---
title: Attractor Ruins in Neural & Coupled Maps
url: https://www.emergentmind.com/topics/attractor-ruins
type: topic
---

# Attractor Ruins in Neural & Coupled Maps

Searching arXiv for papers on attractor ruins and related chaotic itinerancy.
Attractor ruins, also termed attractor relicts, are dynamical structures obtained when attractors lose strict stability but continue to organize trajectories transiently or over a basin of positive measure. In a continuous neural-network setting, coupling a Hopfield-type attractor network to slow adaptation of intrinsic parameters destabilizes all fixed-point attractors and converts them into transiently visited “ruins,” yielding autonomous latching dynamics rather than convergence to a stationary state [1212.5054]. In a high-dimensional discrete setting, notably globally coupled logistic maps, attractor-ruins are described as Milnor attractors that have been “ruined,” i.e. they have lost strict asymptotic stability while retaining a nonzero measure basin; chaotic itinerancy then appears as motion among such quasi-stable cluster manifolds [2510.00538].

## 1. Definition and conceptual scope

In the standard attractor-network formulation, the dynamics descends in an energy or Lyapunov functional, and stable fixed points are minima of that functional. Once the state enters such an attractor, it remains there. A canonical example is the Hopfield energy
$$
E(s)= -\frac12\sum_{i,j}w_{ij}s_i s_j + \sum_i \theta_i s_i,
$$
or, for graded units and leaky integrators,
$$
E[x] = -\tfrac12\sum_{i,j}(y_i w_{ij}y_j)^2 + \tfrac{\Gamma}{2}\sum_i x_i^2.
$$
The attractor-ruin construction begins from this attractor-based organization but augments each neuron with slow adaptive intrinsic parameters such as thresholds or gains. Because the slow adaptation is driven by an information-theoretic objective that is incompatible with stationary firing-rate distributions, the original fixed points lose strict stability and survive only as transient slow regions in phase space [1212.5054].

In the globally coupled map formulation, attractor-ruins are introduced through invariant cluster manifolds. For a permutation $\sigma \in S_N$,
$$
H_\sigma := \{ x = (x(1),\dots,x(N)) \in [0,1]^N \mid x(i)=x(\sigma(i)) \text{ for all } i=1,\dots,N \}.
$$
If the Lebesgue-measure basin
$$
B(H_\sigma) := \{ x \in [0,1]^N \mid \mathrm{dist}(F^n(x),H_\sigma)\to 0 \text{ as } n\to\infty \}
$$
has positive volume, then $H_\sigma$ is a Milnor attractor. If, under a small change of parameters, $H_\sigma$ ceases to be asymptotically stable but still attracts a positive-measure set of orbits, it is called an attractor-ruin [2510.00538].

A common misconception is to treat an attractor ruin as the disappearance of dynamical organization. In both formulations, the opposite is true: the system remains strongly structured, but the structure no longer supports permanent resting states. This suggests a unifying interpretation in which “ruin” denotes persistence without strict asymptotic trapping.

## 2. Generating-function construction in neural systems

The neural attractor-relict framework is built from two generating functionals. The first is the Hopfield energy, which generates a neural attractor network. The second is polyhomeostatic optimization, an information-theoretic generating functional encoding the information content of neural firing statistics [1212.5054].

A convenient continuous-time formulation uses membrane potentials $x_i(t)$ and a sigmoidal transfer function
$$
y_i = g_i(x_i) = \frac{1}{1+\exp(a_i(b_i-x_i))},
$$
where $a_i$ is the gain and $b_i$ the threshold. Minimizing the attractor-generating functional yields the fast dynamics
$$
\dot x_i = -\Gamma x_i + \sum_{j=1}^N w_{ij} y_j,
\qquad y_i = g_i(x_i).
$$
With fixed $a_i$, $b_i$, and symmetric weights $w_{ij}=w_{ji}$, the fixed points of these equations correspond to minima of the Hopfield energy.

The second functional specifies a target long-term firing-rate distribution $q(y)$ and measures mismatch via the Kullback–Leibler divergence
$$
D_{\rm KL}[p_i||q] = \int_0^1 p_i(y)\,\ln\frac{p_i(y)}{q(y)}\,dy \ge 0,
$$
where
$$
p_i(y) = \lim_{T\to\infty}\tfrac1T\int_0^T \delta(y-y_i(t))\,dt
$$
is the empirical firing-rate distribution of neuron $i$. The target distribution is chosen as
$$
q(y)\propto \exp(\lambda_1 y + \lambda_2 y^2),
$$
which fixes mean and variance. Stochastic gradient descent on $D_{\rm KL}$ with respect to the intrinsic parameters produces the polyhomeostatic adaptation rules
$$
\dot a_i = \epsilon_a\Bigl[\tfrac1{a_i} +(x_i-b_i)\bigl(1-2y_i + (\lambda_1+2\lambda_2 y_i)y_i(1-y_i)\bigr)\Bigr],
$$
$$
\dot b_i = \epsilon_b\Bigl[-a_i\bigl(1-2y_i + (\lambda_1+2\lambda_2 y_i)y_i(1-y_i)\bigr)\Bigr],
$$
with slow rates $\epsilon_a,\epsilon_b \ll 1$.

The technical significance of this construction is that the two functionals do not simply coexist; they compete. The Hopfield functional favors stationary pattern retrieval, whereas the polyhomeostatic functional favors firing statistics close to a broad target distribution. The resulting fast–slow system is therefore organized by incompatible optimization pressures.

## 3. Fast–slow destabilization and latching dynamics

Collecting the two generating functionals yields the coupled system
$$
\dot x_i = -\Gamma x_i + \sum_j w_{ij}y_j,\qquad
y_i = \frac{1}{1+e^{a_i(b_i-x_i)}},
$$
together with the slow adaptation equations for $a_i$ and $b_i$. Because stationary firing patterns $y_i(t)=\mathrm{const}$ are highly non-entropic, the Kullback–Leibler objective remains large and the adaptation gradients do not vanish. Consequently, no fixed point of the fast Hopfield subsystem survives indefinitely; every original attractor is continuously “chased away” by the slow evolution of intrinsic parameters [1212.5054].

The resulting dynamics is latching. Former attractors persist only as transient slowdowns, and the trajectory moves from one such ruin to another. Autonomous dynamics involving sequences of transiently stable states have been termed associative latching in the context of grammar generation, and the same mechanism is proposed as relevant for motor control and associative neural computation in the brain. In larger Hopfield-encoded networks with $N \gg 1$ and stored patterns $\xi^p$, each encoded pattern becomes an unstable slow region in phase space, and the trajectory visits these regions sequentially.

A minimal three-neuron example makes the mechanism explicit. For a symmetric network with one central neuron excitatory to two side neurons and mutual inhibition between the side neurons, the fixed-parameter system has a two-attractor region in the $(a,b)$ plane. Its lower boundary is a supercritical, second-order bifurcation,
$$
1 = a\,\tilde y(1-\tilde y),
$$
where $\tilde y$ is the firing rate of the active side-sites. Since $\tilde y(1-\tilde y)\le \tfrac14$, the critical gain is $a_c=4$. The upper boundary is a fold, or first-order, bifurcation. Under slow adaptation, the trajectories of $(a_i(t),b_i(t))$ numerically “sit” near the second-order boundary and repeatedly cross it, alternately destabilizing and re-stabilizing the former attractors.

This phase-boundary picture is important because it identifies attractor ruins not as an ad hoc perturbation effect but as a systematically maintained near-bifurcation regime.

## 4. Objective-function stress and temporal organization

The neural construction distinguishes two forms of mismatch between the generating functionals. The first is functional stress: Hopfield attractors imply sharply peaked firing-rate distributions, whereas polyhomeostatic adaptation demands a broad target distribution. This irreconcilable aim destroys the fixed points. The second is scalar, or mean-rate, stress: the Hopfield patterns have mean activity
$$
\alpha = \frac1N \sum_i \xi_i^p,
$$
while the target distribution has mean
$$
\mu = \int_0^1 y\,q(y)\,dy.
$$
When $\mu \ne \alpha$, the network must either reduce or boost its firing rate relative to the stored-pattern level in order to optimize $D_{\rm KL}$ [1212.5054].

If $\mu=\alpha$, objective-function stress is absent at the level of mean activity. The slow adaptation then drives the system onto a near-periodic cycle that visits every attractor relict in turn, yielding regular latching with constant burst intervals. If $\mu\ne\alpha$, the dynamics separates into long laminar phases far from any attractor relict and short bursting episodes of rapid latching among attractor relicts. For $\mu<\alpha$, these bursts transiently raise the average activity; for $\mu>\alpha$, they lower it. The observed regime is intermittent bursting latching with irregular intervals.

This distinction is not merely phenomenological. It ties a precise scalar mismatch between target statistics and encoded patterns to a qualitative change in temporal organization. A plausible implication is that attractor ruins can act as mediators between memory structure and homeostatic constraints, rather than as residual artifacts of destabilization.

## 5. Attractor-ruins in globally coupled maps

In the globally coupled map setting, the system is the $N$-dimensional map
$$
x_{n+1}(i) = (1-\epsilon)\,f(x_n(i)) + (\epsilon/N)\sum_{j=1}^N f(x_n(j)),\qquad i=1,\dots,N,
$$
with
$$
f(x)=\alpha x(1-x),\qquad x_n\in[0,1]^N,\qquad (\alpha,\epsilon)\in[0,4]\times[0,1].
$$
For fixed $(\alpha,\epsilon)$, orbits often collapse into clusters defined by some permutation $\sigma$, meaning that they approach one of the subspaces $H_\sigma$. These cluster manifolds act as attractors, or Milnor attractors, for many initial conditions [2510.00538].

The long-time behavior exhibits four characteristic phases in the $(\alpha,\epsilon)$ plane: coherent, ordered, partially ordered, and turbulent. In the coherent phase, all components synchronize into a single cluster of dimension $1$. In the ordered phase, a small finite number of clusters forms and remains fixed. In the turbulent phase, the components remain desynchronized. The partially ordered phase is the regime in which the orbit does not settle on any single $H_\sigma$ but wanders intermittently among a family of cluster manifolds.

It is precisely this partially ordered region that displays chaotic itinerancy. Orbits spend long epochs near one cluster manifold, undergo bursts of desynchronization, then settle near another cluster manifold. In this formulation, chaotic itinerancy is described directly as itineration among attractor-ruins. The characterization differs from the neural fast–slow construction, but the phenomenology is closely aligned: transient residence, loss of strict stability, and recurrent migration among structured states.

## 6. Quantification, simulation regimes, and interpretive issues

The globally coupled map study introduces two complementary quantitative measures. The first is the optimal-transport distance (OTD) between consecutive clustering patterns. At each time $n$, a clustering pattern $(N_1,\dots,N_k)$ with $\sum_{j=1}^k N_j=N$ is mapped to a discrete distribution $p=(p_1,\dots,p_N)\in\Sigma_N$ via
$$
p_i = (1/N)\sum_{j: N_j=i} N_j,
$$
so that $p_i$ is the fraction of points that form clusters of size $i$. Given distributions $p^n$ and $p^{n+1}$ and a cost matrix $C_{ij}=|i-j|$, the OTD is
$$
\mathrm{OTD}(p^n,p^{n+1}) = \min_{P\ge 0}\sum_{i,j} C_{ij}P_{ij}
$$
subject to the transport constraints
$$
\sum_j P_{ij}=p_i^n,\qquad \sum_i P_{ij}=p_j^{n+1}.
$$
When the clustering pattern has converged, OTD is identically zero after a transient; in the partially ordered phase it continues to fluctuate. The second measure is the Shannon-entropy strength of attractor-ruins,
$$
H=-\sum_d q_d\log q_d,
$$
where $q_d$ is the empirical distribution of effective dimensions $d_n=\dim H_\sigma$ at times satisfying $\mathrm{OTD}(p^n,p^{n+1})=0$. Higher $H$ indicates that more types of cluster manifolds are visited, and more evenly visited, during itinerant motion. Both $\langle \mathrm{OTD}\rangle$ and $H$ are approximately zero in the coherent, ordered, and turbulent phases, and both rise in the partially ordered region, with $\langle \mathrm{OTD}\rangle$ peaking in “Partially Ordered II” [2510.00538].

The neural study reports representative simulation regimes with $N=100\ldots 10^5$ and $N_p$ up to $20$, leak rate $\Gamma=1$, Hopfield weights
$$
w_{ij} = \frac{1}{\alpha(N-1)}\sum_{p=1}^{N_p}(\xi_i^p-\alpha)(\xi_j^p-\alpha),
$$
pattern sparseness $\alpha\in[0.1,0.5]$, and $N_p\lesssim 0.2N$ to avoid severe interference. Typical adaptation rates are $\epsilon_b\simeq 0.01$ and $\epsilon_a\simeq 0.1\ldots 1$, and numerical integration uses 4th-order Runge–Kutta with $\Delta t=0.1$. For $\mu=\alpha$, overlap traces $O_p(t)$ or $A_p(t)$ show a clean, regular latching cycle through all patterns; for $\mu<\alpha$, long low-activity laminar epochs are interrupted by fast bursts of pattern visits [1212.5054].

Two interpretive cautions follow directly from these results. First, attractor ruins are not synonymous with noise-driven switching: in the neural construction they are generated by deterministic coupling of fast attractor dynamics to slow polyhomeostatic adaptation, and in the globally coupled map they are organized by cluster-manifold geometry and parameter-dependent loss of asymptotic stability. Second, attractor ruins are not equivalent to generic disorder. In the neural case, trajectories remain shaped by encoded Hopfield patterns; in the globally coupled map, coherent, ordered, and turbulent phases are all distinguished from the partially ordered itinerant regime by the behavior of OTD and $H$. The available evidence therefore supports a narrow technical usage: an attractor ruin is a residual organizing structure that persists after strict attractor stability has been lost.

Source: https://www.emergentmind.com/topics/attractor-ruins