Another prominent model for non-Newtonian fluids, e.g., in polymer processing, is the power-law model, see, e.g., [
20, Ch. 3.3]. For this model, the weak formulation (
4) of the boundary-value problem under consideration is as follows:
$$\begin{aligned} \text {find } \mathbf {u}\in X \ \text {such that} \qquad \int _\Omega |\underline{e}(\mathbf {u})|^{r-2} \underline{e}(\mathbf {u}):\underline{e}(\mathbf {v})\,\mathsf {d}{\varvec{x}}=\ell (\mathbf {v}) \qquad \text {for all } \mathbf {v}\in X; \end{aligned}$$
(37)
here
\(X:=\{\mathbf {u}\in \mathrm {W}_0^{1,r}(\Omega )^d: \nabla \cdot \mathbf {u}=0\}\) and
\(\ell \in X^\star \), where, for shear-thinning fluids,
\(r \in (1,2)\). In particular, the viscosity coefficient is given by
$$\begin{aligned} \mu (t)=t^{\frac{r-2}{2}}. \end{aligned}$$
Clearly,
\(\mu :{\mathbb {R}}_{> 0} \rightarrow {\mathbb {R}}_{> 0}\) is neither bounded away from zero nor bounded from above, i.e., (A2) is not satisfied. Therefore, as was proposed in the work [
4], we consider a relaxed version of
\(\mu \): for
\(0<\varepsilon _{-}<\varepsilon _{+}<\infty \) we define the viscosity coefficient
$$\begin{aligned} \mu _{\varepsilon }(t):= {\left\{ \begin{array}{ll} \varepsilon _{-}^{r-2} &{}\quad \, \text {for}\ 0 \le t < \varepsilon _{-}^2, \\ t^{\frac{r-2}{2}} &{}\quad \, \text {for}\ \varepsilon _{-}^2 \le t \le \varepsilon _{+}^2, \\ \varepsilon _{+}^{r-2} &{}\quad \, \text {for}\ t \ge \varepsilon _{+}^2. \end{array}\right. } \end{aligned}$$
(38)
The function
\(\mu _\varepsilon \) is decreasing, strictly positive, bounded, globally Lipschitz continuous, and satisfies (A2) with
$$\begin{aligned}(r-1)\varepsilon _{+}^{r-2}(t-s) \le \mu (t^2)t-\mu (s^2)s \le \varepsilon _{-}^{r-2}(t-s), \qquad t \ge s \ge 0;\end{aligned}$$
it is, furthermore, differentiable at all
\(t \in [0,\infty )\setminus \{\varepsilon _{-}^2,\varepsilon _{+}^2\}\) and has finite left- and right-derivatives at
\(t=\varepsilon _{-}^2\) and
\(t=\varepsilon _{+}^2\), respectively. Hence, even though
\(\mu _{\varepsilon }\) is not continuously differentiable on
\([0,\infty )\), Theorem
3.2 can, nevertheless, be applied in the given setting. Moreover, in the generic case when the set
\(\Omega _S^n:=\{{\varvec{x}}\in \Omega : |\underline{e}(\mathbf {u}^n({\varvec{x}}))| \in \{\varepsilon _{-},\varepsilon _{+}\}\}\), for every
\(n \ge 0\), has Lebesgue measure zero, the operator
\(\omega \) from the proof of Theorem
3.4 is Fréchet differentiable at
\(\mathbf {u}^n \in W\). Thus, in turn, Theorem
3.4 can then also be applied to the relaxed power-law model
1. A simple calculation reveals that the computable contraction factor from (
33) can again be bounded; indeed,
$$\begin{aligned} q_A(n) \le 1-\frac{1}{4}(r-1). \end{aligned}$$
(39)
Moreover, one even has that
\(q_A(n)=1-4^{-1}(r-1)\) if the set
\(\{{\varvec{x}}\in \Omega : \varepsilon _{-} \le |\underline{e}(\mathbf {u}^n({\varvec{x}}))| \le \varepsilon _{+}\}\) is of positive Lebesgue measure. We further remark that
$$\begin{aligned} \frac{m_\mu }{M_\mu }=\frac{(r-1) \varepsilon _{+}^{r-2}}{\varepsilon _{-}^{r-2}} < (r-1), \end{aligned}$$
since
\(r \in (1,2)\). This shows that the bound (
39) is, for every value
\(r \in (1,2)\), sharper than the bound from Remark
3.3. Furthermore, this bound predicts that it is the physical parameter
r that affects the convergence rate of the iteration, in the finite-dimensional setting at least, rather than the quotient
\(\nicefrac {\varepsilon _{+}^{r-2}}{\varepsilon _{-}^{r-2}}\) implied by existing bounds on the contraction factor, cf. [
4, Cor. 19]. Significantly, the upper bound
\((r-1)\) on the contraction factor appearing of the right-hand side of (
39) is independent of the relaxation parameters
\(\varepsilon _{\pm }\). This is of importance as we are interested in the power-law model (
37) and we thus need to let
\(\varepsilon _{-} \rightarrow 0\) and
\(\varepsilon _{+} \rightarrow \infty \). We note that the existence of a bound independent of
\(\varepsilon _{\pm }\) on the contraction factor of the relaxed Kačanov iteration applied to the power-law model with
\(r \in (1,2)\) was stated in the infinite-dimensional case as an open problem in [
4, Ex. 20].
We further note that the energy functional
\(\mathsf {E}_{\varepsilon }\) corresponding to the viscosity from (
38) coincides with the energy functional
\({\mathcal {J}}_\epsilon \) from [
4] up to a constant shift depending on
\(\varepsilon _{-}\). To be precise, one has that
$$\begin{aligned} \mathsf {E}_{\varepsilon }(\mathbf {u})={\mathcal {J}}_\epsilon (\mathbf {u})+\left( \frac{1}{2}-\frac{1}{r}\right) \varepsilon _{-}^r, \qquad \mathbf {u}\in V, \end{aligned}$$
and thus the results established in [
4] may be directly applied in our setting. In particular, this implies that the sequence of unique minimisers
\(\mathbf {u}^\star _\varepsilon \in V\) of
\(\mathsf {E}_{\varepsilon }\) converges in
\(\mathrm {W}^{1,r}_{0}(\Omega )^d\) to the unique minimiser
\(\mathbf {u}^\star \in X\) of
$$\begin{aligned}\mathsf {E}(\mathbf {u})=\frac{1}{r}\int _\Omega |\underline{e}(\mathbf {u})|^r\,\mathsf {d}{\varvec{x}}-\ell (\mathbf {u}),\end{aligned}$$
cf. [
4, Cor. 10].