The univariate multinode Shepard method for the Caputo fractional derivatives: from Approximation to the solution of Bagley-Torvik equation Journal: Fract. Calc. Appl. Anal. Francesco Dell’Accio 1, Filomena Di Tommaso 1, Ilde Ferrara 1 Abstract In this paper, we approximate the fractional derivative of a given function using the univariate multinode Shepard method through the Gauss-Jacobi quadrature formula. Subsequently, the proposed method is applied to the numerical solution of boundary value problems (BVPs) and initial value problems (IVPs), specifically addressing the Bagley-Torvik equations. Experimental results confirm the method’s effectiveness, particularly in accurately approximating the Bagley-Torvik equation for both BVPs and IVPs. ∎ MSC 65M70 26A33 34A08 1 Introduction Fractional calculus is a very old mathematical concept, dating back to the 17th century, that deals with the integration and differentiation of arbitrary order Fractional_derivative_BTE ; math7050407 . In the early stages of the history of integer-order calculus, several mathematicians, including L’Hospital, Leibniz, and others, began to explore the idea of fractional calculus, but the concept did not gain much attention or further development due to the lack of practical applications Fractional_derivative_book . In recent decades, however, fractional calculus has experienced significant theoretical and practical advancements esempi_numerici , with applications emerging in a wide range of fields such as physics and engineering Fractional_derivative_BTE . The numerical solution of fractional differential equations (FDEs) has become an area of growing interest, as these equations offer a more accurate explanation of many natural physical phenomena, facilitating the modeling of complex dynamic systems esempi_numerici . In more recent years, both mathematicians and physicists have made considerable efforts to develop robust and stable numerical and analytical methods for solving fractional differential equations (FDEs) of physical relevance, driven by the increasing number of applications of FDEs Application_BTE . In particular, FDEs are used to characterize complex systems with memory and hereditary properties, such as anomalous diffusion, signal processing, stock market dynamics, vibration theory, electrical circuits, stability analysis, control theory, stochastic analysis, time-delay systems, and bioengineering (see article and the references therein). They also have applications in thermodynamics, biophysics, blood flow phenomena, rheology, electrodynamics, electrochemistry, electromagnetism, continuum and statistical mechanics, and dynamical systems Fractional_derivative_BTE ; Fractional_derivative_book ; article2 ; article3 . One of the main engineering applications of fractional calculus is in the study of the viscoelasticity of materials Fractional_derivative_legendre , where the first significant application seems to have been made by Bagley and Torvik, in modeling the motion of a rigid plate immersed in a Newtonian fluid entropy ; article4 . As well known, fractional derivatives have various definitions, among which the most commonly used in practical applications are the Riemann-Liouville and the Caputo fractional derivatives. The Riemann–Liouville definition is historically the first, and the theory behind it is now well established; however, there are some difficulties in applying it to real-world problems DEMIRCI20122754 . To solve these difficulties, Caputo reformulated the traditional Riemann–Liouville definition, allowing for the use of integer-order initial conditions when solving fractional-order differential equations article2 . The Caputo’s fractional derivative of order \alpha>0 of the function f(x) is defined as (D^{\alpha}f)(x)=\frac{1}{\Gamma(m-\alpha)}\int_{0}^{x}(x-t)^{m-\alpha-1}f^{(m)}(t)\,dt,\hskip 10.00002ptx>0, (1.1) with m-1<\alpha0 is a parameter related to the differentiability class of B_{\mu,k}(x) DellAccio:2018 . The multinode functions are non-negative and form a partition of unity, and instead of being cardinal, they satisfy the following properties DellAccio:2018 : (i) B_{\mu,k}\left(x_{j}\right)=0 , for \mu>0 ; (ii) \sum\limits_{k\in\mathcal{K}_{i}}B_{\mu,k}\left(x_{i}\right)=1 , for \mu>0 ; (iii) B_{\mu,k}^{\left(m\right)}\left(x_{j}\right)=0 , for \mu>m\geq 1 ; (iv) \sum\limits_{k\in\mathcal{K}_{i}}B_{\mu,k}^{\left(m\right)}\left(x_{i}\right)=0 , for \mu>m\geq 1 ; where \mathcal{K}_{i}=\{k\in\{1,...,s\}:i\in\{k_{1},...,k_{p}\}\}\neq\emptyset, (2.3) is the set of indices of all subsets of \mathcal{F} that contain x_{i} . 2.2 First and second order derivatives of the multinode functions In this Section, we compute the first and second order derivatives of the multinode functions in view of the application of the multinode Shepard operator to the approximation of fractional derivatives. Proposition 1 For \mu>1 and even, we have B_{\mu,k}^{\prime}(x)=\mu B_{\mu,k}(x)\left(\sum_{\ell=1}^{s}B_{\mu,\ell}(x)\sum_{i=1}^{p}\frac{1}{x-x_{\ell_{i}}}-\sum_{i=1}^{p}\frac{1}{x-x_{k_{i}}}\right). (2.4) Proof Let be x_{j}\in X . If we multiply both the numerator and the denominator of ( 2.2 ) by \left|x-x_{j}\right|^{\mu} , we get B_{\mu,k}\left(x\right)=\frac{C_{k}\left(x\right)}{\sum\limits_{\ell=1}^{s}C_{\ell}\left(x\right)}, (2.5) where \begin{array}[c]{cc}C_{\ell}\left(x\right)=\left|x-x_{j}\right|^{\mu}\prod\limits_{i=1}^{p}{\left|{x-x_{\ell_{i}}}\right|^{-\mu}},&\ell=1,\dots,s.\end{array} (2.6) Since \mu>1 and even, we can rewrite ( 2.6 ) as C_{\ell}(x)=\left(x-x_{j}\right)^{\mu}D_{\ell}(x), (2.7) where D_{\ell}(x)=\prod_{i=1}^{p}(x-x_{\ell_{i}})^{-\mu}. (2.8) By differentiating the right-hand side of ( 2.5 ), we get B_{\mu,k}^{\prime}(x)=\frac{C_{k}^{\prime}(x)\displaystyle\sum_{\ell=1}^{s}C_{\ell}(x)-C_{k}(x)\sum_{\ell=1}^{s}C_{\ell}^{\prime}(x)}{\left(\displaystyle\sum_{\ell=1}^{s}C_{\ell}(x)\right)^{2}}=\frac{C_{k}^{\prime}(x)}{\displaystyle\sum_{\ell=1}^{s}C_{\ell}(x)}-\frac{\displaystyle\sum_{\ell=1}^{s}C_{\ell}^{\prime}(x)}{\displaystyle\sum_{\ell=1}^{s}C_{\ell}(x)}B_{\mu,k}(x). (2.9) The final step is the computation of C_{\ell}^{\prime}(x) in ( 2.7 ), for which we get C_{\ell}^{\prime}(x)=\mu(x-x_{j})^{\mu-1}D_{\ell}(x)+(x-x_{j})^{\mu}D_{\ell}^{\prime}(x). (2.10) To compute D_{\ell}^{\prime}(x) , we notice that, since \mu is even, it is possible to compute the logarithm of both sides of ( 2.8 ), that is \log D_{\ell}(x)=\log\prod_{i=1}^{p}(x-x_{\ell_{i}})^{-\mu}=\sum_{i=1}^{p}(-\mu)\log(x-x_{\ell_{i}}). (2.11) By differentiating both sides of ( 2.11 ), we get \frac{D_{\ell}^{\prime}(x)}{D_{\ell}(x)}=-\mu\sum_{i=1}^{p}\frac{1}{x-x_{\ell_{i}}} and then D_{\ell}^{\prime}(x)=-\mu D_{\ell}(x)\sum_{i=1}^{p}\frac{1}{x-x_{\ell_{i}}}. (2.12) By substituting ( 2.12 ) in ( 2.10 ) and by some computations, we obtain C_{\ell}^{\prime}(x)=\mu C_{\ell}(x)\left(\frac{1}{x-x_{j}}-\sum_{i=1}^{p}\frac{1}{x-x_{\ell_{i}}}\right). (2.13) Finally, by substituting ( 2.13 ) in ( 2.9 ) and by the partition of unity property, we get \begin{split}B_{\mu,k}^{\prime}(x)&=\frac{\displaystyle\mu C_{k}(x)\left(\frac{1}{x-x_{j}}-\sum_{i=1}^{p}\frac{1}{x-x_{k_{i}}}\right)}{\displaystyle\sum_{\ell=1}^{s}C_{\ell}(x)}-\frac{\displaystyle\sum_{\ell=1}^{s}\mu C_{\ell}(x)\left(\frac{1}{x-x_{j}}-\sum_{i=1}^{p}\frac{1}{x-x_{\ell_{i}}}\right)}{\displaystyle\sum_{\ell=1}^{s}C_{\ell}(x)}B_{\mu,k}(x)=\\ &=\mu B_{\mu,k}(x)\left(\sum_{\ell=1}^{s}B_{\mu,\ell}(x)\sum_{i=1}^{p}\frac{1}{x-x_{\ell_{i}}}-\sum_{i=1}^{p}\frac{1}{x-x_{k_{i}}}\right).\\ \end{split} Proposition 2 For \mu>2 and even, we have \begin{split}B_{\mu,k}^{\prime\prime}(x)&=-\mu B_{\mu,k}(x)\left(\sum_{\ell=1}^{s}B_{\mu,\ell}(x)\sum_{i=1}^{p}\frac{1}{(x-x_{\ell_{i}})^{2}}-\sum_{i=1}^{p}\frac{1}{(x-x_{k_{i}})^{2}}\right)\\ &+\mu B_{\mu,k}(x)\sum_{\ell=1}^{s}B_{\mu,\ell}^{\prime}(x)\sum_{i=1}^{p}\frac{1}{x-x_{\ell_{i}}}\\ &+\mu B_{\mu,k}^{\prime}(x)\left(\sum_{\ell=1}^{s}B_{\mu,\ell}(x)\sum_{i=1}^{p}\frac{1}{x-x_{\ell_{i}}}-\sum_{i=1}^{p}\frac{1}{x-x_{k_{i}}}\right).\\ \end{split} (2.14) Proof Let us fix \mu>2 and even. By the equations ( 2.5 ), ( 2.7 ) and ( 2.8 ), by differentiating the right-hand side of ( 2.4 ) and by rearranging the resulting terms, we get B_{\mu,k}^{\prime\prime}(x)=\frac{\displaystyle C_{k}^{\prime\prime}(x)}{\displaystyle\sum_{\ell=1}^{s}C_{\ell}(x)}-\frac{\displaystyle\sum_{\ell=1}^{s}C_{\ell}^{\prime\prime}(x)}{\displaystyle\sum_{\ell=1}^{s}C_{\ell}(x)}B_{\mu k}(x)-2\frac{\displaystyle\sum_{\ell=1}^{s}C_{\ell}^{\prime}(x)}{\displaystyle\sum_{\ell=1}^{s}C_{\ell}(x)}B_{\mu k}^{\prime}(x). (2.15) In line with the proof of Proposition 1 , we need to compute the second order derivative of C_{\ell}(x) , that is C_{\ell}^{\prime\prime}(x)=\mu(\mu-1)(x-x_{j})^{\mu-2}D_{\ell}(x)+2\mu(x-x_{j})^{\mu-1}D_{\ell}^{\prime}(x)+(x-x_{j})^{\mu}D_{\ell}^{\prime\prime}(x) (2.16) and, consequently, the second order derivative of D_{\ell}(x) D_{\ell}^{\prime\prime}(x)=-\mu D_{\ell}^{\prime}(x)\sum_{i=1}^{p}\frac{1}{x-x_{\ell_{i}}}+\mu D_{\ell}(x)\sum_{i=1}^{p}\frac{1}{(x-x_{\ell_{i}})^{2}}. (2.17) By substituting ( 2.12 ) and ( 2.17 ) in ( 2.16 ), after some computation, we obtain \begin{split}C_{\ell}^{\prime\prime}(x)=\mu\left(-\frac{C_{\ell}(x)}{(x-x_{j})^{2}}+\sum_{i=1}^{p}\frac{C_{\ell}(x)}{(x-x_{\ell_{i}})^{2}}+\frac{C_{\ell}^{\prime}(x)}{x-x_{j}}-\sum_{i=1}^{p}\frac{C_{\ell}^{\prime}(x)}{x-x_{\ell_{i}}}\right).\end{split} (2.18) Finally, by substituting ( 2.18 ) in ( 2.15 ), by the partition of unity property and by some computations, we get \begin{split}B_{\mu,k}^{\prime\prime}(x)&=-\mu B_{\mu,k}(x)\left(\sum_{\ell=1}^{s}B_{\mu,\ell}(x)\sum_{i=1}^{p}\frac{1}{(x-x_{\ell_{i}})^{2}}-\sum_{i=1}^{p}\frac{1}{(x-x_{k_{i}})^{2}}\right)\\ &+\mu B_{\mu,k}(x)\sum_{\ell=1}^{s}B_{\mu,\ell}^{\prime}(x)\sum_{i=1}^{p}\frac{1}{x-x_{\ell_{i}}}\\ &+\mu B_{\mu,k}^{\prime}(x)\left(\sum_{\ell=1}^{s}B_{\mu,\ell}(x)\sum_{i=1}^{p}\frac{1}{x-x_{\ell_{i}}}-\sum_{i=1}^{p}\frac{1}{x-x_{k_{i}}}\right).\\ \end{split} By following an analogous procedure to that of Proposition 1 and Proposition 2 , we obtain the general formula B_{\mu,k}^{(r)}(x)=\frac{C_{k}^{(r)}(x)}{\displaystyle\sum_{\ell=1}^{s}C_{\ell}(x)}-\sum_{i=0}^{n-1}\binom{r}{i}\left(\frac{\displaystyle\sum_{\ell=1}^{s}C_{\ell}^{(r-i)}(x)}{\displaystyle\sum_{\ell=1}^{s}C_{\ell}(x)}B_{\mu,k}^{(i)}(x)\right). 2.3 The Multinode Shepard operator Let f:I\rightarrow\mathbb{R} be a function of class C^{m} in I , m\geq 0 , sampled at the point x_{i}\in X , and let us denote such values by f_{i}=f(x_{i}),\ i=1,...,n . Moreover, we consider the interpolation polynomials P_{k}\left[f\right]\left(x\right) of degree p-1 on the points in F_{k} , k=1,\dots,s , and we combine them with the multinode functions B_{\mu,k}(x) to get the multinode Shepard operator \mathcal{M}_{\mu}\left[f\right](x)=\sum\limits_{k=1}^{s}B_{\mu,k}(x)P_{k}[f](x),\hskip 10.00002pt\mu>0,\,x\in I. (2.19) The multinode Shepard operator is an interpolation operator at the points of X and reproduces polynomials up to the degree p-1 ( DellAccio:2018 , Proposition 4) . In view of the following applications, it is convenient to rewrite \mathcal{M}_{\mu}\left[f\right] as \displaystyle\mathcal{M}_{\mu}\left[f\right](x)=\sum_{k=1}^{s}\sum_{i=1}^{p}B_{\mu,k}(x)l_{k,i}(x)f_{k_{i}}, where l_{k,i}(x) are the fundamental Lagrange polynomials on the nodes of F_{k} , that is l_{k,i}(x)=\prod_{\begin{subarray}{c}j=1\\ j\neq k_{i}\end{subarray}}^{p}\frac{x-x_{j}}{x_{k_{i}}-x_{j}}, (2.20) satisfying the well-known Kronecker delta property \displaystyle l_{k,i}(x_{k_{j}})=\delta_{ij}=\begin{cases}1,\qquad i=j\\ 0,\qquad\text{otherwise}.\end{cases} (2.21) Moreover, if we identify the subsets \mathcal{K}_{i} as in ( 2.3 ), we can rewrite \mathcal{M}_{\mu}\left[f\right](x) as \mathcal{M}_{\mu}\left[f\right](x)=\sum_{i=1}^{n}\sum_{j\in\mathcal{K}_{i}}B_{\mu,j}(x)l_{j,i}(x)f_{i}. (2.22) After these rearrangements, it is easy to compute the successive derivatives of \mathcal{M}_{\mu} by the formula \mathcal{M}_{\mu}^{(m)}\left[f\right](x)=\sum_{i=1}^{n}g_{i}^{(m)}(x)f_{i} (2.23) where g_{i}^{(m)}(x)=\sum_{j\in\mathcal{K}_{i}}\left(B_{\mu,j}(x)l_{j,i}(x)\right)^{(m)}. (2.24) 3 Approximation of the Caputo fractional derivative The goal of this Section is to approximate, for any x>0 and m-1<\alpha0. (6.49) We fix n_{e}=9 . The numerical results are reported in Figure 4 and Figure 5 . Example 2 Let f(x)=x^{\frac{9}{2}} . The analytical form of the Caputo fractional derivative for f(x) is given by (D^{\alpha}f)(x)=\frac{\Gamma(\frac{9}{2}+1)}{\Gamma(\frac{9}{2}+1-\alpha)}x^{\frac{9}{2}-\alpha}. We fix n_{e}=20 . The numerical results are reported in Figure 6 and Figure 7 . Example 3 Let f(x)=e^{2x} . The analytical form of the Caputo fractional derivative for f(x) is given by some_fractional_derivatives \left(D^{\alpha}f\right)(x)=\displaystyle\sum_{k=0}^{\infty}\frac{2^{k+m}}{{\Gamma(k+1+m-\alpha)}}x^{k+m-\alpha}=2^{m}x^{m-\alpha}E_{1,m-\alpha+1}\left(2x\right), where m-1<\alpha