Advanced Search
Article Contents
Article Contents

A convergence analysis of the perturbed compositional gradient flow: Averaging principle and normal deviations

  • * Corresponding author: Wenqing Hu

    * Corresponding author: Wenqing Hu 
The first author is supported by an University of Missouri Research Board (UMRB) grant.
Abstract Full Text(HTML) Related Papers Cited by
  • We consider in this work a system of two stochastic differential equations named the perturbed compositional gradient flow. By introducing a separation of fast and slow scales of the two equations, we show that the limit of the slow motion is given by an averaged ordinary differential equation. We then demonstrate that the deviation of the slow motion from the averaged equation, after proper rescaling, converges to a stochastic process with Gaussian inputs. This indicates that the slow motion can be approximated in the weak sense by a standard perturbed gradient flow or the continuous-time stochastic gradient descent algorithm that solves the optimization problem for a composition of two functions. As an application, the perturbed compositional gradient flow corresponds to the diffusion limit of the Stochastic Composite Gradient Descent (SCGD) algorithm for minimizing a composition of two expected-value functions in the optimization literatures. For the strongly convex case, such an analysis implies that the SCGD algorithm has the same convergence time asymptotic as the classical stochastic gradient descent algorithm. Thus it validates, at the level of continuous approximation, the effectiveness of using the SCGD algorithm in the strongly convex case.

    Mathematics Subject Classification: 34C29, 60J60, 62L20, 90C30.


    \begin{equation} \\ \end{equation}
  • 加载中
  • [1] L. Arnold, Hasslemann's program revisited: The analysis of stochasticity in deterministic climate models, in Progress in Probability Book Series, Springer, 49, Stochastic Climate Models, 141-157.
    [2] V. I. Bakhtin, Asymptotics of superregular perturbations of fiber ergodic semigroups, Stochastics and Stochastic Reports, 75 (2003), 295-318.  doi: 10.1080/1045112031000155678.
    [3] V. Bakhtin and Y. Kifer, Diffusion approximation for slow motion in fully coupled averaging, Probability Theory and Related Fields, 129 (2004), 157-181.  doi: 10.1007/s00440-003-0326-7.
    [4] A. Benveniste, M. Metivier and P. Priouret, Adaptive algorithms and stochastic approximations, Applications of Mathematics, Springer, 22 (1990), xii+365pp. doi: 10.1007/978-3-642-75894-2.
    [5] V. S. Borkar, Stochastic Approximation: A Dynamical Systems Viewpoint, Cambridge University Press, 2008.
    [6] S. Cerrai, A Khasminskii's averaging principle for stochastic reaction-diffusion equations, Annals of Applied Probability, 19 (2009), 899-948.  doi: 10.1214/08-AAP560.
    [7] S. Cerrai, Normal deviations from the averaged motion for some reaction-diffusion equations with fast oscillating perturbation, Journal de Mathématiques Pures et Appliquées, 91 (2009), 614-647.  doi: 10.1016/j.matpur.2009.04.007.
    [8] Y. Ermoliev, Methods of Stochastic Programming. Monographs in Optimization and OR, Nauka, Moscow, 1976.
    [9] M. I. Freidlin, Functional Integration and Partial Differential Equations, Princeton University Press, Princeton, 1985. doi: 10.1515/9781400881598.
    [10] M. Freidlin and A. Wentzell, Random Perturbations of Dynamical Systems, 2$^{nd}$ Edition, Springer, 1998. doi: 10.1007/978-1-4612-0611-8.
    [11] M. Hairer and G. Pavliotis, Periodic homogenization for hypoelliptic diffusions, Journal of Statistical Physics, 117 (2004), 261-279.  doi: 10.1023/B:JOSS.0000044055.59822.20.
    [12] K. Hasselmann, Stochastic climate models, Part Ⅰ, Theory, Tellus, 28 (1976), 473-485. 
    [13] W. Hu and C. J. Li, On the fast convergence of random perturbations of the gradient flow, preprint, arXiv: 1706.00837
    [14] W. Hu, C. J. Li, L. Li and J. Liu, On the diffusion approximation of nonconvex stochastic gradient descent, Annals of Mathematical Science and Applications, to appear. arXiv: 1705.07562
    [15] R. Khasminskii, On stochastic processes defined by differential equations with a small parameter, Theory of Probability and its Applications, 11 (1966), 211-228. 
    [16] R. Khasminskii, On the principle of averaging the It ô's stochastic differential equations, Kybernetika (Prague), 4 (1968), 260-279. 
    [17] Y. Kifer, The exit problem for small random perturbations of dynamical systems with a hyperbolic fixed point, Israel Journal of Mathematics, 40 (1981), 74-96.  doi: 10.1007/BF02761819.
    [18] C. Kipnis and S. R. S. Varadhan, Central limit theorem for additive functionals of reversible Markov processes and applications to simple exclusions, Communications in Mathematical Physics, 104 (1986), 1-19.  doi: 10.1007/BF01210789.
    [19] H. Kushner and G. George Yin, Stochastic approximations and recursive algorithms and applications, Applications of Mathematics (Stochastic Modeling and Applied Probability), 35, 2$^{nd}$ Edition, Springer, 2003.
    [20] R. J. Muirhead, Aspects of Multivariate Statistical Theory, John Wiley & Sons, Inc., New York, 1982. doi: 10.1109/IGARSS.2010.5653437.
    [21] B. Oksendal, Stochastic Differential Equations, Springer-Verlag, 1992. doi: 10.1007/978-3-662-02847-6.
    [22] E. Pardoux and A. Yu. Veretennikov, On the Possion equation and diffusion approximation 1, Annals of Probability, 29 (2001), 1061-1085.  doi: 10.1214/aop/1015345596.
    [23] E. Pardoux and A. Yu. Veretennikov, On the Possion equation and diffusion approximation 2, Annals of Probability, 31 (2003), 1166-1192.  doi: 10.1214/aop/1055425774.
    [24] E. Pardoux and A. Yu. Veretennikov, On the Possion equation and diffusion approximation 3, Annals of Probability, 33 (2005), 1111-1133.  doi: 10.1214/009117905000000062.
    [25] D. Revuz and M. Yor, Continuous martingales and Brownian motion, Grundlehren der Mathematischen Wissenschaften, 293, 3rd edition, Springer-Verlag, Berlin, 1999, xiv+602 pp. doi: 10.1007/978-3-662-06400-9.
    [26] N. G. Van Kampen, The diffusion approximation for markov processes, Reprinted from: Thermodynamics & kinetics of biological processes (eds. I. Lamprecht and A. I. Zotin) Walter de Gruyter & Co., New York (1982). 181-195.
    [27] V. N. Vapnik, The Nature of Statistical Learning Theory, Springer, 1995. doi: 10.1007/978-1-4757-2440-0.
    [28] A. Yu. Veretennikov, On polynomial mixing and convergence rate for stochastic difference and differential equations, Theory of Probability and its Applications, 44 (2000), 361-374.  doi: 10.1137/S0040585X97977550.
    [29] M. WangE. X. Fang and H. Liu, Stochastic compositional gradient descent: Algorithms for minimizing compositions of expected-value functions, Mathematical Programming, 161 (2016), 419-449.  doi: 10.1007/s10107-016-1017-3.
    [30] M. Wang, J. Liu and E. X. Fang, Accelerating stochastic composition optimization, Advances in Neural Information Processing Systems, 2016. arXiv: 1607.07329
  • 加载中

Article Metrics

HTML views(1053) PDF downloads(255) Cited by(0)

Access History

Other Articles By Authors



    DownLoad:  Full-Size Img  PowerPoint