When Scale and Time Agree
In Brief
We did not go looking for a universal vortex shape, and we did not find one. What we found instead is relational: for a structural object inside a turbulent field, the route obtained by changing scale and then advancing time is, for the overwhelming majority of objects, the same route obtained by advancing time first and then changing scale. We call this scale–time commutation, and we measure the disagreement between the two routes with a single normalized quantity, D□.
The result was not accepted on a single sample. A second, preregistered, spatially and temporally disjoint cutout of the same simulation — Pilot B — was run through the identical frozen pipeline. The low-disagreement architecture reappeared, and reappeared more strongly. Two of eight preregistered criteria did not replicate, and those two failures are as informative as the six that passed: they tell us which parts of the original signature are the stable core, and which parts are sample-dependent expression. The result is Structural Partial Replication — a more precise, and more honest, finding than a flat pass or fail.
🌀 Turbulence Became a Transformation Problem
Turbulence is usually studied by asking what a structure looks like, or how it changes as you move to a coarser scale — the cascade question. We started somewhere adjacent to that, with an object X inside a forced-isotropic velocity field, and two ways of transforming it: a scale transformation S, and a step forward in time T. There are two routes from X to the same scale-and-time destination:
- Scale then time (ST): shrink or coarsen the object first, then advance it through time.
- Time then scale (TS): advance it through time first, then change its scale.
If turbulence is well-behaved under this pair of transformations, the two routes should land on statistically indistinguishable destinations. If it isn't, they should diverge. We built a single number to measure exactly that disagreement — the normalized scale–time stitching defect:
PST and PTS are the routing distributions produced by the two orderings; JSD is the Jensen–Shannon divergence between them, normalized so that D□ runs from 0 (the two routes are indistinguishable) to 1 (maximal disagreement under the normalized Jensen–Shannon measure). A small D□ means scale and time compatibly agree on where a structure goes. This reframes the question. It's no longer just what structures exist in turbulence — it becomes: does a turbulent object's structural history depend on whether you change scale first or advance time first?
This is why "scale–time commutation" is a stronger claim than saying structures merely persist. Persistence is about survival along one direction. Commutation is about whether two different transformations of the same object agree with each other — a question about the field's transformation grammar, not just its inventory of shapes.
📊 A Signature That Reproduces — and Sharpens
A single low D□ in one cube and one short time window proves very little. It could be a property of turbulence, or it could be an accident of that particular cube and that particular ten-frame window. So before treating this as a finding about the flow, we needed to ask whether it would reappear in a sample sharing no spatial grid points and no stored frames with the first.
Pilot B is that sample: a 256³ cutout translated to the opposite half of the domain and drawn from frames roughly a full time unit later, run through the exact same frozen pipeline — same segmentation rules, same thresholds, same null ensembles, same seeds, all locked before Pilot B was ever evaluated.
The result was never going to be "another low number" by itself — what makes it a finding is that the same ordering reappeared: the real defect sits far below the hardest of three progressively-stricter randomized controls, in both pilots, even as the null levels themselves barely moved between samples and the real defect became noticeably smaller. That combination — a stable null scaffold with a real value that separates from it more, not less, on replication — supports a reproducible effect within the tested representation, rather than one confined to a single sampled region or time window.
It also tells us something about what kind of quantity we should be looking for. If D□ were a universal physical constant, we would have expected the same number in both pilots. It wasn't. What reproduced instead was the relation between the real system and its controls:
Searching for one magic value of D□ would have missed this. The architecture of separation is the invariant, not the number.
🔬 What Explains It — and What Still Doesn't
"Real differs from random" is a weak description once you have a null hierarchy that can be peeled apart layer by layer. Each null in the sequence restores a bit more of the real structure's organization, and each step tells you how much of the commutation that layer alone can account for.
Directed degree/layer structure
Only the basic combinatorics of the routing graph are preserved. That alone gives almost no scale–time compatibility — the raw connectivity rules don't generate the effect.
+ Local spatial geometry
Constraining candidate matches to nearby positions drops the defect by two-thirds. Geometry explains a substantial share of what the naive null destroyed.
+ Feature similarity
Adding local structural-state similarity on top of geometry explains still more. Both N1 and N2 pass their mobility, uniqueness, and real-match quality gates.
Remaining route organization
Even after restoring degree structure, geometry, and feature similarity, the real system sits roughly 93% below the hardest control. That residual is presently unexplained.
That unexplained residual is itself one of the strongest products of the analysis. We haven't only shown that organization exists — we've shown, progressively, what it is not reducible to among the controls we tested. That turns a significance test into a mechanism question: degree structure, then local geometry, then feature compatibility, then — something else, not yet identified.
🌊 A Commuting Bulk, and a Sparse Exceptional Set
The mean D□ understates what's actually happening in the population. Looking at every eligible object rather than just the average reveals a distribution that is not "generally quite good" — it is sharply two-part.
So the population is better described as a large zero-defect bulk plus a sparse noncommuting remainder, not a single cloud of objects each carrying a small, evenly-distributed error. Under the frozen representation, a very large majority of objects show exact or numerically indistinguishable ST/TS agreement, while a comparatively small population carries essentially all of the disagreement. (It is worth being precise about what "exact" means here: D□ = 0 signals agreement within this routing representation, not a proof that the Navier–Stokes operators themselves commute.)
And the exceptions are not scattered randomly. In Pilot B, both scale branching and time branching carry significantly elevated D□ (p ≈ 0.0002), and the same branching association appears in Pilot A. The corresponding merging associations from Pilot A did not reproduce — so the replicated statement is narrower and, for that reason, stronger:
Dominant Route Compatibility + Branching-Associated Noncommutation
Structural splitting looks like a preferential location where scale and time stop telling the same story. A commutation square stays closed for the overwhelming bulk of objects; when an object branches, the two routes are more likely to disagree. We haven't shown causality, and we haven't shown that a rising D□ predicts an upcoming branch — but we have identified a reproducible location where the transformation grammar loosens.
⚖️ Two Failures That Sharpened the Result
Pilot B did not reproduce every number from Pilot A — and that turned out to be one of the most useful things it did. Two of the eight preregistered criteria, R3 and R7, failed on replication. Neither failure means the association disappeared; each one refines what kind of object we're actually claiming to have found.
R3 — Scalar Admissibility (STRUC-I)
Pilot A's route ladder reached Geometric Persistence (Aκ ≈ 0.972). Pilot B reaches only Structural Boundary (Aκ ≈ 0.852) — while simultaneously reproducing constrained routing, the exceptionally low real D□, and full survival against N1 and N2. Route organization is not the same thing as fixed scalar perturbative robustness. We did not know that after Pilot A alone.
R7 — Physical-Intensity Tail Localization
The frozen test asks whether the top decile of D□ is enriched in enstrophy and dissipation. In Pilot B, more than 90% of objects sit at D□ = 0, so the 90th-percentile cutoff collapses onto zero and selects nearly the whole population — the selector becomes degenerate, and the criterion correctly fails. Alternative descriptive tails — including fixed-fraction and fixed-D□ threshold definitions — still show positive, but weaker, enrichment (≈1.13–1.32×, versus Pilot A's 2.50× / 1.87×).
That last pairing is the point: Pilot B has weaker scalar admissibility while showing complete value-space connectivity and an even stronger real-versus-null separation. That is only visible because Pilot B disagreed with Pilot A on the specific number. Route organization, value-space connectivity, and perturbative admissibility strength are turning out to be three separable structural layers, not three descriptions of the same thing.
Why This Is Good News, Not a Setback
Neither R3 nor R7 were rescued after the fact. No threshold was retuned, no upper tail was redefined once it became inconvenient, and Pilot B remains frozen exactly as observed: Structural Partial Replication. That refusal to patch a disagreeing result is what makes the six criteria that did replicate worth trusting.
🧭 A Hierarchy of Structural Invariance
Put the full R1–R8 record together and a hierarchy falls out on its own: some properties of the original signature are stable across both operationally independent samples, and some are specific to the sample that produced them.
Stable Core — Replicates Strongly
✓ Constrained routing (scale–time compatibility)
✓ Low D□ vs. random (N0)
✓ Survival against N1 and N2
✓ Full value-space connectivity
✓ Branching-associated noncommutation
Variable Periphery — Sample-Dependent
○ Exact STRUC-I regime strength (Geometric Persistence vs. Structural Boundary)
○ Extreme-tail physical-intensity enrichment
○ Merging localization
A turbulent field can be highly irregular in state, yet strongly constrained in how its structural transformations relate. That is a deeper result than "Pilot B reproduced Pilot A." Replication didn't just confirm a package of observations — it factorized the package, and told us which pieces move together and which don't.
♾️ Unpredictability and Constraint Can Coexist
This gives the project a broader physical interpretation, and it's worth stating carefully. A turbulent velocity field can be extraordinarily high-dimensional and locally hard to predict, while still obeying comparatively strong constraints on how transformations of its structural representation relate to each other. Those are not contradictory statements — they describe two different things.
The manuscript calls the resulting object a possible transformation-space skeleton. It sits apart from familiar coherent-structure ideas: exact coherent structures organize states or trajectories, Lagrangian coherent structures organize transport, and this result — if it holds up under broader testing — organizes transformations. The elementary object is no longer only a point, a ladder, or a trajectory; it can be the commutation square itself, and the scientific question becomes how nearly that square closes.
An Orthogonal Axis for Cascade Thinking
Classic cascade analysis asks what changes as scale changes — energy transfer, structure functions, scaling exponents, intermittency. Our question sits next to that one, not in place of it:
The Cascade Question
What happens across scale? A Markov-in-scale treatment asks about transition statistics P(Xℓ2 | Xℓ1) — how a quantity's distribution changes as you move between scales.
The Commutation Question
Is what happens across scale compatible with what happens through time? Our construction compares the distributions PST and PTS directly — operator compatibility as an added coordinate of cascade organization.
That's why this result is adjacent to Kolmogorov scaling, intermittency theory, Germano filtering, and Friedrich–Peinke stochastic scale dynamics, rather than a replacement for any of them.
🔗 A Falsifiable Quantity, Not a Metaphor
"Structural grammar" could sound like a figure of speech. Here it is operational. The adapter constructs 3D rotation-dominated objects using frozen segmentation rules; scale and time relations are generated from measured overlap; two-step routing distributions are built from those relations; their difference is quantified through normalized Jensen–Shannon divergence; and the resulting population is compared against explicitly generated null graphs, not eyeballed.
Grammar, here, means measurable compatibility among transformation routes — not an impression that two structures "look similar." Both heavy source cutouts and the High-Reynolds-number export are archived under a single versioned Zenodo record, alongside exact source hashes, selection coordinates, frozen instrument versions, and canonical result files. An external investigator doesn't need to rediscover the JHTDB acquisition procedure to check this work — they can retrieve the exact byte-identified cutouts, verify them, and run the frozen pipeline themselves. The dashboard above exposes all four layers of that chain: manuscript, analytics, project corpus, and archival source data.
🔭 Where This Can Lead
None of the following are conclusions of the present two-pilot experiment — they are the open questions this result makes it possible to ask precisely, several of them with a natural next step already sitting in the archived High-Reynolds-number export.
Does D□ rise first?
Does D□ climb before a branching event, rather than merely being elevated around it?
Who carries the tail?
Does the sparse noncommuting population disproportionately carry dissipation, enstrophy, or high-order anomalous statistics?
A restricted transition set?
Does the overwhelmingly commuting bulk permit structural evolution to be described through a restricted set of allowed route transitions, rather than full-field prediction?
Does a reduced model preserve this?
Does an LES or reduced model preserve not only spectra and stresses, but the same scale–time transformation compatibility?
Is the hierarchy the constant?
Is the invariant not a universal D□ value but the hierarchy D□(REAL) ≪ D□(N2) < D□(N1) ≪ D□(N0), together with a commuting bulk and branching-localized defect?
Does it hold at higher Reynolds number?
The archived cutouts and High-Re export give a direct route to testing this without redesigning the present result.
The Central Finding
This is not a proof that scale and time commute in the Navier–Stokes equations. It is a reproducible empirical property of one frozen multiscale structural representation — now ready to be challenged at higher Reynolds number, under alternative representations, and with stronger field-level controls.
The Underlying Proposition
A turbulent field can be highly irregular in state, yet strongly constrained in how its structural transformations relate. For UNNS, that reframes the search for invariants: perhaps not primarily in values, but in relations among operations — the strongest invariant need not be a state, it may be a relation between transformations.
Resources & References
-
Primary Manuscript (PDF):
Reproducible Scale–Time Commutation Structure in Forced Isotropic Turbulence
Full derivation of D□, the frozen adapter/routing/null-hierarchy pipeline, and the complete two-pilot R1–R8 record. UNNS Substrate Research Program · Turbulence Branch · 2026. -
Interactive Dashboard:
JHTDB Turbulence Dashboard
One hub for the manuscript, analytics, corpus, and archival data — embedded above. -
JHTDB Turbulence Analytics (Interactive HTML):
Pilot-B Replication — Full Analytical Trace
Fourteen sections: independent-sample audit, primary ROUTE-I result, STRUC-I / STRUC-PERC-I, the N0→N1→N2→real null collapse, R7/R8 divergence detail, and the direct Pilot A ↔ B comparison. -
Data and Corpus Construction:
UNNS_TURB_JHTDB_v0_1.zip
Complete branch archive: route adapter, frozen chamber specs, both pilots' route projects and object tables, and every locked run with its reproducibility hashes. -
Archival Source Data (Reproducibility):
UNNS Turbulence — JHTDB Analysis Cutouts and High-Re Export (DOI 10.5281/zenodo.22650769)
Permanent archival record of the raw velocity cutouts and the high-Reynolds-number export, versioned outside the working repository for independent verification.
Established Turbulence Context
This result sits alongside, not in place of, established multiscale and cascade theory — the filtering, closure, and stochastic-cascade concepts it is adjacent to (see "Unpredictability and Constraint Can Coexist," above) are defined in the following works.
-
JHTDB Forced Isotropic Turbulence Dataset:
isotropic1024coarse, Johns Hopkins Turbulence Database (DOI 10.7281/T1KK98XB)
The source DNS dataset both pilots' cutouts are drawn from. -
Germano, M. (1992).
"Turbulence: the filtering approach." Journal of Fluid Mechanics, 238, 325–336.
Foundational scale-filtering formalism underlying multiscale decomposition of turbulent fields. -
Ghosal, S., & Moin, P. (1995).
"The basic equations for the large eddy simulation of turbulent flows in complex geometry." Journal of Computational Physics, 118(1), 24–37.
Shows explicitly that filtering and differentiation need not commute for variable filter widths, providing an established turbulence precedent for treating operator-order effects as measurable quantities. -
Friedrich, R., & Peinke, J. (1997).
"Description of a turbulent cascade by a Fokker–Planck equation." Physical Review Letters, 78(5), 863–866.
The stochastic scale-dynamics framework referenced in the Markov-in-scale comparison above.