Skip to main content
Log in

Importance sampling for a robust and efficient multilevel Monte Carlo estimator for stochastic reaction networks

  • Published:
Statistics and Computing Aims and scope Submit manuscript

Abstract

The multilevel Monte Carlo (MLMC) method for continuous-time Markov chains, first introduced by Anderson and Higham (SIAM Multiscal Model Simul 10(1):146–179, 2012), is a highly efficient simulation technique that can be used to estimate various statistical quantities for stochastic reaction networks, in particular for stochastic biological systems. Unfortunately, the robustness and performance of the multilevel method can be affected by the high kurtosis, a phenomenon observed at the deep levels of MLMC, which leads to inaccurate estimates of the sample variance. In this work, we address cases where the high-kurtosis phenomenon is due to catastrophic coupling (characteristic of pure jump processes where coupled consecutive paths are identical in most of the simulations, while differences only appear in a tiny proportion) and introduce a pathwise-dependent importance sampling (IS) technique that improves the robustness and efficiency of the multilevel method. Our theoretical results, along with the conducted numerical experiments, demonstrate that our proposed method significantly reduces the kurtosis of the deep levels of MLMC, and also improves the strong convergence rate from \(\beta =1\) for the standard case (without IS), to \(\beta =1+\delta \), where \(0<\delta <1\) is a user-selected parameter in our IS algorithm. Due to the complexity theorem of MLMC, and given a pre-selected tolerance, \(\text {TOL}\), this results in an improvement of the complexity from \({\mathcal {O}}\left( \text {TOL}^{-2} \log (\text {TOL})^2\right) \) in the standard case to \({\mathcal {O}}\left( \text {TOL}^{-2}\right) \), which is the optimal complexity of the MLMC estimator. We achieve all these improvements with a negligible additional cost since our IS algorithm is only applied a few times across each simulated path.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5
Fig. 6
Fig. 7
Fig. 8
Fig. 9
Fig. 10
Fig. 11
Fig. 12
Fig. 13
Fig. 14
Fig. 15
Fig. 16
Fig. 17
Fig. 18
Fig. 19

Similar content being viewed by others

Notes

  1. Hereafter, we use \(\text {Prob}\left( A;B\right) \) and \(\mathrm {E}\left[ A;B\right] \) to denote the conditional probability and conditional expectation of A given B, respectively.

  2. \(\alpha _{j,i}\) molecules of the species \(S_i\) are consumed and \(\beta _{j,i}\) are produced. Thus, \((\alpha _{j,i},\beta _{j,i}) \in {\mathbb {N}}^2\) but \(\beta _{j,i}-\alpha _{j,i}\), can be a negative integer, constituting the vector \(\varvec{\nu }_j=\left( \beta _{j,1}-\alpha _{j,1},\dots ,\beta _{j,d}-\alpha _{j,d}\right) \in {\mathbb {Z}}^d\).

  3. We refer to Li (2007) for the underlying assumptions and proofs of this statement, in the context of the TL scheme.

  4. We set \(L_0=0\) unless otherwise stated. In our numerical experiments, we select \(L_0\) such that \(\mathrm {Var}\left[ g_{L_0+1}{-}g_{L_0}\right] \ll \mathrm {Var}\left[ g_{L_0}\right] \), in order to ensure the stability of the variance of the coupled paths of our MLMC estimator.

References

  • Abdulle, A., Yucheng, H., Li, T.: Chebyshev methods with discrete noise: the \(\tau \)-rock methods. J. Comput. Math. (2010). https://doi.org/10.4208/jcm.2009.10-m1004

    Article  MathSciNet  MATH  Google Scholar 

  • Ahn, T.-H., Sandu, A., Han, X.: Implicit simulation methods for stochastic chemical kinetics (2013). arXiv preprintarXiv:1303.3614

  • Anderson, D., Higham, D.: Multilevel Monte Carlo for continuous Markov chains, with applications in biochemical kinetics. SIAM Multiscal Model. Simul. 10(1), 146–179 (2012)

    Article  MathSciNet  MATH  Google Scholar 

  • Anderson, D.F.: A modified next reaction method for simulating chemical systems with time dependent propensities and delays. J. Chem. Phys. 127(21), 214107 (2007)

    Article  Google Scholar 

  • Anderson, D.F., Ganguly, A., Kurtz, T.G., et al.: Error analysis of tau-leap simulation methods. Ann. Appl. Probab. 21(6), 2226–2262 (2011)

    Article  MathSciNet  MATH  Google Scholar 

  • Anderson, D.F., Kurtz, T.G.: Stochastic Analysis of Biochemical Systems, vol. 1. Springer, Berlin (2015)

    Book  MATH  Google Scholar 

  • Aparicio, J.P., Solari, H.G.: Population dynamics: Poisson approximation and its relation to the langevin process. Phys. Rev. Lett. 86(18), 4183 (2001)

    Article  Google Scholar 

  • Bayer, C., Hammouda, C.B., Tempone, R.: Numerical smoothing and hierarchical approximations for efficient option pricing and density estimation (2020). arXiv preprint arXiv:2003.05708

  • Brauer, F., Castillo-Chavez, C.: Mathematical Models in Population Biology and Epidemiology, vol. 40. Springer, Berlin (2001)

    Book  MATH  Google Scholar 

  • Cao, Y., Petzold, L.: Trapezoidal tau-leaping formula for the stochastic simulation of biochemical systems. In: Proceedings of Foundations of Systems Biology in Engineering (FOSBE 2005), pp. 149–152 (2005)

  • Cao, Y., Liang, J.: Adaptively biased sequential importance sampling for rare events in reaction networks with comparison to exact solutions from finite buffer dCME method. J. Chem. Phys. 139(2), 07B605_1 (2013)

    Article  Google Scholar 

  • Çınlar, E.: Probability and Stochastics, vol. 261. Springer, Berlin (2011)

    Book  MATH  Google Scholar 

  • Cliffe, K.A., Giles, M.B., Scheichl, R., Teckentrup, A.L.: Multilevel Monte Carlo methods and applications to elliptic PDEs with random coefficients. Comput. Vis. Sci. 14(1), 3 (2011)

    Article  MathSciNet  MATH  Google Scholar 

  • Daigle Jr, B.J., Roh, M.K., Gillespie, D.T., Petzold, L.R.: Automated estimation of rare event probabilities in biochemical systems. J. Chem. Phys. 134(4), 01B628 (2011)

    Google Scholar 

  • Duffie, D., Glynn, P., et al.: Efficient Monte Carlo simulation of security prices. Ann, Appl. Probab. 5(4), 897–905 (1995)

    Article  MathSciNet  MATH  Google Scholar 

  • Engblom, S.: On the stability of stochastic jump kinetics (2012). arXiv preprint arXiv:1202.3892

  • Ethier, S.N., Kurtz, T.G.: Markov Processes: Characterization and Convergence. Wiley Series in Probability and Mathematical Statistics. Wiley, New York (1986)

    Book  Google Scholar 

  • Giles, M.B.: Multilevel Monte Carlo path simulation. Oper. Res. 56(3), 607–617 (2008)

    Article  MathSciNet  MATH  Google Scholar 

  • Giles, M.B.: Multilevel Monte Carlo methods. Acta Numerica 24, 259–328 (2015)

    Article  MathSciNet  MATH  Google Scholar 

  • Gillespie, D.T.: A general method for numerically simulating the stochastic time evolution of coupled chemical reactions. J. Comput. Phys. 22(4), 403–434 (1976)

    Article  MathSciNet  Google Scholar 

  • Gillespie, D.T.: The chemical Langevin equation. J. Chem. Phys. 113(1), 297–306 (2000)

    Article  Google Scholar 

  • Gillespie, D.T.: Approximate accelerated stochastic simulation of chemically reacting systems. J. Chem. Phys. 115(4), 1716–1733 (2001). Please provide complete details for reference Gou (2016)

    Article  Google Scholar 

  • Gou, W.: Estimating value-at-risk using multilevel Monte Carlo maximum entropy method. Master’s thesis, University of Oxford (2016)

  • Gupta, A., Briat, C., Khammash, M.: A scalable computational framework for establishing long-term behavior of stochastic reaction networks. PLoS Comput. Biol. 10(6), e1003669 (2014)

    Article  Google Scholar 

  • Hammouda, C.B., Moraes, A., Tempone, R.: Multilevel hybrid split-step implicit tau-leap. Numer. Algorithms 74(2), 527–560 (2017)

    Article  MathSciNet  MATH  Google Scholar 

  • Hensel, S.C., Rawlings, J.B., Yin, J.: Stochastic kinetic modeling of vesicular stomatitis virus intracellular growth. Bull. Math. Biol. 71(7), 1671–1692 (2009)

    Article  MathSciNet  MATH  Google Scholar 

  • Kebaier, A., et al.: Statistical Romberg extrapolation: a new variance reduction method and applications to option pricing. Ann. Appl. Probab. 15(4), 2681–2705 (2005)

    Article  MathSciNet  MATH  Google Scholar 

  • Kuwahara, H., Mura, I.: An efficient and exact stochastic simulation method to analyze rare events in biochemical systems. J. Chem. Phys. 129(16), 10B619 (2008)

    Article  Google Scholar 

  • Lester, C., Baker, R.E., Giles, M.B., Yates, C.A.: Extending the multi-level method for the simulation of stochastic biological systems. Bull. Math. Biol. 78(8), 1640–1677 (2016)

    Article  MathSciNet  MATH  Google Scholar 

  • Lester, C., Yates, C.A., Baker, R.E.: Robustly simulating biochemical reaction kinetics using multi-level Monte Carlo approaches. J. Comput. Phys. 375, 1401–1423 (2018)

    Article  MathSciNet  MATH  Google Scholar 

  • Lester, C., Yates, C.A., Giles, M.B., Baker, R.E.: An adaptive multi-level simulation algorithm for stochastic biological systems. J. Chem. Phys. 142(2), 01B612_1 (2015)

    Article  Google Scholar 

  • Li, T.: Analysis of explicit tau-leaping schemes for simulating chemically reacting systems. Multiscale Model. Simul. 6(2), 417–436 (2007)

    Article  MathSciNet  MATH  Google Scholar 

  • Moraes, A., Tempone, R., Vilanova, P.: A multilevel adaptive reaction-splitting simulation method for stochastic reaction networks. SIAM J. Sci. Comput. 38(4), A2091–A2117 (2016)

    Article  MathSciNet  MATH  Google Scholar 

  • Moraes, A., Tempone, R., Vilanova, P.: Multilevel hybrid Chernoff tau-leap. BIT Numer. Math. 56(1), 189–239 (2016)

    Article  MathSciNet  MATH  Google Scholar 

  • Rao, C.V., Arkin, A.P.: Stochastic chemical kinetics and the quasi-steady-state assumption: application to the Gillespie algorithm. J. Chem. Phys. 118(11), 4999–5010 (2003)

    Article  Google Scholar 

  • Rathinam, M.: Moment growth bounds on continuous time Markov processes on non-negative integer lattices (2013). arXiv preprint arXiv:1304.5169

  • Rathinam, M., El Samad, H.: Reversible-equivalent-monomolecular tau: a leaping method for “small number and stiff” stochastic chemical systems. J. Comput. Phys. 224(2), 897–923 (2007)

    Article  MathSciNet  MATH  Google Scholar 

  • Srivastava, R., You, L., Summers, J., Yin, J.: Stochastic vs. deterministic modeling of intracellular viral kinetics. J. Theor. Biol. 218(3), 309–321 (2002)

    Article  MathSciNet  Google Scholar 

Download references

Acknowledgements

This work was supported by the KAUST Office of Sponsored Research (OSR) under Award No. URF/1/2584-01-01 and the Alexander von Humboldt Foundation. C. Ben Hammouda and R. Tempone are members of the KAUST SRI Center for Uncertainty Quantification in Computational Science and Engineering. The authors would like to thank Dr. Alvaro Moraes and Sophia Franziska Wiechert for their helpful and constructive comments. The authors are also very grateful to the anonymous referees for their valuable comments and suggestions that greatly contributed to shape the final version of the work.

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Chiheb Ben Hammouda.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendices

Appendix A: Proofs of Lemma 4.1 and Theorems 4.1 and 4.2

Proof of Lemma 4.1

We denote by \(K=\sum _{n \in S} k_n\), \(L_{\ell }\left( j\right) \) the likelihood evaluated at \(K=j\). Then, for \(p\ge 1\) and \(0\le \delta <1\), and using relation (4.4), we write

$$\begin{aligned}&\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left|\Delta g_{\ell }\right|^p(T) L_{\ell }^p ; \left( {\mathcal {F}}_{N_{\ell }-1}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right] = \sum _{\underset{ i \in {\mathbb {N}}\setminus \{0\}}{\left|\Delta g_{\ell }(T)\right|=K =i}} i ^p L_{\ell }^p \left( i\right) {\overline{\pi }}_{\ell } \left( \left|\Delta g_{\ell }(T)\right|=i, K= i ; \left( {\mathcal {F}}_{N_{\ell }-1}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right) \nonumber \\&\qquad + \sum _{\underset{i \ne j, \, (i,j) \in {\mathbb {N}}^2}{\left|\Delta g_{\ell }(T)\right|=i,\, K =j}} i ^p L_{\ell }^p \left( j\right) {\overline{\pi }}_{\ell } \left( \left|\Delta g_{\ell }(T)\right|=i, K= j ; \left( {\mathcal {F}}_{N_{\ell }-1}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right) \nonumber \\&\quad = \sum _{\underset{ i \in {\mathbb {N}}\setminus \{0\}}{\left|\Delta g_{\ell }\right|=K =i}} \underset{A_{i}}{\underbrace{ i ^p e^{p(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} \Delta t_{\ell }^{i p \delta } {\overline{\pi }}_{\ell } \left( \left|\Delta g_{\ell }(T)\right|=i, K= i ; \left( {\mathcal {F}}_{N_{\ell }-1}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right) }}\nonumber \\&\qquad + \sum _{\underset{\underset{i \ne j, \, (i,j) \in {\mathbb {N}}^2}{K =j}}{\left|\Delta g_{\ell }(T)\right|=i}} \underset{B_{ij}}{\underbrace{i^p e^{p(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} \Delta t_{\ell }^{j p \delta } {\overline{\pi }}_{\ell } \left( \left|\Delta g_{\ell }(T)\right|=i, K= j ; \left( {\mathcal {F}}_{N_{\ell }-1}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right) }} =\sum _{\underset{ i \in {\mathbb {N}}\setminus \{0\}}{\left|\Delta g_{\ell }\right|=K =i}} A_i + \sum _{\underset{i \ne j, \, (i,j) \in {\mathbb {N}}^2}{\left|\Delta g_{\ell }\right|=i,\, K =j}} B_{ij}. \end{aligned}$$
(A.1)

Using Assumption 4.1(a), we have

$$\begin{aligned} 0 \le \frac{\sum _{\underset{ i \in {\mathbb {N}}\setminus \{0,1\}}{\left|\Delta g_{\ell }(T)\right|=K =i}} A_i}{A_1} \le \underset{\underset{\Delta t_{\ell } \rightarrow 0}{\longrightarrow 0 }}{\underbrace{\sum _{i \in {\mathbb {N}}\setminus \{0,1\}} i ^{p} \Delta t_{\ell }^{(i-1) p \delta }}}. \end{aligned}$$
(A.2)

Now, let us examine the second sum in the right-hand side of (A.1). First, observe that \(B_{0j}=0, \, \forall j \ge 1\) and \(B_{i0}=0, \, \forall i \ge 1\). Although the first observation is clear, we need to explain the second observation, which is mainly due to the fact that \({\overline{\pi }}\left( \left|\Delta g_{\ell }(T)\right|=i, K= 0 ; \left( {\mathcal {F}}_{N_{\ell }-1}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right) =0\), \(\forall i \ge 1\). For the purpose of simplification, let us consider \(g_{\ell }={\overline{X}}_{\ell }\); then considering the first interval in the coarse level, and using the coupling equation (2.6), we have: (i) At \(t=0\): \({\overline{X}}_{\ell }(0)={\overline{X}}_{\ell -1}(0)\) and \(\Delta a^1_{\ell -1,0}=0\). (ii) At \(t=\Delta t_{\ell }\): \({\overline{X}}_{\ell }(\Delta t_{\ell })={\overline{X}}_{\ell -1}(\Delta t_{\ell })\) and \(\Delta a^2_{\ell -1,0}=a({\overline{X}}_{\ell }(\Delta t_{\ell }))-a({\overline{X}}_{\ell -1}(0))\). (iii) At \(t=t_1=2\Delta t_{\ell }\): if \(\Delta a^2_{\ell -1,0}=0\), then we simulate this step under the old measure and consequently we will have \({\overline{X}}_{\ell }(t_1)={\overline{X}}_{\ell -1}(t_1)\) otherwise if \(\Delta a^2_{\ell -1,0} \ne 0\), then we simulate this step under the IS measure, but since \(j=0\), then we will have \({\overline{X}}_{\ell }(t_1)={\overline{X}}_{\ell -1}(t_1)\). Therefore, in both scenarios, we will have the same situation at the start, \(t_0=0\). Therefore, we conclude that \({\overline{\pi }}\left( \left|\Delta g_{\ell }(T)\right|=i, K= 0 ; \left( {\mathcal {F}}_{N_{\ell }-1}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right) =0\), \(\forall \, i \ge 1\) and \(B_{i0}=0\), \(\forall \, i \ge 1\).

Then, using Assumptions 4.1(b) and 4.1(c), we obtain

$$\begin{aligned}&0 \le \frac{\sum _{\underset{i \ne j, \, (i,j) \in {\mathbb {N}}^2}{\left|\Delta g_{\ell } (T)\right|=i,\, K =j}} B_{ij}}{A_1} \nonumber \\&\quad \le \frac{ \sum _{i \ne j, \, 1 \le i,j} i^p \Delta t_{\ell }^{j p\delta } {\overline{\pi }}_{\ell } \left\{ \left|\Delta g_{\ell } (T)\right|= i ; \left( {\mathcal {F}}_{N_{\ell }-1}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right\} }{\left( \Delta t_{\ell }^{p \delta } \right) e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*}\right) \left( 1+{o}\left( 1\right) \right) }\nonumber \\&\quad \le \frac{ \sum _{i \ne j, \, 1 \le i,j} \eta _{i,\ell } i^p \Delta t_{\ell }^{j p\delta } \Delta t_{\ell }^{i (1-\delta )} }{\left( \Delta t_{\ell }^{p \delta } \right) e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*}\right) \left( 1+{o}\left( 1\right) \right) }\nonumber \\&\quad = \underset{\underset{\Delta t_{\ell } \rightarrow 0}{\longrightarrow 0}}{\underbrace{ \left( 1+{o}\left( 1\right) \right) ^{-1} \left( e^{ \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \Delta a_{\ell ,n^*}^{-1} \sum _{i \ne j, \, 1 \le i,j} \eta _{i,\ell } i^p \Delta t_{\ell }^{p\delta ( j-1)} \Delta t_{\ell }^{(1-\delta ) ( i-1)} \right) }}. \end{aligned}$$
(A.3)

Therefore, using (A.1), (A.2) and (A.3), we conclude Lemma 4.1, that is

$$\begin{aligned}&\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left|\Delta g_{\ell }\right|^p(T) L_{\ell }^p ; \left( {\mathcal {F}}_{N_{\ell }-1}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right] \\&\quad = \Delta t_{\ell }^{(p -1)\delta +1} \left( \Delta a_{\ell ,n^*}\right) e^{p(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} \\&\qquad e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \left( 1+h_{p,\ell }\right) , \end{aligned}$$

such that \(h_{p,\ell } \underset{\Delta t_{\ell } \rightarrow 0}{\longrightarrow 0}\). \(\square \)

Proof of Theorem 4.1

Let \(0 \le \delta < 1\). In the first step of the proof, we want to show that

$$\begin{aligned} \kappa _{\ell }:=\frac{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left( Y_{\ell }-\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^4\right] }{\left( \text {Var}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^2} \underset{\Delta t_{\ell } \rightarrow 0}{\sim } \frac{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] }{\left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] \right) ^2} \end{aligned}$$

Let us first show that \(\text {Var}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \underset{\Delta t_{\ell } \rightarrow 0}{\sim } \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] \). In fact,

$$\begin{aligned} \frac{\text {Var}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] }{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] }=\frac{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] - \left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^2}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] }=1- \frac{ \left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^2}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] }. \end{aligned}$$

Therefore, we need to show that \(I_1:=\frac{ \left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^2}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] } \underset{\Delta t_{\ell }\rightarrow 0}{\longrightarrow }0\).

Due to the order one weak error convergence, there exists a constant \(d_1>0\) such that \(\left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) \le d_1 \Delta t_{\ell }\). Therefore, using Lemma 4.1 and Assumption 4.2, we obtain

$$\begin{aligned}&0 \le I_1\le \frac{d_1^2 \Delta t_{\ell }^2}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] }\\&\quad = \frac{d_1^2 \Delta t_{\ell }^2}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2; \left( {\mathcal {F}}_{N_{\ell }}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right] \right] }\\&\quad =\frac{d_1^2 \Delta t_{\ell }^2}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left( \Delta t_{\ell }^{\delta +1} \right) e^{2(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \left( \Delta a_{\ell ,n^*}\right) \left( 1+ h_{2,\ell }\right) \right] }\\&\quad \le \frac{d_1^2 \Delta t_{\ell }^{1-\delta }}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \Delta a_{\ell ,n^*} \right] }\le \frac{d_1^2 \Delta t_{\ell }^{1-\delta }}{C_1} \underset{\Delta t_{\ell }\rightarrow 0}{\longrightarrow }0. \end{aligned}$$

Therefore, we conclude that

$$\begin{aligned} \text {Var}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \underset{\Delta t_{\ell } \rightarrow 0}{\sim } \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] . \end{aligned}$$
(A.4)

Now, let us show that \(\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left( Y_{\ell }-\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^4\right] \underset{\Delta t_{\ell } \rightarrow 0}{\sim } \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] \). In fact,

$$\begin{aligned}&\frac{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left( Y_{\ell }-\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^4\right] }{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] }\\&\quad =\frac{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] -4 \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^3\right] \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] + 6 \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] \left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^2 -3 \left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^4}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] }\\&\quad =1-4\frac{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^3\right] \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] }{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] }+ 6\frac{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] \left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^2}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] }- 3 \frac{\left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^4}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] }\\&\quad =1-4 I_2+ 6 I_3- 3 I_4 \end{aligned}$$

Therefore, we need to show that \(I_2, I_3, I_4 \underset{\Delta t_{\ell }\rightarrow 0}{\longrightarrow }0\).

Using Lemma 4.1 and Assumption 4.2, we obtain

$$\begin{aligned}&0 \le I_4\le \frac{d_1^4 \Delta t_{\ell }^4}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] }\\&\quad = \frac{d_1^4 \Delta t_{\ell }^4}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4; \left( {\mathcal {F}}_{N_{\ell }}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right] \right] }\\&\quad =\frac{d_1^4 \Delta t_{\ell }^4}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left( \Delta t_{\ell }^{3\delta +1} \right) e^{4(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \left( \Delta a_{\ell ,n^*}\right) \left( 1+ h_{4,\ell }\right) \right] }\\&\quad \le \frac{d_1^4 \Delta t_{\ell }^{3(1-\delta )}}{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \Delta a_{\ell ,n^*} \right] }\le \frac{d_1^4 \Delta t_{\ell }^{3(1-\delta )}}{C_1}\underset{\Delta t_{\ell }\rightarrow 0}{\longrightarrow }0. \end{aligned}$$

Similarly for \(I_2\), using Lemma 4.1 and Assumptions 4.2 and 4.3, we obtain

$$\begin{aligned} 0 \le I_2&\le \frac{d_1 \Delta t_{\ell } \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left|Y_{\ell }\right|^3\right] }{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] }\\&= \frac{d_1 \Delta t_{\ell } \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^3; \left( {\mathcal {F}}_{N_{\ell }}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right] \right] }{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4; \left( {\mathcal {F}}_{N_{\ell }}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right] \right] }\\&=\frac{d_1 \Delta t_{\ell } \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left( \Delta t_{\ell }^{2 \delta +1} \right) e^{3(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \left( \Delta a_{\ell ,n^*}\right) \left( 1+ h_{3,\ell }\right) \right] }{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left( \Delta t_{\ell }^{3 \delta +1} \right) e^{4(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} e^{- \left( \Delta a_{\ell ,n^*} \right) } \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*}\right) \left( 1+h_{4,\ell }\right) \right] }\\&\le \frac{d_1 \Delta t_{\ell }^{1-\delta } \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ e^{3(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} \Delta a_{\ell ,n^*}\left( 1+ h_{3,\ell }\right) \right] }{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \Delta a_{\ell ,n^*} \right] }\\&\le C \Delta t_{\ell }^{1-\delta } \underset{\Delta t_{\ell }\rightarrow 0}{\longrightarrow } 0. \end{aligned}$$
Fig. 20
figure 20

Example 5.1 with IS (with \(\delta =0.5\)): Histogram of \(\# \{n \in {\mathcal {S}}: k_n>0\}\), for number of samples \(M_{\ell }=10^5\). a \(\ell =6\). b \(\ell =10\)

Finally, for \(I_3\), using Lemma 4.1 and Assumptions 4.2 and 4.3, we obtain

$$\begin{aligned}&0 \le I_3\le \frac{d_1^2 \Delta t_{\ell }^2 \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left|Y_{\ell }\right|^2\right] }{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] }\\&\quad = \frac{d_1^2 \Delta t_{\ell }^2 \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2; \left( {\mathcal {F}}_{N_{\ell }}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right] \right] }{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4; \left( {\mathcal {F}}_{N_{\ell }}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right] \right] }\\&\quad =\frac{d_1^2 \Delta t_{\ell }^2 \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left( \Delta t_{\ell }^{ \delta +1} \right) e^{2(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \left( \Delta a_{\ell ,n^*}\right) \left( 1+ h_{2,\ell }\right) \right] }{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left( \Delta t_{\ell }^{3 \delta +1} \right) e^{4(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \left( \Delta a_{\ell ,n^*}\right) \left( 1+ h_{4,\ell }\right) \right] }\\&\quad \le \frac{d_1^2 \Delta t_{\ell }^{2(1-\delta )} \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ e^{2(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} \Delta a_{\ell ,n^*} \left( 1+ h_{2,\ell }\right) \right] }{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \Delta a_{\ell ,n^*} \right] }\\&\quad \le {\widetilde{C}} \Delta t_{\ell }^{2(1-\delta )} \underset{\Delta t_{\ell }\rightarrow 0}{\longrightarrow } 0. \end{aligned}$$
Fig. 21
figure 21

Example 5.2 with IS (with \(\delta =0.5\)): Histogram of \(\# \{ \text {IS steps : s.t.}\, {\overline{K}}:=\sum _{j \in {\mathcal {J}}_1} \sum _{n \in {\mathcal {S}}_j}k^j_n>0\}\), for number of samples \(M_{\ell }=10^5\). a \(\ell =4\). b \(\ell =8\)

Fig. 22
figure 22

Example 5.3 with IS (with \(\delta =0.5\)): Histogram of \(\# \{ \text {IS steps : s.t.}\, {\overline{K}}:=\sum _{j \in {\mathcal {J}}_1} \sum _{n \in {\mathcal {S}}_j}k^j_n>0\}\), for number of samples \(M_{\ell }=10^5\). a) \(\ell =6\). b) \(\ell =10\)

Therefore, we conclude that

$$\begin{aligned} \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left( Y_{\ell }-\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^4\right] \underset{\Delta t_{\ell } \rightarrow 0}{\sim } \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] . \end{aligned}$$
(A.5)

Finally, using (A.4), (A.5), Lemma 4.1 and Assumptions 4.2 and 4.3, we obtain

$$\begin{aligned}&\kappa _{\ell }:=\frac{ \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \left( Y_{\ell }-\mathrm {E}\left[ Y_{\ell }\right] \right) ^4\right] }{\left( \text {Var}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }\right] \right) ^2} \underset{\Delta t_{\ell } \rightarrow 0}{\sim } \frac{ \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4\right] }{\left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] \right) ^2}\nonumber \\&\quad =\frac{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^4(T) ; \left( {\mathcal {F}}_{N_{\ell }}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right] \right] }{\left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2(T) ; \left( {\mathcal {F}}_{N_{\ell }}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right] \right] \right) ^2} \nonumber \\&\quad = \Delta t_{\ell }^{\delta -1} \frac{\mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ e^{4(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \Delta a_{\ell ,n^*} \left( 1+ h_{4,\ell }\right) \right] }{\left( \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ e^{2(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \Delta a_{\ell ,n^*} \left( 1+ h_{2,\ell }\right) \right] \right) ^2} \nonumber \\&\quad ={\mathcal {O}}\left( \Delta t_{\ell }^{\delta -1}\right) . \end{aligned}$$
(A.6)

\(\square \)

Proof of Theorem 4.2

Let \(0< \delta < 1\). Then, using (A.4), Lemma 4.1 and Assumptions 4.2 and 4.3, we obtain

$$\begin{aligned}&\text {Var}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell } \right] \underset{\Delta t_{\ell } \rightarrow 0}{\sim } \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2\right] \nonumber \\&\quad = \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ \mathrm {E}_{{\overline{\pi }}_{\ell }}\left[ Y_{\ell }^2; \left( {\mathcal {F}}_{N_{\ell }-1}, {\mathcal {I}}^s_{\ell }={\mathcal {S}}\right) \right] \right] \nonumber \\&\quad =\left[ \left( \Delta t_{\ell }^{\delta +1} \right) e^{2(\Delta t_{\ell }^{1-\delta } -\Delta t_{\ell }) \sum _{n \in {\mathcal {S}}} \Delta a_{\ell ,n}} \right. \nonumber \\&\qquad \left. e^{- \left( \Delta t_{\ell }^{1-\delta } \Delta a_{\ell ,n^*} \right) } \left( \Delta a_{\ell ,n^*}\right) \left( 1+ h_{2,\ell }\right) \right] \nonumber \\&\quad ={\mathcal {O}}\left( \Delta t_{\ell }^{1+\delta }\right) . \end{aligned}$$
(A.7)

\(\square \)

Appendix B: Numerical evidence of Assumption 4.1

In Figs. 2021 and 22, we plot the histograms, for Examples 5.15.2 and 5.3, with \(\delta =0.5\), corresponding to \(\# \{ \text {IS steps : s.t.}\, {\overline{K}}:=\sum _{j \in {\mathcal {J}}_1} \sum _{n \in {\mathcal {S}}_j}k^j_n>0\}\), that is the number of times where we perform IS and succeeded to separate the two paths. These Figs. show that our assumption 4.1 (c) is valid since for small values of \(\Delta t_{\ell }\), we have at most one jump created by IS such that it separates the two paths.

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Ben Hammouda, C., Ben Rached, N. & Tempone, R. Importance sampling for a robust and efficient multilevel Monte Carlo estimator for stochastic reaction networks. Stat Comput 30, 1665–1689 (2020). https://doi.org/10.1007/s11222-020-09965-3

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s11222-020-09965-3

Keywords

Mathematics Subject Classification

Navigation