In this paper, the least squares differential quadrature method for computing approximate analytical solutions for the generalized Bagley–Torvik fractional differential equation is presented. This new method is introduced as a straightforward and accurate method, fact proved by the examples included, containing a comparison with previous results obtained by using other methods.
1. Introduction
Even though the fractional calculus is a relatively old concept, dating back from the time of the mathematicians L’Hospital and Leibniz in the 17th century, it has seen a significant development in the last decades only. The fractional differential equations have numerous applications in biology, physics, chemistry, and engineering, which is why many scientists are concerned about finding more effective ways to solve them. Unfortunately, finding analytical solutions for fractional differential equations is often difficult or even impossible, which is why in recent years various numerical and approximate methods have been developed. One of the fractional differential equations with wide applicability in engineering is the Bagley–Torvik equation. This equation was introduced by Bagley and Torvik in [1] in 1984 to “model a motion of a rigid plate immersed in a Newtonian fluid” and in [2] for modeling the damping properties of the polymers. In engineering (construction, biotechnology, and chemistry), condensing polymers are widely used, and in their study, the Bagley–Torvik equation is the one employed [3, 4].
Bagley–Torvik-type fractional differential equations have been studied both numerically and analytically in numerous articles. Among the methods used to solve this equation, we mention the following: numerical methods [5], Adomian decomposition method [6], discrete spline methods [7], Haar wavelet method [8, 9], homotopy perturbation method [10], sinc collocation method [11, 12], cubic spline method [13], quadratic spline solution [14], B-spline collocation method [15, 16], Chebyshev collocation method [17], hybrid functions approximation [18], harmonic wavelets [19], predictor-corrector method of Adams type [20], spectral methods [21], fractional natural decomposition method [22], and finite element method [23].
The motivation of this paper is to introduce a new method for obtaining analytical approximate solutions for fractional differential equations. This method is obtained as a combination of the differential quadrature method (DQM) ([24]) and the least squares method (LSM), and we named it the least squares differential quadrature method (LSDQM). We applied this new method to find solutions of the following generalized Bagley–Torvik fractional differential equation [25]:(1)A⋅y′′x+Bx⋅Dxαyx=fx,yx,Dxβyx,y′x,where yx:a,b⟶ℝ, α∈1,2,β∈0,1, together with the conditions:(2)ya=μ0,ν1⋅y′a+ν2⋅yb=μ1,where μ0, μ1, ν0, and ν1 are constants (for ν1=0, we obtain boundary conditions, while for ν2=0, we obtain initial conditions). A∈ℝ, Bx and fx are given such that problems (1) and (2) satisfy the existence and uniqueness conditions for a continuous solution [25].
In the next section, we will present the least squares differential quadrature method (LSDQM), and in the third section, some numerical examples are presented followed by the conclusions.
2. The Least Squares Differential Quadrature Method
First, we will consider a numerical mesh of the interval I=a,b by means of a partition ΔM consisting of M+1 points: a=x0<x1<x2<⋯<xM−1<xM=bp=0.007. In order to find an approximate solution of the generalized Bagley–Torvik problems (1) and (2), we will compute the values of a certain functional (the functional (12) introduced in the following) at the point xi.
To equation (1), we attach the following operator:(4)Dαyx=A⋅y″x+Bx⋅Dxαyx−fx,yx,Dxβyx,y′x.
We denote by y˜x an approximate solution of equation (1). By replacing in Dα the exact solution yx with this approximate solution, we obtain the following reminder:(5)ℛx,y˜x=Dαy˜x.
Definition 1.
We call an ϵ-approximate solution of problems (1) and (2) related to the partition ΔM an approximate polynomial solution, which satisfies the following relations:(6)ℛxi,y˜xi|<ε,i=0,M¯,(7)y˜a=μ0,ν1⋅y˜′a+ν2⋅y˜b=μ1.
Definition 2.
We consider the sequence of polynomials:(8)PNx=∑k=0Nckxk,ck∈ℝ,k=0,N¯.
We call the sequence of polynomials PNx convergent to the solution of problems (1) and (2) if(9)limN⟶∞DαPNx=0.
Taking into account the above definitions, we will compute ϵ-approximate polynomial solutions for a Bagley–Torvik problem of the types (1) and (2) by taking the steps described in the following algorithm:
Step 1. we will compute an approximate solution of the type
(10)TNx=∑k=0N˜ckxk,
where the sequence of polynomials must also satisfy the boundary conditions:
(11)TNa=μ0,ν1⋅TN′a+ν2⋅TNb=μ1,
and the constants c˜k are calculated in the following steps.
Step 2. From the boundary conditions, we obtain ˜c0 and ˜c1 as functions of ˜c2,˜c3,…,c˜N. Thus, the expression of TNx from now on will be a function of x and c˜2,c˜3,…,c˜N only.
Step 3. We attach to problems (1) and (2) the functional ℱc˜2,c˜3,…,c˜N:
Step 4. By minimizing the functional (12), we obtain the values of the coefficients c˜20,c˜30,…,c˜N0 which give the minimum of (12).
Step 5. After computing c˜20,c˜30,…,c˜N0, we return to the boundary conditions and obtain the final value of c0˜ and c1˜.
Step 6. Finally, we replace the values c˜00,c˜10,c˜20,c˜30,…,c˜N0, calculate in the expression of TNx, and denote by TN0x=∑k=0Nc˜k0xk the analytical approximate polynomial solution by LSDQM of problems (1) and (2).
The following convergence theorem is satisfied:
Theorem 1.
The sequence of polynomials TNx satisfies the following relation:(13)limN⟶∞ℛ2xi,TNxi=0,i=0,M¯.
Proof.
Let yx be an exact solution of problems (1) and (2), which means from hypothesis that there exists a sequence of polynomials PNx=∑k=0Nckxk,\\ck∈ℝ,k=0,N¯ converging to yx (according to the Weierstrass theorem on Polynomial Approximation, see [24]), i.e., limN⟶∞PNx=yx,∀x∈I. Taking into account Definition 2, this means that PNx satisfies(14)limN⟶∞DαPNx=0.
Denoting by c20,c30,…,cN0 the values of the coefficients ck which give the minimum of the functional (12), it follows that ℱc20,c30,…,cN0≤ℱc2,c3,…,cN. If TN0x=∑k=0Nc˜k0xk is the approximate solution by LSDQM (with c˜00 and c˜10 computed by using the initial conditions), we obtain(15)∑i=0Mℛ2xi,TN0xi≤∑i=0Mℛ2xi,PNxi.Hence, limN⟶∞∑i=0Mℛ2xi,TN0xi≤≤limN⟶∞∑i=0Mℛ2xi,PNxi.
We conclude that limN⟶∞ℛ2xi,TN0xi=0.
2.1. Error Estimation
We can rewrite equation (1) as(16)A⋅y″x+Bx⋅Dxαyx−f1yx,Dxβyx,yx=f2x.
We denote with ℒ the following operator:(17)ℒyx=A⋅y″x+Bx⋅Dxαyx−f1yx,Dxβyx,y′x.
Thus, equation (1) becomes(18)ℒyx=f2x.
Using (for simplicity) the notation: ℛx=ℛx,y˜x=Dαy˜x, in (4), we obtain(19)ℛx=ℒy˜x−f2x,with y˜ as an approximate solution for problems (1) and (2), which means as y˜ satisfies(20)ℒy˜x=A⋅y″x+Bx⋅Dxαyx−f1yx,Dxβyx,y′x=f2x+ℛx,together with the conditions:(21)y˜a=μ0,ν1⋅y˜′a+ν2⋅y˜b=μ1.
We define the error function in the following way: e˜x=yx−y˜x, where yx is the exact solution for problems (1) and (2) and we obtain (using (18) and (19)) the differential equation for the error function:(22)ℒe˜x=ℒyx−ℒy˜x=−ℛx,with the conditions:(23)e˜a=0,ν1⋅e˜′a+ν2⋅e˜b=0.
The problem for the error function becomes(24)ℒe˜x=−ℛx.
Solving (24) together with (23) in the same manner as described above for problems (1) and (2), we obtain the approximation e˜x. We will be able to determine the absolute maximum error:(25)E=maxe˜x,0≤x≤1.
In this manner, we can estimate the error without knowing the exact solution of the initial problems (1) and (2).
Remark 1.
We remark that the above theorem proves the convergence of LSDQM since as the degree of the polynomial increases the remainder corresponding to the approximation tends to zero. This fact is illustrated in the second example.
3. Numerical Examples
In this section, we will present some examples of problems including generalized Bagley–Torvik fractional differential equations together with boundary or initial conditions and problems solved using the least squares differential quadrature method (LSDQM).
3.1. Example 1
We consider the generalized Bagley–Torvik equation [26, 27]:(26)y″x+x1/2⋅Dαyx+Byx=−x1/3⋅yx−x1/4⋅Dβyx−x1/5yx+f2x,together with the initial conditions: y0=2,y′0=0, where f2x=1−x1/2/Γ3−α⋅x2−α−x1/3⋅x−x1/4/Γ3−β⋅x2−β+x1/5⋅2−1/2⋅x2, α=1.234, and β=0.333. The exact solution of this problem is yx=2−1/2⋅x2.
Approximate solutions for this problem with absolute errors larger than 10−4 were proposed by El-Mesiry et al. in [26] and with absolute errors larger than 10−5 by Li and Zhao in [27].
We applied LSDQM to find an approximate analytical solution on the 0,1 interval for (26).
In Step 1, we choose an approximate solution y˜x of the following type:(27)y˜x=c˜0+c˜1x+c˜2x2.
In Step 2, by using the initial conditions, we find c˜0=2 and c˜1=0, and the approximate solution becomes(28)y˜x=2+c˜2x2.
In Step 3, the corresponding reminder (5) is(29)ℛy˜=c˜22x4/3+x11/5+1000x633/500383Γ383/500+2000000x1917/10001111889Γ667/1000+2+x4/3+x11/52+x633/500Γ883/500+x1917/1000Γ2667/1000+1.
We consider the simplest possible partition Δ1: 0=x0<x1=1, and we attach to the equation the functional ℱc˜2=∑i=01ℛ2xi,T1xi (too long to be included here).
In Step 4, in order to find the minimum of this functional, we can compute the stationary points by equating to zero its first derivative (the computations are performed using the software Wolfram Mathematica). We obtain c˜2=1/2, and it is easy to show (by means of the second derivative) that this stationary point is indeed the minimum.
In the case of this problem, Step 5 is not needed since we already have the final values of c˜0 and c˜1.
Finally, in Step 6, we replace c˜0=2, c˜1=0, and c˜2=1/2 in the initial expression of y˜x, thus obtaining the exact solution of the equation: y˜x=2−1/2⋅x2.
We remark that, in the case of this example, we could choose any partition ΔM since the particular form of the remainder (5) assures the fact that the stationary point is always c˜2=1/2 no matter how large M is.
3.2. Example 2
We consider the fractional Bagley–Torvik boundary value problem [21]:(30)y″x+Dαyx+yx=f2x,(31)f2x=−4c3x13/2F21;5/4,7/4;−1/4c2x23π−c2sincx+sincx,(32)y0=0,y1=sinc,where a,b=0,1, α=1.5, and F is the generalized hypergeometric function.
The exact solution of this problem is yx=sinc⋅x.
We remark the fact that if the exact solutions of the problem is a polynomial function (as in the case of the first example), then, naturally, in applying the first step of our method, we have chosen an approximate polynomial solution (10) of the same degree as the exact solution. Since, in this example, the exact solution is not a polynomial, we computed approximate polynomial solutions of several degrees, and for each degree, we calculated the maximal value of the absolute error on the whole interval 0,1 for c=1. The results presented in Table 1 together with a comparison with previous results from [21] clearly illustrate the convergence of the method.
Maximal absolute errors for approximate polynomial solutions of various degrees obtained using LSDQM for example 2.
Degree
Maximal error [21]
Maximal error LSDQM
4
3.4⋅10−4
6.3979⋅10−5
5
—
1.3366⋅10−6
6
—
8.2879⋅10−8
7
—
1.3333⋅10−9
8
4.3⋅10−7
7.3808⋅10−11
9
—
1.0573⋅10−12
10
—
6.5171⋅10−14
The approximate polynomial solutions computed by LSDQM using an equidistant partition Δ100 are(33)y˜4x=0.0196467x4−0.181975x3+0.00382023x2+0.999979x,(34)y˜5x=0.00720937x5+0.00159936x4−0.167467x3+0.000129252x2+1.x,(35)y˜6x=−0.000658967x6+0.00918884x5−0.000526231x4−0.166518x3−0.0000146541x2+1.x,(36)y˜7x=−0.000172399x7−0.0000548963x6+0.00838255x5−0.0000217806x4−0.166662x3−3.209289⋅10−7x2+1.x,(37)y˜8x=0.0000117989x8−0.000219638x7+0.0000202955x6+0.00832241x5+3.233537⋅10−6x4−0.166667x3+2.3498896⋅10−8x2+1.x,(38)y˜9x=2.40043968⋅10−6x9+9.888419⋅10−7x8−0.000199677x7+9.028126⋅10−7x6+0.00833296x5+8.6069828⋅10−8x4−0.166667x3(39)+3.73299399⋅10−10x2+1.x,(40)y˜10x=−1.3736474⋅10−7x10+3.0878011⋅10−6x9−4.5494814⋅10−7x8−0.000198028x7−2.0494998⋅10−7x6+0.0083334x5−1.31793399⋅10−8x4−0.166667x3−4.5065671⋅10−11x2+1.x.
3.3. Example 3
Choosing in (1) A=0, Bx=0, fx,yx,Dxβyx,y′x=Dxβyx+y2x−1, a,b=0,1, μ0=μ1=ν0=ν1=0, and α∈0,1, we obtain the problem [5, 28]:(41)Dxβyx+y2x−1=0,(42)y0=0.
This equation is actually a Riccati-type equation, and for α=1, the exact solution of the problem is yx=e2x−1/e2x+1.
Following the LSDQM steps presented in the previous section, we computed approximate polynomial solutions y˜x of the 9th degree for problems (41) and (42) choosing in turn β=1, β=0.9, β=0.8, and β=0.7. The solutions are computed on the 0,1 interval using again an equidistant partition Δ100. The results are presented in Figure 1.
Approximate solutions for problems (41) and (42) corresponding to several values of β.
For example, the approximate polynomial solutions corresponding to β=1 is(43)y˜x=−0.0106498⋅x9+0.0567197⋅x8−0.103159⋅x7+0.0253074⋅x6++0.125356⋅x5+0.00150329⋅x4−0.333491⋅x3+7.88553637⋅10−6⋅x2+1⋅x.
The maximal absolute error corresponding to this approximation is 4.11⋅10−9.
We remark that the results presented in Figure 1 are in good agreement with previously obtained results [5].
3.4. Example 4
While LSDQM, as most methods of this type, works best on small intervals, it can also be successfully employed in the case of large intervals. In order to illustrate this feature, we consider the Bagley–Torvik problem [6]:(44)y″x+Dxαyx+yx=0,(45)y0=1,y′0=0,where α∈1,2.
The exact solution of this problem is not known.
An approximate solution obtained by using the Adomian decomposition method was proposed in [6] by V. Daftardar-Gejji and H. Jafari.
For α=1.5, using an equidistant partition Δ10 on the interval I=0,20, we obtain the approximate analytical solution, presented in Figure 2:(46)y˜=−4.24381⋅10−9⋅x9+3.757992⋅10−7⋅x8−0.0000130166⋅x7+0.00021244,(47)x6−0.00133775⋅x5−0.00503943⋅x4+0.108157⋅x3−0.404114⋅x2+1.
Approximate solution for the problems (44) and (45) corresponding to α=1.5.
We remark that our result is in very good agreement with the one in [6].
4. Conclusions
In this paper, the application of the least squares differential quadrature method (LSDQM) to the Bagley–Torvik fractional differential equation is presented.
Due to the fact that the method is relative straightforward, the approximations may be obtained in a quick and simple manner. The numerical examples included clearly illustrate the accuracy of the method by means of comparisons with solutions previously computed by other methods. If the exact solution of the problem is itself a polynomial, then usually LSDQM is able to find the exact solution, as in the case of the first example. If not, in most of the cases, the solutions provided by LSDQM are not only more precise, but being of the simplest form (i.e., polynomial solutions), they are much easier to use in subsequent computation than many of the solutions provided by other methods.
Data Availability
All the necessary data are included in the paper.
Conflicts of Interest
The authors declare that there are no conflicts of interest.
TorvikP. J.BagleyR. L.On the appearance of the fractional derivative in the behavior of real materials198451229429810.1115/1.31676152-s2.0-0021439912BagleyR. L.TorvikP. J.Fractional calculus-a different approach to the analysis of viscoelastically damped structures198321574174810.2514/3.81422-s2.0-0020751558AleroevT.ErokhinS.KekharsaevaE.Modeling of deformation-strength characteristics of polymer concrete using fractional calculus201836503200410.1088/1757-899X/365/3/0320042-s2.0-85049477822IbrahimB. H. E.DongQ.FanZ.Existence for boundary value problems of two-term Caputo fractional differential equations201710251152010.22436/jnsa.010.02.16JafariH.TajadodiH.NazariH.KhaliqueC. M.Numerical solution of non-linear Riccati differential equations with fractional order20101117918210.1515/ijnsns.2010.11.s1.179Daftardar-GejjiV.JafariH.Adomian decomposition: a tool for solving a system of fractional differential equations2005301250851810.1016/j.jmaa.2004.07.0392-s2.0-10344238128ZahraW. K.Van DaeleM.Discrete spline methods for solving two point fractional bagley-torvik equation2017296425610.1016/j.amc.2016.09.0162-s2.0-84993995820RehmanM. U.KhanR. A.A numerical method for solving boundary value problems for fractional differential equations2012363894907MiculaS.CattaniC.On a numerical method based on wavelets for fredholm-hammerstein integral equations of the second kind201841189103911510.1002/mma.49522-s2.0-85046039566WangY.-G.SongH.-F.LiD.Solving two-point boundary value problems using combined homotopy perturbation method and green’s function method2009212236637610.1016/j.amc.2009.02.0362-s2.0-67349176680AlkanS.YildirimK.SecerA.An efficient algorithm for solving fractional differential equations with boundary conditions201614161410.1515/phys-2015-00482-s2.0-84959313987YüzbasiS.Numerical solution of the Bagley-Torvik equation by the bessel collocation method201336300312ZahraW. K.ElkholyS. M.Cubic spline solution of fractional Bagley Torvik equation201312230241ZahraW. K.ElkholyS. M.Quadratic spline solution for boundary value problem of fractional order201259337339110.1007/s11075-011-9495-02-s2.0-84856757008ThulaK.RoulP.A high-order B-spline collocation method for solving nonlinear singular boundary value problems arising in engineering and applied science201815AlloucheH.TazdayteA.TigmaK.Highly accurate method for solving singular boundary-value problems via padé approximation and two-step quartic B-spline collocation20191910.1007/s00009-019-1342-x1660-5446/19/030001-27JafariH.BabaeiA.BanihashemiS.A novel approach for solving an inverse reaction-diffusion-convection problem2019183268870410.1007/s10957-019-01576-x2-s2.0-85070909720MashayekhiS.RazzaghiM.Numerical solution of the fractional bagley-torvik equation by using hybrid functions approximation201639335336510.1002/mma.34862-s2.0-84956876682CattaniC.A review on harmonic wavelets and their fractional extension20182422423810.25073/jaec.201824.225DiethelmK.FordJ.Numerical solution of the bagley-torvik equation2002423490507DohaE. H.BhrawyA. H.Ezz-EldienS. S.Efficient chebyshev spectral methods for solving multi-term fractional orders differential equations201135125662567210.1016/j.apm.2011.05.0112-s2.0-79960555480RawashdehM. S.Al-JammalH.Numerical solutions for systems of nonlinear fractional ordinary differential equations using the FNDM20161364661467710.1007/s00009-016-0768-72-s2.0-84977151023BadrM.YazdaniA.JafariH.Stability of a finite volume element method for the time-fractional advection-diffusion equation20183451459147110.1002/num.222432-s2.0-85042213202ShuC.2000New York, NY, USA,SpringerStanekS.Two-point boundary value problems for the generalized bagley-torvik fractional differential equation2013113574593El-MesiryA. E. M.El-SayedA. M. A.El-SakaH. A. A.Numerical methods for multi-term fractional (arbitrary) orders differential equations2005160368369910.1016/j.amc.2003.11.0262-s2.0-9744276851LiY.ZhaoW.Haar wavelet operational matrix of fractional order integration and its applications in solving the fractional order differential equations201021682276228510.1016/j.amc.2010.03.0632-s2.0-77953135236CăruntuB.BotaC.PaşcaM. S.LăpădatM.Approximate solutions for Riccati differential equation of fractional order using the Least Squares differential quadrature method201964781828