Abstract
In the simplest sequential decision problem for an ergodic stochastic process X, at each time n a decision u n is made as a function of past observations \(X_{0},\ldots,X_{n-1}\), and a loss l(u n , X n ) is incurred. In this setting, it is known that one may choose (under a mild integrability assumption) a decision strategy whose pathwise time-average loss is asymptotically smaller than that of any other strategy. The corresponding problem in the case of partial information proves to be much more delicate, however: if the process X is not observable, but decisions must be based on the observation of a different process Y, the existence of pathwise optimal strategies is not guaranteed. The aim of this paper is to exhibit connections between pathwise optimal strategies and notions from ergodic theory. The sequential decision problem is developed in the general setting of an ergodic dynamical system \((\varOmega,\mathcal{B},\mathbf{P},T)\) with partial information \(\mathcal{Y}\subseteq \mathcal{B}\). The existence of pathwise optimal strategies grounded in two basic properties: the conditional ergodic theory of the dynamical system, and the complexity of the loss function. When the loss function is not too complex, a general sufficient condition for the existence of pathwise optimal strategies is that the dynamical system is a conditional K-automorphism relative to the past observations \(\bigvee _{n\geq 0}T^{n}\mathcal{Y}\). If the conditional ergodicity assumption is strengthened, the complexity assumption can be weakened. Several examples demonstrate the interplay between complexity and ergodicity, which does not arise in the case of full information. Our results also yield a decision-theoretic characterization of weak mixing in ergodic theory, and establish pathwise optimality of ergodic nonlinear filters.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
- 2.
Non-dominated loss functions may also be of significant interest, see [24] for example. We will restrict attention to dominated loss functions, however, which suffice in many cases of interest.
- 3.
In [1, Appendix II.B] it is shown that under a continuity assumption on the loss function l, the optimal asymptotic loss in the full information setting is given by E[inf u E[l(u, X 1) | X 0, X −1, …]]. However, a counterexample is given of a discontinuous loss function for which this expression does not yield the optimal asymptotic loss. The key difference with the expression for L ⋆ given in Theorem 2.6 is that in the latter the essential infimum runs over \(u \in \mathbb{U}_{0}\), while it is implicit in [1] that the infimum in the above expression is an essential infimum over \(u \in \mathbb{U}_{-\infty,0}\). As the counterexample in [1] shows, these quantities need not coincide in the absence of continuity assumptions.
- 4.
- 5.
- 6.
As particle filters employ a random sampling mechanism, the strategy \(\mathbf{\tilde{u}}^{N}\) is technically speaking not admissible in the sense of this paper: Π k N (and therefore \(\tilde{u}_{k}^{N}\)) depends also on auxiliary sampling variables ξ 0, …, ξ k that are independent of Y 0, …, Y k . However, it is easily seen that all our results still hold when such randomized strategies are considered. Indeed, it suffices to condition on (ξ k ) k ≥ 0, so that all our results apply immediately under the conditional distribution.
References
P.H. Algoet, The strong law of large numbers for sequential decisions under uncertainty. IEEE Trans. Inform. Theory 40(3), 609–633 (1994)
A. Bellow, V. Losert, The weighted pointwise ergodic theorem and the individual ergodic theorem along subsequences. Trans. Am. Math. Soc. 288(1), 307–345 (1985)
D. Berend, V. Bergelson, Mixing sequences in Hilbert spaces. Proc. Am. Math. Soc. 98(2), 239–246 (1986)
O. Cappé, E. Moulines, T. Rydén, Inference in Hidden Markov Models (Springer, New York, 2005)
P. Chigansky, R. van Handel, A complete solution to Blackwell’s unique ergodicity problem for hidden Markov chains. Ann. Appl. Probab. 20(6), 2318–2345 (2010)
J.P. Conze, Convergence des moyennes ergodiques pour des sous-suites. In: Contributions au calcul des probabilités. Bull. Soc. Math. France, Mém. No. 35 (Soc. Math. France, Paris, 1973), pp. 7–15
D. Crisan, B. Rozovskiĭ (eds.), The Oxford Handbook of Nonlinear Filtering (Oxford University Press, Oxford, 2011)
P. Del Moral, M. Ledoux, Convergence of empirical processes for interacting particle systems with applications to nonlinear filtering. J. Theor. Probab. 13(1), 225–257 (2000)
C. Dellacherie, P.A. Meyer, Probabilities and Potential. C (North-Holland, Amsterdam, 1988)
R.M. Dudley, Uniform Central Limit Theorems (Cambridge University Press, Cambridge, 1999)
N. Etemadi, An elementary proof of the strong law of large numbers. Z. Wahrsch. Verw. Gebiete 55(1), 119–122 (1981)
A. Grothendieck, Produits tensoriels topologiques et espaces nucléaires. Mem. Am. Math. Soc. 1955(16), 140 (1955)
P.R. Halmos, In general a measure preserving transformation is mixing. Ann. Math. (2) 45, 786–792 (1944)
R. van Handel, The stability of conditional Markov processes and Markov chains in random environments. Ann. Probab. 37(5), 1876–1925 (2009)
R. van Handel, Uniform time average consistency of Monte Carlo particle filters. Stoch. Process. Appl. 119(11), 3835–3861 (2009)
R. van Handel, On the exchange of intersection and supremum of σ-fields in filtering theory. Isr. J. Math. 192, 763–784 (2012)
van Handel, R.: The universal Glivenko-Cantelli property. Probab. Theor. Relat. Fields 155(3–4), 911–934 (2013)
J. Hoffmann-Jørgensen, Uniform Convergence of Martingales. In: Probability in Banach spaces, 7 (Oberwolfach, 1988), Progr. Probab., vol. 21 (Birkhäuser Boston, Boston, 1990), pp. 127–137
O. Kallenberg, Foundations of Modern Probability, 2nd edn. (Springer, New York, 2002)
H. Kunita, Asymptotic behavior of the nonlinear filtering errors of Markov processes. J. Multivariate Anal. 1, 365–393 (1971)
T. Lindvall, Lectures on the Coupling Method (Dover Publications, Mineola, 2002). Corrected reprint of the 1992 original
S. Meyn, R.L. Tweedie, Markov Chains and Stochastic Stability, 2nd edn. (Cambridge University Press, Cambridge, 2009)
J. Neveu, Discrete-Parameter Martingales (North-Holland, Amsterdam, 1975)
A.B. Nobel, On optimal sequential prediction for general processes. IEEE Trans. Inform. Theory 49(1), 83–98 (2003)
D. Pollard, A User’S Guide to Measure Theoretic Probability (Cambridge University Press, Cambridge, 2002)
D.J. Rudolph, Pointwise and L 1 mixing relative to a sub-sigma algebra. Ill. J. Math. 48(2), 505–517 (2004)
X.T. Tong, R. van Handel, Conditional ergodicity in infinite dimension (2012). Preprint
H. Totoki, On a class of special flows. Z. Wahrscheinlichkeitstheorie und Verw. Gebiete 15, 157–167 (1970)
A.W. van der Vaart, J.A. Wellner, Weak Convergence and Empirical Processes (Springer, New York, 1996)
V.A. Volkonskiĭ, Y.A. Rozanov, Some limit theorems for random functions. I. Theor. Probab. Appl. 4, 178–197 (1959)
P. Walters, An Introduction to Ergodic Theory (Springer, New York, 1982)
T. Weissman, N. Merhav, Universal prediction of random binary sequences in a noisy environment. Ann. Appl. Probab. 14(1), 54–89 (2004)
H. von Weizsäcker, Exchanging the order of taking suprema and countable intersections of σ-algebras. Ann. Inst. H. Poincaré Sect. B (N.S.) 19(1), 91–100 (1983)
Acknowledgements
This work was partially supported by NSF grant DMS-1005575.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2014 Springer International Publishing Switzerland
About this chapter
Cite this chapter
van Handel, R. (2014). Ergodicity, Decisions, and Partial Information. In: Donati-Martin, C., Lejay, A., Rouault, A. (eds) Séminaire de Probabilités XLVI. Lecture Notes in Mathematics(), vol 2123. Springer, Cham. https://doi.org/10.1007/978-3-319-11970-0_18
Download citation
DOI: https://doi.org/10.1007/978-3-319-11970-0_18
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-11969-4
Online ISBN: 978-3-319-11970-0
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)