Skip to Content
AxiomsAxioms
  • Article
  • Open Access

2 January 2026

16 Pages

Robust Synchronization of Time-Fractional Memristive Hopfield Neural Networks

Department of Mathematics & Statistics, University of South Florida, Tampa, FL 33620, USA
†
Retied Professor Emeritus.
This article belongs to the Section Mathematical Analysis

Abstract

We introduce and study robust synchronization of time-fractional Hopfield neural networks with memristive synapses and Hebbian learning. This novel model of artificial neural networks exhibits strong memory and long-range path dependence. By scaled group estimates and analysis of fractional differencing equations, it is proved that under rather general assumptions the solution dynamics are globally dissipative and there exists a threshold condition for achieving robust synchronization of the entire neural networks if this condition is satisfied by the interneuron coupling strength. The synchronizing threshold is explicitly expressed in terms of the original parameters in the model equations and strictly decreasing for the fractional order α ∈ ( 0 , 1 ) . This result makes a breakthrough in the exploration of fractional global and longtime dynamics for AI mathematical models.

1. Introduction

Dynamical analysis of artificial neural networks plays an essential role in deep learning and all-around AI applications. In the past two decades, many biological and artificial neural network models of differential equations have been proposed and studied on topics of stability and pattern formation, mainly by approaches of matrix spectrum estimation and computational simulations [1,2,3,4]. As a most important type of recurrent neural networks, Hopfield neural networks [5,6,7] with memristive synapses [8,9] are extensively used for the preeminent performances in many frontier research applications [10,11,12,13].
For biological neural network models in terms of diffusive Hindmarsh–Rose equations [14], diffusive FitzHugh-Nagumo equations [15], and general reaction-diffusion equations [16] with memristive synapses, the exponential synchronization results have been proved by the author’s group. For the memristive Hopfield neural networks, very recently it was shown [17] that only approximate synchronization can be achieved instead of complete synchronization, due to mismatched weight dynamics and heterogeneous activation functions for the artificial neurons.
In the past two decades, along with fractional-order derivatives and related calculus, research on time-fractional differential equations has been booming in modeling the dynamics of physical and interdisciplinary processes as well as Hopfield neural networks [18,19], ranging from viscoelastic liquids [20], electrochemical processes [21], dielectric polarization [22], image processing and encryption [23], mathematical finance [24,25], to machine learning [26]. The fractional differential equation models have the advantage of continuum long-term memory properties in comparison to classical integer-order differential equations with discrete time-delay terms [27,28,29].
From the biosimulation point of view, fractional-order formulation of artificial neural networks can be analogously justified by some reported researches on biological neural systems, such as the oculomotor integrator [30], which is a neural network in the brainstem that quickly converts a temporary eye velocity signal to a persistent eye position command of a fractional order less than one. Another piece of evidence is that the fractional-order differential model can be used to efficiently emulate stimulus anticipation and oscillatory neuron firing dynamics [31], which is a cognitive preparatory measure triggered by the emotional potentials called SPN (stimulus preceding negativity). In this sense, fractional model dynamics is an important aspect for the frontier BMI (brain–machine interface) technology.
Synchronization dynamics of fractional-order memristive Hopfield neural networks is an important but widely open topic, which is more diversified than the classical stability properties converging limited to equilibrium states [32,33], because it also includes the chaotic synchronization and the extensive approximate synchronization.
In this paper, we propose a new model of Caputo-type time-fractional memristive Hopfield neural networks with the Hebbian learning rule. It can also be called fractional memristive Hopfield–Hebbian neural networks (briefly fmHHNN). We shall investigate robust synchronization of the solution trajectories for this model.
Consider a fractional-order memristive Hopfield neural network composed of m neuron nodes and denoted by NW = { N i : i = 1 , 2 , ⋯ , m } . The state variable u i ( t ) for each neuron N i , 1 ≤ i ≤ m , the symmetric weights w i j ( t ) updated by the Hebbian learning rule [34,35] before synaptic activation functions f j ( u j ) , and the memristor’s window function ρ ( t ) are governed by fractional differential equations of the same fractional order α in the following model:
D c α u i ( t ) = − a i u i + ∑ j = 1 m w i j f j ( u j ) + k ψ i ( ρ ) u i + J i − P ∑ j = 1 m ( u i − u j ) , 1 ≤ i ≤ m , D c α w i j ( t ) = − c i j w i j + λ i j f i ( u i ) f j ( u j ) , 1 ≤ i , j ≤ m , D c α ρ ( t ) = ∑ i = 1 m γ i u i − b ρ , where t > 0 , α ∈ ( 0 , 1 ) .
In this model, the Caputo-type fractional time derivative of order α ∈ ( 0 , 1 ) denoted by D c α of a scalar function y ( t ) is defined by the following:
D c α y ( t ) = 1 Γ ( 1 − α ) ∫ 0 t ( t − s ) − α y ′ ( s ) d s , t > 0 ,
and Γ ( z ) = ∫ 0 ∞ t z − 1 e − t d t , z > 0 , is the Gamma function, for y ( · ) ∈ C [ 0 , ∞ ) and y ′ ( · ) ∈ L l o c 1 [ 0 , ∞ ) . The coefficient P > 0 in (1) is called the interneuron coupling strength. The state variable u i ( t ) simulates the electric potential of a biological neuron. The initial states of the system (1) will be denoted by the following:
u i 0 = u i ( 0 ) , w i j 0 = w i j ( 0 ) , ρ 0 = ρ ( 0 ) , 1 ≤ i , j ≤ m .
Here w i j 0 = w j i 0 for any 1 ≤ i , j ≤ m . The memristive feature of this model is reflected by the memductance-potential synapse term k ψ i ( ρ ) u i for each neuron N i . The Hebbian learning rule used in (1) for synaptic weights w i j ( t ) demonstrated good performance in unsupervised machine learning tasks on certain datasets like MNIST datasets and CIFAR datasets in image recognition and classification [36].
We assume that the scalar functions f i ( s ) , 1 ≤ i ≤ m are locally Lipschitz continuous functions and the memristor window’s functions are ψ i ( s ) = s ( η i − s ) , 1 ≤ i ≤ m . The parameters a i ,   b ,   k ,   η i ,   c i j can be any given positive constants, while input/bias constants J i and coefficients λ i j , γ i can be any given real numbers. More specific assumptions on parameters and element functions will be made in Section 3.
In this paper we shall define robust synchronization for artificial neural network models and investigate this new model (1) of fractional-order memristive Hopfield neural networks to rigorously prove that robust synchronization of this fmHHNN will be triggered if the network coupling strength coefficient P satisfies a threshold condition explicitly expressed in terms of the parameters in the model equations.
The rest of the paper is organized as follows: In Section 2 we recall preliminary basics of fractional calculus and time-fractional differential equations. In particular, we shall present/prove three useful tool Lemmas, especially a new form of fractional Gronwall inequality. A definition of robust synchronization will then be introduced. In Section 3 we show the dissipative dynamics of global solutions to the initial value problem (1)–(2) in terms of the existence of an absorbing set. In Section 4 we shall prove the main result on robust synchronization of the fractional memristive Hopfield–Hebbian neural networks by the approach of direct uniform estimates of the interneuron differencing equations. Section 5 includes the Conclusions at the end of the article.

2. Fractional Analysis Lemmas

Fractional-order derivatives and fractional calculus broadly expanded the landscape of mathematical fields in modeling physical processes and phenomenological dynamics by fractional differential equations (FDE). The framework of FDE provides a flexible incorporation of long-term memory of continuum past states in the trajectories of real-time evolution. The basic theory and some applications of fractional calculus and FDE can be found in books [37,38,39].
The Caputo fractional derivative D c α y ( t ) defined in Section 1 is strictly called the left-sided Djrbashian-Caputo fractional derivative [39,40], usually for absolutely continuous functions y ∈ A C [ 0 , T ) that annihilated the singularity at the initial point t = 0 . Compared with the Riemann–Liouville fractional derivative and Grünwald–Letnikov fractional derivative, the Caputo fractional derivative can easily handle the initial conditions in a way that matches the physically observable ordinary derivatives at t = 0 , making it most suitable and popular for modeling initial value problems of evolutionary equations in real-world applications. But the product rule and the chain rule for the ordinary derivatives are not valid for Caputo fractional derivatives in analysis.
First we present a useful Lemma [41] with a brief proof for the sake of completeness, which will be used in conducting a priori estimates of solutions for FDE.
Lemma 1.
For α ∈ ( 0 , 1 ) and any given T > 0 , if f ∈ A C [ 0 , T ] , then the following holds:
1 2 D c α f 2 ( t ) ≤ f ( t ) D c α f ( t ) , for t ∈ ( 0 , T ) .
Proof. 
By changing the order of integration and then integrating by parts, we can see
Γ ( 1 − α ) f ( t ) D c α f ( t ) − 1 2 D c α f 2 ( t ) = f ( t ) ∫ 0 t ( t − s ) − α f ′ ( s ) d s − ∫ 0 t ( t − s ) − α f ( s ) f ′ ( s ) d s = ∫ 0 t ( t − s ) − α f ′ ( s ) [ f ( t ) − f ( s ) ] d s = ∫ 0 t ( t − s ) α f ′ ( s ) ∫ s t f ′ ( x ) d x d s = CoI ∫ 0 t f ′ ( x ) ∫ 0 x ( t − s ) α f ′ ( s ) d s d x = ∫ 0 t ( t − x ) α f ′ ( x ) ( t − x ) α ∫ 0 x ( t − s ) α f ′ ( s ) d s d x = 1 2 ∫ 0 t ( t − x ) α d d x ∫ 0 x ( t − s ) − α f ′ ( s ) d s 2 d x = IbP α 2 ∫ 0 t ( t − x ) α − 1 ∫ 0 x ( t − s ) − α f ′ ( s ) d s 2 d x ≥ 0 , t ∈ ( 0 , T ) .
Since Γ ( 1 − α ) > 0 for α ∈ ( 0 , 1 ) , the inequality (3) is proved. □
The Mittag–Leffler function E α ( z ) and the two-parameter Mittag–Leffler function E α 1 , α 2 ( z ) [42,43] defined by
E α ( z ) = ∑ n = 0 ∞ z n Γ ( n α + 1 ) , E α 1 , α 2 ( z ) = ∑ n = 0 ∞ z n Γ ( n α 1 + α 2 ) , z ∈ C ,
have an essential role to play in fractional differential equations and analysis. E α ( z ) is an entire function and E 1 ( z ) = e z . The function E α ( − x ) with real variable x > 0 is especially more involved in treatment of FDE. It is completely monotone shown in ([39], Corollary 3.2) meaning the derivatives ( − 1 ) n E α ( n ) ( − x ) ≥ 0 for x > 0 and all positive integer orders n.
The following Lemma provides a very useful tool in the study of dynamics with respect to global solutions of fractional-order differential equations.
Lemma 2
(Fractional Gronwall Inequality). If a nonnegative function x ( t ) is absolutely continuous for t ≥ 0 and satisfies the following inequality:
D c α x ( t ) ≤ p − q x ( t ) , for t > 0 ,
where the fractional order is α ∈ ( 0 , 1 ) , p and q are positive constants, then it holds that
x ( t ) ≤ x ( 0 ) E α ( − q t α ) + p q Γ ( α ) 1 − E α ( − q t α ) < x ( 0 ) Γ ( 1 + α ) Γ ( 1 + α ) + q t α + p q Γ ( α ) , for t > 0 .
Proof. 
Parallel to the variation of constants formula for linear ODEs, it can be shown ([39], Proposition 4.5) that x ( t ) in (4) satisfies the following inequality
x ( t ) = x ( 0 ) E α ( − q t α ) + p ∫ o t 1 ( t − s ) 1 − α E α , α ( − q ( t − s ) α ) d s − ∫ 0 t R ( t ) ( t − s ) 1 − α E α , α ( − q ( t − s ) α ) d s ≤ x ( 0 ) E α ( − q t α ) + p ∫ o t 1 ( t − s ) 1 − α E α , α ( − q ( t − s ) α ) d s , t > 0 ,
where − R ( t ) is the nonpositive gap function between the two sides of the inequality (4) and for 0 < α < 1 the function E α , α ( − x ) is always positive on x > 0 due to the complete monotonicity as mentioned.
By the integral relation with the two-parameter Mittag–Leffler functions ([43], Section 4.4), we have
1 Γ ( α ) ∫ o t 1 ( t − s ) 1 − α E α , α ( − q ( t − s ) α ) d s = t α E α , α + 1 ( − q t α ) .
From (6) and (7) the following is calculated:
x ( t ) ≤ x ( 0 ) E α ( − q t α ) + p Γ ( α ) t α E α , α + 1 ( − q t α ) , t > 0 .
Note that ([38], Theorem 4.2)
E α , β ( x ) = x E α , α + β ( x ) + 1 Γ ( β ) .
Then we have the following:
t α E α , α + 1 ( − q t α ) = 1 q 1 Γ ( 1 ) − E α , 1 ( − q t α ) = 1 q 1 − E α ( − q t α ) .
Moreover, Theorem 3.6 in [39] shows that the Mittag–Leffler function E α ( − x ) admits the two-side bounds as follows,
1 1 + Γ ( 1 − α ) x ≤ E α ( − x ) ≤ Γ ( 1 + α ) Γ ( 1 + α ) + x , x > 0 .
Finally substitute (9) and then (10) with x = q t α into (8), we end up with the result
x ( t ) ≤ x ( 0 ) E α ( − q t α ) + p Γ ( α ) t α E α , α + 1 ( − q t α ) = x ( 0 ) E α ( − q t α ) + p q Γ ( α ) 1 − E α ( − q t α ) ≤ x ( 0 ) Γ ( 1 + α ) Γ ( 1 + α ) + q t α + p q Γ ( α ) Γ ( 1 − α ) q t α 1 + Γ ( 1 − α ) q t α < x ( 0 ) Γ ( 1 + α ) Γ ( 1 + α ) + q t α + p q Γ ( α ) , t > 0 .
The fractional Gronwall inequality (5) is proved. □
In regard to fractional-order differential equations, the existence of global solutions for an initial value problem needs to be carefully addressed. Continuation of a local solution in time is a subtle issue because two local solution segments in time may not be simply concatenated like ODEs. In fact two different solution trajectories of a Caputo FDE can intersect ([44], Theorem 6.1).
The next lemma presents an implementable sufficient condition, other than the stringent global Lipschitz condition, for the existence of global solutions in time with respect to initial value problems of autonomous fractional differential equations.
Lemma 3.
For an initial value problem of autonomous Caputo fractional differential equation(s)
D c α x ( t ) = f ( x ) , x ( 0 ) = x 0 ∈ Ω ⊂ R n ,
if the scalar or vector function f ( x ) is locally Lipschitz on its bounded or unbounded domain Ω ⊂ R n and a solution of this initial value problem satisfies
∥ x ( t ) ∥   ≤   ∥ x 0 ∥ C ( t ) + Q , t ∈ I m a x = [ 0 , T m a x ) ,
where C ( t ) is a positive non-increasing continuous function, C ( t ) → 0 as t → ∞ , and Q > 0 is a constant, then the maximal existence interval of this solution is I m a x = [ 0 , ∞ ) so that x ( t ) is a global solution in time.
Proof. 
According to the dissipative condition (11), any solution trajectory satisfies { x ( t ) : t ∈ I m a x } ⊂ B x 0 = { x ∈ Ω : ∥ x ∥   ≤   ∥ x 0 ∥ C ( 0 ) + Q + 1 } , which is a bounded ball in the state space R n . By the locally Lipschitz continuous condition, the vector field f ( x ) is globally Lipschitz within the bounded set B x 0 . Then we can apply the standard existence and uniqueness theorems [38,39,45] to assert that this solution x ( t ) is a unique global solution, I m a x = [ 0 , ∞ ) , of this initial value problem. □
The absorbing set [46] defined below is an important concept in the study of dissipative evolutionary dynamics.
Definition 1.
For a d-dimensional system S of time-fractional differential equations, a bounded subset B * ⊂ R d is called an absorbing set if for any given bounded set B in the state space R d there is a finite time T B ≥ 0 such that all the solution trajectories of this system S initially started in the set B will permanently enter the set B * when t > T B . A system of fractional differential equations is called dissipative if there exists an absorbing set.
Young’s inequality is widely seen in the management of differential inequalities. It states that for any positive numbers x and y, if 1 p + 1 q = 1 and p > 1 ,   q > 1 are constants, then it holds that
x y ≤ 1 p ε x p + 1 q C ( ε , p ) y q ≤ ε x p + C ( ε , p ) y q ,
where C ( ε , p ) = ε − q / p and the constant ε > 0 can be arbitrarily given.
Finally in this section, we introduce the definition of robust synchronization.
Definition 2.
For an artificial neural network model NW in terms of time-fractional differential equations such as (1), define its synchronous degree to be
deg s ( NW ) = sup u 0 ∈ Ω ⊂ R d max 1 ≤ i < j ≤ m lim sup t → ∞ | u i ( t ) − u j ( t ) | .
where u 0 = ( u 1 ( 0 ) , ⋯ , u d ( 0 ) ) is any initial state of the neurons in the domain Ω of the state space R d . The neural network NW is called robust synchronizable if for any small ε > 0 , there exists an interneuron coupling threshold P * ( ε ) > 0 such that d e g s ( NW ) < ε for all P > P * ( ε ) .

3. Dissipative Dynamics of Fractional Hopfield Neural Networks

The mathematical model (1) with (2) of the Caputo fractional neural network NW with Hebbian learning rule can be formulated into an initial value problem of the evolutionary equation:
D c α g ( t ) = F ( g ) , t > 0 , g ( 0 ) = g 0 ∈ X = R m ( 1 + m ) + 1 .
In (14) the column vector function g ( t ) = col u i ( t ) , w i j ( t ) , ρ ( t ) ) : 1 ≤ i , j ≤ m and the initial state vector g ( 0 ) = g 0 = col ( ( u i 0 , w i j 0 , ρ 0 ) : 1 ≤ i , j ≤ m ) . The m ( 1 + m ) + 1 -dimensional nonlinear vector function, written as follows:
F ( g ) = − a i u i + ∑ j = 1 m w i j f j ( u j ) + k ψ i ( ρ ) u i + J i − P ∑ j = 1 m ( u i − u j ) : 1 ≤ i ≤ m − c i j w i j + λ i j f i ( u i ) f j ( u j ) : 1 ≤ i , j ≤ m − b ρ + ∑ i = 1 m γ i u i
is locally Lipschitz continuous. The Equation (14) is the concise version of the model (1)–(2) of the fractional memristive Hopfield–Hebbian neural networks. Here are some remarks on the versatility and characteristics:
[1] The activation functions { f i ( x ) : 1 ≤ i ≤ m } can be heterogeneous for different neurons, including Softplus Sp ( x ) = ln ( 1 + e x ) , ReLU ( x ) , GeLU ( x ) = ( x / 2 ) [ 1 + Erf ( x / 2 ) ] , which are plausibly assumed to be uniformly bounded functions since the analogous potential of biological neurons is bounded from the resting level up to the bursting limit.
[2] This model is compatible with typical memristor (chips or circuits) window functions such as the sigmoidal function tanh ( s ) , the linear function 1 − η | s | , the quadratic function 1 − η s 2 , the Jogelker memristor 1 − ( 2 s − 1 ) 2 q , and the Strukov–Williams memristor s ( η − s ) , which is taken in this model and used in CMOS technology.
[3] The synaptic weights w i j ( t ) , which are updated in real time by the Hebbian rule actually incarnate the key biological neuron factors of locality, synaptic dissipativity, and cooperative ( λ i j > 0 ) or competitive ( λ i j < 0 ) attributes. The matrix of Hebbian coefficients { λ i j } describes topological pathways of information through the network and can be sparse in training or learning networks.
We make the following Assumption and set up new notations:
a = min { a i : 1 ≤ i ≤ m } > 1 2 k η 2 , | f i ( s ) | ≤ β , J = max { | J i | : 1 ≤ i ≤ m } , ψ i ( s ) = s ( η i − s ) , η = max { η i : 1 ≤ i ≤ m } , γ = max { | γ i | : 1 ≤ i ≤ m } , λ = max { | λ i j | : 1 ≤ i , j ≤ m } , c = min { c i j : 1 ≤ i , j ≤ m } > 0 .
First we address the existence and uniqueness of global solutions in time for this initial value problem (14).
Theorem 1.
Given any initial state g 0 ∈ X , there exists a unique global solution g ( t ; g 0 ) , t ∈ [ 0 , ∞ ) , for the initial value problem (14) of the fractional memristive Hopfield neural network NW described by the model (1) with the Assumption (16).
Proof. 
By the standard theorems on solutions of Caputo fractional differential equations [38,39] and the locally Lipschitz continuous condition here in this model, for any given initial state g 0 ∈ X there exists a unique local solution g ( t ; g 0 ) of (14), whose maximum existence interval is I m a x = [ 0 , T m a x ( g 0 ) ) , which may depend on the initial state g 0 .
We first treat the dynamic equations governing the time-varying weights in the neural network NW . Multiply the w i j -equation in (1) by w i j ( t ) . Lemma 1 with the Assumption (16) shows that
1 2 D c α w i j 2 ( t ) ≤ w i j ( t ) D c α w i j ( t ) = − c i j w i j 2 ( t ) + λ i j w i j ( t ) f i ( u i ( t ) ) f j ( u j ( t ) ) ≤ − c w i j 2 ( t ) + λ β 2 w i j ( t ) ≤ − 1 2 c w i j 2 ( t ) + 1 2 c λ 2 β 4 , t ∈ I m a x .
Then, by the fractional Gronwall inequality in Lemma 2, (17) implies that
w i j 2 ( t ) ≤ w i j 2 ( 0 ) Γ ( 1 + α ) Γ ( 1 + α ) + c t α + λ 2 β 4 c 2 Γ ( α ) ≤ w i j 2 ( 0 ) + λ 2 β 4 c 2 Γ ( α ) , 1 ≤ i , j ≤ m .
Denote by W 0 = [ max { ∑ j = 1 m w i j 2 ( 0 ) : i = 1 , ⋯ , m } ] 1 / 2 . From (18) we get
∑ i = 1 m ∑ j = 1 m w i j 2 ( t ) ≤ m W 0 2 Γ ( 1 + α ) Γ ( 1 + α ) + c t α + m 2 λ 2 β 4 c 2 Γ ( α ) , t ∈ I m a x .
The key in this proof is conducting a priori scaled group estimates to treat the main fractional differential equations for the state variables u i ( t ) , 1 ≤ i ≤ m , of the neurons. It is an effective approach in the study of global and asymptotic dynamics. Multiply the u i -equation in (1) by C 1 u i ( t ) for 1 ≤ i ≤ m and sum them up. Using (16) and (19), we obtain the following fractional differential inequality
1 2 D c α ∑ i = 1 m C 1 u i 2 + 1 2 C 1 P ∑ i = 1 m ∑ j = 1 m ( u i − u j ) 2 ≤ C 1 u i D c α u i + 1 2 C 1 P ∑ i = 1 m ∑ j = 1 m ( u i − u j ) 2 = ∑ i = 1 m C 1 − a i u i 2 + ∑ j = 1 m w i j ( t ) f j ( u j ) u i + k ψ i ( ρ ) u i 2 + J i u i ≤ ∑ i = 1 m C 1 − a u i 2 + ∑ j = 1 m w i j 2 ( 0 ) G α ( t ) + m λ 2 β 4 c 2 Γ ( α ) β | u i | + k ρ ( η i − ρ ) u i 2 + J | u i | ≤ − ∑ i = 1 m C 1 a − 1 2 k η 2 u i 2 ( t ) + ∑ i = 1 m C 1 J + β W 0 G α ( t ) + 1 c m Γ ( α ) λ β 2 | u i ( t ) | ≤ − 1 2 ∑ i = 1 m C 1 a − 1 2 k η 2 u i 2 ( t ) + 1 2 C 1 m J + β W 0 G α ( t ) + 1 c m Γ ( α ) λ β 2 2 a − 1 2 k η 2 ,
for t ∈ I m a x , where C 1 > 0 is a scaling constant yet to be specified, the double sum of interneuron coupling terms in the first inequality of (20) comes from
− C 1 P ∑ i = 1 m ∑ j = 1 m ( u j − u i ) u i + ( u i − u j ) u j = − C 1 P ∑ 1 ≤ i < j ≤ m ( u i − u j ) 2 = − 1 2 C 1 P ∑ i = 1 m ∑ j = 1 m ( u i − u j ) 2
and is moved from the right-hand side to the left-hand side with the derivative terms. In the third inequality of (20), we note
0 < G α ( t ) = Γ ( 1 + α ) Γ ( 1 + α ) + c t α ≤ 1 and lim t → ∞ G α ( t ) = 0 .
The memductance synaptic term in the third inequality of (20) is treated as follows,
k ρ ( η i − ρ ) u i 2 = k ( ρ u i ) ( η u i ) − k ρ 2 u i 2 ≤ 1 2 k η 2 u i 2 + ρ 2 u i 2 − k ρ 2 u i 2 ≤ 1 2 k η 2 u i 2 .
Moreover, the Cauchy inequality is used for the second sum to reach the last inequality of (20).
Next we multiply the ρ -equation in (1) with ρ ( t ) . Using Young’s inequality (12) we obtain
1 2 D c α ρ 2 ( t ) ≤ ρ ( t ) D c α ρ ( t ) = ∑ i = 1 m γ i u i ( t ) ρ ( t ) − b ρ 2 ( t ) ≤ ∑ i = 1 m m γ i 2 2 b u i 2 ( t ) + b 2 m ρ 2 ( t ) − b ρ 2 ( t ) = ∑ i = 1 m m γ i 2 2 b u i 2 ( t ) − b 2 ρ 2 ( t ) , t ∈ I m a x .
Combine the two differential inequalities (20) and (22) together. It yields
D c α ∑ i = 1 m C 1 u i 2 + ρ 2 ( t ) + C 1 P ∑ i = 1 m ∑ j = 1 m ( u i − u j ) 2 ≤ − ∑ i = 1 m C 1 a − 1 2 k η 2 u i 2 ( t ) + C 1 m J + β W 0 G α ( t ) + 1 c m Γ ( α ) λ β 2 2 a − 1 2 k η 2 + ∑ i = 1 m m γ 2 b u i 2 ( t ) − b ρ 2 ( t ) , t ∈ I m a x .
Now we choose C 1 to be the positive constant, written as follows:
C 1 = m γ 2 b + b a − 1 2 k η 2 which means C 1 a − 1 2 k η 2 − m γ 2 b = b .
Substitute (24) in (23). Since C 1 P ∑ i = 1 m ∑ j = 1 m ( u i − u j ) 2 ≥ 0 and 0 < G α ( t ) ≤ 1 , it results in the following:
D c α ∑ i = 1 m C 1 u i 2 ( t ) + ρ 2 ( t ) + b ∑ i = 1 m u i 2 ( t ) + ρ 2 ( t ) ≤ C 2 ( W 0 ) , t ∈ I m a x ,
where
C 2 ( W 0 ) = m m γ 2 b + b J + β W 0 + 1 c m Γ ( α ) λ β 2 2 a − 1 2 k η 2 − 2
is a uniform constant independent of any initial state ( u 1 0 , ⋯ , u m 0 , ρ 0 ) of the network neurons. The fractional differential inequality (25) infers that
D c α ∑ i = 1 m C 1 u i 2 ( t ) + ρ 2 ( t ) + δ ∑ i = 1 m C 1 u i 2 ( t ) + ρ 2 ( t ) ≤ C 2 ( W 0 ) , t ∈ I m a x ,
where
δ = b min 1 C 1 , 1 .
Now we can apply the fractional Gronwall inequality in Lemma 2 to (27). Thus we obtain
min { C 1 , 1 } ∑ i = 1 m u i 2 ( t ) + ρ 2 ( t ) ≤ ∑ i = 1 m C 1 u i 2 ( t ) + ρ 2 ( t ) ≤ ∑ i = 1 m C 1 u i 2 ( 0 ) + ρ 2 ( 0 ) E α ( − δ t α ) + C 2 ( W 0 ) δ Γ ( α ) 1 − E α ( − δ t α ) ≤ ∑ i = 1 m C 1 u i 2 ( 0 ) + ρ 2 ( 0 ) Γ ( 1 + α ) Γ ( 1 + α ) + δ t α + C 2 ( W 0 ) δ Γ ( α ) , for t ∈ I m a x .
The inequalities (18) for the network weights and (29) for the network neuron states together show that the condition (11) in Lemma 3 is satisfied by all the solutions g ( t ; g 0 ) of the initial value problem (14) for any initial state g 0 . Therefore, the existence time interval I m a x = [ 0 , ∞ ) for every local solution g ( t ; g 0 ) = col ( u i ( t ) , w i j ( t ) , ρ ( t ) ) of this initial value problem (14). Therefore, for any given initial state there exists a unique global solution of this neural network NW . The proof is completed. □
Based on the proof of Theorem 1, the next theorem exhibits dissipative dynamics of this fractional memristive Hopfield–Hebbian neural network.
Theorem 2.
For time-fractional memristive Hopfield neural networks represented by the model (1) with the Assumption (16), there exists a bounded absorbing set in the state space R m + 1 of neuron variables ( u 1 , ⋯ , u m , ρ ) , which is the ball
B α * = { ( u 1 , ⋯ , u m , ρ ) ∈ R m + 1 : ∥ ( u 1 , ⋯ , u m , ρ ) ∥ 2 ≤ M ( α ) } ,
where
M ( α ) = 1 + m Γ ( α ) m γ 2 b + b J + β 1 + 1 c m Γ ( α ) λ β 2 2 δ min m γ 2 b + b , a − 1 2 k η 2 a − 1 2 k η 2 ,
and M ( α ) is a uniform positive constant independent of any initial state.
Proof. 
This is the consequence of the uniform estimate (23) with (24) for all the solutions shown in the proof of Theorem 1. From (26) and (29) we obtain the bounding inequality
min { C 1 , 1 } ∑ i = 1 m u i 2 ( t ) + ρ 2 ( t ) ≤ ∑ i = 1 m C 1 u i 2 ( 0 ) + ρ 2 ( 0 ) Γ ( 1 + α ) Γ ( 1 + α ) + δ t α + 1 δ m Γ ( α ) m γ 2 b + b J + β W 0 G α ( t ) + 1 c m Γ ( α ) λ β 2 2 a − 1 2 k η 2 − 2
for t > 0 . Note that
lim t → ∞ G α ( t ) = lim t → ∞ Γ ( 1 + α ) Γ ( 1 + α ) + c t α = 0 and lim t → ∞ Γ ( 1 + α ) Γ ( 1 + α ) + δ t α = 0 .
Thus the summation term on the right-hand side and the term W 0 G α ( t ) in (32) converge to zero ( < 1 ) as t → ∞ . Divide the inequality (32) by min { C 1 , 1 } = min m γ 2 b + b a − 1 2 k η 2 , 1 . Then we obtain the asymptotic bound
lim sup t → ∞ ∑ i = 1 m u i 2 ( t ) + ρ 2 ( t ) < M ( α ) .
Moreover, it is easy to see that the limsup convergence is uniform for all initial states of the neurons confined in any given bounded set of R m + 1 . By Definition 1, the ball B α * described in (30) and (31) is an absorbing set so that the fractional memristive Hopfield neural network NW is a globally dissipative dynamical system. □

4. Main Result on Robust Synchronization

In this section we shall present and prove the main result on robust synchronization of the fractional memristive Hopfield–Hebbian neural networks described in the model (1). The methodology is to tackle directly the fractional differencing equations of the pairwise gap functions of the neuron state variables between any two nodes in the network. Then conduct a priori estimates to sharply derive a computable threshold condition in terms of the known parameters in the model equations, which will trigger occurrence of robust synchronization of the entire neural network once this threshold condition is satisfied by the interneuron coupling strength.
The gap functions between any two neuron nodes N i and N j are denoted by
U i j ( t ) = u i ( t ) − u j ( t ) , for 1 ≤ i ≠ j ≤ m .
These gap functions satisfy the following fractional differencing equations:
D c α U i j ( t ) = − a i U i j + ( a i − a j ) u j + ∑ ℓ = 1 m w i ℓ − w j ℓ f ℓ ( u ℓ ) + k ( ψ i ( ρ ) u i − ψ j ( ρ ) u j ) + ( J i − J j ) − m P U i j ,
where the last term is
− P ∑ ℓ = 1 m [ ( u i − u ℓ ) − ( u j − u ℓ ) ] = − P ∑ ℓ = 1 m ( u i − u j ) = − m P U i j .
We introduce the secondary parameters which will be involved in the estimates of the fractional differencing Equation (34) of the gap functions as follows:
a * = max { | a i − a j | : 1 ≤ i , j ≤ m } , η * = max { | η i − η j | : 1 ≤ i , j ≤ m } , J * = max { | J i − J j | : 1 ≤ i , j ≤ m } .
The main result is the next theorem on the threshold condition for the interneuron coupling strength to achieve a robust synchronization of the entire fractional memristive Hopfield–Hebbian neural network NW described in (1). The proof is based on the global dissipative dynamics of this neural network and through analytic a priori estimation of the fractional differencing Equation (34) leveraged by the fractional Granwall inequality in Lemma 2.
Theorem 3.
The time-fractional memristive Hopfield–Hebbian neural network NW presented in the model (1) with the Assumption (16) is robust synchronizable. For any prescribed small gap ε > 0 , there exists a constant threshold
P α * ( ε ) = 1 m ε a * M 1 / 2 ( α ) + k η * M 3 / 2 ( α ) + 2 β 1 + 1 c m Γ ( α ) λ β 2 + J * > 0 ,
where M ( α ) is shown in (31), such that if the interneuron coupling strength coefficient P > P α * ( ε ) , then the neural network NW is robust synchronized. That is, for any prescribed small gap ε > 0 ,
deg s ( NW ) = sup ( u 1 0 , ⋯ , u m 0 , ρ 0 ) ∈ R m + 1 max 1 ≤ i < j ≤ m lim sup t → ∞ | u i ( t ) − u j ( t ) | < ε .
Moreover, the synchronization threshold P α * ( ε ) is strictly decreasing with respect to the fractional-order α ∈ ( 0 , 1 ) .
Proof. 
In the sequel of this proof we can denote U i j ( t ) by U ( t ) for notational simplicity. For any given 1 ≤ i ≠ j ≤ m , multiply the U i j -equation (34) by U i j ( t ) . With (18) and (33) we have the following fractional differential inequality
1 2 D c α U 2 ( t ) ≤ U ( t ) D c α U ( t ) = − a i U 2 ( t ) − ( a i − a j ) u j ( t ) U ( t ) + ∑ ℓ = 1 m w i ℓ ( t ) − w j ℓ ( t ) f ℓ ( u ℓ ( t ) ) U ( t ) + k ( ψ i ( ρ ( t ) ) − ψ j ( ρ ( t ) ) ) u i ( t ) U ( t ) + k ψ j ( ρ ( t ) ) U 2 ( t ) + ( J i − J j ) U ( t ) − m P U 2 ( t ) ≤ − a U 2 ( t ) + a * | u j ( t ) | | U ( t ) | + 2 β W 0 G α ( t ) + 1 c m Γ ( α ) λ β 2 | U ( t ) | + k | η i − η j | ρ ( t ) | u i ( t ) | | U ( t ) | + k η j ρ ( t ) − ρ 2 ( t ) U 2 ( t ) + J * | U ( t ) | − m P U 2 ( t ) ≤ − a U 2 ( t ) + a * | u j ( t ) | | U ( t ) | + 2 β W 0 G α ( t ) + 1 c m Γ ( α ) λ β 2 | U ( t ) | + k η * | ρ ( t ) | | u i ( t ) | | U ( t ) | + 1 2 k ( η 2 − ρ 2 ( t ) ) U 2 ( t ) + J * | U ( t ) | − m P U 2 ( t ) ≤ − a − 1 2 k η 2 + m P U 2 ( t ) + a * | u j ( t ) | + k η * | ρ ( t ) | | u i ( t ) | + 2 β W 0 G α ( t ) + 1 c m Γ ( α ) λ β 2 + J * | U ( t ) | .
According to the ultimate bounding estimate (21) and (33), for any given initial state and weights g 0 = ( u i 0 , w i j 0 , ρ 0 ) , there is a finite time T ( g 0 ) > 0 such that the following is written:
∑ i = 1 m u i 2 ( t ) + ρ 2 ( t ) < M ( α ) and W 0 G α ( t ) < 1 , for t > T ( g 0 ) .
Therefore the above differential inequality (37) implies that, for t > T ( g 0 ) ,
D c α U 2 ( t ) + 2 a − 1 2 k η 2 + m P U 2 ( t ) ≤ 2 a * M 1 / 2 ( α ) + k η * M 3 / 2 ( α ) + 2 β 1 + 1 c m Γ ( α ) λ β 2 + J * | U ( t ) | .
We can use the Cauchy inequality to split the product on the right-hand side of the inequality (38) into two terms. Then we obtain the inequality
D c α U 2 ( t ) + a − 1 2 k η 2 + m P U 2 ( t ) ≤ a * M 1 / 2 ( α ) + k η * M 3 / 2 ( α ) + 2 β 1 + 1 c m Γ ( α ) λ β 2 + J * 2 a − 1 2 k η 2 + m P , t > T ( g 0 ) .
Denote by
μ = a − 1 2 k η 2 + m P .
The fractional differential inequality (39) can be written as
D c α U 2 ( t ) + μ U 2 ( t ) ≤ 1 μ a * M 1 / 2 ( α ) + k η * M 3 / 2 ( α ) + 2 β 1 + 1 c m Γ ( α ) λ β 2 + J * 2
for t > T ( g 0 ) .
Apply the fractional Gronwall inequality (5) in Lemma 2 to the fractional differential inequality (41) on the time interval [ T ( g 0 ) , ∞ ) . Then it shows that for any two different neuron nodes N i and N j in the neural network, the gap function U i j ( t ) ( = U ( t ) ) of their state trajectories satisfies the following:
U i j 2 ( t ) ≤ U i j 2 ( T g 0 ) Γ ( 1 + α ) Γ ( 1 + α ) + μ ( t − T ( g 0 ) ) α + Γ ( α ) μ 2 a * M 1 / 2 ( α ) + k η * M 3 / 2 ( α ) + 2 β 1 + 1 c m Γ ( α ) λ β 2 + J * 2 , t > T ( g 0 ) .
Finally, for any prescribed small gap ε > 0 , if the threshold condition P > P α * ( ε ) in (35) is satisfied, then by (42) and because a − 1 2 k η 2 > 0 due to (16) we conclude that for any 1 ≤ i ≠ j ≤ m , the asymptotic bound is proved:
lim sup t → ∞ | u i ( t ) − u j ( t ) | = lim sup t → ∞ | U i j ( t ) | ≤ Γ ( α ) μ a * M 1 / 2 ( α ) + k η * M 3 / 2 ( α ) + 2 β 1 + 1 c m Γ ( α ) λ β 2 + J * = 1 a − 1 2 k η 2 + m P a * M 1 / 2 ( α ) + k η * M 3 / 2 ( α ) + 2 β 1 + 1 c m Γ ( α ) λ β 2 + J * < 1 m P α * ( ε ) a * M 1 / 2 ( α ) + k η * M 3 / 2 ( α ) + 2 β 1 + 1 c m Γ ( α ) λ β 2 + J * = ε .
Thus (36) is proved. Since Γ ′ ( α ) < 0 for α ∈ ( 0 , 1 ) , Γ ( α ) is strictly decreasing in this interval. Therefore, from (31) and (35), it is seen that the synchronization threshold P α * ( ε ) is strictly decreasing with respect to the fractional order α ∈ ( 0 , 1 ) . The proof is completed. □
Remark. For fractional order 0 < α < 1 , the Mittag–Leffler function E α ( − μ t α ) as t → ∞ does not have an exponential decay rate. According to (5) and (42), we can see that the robust synchronization in Theorem 3 admits a slower fractional power convergence rate O ( [ μ t α ] − 1 ) . This characteristic of the fractional robust synchronization is different from the approximate exponential synchronization rate of the ODE-modeled memristive Hopfield neural networks in the author’s earlier paper [17].
Corollary 1.
Any recursive fractional memristive Hopfield neural network presented in the model (1) with the assumption (16), except that all learning weights w i j , 1 ≤ i , j ≤ m , are fixed constants, is robustly synchronizable. For any prescribed small gap ε > 0 , there is a constant threshold
P ^ α ( ε ) = 1 m ε a * L 1 / 2 ( α ) + k η * L 3 / 2 ( α ) + 2 β W + J * > 0 ,
where W = max { | w i j | : ≤ i , j ≤ m } and
L ( α ) = 1 + m Γ ( α ) m γ 2 b + b ( β W + J ) 2 δ min m γ 2 b + b , a − 1 2 k η 2 a − 1 2 k η 2 ,
such that if the interneuron coupling strength coefficient P > P ^ α ( ε ) , then the neural network is robust synchronized.
Proof. 
Since all the weights { w i j } in the neural network are fixed constants, we should let W = max { | w i j | : 1 ≤ i , j ≤ m } replace W 0 G α ( t ) + 1 c m Γ ( α ) λ β 2 in (20), (23), (26) and later replace 1 + 1 c m Γ ( α ) λ β 2 in (31), (35).
Then we reach the conclusion that any such recursive fractional memristive Hopfield neural network will be robustly synchronized if the threshold condition P > P ^ α ( ε ) shown in (44) is satisfied by the network coupling strength coefficient. □
In particular, by the strict decreasing property of Γ ( α ) on the interval ( 0 , 1 ) and the known value Γ ( 0.5 ) = π , we have
L ( α ) = 1 + m π m γ 2 b + b ( β W + J ) 2 δ min m γ 2 b + b , a − 1 2 k η 2 a − 1 2 k η 2 , if 0.5 ≤ α < 1 .
Although Hopfield neural networks are cross-coupled networks but not multilayer or deep neural networks, the procedures and techniques in this work can be applied to the study of interlayer and intralayer synchronization of multilayer learning neural networks as well as the cluster synchronization of many multiplex networks.

5. Conclusions and Future Work

The following summarizes the contributions in this paper and outlook for further extensions.
First, I proposed and investigated a new model of Caputo-type time-fractional memristive Hopfield–Hebbian neural networks. This model incorporates multiple features of long-term memory, efficient stimulus responses in brain–machine learning, and Hebbian learning rule in unsupervised learning.
The fractional analysis of three Lemmas is presented and proved as a foundation for conducting the scaled a priori estimates of solutions, especially the fractional Gronwall inequality proved in Lemma 2, not just in the form of Mittag–Lefflrter functions but also in the explicit form in terms of Gamma functions, which enables us to effectively establish the global solution existence in Theorem 1 and the ultimate absorbing bound in Theorem 2. It paves the way to explore the asymptotic dynamics of the time-fractional neural networks.
The main contribution of this paper is Theorem 3. It rigorously proved a sufficient threshold condition to be satisfied by a single interneuron coupling strength coefficient for achieving robust synchronization of the Caputo time-fractional neural networks. The versatility and merit of this robust synchronization result lie in the explicit expression of the threshold and the convergence rate, directly computable in terms of the original parameters in the model equations. Moreover, the synchronizing threshold is shown to be decreasing for the fractional order α ∈ ( 0 , 1 ) .
The methodology and mathematical analysis in this work potentially can be generalized to study other types of time-fractional artificial neural networks, especially for multilayer forward neural networks with semi-supervised data sets and for multiplex networks with fractional dynamic models.
Mathematical research on predictable or complex asymptotic dynamics, including synchronization of artificial neural networks, is meaningful and essential in advancing AI development but remains challenging and widely open, especially for training and learning with backpropagation algorithms.

Funding

This research received no external funding.

Data Availability Statement

The original contributions in this study are presented in the article. Further inquiries can be directed to the author.

Acknowledgments

The author would like to thank the editor and the reviewers for their very helpful comments and suggestions.

Conflicts of Interest

The author declares no conflicts of interest.

References

  1. Zhang, H.; Wang, Z.; Liu, D. A comprehensive review of stability analysis of continuous-time recurrent neural networks. IEEE Trans. Neural Netw. Learn. Syst. 2014, 25, 1229–1262. [Google Scholar] [CrossRef] [Scilit]
  2. Kim, K.K.; Patron, E.R.; Braatz, R.D. Standard representation and unified stability for dynamic artificial neural network models. Neural Netw. 2018, 98, 251–262. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  3. Phan, C.; You, Y. Synchronization of boundary coupled Hindmarsh-Rose neural networks. Nonlinear Anal. Real World Appl. 2020, 55, 103139. [Google Scholar] [CrossRef] [Scilit]
  4. Majarahan, C.; Sowmiya, C.; Xu, C. Lagrange stability of inertial type neural networks: A Lyapunov-Krasovskii functional approach. Evol. Syst. 2025, 16, 63. [Google Scholar] [CrossRef] [Scilit]
  5. Hopfield, J.J. Neural networks and physical systems with emergent collective computational abilities. Proc. Nat. Acad. Sci. USA 1982, 79, 2554–2558. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  6. Hopfield, J.J. Brain, neural networks, and computation. Rev. Mod. Phys. 1999, 71, S431. [Google Scholar] [CrossRef] [Scilit]
  7. Calin, O. Deep Learning Architectures—A Mathematical Approach; Springer: Cham, Switzerland, 2020. [Google Scholar]
  8. Chua, L. Memristor—The missing circuit element. IEEE Trans. Circuit Theory 1971, 18, 507–519. [Google Scholar] [CrossRef] [Scilit]
  9. Vaidyanathan, S.; Volos, C.; Kyprianidis, I.; Stouboulos, I.N.; Tlelo-Cuautle, E. Memristor: A new concept in synchronization of coupled neuromorphic circuits. J. Eng. Sci. Technol. Rev. 2015, 8, 157–173. [Google Scholar]
  10. Mladenov, V. Synthesis and analysis of a memristor-based artificial neuron. In Proceedings of the 16th International Workshop on Cellular Nanoscale Networks and Their Applications (CNNA), Budapest, Hungary, 8–30 August 2018; ISBN 978-3-8007-4766-5. [Google Scholar]
  11. Li, Y. Simulation of memristive synapses and neuromorphic computing on a quantum computer. Phys. Rev. Res. 2021, 3, 023146. [Google Scholar] [CrossRef] [Scilit]
  12. Eshraghian, J.K.; Wang, X.X.; Lu, W.D. Memristor-based binarized spiking neural networks: Challenges and applications. IEEE Nanotechnol. Mag. 2022, 16, 14–23. [Google Scholar] [CrossRef] [Scilit]
  13. Deng, Q.; Wang, C.; Lin, H. Memristive Hopfield neural network dynamics with heterogeneous activations and its applications. Chaos Solitons Fractals 2024, 178, 114387. [Google Scholar] [CrossRef] [Scilit]
  14. You, Y. Exponential Synchronization of Memristive Hindmarsh-Rose Neural Networks. Nonlinear Anal. Real World Appl. 2023, 73, 103909. [Google Scholar] [CrossRef] [Scilit]
  15. You, Y.; Tian, J.; Tu, J. Synchronization of memristive FitzHugh-Nagumo neural networks. Commun. Nonlinear Sci. Numer. Simul. 2023, 125, 107405. [Google Scholar] [CrossRef] [Scilit]
  16. You, Y.; Tu, J. Dynamics and synchronization of weakly coupled memristive reaction-diffusion neural networks. Dyn. Partial Differ. Equ. 2025, 22, 1–28. [Google Scholar] [CrossRef] [Scilit]
  17. You, Y. Approximate synchronization of memristive Hopfield Neural networks. arXiv 2025, arXiv:2506.23279v1. [Google Scholar] [CrossRef] [Scilit]
  18. Boroomand, A.; Menhaj, M. Fractional-order Hopfield neural networks. Lect. Notes Comput. Sci. 2009, 5506, 883–890. [Google Scholar]
  19. Kaslik, C.; Sivasundaram, S. Nonlinear dynamics and chaos in fractional-order neural networks. Neural Netw. 2012, 32, 245–256. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  20. Bagley, R.L.; Calico, R.A. Fractional-order state equations for the control of viscoelastically damped structures. J. Guid. Control Dyn. 1991, 14, 304–311. [Google Scholar] [CrossRef] [Scilit]
  21. Hartley, T.T.; Lorenzo, C.E. Dynamics and control of initialized fractional-order systems. Nonlinear Anal. 2002, 29, 201–233. [Google Scholar] [CrossRef] [Scilit]
  22. Reyes-Melo, E.; Guerrero-Salazar, C.; Ortiz-Mendez, U. Application of fractional calculus to the modeling of dielectric relaxation phenomena in polymeric materials. J. Appl. Polym. Sci. 2005, 98, 923–935. [Google Scholar] [CrossRef] [Scilit]
  23. Venkatesh, J.; Pchelintsev, A.N.; Karthikeyan, A.; Parastesh, F.; Jafari, S. A fractional-order memristive two-neuron-based Hopfield neural network: Dynamical analysis and applications for image encryption. Mathematics 2023, 11, 4470. [Google Scholar] [CrossRef] [Scilit]
  24. Scalas, E.; Gorenflo, R.; Mainardi, F. Fractional calculus and continuous-time finance. Phys. A Stat. Mech. Its Appl. 2000, 284, 376–384. [Google Scholar] [CrossRef] [Scilit]
  25. Akrami, M.H.; Erjaee, G.H. Examples of analytic solutions by means of Mittag-Leffner functions of fractional Black-Scholes option pricing equation. Fract. Calc. Appl. Anal. 2015, 18, 38–47. [Google Scholar] [CrossRef] [Scilit]
  26. Raubitzek, S.; Mallinger, K.; Neubauer, T. Combining fractional derivatives and machine learning: A review. Entropy 2022, 25, 35. [Google Scholar] [CrossRef] [Scilit]
  27. Chen, W.; Sun, H.; Li, X. Fractional Derivative Modeling in Mechanics and Engineering; Springer: Berlin/Heidelberg, Germany; Science Press: Beijing, China, 2022. [Google Scholar]
  28. Li, P.; Gao, R.; Xu, C.; Li, Y.; Akgül, A.; Baleanu, D. Dynamics exploration for a fractional-order delayed zooplankton-phytoplankton system. Chaos Solitons Fractals 2023, 166, 112975. [Google Scholar] [CrossRef] [Scilit]
  29. Shukla, V.K.; Joshi, M.C.; Mishra, P.K.; Xu, C. Generalized synchronization between coupled integer, fractional and delay chaotic systems. Evol. Syst. 2025, 16, 132. [Google Scholar] [CrossRef] [Scilit]
  30. Anastasio, T. The fractional-order dynamics of brainstem vestibulo-oculomotor neurons. Biol. Cybern. 1994, 72, 69–79. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  31. Lundstrom, B.N.; Higgs, M.H.; Spain, W.J.; Fairhall, A.L. Fractional differentiation by neocortical pyramidal neurons. Nat. Neurosci. 2008, 11, 1335–1342. [Google Scholar] [CrossRef] [Scilit]
  32. Pu, Y.; Yi, Z.; Zhou, J. Fractional Hopfield neural networks: Fractional dynamic associative recurrent neural networks. IEEE Trans. Neural Netw. Learn. Syst. 2017, 28, 2319–2333. [Google Scholar] [CrossRef] [Scilit]
  33. Wang, F.; Liu, X.; Tang, M.; Chen, L. Further results on stability and synchronization of fractional-order Hopfield neural networks. Neurocomputing 2019, 346, 12–19. [Google Scholar] [CrossRef] [Scilit]
  34. Hebb, D.O. The Organization of Behavior: A Neuropsychological Theory; John Wiley and Sons: Hoboken, NJ, USA, 1949. [Google Scholar]
  35. Centorrino, V.; Bullo, F.; Russo, G. Contraction analysis of Hopfield neural networks with Hebbian learning. In Proceedings of the IEEE 61st Conference on Decision and Control, Cancun, Mexico, 6–9 December 2022; pp. 622–627. [Google Scholar]
  36. Centorrino, V.; Bullo, F.; Russo, G. Modeling and contractility of neural-synaptic networks with Hebbian learning. Automatica 2024, 164, 111636. [Google Scholar] [CrossRef] [Scilit]
  37. Kilbas, A.A.; Srivastava, H.M.; Trujillo, J.J. Theory and Applications of Fractional Differential Equations; Elsevier: Amsterdam, The Netherlands, 2006. [Google Scholar]
  38. Diethelm, K. The Analysis of Fractional Differential Equations; Springer: Berlin/Heidelberg, Germany, 2010. [Google Scholar]
  39. Jin, B. Fractional Differential Equations—An Approach via Fractional Derivatives; Springer: Cham, Switzerland, 2021. [Google Scholar]
  40. Caputo, M. Elasticitàe Dissipazione; Zanichelli: Bologna, Italy, 1969. [Google Scholar]
  41. Alikhanov, A.A. A priori estimates for solutions of boundary value problems for fractional-order equations. Differ. Equ. (Springer Nat.) 2010, 46, 660–666. [Google Scholar] [CrossRef] [Scilit]
  42. Mittag-Leffler, G.M. Sur la nouvelle fonction Eα(x). C. R. Acad. Sci. Paris (Ser. II) 1903, 137, 554–558. [Google Scholar]
  43. Gorenflo, R.; Kilbas, A.A.; Mainardi, F.; Rogosin, S.V. Mittag-Leffler Functions, Related Topics and Applications; Springer: Berlin/Heidelberg, Germany, 2014. [Google Scholar]
  44. Cong, N.D.; Tuan, H.T. Generation of nonlocal fractional dynamical systems by fractional differential equations. J. Integral. Equ. Appl. 2017, 29, 585–608. [Google Scholar] [CrossRef] [Scilit]
  45. Li, C.; Sarwar, S. Existence and continuation of solutions for Caputo type fractional differential equations. Electron. J. Differ. Equ. 2016, 1–14. Available online: https://ejde.math.txstate.edu/Volumes/2016/207/li.pdf (accessed on 29 December 2025).
  46. Sell, G.R.; You, Y. Dynamics of Evolutionary Equations; Applied Mathematical Sciences; Springer: New York, NY, USA, 2002; Volume 143. [Google Scholar]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.