Skip to main content
Log in

Multi-level Bayes and MAP Monotonicity Testing

  • Published:
Mathematical Methods of Statistics Aims and scope Submit manuscript

Abstract

In this paper, we develop Bayes and maximum a posteriori probability (MAP) approaches to monotonicity testing. In order to simplify this problem, we consider a simple white Gaussian noise model and with the help of the Haar transform we reduce it to the equivalent problem of testing positivity of the Haar coefficients. This approach permits, in particular, to understand links between monotonicity testing and sparse vectors detection, to construct new tests, and to prove their optimality without supplementary assumptions. The main idea in our construction of multi-level tests is based on some invariance properties of specific probability distributions. Along with Bayes and MAP tests, we construct also adaptive multi-level tests that are free from the prior information about the sizes of non-monotonicity segments of the function.

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

Similar content being viewed by others

REFERENCES

  1. Y. Baraud, S. Huet, and B. Laurent, ‘‘Testing convex hypotheses on the mean of a gaussian vector, application to testing qualitative hypotheses on a regression function,’’ Ann. Statist. 33 (1), 214–257 (2005).

    Article  MathSciNet  Google Scholar 

  2. A. W. Bowman, M. C. Jones, and I. Gijbels, ‘‘Testing monotonicity of regression,’’ J. Comput. Graph. Statist. 7 (4), 489–500 (1998).

    Google Scholar 

  3. M. Burnashev and I. Bergamov, ‘‘On a problem of signal detection leading to stable distributions,’’ Theory Probab. Appl. 35 (3), 556–560 (1990).

    Article  MathSciNet  Google Scholar 

  4. D. Chetverikov, ‘‘Testing regression monotonicity in econometric models,’’ Econometric Theory 35 (4), 729–776 (2019).

    Article  MathSciNet  Google Scholar 

  5. R. Dobrushin, ‘‘A statistical problem arising in the theory of detection of signals in the presence of noise in a multi-channel system and leading to stable distribution laws,’’ Theory Probab. Appl. 3 (2), 161–173 (1958).

    Article  Google Scholar 

  6. L. Dümbgen and V. G. Spokoiny, ‘‘Multiscale testing of qualitative hypotheses,’’ Ann. Statist., 29 (1), 124–152 (2001).

    Article  MathSciNet  Google Scholar 

  7. C. Durot, ‘‘A Kolmogorov-type test for monotonicity of regression,’’ Statist. Probab. Lett. 63 (4), 425–433 (2003).

    Article  MathSciNet  Google Scholar 

  8. S. Ghosal, A. Sen, and A. van der Vaart, ‘‘Testing monotonicity of regression,’’ Ann. Statist. 28 (4), 1054–1082 (2000).

    MathSciNet  MATH  Google Scholar 

  9. I. Gijbels, P. Hall, M. Jones, and I. Koch, ‘‘Tests for monotonicity of a regression mean with guaranteed level,’’ Biometrika 87 (3), 663–673 (2000).

    Article  MathSciNet  Google Scholar 

  10. P. Hall and N. Heckman, ‘‘Testing for monotonicity of a regression mean by calibrating for linear functions,’’ Ann. Statist. 28 (1), 20–39 (2000).

    MathSciNet  MATH  Google Scholar 

  11. Yu. I. Ingster and I. A. Suslina, Nonparametric Goodness-of-Fit Testing under Gaussian Models (Lecture Notes in Statistics 169, Springer, New York, 2003).

  12. G. S. Mudholkar, ‘‘The integral of an invariant unimodal function over an invariant convex set—an inequality and applications’’, Proc. Amer. Math. Soc 17 (6), 1327–1333 (1966).

    MathSciNet  MATH  Google Scholar 

  13. J. P. Nolan, Stable Distributions: Models for Heavy-Tailed Data (Springer, New York, 2016).

    Google Scholar 

  14. R. Pyke, ‘‘Spacings (with discussion)’’, J. Roy. Statist. Soc. Ser. B 27 (3), 395–449 (1965).

    MathSciNet  MATH  Google Scholar 

  15. W. Schlee, Nonparametric Tests of the Monotony and Convexity of Regression (In Non-parametric Statistical Inference, North-Holland, Amsterdam, 1982).

  16. Y. L. Tong, Probability Inequalities in Multivariate Distributions., 1st ed. (Academic Press, London, 1980).

    MATH  Google Scholar 

  17. J. C. Wang and M. C. Meyer, ‘‘Testing the monotonicity or convexity of a function using regression splines’’, Canad. J. Statist. 39 (1), 89–107 (2011).

    Article  MathSciNet  Google Scholar 

Download references

Author information

Authors and Affiliations

Authors

Corresponding authors

Correspondence to Yu. Golubev or C. Pouet.

APPENDIX

A.1. Proof of Theorem 1

With a simple algebra we obtain

$$\log\bigl{\{}\mathbf{E}\exp\bigl{[}{\textrm{i}}tB_{h}(\xi)\bigr{]}\bigr{\}}=n_{h}\log\left\{\frac{1}{\sqrt{2\pi}}\int\limits_{-\infty}^{\infty}\cos\biggl{[}\frac{tS(x)}{n_{h}}\biggr{]}{\textrm{e}}^{-x^{2}/2}\ dx+\frac{\textrm{i}}{\sqrt{2\pi}}\int\limits_{-\infty}^{\infty}\sin\biggl{[}\frac{tS(x)}{n_{h}}\biggr{]}{\textrm{e}}^{-{x^{2}}/{2}}\ dx\right\}$$

$${}=n_{h}\log\left\{1+\int\limits_{-\infty}^{\infty}\biggr{[}\cos\biggl{[}\frac{t}{n_{h}}\biggl{(}\frac{1}{\Phi(x)}-1\biggr{)}\biggr{]}-1\biggl{]}\ d\Phi(x)+{\textrm{i}}\int\limits_{-\infty}^{\infty}\sin\biggl{[}\frac{t}{n_{h}}\biggl{(}\frac{1}{\Phi(x)}-1\biggr{)}\biggr{]}\ d\Phi(x)\right\}$$

$${}=n_{h}\log\left\{1+\int\limits_{0}^{1}\biggl{\{}\cos\biggl{[}\frac{t}{n_{h}}\biggl{(}\frac{1}{u}-1\biggr{)}\biggr{]}-1\biggr{\}}\ du+{\textrm{i}}\int\limits_{0}^{1}\sin\biggl{[}\frac{t}{n_{h}}\biggl{(}\frac{1}{u}-1\biggr{)}\biggr{]}\ du\right\}$$

$${}=n_{h}\log\left\{1+\int\limits_{0}^{\infty}\frac{1}{(1+u)^{2}}\biggl{[}\cos\biggl{(}\frac{tu}{n_{h}}\biggr{)}-1\biggr{]}\ du+{\textrm{i}}\int\limits_{0}^{\infty}\frac{1}{(1+u)^{2}}\sin\biggl{(}\frac{tu}{n_{h}}\biggr{)}\ du\right\}$$

$${}=n_{h}\log\left\{1+\frac{|t|}{n_{h}}\int\limits_{0}^{\infty}\frac{\cos(u)-1}{(|t|/n_{h}+u)^{2}}\ du+\frac{{\textrm{i}}t}{n_{h}}\int\limits_{0}^{\infty}\frac{\sin(u)}{(|t|/n_{h}+u)^{2}}\ du\right\}.$$
(A.1)

It is clear that as \(n_{h}\rightarrow\infty\)

$$\int\limits_{0}^{\infty}\frac{\cos(u)-1}{(|t|/n_{h}+u)^{2}}\ du\rightarrow\int\limits_{0}^{\infty}\frac{\cos(u)-1}{u^{2}}\ du=-\frac{\pi}{2}.$$
(A.2)

Let us choose \(\epsilon_{h}<1\) such that

$$\lim_{h\rightarrow 0}\epsilon_{h}=0,\quad\lim_{h\rightarrow 0}\epsilon_{h}n_{h}=\infty.$$

Then we get by the Taylor formula

$$\int\limits_{0}^{\infty}\frac{\sin(u)}{(|t|/n_{h}+u)^{2}}\ du=\int\limits_{0}^{\epsilon_{h}}\frac{\sin(u)}{(|t|/n_{h}+u)^{2}}\ du+\int\limits_{\epsilon_{h}}^{\infty}\frac{\sin(u)}{(|t|/n_{h}+u)^{2}}\ du$$

$${}=\int\limits_{0}^{\epsilon_{h}}\frac{u}{(|t|/n_{h}+u)^{2}}\ du+O\left[\int\limits_{0}^{\epsilon_{h}}\frac{u^{3}}{(|t|/n_{h}+u)^{2}}\ du\right]+(1+o(1))\int\limits_{\epsilon_{h}}^{\infty}\frac{\sin(u)}{u^{2}}\ du$$

$${}=\log\bigg{(}1+\frac{\epsilon_{h}n_{h}}{|t|}\biggr{)}-\frac{\epsilon_{h}n_{h}}{(|t|+\epsilon_{h}n_{h})}+(1+o(1))\int\limits_{\epsilon_{h}}^{\infty}\frac{\sin(u)}{u^{2}}\ du+O(\epsilon_{h}^{2}).$$
(A.3)

Next, integrating by parts, we obtain

$$\int\limits_{x}^{\infty}\frac{\sin(z)}{z^{2}}\ dz=-\int\limits_{x}^{\infty}\frac{\sin(z)}{z}\ d\biggl{(}\log\frac{1}{z}\biggr{)}=\frac{\sin(x)}{x}\log\frac{1}{x}+\int\limits_{x}^{\infty}\frac{z\cos(z)-\sin(z)}{z^{2}}\log\frac{1}{z}\ dz.$$

Hence, as \(x\rightarrow 0\)

$$\int\limits_{x}^{\infty}\frac{\sin(z)}{z^{2}}\ dz=\log\frac{1}{x}+\int\limits_{0}^{\infty}\frac{z\cos(z)-\sin(z)}{z^{2}}\log\frac{1}{z}\ dz+O\bigg{(}x^{2}\log\frac{1}{x}\biggr{)}$$

$${}=\log\frac{1}{x}+(1-\gamma)+O\bigg{(}x^{2}\log\frac{1}{x}\biggr{)},$$

where \(\gamma\) is Euler’s constant.

With this equation we continue (44) as follows:

$$\int\limits_{0}^{\infty}\frac{\sin(u)}{(|t|/n_{h}+u)^{2}}\ du=\log\frac{1}{|t|}+\log(n_{h})-\gamma+O\biggl{(}\frac{|t|}{\epsilon_{h}n_{h}}\biggr{)}+O\biggl{(}\epsilon_{h}^{2}\log\frac{1}{\epsilon_{h}}\biggr{)}.$$

Substituting this equation and (43) in (42), we get

$$\log\bigl{\{}\mathbf{E}\exp\bigl{[}{\textrm{i}}tB_{h}(\xi)\bigr{]}\bigr{\}}=-\frac{\pi|t|}{2}+{\textrm{i}}t\biggl{(}\log\frac{1}{|t|}+\log(n_{h})-\gamma\biggr{)}+o(1),$$

thus, proving the theorem.

A.2. Proof of Theorem 3

I. A lower bound. By (22) we have for any given \(x\)

$$\beta^{M}_{\rho,\tau}(A)\geqslant\mathbf{P}\biggl{\{}\biggl{[}\log\biggl{[}S\biggl{(}-\frac{A}{\sigma_{\rho}}+\xi_{\rho,\tau}\biggr{)}\biggr{]}-\log\frac{n_{\rho}}{\bar{\pi}_{\rho}}\biggr{]}$$

$${}\bigvee\biggl{[}\zeta_{\rho}-\log\frac{1}{\bar{\pi}_{\rho}}\biggr{]}\leqslant q_{\alpha}^{\varkappa}\biggr{\}}\mathbf{P}\biggl{\{}\max_{h\in\mathcal{H}}\biggl{[}\zeta_{h}-\log\frac{1}{\bar{\pi}_{h}}\biggr{]}\leqslant q_{\alpha}^{\varkappa}\biggr{\}}$$

$${}=\mathbf{P}\biggl{\{}S\biggl{(}-\frac{A}{\sigma_{\rho}}+\xi_{\rho,\tau}\biggr{)}\leqslant\exp\biggl{[}q_{\alpha}^{\varkappa}+\log\frac{n_{\rho}}{\bar{\pi}_{\rho}}\biggr{]}\biggr{\}}\mathbf{P}\biggl{\{}\zeta_{\rho}\leqslant q_{\alpha}^{\varkappa}+\log\frac{1}{\bar{\pi}_{\rho}}\biggr{\}}\mathbf{P}\bigl{\{}\zeta\leqslant q_{\alpha}^{\varkappa}\bigr{\}}$$

$${}\geqslant(1-\alpha)^{1+\bar{\pi}_{\rho}}\mathbf{P}\biggl{\{}S\biggl{(}-\frac{A}{\sigma_{\rho}}+\xi_{\rho,\tau}\biggr{)}\leqslant\exp\biggl{[}q_{\alpha}^{\varkappa}+\log\frac{n_{\rho}}{\bar{\pi}_{\rho}}\biggr{]};\xi_{\rho,\tau}\geqslant x\biggr{\}}$$

$${}\geqslant(1-\alpha)^{1+\bar{\pi}_{\rho}}\mathbf{P}\bigl{\{}\xi_{\rho,\tau}\geqslant x\bigr{\}}\mathbf{1}\biggl{\{}S\biggl{(}-\frac{A}{\sigma_{\rho}}+x\biggr{)}\leqslant\exp\biggl{[}q_{\alpha}^{\varkappa}+\log\frac{n_{\rho}}{\bar{\pi}_{\rho}}\biggr{]}\biggr{\}}.$$
(A.4)

Let \(R(z)\geqslant 0\) be a solution to

$$S\Bigl{[}-\sqrt{R(z)}\Bigr{]}=z.$$

It is easy to check with the help of (7) that as \(z\rightarrow\infty\)

$$R(z)=2\log(z)-\log[4\pi\log(z)]+o(1).$$
(A.5)

Denote for brevity

$$r_{h}(q,u)=2\biggl{(}q+\log\frac{n_{h}}{u}\biggr{)}-\log\biggl{[}4\pi\biggl{(}q+\log\frac{n_{h}}{u}\biggr{)}\biggr{]}.$$

With (A.5) and the Markov inequality we obtain for any \(\epsilon>0\)

$$\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{1}\biggl{\{}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+x\biggr{)}\leqslant\exp\biggl{[}q_{\alpha}^{\varkappa}+\log\frac{n_{h}}{\bar{\pi}_{h}}\biggr{]}\biggr{\}}\biggr{]}=1-\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{1}\biggl{\{}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+x\biggr{)}>\exp\biggl{[}q_{\alpha}^{\varkappa}+\log\frac{n_{h}}{\bar{\pi}_{h}}\biggr{]}\biggr{\}}\biggr{]}$$

$${}=1-\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{1}\biggl{\{}\biggl{(}\frac{A_{h}}{\sigma_{h}}-x\biggr{)}^{2}-r_{h}(q_{\alpha}^{\varkappa},\bar{\pi}_{h})>o(1)\biggr{\}}\biggr{]}$$

$${}\geqslant 1-\frac{1}{\epsilon H^{*}_{\bar{\pi}}}\mathbf{E}_{\bar{\pi}}\biggl{[}\biggl{(}\frac{A_{h}}{\sigma_{h}}-x\biggr{)}^{2}-r_{h}(q_{\alpha}^{\varkappa},\bar{\pi}_{h})+\epsilon H^{*}_{\bar{\pi}}\biggr{]}_{+}.$$
(A.6)

With condition (25) and simple algebra it is easy to check that

$$\lim_{\bar{\pi}\rightarrow 0}\frac{1}{H^{*}_{\bar{\pi}}}\mathbf{E}_{\bar{\pi}}\biggl{|}r_{h}(q_{\alpha}^{\varkappa},\bar{\pi})-R_{h}(q_{\alpha}^{\varkappa},H_{\bar{\pi}})\biggr{|}=0.$$

Therefore (29) follows from (45), (47) and the above equation.

II. An upper bound. Since \(S(x)\) is a decreasing function, we have obviously for any \(x\geqslant 0\)

$$\bar{\beta}^{M}_{\bar{\pi}}(\mathbf{A})\leqslant\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{P}\biggl{\{}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+\xi\biggr{)}\leqslant\exp\biggl{[}q_{\alpha}^{\varkappa}+\log\frac{n_{h}}{\bar{\pi}_{h}}\biggr{]}\biggr{\}}\biggr{]}$$

$${}\leqslant\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{P}\biggl{\{}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+\xi\biggr{)}\leqslant\exp\biggl{[}q_{\alpha}^{\varkappa}+\log\frac{n_{h}}{\bar{\pi}_{h}}\biggr{]};\xi\leqslant x\biggr{\}}\biggr{]}$$

$${}+\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{P}\biggl{\{}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+\xi\biggr{)}\leqslant\exp\biggl{[}q_{\alpha}^{\varkappa}+\log\frac{n_{h}}{\bar{\pi}_{h}}\biggr{]};\xi>x\biggr{\}}\biggr{]}$$

$${}\leqslant\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{1}\biggl{\{}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+x\biggr{)}\leqslant\exp\biggl{[}q_{\alpha}^{\varkappa}+\log\frac{n_{h}}{\bar{\pi}_{h}}\biggr{]}\biggr{\}}\biggr{]}+\mathbf{P}\bigl{\{}\xi>x\bigr{\}}.$$

Next, with (46), the Markov inequality, and Condition (25) we get for any \(\epsilon>0\)

$$\sum_{h\in\mathcal{H}}\bar{\pi}_{h}\mathbf{1}\biggl{\{}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+x\biggr{)}\leqslant\exp\biggl{[}q_{\alpha}^{\varkappa}+\log\frac{n_{h}}{\bar{\pi}_{h}}\biggr{]}\biggr{\}}$$

$${}\leqslant\sum_{h\in\mathcal{H}}\bar{\pi}_{h}\mathbf{1}\biggl{\{}r_{h}(q_{\alpha}^{\varkappa},\bar{\pi}_{h})-\biggl{(}\frac{A_{h}}{\sigma_{h}}-x\biggr{)}^{2}>o(1)\biggr{\}}$$

$${}\leqslant\frac{1}{\epsilon H^{*}_{\bar{\pi}}}\sum_{h\in\mathcal{H}}\bar{\pi}_{h}\biggr{[}r_{h}(q_{\alpha}^{\varkappa},\bar{\pi}_{h})-\biggl{(}\frac{A_{h}}{\sigma_{h}}-x\biggr{)}^{2}+\epsilon H^{*}_{\bar{\pi}}\biggr{]}_{+}$$

$${}\leqslant\frac{1}{\epsilon H^{*}_{\bar{\pi}}}\sum_{h\in\mathcal{H}}\bar{\pi}_{h}\biggr{[}R_{h}(q_{\alpha}^{\varkappa},H_{\bar{\pi}})-\biggl{(}\frac{A_{h}}{\sigma_{h}}-x\biggr{)}^{2}+\epsilon H^{*}_{\bar{\pi}}\biggr{]}_{+}+o(1).$$

To complete the proof, let us choose \(x=\sqrt{\epsilon H^{*}_{\bar{\pi}}}\).

A.3. Proof of Theorem 4

A lower bound. For given \(x\) and \(\delta>0\) by (17) and (13), we obtain

$$\beta^{B}_{\rho,\tau}(A)\geqslant\mathbf{P}\biggl{\{}\frac{\bar{\pi}_{\rho}}{n_{\rho}}S\biggl{(}-\frac{A}{\sigma_{\rho}}+\xi_{\rho,\tau}\biggr{)}+\zeta^{\circ}\leqslant q_{\alpha}^{\circ};\xi_{\rho,\tau}\geqslant x\biggr{\}}$$

$${}\geqslant\mathbf{P}\biggl{\{}\frac{\bar{\pi}_{\rho}}{n_{\rho}}S\biggl{(}-\frac{A}{\sigma_{\rho}}+x\biggr{)}+\zeta^{\circ}\leqslant q_{\alpha}^{\circ};\xi_{\rho,\tau}\geqslant x\biggr{\}}$$

$${}=\mathbf{P}\bigl{\{}\xi_{\rho,\tau}\geqslant x\bigr{\}}\mathbf{P}\biggl{\{}\zeta^{\circ}\leqslant q_{\alpha}^{\circ}-\frac{\bar{\pi}_{\rho}}{n_{\rho}}S\biggl{(}-\frac{A}{\sigma_{\rho}}+x\biggr{)}\biggr{\}}$$

$${}\geqslant\mathbf{P}\bigl{\{}\xi_{\rho,\tau}\geqslant x\bigr{\}}\mathbf{P}\bigl{\{}\zeta^{\circ}\leqslant(1-\delta)q_{\alpha}^{\circ}\bigr{\}}\mathbf{1}\biggl{\{}\frac{\bar{\pi}_{\rho}}{n_{\rho}}S\biggl{(}-\frac{A}{\sigma_{\rho}}+x\biggr{)}\leqslant\delta q_{\alpha}^{\circ}\biggr{\}}.$$

Similarly to (47), since \(\bar{\pi}\rightarrow 0\), we get

$$\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{1}\biggl{\{}\frac{\bar{\pi}_{h}}{n_{h}}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+x\biggr{)}\leqslant\delta q_{\alpha}^{\circ}\biggr{\}}\biggr{]}$$

$${}\geqslant 1-\frac{1}{\epsilon H^{*}_{\bar{\pi}}}\mathbf{E}_{\bar{\pi}}\biggl{[}\biggl{(}\frac{A_{h}}{\sigma_{h}}-x\biggr{)}^{2}+\epsilon H^{*}_{\bar{\pi}}-R_{h}[\log(\delta q_{\alpha}^{\circ}),H_{\bar{\pi}}]\biggr{]}_{+}$$

$${}\geqslant 1-\frac{1}{\epsilon H^{*}_{\bar{\pi}}}\mathbf{E}_{\bar{\pi}}\biggl{[}\biggl{(}\frac{A_{h}}{\sigma_{h}}-x\biggr{)}^{2}+\epsilon H^{*}_{\bar{\pi}}-R_{h}[\log(q_{\alpha}^{\circ}),H_{\bar{\pi}}]\biggr{]}_{+}v\geqslant 1+o(1).$$

So, since \(\delta\) is arbitrary, (33) follows from the above inequalities.

An upper bound. Since \(S(x)\) is decreasing, we get

$$\beta^{B}_{\rho,\tau}(A)\leqslant\mathbf{P}\biggl{\{}\frac{\bar{\pi}_{\rho}}{n_{\rho}}S\biggl{(}-\frac{A}{\sigma_{\rho}}+\xi_{\rho,\tau}\biggr{)}+\zeta^{\circ}\leqslant q_{\alpha}^{\circ};\xi_{\rho,\tau}\geqslant x\biggr{\}}$$

$${}+\mathbf{P}\biggl{\{}\frac{\bar{\pi}_{\rho}}{n_{\rho}}S\biggl{(}-\frac{A}{\sigma_{\rho}}+\xi_{\rho,\tau}\biggr{)}+\zeta^{\circ}\leqslant q_{\alpha}^{\circ};\xi_{\rho,\tau}\leqslant x\biggr{\}}$$

$${}\leqslant\mathbf{P}\bigl{\{}\xi_{\rho,\tau}\geqslant x\bigr{\}}+\mathbf{P}\biggl{\{}\frac{\bar{\pi}_{\rho}}{n_{\rho}}S\biggl{(}-\frac{A}{\sigma_{\rho}}+x\biggr{)}+\zeta^{\circ}\leqslant q_{\alpha}^{\circ}\biggr{\}}.$$
(A.7)

Next, for any given \(x^{\circ}<q_{\alpha}^{\circ}\) we obtain with the help of the Markov inequality and (46)

$$\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{P}\biggl{\{}\frac{\bar{\pi}_{h}}{n_{h}}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+x\biggr{)}+\zeta^{\circ}\leqslant q_{\alpha}^{\circ}\biggr{\}}\biggr{]}$$

$${}\leqslant\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{P}\biggl{\{}\frac{\bar{\pi}_{h}}{n_{h}}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+x\biggr{)}+\zeta^{\circ}\leqslant q_{\alpha}^{\circ};\zeta^{\circ}<x^{\circ}\biggr{\}}\biggr{]}$$

$${}+\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{P}\biggl{\{}\frac{\bar{\pi}_{h}}{n_{h}}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+x\biggr{)}+\zeta^{\circ}\leqslant q_{\alpha}^{\circ};\zeta^{\circ}\geqslant x^{\circ}\biggr{\}}\biggr{]}$$

$${}\leqslant\mathbf{P}\bigl{\{}\zeta^{\circ}<x^{\circ}\}+\mathbf{E}_{\bar{\pi}}\biggl{[}\mathbf{1}\biggl{\{}\frac{\bar{\pi}_{h}}{n_{h}}S\biggl{(}-\frac{A_{h}}{\sigma_{h}}+x\biggr{)}\leqslant q_{\alpha}^{\circ}-x^{\circ}\biggr{\}}\biggr{]}$$

$${}\leqslant\mathbf{P}\bigl{\{}\zeta^{\circ}<x^{\circ}\}+\frac{1}{\epsilon H^{*}_{\bar{\pi}}}\mathbf{E}_{\bar{\pi}}\biggl{[}R_{h}[\log(q_{\alpha}^{\circ}-x^{\circ}),H_{\bar{\pi}}]+\epsilon H^{*}_{\bar{\pi}}-\biggl{(}\frac{A_{h}}{\sigma_{h}}-x\biggr{)}^{2}\biggr{]}_{+}.$$

Finally choosing \(x=\sqrt{\epsilon H^{*}_{\bar{\pi}}}\) and combining this equation with (48), we complete the proof of (35).

A.4. Proof of Theorem 5

Let us consider

$$H\bigl{(}\bar{\pi}^{\omega,\nu},{\Pi}\bigr{)}=\sum_{h\in\mathcal{H}}\bar{\pi}^{\omega,\nu}_{h}\log\frac{1}{{\Pi}_{h}}.$$

One can check easily with (38)–(40) and (26) that as \(\omega\rightarrow\infty\)

$$H\bigl{(}\bar{\pi}^{\omega,\nu},{\Pi}\bigr{)}=\log\frac{1}{\varepsilon}+\frac{1}{\omega}\sum_{k=1}^{\infty}\nu\biggl{(}\frac{k}{\omega}\biggr{)}L^{m,\varepsilon}(k)+o(1)=\log\frac{1}{\varepsilon}+o(1)$$

$${}+\int\limits_{0}^{\infty}\nu(x)\biggl{[}\log(x\omega+1)+\sum_{s=0}^{m-1}\log[\psi_{s}(x\omega)]+(1+\varepsilon)\log[\psi_{m}(x\omega)]\biggr{]}\ dx.$$

It is also clear in view of (27) that

$$\int\limits_{0}^{\infty}\nu(x)\log(x\omega+1)\ dx=\log(\omega)+O(1)$$

and for any integer \(s\geqslant 0\)

$$\int\limits_{0}^{\infty}\nu(x)\log[\psi_{s}(x\omega)]\ dx=\log[\psi_{s}(\omega)]+O(1).$$

Therefore as \(\omega\rightarrow\infty\)

$$H\bigl{(}\bar{\pi}^{\omega,\nu},{\Pi}\bigr{)}=\log(\omega)+(1+o(1))\log^{*}(\omega).$$
(A.8)

Next, with similar arguments we obtain

$$\sum_{h\in\mathcal{H}}\bar{\pi}^{\omega,\nu}_{h}\biggl{[}H(\bar{\pi}^{\omega,\nu},{\Pi})-\log\frac{1}{{\Pi}_{h}}\biggr{]}_{+}=\int\limits_{0}^{\infty}\nu(x)\Bigl{[}H(\bar{\pi}^{\omega,\nu},{\Pi})-\log(x\omega+1)$$

$${}-(1+o(1))\log[\log(x\omega+1)+1]\Bigr{]}_{+}dx$$

$${}=\int\limits_{0}^{\infty}\nu(x)\Bigl{[}\log(\omega)+(1+o(1))\log^{*}(\omega)-\log(x\omega+1)$$

$${}-(1+o(1))\log[\log(x\omega+1)+1]\Bigr{]}_{+}dx$$

$${}=\int\limits_{0}^{\infty}\nu(x)\bigl{[}-\log(x+1)+o(1)\log^{*}(\omega)\bigr{]}_{+}dx=o(1)\log^{*}(\omega)].$$
(A.9)

In view of (49) and (50) the rest of the proof is similar to the one of Theorem 3 and therefore omitted.

CONFLICT OF INTEREST

The authors declare that they have no conflict of interest.

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Golubev, Y., Pouet, C. Multi-level Bayes and MAP Monotonicity Testing. Math. Meth. Stat. 29, 57–75 (2020). https://doi.org/10.3103/S1066530720010032

Download citation

  • Received:

  • Revised:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.3103/S1066530720010032

Keywords:

Navigation