Multi-Scale Coherence of Represented Flows

This paper introduces a representation-dependent diagnostic called the "coherence matrix" to test the stability of finite-separation flow geometry across observational resolutions, demonstrating through synthetic fields, Lorenz dynamics, and renormalization-group flows that this metric reveals structural inconsistencies in representations, models, and truncations that standard local or spectral diagnostics may miss.

The Gist

Imagine you are trying to understand a complex, flowing river. Usually, scientists look at the river in two main ways:

1. The "What": How much water is there? How fast is it moving on average? (This is like looking at the river's speedometer or measuring the total volume).
2. The "Where": If you drop two leaves in the river, how far apart do they end up after a minute? (This is like looking at local turbulence or how much the water stretches).

This paper introduces a third way to look at the river. It asks a specific question: "If we look at the river through different-sized glasses (resolutions), does the direction the water is pushing two points apart stay consistent?"

The authors call this "Multi-Scale Coherence." Think of it as a "consistency check" for how a system behaves when you zoom in and out.

Here is a breakdown of their findings using simple analogies:

1. The Core Idea: The "Zoom Lens" Test

Imagine you have a map of a city.

• Resolution A is a high-definition satellite image where you can see individual cars.
• Resolution B is a blurry, low-resolution map where you only see neighborhoods.

The authors take two points on the map (say, two houses) and ask: "If I draw an arrow showing the direction of traffic between these two houses, does that arrow point the same way on the high-def map as it does on the blurry map?"

If the answer is "Yes, the arrow points the same way," the system has High Coherence.
If the answer is "No, the arrow points in a totally different direction," the system has Low Coherence.

The paper argues that standard tools (like measuring average speed or total traffic volume) often miss this. You can have two cities with the exact same amount of traffic and the same average speed, but if the directions of the traffic flow change differently when you zoom in and out, they are actually very different cities.

2. The Three Experiments (The "Proofs")

The authors tested this idea in three different "worlds":

A. The "Identical Twins" (Synthetic Fields)

They created two computer-generated wind patterns.

• The Setup: They made sure these two winds were "twins." They had the exact same speed at every point, the exact same energy distribution, and the exact same statistical correlations. By all standard measurements, they were identical.
• The Twist: They arranged the "phases" (the timing of the wind gusts) differently.
• The Result: When they applied their "Zoom Lens" test, the two winds looked completely different. One stayed consistent when zoomed; the other got messy.
• The Lesson: Just because two things look the same on a standard checklist (speed, energy), it doesn't mean they behave the same way when you look at the geometry of their flow from different distances.

B. The "Distorted Mirror" (Lorenz System)

They looked at the famous "Lorenz System," a mathematical model of chaotic weather (like a butterfly effect).

• The Setup: They took the weather model and then "wrinkled" the coordinate system (like looking at the weather map through a funhouse mirror). The actual weather physics didn't change; only the way we described it changed.
• The Result: The "Zoom Lens" test showed a big drop in coherence. The map looked messy because the "wrinkles" in the paper distorted how the arrows pointed between two points.
• The Lesson: This tool is sensitive to how you represent the data. If you change the map or the coordinates, the "directional consistency" changes, even if the underlying reality is the same.

C. The "Rough Draft vs. Final Draft" (Renormalization Group)

In physics, scientists often try to solve complex equations by simplifying them (truncating them). Imagine writing a novel:

• Draft 1 (M=4): You only write the first 4 chapters.
• Draft 2 (M=6): You write the first 6 chapters.
• The Question: If you look at the story's direction in the first 4 chapters, does it match the direction in the first 6 chapters?
• The Result: When the story was simple, the drafts matched perfectly. But as they added more complex "higher-order" details (chapters 5 and 6), the direction of the plot in the shorter draft started to drift away from the longer draft.
• The Lesson: This tool helps physicists see if their simplified models (shorter drafts) are losing the "shape" of the full story when they ignore complex details.

3. What This Means (In Simple Terms)

The paper concludes that this "Coherence Matrix" is a new kind of ruler.

• Old Rulers: Measure speed, energy, and local stretching.
• New Ruler: Measures geometric consistency across different levels of detail.

It tells us that you can have two systems that look identical on a standard report card (same stats, same local behavior) but are actually organizing their "flow" in completely different ways when you look at the bigger picture.

The Bottom Line:
This isn't a magic wand that fixes physics or predicts the weather. It's a diagnostic tool. It's like a mechanic who, instead of just checking the engine's horsepower, checks if the gears mesh together smoothly whether you look at them with a magnifying glass or a telescope. If the gears don't mesh consistently across those views, the engine (or the model) has a hidden geometric flaw that standard tests missed.

Technical Summary

Technical Summary: Multi-Scale Coherence of Represented Flows

Problem Statement
Many problems in nonlinear and statistical physics are formulated through "represented flows," including physical-space vector fields (e.g., velocity, magnetic fields), phase-space drift fields (e.g., dynamical systems), and truncated renormalization-group (RG) beta functions. Standard diagnostics for these systems typically address specific aspects: spectra and correlation functions measure amplitudes and low-order statistical structure; Lyapunov exponents and local Jacobians measure local deformation and instability; and fixed-point coordinates describe local behavior near special theories.

However, these standard quantities do not generally determine the directional organization of finite vector-field increments after coarse graining. Fields with identical second-order statistics (spectra, correlations) can possess different phase organizations; dynamical systems with similar local stretching can differ in finite-separation drift geometry; and nearby truncations of an RG flow can agree near a fixed point while organizing the surrounding beta field differently. There is a need for a diagnostic that tests whether the finite-separation flow geometry remains stable across observational resolutions and representations.

Methodology
The paper introduces a representation-dependent diagnostic called the multi-scale coherence matrix. This method tests the stability of flow geometry by comparing the direction of vector-field increments across a fixed separation $r$ when the field is observed at two different resolutions, $\ell$ and $L$ (where $\ell, L < r$).

1. Definition: Given two vector fields $A$ and $B$ (or a single field compared at two resolutions), the fields are smoothed at resolutions $\ell$ and $L$ using a smoothing kernel $G$. Finite-resolution increments are defined as $\delta_r A_\ell(x) = A_\ell(x+r) - A_\ell(x)$.
2. Local Coherence: The local two-field, two-resolution coherence is the normalized inner product (cosine) of these increments:
   $$S_{\ell L}^{A,B}(x, r) = \frac{\delta_r A_\ell(x) \cdot \delta_r B_L(x)}{\|\delta_r A_\ell(x)\| \|\delta_r B_L(x)\|}$$
   This metric is defined relative to a specific representation, metric, and sampling protocol.
3. Averaging: The statistic is averaged over base points, displacement directions, and sampled separations to produce a coherence matrix $S_{\ell L}(r)$. The analysis uses dimensionless ratios $a = L/r$ and $b = \ell/r$ to bin the data.
4. Variants: The paper defines both signed (retaining orientation) and unsigned (measuring alignment strength independent of sign) variants. A local windowed average is also defined to identify spatially localized coherence.

Key Contributions and Results
The diagnostic is demonstrated in three distinct settings to show its ability to detect geometric organization not fixed by standard diagnostics:

1. Synthetic Vector Fields (Physical Space):
   Two two-dimensional periodic, divergence-free vector fields were constructed with identical Fourier amplitudes, shell spectra, and scalar two-point correlations. Field A had organized phases, while Field B had randomized phases.

   • Result: Despite being indistinguishable by second-order statistics, the fields produced distinct coherence matrices. Field A showed higher coherence (0.903 signed average) than Field B (0.839). This demonstrates that second-order statistics do not determine cross-resolution increment geometry.

2. Lorenz Phase-Space Flow (Dynamical Systems):
   The diagnostic was applied to the Lorenz vector field on a sampled phase-space region.

   • Coordinate Wrinkling: A smooth coordinate transformation (wrinkling) was applied. While the underlying dynamics remained conjugate (unchanged), the represented drift geometry changed. The coherence matrix decreased significantly (from 0.9987 to 0.9543), showing the diagnostic is sensitive to the representation, not just the abstract flow.
   • Model Mismatch: A weakly perturbed model was compared to the true Lorenz field. Local stretching proxies (largest singular value of the Jacobian) remained highly correlated (Pearson $r=0.979$), yet the true-model cross-coherence was systematically reduced compared to the true self-coherence. This indicates that local linear agreement does not guarantee agreement in finite-separation drift geometry.

3. Functional Renormalization-Group (FRG) Flows (Theory Space):
   The method was applied to beta-function vector fields in the three-dimensional $O(1)$ scalar theory using Local Potential Approximation (LPA) with polynomial truncations $M=4, 5, 6$.

   • Internal vs. Cross-Truncation Coherence: Projected low-order beta fields ($M=4, 5, 6$) remained internally coherent under smoothing. However, cross-truncation coherence decreased as higher-order coupling directions ($\lambda_5, \lambda_6$) were activated (controlled by a width factor $\chi$).
   • Hierarchy: The geometric mismatch was ordered by truncation distance; the $M=4$ and $M=6$ truncations showed the largest coherence loss, while neighboring truncations ($M=5$ vs. $M=6$) remained closer. This provides a direct measure of how projected low-order beta-field geometry is sensitive to the higher-order sector.

Significance and Claims
The paper claims that the multi-scale coherence diagnostic provides a practical field-level consistency test for represented flows. Its significance lies in:

• Complementarity: It does not replace standard local, spectral, or fixed-point diagnostics (such as critical exponents or Lyapunov exponents) but complements them.
• Geometric Sensitivity: It detects geometric organization—specifically the directional stability of finite increments across resolutions—that is invisible to isotropically averaged second-order diagnostics and local linear data.
• Representation Awareness: It explicitly tests whether a chosen representation, model, truncation, or regulator preserves the finite-separation geometry of a flow. In the context of FRG, it offers a way to compare truncations beyond fixed-point coordinates, probing the extended structure of the beta field in theory space.

The authors emphasize that the statistic is protocol-dependent (relying on variables, smoothing, metric, and sampling), making it most useful for comparative studies where these protocols are fixed to assess the fidelity of different representations or models.