Time-Fractional Diffusion

• Time-fractional diffusion is a generalization of classical diffusion where the time derivative is replaced by a fractional derivative to model subdiffusion and memory effects.
• It unifies stochastic models like continuous-time random walks with deterministic fractional calculus, using Caputo and Riemann–Liouville derivatives to describe anomalous transport.
• Its applications span materials science, biology, and network science, with advanced numerical methods addressing the challenges of nonlocal temporal dynamics.

Time-fractional diffusion refers to a family of diffusion processes and partial differential equations (PDEs) in which the standard first-order time derivative is replaced by a fractional-order derivative, typically of Caputo or Riemann–Liouville type. This generalization, which models subdiffusion and memory effects, is fundamentally motivated by the breakdown of Brownian scaling in systems exhibiting power-law waiting times or memory-driven anomalous transport. The mathematical theory of time-fractional diffusion unifies stochastic process models (subordinators, continuous-time random walks), deterministic fractional calculus (Caputo, Riemann–Liouville, and related derivatives), spectral analysis (Mittag-Leffler functions), and a broad set of numerical and computational frameworks for complex, multiscale, or heterogeneous media.

1. Mathematical Foundations and Model Equations

The prototypical time-fractional diffusion equation in one spatial dimension is

$\prescript{C}{}{D}_t^{\alpha} u(x, t) = D\,\frac{\partial^2 u}{\partial x^2}(x, t),\quad 0 < \alpha < 1$

where $\prescript{C}{}{D}_t^{\alpha}$ is the Caputo fractional derivative of order $\alpha$, and $D$ is a generalized diffusion coefficient with physical dimension $L^2/T^\alpha$ (Hermann et al., 16 Jun 2025). The Caputo derivative is defined as

$\prescript{C}{}{D}_t^{\alpha} f(t) = \frac{1}{\Gamma(1-\alpha)} \int_0^t (t-\tau)^{-\alpha} f'(\tau)\,d\tau$

and generalizes naturally to higher dimensions and complex domains.

On bounded domains, the equation is typically supplemented by Dirichlet or Neumann boundary conditions and initial data. For $1<\alpha<2$, the Caputo derivative encodes a fractional diffusion-wave regime and requires two initial conditions: $\prescript{C}{}{D}_t^{\alpha} u(x,t) = D\,\Delta u(x, t) + f(x, t),\quad u(x,0) = a(x),\ \partial_t u(x,0) = b(x)$ (Li et al., 13 Jun 2025, Bai et al., 13 Feb 2025, Manapany et al., 2024).

The extension to spatially variable fractional order, space-time fractional equations (mixing time-fractional and space-fractional operators), stochastic noise, and graph or network domains broadens the theory's scope and introduces new technical challenges (Angstmann et al., 26 Apr 2025, Majdoub et al., 12 Sep 2025, Deniskin et al., 21 Jan 2026, Jin et al., 2018).

2. Stochastic Origins and Subordination Principle

Time-fractional diffusion equations rigorously arise as scaling limits of continuous-time random walks (CTRWs) with heavy-tailed waiting time distributions. If waiting times $\psi(t) \sim t^{-\alpha-1}$ ($0 < \alpha < 1$), the corresponding CTRW converges in distribution to a process subordinated in time by an inverse $\alpha$-stable subordinator $E_t$ (0904.1176, Majdoub et al., 12 Sep 2025, Angstmann et al., 26 Apr 2025). The solution to the time-fractional diffusion equation can be represented as an integral (subordination formula): $$u(x, t) = \int_0^\infty p(x, \tau) h(\tau, t)\, d\tau$$ where $p(x,\tau)$ is the transition density of classical diffusion, and $h(\tau, t)$ is the probability density of the inverse stable subordinator (Mittag-Leffler waiting law) (Hermann et al., 16 Jun 2025, Cui et al., 28 Aug 2025).

On graphs, subdiffusive time-fractional diffusion is generated by randomizing the operational time with a subordinator, leading to a convex superposition of classical heat propagators with a memory-weighted time-change density (Deniskin et al., 21 Jan 2026).

3. Analytical Solution Structure: Green's Functions and Spectral Analysis

Fundamental solutions (Green's functions) of the time-fractional diffusion equation exhibit non-Gaussian self-similar structure, governed by the M–Wright (Mainardi) function and Mittag-Leffler temporal scaling: $$G_\alpha(x, t) = \frac{1}{2\sqrt{D} t^{\alpha/2}}\, M_{\alpha/2}\left(\frac{|x|}{\sqrt{D} t^{\alpha/2}}\right)$$

$$E_\alpha(-D k^2 t^\alpha) = \sum_{n=0}^{\infty} \frac{[-Dk^2t^\alpha]^n}{\Gamma(\alpha n + 1)}$$

(Hermann et al., 16 Jun 2025, Manapany et al., 2024). For $\alpha=1$, $M_{1/2}(\zeta)$ reduces to the Gaussian kernel; for $0<\alpha<1$, solutions display "fatter" tails and algebraic decay due to the $t^{-\alpha/2}$ scaling.

In bounded domains, modal expansions show algebraic (power-law) decay of all nonzero modes, controlled by asymptotics of the Mittag-Leffler function: $$E_\alpha(-\lambda t^\alpha) \sim \frac{1}{\lambda \Gamma(1-\alpha)} t^{-\alpha}\quad (t \to \infty)$$ (Deniskin et al., 21 Jan 2026). This implies that memory slows relaxation below the exponential rate of ordinary diffusion.

Operator-theoretic subordination provides an exact representation: $$E_\alpha(-t^\alpha L) = \int_0^\infty e^{-\tau L} g_\alpha(\tau, t)\, d\tau$$ where $L$ is the (possibly graph) Laplacian, and $g_\alpha(\tau, t)$ is the density of the inverse-stable subordinator (Deniskin et al., 21 Jan 2026).

4. Physical and Structural Consequences

Key physical implications and mathematical features include:

• Subdiffusion and Memory: Time-fractional models with $0<\alpha<1$ capture sublinear mean-square displacement, $\langle x^2(t)\rangle \sim t^\alpha$, reflecting trapping and retention phenomena (Hermann et al., 16 Jun 2025, Majdoub et al., 12 Sep 2025, Manapany et al., 2024).
• Non-Markovianity: The Caputo or Riemann–Liouville time derivative is a convolution over the entire past, introducing "aging" and semi-Markovian dynamics; systems exhibit slow, algebraic (not exponential) relaxation (Deniskin et al., 21 Jan 2026).
• Vertex-Dependent Waiting and Aging in Networks: On networks, holding-times at high-degree vertices decay faster, but still exhibit infinite mean (for $\alpha<1$), and enable transport biases favoring high-connectivity regions (Deniskin et al., 21 Jan 2026).
• Wave–Diffusion Interpolation: For $1<\alpha<2$ (diffusion-wave equations), short-time responses are wave-like, but the memory kernel imposes diffusion-like algebraic decay at longer times, interpolating between undamped waves and damping (Manapany et al., 2024).
• Maximum Principles: Maximum and minimum principles hold in both classical and weak (variational) forms, generalizing parabolic and elliptic theory to the time–space fractional setting, provided appropriate regularity on the memory kernel and the fractional Laplacian (Jia et al., 2016).

5. Numerical Methods and Computational Approaches

Fractional diffusion equations present unique numerical difficulties due to their nonlocal temporal structure and potential for spatial nonlocality:

• Memory-efficient Time-stepping: Fast L1 convolution—in particular, with sum-of-exponentials acceleration—reduces computational complexity versus naive history summation, achieving cost $O(N^{-(2-\alpha)})$ in time with optimal mesh grading (Yang et al., 2022).
• Mapping from Integer-order Solutions: The solution can be linearly mapped from its Brownian (integer-order) counterpart via explicit subordination formulae, dramatically reducing computational costs for long-time integration (Stokes et al., 2014).
• Physics-Informed Neural Networks: Transform-based fPINNs (fractional PINNs) for $\alpha\in(1,2)$ achieve scalable mesh-free solutions by eliminating costly derivative evaluations at off-collocation points, favoring analytic integrand transformation and efficient quadrature (Li et al., 13 Jun 2025).
• Stochastic Monte Carlo: Feynman–Kac formulas enable meshless stochastic simulation (by subordinator and Lévy-flight sampling) that is robust in high dimensions and complex geometries, avoiding the curse of dimensionality (Cui et al., 28 Aug 2025).
• Space–Time Finite Elements and Explicit/Implicit Methods: Space-time Petrov–Galerkin FEM schemes, partially-explicit multiscale time-stepping, and coarse active subspace splitting strategies address stability and efficiency for highly heterogeneous or multiscale media (Duan et al., 2017, Hu et al., 2021, Bai et al., 13 Feb 2025).

6. Extensions: Variable Order, Tempered, and Non-Singular Kernels

The classical Caputo and Riemann–Liouville operators can be generalized in several ways:

• Variable-Order Fractional Diffusion: When the order $\alpha(x)$ varies in space, variable-order models capture spatial heterogeneity in trapping (e.g., obstacles, localized defects). Proper global scaling and additional time scales are essential to ensure well-posedness and meaningful micro–macro correspondence (Angstmann et al., 26 Apr 2025).
• Tempered Fractional Diffusion: Tempered fractional derivatives incorporate an exponential cutoff in the memory kernel, truncating power-law waiting times, leading to finite moments and enabling the modeling of transient anomalous regimes in material science (Morgado et al., 2018).
• Non-singular Kernels: Caputo–Fabrizio and Atangana–Baleanu derivatives employ regular (exponential or Mittag-Leffler) kernels, avoiding $t=0$ singularities. They offer well-posed models for confined, transient, or cross-over diffusion phenomena, linking to stochastic resetting and distributed-order equations (Tateishi et al., 2017).

7. Applications and Conservation Laws

Time-fractional diffusion is a fundamental tool in:

• Materials Science and Charge Transport: Modeling dispersive transport in disordered media, time-of-flight experiments, and calibration of anomalous diffusion parameters (Hermann et al., 16 Jun 2025, Angstmann et al., 26 Apr 2025).
• Biological and Cellular Systems: Describing cell migration with non-exponential run and waiting time distributions; fractional PDEs represent emergent behaviors in search dynamics and chemokinesis (Estrada-Rodriguez et al., 2018).
• Thermal and Bioheat Models: Refining classical models (e.g., Pennes model) to describe wave-like and diffusive thermal transport in tissue (Manapany et al., 2024).
• Network Science: Elucidating subdiffusive navigation, memory effects, and optimal transport on large or heterogeneous graphs (Deniskin et al., 21 Jan 2026).

Moreover, time-fractional equations admit conservation laws (via generalized Noether-type operators and nonlinear self-adjointness) even in non-Euler–Lagrange forms, ensuring mass or energy conservation on both linear and nonlinear, subdiffusive and diffusion-wave regimes (Lukashchuk, 2014).

---

References:

• "Fractional Diffusion on Graphs: Superposition of Laplacian Semigroups and Memory" (Deniskin et al., 21 Jan 2026)
• "Anomalous diffusion for mass transport phenomena I: Analytic solutions to time fractional diffusion" (Hermann et al., 16 Jun 2025)
• "Numerical Method for Space-Time Fractional Diffusion: A Stochastic Approach" (Cui et al., 28 Aug 2025)
• "Efficient numerical solution of the time fractional diffusion equation by mapping from its Brownian counterpart" (Stokes et al., 2014)
• "Space-Time Petrov-Galerkin FEM for Fractional Diffusion Problems" (Duan et al., 2017)
• "Maximum principles for a time-space fractional diffusion equation" (Jia et al., 2016)
• "The role of fractional time-derivative operators on anomalous diffusion" (Tateishi et al., 2017)
• "Conservation laws for time-fractional subdiffusion and diffusion-wave equations" (Lukashchuk, 2014)
• "Space-time duality for fractional diffusion" (0904.1176)
• "Fractional diffusion equations interpolate between damping and waves" (Manapany et al., 2024)
• Additional references as embedded.