Time-Fractional SEIR Model Overview

• The time-fractional SEIR model generalizes the classical epidemic model by replacing standard derivatives with fractional ones to incorporate nonlocal memory effects.
• It captures delayed incidence, heavy-tailed transmission, and sub-exponential growth, providing a flexible framework for analyzing complex epidemic dynamics.
• Numerical methods like the L1 finite difference and ABM schemes enhance simulation accuracy and support effective parameter inference for outbreak prediction.

A time-fractional SEIR model generalizes the classical SEIR (Susceptible-Exposed-Infectious-Recovered) epidemic framework by replacing ordinary time derivatives with fractional-order derivatives, typically of Caputo, Caputo-Fabrizio, or Caputo-Hadamard type. This extension incorporates nonlocal memory effects and hereditary influences in the disease dynamics, reflecting empirical observations of sub-exponential growth and persistence in diverse epidemic datasets. The fractional order $\alpha$ offers both theoretical generality and practical flexibility, functioning as a tunable parameter to capture anomalous temporal features such as delayed incidence, heavy-tailed transmission, and transient heterogeneity in contact or behavioral patterns.

1. Mathematical Formulation and Compartmental Extensions

The standard time-fractional SEIR system with Caputo derivative of order $0 < \alpha \leq 1$ is written as

$$\begin{cases} D_t^{\alpha} S(t) = \Lambda - \beta S(t) I(t)/N(t) - (\nu + \mu) S(t), \ D_t^{\alpha} E(t) = \beta S(t) I(t)/N(t) - (\gamma + \mu) E(t), \ D_t^{\alpha} I(t) = \gamma E(t) - (\alpha_r + \sigma + \mu) I(t), \ D_t^{\alpha} R(t) = \nu S(t) + \alpha_r I(t) - \mu R(t), \end{cases}$$

where $D_t^{\alpha}$ indicates the fractional derivative (typically Caputo), $\Lambda$ represents recruitment, $\beta$ the transmission rate, $\nu$ the vaccination rate, $\gamma$ the progression $E \to I$, $\alpha_r$ the recovery rate, $\mu$ the natural death rate, and $\sigma$ disease-induced mortality. Compartmental extensions include the addition of quarantine, protection, asymptomatic, hospitalization, and cumulative death states as required for greater biological realism or to fit particular disease contexts such as COVID-19 or Mpox (Saini et al., 14 Jan 2026, Santos et al., 2021, Xu et al., 2020, Biala et al., 2020, Cai et al., 2022).

Fractional SEIR models based on non-singular kernels (Caputo-Fabrizio) (Zinihi et al., 2024), Caputo-Hadamard (Cai et al., 2022), or with generalized incidence functions (Xu et al., 2020) have been developed to accommodate diverse forms of memory, mitigate singularities at the origin, and support more flexible parameterization.

2. Fractional Derivatives and Memory Effects

The Caputo fractional derivative of order $\alpha$ ($0<\alpha\le1$) is defined as

$$D_t^\alpha f(t) = \frac{1}{\Gamma(1-\alpha)} \int_0^t (t-s)^{-\alpha} f'(s) ds,$$

with $D_t^1 f(t) = df/dt$. For the Caputo-Fabrizio variant,

$${}^{CFC}D^\alpha f(t) = \frac{M(\alpha)}{1-\alpha} \int_0^t f'(s) e^{-\gamma (t-s)} ds,$$

using an exponential kernel, while the Caputo-Hadamard derivative employs a logarithmic kernel (Cai et al., 2022).

The physical role of $\alpha$: for $\alpha = 1$, the model reduces to the standard Markovian SEIR ODE; values $\alpha<1$ encode long-memory kernels, leading to a "history-dependent" rate of change. Empirically, lower $\alpha$ produces delayed and lower epidemic peaks, extended tails in infection curves, and heavy persistence, as confirmed by simulations and data fits (Saini et al., 14 Jan 2026, Santos et al., 2021, Cai et al., 2022, Gabrick et al., 2024).

3. Existence, Stability, and Qualitative Properties

Well-posedness of the fractional SEIR system follows from classical fixed-point and contraction mapping theorems adapted using fractional Volterra integral reformulations (Saini et al., 14 Jan 2026). Hyers–Ulam stability holds under natural Lipschitz conditions, ensuring that solutions subject to small perturbations remain close to true trajectories. Theoretical results guarantee existence and uniqueness of solutions in appropriate Banach spaces for both ODE (Santos et al., 2021, Gabrick et al., 2024) and PDE (Zinihi et al., 2024) variants.

The basic reproduction number $R_0$ for fractional SEIR is formally identical to its classical counterpart (e.g., $R_0 = \beta / \gamma$ in the simple compartmental case (Gabrick et al., 2024)), although in some extended systems $R_0$ incorporates $\alpha$-dependent rate exponents (Santos et al., 2021, Biala et al., 2020). Stability analyses extend the next-generation matrix approach and Routh–Hurwitz criteria to fractional settings: the disease-free equilibrium is locally asymptotically stable if $R_0 < 1$ and $\alpha$ satisfies the fractional angle condition $|\arg \lambda_i| > \alpha \pi / 2$ for all eigenvalues $\lambda_i$ (Biala et al., 2020, Xu et al., 2020).

4. Numerical Methods and Simulation Accuracy

Numerical integration of time-fractional SEIR systems employs specialized discretizations for fractional derivatives. The L1 finite difference scheme,

$$D_N^{\alpha} X_j \approx \frac{\Delta t^{-\alpha}}{\Gamma(2-\alpha)} \sum_{k=0}^{j-1} (X_{k+1} - X_k) d_{j-k},$$

with weights $d_p = p^{1-\alpha} - (p-1)^{1-\alpha}$ achieves algebraic convergence of order $\alpha$ (Saini et al., 14 Jan 2026). Alternative schemes include the Grünwald–Letnikov method, fractional Euler (FMEM), Gorenflo–Mainardi–Moretti–Paradisi (GMMP), and the Adams–Bashforth–Moulton predictor–corrector (Santos et al., 2021, Saini et al., 14 Jan 2026). L1 and ABM methods provide strongly superior accuracy to FMEM and can handle stiff, long-time integration with memory efficiently (Saini et al., 14 Jan 2026).

Parameter estimation is performed by nonlinear least-squares fitting to epidemiological data (cases, deaths, recovered), frequently incorporating piecewise-constant or neural network parametrizations of transmission, recovery, and removal rates (Cai et al., 2022, Santos et al., 2021). Inverse solvers adapt the loss functional to combine data fidelity with dynamical consistency (supervised and physics-informed residual losses in fPINNs (Cai et al., 2022)). Calibration yields typical values for $\alpha$ in the range $0.7 < \alpha < 1$ depending on disease phase, region, and intervention regime, with lower $\alpha$ reflecting stronger memory or hidden subdiffusive dynamics.

5. Modeling Findings and Epidemiological Insights

Time-fractional SEIR models consistently demonstrate that:

• For fixed epidemic parameters, lowering $\alpha$ delays and reduces the infectious peak, extends the outbreak duration, and increases the total number of cases in heavy-tail scenarios (Saini et al., 14 Jan 2026, Santos et al., 2021, Gabrick et al., 2024).
• Sub-exponential and ultra-slow initial growth phases observed in diseases with concealed transmission (e.g., COVID-19 Omicron) are accurately captured for $\alpha$ well below 1 (e.g., $\alpha=0.75$ early in the outbreak) (Cai et al., 2022).
• Incorporating optimal control (vaccination) in spatially extended, fractional PDE models demonstrates that fractional memory gives more time for control to act and further suppresses peaks compared to integer-order analogues (Zinihi et al., 2024).

Lockdown and policy measures appear as time-varying transmission rates $\beta(t)$, with fractional models able to replicate multiple epidemic waves or resurgence after relaxation of interventions (Santos et al., 2021, Xu et al., 2020). Fitting to regional or national COVID-19 datasets yields out-of-sample forecast errors below 5% for cumulative and active cases, affirming the practical predictive utility of the fractional framework (Xu et al., 2020).

6. Extensions: Spatial, Behavioral, and Data-Driven Models

Fractional SEIR models are routinely extended in three directions:

1. Spatial Reaction–Diffusion: Inclusion of diffusion terms and spatial Laplacians coupled with fractional temporal derivatives to model the spread of epidemic fronts and the effect of spatial heterogeneity (Zinihi et al., 2024).
2. Behavioral and Policy Modules: Addition of compartments for protected, quarantined, or isolated individuals, piecewise or time-dependent parameters for government measures, and explicit vaccination control (Xu et al., 2020, Zinihi et al., 2024).
3. Machine-Learned Fractional Dynamics: Use of physics-informed neural networks (fPINNs) for simultaneous inference of $\alpha(t)$, rate functions, and latent compartments directly from noisy surveillance data under fractional ODE/PDE constraints (Cai et al., 2022).

7. Comparative Model Performance and Practical Recommendations

Comparative investigations demonstrate that:

• Fractional models generally outperform both integer-order and fractal-derivative (Hausdorff-type) analogues in fitting slow-evolving outbreaks, but for rapidly equilibrating epidemics (e.g., classic SEIR dynamics for India's COVID-19 data) the standard model may sometimes suffice (Gabrick et al., 2024).
• The L1 finite difference and ABM predictor–corrector schemes are preferred for accurate and robust simulation, showing orders-of-magnitude smaller errors than FMEM at moderate step sizes (Saini et al., 14 Jan 2026, Santos et al., 2021).
• For parameter inference and prediction, fractional SEIR models offer greater flexibility and interpretability, particularly in regimes exhibiting history dependence, anomalous incubation, or behavioral memory, and should be calibrated using surveillance data and cross-validated for predictive accuracy (Santos et al., 2021, Cai et al., 2022).

The fractional order $\alpha$ itself is interpreted as an epidemiological marker of memory and heterogeneity; inference of time-varying or regime-specific $\alpha$ offers insights into transitional phases of epidemics and the impact of unobserved factors such as compliance decay or transmission delays.

---

References:

(Saini et al., 14 Jan 2026, Zinihi et al., 2024, Santos et al., 2021, Gabrick et al., 2024, Cai et al., 2022, Xu et al., 2020, Biala et al., 2020)