Introduction
The description of the flow of non-Newtonian fluids rests on constitutive laws . They link the stress tensor that enters the coarse-grained macroscopic description based on the Navier-Stokes equations, with the microscopic material models. Recall the simplest example of a constitutive law of a fluid, the Newtonian fluid model: It links the velocity-gradient tensor γ ˙ = ∇ u ⃗ \dot{\boldsymbol{\gamma}} = \nabla\vec{u} γ ˙ = ∇ u of the fluid velocity field u ⃗ \vec{u} u , linearly to the Euler stress tensor σ \boldsymbol{\sigma} σ . Respecting the expected symmetries of the stress tensor, one obtains for incompressible flow, ∇ ⋅ u ⃗ = 0 \nabla\cdot\vec{u} = 0 ∇ ⋅ u = 0 , the relation σ = η ( γ ˙ + γ ˙ T ) \boldsymbol{\sigma} = \eta(\dot{\boldsymbol{\gamma}} + \dot{\boldsymbol{\gamma}}^{T}) σ = η ( γ ˙ + γ ˙ T ) , where η \eta η is the Newtonian viscosity.1
Many materials of technological interest go far beyond the simple Newtonian fluid. The stress-response to a flow can be significantly delayed in time due to slow internal relaxation of the fluid’s microstructure, and it can be nonlinear especially in soft materials where the intrinsic forces are comparable to deformation forces. In principle, also spatial non-locality could be considered, but for simplicity we do not deal with this in the present paper.
A natural form to write a constitutive equation for such a material is thus an integral constitutive equation (ICE):
σ ( t ) = η ( γ ˙ ( t ) + γ ˙ T ( t ) ) + ∫ t 0 t G ( t , t ′ , [ γ ˙ ] ) : γ ˙ ( t ′ ) d t ′ (1) \sigma(t) = \eta\left(\dot{\gamma}(t) + \dot{\gamma}^{T}(t)\right) + \int_{t_0}^{t} G\left(t,t',[\dot{\gamma}]\right) : \dot{\gamma}(t')\,\mathrm{d}t' \tag*{(1)} σ ( t ) = η ( γ ˙ ( t ) + γ ˙ T ( t ) ) + ∫ t 0 t G ( t , t ′ , [ γ ˙ ] ) : γ ˙ ( t ′ ) d t ′ (1)
if the flow is started at some time t 0 < t t_0 < t t 0 < t in the past. The delayed and potentially non-linear response is encoded in a history integral whose kernel is a response function G G G . It is a fourth-rank tensor field, obeying certain symmetries that we will not elaborate in detail here. In the presence of nonlinear response, it has a functional dependence on the velocity-gradient tensor γ ˙ \dot{\gamma} γ ˙ , denoted by square brackets. The first term on the r.h.s. of Eq. (1) is an instantaneous Newtonian contribution, representing degrees of freedom that can be assumed to relax instantaneously on the time scale of change of the considered flow. In principle it should be encoded in the integral kernel as a fast-decaying contribution followed by this time-scale separation argument; it is common practice to represent it as a singular t = t ′ t=t' t = t ′ contribution leading to the form of Eq. (1) .
A common setup is shear flow, where the flow gradient can be considered to be only along one axis, so that γ ˙ \dot{\gamma} γ ˙ has only one non-trivial element, [ γ ˙ ( t ) ] i j = δ i , x δ j , y γ ˙ ( t ) [\dot{\gamma}(t)]_{ij} = \delta_{i,x}\delta_{j,y}\dot{\gamma}(t) [ γ ˙ ( t ) ] ij = δ i , x δ j , y γ ˙ ( t ) , where the scalar function γ ˙ ( t ) \dot{\gamma}(t) γ ˙ ( t ) is called the shear-rate of the flow. In this case, if one restrcits to calculating just the corresponding shear-stress entry of the stress tensor, one can reduce Eq. (1) to a scalar equation,
σ ( t ) = η γ ˙ ( t ) + ∫ t 0 t G ( t , t ′ , [ γ ˙ ] ) γ ˙ ( t ′ ) d t ′ (2) \sigma(t) = \eta\dot{\gamma}(t) + \int_{t_0}^{t} G\left(t,t',[\dot{\gamma}]\right)\dot{\gamma}(t')\,\mathrm{d}t' \tag*{(2)} σ ( t ) = η γ ˙ ( t ) + ∫ t 0 t G ( t , t ′ , [ γ ˙ ] ) γ ˙ ( t ′ ) d t ′ (2)
The scalar function G G G is then called the generalized non-linear shear modulus. We work with this scalar equation in the following for simplicity; we expect the generalization to tensorial quantities to be straight-forward.
Equations of this form in particular arise in a microscopic approach from statistical physics, the integration through transients (ITT) formalism [1 , 2 ] . In general, the principles of perturbation theory in physics suggest that the stress-response σ ( t ) \sigma(t) σ ( t ) to a shear perturbation γ ˙ ( t ) \dot{\gamma}(t) γ ˙ ( t ) is given by a generalized convolution integral such as the one appearing in Eq. (2) . The generalized shear modulus G G G models the material’s response.
Most common rheological models for viscoelasticity can be seen as specializations of Eq. (2) with specific kernels; often such that the resulting integral equation can be converted into a differential equation for the stress or the strain. For example, the Maxwell model [3 ] is obtained by letting G ( t , t ′ , [ γ ˙ ] ) = G ∞ exp ( − ( t − t ′ ) / τ ) G(t,t',[\dot{\gamma}]) = G_{\infty}\exp(-(t-t')/\tau) G ( t , t ′ , [ γ ˙ ]) = G ∞ exp ( − ( t − t ′ ) / τ ) (and η = 0 \eta= 0 η = 0 ). We are however interested in models where the response is more complex, and the larger class of integral equations characterizes the response. This includes a number of “generalized” Maxwell-type models, but also models derived from the mode-coupling theory of the glass transition (MCT), where Eq. (2) is supplemented by another integral equation for the evolution of G G G [4 , 5 ] .
The form of Eq. (2) as the result of microscopic modeling is dictated by the fact that a shear-flow perturbation is easily modeled as a driving term in the microscopic equations of motion – this is particularly true for colloidal suspensions [4 , 5 ] , but also for molecular fluids (save some thermostatting issues) [6 , 7 ] . One thus imagines external forces to act such that a given shear-rate γ ˙ ( t ) \dot{\gamma}(t) γ ˙ ( t ) is prescribed, and the material is characterized by the resulting shear stress σ ( t ) \sigma(t) σ ( t ) . This decouples the Navier-Stokes equation from Eq. (2) , and the latter is solved by numerical quadrature. For example, linear viscoelasticity [8 ] , constant shear flows [9 ] , or startup-flow experiments involving so-called stress overshoots [10 , 11 ] have been studied using various MCT-derived models, or simple generalizations of the Maxwell model [12 ] .
We are concerned with numerical solutions of the inverse problem defined by Eq. (2) : given a certain material stress σ ( t ) \sigma(t) σ ( t ) , what is the strain rate γ ˙ ( t ) \dot{\gamma}(t) γ ˙ ( t ) that corresponds to that imposed stress? An especially important example of such stress-controlled rheology is creep, where a constant stress σ ( t ) = σ 0 Θ ( t ) \sigma(t)=\sigma_0\Theta(t) σ ( t ) = σ 0 Θ ( t ) is fixed for all t > 0 t>0 t > 0 . The asymptotic behavior for γ ˙ ( t → ∞ ) \dot{\gamma}(t\to\infty) γ ˙ ( t → ∞ ) is then of interest, as it allows to distinguish flowing states (γ ˙ → γ ˙ 0 > 0 \dot{\gamma}\to\dot{\gamma}_0>0 γ ˙ → γ ˙ 0 > 0 ) from non-flowing ones (γ ˙ → 0 \dot{\gamma}\to0 γ ˙ → 0 ), where the latter can be further classified into solid-like states with finite deformation (γ = ∫ γ ˙ ( t ) d t < ∞ \gamma=\int\dot{\gamma}(t)\,\mathrm{d}t<\infty γ = ∫ γ ˙ ( t ) d t < ∞ ) from those showing creeping flow (γ → ∞ \gamma\to\infty γ → ∞ even though γ ˙ → 0 \dot{\gamma}\to0 γ ˙ → 0 ). This asymptotic behavior is an important predictor of material failure. Also the more general setting of imposed oscillatory stress, σ ( t ) = σ 0 sin ( ω t ) \sigma(t)=\sigma_0\sin(\omega t) σ ( t ) = σ 0 sin ( ω t ) , is a commonly used and extremely powerful protocol to understand materials.
The solution of this inverse problem is numerically difficult. This is especially true if the integral kernel G G G decays very slowly, so that γ ˙ ( t ) \dot{\gamma}(t) γ ˙ ( t ) might change non-trivially over many decades in time, requiring adaptive-time-grid solvers. In principle, one would also need to take into account the Navier-Stokes equation in this case, but we will concentrate on slowly varying flows where the inertial term can be neglected. Recall that for a flow of average density ϱ \varrho ϱ and of lateral extension L L L , one expects these inertial terms to decay on the timescale of hydrodynamic momentum diffusion, τ u = ϱ L 2 / η \tau_u=\varrho L^2/\eta τ u = ϱ L 2 / η .
We discuss in this paper a robust and versatile method for the numerical inversion of ICEs, using creep and a specific toy model of glass-forming colloidal suspensions as an illustrative example. The paper is organized as follows: We first (§ 2 ) provide a set of relevant examples of ICE that necessitate our approach. In § 3 we summarize some analytical background on the type of equations given by Eq. (1) . In § 4 we demonstrate the use of collocation methods as a powerful numerical technique to solve the problem. Some numerical experiments to demonstrate convergence orders and the computational effort, together with a brief application example constitute § 5 . We conclude the paper with the discussion in § 6 . More rigorous proofs for existence and uniqueness of the solution, and convergence of the collocation method in some generic rheological cases are collected in the appendix.
Integral Models in Glassy Rheology
Let us give some examples with physical applications that fit the class of models we consider.
Example 1 (Kelvin-Voigt model). This is the trivial case where
G K V ( t , t ′ , [ γ ˙ ] ) = G ∞ (3) G_{KV}(t,t',[\dot{\gamma}]) = G_{\infty} \tag*{(3)} G K V ( t , t ′ , [ γ ˙ ]) = G ∞ (3) is constant. In the literature this model is usually given in the form of an equivalent differential equation for γ ( t ) = ∫ t 0 t γ ˙ ( s ) d s \gamma(t) = \int_{t_0}^{t}\dot{\gamma}(s)\,ds γ ( t ) = ∫ t 0 t γ ˙ ( s ) d s :
σ ( t ) = G ∞ γ ( t ) + η γ ˙ ( t ) ; γ ( t 0 ) = 0. (4) \sigma(t) = G_{\infty}\gamma(t) + \eta\dot{\gamma}(t); \qquad\gamma(t_0) = 0. \tag*{(4)} σ ( t ) = G ∞ γ ( t ) + η γ ˙ ( t ) ; γ ( t 0 ) = 0. (4) The model describes a visco-elastic solid with shear modulus G ∞ G_{\infty} G ∞ . Its creep response is an exponentially relaxing shear rate with time scale τ V = η / G ∞ \tau_V = \eta/G_{\infty} τ V = η / G ∞ .
γ ˙ c r e e p , K V ( t ) = σ 0 η e − t / τ V . (5) \dot{\gamma}_{\mathrm{creep},KV}(t) = \frac{\sigma_0}{\eta}e^{-t/\tau_V}. \tag*{(5)} γ ˙ creep , K V ( t ) = η σ 0 e − t / τ V . (5) Since the integrated strain γ < ∞ \gamma< \infty γ < ∞ in this case, this is the response of a solid.
Example 2 (Jeffreys model). Another central model in the literature is the Jeffreys model for a visco-elastic fluid, given by
G J ( t , s , [ γ ˙ ] ) = G ∞ e − t − s τ (6) G_J(t,s,[\dot{\gamma}]) = G_{\infty}e^{-\frac{t-s}{\tau}} \tag*{(6)} G J ( t , s , [ γ ˙ ]) = G ∞ e − τ t − s (6) Also this model can be formulated in terms of a a differential equation,
σ ˙ ( t ) + 1 τ σ ( t ) = ( G ∞ + η τ ) γ ˙ ( t ) + η γ ¨ ( t ) ; γ ˙ ( 0 ) = σ ( 0 ) η . (7) \dot{\sigma}(t) + \frac{1}{\tau}\sigma(t) = \left(G_{\infty}+\frac{\eta}{\tau}\right)\dot{\gamma}(t) + \eta\ddot{\gamma}(t); \qquad\dot{\gamma}(0) = \frac{\sigma(0)}{\eta}. \tag*{(7)} σ ˙ ( t ) + τ 1 σ ( t ) = ( G ∞ + τ η ) γ ˙ ( t ) + η γ ¨ ( t ) ; γ ˙ ( 0 ) = η σ ( 0 ) . (7) In the case η = 0 \eta=0 η = 0 this is Maxwell’s model. The creep response of the Jeffreys model is
γ ˙ c r e e p , J ( t ) = σ 0 η M + σ 0 η ˉ e − t / τ J (8) \dot{\gamma}_{\mathrm{creep},J}(t) = \frac{\sigma_0}{\eta_M} + \frac{\sigma_0}{\bar{\eta}}e^{-t/\tau_J} \tag*{(8)} γ ˙ creep , J ( t ) = η M σ 0 + η ˉ σ 0 e − t / τ J (8) with τ J = η τ / η M \tau_J = \eta\tau/\eta_M τ J = η τ / η M , η M = G ∞ τ + η \eta_M = G_{\infty}\tau+\eta η M = G ∞ τ + η , and η ˉ = η ( 1 + η / G ∞ τ ) \bar{\eta}=\eta(1+\eta/G_{\infty}\tau) η ˉ = η ( 1 + η / G ∞ τ ) . In the case η → 0 \eta\to0 η → 0 , the last term becomes a singular contribution around t = 0 t=0 t = 0 , and Maxwell’s result γ ˙ = σ 0 / G ∞ τ \dot{\gamma}=\sigma_0/G_{\infty}\tau γ ˙ = σ 0 / G ∞ τ is recovered. Since γ ˙ ( t ) \dot{\gamma}(t) γ ˙ ( t ) approaches a constant, this is the response of a viscous fluid.
Many complex materials do not show the clear single-exponential relaxation that is predicted by the Jeffreys or Kelvin-Voigt models, but instead show broad relaxation-rate spectra that are well described by power laws. This motivates the introduction of fractional viscoelastic models, often formulated in terms of the Caputo fractional derivative (n ∈ N 0 n\in\mathbb{N}_0 n ∈ N 0 and 0 < α < 1 0<\alpha<1 0 < α < 1 ),
∂ n + α ∂ t n + α f ( t ) : = 1 Γ ( 1 − α ) ∫ t 0 t f ( n + 1 ) ( t ′ ) d t ′ ( t − t ′ ) α . (9) \frac{\partial^{n+\alpha}}{\partial t^{n+\alpha}}f(t) := \frac{1}{\Gamma(1-\alpha)}\int_{t_0}^{t}\frac{f^{(n+1)}(t')\,dt'}{(t-t')^{\alpha}}. \tag*{(9)} ∂ t n + α ∂ n + α f ( t ) := Γ ( 1 − α ) 1 ∫ t 0 t ( t − t ′ ) α f ( n + 1 ) ( t ′ ) d t ′ . (9)
Example 3 (Fractional Viscoelasticity). The most basic model of fractional viscoelasticity interpolates between fluid- and solid-like response by letting [13 ]
σ ( t ) = G ~ ∂ β ∂ ( t / τ ) β γ ( t ) (10) \sigma(t)=\widetilde{G}\frac{\partial^{\beta}}{\partial(t/\tau)^{\beta}}\gamma(t) \tag*{(10)} σ ( t ) = G ∂ ( t / τ ) β ∂ β γ ( t ) (10) with some internal relaxation time τ \tau τ , such that β = 0 \beta=0 β = 0 corresponds to the Hookean solid, and β = 1 \beta=1 β = 1 to the Newtonian fluid. This is, in fact, an integral model, setting
G F V E ( t , s , [ γ ˙ ] ) = G ~ ( ( t − s ) / τ ) − β Γ ( 1 − β ) (11) G_{FVE}(t,s,[\dot{\gamma}])=\widetilde{G}\frac{\left((t-s)/\tau\right)^{-\beta}}{\Gamma(1-\beta)} \tag*{(11)} G F V E ( t , s , [ γ ˙ ]) = G Γ ( 1 − β ) ( ( t − s ) / τ ) − β (11) (and η = 0 \eta=0 η = 0 although a background viscosity can be easily included). The creep response of this model (for η = 0 \eta=0 η = 0 ) is easily calculated from the Laplace transform; one obtains (for 0 < β < 1 0<\beta<1 0 < β < 1 )
γ ˙ c r e e p , F V E ( t ) = σ 0 Γ ( β ) 1 G ~ τ β t β − 1 . (12) \dot{\gamma}_{\mathrm{creep},FVE}(t)=\frac{\sigma_{0}}{\Gamma(\beta)}\frac{1}{\widetilde{G}\tau^{\beta}}t^{\beta-1}. \tag*{(12)} γ ˙ creep , F V E ( t ) = Γ ( β ) σ 0 G τ β 1 t β − 1 . (12)
The models discussed so far are all linear rheological models, in the sense that they do not incorporate a nonlinear dependence of the stress response to a shear-rate perturbation. The description of materials with non-linear response, in particular yield-stress materials, motivates modifications of the simple viscoelastic models. In particular, the modification of Eq. (7) with a strain-rate-dependent (or stress-dependent) relaxation rate 1 / τ 1/\tau 1/ τ defines a class of models, called White-Metzner models.
Example 4 (Accumulated strain model). This White-Metzner-type model, originally suggested by Bogue [14 ] , lets the stress relaxation depend on the accumulated strain γ ˉ t t ′ = ∫ t ′ t ∣ γ ˙ ( s ) ∣ d s \bar{\gamma}_{tt'}=\int_{t'}^{t}|\dot{\gamma}(s)|ds γ ˉ t t ′ = ∫ t ′ t ∣ γ ˙ ( s ) ∣ d s ,
G A S ( t , t ′ , [ γ ˙ ] ) = G ∞ e − ( t − t ′ ) / τ e − ∫ t ′ t ∣ γ ˙ ( s ) ∣ γ c d s (13) G_{AS}(t,t',[\dot{\gamma}])=G_{\infty}e^{-(t-t')/\tau}e^{-\int_{t'}^{t}\frac{|\dot{\gamma}(s)|}{\gamma_{c}}\,ds} \tag*{(13)} G A S ( t , t ′ , [ γ ˙ ]) = G ∞ e − ( t − t ′ ) / τ e − ∫ t ′ t γ c ∣ γ ˙ ( s ) ∣ d s (13) with a material parameter τ ∈ R + ∪ ∞ \tau\in\mathbb{R}_{+}\cup\infty τ ∈ R + ∪ ∞ that sets the history-independent stress relaxation, and a strain-scale parameter γ c > 0 \gamma_{c}>0 γ c > 0 . The model describes shear thinning, since the effective relaxation rate of stress fluctuations (quantified through G ( t , t ′ ) G(t,t') G ( t , t ′ ) ), increases with increasing shear rate. It was rediscovered in this context as a sketch of MCT and termed “generalized nonlinear Maxwell model” [12 ] . The corresponding ordinary differential equation (ODE) reads
σ ˙ ( t ) + ( 1 τ + ∣ γ ˙ ( t ) ∣ γ c ) σ ( t ) = ( G ∞ + η τ + η ∣ γ ˙ ( t ) ∣ γ c ) γ ˙ ( t ) + η γ ¨ ( t ) ; (14) \dot{\sigma}(t)+\left(\frac{1}{\tau}+\frac{|\dot{\gamma}(t)|}{\gamma_{c}}\right)\sigma(t)=\left(G_{\infty}+\frac{\eta}{\tau}+\eta\frac{|\dot{\gamma}(t)|}{\gamma_{c}}\right)\dot{\gamma}(t)+\eta\ddot{\gamma}(t); \tag*{(14)} σ ˙ ( t ) + ( τ 1 + γ c ∣ γ ˙ ( t ) ∣ ) σ ( t ) = ( G ∞ + τ η + η γ c ∣ γ ˙ ( t ) ∣ ) γ ˙ ( t ) + η γ ¨ ( t ) ; (14) with γ ˙ ( 0 ) = σ ( 0 ) η \dot{\gamma}(0)=\frac{\sigma(0)}{\eta} γ ˙ ( 0 ) = η σ ( 0 ) . Note, in the limits γ c → ∞ \gamma_{c}\to\infty γ c → ∞ and γ c → 0 \gamma_{c}\to0 γ c → 0 one recovers the Maxwell model and the Newtonian fluid respectively.
This example of a nonlinear model in fact reduces the ICE (2) to an ODE. However, this might not be sufficiently generic to capture complex-flow phenomena in glass-forming fluids. For example, Siebenbürger et al. [15 ] studied the emergence of logarithmic creep in colloidal suspensions and suggested a model where the stress relaxation is driven by the instantaneous shear rate rather than the accumulated strain.
Example 5 (Strainrate Relaxation model). Following Ref. [15 ] we consider
G S R R ( t , t ′ , [ γ ˙ ] ) = G ∞ e − t − t ′ τ e − ( t − t ′ ) ∣ γ ˙ ( t ′ ) ∣ γ c . (15) G_{SRR}(t,t',[\dot{\gamma}]) = G_{\infty} e^{-\frac{t-t'}{\tau}} e^{-\frac{(t-t')|\dot{\gamma}(t')|}{\gamma_c}} . \tag*{(15)} G S R R ( t , t ′ , [ γ ˙ ]) = G ∞ e − τ t − t ′ e − γ c ( t − t ′ ) ∣ γ ˙ ( t ′ ) ∣ . (15) This model has a non-analytic dependence on the strain rate at the point γ ˙ = 0 \dot{\gamma}=0 γ ˙ = 0 .
A remark on the appearance of ∣ γ ˙ ∣ |\dot{\gamma}| ∣ γ ˙ ∣ in Eq. (15) might be in order. One expects the full strain-rate tensor to enter any rheological model through its tensorial invariants, in order to ensure that the resulting model is compatible with the frame-invariance principles of classical continuum mechanics. The absolute value ∣ γ ˙ ( t ) ∣ |\dot{\gamma}(t)| ∣ γ ˙ ( t ) ∣ arises in fact from the second invariant of the tensor γ ˙ ( t ) \dot{\gamma}(t) γ ˙ ( t ) under stationary shear.
The examples given above are in particular relevant for the rheology of yield-stress fluids: in the case of stationary shear, γ ˙ ( t ) = γ ˙ = c o n s t . \dot{\gamma}(t)=\dot{\gamma}=\mathrm{const.} γ ˙ ( t ) = γ ˙ = const. (with t 0 → − ∞ t_0 \to-\infty t 0 → − ∞ ), the resulting stationary stress σ ( γ ˙ ) \sigma(\dot{\gamma}) σ ( γ ˙ ) approaches a finite constant, the dynamical yield stress, σ ( γ ˙ ) → σ y > 0 \sigma(\dot{\gamma}) \to\sigma_y > 0 σ ( γ ˙ ) → σ y > 0 , as γ ˙ → 0 \dot{\gamma} \to0 γ ˙ → 0 . This immediately arises from Eq. (2) if G ( t , t ′ ; γ ˙ ) G(t,t';\dot{\gamma}) G ( t , t ′ ; γ ˙ ) becomes a function of γ ˙ ( t − t ′ ) \dot{\gamma}(t-t') γ ˙ ( t − t ′ ) under these circumstances. Both the accumulated strain, and the strainrate-relaxation model fulfill this criterion. In fact, they are equivalent under the imposition of constant strain rate, and offer different physically motivated generalizations to time-dependent shear.
Another hallmark of glass-forming fluids is the appearance of stress overshoots in the startup flow: after imposing a constant shear rate, γ ˙ ( t ) = γ ˙ 0 Θ ( t ) \dot{\gamma}(t)=\dot{\gamma}_0\Theta(t) γ ˙ ( t ) = γ ˙ 0 Θ ( t ) , one observes the stress σ ( t ) \sigma(t) σ ( t ) to increase non-monotonically to its stationary value. Crucially, such stress overshoots can be observed in homogeneous-flow conditions, without the necessity of shear bands [10 ] . In terms of the ICE framework, they thus require to model a dynamical shear modulus that becomes transiently negative as a function of the imposed strain. Theories like MCT suggest to separate this as an effect of a nonlinear response of the elastic coefficients, from the plastic relaxation contributions; effectively one keeps the exponential relaxation for the latter and imposes an empirical function for the strain-dependent G ∞ ( γ ) G_{\infty}(\gamma) G ∞ ( γ ) .
Example 6 (Stressovershoot model). In the empirical model proposed by Amann et al. [16 ] and Frahsa et al. [11 ] one sets
G S O ( t , t ′ , [ γ ˙ ] ) = G ∞ ( γ t t ′ [ γ ˙ ] ) G ∞ G S R R ( t , t ′ , [ γ ˙ ] ) (16) G_{SO}(t,t',[\dot{\gamma}]) = \frac{G_{\infty}(\gamma_{tt'}[\dot{\gamma}])}{G_{\infty}} G_{SRR}(t,t',[\dot{\gamma}]) \tag*{(16)} G S O ( t , t ′ , [ γ ˙ ]) = G ∞ G ∞ ( γ t t ′ [ γ ˙ ]) G S R R ( t , t ′ , [ γ ˙ ]) (16)
where γ t t ′ [ γ ˙ ] = ∫ t ′ t γ ˙ ( s ) d s \gamma_{tt'}[\dot{\gamma}] = \int_{t'}^{t} \dot{\gamma}(s)\,ds γ t t ′ [ γ ˙ ] = ∫ t ′ t γ ˙ ( s ) d s is the total strain2 and the strain-softening function
G ∞ ( γ ) = G ∞ ( 1 − ( γ γ ∗ ) 2 n ) e − ( γ γ ∗ ∗ ) 2 n . (17) G_{\infty}(\gamma) = G_{\infty}\left(1-\left(\frac{\gamma}{\gamma_*}\right)^{2n}\right)e^{-\left(\frac{\gamma}{\gamma_{**}}\right)^{2n}}. \tag*{(17)} G ∞ ( γ ) = G ∞ ( 1 − ( γ ∗ γ ) 2 n ) e − ( γ ∗∗ γ ) 2 n . (17)
Here γ ∗ \gamma_* γ ∗ , γ ∗ ∗ \gamma_{**} γ ∗∗ , and n ∈ N n \in\mathbb{N} n ∈ N are parameters that are adjusted empirically to match position, strength and shape of the stress overshoot in experimental data [16 ] .
We will be using the stress-overshoot (SO) model to elucidate the numerical scheme that we develop in the following.
For completeness, let us briefly mention another widely used class of nonlinear rheological models, the K-BKZ models. They are commonly formulated in terms of the strain, and not the strain rate, as a fundamental model function. However, they fall into the same category as the models we discuss.
Example 7 (Scalar K-BKZ model). The standard K-BKZ model [17 ] in simple shear is given by
σ ( t ) = ∫ t 0 t m ( t − s , γ ( s ) ) γ ( s ) d s . (18) \sigma(t) = \int_{t_0}^{t} m(t-s,\gamma(s))\,\gamma(s)\,ds. \tag*{(18)} σ ( t ) = ∫ t 0 t m ( t − s , γ ( s )) γ ( s ) d s . (18) Identifying G ( t , s , [ γ ˙ ] ) = ∫ s t m ( t − s ′ , γ ( s ′ ) ) d s ′ G(t,s,[\dot{\gamma}]) = \int_s^t m(t-s',\gamma(s'))\,ds' G ( t , s , [ γ ˙ ]) = ∫ s t m ( t − s ′ , γ ( s ′ )) d s ′ allows to cast the model into the form of Eq. (2) , assuming that integration by parts is allowed. (In this definition, G ( t , t ) = 0 G(t,t) = 0 G ( t , t ) = 0 , which is different than the Maxwell-type shear modulus that would imply G ( t , t ) = G ∞ G(t,t) = G_{\infty} G ( t , t ) = G ∞ , since an instantaneous strain contribution is treated differently in the K-BKZ model.) The assumption of a proper convolution in the K-BKZ models, m ( t − s , γ ( s ) ) m(t-s,\gamma(s)) m ( t − s , γ ( s )) instead of m ( t , s , γ ( s ) ) m(t,s,\gamma(s)) m ( t , s , γ ( s )) , rules out some nonlinear rheological features in comparison to the models that we discussed above [18 ] .
Classification
As shown by the examples above, a number of relevant rheological models cannot be formulated as ODEs. Thus, the fundamental evolution equation Eq. (2) remains an integral equation. Specifically, if the dependence of the generalized shear modulus on the shear rate is time-local, G ( t , t ′ , [ γ ˙ ] ) = G ( t , t ′ , γ ˙ ( t ′ ) ) G(t,t',[\dot{\gamma}]) = G(t,t',\dot{\gamma}(t')) G ( t , t ′ , [ γ ˙ ]) = G ( t , t ′ , γ ˙ ( t ′ )) , one obtains a so-called Volterra integral equation. The theory of Volterra equations is rich; we refer the reader to the books by Hackbusch [19 ] and by Wazwaz [20 ] as introductory texts, and to Gripenberg et al. [21 ] for a more thorough and abstract treatment. One distinguishes three types of Volterra integral equations: denoting the unknown function to be determined as the solution of the problem by u ( t ) u(t) u ( t ) , the general Volterra equation reads
a ( t ) u ( t ) = f ( t ) + ∫ t 0 t k ( t , s , u ( s ) ) d s (19) a(t)u(t) = f(t) + \int_{t_0}^{t} k(t,s,u(s))\,ds \tag*{(19)} a ( t ) u ( t ) = f ( t ) + ∫ t 0 t k ( t , s , u ( s )) d s (19)
In the case where a ( t ) ≡ 0 a(t) \equiv0 a ( t ) ≡ 0 , one speaks of a Volterra equation of the first kind. Setting a ( t ) = c o n s t . a(t) = \mathrm{const.} a ( t ) = const. (and thus a = 1 a = 1 a = 1 without loss of generality) one obtains a Volterra equation of second kind, while the fully general case is referred to as a Volterra equation of the third kind. Here k ( t , s , y ) k(t,s,y) k ( t , s , y ) is the kernel of the equation and f ( t ) f(t) f ( t ) is sometimes referred to as the forcing function. In the Volterra equation of third kind the factor a ( t ) a(t) a ( t ) needs to be a continuous function with only a finite number of zeros in the interval [ t 0 , T ] [t_0,T] [ t 0 , T ] over which the solution is sought. We speak of linear (nonlinear) Volterra equations if the kernel k k k is linear (nonlinear) in u ( s ) u(s) u ( s ) . Some nonlinear equations that have been studied include the specific nonlinear version (see, e.g., Ref. [22 ] )
u ( t ) = f ( t ) + ∫ t 0 t k ( t , s , u ( t ) , u ( s ) ) d s (20) u(t) = f(t) + \int_{t_0}^{t} k(t,s,u(t),u(s))\,\mathrm{d}s \tag*{(20)} u ( t ) = f ( t ) + ∫ t 0 t k ( t , s , u ( t ) , u ( s )) d s (20)
or the equations of autocorrelation type ,
u ( t ) = f ( t ) + ∫ t 0 t k ( t , s ) u ( t − s ) u ( s ) d s (21) u(t) = f(t) + \int_{t_0}^{t} k(t,s)u(t-s)u(s)\,\mathrm{d}s \tag*{(21)} u ( t ) = f ( t ) + ∫ t 0 t k ( t , s ) u ( t − s ) u ( s ) d s (21)
termed autoconvolution Volterra integral equation (AVIE). Such autoconvolution equations are known to be ill-posed, in the sense that the solution might change drastically with small changes in the parameters (i.e., they are prone to show bifurcations).
To connect the rheological models to mathematical theory, we identify the strain rate as γ ˙ ( t ) = u ( t ) \dot{\gamma}(t) = u(t) γ ˙ ( t ) = u ( t ) , and the stress as the forcing function, σ ( t ) = f ( t ) \sigma(t) = f(t) σ ( t ) = f ( t ) . Thus,
η u ( t ) = σ ( t ) − ∫ t 0 t G ( t , s , [ u ] ) u ( s ) d s : = H [ u ] ( t ) . (22) \eta u(t) = \sigma(t) - \int_{t_0}^{t} G(t,s,[u])u(s)\,\mathrm{d}s := H[u](t). \tag*{(22)} η u ( t ) = σ ( t ) − ∫ t 0 t G ( t , s , [ u ]) u ( s ) d s := H [ u ] ( t ) . (22)
Adopting the classification of Volterra equations, we distinguish models without background viscosity (η = 0 \eta= 0 η = 0 ) as ICE of first kind, and those with a Newtonian contribution (η ≠ 0 \eta\ne0 η = 0 ) as ICE of second kind. In rheology terms, this amounts to the distinction between the Maxwell and the Jeffreys model. An ICE of the first kind can in principle be transformed into one of second kind by taking a derivative, re-identifying the forcing function with σ ˙ ( t ) \dot{\sigma}(t) σ ˙ ( t ) ; we do not discuss this case separately.
In the general rheological ICE, the dependence of the shear modulus on the strain rate is not time-local; in this general case of a functional dependence, one speaks of H [ u ] H[u] H [ u ] in Eq. (22) as an abstract Volterra operator. Causality implies that it depends only on the past values u ( s ) u(s) u ( s ) with s ∈ [ t 0 , t ] s \in[t_0,t] s ∈ [ t 0 , t ] ; such operators H H H are also called causal or anticipative (cf. [21 Ch.12 sec. 2 and 8] ).
Let us elucidate the classification of the relevant examples given above: The fractional viscoelasticity (FVE) model, Example 3, defines a linear Volterra equation (of first kind in the absence of a Newtonian background viscosity). Its integral kernel is weakly singular; such equations are also called of integral equations of Abel type. The K-BKZ model, Example 7, presents a Volterra equation, either by the transformation discussed there, or directly by identifying the strain as the function to solve for, u ( t ) = γ ( t ) u(t) = \gamma(t) u ( t ) = γ ( t ) . These models are thus subject to the same mathematical analysis and the same numerical treatment.
The strain-rate relaxation (SRR) model, Example 5, with (without) a Newtonian background viscosity defines a nonlinear Volterra equation of second (first) kind. The SO model, Example 6, does not define a Volterra integral equation due to the dependence of the shear modulus on the accumulated strain, and not just the instantaneous strain rate. But it does define a causal abstract Volterra operator, as introduced above.
In particular, we can leverage the theory of Volterra integral equations to assert existence and uniqueness of solutions of the rheological models. For the models that lead to standard Volterra equations, such as the SRR model, this follows directly from the standard theorems for Volterra equations of the second kind with regular kernels. We briefly summarize the relevant results and apply them to the SRR model as an example in Appendix § 11 .§ 1 . For the FVE model, whose kernel is weakly singular, an existence and uniqueness theorem can be found, for example, in the book by Brunner and van der Houwen [23 comments below Theorems 1.3.8 and 1.3.9] . There, also existence and uniqueness theorems for Volterra equations of the first kind with regular and weakly singular kernels can be found [23 Theorems 1.3.11 and 1.3.12] .
As noted, the SO model is more general; here one can however rely on the fixpoint theorems for general functional equations, see for example Gripenberg et al. [21 Ch. 12, Theorem 12.1, Corollaries 12.2 and 12.3] . Since the model (and its variants) have some relevance in the study of glass-forming fluids, we establish existence and uniqueness of (continuous) solutions for this model in Appendix § 11 .§ 2 .
Numerical methods
We examine two numerical approaches to equations of the type (22) , using as examples the SRR model (see example 5) and the SO model (see example 6). In particular we extend standard methods that were originally developed for Volterra equations to the nonlinear Volterra operators for which the SO model is a physically motivated example. A particular emphasis is on collocation methods, introduced below in § 4.3 . The models that we study have a unique solution, as ascertained by the standard techniques for Volterra equations (see Appendix § 11 .§ 1 ), extended to the functional equations of type (22) in Appendix § 11 .§ 2 . For the collocation methods, we are able to prove convergence of the numerical scheme (see Appendix § 12 ), and this serves as a reference point also for simpler, more straightforward quadrature schemes. To our knowledge, such a proof has not appeared yet in the mathematical literature.
Preliminaries
The theory for the numerical solution of Volterra-integral equations is well developed and we follow these standard techniques. We refer to the introductory texts by Hackbusch [19 ] or by Linz [24 ] ; a more extensive treatment is given by Brunner and van der Houwen [23 ] .
Our goal is to numerically approximate the solution of Eq. (22) of second type (with η ≠ 0 \eta\ne0 η = 0 ) on an Intervall [ t 0 , T ] [t_0,T] [ t 0 , T ] . To do that we replace the interval by a grid ( t i ) i = 0 , … , N (t_i)_{i=0,\ldots,N} ( t i ) i = 0 , … , N of points,
t 0 < t 1 < ⋯ < t N ≤ T , (23) t_0 < t_1 < \cdots< t_N \le T , \tag*{(23)} t 0 < t 1 < ⋯ < t N ≤ T , (23)
h i : = t i + 1 − t i . (24) h_i := t_{i+1} - t_i . \tag*{(24)} h i := t i + 1 − t i . (24)
Now, we need to determine the approximations u i ≈ u ( t i ) u_i \approx u(t_i) u i ≈ u ( t i ) to the solution on the grid points t i t_i t i . (For any function, we will denote f ( t i ) = f i f(t_i) = f_i f ( t i ) = f i for brevity if convenient.) The most direct ansatz for a numerical method is to approximate the time-integrals ∫ t 0 t i G ( t i , s , [ u ] ) u ( s ) d s \int_{t_0}^{t_i} G(t_i,s,[u])u(s)\,\mathrm{d}s ∫ t 0 t i G ( t i , s , [ u ]) u ( s ) d s by some quadrature rule using the grid points as nodes. This approach will be shown in § 4.2 . Another approach that has proven fruitful in many areas of numerics, is the collocation ansatz, where the solution is approximated by piecewise polynomial functions. This approach will be followed in § 4.3 . Both approaches can be refined by applying Richardson extrapolation, as found in introductory numerics books (e.g. [25 Ch.3.5] ). For convenience we will outline the main ideas in subsection 4.2 with view towards an application together with direct quadrature methods. The approaches we show here are all formulated for regular kernels G ( t , s , [ u ] ) G(t,s,[u]) G ( t , s , [ u ]) . Fractional models of the form G ( t , s , [ u ] ) = τ β ( t − s ) − β G ~ ( t , s , [ u ] ) G(t,s,[u])=\tau^\beta(t-s)^{-\beta}\widetilde{G}(t,s,[u]) G ( t , s , [ u ]) = τ β ( t − s ) − β G ( t , s , [ u ]) require some extra care, although the ideas of using a quadrature or collocation ansatz survive at least in the case of Volterra equations and are variously called fractional convolutional quadrature, fractional VLM methods, collocation on graded meshes, or collocation with nonpolynomial spline spaces; see for example [23 Ch. 6] for more details.
The numerical methods presented here require the exact solution u u u to be at least continuously differentiable (u ∈ C 1 ( [ 0 , ∞ ) ) u \in C^1([0,\infty)) u ∈ C 1 ([ 0 , ∞ )) ). For this it is necessary, that the forcing function σ ( t ) \sigma(t) σ ( t ) is also at least C 1 C^1 C 1 . Especially jumps in σ ( t ) \sigma(t) σ ( t ) , t ∈ [ 0 , ∞ ) t \in[0,\infty) t ∈ [ 0 , ∞ ) , are strictly speaking not allowed and need to be treated separately.
Direct Quadrature ansatz
η u n = σ n − ∫ t 0 t G ( t , s , [ u ] ) u ( s ) d s (25) \eta u_n = \sigma_n - \int_{t_0}^{t} G(t,s,[u])u(s)\,\mathrm{d}s \tag*{(25)} η u n = σ n − ∫ t 0 t G ( t , s , [ u ]) u ( s ) d s (25)
≈ σ n − ∑ i = 0 n w n , i G ( t n , t i , [ u ] ) u i (26) \approx\sigma_n - \sum_{i=0}^{n} w_{n,i}G(t_n,t_i,[u])u_i \tag*{(26)} ≈ σ n − i = 0 ∑ n w n , i G ( t n , t i , [ u ]) u i (26)
where w n , i w_{n,i} w n , i are the weights of the quadrature rule. We keep the notation [ u ] [u] [ u ] to denote a potentially functional dependence on u u u . Since we assume G G G to be causal, for any given quadrature point this implies a potential dependence on all ( u j ) j = 0 , … , n (u_j)_{j=0,\ldots,n} ( u j ) j = 0 , … , n . The three most elementary examples are
Example 8 (Quadrature rules). Setting Δ t i = t i + 1 − t i \Delta t_i = t_{i+1}-t_i Δ t i = t i + 1 − t i , the well known left-point (LP), right-point (RP), and trapezoidal (TZ) quadrature rules,
∫ t i t i + 1 G ( t n , s , [ u ] ) u ( s ) d s ≈ Δ t i G ( t n , t i , [ u ] ) u i (27) \int_{t_i}^{t_{i+1}} G(t_n,s,[u])\,u(s)\,\mathrm{d}s \approx\Delta t_i G(t_n,t_i,[u])\,u_i \tag*{(27)} ∫ t i t i + 1 G ( t n , s , [ u ]) u ( s ) d s ≈ Δ t i G ( t n , t i , [ u ]) u i (27) ∫ t i t i + 1 G ( t n , s , [ u ] ) u ( s ) d s ≈ Δ t i G ( t n , t i + 1 , [ u ] ) u i + 1 (28) \int_{t_i}^{t_{i+1}} G(t_n,s,[u])\,u(s)\,\mathrm{d}s \approx\Delta t_i G(t_n,t_{i+1},[u])\,u_{i+1} \tag*{(28)} ∫ t i t i + 1 G ( t n , s , [ u ]) u ( s ) d s ≈ Δ t i G ( t n , t i + 1 , [ u ]) u i + 1 (28) ∫ t i t i + 1 G ( t n , s , [ u ] ) u ( s ) d s ≈ Δ t i 2 ( G ( t n , t i + 1 , [ u ] ) u i + 1 − G ( t n , t i , [ u ] ) u i ) (29) \int_{t_i}^{t_{i+1}} G(t_n,s,[u])\,u(s)\,\mathrm{d}s \approx\frac{\Delta t_i}{2}\left(G(t_n,t_{i+1},[u])\,u_{i+1} - G(t_n,t_i,[u])\,u_i\right) \tag*{(29)} ∫ t i t i + 1 G ( t n , s , [ u ]) u ( s ) d s ≈ 2 Δ t i ( G ( t n , t i + 1 , [ u ]) u i + 1 − G ( t n , t i , [ u ]) u i ) (29) yield, respectively,
η u n = σ n − ∑ i = 0 n − 1 h i G ( t n , t i , [ u ] ) u i (30) \eta u_n = \sigma_n - \sum_{i=0}^{n-1} h_i G(t_n,t_i,[u])\,u_i \tag*{(30)} η u n = σ n − i = 0 ∑ n − 1 h i G ( t n , t i , [ u ]) u i (30) η u n = σ n − ∑ i = 0 n − 1 h i G ( t n , t i + 1 , [ u ] ) u i + 1 (31) \eta u_n = \sigma_n - \sum_{i=0}^{n-1} h_i G(t_n,t_{i+1},[u])\,u_{i+1} \tag*{(31)} η u n = σ n − i = 0 ∑ n − 1 h i G ( t n , t i + 1 , [ u ]) u i + 1 (31) η u n = σ n − ∑ i = 0 n − 1 h i 2 ( G ( t n , t i + 1 , [ u ] ) u i + 1 + G ( t n , t i , [ u ] ) u i ) (32) \eta u_n = \sigma_n - \sum_{i=0}^{n-1} \frac{h_i}{2}\left(G(t_n,t_{i+1},[u])\,u_{i+1} + G(t_n,t_i,[u])\,u_i\right) \tag*{(32)} η u n = σ n − i = 0 ∑ n − 1 2 h i ( G ( t n , t i + 1 , [ u ]) u i + 1 + G ( t n , t i , [ u ]) u i ) (32) In the case of a Volterra equation of second kind, G ( t , s , [ u ] ) = G ( t , s , u ( s ) ) G(t,s,[u]) = G(t,s,u(s)) G ( t , s , [ u ]) = G ( t , s , u ( s )) , the choice of quadrature rule is sufficient to define the numerical scheme. Higher-order quadrature rules could in principle be used, but they necessarily lead to multistep methods that cannot start on their own, and they become more intricate on nonlinear grids ( t i ) i = 0 , … , N (t_i)_{i=0,\ldots,N} ( t i ) i = 0 , … , N . For this reason we do not consider higher-order direct quadrature methods here.
Note that for Volterra equations, there is also a close correspondence of these quadrature methods to linear multi-step methods for ODEs: the leftpoint method corresponds to an explicit Euler method, the rightpoint method to an implicit Euler method, and the trapezoidal rule corresponds to the trapezoidal method for ODEs.3 Such quadrature rules are also called ( ρ , σ ) (\rho,\sigma) ( ρ , σ ) -reducible. At least for linear Volterra equations, G ( t , s , u ( s ) ) u ( s ) = G ( t , s ) u ( s ) G(t,s,u(s))u(s) = G(t,s)u(s) G ( t , s , u ( s )) u ( s ) = G ( t , s ) u ( s ) , one can show that if the underlying ODE method has certain properties, the resulting algorithm applied to the integral equation of first kind (η = 0 \eta= 0 η = 0 ) converges with the same order [26 ] . The three methods outlined above fulfill the required properties. The order of convergence p p p is given by the leading order of the error in terms of the step size, h p h^p h p . The leftpoint rule for a Volterra equation of second kind leads to a method of order p = 1 p = 1 p = 1 , while the trapezoidal rule results in a method of order p = 2 p = 2 p = 2 [23 Corrolary 3.2.1] .
For a Volterra equation it is sufficient to choose a quadrature rule. However, in the case of more general operators G ( t , s , [ u ] ) G(t,s,[u]) G ( t , s , [ u ]) one needs to additionally chose how to approximate G ( t n , t i , [ u ] ) G(t_n,t_i,[u]) G ( t n , t i , [ u ]) at the corresponding grid points, through representing the functional dependence on the numerical grid. This is a potentially crucial detail: G ( t n , t i , [ u ] ) G(t_n,t_i,[u]) G ( t n , t i , [ u ]) might involve u n u_n u n so that the chosen method necessarily becomes an implicit method. Nonlinear integral equations with such functional dependence can show bifurcations in the t = ∞ t=\infty t = ∞ limit so that a proper treatment of the implicit term is crucial to obtain solutions that are asymptotically correct as t t t becomes arbitrarily large – in a related context, MCT gives a mathematically intricate example [27 ] .
The direct quadrature methods for models like the SO model are characterized by their “outer” quadrature scheme (used to approximate ∫ t 0 t G ( t , s , [ u ] ) d s \int_{t_0}^{t}G(t,s,[u])\,\mathrm{d}s ∫ t 0 t G ( t , s , [ u ]) d s ) supplemented by the “inner” scheme (used to approximate z ( t , s ) = ∫ s t u ( t ′ ) d t ′ z(t,s)=\int_s^t u(t')\,\mathrm{d}t' z ( t , s ) = ∫ s t u ( t ′ ) d t ′ . These need not be the same, and we name them outer–inner accordingly (for example, leftpoint–leftpoint (LP-LP) etc.)
Example 9 (Stressovershoot model). In the SO model, Example 6, G ( t , s , [ u ] ) G(t,s,[u]) G ( t , s , [ u ]) for any interval [ s , t ] [s,t] [ s , t ] is of the form
G ( t , s , [ u ] ) = G ( t , s , ∫ s t u ( t ′ ) d t ′ , u ( s ) ) (33) G(t,s,[u])=G\left(t,s,\int_s^t u(t')\,\mathrm{d}t',u(s)\right) \tag*{(33)} G ( t , s , [ u ]) = G ( t , s , ∫ s t u ( t ′ ) d t ′ , u ( s ) ) (33)
We can choose a direct quadrature approximation for the strain integral z ( t , s ) z(t,s) z ( t , s ) ,
z n , i = ∑ l = i n w n , l u l ≈ z ( t n , t i ) = ∫ t i t n u ( t ′ ) d t ′ (34) z_{n,i}=\sum_{l=i}^{n}w_{n,l}u_l\approx z(t_n,t_i)=\int_{t_i}^{t_n}u(t')\,\mathrm{d}t' \tag*{(34)} z n , i = l = i ∑ n w n , l u l ≈ z ( t n , t i ) = ∫ t i t n u ( t ′ ) d t ′ (34)
and the trapezoidal rule for the outer integral. Thus, the trapezoidal–leftpoint scheme is obtained,
η u n = σ n − ∑ i = 0 n − 1 h i 2 ( G ( t n , t i + 1 , z n , i + 1 , u i + 1 ) u i + 1 + G ( t n , t i , z n , i , u i ) u i ) , (35) \begin{aligned}
\eta u_n=\sigma_n-\sum_{i=0}^{n-1}\frac{h_i}{2}(G(t_n,t_{i+1},z_{n,i+1},u_{i+1})u_{i+1} \\
&+G(t_n,t_i,z_{n,i},u_i)u_i), \tag*{(35)}
\end{aligned} η u n = σ n − i = 0 ∑ n − 1 2 h i ( G ( t n , t i + 1 , z n , i + 1 , u i + 1 ) u i + 1 + G ( t n , t i , z n , i , u i ) u i ) , (35)
z n , i = ∑ j = i n − 1 h j u j , (36) z_{n,i}=\sum_{j=i}^{n-1}h_j u_j, \tag*{(36)} z n , i = j = i ∑ n − 1 h j u j , (36)
and similar schemes follow by analogy. The leftpoint–leftpoint method for u n u_n u n is fully explicit, while the trapezoidal–leftpoint method is implicit but leads to a linear system for u n u_n u n . The trapezoidal–trapezoidal method is fully implicit and u n u_n u n is determined by the solution of a (typically) nonlinear equation. Our experimental convergence analysis (see Figure 1 ) indicates that the former two methods are first order methods, while the third method appears to be of second order as expected from the orders of the underlying quadrature rules.4
Figure 1. Convergence test for the direct quadrature methods, applied to the test case of the stressovershoot model described in § 5.1 . The error norm relative to the reference exact solution (see § 5.1 ) is shown as a function of the inverse steplength 1 / h 1/h 1/ h . The dotted lines represent the slopes h p h^p h p , p = 1 , 2 , 3 , 4 p = 1, 2, 3, 4 p = 1 , 2 , 3 , 4 respectively. The methods are labeled by their quadrature rule for the outer integral followed by that used for the inner integral, cf. Eq. (31) as an example.
The convergence order of the given quadrature rule can be improved by a standard method called Richardson extrapolation (see, e.g., Ref. [25 Ch. 3.5] ). The idea rests on the observation that in many cases the error of the numerical solution u ~ h ( t ) \tilde{u}_h(t) u ~ h ( t ) with respect to the true solution u ( t ) u(t) u ( t ) can be expanded as a polynomial in the step size h h h as h → 0 h \to0 h → 0 :
u ( t ) − u ~ h ( t ) = c 1 ( t ) h p + c 2 ( t ) h p + 1 + … (37) u(t)-\tilde{u}_h(t)=c_1(t)h^p+c_2(t)h^{p+1}+\ldots \tag*{(37)} u ( t ) − u ~ h ( t ) = c 1 ( t ) h p + c 2 ( t ) h p + 1 + … (37)
where p p p is the order of convergence. Interpreting u ~ h ( t ) \tilde{u}_h(t) u ~ h ( t ) as a polynomial P R ( h ) P_R(h) P R ( h ) in h h h , the exact solution is P R ( 0 ) P_R(0) P R ( 0 ) and can be approximated from knowledge of the values P R ( h ) P_R(h) P R ( h ) and P R ( h / 2 ) P_R(h/2) P R ( h /2 ) . This defines the improved numerical solution
u R ( t ) = P R ( 0 ) = α u ~ h ( t ) + ( 1 − α ) u ~ h / 2 ( t ) , α = 1 1 − 2 p (38) u_R(t)=P_R(0)=\alpha\tilde{u}_h(t)+(1-\alpha)\tilde{u}_{h/2}(t), \qquad\alpha=\frac{1}{1-2^p} \tag*{(38)} u R ( t ) = P R ( 0 ) = α u ~ h ( t ) + ( 1 − α ) u ~ h /2 ( t ) , α = 1 − 2 p 1 (38)
as an approximation of order O ( h p + 1 ) O(h^{p+1}) O ( h p + 1 ) to the exact solution. This scheme can be iterated if more higher-order error terms are known. For the trapezoidal rule, the error can be developed in terms of h 2 n h^{2n} h 2 n (n ∈ N n \in\mathbb{N} n ∈ N ), however this is only true for the quadrature method if the integrand is smooth enough (see Appendix Appendix C). In the physical problem at hand, this can be guaranteed for problems where the shear rate never changes sign, due to the appearance of ∣ γ ˙ ( t ) ∣ |\dot{\gamma}(t)| ∣ γ ˙ ( t ) ∣ in some of the models.
Richardson extrapolation can be generalized to non-uniform grids: it is sufficient that the expression for the error in terms of the step length factorizes the time- and stepsize-dependence, so that the Richardson scheme can be applied by using the solution on a given grid, and the grid that is defined by subdividing each grid intervall by two.
Figure 1 shows the performance of several direct quadrature methods for the SO model, i.e., a model of generalized Volterra-operator type. Numerical details for the model and the used grids are given below in § 5.1 . One observes the expected orders of convergence, p = 1 p=1 p = 1 for the LP rule, p = 2 p=2 p = 2 for the TZ rule, with the quadrature of the lower order dictating the convergence for the combined scheme. Richardson extrapolation increases the order of convergence by one for methods involving the LP rule, and by two if the extrapolation is iterated; for the TZ rule, each extrapolation gains two orders of convergence. Thus the TZ-TZ scheme with Richardson extrapolation is a fourth-order scheme.
Collocation ansatz
Collocation methods are close cousins of the well-known Runge-Kutta schemes for the numerical solution of (stiff) ODEs. An extensive introduction into collocation methods applied to Volterra equations can be found in the book by Brunner [28 ] .
In the following we will outline the principal idea and show how to apply this to derive methods for more general types of evolutionary integral equations. The essential idea of an collocation ansatz is to approximate the exact solution u u u by a spline function p p p , e.g. by a piecewise polynomial function
u ( t ) ≈ p ( t ) = p i ( t ) = ∑ j = 0 m a i , j P j ( t ) , t ∈ ( t i , t i + 1 ] (39) u(t) \approx p(t) = p_i(t) = \sum_{j=0}^{m} a_{i,j}P_j(t), \quad t \in(t_i,t_{i+1}] \tag*{(39)} u ( t ) ≈ p ( t ) = p i ( t ) = j = 0 ∑ m a i , j P j ( t ) , t ∈ ( t i , t i + 1 ] (39)
u ( t i ) ≈ p i − 1 ( t i ) (40) u(t_i) \approx p_{i-1}(t_i) \tag*{(40)} u ( t i ) ≈ p i − 1 ( t i ) (40)
of some fixed maximum degree m m m , where the knots correspond to the grid-points t i t_i t i , i = 0 , … , N i = 0,\ldots,N i = 0 , … , N and P j ( t ) P_j(t) P j ( t ) , j = ( 1 , … m ) j = (1,\ldots m) j = ( 1 , … m ) , is some polynomial basis. A usual choice are the monomials, P j ( t ) = t j P_j(t) = t^j P j ( t ) = t j , but in the present case we found it more stable to choose a Lagrange-polynomial basis P j ( t ) = L j + 1 ( t ) P_j(t) = L_{j+1}(t) P j ( t ) = L j + 1 ( t ) .
Additionally, for each interval [ t i , t i + 1 ] [t_i,t_{i+1}] [ t i , t i + 1 ] one defines a set of collocation points t ~ i , k = t i + c k h i \tilde{t}_{i,k} = t_i + c_kh_i t ~ i , k = t i + c k h i , k = 1 , … , m + 1 k = 1,\ldots,m+1 k = 1 , … , m + 1 based on the collocation parameters
0 ≤ c 1 < ⋯ < c m + 1 ≤ 1. (41) 0 \le c_1 < \cdots< c_{m+1} \le1. \tag*{(41)} 0 ≤ c 1 < ⋯ < c m + 1 ≤ 1. (41)
In the numerical analysis of Volterra equations it is common to use discontinuous splines that may have jump discontinuities at the knots t i t_i t i . Thus, on each interval [ t i , t i + 1 ] [t_i,t_{i+1}] [ t i , t i + 1 ] the collocation polynomial is uniquely determined by its values on the m + 1 m+1 m + 1 collocation points t ~ i , k \tilde{t}_{i,k} t ~ i , k , k = 1 … m + 1 k = 1\ldots m+1 k = 1 … m + 1 . Demanding that the collocation polynomials solve the given problem exactly on the collocation points defines a nonlinear system of equations for the coefficients a i , j a_{i,j} a i , j of the polynomials. Applied to (22) , the collocation equation reads
σ ( t ~ i , k ) = η p i ( t ~ i , k ) + ∫ t 0 t ~ i , k G ( t , s , [ p ] ) p i ( s ) d s , i ∈ { 0 , … , N − 1 } ; k ∈ { 1 , … , m + 1 } . (42) \sigma(\tilde{t}_{i,k}) = \eta p_i(\tilde{t}_{i,k}) + \int_{t_0}^{\tilde{t}_{i,k}} G(t,s,[p])p_i(s)\,\mathrm{d}s,\quad i \in\{0,\ldots,N-1\};\ k \in\{1,\ldots,m+1\}. \tag*{(42)} σ ( t ~ i , k ) = η p i ( t ~ i , k ) + ∫ t 0 t ~ i , k G ( t , s , [ p ]) p i ( s ) d s , i ∈ { 0 , … , N − 1 } ; k ∈ { 1 , … , m + 1 } . (42)
Note that depending on the choice of the p i ( t ) p_i(t) p i ( t ) , it can even be possible to evaluate the integral analytically, as demonstrated by the example below. This offers a nice advantage for some models and can be helpful to capture properties of the integral that might be difficult to restore purely by numerical quadrature unless the time step is chosen very small. Such applications of simple variants of collocation methods are already well known from, e.g., Fouier-Laplace transforms of numerical data over wide frequency bands [29 ] , or for the evaluation of MCT memory kernels in two spatial dimensions where singular Jacobians appear due to the desired variable transform [30 ] .
Example 10 (Collocation in the SRR model using piecewise constant splines). A particularly simple collocation method is obtained by letting p i ( t ) = p i p_i(t) = p_i p i ( t ) = p i be constant. Then, only one collocation parameter c 1 = θ ∈ ( 0 , 1 ] c_1 = \theta\in(0,1] c 1 = θ ∈ ( 0 , 1 ] is needed. For the SRR model, Example 5, with τ = ∞ \tau= \infty τ = ∞ , the integrals can then be performed
analytically and lead to the collocation equation
σ ( t i + θ h i ) = η p i + ∫ t i t i + θ h i G ∞ e − ( t i + θ h i − s ) ∣ p i ∣ γ c p i d s + ∑ j = 0 i − 1 ∫ t j t j + 1 G ∞ e − ( t i + θ h i − s ) ∣ p j ∣ γ c p j d s = η p i + G ∞ γ c p i ∣ p i ∣ ( 1 − e − θ h i ∣ p i ∣ γ c ) + ∑ j = 0 i − 1 G ∞ γ c p j ∣ p j ∣ ( e − ( t i + θ h i − t j + 1 ) ∣ p j ∣ γ c − e − ( t i + θ h i − t j ) ∣ p j ∣ γ c ) , (43) \begin{aligned}
\sigma(t_i+\theta h_i)={}&\eta p_i+\int_{t_i}^{t_i+\theta h_i}G_\infty e^{-\frac{(t_i+\theta h_i-s)|p_i|}{\gamma_c}}p_i\,\mathrm{d}s+\sum_{j=0}^{i-1}\int_{t_j}^{t_{j+1}}G_\infty e^{-\frac{(t_i+\theta h_i-s)|p_j|}{\gamma_c}}p_j\,\mathrm{d}s\\
={}&\eta p_i+G_\infty\gamma_c\frac{p_i}{|p_i|}\left(1-e^{-\frac{\theta h_i|p_i|}{\gamma_c}}\right)\\
&+\sum_{j=0}^{i-1}G_\infty\gamma_c\frac{p_j}{|p_j|}\left(e^{-\frac{(t_i+\theta h_i-t_{j+1})|p_j|}{\gamma_c}}-e^{-\frac{(t_i+\theta h_i-t_j)|p_j|}{\gamma_c}}\right), \tag*{(43)}
\end{aligned} σ ( t i + θ h i ) = = η p i + ∫ t i t i + θ h i G ∞ e − γ c ( t i + θ h i − s ) ∣ p i ∣ p i d s + j = 0 ∑ i − 1 ∫ t j t j + 1 G ∞ e − γ c ( t i + θ h i − s ) ∣ p j ∣ p j d s η p i + G ∞ γ c ∣ p i ∣ p i ( 1 − e − γ c θ h i ∣ p i ∣ ) + j = 0 ∑ i − 1 G ∞ γ c ∣ p j ∣ p j ( e − γ c ( t i + θ h i − t j + 1 ) ∣ p j ∣ − e − γ c ( t i + θ h i − t j ) ∣ p j ∣ ) , (43)
defining an implicit equation for the coefficients p i p_i p i . By Identifying
u ( t i + 1 ) ≈ u i + 1 = p ( t i + 1 ) = p i ( t i + 1 ) = p i (44) u(t_{i+1})\approx u_{i+1}=p(t_{i+1})=p_i(t_{i+1})=p_i \tag*{(44)} u ( t i + 1 ) ≈ u i + 1 = p ( t i + 1 ) = p i ( t i + 1 ) = p i (44)
and rearranging some terms the derived collocation method reads
u i + 1 = σ ( t i + θ h i ) η − G ∞ γ c η u i + 1 ∣ u i + 1 ∣ ( 1 − e − θ h i ∣ u i + 1 ∣ γ c ) (45) u_{i+1}=\frac{\sigma(t_i+\theta h_i)}{\eta}-\frac{G_\infty\gamma_c}{\eta}\frac{u_{i+1}}{|u_{i+1}|}\left(1-e^{-\frac{\theta h_i|u_{i+1}|}{\gamma_c}}\right) \tag*{(45)} u i + 1 = η σ ( t i + θ h i ) − η G ∞ γ c ∣ u i + 1 ∣ u i + 1 ( 1 − e − γ c θ h i ∣ u i + 1 ∣ ) (45)
+ G ∞ γ c η ∑ j = 0 i − 1 u j + 1 ∣ u j + 1 ∣ ( e − ( t i + θ h i − t j + 1 ) ∣ u j + 1 ∣ γ c − e − ( t i + θ h i − t j ) ∣ u j + 1 ∣ γ c ) (46) +\frac{G_\infty\gamma_c}{\eta}\sum_{j=0}^{i-1}\frac{u_{j+1}}{|u_{j+1}|}\left(e^{-\frac{(t_i+\theta h_i-t_{j+1})|u_{j+1}|}{\gamma_c}}-e^{-\frac{(t_i+\theta h_i-t_j)|u_{j+1}|}{\gamma_c}}\right) \tag*{(46)} + η G ∞ γ c j = 0 ∑ i − 1 ∣ u j + 1 ∣ u j + 1 ( e − γ c ( t i + θ h i − t j + 1 ) ∣ u j + 1 ∣ − e − γ c ( t i + θ h i − t j ) ∣ u j + 1 ∣ ) (46)
This yields a first order method for all choices of θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] (while θ = 0 \theta=0 θ = 0 would generate an explicit method).
In general, for the solution of Volterra equations of second kind (η ≠ 0 \eta\ne0 η = 0 ) it is known, that for any choice of collocation parameters 0 ≤ c 1 < ⋯ < c m + 1 ≤ 1 0\le c_1<\cdots<c_{m+1}\le1 0 ≤ c 1 < ⋯ < c m + 1 ≤ 1 the resulting method is at least of order O ( h ) = h m + 1 O(h)=h^{m+1} O ( h ) = h m + 1 [23 Theorem 5.3.2] . However, the maximum convergence rate can only be achieved if the kernel G ( t , s , y ) G(t,s,y) G ( t , s , y ) is smooth enough. In the above example of the SRR model, where a non-analytic dependence on ∣ γ ˙ ( t ) ∣ |\dot{\gamma}(t)| ∣ γ ˙ ( t ) ∣ appears in the model definition, this implies that the maximum possible order of convergence by these methods is limited by O ( h ) = h O(h)=h O ( h ) = h if the exact solution u u u changes sign (e.g. if u ( t ∗ ) = 0 u(t_\ast)=0 u ( t ∗ ) = 0 for some t ∗ ∈ [ t 0 , t ] t_\ast\in[t_0,t] t ∗ ∈ [ t 0 , t ] ). The collocation method still works since it only requires the solution to be C 1 C^1 C 1 (once continuously differentiable), which is the case if σ \sigma σ is sufficiently smooth. If the exact solution does not change sign numerical approximation schemes with higher orders are possible (see, e.g., Sec. 5.2 for the example of creep).
In general, the integrals remaining in the collocation equation will have to be evaluated numerically. In principle, one can refine separate numerical grids to solve the integrals on each collocation interval to the maximum available numerical precision. However, it is more straightforward to fix a quadrature rule with a given numerical precision. In this latter case it is common to use the collocation points t i , k t_{i,k} t i , k as the abscissas for the quadrature rule [28 ] . We adopt this practice here.
A strong merit of the collocation ansatz is the flexibility in dealing with generalized Volterra operators: it leads to a self-consistent ansatz for elastic moduli of the form G ( t , s , [ u ] ) G(t,s,[u]) G ( t , s , [ u ]) , where the approximation replaces that with G ( t , s , [ p ] ) G(t,s,[p]) G ( t , s , [ p ]) for each collocation interval.
Example 11 (Collocation in the SO model using constant splines). As in the previous example we choose the collocation polynomials p i ( t ) = p i p_i(t)=p_i p i ( t ) = p i to be constant, and we set as collocation parameter c 1 = θ ∈ ( 0 , 1 ] c_1=\theta\in(0,1] c 1 = θ ∈ ( 0 , 1 ] . Recall that in the SO model (Example 6) the strain-history dependent elastic modulus is of the form
G ( t , s , [ u ] ) = G ( t , s , ∫ s t u ( t ′ ) d t ′ , u ( s ) ) . G(t,s,[u])=G\left(t,s,\int_s^t u(t')\mathrm{d}t',u(s)\right). G ( t , s , [ u ]) = G ( t , s , ∫ s t u ( t ′ ) d t ′ , u ( s ) ) . Inserting the collocation ansatz yields
G ( t i + θ h i , s , [ p ] ) = G ( t i + θ h i , s , ∫ s t i + θ h i p ( t ′ ) d t ′ , p j ) , s ∈ [ t j , t j + 1 ] (47) G(t_i+\theta h_i,s,[p])=G\left(t_i+\theta h_i,s,\int_s^{t_i+\theta h_i}p(t')\mathrm{d}t',p_j\right),\qquad s\in[t_j,t_{j+1}] \tag*{(47)} G ( t i + θ h i , s , [ p ]) = G ( t i + θ h i , s , ∫ s t i + θ h i p ( t ′ ) d t ′ , p j ) , s ∈ [ t j , t j + 1 ] (47) Since the integral over the collocation polynomial can be performed exactly,
∫ s t i + θ h i p ( t ′ ) d t ′ = { ( t i + θ h i − s ) p i , s ∈ [ t i , t i + θ h i ] θ h i p i + ∑ k = j + 1 i − 1 h k p k + ( t j + 1 − s ) p j , s ∈ [ t j , t j + 1 ] (48) \int_s^{t_i+\theta h_i}p(t')\mathrm{d}t'=
\begin{cases}
(t_i+\theta h_i-s)p_i, & s\in[t_i,t_i+\theta h_i]\\
\theta h_i p_i+\sum_{k=j+1}^{i-1}h_kp_k+(t_{j+1}-s)p_j, & s\in[t_j,t_{j+1}]
\end{cases} \tag*{(48)} ∫ s t i + θ h i p ( t ′ ) d t ′ = { ( t i + θ h i − s ) p i , θ h i p i + ∑ k = j + 1 i − 1 h k p k + ( t j + 1 − s ) p j , s ∈ [ t i , t i + θ h i ] s ∈ [ t j , t j + 1 ] (48) this leads to the collocation equation
σ ( t i + θ h i ) = η p i + ∫ t i t i + θ h i G ( t i + θ h i , s , ( t i + θ h i − s ) p i , p i ) p i d s + ∑ j = 0 i − 1 ∫ t j t j + 1 G ( t i + θ h i , s , ∫ s t i + θ h i p ( t ′ ) d t ′ , p j ) p j d s (49) \begin{aligned}
\sigma(t_i+\theta h_i)
&=\eta p_i+\int_{t_i}^{t_i+\theta h_i}G\left(t_i+\theta h_i,s,(t_i+\theta h_i-s)p_i,p_i\right)p_i\mathrm{d}s\\
&\quad+\sum_{j=0}^{i-1}\int_{t_j}^{t_{j+1}}G\left(t_i+\theta h_i,s,\int_s^{t_i+\theta h_i}p(t')\mathrm{d}t',p_j\right)p_j\mathrm{d}s \tag*{(49)}
\end{aligned} σ ( t i + θ h i ) = η p i + ∫ t i t i + θ h i G ( t i + θ h i , s , ( t i + θ h i − s ) p i , p i ) p i d s + j = 0 ∑ i − 1 ∫ t j t j + 1 G ( t i + θ h i , s , ∫ s t i + θ h i p ( t ′ ) d t ′ , p j ) p j d s (49) We fix a quadrature rule for the remaining integrals based on the collocation parameter, using the rectangular rule with abscissas t j , 1 = t j + θ h j t_{j,1}=t_j+\theta h_j t j , 1 = t j + θ h j ,
∫ t j t j + 1 G ( t i + θ h i , s , ∫ s t i + θ h i p ( t ′ ) d t ′ , p j ) p j d s ≈ h j p j G ( t i + θ h i , t j + θ h j , z i j , p j ) (50) \int_{t_j}^{t_{j+1}}G\left(t_i+\theta h_i,s,\int_s^{t_i+\theta h_i}p(t')\mathrm{d}t',p_j\right)p_j\mathrm{d}s\approx h_jp_jG\left(t_i+\theta h_i,t_j+\theta h_j,z_{ij},p_j\right) \tag*{(50)} ∫ t j t j + 1 G ( t i + θ h i , s , ∫ s t i + θ h i p ( t ′ ) d t ′ , p j ) p j d s ≈ h j p j G ( t i + θ h i , t j + θ h j , z ij , p j ) (50) z i j = ∫ t j + θ h j t i + θ h i p ( t ′ ) d t ′ (51) z_{ij}=\int_{t_j+\theta h_j}^{t_i+\theta h_i}p(t')\mathrm{d}t' \tag*{(51)} z ij = ∫ t j + θ h j t i + θ h i p ( t ′ ) d t ′ (51) Identifying again u i + 1 = p i u_{i+1}=p_i u i + 1 = p i and rearranging some terms leads to the following method:
u i + 1 = σ ( t i + θ h i ) η + θ h i u i + 1 η G ( t i + θ h i , t i + θ h i , 0 , u i + 1 ) + ∑ j = 0 i − 1 h j u j + 1 η G ( t i + θ h i , t j + θ h j , z i j , u j + 1 ) (52) \begin{aligned}
u_{i+1}
&=\frac{\sigma(t_i+\theta h_i)}{\eta}
+\frac{\theta h_i u_{i+1}}{\eta}G\left(t_i+\theta h_i,t_i+\theta h_i,0,u_{i+1}\right)\\
&\quad+\sum_{j=0}^{i-1}\frac{h_j u_{j+1}}{\eta}G\left(t_i+\theta h_i,t_j+\theta h_j,z_{ij},u_{j+1}\right) \tag*{(52)}
\end{aligned} u i + 1 = η σ ( t i + θ h i ) + η θ h i u i + 1 G ( t i + θ h i , t i + θ h i , 0 , u i + 1 ) + j = 0 ∑ i − 1 η h j u j + 1 G ( t i + θ h i , t j + θ h j , z ij , u j + 1 ) (52) z i j = θ h i u i + 1 + ∑ k = j + 1 i − 1 h k u k + 1 + h j ( 1 − θ ) u j + 1 (53) z_{ij}=\theta h_i u_{i+1}+\sum_{k=j+1}^{i-1}h_k u_{k+1}+h_j(1-\theta)u_{j+1} \tag*{(53)} z ij = θ h i u i + 1 + k = j + 1 ∑ i − 1 h k u k + 1 + h j ( 1 − θ ) u j + 1 (53)
Note, that the choice θ = 1 \theta= 1 θ = 1 leads to a method that could alternatively also have been obtained by direct quadrature (see section 4.2) with the rightpoint–rectangular rule to approximate ∫ G ( t , s , z ( t , s ) , u ( s ) ) u ( s ) d s \int G(t,s,z(t,s),u(s))u(s)\,\mathrm{d}s ∫ G ( t , s , z ( t , s ) , u ( s )) u ( s ) d s and z ( t , s ) = ∫ s t u ( t ′ ) d t ′ z(t,s) = \int_s^t u(t')\,\mathrm{d}t' z ( t , s ) = ∫ s t u ( t ′ ) d t ′ . One expects this method to be of first order in analogy to example 10, which we indeed observe in our test case (Sec. 5.1).
There exist typical choices of collocation parameters, and we list six common ones in Table 1 together with their order of convergence when applied to standard Volterra equations. Note, that for these choices of collocation parameters the order of convergence (at the grid points t i t_i t i ) is larger than the previously cited minimum order of convergence. This phenomenon is known as local superconvergence [23 Section 5.3.3] .
Numerical Experiments
Test case for experimental convergence analysis
We apply different variants of the collocation methods to the SO model as a test case with model parameters G = 1 G = 1 G = 1 , η = 0.1 \eta= 0.1 η = 0.1 , τ = ∞ \tau= \infty τ = ∞ , γ c = 0.1 \gamma_c = 0.1 γ c = 0.1 , γ ∗ = 0.05 \gamma_* = 0.05 γ ∗ = 0.05 and γ ∗ ∗ = 0.1 \gamma_{**} = 0.1 γ ∗∗ = 0.1 . That model suggests itself because its structure makes it the numerically most demanding of the examples given in § 2 , while its result are of relevance for the creep and “sudden yielding” of glass-forming colloidal suspensions [11 ] .
To obtain error estimates, we first construct an “exact solution”. To this end, we impose the strain rate
γ ˙ ( t ) = u r e f ( t ) = γ c τ V ( 1 + 0.5 sin ( 2 π τ V t ) ) , (54) \dot{\gamma}(t) = u_{\mathrm{ref}}(t) = \frac{\gamma_{\mathrm{c}}}{\tau_{\mathrm{V}}}\left(1 + 0.5\sin\left(\frac{2\pi}{\tau_{\mathrm{V}}}t\right)\right), \tag*{(54)} γ ˙ ( t ) = u ref ( t ) = τ V γ c ( 1 + 0.5 sin ( τ V 2 π t ) ) , (54)
for which we obtain σ ( t ) \sigma(t) σ ( t ) by direct quadrature with high accuracy. The collocation method is then applied to Eq. (2) with Eq. (16) , in order to reconstruct γ ˙ ( t ) \dot{\gamma}(t) γ ˙ ( t ) with the imposed σ ( t ) \sigma(t) σ ( t ) . The imposed solution does not change sign, and thus the potential nonanalyticity of the SO-model kernel is not limiting the convergence order of the methods. We have implemented the methods in the Python programming language, making heavy use of the standard libraries numpy and scipy. The methods all lead to implicit equations for the values of the desired function u ( t ) u(t) u ( t ) on the grid points; for these implicit equations a numerical solver (Newton method implementation in the scipy.optimize package) has been used with a pre-set tolerance of 10 − 14 10^{-14} 1 0 − 14 . We employed equidistant grids with N + 1 N + 1 N + 1 gridpoints:
t 0 = 0 < ⋯ < t i = i 10 τ V N < ⋯ < t N = 10 τ V , where τ V = η G (55) t_0 = 0 < \cdots< t_i = i\frac{10\tau_{\mathrm{V}}}{N} < \cdots< t_N = 10\tau_{\mathrm{V}}, \quad\text{where } \tau_{\mathrm{V}} = \frac{\eta}{G} \tag*{(55)} t 0 = 0 < ⋯ < t i = i N 10 τ V < ⋯ < t N = 10 τ V , where τ V = G η (55)
With our choice of model parameters we thus have h = 1 / N h = 1/N h = 1/ N .
In our numerical experiment, we observe the same convergence rates as for standard Volterra equations (see Figure 2 ): the Radau II–constant method (explicitly written out in Example 11 ) is of order p = 1 p = 1 p = 1 , while the Lobatto–cubic method reaches p = 6 p = 6 p = 6 . Note that in Figure 2 the minimum possible error is dictated by the accuracy of the implicit equation solver, which explains the deviations from scaling that appear around an error of 10 − 13 10^{-13} 1 0 − 13 .
Figure 2. Convergence rate of different collocation methods applied to the SO-model in the test case described in § 5.1 . The corresponding collocation parameters are listed in Table 1 . The dotted lines represent h p h^p h p with p = 1 , … , 6 p = 1, \ldots, 6 p = 1 , … , 6 respectively.
Table 1. Exemplary choices of collocation parameters (nomenclature follows the Ref. [23 ] ), together with the order of local superconvergence when applied to Volterra equations, and the quadrature weights of the corresponding quadrature rules.
A usual trade-off in numerics is between accuracy and speed. Note that in our setting, the asymptotic numerical complexity is determined by the number of grid points N N N on the time axis, and due to the convolution nature of the integral, is expected to scale as N 2 N^2 N 2 for large N N N . This is indeed seen in Figure 3 for N ≳ 1000 N \gtrsim1000 N ≳ 1000 in most methods. For smaller N N N our implementation shows linear scaling, which we attribute to a large overhead of the Python code (as opposed to the highly optimized numpy routines) in this parameter regime.
Figure 3. Computational effort of different collocation methods, as in Figure 2 with respect to the number of grid points N N N . Times are given as total run time of our Python implementation on a standard laptop (Intel® Core™ i7-1335U).
To examine which of the implemented methods is the most efficient in solving the testproblem, we show the required computational effort versus the desired accuracy in Figure 4 . Note that for a method of convergence order p p p , we expect the error to scale as ε ∼ h p \varepsilon\sim h^p ε ∼ h p ; for a discretization of the time integral on a linear grid with N N N points we expect the running time to scale as T ∼ N 2 T \sim N^2 T ∼ N 2 and with N ∼ 1 / h N \sim1/h N ∼ 1/ h as T ∼ ( 1 / ε ) 2 / p T \sim(1/\varepsilon)^{2/p} T ∼ ( 1/ ε ) 2/ p . This is indeed seen in Figure 4 for those cases where the running time exceeds roughly 10 s; for shorter runs, the pre-asymptotic implementation details dominate; for the largest 1 / ε 1/\varepsilon 1/ ε the finite precision of the floating-point implementation causes deviations.
Figure 4. Scaling of the needed computational effort to achieve a desired accuracy 1 / ϵ 1/\epsilon 1/ ϵ in solving the testproblem for the SO model with different methods described in the text.
Note that the trapezoidal–trapezoidal (TZ-TZ) direct quadrature yields a method equivalent to Lobatto–linear; indeed both show the same scaling, with differences solely attributed to the difference in our respective implementations. Figure 4 shows that the Lobatto–cubic method is extremely efficient for our test case, surpassed only by the TZ-TZ quadrature with two-stage Richardson extrapolation. Note that Richardson extrapolation is possible in this test case; in general for the models with non-analytic dependence on the sought-after solution, such as the SRR and SO model, Richardson extrapolation cannot be applied.
For applications to physical problems, it is often desirable to cover large time intervals in the solution; in this case, it is advisable to choose a method whose error is small even for relatively large step sizes, as they might be needed in a non-uniform grid. For this reason especially, the Lobatto–cubic method is an efficient choice in our setup.
Application to Creep
As an exemplary application for our numerical methods, we demonstrate the creep response of the SO model. Recall that we define creep as the response in shear rate to an applied constant stress σ ( t ) = σ 0 Θ ( t ) \sigma(t) = \sigma_0\Theta(t) σ ( t ) = σ 0 Θ ( t ) that is instantaneously switched on at t = 0 t = 0 t = 0 .
Fig. 5 shows the numerically obtained solutions for creep in the SO-model for five different applied stresses σ 0 \sigma_0 σ 0 . For reference, we fix the numerical parameters of the model as G ∞ = 1 G_\infty= 1 G ∞ = 1 , η = 0.1 \eta= 0.1 η = 0.1 , γ c = 0.25 \gamma_c = 0.25 γ c = 0.25 , γ ∗ = 0.1 \gamma_* = 0.1 γ ∗ = 0.1 , and γ ∗ ∗ = 0.133 \gamma_{**} = 0.133 γ ∗∗ = 0.133 . This results in a numerically evaluated yield stress of σ y ≈ 0.03 \sigma_y \approx0.03 σ y ≈ 0.03 ; applied stresses are given in relation to this value.
These solutions were obtained on a non-uniform time grid specifically adapted to a prior estimate of the result, in order to cover a time interval of up to five orders of magnitude. The grid balances small time steps at short times where the solution changes rapidly, with larger steps where possible, reflecting the slow evolution of the solution on exponentially large times. Specifically, we started by defining a basegrid t i b a s e = i h s t_i^{\mathrm{base}} = ih_s t i base = i h s with h s = 10 − 4 τ V h_s = 10^{-4}\tau_{\mathrm{V}} h s = 1 0 − 4 τ V for i = 0 , … , 10 i = 0,\ldots,10 i = 0 , … , 10 ; t i b a s e = 10 − 3 + ( i − 10 ) h l τ V t_i^{\mathrm{base}} = 10^{-3+(i-10)h_l}\tau_{\mathrm{V}} t i base = 1 0 − 3 + ( i − 10 ) h l τ V with h l = 0.01 h_l = 0.01 h l = 0.01 for i = 11 , … , 510 i = 11,\ldots,510 i = 11 , … , 510 ; t i b a s e = 10 2 τ V + ( i − 510 ) h e t_i^{\mathrm{base}} = 10^2\tau_{\mathrm{V}} + (i-510)h_e t i base = 1 0 2 τ V + ( i − 510 ) h e with h e = 0.495 τ V h_e = 0.495\tau_{\mathrm{V}} h e = 0.495 τ V for i = 511 , … , 20510 i = 511,\ldots,20510 i = 511 , … , 20510 which we then refined two times by adding points yielding t 4 i + k = t i b a s e + k ( t i + 1 b a s e − t i b a s e ) / 4 t_{4i+k} = t_i^{\mathrm{base}} + k(t_{i+1}^{\mathrm{base}} - t_i^{\mathrm{base}})/4 t 4 i + k = t i base + k ( t i + 1 base − t i base ) /4 , k = 0 , 1 , 2 , 3 k = 0,1,2,3 k = 0 , 1 , 2 , 3 . The smallest and largest step lengths thus were h min = t 40 − t 41 ≈ 0.58 ⋅ 10 − 7 τ V h_{\min} = t_{40} - t_{41} \approx0.58 \cdot10^{-7}\tau_{\mathrm{V}} h m i n = t 40 − t 41 ≈ 0.58 ⋅ 1 0 − 7 τ V and h max = t 2040 − t 2039 ≈ 0.57 τ V h_{\max} = t_{2040} - t_{2039} \approx0.57\tau_{\mathrm{V}} h m a x = t 2040 − t 2039 ≈ 0.57 τ V , both corresponding to the first and last step in the refined logarithmic part of the grid.
For the example, the Lobatto–quadratic method was deemed sufficiently accurate. We have checked that the solutions shown in Figure 5 are also obtained with other variants of the numerical methods described above, which suggests that they are free from numerical artifacts.
Figure 5. Creep solutions of the stressovershoot model (Example 6) with exemplary parameters as given in the text, for different values of imposed stress σ 0 \sigma_0 σ 0 in relation to the yield stress σ y \sigma_y σ y of the model. The time unit is set by the relaxation time of the Kelvin-Voigt model, τ V = G ∞ / η \tau_{\mathrm{V}} = G_{\infty}/\eta τ V = G ∞ / η . A dashed line denotes the asymptote t − 1 t^{-1} t − 1 obtained at the yield stress, σ 0 = σ y \sigma_0 = \sigma_y σ 0 = σ y .
In particular, the numerical method is able to resolve the various qualitatively different asymptotic regimes for the shear rate γ ˙ ( t ) \dot{\gamma}(t) γ ˙ ( t ) as the imposed stress σ 0 \sigma_0 σ 0 is varied across the yield stress. For σ 0 ≪ σ y \sigma_0 \ll\sigma_y σ 0 ≪ σ y , the solution is a sequence of increasingly smeared-out exponential decays, with the initial decay given by the Kelvin-Voigt relaxation rate. For σ 0 ≫ σ y \sigma_0 \gg\sigma_y σ 0 ≫ σ y on the other hand, the solutions are asymptotically constant, γ ˙ ( t ) ∼ const. \dot{\gamma}(t) \sim\operatorname{const.} γ ˙ ( t ) ∼ const. for t → ∞ t \to\infty t → ∞ ; these two cases represent solid-like and fluid-like response in physical terms.
The region around the yield stress is of particular interest. For σ 0 / σ y ≈ 1 \sigma_0/\sigma_y \approx1 σ 0 / σ y ≈ 1 , an asymptotic decay γ ˙ ( t ) ∼ 1 / t \dot{\gamma}(t) \sim1/t γ ˙ ( t ) ∼ 1/ t is observed (cf. dashed line in Figure 5 ). The model thus predicts so-called logarithmic creep (since γ ( t ) ∼ ln t \gamma(t) \sim\ln t γ ( t ) ∼ ln t for t → ∞ t \to\infty t → ∞ ), i.e., a state in-between the finitely deformed solid and a stationary flowing liquid. For σ 0 ≳ σ y \sigma_0 \gtrsim\sigma_y σ 0 ≳ σ y , a cross-over from the creep-asymptote to constant flow is observed, and due to the presence of the stress overshoot in the model (i.e., a non-monotonic, not strictly positive function G G G ), there is a relatively rapid crossover that also involves a non-monotonic time dependence of the shear rate. The rapid crossover is also termed “sudden yielding” in the physics literature and in particular associated to the sudden failure of a large class of glasses and gels (see, e.g., Ref. [31, Fig. 1]). We postpone a more in-depth discussion of the features of this model to a future publication.
Conclusion
We developed a numerical scheme to solve common integral constitutive equations (ICEs) of nonlinear rheology that are formulated as generalized Volterra-type integral equations, Eq. (2) . In particular, we transferred standard collocation methods that have been developed for classical Volterra integral equations of second kind, and demonstrate their convergence when applied to a set of physically motivated examples. While simpler direct-quadrature methods can give similar performance when combined with Richardson extrapolation (where allowed by the structure of the integral kernel), for the case of the collocation methods we are able to prove convergence to the unique solution of the problem. The proof is an adaptation of the techniques used to prove uniqueness, existence of, and convergence to the solutions in the case of ordinary Volterra integral equations.
Physically, the models we discuss here are of interest when the entire strain-rate history enters the nonlinear shear modulus, and one seeks a a solution to the stress-controlled rheology problem, i.e., to determine the strain and strain rates resulting from an applied stress. As a particular example we demonstrate the applicability of our collocation method to determine the creep-response of a glass-forming model that incorporates stress overshoots and with it, “sudden yielding” phenomena.
Although not the center of interest in the present application, our proofs extend also to certain autoconvolution Volterra integral equations (AVIEs), Eq. (21) , as long as a finite time interval is considered and a continuous solution exists. In this regard, we extend the global convergence proof of Zhang et al. [22 ] , although we do not prove superconvergence of the collocation method; we also do not prove existence of a unique continuous solution here. Some comments can be found in Appendix Appendix § 12 .
For now, the methods that we presented are all “onestep” methods. Multistep methods might be required in cases where the solutions are sought over extremely large time intervals and still might involve rapid changes at late times. The error analysis of multi-step methods is more intricate and has to be left for future work. For now, the salient assumption is that in many physical problems, the solutions will be slowly varying functions, or moderate deviations from that (as in the case of “sudden” yielding). One relevant application for improved multi-step methods might be given by large-amplitude oscillatory tests, where the imposed stress σ ( t ) ∝ sin ( ω t ) \sigma(t) \propto\sin(\omega t) σ ( t ) ∝ sin ( ω t ) imposes a time scale 1 / ω 1/\omega 1/ ω that needs to be honored by the numerical grid at all times.
With the groundwork laid down here, a future task would be to apply the methods to more intricate models as the schematic ITT-MCT model [5 ] . There, the integral constitutive equation (2) is complemented by a closure that relates the kernel G ( t , t ′ , [ γ ˙ ] ) G(t,t',[\dot{\gamma}]) G ( t , t ′ , [ γ ˙ ]) to dynamical density correlation functions ϕ ( t , t ′ , [ γ ˙ ] ) \phi(t,t',[\dot{\gamma}]) ϕ ( t , t ′ , [ γ ˙ ]) that in turn are solutions of a set of nonlinear Volterra equations of autoconvolution type for u t ′ ( t ) = ϕ ( t , t ′ , [ γ ˙ ] ) u_{t'}(t)=\phi(t,t',[\dot{\gamma}]) u t ′ ( t ) = ϕ ( t , t ′ , [ γ ˙ ]) —with additional parametric dependence on t ′ t' t ′ . While the methods developed here will still be applicable, the solution of such a one-parameter family of equations together with an appropriate choice of ( t , t ′ ) (t,t') ( t , t ′ ) grid to cover exponentially large times and time differences, is numerically extremely demanding.
Acknowledgements
We thank Jürgen Saal for continued valuable input in the preparation of the analytical background.
Data availability
Python code to implement the collocation methods discussed here, a Jupyter notebook demonstrating the calculation of creep response for the stress-overshoot model, and the data corresponding to Figure 5 are publicly available as a Zenodo repository [32 ] .
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
CRediT authorship contribution statement
Stephan Domann: Writing - Original Draft, Investigation, Software. Thomas Voigtmann: Writing - Review & Editing, Supervision, Software.
Existence and uniqueness of solutions
Nonlinear Volterra equations and the SRR model
In the following let I = [ t 0 , T ] ⊂ R I=[t_0,T]\subset\mathbb{R} I = [ t 0 , T ] ⊂ R or I = [ t 0 , ∞ ) I=[t_0,\infty) I = [ t 0 , ∞ ) be the time-interval on which we want to know the solution u u u . Define D = { ( t , s ) ∈ I × I ; s ≤ t } D=\{(t,s)\in I\times I;\ s\leq t\} D = {( t , s ) ∈ I × I ; s ≤ t } . For nonlinear Volterra equations of second kind
u ( t ) = f ( t ) + ∫ t 0 t k ( t , s , u ( s ) ) d s (56) u(t)=f(t)+\int_{t_0}^{t} k(t,s,u(s))\mathrm{d}s \tag*{(56)} u ( t ) = f ( t ) + ∫ t 0 t k ( t , s , u ( s )) d s (56)
one can find the following theorem [19 ] [Ch. 2.1.1]:
Theorem 12. Assume f f f and k k k to be continuous functions (f ∈ C ( I ) f\in C(I) f ∈ C ( I ) and k ∈ C ( D × R ) k\in C(D\times\mathbb{R}) k ∈ C ( D × R ) ). Furthermore let the kernel k k k fulfill the following Lipschitz condition: There is a function L ∈ C ( D ) L\in C(D) L ∈ C ( D ) such that
∣ k ( t , s , x ) − k ( t , s , y ) ∣ ≤ L ( t , s ) ∣ x − y ∣ for all ( t , s ) ∈ D ; x , y ∈ R (57) \lvert k(t,s,x)-k(t,s,y)\rvert\leq L(t,s)\lvert x-y\rvert\quad\text{for all }(t,s)\in D;\ x,y\in\mathbb{R} \tag*{(57)} ∣ k ( t , s , x ) − k ( t , s , y )∣ ≤ L ( t , s ) ∣ x − y ∣ for all ( t , s ) ∈ D ; x , y ∈ R (57) Then, there is exactly one solution u ∈ C ( I ) u\in C(I) u ∈ C ( I ) of the integral equation (A.1).
Example 13 (Strainrate relaxation model). In the case of the strain-rate relaxation (SRR) model (Example 5 with η ≠ 0 \eta\ne0 η = 0 ) the forcing function f f f is given by f ( t ) = σ ( t ) / η f(t)=\sigma(t)/\eta f ( t ) = σ ( t ) / η and the kernel k k k is given by
k ( t , s , y ) = G ∞ η e − t − s τ e − ( t − s ) ∣ y ∣ γ c y (58) k(t,s,y)=\frac{G_\infty}{\eta}e^{-\frac{t-s}{\tau}}e^{-\frac{(t-s)\lvert y\rvert}{\gamma_c}}y \tag*{(58)} k ( t , s , y ) = η G ∞ e − τ t − s e − γ c ( t − s ) ∣ y ∣ y (58)
All we need to do is to check the Lipschitz-type condition for k ( t , s , y ) k(t,s,y) k ( t , s , y ) :
∣ k ( t , s , x ) − k ( t , s , y ) ∣ = C ∞ η e − t − s τ ∣ e − ( t − s ) ∣ x ∣ γ c x − e − ( t − s ) ∣ y ∣ γ c y ∣ (59) |k(t,s,x)-k(t,s,y)|=\frac{C_{\infty}}{\eta}e^{-\frac{t-s}{\tau}}\left|e^{-\frac{(t-s)|x|}{\gamma_c}}x-e^{-\frac{(t-s)|y|}{\gamma_c}}y\right| \tag*{(59)} ∣ k ( t , s , x ) − k ( t , s , y ) ∣ = η C ∞ e − τ t − s e − γ c ( t − s ) ∣ x ∣ x − e − γ c ( t − s ) ∣ y ∣ y (59)
But h ( x ) = x e − α ∣ x ∣ h(x)=xe^{-\alpha|x|} h ( x ) = x e − α ∣ x ∣ , α ≥ 0 \alpha\geq0 α ≥ 0 , is Lipschitz; as can be observed from the mean-value theorem in integral form
∣ h ( x ) − h ( y ) ∣ ≤ ∫ x y ∣ h ′ ( r ) ∣ d r ≤ ∣ x − y ∣ sup r ∈ R \ { 0 } ∣ h ′ ( r ) ∣ = ∣ x − y ∣ , (60) |h(x)-h(y)|\leq\int_x^y|h'(r)|dr\leq|x-y|\sup_{r\in\mathbb{R}\backslash\{0\}}|h'(r)|=|x-y|, \tag*{(60)} ∣ h ( x ) − h ( y ) ∣ ≤ ∫ x y ∣ h ′ ( r ) ∣ d r ≤ ∣ x − y ∣ r ∈ R \ { 0 } sup ∣ h ′ ( r ) ∣ = ∣ x − y ∣ , (60)
i.e., a Lipschitz condition with L = 1 L=1 L = 1 . Inserting this result into Eq. (A.4) yields that the SRR model has a unique solution.
General functional equations and the stressovershoot model
Due to the more abstract viewpoint of functional equations the existence and uniqueness theorem we cite here is more intricate and requires substantially more effort to check. Before citing the theorem itself some introduction of notation and explanatory comments are in order.
We want to find a solution of the functional equation given by
u ( t ) = H [ u ] ( t ) , t ∈ [ t 0 , T max ) ⊂ R (61) u(t)=H[u](t),\quad t\in[t_0,T_{\max})\subset\mathbb{R} \tag*{(61)} u ( t ) = H [ u ] ( t ) , t ∈ [ t 0 , T m a x ) ⊂ R (61)
where H H H is some nonlinear operator and where T max T_{\max} T m a x is the a priori unknown maximum time of existence. This is a fixpoint equation. For our purposes we are interested in continuous solutions, so we need to H H H to be an operator that maps a certain set of continuous functions to a continuous function, e.g. ϕ ↦ H [ ϕ ] ∈ C ( [ t 0 , T ] ; R n ) \phi\mapsto H[\phi]\in C([t_0,T];\mathbb{R}^n) ϕ ↦ H [ ϕ ] ∈ C ([ t 0 , T ] ; R n ) for ϕ ∈ A ⊂ C ( [ t 0 , T ] ; R n ) \phi\in A\subset C([t_0,T];\mathbb{R}^n) ϕ ∈ A ⊂ C ([ t 0 , T ] ; R n ) for all T < T max T<T_{\max} T < T m a x . To have a notion of continuity of the nonlinear operator H : A ⊂ C ( [ t 0 , T ] ; R n ) → C ( [ t 0 , T ] ; R n ) H:A\subset C([t_0,T];\mathbb{R}^n)\to C([t_0,T];\mathbb{R}^n) H : A ⊂ C ([ t 0 , T ] ; R n ) → C ([ t 0 , T ] ; R n ) we need to fix a norm on C ( [ t 0 , T ] ; R n ) C([t_0,T];\mathbb{R}^n) C ([ t 0 , T ] ; R n ) . As norm we use the supremum norm
∥ ϕ ∥ ∞ = sup t ∈ [ t 0 , T ] ∣ ϕ ( t ) ∣ (62) \|\phi\|_{\infty}=\sup_{t\in[t_0,T]}|\phi(t)| \tag*{(62)} ∥ ϕ ∥ ∞ = t ∈ [ t 0 , T ] sup ∣ ϕ ( t ) ∣ (62)
where ∣ x ∣ |x| ∣ x ∣ , x ∈ R n x\in\mathbb{R}^n x ∈ R n denotes the norm in R n \mathbb{R}^n R n . To ensure the pointwise continuity of the operator H H H will be the role of condition 1) in the following theorem.
The solution u u u defines curves ( t , u ( t ) ) t ∈ [ t 0 , T ] (t,u(t))_{t\in[t_0,T]} ( t , u ( t ) ) t ∈ [ t 0 , T ] in R ≥ t 0 × R n \mathbb{R}_{\geq t_0}\times\mathbb{R}^n R ≥ t 0 × R n , T < T max T<T_{\max} T < T m a x . To capture cases where H H H is not defined for functions ϕ \phi ϕ that exceed a certain bound or where H H H is not defined for some time-instances t ≥ T H t\geq T_H t ≥ T H , we have to restrict ourselves to functions ϕ \phi ϕ where the induced curves ( t , ϕ ( t ) ) t ∈ [ t 0 , T ] (t,\phi(t))_{t\in[t_0,T]} ( t , ϕ ( t ) ) t ∈ [ t 0 , T ] lie within some open, connected subset M ⊆ R ≥ t 0 × R n M\subseteq\mathbb{R}_{\geq t_0}\times\mathbb{R}^n M ⊆ R ≥ t 0 × R n . This M M M has to be adjusted according to the operator H H H at hand. Thus, the function space where we search for a solution is given by
A = C ξ , M ( [ t 0 , T ] ; R n ) = { ϕ ∈ C ( [ t 0 , T ] ; R n ) ∣ ϕ ( t 0 ) = ξ , ( t , ϕ ( t ) ) ∈ M for t ∈ [ t 0 , T ] } . (63) A=C_{\xi,M}([t_0,T];\mathbb{R}^n)=\left\{\phi\in C([t_0,T];\mathbb{R}^n)\mid\phi(t_0)=\xi,\ (t,\phi(t))\in M\ \text{for }t\in[t_0,T]\right\}. \tag*{(63)} A = C ξ , M ([ t 0 , T ] ; R n ) = { ϕ ∈ C ([ t 0 , T ] ; R n ) ∣ ϕ ( t 0 ) = ξ , ( t , ϕ ( t )) ∈ M for t ∈ [ t 0 , T ] } . (63)
The choice of M M M naturally defines an upper bound for the length of the time-intervals [ t 0 , T ] [t_0,T] [ t 0 , T ] given by
T ∞ M = sup { T ∈ R > t 0 | C ξ , M ( [ t 0 , T ] , R n ) ≠ ∅ } (64) T_\infty^M = \sup\left\{ T \in\mathbb{R}_{>t_0} \,\middle|\, C_{\xi,M}([t_0,T],\mathbb{R}^n) \ne\varnothing\right\} \tag*{(64)} T ∞ M = sup { T ∈ R > t 0 ∣ C ξ , M ([ t 0 , T ] , R n ) = ∅ } (64)
Note, T ∞ M = ∞ T_\infty^M = \infty T ∞ M = ∞ is allowed.
We resort to the Banach fixpoint theorem to establish uniqueness of the solution of the functional equation (A.6). This will be the origin of condition 4) in the theorem, where we in principle demand that H H H defines a contraction on some chosen subset of C ξ , M ( [ 0 , T ] , R n ) C_{\xi,M}([0,T],\mathbb{R}^n) C ξ , M ([ 0 , T ] , R n ) .
Since we cannot expect H H H to be a contraction on the whole space C ξ , M ( [ t 0 , T ] ; R n ) C_{\xi,M}([t_0,T];\mathbb{R}^n) C ξ , M ([ t 0 , T ] ; R n ) we restrict this condition to functions that are equal up to some time-instance T e q < T T_{\mathrm{eq}} < T T eq < T which is sufficiently close to T T T . The full unique solution is then constructed by starting to construct a solution on a small time-interval [ t 0 , T 1 ] [t_0,T_1] [ t 0 , T 1 ] and then subsequently constructing continuations on the time-intervalls [ T i , T i + 1 ] [T_i,T_{i+1}] [ T i , T i + 1 ] . This is the origin of the assumptions 4a) and 4b) in the theorem below. A possibly infinite sequence of such steps can assert uniqueness on the right-open interval [ t 0 , T max ] [t_0,T_{\max}] [ t 0 , T m a x ] . In the case T max < T ∞ M T_{\max} < T_\infty^M T m a x < T ∞ M , one can show with some additional hypotheses that then the solution u u u has to either diverge (∣ u ( t ) ∣ → ∞ |u(t)| \to\infty ∣ u ( t ) ∣ → ∞ as t ↑ T max t \uparrow T_{\max} t ↑ T m a x ) or approach the boundary of M M M (e.g. ( T max , u ( t ↑ T max ) ) ∈ ∂ M (T_{\max},u(t \uparrow T_{\max})) \in\partial M ( T m a x , u ( t ↑ T m a x )) ∈ ∂ M ). The latter case may appear if the curve ( t , u ( t ) ) (t,u(t)) ( t , u ( t )) , t ∈ [ t 0 , T max ) t \in[t_0,T_{\max}) t ∈ [ t 0 , T m a x ) , takes values such that H [ u ] ( T max ) H[u](T_{\max}) H [ u ] ( T m a x ) is not defined anymore.
With these qualifications, the following theorem given by Gripenberg et al. [21 Ch. 12 Theorem 2.1 and Corollary 2.2] applies (with minor change in notation):
Theorem 14. Assume that the following conditions are satisfied:
For each T ∈ [ t 0 , T ∞ M ) T \in[t_0,T_\infty^M) T ∈ [ t 0 , T ∞ M ) , H H H maps C ξ , M ( [ t 0 , T ] ; R n ) C_{\xi,M}([t_0,T];\mathbb{R}^n) C ξ , M ([ t 0 , T ] ; R n ) into C ( [ t 0 , T ] ; R n ) C([t_0,T];\mathbb{R}^n) C ([ t 0 , T ] ; R n ) and for each t ∈ [ t 0 , T ] t \in[t_0,T] t ∈ [ t 0 , T ] the mapping ϕ ↦ H [ ϕ ] ( t ) \phi\mapsto H[\phi](t) ϕ ↦ H [ ϕ ] ( t ) , ϕ ∈ C ξ , M ( [ t 0 , T ] ; R n ) \phi\in C_{\xi,M}([t_0,T];\mathbb{R}^n) ϕ ∈ C ξ , M ([ t 0 , T ] ; R n ) , is continous with respect to the supremum norm ∥ ⋅ ∥ ∞ \|\cdot\|_\infty ∥ ⋅ ∥ ∞ .
H H H is causal, e.g. if t ~ ∈ [ t 0 , T ) \tilde{t} \in[t_0,T) t ~ ∈ [ t 0 , T ) and if ϕ ( t ) = ψ ( t ) \phi(t)=\psi(t) ϕ ( t ) = ψ ( t ) for t ≤ t ~ t \le\tilde{t} t ≤ t ~ then H [ ϕ ] ( t ) = H [ ψ ] ( t ) H[\phi](t)=H[\psi](t) H [ ϕ ] ( t ) = H [ ψ ] ( t ) .
H [ ξ ] ( t 0 ) = ξ H[\xi](t_0)=\xi H [ ξ ] ( t 0 ) = ξ (initial value compatibility)
For each compact connected subset K ⊂ M K \subset M K ⊂ M containing the starting point ( t 0 , ξ ) (t_0,\xi) ( t 0 , ξ ) there exists constants δ K > 0 \delta_K > 0 δ K > 0 and γ K ∈ ( 0 , 1 ) \gamma_K \in(0,1) γ K ∈ ( 0 , 1 ) such that the following impliciation holds: if for some functions ϕ , ψ ∈ C ξ , K ( [ t 0 , T ] ; R n ) \phi,\psi\in C_{\xi,K}([t_0,T];\mathbb{R}^n) ϕ , ψ ∈ C ξ , K ([ t 0 , T ] ; R n ) we have
(a) ψ ( s ) = ϕ ( s ) \psi(s)=\phi(s) ψ ( s ) = ϕ ( s ) for s ∈ [ t 0 , T e q ] s \in[t_0,T_{\mathrm{eq}}] s ∈ [ t 0 , T eq ] with T e q = max ( t 0 , T − δ K ) T_{\mathrm{eq}}=\max(t_0,T-\delta_K) T eq = max ( t 0 , T − δ K )
(b) ∣ ψ ( s ) − ϕ ( s ) ∣ ≤ δ K |\psi(s)-\phi(s)| \le\delta_K ∣ ψ ( s ) − ϕ ( s ) ∣ ≤ δ K for s ∈ [ T e q , T ] s \in[T_{\mathrm{eq}},T] s ∈ [ T eq , T ]
then there holds
sup s ∈ [ T e q , T ] ∣ H [ ϕ ] ( s ) − H [ ψ ] ( s ) ∣ ≤ γ K sup s ∈ [ T e q , T ] ∣ ψ ( s ) − ϕ ( s ) ∣ (65) \sup_{s\in[T_{\mathrm{eq}},T]} |H[\phi](s)-H[\psi](s)| \le\gamma_K \sup_{s\in[T_{\mathrm{eq}},T]} |\psi(s)-\phi(s)| \tag*{(65)} s ∈ [ T eq , T ] sup ∣ H [ ϕ ] ( s ) − H [ ψ ] ( s ) ∣ ≤ γ K s ∈ [ T eq , T ] sup ∣ ψ ( s ) − ϕ ( s ) ∣ (65)
Then there exists a unique noncontinuable solution u u u of Eq. (A.6) fulfilling the initial condition enforced by H H H . The maximal interval of existence is open to the right. e.g. it is given by [ t 0 , T max ) [t_0,T_{\max}) [ t 0 , T m a x ) for some T max > t 0 T_{\max}>t_0 T m a x > t 0 . Moreover, if γ K ≤ 1 2 \gamma_K\leq\frac{1}{2} γ K ≤ 2 1 and if the implication 4) is true without the assumption b), then the curve ( t , u ( t ) ) (t,u(t)) ( t , u ( t )) leaves every compact subset of M M M as t ↑ T max t\uparrow T_{\max} t ↑ T m a x .
Example 15 (Stressovershoot model). The stress-overshoot (SO) model can be written as
η u ( t ) = σ ( t ) − ∫ t 0 t G ∞ ( 1 − ( z ( t , s , [ u ] ) γ ∗ ) 2 n ) e − ( z ( t , s , [ u ] ) γ ∗ ∗ ) 2 n e − t − s τ e − ( t − s ) ∣ u ( s ) ∣ γ c u ( s ) d s (66) \eta u(t)=\sigma(t)-\int_{t_0}^{t}G_{\infty}\left(1-\left(\frac{z(t,s,[u])}{\gamma_*}\right)^{2n}\right)e^{-\left(\frac{z(t,s,[u])}{\gamma_{**}}\right)^{2n}}e^{-\frac{t-s}{\tau}}e^{-\frac{(t-s)|u(s)|}{\gamma_c}}u(s)\,\mathrm{d}s \tag*{(66)} η u ( t ) = σ ( t ) − ∫ t 0 t G ∞ ( 1 − ( γ ∗ z ( t , s , [ u ]) ) 2 n ) e − ( γ ∗∗ z ( t , s , [ u ]) ) 2 n e − τ t − s e − γ c ( t − s ) ∣ u ( s ) ∣ u ( s ) d s (66)
G ( t , s , [ u ] ) = G ∞ ( 1 − ( z ( t , s , [ u ] ) γ ∗ ) 2 n ) e − ( z ( t , s , [ u ] ) γ ∗ ∗ ) 2 n e − t − s τ e − ( t − s ) ∣ u ( s ) ∣ γ c (67) G(t,s,[u])=G_{\infty}\left(1-\left(\frac{z(t,s,[u])}{\gamma_*}\right)^{2n}\right)e^{-\left(\frac{z(t,s,[u])}{\gamma_{**}}\right)^{2n}}e^{-\frac{t-s}{\tau}}e^{-\frac{(t-s)|u(s)|}{\gamma_c}} \tag*{(67)} G ( t , s , [ u ]) = G ∞ ( 1 − ( γ ∗ z ( t , s , [ u ]) ) 2 n ) e − ( γ ∗∗ z ( t , s , [ u ]) ) 2 n e − τ t − s e − γ c ( t − s ) ∣ u ( s ) ∣ (67)
z ( t , t ′ , [ u ] ) = ∫ t ′ t u ( s ) d s (68) z(t,t',[u])=\int_{t'}^{t}u(s)\,\mathrm{d}s \tag*{(68)} z ( t , t ′ , [ u ]) = ∫ t ′ t u ( s ) d s (68)
Assuming η > 0 \eta>0 η > 0 we can check the conditions of theorem 14 for
H [ u ] ( t ) = σ ( t ) η − ∫ t 0 t G ∞ η ( 1 − ( z ( t , s , [ u ] ) γ ∗ ) 2 n ) e − ( z ( t , s , [ u ] ) γ ∗ ∗ ) 2 n e − t − s τ e − ( t − s ) ∣ u ( s ) ∣ γ c u ( s ) d s (69) H[u](t)=\frac{\sigma(t)}{\eta}-\int_{t_0}^{t}\frac{G_{\infty}}{\eta}\left(1-\left(\frac{z(t,s,[u])}{\gamma_*}\right)^{2n}\right)e^{-\left(\frac{z(t,s,[u])}{\gamma_{**}}\right)^{2n}}e^{-\frac{t-s}{\tau}}e^{-\frac{(t-s)|u(s)|}{\gamma_c}}u(s)\,\mathrm{d}s \tag*{(69)} H [ u ] ( t ) = η σ ( t ) − ∫ t 0 t η G ∞ ( 1 − ( γ ∗ z ( t , s , [ u ]) ) 2 n ) e − ( γ ∗∗ z ( t , s , [ u ]) ) 2 n e − τ t − s e − γ c ( t − s ) ∣ u ( s ) ∣ u ( s ) d s (69)
The initial condition ξ \xi ξ fixed by H H H is ξ = u ( t 0 ) = σ ( t 0 ) / η \xi=u(t_0)=\sigma(t_0)/\eta ξ = u ( t 0 ) = σ ( t 0 ) / η and one can directly see that H H H is causal. Functions in the domain of H H H do not need to be bounded in any fashion, so we can choose M = R ≥ t 0 × R M=\mathbb{R}_{\geq t_0}\times\mathbb{R} M = R ≥ t 0 × R , such that T ∞ M = ∞ T_{\infty}^{M}=\infty T ∞ M = ∞ . Lets start by checking condition 1.) of the theorem and fix a T ∈ [ t 0 , ∞ ) T\in[t_0,\infty) T ∈ [ t 0 , ∞ ) . As long as σ ( t ) \sigma(t) σ ( t ) is continuous, H H H will map continuous functions ϕ ∈ C ξ , M ( [ t 0 , T ] ) \phi\in C_{\xi,M}([t_0,T]) ϕ ∈ C ξ , M ([ t 0 , T ]) to continuous functions on [ t 0 , T ] [t_0,T] [ t 0 , T ] . For any fixed t ∈ [ t 0 , T ] t\in[t_0,T] t ∈ [ t 0 , T ] the continuity of the mapping ϕ ↦ H [ ϕ ] ( t ) \phi\mapsto H[\phi](t) ϕ ↦ H [ ϕ ] ( t ) , ϕ ∈ C ξ , M ( [ t 0 , T ] ) \phi\in C_{\xi,M}([t_0,T]) ϕ ∈ C ξ , M ([ t 0 , T ]) , is given since this mapping can be seen as a product and chain of continuous operations.
This leaves checking whether condition 4.) holds: Let K ⊂ M K\subset M K ⊂ M be compact and connected with ( t 0 , ξ ) ∈ K (t_0,\xi)\in K ( t 0 , ξ ) ∈ K . For later usage we define
α K = max ( t , x ) ∈ K ∣ x ∣ (70) \alpha_K=\max_{(t,x)\in K}|x| \tag*{(70)} α K = ( t , x ) ∈ K max ∣ x ∣ (70)
In the following we will determine δ K > 0 \delta_K>0 δ K > 0 such that γ K = 1 3 \gamma_K=\frac{1}{3} γ K = 3 1 is possible. Take two functions ϕ , ψ ∈ C ξ , K ( [ 0 , T ] ) \phi,\psi\in C_{\xi,K}([0,T]) ϕ , ψ ∈ C ξ , K ([ 0 , T ]) that fulfill condition 4a). For t ∈ [ T e q , T ] t\in[T_{\mathrm{eq}},T] t ∈ [ T eq , T ] we have
∣ H [ ϕ ] ( t ) − H [ ψ ] ( t ) ∣ = ∣ ∫ t 0 t G ( t , s , [ ϕ ] ) ϕ ( s ) − G ( t , s , [ ψ ] ) ψ ( s ) d s ∣ (71) \left|H[\phi](t)-H[\psi](t)\right|=\left|\int_{t_0}^{t}G(t,s,[\phi])\phi(s)-G(t,s,[\psi])\psi(s)\,ds\right| \tag*{(71)} ∣ H [ ϕ ] ( t ) − H [ ψ ] ( t ) ∣ = ∫ t 0 t G ( t , s , [ ϕ ]) ϕ ( s ) − G ( t , s , [ ψ ]) ψ ( s ) d s (71)
≤ ∫ t 0 t ∣ G ( t , s , [ ϕ ] ) ∣ ∣ ϕ ( s ) − ψ ( s ) ∣ + ∣ ψ ( s ) ∣ ∣ G ( t , s , [ ϕ ] ) − G ( t , s , [ ψ ] ) ∣ d s (72) \leq\int_{t_0}^{t}\left|G(t,s,[\phi])\right|\left|\phi(s)-\psi(s)\right|+\left|\psi(s)\right|\left|G(t,s,[\phi])-G(t,s,[\psi])\right|\,ds \tag*{(72)} ≤ ∫ t 0 t ∣ G ( t , s , [ ϕ ]) ∣ ∣ ϕ ( s ) − ψ ( s ) ∣ + ∣ ψ ( s ) ∣ ∣ G ( t , s , [ ϕ ]) − G ( t , s , [ ψ ]) ∣ d s (72)
≤ ( T − T e q ) max t , s , ϕ ∣ G ( t , s , [ ϕ ] ) ∣ ∥ ϕ − ψ ∥ ∞ + ∫ t 0 t α K ∣ G ( t , s , [ ϕ ] ) − G ( t , s , [ ψ ] ) ∣ d s (73) \begin{aligned}
\leq(T-T_{eq})\max_{t,s,\phi}\left|G(t,s,[\phi])\right|\|\phi-\psi\|_{\infty} \\
&+\int_{t_0}^{t}\alpha_K\left|G(t,s,[\phi])-G(t,s,[\psi])\right|\,ds \tag*{(73)}
\end{aligned} ≤ ( T − T e q ) t , s , ϕ max ∣ G ( t , s , [ ϕ ]) ∣ ∥ ϕ − ψ ∥ ∞ + ∫ t 0 t α K ∣ G ( t , s , [ ϕ ]) − G ( t , s , [ ψ ]) ∣ d s (73)
Note, that
max t , s , ϕ ∣ G ( t , s , [ ϕ ] ) ∣ = max { ∣ G ( t , s , [ ϕ ] ) ∣ ∣ s ≤ t ; t , s ∈ [ t 0 , T ] and ϕ ∈ C ξ , K ( [ 0 , T ] ) } < ∞ (74) \max_{t,s,\phi}\left|G(t,s,[\phi])\right|=\max\left\{\left|G(t,s,[\phi])\right|\mid s\leq t;\ t,s\in[t_0,T]\text{ and }\phi\in C_{\xi,K}([0,T])\right\}<\infty \tag*{(74)} t , s , ϕ max ∣ G ( t , s , [ ϕ ]) ∣ = max { ∣ G ( t , s , [ ϕ ]) ∣ ∣ s ≤ t ; t , s ∈ [ t 0 , T ] and ϕ ∈ C ξ , K ([ 0 , T ]) } < ∞ (74)
is finite. Next, we have to deal with the second term. Observe that G ( t , s , [ ϕ ] ) G(t,s,[\phi]) G ( t , s , [ ϕ ]) is of the form
G ( t , s , [ ϕ ] ) = G ( t , s , ϕ ( s ) , z ( t , s , [ ϕ ] ) ) (75) G(t,s,[\phi])=G(t,s,\phi(s),z(t,s,[\phi])) \tag*{(75)} G ( t , s , [ ϕ ]) = G ( t , s , ϕ ( s ) , z ( t , s , [ ϕ ])) (75)
where G ( t , s , x , z ) G(t,s,x,z) G ( t , s , x , z ) is Lipschitz-continuous both in x x x and z z z :
∣ G ( t , s , x , z ) − G ( t , s , x ′ , z ′ ) ∣ ≤ L 1 ( t , s ) ∣ x − x ′ ∣ + L 2 ( t , s ) ∣ z − z ′ ∣ (76) \left|G(t,s,x,z)-G(t,s,x',z')\right|\leq L_1(t,s)\left|x-x'\right|+L_2(t,s)\left|z-z'\right| \tag*{(76)} ∣ G ( t , s , x , z ) − G ( t , s , x ′ , z ′ ) ∣ ≤ L 1 ( t , s ) ∣ x − x ′ ∣ + L 2 ( t , s ) ∣ z − z ′ ∣ (76)
L 1 ( t , s ) = G η e − t − s τ sup x ∈ R n e − ( t − s ) ∣ x ∣ γ = G η e − t − s τ (77) L_1(t,s)=\frac{G}{\eta}e^{-\frac{t-s}{\tau}}\sup_{x\in\mathbb{R}^n}e^{-\frac{(t-s)|x|}{\gamma}}=\frac{G}{\eta}e^{-\frac{t-s}{\tau}} \tag*{(77)} L 1 ( t , s ) = η G e − τ t − s x ∈ R n sup e − γ ( t − s ) ∣ x ∣ = η G e − τ t − s (77)
L 2 ( t , s ) = G η e − t − s τ sup z ∈ R n ∣ ( 1 − ( z / γ ∗ ) 2 n ) e − ( z / γ ∗ ∗ ) 2 n ∣ . (78) L_2(t,s)=\frac{G}{\eta}e^{-\frac{t-s}{\tau}}\sup_{z\in\mathbb{R}^n}\left|(1-(z/\gamma_*)^{2n})e^{-(z/\gamma_{**})^{2n}}\right|. \tag*{(78)} L 2 ( t , s ) = η G e − τ t − s z ∈ R n sup ( 1 − ( z / γ ∗ ) 2 n ) e − ( z / γ ∗∗ ) 2 n . (78)
Thus, when we can estimate ∣ z − z ′ ∣ = ∣ z ( t , s , [ ϕ ] ) − z ( t , s , [ ψ ] ) ∣ \left|z-z'\right|=\left|z(t,s,[\phi])-z(t,s,[\psi])\right| ∣ z − z ′ ∣ = ∣ z ( t , s , [ ϕ ]) − z ( t , s , [ ψ ]) ∣ with respect to δ K \delta_K δ K or T − T e q T-T_{eq} T − T e q , we have everything we need to finish the proof.
∣ z ( t , s , [ ϕ ] ) − z ( t , s , [ ψ ] ) ∣ = ∣ ∫ s t ϕ ( t ′ ) − ψ ( t ′ ) d t ′ ∣ (79) \left|z(t,s,[\phi])-z(t,s,[\psi])\right|=\left|\int_s^t\phi(t')-\psi(t')\,dt'\right| \tag*{(79)} ∣ z ( t , s , [ ϕ ]) − z ( t , s , [ ψ ]) ∣ = ∫ s t ϕ ( t ′ ) − ψ ( t ′ ) d t ′ (79)
≤ ( T − T e q ) ∥ ϕ − ψ ∥ ∞ (80) \leq(T-T_{eq})\|\phi-\psi\|_{\infty} \tag*{(80)} ≤ ( T − T e q ) ∥ ϕ − ψ ∥ ∞ (80)
Thus, we have (defining L i = max s < t L i ( t , s ) L_i=\max_{s<t}L_i(t,s) L i = max s < t L i ( t , s ) )
∫ t 0 t ∣ G ( t , s , [ ϕ ] ) − G ( t , s , [ ψ ] ) ∣ d s ≤ ∫ t 0 t ( L 1 ( t , s ) ∣ ϕ ( s ) − ψ ( s ) ∣ + L 2 ( t , s ) ∣ z ( t , s , [ ϕ ] ) − z ( t , s , [ ψ ] ) ∣ ) d s (81) \int_{t_0}^{t}\left|G(t,s,[\phi])-G(t,s,[\psi])\right|\,ds\leq\int_{t_0}^{t}\left(L_1(t,s)\left|\phi(s)-\psi(s)\right|+L_2(t,s)\left|z(t,s,[\phi])-z(t,s,[\psi])\right|\right)\,ds \tag*{(81)} ∫ t 0 t ∣ G ( t , s , [ ϕ ]) − G ( t , s , [ ψ ]) ∣ d s ≤ ∫ t 0 t ( L 1 ( t , s ) ∣ ϕ ( s ) − ψ ( s ) ∣ + L 2 ( t , s ) ∣ z ( t , s , [ ϕ ]) − z ( t , s , [ ψ ]) ∣ ) d s (81)
≤ L 1 ( T − T e q ) ∥ ϕ − ψ ∥ ∞ + L 2 ( T − t 0 ) ( T − T e q ) ∥ ϕ − ψ ∥ ∞ (82) \leq L_1(T-T_{eq})\|\phi-\psi\|_{\infty}+L_2(T-t_0)(T-T_{eq})\|\phi-\psi\|_{\infty} \tag*{(82)} ≤ L 1 ( T − T e q ) ∥ ϕ − ψ ∥ ∞ + L 2 ( T − t 0 ) ( T − T e q ) ∥ ϕ − ψ ∥ ∞ (82)
Going back to equation A.18 we have shown
∣ H [ ϕ ] ( t ) − H [ ψ ] ( t ) ∣ ≤ ( T − T e q ) ( max t , s , ϕ ∣ G ( t , s , [ ϕ ] ) ∣ + L 1 α K + L 2 α K ( T − t 0 ) ) ∥ ϕ − ψ ∥ ∞ (83) |H[\phi](t)-H[\psi](t)| \le(T-T_{eq})\left(\max_{t,s,\phi}|G(t,s,[\phi])|+L_1\alpha_K+L_2\alpha_K(T-t_0)\right)\|\phi-\psi\|_\infty \tag*{(83)} ∣ H [ ϕ ] ( t ) − H [ ψ ] ( t ) ∣ ≤ ( T − T e q ) ( t , s , ϕ max ∣ G ( t , s , [ ϕ ]) ∣ + L 1 α K + L 2 α K ( T − t 0 ) ) ∥ ϕ − ψ ∥ ∞ (83)
≤ δ K ( max t , s , ϕ ∣ G ( t , s , [ ϕ ] ) ∣ + L 1 α K + L 2 α K ( T ∞ K − t 0 ) ) ∥ ϕ − ψ ∥ ∞ (84) \le\delta_K\left(\max_{t,s,\phi}|G(t,s,[\phi])|+L_1\alpha_K+L_2\alpha_K(T^K_\infty-t_0)\right)\|\phi-\psi\|_\infty \tag*{(84)} ≤ δ K ( t , s , ϕ max ∣ G ( t , s , [ ϕ ]) ∣ + L 1 α K + L 2 α K ( T ∞ K − t 0 ) ) ∥ ϕ − ψ ∥ ∞ (84)
Thus, we have ∣ H [ ϕ ] ( t ) − H [ ψ ] ( t ) ∣ ≤ 1 3 ∥ ϕ − ψ ∥ ∞ |H[\phi](t)-H[\psi](t)|\le\frac{1}{3}\|\phi-\psi\|_\infty ∣ H [ ϕ ] ( t ) − H [ ψ ] ( t ) ∣ ≤ 3 1 ∥ ϕ − ψ ∥ ∞ with
δ K = 1 3 ( max t , s , ϕ ∣ G ( t , s , [ ϕ ] ) ∣ + L 1 α K + L 2 α K ( T ∞ K − t 0 ) ) (85) \delta_K=\frac{1}{3\left(\max_{t,s,\phi}|G(t,s,[\phi])|+L_1\alpha_K+L_2\alpha_K(T^K_\infty-t_0)\right)} \tag*{(85)} δ K = 3 ( max t , s , ϕ ∣ G ( t , s , [ ϕ ]) ∣ + L 1 α K + L 2 α K ( T ∞ K − t 0 ) ) 1 (85)
Thus, the stressovershoot model fulfills all conditions of theorem 14 and has a unique solution. Furthermore it fulfills the additional conditions in the theorem. Therefore, if the solution u ( t ) u(t) u ( t ) stays bounded, we have T max = T ∞ M = ∞ T_{\max}=T^M_\infty=\infty T m a x = T ∞ M = ∞ .
In the essential steps in the proof of the example the specific form of G ( t , s , x , z ) G(t,s,x,z) G ( t , s , x , z ) was not of importance. We only needed the Lipschitz-type condition for G ( t , s , x , z ) G(t,s,x,z) G ( t , s , x , z ) in both components and the specific estimate for ∣ z ( t , s , [ ϕ ] ) − z ( t , s , [ ψ ] ) ∣ |z(t,s,[\phi])-z(t,s,[\psi])| ∣ z ( t , s , [ ϕ ]) − z ( t , s , [ ψ ]) ∣ . Thus, we also have proven:
Proposition 16. Let σ ( t ) \sigma(t) σ ( t ) be continuous. Assume
G ( t , s , [ ϕ ] ) = G ( t , s , ϕ ( s ) , z ( t , s , [ ϕ ] ) ) (86) G(t,s,[\phi])=G(t,s,\phi(s),z(t,s,[\phi])) \tag*{(86)} G ( t , s , [ ϕ ]) = G ( t , s , ϕ ( s ) , z ( t , s , [ ϕ ])) (86) z ( t , s , [ ϕ ] ) = ∫ s t k ( t , t ′ , s , ϕ ( t ′ ) ) d t ′ (87) z(t,s,[\phi])=\int_s^t k(t,t',s,\phi(t'))\,\mathrm{d}t' \tag*{(87)} z ( t , s , [ ϕ ]) = ∫ s t k ( t , t ′ , s , ϕ ( t ′ )) d t ′ (87) to fulfill the following Lipschitz-type conditions: There are continuous functions L 1 , L 2 ∈ C ( D ) L_1,L_2\in C(D) L 1 , L 2 ∈ C ( D ) , D = { ( t , s ) ∈ [ t 0 , ∞ ) × [ t 0 , ∞ ) ∣ s ≤ t } D=\{(t,s)\in[t_0,\infty)\times[t_0,\infty)\mid s\le t\} D = {( t , s ) ∈ [ t 0 , ∞ ) × [ t 0 , ∞ ) ∣ s ≤ t } , such that
∣ G ( t , s , x , z ) − G ( t , s , x ′ , z ′ ) ∣ ≤ L 1 ( t , s ) ∣ x − x ′ ∣ + L 2 ( t , s ) ∣ z − z ′ ∣ (88) |G(t,s,x,z)-G(t,s,x',z')|\le L_1(t,s)|x-x'|+L_2(t,s)|z-z'| \tag*{(88)} ∣ G ( t , s , x , z ) − G ( t , s , x ′ , z ′ ) ∣ ≤ L 1 ( t , s ) ∣ x − x ′ ∣ + L 2 ( t , s ) ∣ z − z ′ ∣ (88) and there is a continuous function L k ∈ C ( D ~ ) L_k\in C(\widetilde{D}) L k ∈ C ( D ) , D ~ = { ( t , t ′ , s ) ∈ [ t 0 , ∞ ) 3 ∣ s ≤ t ′ ≤ t } \widetilde{D}=\{(t,t',s)\in[t_0,\infty)^3\mid s\le t'\le t\} D = {( t , t ′ , s ) ∈ [ t 0 , ∞ ) 3 ∣ s ≤ t ′ ≤ t } , such that
∣ k ( t , t ′ , s , x ) − k ( t , t ′ , s , x ′ ) ∣ ≤ L k ( t , t ′ , s ) ∣ x − x ′ ∣ . (89) |k(t,t',s,x)-k(t,t',s,x')|\le L_k(t,t',s)|x-x'|. \tag*{(89)} ∣ k ( t , t ′ , s , x ) − k ( t , t ′ , s , x ′ ) ∣ ≤ L k ( t , t ′ , s ) ∣ x − x ′ ∣. (89) Assume furthermore, that G ( t , s , x , z ) G(t,s,x,z) G ( t , s , x , z ) is bounded for ( t , s ) ∈ D (t,s)\in D ( t , s ) ∈ D and x , z ∈ R n x,z\in\mathbb{R}^n x , z ∈ R n . Then the functional equation
η u ( t ) = σ ( t ) − ∫ t 0 t G ( t , s , ϕ ( s ) , z ( t , s , [ ϕ ] ) ) ϕ ( s ) d s , η > 0 (90) \eta u(t)=\sigma(t)-\int_{t_0}^t G(t,s,\phi(s),z(t,s,[\phi]))\phi(s)\,\mathrm{d}s,\qquad\eta>0 \tag*{(90)} η u ( t ) = σ ( t ) − ∫ t 0 t G ( t , s , ϕ ( s ) , z ( t , s , [ ϕ ])) ϕ ( s ) d s , η > 0 (90) has a unique (noncontinuable) solution u ( t ) u(t) u ( t ) , t ∈ [ t 0 , T max ) t\in[t_0,T_{\max}) t ∈ [ t 0 , T m a x ) , and if u ( t ) u(t) u ( t ) stays bounded we have T max = ∞ T_{\max}=\infty T m a x = ∞ .
Convergence proof for collocation methods applied to Volterra equations with Fréchet-differentiable functional kernels
Theorem 17. Assume that the evolutionary integral equation
u ( t ) = g ( t ) + ∫ t 0 t K ( t , s , [ u ] ) d s , t ∈ [ t 0 , T ] (91) u(t) = g(t) + \int_{t_0}^{t} K(t,s,[u])\,\mathrm{d}s,\qquad t \in[t_0,T] \tag*{(91)} u ( t ) = g ( t ) + ∫ t 0 t K ( t , s , [ u ]) d s , t ∈ [ t 0 , T ] (91) has a unique m m m -times continuously differentiable solution y ∈ C ( m ) ( [ t 0 , T ] ) y \in C^{(m)}([t_0,T]) y ∈ C ( m ) ([ t 0 , T ]) , m ≥ 1 m \ge1 m ≥ 1 , where g ( t ) g(t) g ( t ) is the forcing function and the kernel K ( t , s , [ u ] ) K(t,s,[u]) K ( t , s , [ u ]) is a (nonlinear) causal functional, e.g. K ( t , s , [ u ] ) = K ( t , s , u ∣ [ t 0 , t ] ) K(t,s,[u]) = K(t,s,u|_{[t_0,t]}) K ( t , s , [ u ]) = K ( t , s , u ∣ [ t 0 , t ] ) . Assume K ( t , s , [ u ] ) K(t,s,[u]) K ( t , s , [ u ]) is continuously Fréchet-differentiable (in u u u ) in C ( [ t 0 , T ] ) C([t_0,T]) C ([ t 0 , T ]) for all s , t ∈ [ t 0 , T ] s,t \in[t_0,T] s , t ∈ [ t 0 , T ] , s ≤ t s \le t s ≤ t , and assume that the Fréchet-derivative D u K D_uK D u K is of the form
D u K ( t , s , [ u ] ) δ u = k 1 ( t , s , [ u ] ) δ u ( s ) + ∫ t 0 t k 2 ( t , s , τ , [ u ] ) δ u ( τ ) d τ , δ u ∈ C ( [ t 0 , T ] ) (92) D_uK(t,s,[u])\delta u = k_1(t,s,[u])\delta u(s) + \int_{t_0}^{t} k_2(t,s,\tau,[u])\delta u(\tau)\,\mathrm{d}\tau,\qquad\delta u \in C([t_0,T]) \tag*{(92)} D u K ( t , s , [ u ]) δ u = k 1 ( t , s , [ u ]) δ u ( s ) + ∫ t 0 t k 2 ( t , s , τ , [ u ]) δ u ( τ ) d τ , δ u ∈ C ([ t 0 , T ]) (92) where k 1 k_1 k 1 and k 2 k_2 k 2 are bounded in the following sense:
sup ∥ k 1 ∥ L 1 : = sup u ∈ C ( [ t 0 , T ] ) , t ∈ [ t 0 , T ] ∫ t 0 t ∣ k 1 ( t , s , [ u ] ) ∣ d s < ∞ (93) \sup\lVert k_1\rVert_{L^1} := \sup_{\substack{u\in C([t_0,T]),\,t\in[t_0,T]}} \int_{t_0}^{t} \lvert k_1(t,s,[u])\rvert\,\mathrm{d}s < \infty \tag*{(93)} sup ∥ k 1 ∥ L 1 := u ∈ C ([ t 0 , T ]) , t ∈ [ t 0 , T ] sup ∫ t 0 t ∣ k 1 ( t , s , [ u ])∣ d s < ∞ (93) sup ∥ k 2 ∥ L 1 : = sup u ∈ C ( [ t 0 , T ] ) , t ∈ [ t 0 , T ] ∫ t 0 t ∫ t 0 t ∣ k 2 ( t , s , τ , [ u ] ) ∣ d τ d s < ∞ . (94) \sup\lVert k_2\rVert_{L^1} := \sup_{\substack{u\in C([t_0,T]),\,t\in[t_0,T]}} \int_{t_0}^{t}\int_{t_0}^{t} \lvert k_2(t,s,\tau,[u])\rvert\,\mathrm{d}\tau\,\mathrm{d}s < \infty. \tag*{(94)} sup ∥ k 2 ∥ L 1 := u ∈ C ([ t 0 , T ]) , t ∈ [ t 0 , T ] sup ∫ t 0 t ∫ t 0 t ∣ k 2 ( t , s , τ , [ u ])∣ d τ d s < ∞. (94) Let I h = { t 0 , … , t N } ⊂ [ t 0 , T ] I_h = \{t_0,\ldots,t_N\} \subset[t_0,T] I h = { t 0 , … , t N } ⊂ [ t 0 , T ] , h = max i ( t i + 1 − t i ) h = \max_i(t_{i+1}-t_i) h = max i ( t i + 1 − t i ) , be a mesh and let
0 = c 1 < ⋯ < c m = 1 (95) 0 = c_1 < \cdots< c_m = 1 \tag*{(95)} 0 = c 1 < ⋯ < c m = 1 (95) be the set of collocation parameters. Then there exists a critical steplength h c h_c h c such that
∥ y ( t ) − u h ( t ) ∥ ∞ ≤ C ∥ y ( m ) ∥ ∞ h m (96) \lVert y(t)-u_h(t)\rVert_\infty\le C\lVert y^{(m)}\rVert_\infty h^m \tag*{(96)} ∥ y ( t ) − u h ( t ) ∥ ∞ ≤ C ∥ y ( m ) ∥ ∞ h m (96) holds for the collocation solution u h u_h u h on any mesh with mesh diameter h < h c h<h_c h < h c . If u ↦ K ( t , s , [ u ] ) u\mapsto K(t,s,[u]) u ↦ K ( t , s , [ u ]) and u ↦ D u K ( t , s , [ u ] ) u\mapsto D_uK(t,s,[u]) u ↦ D u K ( t , s , [ u ]) can be extended to piecewise continuous and bounded functions u ∈ P C b ( [ 0 , T ] ) u\in PC_b([0,T]) u ∈ P C b ([ 0 , T ]) (with L ∞ L^\infty L ∞ -norm) such that sup ∥ k 1 ∥ L 1 < ∞ \sup\lVert k_1\rVert_{L^1}<\infty sup ∥ k 1 ∥ L 1 < ∞ and sup ∥ k 2 ∥ L 1 < ∞ \sup\lVert k_2\rVert_{L^1}<\infty sup ∥ k 2 ∥ L 1 < ∞ also holds for u ∈ P C b ( [ t 0 , T ] ) u\in PC_b([t_0,T]) u ∈ P C b ([ t 0 , T ]) , then the allowed range of collocation parameters can be extended to
0 ≤ c 1 < ⋯ < c m ≤ 1. (97) 0 \le c_1 < \cdots< c_m \le1. \tag*{(97)} 0 ≤ c 1 < ⋯ < c m ≤ 1. (97)
Remark 18. Consider that the exact solution y y y is known to lie in some open convex subset F ⊂ C ( [ t 0 , T ] ) F \subset C([t_0,T]) F ⊂ C ([ t 0 , T ]) . This is for example the case if y ( t ) > 0 y(t)>0 y ( t ) > 0 or if ∣ y ( t ) ∣ < c \lvert y(t)\rvert<c ∣ y ( t )∣ < c , c ∈ R > 0 c\in\mathbb{R}_{>0} c ∈ R > 0 , for all t ∈ [ t 0 , T ] t\in[t_0,T] t ∈ [ t 0 , T ] . Then, the full space C ( [ t 0 , T ] ) C([t_0,T]) C ([ t 0 , T ]) in the theorem above can be replaced by F F F , e.g. K ( t , s , [ u ] ) K(t,s,[u]) K ( t , s , [ u ]) only needs to be Fréchet-differentiable in F F F and the bounds sup ∥ k 1 ∥ L 1 \sup\lVert k_1\rVert_{L^1} sup ∥ k 1 ∥ L 1 and sup ∥ k 2 ∥ L 1 \sup\lVert k_2\rVert_{L^1} sup ∥ k 2 ∥ L 1 only need to be finite for u ∈ F u\in F u ∈ F . Note, the convexity of the subset F F F is needed in the application of the vector-valued mean-value theorem in the proof and thus is a crucial condition.
The derived order h m h^m h m is a global convergence order. For collocation methods, one often observes the phenomenon of local superconvergence, i.e., that on the gridpoints the error converges faster than guaranteed by the minimal convergence rate. We leave as an open question under which conditions local superconvergence can be guaranteed. Before we give the proof of the theorem, consider the following three applications:
Example 19 (Nonlinear Volterra equations of second kind). In this case we have
K ( t , s , [ u ] ) = K ( t , s , u ( s ) ) (98) K(t,s,[u]) = K(t,s,u(s)) \tag*{(98)} K ( t , s , [ u ]) = K ( t , s , u ( s )) (98)
⇒ D u K ( t , s , [ u ] ) δ u = ∂ y K ( t , s , y ) ∣ y = u ( s ) δ u ( s ) (99) \Rightarrow D_uK(t,s,[u])\delta u = \left.\partial_yK(t,s,y)\right|_{y=u(s)}\delta u(s) \tag*{(99)} ⇒ D u K ( t , s , [ u ]) δ u = ∂ y K ( t , s , y ) ∣ y = u ( s ) δ u ( s ) (99)
Thus, according to the theorem convergence can be guaranteed if
sup ∥ k 1 ∥ L 1 = sup u ∈ C ( [ t 0 , T ] ) , t ∈ [ t 0 , T ] ∫ t 0 t ∣ ∂ y K ( t , s , u ( s ) ) ∣ d s < ∞ (100) \sup\lVert k_1\rVert_{L^1} =
\sup_{u\in C([t_0,T]),\,t\in[t_0,T]}
\int_{t_0}^{t}\left|\partial_yK(t,s,u(s))\right|\,\mathrm{d}s < \infty \tag*{(100)} sup ∥ k 1 ∥ L 1 = u ∈ C ([ t 0 , T ]) , t ∈ [ t 0 , T ] sup ∫ t 0 t ∣ ∂ y K ( t , s , u ( s )) ∣ d s < ∞ (100)
This is for example the case if the partial derivative ∂ y K ( t , s , y ) \partial_yK(t,s,y) ∂ y K ( t , s , y ) is bounded for all y ∈ R y\in\mathbb{R} y ∈ R and t , s ∈ [ t 0 , T ] t,s\in[t_0,T] t , s ∈ [ t 0 , T ] with s < t s<t s < t , see also [28 Ch. 2.3.1] . In this case also local superconvergence has been shown [28 Ch. 2.3.2] .
Example 20 (Stressovershoot model). In this case the functional kernel is given by
K ( t , s , [ u ] ) = G ( t , s , [ u ] ) u ( s ) (101) K(t,s,[u]) = G(t,s,[u])u(s) \tag*{(101)} K ( t , s , [ u ]) = G ( t , s , [ u ]) u ( s ) (101)
G ( t , s , [ u ] ) = G ∞ ( 1 − ( ∫ s t u ( τ ) d τ γ ∗ ) 2 ) e − ( ∫ s t u ( τ ) d τ / γ ∗ ∗ ) 2 e − ∣ u ( s ) ∣ γ c (102) G(t,s,[u]) = G_\infty\left(1-\left(\frac{\int_s^t u(\tau)\,\mathrm{d}\tau}{\gamma_*}\right)^2\right)e^{-\left(\int_s^t u(\tau)\,\mathrm{d}\tau/\gamma_{**}\right)^2}e^{-\frac{|u(s)|}{\gamma_c}} \tag*{(102)} G ( t , s , [ u ]) = G ∞ 1 − ( γ ∗ ∫ s t u ( τ ) d τ ) 2 e − ( ∫ s t u ( τ ) d τ / γ ∗∗ ) 2 e − γ c ∣ u ( s ) ∣ (102)
Under the assumption, that u u u does not change sign we can calculate the first Fréchet-derivative:
D u K ( t , s , [ u ] ) δ u = G ( t , s , [ u ] ) δ u ( s ) + u ( s ) D u G ( t , s , [ u ] ) δ u (103) D_uK(t,s,[u])\delta u = G(t,s,[u])\delta u(s)+u(s)D_uG(t,s,[u])\delta u \tag*{(103)} D u K ( t , s , [ u ]) δ u = G ( t , s , [ u ]) δ u ( s ) + u ( s ) D u G ( t , s , [ u ]) δ u (103)
D u G ( t , s , [ u ] ) δ u = G ∞ − 2 γ ∗ ( ∫ s t u ( τ ) d τ γ ∗ ) e − ( ∫ s t u ( τ ) d τ / γ ∗ ∗ ) 2 e − ∣ u ( s ) ∣ γ c ∫ s t δ u ( τ ) d τ (104) D_uG(t,s,[u])\delta u =
G_\infty\frac{-2}{\gamma_*}\left(\frac{\int_s^t u(\tau)\,\mathrm{d}\tau}{\gamma_*}\right)e^{-\left(\int_s^t u(\tau)\,\mathrm{d}\tau/\gamma_{**}\right)^2}e^{-\frac{|u(s)|}{\gamma_c}}\int_s^t\delta u(\tau)\,\mathrm{d}\tau \tag*{(104)} D u G ( t , s , [ u ]) δ u = G ∞ γ ∗ − 2 ( γ ∗ ∫ s t u ( τ ) d τ ) e − ( ∫ s t u ( τ ) d τ / γ ∗∗ ) 2 e − γ c ∣ u ( s ) ∣ ∫ s t δ u ( τ ) d τ (104)
+ G ∞ ( 1 − ( ∫ s t u ( τ ) d τ γ ∗ ) 2 ) ( − 2 γ ∗ ∗ 2 ∫ s t u ( τ ) d τ ) e − ( ∫ s t u ( τ ) d τ / γ ∗ ∗ ) 2 e − ∣ u ( s ) ∣ γ c ∫ s t δ u ( τ ) d τ + G ∞ ( 1 − ( ∫ s t u ( τ ) d τ γ ∗ ) 2 ) e − ( ∫ s t u ( τ ) d τ / γ ∗ ∗ ) 2 − 1 γ c e − ∣ u ( s ) ∣ γ c δ u ( s ) (105) \begin{aligned}
&+G_\infty\left(1-\left(\frac{\int_s^t u(\tau)\,\mathrm{d}\tau}{\gamma_*}\right)^2\right)\left(\frac{-2}{\gamma_{**}^2}\int_s^t u(\tau)\,\mathrm{d}\tau\right)e^{-\left(\int_s^t u(\tau)\,\mathrm{d}\tau/\gamma_{**}\right)^2}e^{-\frac{|u(s)|}{\gamma_c}}\int_s^t\delta u(\tau)\,\mathrm{d}\tau\\
&+G_\infty\left(1-\left(\frac{\int_s^t u(\tau)\,\mathrm{d}\tau}{\gamma_*}\right)^2\right)e^{-\left(\int_s^t u(\tau)\,\mathrm{d}\tau/\gamma_{**}\right)^2}\frac{-1}{\gamma_c}e^{-\frac{|u(s)|}{\gamma_c}}\delta u(s) \tag*{(105)}
\end{aligned} + G ∞ 1 − ( γ ∗ ∫ s t u ( τ ) d τ ) 2 ( γ ∗∗ 2 − 2 ∫ s t u ( τ ) d τ ) e − ( ∫ s t u ( τ ) d τ / γ ∗∗ ) 2 e − γ c ∣ u ( s ) ∣ ∫ s t δ u ( τ ) d τ + G ∞ 1 − ( γ ∗ ∫ s t u ( τ ) d τ ) 2 e − ( ∫ s t u ( τ ) d τ / γ ∗∗ ) 2 γ c − 1 e − γ c ∣ u ( s ) ∣ δ u ( s ) (105)
Thus, the functions k 1 k_1 k 1 and k 2 k_2 k 2 from the theorem are given by
k 1 ( t , s , [ u ] ) = G ( t , s , [ u ] ) − G ( t , s , [ u ] ) u ( s ) γ c (106) k_1(t,s,[u]) = G(t,s,[u]) - \frac{G(t,s,[u])u(s)}{\gamma_c} \tag*{(106)} k 1 ( t , s , [ u ]) = G ( t , s , [ u ]) − γ c G ( t , s , [ u ]) u ( s ) (106)
k 2 ( t , s , τ , [ u ] ) = − 2 γ ∗ ∗ 2 ( ∫ s t u ( τ ) d τ ) G ( t , s , [ u ] ) u ( s ) χ [ s , t ] ( τ ) − 2 G ∞ γ ∗ ( ∫ s t u ( τ ) d τ γ ∗ ) e − ( ∫ s t u ( τ ) d τ / γ ∗ ∗ ) 2 e − ∣ u ( s ) ∣ γ c u ( s ) χ [ s , t ] ( τ ) (107) \begin{aligned}
k_2(t,s,\tau,[u]) ={}& -\frac{2}{\gamma_{**}^{2}}\left(\int_s^t u(\tau)\,\mathrm{d}\tau\right)G(t,s,[u])u(s)\chi_{[s,t]}(\tau)\\
&-\frac{2G_\infty}{\gamma_*}\left(\frac{\int_s^t u(\tau)\,\mathrm{d}\tau}{\gamma_*}\right)e^{-\left(\int_s^t u(\tau)\,\mathrm{d}\tau/\gamma_{**}\right)^2}e^{-\frac{|u(s)|}{\gamma_c}}u(s)\chi_{[s,t]}(\tau) \tag*{(107)}
\end{aligned} k 2 ( t , s , τ , [ u ]) = − γ ∗∗ 2 2 ( ∫ s t u ( τ ) d τ ) G ( t , s , [ u ]) u ( s ) χ [ s , t ] ( τ ) − γ ∗ 2 G ∞ ( γ ∗ ∫ s t u ( τ ) d τ ) e − ( ∫ s t u ( τ ) d τ / γ ∗∗ ) 2 e − γ c ∣ u ( s ) ∣ u ( s ) χ [ s , t ] ( τ ) (107)
where χ [ s , t ] ( τ ) \chi_{[s,t]}(\tau) χ [ s , t ] ( τ ) is the characteristic function, e.g. χ [ t , s ] ( τ ) = 1 \chi_{[t,s]}(\tau)=1 χ [ t , s ] ( τ ) = 1 if τ ∈ [ s , t ] \tau\in[s,t] τ ∈ [ s , t ] and zero elsewhere. Note, due to the exponential terms in G ( t , s , [ u ] ) G(t,s,[u]) G ( t , s , [ u ]) both k 1 k_1 k 1 and k 2 k_2 k 2 are globally bounded and thus the collocation methods converge for the stressovershoot model. Since G ( t , s , [ u ] ) G(t,s,[u]) G ( t , s , [ u ]) can be defined also for u ∈ P C b u\in PC_b u ∈ P C b (and since G ( t , ⋅ , [ u ] ) ∈ L ∞ ( [ t 0 , t ] ) G(t,\cdot,[u])\in L^\infty([t_0,t]) G ( t , ⋅ , [ u ]) ∈ L ∞ ([ t 0 , t ]) ), collocation with discontinuous splines is possible.
The additional assumptions placed on the first Frèchet derivative D u K D_uK D u K to have a specific form may seem strong. However, note that D u K D_uK D u K is a linear operator acting on the space of continuous functions C ( [ t 0 , T ] ) C([t_0,T]) C ([ t 0 , T ]) (or on P C b ( [ t 0 , T ] ) PC_b([t_0,T]) P C b ([ t 0 , T ]) respectively). According to the theorem of Riesz and Markow (see for example [33 Theorem 12.6] ) any linear operator A A A acting on continuous functions δ u ∈ C ( [ t 0 , T ] ) \delta u\in C([t_0,T]) δ u ∈ C ([ t 0 , T ]) can be represented by a measure ν \nu ν such that
A δ u = ∫ δ u d ν , δ u ∈ C ( [ t 0 , T ] ) . (108) A\delta u=\int\delta u\,\mathrm{d}\nu,\qquad\delta u\in C([t_0,T]). \tag*{(108)} A δ u = ∫ δ u d ν , δ u ∈ C ([ t 0 , T ]) . (108)
According to Lebesgue’s decomposition theorem (see e.g. [33 Corollary 4.17] ), this measure can be decomposed into three parts: a part ν c \nu_c ν c which can be described by a density k k k , a discrete part ν d \nu_d ν d (point evaluations/Dirac-deltas δ ( t i − t ) \delta(t_i-t) δ ( t i − t ) ) and a singular continuous part ν s c \nu_{sc} ν sc . Neglecting the singular continuous part we set ν s c = 0 \nu_{sc}=0 ν sc = 0 , and we have
A δ u = ∫ δ u d ν c + ∫ δ u d ν d , δ u ∈ C ( [ t 0 , T ] ) , = ∫ [ t 0 , T ] k ( t ) δ u ( t ) d t + ∑ i ∈ N c i δ u ( t i ) , t i ∈ [ t 0 , T ] . (109) \begin{aligned}
A\delta u &= \int\delta u\,\mathrm{d}\nu_c+\int\delta u\,\mathrm{d}\nu_d,\qquad\delta u\in C([t_0,T]),\\
&= \int_{[t_0,T]} k(t)\delta u(t)\,\mathrm{d}t+\sum_{i\in\mathbb{N}}c_i\delta u(t_i),\qquad t_i\in[t_0,T]. \tag*{(109)}
\end{aligned} A δ u = ∫ δ u d ν c + ∫ δ u d ν d , δ u ∈ C ([ t 0 , T ]) , = ∫ [ t 0 , T ] k ( t ) δ u ( t ) d t + i ∈ N ∑ c i δ u ( t i ) , t i ∈ [ t 0 , T ] . (109)
Arguing that in physically motivated equations the point evaluations in D u K ( t , s , [ u ] ) δ u D_uK(t,s,[u])\delta u D u K ( t , s , [ u ]) δ u should solely depend on t t t and s s s (and maybe t 0 t_0 t 0 ) and not on some other arbitrary time-instance, suggests to only consider point evaluations at these points6 . This implies to consider
D u K ( t , s , [ u ] ) δ u = ∫ t 0 t k 2 ( t , s , τ , [ u ] ) δ u ( τ ) d τ + k 1 ( t , s , [ u ] ) δ u ( s ) (110) D_uK(t,s,[u])\delta u=\int_{t_0}^{t}k_2(t,s,\tau,[u])\delta u(\tau)\,\mathrm{d}\tau+k_1(t,s,[u])\,\delta u(s) \tag*{(110)} D u K ( t , s , [ u ]) δ u = ∫ t 0 t k 2 ( t , s , τ , [ u ]) δ u ( τ ) d τ + k 1 ( t , s , [ u ]) δ u ( s ) (110)
Note, terms of the form k ~ 1 ( t , s , [ u ] ) δ u ( t 0 + ( t − s ) ) \widetilde{k}_1(t,s,[u])\,\delta u(t_0+(t-s)) k 1 ( t , s , [ u ]) δ u ( t 0 + ( t − s )) can be absorbed by the k 1 k_1 k 1 -term using
∫ t 0 t k ~ 1 ( t , s , [ u ] ) δ u ( t 0 + ( t − s ) ) d s = ∫ t 0 t k ~ 1 ( t , t 0 + ( t − s ) , [ u ] ) δ u ( s ) d s . \int_{t_0}^{t}\widetilde{k}_1(t,s,[u])\delta u(t_0+(t-s))\,\mathrm{d}s=\int_{t_0}^{t}\widetilde{k}_1(t,t_0+(t-s),[u])\delta u(s)\,\mathrm{d}s. ∫ t 0 t k 1 ( t , s , [ u ]) δ u ( t 0 + ( t − s )) d s = ∫ t 0 t k 1 ( t , t 0 + ( t − s ) , [ u ]) δ u ( s ) d s .
Example 21 (Generalized autoconvolution equation). The generalized autoconvolution Volterra integral equation (AVIE) is given by
u ( t ) = g ( t ) + ∫ 0 t M ( t , s ) u ( t − s ) u ( s ) d s (111) u(t)=g(t)+\int_{0}^{t}M(t,s)u(t-s)u(s)\,\mathrm{d}s \tag*{(111)} u ( t ) = g ( t ) + ∫ 0 t M ( t , s ) u ( t − s ) u ( s ) d s (111)
It has been shown by Zhang et al. that this equation has a unique solution y ∈ C m ( [ 0 , T ] ) y\in C^m([0,T]) y ∈ C m ([ 0 , T ]) if g ∈ C m ( [ 0 , T ] ) g\in C^m([0,T]) g ∈ C m ([ 0 , T ]) and M ∈ C m ( D ) M\in C^m(D) M ∈ C m ( D ) [22 ] . Furthermore, these authors also gave global convergence and local superconvergence results for collocation methods applied to the generalized AVIE.
Our theorem provides an alternative proof for the global convergence of collocation methods applied to the generalized AVIE (but not for the local superconvergence): The functional kernel of the AVIE is given by
K ( t , s , [ u ] ) = M ( t , s ) u ( t − s ) u ( s ) (112) K(t,s,[u])=M(t,s)u(t-s)u(s) \tag*{(112)} K ( t , s , [ u ]) = M ( t , s ) u ( t − s ) u ( s ) (112)
Calculating the Fréchet-derivative D u K D_uK D u K yields
D u K ( t , s , [ u ] ) δ u = M ( t , s ) u ( s ) δ u ( t − s ) + M ( t , s ) u ( t − s ) δ u ( s ) , δ u ∈ C ( [ 0 , T ] ) (113) D_uK(t,s,[u])\delta u=M(t,s)u(s)\delta u(t-s)+M(t,s)u(t-s)\delta u(s),\quad\delta u\in C([0,T]) \tag*{(113)} D u K ( t , s , [ u ]) δ u = M ( t , s ) u ( s ) δ u ( t − s ) + M ( t , s ) u ( t − s ) δ u ( s ) , δ u ∈ C ([ 0 , T ]) (113)
Thus, the corresponding k 1 ( t , s , [ u ] ) k_1(t,s,[u]) k 1 ( t , s , [ u ]) from the theorem is given by
k 1 ( t , s , [ u ] ) = M ( t , t − s ) u ( t − s ) + M ( t , s ) u ( t − s ) (114) k_1(t,s,[u])=M(t,t-s)u(t-s)+M(t,s)u(t-s) \tag*{(114)} k 1 ( t , s , [ u ]) = M ( t , t − s ) u ( t − s ) + M ( t , s ) u ( t − s ) (114)
Since we know a solution y ( t ) y(t) y ( t ) exists in C m ( [ 0 , T ] ) C^m([0,T]) C m ([ 0 , T ]) , ∣ y ( t ) ∣ |y(t)| ∣ y ( t ) ∣ assumes a maximum y max y_{\max} y m a x somewhere. Thus, there is a constant c > y max c>y_{\max} c > y m a x . From now on only consider functions u ∈ F c = { f ∈ C ( [ 0 , T ] ) ; ∣ f ( t ) ∣ < c } u\in F_c=\{f\in C([0,T]); |f(t)|<c\} u ∈ F c = { f ∈ C ([ 0 , T ]) ; ∣ f ( t ) ∣ < c } . Then y ∈ F c y\in F_c y ∈ F c and
sup ∥ k 1 ∥ L 1 = sup u ∈ F c , t ∈ [ t 0 , T ] ∫ t 0 t ∣ k 1 ( t , s , [ u ] ) ∣ d s (115) \sup\lVert k_1\rVert_{L^1}=\sup_{u\in F_c,t\in[t_0,T]}\int_{t_0}^{t}|k_1(t,s,[u])|\,\mathrm{d}s \tag*{(115)} sup ∥ k 1 ∥ L 1 = u ∈ F c , t ∈ [ t 0 , T ] sup ∫ t 0 t ∣ k 1 ( t , s , [ u ]) ∣ d s (115)
≤ c T max 0 ≤ s ≤ t ≤ T ∣ M ( t , s ) ∣ < ∞ (116) \leq cT\max_{0\leq s\leq t\leq T}|M(t,s)|<\infty \tag*{(116)} ≤ c T 0 ≤ s ≤ t ≤ T max ∣ M ( t , s ) ∣ < ∞ (116)
Applying theorem 17 and the remark after it then yields the same global convergence order as also has been found by Zhang et al. [22 ] .
Proof of Theorem 17. We adapt the proof given by Brunner for nonlinear Volterra integral equations [28 Ch. 2.3.1] , generalizing the underlying idea to the case of functional kernels K ( t , s , [ u ] ) K(t,s,[u]) K ( t , s , [ u ]) . The underlying idea is to linearize the kernel using a mean-value theorem and then proceed as in the linear case by deriving a discrete Gronwall inequality (compare theorem 24 below) which will establish an upper bound of order O ( h m ) O(h^m) O ( h m ) in combination with some kind of boundedness properties of the Fréchet-derivative D u K D_uK D u K .
Assume the problem has a unique solution y ( t ) y(t) y ( t ) with y ∈ C m [ 0 , T ] y \in C^{m}[0,T] y ∈ C m [ 0 , T ] . Let t 0 < ⋯ < t N = T t_0 < \cdots< t_N = T t 0 < ⋯ < t N = T be the gridpoints and denote the collocation points by t n , k = t n + c k h n t_{n,k}=t_n+c_kh_n t n , k = t n + c k h n with the collocation parameters 0 ≤ c 1 < ⋯ < c m ≤ 1 0 \le c_1 < \cdots< c_m \le1 0 ≤ c 1 < ⋯ < c m ≤ 1 . If the functional K ( t , s , [ u ] ) K(t,s,[u]) K ( t , s , [ u ]) cannot be extended to piecewise continuous and bounded functions u ∈ P C b ( [ t 0 , T ] ) u \in PC_b([t_0,T]) u ∈ P C b ([ t 0 , T ]) , set c 1 = 0 c_1=0 c 1 = 0 and c m = 1 c_m=1 c m = 1 . We denote the collocation solution by
u ( t ) = ∑ n = 0 N − 1 χ ( t n , t n + 1 ] ( t ) ∑ k = 1 m u ( t n , k ) L k ( t − t n h n ) (117) u(t)=\sum_{n=0}^{N-1}\chi_{(t_n,t_{n+1}]}(t)\sum_{k=1}^{m}u(t_{n,k})L_k\left(\frac{t-t_n}{h_n}\right) \tag*{(117)} u ( t ) = n = 0 ∑ N − 1 χ ( t n , t n + 1 ] ( t ) k = 1 ∑ m u ( t n , k ) L k ( h n t − t n ) (117)
where L k L_k L k are the Lagrange-Polynomials corresponding to the abscissas c k c_k c k on the intervall [ 0 , 1 ] [0,1] [ 0 , 1 ] . Note, in the case c 1 = 0 c_1=0 c 1 = 0 and c m = 1 c_m=1 c m = 1 the collocation solution u ( t ) u(t) u ( t ) is continuous by construction such that K ( t , s , [ u ] ) K(t,s,[u]) K ( t , s , [ u ]) is well-defined even if K ( t , s , [ u ] ) K(t,s,[u]) K ( t , s , [ u ]) cannot be extended to u ∈ P C b u \in PC_b u ∈ P C b .
It will be convenient to represent u ( t ) u(t) u ( t ) as a sum of some kind of base functions. We choose7
L ~ i , j ( t ) = χ ( t i , t i + 1 ] ( t ) L j ( t − t i h i ) , i = 0 , … , N − 1 ; j = 1 , … , m (118) \widetilde{L}_{i,j}(t)=\chi_{(t_i,t_{i+1}]}(t)L_j\left(\frac{t-t_i}{h_i}\right),\qquad i=0,\ldots,N-1;\ j=1,\ldots,m \tag*{(118)} L i , j ( t ) = χ ( t i , t i + 1 ] ( t ) L j ( h i t − t i ) , i = 0 , … , N − 1 ; j = 1 , … , m (118)
Thus we have for t ∈ [ t 0 , t N ] t \in[t_0,t_N] t ∈ [ t 0 , t N ]
u ( t ) = ∑ n = 0 N − 1 ∑ j = 1 m U n , j L ~ n , j ( t ) (119) u(t)=\sum_{n=0}^{N-1}\sum_{j=1}^{m}U_{n,j}\widetilde{L}_{n,j}(t) \tag*{(119)} u ( t ) = n = 0 ∑ N − 1 j = 1 ∑ m U n , j L n , j ( t ) (119)
Our first task will be to show that the collocation equation actually has a (unique) solution. Then we will show in several steps that this solution indeed converges globally(!) to the exact solution y ( t ) y(t) y ( t ) as the grid gets more refined (h → 0 h \to0 h → 0 ).
Step 1: The collocation equation has a solution for h ≤ h c h \le h_c h ≤ h c
The collocation equations are given by
U n , k = u ( t n , k ) = g ( t n , k ) + ∫ t 0 t n , k K ( t n , k , s , [ u ] ) d s (120) U_{n,k}=u(t_{n,k})=g(t_{n,k})+\int_{t_0}^{t_{n,k}}K(t_{n,k},s,[u])\,ds \tag*{(120)} U n , k = u ( t n , k ) = g ( t n , k ) + ∫ t 0 t n , k K ( t n , k , s , [ u ]) d s (120)
Assume we already calculated the values U i , j U_{i,j} U i , j for i = 0 , … , n − 1 i=0,\ldots,n-1 i = 0 , … , n − 1 and j = j= j = 1 , … , m 1,\ldots,m 1 , … , m . The right hand side of the collocation equation then defines a function
f : R m → R m (121) f:\mathbb{R}^{m}\to\mathbb{R}^{m} \tag*{(121)} f : R m → R m (121)
( U n , 1 ⋮ U n , m ) ↦ ( g ( t n , 1 ) + ∫ t 0 t n , 1 K ( t n , m , s , [ u ] ) d s ⋮ g ( t n , m ) + ∫ t 0 t n , m K ( t n , m , s , [ u ] ) d s ) (122) \begin{pmatrix}
U_{n,1}\\
\vdots\\
U_{n,m}
\end{pmatrix}
\mapsto
\begin{pmatrix}
g(t_{n,1})+\int_{t_0}^{t_{n,1}}K(t_{n,m},s,[u])\,\mathrm{d}s\\
\vdots\\
g(t_{n,m})+\int_{t_0}^{t_{n,m}}K(t_{n,m},s,[u])\,\mathrm{d}s \tag*{(122)}
\end{pmatrix} U n , 1 ⋮ U n , m ↦ g ( t n , 1 ) + ∫ t 0 t n , 1 K ( t n , m , s , [ u ]) d s ⋮ g ( t n , m ) + ∫ t 0 t n , m K ( t n , m , s , [ u ]) d s (122)
u ( t ) = ∑ j = 1 m U n , j L ~ n , j ( t ) + ∑ i = 0 n − 1 ∑ j = 1 m U i , j L ~ i , j ( t ) (123) u(t)=\sum_{j=1}^{m}U_{n,j}\widetilde{L}_{n,j}(t)+\sum_{i=0}^{n-1}\sum_{j=1}^{m}U_{i,j}\widetilde{L}_{i,j}(t) \tag*{(123)} u ( t ) = j = 1 ∑ m U n , j L n , j ( t ) + i = 0 ∑ n − 1 j = 1 ∑ m U i , j L i , j ( t ) (123)
Note, the fixpoints of f f f are the solutions to the collocation equations. We will now use Banachs fixpoint theorem to show that f f f indeed has a unique fixpoint as long as h n h_n h n is smaller than some critical steplength h c h_c h c . Let U , V ∈ R m U,V\in\mathbb{R}^{m} U , V ∈ R m be two vectors. Then
( f ( U ) − f ( V ) ) k = ∫ t 0 t n , k ( K ( t n , k , s , [ u 1 ] ) − K ( t n , k , s , [ u 2 ] ) ) d s (124) \left(f(U)-f(V)\right)_k=\int_{t_0}^{t_{n,k}}\left(K(t_{n,k},s,[u_1])-K(t_{n,k},s,[u_2])\right)\,\mathrm{d}s \tag*{(124)} ( f ( U ) − f ( V ) ) k = ∫ t 0 t n , k ( K ( t n , k , s , [ u 1 ]) − K ( t n , k , s , [ u 2 ]) ) d s (124)
with
u 1 ( t ) = ∑ j = 1 m U j L ~ n , j ( t ) + u p a s t ( t ) (125) u_1(t)=\sum_{j=1}^{m}U_j\widetilde{L}_{n,j}(t)+u_{\mathrm{past}}(t) \tag*{(125)} u 1 ( t ) = j = 1 ∑ m U j L n , j ( t ) + u past ( t ) (125)
u 2 ( t ) = ∑ j = 1 m V j L ~ n , j ( t ) + u p a s t ( t ) (126) u_2(t)=\sum_{j=1}^{m}V_j\widetilde{L}_{n,j}(t)+u_{\mathrm{past}}(t) \tag*{(126)} u 2 ( t ) = j = 1 ∑ m V j L n , j ( t ) + u past ( t ) (126)
We can linearize the equation by employing the vector-valued mean-value theorem (see theorem 22 below)
K ( t n , k , s , [ u 1 ] ) − K ( t n , k , s , [ u 2 ] ) = ∫ 0 1 D u K ( t n , k , s , [ u 2 + α ( u 1 − u 2 ) ] ) ( u 1 − u 2 ) d α (127) K(t_{n,k},s,[u_1])-K(t_{n,k},s,[u_2])=\int_{0}^{1}D_uK(t_{n,k},s,[u_2+\alpha(u_1-u_2)])(u_1-u_2)\,\mathrm{d}\alpha \tag*{(127)} K ( t n , k , s , [ u 1 ]) − K ( t n , k , s , [ u 2 ]) = ∫ 0 1 D u K ( t n , k , s , [ u 2 + α ( u 1 − u 2 )]) ( u 1 − u 2 ) d α (127)
u 1 ( t ) − u 2 ( t ) = ∑ j = 1 m ( U j − V j ) L ~ n , j ( t ) (128) u_1(t)-u_2(t)=\sum_{j=1}^{m}(U_j-V_j)\widetilde{L}_{n,j}(t) \tag*{(128)} u 1 ( t ) − u 2 ( t ) = j = 1 ∑ m ( U j − V j ) L n , j ( t ) (128)
Inserting our assumed form of the Fréchet-derivative
∣ ( f ( U ) − f ( V ) ) k ∣ ≤ ∣ ∫ t 0 t n , k ∫ 0 1 k 1 ( t n , k , s , [ u 1 + α ( u 2 − u 1 ) ] ) ( u 1 ( s ) − u 2 ( s ) ) d α d s ∣ (129) \left|(f(U)-f(V))_k\right| \leq\left|\int_{t_0}^{t_{n,k}}\int_0^1 k_1(t_{n,k},s,[u_1+\alpha(u_2-u_1)])(u_1(s)-u_2(s))\,\mathrm{d}\alpha\mathrm{d}s\right| \tag*{(129)} ∣ ( f ( U ) − f ( V ) ) k ∣ ≤ ∫ t 0 t n , k ∫ 0 1 k 1 ( t n , k , s , [ u 1 + α ( u 2 − u 1 )]) ( u 1 ( s ) − u 2 ( s )) d α d s (129)
+ ∣ ∫ t 0 t n , k ∫ 0 1 ∫ t 0 t k 2 ( t n , k , s , τ , [ u 1 + α ( u 2 − u 1 ) ] ) ( u 1 ( τ ) − u 2 ( τ ) ) d τ d α d s ∣ (130) +\left|\int_{t_0}^{t_{n,k}}\int_0^1\int_{t_0}^{t} k_2(t_{n,k},s,\tau,[u_1+\alpha(u_2-u_1)])(u_1(\tau)-u_2(\tau))\,\mathrm{d}\tau\mathrm{d}\alpha\mathrm{d}s\right| \tag*{(130)} + ∫ t 0 t n , k ∫ 0 1 ∫ t 0 t k 2 ( t n , k , s , τ , [ u 1 + α ( u 2 − u 1 )]) ( u 1 ( τ ) − u 2 ( τ )) d τ d α d s (130)
= ∣ h n ∑ j = 1 m ( U j − V j ) ∫ 0 c k ∫ 0 1 k 1 ( t n , k , t n + s h n , [ u 1 + α ( u 2 − u 1 ) ] ) L j ( s ) d α d s ∣ (131) =\left|h_n\sum_{j=1}^{m}(U_j-V_j)\int_0^{c_k}\int_0^1 k_1(t_{n,k},t_n+sh_n,[u_1+\alpha(u_2-u_1)])L_j(s)\,\mathrm{d}\alpha\mathrm{d}s\right| \tag*{(131)} = h n j = 1 ∑ m ( U j − V j ) ∫ 0 c k ∫ 0 1 k 1 ( t n , k , t n + s h n , [ u 1 + α ( u 2 − u 1 )]) L j ( s ) d α d s (131)
+ ∣ h n ∑ j = 1 m ( U j − V j ) ∫ t 0 t n , k ∫ 0 1 ∫ 0 1 k 2 ( t n , k , s , t n + τ h n , [ u 1 + α ( u 2 − u 1 ) ] ) L j ( τ ) d τ d α d s ∣ (132) +\left|h_n\sum_{j=1}^{m}(U_j-V_j)\int_{t_0}^{t_{n,k}}\int_0^1\int_0^1 k_2(t_{n,k},s,t_n+\tau h_n,[u_1+\alpha(u_2-u_1)])L_j(\tau)\,\mathrm{d}\tau\mathrm{d}\alpha\mathrm{d}s\right| \tag*{(132)} + h n j = 1 ∑ m ( U j − V j ) ∫ t 0 t n , k ∫ 0 1 ∫ 0 1 k 2 ( t n , k , s , t n + τ h n , [ u 1 + α ( u 2 − u 1 )]) L j ( τ ) d τ d α d s (132)
≤ h n C k ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) ∥ U − V ∥ (133) \leq h_n C_k\left(\sup\lVert k_1\rVert_{L^1}+\sup\lVert k_2\rVert_{L^1}\right)\lVert U-V\rVert \tag*{(133)} ≤ h n C k ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) ∥ U − V ∥ (133)
From this follows, that for h < h c h<h_c h < h c f f f is a contraction and thus the collocation equations have unique solutions U n , k U_{n,k} U n , k , k = 1 , … , m k=1,\ldots,m k = 1 , … , m .
Step 2: Identifying the collocation error
The collocation error is defined as
e ( t ) = y ( t ) − u ( t ) (134) e(t)=y(t)-u(t) \tag*{(134)} e ( t ) = y ( t ) − u ( t ) (134)
To get a better representation, recall Peano’s theorem for t ∈ [ t n , t n + 1 ] t\in[t_n,t_{n+1}] t ∈ [ t n , t n + 1 ] for the error of polynomial interpolation (theorem 23 below):
y ( t n + z h n ) = ∑ k = 1 m y ( t n , k ) L k ( z ) + h n m ∫ 0 1 K m ( z , τ ) f ( m ) ( t n + τ h n ) d τ (135) y(t_n+zh_n)=\sum_{k=1}^{m}y(t_{n,k})L_k(z)+h_n^m\int_0^1 K_m(z,\tau)f^{(m)}(t_n+\tau h_n)\,\mathrm{d}\tau \tag*{(135)} y ( t n + z h n ) = k = 1 ∑ m y ( t n , k ) L k ( z ) + h n m ∫ 0 1 K m ( z , τ ) f ( m ) ( t n + τ h n ) d τ (135)
= ∑ k = 1 m y ( t n , k ) L k ( z ) + h n m ρ n ( z ) (136) =\sum_{k=1}^{m}y(t_{n,k})L_k(z)+h_n^m\rho_n(z) \tag*{(136)} = k = 1 ∑ m y ( t n , k ) L k ( z ) + h n m ρ n ( z ) (136)
Writing Y n , k = y ( t n , k ) Y_{n,k}=y(t_{n,k}) Y n , k = y ( t n , k ) we have for t ∈ ( t 0 , T ] t\in(t_0,T] t ∈ ( t 0 , T ] :
y ( t ) = ∑ n = 0 N − 1 χ ( t n , t n + 1 ] ∑ k = 1 m y ( t n , k ) L k ( t − t n h n ) + ∑ n = 0 N χ ( t n , t n + 1 ] h n m ρ n ( t − t n h n ) (137) y(t)=\sum_{n=0}^{N-1}\chi_{(t_n,t_{n+1}]} \sum_{k=1}^{m}y(t_{n,k})L_k\left(\frac{t-t_n}{h_n}\right)+\sum_{n=0}^{N}\chi_{(t_n,t_{n+1}]}h_n^m\rho_n\left(\frac{t-t_n}{h_n}\right) \tag*{(137)} y ( t ) = n = 0 ∑ N − 1 χ ( t n , t n + 1 ] k = 1 ∑ m y ( t n , k ) L k ( h n t − t n ) + n = 0 ∑ N χ ( t n , t n + 1 ] h n m ρ n ( h n t − t n ) (137)
= ∑ n = 0 N − 1 ∑ k = 1 m Y n , k L ~ n , k + ∑ n = 0 N h n m ρ ~ n ( t ) (138) =\sum_{n=0}^{N-1}\sum_{k=1}^{m}Y_{n,k}\widetilde{L}_{n,k}+\sum_{n=0}^{N}h_n^m\widetilde{\rho}_n(t) \tag*{(138)} = n = 0 ∑ N − 1 k = 1 ∑ m Y n , k L n , k + n = 0 ∑ N h n m ρ n ( t ) (138)
where we defined
ρ ~ n ( t ) = χ ( t n , t n + 1 ] ( t ) ρ n ( t − t n h n ) . (139) \widetilde{\rho}_n(t)=\chi_{(t_n,t_{n+1}]}(t)\rho_n\left(\frac{t-t_n}{h_n}\right). \tag*{(139)} ρ n ( t ) = χ ( t n , t n + 1 ] ( t ) ρ n ( h n t − t n ) . (139)
Thus, the collocation error is given by
e ( t ) = ∑ n = 0 N − 1 ∑ k = 1 m E n , k L ~ n , k ( t ) + ∑ n = 0 N − 1 h n m ρ ~ n ( t ) (140) e(t) = \sum_{n=0}^{N-1}\sum_{k=1}^{m} E_{n,k}\widetilde{L}_{n,k}(t) + \sum_{n=0}^{N-1}h_n^m\widetilde{\rho}_n(t) \tag*{(140)} e ( t ) = n = 0 ∑ N − 1 k = 1 ∑ m E n , k L n , k ( t ) + n = 0 ∑ N − 1 h n m ρ n ( t ) (140)
E n , k = Y n , k − U n , k (141) E_{n,k}=Y_{n,k}-U_{n,k} \tag*{(141)} E n , k = Y n , k − U n , k (141)
To show convergence, we thus have to show on one hand that the ρ n ( t ) \rho_n(t) ρ n ( t ) are bounded by a common constant and that the error E n , k = e ( t n , k ) E_{n,k}=e(t_{n,k}) E n , k = e ( t n , k ) at the collocation point scales with the maximum steplength h h h . As Peano’s theorem shows, the interpolation error ρ n \rho_n ρ n will be bounded if ∥ y ( m ) ∥ ∞ < ∞ \lVert y^{(m)}\rVert_\infty<\infty ∥ y ( m ) ∥ ∞ < ∞ . We will deal with E n , k E_{n,k} E n , k by using a Gronwall inequality.
Step 3: Deriving an inequality for the evolution of the collocation error
Since the collocation solution solves the equation exactly at the collocation points and since the interpolation error ρ n ( t n , k ) = 0 \rho_n(t_{n,k})=0 ρ n ( t n , k ) = 0 vanishes on the collocation points, the collocation error solves the following equation:
e ( t n , k ) = E n , k = ∫ t 0 t n , k ( K ( t n , k , s , [ y ] ) − K ( t n , k , s , [ u ] ) ) d s (142) e(t_{n,k})=E_{n,k}=\int_{t_0}^{t_{n,k}}\left(K(t_{n,k},s,[y])-K(t_{n,k},s,[u])\right)\,\mathrm{d}s \tag*{(142)} e ( t n , k ) = E n , k = ∫ t 0 t n , k ( K ( t n , k , s , [ y ]) − K ( t n , k , s , [ u ]) ) d s (142)
Again, we can ’linearize’ using the mean-value theorem (theorem 22):
E n , k = ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α ( y − u ) ] ) ( y − u ) d α ) d s (143) E_{n,k}=\int_{t_0}^{t_{n,k}}\left(\int_0^1D_uK(t_{n,k},s,[u+\alpha(y-u)])(y-u)\,\mathrm{d}\alpha\right)\,\mathrm{d}s \tag*{(143)} E n , k = ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α ( y − u )]) ( y − u ) d α ) d s (143)
= ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ] ) e d α ) d s (144) =\int_{t_0}^{t_{n,k}}\left(\int_0^1D_uK(t_{n,k},s,[u+\alpha e])e\,\mathrm{d}\alpha\right)\,\mathrm{d}s \tag*{(144)} = ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ]) e d α ) d s (144)
Next, insert the representation of the collocation error, Eq. B.50,
E n , k = ∑ j = 1 m E n , j ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ] ) ( L ~ n , j ) d α ) d s (145) E_{n,k}=\sum_{j=1}^{m}E_{n,j}\int_{t_0}^{t_{n,k}}\left(\int_0^1D_uK(t_{n,k},s,[u+\alpha e])(\widetilde{L}_{n,j})\,\mathrm{d}\alpha\right)\,\mathrm{d}s \tag*{(145)} E n , k = j = 1 ∑ m E n , j ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ]) ( L n , j ) d α ) d s (145)
+ ∑ i = 0 n − 1 ∑ j = 1 m E i , j ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ] ) ( L ~ i , j ) d α ) d s (146) +\sum_{i=0}^{n-1}\sum_{j=1}^{m}E_{i,j}\int_{t_0}^{t_{n,k}}\left(\int_0^1D_uK(t_{n,k},s,[u+\alpha e])(\widetilde{L}_{i,j})\,\mathrm{d}\alpha\right)\,\mathrm{d}s \tag*{(146)} + i = 0 ∑ n − 1 j = 1 ∑ m E i , j ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ]) ( L i , j ) d α ) d s (146)
+ ∑ i = 0 n h i m ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ] ) ( ρ ~ i ) d α ) d s (147) +\sum_{i=0}^{n}h_i^m\int_{t_0}^{t_{n,k}}\left(\int_0^1D_uK(t_{n,k},s,[u+\alpha e])(\widetilde{\rho}_i)\,\mathrm{d}\alpha\right)\,\mathrm{d}s \tag*{(147)} + i = 0 ∑ n h i m ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ]) ( ρ i ) d α ) d s (147)
Here, we already used the causality of the operator K K K to rule out contributions from future times t > t n , k t>t_{n,k} t > t n , k . Defining the vector E n = ( E n , 1 , … , E n , m ) E_n=(E_{n,1},\ldots,E_{n,m}) E n = ( E n , 1 , … , E n , m ) we have a linear equation for E n E_n E n of the form
( 1 − B n ) E n = ∑ i = 0 n − 1 B i E i + ∑ i = 0 n h i m P i (148) (1-B_n)E_n=\sum_{i=0}^{n-1}B_iE_i+\sum_{i=0}^{n}h_i^mP_i \tag*{(148)} ( 1 − B n ) E n = i = 0 ∑ n − 1 B i E i + i = 0 ∑ n h i m P i (148)
Assuming that ( 1 − B n ) (1-B_n) ( 1 − B n ) is invertible and uniformly bounded for all n ≥ 0 n \ge0 n ≥ 0 (∥ ( 1 − B n ) − 1 ∥ ≤ D 0 \|(1-B_n)^{-1}\| \le D_0 ∥ ( 1 − B n ) − 1 ∥ ≤ D 0 ), we have the following inequality for the norm of E n E_n E n :
∥ E 0 ∥ ≤ h 0 m ∥ ( 1 − B 0 ) − 1 ∥ ∥ P 0 ∥ ≤ h 0 m D 0 ∥ P 0 ∥ (149) \|E_0\| \le h_0^m\|(1-B_0)^{-1}\|\|P_0\| \le h_0^mD_0\|P_0\| \tag*{(149)} ∥ E 0 ∥ ≤ h 0 m ∥ ( 1 − B 0 ) − 1 ∥∥ P 0 ∥ ≤ h 0 m D 0 ∥ P 0 ∥ (149)
∥ E n ∥ ≤ ∑ i = 0 n − 1 ∥ ( 1 − B n ) − 1 ∥ ∥ B i ∥ ∥ E i ∥ + ∑ i = 0 n h i m ∥ ( 1 − B n ) − 1 ∥ ∥ P i ∥ , n ≥ 1 (150) \|E_n\| \le\sum_{i=0}^{n-1}\|(1-B_n)^{-1}\|\|B_i\|\|E_i\|+\sum_{i=0}^{n}h_i^m\|(1-B_n)^{-1}\|\|P_i\|,\qquad n\ge1 \tag*{(150)} ∥ E n ∥ ≤ i = 0 ∑ n − 1 ∥ ( 1 − B n ) − 1 ∥∥ B i ∥∥ E i ∥ + i = 0 ∑ n h i m ∥ ( 1 − B n ) − 1 ∥∥ P i ∥ , n ≥ 1 (150)
≤ ∑ i = 0 n − 1 D 0 ∥ B i ∥ ∥ E i ∥ + ∑ i = 0 n h i m D 0 ∥ P i ∥ , n ≥ 1 (151) \le\sum_{i=0}^{n-1}D_0\|B_i\|\|E_i\|+\sum_{i=0}^{n}h_i^mD_0\|P_i\|,\qquad n\ge1 \tag*{(151)} ≤ i = 0 ∑ n − 1 D 0 ∥ B i ∥∥ E i ∥ + i = 0 ∑ n h i m D 0 ∥ P i ∥ , n ≥ 1 (151)
Next, we apply Gronwalls inequality (theorem 24) to the sequence ϵ n = ∥ E n ∥ \epsilon_n=\|E_n\| ϵ n = ∥ E n ∥ , k j = D 0 ∥ B j ∥ k_j=D_0\|B_j\| k j = D 0 ∥ B j ∥ , q j = h j + 1 m D 0 ∥ P j + 1 ∥ q_j=h_{j+1}^mD_0\|P_{j+1}\| q j = h j + 1 m D 0 ∥ P j + 1 ∥ , ρ 0 = h 0 m D 0 ∥ P 0 ∥ \rho_0=h_0^mD_0\|P_0\| ρ 0 = h 0 m D 0 ∥ P 0 ∥ , and get the inequality
∥ E n ∥ ≤ ( D 0 ∑ i = 0 n h i m ∥ P i ∥ ) exp ( D 0 ∑ i = 0 n − 1 ∥ B i ∥ ) (152) \|E_n\| \le\left(D_0\sum_{i=0}^{n}h_i^m\|P_i\|\right)\exp\left(D_0\sum_{i=0}^{n-1}\|B_i\|\right) \tag*{(152)} ∥ E n ∥ ≤ ( D 0 i = 0 ∑ n h i m ∥ P i ∥ ) exp ( D 0 i = 0 ∑ n − 1 ∥ B i ∥ ) (152)
To establish convergence we now have to show that the terms appearing on the right hand side are bounded and vanish as h → 0 h\to0 h → 0 as will be done in the next few steps of the proof.
Step 4: ( 1 − B n ) (1-B_n) ( 1 − B n ) is invertible and the inverse stays bounded as h → 0 h\to0 h → 0
The idea is to use the Neumann-series to establish the existence of an inverse. Thus, we need to show ∥ B n ∥ → 0 \|B_n\|\to0 ∥ B n ∥ → 0 as h → 0 h\to0 h → 0 , since then there is a critical steplength h c h_c h c such that ( 1 − B n ) (1-B_n) ( 1 − B n ) is invertible and the inverse is bounded by ∥ ( 1 − B n ) − 1 ∥ ≤ D 0 \|(1-B_n)^{-1}\|\le D_0 ∥ ( 1 − B n ) − 1 ∥ ≤ D 0 for all h ≤ h c h\le h_c h ≤ h c for a D 0 ∈ R + D_0\in\mathbb{R}_+ D 0 ∈ R + .
The matrix B n B_n B n is defined by
( B n ) k , j = ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ] ) ( L ~ n , j ) d α ) d s (153) (B_n)_{k,j}=\int_{t_0}^{t_{n,k}}\left(\int_0^1D_uK(t_{n,k},s,[u+\alpha e])(\widetilde{L}_{n,j})\,d\alpha\right)\,ds \tag*{(153)} ( B n ) k , j = ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ]) ( L n , j ) d α ) d s (153)
Inserting the assumed form of the Frechet-derivative D u K D_uK D u K yields
( B n ) k , j = ∫ t 0 t n , k ( ∫ 0 1 k 1 ( t n , k , s , [ u + α e ] ) d α ) L ~ n , j ( s ) d s (154) (B_n)_{k,j}=\int_{t_0}^{t_{n,k}}\left(\int_0^1k_1(t_{n,k},s,[u+\alpha e])\,d\alpha\right)\widetilde{L}_{n,j}(s)\,ds \tag*{(154)} ( B n ) k , j = ∫ t 0 t n , k ( ∫ 0 1 k 1 ( t n , k , s , [ u + α e ]) d α ) L n , j ( s ) d s (154)
+ ∫ t 0 t n , k ( ∫ 0 1 ( ∫ t 0 t n , k k 2 ( t n , k , s , τ , [ u + α e ] ) L ~ n , j ( τ ) d τ ) d α ) d s (155) +\int_{t_0}^{t_{n,k}}\left(\int_0^1\left(\int_{t_0}^{t_{n,k}}k_2(t_{n,k},s,\tau,[u+\alpha e])\widetilde{L}_{n,j}(\tau)\,d\tau\right)d\alpha\right)\,ds \tag*{(155)} + ∫ t 0 t n , k ( ∫ 0 1 ( ∫ t 0 t n , k k 2 ( t n , k , s , τ , [ u + α e ]) L n , j ( τ ) d τ ) d α ) d s (155)
Restricting the integrals to the support of L ~ j , k \widetilde{L}_{j,k} L j , k and applying the transformation t → t − t j h j t\to\frac{t-t_j}{h_j} t → h j t − t j yields:
( B n ) k , j = h n ∫ 0 c k ( ∫ 0 1 k 1 ( t n , k , t n + s h n , [ u + α e ] ) d α ) L j ( s ) d s (156) (B_n)_{k,j}=h_n\int_0^{c_k}\left(\int_0^1k_1(t_{n,k},t_n+sh_n,[u+\alpha e])\,d\alpha\right)L_j(s)\,ds \tag*{(156)} ( B n ) k , j = h n ∫ 0 c k ( ∫ 0 1 k 1 ( t n , k , t n + s h n , [ u + α e ]) d α ) L j ( s ) d s (156)
+ h n ∫ t 0 t n , k ∫ 0 1 ( ∫ 0 c k k 2 ( t n , k , s , t n + τ h n , [ u + α e ] ) L j ( τ ) d τ ) d α d s (157) +h_n\int_{t_0}^{t_{n,k}}\int_0^1\left(\int_0^{c_k}k_2(t_{n,k},s,t_n+\tau h_n,[u+\alpha e])L_j(\tau)\,d\tau\right)d\alpha\,ds \tag*{(157)} + h n ∫ t 0 t n , k ∫ 0 1 ( ∫ 0 c k k 2 ( t n , k , s , t n + τ h n , [ u + α e ]) L j ( τ ) d τ ) d α d s (157)
Since sup ∣ k 1 ∣ , sup ∣ k 2 ∣ < ∞ \sup|k_1|,\sup|k_2| < \infty sup ∣ k 1 ∣ , sup ∣ k 2 ∣ < ∞ are bounded, we have
∣ ( B n ) k , j ∣ ≤ h n sup ∥ k 1 ∥ L 1 max s ∈ [ 0 , 1 ] ∣ L j ( s ) ∣ + h n sup ∥ k 2 ∥ L 1 max s ∈ [ 0 , 1 ] ∣ L j ( s ) ∣ (158) |(B_n)_{k,j}| \le h_n \sup\|k_1\|_{L^1} \max_{s\in[0,1]} |L_j(s)| + h_n \sup\|k_2\|_{L^1} \max_{s\in[0,1]} |L_j(s)| \tag*{(158)} ∣ ( B n ) k , j ∣ ≤ h n sup ∥ k 1 ∥ L 1 s ∈ [ 0 , 1 ] max ∣ L j ( s ) ∣ + h n sup ∥ k 2 ∥ L 1 s ∈ [ 0 , 1 ] max ∣ L j ( s ) ∣ (158)
≤ h n C j ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) (159) \le h_n C_j\left(\sup\|k_1\|_{L^1}+\sup\|k_2\|_{L^1}\right) \tag*{(159)} ≤ h n C j ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) (159)
Thus, all elements of B n B_n B n are bounded by a factor proportional to h n h_n h n and there exists an h c h_c h c such that ( 1 − B n ) (1-B_n) ( 1 − B n ) is invertible with bounded inverse for all h < h c h<h_c h < h c .
Step 5: Bounds for ∥ B i ∥ \|B_i\| ∥ B i ∥
The elements of the matrices B i B_i B i , i = 0 , … , n − 1 i=0,\ldots,n-1 i = 0 , … , n − 1 were defined by
( B i ) k , j = ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ y + α e ] ) ( L ~ i , j ) d α ) d s (160) (B_i)_{k,j}=\int_{t_0}^{t_{n,k}}\left(\int_0^1 D_uK(t_{n,k},s,[y+\alpha e])(\widetilde L_{i,j})\,d\alpha\right)ds \tag*{(160)} ( B i ) k , j = ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ y + α e ]) ( L i , j ) d α ) d s (160)
= h i ∫ 0 1 ( ∫ 0 1 k 1 ( t n , k , t i + s h i , [ y + α e ] ) d α ) L j ( s ) d s (161) =h_i\int_0^1\left(\int_0^1 k_1(t_{n,k},t_i+sh_i,[y+\alpha e])\,d\alpha\right)L_j(s)\,ds \tag*{(161)} = h i ∫ 0 1 ( ∫ 0 1 k 1 ( t n , k , t i + s h i , [ y + α e ]) d α ) L j ( s ) d s (161)
+ h i ∫ t 0 t n , k ∫ 0 1 ( ∫ 0 1 k 2 ( t n , k , s , t i + τ h i , [ y + α e ] ) L j ( τ ) d τ ) d α d s (162) +h_i\int_{t_0}^{t_{n,k}}\int_0^1\left(\int_0^1 k_2(t_{n,k},s,t_i+\tau h_i,[y+\alpha e])L_j(\tau)\,d\tau\right)d\alpha\,ds \tag*{(162)} + h i ∫ t 0 t n , k ∫ 0 1 ( ∫ 0 1 k 2 ( t n , k , s , t i + τ h i , [ y + α e ]) L j ( τ ) d τ ) d α d s (162)
Thus, we have
∣ ( B i ) j , k ∣ ≤ h i C j ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) (163) |(B_i)_{j,k}|\le h_i C_j\left(\sup\|k_1\|_{L^1}+\sup\|k_2\|_{L^1}\right) \tag*{(163)} ∣ ( B i ) j , k ∣ ≤ h i C j ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) (163)
Step 6: Bounds for ∥ P i ∥ \|P_i\| ∥ P i ∥
The elements of P i P_i P i are given by
( P i ) k = ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ] ) ρ ~ i d α ) d s (164) (P_i)_k=\int_{t_0}^{t_{n,k}}\left(\int_0^1 D_uK(t_{n,k},s,[u+\alpha e])\widetilde{\rho}_i\,d\alpha\right)ds \tag*{(164)} ( P i ) k = ∫ t 0 t n , k ( ∫ 0 1 D u K ( t n , k , s , [ u + α e ]) ρ i d α ) d s (164)
= ∫ t 0 t n , k ( ∫ 0 1 k 1 ( t n , k , s , [ u + α e ] ) d α ) ρ ~ i ( s ) d s + ∫ t 0 t n , k ∫ 0 1 ( ∫ t 0 t n , k k 2 ( t n , k , s , τ , [ u + α e ] ) ρ ~ i ( τ ) d τ ) d α d s (165) \begin{aligned}
=\int_{t_0}^{t_{n,k}}\left(\int_0^1 k_1(t_{n,k},s,[u+\alpha e])\,d\alpha\right)\widetilde{\rho}_i(s)\,ds \\
&+\int_{t_0}^{t_{n,k}}\int_0^1\left(\int_{t_0}^{t_{n,k}}k_2(t_{n,k},s,\tau,[u+\alpha e])\widetilde{\rho}_i(\tau)\,d\tau\right)d\alpha\,ds \tag*{(165)}
\end{aligned} = ∫ t 0 t n , k ( ∫ 0 1 k 1 ( t n , k , s , [ u + α e ]) d α ) ρ i ( s ) d s + ∫ t 0 t n , k ∫ 0 1 ( ∫ t 0 t n , k k 2 ( t n , k , s , τ , [ u + α e ]) ρ i ( τ ) d τ ) d α d s (165)
Inserting the definition of ρ ~ i \widetilde{\rho}_i ρ i (equ. B.49) yields
( P i ) k = h i ∫ 0 1 ( ∫ 0 1 k 1 ( t n , k , t i + s h i , [ u + α e ] ) d α ) ρ i ( s ) d s + h i ∫ t 0 t n , k ∫ 0 1 ( ∫ 0 1 k 2 ( t n , k , s , t i + τ h i , [ u + α e ] ) ρ i ( τ ) d τ ) d α d s (166) \begin{aligned}
(P_i)_k=h_i\int_0^1\left(\int_0^1 k_1(t_{n,k},t_i+sh_i,[u+\alpha e])\,d\alpha\right)\rho_i(s)\,ds \\
&+h_i\int_{t_0}^{t_{n,k}}\int_0^1\left(\int_0^1 k_2(t_{n,k},s,t_i+\tau h_i,[u+\alpha e])\rho_i(\tau)\,d\tau\right)d\alpha\,ds \tag*{(166)}
\end{aligned} ( P i ) k = h i ∫ 0 1 ( ∫ 0 1 k 1 ( t n , k , t i + s h i , [ u + α e ]) d α ) ρ i ( s ) d s + h i ∫ t 0 t n , k ∫ 0 1 ( ∫ 0 1 k 2 ( t n , k , s , t i + τ h i , [ u + α e ]) ρ i ( τ ) d τ ) d α d s (166)
Thus, we have
∣ ( P i ) k ∣ ≤ h i ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) max τ ∈ [ 0 , 1 ] ∣ ρ i ( τ ) ∣ (167) |(P_i)_k|\le h_i\left(\sup\|k_1\|_{L^1}+\sup\|k_2\|_{L^1}\right)\max_{\tau\in[0,1]}|\rho_i(\tau)| \tag*{(167)} ∣ ( P i ) k ∣ ≤ h i ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) τ ∈ [ 0 , 1 ] max ∣ ρ i ( τ ) ∣ (167)
Recall from Peano’s theorem (theorem 23) that ρ n ( τ ) = ∫ 0 1 K m ( z , τ ) f ( m ) ( t n + z h n ) d z \rho_n(\tau)=\int_0^1 K_m(z,\tau)f^{(m)}(t_n+zh_n)\mathrm{d}z ρ n ( τ ) = ∫ 0 1 K m ( z , τ ) f ( m ) ( t n + z h n ) d z . This yields
∣ ( P i ) k ∣ ≤ h i C m ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) ∥ y ( m ) ∥ ∞ (168) \lvert(P_i)_k\rvert\le h_i C_m\left(\sup\lVert k_1\rVert_{L^1}+\sup\lVert k_2\rVert_{L^1}\right)\lVert y^{(m)}\rVert_\infty \tag*{(168)} ∣( P i ) k ∣ ≤ h i C m ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) ∥ y ( m ) ∥ ∞ (168)
Step 7: Bound for ∥ E 0 ∥ \lVert E_0\rVert ∥ E 0 ∥
As previously derived when establishing the Gronwall inequality (equ. B.59), we have
∥ E 0 ∥ ≤ h 0 m D 0 ∥ P 0 ∥ (169) \lVert E_0\rVert\le h_0^m D_0\lVert P_0\rVert \tag*{(169)} ∥ E 0 ∥ ≤ h 0 m D 0 ∥ P 0 ∥ (169)
Inserting the bounds for the ∣ ( P 0 ) k ∣ \lvert(P_0)_k\rvert ∣( P 0 ) k ∣ derived in the previous step yields
∥ E 0 ∥ ≤ h 0 m + 1 C ~ m ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) ∥ y ( m ) ∥ ∞ (170) \lVert E_0\rVert\le h_0^{m+1}\widetilde C_m\left(\sup\lVert k_1\rVert_{L^1}+\sup\lVert k_2\rVert_{L^1}\right)\lVert y^{(m)}\rVert_\infty \tag*{(170)} ∥ E 0 ∥ ≤ h 0 m + 1 C m ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) ∥ y ( m ) ∥ ∞ (170)
The exact definition of C ~ m \widetilde C_m C m now depends on the choice of the vector norm ∥ E 0 ∥ \lVert E_0\rVert ∥ E 0 ∥ . If the 1-norm has been choosen we have C ~ m = m D 0 C m \widetilde C_m=mD_0C_m C m = m D 0 C m .
Last step: Inserting all previous results into the discrete Gronwall Inequality
Inserting the previous bounds into the Gronwall inequality, Eq. (B.61), yields
∥ E n ∥ ≤ ( ∑ i = 0 n h i m C ~ m h i ∥ y ( m ) ∥ ∞ ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) ) e ∑ i = 0 n h i C ˉ ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) (171) \lVert E_n\rVert\le\left(\sum_{i=0}^{n}h_i^m\widetilde C_mh_i\lVert y^{(m)}\rVert_\infty\left(\sup\lVert k_1\rVert_{L^1}+\sup\lVert k_2\rVert_{L^1}\right)\right)e^{\sum_{i=0}^{n}h_i\bar C\left(\sup\lVert k_1\rVert_{L^1}+\sup\lVert k_2\rVert_{L^1}\right)} \tag*{(171)} ∥ E n ∥ ≤ ( i = 0 ∑ n h i m C m h i ∥ y ( m ) ∥ ∞ ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) ) e ∑ i = 0 n h i C ˉ ( s u p ∥ k 1 ∥ L 1 + s u p ∥ k 2 ∥ L 1 ) (171)
≤ h m T ∥ y ( m ) ∥ ∞ ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) e C ˉ T ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) (172) \le h^mT\lVert y^{(m)}\rVert_\infty\left(\sup\lVert k_1\rVert_{L^1}+\sup\lVert k_2\rVert_{L^1}\right)e^{\bar C T\left(\sup\lVert k_1\rVert_{L^1}+\sup\lVert k_2\rVert_{L^1}\right)} \tag*{(172)} ≤ h m T ∥ y ( m ) ∥ ∞ ( sup ∥ k 1 ∥ L 1 + sup ∥ k 2 ∥ L 1 ) e C ˉ T ( s u p ∥ k 1 ∥ L 1 + s u p ∥ k 2 ∥ L 1 ) (172)
≤ C ( T ) ∥ y ( m ) ∥ ∞ h m (173) \le C(T)\lVert y^{(m)}\rVert_\infty h^m \tag*{(173)} ≤ C ( T ) ∥ y ( m ) ∥ ∞ h m (173)
This prooves the convergence theorem. □
Collection of theorems from Analysis needed for the proof
Theorem 22 (Mean-Value theorem). [34 Ch.VII Theorem 3.10] Let f ∈ C 1 ( X , F ) f\in C^1(X,F) f ∈ C 1 ( X , F ) be a continuous (Fréchet-)differentiable function between two Banach-spaces E , F E,F E , F with X ⊆ E X\subseteq E X ⊆ E . Then:
f ( y ) − f ( x ) = ∫ 0 1 D f ( x + α ( y − x ) ) ( y − x ) d α (174) f(y)-f(x)=\int_0^1 Df\left(x+\alpha(y-x)\right)(y-x)\,\mathrm{d}\alpha \tag*{(174)} f ( y ) − f ( x ) = ∫ 0 1 D f ( x + α ( y − x ) ) ( y − x ) d α (174) as long as [ x , y ] = { x + α ( y − x ) ∣ α ∈ [ 0 , 1 ] } ⊂ X [x,y]=\{x+\alpha(y-x)\mid\alpha\in[0,1]\}\subset X [ x , y ] = { x + α ( y − x ) ∣ α ∈ [ 0 , 1 ]} ⊂ X .
Theorem 23 (Peano’s Theorem). [28 Theorem 1.8.1 and corollary 1.8.2] For given abscissas a ≤ ξ 1 < ⋯ < ξ m ≤ b a\le\xi_1<\cdots<\xi_m\le b a ≤ ξ 1 < ⋯ < ξ m ≤ b with ξ i = a + c i h \xi_i=a+c_i h ξ i = a + c i h , h = b − a h=b-a h = b − a , let
e m ( f ; t ) = f ( t ) − ∑ j = 1 m L j ( t − a h ) f ( ξ j ) , t ∈ [ a , b ] (175) e_m(f;t)=f(t)-\sum_{j=1}^{m}L_j\left(\frac{t-a}{h}\right)f(\xi_j),\qquad t\in[a,b] \tag*{(175)} e m ( f ; t ) = f ( t ) − j = 1 ∑ m L j ( h t − a ) f ( ξ j ) , t ∈ [ a , b ] (175)
the error of the polynomial interpolation, with L j L_j L j the Lagrange-polynomials on the intervall [ 0 , 1 ] [0,1] [ 0 , 1 ] with abscissas c j c_j c j . Assume that f ∈ C d ( [ a , b ] ) f \in C^d([a,b]) f ∈ C d ([ a , b ]) is d d d -times continuously differentiable with 1 ≤ d ≤ m 1 \le d \le m 1 ≤ d ≤ m . Then the interpolation error e m ( f ; t ) e_m(f;t) e m ( f ; t ) posseses the integral representation
e m ( f ; a + z h ) = h d ∫ 0 1 K d ( z , s ) f ( d ) ( a + s h ) d s (176) e_m(f;a+zh)=h^d\int_0^1 K_d(z,s)f^{(d)}(a+sh)\,ds \tag*{(176)} e m ( f ; a + z h ) = h d ∫ 0 1 K d ( z , s ) f ( d ) ( a + s h ) d s (176)
where the Peano-kernel K d K_d K d is given by
K d ( z , s ) = 1 ( d − 1 ) ! ( ( z − s ) + d − 1 − ∑ k = 1 m L k ( z ) ( c k − s ) + d − 1 ) (177) K_d(z,s)=\frac{1}{(d-1)!}\left((z-s)_+^{d-1}-\sum_{k=1}^{m}L_k(z)(c_k-s)_+^{d-1}\right) \tag*{(177)} K d ( z , s ) = ( d − 1 )! 1 ( ( z − s ) + d − 1 − k = 1 ∑ m L k ( z ) ( c k − s ) + d − 1 ) (177)
( z − s ) + p = { 0 , if z < s ( z − s ) p , if z ≥ s (178) (z-s)_+^p=
\begin{cases}
0, & \text{if } z<s \\
(z-s)^p, & \text{if } z\ge s
\end{cases} \tag*{(178)} ( z − s ) + p = { 0 , ( z − s ) p , if z < s if z ≥ s (178)
Theorem 24 (Discrete Gronwall Inequality). [28 Theorem 2.1.18] Assume that { k j } j ≥ 0 \{k_j\}_{j\ge0} { k j } j ≥ 0 is a given non-negative sequence (k j ≥ 0 k_j\ge0 k j ≥ 0 ). If a sequence { ϵ n } n ≥ 0 \{\epsilon_n\}_{n\ge0} { ϵ n } n ≥ 0 satisfies
ϵ 0 ≤ ρ 0 \epsilon_0\le\rho_0 ϵ 0 ≤ ρ 0
ϵ n ≤ ρ 0 + ∑ j = 0 n − 1 q j + ∑ j = 0 n − 1 k j ϵ j \epsilon_n\le\rho_0+\sum_{j=0}^{n-1}q_j+\sum_{j=0}^{n-1}k_j\epsilon_j ϵ n ≤ ρ 0 + ∑ j = 0 n − 1 q j + ∑ j = 0 n − 1 k j ϵ j
with ρ 0 ≥ 0 \rho_0\ge0 ρ 0 ≥ 0 , q j ≥ 0 q_j\ge0 q j ≥ 0 for j ≥ 0 j\ge0 j ≥ 0 , then it fulfills the discrete Gronwall inequality:
ϵ n ≤ ( ρ 0 + ∑ j = 0 n − 1 q j ) e ∑ j = 0 n − 1 k j (179) \epsilon_n\le\left(\rho_0+\sum_{j=0}^{n-1}q_j\right)e^{\sum_{j=0}^{n-1}k_j} \tag*{(179)} ϵ n ≤ ( ρ 0 + j = 0 ∑ n − 1 q j ) e ∑ j = 0 n − 1 k j (179)
Trapezoidal rule applied to functions with jump discontinuities and kinks
We want to understand how isolated jump-discontinuities and kinks in the integrand f ( t ) f(t) f ( t ) influence the convergence behaviour of the (iterated) trapezoidal rule for numerical integration of ∫ a b f ( t ) d t \int_a^b f(t)dt ∫ a b f ( t ) d t . Amat et al. [35 Theorem 1] showed, that the error of the trapezoidal rule depends on the jump terms
[ f ] ( t ∗ ) = f ( t ∗ + ) − f ( t ∗ − ) (180) [f](t^*)=f(t^{*+})-f(t^{*-}) \tag*{(180)} [ f ] ( t ∗ ) = f ( t ∗+ ) − f ( t ∗− ) (180)
[ f ′ ] ( t ∗ ) = f ′ ( t ∗ + ) − f ′ ( t ∗ − ) (181) [f'](t^*)=f'(t^{*+})-f'(t^{*-}) \tag*{(181)} [ f ′ ] ( t ∗ ) = f ′ ( t ∗+ ) − f ′ ( t ∗− ) (181)
where t ∗ t^* t ∗ denotes the position of the jump discontinuity or kink of f f f . Let
a = t 0 < ⋯ < t n = b (182) a=t_0<\cdots<t_n=b \tag*{(182)} a = t 0 < ⋯ < t n = b (182)
denote the grid used for the numerical integration. The local steplength is denoted as
h j = t j − t j − 1 j ∈ { 0 , … , n − 1 } (183) h_j=t_j-t_{j-1}\quad j\in\{0,\ldots,n-1\} \tag*{(183)} h j = t j − t j − 1 j ∈ { 0 , … , n − 1 } (183)
Let t k ∗ ∈ [ a , b ] t_k^* \in[a,b] t k ∗ ∈ [ a , b ] , k ∈ { 1 , … , m } k \in\{1,\ldots,m\} k ∈ { 1 , … , m } , denote the position of the (isolated) jumps or kinks of f f f in [ a , b ] [a,b] [ a , b ] . Denote the intervall where the jumps or kinks are located by t k ∗ ∈ [ t j k , t j k + 1 ] t_k^* \in[t_{j_k},t_{j_k+1}] t k ∗ ∈ [ t j k , t j k + 1 ] and assume that each such interval contains only one such point. The distance from the jump or kink to the lower endpoint of the interval is given by
α k = t k ∗ − t j k , k ∈ { 1 , … , m } . (184) \alpha_k=t_k^*-t_{j_k},\quad k\in\{1,\ldots,m\}. \tag*{(184)} α k = t k ∗ − t j k , k ∈ { 1 , … , m } . (184)
The local error of the trapezoidal approximation given in the paper from Amat et al. [35 ] then reads
∫ t j k t j k + 1 f ( t ) d t − h j k 2 ( f ( t j k + 1 ) + f ( t j k ) ) = h j k − 2 α k 2 [ f ] ( t k ∗ ) + α k 2 − h j k α k 2 [ f ′ ] ( t k ∗ ) − 1 12 α k 3 f ′ ′ ( ξ 1 ) − 1 12 ( h j k − α k ) 3 f ′ ′ ( ξ 2 ) + O ( h j 4 ) (185) \int_{t_{j_k}}^{t_{j_k+1}}f(t)\,\mathrm{d}t-\frac{h_{j_k}}{2}\left(f(t_{j_k+1})+f(t_{j_k})\right)=\frac{h_{j_k}-2\alpha_k}{2}[f](t_k^*)+\frac{\alpha_k^2-h_{j_k}\alpha_k}{2}[f'](t_k^*)-\frac{1}{12}\alpha_k^3 f''(\xi_1)-\frac{1}{12}(h_{j_k}-\alpha_k)^3f''(\xi_2)+O(h_j^4) \tag*{(185)} ∫ t j k t j k + 1 f ( t ) d t − 2 h j k ( f ( t j k + 1 ) + f ( t j k ) ) = 2 h j k − 2 α k [ f ] ( t k ∗ ) + 2 α k 2 − h j k α k [ f ′ ] ( t k ∗ ) − 12 1 α k 3 f ′′ ( ξ 1 ) − 12 1 ( h j k − α k ) 3 f ′′ ( ξ 2 ) + O ( h j 4 ) (185)
where ξ 1 ∈ [ t j k , t k ∗ ] \xi_1\in[t_{j_k},t_k^*] ξ 1 ∈ [ t j k , t k ∗ ] and ξ 2 ∈ [ t k ∗ , t j k + 1 ] \xi_2\in[t_k^*,t_{j_k+1}] ξ 2 ∈ [ t k ∗ , t j k + 1 ] . Thus, if the positions of the kink t k ∗ t_k^* t k ∗ and the jump terms [ f ] ( t k ∗ ) [f](t_k^*) [ f ] ( t k ∗ ) and [ f ′ ] ( t k ∗ ) [f'](t_k^*) [ f ′ ] ( t k ∗ ) are known, one can calculate corrections to the integrated trapezoidal rule to ensure that the (global) approximation can be bounded by O ( h max 2 ) O(h_{\max}^2) O ( h m a x 2 ) for a sequence of quasiuniform grids8 :
∣ ∫ a b f ( t ) d t − ∑ j = 0 n − 1 h j 2 ( f ( t j + 1 ) + f ( t j ) ) + ∑ k = 1 m C k + C k ′ ∣ ≤ ( b − a ) 12 h max h min h max 2 max ξ ∈ [ a , b ] ∣ f ′ ′ ( ξ ) ∣ (186) \left|\int_a^b f(t)\,\mathrm{d}t-\sum_{j=0}^{n-1}\frac{h_j}{2}\left(f(t_{j+1})+f(t_j)\right)+\sum_{k=1}^{m}C_k+C'_k\right|\leq\frac{(b-a)}{12}\frac{h_{\max}}{h_{\min}}h_{\max}^2\max_{\xi\in[a,b]}|f''(\xi)| \tag*{(186)} ∫ a b f ( t ) d t − j = 0 ∑ n − 1 2 h j ( f ( t j + 1 ) + f ( t j ) ) + k = 1 ∑ m C k + C k ′ ≤ 12 ( b − a ) h m i n h m a x h m a x 2 ξ ∈ [ a , b ] max ∣ f ′′ ( ξ ) ∣ (186)
C k = h j k − 2 α k 2 [ f ] ( t k ∗ ) (187) C_k=\frac{h_{j_k}-2\alpha_k}{2}[f](t_k^*) \tag*{(187)} C k = 2 h j k − 2 α k [ f ] ( t k ∗ ) (187)
C k ′ = α k 2 − h j k α k 2 [ f ′ ] ( t k ∗ ) (188) C'_k=\frac{\alpha_k^2-h_{j_k}\alpha_k}{2}[f'](t_k^*) \tag*{(188)} C k ′ = 2 α k 2 − h j k α k [ f ′ ] ( t k ∗ ) (188)
Note, during refinement of the grid { t j } \{t_j\} { t j } the distances
α k = t k ∗ − t j k ≤ h j k (189) \alpha_k=t_k^*-t_{j_k}\leq h_{j_k} \tag*{(189)} α k = t k ∗ − t j k ≤ h j k (189)
between the kink positions t k ∗ t_k^* t k ∗ and the nearest smaller gridpoints t j k t_{j_k} t j k fluctuate arbitrarily in general.Thus the error will fluctuate too and even with the correction one will only see a ’rugged’ second order convergence which prohibits the use of extrapolation methods.
To apply this result to the numerical study of the SRR and the SO model it is necessary to look at the jump terms [ G ( t , s , [ u ] ) ] u ( s ) [G(t,s,[u])]u(s) [ G ( t , s , [ u ])] u ( s ) and [ ∂ s ( G ( t , s , [ u ] ) u ( s ) ) ] [\partial_s(G(t,s,[u])u(s))] [ ∂ s ( G ( t , s , [ u ]) u ( s ))] at the points t ∗ t^* t ∗ where u ( t ∗ ) = 0 u(t^*)=0 u ( t ∗ ) = 0 . Since G ( t , s , [ u ] ) G(t,s,[u]) G ( t , s , [ u ]) is continuous and smooth everywhere where u ( t ) u(t) u ( t ) does not vanish, we have
[ G ( t , t , [ u ] ) u ( t ) ] = 0 (190) [G(t,t,[u])u(t)]=0 \tag*{(190)} [ G ( t , t , [ u ]) u ( t )] = 0 (190)
To calculate the jump in the first derivative note
∂ s ( G ( t , s , [ u ] ) u ( s ) ) = G ( t , s , [ u ] ) ∂ s u ( s ) + ( ∂ s G ( t , s , [ u ] ) ) u ( s ) (191) \partial_s\left(G(t,s,[u])u(s)\right) = G(t,s,[u])\partial_s u(s) + \left(\partial_s G(t,s,[u])\right)u(s) \tag*{(191)} ∂ s ( G ( t , s , [ u ]) u ( s ) ) = G ( t , s , [ u ]) ∂ s u ( s ) + ( ∂ s G ( t , s , [ u ]) ) u ( s ) (191)
= u ( s ) = 0 G ( t , s , [ u ] ) ∂ s u ( s ) (192) \overset{u(s)=0}{=}G(t,s,[u])\partial_s u(s) \tag*{(192)} = u ( s ) = 0 G ( t , s , [ u ]) ∂ s u ( s ) (192)
Thus, for the SRR and the SO model we have
[ ∂ s = t ∗ ( G ( t , t ∗ , [ u ] ) u ( s ) ) ] = 0 (193) \left[\partial_{s=t^*}\left(G(t,t^*,[u])u(s)\right)\right] = 0 \tag*{(193)} [ ∂ s = t ∗ ( G ( t , t ∗ , [ u ]) u ( s ) ) ] = 0 (193)
Therefore, for both models the approximation to the integral term using the iterated trapezoidal rule is of second order by equation C.7 even if u ( t ) u(t) u ( t ) changes sign, e.g. if u ( t ∗ ) = 0 u(t^*)=0 u ( t ∗ ) = 0 at isolated points t ∗ t^* t ∗ .