The role of the Havriliak-Negami relaxation in the description of local structure of Kohlrausch's function in the frequency domain. Part II
TThe role of the Havriliak-Negami relaxation inthe description of local structure of Kohlrausch’sfunction in the frequency domain. Part II
J.S. Medina, ∗ R. Prosmiti, and J.V. Alemán October 14, 2015 Instituto de Física Fundamental, IFF-CSIC, Serrano 123, Madrid ES-28006, Spain Departamento de Química, Facultad de Ciencias del Mar, ULPGC, Campus Universitariode Tafira, Las Palmas de G. Canaria ES-35017, Spain
Abstract
Two new sets of models for describing compactly the Fourier Transformof Kohlrausch-Williams-Watts, both based on the adiabatic variation ofparameters of a double Havriliak-Negami approximation along the wholeinterval of frequencies, are presented. One of them is relying, obviously,on the use of a well-behaved-pair of patches of the mentioned type ofapproximants, A p HN ( ω ) . The other is obtained by altering the simplefunctions HN ( ω ) and making dissimilar the couple. They are proposedthe guidelines of a new and systematic approach with extended Havriliak-Negami functions which is global, (non local), and of constant parameters.The latter at the cost of a more complicated dependency with the lowfrequencies than iωτ HN ) α . Introduction
To the extent that the object of study of soft matter and fluids has been passingfrom simple polar liquids to polymer, glasses and quasi-amorphous materials, thephenomenology of rheological, or dielectric, relaxations of physical systems hasbecome increasingly complicated. And it is not only that several types of thoserelaxations are superposed along frequency space making difficult to distinguishamong them, but that employed functional form evolves from an easy one asDebye [1], iω , to other more complex as Havriliak-Negami [2, 3], iω ) α ) γ ,after experimenting with intermediate stages as Cole-Cole [4], iω ) α , and Cole-Davidson [5], iω ) γ . ∗ [email protected] a r X i v : . [ c ond - m a t . s o f t ] O c t imultaneously something similar happens in time description while we con-sider the different temporal scales implied, so certain habit to model, –imposedby the mentioned physical phenomena and the discriminatory ability of ex-perimental equipment–, shows a methodological exhaustion. In this sense anytesting for the use of new relaxation functions, giving account of the novel ex-perimental records, is fully justified [6, 7].However, as has been quoted previously ([8, 9, 10, 11, 12, 13, 14, 15]), theswapping from one functional space to the other still continues to be hard,often drawing upon, the researcher, efficient numerical methods to perform suchdevious change [16, 17, 18, 19].Nevertheless it turns to be sometimes unsatisfactory this capability for blindcalculations as it does not provide many times of a general view allowing for theinterpretation and identification of these models whose available informationis fragmentary or incomplete. In this sense a catalogue for formulae linkingfrequency and time realms [14], is a precious help while are appraised significantsystem parameters. Besides it is also an essential partner of the numericalanalysis to bound errors and expose constructs [19, 20, 21], both characteristicsof computer techniques.We already have shown as a set of Weibull distributions [22], ( βt β − exp − t β , < β ≤ ), in Fourier space, ψ β ( ω ) , admit a good approximate description bysums of Havriliak-Negami functions [23, 24, 25]. Additionally in a precedentpaper to this work [25], it is established the local character of the approxima-tion and how, with slight variation of the parameters { α , , γ , , τ , , λ } withfrequency ω , Ap HN can describe a perfect fit with the objective function, ψ β .Such adiabatic behavior is commonly misunderstood as an argument againstthe approximation by means of basic relaxation functions as Havriliak-Negami.This fact it is best interpreted as the need for a wider family of relaxations witha known local portrayal.Thus in this job we will focus on taking advantage of such “local” informationto build “global” functions that ameliorate the preceding approximation in thewhole range of frequencies, [0 , ∞ ) . Also the relative error of all proposed modelsin the present and previous report [25] is depicted and tested against the realdata obtained from Fourier integrals. In short we have constructed two approximants of type [25], A p HN α,γ,τ,λ ( ω ) = (cid:88) s =1 λ s (1 + ( iτ s ω ) α s ) γ s , (1)( λ + λ = 1 ), for two different overlapping intervals ν ∈ [0 , . and ν ∈ [1 . , ] , (or ν ∈ [1 . , ] if β > ). Besides if we consider how the2elative error between moduli of approximant and function behaves as frequencyvaries, ( i.e. it stabilizes at an almost constant value never greater than 0.2%for high frequencies and < β ≤ ), the upper bound of second interval can beextended without a big amount of error to an unlimited frequency. It means thatthere are two charts ( A p ,l HN ( ω ) , Ω l ) and ( A p ,h HN ( ω ) , Ω ∗ h ) with Ω l ≡ π × (0 , . and Ω ∗ h ≡ π × (1 , ∞ ) that reconstruct in an acceptable way thefunction ψ β ( ω ) in the whole interval (0 , ∞ ) , plus the value at ψ β (0) = 1 as animposed condition.Now, all that is needed to obtain a global solution is a way to merge bothcharts without overlaps, following a standard procedure. We will resort to ansmooth, and monotonously increasing, function defined as: W i,s ( ω ) = ,
1) :1 : ω ≤ ω i ω i < ω < ω s ω s ≤ ω , with ω i and ω s chosen arbitrarily. So starting from locally adjusted functionswhich are properly selected it is possible to write a suitable approximation tofunction ψ β , in the whole interval [0 , ∞ ) , as: A p HN ( ω ) = A p ,l HN ( ω ) × (1 − W i,s ) + A p ,h HN ( ω ) × W i,s . (2)With our particular choice ω i = 2 π and ω s = 4 π , (or ω s = 2 . π if β > ),we have finally laid down a global surrogate of Havriliak-Negami type for theFourier Transform of Weibull function ψ β ( ω ) , whenever < β ≤ . β < While the chart for low frequencies is a quite good approximation to function ψ β ( ω ) it is not however a perfect matching in the mentioned range. Thus,there still is room for improving the fit as the difference reaches a peak of 1.3%for some of the lower frequencies, ω < δω , at medium values of beta, i.e. β ≈ . , and working with an r sampling, ( δω = 1 / . ). So the logicalnext step would be to add another Havriliak-Negami term to the approximantto fill the gap, although using some restrictions, ( v. gr. over α · γ products),to avoid proliferation of parameters and to hold the resemblance of a truncatedseries. The procedure works, however one should cope with other problemsof the multi-parameter optimization as the non isolated loci of minima, themultiplicity of them, and the competition among coefficients for some regions,(which prevents a broad and proper allocation of them in the search space). Itis then a good idea to analyze why a small gap befalls precisely at very lowfrequencies in order to correct it or design a strategy for the restrictions of new3 ormulae βA exp[ (cid:80) s =1 as (1 − β ) s ] exp[ − ( B − β + (cid:15) )2] + (cid:80) s =1 ds (1 − β ) s ParametersConstants Aa a a a a Corr. α ˆ α ParameterConstants − λB d d -4.51252 d d -0.916873Corr. Table 1:
Optimization parameters of the Eq. 3 approximant, ( case β ≤ ): Formulas, (with (cid:15) = 10 − ), and their constants for α , ˆ α and − λ . incoming terms of a series, or an approximant.The first derivative of ψ β ( ω ) = ´ ∞ βt β − e − t β − iωt d t with respect the angularfrequency is written as: lim ω → + d ψ β ( ω )d ω = lim ω → + − i ˆ ∞ βt β e − t β − iωt d t = − i ˆ ∞ βt β e − t β d t = − i Γ( 1 β +1) , while the first derivative for the Havriliak-Negami function is expressed as: dd ω iωτ ) α ) γ = − γα ( iτ ) α ω α − (1 + ( iωτ ) α ) γ +1 −→ ω → + − iγτ : − i α ∞ : α > α = 1 α < , which implies, as − i α = − cos πα − i sin πα takes values in the third quadrant,that no lineal combination of two vectors ( d HN ( ω )d ω /ω α − ) ω =0 and ( d HN ( ω )d ω /ω α − ) ω =0 can give a vector parallel to − i when both α , < . Nor it is possible withonly one α < . In such a situation, i.e. α < or α < , the magnitude ofat least one modulus will be infinity due to factor ω α , − what gives a reason,–the momentary diminishing of any | HN , ( ω ) | in the vicinity of ω ≈ + isquicker than that of | ψ ( ω ) | –, for the underestimation of ψ β by A p HN α,γ,τ,λ inthe range of low frequencies.All this confront us with the fact of finding alternatives to α , < , andunder the conditions of module and direction of the derivative of ψ β these aresimply α , = 1 , or α = 1 and α > , (or the converse pair of indices).That contradicts the empirical finding for parameters α , we made with the r sampling in the range of medium to low frequencies which supports heavily thecondition α , < when β < . and the conditions α < and α ∼ when A practical production of function W i,s ( ω ) can be made following instructions found inref. [26], lemma 1.10, p. 10. − β ) (cid:8) (cid:80) r =0 brβr + (cid:80) s =1 Cs cos( ζsβ + φs ) (cid:9) exp (cid:2) ( 1 β )2 (cid:80) s =0 cs (1 − β ) s (cid:3) exp[ − Mβd ] (cid:80) r =0 brβr ParameterConstants log τ b b b -0.228709 C -0.142877 ζ φ -7.14011 C ζ -8.86238 φ ParameterConstants τ c -0.796921 c c -2.98386 c -14.9068 c c -26.9545 c ParameterConstants log τ M d b -9.15927 × − b b b -0.641584Corr. Table 2:
Optimization parameters of the Eq. 3 approximant, ( case β ≤ ): Formulas andtheir constants for τ , τ and τ . β > . . Consequently the double approximant will always hold this gap inthe region of small frequencies unless of course we add some local modificationto parameters { α, γ, τ, λ } , near ω ≈ + .In summary there are three zones in the ω -space where the coefficients { α, γ, τ, λ } , are similar in magnitude, –comparing equal symbols–, althoughwith a different behaviour as functions of β . And they change from one toanother conduct when they are shifted across the intervals of frequency. Thismeans the set of parameters depends on ω , i.e. { α, γ, τ, λ } , = { α, γ, τ, λ } , ( ω ) ,but they are of a very slow variation through the whole interval of frequencies [0 , ∞ ) . They are adiabatic coefficients of the approximant A p HN α,γ,τ,λ ( ω ) .Therefore we are looking for a function reproducing the main traits of ψ β ( ω ) ,namely: ( d ψ β d ω ) ω =0 (cid:107) − i , ω β ∗ | ψ β ( ω ) | ω →∞ −→ O (1) , and ψ β ( ω ) ≈ O ( A p HN ( ω )) locally ∀ ω ∈ [0 , ∞ ) with adiabatic coefficients as parameters in the fit. So withthat in mind, and respecting conditions ( α · γ ) , = β for β < , our candidatewill take the form: ψ β ( ω ) ∼ = AM l, HN ( ω ) ≡ λ (1 + iωτ ) β + 1 − λ (1 + M l ( ω )( iωτ ) α ) βα , (3)with M l ( ω ) , the mollifier of the Havriliak-Negami function, verifying the fol-lowing boundary conditions lim ω →∞ M l ( ω ) = 1 and M l ( ω ) ≈ ( ωτ ) α , if ω ≈ , where α + α ≥ .Undoubtedly to determine the mollifier it is not an easy task and surely itsexpression as a series could be as difficult of obtaining as the one of ψ β , and yetthe conditions we have imposed on M l ( ω ) may allow an easy estimate of it. So5 β l og τ , τ , l og τ , −λ log τ τ τ β α , α α a) b) Figure 1:
Optimization parameters of modified Havriliak-Negami approximant that is shownin Eq. 3. They, α , τ , τ and − λ , follow strictly Eq. 3 while τ and α , ( ˆ α in text), are asin Eq. 4, the latter a simple estimation for the mollifier . β ≤ . Solid lines are mathematicaladjustments for parameters, which are given in tables 1 and 2. we will put forward as an estimator of mollifier M l ( ω ) the function ˆ M l ( ω ) ≡ [ 2 π arctan(( ωτ ) ˆ α )] N , (4)setting N = 3 as an appropriate average after a timely optimization for somevalues of β . β > As α > and α < for ≥ β > , when the adjustment to r sam-pling is done, it is arguable that a lineal combination of both direction vectors, ( d HN , ( ω )d ω /ω α , − ) , at ω = 0 can be parallel to − i , since they take values inopposite quadrants ( i.e. second and third ones). Nevertheless the magnitude ofthe derivatives, (zero and infinity respectively), avoids such result, and again theadmissible options for α , are those of the case β < which as we pointed outcontradict the numerical findings. We must proceed again with some modifica-tions of the approximant, however it is not possible now to extend the modelof equation 3 to this situation. Firstly the logarithmic second derivatives of log | ψ β | are qualitatively different for β < and β > cases. (See figures1 and 2 in [25]). Second only the α · γ = β condition holds while the otherbecomes α · γ (cid:39) β for the asymptotic behaviour of tails. Besides in the lowto medium frequencies range the products ( α · γ ) , present a nonlinear trendgreater in many cases than the corresponding linear condition of high frequency.(See figure 9 in [25]). In consequence we introduce a peculiar model with two6 β τ , τ , τ , τ , λ τ τ τ τ λ β α , , , γ α α α γ a) b) Figure 2:
For β ≥ an approximation to ψ β is done using the modified approximant AM g, HN set in Eq. 5. At left panel parameters: τ , τ , τ , τ and λ . At right panel: α , α , α and γ . Except for τ and α , ( ˆ α in text), all of them are defined in Eq. 5. Theformer are established in Eq. 6 which describes an estimator of the current mollifier . As inprevious graphics the parameters are adjusted with suitable mathematical expressions, (solidlines), given in tables 3 and 4. characteristic times in the first Havriliak-Negami relaxation which jointly withthe two terms structure of the whole approximant will retain the mentionedcharacteristics and restrictions for tails, with the sole exception of α · γ whichis to be determined by fitting. The latter in light of the numerical results ap-pears a spurious property or at least a virtual one conceived to explain thesudden change of curvature in a small interval of frequencies, ν (cid:46) .Thus when β > the global formula for the approximant it reads: ψ β ( ω ) ∼ = AM g, HN ( ω ) ≡ λ (1 + iωτ + ( iωτ ) α ) γ + 1 − λ (1 + M g ( ω )( iωτ ) α ) βα , (5)with the boundary conditions for the mollifier as lim ω →∞ M g ( ω ) = 1 and M g ( ω ) ∼ O (1) , if ω ≈ , where α > and α (cid:38) .Again an exact expression for the mollifier is out of scope of present studyand we settle for an estimator like ˆ M g ( ω ) ≡ [ 1 n √ { n √ −
1) 2 π arctan( ωτ ) ˆ α } ] n , (6)and with an ad hoc choice of n = 3 . With such estimator we will obtain a goodapproximation to | ψ β | which deviates slightly in a neighborhood of ω (cid:39) π , (the7one where the maximum of curvature happens in logarithmic scale), althoughit describes fairly the body of function and quite well the trend and values oftail. (See figure 2 in [25]). Finally in figures 1 and 2 we display the parameters of estimated functions ˆ AM l, HN ( ω ) and ˆ AM g, HN ( ω ) and their adjustments as functions of β (re-spectively β < and β > ). In this occasion the interval of frequencies is upto ν = 10 , and up to ν = 10 depending upon choice of β , and the samplingof frequencies is what we called logarithmically homogeneous, i.e. r sl . In bothcases all the curves have a break, or turnaround, more or less evident accordingto each one. This occurs for each curve, –within same β case–, in the same point( β ≈ . and β ≈ . ). Also during the course of several optimizations suchpoints have changed marginally and the breaks have increased or diminishedtheir sharpness according as we changed the average exponent ( N , n ) of equa-tions 4 and 6, data weights or samplings ( r , r sl ). So in conclusion we interpretthat such abnormalities are a result of the shape of estimators. Stretched instance, β < The best option in case β < is to weigh, –while using xmgrace to get a fit[27]–, the tails with option /Y to soften the jump and obtain an even ad-justment all the way in the interval of frequencies. The value of N also couldbe lowered but the price to pay is an increasing error for all the matching be-tween both functions, (approximant and ψ β ), around values of β ∈ (0 . , . and β ∈ (0 . , . . On the other hand the ability of the new function AM l, HN ( ω ) for describing the effect that slow variation parameters { α, γ, τ } , would havein the original Havriliak-Negami functions fully justifies the introduction of mol-lifier M l ( ω ) . Unfortunately the expression of its estimator does not seems goodenough in the vicinity of ω ≈ since the results do not fulfill the required condi-tion α + N ˆ α ≥ at all when β → + . This is a consequence of having frozenthe exponent at N = 3 , we should increase its value till infinity to compensatethe empirical trends of α and ˆ α to be zero when β → + . However the firstterm of ˆ AM l, HN ( ω ) , an almost residual one since λ ≈ for β < . , seems tobalance numerically this mathematical unsuitability of the second term of theapproximant in the description of ψ β ( ω ≈ . And that is possible since thereis no conflict in accounting for a slow diminishing | ψ β | in the neighborhood of ω ≈ , ( β < . ), using a fast decaying Havriliak-Negami type function, ( i.e. of large τ ), with the sampling step that we used. For such small values of β a neighborhood of zero where | ψ β | ∼ is so elusive that a frequency step of δν = 10 − is too large for considering a description of the modulus gradualdecay. (In tables 1 and 2 we wrote the mathematical expressions for the sixparameters of ˆ AM l, HN ( ω ) as curves depending of variable β ).8 ormulae exp (cid:2) { (cid:80) s =0 as ( β − s } exp( − Mβ ) (cid:3) B + exp[ − M ( β − (cid:80) s = bs ( β − s Cβ (cid:80) s =1 cs ( β − s ] ParametersConstants Ma a a a a Corr. α (cid:39) α (cid:39) ParameterConstants ˆ α M B b -9.16895 b b -41.314 b ParameterConstants γ C c c -34.1656 c c -373.845 c c -359.883 c Table 3:
Optimization parameters of the Eq. 5 approximant, ( case β > ): Formulas andtheir constants for α , α , ˆ α and γ . exp (cid:2) ( 1 β ) p (cid:80) s =0 ds ( β − s (cid:3) exp (cid:2) − M ( β − . (cid:3) (cid:80) s =1 qs ( b − s ParametersConstants τ τ τ τ p ≡ . ≡ . ≡ . ≡ . d d -38.3857 -27.9873 -4.7472 -30.3907 d d -3368.51 -2810.17 -31.8961 -1955.88 d d -9737.49 -7533.20 -28.0215 -2939.76Corr. ParameterConstants λM q q -407.192 q -3.32942 q Table 4:
Optimization parameters of the Eq. 5 approximant, ( case β > ): Formulas andtheir constants for τ , τ , τ , τ and λ . See figure 2. Squeezed instance, β > Case β > is instead more difficult to adjust in the whole interval of frequenciessince no additional weight is possible to use. The kink of log | ψ β | near ν ∼ claims for a body not overlooked which would be the case if tails were given moreimportance by weighing them as in previous procedure. Besides, the mollifierof Havriliak-Negami function is not enough elaborated and as a consequenceappears a bifurcation for each curve of parameters, corresponding the lowerbranch to the best adjustment to data. Nevertheless if the latter is employed fordescribing the curves, an abrupt change in trend for them is evident and makesmore difficult the handled mathematical expressions in parameter adjustment.We show here only the upper branch of all curves, this leads to a smooth andnice interpolation line for each parameter as seen in tables 3 and 4.Although we started with a nine parameters ansatz for the approximant ˆ AM g, HN ( ω ) is clear from the graphs, (right panel of figure 2), that α and α are almost constants. Now the written requirements over them are amply9ulfilled. Only there is a small disagree of order − from condition α ( β ) (cid:39) for some values of β . This is entirely due to competition among parametersand subsequent numerical errors. Meanwhile α ( β ) ≈ for all betas, and anyof them differ from this number less than 1.5% for β > and only with somesignificance for beta 1.00 and 1.02, and for β > . . Thus with a slight setting of ˆ M g ( ω ) has to be possible to write AM g, HN ( ω ) as a seven parameter functionwhich is more economic computationally. (All the adjustments to this set ofparameters are given in tables 3 and 4). Apart from already explained conditions in the onset of frequencies which makesa Cole-Davidson relaxation suitable to describe the boundary condition of dψ β dω ,it is obvious that in the interval β ∈ (0 . , . the first term of approximantdescribed in Eq. 3 plays an important role in the approximation since λ is notat all negligible. However for the interval β ∈ [0 . , . the story is quitedifferent, almost all its contribution is forced by theoretical considerations asnow the share coefficient is really small. To extend this situation and make anadjustment with a one-term approximant in the whole interval β ∈ (0 , , weshould prepare a more flexible second term of Havriliak-Negami type in Eq. 3.And with this goal in mind we establish the exponent N of ˆ M l ( ω ) as a newparameter of the optimization. Namely we shall do the following setting: ψ β ( ω ) ∼ = AM l, HN ( ω ) ≡ M l ( ω )( iωτ ) α ) βα , (7)with an estimator to M l ( ω ) similar to that of Eq. 4 though now N = N ( β ) isnon constant.The results, ( i.e. the parametric curves of β ), are shown in figure 3, therewe note two important features about the behaviour of parameters α and ˆ α and the shape peculiarities of N ( β ) . The first characteristic, in the interval β ∈ (0 . , . , is that we do not recover the functional form of a Debye relaxationas β → . For such a requirement it should happen at least N → and α → as an strong condition, or α + N ˆ α (cid:39) as a weaker one, in that limit. Neitherthe strong nor weak conditions are fulfilled by the parameters as can be seen inright and left panels of figure 3.In light of the share coefficient behaviour ( λ ) remains an important question:if the auxiliary term dominated by it in the modified approximant is reallynecessary. (See left panel in figure 1). Or if instead it is only needed to ’unfreeze’the exponent N in the estimator ˆ M l ( ω ) of the mollifier, (see Eq. 4), to adjust ψ β ( ω ) properly with only one term: the mollified Havriliak-Negami function.The second flaw is patent when we realize that it is not possible, in theinterval β ∈ (0 . , . , to hold the condition α + N ˆ α ≥ when α → + and ˆ α → + since N is finite and decreasing as β → . These trends ofalpha parameters are attested, jointly with the N one, and depicted in figure10 β τ , l og τ , N τ τ N/3 0 0.2 0.4 0.6 0.8 1 β α , α α a) b) Figure 3:
Fitting parameters of Eq. 7 as functions of variable β ≤ . Left panel: character-istic times τ , τ and exponent N ( β ) , (see Eq. 4). Right panel: frequency exponents α and α , ( ˆ α in Eq. 4). The dotted lines are just guides for the eye. β interval and it is not hard to imagine thedifficulties it has to describe an environment of | ψ β | ∼ , ( i.e. ω ≈ ), witha poor sampling of very low frequencies, (as is the case of ours for so smallvalues of beta). The tails obviously, in such a situation, lead the adjustmentand the mentioned requirements about the behaviour of dψ β dω near ω ≈ shouldbe imposed externally. At this event the best option to save both flaws is tomaintain the optimization with a two-terms approximant like that of Eq. 3.At last a further reason to keep the expression of Eq. 3 is to have a formalsimilarity with Eq. 5 for the case β > which makes more manageable thetreatment and analysis of the problem in the whole interval β ∈ (0 , . In light of these circumstances we will consider the frequency-averaged relativeerror, (among data ψ β ( ω ) and test functions), as an indicator of reconstructioncapability for any of the proposed Havriliak-Negami approximants. As it hasbeen patent till now most of the present discussion here refers to the suitabilityof pairs proposed to describe the modulus of data | ψ β | . We must be aware thataside from the results here discussed some additional tuning of phase shouldbe sought. Different one with each model for approximation we use. Even so,without all the benefits of the phase, an accurate adjustment between data and11 -8 -6 -4 -2 < ε r e l > AtAdjAtIntHTIt -10 -8 -6 -4 -2 GlbAdjGlbInt -10 -8 -6 -4 -2 AACWutt -9 -6 -3 β -9 -6 -3 < ε r e l > β -12 -9 -6 -3 β -9 -6 -3 β -8 -6 -4 -2 (a) (b) (c) (d)(e) (f) (g) (h) Figure 4:
Frequency-averaged relative error, (cid:104) ε rel ( ψ ( a ) β , ψ ( b ) β , ω m , ω x ) (cid:105) , for data ψ β ≡ ψ ( a ) β obtained with Mathematica TM and function values ψ ( b ) β of models AAC (circle), Wutt (dot-dashed), AtAdj (plusses), AtInt (squares), HTIt (light solid), GlbAdj (dark solid) and GlbInt(diamonds). Eight frequency windows, whose details are written in the text, and two typesof samplings –linear from (a) to (d) and logarithmic from (e) to (h)– are shown for interval < β ≤ . approximants makes this methodology of multiple Havriliak-Negami summandsuseful to determine form parameters in dielectric spectroscopy experiments, orin systematic search of them by means of genetics algorithms.Previously, a frequency-averaged relative error for the moduli of functionswas defined as: (cid:104) ε rel ( ψ ( a ) β , ψ ( b ) β , ω m , ω x ) (cid:105) = ´ ω x ω m | − | ψ ( b ) β || ψ ( a ) β | | d ωω x − ω m . Such error function is depicted in graphics 4 and 5. There it was calcu-lated for the different models described in a previous paper, [25], and in thepresent one, and also for other two models found in literature (see references[19, 20]). In particular, the errors for a Havriliak-Negami function, the threemodels described above (double H-N, atlas of aproximants and modified H-N),two variants of the last ones with parameters calculated via formulae given intables 1 to 4 in [25], and the numerical solution obtained from the C code inreference [19] are given, when β ≤ , in figure 4. Same models with β > ,except 1HN of AAC (Ref. [20]), are portrayed in figure 5.The frequency windows studied are: (a) [ ω m , ω x ) / π = [0 , , (b) [1 , ,(c) [10 , , (d) [100 , in the upper row of both figures. They followed a12inear sampling with δν = . , conversely the lower rows were logarithmic,homogeneous in each decade in the way we already explained. Their intervals offrequencies are: (e) [100 , ) , (f) [10 , ) , for both graphs, but (g) [10 , ) ,(h) [10 , ] with β ≤ , and (g’) [10 , ] when β > .For a quick sight inside the plots in 4 and 5 we have tagged the modelsalready explained. We remind that ψ ( a ) β ( ω ) ≡ ψ β ( ω ) was obtained from thedirect calculation of Fourier integral and is the same reference for all errors (cid:104) ε rel (cid:105) calculated with different test models ψ ( b ) β which are now listed as: AAC ,the Havriliak-Negami approximation cited in ref. [20].
Wutt, the C library ofreference [19] which employs the power series for low and high frequencies andan effective numerical method for the intermediary frequencies in the interval β ∈ [0 . , . . AtAdj is the label assigned to the model of equation 2, and thesame is true for the symbol
AtInt . The distinction is that while the parameters { α , , γ , , τ , , λ } in the first case are calculated following the formulas in tables1 to 8 of [25], in the last one are obtained from the points directly obtained oferror minimization and depicted in graphics 5, 6, 7 and 8 of [25]. For thesake of clarity we have repeated the latter results showing separately each partwhich follows the eq. 1 for low or high frequencies (head and tail functionsof equation 2). It allows to appreciate where in the frequency interval theindividual approximant diverges from the data, and which one is exactly itscontribution to the atlas of approximants. The transition from a plain downhillto a potential tail ( ν − β ) it is then quite clear. This is called HTIt . Besides allthe three previous models refer to the range [0 . , . of shape parameter β .The last assertion it is also true for labels GlbAdj and
GlbInt though twodifferent formulas and their respective implementations are employed, the equa-tions 3 and 4 for β ≤ and the equations 5 and 6 for β > . Again the first tagrefers to the adjusted parameters (see tables 1, 2, 3 and 4) and the second tothe original points as depicted in figures 1 and 2. In figure 4,
AAC , the approximation with only one HN function, shows thebiggest of all errors for the models here presented and the interval < β ≤ .The best result is of course that of Wutt which combine analytics and numerics.Meanwhile the atlas described by equation 2, ( i. e. AtInt ), works quite well evenin the interval of very low frequencies where we demonstrated that one of thefunctions of the approximant should be a Cole-Davidson relaxation or a modifiedversion of it, –this depends on if β is less than or greater than unit. (See Eqs. 3and 5). Besides, it holds good terms in the medium range thanks to the changeof describer function, ( i.e. from head, A p ,l HN , to tail, A p ,h HN ; see figure4 panel (b)). The performance of this swap is even better for high frequenciesbecause the matching with data exceeds expectations and the approximation hasnot required any restriction in the product ( α · γ ) , . This reinforces our previousconclusion [24], and points to the true nature of ψ β as a sum of a Havriliak-Negami pair with almost constant coefficients which only change significatively13s > ω → [28]. Such important transition is highlighted in figures 4 and 5 bythe discrepancy between model AtInt (squares) and
HTIt (light solid line) andreminds us the need for at least two charts (one for ν < the other for ν > )in the description of the whole function ψ β . Usually this approach is not takenin consideration in the literature since only one set of parameters { α, γ, τ, λ } isemployed to match the data, or if considered is misinterpreted due to the usageof a time scale factor in the stretched exponential ( i.e. exp − ( t/τ KW W ) β ) [29].See panels (a) to (d) in figure 4.It is worth to note how the Havriliak-Negami approach AAC starts to workbetter than the double approximant
HTIt in the range of medium-to-high fre-quencies, according as . < β → . It sounds logical since β → implies λ → ,and this final value is a pathology for the double sum of HN functions quite dif-ficult to treat numerically. Such problem is not evident using the atlas AtInt since the approximant of tail A p ,h HN takes the control over A p ,l HN in suchfrequency interval. The former function besides, at very high ν ’s and with β near . , shows similar errors to the results of numerical-analytic method Wutt .Now the major problem for both of them will be the numerical oscillations ofreference data. See panels (g) and (h) in figure 4.The differences of error between model
AtAdj and
AtInt present clearly tworegions. One in the low frequencies zone ( ν < ) the other in the high frequencyone ( ν > ), as it is usual coinciding with the onset of potential behaviour fortails. In the first case it is the lack of ability of the double approximant withconstant parameters to approach data, what gets closer both models. This isshown in figure 4a and is more conspicuous when β ∼ . . In the second casewhere both models split apart more than one order of magnitude the reason forthis is more subtle because the correspondence between ψ ( a ) β and ψ ( b ) β is tighteras both functions follow a quite similar potential decaying. (See graph 4, panels(b) to (h), and figure 9 in [25]). What makes such difference between AtInt and
AtAdj is an extra error provided for each of the parameters α , , γ , , τ , and λ . An additional contribution which is consequence of the fitting of parametriccurves to optimized points. Then it would be desirable to reduce those inputsbinding the parameters to relationships as the already mentioned ( α · γ ) , ∼ β .Nevertheless, we feel that some further work should be done to link the conductof τ ( β ) β , and λ to the coefficients of the analytical series for ψ β , and so in theabsence of them we have presented the models drawn from eq. 2 free of anyexternal constraints.As a novelty we introduced here two global models to simulate the data,namely GblInt (optimized points) and
GblAdj (adjusted parameters). Watch-ing them carefully in the various intervals of frequency one can realize how theperformances are similar to the models of Havriliak-Negami double approxi-mants,
AtInt and
AtAdj , respectively. Also is possible to observe how for thelow frequency range, ν < , the model GblInt outperforms to
AtInt , althoughthis one later improves and is usually better for the high frequencies in accuracy.(See in panels (a) to (h) of figure 4 the square and diamond curves). Also forthe global model, the aggregated error of several independent parameters spoils14 -10 -8 -6 -4 -2 < ε r e l > -12 -9 -6 -3 -12 -9 -6 -3 β -12 -9 -6 -3 AtAdjAtIntHTItGlbAdjGlbIntWutt1 1.2 1.4 1.6 1.8 2 β -12 -9 -6 -3 < ε r e l > β -9 -6 -3 β -9 -6 -3 (a)(e) (b) (c) (d)(f) (g) Figure 5:
Relative error, (cid:104) ε rel ( ψ ( a ) β , ψ ( b ) β , ω m , ω x ) (cid:105) , for data ψ β ≡ ψ ( a ) β obtained withMathematica TM and test values ψ ( b ) β of models: Wutt (dot-dashed), AtAdj (plusses), AtInt(squares), HTIt (light solid), GlbAdj (dark solid) and GlbInt (diamonds). Seven frequencywindows and two types of samplings –linear from (a) to (d) and logarithmic from (e) to (g)–are shown. Maximum frequency in panel (g) is reduced to ν x = 10 due to glitches andnumerical noise of data beyond this point. < β ≤ . AtAdj and
GblAdj , (plussesand dark solid lines in graphics). Therefore the drastic displacement of rela-tive error lines towards a lesser precision for models with adjusted parameters { α , , γ , , τ , , λ } ( β ) points to the need for strict relationships among them andgood descriptions of functional dependence with the shape parameter β .However it is important to emphasize how proposed estimators of mollifiers,( ˆ M l,g in Eqs. 4 and 6), are subject to many “ ad hoc ” restrictions, mainlydeduced of data traits and information obtained from the behavioral changesof α , , γ , , τ , , λ curves while changing the regime of frequencies from low tohigh. The most significative restraint here is the fact the ˆ M l,g functions areset as real ones when they should value in the complex field. As we have seenbefore, in Eqs. 3 and 5, near ω ≈ + the dominant demeanor of ˆ M l,g ( ω ) isdetermined without consider further modifications to the phase i α since thefrequency factor ω α + ∗ extinguishes such contribution quickly and only remainsthe one of extended Cole-Davidson term. Nevertheless the role the phase playsis over the entire interval of frequencies and, although to our present purposeof describing the modulus accurately it is not crucial this bias in the argumentof the approximation AM l,g, HN , it is important to point out the need of amollifier in complex series for proper description of ψ β .It seems promising, from an analytical point of view, that a strategy to sumup a quite difficult series in the neighborhood of ω ≈ + comes from the help ofan extended Havriliak-Negami pair. Perhaps could be interesting to decide themollifier’s functional form with the aid of series, integrals or equations whichdetermine ψ β . Mainly when ω → , for β < , or when ω → ∞ for β > , themost difficult cases for power series involved [8, 16, 29].Obviously, in the light of the problems we face when using unsuitable func-tions as estimators of M l,g ( ω ) , it is suggested that a similar ’loss’ of phase couldhappen in the atlas approximation of eq. 2 to data. And consequently foreseea mild shift in the argument of whole function A p HN ( ω ) at high frequencies.This displacement is due to the way the tail functions A p ,h HN ( ω ) are deter-mined. Data are pruned in a logarithmic pace and the important informationat low frequencies, –the plateau–, is removed when optimizing tail parameters,something is not made in the case of head functions, A p ,l HN ( ω ) . All this doesnot impact very much on the approach to the modulus, as we saw in graphics offigures 4 and 5, but suggests a share parameter fully complex, i.e. λ ∈ C − R .Unfortunately it would add a new degree of freedom and would overshadow thediscussion about modulus characteristics in absence of a thorough treatment ofthe data phase.Apart its mentioned inability to describe the very low frequency range, thedouble Havriliak-Negami approximant, A p HN α,γ,τ,λ ( ω ) , will not present suchdifficulties while describing the argument, –much less the modulus–, of data ψ β . The last stay in this discussion is the figure 5 that shows relative errors of thesix previous models for shape parameter interval < β ≤ . As predicted by16rst and second logarithmic derivative, (see frames a) and b) of fig. 2 in [25]),there is a sudden change of behaviour in the slope of | ψ β | from flatness to apotential decline ω − β in a relatively small interval of frequencies. (See frame c)of same graphics in [25]). This is a much more sharper and distinct transitionthan in case β ≤ , which forces the existence of a different set of constraints forproducts α · γ in the double Hav.-Neg. approximation, as clearly shows figure 9in [25]. All this oblige to abandon quickly the HTIt model , (light solid line in5, panels (a) to (d)), in favor of AtInt because the latter holds itself quite closeto the potential tail and shares same description of | ψ β | with the former at verylow frequencies. (See squares inside panels (a), and (b) to (g) in figure 5).Again, as with β ≤ , a big distance in terms of relative error separates AtInt and
AtAdj , and as before this gap is attributed to a collective error subscribed byeach individual parameter, whenever every uncertainty is caused by obtainingthe appropriate parameter from a pertinent fitting function along all valuesof β . However for the models GlbInt and
GlbAdj such distance doesn’t existat medium frequencies and thereafter, i.e. ν > . (See diamonds and darksolid line in graphics of figure 5). It seems that model GlbInt it is not able tokeep track of data tail so close as
AtInt does. Surely the ’bi-chronicity’ or themollifier in functional form of Eq. 5 should be revisited to give a proper accountof directional twist of data near ν ∼ , and thus to diminish the error below thecollective contribution of parameters. As in fact it happens at low frequencies,(see figure 5a). Nevertheless, as far we know, this is one of the few attempts todescribe globally the Fourier transform ψ β for β > with an analytical modelalbeit approximate, so each piece of formula AM g, HN ( ω ) has great value forfuture mathematical inquiries.Finally we should point out how, at very large frequencies, the potentialdescription overcomes some numerical difficulties experienced by Wutt modelfor large β values, i.e. β ≥ . . (See figures 5f and 5g). The present work is devoted to a compact description of the Fourier Transformof the Kohlrausch relaxation. As any reconstruction of this function as fromspectral data should heavily depend on the information of frequencies near zerosince tails will be surely corrupted by noise, an extra effort has to be made tomanufacture a global function mimicking all aspects of this transform from lowto high frequencies. Thus, two new sets of such models are proposed, and adetailed discussion on their errors compared to numerical control calculationsgenerated directly by evaluating the Fourier integrals is made.We found how the approximation with a double H-N function always under-estimates ψ β in the low frequency range, ( ν < . ). And although it is a smalldifference forces us to change those values of parameters { ( α, γ, τ ) , , λ } alreadyobtained in the range of medium frequencies.This repeats again in the transition from medium to high, or very high,frequencies. Nevertheless the double Havriliak-Negami approximation is close17nough to the original function as to describe it along a wide range of frequenciesbefore the variation can be noticed. Moreover the parameters should not beregarded as varying, if the interval where the approximation is performed onlycomprises very high frequencies.Thus, due to the slow variation of parameters of the approximant with thefrequency, ( adiabatic parameters ), instead of a global function with ω − dependent { ( α, γ, τ ) , , λ } , we employed different charts of double Havriliak-Negami sumsto describe locally ψ β ( ω ) . We found that using two charts is a good approxi-mation, enough to establish an atlas, however the differences at low frequenciesstill persist with such number of maps. The inclusion of a third one in theneighborhood of zero, i.e. ν ≈ , should be convenient, though the existenceof an exact analytical series of powers in terms of Cole-Cole relaxations [4, 28],( i.e. / (1 + ( iω ) β ) ), for ψ β suggests the radius of such chart will depend on β [29]. What it makes difficult and cumbersome working with an atlas of threemaps.The question whether it is possible to sum up the series of Cole-Cole termsat ν = 0 , or it is possible to write the atlas of Havriliak-Negami charts just ina global way, seems to have a positive answer. We presented two ansätze forextending the double Havriliak-Negami approximation, – that has proved to besuccessful locally –, which describe along several decades in frequency, and withenough functional proximity, the data of | ψ β | . By means of rough estimatesof the mollifiers of these new testing relaxations, we have found a good agree-ment with data moduli signaling a path for a future close approximation in thecomplex domain to the series, or numerical integrals, of the Fourier Transform ψ β . References [1] P. Debye, Ver. Deut. Phys. Gesell. , 777 (1913).[2] S. Havriliak and S. Negami, J. Polym. Sci. Part C , 99 (1966).[3] S. Havriliak and S. Negami, Polymer , 161 (1967).[4] K. S. Cole and R. H. Cole, J. Phys. Chem. , 341 (1941).[5] D. W. Davidson and R. H. Cole, J. Chem. Phys. , 1484 (1951).[6] R. Kahlau, D. Kruk, Th. Blochowicz, V. N. Novikov, and E. A. Rössler, J.Phys.: Codens. Matter , 365101 (2010).[7] A. Stanislavsky, K. Weron and J. Trzmiel, Eur. Phys. Lett. , 40003(2010).[8] A. Wintner, Duke Math. J. , 678 (1941).[9] P. Humbert, Bull. Sci. Math. , 121 (1945).1810] H. Pollard, Bull. Am. Math. Soc. , 908 (1946).[11] G. Williams and D. C. Watts, Trans. Faraday Soc. , 80 (1970).[12] G. Williams, D. C. Watts, S. B. Dev, and A. M. North, Trans. FaradaySoc. , 1323 (1971).[13] C. P. Lindsey and G. D. Patterson, J. Chem. Phys. , 3348 (1980).[14] R. Hilfer, Phys. Rev. E , 061510 (2002).[15] E. Capelas de Oliveira, F. Mainardi, and J. Vaz Jr., Eur. Phys. J. SpecialTopics , 161 (2011).[16] M. Dishon, G.H. Weiss, and J.T. Bendler, J. Res. Natl. Bur. Stand. , 27(1985).[17] H. Schaefer, E. Sternin, R. Stannarius, M. Arndt, and F. Kremer, Phys.Rev. L. , 2177 (1996).[18] Ch. R. Snyder and F. I. Mopsik, Phys. Rev. B , 984 (1999).[19] J. Wuttke, preprint: arXiv:0911.4796v1, (2009); Algorithms , 604 (2012).[20] F. Alvarez, A. Alegria, and J. Colmenero, Phys. Rev. B , 7306 (1991).[21] F. Alvarez, A. Alegria, and J. Colmenero, Phys. Rev. B , 125 (1993).[22] W. Weibull, J. Appl. Mech. Trans. ASME. , 293 (1951).[23] J. R. Macdonald and R. L. Hurt, J. Chem. Phys. , 496 (1986).[24] J.S. Medina, R. Prosmiti, P. Villareal, G. Delgado-Barrio, and J. V.Alemán, Phys. Rev. E , 066703 (2011).[25] J.S. Medina, R. Prosmiti, and J. V. Alemán, preprint: arXiv:1509.07893,(2015).[26] F. W. Warner, Foundations of Differentiable Manifolds and Lie Groups ,Scott, Foresman and Co. (1971); Springer, ISBN: 978-0-387-90894-6 (1983).[27] P. Turner, E. Stambulchik, and Grace Development Team, http://plasma-gate.weizmann.ac.il/Grace[28] G. H. Weiss, M. Dishon, A. M. Long, J. T. Bendler, A. A. Jones, P. T.Inglefield, and A. Bandis, Polymer , 1880 (1994).[29] Ch. Schroeder and O. Steinhauser, J. Chem. Phys.132