Abstract
We consider the numerical approximation of the stochastic Darcy problem with log-normal permeability field and propose a novel Multi Level Monte Carlo (MLMC) approach with a control variate variance reduction technique on each level. We model the log-permeability as a stationary Gaussian random field with a covariance function belonging to the so called Matérn family, which includes both fields with very limited and very high spatial regularity. The control variate is obtained starting from the solution of an auxiliary problem with smoothed permeability coefficient and its expected value is effectively computed with a Stochastic Collocation method on the finest level in which the control variate is applied. We analyze the variance reduction induced by the control variate, and the total mean square error of the new estimator. To conclude we present some numerical examples and a comparison with the standard MLMC method, which shows the effectiveness of the proposed method.
Similar content being viewed by others
References
Chilès, J.-P., Delfiner, P.: Geostatistics: Modeling Spatial Uncertainty. Wiley series in probability and statistics, 2nd edn. Wiley, Hoboken (2012)
Babuška, I., Nobile, F., Tempone, R.: A stochastic collocation method for elliptic partial differential equations with random input data. SIAM J. Numer. Anal. 45(3), 1005–1034 (2007)
Babuška, I., Tempone, R., Zouraris, G.E.: Galerkin finite element approximations of stochastic elliptic partial differential equations. SIAM J. Numer. Anal. 42(2), 800–825 (2004)
Beck, J., Nobile, F., Tamellini, L., Tempone, R.: Stochastic spectral Galerkin and collocation methods for PDEs with random coefficients: a numerical comparison. In: Spectral and High Order Methods for Partial Differential Equations, Lecture Notes in Computational Science and Engineering, vol. 76, pp. 43–62 (2011)
Le Maitre, O.P., Knio, O.M.: Spectral Methods for Uncertainty Quantification. Springer, Berlin (2010)
Nobile, F., Tempone, R., Webster, C.G.: A sparse grid stochastic collocation method for partial differential equations with random input data. SIAM J. Numer. Anal. 46(5), 2309–2345 (2008)
Barth, A., Schwab, C., Zollinger, N.: Multi-level Monte Carlo finite element method for elliptic PDEs with stochastic coefficients. Numerische Mathematik 119(1), 123–161 (2011)
Xiu, D., Karniadakis, G.E.: Modeling uncertainty in flow simulations via generalized polynomial chaos. J. Comput. Phys. 187(1), 137–167 (2003)
Beck, J., Nobile, F., Tamellini, L., Tempone, R.: A quasi-optimal sparse grids procedure for groundwater flows, in spectral and high order methods for partial differential equations - ICOSAHOM 2012. In: Lecture Notes in Computational Science and Engineering, vol. 95, pp. 1–16 (2014)
Ernst, O.G., Sprungk, B.: Stochastic collocation for elliptic PDEs with random data: the lognormal case. In: Sparse Grids and Applications - Munich 2012, Lecture Notes in Computational Science and Engineering, vol. 97, pp. 29–53 (2014)
Charrier, J.: Analyse numérique d’équations aux dérivées partielles à coefficients aléatoires, applications à l’hydrologéologie, thèse de doctorat, ENS Cachan-Bretagne, (2011)
Cliffe, K.A., Giles, M.B., Scheichl, R., Teckentrup, A.L.: Multi Level Monte Carlo methods and applications to elliptic PDEs with random coefficients. Comput. Vis. Sci. 14(1), 3–15 (2011)
Graham, I.G., Kuo, F.Y., Nuyens, D., Scheichl, R., Sloan, I.H.: Quasi-Monte Carlo methods for elliptic PDEs with random coefficients and applications. J. Comput. Phys. 230(10), 3668–3694 (2011)
Charrier, J., Scheichl, R., Teckentrup, A.: Finite element error analysis of elliptic PDEs with random coefficients and its application to multilevel Monte Carlo methods. SIAM J. Numer. Anal. 51(1), 332–352 (2013)
Teckentrup, A.L., Scheichl, R., Giles, M.B., Ullmann, E.: Further analysis of multilevel Monte Carlo methods for elliptic PDEs with random coefficients. Numer. Math. 125(3), 569–600 (2013)
Gittelson, C.J.: Stochastic Galerkin discretization of the log-normal isotropic diffusion problem. Math. Models Methods Appl. Sci. 20(2), 237–263 (2010)
Galvis, J., Sarkis, M.: Approximating infinity-dimensional stochastic Darcy’s equations without uniform ellipticity. SIAM J. Numer. Anal. 47, 3624–3651 (2009)
Grisvard, P.: Elliptic Problems in Nonsmooth Domains, Classics in Applied Mathematics. SIAM, Philadelphia (2011)
Charrier, J., Debussche, A.: Weak truncation error estimates for elliptic PDEs with lognormal coefficients. Stoch. Partial. Differ. Equ. 1(1), 63–93 (2013)
Giles, M.B.: Multi Level Monte Carlo path simulation. Oper. Res. 56(3), 607–617 (2008)
Collier, N., Haji-Ali, A.L., Nobile, F., von Schwerin, E., Tempone, R.: A continuation Multi Level Monte Carlo algorithm. BIT Numer. Math. 55, 1–34 (2014)
Robert, Christian, Casella, George: Monte Carlo Statistical Methods. Springer texts in statistics. Springer, New York (2005)
Kronrod, A.S.: Nodes and weights of quadrature formulas. Sixteen-place tables. Authorized translation from the Russian. Consultants Bureau, New York (1965)
Patterson, T.N.L.: The optimum addition of points to quadrature formulae. Math. Comput. 22, 847–856 (1968)
Nobile, F., Tamellini, L., Tesei, F., Tempone, R.: An adaptive sparse grid algorithm for elliptic PDEs with lognormal diffusion coefficient, Mathicse report 04/2015, (2015)
Chkifa, A., Cohen, A., Schwab, C.: High-dimensional adaptive sparse polynomial interpolation and applications to parametric PDEs. Found. Comput. Math. 14(4), 601–633 (2014)
Lang, A., Potthoff, J.: Fast simulation of Gaussian random field. Monte Carlo Methods Appl. 17, 195–214 (2011)
Haji-Ali, A.L., Nobile, F., Tempone, R.: Multi Index Monte Carlo: when sparsity meets sampling, Mathicse Report (2014)
Da Prato, G.: Encyclopedia of Mathematics and Its Applications. Stochastic equations in infinite dimensions, vol. 44. Cambridge University Press, Cambridge (1992)
Acknowledgments
F. Nobile and F. Tesei have been partially supported by the Swiss National Science Foundation under the Project No. 140574 “Efficient numerical methods for flow and transport phenomena in heterogeneous random porous media” and by the Center for ADvanced MOdeling Science (CADMOS).
Author information
Authors and Affiliations
Corresponding author
Appendices
Appendix 1: Proof of Lemma 4
We consider a fixed \(\omega \in \varOmega \) and we will not specify the dependence on \(\omega \) in the proof. In order to prove Lemma 4 we will need the following three preliminary lemmas.
Lemma 7
Let \(\gamma (x)\) be a deterministic function in \({\mathrm {C}}^{\alpha }({\mathbb {R}}^d)\) and, \(\forall h\in {{\mathbb {R}}^d}\), let us define \(D_{h,\beta } \gamma (x)\) as
Then, \(\forall \) \( 0<\beta \le \min (1,\alpha )\) it holds
and in particular
Proof
Let us denote \(\alpha = A + s\), with \(A\in {\mathbb {N}}\) and \(s\in (0,1]\).
\(\bullet \) We first consider the case \(0\le \beta \le \alpha \le 1\) so that \(A=0\) and \(s=\alpha \). The norm we want to bound can be written as
The first term can be bounded as
The second term can be bounded as
Hence we get \(\Vert D_{h,\beta }\gamma \Vert _{{\mathrm {C}}^{\alpha -\beta }({\mathbb {R}}^d)}\le 3 \Vert \gamma \Vert _{{\mathrm {C}}^{\alpha }({\mathbb {R}}^d)}\) and \(| D_{h,\beta }\gamma |_{{\mathrm {C}}^{\alpha -\beta }({\mathbb {R}}^d)}\le 2 |\gamma |_{{\mathrm {C}}^{\alpha }({\mathbb {R}}^d)}\).
\(\bullet \) Let us consider now the case \(0<\beta \le 1<\alpha \) so that \(\alpha =A+s\) with \(A\ge 1\). The proof can be further divided in two parts: \(s>\beta \) and \(s<\beta \) since for \(s=\beta \) the result is obvious. We start with the case \(s> \beta \). The norm that we want to bound can be written as
In what follows, we denote \(\xi _x^{y}\) a point of the segment \(\overline{xy}\), i.e. \(\xi _x^{y}=\theta x + (1-\theta )y\) for some \(\theta \in [0,1]\). The first term (i) can be bounded as
Each term of (ii), for \(k=1,\ldots ,A-1\), can be bounded as
The last term of (ii) for \(k=A\), analogously to what we did in the case \(0<\beta \le \alpha \le 1\), can be bounded as
Hence the term (ii) can be bounded as
The last term (iii) can be bounded as
So the norm of \(\Vert D_{h,\beta }\gamma \Vert _{{\mathrm {C}}^{\alpha -\beta }({\mathbb {R}}^d)}\) can we bounded as
and \(| D_{h,\beta }\gamma |_{{\mathrm {C}}^{\alpha -\beta }({\mathbb {R}}^d)}\le 2 | \gamma |_{{\mathrm {C}}^{\alpha }({\mathbb {R}}^d)} \).
\(\bullet \) Let us consider now the case \(s<\beta \). The quantity we want to bound becomes
We have already derived a bound for the terms (i) and (ii). The term (iii) can be bounded as follows:
By bounding separately the two terms we get
Similarly for the term \(II_i\), we have:
and then the term (iii) can be bounded as \((iii)\le 2\sqrt{d}| \gamma |_{{\mathrm {C}}^{\alpha }({\mathbb {R}}^d)} \). Finally the norm of \(\Vert D_{h,\beta }\gamma \Vert _{{\mathrm {C}}^{\alpha -\beta }({\mathbb {R}}^d)}\) can we bounded as
By comparing this expression with the one obtained in the previous case it is possible to conclude that \(\forall h\in {{\mathbb {R}}^d}\), \(0<\beta \le \min \{\alpha ,1\}\) it holds
\(\square \)
Lemma 8
Let \(\kappa (x)\in {\mathrm {C}}^{\alpha }({\mathbb {R}}^d)\) be a deterministic function as in Lemma 7, and let \(\kappa ^{\epsilon }(x)\) be a smoothed version of \(\kappa (x)\) defined as \(\kappa ^{\epsilon }(x)=(\kappa *\phi _{\epsilon })(x)\) with \( \phi _{\epsilon }(x)\) as in (11). It holds:
where \(C(\alpha ,d)=\frac{1}{\sqrt{2\pi }^d}\int _{{\mathbb {R}}^d}|y|^{\alpha } e^{-\frac{|y|^2}{2}} \ dy\).
Proof
By definition we have
\(\bullet \) Let us start with the case \({0<\alpha \le 1}\): if \(\kappa (x)\in {\mathrm {C}}^{\alpha }({\mathbb {R}}^d)\) then we obtain
\(\bullet \) If \(1<\alpha \le 2\) we consider a Taylor expansion of \(\kappa (x+y)\) around x and set \(\alpha = 1 +s\) with \(s\in (0,1]\). Since odd moments of a normal distribution vanish, we get:
\(\bullet \) Finally by considering \(2<\alpha \) and by expanding further the function \(\kappa \), since the second moment of a normal distribution does not vanish, we get:
\(\square \)
Lemma 9
Let \(\gamma (x)\in {\mathrm {C}}^{\alpha }(\overline{D})\) and \(\gamma ^{\epsilon }(x)\in {\mathrm {C}}^{\alpha }(\overline{D})\) be two deterministic functions and let \(a(x)=e^{\gamma (x)}\) and \(a^{\epsilon }(x)=e^{\gamma ^{\epsilon }(x)}\). For any \(0<\beta \le \min (1,\alpha )\) it holds
Proof
We bound separately the terms coming from the definition of the \({\mathrm {C}}^{\beta }\) norm, namely \(\Vert a- a^{\epsilon }\Vert _{{\mathrm {C}}^{\beta }( \overline{D} )}= \Vert a- a^{\epsilon }\Vert _{{\mathcal {C}}^{0}( \overline{D} )} +| a- a^{\epsilon }|_{{\mathrm {C}}^{\beta }( \overline{D} )}\). For the first one we simply observe that
For the second term we start by considering the inequality
The terms in the above equation can be bounded as follows:
By putting everything together we obtain
which is the desired result. \(\square \)
Thanks to these results we can prove Lemma 4.
Proof
(of Lemma 4.) From lemma 7 we have that \(\tilde{\gamma } \in {\mathrm {C}}^{\alpha }({\mathbb {R}}^d)\) implies \(D_{h,\beta } \tilde{\gamma } \in {\mathrm {C}}^{\alpha -\beta }({\mathbb {R}}^d)\) \(\forall \beta \le \min (\alpha ,1)\). By using the definitions given in (17), thanks to Lemmas 8 and 7 we get
Since \(D_{h,\beta } \gamma ^{\epsilon } = \left( D_{h,\beta } \tilde{\gamma } \right) ^{\epsilon }\) and thanks to the fact that the previous estimate is valid uniformly in h, we can take the supremum of \(\left\| D_{h,\beta } \gamma - \left( D_{h,\beta } \gamma \right) ^{\epsilon } \right\| _{{\mathcal {C}}^{0}({\mathbb {R}}^d)} \) with respect to h. By doing this we get
Now we get the desired result by observing that \( \left| \tilde{ \gamma } \right| _{C^{\alpha }({\mathbb {R}}^d)}\le \Vert \varphi \Vert _{C^{\alpha }(\overline{D}_{1})} \Vert \gamma \Vert _{C^{\alpha }(\overline{D}_{1})}\) . In fact, since \(\varphi \) vanishes on \(D_1^c\), we obtain
Hence, by considering the inequality given in Lemma 9:
since in D it holds \(\gamma =\tilde{\gamma }\) we get \(\Vert \gamma - \gamma ^{\epsilon } \Vert _{{\mathrm {C}}^{\beta }( \overline{D} )}=\Vert \tilde{\gamma } - \gamma ^{\epsilon } \Vert _{{\mathrm {C}}^{\beta }( \overline{D} )}\le \Vert \tilde{\gamma } - \gamma ^{\epsilon } \Vert _{{\mathrm {C}}^{\beta }( {\mathbb {R}}^d )}\) and we can conclude that
To prove the second bound concerning the \({\mathcal {C}}^1\) norm when \(\alpha >1\) we start again from the definition:
The first term, thanks to Lemma 8, can be bounded as
For the second term, since the derivatives and the convolution commute, we obtain
therefore we get
which implies
Finally, by putting everything together, and by recalling that \( \left| \tilde{ \gamma } \right| _{C^{\beta }({\mathbb {R}}^d)}\le \Vert \varphi \Vert _{C^{\beta }(\overline{D}_{1})} \Vert \gamma \Vert _{C^{\beta }(\overline{D}_{1})}\) \(\forall \beta \in {\mathbb {R}}_+\), we get the desired result:
\(\square \)
Appendix 2: on products of Hölder and Sobolev functions
Lemma 10
Let \(b\in {\mathrm {C}}^{\alpha }(\overline{D})\) and let v be a function in \(H^{\beta }(D)\) for some \(0<\beta <\alpha \le 1\). It holds
Proof
By definition the \(H^{\beta }\) norm of the function bv is
The first term can be easily bounded as \(\Vert bv\Vert _{L^2(D)}^2\le \Vert b\Vert _{{\mathcal {C}}^0(\overline{D})}^2\Vert v\Vert _{L^2(D)}^2\). For the second term we obtain
If we extend v by 0 in \({\mathbb {R}}^d \setminus D\), and denote \(\tilde{v}\) this extension and \(\rho =\max _{x\in D} |x|\), the integral appearing in the right hand side of the above inequality can be bounded as
By putting everything together we obtain
which is the desired result. \(\square \)
Appendix 3: regularity of Gaussian random fields with Matérn covariance
Lemma 11
Let \(\text {cov}_{\gamma }\) be a covariance function belonging to the Matérn family defined in (5) on an open bounded convex domain D. Then, if \(\nu \) is not an integer, \(\text {cov}_{\gamma }\in {\mathrm {C}}^{2\nu }(\bar{D}\times \bar{D})\), otherwise \(\text {cov}_{\gamma }\in {\mathrm {C}}^{\alpha }(\bar{D}\times \bar{D})\) for any \(\alpha <2\nu \) if \(\nu \in {\mathbb {N}}_+\) .
Proof
By definition we have
\(K_{\nu }:{\mathbb {R}}_{+}\rightarrow {\mathbb {R}}_+\) is given by \(K_{\nu }(\rho )=\frac{\pi }{2\sin {\pi \nu }}(I_{-\nu }(\rho )-I_{\nu }(\rho ))\) where \(I_{\alpha }(\rho )=\sum _{m=0}^{\infty }\frac{1}{m! \varGamma (m+\alpha +1)}\left( \frac{\rho }{2}\right) ^{2m+\alpha }\). This formula is valid when \(\nu \) is not an integer, i.e. \(\nu =n+s\) with \(n\in {\mathbb {N}}\) and \(s\in (0,1)\). Since \(\forall \epsilon >0\) the function \(K_{\nu }\in {\mathcal {C}}^{\infty }[\epsilon ,+\infty )\), and consequently \(\widetilde{\text {cov}}_{\gamma }\) as well, in order to prove the result we focus on the asymptotic behavior of the function \(\widetilde{\text {cov}}_{\gamma }(|x-y|)\) in a neighborhood of \(|x-y|=0\). By denoting \(\lambda _{\nu }=\frac{\sqrt{2\nu }}{2L_c}\) and by recalling that, for any \(x\in {\mathbb {R}}\setminus {\mathbb {Z}}\) it holds \(\varGamma (-x)=\frac{-\pi }{\sin \pi x \varGamma (x+1)}\), it is possible to obtain
Hence, the asymptotic behavior is
Since the function \(f(z)=|z|^{2\nu }:{\mathbb {R}}^d\rightarrow {\mathbb {R}}\) belongs to the space \({\mathrm {C}}^{2\nu }(A)\) for any bounded set \(A\in {\mathbb {R}}^d\) we can conclude that \( \text {cov}_{\gamma }\in {\mathrm {C}}^{2\nu }(\bar{D}\times \bar{D})\).
When \(\nu =n\in {\mathbb {N}}_+\) the previous definition gives removable indeterminate values of the form \(\frac{0}{0}\); in this case the Bessel function \(K_{\nu }\) can be defined through the limit \(K_{n}(\rho )=\lim _{\nu \rightarrow n} K_{\nu }(\rho )\). The covariance function becomes:
Again we focus on the asymptotic behavior of the function \(\widetilde{\text {cov}}_{\gamma }(|x_1-x_2|)\) in a neighborhood of \(|x_1-x_2|=0\). We obtain
Since the function \(f(z)=|z|^{2n}\log (|z|):{\mathbb {R}}^d\rightarrow {\mathbb {R}}\) belongs to the space \({\mathrm {C}}^{\alpha }(A)\) for any \(\alpha <2\nu \) and for any bounded set \(A\in {\mathbb {R}}^d\) we can conclude that \({\mathrm{cov}}_{\gamma }\in {\mathrm {C}}^{\alpha }(\bar{D}\times \bar{D})\) for any \(\alpha <2\nu \). \(\square \)
Remark 7
Let \(\widetilde{\text {cov}}_{\gamma }(|x-y|)\) be a covariance function belonging to the Matérn family defined in (5) and let \(\gamma (x,\omega )\) be a centered Gaussian random field defined on \(\bar{D}\). Denote \(\nu =n+\alpha \) with \(n\in {\mathbb {N}}\) and \(\alpha \in (0,1]\). Then, for any multi-index \(i\in {\mathbb {N}}^d\) such that \(|i|_1\le n\), it holds:
Lemma 12
Let \(\gamma (x,\omega )\) be a centered Gaussian random field with covariance function \(\text {cov}_{\gamma }\) as in Lemma 11. Then \(\gamma \) admits a version with trajectories a.s. in \({\mathrm {C}}^{\alpha }(\overline{D})\) for any \(0<\alpha < \nu \).
Proof
Let us start with the case in which \(\nu \) is not an integer. Lemma 11 tells us that \(\widetilde{\text {cov}}_{\gamma }\in {\mathrm {C}}^{2\nu }(A)\) for any bounded set \(A\in {\mathbb {R}}^d\). Therefore, thanks to (30), by writing \(\nu =n+s\) with \(n\in {\mathbb {N}}\) and \(s\in (0,1)\), for any multi-index \(i\in {\mathbb {N}}^d\) such that \(|i|_1= n\), we obtain
where the last inequality comes from the fact that the coefficients appearing in the covariance function decay sufficiently fast. Since for any positive integer p it holds \({\mathbb {E}}[(D^i\gamma (x,\cdot )-D^i\gamma (y,\cdot ))^{2p}] \le c_p{\mathbb {E}}[(D^i\gamma (x,\cdot )-D^i\gamma (y,\cdot ))^2] ^p\) with \(c_p=\frac{1}{\sqrt{2\pi }}\int _{{\mathbb {R}}}x^{2p} e^{-\frac{x^2}{2}}dx\) we have
Thanks to the Kolmogorov continuity theorem (see e.g. [29]) we can deduce that there exists a version of \(D^i\gamma \) which belongs to \({\mathrm {C}}^{a}(\bar{D})\) for any \(a<\frac{2ps-d}{2p}\); by taking the limit for \(p\rightarrow +\infty \) we can conclude that there exist a version of \(D^i\gamma \) which belongs to \({\mathrm {C}}^{a}(\bar{D})\) for any a strictly smaller than s. Consequently, since this reasoning can be repeated for every \(k<n\), \(k\in {\mathbb {N}}\), by picking 1 instead of s, we deduce that there exist a version of \(\gamma \) which belongs to \({\mathrm {C}}^{\alpha }(\bar{D})\) for any \(\alpha \) strictly smaller than \(\nu \). The proof in the case \(\nu \in {\mathbb {N}}\) is similar. By writing \(\nu =n+1\) in this case, thanks to (30), for any \(\epsilon >0\) and for any multi-index \(i\in {\mathbb {N}}^d\) such that \(|i|_1= n\) we obtain
Again, thanks to the Kolmogorov continuity theorem we can deduce that there exists a version of \(D^i\gamma \) which belongs to \({\mathrm {C}}^{a}(\bar{D})\) for any \(a<\frac{p(2-\epsilon )-d}{2p}\); thanks to the arbitrariness of \(\epsilon \) by taking the limit for \(p\rightarrow +\infty \) we can conclude that there exist a version of \(D^i\gamma \) which belongs to \({\mathrm {C}}^{a}(\bar{D})\) for any a strictly smaller than 1. Consequently we deduce that there exist a version of \(\gamma \) which belongs to \({\mathrm {C}}^{\alpha }(\bar{D})\) for any \(\alpha \) strictly smaller than \(\nu \). \(\square \)
Appendix 4: optimal rates in Theorem 1
Here we present a sharper bound for the infimum \(\inf _{\begin{array}{c} 0\le \beta \le \min (\alpha ,1)\\ 0<\eta +\beta \le \alpha \end{array} } \frac{h^{\beta } \epsilon ^{\min (\alpha -\beta -\eta ,2)}}{(\alpha -\beta )^2\sqrt{\eta }} \) than the one presented in Proposition 1 that can be obtained with very tedious calculations.
Lemma 13
Let \(\epsilon \le e^{-\frac{1}{2\alpha }}\); the following bounds hold:
-
\(1/2\le \alpha \le 1\):
$$\begin{aligned} \inf _{\begin{array}{c} 0\le \beta <\alpha \\ 0<\eta +\beta \le \alpha \end{array} } \frac{h^{\beta } \epsilon ^{\alpha -\beta -\eta }}{(\alpha -\beta )^2\sqrt{\eta }} \lesssim \,{\left\{ \begin{array}{ll} \,{\left\{ \begin{array}{ll} h^{\alpha }\left| \log \dfrac{h}{\epsilon } \right| ^2|\log \epsilon |^{\frac{1}{2}}, &{} \quad h\ge \epsilon ^5,\\ h^{\alpha }|\log h|^{\frac{5}{2}}, &{}\quad h\le \epsilon ^5\ \end{array}\right. } , &{}\quad h\le e^{-\frac{2}{\alpha }}\epsilon ,\\ \epsilon ^{\alpha }|\log \epsilon |^{\frac{1}{2}}\,&{} \quad h\ge e^{-\frac{2}{\alpha }}\epsilon .\ \end{array}\right. } \end{aligned}$$ -
\(1< \alpha \le 2\):
$$\begin{aligned}&\inf _{\begin{array}{c} 0\le \beta \le 1 \\ 0<\eta +\beta \le \alpha \end{array} } \frac{h^{\beta } \epsilon ^{\alpha -\beta -\eta }}{(\alpha -\beta )^2\sqrt{\eta }} \\&\lesssim {\left\{ \begin{array}{ll} h\epsilon ^{\alpha -1}|\log h|^{\frac{1}{2}}, \, \quad \quad h\le \min (e^{-\frac{5}{2(\alpha -1)}},e^{-\frac{2}{\alpha -1}}\epsilon ),\\ h^{\alpha }|\log h|^{\frac{5}{2}}, \, \quad \quad \min (e^{-\frac{5}{2(\alpha -1)}},e^{-\frac{2}{\alpha -1}}\epsilon )\le h\le e^{-\frac{2}{\alpha -1}}\epsilon ,\\ \,{\left\{ \begin{array}{ll} \,{\left\{ \begin{array}{ll} h^{\alpha }|\log h|^{\frac{5}{2}}, &{}\quad h\le \epsilon ^5,\\ h^{\alpha }\left| \log \dfrac{h}{\epsilon } \right| ^2|\log \epsilon |^{\frac{1}{2}}, &{}\quad h\ge \epsilon ^5,\ \end{array}\right. } , &{}\quad \text { if } e^{-\frac{2}{\alpha -1}}\le \epsilon ^5,\\ h^{\alpha }\left| \log \dfrac{h}{\epsilon } \right| ^2|\log \epsilon |^{\frac{1}{2}}, &{}\quad \text { if } e^{-\frac{2}{\alpha -1}}\ge \epsilon ^5,\ \end{array}\right. } \quad e^{-\frac{2}{\alpha -1}}\epsilon \le h\le e^{-\frac{2}{\alpha -1}}\epsilon ,\\ \epsilon ^{\alpha }|\log \epsilon |^{\frac{1}{2}}, \quad h\ge \! e^{-\frac{2}{\alpha }}\epsilon .\ \end{array}\right. } \end{aligned}$$ -
\(2< \alpha \le 3\):
$$\begin{aligned}&\inf _{\begin{array}{c} 0\le \beta \le 1 \\ 0<\eta +\beta \le \alpha \end{array} } \frac{h^{\beta } \epsilon ^{\alpha -\beta -\eta }}{(\alpha -\beta )^2\sqrt{\eta }} \\&\lesssim {\left\{ \begin{array}{ll} h\epsilon ^{\alpha -1}|\log h|^{\frac{1}{2}}, \, \quad \quad h\le \min (e^{-\frac{5}{2(\alpha -1)}},e^{-\frac{2}{\alpha -1}}\epsilon ),\\ h^{\alpha }|\log h|^{\frac{5}{2}}, \, \quad \quad \min (e^{-\frac{5}{2(\alpha -1)}},e^{-\frac{2}{\alpha -1}}\epsilon )\le h\le e^{-\frac{2}{\alpha -1}}\epsilon ,\\ \,{\left\{ \begin{array}{ll} \,{\left\{ \begin{array}{ll} h^{\alpha }|\log h|^{\frac{5}{2}}, &{}\quad h\le \epsilon ^5,\\ h^{\alpha }\left| \log \dfrac{h}{\epsilon } \right| ^2|\log \epsilon |^{\frac{1}{2}}, &{}\quad h\ge \epsilon ^5,\ \end{array}\right. } , &{}\quad \text { if } e^{-\frac{2}{\alpha -1}}\le \epsilon ^5,\\ h^{\alpha }\left| \log \dfrac{h}{\epsilon } \right| ^2|\log \epsilon |^{\frac{1}{2}}, &{}\quad \text { if } e^{-\frac{2}{\alpha -1}}\ge \epsilon ^5,\ \end{array}\right. } ,\quad e^{-\frac{2}{\alpha -1}}\epsilon \le h\le e^{-\frac{2}{\alpha }}\epsilon ,\\ \,{\left\{ \begin{array}{ll} \epsilon ^{\alpha }|\log \epsilon |^{\frac{1}{2}}, &{}\quad \text { if } \epsilon \ge e^{-\frac{1}{2(\alpha -2)}},\\ h^{\alpha -2}\epsilon ^{2}|\log \epsilon |^{\frac{1}{2}}, &{}\quad \text { if } \epsilon \le e^{-\frac{1}{2(\alpha -2)}},\ \end{array}\right. } \, \quad \quad h\ge e^{-\frac{2}{\alpha }}\epsilon .\ \end{array}\right. } \end{aligned}$$ -
\(\alpha > 3\):
$$\begin{aligned} \inf _{\begin{array}{c} 0\le \beta \le 1 \\ 0<\eta +\beta \le \alpha \end{array} } \frac{h^{\beta } \epsilon ^{\alpha -\beta -\eta }}{(\alpha -\beta )^2\sqrt{\eta }} \lesssim h\epsilon ^2 . \end{aligned}$$
Rights and permissions
About this article
Cite this article
Nobile, F., Tesei, F. A Multi Level Monte Carlo method with control variate for elliptic PDEs with log-normal coefficients. Stoch PDE: Anal Comp 3, 398–444 (2015). https://doi.org/10.1007/s40072-015-0055-9
Received:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s40072-015-0055-9
Keywords
- Log-normal random-fields
- Multi Level Monte Carlo
- Control variate
- Stochastic Collocation
- Matérn covariance
- Stochastic Darcy Problem