Behavioral State Decay: Key Dynamics

• Behavioral State Decay is the progressive loss of a system's ability to sustain meaningful behavioral states, measured through various approaches including autocorrelation decay and metastable basin residence times.
• Methodologies encompass analyses from EEG long-range temporal correlations, survival analysis in state-space clustering, to Lyapunov exponents in dynamical systems, each revealing distinct decay signatures.
• The implications extend from understanding cognitive impairments in sleep deprivation to enhancing memory mechanisms in state-space language models, highlighting practical insights across neuroscience and machine learning.

Searching arXiv for the cited works and topic coverage on behavioral state decay. to=arxiv_search.search 娱乐彩票注册 天天中彩票软件_json {"query":"\"Behavioral State Decay\" OR \"long-range temporal correlations\" sleep deprivation OR metastable geometry of behavior OR Mamba primacy recency OR C. elegans continuous complexity OR non-Markovian animal behavior", "max_results": 10} to=arxiv_search.search 彩神争霸怎么样ెన്റ് quotelev_json {"query":"(Meisel et al., 2017)", "max_results": 5} to=arxiv_search.search 在天天中彩票json {"query":"(Kaur et al., 22 May 2026)", "max_results": 5} to=arxiv_search.search 天天中彩票开奖json {"query":"(Ahamed et al., 2019, Airlangga et al., 18 Jun 2025, Margolis et al., 2023, Alba et al., 2020, Lee et al., 26 Jun 2025)", "max_results": 10} Behavioral state decay denotes the progressive erosion of a system’s ability to remain in, transmit, or reconstruct a behaviorally meaningful state over time. Across the cited literature, it is operationalized as diminished long-range temporal correlations in cortical EEG during sustained wakefulness, loss of residence within metastable behavioral basins, exponential contraction of perturbations toward stereotyped locomotor motifs, attenuation of past information in recurrent state-space LLMs, smoothly decaying reorientation hazards that can mimic abrupt state changes, slow long-tailed reduction of correlations in compressed fly behavior, and net decline of user and addicted states in adaptive social networks (Meisel et al., 2017, Kaur et al., 22 May 2026, Ahamed et al., 2019, Airlangga et al., 18 Jun 2025, Margolis et al., 2023, Alba et al., 2020, Lee et al., 26 Jun 2025). The term therefore has no single universal observable; its mathematical form depends on the state representation and on whether decay is defined in terms of correlations, residence times, contraction rates, memory kernels, hazards, or population flows.

1. Operationalizations and observables

Domain	What decays	Primary observables
Human EEG during sleep deprivation (Meisel et al., 2017)	Long-range temporal correlations and intrinsic cortical timescales	Autocorrelation decay, DFA exponent, PSD slope
Multi-timescale animal behavior (Kaur et al., 22 May 2026)	Residence within metastable basins	Survival $S(t)$, hazard $h(t)$, transfer-operator spectrum
C. elegans posture dynamics (Ahamed et al., 2019)	Perturbations away from cyclic locomotor motifs	Negative Lyapunov exponents, phase-space contraction
Mamba state-space LLMs (Airlangga et al., 18 Jun 2025)	Contribution of earlier inputs to current recurrent state	Memory kernel, primacy/recency profile, $\Delta_t$ gating
Off-food C. elegans foraging (Margolis et al., 2023)	Reorientation rate after food removal	Decaying hazard $\lambda(t)$, changepoint statistics
Freely walking flies (Alba et al., 2020)	Correlations in compressed behavioral sequence	Return probability, mutual information, Ising couplings
Adaptive drug-use networks (Lee et al., 26 Jun 2025)	User and addicted prevalence	ODE flows, cessation/initiation rates, rewiring-driven relaxation

These operationalizations separate at least three distinct questions. First, decay can refer to shortening of temporal memory in a signal, as in EEG autocorrelation and DFA. Second, it can refer to escape from metastable regions of a state space, where the central objects are dwell-time distributions, hazards, and spectral gaps. Third, it can refer to net contraction or attenuation in a dynamical system, whether in the Lyapunov sense, in the recurrence kernel of a learned model, or in the compartmental flow of a population process. A recurrent theme is that the choice of state representation largely determines whether decay appears exponential, heavy-tailed, non-Markovian, or apparently abrupt.

2. Cortical timescale shortening under sustained wakefulness

In human neurophysiology, behavioral state decay is defined as the progressive loss of the brain’s capacity to sustain coordinated, temporally integrated activity patterns during prolonged wakefulness. In a 40-hour sleep deprivation protocol, eight healthy young right-handed males (mean age $23.0 \pm 0.46$ years, SEM) underwent 14 waking EEG sessions every three hours starting at 07:00, plus one waking EEG session after a night of recovery sleep. Waking EEG was recorded from 27 scalp derivations, re-referenced to the average reference, sampled at 256 Hz, and analyzed in artifact-free 20-second eyes-open segments. Bandpassed signals were studied in theta (4–8 Hz), alpha (8–12 Hz), and beta/low gamma (12–30 Hz), with the amplitude envelope $x(t)$ obtained as the absolute value of the Hilbert transform of the bandpassed signal (Meisel et al., 2017).

Two complementary measures were used to quantify long-range temporal correlations (LRTCs). For the amplitude envelope, the normalized autocorrelation was defined as

$$R(\tau) = \frac{\mathbb{E}[(x_t)(x_{t+\tau})]}{\mathbb{E}[x_t^2]}.$$

Decay was quantified by the lag-1 autocorrelation and by the first lag at which autocorrelation fell to $\le 0.5$. Detrended fluctuation analysis (DFA) was then applied by forming the profile

$y(k) = \sum_{i=1}^{k} (x_i - \bar{x}),$

computing the root-mean-square fluctuation $F(n)$ after polynomial detrending in windows of size $h(t)$0, and fitting the scaling relation $h(t)$1. Because bandpass filtering inflates $h(t)$2 at short window sizes, white-noise simulations were used to identify the filter-induced “kink,” and the fit was restricted to 2–16 seconds (512–4096 samples). After showing similar trends for detrending orders $h(t)$3, the core analyses used 4th-order detrending (DFA-4).

The central empirical result was a decline of LRTCs as sleep deprivation progressed, but only after changes in signal power were brought under control. In the alpha band, which showed no net power increase across deprivation (linear regression $h(t)$4, $h(t)$5; early vs late power $h(t)$6), autocorrelation functions decayed faster at the end than at the beginning of deprivation. The lag-1 autocorrelation decreased progressively across sessions and increased again after recovery sleep; the alternative metric based on the first lag at which autocorrelation was $h(t)$7 shortened similarly. Alpha-band DFA scaling exponents, typically $h(t)$8–1 at baseline, also decreased with time awake, with significance for detrending orders $h(t)$9 and persistence after stable-power channel selection (87/189 channels remained after control, $\Delta_t$0). By contrast, theta and beta power both increased with time awake (theta: $\Delta_t$1, $\Delta_t$2; beta: $\Delta_t$3, $\Delta_t$4), and the apparent increases in DFA exponents in those bands vanished when analyses were restricted to channels without significant power change (theta: 7/189 channels; beta: 31/189 channels; no significant DFA increase, $\Delta_t$5).

The spatial pattern of decline was broad but most prominent over frontal and parieto-occipital regions. In addition, the power spectral density became more shallow after sleep deprivation and reversed after recovery sleep, consistent with the relation $\Delta_t$6 for scale-free processes. The interpretation advanced in the study is that sustained wakefulness systematically moves cortical dynamics away from near-criticality. Because LRTCs are linked to decision-making and working memory and are regarded as hallmarks of systems near criticality, their decline implies shorter intrinsic timescales, weaker long-range memory, reduced temporal integration, and a mechanistic route from extended wakefulness to cognitive impairment. The study also identifies important confounds: LRTC estimates are sensitive to power and signal-to-noise ratio, circadian modulation remains visible in alpha-band DFA, and the small, all-male sample limits generalizability.

3. Metastability, heavy tails, and slow behavioral basins

A distinct formulation treats behavioral state decay as gradual loss of residence within long-lived metastable basins. In this framework, a basin is a region of effective state space in which the system spends long contiguous epochs, with slow escape pathways producing a spectral gap between slow and fast modes. If $\Delta_t$7 is the dwell time, then the survival probability is $\Delta_t$8, the density is $\Delta_t$9, and the hazard is $\lambda(t)$0. Exponential decay corresponds to constant hazard, $\lambda(t)$1 and $\lambda(t)$2, whereas heavy-tailed decay is slower than exponential and includes power laws, stretched exponentials, and truncated power laws. Mixtures of exponentials,

$\lambda(t)$3

produce decreasing hazards and are identified as residence-time signatures of multi-timescale modulation (Kaur et al., 22 May 2026).

The methodological advance in this literature is to treat timescale itself as an explicit coordinate of the state representation. Each observed time series is transformed by a complex Morlet continuous wavelet into amplitudes $\lambda(t)$4 across dyadically spaced frequency bands; PCA above shuffled baselines and short delay embedding then define a time–frequency state space in which fast motion and slow modulation coexist. Clustering this space into $\lambda(t)$5 states and counting transitions at lag $\lambda(t)$6 yields a row-stochastic transfer operator $\lambda(t)$7. Its leading non-trivial eigenvectors capture slow collective modes, with implied timescales

$\lambda(t)$8

Because behavioral dynamics are typically non-reversible, metastable basins are extracted with Generalized PCCA (G-PCCA), a Schur-based extension of PCCA+ that does not assume detailed balance. Diagnostics include a ratio gap after $\lambda(t)$9 modes, high participation ratios, simplex-like “arms-and-hub” geometry in slow eigenvector space, and held-out predictive information that plateaus at the chosen number of basins.

Three empirical systems illustrate this definition. In a synthetic stochastically driven Lorenz system, the multi-timescale operator recovered a hidden bistable driver across mean dwell times from approximately 7.5 seconds to 80 minutes using 25 frequency channels and delay embedding $23.0 \pm 0.46$0; a single eigenvalue remained near 1 across lag, and the slow eigenvector $23.0 \pm 0.46$1 tracked the hidden variable with Pearson $23.0 \pm 0.46$2–0.92. A fixed-timescale delay embedding of the raw trajectory failed, with all eigenvalues decaying to zero and $23.0 \pm 0.46$3. In C. elegans locomotion, 12 worms recorded for 35 minutes at 16 Hz were represented by 5 eigenworm coefficients, wavelet-decomposed over 0.1–8 Hz with 25 channels, reduced to 4 PCA components, embedded with $23.0 \pm 0.46$4, clustered into $23.0 \pm 0.46$5 states, and analyzed at lag $23.0 \pm 0.46$6 s. The largest ratio gap, $23.0 \pm 0.46$7, selected $23.0 \pm 0.46$8 basins corresponding to run and pirouette, with strong non-reversibility (approximately 89% flux violating detailed balance). Pooled residence-time complementary cumulative distribution functions were best fit by truncated power laws, with pirouette $23.0 \pm 0.46$9 and cutoff $x(t)$0 s, and run $x(t)$1 with cutoff approximately 38 s. In freely moving D. melanogaster, 30 flies recorded for 1 hour at 100 Hz were represented by 22 joint-angle time series, wavelet-transformed over 1–50 Hz with 25 channels, reduced to 15 PCA components, embedded with $x(t)$2, clustered into $x(t)$3 states, and analyzed at lag $x(t)$4 s. The largest spectral gap, $x(t)$5, selected $x(t)$6 basins. In $x(t)$7 space, clusters formed four linear arms from a stationary-weighted hub, corresponding to Idle/Slow, Anterior Movements, Posterior/Wing Movements, and Locomotion. Apparent decay rate $x(t)$8 decreased by approximately $x(t)$9 from 0.01 s to 200 s, and all four basins exhibited truncated-power-law dwell-time distributions with exponents $$R(\tau) = \frac{\mathbb{E}[(x_t)(x_{t+\tau})]}{\mathbb{E}[x_t^2]}.$$0–1.7 and cutoffs $$R(\tau) = \frac{\mathbb{E}[(x_t)(x_{t+\tau})]}{\mathbb{E}[x_t^2]}.$$1–303 s.

This framework changes the meaning of decay from “loss of state” to “loss of residence.” It also places strong constraints on interpretation. Fixed-timescale analyses can miss slow modes when fast fluctuations dominate delay-embedded geometry. Heavy-tailed residence times imply slow latent drivers or multi-timescale modulation of escape barriers, but heavy tails alone cannot distinguish slow barrier modulation from a renewal process with broad dwell marginals. In non-reversible systems, theoretical simplex guarantees are weaker than in reversible chains and must be validated empirically.

4. Lyapunov contraction and relaxation toward locomotor motifs

In high-resolution C. elegans posture dynamics, behavioral state decay is defined not through residence times but through contraction along stable dynamical directions. Posture was parameterized by the first five “eigenworms,” the leading singular vectors of the body-angle matrix $$R(\tau) = \frac{\mathbb{E}[(x_t)(x_{t+\tau})]}{\mathbb{E}[x_t^2]}.$$2, yielding a multivariate observation time series $$R(\tau) = \frac{\mathbb{E}[(x_t)(x_{t+\tau})]}{\mathbb{E}[x_t^2]}.$$3 sampled at 16 Hz. Short-time dynamics were encoded by stacking $$R(\tau) = \frac{\mathbb{E}[(x_t)(x_{t+\tau})]}{\mathbb{E}[x_t^2]}.$$4 contiguous frames into a delay-space representation,

$$R(\tau) = \frac{\mathbb{E}[(x_t)(x_{t+\tau})]}{\mathbb{E}[x_t^2]}.$$5

then projecting to an $$R(\tau) = \frac{\mathbb{E}[(x_t)(x_{t+\tau})]}{\mathbb{E}[x_t^2]}.$$6-dimensional state space using SVD followed by ICA,

$$R(\tau) = \frac{\mathbb{E}[(x_t)(x_{t+\tau})]}{\mathbb{E}[x_t^2]}.$$7

The embedding parameters were chosen by maximizing the predictability measure

$$R(\tau) = \frac{\mathbb{E}[(x_t)(x_{t+\tau})]}{\mathbb{E}[x_t^2]}.$$8

where $$R(\tau) = \frac{\mathbb{E}[(x_t)(x_{t+\tau})]}{\mathbb{E}[x_t^2]}.$$9 is nearest-neighbor prediction error and $\le 0.5$0 is the saturation error. For a representative worm, $\le 0.5$1 peaked at $\le 0.5$2 frames, approximately 0.75 s, and saturated at $\le 0.5$3; in the ensemble embedding across 12 worms, $\le 0.5$4 and $\le 0.5$5 (Ahamed et al., 2019).

The resulting seven ICA modes grouped into three families: forward locomotion $\le 0.5$6, reversal/backward $\le 0.5$7, and turning $\le 0.5$8. Projected trajectories formed cyclic bands and spirals: a circular ring for forward waves, spiral-out/in cycles for reversals, and large transient loops for deep bends. Against this geometric background, Lyapunov exponents

$\le 0.5$9

quantified expansion and contraction. Positive exponents indicated instability and variability; negative exponents indicated exponential decay of perturbations. The maximal Lyapunov exponent was $y(k) = \sum_{i=1}^{k} (x_i - \bar{x}),$0. In the 7D embedding, bootstrapped estimates yielded $y(k) = \sum_{i=1}^{k} (x_i - \bar{x}),$1, $y(k) = \sum_{i=1}^{k} (x_i - \bar{x}),$2, and $y(k) = \sum_{i=1}^{k} (x_i - \bar{x}),$3, with total phase-space contraction

$y(k) = \sum_{i=1}^{k} (x_i - \bar{x}),$4

The Kolmogorov–Sinai entropy rate satisfied $y(k) = \sum_{i=1}^{k} (x_i - \bar{x}),$5 nats/s, and the Kaplan–Yorke dimension was $y(k) = \sum_{i=1}^{k} (x_i - \bar{x}),$6.

The key structural result was conjugate pairing,

$y(k) = \sum_{i=1}^{k} (x_i - \bar{x}),$7

with symmetry about $y(k) = \sum_{i=1}^{k} (x_i - \bar{x}),$8. This links variability and decay in a damped, driven Hamiltonian picture: unstable directions generate divergence, while conjugate stable directions contract perturbations back toward cyclic manifolds. Using the measured positive exponents and $y(k) = \sum_{i=1}^{k} (x_i - \bar{x}),$9, the inferred negative rates were approximately $F(n)$0, $F(n)$1, $F(n)$2, and $F(n)$3, corresponding to decay times of roughly 1.1 s, 1.8 s, 3.1 s, and 7.4 s. These multiple timescales match the observed rapid re-centering of forward and reversal cycles and slower relaxation following large turning excursions.

The same state space also showed deterministic chaos. Unstable periodic orbits (UPOs) were identified by close recurrences and local Jacobians; their count peaked at integer multiples of the minimal body-wave period, and their maximal Floquet exponents were predominantly positive. Period-1 UPOs corresponded to forward and backward waves, whereas longer UPOs encoded composites such as reversal, deep bend, and forward re-entry. In this account, behavioral state decay is the relaxation of posture perturbations toward a structured set of cyclic attractors, not the disappearance of state per se. Variability and stereotypy are therefore co-generated: positive exponents ensure that successive cycles are not identical, while negative exponents keep trajectories organized around reproducible locomotor motifs.

5. Exponential forgetting in selective state-space LLMs

In state-space LLMs, behavioral state decay refers to attenuation of earlier inputs in the recurrent state. For Mamba, the per-dimension selective state-space recurrence is

$F(n)$4

When unrolled, the contribution of input $F(n)$5 to state $F(n)$6 is weighted by the product $F(n)$7, which decays approximately exponentially with lag because $F(n)$8 is diagonal and $F(n)$9 along most channels. In the continuous-time view, discretization is governed by an input-dependent step size $h(t)$00, computed empirically from $h(t)$01 via a low-rank projection and softplus nonlinearity, jointly modulating both forgetting and input injection. The behavioral signature in structured recall is a robust U-shaped accuracy profile over positions: primacy for early tokens, recency for very recent tokens, and lost-in-the-middle for intermediate items, consistently across $h(t)$02 (Airlangga et al., 18 Jun 2025).

The first mechanistic component is long-term memory via sparse channels. The paper defines a per-channel memory coefficient over a context of length $h(t)$03,

$h(t)$04

and a thresholded long-term memory probability

$h(t)$05

Channels with $h(t)$06 are labeled long-term-memory channels. In Falcon Mamba 7B, these channels clustered in particular layers, with Layer 17 standing out at $h(t)$07 and $h(t)$08. Causal ablation was then performed by zeroing $h(t)$09 on identified channels at the first triplet. This sharply reduced first-position recall, whereas ablation of a matched number of random channels had negligible effect. The effect strengthened with longer sequences and stricter selection criteria such as $h(t)$10. A complementary initialization experiment showed that uniform $h(t)$11 initialization at Layer 31 mitigated lost-in-the-middle and flattened the U-shaped recall curve.

The second component is short-term memory via delta-modulated recurrence. Because recent inputs have undergone fewer multiplicative recurrences, they are least decayed. Yet this recency advantage is fragile under interference. Inserting $h(t)$12 random distractor tokens between list and query degraded recall at all positions and collapsed recency first; qualitative breakdown was pronounced for $h(t)$13 and $h(t)$14. This establishes a finite short-term memory depth: recency persists only while the recurrent state is not saturated by intervening updates.

The third component is dynamic memory allocation modulated by semantic regularity. Periodic input sequences were constructed by repeating a token every $h(t)$15 positions. Longer repetition periods increased average $h(t)$16 across layers and channels, while deeper layers amplified $h(t)$17, especially for lower-frequency inputs. At the level of contribution kernels, frequent repetition weakened the influence of earlier positions, and repeated relations produced a sharper U-shaped accuracy curve than random relations. The interpretation is that repeated relations bias $h(t)$18 toward regimes that increase forgetting and input injection, thereby accentuating lost-in-the-middle. The comparison between Falcon Mamba 7B and Mamba 1.4B further showed that long-term memory is more localized and more targetable in the larger model: in 7B, top-1-layer intervention with $h(t)$19 and $h(t)$20 strongly reduced first-position recall, whereas in 1.4B broader interventions over top-3 layers with $h(t)$21 were needed and effects weakened for $h(t)$22.

This formulation makes decay an architectural property of exponential recurrence rather than a pathology. It also identifies a design tension: $h(t)$23 jointly controls forgetting and input responsiveness. The paper therefore proposes that decoupling $h(t)$24 from $h(t)$25, or adding a separate injection gate, could reduce lost-in-the-middle without sacrificing recency. At the same time, the evidence comes from synthetic structured recall, so extension to question answering, summarization, or code remains an open empirical issue.

6. Null models and long-memory alternatives to discrete switching

One recurrent controversy is whether abrupt behavioral traces imply abrupt latent state transitions. Off food, C. elegans typically shows an initial local-search phase with frequent reorientations followed by a global-search phase with sparse reorientations. Population-average reorientation rate, computed in rolling 2-minute windows, is well fit by

$h(t)$26

with $h(t)$27 and $h(t)$28, implying a half-time $h(t)$29. The proposed null model introduces a decaying latent signaling factor $h(t)$30 satisfying $h(t)$31 and generates reorientation events $h(t)$32 with rate $h(t)$33, using Gillespie sampling with $h(t)$34 and 45-minute trajectories. Under the equivalent nonhomogeneous Poisson process, hazard is $h(t)$35, cumulative intensity is $h(t)$36, and survival is $h(t)$37. Despite the absence of discrete switching, the model reproduces both abrupt-looking and gradual individual trajectories. In the experimental dataset of $h(t)$38 worms, abruptness was operationalized by MATLAB’s findchangepts as two local linear regressions on each reorientation-rate trajectory, with slope difference $h(t)$39 and a “decision time” defined by the intersection. The distributions of slope differences and transition times were continuous, and the model reproduced them qualitatively. Sudden changes also occurred with constant rate $h(t)$40, showing that apparent switches can arise from stochastic event clustering alone (Margolis et al., 2023).

A different challenge to simple state-switching comes from freely walking flies. Behavior was first partitioned into $h(t)$41 stereotyped states, extracted from a low-dimensional embedding of short posture/motion snippets. Across 59 flies, each hour-long recording yielded on the order of $h(t)$42 transitions. Transition-matrix tests, return probabilities, and mutual information showed strong non-Markovianity: a matched one-step Markov model predicted memory loss by about 30 transitions, whereas empirical persistence extended to approximately 1000 transitions. To compress this large alphabet while preserving temporal structure, a generalized information bottleneck was introduced,

$h(t)$43

and optimized at $h(t)$44 with $h(t)$45 and $h(t)$46, producing a deterministic binary mapping $h(t)$47. The resulting autocorrelation,

$h(t)$48

had a short-lag exponential guide-to-the-eye with correlation time $h(t)$49 transitions, but remained significantly nonzero out to $h(t)$50. A pairwise maximum-entropy Ising model,

$h(t)$51

matched the empirical mean and correlations, with inferred couplings approaching $h(t)$52 over more than two decades and a one-parameter conditional prediction yielding $h(t)$53. The inferred field was near zero, $h(t)$54, despite magnetization $h(t)$55 (Alba et al., 2020).

Taken together, these studies suggest two cautionary principles. First, abrupt trajectories do not by themselves license an inference of abrupt hidden state changes; a monotone decaying hazard can generate changepoint-like structure through stochastic sampling. Second, extreme compression of a behavioral alphabet need not destroy the slow variables of interest; a binary sequence can retain long-range structure if the compression aligns with slowly varying internal biases. In this sense, “behavioral state decay” can reflect either decay of an event rate or slow decay of correlations, and the two are