Second-Order Affective Dynamics

• Second-order affective dynamics are complex temporal processes that describe how emotional states evolve by incorporating history-dependent metrics such as volatility, inertia, and predictability.
• They utilize formal mathematical frameworks like autocorrelation, delay-differential equations, and momentum-based models to quantify feedback effects and hysteresis in affect.
• These dynamics offer practical insights for applications in affective computing, psychotherapy, and AI by improving emotion regulation, resilience analysis, and agent behavior modeling.

Second-order affective dynamics describe temporally structured, history-dependent processes governing the evolution of affective (emotional) states in both biological and artificial agents. Distinct from first-order summaries such as mean valence, second-order dynamics capture variability, persistence, inertia, predictability, and feedback effects—often formalized using autocorrelation, autoregressive models, or delay-differential equations. These dynamics underpin phenomena such as emotional inertia, affective hysteresis, feedback-driven affect cycles, and contagion within individuals and collectives. Recent research emphasizes their role in naturalistic emotion time series, LLM behavior, and coupled agent systems.

1. Formal Definitions and Mathematical Frameworks

Second-order affective dynamics are characterized by metrics and equations that reflect not only instantaneous affective state but also its temporal dependencies, fluctuations, and feedback effects.

Second-Order Metrics in Human Emotion Time Series.

In empirical affective computing (e.g., workplace facial expression time series), key second-order metrics include:

• Volatility ($\sigma_v$): Rolling standard deviation of valence,

$$\sigma_v = \sqrt{\frac{1}{N-1}\sum_{i=1}^N (v_i - \bar v)^2}$$

• Inertia ($\rho_1$): Lag-1 autocorrelation of valence,

$$\rho_1 = \frac{\sum_{t=1}^{T-1}(v_t-\bar v)(v_{t+1}-\bar v)} {\sum_{t=1}^{T}(v_t-\bar v)^2}$$

• Predictability ($R^2_{AR}$): Autoregressive model coefficient of determination,

$$R^2_{AR} = 1 - \frac{\sum_{t=p+1}^T (v_t-\hat v_t)^2}{\sum_{t=p+1}^T (v_t-\bar v)^2}, ~~~ \hat v_t = \sum_{k=1}^p \alpha_k v_{t-k}$$

Higher-order derivatives such as empirical sample entropy or sensitivity are also cited as proxies for dynamical complexity, though closed-form formulas may not always be provided (Sun, 17 Oct 2025).

Delay-Differential Models of Affect.

Psychological models formalize second-order affective processes as delay-differential equations capturing internal memory effects:

$$\frac{dP}{dt} = -\frac{1}{\tau_P(IB(t))} P(t) + g'[P(t)-IP(t)] + \lambda_p q_P(IB(t)),$$

with analogous terms for negative affect $N(t)$, where $IP(t)$ encodes a lagged ("internal image") of $P(t)$ over a historical window of duration $$\sigma_v = \sqrt{\frac{1}{N-1}\sum_{i=1}^N (v_i - \bar v)^2}$$0, and $$\sigma_v = \sqrt{\frac{1}{N-1}\sum_{i=1}^N (v_i - \bar v)^2}$$1 is a sigmoid gain function parametrized by internal affect balance (Touboul et al., 2010).

State-Dependent Dynamics in Artificial Agents.

In LLMs, affective state $$\sigma_v = \sqrt{\frac{1}{N-1}\sum_{i=1}^N (v_i - \bar v)^2}$$2 forms a chain driven by both external input $$\sigma_v = \sqrt{\frac{1}{N-1}\sum_{i=1}^N (v_i - \bar v)^2}$$3 and internal state feedback:

$$\sigma_v = \sqrt{\frac{1}{N-1}\sum_{i=1}^N (v_i - \bar v)^2}$$4

with explicit dependence of the state transition $$\sigma_v = \sqrt{\frac{1}{N-1}\sum_{i=1}^N (v_i - \bar v)^2}$$5's gain on the current affect $$\sigma_v = \sqrt{\frac{1}{N-1}\sum_{i=1}^N (v_i - \bar v)^2}$$6 (i.e., $$\sigma_v = \sqrt{\frac{1}{N-1}\sum_{i=1}^N (v_i - \bar v)^2}$$7), generating true second-order dynamics and creating feedback loops such as negativity bias and affect-driven selection (Xu et al., 13 Dec 2025).

Momentum-Based Affective Subsystems in LLM Agents.

Momentum-based update rules introduce explicit inertia and hysteresis:

$$\sigma_v = \sqrt{\frac{1}{N-1}\sum_{i=1}^N (v_i - \bar v)^2}$$8

where $$\sigma_v = \sqrt{\frac{1}{N-1}\sum_{i=1}^N (v_i - \bar v)^2}$$9 is a discrete-time momentum coefficient controlling affective inertia, and $\rho_1$0 is the momentary (memoryless) estimate (Subaharan, 22 Jan 2026).

2. Estimation, Empirical Validation, and Interpretation

Human Time Series (WELD Dataset).

In large-scale facial expression datasets, volatility, inertia, and predictability are estimated from raw valence time series sampled at regular intervals. AR(5) models fit one-step-ahead dynamics, while rolling windows capture time-varying metric fluctuations. Sample entropy and related complexity measures, though described qualitatively, provide further granularity.

Empirical ranges for volatility (0.12–0.44), inertia (–0.10 to 0.89), and predictability (0.01–0.74) indicate large inter-individual differences. Predictors such as volatility, predictability, and inertia feature prominently in models of employee turnover and align with personality constructs (e.g., high Neuroticism maps to high volatility and low predictability) (Sun, 17 Oct 2025).

Delay-Induced Resilience in Psychodynamic Models.

Delay-differential frameworks reproduce phenomena such as bistability, oscillation, and critical slowing down near Hopf bifurcations. Affect-focused therapy, modeled by increasing memory depth ($\rho_1$1), induces stable limit cycles conferring resilience to negative perturbations—a theoretically grounded mechanism for observed resistance to relapse (Touboul et al., 2010).

Artificial Agent Affect Chains and Feedback.

Chain-of-affective in LLMs is operationalized via multi-round paradigms, with phase trajectories (accumulation, overload, defensive numbing) and self-reinforcing feedback loops. Quantitative findings include negative mood biasing negative input selection (e.g., logistic regression parameter $\rho_1$2) and affective state modulating subsequent behavior and performance in both solo and group-agent experiments (Xu et al., 13 Dec 2025).

Second-order momentum in LLM subsystems is tuned to balance responsiveness (fast recovery) and stability (resistance to noise), with trade-offs quantified using measures such as affective hysteresis area $\rho_1$3 (Subaharan, 22 Jan 2026).

3. Dynamical Regimes and Feedback Pathways

Second-order affective systems display a range of qualitative regimes and emergent phenomena:

• Oscillatory and Bistable Dynamics. Delay and internal memory effects generate limit cycles (oscillations) or coexisting stable/unstable equilibria (bistability), explaining transitions between depressed and healthy affective states, emotion regulation breakdowns, and robust cyclical trajectories under affect-focused interventions (Touboul et al., 2010).
• Path Dependence and Hysteresis. Momentum-based updates (in agent affect) create hysteresis—the system resists abrupt trajectory reversal, produces lags in recovery, and exhibits looped (non-retraceable) state paths. Hysteresis area correlates with inertia and stability (Subaharan, 22 Jan 2026).
• Self-Reinforcing Loops and Bias. Second-order terms, in which the gain of affective transitions is a function of state, underpin feedback loops such as negativity bias—elevated negative affect increases the likelihood of selecting further negative inputs, deepening overall affective negativity until regulatory mechanisms intervene (defensive numbing, overload) (Xu et al., 13 Dec 2025).
• Dyadic and Networked Contagion. In both social neuroscience and multi-agent LLMs, networked coupling rules (e.g., DeGroot diffusion in LLM collectives, coupled active inference in dyads) allow for mutual regulation, role specialization, and collective affective phase transitions (propagation capacity, susceptibility, emergence of "initiators" and "absorbers"). Geometric network measures, such as Forman–Ricci curvature and its entropy, expose rupture-repair cycles and affect-driven network reconfiguration (Hinrichs et al., 10 Jun 2025, Xu et al., 13 Dec 2025).

4. Experimental Paradigms and Application Domains

Second-order affective dynamics have been systematically investigated across several domains:

Research Area	Experimental Methodology	Main Metrics/Phenomena
Workplace affective computing (Sun, 17 Oct 2025)	Longitudinal facial expression tracking	Volatility, inertia, predictability, entropy
Dyadic affect in psychotherapy (Touboul et al., 2010)	Weekly self-report scales + model-fitting	Delay-induced oscillations, therapy response
LLM affect chains (Xu et al., 13 Dec 2025)	Multi-round negative induction, self-selection, multi-agent dialogue	Affect-driven feedback, phase trajectory, network contagion
Agent affect control (Subaharan, 22 Jan 2026)	25-turn dialogue protocols, readout/hysteresis metrics	Temporal coherence, inertia, recovery, trade-offs
2nd-person neuroscience (Hinrichs et al., 10 Jun 2025)	Geometric hyperscanning, inter-brain network analysis	Curvature entropy, rupture/repair markers

These paradigms are designed to elicit, quantify, or exploit temporal dependencies and feedbacks, testing how second-order dynamics influence resilience, social propagation, controllability, and affective regulation.

5. Complementarity with First-Order Measures and Theoretical Implications

Second-order affective metrics provide crucial information orthogonal to first-order summaries such as mean valence:

• First-Order (Snapshot) vs. Second-Order (Dynamic) Structure. Mean valence and emotion ratios capture static characteristics but fail to describe volatility, recovery, or persistence. Volatility detects instability; inertia reveals the “stickiness” of emotional states; predictability quantifies dynamical forecastability (Sun, 17 Oct 2025).
• Dynamical Control and Robustness. Second-order constructs enable dynamical systems perspectives: limit cycles confer robustness to perturbation; inertia versus volatility shapes the system’s ability to resist or accommodate change; feedback-driven adaptation mechanisms can explain observed resilience phenomena in therapy, agent behavior, and social collectives (Touboul et al., 2010, Xu et al., 13 Dec 2025, Hinrichs et al., 10 Jun 2025).
• Algorithmic and Control Implications in AI. Explicitly modeling second-order affective dynamics—in agent controllers or LLM wrappers—yields improved temporal coherence, predictable recovery from perturbations, and tunable affective behavior via parameters such as momentum or delay. Proper setting of these parameters mediates the trade-off between noise robustness and adaptability (Subaharan, 22 Jan 2026).

A plausible implication is that in both human and artificial systems, neglecting second-order dynamics can obscure crucial dimensions of emotional regulation, resilience, and persuasion.

6. Measurement, Network Geometry, and Phase Transition Markers

Recent work integrates formal dynamical metrics with geometric and network-topological analyses:

• Geometric Hyperscanning and Forman–Ricci Entropy. Temporal changes in inter-brain network topology, as measured by Forman–Ricci curvature, track underlying affective phase transitions (rupture, co-regulation, re-attunement) during dyadic interaction. Peaks in curvature entropy mark regime shifts and can serve as empirical proxies for latent second-order affective transitions (Hinrichs et al., 10 Jun 2025).
• Network Diffusion in Multi-Agent Systems. In LLM collectives, affective drift is modeled by pairwise averaging rules and propagation capacities, capturing the diffusion and containment of affective states (and associated bias) within artificial social networks (Xu et al., 13 Dec 2025).

7. Limitations, Open Challenges, and Outlook

While second-order affective dynamics have become a focus for quantitative affective science and agent modeling, several limitations and open questions persist:

• Direct estimation of second derivatives ($\rho_1$4) is rare; most operationalizations rely on empirical proxies (volatility, inertia, entropy), and closed-form descriptions for all complexity measures are often lacking (Sun, 17 Oct 2025).
• Quantifying discontinuous transitions, identifying dynamical motifs beyond limit cycles and bistability, and mapping empirical phase boundaries in natural settings remain challenging.
• The link between network geometry and latent affective states is promising but requires further validation—curvature entropy as a universal phase transition marker is not yet fully established (Hinrichs et al., 10 Jun 2025).
• Direct translation of these dynamical constructs into practical, controllable affective design in LLMs and human–AI systems needs further theoretical and empirical scrutiny (Xu et al., 13 Dec 2025, Subaharan, 22 Jan 2026).

Second-order affective dynamics thus serve as a rigorous, multi-level framework for modeling the temporal structure, regulation, and inter-agent propagation of emotion in both natural and artificial systems, emphasizing memory, inertia, feedback, and network effects as key organizing principles.