Coherent Perturbations in Chaotic Network Reductions Should Produce Lyapunov-Scaled Residual Power

Authors: Sterling Geisel, QBist Lab; Dr. Hideo Tanaka

Abstract

Li, Wu, Ling, and Tu review dimensionality reduction in dynamical networks as a problem of controlled information loss. A reduced network is useful only when it preserves the macroscopic quantity under study. This Working Paper applies one Postulate of Pudding Theory: Chaos Susceptibility. The claim is not that chaotic systems amplify all errors. That is standard dynamics. The claim is narrower. If a small coherent perturbation is injected with fixed microscopic energy, phase schedule, and coupling channel before reduction, then structured macroscopic residual power should scale with the full system's maximal Lyapunov exponent after ordinary solver and architecture controls. The test compares full and reduced dynamics under matched coherent and incoherent perturbation ensembles. The observable is a preregistered slope between positive Lyapunov exponent and residual spectral power. The paper defines the perturbation protocol, null ensemble, residual formula, and falsifier.

Source Synopsis

Li, Wu, Ling, and Tu synthesize dimensionality reduction methods for high-dimensional dynamical networks. Their review treats a common problem in biology, neuroscience, infrastructure, social dynamics, and power systems. The microscopic graph and state space can be too large to simulate, interpret, or control directly. Reduction seeks a lower-dimensional object that preserves a selected macroscopic behavior.

The source paper organizes the field into structural, analytical, and data-driven approaches. Structural coarse graining reduces the graph. It groups nodes and edges while preserving selected topological, spectral, or dynamical quantities. Spectral coarse graining, including the work of Gfeller and De Los Rios, preserves dominant operator modes relevant to diffusion, synchronization, and relaxation.

Analytical reduction compresses the dynamical state. Watanabe-Strogatz and Ott-Antonsen theory show that special oscillator populations can collapse onto low-dimensional manifolds. These reductions are powerful because they preserve mechanism. Their limits are also clear. They require symmetry, coupling regularity, or special frequency structure.

Data-driven reduction infers latent coordinates from trajectories. Dynamic Mode Decomposition, Koopman methods, Neural ODEs, manifold learning, and scientific machine learning extend reduction to systems without closed-form equations. These methods scale well, but the review stresses the risk of phenomenology. A learned manifold may predict short trajectories while hiding causal variables.

The paper's central lesson is a trade-off. Exact reductions preserve mechanism but need restrictive assumptions. Spectral methods preserve selected invariants but can discard relevant modes. Learned models generalize poorly when trained outside the active regime. The review also emphasizes higher-order interactions, simplicial structure, and hybrid models that embed physical priors.

Postulate Lens

Chaos Susceptibility applies because the source paper studies when microscopic loss does or does not affect macroscopic dynamics. The Postulate states that systems with positive Lyapunov exponents amplify small coherent inputs into observable shifts. That statement only has content here if the coherent input is treated as an intervention, not as ordinary model error.

The proposed intervention is a deterministic, low-amplitude perturbation field applied to a specified subset of microscopic state variables before reduction. Its amplitude, phase schedule, timing, and coupling channel are fixed across networks. The same perturbation energy is injected into stable and unstable systems. The full model is the reference. The reduced model is then tested on the same initial condition and intervention record.

The source paper supplies the setting. It compares reductions that preserve spectra, manifolds, collective variables, or learned latent coordinates. Pudding Theory adds a diagnostic. If discarded coordinates sit near an instability channel, then coherent microscopic bias should not average away. It should appear as structured residual power in macroscopic observables.

The distinction from ordinary Lyapunov error growth is the matched null. Random-phase perturbations with the same energy should produce weaker or spectrally flatter residuals than coherent perturbations when the positive Lyapunov exponent is large.

Pudding Theory Prediction

Consider a family of dynamical networks with matched node count, degree distribution, coupling law, observation window, compression ratio, and noise amplitude. For each full system, estimate the maximal Lyapunov exponent $\lambda_{\max}$ by tangent-space integration or a standard two-trajectory renormalization method over the same time window used for reduction testing. Only estimates with stable convergence under window doubling enter the primary analysis.

The coherent perturbation is defined as \[ \delta x_i(t)=a\,c_i\,\cos(\omega t+\phi)\,u_i , \] applied for $t\in[t_0,t_1]$ to a preregistered microscopic channel $u_i$ on a fixed fraction of nodes. The vector $c_i$ sets the spatial mask. The amplitude $a$ is chosen so that microscopic perturbation energy \[ E_\delta=\int_{t_0}^{t_1}\|\delta x(t)\|_2^2\,dt \] is identical across networks. Phase $\phi$, frequency $\omega$, timing, and mask are held fixed.

The null ensemble uses the same $E_\delta$, mask norm, and timing, but randomizes phase independently across nodes or across short time blocks. It therefore preserves perturbation energy while destroying coherence. A second null uses no injected perturbation and measures baseline reduction error.

For a macroscopic observable $M(t)$, the normalized residual is \[ R=\frac{\int_T |M_{\mathrm{full}}(t)-M_{\mathrm{red}}(t)|^2dt} {\int_T |M_{\mathrm{full}}(t)-\bar{M}_{\mathrm{full}}|^2dt+\epsilon}. \] The structured residual power is the fraction of residual spectrum concentrated at $\omega$ and its first harmonics after subtracting the incoherent null mean.

Pudding Theory predicts a positive slope $\beta$ in \[ P_{\mathrm{struct}}=\alpha+\beta\lambda_{\max}+\Gamma Z+\varepsilon , \] where $Z$ includes solver tolerance, compression ratio, training length, architecture, and noise. The minimum detectable effect is set before analysis, for example $\beta_{\min}$ equal to a 10 percent increase in structured residual power per unit positive $\lambda_{\max}$.

Falsifiable Observable

The observable is the preregistered coefficient $\beta$ relating positive $\lambda_{\max}$ to coherent-minus-incoherent structured residual power across at least three reduction families and two macroscopic observables, with equal microscopic perturbation energy verified in every trial. If coherent-minus-incoherent structured residual power were measured to have $\beta=0$ within a 95 percent confidence interval narrower than $\pm\beta_{\min}$, this Postulate would be falsified.

Editorial Dialogue

Tanaka critique, verbatim:

Chaos Susceptibility is the only postulate doing any work, and even that work is not yet separated from ordinary Lyapunov error growth. What measurement would distinguish this from the null? If the coherent perturbation is not specified as an intervention with amplitude, phase, timing, and coupling channel, the prediction collapses into standard sensitivity to discarded modes. Would removing Pudding Theory change the prediction? Mostly no. A numerical analyst could predict Lyapunov-ordered residuals without Xi, Lumina, or information pressure.

The falsifiability sentence is close to operational. It names beta, residual variance, lambda max, controls, reduction families, and observables. That is better than analogy. But the regression alone is too weak. Define the normalized residual formula. Define how lambda max is estimated from the full dynamics. Define the null ensemble. Define the coherent perturbation protocol. Define the minimum detectable beta. Without those, beta equals zero may mean poor power, poor matching, or a bad Lyapunov estimator. The stronger falsifier is this: inject matched perturbations into stable and unstable networks, verify equal microscopic perturbation energy, then test whether structured macroscopic residual power scales with lambda max after solver and architecture controls.

Citation risk is moderate. The DOI for the source paper resolves at https://doi.org/doi:10.1080/23746149.2026.2645192, but the author list appears incomplete. It includes Zebiao Li, XueYing Wu, Xiaodong Ling, and Chengyi Tu, not only Li, Wu, and Tu. The arXiv identifier should be verified or removed. Gfeller and De Los Rios resolves at https://doi.org/doi:10.1103/PhysRevLett.100.174104. The draft stays within fair use because it paraphrases rather than quotes.

Not yet. Add a perturbation protocol and null ensemble that separate PT-specific coherent bias from ordinary Lyapunov amplification.

Sterling response:

The critique is accepted as a constraint on the final claim. The paper no longer treats residual growth alone as the observable. The intervention now has amplitude, phase, timing, and coupling channel. The null preserves perturbation energy and removes coherence.

The author list has been corrected. The arXiv identifier has been removed because the DOI is sufficient for the source citation here. The falsifier now requires a confidence interval narrower than a preregistered minimum detectable slope. A null result without power is not allowed to masquerade as falsification.

The remaining Pudding Theory content is narrow. It predicts a coherent-minus-incoherent residual effect ordered by measured instability. Without coherence, the prediction reduces to standard numerical analysis. With coherence controlled against an equal-energy null, it becomes a distinct test.

Discussion

This Working Paper does not claim that dimensionality reduction evidence already supports Pudding Theory. The source article is a review. Its value is methodological. It identifies a mature setting in which full models, reduced models, and macroscopic observables can be compared.

The main risk is attribution. Structured residual power can arise from bad closure, insufficient training data, solver error, or a poor reduction architecture. The protocol therefore requires stable and unstable networks, multiple reductions, equal perturbation energy, and a phase-randomized null. It also requires that $\lambda_{\max}$ be estimated from full dynamics rather than inferred from reduction failure.

Exact reductions may suppress the effect. That result would be informative, not fatal, if the unstable channel is represented analytically. The stronger test is across empirical or simulated networks with tunable instability and known microscopic equations.

The conclusion would change if coherent perturbations show no excess structured residual power over equal-energy incoherent perturbations after the stated controls. It would also change if residual power tracks only architecture or training horizon and not measured instability.

References

1. Li, Z., Wu, X., Ling, X., and Tu, C. (2026). Compressing Complexity: A Critical Synthesis of Structural, Analytical, and Data-Driven Dimensionality Reduction in Dynamical Networks. DOI: doi:10.1080/23746149.2026.2645192.

2. Geisel, S. (2025). Pudding Theory: A Topological Theory of Information Fields. QBist Lab Working Paper, September 10, 2025.

3. Gfeller, D., and De Los Rios, P. (2008). “Spectral coarse graining and synchronization in oscillator networks.” Physical Review Letters, 100, 174104. DOI: doi:10.1103/PhysRevLett.100.174104.

4. Watanabe, S., and Strogatz, S. H. (1994). “Constants of motion for superconducting Josephson arrays.” Physica D, 74, 197-253. DOI: doi:10.1016/0167-2789(94)90196-1.

5. Ott, E., and Antonsen, T. M. (2008). “Low dimensional behavior of large systems of globally coupled oscillators.” Chaos, 18, 037113. DOI: doi:10.1063/1.2930766.

6. Gao, J., Barzel, B., and Barabasi, A.-L. (2016). “Universal resilience patterns in complex networks.” Nature, 530, 307-312. DOI: doi:10.1038/nature16948.

7. Pikovsky, A., and Rosenblum, M. (2011). “Dynamics of heterogeneous oscillator ensembles in terms of collective variables.” Physica D, 240, 872-881. DOI: doi:10.1016/j.physd.2011.01.002.

8. Brunton, S. L., and Kutz, J. N. (2022). Data-Driven Science and Engineering: Machine Learning, Dynamical Systems, and Control. Cambridge University Press.