Abstract

This paper combines the synergies of three mathematical and computational

generalizations. The concepts of fractional calculus, memristor and information
visualization extend the classical ideas of integro-differential calculus, electrical
elements and data representation, respectively. The study embeds these notions in a
common framework, with the objective of organizing and describing the
"continuum" of fractional order elements (FOE). Each FOE is characterized by its
behavior, either in the time or in the frequency domains, and the differences
between the FOE are captured by a variety of distinct indices, such as the
Arccosine, Canberra, Jaccard and Sørensen distances. The dissimilarity
information is processed by the multidimensional scaling (MDS) computational
algorithm to unravel possible clusters and to allow a direct pattern visualization.
The MDS yields 3-dimensional loci organized according to the FOE characteristics
both for linear and nonlinear elements. The new representation generalizes the
standard Cartesian 2-dimensional periodic table of elements.

Graphical abstract

1. Download : Download high-res image (89KB)

2. Download : Download full-size image
 Previous article in issue
 Next article in issue

Keywords
Fractional calculus
Memristor
Information visualization
Multidimensional scaling
Procrustes analysis

Introduction
Leibniz (1646–1716) extended the differential calculus to the paradigm known as
"Fractional Calculus" (FC) [1], [2]. However, the FC remained an abstract tool
restricted to the area of mathematics. The first application of FC is usually credited
to Abel (1802–1829) with the so-called tautochrone curve problem [3], [4]. Later
we find the work of Heaviside (1850–1925), who fist applied such ideas in the
scope of the operational calculus and electromagnetism [5], [6]. Nonetheless, it
was during the last two decades that FC was recognized as a good tool to
characterize complex phenomena, due to the ability of describing adequately non-
locality and long-range memory effects [7], [8], [9], [10], [11].
Paynter (1923–2002) formulated one systematic approach to modeling and
invented the so-called bond graphs [12]. He considered 4 generalized variables,
namely the effort, flow, momentum and displacement {e,f,p,q} so
that p=∫etdt and q=∫ftdt. In page 136 of his class notes [12] he designed a diagram
including the 4 state variables with vertex of a "tetrahedron of state". Paynter
characterized the functional relationship between the 4 variables that are associated
with the edges of the tetrahedron. The relations for e-f, e-q and p-f (for resistance,
capacitance and inductance, respectively) were marked by continuous lines.
Identically for p-e and q-f (for the integral/differential relationships). However, the
relation p-q was merely marked with a dashed line and no particular importance
was given to it.
In 1971 Chua [13] noticed again the symmetries in the electrical integer order
elements (IOE) and variables. Chua speculated that 4 elements were necessary to
preserve a Cartesian arrangement. By other words, in his opinion, besides the
standard 3 elements represented by the resistor, capacitor and inductor, a 4-th one,
the so-called "memristor" or resistor with memory, was also necessary. In 2008
these ideas were brought to light in the scope of a laboratory experiment [14] and
the topic became popular in a variety of applications.
The Chua [15] periodic table of elements (PTE) organizes two-terminal IOE in a 2-
dimensional Cartesian matrix. Besides including IOE the generalization to real-
and complex-order elements was also proposed [16], [17]. However, the necessity
of the 4-th element and the Cartesian layout of the PTE is still under
debate [18], [19]. In fact, this type of organization may not be the best one to
accommodate the elements. It is out of the scope of the present paper to address the
problem of writing systems, that is the method of visually representing
communication. We recall that the Greek alphabet and consequent systems, settled
on a left-to-right pattern, from the top to the bottom of the page. Nonetheless,
Arabic and Hebrew scripts are written right-to-left, while those including Chinese
characters were traditionally written vertically top-to-bottom and from the right to
the left of the page. Therefore, we can question up to what point are we "prisoners"
of our cultural heritage
(https://en.wikipedia.org/wiki/Writing_system#Directionality). Furthermore,
present day computational techniques for data processing and information
visualization can provide superior forms of representation.
Information visualization involves the computer construction of some type of
graphical representations, that otherwise would require more efforts to be
interpreted, and helps to unravel patterns embedded in the data [20], [21]. Due to
the multidimensional nature of most data, the information visualization can take
advantage of dimensionality reduction [22] and clustering [23] techniques.
This paper adopts information visualization to organize two-terminal fractional
order elements (FOE). The new representation generalizes the 2-dimensional PTE
by means of 3-dimensional loci of FOE. We verify that the FOE form a
"continuum" where the IOE are special cases, and not the opposite, as often
assumed. Therefore, without lack of generality, in the follow-up we shall mention
as FOE to all elements. The proposed numerical and computational approach
includes 2 phases. First, we characterize the FOE either in the time or in the
frequency domains. The comparison of the FOE characteristics is performed by
means of four metrics, namely the Arccosine, Canberra, Jaccard and Sørensen
distances. Second, we process the dissimilarities through the multidimensional
scaling (MDS) visualization computational method, that produces loci
representative of the input information. The computational portraits are not
restricted neither to 2-dimensional nor to Cartesian concepts based on human
notions. Indeed, the FOE loci reveal distinct patterns that are built upon the
distance metrics properties.

Following these thoughts, the paper has the following organization. Section 2
presents the concepts supporting the mathematical and computational methods.
Section 3 characterizes the FOE by distinct methods, namely in the time and
frequency domains. Additionally the FOE are compared with four distances and
the information is processed by means of the MDS technique. Section 4 compares
the effect of nonlinearities by means of Procrustes analysis. Finally, Section 5
draws the most important conclusions.

Mathematical and computational concepts

Fractional calculus
FC generalizes the concept of differentiation and integration to non integer and
complex orders [24], [25]. We find a variety of applications of FC, such as in
control, physics, anomalous diffusion, and many others [26], [27], [28], [29].
Fractional derivatives and integrals are non-local operators that capture the history
dynamics, contrary to what happens with integer derivatives. Fractional systems
have a memory of the dynamical evolution and many natural and artificial
phenomena revealed these characteristics [30], [31], [7], [8], [9], [10], [32].
The most used definitions of fractional derivative are the Riemann-Liouville,
Grünwald-Letnikov and Caputo formulations [33], [34], [35]. For certain
functions, the fractional derivative follows closely their integer order version. For
example, at steady state, a sinusoidal function with amplitude A and phase Φ has
the derivative of order α∈R given by [36]:
(1)dαdtαAcos(ωt+Φ)=Aωαcosωt+Φ+απ2,where dαdtα denotes the fractional derivative
or order α, t represents time, and f and ω=2πf are the frequency and angular
frequency, respectively.
In the frequency domain, for zero initial conditions and the function x(t), we can
write:(2)Ldαdtαx(t)=sαL{x(t)},(3)Fdαdtαx(t)=(jω)αF{x(t)},where L{·} and F{·} represe
nt the Fourier and Laplace operators, s stands for the Laplace variable and j=-1.
The frequency dependent negative conductance and negative resistance
The frequency dependent negative conductance and frequency dependent negative
resistance (FDNC and FDNR) were introduced in 1969 and 1971 by
Bruton [37] and Antoniou [38]. The electronic implementation of these elements
have been under progress during the last decades [39], [40], [41], [42]. The FDNC
and FDNR are denoted by D- and N-elements and are usually considered with
linear behavior, having admittance and impedance Ys=Z-1s=Ds2 and Zs=Ns2,
respectively. These devices are often adopted in ladder filters without
inductors [37] and chaotic oscillators [43].
The FDNC and FDNR require an implementation using active devices and,
although not passive, demonstrate that they are feasible and useful. Moreover, for
circuit branch impedances Zis=kisni,ki∈R, ni∈ℕ, i=1,2,⋯, of integer order we obtain
also an integer order input impedance Zs. On the other hand, if we use fractional
impedances Zis=kisαi,αi∈R, then we can obtain Zs both integer and fractional [44].
The memristor
The magnetic flux and the electrical charge, ϕ(t) and q(t), are related to the voltage
and current, vt and it, by:(4)ϕt=∫-∞tvτdτ,qt=∫-∞tiτdτ.
In linear circuits, the resistor, inductor and capacitor, R, L and C, follow the
relations:(5)vt=Rit,ϕt=Lit,qt=Cvt.
The "memristor" M is the element verifying the
relation [45], [46], [47], [48], [49], [50]:(6)ϕt=M(q)qt.
If we have a linear relationship between ϕ and q, then M(q)=M similarly to a
resistance, since dϕdt=Mdqdt⟺v=Mi.
The generalization of the memristor concept to a larger class, the so-called
"memristive systems", is also possible [51], [52], [53], [15]. The charge-controlled
memristor and flux-controlled memconductance are modeled by the expressions:
(7)ϕ=ϕ^q,q=q^ϕ,and their time derivatives yield:
(8)v=∂ϕ^q∂qi,i=∂q^ϕ∂ϕv,where Miq=∂ϕ^q∂q and Mvϕ=∂q^ϕ∂ϕ are the incremental
memristor and memconductance, respectively.
The expressions (8) establish that i=0⇒v=0 and v=0⇒i=0, independently of q and ϕ,
respectively. Again, if the models (7) are linear, then we obtain the
resistance R and conductance G, respectively.
If we consider the generalized relations:(9) σt=∫t-∞qτdτ,ρt=∫t-∞ϕτdτ,then we
have [54]:(10)σ=σ^ϕ,ρ=ρ^q,(11)q=CMϕv,ϕ=LMqi,where CMϕ=∂σ^ϕ∂ϕ and LMq=∂ρ^q∂
q stand for the incremental memcapacitor and meminductor, respectively. Similarly
to what occurs with Mi(q) and Mv(ϕ), the elements CM(ϕ) and LM(q) "remember"
the flux and charge previously applied. These ideas support the so-called one-port
higher order element, establishing a relation between vt and it, such that:
(12)dmdtmvt=ψdndtnit,m,n∈Z.
Linearizing expression (12) around some operating point Q=dmdtmvQ,dndtniQ on
the element characteristic (12), we obtain:(13)dndtnv-vQ=mQ·dmdtmi-
iQ,where mQ denotes the slope of the line tangent to the
characteristic dmdtmvt=ψdndtnit at point Q.
In the frequency domain, expression (13) yields:
(14)V(jω)=Z(jω)I(jω),where(15)Z(jω)=(jω)n-m·mQ,is the small-signal impedance of
the element at the operating point Q.
 Based on those concepts, the PTE was proposed [15] as represented in Fig. 1. Each
point (m,n) represents an IOE and we verify that: (i) there are four element
categories that repeat ad infinitum along the Γ-diagonal lines; (ii) if we take
any (m,n) IOE and add (subtract) a multiple of four to either m or n, or to
both m and n, then we obtain a higher (lower) order IOE of the same category; (iii)
if we add or subtract 1 to both m and n, then we move the (m,n) IOE along its Γ-
diagonal line to the new position (m+1,n+1), maintaining the IOE category; (iv)
both the local and the global properties of all IOE on any Γ-diagonal line are
preserved; (v) the cases m,n= {(0,0), (–1,0), (–2,0), (0,–1), (–1,–1), (–2,–1), (–1,–
2)} stand for {R,L,D,C,M,LM,CM}. The location of the elements in the PTE may
also be specified by other types of coordinates. For example, if we choose the
coordinates (σ,γ), where σ=n+m and γ=n-m, then all IOE with the same value σ/γ lie
on one of the Σ/Γ-diagonals and the element of coordinates (σ,γ) is at the
intersection of both lines. Moreover, the Σ-diagonals are occupied either by
resistive or reactive IOE, for even or odd values of σ, respectively [55].

1. Download : Download high-res image (230KB)

2. Download : Download full-size image
Fig. 1. Simplified Cartesian representation of the PTE of two-terminal IOE. The
acronyms stand for resistor, inductor, frequency dependent negative conductance,
capacitor, memristor, meminductor, memcapacitor.
 The classical PTE represents only IOE and, therefore, we are restricted to γ∈Z.
However, other non-planar and non-Cartesian arrangements are possible for
representing the IOE. If we apply the coordinate transformation m,n→φ,σ,z, such
that(16)φ=m-n·π2,σ=m+n,z=φ,then we obtain the 3-dimensional PTE illustrated
in Fig. 2. With this representation, R is located at the center of the spiral-like locus
and the elements in the diagonals are represented at horizontal lines.

1. Download : Download high-res image (188KB)

2. Download : Download full-size image
Fig. 2. The 3-dimensional representation of the PTE of two-terminal IOE using the
coordinate transformation (16). The acronyms {R,L,D,C,M,LM,CM} stand
for {resistor, inductor, frequency dependent negative conductance, capacitor,
memristor, meminductor, memcapacitor}.
In another point of view, a closer look to the standard PTE reveals that the space in
the middle of the grid lines is, in fact, the locus for the FOE. Indeed, it is known
the existence of fractional inductors and capacitors [56], [57], [58], [59], [60] and,
therefore, the generalization of the PTE to a "continuum" of FOE is the logical step
to follow [61].
Distance functions
We adopt a set of 4 distances, {δA,δC,δJ,δS}, to measure the dissimilarity between
pairs (Pn,Pp) of objects with real and imaginary components [62]. Therefore, the
items are characterized by K×2 dimensional matrices Pn=RePn1⋯,RePnKT,
[ImPn1,⋯,ImPnK]T and Pp=RePp1,⋯,RePpKT,[ImPp1,⋯,ImPpK]T,
where Re· and Im· stand for the real and imaginary parts. The distances are given
by the expressions:
(17)δAPn,Pp=arccos∑k=1KRePnkRePpk+∑k=1KImPnkImPpk∑k=1KRePnk2+ImPnk2∑k=
1KRePpk2+ImPpk2,(18)δCPn,Pp=∑k=1K|RePnk-RePpk||RePnk|+|RePpk|+∑k=1K|ImPnk-
ImPpk||ImPnk|+|ImPpk|,(19)δJPn,Pp=∑k=1KRePnk-
RePpk2∑k=1KRePnk2+∑k=1KRePpk2-∑k=1KRePnkRePpk+∑k=1KImPnk-
ImPpk2∑k=1KImPnk2+∑k=1KImPpk2-∑k=1KImPnkImPpk.(20)δSPn,Pp=∑k=1KRePnk-
RePpk+∑k=1KImPnk-ImPpk∑k=1KRePnk+RePpk+∑k=1KImPnk+ImPpk.
If the objects to be compared have no imaginary part, then the
vectors Pn and Pp are K×1 dimensional, and the set {δA,δC,δJ,δS} corresponds to the
standard {Arccosine, Canberra, Jaccard, Sørensen} distances [63].
We must note that other distances are possible [63] and that several of them were
also tested. By other words, we are not restricted to the standard Cartesian
concepts, neither for in the chart nor for the difference measurements. It is well
known that the Cartesian perspective is a particular case of the Minkowski
formulation and that this is just a family of distances within a plethora of
generalized expressions [63], [62]. However, further distances are not included
herein for sake of parsimony, since {δA,δC,δJ,δS} illustrate adequately the proposed
concepts.
Multidimensional scaling
The MDS is a computational recursive method that provides dimensionality
reduction and envisages to produce a locus with clusters and, possibly, some data
organization capable of being visualized and interpreted [64], [65], [66]. Given a
set of N objects K-dimensional and a dissimilarity index, we calculate
a N×N matrix, Δ=[δnp], n,p=1,⋯,N, of object-to-object dissimilarities, such
that δnp=δpn and δnn=0. This information represents the input of the visualization
algorithm. The MDS represents the N objects by points in a W-dimensional space,
with W<K, and tries to reproduce the measured dissimilarities. The MDS iterates
the estimate of point configuration for optimizing a given fitness, achieving a
matrix of distances Δ^=[δ^np], n,p=1,⋯,N, that approximates the original
one Δ=[δnp]. A common fitness is the raw stress:(21)S=∑n=2N∑p=1n-1δ̂np-
θ(δnp)2,where θ(·) denotes some kind of linear or nonlinear transformation.
The MDS interpretation is based on the clusters and patterns in the W-dimensional
locus and not in the individual coordinates of the points. Points that are close
(distant) in the W-dimensional locus represent similar (dissimilar) objects in the K-
 dimensional space. We can translate, rotate and magnify the locus to provide a
better visualization. The MDS axes have no units and no special physical meaning.
The MDS quality can be verified through the Shepard and stress plots. The first
compares the resulting and the original distances, δ^np and δnp, for a given value
of W. Therefore, a narrow (large) dispersion of the points represents a good (poor)
fit between δ^np and δnp. On the other hand, the stress plot
represents S versus W and is a monotonically decreasing function. Usually the
values W=2 or W=3 are adopted, because they allow a straightforward
computational visualization and establish a compromise between low values
of S or W.
Visualizing fractional order elements
In this Section we generate several MDS representations both of linear and
nonlinear FOE. Firstly, we describe the FOE by their behavior either in the time or
in the frequency domains. This information will represent the objects P, that is, the
FOE. Secondly, we use the resulting data for comparing the FOE and calculate the
dissimilarity matrix Δ measured by means of a given distance function δ. Finally,
we feed the data into the MDS for constructing the matrix Δ^ and the W-
dimensional loci of FOE.
Let us consider the set of N FOE of orders Γ={γn:γmin≤γn≤γmax,n=1,⋯,N}. To each
FOE we apply a collection of sinusoidal signals xd,e(t)=Aecos(ωdt) with
frequencies Ω={ωd:ωmin≤ωd≤ωmax,d=1,⋯,Nω}. For nonlinear systems the set of
testing amplitudes is given by A={Ae:Amin≤Ae≤Amax,e=1,⋯,NA}, but, obviously,
for the linear case we can use just one value ( NA=1). Then we compute
the Nω·NA system outputs zd,e(t)=ψAeωdγncosωdt+γnπ2, where ψ represents some
kind of linear/nonlinear function, t=l·td, l=0,1,⋯,Nt-1 and td=Np·2πωd(Nt-1),
with td denoting the sampling period, Nt standing for the number of time samples,
and Np representing the number of periods of the signals (Fig. 3).

1. Download : Download high-res image (31KB)

2. Download : Download full-size image
Fig. 3. Block diagram of a FOE where ψ(·) is some linear/nonlinear function:
input i(t) and output v(t) given by xd,et=Aecosωdt and zd,et=ψAeωdγncosωdt+γnπ2,
respectively.
During the experiments some effect of truncating the series of γn values, that is, of
limiting to γmin and γmax was observed on the produced loci. Therefore, to reduce
that effect, all experiments adopted some extra values at both extremes that are not
represented.
Time domain analysis and visualization of linear fractional order elements
In this case we compare linear FOE in the time domain, meaning that we
consider NA=1. Therefore, after collecting the Nω outputs of the n-th FOE, we
construct the K=Nt·Nω dimensional real-valued vectors Pn(t)=z1,1(t),⋯,zNω,1(t) and
calculate the distance matrices Δ=[δ(Pn(t),Pp(t))], n,p=1,⋯,N,
where δ(Pn(t),Pp(t)) denotes one distance of the set {δA,δC,δJ,δS} between the
vectors Pn(t) and Pp(t) given by expressions (17), (18), (19), (20). Finally, we
process each matrix Δ by means of the MDS for constructing the FOE loci.
Fig. 4 depicts the 3-dimensional MDS loci of N=721 linear FOE with order values
spaced linearly in the interval -10≤γn≤10, when adopting Nω=40 test frequencies
spaced logarithmically in the interval 10-1≤ωd≤101, Nt=1000 time samples
and Np=5 periods. Several distinct amplitudes were tested numerically, but, as
expected, the FOE loci do not depend on this parameter. All other parameters were
adjusted by successively increasing their values until the loci are insensitive to
changes. The markers represent the FOE and the colors vary with the FOE
order, γn, to enhance the visualization. In the loci (a) and (b) the IOE of the same
category are connected by dashed lines. Such lines are not included in (c) and (d),
since for their good visualization we need to rotate the charts. For all distances, we
verify that the FOE form smooth patterns, exhibiting regularities that depend on
the FOE categories. Moreover, the representations do not follow the standard
Cartesian arrangement and use efficiently the 3-dimensional visualization space.
1. Download : Download high-res image (737KB)
2. Download : Download full-size image
Fig. 4. The 3-dimensional loci of N=721 linear FOE, characterized in the time domain
by means of the distances: (a) δA; (b) δC; (c) δJ; (d) δS. The markers represent the
FOE and their color varies with the FOE order γn∈[-10,10]. The other parameters
are Nω=40, ωd∈[10-1,101], NA=1, Nt=1000 and Np=5. In the loci (a) and (b) the IOE of
the same category are connected with dashed lines.
Variations to the previous loci are possible to highlight specific aspects of the
organization of the FOE and to capture distinct information provided by the MDS
computational scheme. These possibilities are illustrated in Fig. 5, where we
consider two MDS dimensions for the (x,y) coordinates, while the z coordinate is
calculated by means of radial basis interpolation (RBI) [67] of the FOE order γn.
The thin-plate spline RBI function, ϕ(ε)=ε2logε, is considered, where the
variable ε denotes the Euclidean distance between the points generated by the 2-
dimensional MDS. Nonetheless, we believe that the 3-dimensional visualization of
 the locus is more advantageous than the 2+1-dimensional portrait. Therefore, for
reducing length, in the follow-up we restrict to the richer visualization method.

1. Download : Download high-res image (1MB)

2. Download : Download full-size image
Fig. 5. The 2 + 1-dimensional loci of N=721 linear FOE, characterized in the time
domain by means of the distances: (a) δA; (b) δC; (c) δJ; (d) δS. The z coordinate of
the loci is calculated by means of RBI based on the value of γn∈[-10,10] at each
MDS (x,y) coordinate. The other parameters are Nω=40, ωd∈[10-
1,101], NA=1, Nt=1000 and.Np=5
Frequency domain analysis and visualization of linear fractional order
elements
For the n-th FOE, n=1,⋯,N, we convert the sinusoidal outputs zd,1(t), d=1,⋯,Nω, to
the Fourier domain, yielding F{zd,1(t)}=Zd,1(jω), where ω=ωd. We generate
the Nω×2 dimensional complex-valued matrix Pn(ω)=Re{Zd,1(jω)},Im{Zd,1(jω)} and
 calculate the dissimilarity matrices Δ that feed the MDS algorithm and generate the
FOE loci.
Fig. 6 depicts the 3-dimensional loci of the N=721 linear FOE, characterized in the
frequency domain by means of the distances {δA,δC,δJ,δS}. The values of the
parameters are identical to those adopted in the Subsection 3.1. For all distances
we verify that the linear FOE loci do not depend on the amplitude of the sinusoidal
inputs and that the loci form smooth patterns with regularities that depend on the
FOE categories. The charts are different from those obtained in the time domain,
but follow an identical logic, namely using a non-Cartesian arrangement in a 3-
dimensional space.

1. Download : Download high-res image (852KB)

2. Download : Download full-size image
Fig. 6. The 3-dimensional loci of N=721 linear FOE, characterized in the frequency
domain by means of the distances: (a) δA; (b) δC; (c) δJ; (d) δS. The markers represent
the FOE and their color varies with the FOE order γn∈[-10,10]. The other parameters
 are Nω=40, ωd∈[10-1,101], NA=1, Nt=1000 and Np=5. In the loci (a) and (b) the IOE of
the same category are connected with dashed lines.
Time domain analysis and visualization of nonlinear fractional order elements
In this case we compare nonlinear FOE in the time domain adopting for ψ a cubic
nonlinearity. We must note that other nonlinearities [68] are possible and that
several were tested. However, they are not included herein for sake of parsimony,
since this one illustrates well the proposed ideas.
For the n-th nonlinear FOE, n=1,⋯,N, we collect Nω·NA outputs, zd,e(t),
where d=1,⋯,Nω and e=1,⋯,NA. Then, for comparing the FOE, we construct
the K=Nt·Nω·NA dimensional real-valued vectors Pn(t) = [z1,1(t),…,z1,NA(t),
…,zNω,1(t),…,zNω,NA(t)]. Finally, we calculate the distance
matrices Δ=[δ(Pn,Pp)], n,p=1,⋯,N, and apply the MDS numerical algorithm. Fig.
7 depicts the 3-dimensional MDS loci of the N=721 nonlinear FOE, characterized
in the time domain by means of {δA,δC,δJ,δS}. The values of the parameters are
identical to those adopted in the previous Subsections, but for the nonlinear case
we consider NA=10 amplitudes of the input signal spaced logarithmically in the
interval 10-2≤Ne≤102. Comparing Fig. 7, Fig. 4, we verify that the loci generated
with the distances {δA,δS} do not vary, while those generated with {δC,δJ} vary
considerably with the presence of the nonlinearity. We verify again that we can
adjust the characteristics of the loci, in this case the sensitivity to the
nonlinearity ψ, by a judicious choice of the proper distance.
1. Download : Download high-res image (695KB)
2. Download : Download full-size image
Fig. 7. The 3-dimensional loci of N=721 nonlinear (ψ of degree 3) FOE, characterized
in the time domain by means of the distances: (a) δA; (b) δC; (c) δJ; (d) δS. The
markers represent the FOE and their color varies with the FOE order γn∈[-10,10]. The
other parameters are Nω=40, ωd∈[10-1,101], NA=10, Ne∈[10-2,102], Nt=1000 and Np=5.
In the loci (a) and (c) the IOE of the same category are connected with dashed lines.
Frequency domain analysis and visualization of nonlinear fractional order
elements
We compare N=721 nonlinear FOE in the frequency domain adopting for ψ the
cubic nonlinearity.
In a first phase, for the n-th FOE, n=1,⋯,N, we convert the Nω·NA outputs zd,e(t) to
the Fourier domain, yielding Fzd,et=Zd,ejω, noting that for a cubic
nonlinearity zd,e(t) has the first and third harmonics. In a second phase, for
comparing the FOE, we generate the (2·Nω·NA)×2 dimensional complex-valued
 array Pn(ω)=Re{Zd,e(jω)},Im{Zd,e(jω)}. Finally, we calculate the dissimilarity
matrices Δ, and generate the MDS FOE loci.
Fig. 8 depicts the 3-dimensional MDS loci of the N=721 nonlinear FOE,
characterized in the frequency domain. All values of the parameters are kept
unchanged from the previous Subsections.

1. Download : Download high-res image (725KB)

2. Download : Download full-size image
Fig. 8. The 3-dimensional loci of N=721 nonlinear (ψ of degree 3) FOE, characterized
in the frequency domain by means of the distances: (a) δA; (b) δC; (c) δJ; (d) δS. The
markers represent the FOE and their color varies with the FOE order γ∈[-10,10]. The
other parameters are Nω=40, ωd∈[10-1,101], NA=10, Ne∈[10-2,102], Nt=1000 and Np=5.
In the loci (a) and (d) the IOE of the same category are connected with dashed lines.
Procrustes analysis and visualization of nonlinear
fractional order elements
In this Section, we compare the loci obtained with different nonlinearities by
means of Procrustes analysis [69], [70], [71], [72]. The Procrustes analysis takes a
collection of loci and transforms them for obtaining the "best" superposition. The
algorithm performs four iterative numerical steps: (i) the user chooses a reference
locus (by selecting one of the available instances); (ii) superimposes all other loci
into the current reference by means of linear transformations, namely translation,
reflection, orthogonal rotation and scaling; (iii) computes the mean form of the
current set of superimposed loci; (iv) compares the distance between the mean and
the reference instances to a given threshold value and, if above, sets the reference
to the mean form and continues to step (ii). The result is a global representation of
all loci that best conforms them.
Fig. 9, Fig. 10 depict three superimposed 3-dimensional MDS loci of N=721 FOE
(using Procrustes), characterized in the time and frequency domains, respectively.
Besides the linear case we adopt power law nonlinearities ψ of degree 3 and 5. The
values of all parameters are identical to those used in the previous Section. We
verify an evolution of the loci with n, demonstrating the sensitivity of the technique
to the nonlinearity. For the distances δA and δS this evolution is smooth, while
for δC and δJ we obtain a sharp transition between the linear and the nonlinear
cases, when the FOE are characterized in the time and the frequency domains,
respectively. Therefore, we verify that we can extend the construction of the MDS
loci and their comparison to other types of nonlinearites.
1. Download : Download high-res image (745KB)
2. Download : Download full-size image
Fig. 9. Three superimposed 3-dimensional loci of N=721 FOE (using Procrustes),
characterized in the time domain by means of the distances: (a) δA; (b) δC; (c) δJ;
(d) δS. The functions ψ of degree 1 (linear case), 3 and 5 are adopted.
1. Download : Download high-res image (730KB)
2. Download : Download full-size image
Fig. 10. Three superimposed 3-dimensional loci of N=721 FOE (using Procrustes),
characterized in the frequency domain by means of the distances: (a) δA; (b) δC;
(c) δJ; (d) δS. The functions ψ of degree 1 (linear case), 3 and 5 are adopted.
Conclusions
This paper used clustering and information visualization techniques to organize
and map FOE accordingly to their characteristics. The new representation
generalizes the concept of PTE, revealing that the integer order cases are just a
limited number of cases in the FOE "continuum". The use of the MDS allows
exploring the 3-dimensional space for the representation and the adoption of
distinct measures, so that users can choose the one fitting better their needs. The
technique is effective both in the time and frequency domains and can be extended
from linear to nonlinear elements. Moreover, the study provides a complementary
perspective in the on-going discussion about the properties of the memristor and
fractional-order elements. Indeed, a new form of representation, based in distinct
domains and distances, may shed further light into possible similarities or
dissimilarities between elements.

In summary, this paper did not intend to give responses to a variety of possible
questions such as if there are finite boundaries, or not, to the Chua's PTE, or what
is the physical meaning of fractional elements. The study shows that we are often
conditioned by representations methods that can be bettered by modern computer-
based information visualization algorithms. Furthermore, in the scope of the new
visualization methods, the use of Cartesian concepts, namely for graphical
representations and for distance (or difference) assessment, can be outperformed
by a careful selection of the formulation that fits better a specific application.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or
personal relationships that could have appeared to influence the work reported in
this paper.

Acknowledgement
Fundação para a Ciência e Tecnologia, Portugal, Reference: Projeto LAETA -
UID/EMS/50022/2013.

References
[1]
B. RossFractional calculus
Math Mag, 50 (3) (1977), pp. 115-122
View Record in ScopusGoogle Scholar
[2]
X.-J. Yang, D. Baleanu, H.M. SrivastavaLocal fractional integral transforms and
their applications
Academic Press, London (2015)
Google Scholar
[3]
D. Valério, J.T. Machado, V. KiryakovaSome pioneers of the applications of
fractional calculus
Fract Calculus Appl Anal, 17 (2) (2014), pp. 552-578
View Record in ScopusGoogle Scholar
[4]
J.A. Tenreiro Machado, Virginia Kiryakova, Anatoly Kochubei, Yuri LuchkoRecent
history of the fractional calculus: data and statistics
Anatoly Kochubei, Yuri Luchko (Eds.), Basic theory, De Gruyter, Berlin,
Boston (2019), pp. 1-22, 10.1515/9783110571622-001
CrossRefView Record in ScopusGoogle Scholar
[5]
H.J. JosephsOliver Heaviside papers found at Paignton in 1957
Inst Electric Eng, 319 (1959), pp. 70-76
CrossRefView Record in ScopusGoogle Scholar
[6]
B. MahonOliver Heaviside: maverick mastermind of electricity
Institution of Engineering and Technology, London (2009)
Google Scholar
[7]
J.A. Tenreiro Machado, António M. Lopes, Vasily E. TarasovFractional Van der
Pol oscillator
Vasily E. Tarasov (Ed.), Applications in physics, Part A, De Gruyter, Berlin,
Boston (2019), pp. 1-22, 10.1515/9783110571707-001
CrossRefView Record in ScopusGoogle Scholar
[8]
Valério D, Ortigueira M, Machado JT, Lopes AM. Continuous-time fractional linear
systems: Steady-state behaviour. In: Petráš, I, editor, Handbook of fractional calculus
with applications: applications in engineering, life and social sciences, Part A, vol. 6.
Berlin: De Gruyter. p. 149–74.
Google Scholar
[9]
António M. Lopes, J.A. Tenreiro Machado, Dumitru Bǎleanu, António Mendes
LopesFractional-order modeling of electro-impedance spectroscopy information
Dumitru Bǎleanu, António Mendes Lopes (Eds.), Applications in engineering, life and
social sciences, part A, De Gruyter, Berlin, Boston (2019), pp. 21-
42, 10.1515/9783110571905-002
CrossRefView Record in ScopusGoogle Scholar
[10]
Parsa B, Dabiri A, Machado JAT. Application of variable order fractional calculus in
solid mechanics. In: Baleanu D, Lopes AM, editors. Handbook of fractional calculus
with applications: applications in engineering, life and social sciences, Part A, Vol. 7.
Berlin: De Gruyter. p. 207–24.
Google Scholar
[11]
J.T. Machado, A.M. LopesRelative fractional dynamics of stock markets
Nonlinear Dyn, 86 (3) (2016), pp. 1613-1619
Google Scholar
[12]
H. PaynterAnalysis and design of engineering systems: Class notes for M.I.T.
course 2.751
M.I.T. Press, Boston (1961)
Google Scholar
[13
]
L.O. ChuaMemristor - the missing circuit element
IEEE Trans Circuit Theory, 18 (2) (1971), pp. 507-519
CrossRefView Record in ScopusGoogle Scholar
 [14
]
D.B. Dmitri, B. Strukov, G.S. Snider, D.R. Stewart, R.S. WilliamsThe missing
memristor found
Nature, 97 (2008), pp. 80-83
Google Scholar
[15]
M. Di Ventra, Y.V. Pershin, L.O. ChuaCircuits elements with memory:
memristors, memcapacitors and meminductors
Proc IEEE, 97 (10) (2009), pp. 1717-1724
CrossRefView Record in ScopusGoogle Scholar
[16]
M.-S. Abdelouahab, R. Lozi, L. ChuaMemfractance: A mathematical paradigm
for circuit elements with memory
Int J Bifurcation Chaos, 24 (09) (2014), p. 1430023
CrossRefGoogle Scholar
[17]
F.Z. Wang, L. Shi, H. Wu, N. Helian, L.O. ChuaFractional memristor
Appl Phys Lett, 111 (24) (2017), Article 243502
CrossRefGoogle Scholar
[18]
I. AbrahamThe case for rejecting the memristor as a fundamental circuit element
Sci Rep, 8 (1) (2018), p. 10972
Google Scholar
[19]
F.Z. WangA triangular periodic table of elementary circuit elements
IEEE Trans Circuits Syst I Regul Pap, 60 (3) (2013), pp. 616-623
View Record in ScopusGoogle Scholar
[20]
C. WareInformation visualization: perception for design
Elsevier, Waltham (2012)
Google Scholar
[21]
R. Spence

Information visualization: An introduction, Vol. 1, Springer, Cham (2001)

[22]
L. Van Der Maaten, E. Postma, J. Van den HerikDimensionality reduction: a
comparative review
J Mach Learn Res, 10 (2009), pp. 66-71
View Record in ScopusGoogle Scholar
[23]
A.K. Jain, M.N. Murty, P.J. FlynnData clustering: a review
ACM Comput Surv (CSUR), 31 (3) (1999), pp. 264-323
CrossRefView Record in ScopusGoogle Scholar
[24]
I. PodlubnyFractional differential equations, Volume 198: An Introduction to
Fractional derivatives, fractional differential equations, to methods of their
solution, mathematics in science and engineering
Academic Press, San Diego (1998)
Google Scholar
[25]
A. Kilbas, H. Srivastava, J. TrujilloTheory and applications of fractional
differential equations, Vol. 204, North-Holland mathematics studies
Elsevier, Amsterdam (2006)
Google Scholar
[26]
A. OustaloupSystèmes asservis linéaires d'ordre fractionnaire: Théorie et
pratique
Masson, Paris (1983)
Google Scholar