Abstract
Initially motivated by the study of the non-asymptotic properties of non-parametric tests based on permutation methods, concentration inequalities for uniformly permuted sums have been largely studied in the literature. Recently, Delyon et al. proved a new Bernstein-type concentration inequality based on martingale theory. This work presents a new proof of this inequality based on the fundamental inequalities for random permutations of Talagrand. The idea is to first obtain a rough inequality for the square root of the permuted sum, and then, iterate the previous analysis and plug this first inequality to obtain a general concentration of permuted sums around their median. Then, concentration inequalities around the mean are deduced. This method allows us to obtain the Bernstein-type inequality up to constants, and, in particular, to recovers the Gaussian behavior of such permuted sums under classical conditions encountered in the literature. Then, an application to the study of the second kind error rate of permutation tests of independence is presented.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
We’re sorry, something doesn't seem to be working properly.
Please try refreshing the page. If that doesn't work, please contact support so we can address the problem.
References
M. Albert, Tests of independence by bootstrap and permutation: an asymptotic and non-asymptotic study. Application to Neurosciences. PhD thesis, Nice, 2015
M. Albert, Y. Bouret, M. Fromont, P. Reynaud-Bouret, Bootstrap and permutation tests of independence for point processes. Ann. Stat. 43(6), 2537–2564 (2015)
M. Albert, Y. Bouret, M. Fromont, P. Reynaud-Bouret, Surrogate data methods based on a shuffling of the trials for synchrony detection: the centering issue. Neural Comput. 28(11), 2352–2392 (2016)
R. Adamczak, D. Chafaï, P. Wolff, Circular law for random matrices with exchangeable entries (2014), arXiv preprint arXiv:1402.3660
Y. Baraud, Non-asymptotic minimax rates of testing in signal detection. Bernoulli 8(5), 577–606 (2002)
B. Bercu, B. Delyon, E. Rio, Concentration Inequalities for Sums and Martingales (Springer, Cham, 2015)
E. Bolthausen, An estimate of the remainder in a combinatorial central limit theorem. Probab. Theory Relat. Fields 66(3), 379–386 (1984)
S. Boucheron, G. Lugosi, P. Massart, Concentration Inequalities: A Nonasymptotic Theory of Independence (Oxford University Press, Oxford, 2013)
S. Chatterjee, Stein’s method for concentration inequalities. Probab. Theory Relat. Fields 138(1), 305–321 (2007)
M. Dwass, On the asymptotic normality of some statistics used in non-parametric tests. Ann. Math. Stat. 26, 334–339 (1955)
M. Fromont, B. Laurent, P. Reynaud-Bouret, Adaptive tests of homogeneity for a Poisson process. Ann. Inst. Henri Poincaré Probab. Stat. 47(1), 176–213 (2011)
M. Fromont, B. Laurent, P. Reynaud-Bouret, The two-sample problem for poisson processes: adaptive tests with a nonasymptotic wild bootstrap approach. Ann. Stat. 41(3), 1431–1461 (2013)
J. Hájek, Some extensions of the Wald-Wolfowitz-Noether theorem. Ann. Math. Stat. 32(2), 506–523 (1961)
S.T. Ho, L.H.Y. Chen, An L p bound for the remainder in a combinatorial central limit theorem. Ann. Probab. 6(2), 231–249 (1978)
W. Hoeffding, A combinatorial central limit theorem. Ann. Math. Stat. 22(4), 558–566 (1951)
W. Hoeffding, The large-sample power of tests based on permutation of the observations. Ann. Math. Stat. 23(2), 169–192 (1952)
K. Jogdeo, Asymptotic normality in nonparametric methods. Ann. Math. Stat. 39, 905–922 (1968)
M. Ledoux, The Concentration of Measure Phenomenon, vol. 89 (American Mathematical Society, Providence, 2005)
P. Massart, J. Picard, Concentration Inequalities and Model Selection, vol. 1896 (Springer, Berlin, 2007)
C. McDiarmid, Concentration for independent permutations. Comb. Probab. Comput. 11(02), 163–178 (2002)
M. Motoo, On the Hoeffding’s combinatrial central limit theorem. Ann. Inst. Stat. Math. 8(1), 145–154 (1956)
G.E. Noether, On a theorem by Wald and Wolfowitz. Ann. Math. Stat. 20, 455–458 (1949)
J.P. Romano, Bootstrap and randomization tests of some nonparametric hypotheses. Ann. Stat. 17(1), 141–159 (1989)
L. Sansonnet, C. Tuleau-Malot, A model of Poissonian interactions and detection of dependence. Stat. Comput. 25(2), 449–470 (2015)
W. Schneller, A short proof of Motoo’s combinatorial central limit theorem using stein’s method. Probab. Theory Relat. Fields 78(2), 249–252 (1988)
C.P. Shapiro, L. Hubert, Asymptotic normality of permutation statistics derived from weighted sums of bivariate functions. Ann. Stat. 7, 788–794 (1979)
M. Talagrand, Concentration of measure and isoperimetric inequalities in product spaces. Publ. Math. Inst. Hautes Etudes Sci. 81(1), 73–205 (1995)
A. Wald, J. Wolfowitz, Statistical tests based on permutations of the observations. Ann. Math. Stat. 15(4), 358–372 (1944)
Acknowledgements
The author is grateful to Jean-François Coeurjolly for his insightful discussions. This work was supported in part by the French Agence Nationale de la Recherche (ANR 2011 BS01 010 01 projet Calibration).
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Appendix: A Non-asymptotic Control of the Second Kind Error Rates
Appendix: A Non-asymptotic Control of the Second Kind Error Rates
Consider the notation from Sect. 17.3. Since this section focuses on the study of the second kind error rate of the test, in all the sequel, the observation is assumed to satisfy the alternative \(({\mathcal {H}}_1)\). Let thus P be an alternative, that is P ≠ (P 1 ⊗ P 2), n ≥ 4 and \({\mathbb {X}}_n=(X_i,\dots ,X_n)\) be an i.i.d. sample from distribution P. Fix α and β be two fixed values in (0, 1). Consider T the test statistic introduced in (17.16), the (random) critical value \(q_{1-\alpha }({\mathbb {X}}_n)\) defined in (17.18), and the corresponding permutation test defined in (17.19) by
which precisely rejects independence when \(T({\mathbb {X}}_n) > q_{1-\alpha }({\mathbb {X}}_n)\). Notice that this test is exactly the upper-tailed test by permutation introduced in [2].
The aim of this section is to provide different conditions on the alternative P ensuring a control of the second kind error rate by a fixed value β > 0, that is \({\mathbb {P}\left (\Delta _\alpha ({\mathbb {X}}_n)=0\right )} \leq \beta .\) The following steps constitute the first steps of a general study of the separation rates for the previous independence test, and is worked through in the specific case of continuous real-valued random variables in [1, Chapter 4].
Recall the notation introduced in (17.17) for a better readability. For all real-valued measurable function g on \({\mathcal {X}}^2\), denote respectively
the expectations of g(X) under the alternative P, that is if X ∼ P, and under the null hypothesis \(({\mathcal {H}}_0)\), that is if X ∼ (P 1 ⊗ P 2).
Assume the following moment assumption holds, that is
so that all variance and second-order moments exist. Then, the following statements hold.
-
1.
By Chebychev’s inequality, one has \({\mathbb {P}\left (\Delta _\alpha ({\mathbb {X}}_n)=0\right )} \leq \beta \) as soon as Condition (17.20) is satisfied, that is
$$\displaystyle \begin{aligned}{\mathbb{E}\left[ T({\mathbb{X}}_n) \right]} \geq q^\alpha_{1-\beta/2} + \sqrt{\frac{2}{\beta} {\operatorname{Var}\left( T({\mathbb{X}}_n) \right)}}.\end{aligned}$$ -
2.
On the one hand,
$$\displaystyle \begin{aligned} {\operatorname{Var}\left( T({\mathbb{X}}_n) \right)} \leq \frac{8}{n} {\left( {\mathbb{E}_{P}\left[ \varphi^2 \right]} + {\mathbb{E}_{{\perp\perp}}\left[ \varphi^2 \right]} \right)}, \end{aligned} $$(17.48) -
3.
On the other hand, in order to control the quantile \(q^\alpha _{1-\beta /2}\), let us first upper bound the conditional quantile, following Hoeffding’s approach based on the Cauchy-Schwarz inequality, by
$$\displaystyle \begin{aligned} q_{1-\alpha}({\mathbb{X}}_n) \leq \sqrt{\frac{1-\alpha}{\alpha} {\operatorname{Var}\left( T{\left( {\mathbb{X}}_n^{{{\Pi}}} \right)}\middle| {\mathbb{X}}_n \right)}}. \end{aligned} $$(17.49) -
4.
Markov’s inequality allows us to deduce the following bound for the quantile:
$$\displaystyle \begin{aligned} q^\alpha_{1-\beta/2} \leq 2\sqrt{\frac{1-\alpha}{\alpha}}\sqrt{\frac{2}{\beta}\frac{{\left( {\mathbb{E}_{{\perp\perp}}\left[ \varphi^2 \right]} + {\mathbb{E}_{P}\left[ \varphi^2 \right]} \right)}}{n}}. \end{aligned} $$(17.50) -
5.
Finally, combining (17.20), (17.48) and (17.50) ensures that \({\mathbb {P}\left (\Delta _\alpha ({\mathbb {X}}_n)=0\right )} \leq \beta \) as soon as Condition (17.21) is satisfied, that is
$$\displaystyle \begin{aligned}{\mathbb{E}\left[ T({\mathbb{X}}_n) \right]} \geq \frac{4}{\sqrt{\alpha}}\sqrt{\frac{2}{\beta}\frac{{\mathbb{E}_{P}\left[ \varphi^2 \right]} + {\mathbb{E}_{{\perp\perp}}\left[ \varphi^2 \right]}}{n}}.\end{aligned}$$
This section is divided in five subsections, each one of them respectively proving a point stated above. The first one proves the sufficiency of Condition (17.20) in order to control the second kind error rate. The second, third and fourth ones provide respectively upper-bounds of the variance term, the critical value and the quantile \(q^\alpha _{1-\beta /2}\). Finally, the fifth one provides the sufficiency of Condition (17.21).
17.1.1 A First Condition Ensuing from Chebychev’s Inequality
In this section, we prove the sufficiency of a first simple condition, derived from Chebychev’s inequality in order to control the second error rate. Assume that (17.20) is satisfied, that is
Then,
by definition of the quantile \(q^\alpha _{1-\beta /2}\). Yet, from (17.20) one obtains from Chebychev’s inequality that
Finally, both (17.52) and (17.53) lead to the desired control \({\mathbb {P}\left (\Delta _\alpha ({\mathbb {X}}_n)=0\right )}\leq \beta \) which ends the proof.
17.1.2 Control of the Variance in the General Case
To upper bound the variance term, we apply Lemma 17.3.1 which directly implies that
which directly leads to (17.48).
17.1.3 Control of the Critical Value Based on Hoeffding’s Approach
This section is devoted to the proof the inequality (17.49), namely
The proof of this upper-bound follows Hoeffding’s approach in [16], and relies on a normalizing trick, and the Cauchy-Schwarz inequality. From now on, for a better readability, denote respectively \({\mathbb {E}^*\left [ \cdot \right ]}\) and \({\operatorname {Var}^*\left ( \cdot \right )}\) the conditional expectation and variance given the sample \({\mathbb {X}}_n\).
As in Hoeffding [16], the first step is to center and normalize the permuted test statistic. Yet, by construction the permuted test statistic is automatically centered, that is \({\mathbb {E}^*\left [ T{\left ( {\mathbb {X}}_n^{{{\Pi }}} \right )} \right ]} = 0\), as one can notice that
Therefore, just consider the normalizing term
Two cases appear: either \(\nu (\mathbb {X}_n)=0\) or not.
In the first case, the nullity of the conditional variance implies that all the permutations of the test statistic are equal. Hence, for all permutation τ of \({\left \{ 1,\dots ,n \right \}}\), one has \(T({\mathbb {X}}_n^{{{\tau }}}) = T({\mathbb {X}}_n)\). Since the centering term \({\mathbb {E}^*\left [ \sum _{i=1}^n \varphi {\left ( X_i^1,X_{{{\Pi }}(i)}^2 \right )} \right ]}=n^{-1}\sum _{i,j=1}^n\varphi (X_i^1,X_j^2)\) is permutation invariant, one obtains the equality of the permuted sums, that is
and this for all permutation τ. In particular, the centering term is also equal to \(\sum _{i=1}^n \varphi {\left ( X_i^1,X_i^2 \right )}\). Indeed, by invariance of the sum (applied in the third equality below),
Therefore, \(T({\mathbb {X}}_n)\) is equal to zero, and thus, so is \(q_{1-\alpha }({\mathbb {X}}_n)\). Finally, inequality (17.55) is satisfied since
Consider now the second case, and assume \(\nu (\mathbb {X}_n)>0\). Let us introduce the (centered and) normalized statistic
In particular, the new statistic \(T'{\left ( \mathbb {X}_n \right )}\) satisfies
One may moreover notice that the normalizing term \(\nu ({\mathbb {X}}_n)\) is permutation invariant, that is, for all permutations τ and τ′ in \({\mathfrak {S}_{n}}\),
In particular, since \(\nu (\mathbb {X}_n) >0\),
Therefore, as the test Δα depends only on the comparison of the \({\left \{ T{\left ( {\mathbb {X}}_n^{{{\tau }}} \right )} \right \}}_{{{\tau }}\in {\mathfrak {S}_{n}}}\), the test statistic T can be replaced by T′, and the new critical value becomes
Moreover, following the proof of Theorem 2.1. of Hoeffding [16], one can show (as below) that
Hence, combining (17.55) with (17.54) leads straightforwardly to (17.49).
Finally, remains the proof of (17.55). There are two cases:
First Case
If \(q^{\prime }_{1-\alpha }({\mathbb {X}}_n)\leq 0\), then (17.55) is satisfied.
Second Case
If \(q^{\prime }_{1-\alpha }({\mathbb {X}}_n) > 0\), then introduce \(Y=q^{\prime }_{1-\alpha }({\mathbb {X}}_n) - T'{\left ( {\mathbb {X}}_n^{{{\Pi }}} \right )}.\)
First, since by construction, \({\mathbb {E}^*\left [ T'{\left ( {\mathbb {X}}_n^{{{\Pi }}} \right )} \right ]}=0\), one directly obtains \({\mathbb {E}^*\left [ Y \right ]} = q^{\prime }_{1-\alpha }({\mathbb {X}}_n)\). Hence,
and by the Cauchy-Schwarz inequality,
Yet, on one hand,
since by the normalizing initial step, \({\operatorname {Var}^*\left ( T'{\left ( {\mathbb {X}}_n^{{{\Pi }}} \right )} \right )}\leq 1\).
And, on the other hand,
So finally,
which is equivalent to \({\left ( q^{\prime }_{1-\alpha }({\mathbb {X}}_n) \right )}^2 \leq (1-\alpha )/\alpha ,\) and thus ends the proof of (17.55).
17.1.4 Control of the Quantile of the Critical Value
The control of the conditional quantile allows us to upper bound its own quantile \(q^\alpha _{1-\beta /2}\) as stated in (17.50), that is
Indeed, (17.49) ensures that
and in particular, the (1 − β∕2)-quantile of \(q_{1-\alpha }({\mathbb {X}}_n)\) satisfies
where ζ 1−β∕2 is the (1 − β∕2)-quantile of \({\mathbb {E}\left [ T{\left ( {\mathbb {X}}_n^{{{\Pi }}} \right )}^2\middle | {\mathbb {X}}_n \right ]}\). Yet, from Markov’s inequality, for all positive x,
In particular, the choice of \(x=2{\mathbb {E}\left [ T{\left ( {\mathbb {X}}_n^{{{\Pi }}} \right )}^2 \right ]}/\beta \) leads to the control of the quantile
Moreover, noticing that one can write
the second-order moment in (17.57) can be rewritten
by independence between Π and \({\mathbb {X}}_n\), where
On the one hand, for all 1 ≤ i, j, k, l ≤ n, the Cauchy-Schwarz inequality always ensures
since for all 1 ≤ i, j ≤ n, \({\mathbb {E}\left [ \varphi ^2(X_i^1,X_j^2) \right ]}\leq {\mathbb {E}_{{\perp \perp }}\left [ \varphi ^2 \right ]} + {\mathbb {E}_{P}\left [ \varphi ^2 \right ]}\).
On the other hand, remains to control the sum \((n-1)^{-2} \sum _{i,j=1}^n \sum _{k,l=1}^n E_{i,j,k,l}\). Three cases appear.
First Case
If i ≠ k and j ≠ l (occurring [n(n − 1)]2 times), then
Second Case
If [i ≠ k and j = l] or [i = k and j ≠ l], then
Third Case
If i = k and j = l (occurring n(n − 1) times), then
Therefore,
Finally, both (17.58) and (17.59) imply that
Therefore, combining (17.56), (17.57) and (17.60) ends the proof of (17.50).
17.1.5 A First Condition Ensuing from Hoeffding’s Approach
Back to the condition (17.20) derived from Chebychev’s inequality, both (17.48) and (17.50) imply that
with \(2\sqrt {(1-\alpha )/\alpha }+ \sqrt {8}\leq 4/\sqrt {\alpha }\), since \(\sqrt {1-\alpha }+\sqrt {\alpha }\leq \sqrt {2}\). Finally, the right-hand side of condition (17.20) being upper bounded by
which is exactly the right-hand side of (17.21), this ensures the sufficiency of condition 17.21 to control the second kind error rate by β.
Rights and permissions
Copyright information
© 2019 Springer Nature Switzerland AG
About this paper
Cite this paper
Albert, M. (2019). Concentration Inequalities for Randomly Permuted Sums. In: Gozlan, N., Latała, R., Lounici, K., Madiman, M. (eds) High Dimensional Probability VIII. Progress in Probability, vol 74. Birkhäuser, Cham. https://doi.org/10.1007/978-3-030-26391-1_17
Download citation
DOI: https://doi.org/10.1007/978-3-030-26391-1_17
Published:
Publisher Name: Birkhäuser, Cham
Print ISBN: 978-3-030-26390-4
Online ISBN: 978-3-030-26391-1
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)