Abstract
Let \(X_1,X_2,\ldots ,X_n,\ldots \) be a discrete-time stochastic process. The following optimal stopping problem is considered. We observe \(X_1,X_2,\ldots \) on the one-by-one basis getting successively the data \(x_1,x_2,\ldots \). At each stage \(n\), \(n=1,2,\ldots \), after the data \(x_1,\ldots ,x_n\) have been observed, we may stop, and if we stop, we gain \(g_n(x_1,\ldots ,x_n)\). In this article, we characterize the structure of all optimal (randomized) stopping times \(\tau \) that maximize the average gain value \(G(\tau )=E g_\tau (X_1,\ldots , X_\tau )\) in some natural classes of stopping times \(\tau \) we call truncatable: \(\tau \) is called truncatable if \(G(\tau \wedge N)\rightarrow G(\tau )\) as \(N\rightarrow \infty \). It is shown that under some additional conditions on the structure of \(g_n\) (suitable for statistical applications) every finite (with probability 1) stopping time is truncatable.
Similar content being viewed by others
References
Arrow, K.J., Blackwell, D., Girshik, M.A.: Bayes and minimax solutions of sequential decision problems. Econometrica 17(3/4), 213–244 (1949)
Berger, J.O.: Statistical Decision Theory and Bayesian Analysis, 2nd edn. Springer, New York (1985)
Burkholder, D.L., Wijsman, R.A.: Optimum property and admissibility of sequential tests. Ann. Math. Stat. 34(1), 1–17 (1963)
Chow, Y.S., Robbins, H., Siegmund, D.: Great Expectations: The Theory of Optimal Stopping. Houghton Mifflin Company, Boston (1971)
DeGroot, M.H.: Optimal Statistical Decisions. McGraw-Hill, New York (1970)
Ferguson, T.S.: Mathematical Statistics: A Decision Theoretic Approach. Academic Press, New York (1967)
Ghosh, M., Mukhopadhyay, N., Sen, P.K.: Sequential Estimation. Wiley, New York (1997)
Gordienko, E., Novikov, A.: Characterizations of optimal policies in a general stopping problem and stability estimating. Probab. Eng. Inf. Sci. 28, 335–352 (2014)
Novikov, A.: Optimal sequential tests for two simple hypotheses based on independent observations. Int. J. Pure Appl. Math. 45(2), 291–314 (2008)
Novikov, A.: Optimal sequential multiple hypothesis tests. Kybernetika 45(2), 309–330 (2009)
Novikov, A.: Optimal sequential tests for two simple hypotheses. Seq. Anal. 28(2), 188–217 (2009)
Novikov, A.: Optimal sequential procedures with Bayes decision rules. Kybernetika 46(4), 754–770 (2010)
Novikov, A.: Optimal sequential procedures with Bayes decision rules. Int. Math. Forum 5(43), 2137–2147 (2010)
Novikov, A., Novikov, P.: Locally most powerful sequential tests of a simple hypothesis vs. one-sided alternatives. J. Stat. Plan. Inf. 140(3), 750–765 (2010)
Novikov, A.A., Novikov, P.A.: Locally most powerful sequential tests of a simple hypothesis vs. one-sided alternatives for independent observations. Theory Probab. Appl. 56–3, 420–442 (2012)
Samuel, E.: Randomized sequential tests. A comparison between curtailed single-sampling plans and sequential probability ratio tests. J. Am. Stat. Assoc. 65(329), 431–437 (1970)
Schmitz, N.: Optimal Sequentially Planned Decision Procedures. Lecture Notes in Statistics 79. Springer, New York (1993)
Shiryayev, A.N.: Optimal Stopping Rules. Springer, New York (1978)
Tartakovsky, A.G.: Sequential Methods in the Theory of Informational Systems (in Russian). “Statistical Theory of Communication” Series, no. 33. Radio i svyaz, Moscow (1991)
Wald, A., Wolfowitz, J.: Optimum character of the sequential probability ratio test. Ann. Math. Statistics 19, 326–339 (1948)
Acknowledgments
The author greatly appreciates the support of the Autonomous Metropolitan University, Mexico City, Mexico, where this work was done, and a partial support of the National System of Investigators (SNI) of CONACyT, Mexico.
Author information
Authors and Affiliations
Corresponding author
Appendix: Proofs
Appendix: Proofs
1.1 Proof of Theorem 1
Let
where \(v=v(x_1,\ldots , x_{k+1})\) is any \(\mu ^{k+1}\)-integrable function, \(k=0,1,2,\ldots \).
The proof of Theorem 1 is based on the following simple lemma (see also Lemma A.2 in [14]):
Lemma 5
Let \(k\) be any natural number, and let
be any \(\mu ^{k+1}\)-integrable function. Then
where
is a \(\mu ^k\)-integrable function. There is an equality in (4.1) if and only if
\(\mu ^{k}\)-almost everywhere on \(T_{k}^\psi =\{(x_1,\ldots ,x_{k}): t_{k}^\psi (x_1,\ldots ,x_{k})>0\}\).
By the Fubini theorem, the left-hand side of (4.1) is equal to
Applying Lemma A.1 [14] (for \(\max \) instead of \(\min \)) we see that (4.3) is lesser than or equal to
by the definition of \(v_{k}\).
Moreover, by the same Lemma A.1 [14], (4.3) is equal to (4.4) if and only if (4.2) holds \(\mu ^k\)-almost everywhere on \(T_k^\psi \).\(\Box \)
Let now for any natural \(1\le n\le N\)
It is obvious that \(G_N(\psi )=\widetilde{G}_N^N(\psi )\) (see (2.2)) and that \(\widetilde{G}_1^N(\psi )=Q_0^N\) (see (2.4)). Applying Lemma 5 initially to \(v\equiv V_N^N\) with \(k+1=N\) we have
where the equality is attained if and only if
\(\mu ^{N-1}\)-almost everywhere on \(T_{N-1}^\psi \).
It is not difficult to see, by induction, using Lemma 5 again, that for all \(n=N-1,\ldots ,1\)
with an equality if and only if
\(\mu ^n\)-almost everywhere on \(T_n^\psi \) for all \(n=1,\ldots , N-1\). Thus, we have
with an equality if and only if (4.6) holds \(\mu ^n\)-almost everywhere on \(T_n^\psi \) for all \(n=1,\ldots ,N-1\). \(\Box \)
1.2 Proof of Lemma 3
Let us denote
By Theorem 1, for all \(N=1,2,\ldots \)
Obviously, \(U_N\le U\) for every \(N=1,2,\ldots \), so
Let us show first that there is an equality in (4.7).
Suppose first that \(U<\infty \).
Let us suppose that there is a strict inequality in (4.7), i.e. that \(\lim _{N\rightarrow \infty }U_N= U-3\epsilon \), with some \(\epsilon >0\). We immediately have from this that
for all \(N\).
On the other hand, by the definition of \(U\) there exists a \(\psi \in \fancyscript{F} \) such that \(U-\epsilon \le G(\psi )\le U\).
By the condition of the Lemma, \(\lim _{N\rightarrow \infty }G_N(\psi )= G(\psi )\), so we have that
for all sufficiently large \(N\). Because, by definition, \(G_N(\psi )\le U_N\), we have from (4.9) that
for all sufficiently large \(N\), which contradicts (4.8).
In a very similar way we also get a contradiction in case \(U=\infty \) in (4.7). The contrary to the equality in (4.7) is \(\lim _{N\rightarrow \infty } U_N=k<\infty \). Substituting, in the above proof \(k\), \(2k\) and \(3k\), respectively, for \(U-3\epsilon \), \(U-2\epsilon \), and \(U-\epsilon \), we easily get to a contradiction in the same way as before.
Therefore, in either case
1.3 Proof of Theorem 2
Let for any \(\psi \in \fancyscript{F}\)
Passing to the limit, as \(N\rightarrow \infty \), on both sides of (4.5) we have
and passing to the limit, as \(N\rightarrow \infty \), in (2.3), we obtain
where, \(Q_n=Q_n(x_1,\ldots ,x_n)=\lim _{N\rightarrow \infty }Q_n^N(x_1,\ldots ,x_n)\) satisfies the equation
\(\mu ^n\)-almost everywhere, \(n=0,1,2,\ldots \). Obviously, \(\widetilde{G}_1(\psi )=Q_0\), so it follows from (4.10) that for any fixed natural \(n\)
Let us suppose that there exists some stopping rule \(\psi \in \fancyscript{F}\) such that \(G(\psi )=Q_0<\infty \). Then there are equalities in all of the inequalities in (4.11). Taking into account the necessary condition in Lemma 5, we see that this is only possible when
\(\mu ^n\)-almost everywhere on \(T_n^\psi \), for all natural \(n\). In addition, we have that for all natural \(n\)
It follows from (4.13), in particular that
Since \(\psi \in \fancyscript{F}\), by supposition, we have that \(\lim _{n\rightarrow \infty }G_n(\psi )= G(\psi ) =Q_0\). Thus, (3.6) follows from (4.14).
On the other hand, suppose now that a stopping rule \(\psi \in \fancyscript{F}\) satisfies (4.12) \(\mu ^n\)-almost everywhere on \(T_n^\psi \), for all natural \(n\), and that (3.6) holds.
Then, by Lemma 5, there are equalities in all of the inequalities (4.11) except, probably, for the first one. In particular, we have that (4.13) (and consequently (4.14)) holds for all natural \(n\). Then it follows from (4.14) and (3.6) that \(G_n(\psi )\rightarrow Q_0\), as \(n\rightarrow \infty \). Because, by supposition, \(G(\psi )=\lim _{n\rightarrow \infty }G_n(\psi )=Q_0\), we have from Lemma 3
1.4 Proof of Lemma 4
By Assumption 2, for every \(\psi \in \fancyscript{F}_1\) \(G^+(\psi )=\sum _{n=1}^\infty E s_n^\psi g_n^+\le E\sup _{n\ge 1}g_n^+=M<\infty \), so Condition (a) of Definition 1 is satisfied. Conditions (b) and (c) of Definition 1 are satisfied in an obvious way.
To complete the proof, let us show that Condition (d) of Definition 1 is also satisfied.
Let \(\psi \in \fancyscript{F}_1\) be a stopping rule such that \(G^-(\psi )<\infty \). Thus, by Assumption 1,
Because \(z_n\ge z_N\) for all \(n\ge N\), it follows from (4.15) that
thus,
We now have
The first summand on the right-hand side of (4.18) tends to 0 as \(N\rightarrow \infty \), because
because \(Et_N^\psi =P(\tau _\psi \ge N)\rightarrow 0\) as \(N\rightarrow \infty \), and \(E\sup _{n\ge 1}g_n^+<\infty \).
In a similar way, the second summand on the right-hand side of (4.18) tends to 0 as \(N\rightarrow \infty \), because, by Assumption 1, \(\{r_n^-,n\ge 1\}\) is uniformly integrable, and \(Et_n^\psi \rightarrow 0\) as \(n\rightarrow \infty \).
At last, the third summand on the right-hand side of (4.18) tends to 0 as \(N\rightarrow \infty \) by virtue of (4.16).
Thus,
It follows from (4.17) that
where the first summand converges to 0, as \(N\rightarrow \infty \), because \(G^-(\psi )<\infty \). The second summand on the right-hand side of (4.21) tends to 0 as \(N\rightarrow \infty \) by virtue of (4.19).
This implies that \(G(\psi )\ge \limsup _{N\rightarrow \infty }G_N(\psi )\); thus (see also (4.20)) \(\lim _{N\rightarrow \infty }G_N\) \((\psi )=G(\psi )\) follows.
1.5 Proof of Theorem 3
The necessity part of Theorem 3 follows from that of Theorem 2.
To prove the sufficiency part we only need to verify the condition (3.6) of Theorem 2. As \(0\le V_n-l_n\le V_n^++g_n^-f^n\), it suffices to show that
as \(n\rightarrow \infty \) and that \({ \int } t_n^\psi g_n^-f^n\mathrm{d}\mu ^n\rightarrow 0 \) as \(n\rightarrow \infty \). The latter follows from the proof of Lemma 4 in Section 4.4.
Let us prove (4.22).
Let \(g_N^*=\sup _{1\le n\le N}g_n^+(X_1,\ldots ,X_n)\) and \(g^*=\sup _{N\ge 1} g_N^*\). By Assumption 2, \(Eg^*<\infty \). It is not difficult to see, by induction, that
with probability 1, for all \(n\le N\), for all natural \(N\) (see the definitions in (2.2), (2.3) and (2.4)). Thus,
with probability 1 for all natural \(n\). Because of this, for every \(\psi \) such that \(P(\tau _\psi <\infty )=1\)
as \(n\rightarrow \infty \).
It follows from (4.13) that
Applying (4.29) on the left-hand side of (4.23), we have \( ER^* -E t_n^\psi c_{n} \ge Q_0, \) or
Let \(k\) be any positive number. Then it follows from (4.24) that
Because
and the second summand by virtue of (4.27) tends to 0, as \(n\rightarrow \infty \), we have that the difference between the first summand on the right-hand side of (4.26) and the left-hand side of it, goes to 0 as \(n\rightarrow \infty \). Thus, from (4.25), we have that
Since \(k>0\) is arbitrary, \(P (\tau _\psi =\infty )=0\) follows.
1.6 Proof of Corollary 1
It is easy to see from the proof of the Corollary of Theorem 4.4 in [4] that Assumption 3 implies that Assumption 1 is fulfilled (with \(r_n^*\) and \(c_n^*\) in place of \(r_n\) and \(c_n\) in Assumption 1, respectively). Assumption 2 follows from Assumption 3 as well, because \(g_n^+\le r_n^+\) (see Assumption 3 (a)). Thus, under Assumption 3 the assertions of Theorem 2 (with \(\fancyscript{F}\) defined by (3.7) ) and of Theorem 3 hold. Therefore, Corollary 1 follows.
The only thing to show is that the condition \(c_n\rightarrow \infty \), \(n\rightarrow \infty \), is sufficient for almost sure finiteness of all stopping times corresponding to (3.9). So let us suppose that
(which is equivalent to the almost sure convergence \(c_n\rightarrow \infty \), \(n\rightarrow \infty \)), and let us show that any \(\psi \) satisfying (3.9) \(\mu ^n\)-almost everywhere on \(T_n^\psi \) for all \(n=1,2,\ldots \) has the property \(P(\tau _\psi <\infty )=1\).
For any \(n=1,2,\ldots \) let us define \(W_n^N=V_n^N+c_nf^n\), \(N\ge n\), and \(W_n=V_n+c_nf^n\). It follows from (2.3) that
with \(W_N^N=(g_N+c_N)f^N=r_Nf^N\), \(N=1,2,\ldots \), \(n=1,2,\ldots ,N-1\). Let \(R_N^*=\max _{1\le n\le N}(r_n^N)^+\), \(N=1,2,\ldots \), and \(R^*=\sup _{N\ge 1}R_N^*\). Because \(\int c_{n+1}f^{n+1}\mathrm{d}\mu (x_{n+1})-c_nf^n\ge 0 \) for all \(n=1,2,\ldots \), it is easy to see by induction, using (4.28), that
almost surely for all \(n=1,\ldots , N-1\), \(N=1,2,\ldots \), and hence
almost surely for all natural \(n\).
It follows from (4.13) that
Applying (4.29) on the left-hand side of (4.30), we have \( ER^* -E t_n^\psi c_{n} \ge Q_0,\) or
Let \(k\) be any positive number. Then it follows from (4.31) that
Because
and the second summand by virtue of (4.27) tends to 0, as \(n\rightarrow \infty \), we have that the difference between the first summand on the right-hand side of (4.33) and the left-hand side of it, goes to 0 as \(n\rightarrow \infty \). Thus, from (4.32), we have that
Since \(k>0\) is arbitrary, \(P (\tau _\psi =\infty )=0\) follows.
Rights and permissions
About this article
Cite this article
Novikov, A. Characterization of optimality in classes of “truncatable” stopping rules. Bol. Soc. Mat. Mex. 21, 99–117 (2015). https://doi.org/10.1007/s40590-014-0028-4
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s40590-014-0028-4