Does Information Flow Converge? Band-Wise Grid Verification for Transfer Entropy

Scientists increasingly read causal stories off simulations: which region of a flow drives another, which receiver in a seismic model "hears" a source first, which patch of a reconstructed cortex leads a seizure. A common tool for this is transfer entropy, a measure of how much the past of one signal improves prediction of another signal's future beyond what that signal's own past already tells you. Simulations, however, are only trustworthy after grid independence verification, the routine check that an answer stops changing as the computational mesh is refined. That check has been formalized for drag, pressure and energy. Nobody has applied it to directed information measures. This matters because transfer entropy is unusually exposed to the grid. Refining the mesh forces a smaller time step, which raises the highest frequency the simulation can represent. A transfer entropy estimate computed on each grid at its native sampling rate therefore sees different frequency content on every grid, and differences that come purely from discretization show up looking like causal structure.

We propose to verify transfer entropy one frequency band at a time. Each grid's output is decomposed into wavelet scales without filtering. The contribution of each band to directed transfer is measured with within-band surrogates. Richardson-extrapolation error bands, the Grid Convergence Index, are then computed per band across three mesh levels. The output is a resolved information bandwidth: the highest frequency band in which directed information transfer has provably stopped depending on the mesh, given the estimator's own statistical noise. The same procedure then serves as an audit of neural surrogate solvers, which are known to lose accuracy at high frequency. A shared-memory implementation that reuses neighbor searches across surrogates keeps the many-grid, many-band, many-surrogate computation feasible on one node.

Core Claim

The claim has two parts. First, transfer entropy computed from mesh-based simulations is not grid-independent by default, even when the usual bulk quantities are. The non-convergence concentrates in frequency bands near each grid's resolution limit. Second, a band-wise convergence test can separate those bands from the ones that are physically meaningful. The test must be filter-free, because filtering itself corrupts transfer entropy, and it must be gated by estimator uncertainty, because transfer entropy estimators are noisy. A single extrapolated transfer entropy number cannot make that separation.

The quantity at the center of the method is, for a band $b$ and grids with spacing $h_1 < h_2 < h_3$ refined by a constant ratio $r$, the band-limited transfer value $T_b(h)$. It is defined as the drop in transfer entropy from source to target when the source's wavelet coefficients in band $b$ are shuffled in time. From this we compute the observed order and the relative convergence index:

$$p_b = \frac{\ln\left|(T_b(h_3)-T_b(h_2))/(T_b(h_2)-T_b(h_1))\right|}{\ln r}, \qquad \mathrm{GCI}_b = \frac{F_s\,|T_b(h_1)-T_b(h_2)|}{(r^{p_b}-1)\,|T_b(h_1)|},$$

where $F_s$ is the conventional safety factor of the Grid Convergence Index. A band is declared converged when $p_b$ falls in the monotone asymptotic range and $\mathrm{GCI}_b$ is smaller than the bootstrap confidence width of $T_b(h_1)$. In that case the remaining mesh error is below what the estimator can resolve anyway. The resolved information bandwidth is the highest band, counted contiguously from low frequency, that passes.

Where This Sits in Prior Work

Two papers supply the spectral machinery. Measuring spectrally-resolved information transfer decomposes signals with the maximum overlap discrete wavelet transform. It shuffles coefficients within one band to test that band's contribution, and it shows that bandpass filtering before estimating transfer entropy distorts phase and creates spurious dependence. We reuse its surrogate construction directly. It was built for single recordings, though, and never asks whether the band-wise result survives a change of discretization. Fourier-domain transfer entropy spectrum builds a time-frequency causality map through wavelet transforms and symbolization. Its dependence on hand-tuned symbolization parameters makes it a poor base for a convergence test, since those parameters would themselves need to be re-tuned per grid. We use it only as a secondary comparison.

Two papers motivate the gap from the verification side. Influence of CFD Modelling Parameters on Air Injection Behaviour in Ship Air Lubrication Systems applies the Grid Convergence Index rigorously and shows that time-step and near-wall resolution choices decide whether high-frequency interfacial instabilities are captured or smeared out. This is exactly the mechanism we expect to contaminate transfer entropy, but that study only verifies conventional outputs. AirfRANS shows that a low global error can coexist with poor prediction of local surface forces. We expect the same dissociation between bulk convergence and convergence of information flow.

Two papers constrain the design from the frequency-bias side. Tuning Frequency Bias of State Space Models shows that gradients with respect to high-frequency parameters decay as $O(|s|^{-2})$, and that pushing a model's frequencies past the sampling limit causes aliasing. This tells us to define bands relative to each grid's resolvable range rather than in absolute terms. Reducing Frequency Bias of Fourier Neural Operators in 3D Seismic Wavefield Simulations Through Multistage Training shows that Fourier neural operators lose accuracy at higher wave frequencies and partly recover it with residual stages. That paper never checks the convergence of its own 64³ reference solution, and never asks whether directed relations between receivers are preserved. Our surrogate audit addresses both.

On the estimator side, TRENTOOL supplies the standard k-nearest-neighbor estimator with embedding selection and surrogate testing, which we adopt. TENDE offers a diffusion-model estimator for high-dimensional cases. We do not use it in the main loop because its training cost multiplies badly across grids, bands and surrogates. The computational design follows the shared-memory pattern of FreeContact and JedAI: independent loop iterations are split across cores over in-memory data, avoiding disk and network serialization.

Research Program

The work has three components.

The first is the verification procedure. For each grid level we record probe signals at fixed physical locations, and we set the embedding delay and history length in physical time, not in samples. Each signal is decomposed at its native rate. The band set is defined so that every band considered lies below the coarsest grid's resolvable frequency, and the extra bands on finer grids are reported separately as unverifiable rather than discarded. The convergence index above is then computed for every band and every source-target pair. Bootstrap over time blocks gives the estimator confidence width used for gating. For pairs where transfer is significant only on the finest grid, the procedure reports "direction not verified." It does not report a value.

The second is a theory-light prediction test. For linear advection-dominated problems the numerical dispersion of a given scheme is known. We predict that the resolved information bandwidth should track the wavelength at which the scheme's dispersion error becomes significant, scaling inversely with $h$. This follows because transfer at a band is carried by waves at that band's wavelength, and those waves travel at the wrong speed once they are under-resolved, which shifts the apparent lag. If the measured bandwidth does not follow this scaling, either the test or the hypothesized mechanism is wrong, and we will know which by inspecting the observed orders $p_b$.

The third is the computational kernel. In the within-band surrogate test for source $X$ to target $Y$, only the source's past changes between surrogates. The target-only neighbor structures needed by the estimator, in the target-past space and the target future-plus-past space, are identical across all surrogates and across all sources sharing that target. We build those structures once per target and grid, hold them read-only in shared memory, and parallelize over surrogates and sources with no locks. Only the source-dependent searches are recomputed. The aim is to make the three-grid, multi-band, hundreds-of-surrogates sweep tractable on a single workstation, and to report thread scaling honestly up to the point where memory bandwidth saturates.

Baselines

The naive reference point is what practitioners do today. Transfer entropy is computed at each grid's native rate with a standard nearest-neighbor estimator from TRENTOOL, and the finest two grids are compared by eye or by a relative difference. It runs on the same probe signals, embedding choices and surrogate counts as our method. Beating it means our procedure flags, in the synthetic ground-truth setting, the pairs and bands where the naive comparison accepts a value that is in fact grid-dependent or directionally wrong.

The second baseline is bulk-output verification: the Grid Convergence Index applied, as in the air lubrication study, to quantities such as kinetic energy, drag, or probe signal amplitude. Any directed analysis on a grid that passes is then accepted. This is the strongest competitor in current practice, because it is what reviewers already accept. We beat it if we exhibit configurations where bulk outputs pass the index while band-wise transfer fails in specific bands, and where the failing bands are the ones our dispersion prediction singles out.

The third and strongest methodological competitor is decimation. Every grid's signals are low-pass filtered and resampled to the coarsest grid's rate, ordinary transfer entropy is computed, and the convergence index is applied to that single number. This equalizes frequency content with a few lines of code. It runs on identical data with the same estimator, history length in physical time, and surrogate budget. We beat it if it either misidentifies direction in the ground-truth cases because of filter-induced phase shifts, as the spectrally-resolved transfer paper predicts, or yields a convergence verdict that is wrong for some band it cannot see.

The fourth is linear: band-wise spectral Granger causality from fitted autoregressive models, put through the same convergence index. It runs with model order selected by the same information criterion on every grid. We beat it in the nonlinear coupling cases, where TRENTOOL reports 100% directional accuracy on synthetic autoregressive systems with quadratic and threshold coupling and a false-positive rate below 5%.

Two ablations isolate the proposal's ingredients. In the first, bands are defined by filtering instead of within-band shuffling, which tests whether filter-free band attribution is what matters. In the second, the confidence-width gate is removed, which tests whether estimator noise would otherwise be mistaken for mesh error or mask it.

There is no published number for grid convergence of transfer entropy itself: no reported result found on Lacuna for transfer entropy grid convergence. For the surrogate audit, the strongest existing method is the multistage Fourier neural operator. On the Salvus 3D seismic data its authors report that the residual stage reduced mean L2 loss by "about 1/3" and raised correlation from "average of 0.89 to 0.96", with a further stage giving "minimal (around 4%)" improvement.

Why not just…?

Why not just verify the bulk outputs and trust the directed analysis on a converged grid? Bulk quantities are dominated by low-frequency, high-energy content, while transfer entropy weights information, not energy, so weak high-frequency components can carry substantial directed transfer. AirfRANS already shows that global error and local force accuracy can dissociate. Experiment 2 tests this directly by searching for configurations where the bulk index passes and band-wise transfer fails. If no such configuration appears across the parameter sweep, the bulk check is sufficient in practice, and we will say so.

Why not just decimate everything to the coarsest rate and verify one number? This is the real threat, because it is simple and addresses the changing-bandwidth problem head on. Its weaknesses are specific. The anti-aliasing filter introduces phase distortion, which the spectrally-resolved transfer paper shows corrupts transfer entropy. It also discards exactly the upper bands whose convergence status we want to report. Experiment 1 settles this on ground-truth problems. If decimation recovers the correct directions and convergence verdicts as often as the band-wise test, we will recommend decimation for the single-number use case and confine our method to the bandwidth question.

Why not just use linear spectral Granger causality, whose convergence is smooth and cheap? For linear dynamics it probably suffices, and we include linear advection cases where we expect it to tie. The directed transfer function study on epilepsy source meshes shows how much of the field already relies on such linear measures. The nonlinear cases, such as Burgers-type steepening and turbulent flow, are where linear measures are known to miss coupling, and the comparison there decides whether the extra cost is justified.

Why not just run the estimator on GPUs instead of engineering shared-memory reuse? The saving we exploit is algorithmic: target-side neighbor structures are reused across surrogates and sources. It would apply on any hardware. The neural estimator option, TENDE, carries a training cost per estimate that multiplies across grids, bands and surrogates. We report wall-clock time against a straightforward per-surrogate recomputation, so the benefit of reuse is measured, not asserted.

Proposed Experiments

Experiment 1 uses ground truth. We simulate one-dimensional linear advection-diffusion and viscous Burgers equations with upstream and downstream probes, where the true direction of transport and its lag are known. Each runs on three grids with a constant refinement ratio and a fixed CFL number, so the time step shrinks with the mesh. We add a coupled logistic lattice of the kind used in the Fourier-domain transfer entropy paper as a nonlinear control. The metrics are directional accuracy, false-positive rate against a null with coupling removed, and agreement of the measured resolved bandwidth with the dispersion-based prediction. The decisive check has two outcomes we care about: the naive method should report spurious or reversed transfer in near-cutoff bands on at least one grid, and our method should mark those bands unverified while correctly verifying the low bands.

Experiment 2 moves to two-dimensional incompressible Navier-Stokes and shallow-water problems from the PDEBench family (PDEBench), regenerated with the benchmark's solvers at three resolutions. Probe networks of tens of points give hundreds of directed pairs. Here we compare the bulk-output index, decimation, linear spectral Granger causality and our band-wise test. We report how often each accepts directed edges that fail to persist under refinement, and the resolved bandwidth as a function of resolution and viscosity.

Experiment 3 audits neural surrogates on the Salvus 3D seismic setting. We train a single-stage and a multistage Fourier neural operator following the multistage paper, and compute band-wise transfer between receiver pairs in the surrogate output and in the reference. We also apply the verification procedure to the reference itself at coarser and finer grids where solver access permits; otherwise we restrict the audit to frequencies verified in Experiment 2's analogous regime. The question is whether the reported L2 and correlation gains translate into preserved directed structure. In particular, we test whether the single-stage model's errors in transfer concentrate in the upper bands, as the frequency-bias results predict.

Experiment 4 measures the computational kernel. We compare the cached shared-memory implementation with per-surrogate recomputation and with a process-pool approach in the style of standard toolkits. The metrics are wall-clock time, speedup versus thread count up to the node's core count, and peak memory, using the Experiment 2 workload. A further check is that the cached and naive implementations agree in estimated transfer to within floating-point tolerance on every pair. The overall result counts as decisive if Experiments 1 and 2 show grid-dependent transfer that the bulk and decimation checks miss, and Experiment 4 shows the verification adds tolerable cost relative to the simulations themselves; for reference, the AirfRANS authors report about 25 minutes on 16 CPU cores per CFD run.