Fractional Dynamical Networks

• Fractional dynamical networks are systems whose evolution is governed by noninteger-order differential or difference equations, capturing long memory and heavy-tail behaviors.
• They employ operators like the Caputo and Grünwald–Letnikov derivatives and fractional Laplacians to model nonlocal spatial diffusion and Lévy-flight-like transport.
• This framework enables robust analysis of stability, control, and estimation in applications ranging from neurophysiology to networked control systems.

Fractional dynamical networks are networked dynamical systems in which the governing evolution is fractional in time, fractional in space, or both. In this literature, fractionality appears as Caputo or Grünwald–Letnikov temporal operators that encode long memory and heavy-tailed renewal statistics, as spectral fractional powers of graph Laplacians that encode nonlocal transport and Lévy-flight-like jumps, and as combined space–time constructions obtained by subordinating a spatial walk to a non-Poisson renewal process (West et al., 2014, Riascos et al., 2015, Benzi et al., 2019, Michelitsch et al., 2019). The resulting models are used to connect microscopic interaction rules to macroscopic phase behavior, to represent long-range spatial connectivity, and to formalize estimation, stability, learning, and control in systems whose dynamics are not well described by memoryless Markov evolution.

1. Definitions and scope

The cited literature uses the term fractional dynamical network in two closely related ways. In one usage, it denotes a network whose node dynamics obey fractional differential or fractional difference equations, so that each node retains a weighted history of past states. In the other, it denotes a network whose spatial coupling is fractionalized, typically by replacing the standard graph Laplacian $L$ by a fractional operator $L^\gamma$, or by imposing a power-law rule on long-range edges so that diffusion becomes superdiffusive rather than local (Chatterjee et al., 2021, Riascos et al., 2015, 2002.04351).

A further structural specialization appears under the name fractional networks: networks embedded in a metric space whose long-range connectivity obeys $P_{ij} = c\, d_{ij}^{-\gamma}$, with sufficiently small $\gamma$, so that transport is governed by a fractional operator and not primarily by hubs. In that setting, the effective diffusion exponent is $\beta = \gamma - 1$, normal diffusion corresponds to $\beta = 2$, and $\beta < 2$ yields superdiffusion (Mendes, 2018, 2002.04351).

Regime	Representative operator	Characteristic effect
Time-fractional	$D_t^\alpha x(t)$, $\Delta^\alpha x[k]$	Long memory, non-Markovian relaxation
Space-fractional	$L^\gamma$, $L^\gamma$0	Nonlocal jumps, Lévy-flight-like transport
Space–time fractional	$L^\gamma$1	Coupled temporal memory and nonlocal spreading

In discrete time, the basic operator is the Grünwald–Letnikov fractional difference

$L^\gamma$2

which realizes memory as an infinite convolution with binomial/Gamma coefficients. In continuous time, the Caputo derivative is the dominant representation for state-space modeling and stability analysis, while spectral graph calculus is the dominant representation for spatial nonlocality (Chatterjee et al., 2021, Siami, 2020, Benzi et al., 2019).

2. Temporal fractionality, memory, and non-Markovian network dynamics

Time-fractional network dynamics arise when event times, switching times, or state updates obey heavy-tailed renewal statistics rather than exponential waiting laws. A canonical example is the decision-making model in which agents imperfectly imitate local majorities through transition rates

$L^\gamma$3

In the all-to-all mean-field limit, the order parameter $L^\gamma$4 undergoes a pitchfork bifurcation at $L^\gamma$5; on a two-dimensional nearest-neighbor lattice, the phase transition occurs at $L^\gamma$6 for $L^\gamma$7. For finite $L^\gamma$8, fluctuations induce intermittent switching between consensus wells, and the switching-time density follows an inverse power law rather than an exponential law (West et al., 2014).

The relevant waiting-time density can be written as

$L^\gamma$9

with survival function

$P_{ij} = c\, d_{ij}^{-\gamma}$0

Asymptotically, $P_{ij} = c\, d_{ij}^{-\gamma}$1 and $P_{ij} = c\, d_{ij}^{-\gamma}$2, which encodes long memory, non-Poisson renewal structure, non-ergodicity, and aging. Under subordination, a node that is Markovian in operational time becomes fractional in chronological time, and its ensemble-averaged state obeys a Caputo equation

$P_{ij} = c\, d_{ij}^{-\gamma}$3

The homogeneous solution is Mittag–Leffler,

$P_{ij} = c\, d_{ij}^{-\gamma}$4

with stretched-exponential short-time behavior and algebraic long-time decay. On $P_{ij} = c\, d_{ij}^{-\gamma}$5 lattices, the reported fits yield $P_{ij} = c\, d_{ij}^{-\gamma}$6 at $P_{ij} = c\, d_{ij}^{-\gamma}$7, $P_{ij} = c\, d_{ij}^{-\gamma}$8 at $P_{ij} = c\, d_{ij}^{-\gamma}$9, and $\gamma$0 at $\gamma$1, with $\gamma$2 values near unity (West et al., 2014).

The same mechanism appears in network continuous-time random walks with bursty inter-event times. When the waiting-time density has a diverging mean,

$\gamma$3

the Laplace-domain memory kernel behaves as $\gamma$4, and the asymptotic network equation becomes fractional in time. A decisive consequence is that all non-stationary modes relax with the same power law $\gamma$5, so the dynamics loses the time-scale separation characteristic of Markovian diffusion (Nigris et al., 2016).

For discrete-time fractional-order dynamical networks, memory is built directly into the state equation through infinite convolutions of past states, inputs, and disturbances. Under the invertibility assumption on $\gamma$6, the dynamics can be rewritten as an infinite-memory autoregressive model, which is then truncated to a finite-memory $\gamma$7-approximation for computation (Chatterjee et al., 2021).

3. Spatial fractionality, nonlocal coupling, and graph operators

Spatially fractional dynamical networks are obtained by fractionalizing the graph operator itself. For an undirected connected graph with Laplacian $\gamma$8, the fractional Laplacian is defined spectrally as

$\gamma$9

Because $\beta = \gamma - 1$0 is typically dense even when $\beta = \gamma - 1$1 is sparse, it induces effective long-range interactions: walkers are no longer restricted to nearest-neighbor moves, and transition probabilities acquire power-law tails in graph distance (Riascos et al., 2015).

A standard random-walk construction uses the modified fractional Laplacian

$\beta = \gamma - 1$2

The stationary distribution is proportional to the fractional degree $\beta = \gamma - 1$3. On the ring, the one-step kernel has the exact asymptotic tail

$\beta = \gamma - 1$4

which identifies a Lévy-flight kernel with index $\beta = \gamma - 1$5. The average return probability decays more rapidly than in the normal walk, and the global time

$\beta = \gamma - 1$6

decreases markedly as $\beta = \gamma - 1$7 decreases from $\beta = \gamma - 1$8, especially on large-world topologies such as rings, trees, and lattices (Riascos et al., 2015).

For directed networks, the nonsymmetric Laplacian $\beta = \gamma - 1$9 is treated as a singular $\beta = 2$0-matrix, and $\beta = 2$1 is defined through primary matrix functions or Jordan calculus. This preserves the kernel, yields off-diagonal nonpositivity, and supports continuous-time dynamics

$\beta = 2$2

with convergence to the projection onto $\beta = 2$3. In the normalized form, $\beta = 2$4 generates a valid Markov semigroup with nonlocal jumps that respect directionality (Benzi et al., 2019).

A complementary spatial view starts from metric connectivity. If long-range links obey $\beta = 2$5, then a renormalization argument identifies the effective space-fractional exponent as $\beta = 2$6, so $\beta = 2$7 or equivalently $\beta = 2$8 marks the superdiffusive regime. The mixed diffusion equation studied in this setting is

$\beta = 2$9

with Fourier solution

$\beta < 2$0

For $\beta < 2$1, the fractional term dominates at small $\beta < 2$2, hence at large spatial scales, while the normal term dominates at large $\beta < 2$3, yielding a crossover wavenumber

$\beta < 2$4

This formalizes mixed diffusion: rapid long-range integration coexisting with local smoothing (2002.04351).

The structural literature distinguishes these fractional networks from scale-free networks. In scale-free networks, transport is organized by hubs and degree heterogeneity. In fractional networks, transport is organized by distributed long-range spatial links, and the reported robustness and controllability properties are correspondingly different (Mendes, 2018).

4. Space–time fractional transport and multiplex generalizations

Space–time fractional dynamical networks combine nonlocal jumps with non-Poisson renewal statistics. In the continuous-time random walk framework, the renewal law is generalized by the Laplace-domain waiting-time transform

$\beta < 2$5

whose time-domain density is the Prabhakar–Mittag–Leffler law

$\beta < 2$6

Its memory kernel in Laplace space is

$\beta < 2$7

In the diffusion limit on $\beta < 2$8, the corresponding macroscopic equation becomes

$\beta < 2$9

with Fourier–Laplace solution

$D_t^\alpha x(t)$0

This four-parameter space $D_t^\alpha x(t)$1 unifies standard Poisson diffusion, Laskin’s fractional Poisson process, and more general non-Markovian Lévy transport (Michelitsch et al., 2019).

A distinct but related extension is the fractional continuous-time random walk on multiplex networks. For a two-layer multiplex with supra-Laplacian $D_t^\alpha x(t)$2, the fractional diffusion equation is

$D_t^\alpha x(t)$3

In the node-centric formulation, the transition kernel is built from the normalized fractional supra-Laplacian $D_t^\alpha x(t)$4, and the propagator is

$D_t^\alpha x(t)$5

For two identical circulant layers, the algebraic connectivity $D_t^\alpha x(t)$6 has a nonmonotone dependence on the inter-layer coupling $D_t^\alpha x(t)$7, and the relaxation time $D_t^\alpha x(t)$8 is minimized at the closed-form optimum

$D_t^\alpha x(t)$9

independent of $\Delta^\alpha x[k]$0. Even under enhanced diffusion, however, the mean-square displacement grows linearly in time in finite multiplexes. The fractional effect appears in the prefactor and in the long-range inter-layer transition kernel, whose tail decays as $\Delta^\alpha x[k]$1 (Allen-Perkins et al., 2019).

These constructions show that fractionality is not confined to isolated nodes or single-layer graphs. It extends to multiscale transport processes in which memory, topology, and interlayer coupling enter the same operator-theoretic description.

5. Stability, estimation, and learning

The stability theory of fractional dynamical networks is organized around the spectrum of the effective linear operator and the admissible sector determined by the fractional order. For commensurate Caputo networks

$\Delta^\alpha x[k]$2

Matignon’s criterion states that asymptotic stability holds iff every eigenvalue $\Delta^\alpha x[k]$3 of $\Delta^\alpha x[k]$4 satisfies

$\Delta^\alpha x[k]$5

For cyclic interconnections of scalar fractional subsystems, a generalized secant condition gives a topology-aware bound. Writing

$\Delta^\alpha x[k]$6

the network is asymptotically stable for any positive gains when $\Delta^\alpha x[k]$7, while for $\Delta^\alpha x[k]$8 it is stable if

$\Delta^\alpha x[k]$9

When the $L^\gamma$0 are identical, this condition is also necessary. The same work quantifies robustness by the $L^\gamma$1-norm and shows that, even for stable FLTI systems, the $L^\gamma$2-norm is unbounded for $L^\gamma$3 (Siami, 2020).

For complex-valued fractional-order neural networks, the local analysis reduces to the spectrum of

$L^\gamma$4

and the critical fractional order for stability loss is

$L^\gamma$5

Hub and ring topologies admit explicit eigenvalue formulas, making the Hopf-like threshold computable in closed form from connectivity parameters and activation derivatives (Kaslik et al., 2016).

Estimation theory for discrete-time fractional-order dynamical networks is built on finite-memory approximations of the Grünwald–Letnikov convolution. In the minimum-energy framework, the $L^\gamma$6-truncated augmented system admits the estimator recursion

$L^\gamma$7

with

$L^\gamma$8

and the estimation error is exponentially input-to-state stable under bounded-operator, complete uniform controllability, complete uniform observability, and weight-bound assumptions. The residual $L^\gamma$9, which collects neglected memory terms, decreases as $L^\gamma$00 increases (Chatterjee et al., 2021).

Under partial observability and hidden drivers, latent-state learning has been formulated as a joint estimation problem for observed states $L^\gamma$01, latent states $L^\gamma$02, sparse unknown inputs $L^\gamma$03, and coupling matrices. The model uses the discrete fractional operator

$L^\gamma$04

with $L^\gamma$05, and alternates a fractional Kalman filtering step with sparse input estimation and EM-style parameter updates (Gupta et al., 2018).

Neural network implementations have likewise adopted fractional operators. Fractional-DNNs discretize a Caputo evolution equation layer by layer, with nonuniform learned step sizes $L^\gamma$06 and history coefficients $L^\gamma$07; this yields additive identity and memory terms that the cited work associates with mitigation of vanishing and exploding gradients (Antil et al., 2022). More recently, variable-order neural differential equation networks have replaced a constant fractional order by a learned state-dependent order,

$L^\gamma$08

and on graphs

$L^\gamma$09

Their numerical realization uses variable-order L1 or Adams–Bashforth–Moulton updates whose weights change with the learned order at every step, producing adaptive memory kernels rather than stationary power-law memory (Cui et al., 20 Mar 2025).

6. Control, applications, and open problems

Fractional dynamical networks have been proposed for social consensus, neuronal and brain networks, traffic and infrastructure, biological regulation, networked control systems, and neurophysiological EEG/BOLD modeling (West et al., 2014, 2002.04351, Chatterjee et al., 2021). The unifying rationale is that these systems often display long memory, bursty event timing, multiscale dependence, or long-range spatial coupling that are not naturally represented by integer-order models.

Control theory has moved beyond fixed exponents. In discrete-time linear fractional-order networks $L^\gamma$10, it has been shown that both the coupling matrix and the vector of fractional exponents can be steered by suitable inputs. After defractionalization, the target dynamics take the form

$L^\gamma$11

and the associated fractional reachability matrix

$L^\gamma$12

gives explicit finite-step steering conditions. Under finite-memory truncation, the truncation error decays at least on the order of $L^\gamma$13, and energy-constrained steering can be formulated as a quadratic program with actuator bounds (Varalda et al., 29 May 2026).

The most developed biomedical application in the supplied literature is seizure modeling and suppression. In intracranial EEG-derived fractional dynamical networks, the identified model is

$L^\gamma$14

with electrode-specific orders $L^\gamma$15 estimated from wavelet logscale regressions and stability assessed through $L^\gamma$16. Across 35 spontaneous seizures from 10 patients, the reported medians of the fractional exponent were $L^\gamma$17 for interictal, $L^\gamma$18 for pre-ictal, $L^\gamma$19 for ictal, and $L^\gamma$20 for post-ictal segments, with all pairwise pooled KS tests significant at $L^\gamma$21. A sparsity-promoting convex stabilization program produced simulated amplitude reduction of $L^\gamma$22, reduced amplitude in 34 of 35 seizures, stabilized 27 of 35 seizures overall, and stabilized 17 of the 22 initially unstable seizures (Wang et al., 26 Nov 2025).

Several open problems recur across the literature. One is mechanistic derivation: in decision and social models, the heavy-tailed $L^\gamma$23 is often modeled phenomenologically, and deriving tail exponents directly from agent rules and topology remains open (West et al., 2014). Another is finite memory and tempering: real systems frequently exhibit tail truncation, whereas many current models assume pure power laws and therefore asymptotic Mittag–Leffler behavior (West et al., 2014, Nigris et al., 2016). Heterogeneity and multiplexity remain central technical challenges, since node-specific rates, directed couplings, and multilayer structure alter both spectra and effective fractional orders (Benzi et al., 2019, Allen-Perkins et al., 2019). Finally, numerical cost is intrinsic: variable-order L1 and ABM solvers require history sums over all previous steps, yielding $L^\gamma$24 time over $L^\gamma$25 steps unless short-memory or compression strategies are introduced (Cui et al., 20 Mar 2025).

Taken together, these results establish fractional dynamical networks as a broad operator-theoretic framework for networked systems with memory and nonlocality. Their distinctive feature is not merely the appearance of noninteger exponents, but the systematic replacement of local, memoryless evolution by dynamics in which past states, distant nodes, or both contribute through power-law kernels and fractional operators.