Abstract
Many mathematical models involve input parameters, which are not precisely known. Global sensitivity analysis aims to identify the parameters whose uncertainty has the largest impact on the variability of a quantity of interest. One of the statistical tools used to quantify the influence of each input variable on the quantity of interest are the Sobol’ sensitivity indices. In this paper, we consider stochastic models described by stochastic differential equations (SDE). We focus the study on mean quantities, defined as the expectation with respect to the Wiener measure of a quantity of interest related to the solution of the SDE itself. Our approach is based on a Feynman-Kac representation of the quantity of interest, from which we get a parametrized partial differential equation (PDE) representation of our initial problem. We then handle the uncertainty on the parametrized PDE using polynomial chaos expansion and a stochastic Galerkin projection.
Similar content being viewed by others
References
Brezis H (1983) Analyse fonctionnelle. Collection Mathématiques Appliquées pour la Maîtrise, Masson, Paris, théorie et applications
Cameron RH, Martin WT (1947) The orthogonal development of non-linear functionals in series of fourier-hermite functionals. Ann Math 48(2):385–392
Cooper FC (2017) Optimisation of an idealised primitive equation ocean model using stochastic parameterization. Ocean Model 113:187–200
Dautray R, Lions J (1993) Mathematical analysis and numerical methods for science and technology. No. vol. 6 in Mathematical Analysis and Numerical Methods for Science and Technology. Springer
Ern A, Guermond J (2002) Éléments finis: théorie, applications, mise en oeuvre. Mathématiques et Applications. Springer, Berlin
Friedman A (1975) Stochastic differential equations and applications. No. vol. 1 in Probability and mathematical statistics. Academic Press, New York
Ghanem RG, Spanos PD (1991) Stochastic finite elements: a spectral approach. Springer, Berlin
Gilbarg D, Trudinger NS (1983) Elliptic partial differential equations of second order. Grundlehren der mathematischen Wissenschaften. Springer, Berlin
Gillespie DT, Petzold LR (2003) Improved leap-size selection for accelerated stochastic simulation. J Chem Phys 119(16):8229–8234
Gobet E (2016) Monte-Carlo methods and stochastic processes: from linear to non-linear. Chapman & Hall/CRC
Hart JL, Alexanderian A, Gremaud PA (2017) Efficient computation of sobol’indices for stochastic models. SIAM J Sci Comput 39(4):A1514–A1530
Iooss B, Ribatet M (2009) Global sensitivity analysis of computer models with functional inputs. Reliab Eng Syst Saf 94(7):1194–1204
Janon A, Klein T, Lagnoux A, Nodet M, Prieur C (2014) Asymptotic normality and efficiency of two sobol index estimators. ESAIM Probab Statist 18:342–364
Jimenez MN, Le Maître O, Knio O (2017) Nonintrusive polynomial chaos expansions for sensitivity analysis in stochastic differential equations. SIAM/ASA J Uncertain Quantif 5(1):378–402
Karatzas I, Shreve S (1991) Brownian motion and stochastic calculus, 2nd edn. Graduate Texts in Mathematics, 113. Springer, New York. xxiii
Le Maître O, Knio OM (2010) Spectral methods for uncertainty quantification: with applications to computational fluid dynamics. Springer Science & Business Media
Le Maître O, Knio OM (2015) Pc analysis of stochastic differential equations driven by wiener noise. Reliab Eng Syst Saf 135:107–124
Lieberman G (1996) Second order parabolic differential equations. World Scientific, Singapore
Lions JL, Magenes E (1972) Non-homogeneous boundary value problems and applications, vol I. Springer, New York-Heidelberg. translated from the French by P. Kenneth, Die Grundlehren der mathematischen Wissenschaften, Band 181
Marrel A, Iooss B, Da Veiga S, Ribatet M (2012) Global sensitivity analysis of stochastic computer models with joint metamodels. Stat Comput 22(3):833–847
Monod H, Naud C, Makowski D (2006) Uncertainty and sensitivity analysis for crop models. Working with dynamic crop models: evaluation, analysis, parameterization, and applications, vol 4, pp 55–100
Nouy A (2009) Recent developments in spectral stochastic methods for the numerical solution of stochastic partial differential equations. Arch Comput Methods Eng 16(3):251–285
Nouy A (2017) Low-rank methods for high-dimensional approximation and model order reduction. Model Reduction and Approximation: Theory and Algorithms 15:171
Raviart P, Thomas J (2004) Introduction à l’analyse numérique des équations aux dérivées partielles. Mathématiques appliquées pour la maîtrise, Dunod
Saltelli A, Chan K, Scott EM et al (2000) Sensitivity analysis, vol 1. Wiley, New York
Sobol IM (1993) Sensitivity estimates for nonlinear mathematical models. Math Modeling Comput Experiment 1(4):407–414
Sobol IM (2001) Global sensitivity indices for nonlinear mathematical models and their monte carlo estimates. Math Comput Simul 55(1):271–280
van der Vaart AW (1998) Asymptotic statistics, Cambridge Series in Statistical and Probabilistic Mathematics, vol 3. Cambridge University Press, Cambridge
Wiener N (1938) The homogeneous chaos. Am J Math 60(4):897–936
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix
Appendix
1.1 A.1 Integrability of the Stochastic Quantities of Interest
Proof
(Proof of Lemma 1) Let \(q\in \mathbb {N}^{*}\) and t ≥ 0. Using Problem 5.2.15 in Karatzas and Shreve (1991), we can claim that, under (A3), for any z ∈Ξ
where C = C(t,M). As the constant M in Eq. 1 is uniform in z, the constant C does not depend on z either. Integrating Eq. 39 over Ξ and against ℙξ we get the desired result. □
Proof
(Proof of Lemma 2) Here we revisit the proof of Lemma 5.7.4 in Karatzas and Shreve (1991) and give details for the sake of completeness. Thanks to (A4), we have for example a11(x,z) ≥ λ for any \(x=(x_{1}, {\ldots } , x_{d})\in \mathbb {R}^{d}\), z ∈Ξ. We note \(q=\min _{x\in \bar {D}}x_{1}\) and b a constant s.t. |b(x,z)|≤ b < ∞,∀x ∈ D,∀z ∈Ξ.
Such a finite constant b exists thanks to Eq. 1. Set now ν > 2b/λ and consider the function h(x) = −μ exp(νx1), x ∈ D, where the constant μ will be determined later. This function is of class C∞(D) and satisfies for any x ∈ D and z ∈Ξ,
Choosing μ sufficiently large, we have for any x ∈ D, any z ∈Ξ:
Let us write τD(ω) = τD(𝜃,z). As X0 = x ∈ D, using Itô Formula and taking the expectation against \(\mathbb {W}\) we get:
Note that in the above computation, the expectation of the stochastic integral vanishes, as h is bounded with bounded derivatives. Then, integrating over Ξ and against ℙξ the above inequality we get the result by monotone convergence. □
Proof
(Proof of Lemma 3) According to Theorem 1, the function \(\mathfrak {U}(\cdot ,\xi )\) solves Eq. 16 (for a.e. fixed value of ξ). We now use a maximum principle argument to get an a priori bound on \(\mathfrak {U}(\cdot ,\xi )\). Without loss of generality we assume that D lies in the slab 0 < x1 < δ for a certain 0 < δ < ∞ (the general case can be recovered by translation arguments). According to Theorem 3.7 in Gilbarg and Trudinger (1983) and its proof we have
where C′(ξ) = eα(ξ)δ − 1 with α(ξ) a quantity chosen s.t. \(\alpha (\xi )\geq 1+\frac {1}{\lambda }\sup _{x\in D}|b(x,\xi )|\) (note that the uniform ellipticity constant λ in (A4) does not depend on z). Then, thanks to the uniformity in z of the constant M in Eq. 1 and to the boundedness of D, the quantity α(ξ) may be chosen independent on ξ. Therefore C′(ξ) = C′ does not depend on ξ, and thus we get the result with C = C′/λ. □
1.2 A.2 Results on Stochastic Partial Differential Equations
Proof
(Proof of Lemma 5) As \(\tilde {f}\) is uniformly bounded in x and z, it is easy to prove that the form F is continuous on \(\mathcal {V}\) (with the help of Cauchy-Schwarz’s inequality). Thanks to (A7) and (A9), the bilinear form \(\mathcal {A}\) is continuous on \(\mathcal {V}\). Thanks to (A8)–(A9) one may check the coercivity of \(\mathcal {A}\). Indeed, using Lemma 8.4 in Gilbarg and Trudinger (1983), one gets for any z ∈Ξ,
where \(\nu =\frac {\tilde {M}}{\tilde {\lambda }}\). Using now the definition of \(||\cdot ||_{H^{1}(D)}\), Inequality Eq. 22, and integrating Eq. 41 over Ξ against ℙξ we get
(we recall that C(d,|D|) is defined by Eq. 23). As \(\tilde {M}<\frac {\tilde {\lambda }}{\sqrt {2C(d,|D|)}}\), we have that \(\frac {\tilde {\lambda }}{2C(d,|D|)}-\tilde {\lambda }\nu ^{2}>0\). The existence of a unique weak solution at the stochastic level then follows from the Lax-Milgram theorem.
Now, as F and \(\mathcal {A}\) are continuous, as \(\mathcal {A}\) is coercive and as \(\mathcal {V}^{N,P}\subset \mathcal {V}\), the existence of a unique solution to Eq. 24 follows from Lax-Milgram theorem again, but applied this time on \(\mathcal {V}^{N,P}\). In order to prove the convergence result, we first notice that thanks to the Céa lemma we have for any \(N,P\in \mathbb {N}^{*}\)
where the constant \(\tilde {C}\) depends on \(\tilde {\lambda }\), \(\tilde {{\Lambda } }\), \(\tilde {M}\) and C(d,|D|). Second, we recall that \(\mathcal {V}^{N,P}\subset \mathcal {V}^{N+1,P+1}\) for any N,P, and that \(\bigcup _{N,P}\mathcal {V}^{N,P}\) is dense in \(\mathcal {V}\). This is sufficient in order to prove that \(\min _{v\in \mathcal {V}^{N,P}}||u-v||_{\mathcal {V}}\to 0\), as N,P →∞, and the result follows. □
Proof
(Proof of Theorem 3) By Corollary 1, \(\mathfrak {U}(x,\xi )\) is a classical solution of Eq. 16. Then, using Theorem 6.6 in Gilbarg and Trudinger (1983) we get that, for a.e. value of ξ
with C″ depending on d, M, |D| and λ, but not of ξ. But, as already noticed in the proof of Lemma 3, we have \(\sup _{x\in \bar {D}}|\mathfrak {U}(x,\xi )|\leq C\) with C depending on M, |D| and λ. Thus \(\mathfrak {U}(x,\xi )\) belongs to the space \(\mathcal {V}\).
As noticed in Section 3.2 any classical solution of Eq. 16 is a classical solution of Eq. 20). Let \(v\in {C^{1}_{c}}(D)\otimes L^{2}(\Xi ,\mathbb{P} _{\xi })\). Multiplying the first line of Eq. 20 by v, integrating over D ×Ξ against dx ⊗ℙξ(dz), and performing integration by parts w.r.t the variable x, we get that \(\mathcal {A}(\mathfrak {U},v)=F(v)\). Thanks to Eq. 1 in (A3), the coefficients \(\tilde {a}\) and \(\tilde {b}\) are obviously bounded uniformly in x and z, because D is bounded, and we have therefore (A7) and the continuity of the form \(\mathcal {A}\). Using then density arguments we get \(\mathcal {A}(\mathfrak {U},v)=F(v)\) for all \(v\in \mathcal {V}\). (A8) is a consequence of (A4), and we are under (A9). Thus we may get the approximation result by applying Lemma 5. □
Proof
(Proof of Lemma 6) In the proof of Lemma 5 we have seen that (A7) and the boundedness of \(\tilde {b}\) imply that the form \(\mathcal {A}\) is continuous on \(\mathcal {V}\).
We now prove that we can find μ ≥ 0 large enough such that the form \(\mathcal {A}_{\mu }:(u,v)\mapsto \mathcal {A}(u,v)+\mu \langle u,v\rangle _{\mathcal {H}}\) is coercive on \(\mathcal {V}\). Let us denote \(\tilde {M}\) the upper bound for \(\tilde {b}\) (i.e. \(|\tilde {b}(x,z)|\leq \tilde {M}\) for any \(x\in \bar {D}\), z ∈Ξ). Proceeding as for Eq. 41 (that we integrate over Ξ against ℙξ) we get that for any \(v\in \mathcal {V}\)
Note that if \(\tilde {M}<\frac {\tilde {\lambda }}{\sqrt {2C(d,|D|)}}\) it suffices to take μ = 0 (see the proof of Lemma 5). If this is not the case we take \(\mu >\frac {\tilde {M}^{2}}{\tilde {\lambda }}\) and set \(c=\min \left (\frac {\tilde {\lambda }}{2},\mu -\frac {\tilde {M}^{2}}{\tilde {\lambda }}\right )\). We then have \(\mathcal {A}_{\mu }(v,v)\geq c||v||^{2}_{\mathcal {V}}\) for any \(v\in \mathcal {V}\).
This is sufficient to prove the result (see Lemma 7.1.2 and Theorem 7.1.4 in Ern and Guermond 2002; see also the earlier reference Lions and Magenes 1972). □
Proof
(Proof of Lemma 7) By Corollary 1, \(\mathfrak {V}(t,x,\xi )\) is a classical solution of Eq. 17. Using Theorem 5.14 in Lieberman (1996) we get that for a.e. value of ξ
is bounded above by some constant C that depends on M, |D|, λ and d but not on ξ. We use that the coefficients a and b are Lipschitz continuous and that the initial condition f is in the H ”older space H2+α for α = 1; see pp 46–47 of Lieberman (1996) for more details on Hölder spaces and norms in the parabolic setting. The point is that here we can write these properties of a, b and f with quantities that are uniform w.r.t. the uncertain parameter ξ. Thus the function \(\mathfrak {V}(t,x,\xi )\) belongs to the space \(L^{2}(0,T;\mathcal {V})\cap H^{1}(0;T,\mathcal {V}^{\prime })\). As it is continuous in time and bounded (in particular uniformly w.r.t. ξ) it is also in \(C([0,T];\mathcal {H})\).
As noticed in Section 3.3.1 the function \(\mathfrak {V}(t,x,\xi )\) is a classical solution of Eq. 28. Multiplying the first line of Eq. 28 by a test function in \({C^{1}_{c}}(D)\otimes L^{2}(\Xi ,\mathbb{P} _{\xi })\), integrating over D ×Ξ against dx ⊗ℙξ(dz), and using density arguments we get
The result is proved. □
Proof
(Proof of Lemma 8) Remember that in the proof of Lemma 6 we have seen that we can find μ ≥ 0 large enough such that the form \(\mathcal {A}_{\mu }\) is coercive. Here for simplicity we assume that μ = 0, i.e. \(\mathcal {A}\) is coercive (with constant c). We claim that this is without loss of generality. Indeed if μ > 0 we consider vμ the solution of
(the solution exists because the form \(\mathcal {A}_{\mu }\) is continuous and coercive). Then it is obvious that the function v(t,⋅) = eμtvμ(t,⋅) solves Eq. 29 (in fact Eqs. 29 and 42 are equivalent via the change of variable). Thus one may approach vμ by the 𝜃-scheme and get an approximation of v applying again the change of variable.
Then for any N,P the form \(\mathcal {A}\) defines a scalar product on \(\mathcal {V}^{N,P}\), so that for any \(u\in \mathcal {V}\) there is by the Riesz theorem an element πN,Pu of \(\mathcal {V}^{N,P}\) such that
The application πN,P is linear and continuous from \(\mathcal {V}\) to \(\mathcal {V}^{N,P}\) and is called the elliptic projection operator. It satisfies
Then, following the proof of Céa’s Lemma (Theorem 3.1-2 in Raviart and Thomas 2004) one can prove that
where again \(\tilde {C}\) depends on the continuity and coercivity constants of \(\mathcal {A}\). Using the monotonicity and density assumptions on the \(\mathcal {V}^{N,P}\)’s we thus get that
In the sequel we note π for πN,P in order to lighten notations. For any \(y\in C(0,T;\mathcal {H})\), t ∈ [0,T] we denote y(t) := y(t,⋅,⋅).
Then for any 0 ≤ m ≤ M we set
(here we follow for example Raviart and Thomas (2004) §77.5). Note that Eq. 33 is equivalent to: ∀0 ≤ m ≤ M − 1, \(\forall w\in \mathcal {V}^{N,P}\),
so that by some algebraic computations we see that: ∀0 ≤ m ≤ M − 1, \(\forall w\in \mathcal {V}^{N,P}\),
where \(\varepsilon ^{m}_{N,P}\in \mathcal {V}^{\prime }\) is defined for any 0 ≤ m ≤ M − 1 by: \(\forall w\in \mathcal {V}\),
In the sequel we denote em (resp. ε) for \(e^{m}_{N,P}\) (resp. \(\varepsilon ^{m}_{N,P}\)).
We now aim at showing that for any \(\theta \in [\frac {1}{2},1]\) we have the stability result
Then we will aim at controlling the \(||\varepsilon ^{k}||_{\mathcal {V}^{\prime }}^{2}\)’s (consistency result; we will only treat the case \(\theta =\frac {1}{2}\)). This will allow to get convergence.
Stability. Here we adapt the energy estimate method to be found pp66–67 in Dautray and Lions (1993). Let \(\theta \in [\frac {1}{2},1]\) and fix 0 ≤ k ≤ M − 1. Taking \(w=\theta e^{k+1}+(1-\theta )e^{k}=:\bar {e}^{k}\) in Eq. 44 we get
Using then the algebraic equality
in Eq. 47 we get
and then
But for any \(f\in \mathcal {V}^{\prime }\) and any \(w\in \mathcal {V}\) we have, using Young’s inequality and the coercivity of \(\mathcal {A}\),
Using this in Eq. 48 we get
and then \( ||e^{k+1}||_{\mathcal {H}}^{2}- ||e^{k}||_{\mathcal {H}}^{2}\leq \frac {{\Delta } t}{c}||\varepsilon ^{k}||_{\mathcal {V}^{\prime }}^{2} \). Summing over 0 ≤ k ≤ m − 1 for any 1 ≤ m ≤ M we get Eq. 46.
Consistency. Here we follow the lines of the proof of Lemma 7.5-1 in Raviart and Thomas (2004). Using Eq. 29 and the fact that \(v\in C^{1}(0,T;\mathcal {V})\) we can rewrite Eq. 45 into: \(\forall w\in \mathcal {V}\),
and further: \(\forall w\in \mathcal {V}\),
From this we get that for any 0 ≤ m ≤ M − 1
with
We take now \(\theta =\frac {1}{2}\). Using a Taylor expansion with rest in integral form and the fact that \(v\in C^{3}(0,T;\mathcal {H})\) we get:
Then we have \(||\eta (t_{m})||_{\mathcal {H}}\leq C{\Delta } t{\int }_{t_{m}}^{t_{m+1}}\left |\left |\partial ^{3}_{t^{3}}v(s) \right |\right |_{\mathcal {H}} ds\) where C is some universal constant. To sum up we have for any 0 ≤ m ≤ M − 1
and, using Cauchy-Schwarz inequality,
Convergence. Let 0 ≤ m ≤ M. Using now \(||v^{m,N,P}-v(t_{m})||_{\mathcal {H}}\leq ||e^{m}||_{\mathcal {H}}+||v(t_{m})-{\Pi } v(t_{m})||_{\mathcal {H}}\) and Eqs. 46 and 49 we get
Recall now that we have \(||f^{N,P}-f||_{\mathcal {H}}\to 0\) as N,P →∞ and Eq. 43. If \(y\in C(0,T;\mathcal {V})\) the family (I −π)y(s), s ∈ [0,T], is equicontinuous, and we know thanks to Eq. 43 that for any s ∈ [0,T] we have \(||(I-{\Pi })y(s)||_{\mathcal {H}} \to 0\) as N,P →∞. With the help of Ascoli theorem we can see that we have then
Using this with y = v,∂tv, in Eq. 50 we get the announced convergence (note that it is in order (Δt)2 is time; cf Remark 8). □
Rights and permissions
About this article
Cite this article
Étoré, P., Prieur, C., Pham, D.K. et al. Global Sensitivity Analysis for Models Described by Stochastic Differential Equations. Methodol Comput Appl Probab 22, 803–831 (2020). https://doi.org/10.1007/s11009-019-09732-6
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11009-019-09732-6
Keywords
- Stochastic differential equations
- Sobol’ indices
- Feynman-Kac representation
- Polynomial chaos expansion
- Stochastic Galerkin projection