Gain-Controlled Reservoirs Exhibit Chaos Susceptibility as a Structural Constraint on Computation

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

Abstract

Reservoir computing shows that useful computation in recurrent networks depends on the controlled availability of unstable dynamics. Metzner, Schilling, Maier, Kinfe, and Krauss study the failure mode directly: strong recurrent coupling drives reservoirs into chaotic, oscillatory, or fixed-point attractors that erase input-related information. Pudding Theory reads this phenomenon through Chaos Susceptibility. The reservoir is not a passive random substrate with a tunable operating point. It is a susceptibility medium whose computational value is set by how small input signals are amplified before herd dynamics capture the state space. Weak rows and automatic gain control work because they reshape the susceptibility landscape, preserving receptive microchannels inside or across global instability. The relevant invariant is not the raw weight distribution but the distribution of local amplification paths available to recent input. If the input-conditioned susceptibility spectrum were measured to remain unchanged between homogeneous, weak-row, and gain-controlled reservoirs despite the reported accuracy changes, this Postulate would be falsified.

Source Synopsis

Metzner et al. study a practical instability in reservoir computing. A reservoir computer uses a recurrent neural network with fixed random weights and trains only a simple readout layer. The reservoir should transform input history into a rich state trajectory that the readout can linearly exploit. The difficulty is that increasing recurrent coupling, often done to exploit nonlinear dynamics, can produce spontaneous activity that dominates the input.

The source paper builds reservoirs with tanh units and fully connected recurrent weight matrices. The matrix is controlled by coupling strength, excitation-inhibition balance, and density. Density is fixed at one. In weakly coupled reservoirs, the system can remain useful across balance values. Even when neurons enter global oscillatory or fixed-point regimes, small input-related perturbations can still ride on the large activation pattern and be recovered by the readout.

Strong coupling changes the situation. Balanced reservoirs become chaotic. Strongly negative or positive balance produces saturating oscillations or fixed points. In all three cases, performance collapses except near narrow transition regions. The paper connects this to the familiar edge-of-chaos view, where useful computation occurs near the boundary between ordered and chaotic dynamics.

The authors then test two countermeasures. The first preserves the global distribution of weights but redistributes values within the matrix. The strongest improvement comes from assigning weaker-than-average incoming weights to about 20 percent of rows. This creates a subpopulation of neurons partially insulated from global herd effects. The second countermeasure is automatic gain control. A control unit monitors the running root-mean-square reservoir activation and modulates a single global gain factor to hold activity near a setpoint. This stabilizes the reservoir across the full balance range.

A key result is that standard aggregate dynamical measures can look nearly unchanged while accuracy changes sharply. Principal component analysis reveals that weak rows and gain control preserve input-related pulse structure in components that the readout can use. The source therefore identifies structural heterogeneity and dynamic regulation as routes to prevent runaway excitation while keeping useful reservoir dynamics.

Postulate Lens

This reading applies Chaos Susceptibility. Reservoir computers are designed to exploit unstable recurrent dynamics, and the source paper shows that the same instability can either amplify input into useful high-dimensional structure or amplify internal feedback into runaway saturation. The Postulate fits because the studied phenomenon is not merely chaos as failure. It is susceptibility itself as the computational resource, with failure occurring when amplification is captured by self-generated herd modes rather than by recent input.

Pudding Theory Reading

Pudding Theory reads the reservoir as a susceptibility field whose function is to receive a small coherent input and expand it into a macroscopic state trajectory without letting autonomous recurrence seize the amplification channel. The source paper treats the weight distribution as the nominal substrate and asks how its arrangement or gain can prevent runaway excitation. Pudding Theory shifts the object of analysis. The substrate is the map of amplification paths, not the multiset of weights.

In the homogeneous strongly coupled reservoir, the amplification field is undifferentiated. Small differences, whether from input, bias, initial condition, or recurrent feedback, enter the same global susceptibility channel. The network therefore cannot preserve the distinction between signal and self-excitation. The chaotic balanced regime and the saturating unbalanced regimes are two forms of the same capture. In one, microscopic differences expand into irregular spontaneous trajectories. In the other, collective feedback compresses the state into a locked oscillation or fixed point. Both destroy computation because susceptibility no longer belongs to the input.

The weak-row construction is not just a structural trick. It creates a protected receptive subfield. Because those neurons receive weaker recurrent drive, they remain less available to the global herd mode. Their importance is not that they carry smaller weights in isolation. Their importance is that they preserve a low-gain path through a high-gain medium. The source notes that the global weight distribution remains unchanged. Pudding Theory makes this fact central. The distribution is not the governing invariant. The governing invariant is whether the network contains channels whose local Lyapunov susceptibility remains assigned to the incoming signal instead of to autonomous recurrence.

Automatic gain control performs the same operation dynamically. It does not need detailed knowledge of individual synapses because it acts on the global susceptibility scale. By holding mean activity near a receptive setpoint, it prevents the reservoir from crossing into a regime where internal fluctuations dominate the amplification budget. The control unit is therefore not an add-on regulator outside the computation. It is part of the computational field. It continuously reassigns amplification from runaway recurrence back to input-conditioned state formation.

This reading also changes the status of the edge of chaos. The edge is not a privileged point to be tuned once. It is the local condition in which susceptibility remains available to signal. Weak rows spatially distribute that condition. Gain control temporally maintains it. The computational system is successful when the input-conditioned susceptibility spectrum remains structured, even while the raw network statistics would predict collapse.

Falsifiable Observable

The distinguishing observable is the input-conditioned susceptibility spectrum: the singular spectrum of the finite-time Jacobian mapping small perturbations in the current input episode into subsequent reservoir states, measured separately for homogeneous, weak-row, and gain-controlled reservoirs at matched coupling strength and balance. Pudding Theory predicts that accuracy improvements track preservation of intermediate finite-time amplification modes aligned with input pulses, even when global activity measures remain similar. If the input-conditioned susceptibility spectrum were measured to be invariant across homogeneous, weak-row, and gain-controlled reservoirs while accuracy changed from chance to near-perfect performance, this Postulate would be falsified.

Editorial Dialogue

Tanaka: The reading risks renaming known reservoir facts. The source already says weak rows preserve input-related perturbations and gain control stabilizes activity. Why call this a susceptibility field rather than ordinary dynamical-systems language?

Sterling: Because the ordinary language leaves the wrong invariant in view. It keeps returning to the global weight distribution, the balance parameter, and aggregate measures such as fluctuation and covariance. The paper itself shows those quantities can fail to distinguish a useless reservoir from a useful one. The Pudding Theory reading identifies the missing object: the allocation of amplification. A reservoir computes when amplification remains conditioned on input. It fails when amplification is captured by recurrence.

Tanaka: That still sounds like a restatement of finite-time Lyapunov analysis.

Sterling: It is stricter than that. A global Lyapunov number is insufficient. The weak-row case proves the point. Most of the reservoir can be in a saturating herd state while a subpopulation remains computationally alive. The observable must be input-conditioned and mode-resolved. The theory predicts a structured susceptibility spectrum, not just less chaos.

Tanaka: And gain control?

Sterling: Gain control makes the same constraint dynamic. It keeps the susceptibility field receptive without knowing the task labels or individual weights. That is why the performance becomes insensitive to balance.

Discussion

This reading buys a sharper account of why two different interventions work. Weak rows and automatic gain control look unrelated in the source framing: one is structural heterogeneity, the other dynamic regulation. Pudding Theory makes them instances of one constraint. Both prevent amplification from being monopolized by self-generated recurrent modes.

The source’s aggregate measures are useful regime labels, but they are not computational observables. They can miss the input-related structure that matters to the readout. The PCA results point toward the correct level of description: hidden components that pulse with the input episode and survive inside otherwise hostile dynamics. Pudding Theory predicts that these components correspond to preserved susceptibility channels.

The limitation is clear. This reading must be tested with perturbative measurements of finite-time input sensitivity, not only with task accuracy or PCA visualization. It may also depend on task class. Some tasks may exploit saturated binary-like states rather than mild nonlinear receptivity. In those cases, the relevant susceptibility channel could be discrete rather than quasi-linear. The conclusion would change if high performance appeared without any measurable preservation of input-conditioned amplification modes.

References

1. Claus Metzner, Achim Schilling, Andreas Maier, Thomas Kinfe, and Patrick Krauss. “Structural and dynamical strategies to prevent runaway excitation in reservoir computing.” arXiv:2603.29597, 2026. DOI: doi:10.48550/arxiv.2603.29597.

2. S. Ochs. “Pudding Theory: A Topological Theory of Information Fields.” QBist Lab Working Paper, 2026.

3. Herbert Jaeger. “The ‘echo state’ approach to analysing and training recurrent neural networks, with an erratum note.” GMD Technical Report 148, 2001.

4. Claus Metzner, Achim Schilling, Andreas Maier, and Patrick Krauss. “Nonlinear neural dynamics and classification accuracy in reservoir computing.” Neural Computation 37(8):1469-1504, 2025.

5. Claus Metzner, Achim Schilling, Andreas Maier, and Patrick Krauss. “Organizational regularities in recurrent neural networks.” Frontiers in Complex Systems 3:1636222, 2025.

6. Nils Bertschinger and Thomas Natschläger. “Real-time computation at the edge of chaos in recurrent neural networks.” Neural Computation 16(7):1413-1436, 2004.

7. Robert Legenstein and Wolfgang Maass. “Edge of chaos and prediction of computational performance for neural circuit models.” Neural Networks 20(3):323-334, 2007.

8. Joschka Boedecker, Oliver Obst, Joseph T. Lizier, N. Michael Mayer, and Minoru Asada. “Information processing in echo state networks at the edge of chaos.” Theory in Biosciences 131(3):205-213, 2012.