics are ill-posed. Specifically, a naive time reversal of the governing partial differential equations (PDEs) results in a backward Laplacian, rendering the problem mathematically ill-posed. This study addresses two fundamental questions: (1) What types of strain fields preserve or dissipate the most information about the initial vorticity distribution when viewed backward in time? and (2) How can the "information dissipation rate" associated with a given strain field be quantitatively characterized?

Methodology
The paper proposes a new inverse-time formulation of vortex stretching by bridging the gap between the Navier–Stokes equations and score-based diffusion models from a PDE perspective.

1. PDE Reformulation: The axisymmetric vorticity equation with viscosity is rewritten as a two-dimensional Fokker–Planck equation with a reaction term. By introducing a time-reversal transformation ($\tau = T - t$), the authors derive a backward-time PDE. To resolve the ill-posed nature of the backward Laplacian, the diffusion term is absorbed into a drift term. This drift is expressed using a "score function," defined as the gradient of the log-density ($\nabla \log \rho^*$).
2. Discrete Lagrangian Flow: To make the backward dynamics tractable, the authors employ a discrete Lagrangian flow representation. By analyzing the short-time behavior of the fundamental solution (via a time-direction Taylor expansion of the semigroup), they demonstrate that the backward velocity can be approximated by a function of the current state and time step. This formulation explicitly links the Gaussian noise in the forward trajectory to the score function in the backward trajectory.
3. Machine Learning Integration: Since the drift term depends on the initial density $\rho_0$ (which is unknown in the inverse problem), the authors utilize machine learning to approximate the score function. A neural network is trained to learn the mapping from the current state and time step to the backward drift, effectively constructing a closed-form backward-time Lagrangian flow.
4. Numerical Implementation: The framework is applied to an axisymmetric Navier–Stokes flow with a Burgers-type strain field ($U_r = -ar, U_z = 2az$). Forward trajectories are generated to create training data, and a multilayer perceptron (score network) is trained to predict the backward drift. The reconstruction accuracy is evaluated using the relative mean absolute error (MAE) of the initial positions.

Key Contributions

• Theoretical Bridge: The paper establishes a structural equivalence between the forward-time Fokker–Planck equation governing diffusion models and the vorticity equation describing Burgers vortex tubes, providing a theoretical basis for applying diffusion model concepts to vortex dynamics.
• Ill-Posedness Resolution: It offers a rigorous PDE-based reformulation that converts the ill-posed backward Laplacian into a well-defined drift term dependent on a learned score function, avoiding the mathematical inconsistencies of naive time reversal.
• Information-Theoretic Perspective: The study introduces a novel framework to quantify information dissipation in vortex dynamics, specifically analyzing how strain fields affect the reversibility of the flow.

Results
Numerical experiments were conducted for axisymmetric flows with varying viscosity ($\nu = 1$ and $\nu = 0.01$) and a two-dimensional flow extension. The key findings include:

• Directional Information Loss: In the compressive direction (radial, $R$), information about the initial position is rapidly lost as trajectories approach the axis. This is evidenced by large relative MAE values in the $R$ component during backward reconstruction.
• Information Preservation: In the stretching direction (axial, $Z$), information about the initial position is relatively well-preserved, with smaller and more stable relative MAE values.
• Self-Similarity: The learning results exhibit self-similar behavior regardless of the viscosity value, suggesting that the mechanism of information dissipation is robust across different viscous regimes.
• Generalizability: The trends observed in the axisymmetric case were replicated in the two-dimensional flow experiments, reinforcing the conclusion that compressive strain fields drive irreversible information loss.

Significance
The paper claims to provide a new information-theoretic perspective on vortex dynamics, extending classical vortex stretching theory into a reversible framework inspired by modern diffusion models. By demonstrating that information is rapidly erased in compressive directions but preserved in stretching directions, the study offers a quantitative characterization of the "degree of reversibility" in the inverse problem. The authors posit that this approach lays the groundwork for understanding how coherent vortex tube structures emerge from initial distributions and suggests that future work should investigate whether vortex-axis structures and geometric arrangements govern information loss in more general, non-axisymmetric turbulent flows.