Why Fractional Differential Equations Have Memory

1. Core Distinction: Local vs. Nonlocal

Integer-order derivatives are local. Taking velocity as an example, $v(t) = u'(t)$ depends only on the instantaneous rate of change of $u$ near time $t$; the past trajectory has no direct influence on the current derivative. However, fractional-order derivatives are nonlocal -- their definition inherently involves integration over the entire history interval, meaning that the state of the system at time $t$ depends not only on the present, but also on the full history from the initial moment to $t$.

2. Mathematical Origin: Integral-Form Definitions

The two classical definitions of fractional derivatives -- Riemann-Liouville (RL) and Caputo -- both appear in integral form, and thus naturally possess memory.

Riemann-Liouville Fractional Derivative

The RL fractional derivative first performs an $\alpha$-th order integration of the function, then takes an integer-order derivative:

$$^{RL} D_t^\alpha u(t) = \frac{1}{\Gamma(n-\alpha)} \frac{d^n}{dt^n} \int_0^t \frac{u(s)}{(t-s)^{\alpha-n+1}} \, ds, \quad n-1 < \alpha < n$$

The integration interval is $[0, t]$, and the kernel $(t-s)^{-(\alpha-n+1)}$ in the integrand serves as the memory kernel. Since integration acts first and differentiation acts afterward, the RL derivative's memory kernel acts directly on the function values $u(s)$ themselves, meaning that the values of the function at every past moment contribute to the current derivative with power-law weighting.

Caputo Fractional Derivative

The Caputo derivative, by contrast, first takes an integer-order derivative and then performs an $\alpha$-th order integration:

$$^C D_t^\alpha u(t) = \frac{1}{\Gamma(n-\alpha)} \int_0^t \frac{u^{(n)}(s)}{(t-s)^{\alpha-n+1}} \, ds, \quad n-1 < \alpha < n$$

Unlike the RL derivative, the Caputo derivative's memory kernel acts on the $n$-th derivative $u^{(n)}(s)$ of the function, meaning it remembers the rates of change at each past moment rather than the function values themselves. This allows the Caputo derivative to impose initial conditions in the classical manner (e.g., $u(0)=u_0$), making it more suitable for physical modeling.

Connection Between the Two

The RL and Caputo derivatives are related through the Taylor expansion of the function at the initial moment:

$$^{RL} D_t^\alpha u(t) = ^C D_t^\alpha u(t) + \sum_{k=0}^{n-1} \frac{u^{(k)}(0)}{\Gamma(k-\alpha+1)} \, t^{k-\alpha}$$

When the initial conditions $u^{(k)}(0) = 0$ (for $k=0,1,\dots,n-1$), the two are completely equivalent. Both share the same memory kernel $(t-s)^{-(\alpha-n+1)}$, so memory is a common property of fractional derivatives, regardless of whether the RL or Caputo form is used.

3. Power-Law Decay of the Memory Kernel

The memory kernel $(t-s)^{\alpha-n-1}$ (or $(t-s)^{-\alpha}$ when $n=1$) exhibits power-law decay. Unlike exponential decay $e^{-\lambda(t-s)}$, power-law decay is much slower: even when $t-s$ is large, the influence of the past on the present does not vanish rapidly, but instead diminishes slowly in an algebraic manner.

$$\text{Power-law decay: } \frac{1}{(t-s)^\alpha} \quad \text{vs} \quad \text{Exponential decay: } e^{-\lambda(t-s)}$$

This precisely corresponds to long-range memory in physical systems: past events are not immediately forgotten, but continue to influence the future evolution of the system persistently and slowly.

4. Comparison with Integer-Order Equations

Feature	Integer-Order Equation	Fractional-Order Equation
Derivative definition	Local limit: $\lim_{h\to 0} \frac{u(t+h)-u(t)}{h}$	Global integral: $\int_0^t K(t,s) \, u(s) \, ds$
Memory kernel	None (instantaneous response)	Power-law type: $(t-s)^{-\alpha}$
History dependence	Depends only on initial conditions	Depends on the full history trajectory
Decay type	Exponential decay (Markov process)	Power-law decay (non-Markov process)
Typical applications	Newtonian mechanics, normal diffusion	Viscoelasticity, subdiffusion, anomalous relaxation

5. Physical Intuition: The Example of Viscoelastic Materials

Consider the stress-strain relationship of a viscoelastic rod. Integer-order models (such as Kelvin-Voigt) assume that stress is proportional only to the current strain and strain rate; fractional-order models, by contrast, characterize the material's "memory" by integrating the historical strain:

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

This means that the stress in the material at time $t$ depends not only on the current deformation, but also on the entire deformation history from the start of loading, and the contribution of earlier deformations to the current stress decays more slowly (power-law decay). This is the mathematical manifestation of viscoelastic materials "remembering" their deformation history.

6. Understanding from the Perspective of Stochastic Processes

Integer-order diffusion equations correspond to Brownian motion, where the mean-square displacement of particles satisfies $\langle x^2(t) \rangle \sim t$, and the displacement increments at any moment are independent (memoryless). Fractional-order diffusion equations, on the other hand, correspond to continuous-time random walks (CTRW), where both the particle waiting times and jump lengths follow heavy-tailed distributions, leading to:

The mean-square displacement follows $\langle x^2(t) \rangle \sim t^\gamma$ (where $\gamma < 1$ for subdiffusion),
The particle's current position depends on its waiting history and jump history, meaning the process exhibits long-range correlation.

Therefore, the memory of fractional equations originates at the microscopic level from the heavy-tailed distributions of particle motion statistics, and manifests at the macroscopic level as the integral form of fractional derivatives.

7. Summary

The memory of fractional differential equations is not an artificially imposed assumption, but a natural consequence of their mathematical definition: fractional derivatives are defined through integration over the entire history interval, and the power-law decay of the memory kernel causes past states to influence the current evolution in a slow and persistent manner. This property makes fractional equations an ideal mathematical tool for describing physical, biological, and engineering systems that exhibit memory, heredity, and long-range correlations.