Abstract
There has been a great deal of interest in the last few years for neural networks (NNs) with values in multidimensional domains. The most popular models are complex-valued neural networks (CVNNs), followed by quaternion-valued neural networks (QVNNs), and, more recently, by Clifford-valued neural networks (ClVNNs). However, also very recently, a different type of NNs were put forward, namely octonion-valued neural networks (OVNNs). OVNNs are defined on the 8D octonion algebra, and they are not a special type of ClVNNs, because Clifford numbers are associative, whereas octonions are not. Moreover, beside the complex and quaternion algebras, the only other normed division algebra over the reals is the algebra of octonions, which makes OVNNs a direct generalization of CVNNs and QVNNs from this point of view, raising interest for applications handling high-dimensional data. On the other hand, systems defined on time scales were proposed as a generalization of both discrete time and continuous time systems, or any type of hybrid combination between the two. Finally, time delays appear as a consequence of implementing NNs in real life circuits. Taking all these into consideration, this paper studies the fundamental properties of exponential stability and exponential synchronization for OVNNs with leakage and mixed delays defined on time scales. In order to avoid the problems raised by the non-associativity of the octonion algebra, the OVNN model is decomposed into a real-valued one. Then, two different Lyapunov-type functionals are defined and the particularities of time scale calculus are used in order to deduce sufficient conditions expressed as scalar and linear matrix inequalities (LMIs) for the exponential stability of the proposed models, based on Halanay-type inequalities suitable for time scale systems. Afterwards, a state feedback controller is used to deduce sufficient criteria given as scalar inequalities and LMIs for the exponential synchronization of the same type of models. The generality of the model, given by the definition on the algebra of octonions, the use of different types of delays, and the definition on time scales represent an important advantage. Also, for less general models, it is possible to particularize the obtained results. With the aim to demonstrate each of the paper’s four theorems, four numerical examples are provided.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
The study of NNs with values in multidimensional domains has increased over the last few years. As such, CVNNs, for which the domain is the 2D complex numbers algebra, and QVNNs, for which the domain is the 4D quaternion algebra, have appeared. Their generalization, ClVNNs, for which the domain can be any \(2^{n}\)-dimensional Clifford algebra, with \(n\ge 1\), have also begun to raise interest lately.
A different generalization of CVNNs and QVNNs has also recently appeared, in the form of OVNNs, for which the domain is the 8D octonion algebra. The octonion algebra does not belong to the Clifford algebras family, which is easy to see because all Clifford algebras are associative, whereas the octonion algebra is not. However, the octonion algebra is the only other normed division algebra that can be defined over the reals, beside the complex and quaternion algebras. This property makes it interesting for applications, and, indeed, the octonions have been successfully applied in electrodynamics (Chanyal 2013), gravitational field equations (Demir 2012), or fluid dynamics (Demir and Tanişli 2016), but also in salient object detection (Gao and Lam 2014a, b), signal processing (Snopek 2015; Wang et al. 2016; Błaszczyk and Snopek 2017), or hyperspectral fluorescence data fusion (Bauer and Leon 2016), for example.
Because of their potential applications in problems related to high-dimensional data processing, it was a natural idea to apply the octonion algebra to the domain of NNs, also, which gave rise to OVNNs. Feedforward OVNNs were first proposed in Popa (2016). Then, the dynamics of recurrent Hopfield OVNNs was studied in Popa (2018a), Popa (2018b). Since then, other papers appeared, which study different dynamic properties of OVNNs, for example (Kandasamy and Rajan 2020; Chouhan et al. 2022; Li and Huang 2022; Chouhan et al. 2023; Gao and Dai 2023; Huang and Li 2023; Li et al. 2023; Popa 2023). The majority of these papers are very recent, showing a growing interest from the scientific community in the domain of OVNNs, indicating the likely publication of additional papers in this area in the near future.
When discussing the dynamics of NN models, it’s essential to put forward models with time delays, because, in real-world implementations of NNs, they naturally appear because of the finite switching velocity of circuit components, and can determine undesired behavior. One common type of delays is the leakage delay, which appears in the self-feedback term of the NNs. It has been added, as part of OVNN models, in Popa (2018a), Gao and Dai (2023). Then, the conduction velocities along the pathways of the circuit implementation of NNs can cause dispersion, which gives rise to distributed delays. When they appear in conjunction with time-varying delays, which is the type of delays that was most often discussed in the context of NNs, they are called mixed delays. Mixed delays have been added to OVNN models in Popa (2018a), Gao and Dai (2023), Popa (2023).
On the other hand, to our best knowledge, all the papers involving OVNNs are done in continuous time. However, when NNs are implemented in circuits, discretization is an inevitable step, resulting in a discrete time system, which can differ very much in its dynamic properties from those of the original continuous time system. It is this realization that led to the definition of NNs in discrete time, for the first time in Mohamad and Gopalsamy (2000). Since then, discrete time NNs have developed into a field of study in their own right, with an increasing number of papers appearing each year. As previously mentioned, there are no papers in the available literature discussing discrete time OVNNs.
An elegant solution to the problem of continuous time vs. discrete time systems was proposed in the form of time scale calculus, for the first time in Hilger (1990). With this type of calculus, the properties of differential and difference systems, or any hybrid combination of them, can be studied in a unified manner. Today, the most important references for time scale calculus are the books (Bohner and Peterson 2001; Martynyuk 2016; Adıvar and Raffoul 2020), in which the domain was developed further and summarized. Time scales have also been added to the study of NNs, for the first time in Chen and Du (2008). Similarly to discrete time NNs, time scale NNs have become an established area of research in their own right, with many papers appearing yearly that discuss this type of networks. Also, like for discrete time OVNNs, there are no papers currently discussing time scales OVNNs, to the best of our knowledge.
For systems defined on time scales, the traditional Lyapunov theory is not suitable, since time scale calculus is more broad and must take into account the properties of both differential and difference systems. As an alternative, in order to study the dynamics of such systems, different variations of Halanay inequalities have been employed over time, see (Mohamad and Gopalsamy 2000; Wen et al. 2008; Wang 2010; Wen et al. 2018; Kassim and Tatar 2021). Halanay inequalities for time scales are the result of extending these inequalities to time scales, see (Adıvar and Bohner 2011; Ou et al. 2015, 2016; Ou 2020). They were then applied to studying the dynamics of time scale NNs, for example in Xiao and Zeng (2017), Xiao and Zeng (2018), Xiao et al. (2018), Xiao et al. (2020), Wan and Zeng (2021), Wan and Zeng (2022a), Wan and Zeng (2022b).
The key contributions of the paper can be summarized as follows:
-
1.
A model of OVNNs with leakage, time-varying, and distributed delays defined on time scales is formulated, and then, to avoid the problems related to the non-associativity and non-commutativity of the octonion algebra, the system is transformed into a real-valued one.
-
2.
Lyapunov-like functions pertaining to two general categories are formulated, and Halanay-type inequalities for time scales are used in order to deduce sufficient conditions, expressed both as scalar inequalities and as LMIs, which guarantee the exponential stability of the proposed model.
-
3.
Then, two similar Lyapunov-like functions are used in conjunction with a state feedback controller in order to obtain sufficient criteria expressed in the same way for the exponential synchronization of the OVNNs put forward.
-
4.
Through the means of one numerical simulation, both in discrete time and in continuous time, each of the four theorems is demonstrated.
-
5.
The results obtained in the paper are general enough that they can be particularized for continuous time or for discrete time OVNNs, or any hybrid combination of the two, or even for models of CVNNs or QVNNs, for which, to our knowledge, no comparable results have been reported in the literature.
The remaining part of the research unfolds in the following way. Section 2 is dedicated to emphasizing the most important aspects of the octonion algebra and of the time scale calculus, along with useful lemmas and an assumption that will be used to obtain the main results. Also, the proposed model is given in the same section, and its transformation into a real-valued system is performed. Then, Sect. 3 is dedicated to presenting the results of the paper: two theorems for the stability and two theorems for the synchronization of the proposed model. One numerical simulation is employed to demonstrate each of the four theorems, both in discrete time and in continuous time, in Sect. 4. The last section, Sect. 5, is dedicated to concluding the paper.
Notations: \(\mathbb {R}\) – reals, \(\mathbb {R}^{+}\) – positive reals, \(\mathbb {O}\) – octonions, \(||\cdot ||_{p}\) – \(L_{p}\) norm, \(p\in \{1,2\}\), \(\mathbb {R}^{N}\) (\(\mathbb {O}^{N}\)) – real (octonion) vectors of dimension N, \(\mathbb {R}^{N\times N}\) (\(\mathbb {O}^{N\times N}\)) – real (octonion) \(N\times N\)-dimensional matrices, \(Y<0\) – Y is negative definite, \(\lambda _{\min }(Y)\) – smallest eigenvalue of Y, \(Y^{T}\) – transpose of Y.
2 Preliminaries
First, we provide an overview of the octonion algebra, mostly based on Popa (2018a), Popa (2018b). “The set of octonions is defined as:
where \(e_{q}\) represent the unit octonions, \(\forall 0\le q\le 7\). On this set, we define the octonion addition by \(o+p:=\sum _{q=0}^{7}(o^{q}+p^{q})e_{q}\) and the scalar multiplication by \(\alpha o:=\sum _{q=0}^{7}(\alpha o^{q})e_{q}\). The multiplication of octonions is defined by the multiplication table of the octonion units:
\(\times \) | \(e_{0}\) | \(e_{1}\) | \(e_{2}\) | \(e_{3}\) | \(e_{4}\) | \(e_{5}\) | \(e_{6}\) | \(e_{7}\) |
---|---|---|---|---|---|---|---|---|
\(e_{0}\) | \(e_{0}\) | \(e_{1}\) | \(e_{2}\) | \(e_{3}\) | \(e_{4}\) | \(e_{5}\) | \(e_{6}\) | \(e_{7}\) |
\(e_{1}\) | \(e_{1}\) | \(-e_{0}\) | \(e_{3}\) | \(-e_{2}\) | \(e_{5}\) | \(-e_{4}\) | \(-e_{7}\) | \(e_{6}\) |
\(e_{2}\) | \(e_{2}\) | \(-e_{3}\) | \(-e_{0}\) | \(e_{1}\) | \(e_{6}\) | \(e_{7}\) | \(-e_{4}\) | \(-e_{5}\) |
\(e_{3}\) | \(e_{3}\) | \(e_{2}\) | \(-e_{1}\) | \(-e_{0}\) | \(e_{7}\) | \(-e_{6}\) | \(e_{5}\) | \(-e_{4}\) |
\(e_{4}\) | \(e_{4}\) | \(-e_{5}\) | \(-e_{6}\) | \(-e_{7}\) | \(-e_{0}\) | \(e_{1}\) | \(e_{2}\) | \(e_{3}\) |
\(e_{5}\) | \(e_{5}\) | \(e_{4}\) | \(-e_{7}\) | \(e_{6}\) | \(-e_{1}\) | \(-e_{0}\) | \(-e_{3}\) | \(e_{2}\) |
\(e_{6}\) | \(e_{6}\) | \(e_{7}\) | \(e_{4}\) | \(-e_{5}\) | \(-e_{2}\) | \(e_{3}\) | \(-e_{0}\) | \(-e_{1}\) |
\(e_{7}\) | \(e_{7}\) | \(-e_{6}\) | \(e_{5}\) | \(e_{4}\) | \(-e_{3}\) | \(-e_{2}\) | \(e_{1}\) | \(-e_{0}\) |
For each octonion \(o\in \mathbb {O}\), its conjugate is defined as \(\overline{o}:=o^{0}e_{0}-\sum _{q=1}^{7}o^{q}e_{q}\). Then, we define the norm of octonion \(o\in \mathbb {O}\) as \(|o|:=\sqrt{\overline{o}o}=\sqrt{\sum _{q=0}^{7}(o^{q})^{2}}\) and its inverse as \(o^{-1}:=\frac{\overline{o}}{|o|^{2}}\). With all these operations, it can be proved that \(\mathbb {O}\) is a normed division algebra. Actually, there exists a famous result by Hurwitz, who showed that the real, complex, quaternion, and octonion number sets with their respective operations are the only normed division algebras over the real numbers.
As can be seen from the multiplication table, we have that \(e_{q}e_{l}=-e_{l}e_{q}\ne e_{l}e_{q}\) if \(q\ne l\), \(q\ne 0\), \(l\ne 0\), which means that, like the algebra of quaternions, the above-defined algebra of octonions is not commutative. Moreover, we can observe that \((e_{q}e_{l})e_{m}=-e_{q}(e_{l}e_{m})\ne e_{q}(e_{l}e_{m})\) if q, l, m are non-zero, different, and \(e_{q}e_{l}\ne \pm e_{m}.\) This tells us that, unlike the algebra of quaternions and other Clifford algebras in general, the algebra of octonions is also not associative.”
Next, we present the fundamentals of calculus on time scales, generally based on Bohner and Peterson (2001). “Time scale \(\mathbb {T}\) is defined as an non-empty closed subset of the real number set \(\mathbb {R}\), from which it inherits the ordering and the topology. The following jump operators can be defined \(\forall t\in \mathbb {T}\): forward jump \(\sigma (t):=\inf \{\left. s\in \mathbb {T}\right| s>t\}\) and backward jump \(\rho (t):=\sup \{\left. s\in \mathbb {T}\right| s<t\}\). Function \(\mu :\mathbb {T}\rightarrow [0,+\infty )\), defined as \(\mu (t):=\sigma (t)-t\), \(\forall t\in \mathbb {T}\), represents the forward graininess function. Denote \(\hat{\mu }=\sup \{\left. \mu (t)\right| t\in \mathbb {T}\}\).
Then, a point \(t\in \mathbb {T}\) is left (right)-dense if \(\rho (t)=t\) (\(\sigma (t)=t\)); a point \(t\in \mathbb {T}\) is left (right)-scattered if \(\rho (t)<t\) (\(\sigma (t)>t\)). If there exists a left-scattered maximum for \(\mathbb {T}\) denoted as m, then \(\mathbb {T}^{\kappa }:=\mathbb {T}\backslash \{m\}\), else \(\mathbb {T}^{\kappa }:=\mathbb {T}\). The function \(f:\mathbb {T}\rightarrow \mathbb {R}\) is said to be rd-continuous if, for any right-dense \(t_{1}\in \mathbb {T}\), \(f(t_{1})=\lim _{\zeta \rightarrow t_{1}^{+}}f(\zeta )\), and for any left-dense \(t_{2}\in \mathbb {T}\), \(\lim _{\zeta \rightarrow t_{2}^{-}}f(\zeta )\) exists. The notation \(C_{rd}(\mathbb {T},\mathbb {R})\) is used for the set of rd-continuous functions \(f:\mathbb {T}\rightarrow \mathbb {R}\). The jump operators for a function \(f:\mathbb {T}\rightarrow \mathbb {R}\) are defined as \(f^{\sigma }(t)=f(\sigma (t))\) and \(f^{\rho }(t)=f(\rho (t))\), respectively. A function \(p:\mathbb {T}\rightarrow \mathbb {R}\) is called regressive, and we write \(p\in \mathcal {R}(\mathbb {T},\mathbb {R})\), if \(p\in C_{rd}(\mathbb {T},\mathbb {R})\) and \(1+\mu (t)p(t)\ne 0\), \(\forall t\in \mathbb {T}^{\kappa }\). The function \(p:\mathbb {T}\rightarrow \mathbb {R}\) is called positively regressive, and we write \(p\in \mathcal {R}^{+}(\mathbb {T},\mathbb {R})\), if \(p\in C_{rd}(\mathbb {T},\mathbb {R})\) and \(1+\mu (t)p(t)>0\), \(\forall t\in \mathbb {T}^{\kappa }\). We define, \(\forall p\in \mathcal {R}(\mathbb {T},\mathbb {R})\), \(\ominus p(t):=-p(t)/(1+\mu (t)p(t))\), \(\forall t\in \mathbb {T}\). For any set \(S\subseteq \mathbb {R}\) we define \(S_{\mathbb {T}}=S\cap \mathbb {T}\).
For a function \(f:\mathbb {T}\rightarrow \mathbb {R}\), if \(t\in \mathbb {T}^{\kappa }\), the \(\Delta \)-derivative of f at t is, if it exists, the number denoted by \(f^{\Delta }(t)\) such that \(\forall \varepsilon >0\) there exists a \(\delta >0\) for which the following inequality holds \(\forall s\in (t-\delta ,t+\delta )_{\mathbb {T}}\):
If the \(\Delta \)-derivative exists \(\forall t\in \mathbb {T}^{\kappa }\), the function f is called \(\Delta \)-differentiable.
\(\Delta \)-integration constitutes the inverse operation of \(\Delta \)-differentiation, i.e., if \(F^{\Delta }(t)=f(t)\), then
Lastly, if \(p\in \mathcal {R}(\mathbb {T},\mathbb {R})\) is a regressive function, then the \(\Delta \)-exponential function \(e_{p}:\mathbb {T}\times \mathbb {T}\rightarrow \mathbb {R}\) is defined by the formula:
where \(\xi _{\mu (s)}\) denotes the cylinder transformation, which is defined as:
Related to time scales, we will need the following lemmas for the proofs of our theorems:
Lemma 1
(Bohner and Peterson (2001)) “If \(f,g:\mathbb {T}\rightarrow \mathbb {R}\) are \(\Delta \)-differentiable, then
-
(i)
\((f(t)+g(t))^{\Delta }=f^{\Delta }(t)+g^{\Delta }(t)\);
-
(ii)
\((f(t)g(t))^{\Delta }=f^{\Delta }(t)g(t)+f(\sigma (t))g^{\Delta }(t)=f(t)g^{\Delta }(t)+f^{\Delta }(t)g(\sigma (t)).\)”
Lemma 2
(Xiao and Zeng (2018)) “Let \(g:\mathbb {T}\rightarrow \mathbb {R}^{+}\) be a non-negative rd-continuous function on \(\mathbb {T}\), and
holds for \(t\in [t_{0},+\infty )_{\mathbb {T}}\), where \(\alpha _{1}\) and \(\alpha _{2}\) are positive constants. If \(-\alpha _{1}\in \mathcal {R}^{+}\), \(\alpha _{1}-\alpha _{2}>0\) then
for any \(t\in [t_{0},+\infty )_{\mathbb {T}}\), where \(\lambda >0\) satisfies the inequality
Lemma 3
(Xiao and Zeng (2018)) “If \(-\lambda \in \mathcal {R}^{+}\), \(y\in C_{rd}(\mathbb {T},\mathbb {R})\), \(z\in C_{rd}(\mathbb {T},\mathbb {R})\), then, \(\forall t\in \mathbb {T}\),
implies
We are now prepared to present the model that will be studied. Define on time scale \(\mathbb {T}\) the subsequent OVNN having leakage, time-varying, and distributed delays:
\(\forall i\in \{1,\ldots ,N\}\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\), where \(o_{i}(t)\in \mathbb {O}\) – states at \(t\in [0,+\infty )_{\mathbb {T}}\), \(c_{i}\) – self-feedback weights, \(a_{ij}\in \mathbb {O}\) – weights without delay, \(b_{ij}\in \mathbb {O}\) – weights with delay, \(g_{ij}\in \mathbb {O}\) – distributed delay weights, \(f_{j}:\mathbb {O}\rightarrow \mathbb {O}\) – activation functions, and \(I_{i}\in \mathbb {O}\) – external inputs, \(\forall i,j\in \{1,\ldots ,N\}\), \(\delta >0\) is the leakage delay, \(\tau :\mathbb {R}^{+}\rightarrow \mathbb {R}^{+}\) are the time-varying delays, and we assume that there exists \(\tau >0\) such that \(\tau (t)\le \tau \), \(\forall t\in [0,+\infty )_{\mathbb {T}}\), and \(\eta >0\) is the distributed delay. The notation \(\omega :=\max \{\delta ,\tau ,\eta \}\) is made. Also, assume that the activation functions \(f_{j}\) have the form \(f_{j}(o)=\sum _{q=0}^{7}f_{j}^{q}(o)e_{q}\), \(\forall o\in \mathbb {O}\), where \(f_{j}^{q}:\mathbb {O}\rightarrow \mathbb {R}\), \(\forall 0\le q\le 7\), \(\forall j\in \{1,\ldots ,N\}\).
The initial conditions of NN (1) are:
where \(\varphi _{i}\in \mathcal {C}([-\omega ,0]_{\mathbb {T}},\mathbb {O})\), \(\forall i\in \{1,\ldots ,N\}\). On \(\mathcal {C}([-\omega ,0]_{\mathbb {T}},\mathbb {O}^{N})\) we define the norm as \(||\varphi ||=\sum _{i=1}^{N}\sup _{[-\omega ,0]_{\mathbb {T}}}|\varphi _{i}(t)|\).
System (1) is assumed to possess an unique equilibrium point, denoted by \(\tilde{o}=(\tilde{o}_{1},\ldots ,\tilde{o}_{N})^{T}\) in the following. In this setting, if we put \(\mathfrak {o}_{i}(t)=o_{i}(t)-\tilde{o}_{i}\), system (1) becomes:
\(\forall i\in \{1,\ldots ,N\}\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\), where \(\tilde{f}_{j}(\mathfrak {o}_{j}(t))=f_{j}(\mathfrak {o}_{j}(t)+\tilde{o}_{j})-f_{j}(\tilde{o}_{j})\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\), \(\forall j\in \{1,\ldots ,N\}\).
Now, NN (2) has the following initial conditions:
where \(\chi _{i}\in \mathcal {C}([-\omega ,0]_{\mathbb {T}},\mathbb {O})\), \(\forall i\in \{1,\ldots ,N\}\).
The system of equations (2) will now be converted into 8 real-valued systems. This is accomplished by using the following 8 equations to represent each equation in (2):
\(\forall i\in \{1,\ldots ,N\}\), \(\forall 0\le q\le 7\), where \(o^{ql}\), \(\forall 0\le q,l\le 7\), denotes a component of \(\text {mat}(o)\), which is a matrix defined by:
Now, by denoting \(\text {vec}(o):=(o^{0},\ldots ,o^{7})^{T}\), NN (2) will be transformed as:
\(\forall i\in \{1,\ldots ,N\}\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\).
Finally, with the notations \(\check{C}=\text {diag}(c_{1}I_{8},\ldots ,c_{N}I_{8}),\) \(\check{A}=\left( \text {mat}(a_{ij})\right) _{1\le i,j\le N}\), \(\check{B}=\left( \text {mat}(b_{ij})\right) _{1\le i,j\le N}\), \(\check{G}=\left( \text {mat}(g_{ij})\right) _{1\le i,j\le N}\), \(\check{\mathfrak {o}}(t):=(\text {vec}(\mathfrak {o}_{1}(t))^{T},\ldots ,\text {vec}(\mathfrak {o}_{N}(t))^{T})^{T}\), \(\check{f}(\mathfrak {o}(t)):=(\text {vec}(\tilde{f}_{1}(\mathfrak {o}_{1}(t)))^{T},\ldots ,\text {vec}(\tilde{f}_{N}(\mathfrak {o}_{N}(t)))^{T})^{T}\), NN (2) becomes:
On the other hand, we will define the following system as the response NN and regard NN (1) as the drive NN, for the analysis of synchronization:
\(\forall i\in \{1,\ldots ,N\}\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\), and \(p_{i}(t)\in \mathbb {O}\) – states at \(t\in [0,+\infty )_{\mathbb {T}}\) and \(u_{i}(t)\in \mathbb {O}\) – control inputs at \(t\in [0,+\infty )_{\mathbb {T}}\).
The initial conditions of NN (5) are:
where \(\psi _{i}\in \mathcal {C}([-\omega ,0]_{\mathbb {T}},\mathbb {O})\), \(\forall i\in \{1,\ldots ,N\}\).
In this new setting, by considering (1) and (5), and making the notation \(\varpi _{i}(t)=p_{i}(t)-o_{i}(t)\), the error system will have the following expression:
\(\forall i\in \{1,\ldots ,N\}\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\), where \(\tilde{f}_{j}(\varpi _{j}(t))=f_{j}(\varpi _{j}(t)+o_{j}(t))-f_{j}(o_{j}(t))\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\), \(\forall j\in \{1,\ldots ,N\}\).
The initial conditions of NN (6) are now:
where \(\phi _{i}\in \mathcal {C}([-\omega ,0]_{\mathbb {T}},\mathbb {O})\), \(\forall i\in \{1,\ldots ,N\}\).
Similarly as before, the OVNNs (6) can be converted into a real-valued system of equations, which, in matrix form, can be designated as:
We also need to make the following assumption regarding the activation functions:
Assumption 1
(Popa (2023)) “The activation functions \(f_{j}\) satisfy, \(\forall o,\ o'\in \mathbb {O}\), the following Lipschitz conditions:
\(\forall j\in \{1,\ldots ,N\}\), \(\forall 0\le q\le 7\), where \(l_{j}^{q}>0\), represent the Lipschitz constants. Furthermore, we denote \(L:=\text {diag}(l_{1}^{0},\ldots ,l_{1}^{7},\ldots ,l_{N}^{0},\ldots ,l_{N}^{7})\in \mathbb {R}^{8N\times 8N}\).”
3 Main results
For the first two theorems, system (1) is considered to not possess leakage delay, i.e., \(\delta =0\):
This system can also be written as:
\(\forall t\in [0,+\infty )_{\mathbb {T}}\), \(\forall i\in \{1,\ldots ,8N\}\).
Theorem 4
In the setting of Assumption 1, if there exist positive definite (PD) matrix \(P\in \mathbb {R}^{8N\times 8N}\), diagonal PD matrices \(R_{1}\), \(R_{2}\in \mathbb {R}^{8N\times 8N}\), any matrices \(N_{1}\), \(N_{2}\), \(N_{3}\), \(N_{4}\), \(N_{5}\in \mathbb {R}^{8N\times 8N}\), positive numbers \(\alpha _{1}\), \(\alpha _{21}\), \(\alpha _{22}\), such that \(\alpha _{1}>\alpha _{21}+\alpha _{22}\), \(-\alpha _{1}\in \mathcal {R}^{+}\), and the subsequent LMI holds:
where \(\Omega _{1,1}=-PC-C^{T}P+L^{T}R_{1}L-N_{2}C-C^{T}N_{2}^{T}+\alpha _{1}P\), \(\Omega _{1,2}=-C^{T}N_{1}^{T}-N_{2}\), \(\Omega _{1,5}=PA+N_{2}A+C^{T}N_{3}^{T}\), \(\Omega _{1,6}=PB+N_{2}B+C^{T}N_{4}^{T}\), \(\Omega _{1,7}=PG+N_{2}G+C^{T}N_{5}^{T}\), \(\Omega _{2,2}=\hat{\mu }P-N_{1}-N_{1}^{T}\), \(\Omega _{2,5}=N_{1}A+N_{3}^{T}\), \(\Omega _{2,6}=N_{1}B+N_{4}^{T}\), \(\Omega _{2,7}=N_{1}G+N_{5}^{T}\), \(\Omega _{3,3}=L^{T}R_{2}L-\alpha _{21}P\), \(\Omega _{4,4}=-\alpha _{22}P\), \(\Omega _{5,5}=-R_{1}-N_{3}A-A^{T}N_{3}^{T}\), \(\Omega _{5,6}=-N_{3}B-A^{T}N_{4}^{T}\), \(\Omega _{5,7}=-N_{3}G-A^{T}N_{5}^{T}\), \(\Omega _{6,6}=-R_{2}-N_{4}B-B^{T}N_{4}^{T}\), \(\Omega _{6,7}=-N_{4}G-B^{T}N_{5}^{T}\), \(\Omega _{7,7}=-N_{5}G-G^{T}N_{5}^{T}\), then the equilibrium point of system (1) is exponentially stable.
Proof
We begin by constructing the subsequent Lyapunov-type function:
Now, we take, along the positive half trajectory of system (8), the \(\Delta \)-derivative of V, and, also taking Lemma 1 into consideration, we have that:
From Assumption 1, we have that the subsequent inequalities are true:
for some diagonal PD matrices \(R_{1},R_{2}\in \mathbb {R}^{8N\times 8N}\).
For any matrices \(N_{1},N_{2},N_{3},N_{4},N_{5}\in \mathbb {R}^{8N\times 8N}\), the following equality is also true:
From relations (10)–(13), we obtain that:
where, for the last inequality, we used hypothesis (9) and
From Lemma 2 applied to (14), we get that
which means that
that is to say that
This proves the exponential stability of the equilibrium point of system (1), which is exactly what we wanted to obtain. \(\square \)
Theorem 5
In the setting of Assumption 1, if there exist positive numbers \(\rho _{i}\), \(0\le i\le 8N\), such that \(\alpha _{1}>\alpha _{21}+\alpha _{22}\), \(-\alpha _{1}\in \mathcal {R}^{+}\), where \(\alpha _{1}=\min _{1\le i\le 8N}\left\{ c_{i}-\sum _{j=1}^{8N}\frac{\rho _{j}}{\rho _{i}}l_{i}|\check{a}_{ji}|\right\} \), \(\alpha _{21}=\max _{1\le i\le 8N}\left\{ \sum _{j=1}^{8N}\frac{\rho _{j}}{\rho _{i}}l_{i}|\check{b}_{ji}|\right\} \), \(\alpha _{22}=\max _{1\le i\le 8N} \left\{ \sum _{i=1}^{8N}\frac{\rho _{j}}{\rho _{i}}l_{i}\eta |\check{g}_{ji}|\right\} \), then the equilibrium point of system (1) is exponentially stable.
Proof
Construct the following Lyapunov-like function:
By taking, along the positive half trajectory of system (8), the \(\Delta \)-derivative of V, and, also taking Lemma 3 and Assumption 1 into consideration, we have that:
From Lemma 2 applied to (15), we get that
which means that
that is to say that
This proves the exponential stability of the equilibrium point of system (1), which is exactly what we wanted to obtain. \(\square \)
Remark 1
The exponential stability of OVNNs was studied in the available literature very rarely, for example, in Popa (2018a), Popa (2018b). Theorems 4 and 5 give sufficient conditions expressed as LMIs and algebraic inequalities, respectively, which ensure the exponential stability of OVNNs with time-varying and distributed delays defined on time scales. Such a general model has not been yet considered in the literature, to our awareness, which means that the obtained results are not directly comparable with the ones already available. However, they can be particularized for simpler models, such as for OVNNs defined in continuous time or discrete time, or even for CVNNs or QVNNs defined on time scales, which again proves the generality of our theorems.
Next, we will use the following controller of state feedback type to achieve synchronization between NNs (1) and (5):
in which \(K_{1},\;K_{2},\;K_{3},\;K_{4}\in \mathbb {R}^{8N\times 8N}\) constitute the control gain matrices. Using this controller, NN (7) becomes:
Theorem 6
In the setting of Assumption 1, if there exist PD matrix \(P\in \mathbb {R}^{8N\times 8N}\), diagonal PD matrices \(R_{1}\), \(R_{2}\in \mathbb {R}^{8N\times 8N}\), any matrices \(N_{1}\), \(N_{2}\), \(N_{3}\), \(N_{4}\), \(N_{5}\), \(N_{6}\), \(N_{7}\), \(N_{8}\in \mathbb {R}^{8N\times 8N}\), positive numbers \(\alpha _{1}\), \(\alpha _{21}\), \(\alpha _{22}\), \(\alpha _{23}\), such that \(\alpha _{1}>\alpha _{21}+\alpha _{22}+\alpha _{23}\), \(-\alpha _{1}\in \mathcal {R}^{+}\), and the subsequent LMI holds:
where \(\Omega _{1,1}=-PK_{1}-K_{1}P+L^{T}R_{1}L-N_{2}K_{1}-K_{1}N_{2}^{T}+\alpha _{1}P\), \(\Omega _{1,2}=-K_{1}N_{1}^{T}-N_{2}\), \(\Omega _{1,3}=-P(C+K_{2})-N_{2}(C+K_{2})-K_{1}N_{3}^{T}\), \(\Omega _{1,4}=-PK_{3}-N_{2}K_{3}-K_{1}N_{4}^{T}\), \(\Omega _{1,6}=PA+N_{2}A+K_{1}N_{6}^{T}\), \(\Omega _{1,7}=PB+N_{2}B+K_{1}N_{7}^{T}\), \(\Omega _{1,8}=-PK_{4}-N_{2}K_{4}-K_{1}N_{5}^{T}\), \(\Omega _{1,9}=PG+N_{2}G+K_{1}N_{8}^{T}\), \(\Omega _{2,2}=\hat{\mu }P-N_{1}-N_{1}^{T}\), \(\Omega _{2,3}=-N_{1}(C+K_{2})-N_{3}^{T}\), \(\Omega _{2,4}=-N_{1}K_{3}-N_{4}^{T}\), \(\Omega _{2,6}=N_{1}A+N_{6}^{T}\), \(\Omega _{2,7}=N_{1}B+N_{7}^{T}\), \(\Omega _{2,8}=-N_{1}K_{4}-N_{5}^{T}\), \(\Omega _{2,9}=N_{1}G+N_{8}^{T}\), \(\Omega _{3,3}=-N_{3}(C+K_{2})-(C+K_{2})N_{3}^{T}-\alpha _{21}P\), \(\Omega _{3,4}=-N_{3}K_{3}-(C+K_{2})N_{4}^{T}\), \(\Omega _{3,6}=N_{3}A+(C+K_{2})N_{6}^{T}\), \(\Omega _{3,7}=N_{3}B+(C+K_{2})N_{7}^{T}\), \(\Omega _{3,8}=-N_{3}K_{4}-(C+K_{2})N_{5}^{T}\), \(\Omega _{3,9}=N_{3}G+(C+K_{2})N_{8}^{T}\), \(\Omega _{4,4}=L^{T}R_{2}L-N_{4}K_{3}-K_{3}N_{4}^{T}-\alpha _{22}P\), \(\Omega _{4,6}=N_{4}A+K_{3}N_{6}^{T}\), \(\Omega _{4,7}=N_{4}B+K_{3}N_{7}^{T}\), \(\Omega _{4,8}=-N_{4}K_{4}-K_{3}N_{5}^{T}\), \(\Omega _{4,9}=N_{4}G+K_{3}N_{8}^{T}\), \(\Omega _{5,5}=-\alpha _{23}P\), \(\Omega _{6,6}=-R_{1}-N_{6}A-A^{T}N_{6}^{T}\), \(\Omega _{6,7}=-N_{6}B-A^{T}N_{7}^{T}\), \(\Omega _{6,8}=N_{6}K_{4}+A^{T}N_{5}^{T}\), \(\Omega _{6,9}=-N_{6}G-A^{T}N_{8}^{T}\), \(\Omega _{7,7}=-R_{2}-N_{7}B-B^{T}N_{7}^{T}\), \(\Omega _{7,8}=N_{7}K_{4}+B^{T}N_{5}^{T}\), \(\Omega _{7,9}=-N_{7}G-B^{T}N_{8}^{T}\), \(\Omega _{8,8}=-N_{5}K_{4}-K_{4}N_{5}^{T}\), \(\Omega _{8,9}=N_{5}G+K_{4}N_{8}^{T}\), \(\Omega _{9,9}=-N_{8}G-G^{T}N_{8}^{T}\), then drive NN (1) is exponentially synchronized with response NN (5) based on state feedback controller (16).
Proof
The subsequent Lyapunov-type function will be constructed firstly:
By taking, along the positive half trajectory of system (17), the \(\Delta \)-derivative of V, and also taking Lemma 1 into consideration, we have that:
From Assumption 1, we have that the subsequent inequalities are true:
for some diagonal PD matrices \(R_{1},R_{2}\in \mathbb {R}^{8N\times 8N}\).
For any matrices, \(N_{1},N_{2},N_{3},N_{4},N_{5},N_{6},N_{7},N_{8}\in \mathbb {R}^{8N\times 8N}\) the following equality is also true:
From relations (19)–(22), we obtain that:
where, for the last inequality, we used hypothesis (18) and
From Lemma 2 applied to (23), we get that
which means that
that is to say that
This proves the exponential synchronization of drive NN (1) with response NN (5) under state feedback controller (17), which is exactly what we wanted to obtain. \(\square \)
For the next theorem, we will write system (17) as:
\(\forall t\in [0,+\infty )_{\mathbb {T}}\), \(\forall i\in \{1,\ldots ,8N\}\).
Theorem 7
In the setting of Assumption 1, if there exist positive numbers \(\rho _{i}\), \(0\le i\le 8N\), such that \(\alpha _{1}>\alpha _{21}+\alpha _{22}+\alpha _{23}\), \(-\alpha _{1}\in \mathcal {R}^{+}\), where \(\alpha _{1}=\min _{1\le i\le 8N}\left\{ k_{1i}-\sum _{j=1}^{8N}\frac{\rho _{j}}{\rho _{i}}l_{i}|\check{a}_{ji}|\right\} \), \(\alpha _{21}=\max _{1\le i\le 8N}\left\{ \sum _{j=1}^{8N}\frac{\rho _{j}}{\rho _{i}}l_{i}|\check{b}_{ji}|\right\} \), \(\alpha _{22}=\max _{1\le i\le 8N}\left\{ \check{c_{i}}+k_{2i}\right\} \), \(\alpha _{23}=\max _{1\le i\le 8N}\left\{ k_{4i}l_{i}\eta +\sum _{i=1}^{8N}\frac{\rho _{j}}{\rho _{i}}l_{i}\eta |\check{g}_{ji}|\right\} \), then drive NN (1) is exponentially synchronized with response NN (5) based on state feedback controller (16).
Proof
Construct the following Lyapunov-like function:
By taking, along the positive half trajectory of system (24), the \(\Delta \)-derivative of V, and also taking Lemma 3 and Assumption 1 into consideration, we have that:
From Lemma 2 applied to (25), we get that
which means that
that is to say that
This proves the exponential synchronization of drive NN (1) with response NN (5) under state feedback controller (17), which is exactly what we wanted to obtain. \(\square \)
Remark 2
Results concerning the exponential synchronization of OVNNs are even rarer than the ones concerning exponential stability. For example, paper (Popa 2023) discusses the Mittag–Leffler synchronization of fractional-order OVNNs and paper (Chouhan et al. 2023) discusses the fixed-time synchronization of OVNNs. Theorems 6 and 7 put forward sufficient criteria expressed as LMIs and algebraic inequalities, respectively, for the exponential synchronization of OVNNs with leakage, time-varying, and distributed delays defined on time scales. The model is even more general than the one in Theorems 4 and 5, and also there are no corresponding results in the available literature, to our best knowledge, focusing on such a general model. This means that Theorems 6 and 7 are not directly comparable with already existing ones, but can be particularized for OVNNs defined on time scales with fewer types of delays, for OVNNs defined in continous time or discrete time, or even for CVNNs or QVNNs defined on time scales with the same types of delays.
4 Numerical examples
Example 1
In the first example, take time scale \(\mathbb {T}=0.1\mathbb {Z}\), which means that \(\hat{\mu }=0.1\).
Let the following OVNN with two neurons and leakage, time-varying, and distributed delays be defined on time scale \(\mathbb {T}\):
\(\forall i\in \{1,2\}\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\). Now, suppose system (26) has a unique equilibrium point, designated by \(\tilde{o}=(\tilde{o}_{1},\tilde{o}_{2})^{T}\). In this setting, if we denote \(\mathfrak {o}_{i}(t)=o_{i}(t)-\tilde{o}_{i}\), system (26) becomes:
\(\forall i\in \{1,2\}\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\), where \(\tilde{f}_{j}(\mathfrak {o}_{j}(t))=f_{j}(\mathfrak {o}_{j}(t)+\tilde{o}_{j})-f_{j}(\tilde{o}_{j})\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\), \(\forall j\in \{1,2\}\).
The parameters are as follows:
Thus, Assumption 1 is satisfied by the activation functions, and \(L=\left[ \begin{array}{cc} 0.025I_{8} &{} 0\\ 0 &{} 0.025I_{8} \end{array}\right] \).
Also, we take \(\delta =0\) (no leakage delay), \(\tau (t)=0.4|\sin t|\), \(\eta =0.5\), hence \(\tau =0.4\) and \(\omega =\max \{\delta ,\tau ,\eta \}=0.5\), and \(\alpha _{1}=4.7\), \(\alpha _{21}=2.2\), \(\alpha _{22}=2.3\), which verify \(\alpha _{1}>\alpha _{21}+\alpha _{22}\). Taking all of the above into consideration, we can conclude that all the conditions of Theorem 4 are met, which means that we can solve the LMI in (9), obtaining \(R_{1}=\text {diag}(1.4962I_{8},1.6684I_{8})\), \(R_{2}=\text {diag}(1.7715I_{8},1.7993I_{8})\), (in order to reduce the length of the paper, we do not provide the values of the other matrices). Thus, the unique equilibrium point of system (26) is exponentially stable.
In Figs. 1, 2 are depicted the state trajectories of octonions \(o_{1}\) and \(o_{2}\) for NN (26), starting from 8 initial values.
Example 2
For this example, consider the same time scale \(\mathbb {T}=0.1\mathbb {Z}\) and the same system (26), which is again assumed to have a unique equilibrium point and can consequently be transformed into system (27), but, this time, with the following parameters:
Thus, Assumption 1 is satisfied by the activation functions, and \(L=\left[ \begin{array}{cc} 0.025I_{8} &{} 0\\ 0 &{} 0.025I_{8} \end{array}\right] \).
We also take \(\delta =0\) (no leakage delay), \(\tau (t)=0.3|\cos t|\), \(\eta =0.5\), hence \(\tau =0.3\) and \(\omega =\max \{\delta ,\tau ,\eta \}=0.5\), and compute \(\alpha _{1}=3.74\), \(\alpha _{21}=0.125\), \(\alpha _{22}=0.2013\), which satisfy \(\alpha _{1}>\alpha _{21}+\alpha _{22}\), and \(\rho _{1}=1\), \(\rho _{2}=2\). All of the above allow us to ascertain that all the hypotheses of Theorem 5 are met, which means that the equilibrium point of system (26) is exponentially stable.
Example 3
For the next example, consider time scale \(\mathbb {T}=\mathbb {R}\), from which we deduce that \(\hat{\mu }=0\). In order to study the synchronization property, we will consider system (26) as being the drive NN, and the following system as being the response NN:
\(\forall i\in \{1,2\}\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\). In this new setting, by considering (26) and (28), and making the notation \(\varpi _{i}(t)=p_{i}(t)-o_{i}(t)\), the error system will have the following expression:
\(\forall i\in \{1,2\}\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\), where \(\tilde{f}_{j}(\varpi _{j}(t))=f_{j}(\varpi _{j}(t)+o_{j}(t))-f_{j}(o_{j}(t))\), \(\forall t\in [0,+\infty )_{\mathbb {T}}\), \(\forall j\in \{1,2\}\).
Next, the subsequent controller of state feedback type will be employed to achieve synchronization between systems (26) and (28):
in which \(K_{1},\;K_{2},\;K_{3},\;K_{4}\in \mathbb {R}^{8N\times 8N}\) constitute the control gain matrices. Using this controller, NN (29) becomes:
The parameters, in this example, are taken as:
Thus, Assumption 1 is satisfied by the activation functions, and \(L=\left[ \begin{array}{cc} 0.025I_{8} &{} 0\\ 0 &{} 0.025I_{8} \end{array}\right] \).
The delays are \(\delta =0.08\), \(\tau (t)=0.3|\sin t|\), so \(\tau =0.3\), and \(\eta =0.4\), so \(\omega =\max \{\delta ,\tau ,\eta \}=0.4\), and also take \(\alpha _{1}=7\), \(\alpha _{21}=2.2\), \(\alpha _{22}=2.3\), \(\alpha _{23}=2.4\), which satisfy \(\alpha _{1}>\alpha _{21}+\alpha _{22}+\alpha _{23}\). The control gain matrices are designed as:
Solving LMI (18), we obtain \(R_{1}=\text {diag}(0.46336,1.0832)\), \(R_{2}=\text {diag}(0.36233,0.37425)\), \(R_{3}=\text {diag}(1.9368,2.1147)\), \(R_{4}=\text {diag}(1.7712,2.3323)\) (the other matrices’ values are not displayed for brevity’s sake), which means that all the conditions of Theorem 6 are met, reaching to the conclusion that drive NN (26) is exponentially synchronized with response NN (28) based on state feedback controller (30).
Figures 3, 4 depict graphically the trajectories of the components of the states \(\check{\varpi }_{1}\) and \(\check{\varpi }_{2}\), starting from 8 initial points.
Example 4
We take the same time scale \(\mathbb {T}=\mathbb {R}\) for our last example, and try to study the same problem of realizing synchronization between drive NN (26) and response NN (28) based on controller of state feedback type (30), but now having the following parameters:
Thus, Assumption 1 is satisfied by the activation functions, and \(L=\left[ \begin{array}{cc} 0.025I_{8} &{} 0\\ 0 &{} 0.025I_{8} \end{array}\right] \).
Now, we take \(\delta =0.01\), \(\tau (t)=0.1|\sin t|\), \(\eta =0.05\), from which we have \(\tau =0.1\) and \(\omega =\max \{\delta ,\tau ,\eta \}=0.1\), and we compute \(\alpha _{1}=9.7625\), \(\alpha _{21}=0.4025\), \(\alpha _{22}=5.3\), \(\alpha _{23}=0.0205\), from which we have \(\alpha _{1}>\alpha _{21}+\alpha _{22}+\alpha _{23}\), and \(\rho _{1}=1\), \(\rho _{2}=2\). The conditions of Theorem 7 are all met, reaching to the conclusion that drive NN (26) is exponentially synchronized with response NN (28) based on state feedback controller (30).
5 Conclusions
In this paper, a general model of OVNNs defined on time scales with leakage, time-varying, and distributed delays was put forward. First of all, to avoid the problems related to the non-commutativity and non-associativity of the octonion algebra, the system with octonion values was transformed into one with real values. Then, a Halanay inequality specially designed for time scales was used in order to obtain sufficient conditions, expressed as scalar inequalities and LMIs, which guarantee the exponential stability of the proposed model, using two general Lyapunov-like functions. Then, using the same Halanay inequality and the same Lyapunov-like functions, but also a state feedback controller, sufficient criteria for the exponential synchronization of the NNs put forward were obtained. Finally, one numerical example is provided to illustrate each of the four theorems deduced in the paper.
The results presented in this research are sufficiently general to allow for their particularization for both discrete time and continuous time OVNNs, as well as any hybrid combination of the two. They can also be particularized for CVNNs or QVNNs, for which, as far as we are aware, no similar results have been published in the literature. Additionally, the techniques utilized in the paper can be applied to other types of models defined on time scales, such as NNs with inertial terms, reaction–diffusion terms, Markov jump parameters, or impulsive effects to obtain sufficient criteria for other dynamic properties, such as multistability, dissipativity, passivity, etc. These show promise for future research.
References
Adıvar M, Bohner EA (2011) Halanay type inequalities on time scales with applications. Nonlinear Anal Theory Methods Appl 74(18):7519–7531. https://doi.org/10.1016/j.na.2011.08.007
Adıvar M, Raffoul YN (2020) Stability, periodicity and boundedness in functional dynamical systems on time scales. Springer, Switzerland. https://doi.org/10.1007/978-3-030-42117-5
Bauer S, Leon FP (2016) Hyperspectral fluorescence data fusion using quaternion and octonion phase. In: 2016 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI). IEEE, Baden-Baden, Germany. https://doi.org/10.1109/mfi.2016.7849555
Błaszczyk Ł, Snopek KM (2017) Octonion Fourier transform of real-valued functions of three variables - selected properties and examples. Signal Process 136:29–37. https://doi.org/10.1016/j.sigpro.2016.11.021
Bohner M, Peterson A (2001) Dynamic equations on time scales. Birkhauser, Boston. https://doi.org/10.1007/978-1-4612-0201-1
Chanyal BC (2013) Octonion massive electrodynamics. Gen Relativ Gravit. https://doi.org/10.1007/s10714-013-1646-2
Chen A, Du D (2008) Global exponential stability of delayed BAM network on time scale. Neurocomputing 71(16–18):3582–3588. https://doi.org/10.1016/j.neucom.2008.06.004
Chouhan SS, Kumar R, Sarkar S, Das S (2022) Multistability analysis of octonion-valued neural networks with time-varying delays. Inform Sci 609:1412–1434. https://doi.org/10.1016/j.ins.2022.07.123
Chouhan SS, kumar U, Das S, Cao J (2023) Fixed time synchronization of octonion valued neural networks with time varying delays. Eng Appl Artif Intellig 118:105684. https://doi.org/10.1016/j.engappai.2022.105684
Demir S (2012) Hyperbolic octonion formulation of gravitational field equations. Int J Theor Phys 52(1):105–116. https://doi.org/10.1007/s10773-012-1307-3
Demir S, Tanişli M (2016) Hyperbolic octonion formulation of the fluid Maxwell equations. J Korean Phys Soc 68(5):616–623. https://doi.org/10.3938/jkps.68.616
Gao H-Y, Lam K-M (2014) From quaternion to octonion: Feature-based image saliency detection. In: 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, Florence, Italy. https://doi.org/10.1109/icassp.2014.6854112
Gao H-Y, Lam K-M (2014) Salient object detection using octonion with Bayesian inference. In: 2014 IEEE International Conference on Image Processing (ICIP). IEEE, Paris, France. https://doi.org/10.1109/icip.2014.7025666
Gao J, Dai L (2023) Weighted pseudo almost periodic solutions of octonion-valued neural networks with mixed time-varying delays and leakage delays. AIMS Math 8(6):14867–14893. https://doi.org/10.3934/math.2023760
Hilger S (1990) Analysis on measure chains – a unified approach to continuous and discrete calculus. Results Math 18(1–2):18–56. https://doi.org/10.1007/bf03323153
Huang X, Li Y (2023) Weyl almost periodic solutions of octonion-valued high-order fuzzy neural networks with delays. Comput Appl Math. https://doi.org/10.1007/s40314-023-02294-x
Kandasamy U, Rajan R (2020) Hopf bifurcation of a fractional-order octonion-valued neural networks with time delays. Discrete Contin Dynam Syst S 13(9):2537–2559. https://doi.org/10.3934/dcdss.2020137
Kassim MD, Tatar N-e (2021) A neutral fractional Halanay inequality and application to a Cohen-Grossberg neural network system. Math Methods Appl Sci 44(13):10460–10476. https://doi.org/10.1002/mma.7422
Li Y, Huang X (2022) Almost periodic oscillation of octonion-valued neural networks with delays on time scales. Int J Syst Sci 54(4):780–790. https://doi.org/10.1080/00207721.2022.2145859
Li B, Cao Y, Li Y (2023) Almost automorphic solutions in distribution for octonion-valued stochastic recurrent neural networks with time-varying delays. Int J Syst Sci 55(1):102–118. https://doi.org/10.1080/00207721.2023.2268770
Martynyuk AA (2016) Stability theory for dynamic equations on time scales. Springer, Switzerland. https://doi.org/10.1007/978-3-319-42213-8
Mohamad S, Gopalsamy K (2000) Dynamics of a class of discrete-time neural networks and their continuous-time counterparts. Math Comput Simul 53(1–2):1–39. https://doi.org/10.1016/s0378-4754(00)00168-3
Mohamad S, Gopalsamy K (2000) Continuous and discrete Halanay-type inequalities. Bull Austral Math Soc 61(3):371–385. https://doi.org/10.1017/s0004972700022413
Ou B (2020) Halanay inequality on time scales with unbounded coefficients and its applications. Indian J Pure Appl Math 51(3):1023–1038. https://doi.org/10.1007/s13226-020-0447-z
Ou B, Jia B, Erbe L (2015) An extended Halanay inequality of integral type on time scales. Electron J Qualit Theory Differ Equat 38:1–11. https://doi.org/10.14232/ejqtde.2015.1.38
Ou B, Lin Q, Du F, Jia B (2016) An extended Halanay inequality with unbounded coefficient functions on time scales. J Inequal Appl. https://doi.org/10.1186/s13660-016-1259-x
Popa C-A (2016) Octonion-Valued Neural Networks. In: Artificial Neural Networks and Machine Learning – ICANN 2016, pp. 435–443. https://doi.org/10.1007/978-3-319-44778-0_51
Popa C-A (2018) Global exponential stability of octonion-valued neural networks with leakage delay and mixed delays. Neural Netw 105:277–293. https://doi.org/10.1016/j.neunet.2018.05.006
Popa C-A (2018) Global exponential stability of neutral-type octonion-valued neural networks with time-varying delays. Neurocomputing 309:117–133. https://doi.org/10.1016/j.neucom.2018.05.004
Popa C-A (2023) Asymptotic and mittag-leffler synchronization of fractional-order octonion-valued neural networks with neutral-type and mixed delays. Fractal Fract 7(11):830. https://doi.org/10.3390/fractalfract7110830
Snopek KM (2015) Quaternions and octonions in signal processing – fundamentals and some new results. Przeglad Telekomunikacyjny + Wiadomosci Telekomunikacyjne 6:618–622
Wan P, Zeng Z (2021) Quasisynchronization of delayed neural networks with discontinuous activation functions on time scales via event-triggered control. IEEE Trans Cybernet. https://doi.org/10.1109/tcyb.2021.3088725
Wan P, Zeng Z (2022) Global exponential stability of impulsive delayed neural networks on time scales based on convex combination method. IEEE Trans Syst Man Cybernet Syst 52(5):3015–3024. https://doi.org/10.1109/tsmc.2021.3061971
Wan P, Zeng Z (2022) Lagrange stability of fuzzy memristive neural networks on time scales with discrete time varying and infinite distributed delays. IEEE Trans Fuzzy Syst 30(8):3138–3151. https://doi.org/10.1109/tfuzz.2021.3105178
Wang W (2010) A generalized halanay inequality for stability of nonlinear neutral functional differential equations. J Inequal Appl 2010(1):475019. https://doi.org/10.1155/2010/475019
Wang R, Xiang G, Zhang F (2016) L1-norm minimization for octonion signals. In: 2016 International Conference on Audio, Language and Image Processing (ICALIP). IEEE, Shanghai, China. https://doi.org/10.1109/icalip.2016.7846602
Wen L, Yu Y, Wang W (2008) Generalized Halanay inequalities for dissipativity of Volterra functional differential equations. J Math Anal Appl 347(1):169–178. https://doi.org/10.1016/j.jmaa.2008.05.007
Wen H, Shu S, Wen L (2018) A new generalization of Halanay-type inequality and its applications. J Inequal Appl. https://doi.org/10.1186/s13660-018-1894-5
Xiao Q, Zeng Z (2017) Scale-limited lagrange stability and finite-time synchronization for memristive recurrent neural networks on time scales. IEEE Trans Cybernet 47(10):2984–2994. https://doi.org/10.1109/tcyb.2017.2676978
Xiao Q, Zeng Z (2018) Lagrange stability for T-S fuzzy memristive neural networks with time-varying delays on time scales. IEEE Trans Fuzzy Syst 26(3):1091–1103. https://doi.org/10.1109/tfuzz.2017.2704059
Xiao Q, Huang T, Zeng Z (2018) Passivity and passification of fuzzy memristive inertial neural networks on time scales. IEEE Trans Fuzzy Syst 26(6):3342–3355. https://doi.org/10.1109/tfuzz.2018.2825306
Xiao Q, Huang T, Zeng Z (2020) Stabilization of nonautonomous recurrent neural networks with bounded and unbounded delays on time scales. IEEE Trans Cybernet 50(10):4307–4317. https://doi.org/10.1109/tcyb.2019.2922207
Acknowledgements
This work is supported by a grant of the Ministry of Research, Innovation and Digitization, CNCS - UEFISCDI, project number PN-III-P1\(-\)1.1-PD-2021-0345, within PNCDI III.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The author declares the following financial interests which may be considered as potential Conflict of interest. Călin-Adrian Popa reports financial support was provided by Ministry of Research, Innovation and Digitization, CNCS - UEFISCDI.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Popa, CA. Stability and synchronization of octonion-valued neural networks with leakage and mixed delays on time scales. Comp. Appl. Math. 43, 300 (2024). https://doi.org/10.1007/s40314-024-02820-5
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s40314-024-02820-5
Keywords
- Octonion-valued neural networks (OVNNs)
- Stability analysis
- Synchronization analysis
- Time delays
- Time scales